Linux Audio

Check our new training course

Loading...
v4.10.11
   1/*
   2 * Generic ring buffer
   3 *
   4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
   5 */
   6#include <linux/trace_events.h>
   7#include <linux/ring_buffer.h>
   8#include <linux/trace_clock.h>
   9#include <linux/trace_seq.h>
  10#include <linux/spinlock.h>
  11#include <linux/irq_work.h>
 
  12#include <linux/uaccess.h>
  13#include <linux/hardirq.h>
  14#include <linux/kthread.h>	/* for self test */
  15#include <linux/kmemcheck.h>
  16#include <linux/module.h>
  17#include <linux/percpu.h>
  18#include <linux/mutex.h>
  19#include <linux/delay.h>
  20#include <linux/slab.h>
  21#include <linux/init.h>
  22#include <linux/hash.h>
  23#include <linux/list.h>
  24#include <linux/cpu.h>
 
  25
  26#include <asm/local.h>
  27
  28static void update_pages_handler(struct work_struct *work);
  29
  30/*
  31 * The ring buffer header is special. We must manually up keep it.
  32 */
  33int ring_buffer_print_entry_header(struct trace_seq *s)
  34{
  35	trace_seq_puts(s, "# compressed entry header\n");
  36	trace_seq_puts(s, "\ttype_len    :    5 bits\n");
  37	trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
  38	trace_seq_puts(s, "\tarray       :   32 bits\n");
  39	trace_seq_putc(s, '\n');
  40	trace_seq_printf(s, "\tpadding     : type == %d\n",
  41			 RINGBUF_TYPE_PADDING);
  42	trace_seq_printf(s, "\ttime_extend : type == %d\n",
  43			 RINGBUF_TYPE_TIME_EXTEND);
  44	trace_seq_printf(s, "\tdata max type_len  == %d\n",
  45			 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 
 
  46
  47	return !trace_seq_has_overflowed(s);
  48}
  49
  50/*
  51 * The ring buffer is made up of a list of pages. A separate list of pages is
  52 * allocated for each CPU. A writer may only write to a buffer that is
  53 * associated with the CPU it is currently executing on.  A reader may read
  54 * from any per cpu buffer.
  55 *
  56 * The reader is special. For each per cpu buffer, the reader has its own
  57 * reader page. When a reader has read the entire reader page, this reader
  58 * page is swapped with another page in the ring buffer.
  59 *
  60 * Now, as long as the writer is off the reader page, the reader can do what
  61 * ever it wants with that page. The writer will never write to that page
  62 * again (as long as it is out of the ring buffer).
  63 *
  64 * Here's some silly ASCII art.
  65 *
  66 *   +------+
  67 *   |reader|          RING BUFFER
  68 *   |page  |
  69 *   +------+        +---+   +---+   +---+
  70 *                   |   |-->|   |-->|   |
  71 *                   +---+   +---+   +---+
  72 *                     ^               |
  73 *                     |               |
  74 *                     +---------------+
  75 *
  76 *
  77 *   +------+
  78 *   |reader|          RING BUFFER
  79 *   |page  |------------------v
  80 *   +------+        +---+   +---+   +---+
  81 *                   |   |-->|   |-->|   |
  82 *                   +---+   +---+   +---+
  83 *                     ^               |
  84 *                     |               |
  85 *                     +---------------+
  86 *
  87 *
  88 *   +------+
  89 *   |reader|          RING BUFFER
  90 *   |page  |------------------v
  91 *   +------+        +---+   +---+   +---+
  92 *      ^            |   |-->|   |-->|   |
  93 *      |            +---+   +---+   +---+
  94 *      |                              |
  95 *      |                              |
  96 *      +------------------------------+
  97 *
  98 *
  99 *   +------+
 100 *   |buffer|          RING BUFFER
 101 *   |page  |------------------v
 102 *   +------+        +---+   +---+   +---+
 103 *      ^            |   |   |   |-->|   |
 104 *      |   New      +---+   +---+   +---+
 105 *      |  Reader------^               |
 106 *      |   page                       |
 107 *      +------------------------------+
 108 *
 109 *
 110 * After we make this swap, the reader can hand this page off to the splice
 111 * code and be done with it. It can even allocate a new page if it needs to
 112 * and swap that into the ring buffer.
 113 *
 114 * We will be using cmpxchg soon to make all this lockless.
 115 *
 116 */
 117
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 118/* Used for individual buffers (after the counter) */
 119#define RB_BUFFER_OFF		(1 << 20)
 120
 121#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
 122
 
 
 
 
 
 
 
 
 
 
 
 123#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
 124#define RB_ALIGNMENT		4U
 125#define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 126#define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
 127
 128#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
 129# define RB_FORCE_8BYTE_ALIGNMENT	0
 130# define RB_ARCH_ALIGNMENT		RB_ALIGNMENT
 131#else
 132# define RB_FORCE_8BYTE_ALIGNMENT	1
 133# define RB_ARCH_ALIGNMENT		8U
 134#endif
 135
 136#define RB_ALIGN_DATA		__aligned(RB_ARCH_ALIGNMENT)
 137
 138/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
 139#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
 140
 141enum {
 142	RB_LEN_TIME_EXTEND = 8,
 143	RB_LEN_TIME_STAMP = 16,
 144};
 145
 146#define skip_time_extend(event) \
 147	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
 148
 149static inline int rb_null_event(struct ring_buffer_event *event)
 150{
 151	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
 152}
 153
 154static void rb_event_set_padding(struct ring_buffer_event *event)
 155{
 156	/* padding has a NULL time_delta */
 157	event->type_len = RINGBUF_TYPE_PADDING;
 158	event->time_delta = 0;
 159}
 160
 161static unsigned
 162rb_event_data_length(struct ring_buffer_event *event)
 163{
 164	unsigned length;
 165
 166	if (event->type_len)
 167		length = event->type_len * RB_ALIGNMENT;
 168	else
 169		length = event->array[0];
 170	return length + RB_EVNT_HDR_SIZE;
 171}
 172
 173/*
 174 * Return the length of the given event. Will return
 175 * the length of the time extend if the event is a
 176 * time extend.
 177 */
 178static inline unsigned
 179rb_event_length(struct ring_buffer_event *event)
 180{
 181	switch (event->type_len) {
 182	case RINGBUF_TYPE_PADDING:
 183		if (rb_null_event(event))
 184			/* undefined */
 185			return -1;
 186		return  event->array[0] + RB_EVNT_HDR_SIZE;
 187
 188	case RINGBUF_TYPE_TIME_EXTEND:
 189		return RB_LEN_TIME_EXTEND;
 190
 191	case RINGBUF_TYPE_TIME_STAMP:
 192		return RB_LEN_TIME_STAMP;
 193
 194	case RINGBUF_TYPE_DATA:
 195		return rb_event_data_length(event);
 196	default:
 197		BUG();
 198	}
 199	/* not hit */
 200	return 0;
 201}
 202
 203/*
 204 * Return total length of time extend and data,
 205 *   or just the event length for all other events.
 206 */
 207static inline unsigned
 208rb_event_ts_length(struct ring_buffer_event *event)
 209{
 210	unsigned len = 0;
 211
 212	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
 213		/* time extends include the data event after it */
 214		len = RB_LEN_TIME_EXTEND;
 215		event = skip_time_extend(event);
 216	}
 217	return len + rb_event_length(event);
 218}
 219
 220/**
 221 * ring_buffer_event_length - return the length of the event
 222 * @event: the event to get the length of
 223 *
 224 * Returns the size of the data load of a data event.
 225 * If the event is something other than a data event, it
 226 * returns the size of the event itself. With the exception
 227 * of a TIME EXTEND, where it still returns the size of the
 228 * data load of the data event after it.
 229 */
 230unsigned ring_buffer_event_length(struct ring_buffer_event *event)
 231{
 232	unsigned length;
 233
 234	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
 235		event = skip_time_extend(event);
 236
 237	length = rb_event_length(event);
 238	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 239		return length;
 240	length -= RB_EVNT_HDR_SIZE;
 241	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
 242                length -= sizeof(event->array[0]);
 243	return length;
 244}
 245EXPORT_SYMBOL_GPL(ring_buffer_event_length);
 246
 247/* inline for ring buffer fast paths */
 248static __always_inline void *
 249rb_event_data(struct ring_buffer_event *event)
 250{
 251	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
 252		event = skip_time_extend(event);
 253	BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 254	/* If length is in len field, then array[0] has the data */
 255	if (event->type_len)
 256		return (void *)&event->array[0];
 257	/* Otherwise length is in array[0] and array[1] has the data */
 258	return (void *)&event->array[1];
 259}
 260
 261/**
 262 * ring_buffer_event_data - return the data of the event
 263 * @event: the event to get the data from
 264 */
 265void *ring_buffer_event_data(struct ring_buffer_event *event)
 266{
 267	return rb_event_data(event);
 268}
 269EXPORT_SYMBOL_GPL(ring_buffer_event_data);
 270
 271#define for_each_buffer_cpu(buffer, cpu)		\
 272	for_each_cpu(cpu, buffer->cpumask)
 273
 274#define TS_SHIFT	27
 275#define TS_MASK		((1ULL << TS_SHIFT) - 1)
 276#define TS_DELTA_TEST	(~TS_MASK)
 277
 278/* Flag when events were overwritten */
 279#define RB_MISSED_EVENTS	(1 << 31)
 280/* Missed count stored at end */
 281#define RB_MISSED_STORED	(1 << 30)
 282
 283struct buffer_data_page {
 284	u64		 time_stamp;	/* page time stamp */
 285	local_t		 commit;	/* write committed index */
 286	unsigned char	 data[] RB_ALIGN_DATA;	/* data of buffer page */
 287};
 288
 289/*
 290 * Note, the buffer_page list must be first. The buffer pages
 291 * are allocated in cache lines, which means that each buffer
 292 * page will be at the beginning of a cache line, and thus
 293 * the least significant bits will be zero. We use this to
 294 * add flags in the list struct pointers, to make the ring buffer
 295 * lockless.
 296 */
 297struct buffer_page {
 298	struct list_head list;		/* list of buffer pages */
 299	local_t		 write;		/* index for next write */
 300	unsigned	 read;		/* index for next read */
 301	local_t		 entries;	/* entries on this page */
 302	unsigned long	 real_end;	/* real end of data */
 303	struct buffer_data_page *page;	/* Actual data page */
 304};
 305
 306/*
 307 * The buffer page counters, write and entries, must be reset
 308 * atomically when crossing page boundaries. To synchronize this
 309 * update, two counters are inserted into the number. One is
 310 * the actual counter for the write position or count on the page.
 311 *
 312 * The other is a counter of updaters. Before an update happens
 313 * the update partition of the counter is incremented. This will
 314 * allow the updater to update the counter atomically.
 315 *
 316 * The counter is 20 bits, and the state data is 12.
 317 */
 318#define RB_WRITE_MASK		0xfffff
 319#define RB_WRITE_INTCNT		(1 << 20)
 320
 321static void rb_init_page(struct buffer_data_page *bpage)
 322{
 323	local_set(&bpage->commit, 0);
 324}
 325
 326/**
 327 * ring_buffer_page_len - the size of data on the page.
 328 * @page: The page to read
 329 *
 330 * Returns the amount of data on the page, including buffer page header.
 331 */
 332size_t ring_buffer_page_len(void *page)
 333{
 334	return local_read(&((struct buffer_data_page *)page)->commit)
 335		+ BUF_PAGE_HDR_SIZE;
 336}
 337
 338/*
 339 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
 340 * this issue out.
 341 */
 342static void free_buffer_page(struct buffer_page *bpage)
 343{
 344	free_page((unsigned long)bpage->page);
 345	kfree(bpage);
 346}
 347
 348/*
 349 * We need to fit the time_stamp delta into 27 bits.
 350 */
 351static inline int test_time_stamp(u64 delta)
 352{
 353	if (delta & TS_DELTA_TEST)
 354		return 1;
 355	return 0;
 356}
 357
 358#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
 359
 360/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
 361#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
 362
 363int ring_buffer_print_page_header(struct trace_seq *s)
 364{
 365	struct buffer_data_page field;
 
 366
 367	trace_seq_printf(s, "\tfield: u64 timestamp;\t"
 368			 "offset:0;\tsize:%u;\tsigned:%u;\n",
 369			 (unsigned int)sizeof(field.time_stamp),
 370			 (unsigned int)is_signed_type(u64));
 371
 372	trace_seq_printf(s, "\tfield: local_t commit;\t"
 373			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 374			 (unsigned int)offsetof(typeof(field), commit),
 375			 (unsigned int)sizeof(field.commit),
 376			 (unsigned int)is_signed_type(long));
 377
 378	trace_seq_printf(s, "\tfield: int overwrite;\t"
 379			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 380			 (unsigned int)offsetof(typeof(field), commit),
 381			 1,
 382			 (unsigned int)is_signed_type(long));
 383
 384	trace_seq_printf(s, "\tfield: char data;\t"
 385			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 386			 (unsigned int)offsetof(typeof(field), data),
 387			 (unsigned int)BUF_PAGE_SIZE,
 388			 (unsigned int)is_signed_type(char));
 389
 390	return !trace_seq_has_overflowed(s);
 391}
 392
 393struct rb_irq_work {
 394	struct irq_work			work;
 395	wait_queue_head_t		waiters;
 396	wait_queue_head_t		full_waiters;
 397	bool				waiters_pending;
 398	bool				full_waiters_pending;
 399	bool				wakeup_full;
 400};
 401
 402/*
 403 * Structure to hold event state and handle nested events.
 404 */
 405struct rb_event_info {
 406	u64			ts;
 407	u64			delta;
 408	unsigned long		length;
 409	struct buffer_page	*tail_page;
 410	int			add_timestamp;
 411};
 412
 413/*
 414 * Used for which event context the event is in.
 415 *  NMI     = 0
 416 *  IRQ     = 1
 417 *  SOFTIRQ = 2
 418 *  NORMAL  = 3
 419 *
 420 * See trace_recursive_lock() comment below for more details.
 421 */
 422enum {
 423	RB_CTX_NMI,
 424	RB_CTX_IRQ,
 425	RB_CTX_SOFTIRQ,
 426	RB_CTX_NORMAL,
 427	RB_CTX_MAX
 428};
 429
 430/*
 431 * head_page == tail_page && head == tail then buffer is empty.
 432 */
 433struct ring_buffer_per_cpu {
 434	int				cpu;
 435	atomic_t			record_disabled;
 436	struct ring_buffer		*buffer;
 437	raw_spinlock_t			reader_lock;	/* serialize readers */
 438	arch_spinlock_t			lock;
 439	struct lock_class_key		lock_key;
 440	unsigned long			nr_pages;
 441	unsigned int			current_context;
 442	struct list_head		*pages;
 443	struct buffer_page		*head_page;	/* read from head */
 444	struct buffer_page		*tail_page;	/* write to tail */
 445	struct buffer_page		*commit_page;	/* committed pages */
 446	struct buffer_page		*reader_page;
 447	unsigned long			lost_events;
 448	unsigned long			last_overrun;
 449	local_t				entries_bytes;
 450	local_t				entries;
 451	local_t				overrun;
 452	local_t				commit_overrun;
 453	local_t				dropped_events;
 454	local_t				committing;
 455	local_t				commits;
 456	unsigned long			read;
 457	unsigned long			read_bytes;
 458	u64				write_stamp;
 459	u64				read_stamp;
 460	/* ring buffer pages to update, > 0 to add, < 0 to remove */
 461	long				nr_pages_to_update;
 462	struct list_head		new_pages; /* new pages to add */
 463	struct work_struct		update_pages_work;
 464	struct completion		update_done;
 465
 466	struct rb_irq_work		irq_work;
 467};
 468
 469struct ring_buffer {
 470	unsigned			flags;
 471	int				cpus;
 472	atomic_t			record_disabled;
 473	atomic_t			resize_disabled;
 474	cpumask_var_t			cpumask;
 475
 476	struct lock_class_key		*reader_lock_key;
 477
 478	struct mutex			mutex;
 479
 480	struct ring_buffer_per_cpu	**buffers;
 481
 482	struct hlist_node		node;
 
 
 483	u64				(*clock)(void);
 484
 485	struct rb_irq_work		irq_work;
 486};
 487
 488struct ring_buffer_iter {
 489	struct ring_buffer_per_cpu	*cpu_buffer;
 490	unsigned long			head;
 491	struct buffer_page		*head_page;
 492	struct buffer_page		*cache_reader_page;
 493	unsigned long			cache_read;
 494	u64				read_stamp;
 495};
 496
 497/*
 498 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
 499 *
 500 * Schedules a delayed work to wake up any task that is blocked on the
 501 * ring buffer waiters queue.
 502 */
 503static void rb_wake_up_waiters(struct irq_work *work)
 504{
 505	struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
 506
 507	wake_up_all(&rbwork->waiters);
 508	if (rbwork->wakeup_full) {
 509		rbwork->wakeup_full = false;
 510		wake_up_all(&rbwork->full_waiters);
 511	}
 512}
 513
 514/**
 515 * ring_buffer_wait - wait for input to the ring buffer
 516 * @buffer: buffer to wait on
 517 * @cpu: the cpu buffer to wait on
 518 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
 519 *
 520 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 521 * as data is added to any of the @buffer's cpu buffers. Otherwise
 522 * it will wait for data to be added to a specific cpu buffer.
 523 */
 524int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
 525{
 526	struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
 527	DEFINE_WAIT(wait);
 528	struct rb_irq_work *work;
 529	int ret = 0;
 530
 531	/*
 532	 * Depending on what the caller is waiting for, either any
 533	 * data in any cpu buffer, or a specific buffer, put the
 534	 * caller on the appropriate wait queue.
 535	 */
 536	if (cpu == RING_BUFFER_ALL_CPUS) {
 537		work = &buffer->irq_work;
 538		/* Full only makes sense on per cpu reads */
 539		full = false;
 540	} else {
 541		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 542			return -ENODEV;
 543		cpu_buffer = buffer->buffers[cpu];
 544		work = &cpu_buffer->irq_work;
 545	}
 546
 547
 548	while (true) {
 549		if (full)
 550			prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
 551		else
 552			prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
 553
 554		/*
 555		 * The events can happen in critical sections where
 556		 * checking a work queue can cause deadlocks.
 557		 * After adding a task to the queue, this flag is set
 558		 * only to notify events to try to wake up the queue
 559		 * using irq_work.
 560		 *
 561		 * We don't clear it even if the buffer is no longer
 562		 * empty. The flag only causes the next event to run
 563		 * irq_work to do the work queue wake up. The worse
 564		 * that can happen if we race with !trace_empty() is that
 565		 * an event will cause an irq_work to try to wake up
 566		 * an empty queue.
 567		 *
 568		 * There's no reason to protect this flag either, as
 569		 * the work queue and irq_work logic will do the necessary
 570		 * synchronization for the wake ups. The only thing
 571		 * that is necessary is that the wake up happens after
 572		 * a task has been queued. It's OK for spurious wake ups.
 573		 */
 574		if (full)
 575			work->full_waiters_pending = true;
 576		else
 577			work->waiters_pending = true;
 578
 579		if (signal_pending(current)) {
 580			ret = -EINTR;
 581			break;
 582		}
 583
 584		if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
 585			break;
 586
 587		if (cpu != RING_BUFFER_ALL_CPUS &&
 588		    !ring_buffer_empty_cpu(buffer, cpu)) {
 589			unsigned long flags;
 590			bool pagebusy;
 591
 592			if (!full)
 593				break;
 594
 595			raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
 596			pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 597			raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
 598
 599			if (!pagebusy)
 600				break;
 601		}
 602
 
 
 603		schedule();
 604	}
 605
 606	if (full)
 607		finish_wait(&work->full_waiters, &wait);
 608	else
 609		finish_wait(&work->waiters, &wait);
 610
 611	return ret;
 612}
 613
 614/**
 615 * ring_buffer_poll_wait - poll on buffer input
 616 * @buffer: buffer to wait on
 617 * @cpu: the cpu buffer to wait on
 618 * @filp: the file descriptor
 619 * @poll_table: The poll descriptor
 620 *
 621 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 622 * as data is added to any of the @buffer's cpu buffers. Otherwise
 623 * it will wait for data to be added to a specific cpu buffer.
 624 *
 625 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
 626 * zero otherwise.
 627 */
 628int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
 629			  struct file *filp, poll_table *poll_table)
 630{
 631	struct ring_buffer_per_cpu *cpu_buffer;
 632	struct rb_irq_work *work;
 633
 
 
 
 
 634	if (cpu == RING_BUFFER_ALL_CPUS)
 635		work = &buffer->irq_work;
 636	else {
 637		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 638			return -EINVAL;
 639
 640		cpu_buffer = buffer->buffers[cpu];
 641		work = &cpu_buffer->irq_work;
 642	}
 643
 644	poll_wait(filp, &work->waiters, poll_table);
 645	work->waiters_pending = true;
 646	/*
 647	 * There's a tight race between setting the waiters_pending and
 648	 * checking if the ring buffer is empty.  Once the waiters_pending bit
 649	 * is set, the next event will wake the task up, but we can get stuck
 650	 * if there's only a single event in.
 651	 *
 652	 * FIXME: Ideally, we need a memory barrier on the writer side as well,
 653	 * but adding a memory barrier to all events will cause too much of a
 654	 * performance hit in the fast path.  We only need a memory barrier when
 655	 * the buffer goes from empty to having content.  But as this race is
 656	 * extremely small, and it's not a problem if another event comes in, we
 657	 * will fix it later.
 658	 */
 659	smp_mb();
 660
 661	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
 662	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
 663		return POLLIN | POLLRDNORM;
 664	return 0;
 665}
 666
 667/* buffer may be either ring_buffer or ring_buffer_per_cpu */
 668#define RB_WARN_ON(b, cond)						\
 669	({								\
 670		int _____ret = unlikely(cond);				\
 671		if (_____ret) {						\
 672			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
 673				struct ring_buffer_per_cpu *__b =	\
 674					(void *)b;			\
 675				atomic_inc(&__b->buffer->record_disabled); \
 676			} else						\
 677				atomic_inc(&b->record_disabled);	\
 678			WARN_ON(1);					\
 679		}							\
 680		_____ret;						\
 681	})
 682
 683/* Up this if you want to test the TIME_EXTENTS and normalization */
 684#define DEBUG_SHIFT 0
 685
 686static inline u64 rb_time_stamp(struct ring_buffer *buffer)
 687{
 688	/* shift to debug/test normalization and TIME_EXTENTS */
 689	return buffer->clock() << DEBUG_SHIFT;
 690}
 691
 692u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
 693{
 694	u64 time;
 695
 696	preempt_disable_notrace();
 697	time = rb_time_stamp(buffer);
 698	preempt_enable_no_resched_notrace();
 699
 700	return time;
 701}
 702EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
 703
 704void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
 705				      int cpu, u64 *ts)
 706{
 707	/* Just stupid testing the normalize function and deltas */
 708	*ts >>= DEBUG_SHIFT;
 709}
 710EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
 711
 712/*
 713 * Making the ring buffer lockless makes things tricky.
 714 * Although writes only happen on the CPU that they are on,
 715 * and they only need to worry about interrupts. Reads can
 716 * happen on any CPU.
 717 *
 718 * The reader page is always off the ring buffer, but when the
 719 * reader finishes with a page, it needs to swap its page with
 720 * a new one from the buffer. The reader needs to take from
 721 * the head (writes go to the tail). But if a writer is in overwrite
 722 * mode and wraps, it must push the head page forward.
 723 *
 724 * Here lies the problem.
 725 *
 726 * The reader must be careful to replace only the head page, and
 727 * not another one. As described at the top of the file in the
 728 * ASCII art, the reader sets its old page to point to the next
 729 * page after head. It then sets the page after head to point to
 730 * the old reader page. But if the writer moves the head page
 731 * during this operation, the reader could end up with the tail.
 732 *
 733 * We use cmpxchg to help prevent this race. We also do something
 734 * special with the page before head. We set the LSB to 1.
 735 *
 736 * When the writer must push the page forward, it will clear the
 737 * bit that points to the head page, move the head, and then set
 738 * the bit that points to the new head page.
 739 *
 740 * We also don't want an interrupt coming in and moving the head
 741 * page on another writer. Thus we use the second LSB to catch
 742 * that too. Thus:
 743 *
 744 * head->list->prev->next        bit 1          bit 0
 745 *                              -------        -------
 746 * Normal page                     0              0
 747 * Points to head page             0              1
 748 * New head page                   1              0
 749 *
 750 * Note we can not trust the prev pointer of the head page, because:
 751 *
 752 * +----+       +-----+        +-----+
 753 * |    |------>|  T  |---X--->|  N  |
 754 * |    |<------|     |        |     |
 755 * +----+       +-----+        +-----+
 756 *   ^                           ^ |
 757 *   |          +-----+          | |
 758 *   +----------|  R  |----------+ |
 759 *              |     |<-----------+
 760 *              +-----+
 761 *
 762 * Key:  ---X-->  HEAD flag set in pointer
 763 *         T      Tail page
 764 *         R      Reader page
 765 *         N      Next page
 766 *
 767 * (see __rb_reserve_next() to see where this happens)
 768 *
 769 *  What the above shows is that the reader just swapped out
 770 *  the reader page with a page in the buffer, but before it
 771 *  could make the new header point back to the new page added
 772 *  it was preempted by a writer. The writer moved forward onto
 773 *  the new page added by the reader and is about to move forward
 774 *  again.
 775 *
 776 *  You can see, it is legitimate for the previous pointer of
 777 *  the head (or any page) not to point back to itself. But only
 778 *  temporarially.
 779 */
 780
 781#define RB_PAGE_NORMAL		0UL
 782#define RB_PAGE_HEAD		1UL
 783#define RB_PAGE_UPDATE		2UL
 784
 785
 786#define RB_FLAG_MASK		3UL
 787
 788/* PAGE_MOVED is not part of the mask */
 789#define RB_PAGE_MOVED		4UL
 790
 791/*
 792 * rb_list_head - remove any bit
 793 */
 794static struct list_head *rb_list_head(struct list_head *list)
 795{
 796	unsigned long val = (unsigned long)list;
 797
 798	return (struct list_head *)(val & ~RB_FLAG_MASK);
 799}
 800
 801/*
 802 * rb_is_head_page - test if the given page is the head page
 803 *
 804 * Because the reader may move the head_page pointer, we can
 805 * not trust what the head page is (it may be pointing to
 806 * the reader page). But if the next page is a header page,
 807 * its flags will be non zero.
 808 */
 809static inline int
 810rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
 811		struct buffer_page *page, struct list_head *list)
 812{
 813	unsigned long val;
 814
 815	val = (unsigned long)list->next;
 816
 817	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
 818		return RB_PAGE_MOVED;
 819
 820	return val & RB_FLAG_MASK;
 821}
 822
 823/*
 824 * rb_is_reader_page
 825 *
 826 * The unique thing about the reader page, is that, if the
 827 * writer is ever on it, the previous pointer never points
 828 * back to the reader page.
 829 */
 830static bool rb_is_reader_page(struct buffer_page *page)
 831{
 832	struct list_head *list = page->list.prev;
 833
 834	return rb_list_head(list->next) != &page->list;
 835}
 836
 837/*
 838 * rb_set_list_to_head - set a list_head to be pointing to head.
 839 */
 840static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
 841				struct list_head *list)
 842{
 843	unsigned long *ptr;
 844
 845	ptr = (unsigned long *)&list->next;
 846	*ptr |= RB_PAGE_HEAD;
 847	*ptr &= ~RB_PAGE_UPDATE;
 848}
 849
 850/*
 851 * rb_head_page_activate - sets up head page
 852 */
 853static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
 854{
 855	struct buffer_page *head;
 856
 857	head = cpu_buffer->head_page;
 858	if (!head)
 859		return;
 860
 861	/*
 862	 * Set the previous list pointer to have the HEAD flag.
 863	 */
 864	rb_set_list_to_head(cpu_buffer, head->list.prev);
 865}
 866
 867static void rb_list_head_clear(struct list_head *list)
 868{
 869	unsigned long *ptr = (unsigned long *)&list->next;
 870
 871	*ptr &= ~RB_FLAG_MASK;
 872}
 873
 874/*
 875 * rb_head_page_dactivate - clears head page ptr (for free list)
 876 */
 877static void
 878rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
 879{
 880	struct list_head *hd;
 881
 882	/* Go through the whole list and clear any pointers found. */
 883	rb_list_head_clear(cpu_buffer->pages);
 884
 885	list_for_each(hd, cpu_buffer->pages)
 886		rb_list_head_clear(hd);
 887}
 888
 889static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
 890			    struct buffer_page *head,
 891			    struct buffer_page *prev,
 892			    int old_flag, int new_flag)
 893{
 894	struct list_head *list;
 895	unsigned long val = (unsigned long)&head->list;
 896	unsigned long ret;
 897
 898	list = &prev->list;
 899
 900	val &= ~RB_FLAG_MASK;
 901
 902	ret = cmpxchg((unsigned long *)&list->next,
 903		      val | old_flag, val | new_flag);
 904
 905	/* check if the reader took the page */
 906	if ((ret & ~RB_FLAG_MASK) != val)
 907		return RB_PAGE_MOVED;
 908
 909	return ret & RB_FLAG_MASK;
 910}
 911
 912static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
 913				   struct buffer_page *head,
 914				   struct buffer_page *prev,
 915				   int old_flag)
 916{
 917	return rb_head_page_set(cpu_buffer, head, prev,
 918				old_flag, RB_PAGE_UPDATE);
 919}
 920
 921static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
 922				 struct buffer_page *head,
 923				 struct buffer_page *prev,
 924				 int old_flag)
 925{
 926	return rb_head_page_set(cpu_buffer, head, prev,
 927				old_flag, RB_PAGE_HEAD);
 928}
 929
 930static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
 931				   struct buffer_page *head,
 932				   struct buffer_page *prev,
 933				   int old_flag)
 934{
 935	return rb_head_page_set(cpu_buffer, head, prev,
 936				old_flag, RB_PAGE_NORMAL);
 937}
 938
 939static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
 940			       struct buffer_page **bpage)
 941{
 942	struct list_head *p = rb_list_head((*bpage)->list.next);
 943
 944	*bpage = list_entry(p, struct buffer_page, list);
 945}
 946
 947static struct buffer_page *
 948rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
 949{
 950	struct buffer_page *head;
 951	struct buffer_page *page;
 952	struct list_head *list;
 953	int i;
 954
 955	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
 956		return NULL;
 957
 958	/* sanity check */
 959	list = cpu_buffer->pages;
 960	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
 961		return NULL;
 962
 963	page = head = cpu_buffer->head_page;
 964	/*
 965	 * It is possible that the writer moves the header behind
 966	 * where we started, and we miss in one loop.
 967	 * A second loop should grab the header, but we'll do
 968	 * three loops just because I'm paranoid.
 969	 */
 970	for (i = 0; i < 3; i++) {
 971		do {
 972			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
 973				cpu_buffer->head_page = page;
 974				return page;
 975			}
 976			rb_inc_page(cpu_buffer, &page);
 977		} while (page != head);
 978	}
 979
 980	RB_WARN_ON(cpu_buffer, 1);
 981
 982	return NULL;
 983}
 984
 985static int rb_head_page_replace(struct buffer_page *old,
 986				struct buffer_page *new)
 987{
 988	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
 989	unsigned long val;
 990	unsigned long ret;
 991
 992	val = *ptr & ~RB_FLAG_MASK;
 993	val |= RB_PAGE_HEAD;
 994
 995	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
 996
 997	return ret == val;
 998}
 999
1000/*
1001 * rb_tail_page_update - move the tail page forward
 
 
1002 */
1003static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1004			       struct buffer_page *tail_page,
1005			       struct buffer_page *next_page)
1006{
 
1007	unsigned long old_entries;
1008	unsigned long old_write;
 
1009
1010	/*
1011	 * The tail page now needs to be moved forward.
1012	 *
1013	 * We need to reset the tail page, but without messing
1014	 * with possible erasing of data brought in by interrupts
1015	 * that have moved the tail page and are currently on it.
1016	 *
1017	 * We add a counter to the write field to denote this.
1018	 */
1019	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1020	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1021
1022	/*
1023	 * Just make sure we have seen our old_write and synchronize
1024	 * with any interrupts that come in.
1025	 */
1026	barrier();
1027
1028	/*
1029	 * If the tail page is still the same as what we think
1030	 * it is, then it is up to us to update the tail
1031	 * pointer.
1032	 */
1033	if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1034		/* Zero the write counter */
1035		unsigned long val = old_write & ~RB_WRITE_MASK;
1036		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1037
1038		/*
1039		 * This will only succeed if an interrupt did
1040		 * not come in and change it. In which case, we
1041		 * do not want to modify it.
1042		 *
1043		 * We add (void) to let the compiler know that we do not care
1044		 * about the return value of these functions. We use the
1045		 * cmpxchg to only update if an interrupt did not already
1046		 * do it for us. If the cmpxchg fails, we don't care.
1047		 */
1048		(void)local_cmpxchg(&next_page->write, old_write, val);
1049		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1050
1051		/*
1052		 * No need to worry about races with clearing out the commit.
1053		 * it only can increment when a commit takes place. But that
1054		 * only happens in the outer most nested commit.
1055		 */
1056		local_set(&next_page->page->commit, 0);
1057
1058		/* Again, either we update tail_page or an interrupt does */
1059		(void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
 
 
 
1060	}
 
 
1061}
1062
1063static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1064			  struct buffer_page *bpage)
1065{
1066	unsigned long val = (unsigned long)bpage;
1067
1068	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1069		return 1;
1070
1071	return 0;
1072}
1073
1074/**
1075 * rb_check_list - make sure a pointer to a list has the last bits zero
1076 */
1077static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1078			 struct list_head *list)
1079{
1080	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1081		return 1;
1082	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1083		return 1;
1084	return 0;
1085}
1086
1087/**
1088 * rb_check_pages - integrity check of buffer pages
1089 * @cpu_buffer: CPU buffer with pages to test
1090 *
1091 * As a safety measure we check to make sure the data pages have not
1092 * been corrupted.
1093 */
1094static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1095{
1096	struct list_head *head = cpu_buffer->pages;
1097	struct buffer_page *bpage, *tmp;
1098
1099	/* Reset the head page if it exists */
1100	if (cpu_buffer->head_page)
1101		rb_set_head_page(cpu_buffer);
1102
1103	rb_head_page_deactivate(cpu_buffer);
1104
1105	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1106		return -1;
1107	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1108		return -1;
1109
1110	if (rb_check_list(cpu_buffer, head))
1111		return -1;
1112
1113	list_for_each_entry_safe(bpage, tmp, head, list) {
1114		if (RB_WARN_ON(cpu_buffer,
1115			       bpage->list.next->prev != &bpage->list))
1116			return -1;
1117		if (RB_WARN_ON(cpu_buffer,
1118			       bpage->list.prev->next != &bpage->list))
1119			return -1;
1120		if (rb_check_list(cpu_buffer, &bpage->list))
1121			return -1;
1122	}
1123
1124	rb_head_page_activate(cpu_buffer);
1125
1126	return 0;
1127}
1128
1129static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1130{
 
1131	struct buffer_page *bpage, *tmp;
1132	long i;
1133
1134	for (i = 0; i < nr_pages; i++) {
1135		struct page *page;
1136		/*
1137		 * __GFP_NORETRY flag makes sure that the allocation fails
1138		 * gracefully without invoking oom-killer and the system is
1139		 * not destabilized.
1140		 */
1141		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1142				    GFP_KERNEL | __GFP_NORETRY,
1143				    cpu_to_node(cpu));
1144		if (!bpage)
1145			goto free_pages;
1146
1147		list_add(&bpage->list, pages);
1148
1149		page = alloc_pages_node(cpu_to_node(cpu),
1150					GFP_KERNEL | __GFP_NORETRY, 0);
1151		if (!page)
1152			goto free_pages;
1153		bpage->page = page_address(page);
1154		rb_init_page(bpage->page);
1155	}
1156
1157	return 0;
1158
1159free_pages:
1160	list_for_each_entry_safe(bpage, tmp, pages, list) {
1161		list_del_init(&bpage->list);
1162		free_buffer_page(bpage);
1163	}
1164
1165	return -ENOMEM;
1166}
1167
1168static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1169			     unsigned long nr_pages)
1170{
1171	LIST_HEAD(pages);
1172
1173	WARN_ON(!nr_pages);
1174
1175	if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1176		return -ENOMEM;
1177
1178	/*
1179	 * The ring buffer page list is a circular list that does not
1180	 * start and end with a list head. All page list items point to
1181	 * other pages.
1182	 */
1183	cpu_buffer->pages = pages.next;
1184	list_del(&pages);
1185
1186	cpu_buffer->nr_pages = nr_pages;
1187
1188	rb_check_pages(cpu_buffer);
1189
1190	return 0;
1191}
1192
1193static struct ring_buffer_per_cpu *
1194rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1195{
1196	struct ring_buffer_per_cpu *cpu_buffer;
1197	struct buffer_page *bpage;
1198	struct page *page;
1199	int ret;
1200
1201	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1202				  GFP_KERNEL, cpu_to_node(cpu));
1203	if (!cpu_buffer)
1204		return NULL;
1205
1206	cpu_buffer->cpu = cpu;
1207	cpu_buffer->buffer = buffer;
1208	raw_spin_lock_init(&cpu_buffer->reader_lock);
1209	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1210	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1211	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1212	init_completion(&cpu_buffer->update_done);
1213	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1214	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1215	init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1216
1217	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1218			    GFP_KERNEL, cpu_to_node(cpu));
1219	if (!bpage)
1220		goto fail_free_buffer;
1221
1222	rb_check_bpage(cpu_buffer, bpage);
1223
1224	cpu_buffer->reader_page = bpage;
1225	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1226	if (!page)
1227		goto fail_free_reader;
1228	bpage->page = page_address(page);
1229	rb_init_page(bpage->page);
1230
1231	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1232	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1233
1234	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1235	if (ret < 0)
1236		goto fail_free_reader;
1237
1238	cpu_buffer->head_page
1239		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1240	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1241
1242	rb_head_page_activate(cpu_buffer);
1243
1244	return cpu_buffer;
1245
1246 fail_free_reader:
1247	free_buffer_page(cpu_buffer->reader_page);
1248
1249 fail_free_buffer:
1250	kfree(cpu_buffer);
1251	return NULL;
1252}
1253
1254static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1255{
1256	struct list_head *head = cpu_buffer->pages;
1257	struct buffer_page *bpage, *tmp;
1258
1259	free_buffer_page(cpu_buffer->reader_page);
1260
1261	rb_head_page_deactivate(cpu_buffer);
1262
1263	if (head) {
1264		list_for_each_entry_safe(bpage, tmp, head, list) {
1265			list_del_init(&bpage->list);
1266			free_buffer_page(bpage);
1267		}
1268		bpage = list_entry(head, struct buffer_page, list);
1269		free_buffer_page(bpage);
1270	}
1271
1272	kfree(cpu_buffer);
1273}
1274
 
 
 
 
 
1275/**
1276 * __ring_buffer_alloc - allocate a new ring_buffer
1277 * @size: the size in bytes per cpu that is needed.
1278 * @flags: attributes to set for the ring buffer.
1279 *
1280 * Currently the only flag that is available is the RB_FL_OVERWRITE
1281 * flag. This flag means that the buffer will overwrite old data
1282 * when the buffer wraps. If this flag is not set, the buffer will
1283 * drop data when the tail hits the head.
1284 */
1285struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1286					struct lock_class_key *key)
1287{
1288	struct ring_buffer *buffer;
1289	long nr_pages;
1290	int bsize;
1291	int cpu;
1292	int ret;
1293
1294	/* keep it in its own cache line */
1295	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1296			 GFP_KERNEL);
1297	if (!buffer)
1298		return NULL;
1299
1300	if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1301		goto fail_free_buffer;
1302
1303	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1304	buffer->flags = flags;
1305	buffer->clock = trace_clock_local;
1306	buffer->reader_lock_key = key;
1307
1308	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1309	init_waitqueue_head(&buffer->irq_work.waiters);
1310
1311	/* need at least two pages */
1312	if (nr_pages < 2)
1313		nr_pages = 2;
1314
 
 
 
 
 
 
 
 
 
 
 
1315	buffer->cpus = nr_cpu_ids;
1316
1317	bsize = sizeof(void *) * nr_cpu_ids;
1318	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1319				  GFP_KERNEL);
1320	if (!buffer->buffers)
1321		goto fail_free_cpumask;
1322
1323	cpu = raw_smp_processor_id();
1324	cpumask_set_cpu(cpu, buffer->cpumask);
1325	buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1326	if (!buffer->buffers[cpu])
1327		goto fail_free_buffers;
 
1328
1329	ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1330	if (ret < 0)
1331		goto fail_free_buffers;
 
 
 
1332
1333	mutex_init(&buffer->mutex);
1334
1335	return buffer;
1336
1337 fail_free_buffers:
1338	for_each_buffer_cpu(buffer, cpu) {
1339		if (buffer->buffers[cpu])
1340			rb_free_cpu_buffer(buffer->buffers[cpu]);
1341	}
1342	kfree(buffer->buffers);
1343
1344 fail_free_cpumask:
1345	free_cpumask_var(buffer->cpumask);
 
 
 
1346
1347 fail_free_buffer:
1348	kfree(buffer);
1349	return NULL;
1350}
1351EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1352
1353/**
1354 * ring_buffer_free - free a ring buffer.
1355 * @buffer: the buffer to free.
1356 */
1357void
1358ring_buffer_free(struct ring_buffer *buffer)
1359{
1360	int cpu;
1361
1362	cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
 
 
 
1363
1364	for_each_buffer_cpu(buffer, cpu)
1365		rb_free_cpu_buffer(buffer->buffers[cpu]);
1366
 
 
 
 
1367	kfree(buffer->buffers);
1368	free_cpumask_var(buffer->cpumask);
1369
1370	kfree(buffer);
1371}
1372EXPORT_SYMBOL_GPL(ring_buffer_free);
1373
1374void ring_buffer_set_clock(struct ring_buffer *buffer,
1375			   u64 (*clock)(void))
1376{
1377	buffer->clock = clock;
1378}
1379
1380static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1381
1382static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1383{
1384	return local_read(&bpage->entries) & RB_WRITE_MASK;
1385}
1386
1387static inline unsigned long rb_page_write(struct buffer_page *bpage)
1388{
1389	return local_read(&bpage->write) & RB_WRITE_MASK;
1390}
1391
1392static int
1393rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1394{
1395	struct list_head *tail_page, *to_remove, *next_page;
1396	struct buffer_page *to_remove_page, *tmp_iter_page;
1397	struct buffer_page *last_page, *first_page;
1398	unsigned long nr_removed;
1399	unsigned long head_bit;
1400	int page_entries;
1401
1402	head_bit = 0;
1403
1404	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1405	atomic_inc(&cpu_buffer->record_disabled);
1406	/*
1407	 * We don't race with the readers since we have acquired the reader
1408	 * lock. We also don't race with writers after disabling recording.
1409	 * This makes it easy to figure out the first and the last page to be
1410	 * removed from the list. We unlink all the pages in between including
1411	 * the first and last pages. This is done in a busy loop so that we
1412	 * lose the least number of traces.
1413	 * The pages are freed after we restart recording and unlock readers.
1414	 */
1415	tail_page = &cpu_buffer->tail_page->list;
1416
1417	/*
1418	 * tail page might be on reader page, we remove the next page
1419	 * from the ring buffer
1420	 */
1421	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1422		tail_page = rb_list_head(tail_page->next);
1423	to_remove = tail_page;
1424
1425	/* start of pages to remove */
1426	first_page = list_entry(rb_list_head(to_remove->next),
1427				struct buffer_page, list);
1428
1429	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1430		to_remove = rb_list_head(to_remove)->next;
1431		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1432	}
1433
1434	next_page = rb_list_head(to_remove)->next;
1435
1436	/*
1437	 * Now we remove all pages between tail_page and next_page.
1438	 * Make sure that we have head_bit value preserved for the
1439	 * next page
1440	 */
1441	tail_page->next = (struct list_head *)((unsigned long)next_page |
1442						head_bit);
1443	next_page = rb_list_head(next_page);
1444	next_page->prev = tail_page;
1445
1446	/* make sure pages points to a valid page in the ring buffer */
1447	cpu_buffer->pages = next_page;
1448
1449	/* update head page */
1450	if (head_bit)
1451		cpu_buffer->head_page = list_entry(next_page,
1452						struct buffer_page, list);
1453
1454	/*
1455	 * change read pointer to make sure any read iterators reset
1456	 * themselves
1457	 */
1458	cpu_buffer->read = 0;
1459
1460	/* pages are removed, resume tracing and then free the pages */
1461	atomic_dec(&cpu_buffer->record_disabled);
1462	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1463
1464	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1465
1466	/* last buffer page to remove */
1467	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1468				list);
1469	tmp_iter_page = first_page;
1470
1471	do {
1472		to_remove_page = tmp_iter_page;
1473		rb_inc_page(cpu_buffer, &tmp_iter_page);
1474
1475		/* update the counters */
1476		page_entries = rb_page_entries(to_remove_page);
1477		if (page_entries) {
1478			/*
1479			 * If something was added to this page, it was full
1480			 * since it is not the tail page. So we deduct the
1481			 * bytes consumed in ring buffer from here.
1482			 * Increment overrun to account for the lost events.
1483			 */
1484			local_add(page_entries, &cpu_buffer->overrun);
1485			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1486		}
1487
1488		/*
1489		 * We have already removed references to this list item, just
1490		 * free up the buffer_page and its page
1491		 */
1492		free_buffer_page(to_remove_page);
1493		nr_removed--;
1494
1495	} while (to_remove_page != last_page);
1496
1497	RB_WARN_ON(cpu_buffer, nr_removed);
1498
1499	return nr_removed == 0;
1500}
1501
1502static int
1503rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1504{
1505	struct list_head *pages = &cpu_buffer->new_pages;
1506	int retries, success;
1507
1508	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1509	/*
1510	 * We are holding the reader lock, so the reader page won't be swapped
1511	 * in the ring buffer. Now we are racing with the writer trying to
1512	 * move head page and the tail page.
1513	 * We are going to adapt the reader page update process where:
1514	 * 1. We first splice the start and end of list of new pages between
1515	 *    the head page and its previous page.
1516	 * 2. We cmpxchg the prev_page->next to point from head page to the
1517	 *    start of new pages list.
1518	 * 3. Finally, we update the head->prev to the end of new list.
1519	 *
1520	 * We will try this process 10 times, to make sure that we don't keep
1521	 * spinning.
1522	 */
1523	retries = 10;
1524	success = 0;
1525	while (retries--) {
1526		struct list_head *head_page, *prev_page, *r;
1527		struct list_head *last_page, *first_page;
1528		struct list_head *head_page_with_bit;
1529
1530		head_page = &rb_set_head_page(cpu_buffer)->list;
1531		if (!head_page)
1532			break;
1533		prev_page = head_page->prev;
1534
1535		first_page = pages->next;
1536		last_page  = pages->prev;
1537
1538		head_page_with_bit = (struct list_head *)
1539				     ((unsigned long)head_page | RB_PAGE_HEAD);
1540
1541		last_page->next = head_page_with_bit;
1542		first_page->prev = prev_page;
1543
1544		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1545
1546		if (r == head_page_with_bit) {
1547			/*
1548			 * yay, we replaced the page pointer to our new list,
1549			 * now, we just have to update to head page's prev
1550			 * pointer to point to end of list
1551			 */
1552			head_page->prev = last_page;
1553			success = 1;
1554			break;
1555		}
1556	}
1557
1558	if (success)
1559		INIT_LIST_HEAD(pages);
1560	/*
1561	 * If we weren't successful in adding in new pages, warn and stop
1562	 * tracing
1563	 */
1564	RB_WARN_ON(cpu_buffer, !success);
1565	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1566
1567	/* free pages if they weren't inserted */
1568	if (!success) {
1569		struct buffer_page *bpage, *tmp;
1570		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1571					 list) {
1572			list_del_init(&bpage->list);
1573			free_buffer_page(bpage);
1574		}
1575	}
1576	return success;
1577}
1578
1579static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1580{
1581	int success;
1582
1583	if (cpu_buffer->nr_pages_to_update > 0)
1584		success = rb_insert_pages(cpu_buffer);
1585	else
1586		success = rb_remove_pages(cpu_buffer,
1587					-cpu_buffer->nr_pages_to_update);
1588
1589	if (success)
1590		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1591}
1592
1593static void update_pages_handler(struct work_struct *work)
1594{
1595	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1596			struct ring_buffer_per_cpu, update_pages_work);
1597	rb_update_pages(cpu_buffer);
1598	complete(&cpu_buffer->update_done);
1599}
1600
1601/**
1602 * ring_buffer_resize - resize the ring buffer
1603 * @buffer: the buffer to resize.
1604 * @size: the new size.
1605 * @cpu_id: the cpu buffer to resize
1606 *
1607 * Minimum size is 2 * BUF_PAGE_SIZE.
1608 *
1609 * Returns 0 on success and < 0 on failure.
1610 */
1611int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1612			int cpu_id)
1613{
1614	struct ring_buffer_per_cpu *cpu_buffer;
1615	unsigned long nr_pages;
1616	int cpu, err = 0;
1617
1618	/*
1619	 * Always succeed at resizing a non-existent buffer:
1620	 */
1621	if (!buffer)
1622		return size;
1623
1624	/* Make sure the requested buffer exists */
1625	if (cpu_id != RING_BUFFER_ALL_CPUS &&
1626	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
1627		return size;
1628
1629	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
 
1630
1631	/* we need a minimum of two pages */
1632	if (nr_pages < 2)
1633		nr_pages = 2;
1634
1635	size = nr_pages * BUF_PAGE_SIZE;
1636
1637	/*
1638	 * Don't succeed if resizing is disabled, as a reader might be
1639	 * manipulating the ring buffer and is expecting a sane state while
1640	 * this is true.
1641	 */
1642	if (atomic_read(&buffer->resize_disabled))
1643		return -EBUSY;
1644
1645	/* prevent another thread from changing buffer sizes */
1646	mutex_lock(&buffer->mutex);
1647
1648	if (cpu_id == RING_BUFFER_ALL_CPUS) {
1649		/* calculate the pages to update */
1650		for_each_buffer_cpu(buffer, cpu) {
1651			cpu_buffer = buffer->buffers[cpu];
1652
1653			cpu_buffer->nr_pages_to_update = nr_pages -
1654							cpu_buffer->nr_pages;
1655			/*
1656			 * nothing more to do for removing pages or no update
1657			 */
1658			if (cpu_buffer->nr_pages_to_update <= 0)
1659				continue;
1660			/*
1661			 * to add pages, make sure all new pages can be
1662			 * allocated without receiving ENOMEM
1663			 */
1664			INIT_LIST_HEAD(&cpu_buffer->new_pages);
1665			if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1666						&cpu_buffer->new_pages, cpu)) {
1667				/* not enough memory for new pages */
1668				err = -ENOMEM;
1669				goto out_err;
1670			}
1671		}
1672
1673		get_online_cpus();
1674		/*
1675		 * Fire off all the required work handlers
1676		 * We can't schedule on offline CPUs, but it's not necessary
1677		 * since we can change their buffer sizes without any race.
1678		 */
1679		for_each_buffer_cpu(buffer, cpu) {
1680			cpu_buffer = buffer->buffers[cpu];
1681			if (!cpu_buffer->nr_pages_to_update)
1682				continue;
1683
1684			/* Can't run something on an offline CPU. */
1685			if (!cpu_online(cpu)) {
 
1686				rb_update_pages(cpu_buffer);
1687				cpu_buffer->nr_pages_to_update = 0;
1688			} else {
 
 
 
 
 
1689				schedule_work_on(cpu,
1690						&cpu_buffer->update_pages_work);
 
1691			}
 
1692		}
1693
1694		/* wait for all the updates to complete */
1695		for_each_buffer_cpu(buffer, cpu) {
1696			cpu_buffer = buffer->buffers[cpu];
1697			if (!cpu_buffer->nr_pages_to_update)
1698				continue;
1699
1700			if (cpu_online(cpu))
1701				wait_for_completion(&cpu_buffer->update_done);
1702			cpu_buffer->nr_pages_to_update = 0;
1703		}
1704
1705		put_online_cpus();
1706	} else {
1707		/* Make sure this CPU has been intitialized */
1708		if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1709			goto out;
1710
1711		cpu_buffer = buffer->buffers[cpu_id];
1712
1713		if (nr_pages == cpu_buffer->nr_pages)
1714			goto out;
1715
1716		cpu_buffer->nr_pages_to_update = nr_pages -
1717						cpu_buffer->nr_pages;
1718
1719		INIT_LIST_HEAD(&cpu_buffer->new_pages);
1720		if (cpu_buffer->nr_pages_to_update > 0 &&
1721			__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1722					    &cpu_buffer->new_pages, cpu_id)) {
1723			err = -ENOMEM;
1724			goto out_err;
1725		}
1726
1727		get_online_cpus();
1728
1729		/* Can't run something on an offline CPU. */
1730		if (!cpu_online(cpu_id))
 
1731			rb_update_pages(cpu_buffer);
1732		else {
 
 
 
 
 
1733			schedule_work_on(cpu_id,
1734					 &cpu_buffer->update_pages_work);
1735			wait_for_completion(&cpu_buffer->update_done);
 
1736		}
 
1737
1738		cpu_buffer->nr_pages_to_update = 0;
1739		put_online_cpus();
1740	}
1741
1742 out:
1743	/*
1744	 * The ring buffer resize can happen with the ring buffer
1745	 * enabled, so that the update disturbs the tracing as little
1746	 * as possible. But if the buffer is disabled, we do not need
1747	 * to worry about that, and we can take the time to verify
1748	 * that the buffer is not corrupt.
1749	 */
1750	if (atomic_read(&buffer->record_disabled)) {
1751		atomic_inc(&buffer->record_disabled);
1752		/*
1753		 * Even though the buffer was disabled, we must make sure
1754		 * that it is truly disabled before calling rb_check_pages.
1755		 * There could have been a race between checking
1756		 * record_disable and incrementing it.
1757		 */
1758		synchronize_sched();
1759		for_each_buffer_cpu(buffer, cpu) {
1760			cpu_buffer = buffer->buffers[cpu];
1761			rb_check_pages(cpu_buffer);
1762		}
1763		atomic_dec(&buffer->record_disabled);
1764	}
1765
1766	mutex_unlock(&buffer->mutex);
1767	return size;
1768
1769 out_err:
1770	for_each_buffer_cpu(buffer, cpu) {
1771		struct buffer_page *bpage, *tmp;
1772
1773		cpu_buffer = buffer->buffers[cpu];
1774		cpu_buffer->nr_pages_to_update = 0;
1775
1776		if (list_empty(&cpu_buffer->new_pages))
1777			continue;
1778
1779		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1780					list) {
1781			list_del_init(&bpage->list);
1782			free_buffer_page(bpage);
1783		}
1784	}
1785	mutex_unlock(&buffer->mutex);
1786	return err;
1787}
1788EXPORT_SYMBOL_GPL(ring_buffer_resize);
1789
1790void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1791{
1792	mutex_lock(&buffer->mutex);
1793	if (val)
1794		buffer->flags |= RB_FL_OVERWRITE;
1795	else
1796		buffer->flags &= ~RB_FL_OVERWRITE;
1797	mutex_unlock(&buffer->mutex);
1798}
1799EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1800
1801static __always_inline void *
1802__rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1803{
1804	return bpage->data + index;
1805}
1806
1807static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1808{
1809	return bpage->page->data + index;
1810}
1811
1812static __always_inline struct ring_buffer_event *
1813rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1814{
1815	return __rb_page_index(cpu_buffer->reader_page,
1816			       cpu_buffer->reader_page->read);
1817}
1818
1819static __always_inline struct ring_buffer_event *
1820rb_iter_head_event(struct ring_buffer_iter *iter)
1821{
1822	return __rb_page_index(iter->head_page, iter->head);
1823}
1824
1825static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1826{
1827	return local_read(&bpage->page->commit);
1828}
1829
1830/* Size is determined by what has been committed */
1831static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1832{
1833	return rb_page_commit(bpage);
1834}
1835
1836static __always_inline unsigned
1837rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1838{
1839	return rb_page_commit(cpu_buffer->commit_page);
1840}
1841
1842static __always_inline unsigned
1843rb_event_index(struct ring_buffer_event *event)
1844{
1845	unsigned long addr = (unsigned long)event;
1846
1847	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1848}
1849
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1850static void rb_inc_iter(struct ring_buffer_iter *iter)
1851{
1852	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1853
1854	/*
1855	 * The iterator could be on the reader page (it starts there).
1856	 * But the head could have moved, since the reader was
1857	 * found. Check for this case and assign the iterator
1858	 * to the head page instead of next.
1859	 */
1860	if (iter->head_page == cpu_buffer->reader_page)
1861		iter->head_page = rb_set_head_page(cpu_buffer);
1862	else
1863		rb_inc_page(cpu_buffer, &iter->head_page);
1864
1865	iter->read_stamp = iter->head_page->page->time_stamp;
1866	iter->head = 0;
1867}
1868
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1869/*
1870 * rb_handle_head_page - writer hit the head page
1871 *
1872 * Returns: +1 to retry page
1873 *           0 to continue
1874 *          -1 on error
1875 */
1876static int
1877rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1878		    struct buffer_page *tail_page,
1879		    struct buffer_page *next_page)
1880{
1881	struct buffer_page *new_head;
1882	int entries;
1883	int type;
1884	int ret;
1885
1886	entries = rb_page_entries(next_page);
1887
1888	/*
1889	 * The hard part is here. We need to move the head
1890	 * forward, and protect against both readers on
1891	 * other CPUs and writers coming in via interrupts.
1892	 */
1893	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1894				       RB_PAGE_HEAD);
1895
1896	/*
1897	 * type can be one of four:
1898	 *  NORMAL - an interrupt already moved it for us
1899	 *  HEAD   - we are the first to get here.
1900	 *  UPDATE - we are the interrupt interrupting
1901	 *           a current move.
1902	 *  MOVED  - a reader on another CPU moved the next
1903	 *           pointer to its reader page. Give up
1904	 *           and try again.
1905	 */
1906
1907	switch (type) {
1908	case RB_PAGE_HEAD:
1909		/*
1910		 * We changed the head to UPDATE, thus
1911		 * it is our responsibility to update
1912		 * the counters.
1913		 */
1914		local_add(entries, &cpu_buffer->overrun);
1915		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1916
1917		/*
1918		 * The entries will be zeroed out when we move the
1919		 * tail page.
1920		 */
1921
1922		/* still more to do */
1923		break;
1924
1925	case RB_PAGE_UPDATE:
1926		/*
1927		 * This is an interrupt that interrupt the
1928		 * previous update. Still more to do.
1929		 */
1930		break;
1931	case RB_PAGE_NORMAL:
1932		/*
1933		 * An interrupt came in before the update
1934		 * and processed this for us.
1935		 * Nothing left to do.
1936		 */
1937		return 1;
1938	case RB_PAGE_MOVED:
1939		/*
1940		 * The reader is on another CPU and just did
1941		 * a swap with our next_page.
1942		 * Try again.
1943		 */
1944		return 1;
1945	default:
1946		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1947		return -1;
1948	}
1949
1950	/*
1951	 * Now that we are here, the old head pointer is
1952	 * set to UPDATE. This will keep the reader from
1953	 * swapping the head page with the reader page.
1954	 * The reader (on another CPU) will spin till
1955	 * we are finished.
1956	 *
1957	 * We just need to protect against interrupts
1958	 * doing the job. We will set the next pointer
1959	 * to HEAD. After that, we set the old pointer
1960	 * to NORMAL, but only if it was HEAD before.
1961	 * otherwise we are an interrupt, and only
1962	 * want the outer most commit to reset it.
1963	 */
1964	new_head = next_page;
1965	rb_inc_page(cpu_buffer, &new_head);
1966
1967	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1968				    RB_PAGE_NORMAL);
1969
1970	/*
1971	 * Valid returns are:
1972	 *  HEAD   - an interrupt came in and already set it.
1973	 *  NORMAL - One of two things:
1974	 *            1) We really set it.
1975	 *            2) A bunch of interrupts came in and moved
1976	 *               the page forward again.
1977	 */
1978	switch (ret) {
1979	case RB_PAGE_HEAD:
1980	case RB_PAGE_NORMAL:
1981		/* OK */
1982		break;
1983	default:
1984		RB_WARN_ON(cpu_buffer, 1);
1985		return -1;
1986	}
1987
1988	/*
1989	 * It is possible that an interrupt came in,
1990	 * set the head up, then more interrupts came in
1991	 * and moved it again. When we get back here,
1992	 * the page would have been set to NORMAL but we
1993	 * just set it back to HEAD.
1994	 *
1995	 * How do you detect this? Well, if that happened
1996	 * the tail page would have moved.
1997	 */
1998	if (ret == RB_PAGE_NORMAL) {
1999		struct buffer_page *buffer_tail_page;
2000
2001		buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2002		/*
2003		 * If the tail had moved passed next, then we need
2004		 * to reset the pointer.
2005		 */
2006		if (buffer_tail_page != tail_page &&
2007		    buffer_tail_page != next_page)
2008			rb_head_page_set_normal(cpu_buffer, new_head,
2009						next_page,
2010						RB_PAGE_HEAD);
2011	}
2012
2013	/*
2014	 * If this was the outer most commit (the one that
2015	 * changed the original pointer from HEAD to UPDATE),
2016	 * then it is up to us to reset it to NORMAL.
2017	 */
2018	if (type == RB_PAGE_HEAD) {
2019		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2020					      tail_page,
2021					      RB_PAGE_UPDATE);
2022		if (RB_WARN_ON(cpu_buffer,
2023			       ret != RB_PAGE_UPDATE))
2024			return -1;
2025	}
2026
2027	return 0;
2028}
2029
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2030static inline void
2031rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2032	      unsigned long tail, struct rb_event_info *info)
 
2033{
2034	struct buffer_page *tail_page = info->tail_page;
2035	struct ring_buffer_event *event;
2036	unsigned long length = info->length;
2037
2038	/*
2039	 * Only the event that crossed the page boundary
2040	 * must fill the old tail_page with padding.
2041	 */
2042	if (tail >= BUF_PAGE_SIZE) {
2043		/*
2044		 * If the page was filled, then we still need
2045		 * to update the real_end. Reset it to zero
2046		 * and the reader will ignore it.
2047		 */
2048		if (tail == BUF_PAGE_SIZE)
2049			tail_page->real_end = 0;
2050
2051		local_sub(length, &tail_page->write);
2052		return;
2053	}
2054
2055	event = __rb_page_index(tail_page, tail);
2056	kmemcheck_annotate_bitfield(event, bitfield);
2057
2058	/* account for padding bytes */
2059	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2060
2061	/*
2062	 * Save the original length to the meta data.
2063	 * This will be used by the reader to add lost event
2064	 * counter.
2065	 */
2066	tail_page->real_end = tail;
2067
2068	/*
2069	 * If this event is bigger than the minimum size, then
2070	 * we need to be careful that we don't subtract the
2071	 * write counter enough to allow another writer to slip
2072	 * in on this page.
2073	 * We put in a discarded commit instead, to make sure
2074	 * that this space is not used again.
2075	 *
2076	 * If we are less than the minimum size, we don't need to
2077	 * worry about it.
2078	 */
2079	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2080		/* No room for any events */
2081
2082		/* Mark the rest of the page with padding */
2083		rb_event_set_padding(event);
2084
2085		/* Set the write back to the previous setting */
2086		local_sub(length, &tail_page->write);
2087		return;
2088	}
2089
2090	/* Put in a discarded event */
2091	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2092	event->type_len = RINGBUF_TYPE_PADDING;
2093	/* time delta must be non zero */
2094	event->time_delta = 1;
2095
2096	/* Set write to end of buffer */
2097	length = (tail + length) - BUF_PAGE_SIZE;
2098	local_sub(length, &tail_page->write);
2099}
2100
2101static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2102
2103/*
2104 * This is the slow path, force gcc not to inline it.
2105 */
2106static noinline struct ring_buffer_event *
2107rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2108	     unsigned long tail, struct rb_event_info *info)
 
2109{
2110	struct buffer_page *tail_page = info->tail_page;
2111	struct buffer_page *commit_page = cpu_buffer->commit_page;
2112	struct ring_buffer *buffer = cpu_buffer->buffer;
2113	struct buffer_page *next_page;
2114	int ret;
2115
2116	next_page = tail_page;
2117
2118	rb_inc_page(cpu_buffer, &next_page);
2119
2120	/*
2121	 * If for some reason, we had an interrupt storm that made
2122	 * it all the way around the buffer, bail, and warn
2123	 * about it.
2124	 */
2125	if (unlikely(next_page == commit_page)) {
2126		local_inc(&cpu_buffer->commit_overrun);
2127		goto out_reset;
2128	}
2129
2130	/*
2131	 * This is where the fun begins!
2132	 *
2133	 * We are fighting against races between a reader that
2134	 * could be on another CPU trying to swap its reader
2135	 * page with the buffer head.
2136	 *
2137	 * We are also fighting against interrupts coming in and
2138	 * moving the head or tail on us as well.
2139	 *
2140	 * If the next page is the head page then we have filled
2141	 * the buffer, unless the commit page is still on the
2142	 * reader page.
2143	 */
2144	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2145
2146		/*
2147		 * If the commit is not on the reader page, then
2148		 * move the header page.
2149		 */
2150		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2151			/*
2152			 * If we are not in overwrite mode,
2153			 * this is easy, just stop here.
2154			 */
2155			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2156				local_inc(&cpu_buffer->dropped_events);
2157				goto out_reset;
2158			}
2159
2160			ret = rb_handle_head_page(cpu_buffer,
2161						  tail_page,
2162						  next_page);
2163			if (ret < 0)
2164				goto out_reset;
2165			if (ret)
2166				goto out_again;
2167		} else {
2168			/*
2169			 * We need to be careful here too. The
2170			 * commit page could still be on the reader
2171			 * page. We could have a small buffer, and
2172			 * have filled up the buffer with events
2173			 * from interrupts and such, and wrapped.
2174			 *
2175			 * Note, if the tail page is also the on the
2176			 * reader_page, we let it move out.
2177			 */
2178			if (unlikely((cpu_buffer->commit_page !=
2179				      cpu_buffer->tail_page) &&
2180				     (cpu_buffer->commit_page ==
2181				      cpu_buffer->reader_page))) {
2182				local_inc(&cpu_buffer->commit_overrun);
2183				goto out_reset;
2184			}
2185		}
2186	}
2187
2188	rb_tail_page_update(cpu_buffer, tail_page, next_page);
 
 
 
 
 
 
 
 
2189
2190 out_again:
2191
2192	rb_reset_tail(cpu_buffer, tail, info);
2193
2194	/* Commit what we have for now. */
2195	rb_end_commit(cpu_buffer);
2196	/* rb_end_commit() decs committing */
2197	local_inc(&cpu_buffer->committing);
2198
2199	/* fail and let the caller try again */
2200	return ERR_PTR(-EAGAIN);
2201
2202 out_reset:
2203	/* reset write */
2204	rb_reset_tail(cpu_buffer, tail, info);
2205
2206	return NULL;
2207}
2208
2209/* Slow path, do not inline */
2210static noinline struct ring_buffer_event *
2211rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
 
2212{
2213	event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2214
2215	/* Not the first event on the page? */
2216	if (rb_event_index(event)) {
2217		event->time_delta = delta & TS_MASK;
2218		event->array[0] = delta >> TS_SHIFT;
2219	} else {
2220		/* nope, just zero it */
2221		event->time_delta = 0;
2222		event->array[0] = 0;
2223	}
2224
2225	return skip_time_extend(event);
2226}
2227
2228static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2229				     struct ring_buffer_event *event);
 
 
 
 
 
2230
2231/**
2232 * rb_update_event - update event type and data
2233 * @event: the event to update
2234 * @type: the type of event
2235 * @length: the size of the event field in the ring buffer
2236 *
2237 * Update the type and data fields of the event. The length
2238 * is the actual size that is written to the ring buffer,
2239 * and with this, we can determine what to place into the
2240 * data field.
2241 */
2242static void
2243rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2244		struct ring_buffer_event *event,
2245		struct rb_event_info *info)
2246{
2247	unsigned length = info->length;
2248	u64 delta = info->delta;
2249
2250	/* Only a commit updates the timestamp */
2251	if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2252		delta = 0;
2253
2254	/*
2255	 * If we need to add a timestamp, then we
2256	 * add it to the start of the resevered space.
2257	 */
2258	if (unlikely(info->add_timestamp)) {
2259		event = rb_add_time_stamp(event, delta);
2260		length -= RB_LEN_TIME_EXTEND;
2261		delta = 0;
2262	}
2263
2264	event->time_delta = delta;
2265	length -= RB_EVNT_HDR_SIZE;
2266	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2267		event->type_len = 0;
2268		event->array[0] = length;
2269	} else
2270		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2271}
2272
2273static unsigned rb_calculate_event_length(unsigned length)
2274{
2275	struct ring_buffer_event event; /* Used only for sizeof array */
2276
2277	/* zero length can cause confusions */
2278	if (!length)
2279		length++;
2280
2281	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2282		length += sizeof(event.array[0]);
 
2283
2284	length += RB_EVNT_HDR_SIZE;
2285	length = ALIGN(length, RB_ARCH_ALIGNMENT);
2286
2287	/*
2288	 * In case the time delta is larger than the 27 bits for it
2289	 * in the header, we need to add a timestamp. If another
2290	 * event comes in when trying to discard this one to increase
2291	 * the length, then the timestamp will be added in the allocated
2292	 * space of this event. If length is bigger than the size needed
2293	 * for the TIME_EXTEND, then padding has to be used. The events
2294	 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2295	 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2296	 * As length is a multiple of 4, we only need to worry if it
2297	 * is 12 (RB_LEN_TIME_EXTEND + 4).
2298	 */
2299	if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2300		length += RB_ALIGNMENT;
2301
2302	return length;
2303}
2304
2305#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2306static inline bool sched_clock_stable(void)
2307{
2308	return true;
2309}
2310#endif
2311
2312static inline int
2313rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2314		  struct ring_buffer_event *event)
2315{
2316	unsigned long new_index, old_index;
2317	struct buffer_page *bpage;
2318	unsigned long index;
2319	unsigned long addr;
2320
2321	new_index = rb_event_index(event);
2322	old_index = new_index + rb_event_ts_length(event);
2323	addr = (unsigned long)event;
2324	addr &= PAGE_MASK;
2325
2326	bpage = READ_ONCE(cpu_buffer->tail_page);
2327
2328	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2329		unsigned long write_mask =
2330			local_read(&bpage->write) & ~RB_WRITE_MASK;
2331		unsigned long event_length = rb_event_length(event);
2332		/*
2333		 * This is on the tail page. It is possible that
2334		 * a write could come in and move the tail page
2335		 * and write to the next page. That is fine
2336		 * because we just shorten what is on this page.
2337		 */
2338		old_index += write_mask;
2339		new_index += write_mask;
2340		index = local_cmpxchg(&bpage->write, old_index, new_index);
2341		if (index == old_index) {
2342			/* update counters */
2343			local_sub(event_length, &cpu_buffer->entries_bytes);
2344			return 1;
2345		}
2346	}
2347
2348	/* could not discard */
2349	return 0;
2350}
2351
2352static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2353{
2354	local_inc(&cpu_buffer->committing);
2355	local_inc(&cpu_buffer->commits);
2356}
2357
2358static __always_inline void
2359rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2360{
2361	unsigned long max_count;
2362
2363	/*
2364	 * We only race with interrupts and NMIs on this CPU.
2365	 * If we own the commit event, then we can commit
2366	 * all others that interrupted us, since the interruptions
2367	 * are in stack format (they finish before they come
2368	 * back to us). This allows us to do a simple loop to
2369	 * assign the commit to the tail.
2370	 */
2371 again:
2372	max_count = cpu_buffer->nr_pages * 100;
2373
2374	while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2375		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2376			return;
2377		if (RB_WARN_ON(cpu_buffer,
2378			       rb_is_reader_page(cpu_buffer->tail_page)))
2379			return;
2380		local_set(&cpu_buffer->commit_page->page->commit,
2381			  rb_page_write(cpu_buffer->commit_page));
2382		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2383		/* Only update the write stamp if the page has an event */
2384		if (rb_page_write(cpu_buffer->commit_page))
2385			cpu_buffer->write_stamp =
2386				cpu_buffer->commit_page->page->time_stamp;
2387		/* add barrier to keep gcc from optimizing too much */
2388		barrier();
2389	}
2390	while (rb_commit_index(cpu_buffer) !=
2391	       rb_page_write(cpu_buffer->commit_page)) {
2392
2393		local_set(&cpu_buffer->commit_page->page->commit,
2394			  rb_page_write(cpu_buffer->commit_page));
2395		RB_WARN_ON(cpu_buffer,
2396			   local_read(&cpu_buffer->commit_page->page->commit) &
2397			   ~RB_WRITE_MASK);
2398		barrier();
2399	}
2400
2401	/* again, keep gcc from optimizing */
2402	barrier();
2403
2404	/*
2405	 * If an interrupt came in just after the first while loop
2406	 * and pushed the tail page forward, we will be left with
2407	 * a dangling commit that will never go forward.
2408	 */
2409	if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2410		goto again;
2411}
2412
2413static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2414{
2415	unsigned long commits;
2416
2417	if (RB_WARN_ON(cpu_buffer,
2418		       !local_read(&cpu_buffer->committing)))
2419		return;
2420
2421 again:
2422	commits = local_read(&cpu_buffer->commits);
2423	/* synchronize with interrupts */
2424	barrier();
2425	if (local_read(&cpu_buffer->committing) == 1)
2426		rb_set_commit_to_write(cpu_buffer);
2427
2428	local_dec(&cpu_buffer->committing);
2429
2430	/* synchronize with interrupts */
2431	barrier();
2432
2433	/*
2434	 * Need to account for interrupts coming in between the
2435	 * updating of the commit page and the clearing of the
2436	 * committing counter.
2437	 */
2438	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2439	    !local_read(&cpu_buffer->committing)) {
2440		local_inc(&cpu_buffer->committing);
2441		goto again;
2442	}
2443}
2444
2445static inline void rb_event_discard(struct ring_buffer_event *event)
2446{
2447	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2448		event = skip_time_extend(event);
2449
2450	/* array[0] holds the actual length for the discarded event */
2451	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2452	event->type_len = RINGBUF_TYPE_PADDING;
2453	/* time delta must be non zero */
2454	if (!event->time_delta)
2455		event->time_delta = 1;
2456}
2457
2458static __always_inline bool
2459rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2460		   struct ring_buffer_event *event)
2461{
2462	unsigned long addr = (unsigned long)event;
2463	unsigned long index;
 
 
 
2464
2465	index = rb_event_index(event);
2466	addr &= PAGE_MASK;
2467
2468	return cpu_buffer->commit_page->page == (void *)addr &&
2469		rb_commit_index(cpu_buffer) == index;
2470}
 
 
 
 
 
 
 
 
 
 
 
2471
2472static __always_inline void
2473rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2474		      struct ring_buffer_event *event)
2475{
2476	u64 delta;
2477
2478	/*
2479	 * The event first in the commit queue updates the
2480	 * time stamp.
 
 
 
 
 
2481	 */
2482	if (rb_event_is_commit(cpu_buffer, event)) {
2483		/*
2484		 * A commit event that is first on a page
2485		 * updates the write timestamp with the page stamp
2486		 */
2487		if (!rb_event_index(event))
2488			cpu_buffer->write_stamp =
2489				cpu_buffer->commit_page->page->time_stamp;
2490		else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2491			delta = event->array[0];
2492			delta <<= TS_SHIFT;
2493			delta += event->time_delta;
2494			cpu_buffer->write_stamp += delta;
2495		} else
2496			cpu_buffer->write_stamp += event->time_delta;
2497	}
2498}
2499
2500static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2501		      struct ring_buffer_event *event)
2502{
2503	local_inc(&cpu_buffer->entries);
2504	rb_update_write_stamp(cpu_buffer, event);
2505	rb_end_commit(cpu_buffer);
2506}
2507
2508static __always_inline void
2509rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2510{
2511	bool pagebusy;
2512
2513	if (buffer->irq_work.waiters_pending) {
2514		buffer->irq_work.waiters_pending = false;
2515		/* irq_work_queue() supplies it's own memory barriers */
2516		irq_work_queue(&buffer->irq_work.work);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2517	}
2518
2519	if (cpu_buffer->irq_work.waiters_pending) {
2520		cpu_buffer->irq_work.waiters_pending = false;
2521		/* irq_work_queue() supplies it's own memory barriers */
2522		irq_work_queue(&cpu_buffer->irq_work.work);
2523	}
2524
2525	pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 
2526
2527	if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2528		cpu_buffer->irq_work.wakeup_full = true;
2529		cpu_buffer->irq_work.full_waiters_pending = false;
2530		/* irq_work_queue() supplies it's own memory barriers */
2531		irq_work_queue(&cpu_buffer->irq_work.work);
2532	}
2533}
2534
 
 
2535/*
2536 * The lock and unlock are done within a preempt disable section.
2537 * The current_context per_cpu variable can only be modified
2538 * by the current task between lock and unlock. But it can
2539 * be modified more than once via an interrupt. To pass this
2540 * information from the lock to the unlock without having to
2541 * access the 'in_interrupt()' functions again (which do show
2542 * a bit of overhead in something as critical as function tracing,
2543 * we use a bitmask trick.
2544 *
2545 *  bit 0 =  NMI context
2546 *  bit 1 =  IRQ context
2547 *  bit 2 =  SoftIRQ context
2548 *  bit 3 =  normal context.
2549 *
2550 * This works because this is the order of contexts that can
2551 * preempt other contexts. A SoftIRQ never preempts an IRQ
2552 * context.
2553 *
2554 * When the context is determined, the corresponding bit is
2555 * checked and set (if it was set, then a recursion of that context
2556 * happened).
2557 *
2558 * On unlock, we need to clear this bit. To do so, just subtract
2559 * 1 from the current_context and AND it to itself.
2560 *
2561 * (binary)
2562 *  101 - 1 = 100
2563 *  101 & 100 = 100 (clearing bit zero)
2564 *
2565 *  1010 - 1 = 1001
2566 *  1010 & 1001 = 1000 (clearing bit 1)
2567 *
2568 * The least significant bit can be cleared this way, and it
2569 * just so happens that it is the same bit corresponding to
2570 * the current context.
2571 */
 
2572
2573static __always_inline int
2574trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2575{
2576	unsigned int val = cpu_buffer->current_context;
2577	int bit;
2578
2579	if (in_interrupt()) {
2580		if (in_nmi())
2581			bit = RB_CTX_NMI;
2582		else if (in_irq())
2583			bit = RB_CTX_IRQ;
2584		else
2585			bit = RB_CTX_SOFTIRQ;
2586	} else
2587		bit = RB_CTX_NORMAL;
2588
2589	if (unlikely(val & (1 << bit)))
2590		return 1;
2591
2592	val |= (1 << bit);
2593	cpu_buffer->current_context = val;
2594
2595	return 0;
2596}
2597
2598static __always_inline void
2599trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2600{
2601	cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2602}
2603
2604/**
2605 * ring_buffer_unlock_commit - commit a reserved
2606 * @buffer: The buffer to commit to
2607 * @event: The event pointer to commit.
2608 *
2609 * This commits the data to the ring buffer, and releases any locks held.
2610 *
2611 * Must be paired with ring_buffer_lock_reserve.
2612 */
2613int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2614			      struct ring_buffer_event *event)
2615{
2616	struct ring_buffer_per_cpu *cpu_buffer;
2617	int cpu = raw_smp_processor_id();
2618
2619	cpu_buffer = buffer->buffers[cpu];
2620
2621	rb_commit(cpu_buffer, event);
2622
2623	rb_wakeups(buffer, cpu_buffer);
2624
2625	trace_recursive_unlock(cpu_buffer);
2626
2627	preempt_enable_notrace();
2628
2629	return 0;
2630}
2631EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2632
2633static noinline void
2634rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2635		    struct rb_event_info *info)
2636{
2637	WARN_ONCE(info->delta > (1ULL << 59),
2638		  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2639		  (unsigned long long)info->delta,
2640		  (unsigned long long)info->ts,
2641		  (unsigned long long)cpu_buffer->write_stamp,
2642		  sched_clock_stable() ? "" :
2643		  "If you just came from a suspend/resume,\n"
2644		  "please switch to the trace global clock:\n"
2645		  "  echo global > /sys/kernel/debug/tracing/trace_clock\n");
2646	info->add_timestamp = 1;
2647}
2648
2649static struct ring_buffer_event *
2650__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2651		  struct rb_event_info *info)
2652{
2653	struct ring_buffer_event *event;
2654	struct buffer_page *tail_page;
2655	unsigned long tail, write;
2656
2657	/*
2658	 * If the time delta since the last event is too big to
2659	 * hold in the time field of the event, then we append a
2660	 * TIME EXTEND event ahead of the data event.
2661	 */
2662	if (unlikely(info->add_timestamp))
2663		info->length += RB_LEN_TIME_EXTEND;
2664
2665	/* Don't let the compiler play games with cpu_buffer->tail_page */
2666	tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2667	write = local_add_return(info->length, &tail_page->write);
2668
2669	/* set write to only the index of the write */
2670	write &= RB_WRITE_MASK;
2671	tail = write - info->length;
2672
2673	/*
2674	 * If this is the first commit on the page, then it has the same
2675	 * timestamp as the page itself.
2676	 */
2677	if (!tail)
2678		info->delta = 0;
2679
2680	/* See if we shot pass the end of this buffer page */
2681	if (unlikely(write > BUF_PAGE_SIZE))
2682		return rb_move_tail(cpu_buffer, tail, info);
2683
2684	/* We reserved something on the buffer */
2685
2686	event = __rb_page_index(tail_page, tail);
2687	kmemcheck_annotate_bitfield(event, bitfield);
2688	rb_update_event(cpu_buffer, event, info);
2689
2690	local_inc(&tail_page->entries);
2691
2692	/*
2693	 * If this is the first commit on the page, then update
2694	 * its timestamp.
2695	 */
2696	if (!tail)
2697		tail_page->page->time_stamp = info->ts;
2698
2699	/* account for these added bytes */
2700	local_add(info->length, &cpu_buffer->entries_bytes);
2701
2702	return event;
2703}
2704
2705static __always_inline struct ring_buffer_event *
2706rb_reserve_next_event(struct ring_buffer *buffer,
2707		      struct ring_buffer_per_cpu *cpu_buffer,
2708		      unsigned long length)
2709{
2710	struct ring_buffer_event *event;
2711	struct rb_event_info info;
2712	int nr_loops = 0;
2713	u64 diff;
2714
2715	rb_start_commit(cpu_buffer);
 
2716
2717#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2718	/*
2719	 * Due to the ability to swap a cpu buffer from a buffer
2720	 * it is possible it was swapped before we committed.
2721	 * (committing stops a swap). We check for it here and
2722	 * if it happened, we have to fail the write.
2723	 */
2724	barrier();
2725	if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2726		local_dec(&cpu_buffer->committing);
2727		local_dec(&cpu_buffer->commits);
2728		return NULL;
2729	}
2730#endif
2731
2732	info.length = rb_calculate_event_length(length);
2733 again:
2734	info.add_timestamp = 0;
2735	info.delta = 0;
2736
2737	/*
2738	 * We allow for interrupts to reenter here and do a trace.
2739	 * If one does, it will cause this original code to loop
2740	 * back here. Even with heavy interrupts happening, this
2741	 * should only happen a few times in a row. If this happens
2742	 * 1000 times in a row, there must be either an interrupt
2743	 * storm or we have something buggy.
2744	 * Bail!
2745	 */
2746	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2747		goto out_fail;
2748
2749	info.ts = rb_time_stamp(cpu_buffer->buffer);
2750	diff = info.ts - cpu_buffer->write_stamp;
2751
2752	/* make sure this diff is calculated here */
2753	barrier();
2754
2755	/* Did the write stamp get updated already? */
2756	if (likely(info.ts >= cpu_buffer->write_stamp)) {
2757		info.delta = diff;
2758		if (unlikely(test_time_stamp(info.delta)))
2759			rb_handle_timestamp(cpu_buffer, &info);
2760	}
2761
2762	event = __rb_reserve_next(cpu_buffer, &info);
2763
2764	if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2765		if (info.add_timestamp)
2766			info.length -= RB_LEN_TIME_EXTEND;
2767		goto again;
2768	}
2769
2770	if (!event)
2771		goto out_fail;
2772
2773	return event;
2774
2775 out_fail:
2776	rb_end_commit(cpu_buffer);
2777	return NULL;
2778}
2779
2780/**
2781 * ring_buffer_lock_reserve - reserve a part of the buffer
2782 * @buffer: the ring buffer to reserve from
2783 * @length: the length of the data to reserve (excluding event header)
2784 *
2785 * Returns a reseverd event on the ring buffer to copy directly to.
2786 * The user of this interface will need to get the body to write into
2787 * and can use the ring_buffer_event_data() interface.
2788 *
2789 * The length is the length of the data needed, not the event length
2790 * which also includes the event header.
2791 *
2792 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2793 * If NULL is returned, then nothing has been allocated or locked.
2794 */
2795struct ring_buffer_event *
2796ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2797{
2798	struct ring_buffer_per_cpu *cpu_buffer;
2799	struct ring_buffer_event *event;
2800	int cpu;
2801
 
 
 
2802	/* If we are tracing schedule, we don't want to recurse */
2803	preempt_disable_notrace();
2804
2805	if (unlikely(atomic_read(&buffer->record_disabled)))
2806		goto out;
 
 
 
2807
2808	cpu = raw_smp_processor_id();
2809
2810	if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2811		goto out;
2812
2813	cpu_buffer = buffer->buffers[cpu];
2814
2815	if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2816		goto out;
2817
2818	if (unlikely(length > BUF_MAX_DATA_SIZE))
2819		goto out;
2820
2821	if (unlikely(trace_recursive_lock(cpu_buffer)))
2822		goto out;
2823
2824	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2825	if (!event)
2826		goto out_unlock;
2827
2828	return event;
2829
2830 out_unlock:
2831	trace_recursive_unlock(cpu_buffer);
2832 out:
 
 
 
2833	preempt_enable_notrace();
2834	return NULL;
2835}
2836EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2837
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2838/*
2839 * Decrement the entries to the page that an event is on.
2840 * The event does not even need to exist, only the pointer
2841 * to the page it is on. This may only be called before the commit
2842 * takes place.
2843 */
2844static inline void
2845rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2846		   struct ring_buffer_event *event)
2847{
2848	unsigned long addr = (unsigned long)event;
2849	struct buffer_page *bpage = cpu_buffer->commit_page;
2850	struct buffer_page *start;
2851
2852	addr &= PAGE_MASK;
2853
2854	/* Do the likely case first */
2855	if (likely(bpage->page == (void *)addr)) {
2856		local_dec(&bpage->entries);
2857		return;
2858	}
2859
2860	/*
2861	 * Because the commit page may be on the reader page we
2862	 * start with the next page and check the end loop there.
2863	 */
2864	rb_inc_page(cpu_buffer, &bpage);
2865	start = bpage;
2866	do {
2867		if (bpage->page == (void *)addr) {
2868			local_dec(&bpage->entries);
2869			return;
2870		}
2871		rb_inc_page(cpu_buffer, &bpage);
2872	} while (bpage != start);
2873
2874	/* commit not part of this buffer?? */
2875	RB_WARN_ON(cpu_buffer, 1);
2876}
2877
2878/**
2879 * ring_buffer_commit_discard - discard an event that has not been committed
2880 * @buffer: the ring buffer
2881 * @event: non committed event to discard
2882 *
2883 * Sometimes an event that is in the ring buffer needs to be ignored.
2884 * This function lets the user discard an event in the ring buffer
2885 * and then that event will not be read later.
2886 *
2887 * This function only works if it is called before the the item has been
2888 * committed. It will try to free the event from the ring buffer
2889 * if another event has not been added behind it.
2890 *
2891 * If another event has been added behind it, it will set the event
2892 * up as discarded, and perform the commit.
2893 *
2894 * If this function is called, do not call ring_buffer_unlock_commit on
2895 * the event.
2896 */
2897void ring_buffer_discard_commit(struct ring_buffer *buffer,
2898				struct ring_buffer_event *event)
2899{
2900	struct ring_buffer_per_cpu *cpu_buffer;
2901	int cpu;
2902
2903	/* The event is discarded regardless */
2904	rb_event_discard(event);
2905
2906	cpu = smp_processor_id();
2907	cpu_buffer = buffer->buffers[cpu];
2908
2909	/*
2910	 * This must only be called if the event has not been
2911	 * committed yet. Thus we can assume that preemption
2912	 * is still disabled.
2913	 */
2914	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2915
2916	rb_decrement_entry(cpu_buffer, event);
2917	if (rb_try_to_discard(cpu_buffer, event))
2918		goto out;
2919
2920	/*
2921	 * The commit is still visible by the reader, so we
2922	 * must still update the timestamp.
2923	 */
2924	rb_update_write_stamp(cpu_buffer, event);
2925 out:
2926	rb_end_commit(cpu_buffer);
2927
2928	trace_recursive_unlock(cpu_buffer);
2929
2930	preempt_enable_notrace();
2931
2932}
2933EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2934
2935/**
2936 * ring_buffer_write - write data to the buffer without reserving
2937 * @buffer: The ring buffer to write to.
2938 * @length: The length of the data being written (excluding the event header)
2939 * @data: The data to write to the buffer.
2940 *
2941 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2942 * one function. If you already have the data to write to the buffer, it
2943 * may be easier to simply call this function.
2944 *
2945 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2946 * and not the length of the event which would hold the header.
2947 */
2948int ring_buffer_write(struct ring_buffer *buffer,
2949		      unsigned long length,
2950		      void *data)
2951{
2952	struct ring_buffer_per_cpu *cpu_buffer;
2953	struct ring_buffer_event *event;
2954	void *body;
2955	int ret = -EBUSY;
2956	int cpu;
2957
 
 
 
2958	preempt_disable_notrace();
2959
2960	if (atomic_read(&buffer->record_disabled))
2961		goto out;
2962
2963	cpu = raw_smp_processor_id();
2964
2965	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2966		goto out;
2967
2968	cpu_buffer = buffer->buffers[cpu];
2969
2970	if (atomic_read(&cpu_buffer->record_disabled))
2971		goto out;
2972
2973	if (length > BUF_MAX_DATA_SIZE)
2974		goto out;
2975
2976	if (unlikely(trace_recursive_lock(cpu_buffer)))
2977		goto out;
2978
2979	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2980	if (!event)
2981		goto out_unlock;
2982
2983	body = rb_event_data(event);
2984
2985	memcpy(body, data, length);
2986
2987	rb_commit(cpu_buffer, event);
2988
2989	rb_wakeups(buffer, cpu_buffer);
2990
2991	ret = 0;
2992
2993 out_unlock:
2994	trace_recursive_unlock(cpu_buffer);
2995
2996 out:
2997	preempt_enable_notrace();
2998
2999	return ret;
3000}
3001EXPORT_SYMBOL_GPL(ring_buffer_write);
3002
3003static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3004{
3005	struct buffer_page *reader = cpu_buffer->reader_page;
3006	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3007	struct buffer_page *commit = cpu_buffer->commit_page;
3008
3009	/* In case of error, head will be NULL */
3010	if (unlikely(!head))
3011		return true;
3012
3013	return reader->read == rb_page_commit(reader) &&
3014		(commit == reader ||
3015		 (commit == head &&
3016		  head->read == rb_page_commit(commit)));
3017}
3018
3019/**
3020 * ring_buffer_record_disable - stop all writes into the buffer
3021 * @buffer: The ring buffer to stop writes to.
3022 *
3023 * This prevents all writes to the buffer. Any attempt to write
3024 * to the buffer after this will fail and return NULL.
3025 *
3026 * The caller should call synchronize_sched() after this.
3027 */
3028void ring_buffer_record_disable(struct ring_buffer *buffer)
3029{
3030	atomic_inc(&buffer->record_disabled);
3031}
3032EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3033
3034/**
3035 * ring_buffer_record_enable - enable writes to the buffer
3036 * @buffer: The ring buffer to enable writes
3037 *
3038 * Note, multiple disables will need the same number of enables
3039 * to truly enable the writing (much like preempt_disable).
3040 */
3041void ring_buffer_record_enable(struct ring_buffer *buffer)
3042{
3043	atomic_dec(&buffer->record_disabled);
3044}
3045EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3046
3047/**
3048 * ring_buffer_record_off - stop all writes into the buffer
3049 * @buffer: The ring buffer to stop writes to.
3050 *
3051 * This prevents all writes to the buffer. Any attempt to write
3052 * to the buffer after this will fail and return NULL.
3053 *
3054 * This is different than ring_buffer_record_disable() as
3055 * it works like an on/off switch, where as the disable() version
3056 * must be paired with a enable().
3057 */
3058void ring_buffer_record_off(struct ring_buffer *buffer)
3059{
3060	unsigned int rd;
3061	unsigned int new_rd;
3062
3063	do {
3064		rd = atomic_read(&buffer->record_disabled);
3065		new_rd = rd | RB_BUFFER_OFF;
3066	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3067}
3068EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3069
3070/**
3071 * ring_buffer_record_on - restart writes into the buffer
3072 * @buffer: The ring buffer to start writes to.
3073 *
3074 * This enables all writes to the buffer that was disabled by
3075 * ring_buffer_record_off().
3076 *
3077 * This is different than ring_buffer_record_enable() as
3078 * it works like an on/off switch, where as the enable() version
3079 * must be paired with a disable().
3080 */
3081void ring_buffer_record_on(struct ring_buffer *buffer)
3082{
3083	unsigned int rd;
3084	unsigned int new_rd;
3085
3086	do {
3087		rd = atomic_read(&buffer->record_disabled);
3088		new_rd = rd & ~RB_BUFFER_OFF;
3089	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3090}
3091EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3092
3093/**
3094 * ring_buffer_record_is_on - return true if the ring buffer can write
3095 * @buffer: The ring buffer to see if write is enabled
3096 *
3097 * Returns true if the ring buffer is in a state that it accepts writes.
3098 */
3099int ring_buffer_record_is_on(struct ring_buffer *buffer)
3100{
3101	return !atomic_read(&buffer->record_disabled);
3102}
3103
3104/**
3105 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3106 * @buffer: The ring buffer to stop writes to.
3107 * @cpu: The CPU buffer to stop
3108 *
3109 * This prevents all writes to the buffer. Any attempt to write
3110 * to the buffer after this will fail and return NULL.
3111 *
3112 * The caller should call synchronize_sched() after this.
3113 */
3114void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3115{
3116	struct ring_buffer_per_cpu *cpu_buffer;
3117
3118	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3119		return;
3120
3121	cpu_buffer = buffer->buffers[cpu];
3122	atomic_inc(&cpu_buffer->record_disabled);
3123}
3124EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3125
3126/**
3127 * ring_buffer_record_enable_cpu - enable writes to the buffer
3128 * @buffer: The ring buffer to enable writes
3129 * @cpu: The CPU to enable.
3130 *
3131 * Note, multiple disables will need the same number of enables
3132 * to truly enable the writing (much like preempt_disable).
3133 */
3134void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3135{
3136	struct ring_buffer_per_cpu *cpu_buffer;
3137
3138	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3139		return;
3140
3141	cpu_buffer = buffer->buffers[cpu];
3142	atomic_dec(&cpu_buffer->record_disabled);
3143}
3144EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3145
3146/*
3147 * The total entries in the ring buffer is the running counter
3148 * of entries entered into the ring buffer, minus the sum of
3149 * the entries read from the ring buffer and the number of
3150 * entries that were overwritten.
3151 */
3152static inline unsigned long
3153rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3154{
3155	return local_read(&cpu_buffer->entries) -
3156		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3157}
3158
3159/**
3160 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3161 * @buffer: The ring buffer
3162 * @cpu: The per CPU buffer to read from.
3163 */
3164u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3165{
3166	unsigned long flags;
3167	struct ring_buffer_per_cpu *cpu_buffer;
3168	struct buffer_page *bpage;
3169	u64 ret = 0;
3170
3171	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3172		return 0;
3173
3174	cpu_buffer = buffer->buffers[cpu];
3175	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3176	/*
3177	 * if the tail is on reader_page, oldest time stamp is on the reader
3178	 * page
3179	 */
3180	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3181		bpage = cpu_buffer->reader_page;
3182	else
3183		bpage = rb_set_head_page(cpu_buffer);
3184	if (bpage)
3185		ret = bpage->page->time_stamp;
3186	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3187
3188	return ret;
3189}
3190EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3191
3192/**
3193 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3194 * @buffer: The ring buffer
3195 * @cpu: The per CPU buffer to read from.
3196 */
3197unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3198{
3199	struct ring_buffer_per_cpu *cpu_buffer;
3200	unsigned long ret;
3201
3202	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3203		return 0;
3204
3205	cpu_buffer = buffer->buffers[cpu];
3206	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3207
3208	return ret;
3209}
3210EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3211
3212/**
3213 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3214 * @buffer: The ring buffer
3215 * @cpu: The per CPU buffer to get the entries from.
3216 */
3217unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3218{
3219	struct ring_buffer_per_cpu *cpu_buffer;
3220
3221	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3222		return 0;
3223
3224	cpu_buffer = buffer->buffers[cpu];
3225
3226	return rb_num_of_entries(cpu_buffer);
3227}
3228EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3229
3230/**
3231 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3232 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3233 * @buffer: The ring buffer
3234 * @cpu: The per CPU buffer to get the number of overruns from
3235 */
3236unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3237{
3238	struct ring_buffer_per_cpu *cpu_buffer;
3239	unsigned long ret;
3240
3241	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3242		return 0;
3243
3244	cpu_buffer = buffer->buffers[cpu];
3245	ret = local_read(&cpu_buffer->overrun);
3246
3247	return ret;
3248}
3249EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3250
3251/**
3252 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3253 * commits failing due to the buffer wrapping around while there are uncommitted
3254 * events, such as during an interrupt storm.
3255 * @buffer: The ring buffer
3256 * @cpu: The per CPU buffer to get the number of overruns from
3257 */
3258unsigned long
3259ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3260{
3261	struct ring_buffer_per_cpu *cpu_buffer;
3262	unsigned long ret;
3263
3264	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3265		return 0;
3266
3267	cpu_buffer = buffer->buffers[cpu];
3268	ret = local_read(&cpu_buffer->commit_overrun);
3269
3270	return ret;
3271}
3272EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3273
3274/**
3275 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3276 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3277 * @buffer: The ring buffer
3278 * @cpu: The per CPU buffer to get the number of overruns from
3279 */
3280unsigned long
3281ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3282{
3283	struct ring_buffer_per_cpu *cpu_buffer;
3284	unsigned long ret;
3285
3286	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3287		return 0;
3288
3289	cpu_buffer = buffer->buffers[cpu];
3290	ret = local_read(&cpu_buffer->dropped_events);
3291
3292	return ret;
3293}
3294EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3295
3296/**
3297 * ring_buffer_read_events_cpu - get the number of events successfully read
3298 * @buffer: The ring buffer
3299 * @cpu: The per CPU buffer to get the number of events read
3300 */
3301unsigned long
3302ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3303{
3304	struct ring_buffer_per_cpu *cpu_buffer;
3305
3306	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3307		return 0;
3308
3309	cpu_buffer = buffer->buffers[cpu];
3310	return cpu_buffer->read;
3311}
3312EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3313
3314/**
3315 * ring_buffer_entries - get the number of entries in a buffer
3316 * @buffer: The ring buffer
3317 *
3318 * Returns the total number of entries in the ring buffer
3319 * (all CPU entries)
3320 */
3321unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3322{
3323	struct ring_buffer_per_cpu *cpu_buffer;
3324	unsigned long entries = 0;
3325	int cpu;
3326
3327	/* if you care about this being correct, lock the buffer */
3328	for_each_buffer_cpu(buffer, cpu) {
3329		cpu_buffer = buffer->buffers[cpu];
3330		entries += rb_num_of_entries(cpu_buffer);
3331	}
3332
3333	return entries;
3334}
3335EXPORT_SYMBOL_GPL(ring_buffer_entries);
3336
3337/**
3338 * ring_buffer_overruns - get the number of overruns in buffer
3339 * @buffer: The ring buffer
3340 *
3341 * Returns the total number of overruns in the ring buffer
3342 * (all CPU entries)
3343 */
3344unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3345{
3346	struct ring_buffer_per_cpu *cpu_buffer;
3347	unsigned long overruns = 0;
3348	int cpu;
3349
3350	/* if you care about this being correct, lock the buffer */
3351	for_each_buffer_cpu(buffer, cpu) {
3352		cpu_buffer = buffer->buffers[cpu];
3353		overruns += local_read(&cpu_buffer->overrun);
3354	}
3355
3356	return overruns;
3357}
3358EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3359
3360static void rb_iter_reset(struct ring_buffer_iter *iter)
3361{
3362	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3363
3364	/* Iterator usage is expected to have record disabled */
3365	iter->head_page = cpu_buffer->reader_page;
3366	iter->head = cpu_buffer->reader_page->read;
3367
3368	iter->cache_reader_page = iter->head_page;
3369	iter->cache_read = cpu_buffer->read;
3370
 
 
 
3371	if (iter->head)
3372		iter->read_stamp = cpu_buffer->read_stamp;
3373	else
3374		iter->read_stamp = iter->head_page->page->time_stamp;
 
 
3375}
3376
3377/**
3378 * ring_buffer_iter_reset - reset an iterator
3379 * @iter: The iterator to reset
3380 *
3381 * Resets the iterator, so that it will start from the beginning
3382 * again.
3383 */
3384void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3385{
3386	struct ring_buffer_per_cpu *cpu_buffer;
3387	unsigned long flags;
3388
3389	if (!iter)
3390		return;
3391
3392	cpu_buffer = iter->cpu_buffer;
3393
3394	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3395	rb_iter_reset(iter);
3396	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3397}
3398EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3399
3400/**
3401 * ring_buffer_iter_empty - check if an iterator has no more to read
3402 * @iter: The iterator to check
3403 */
3404int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3405{
3406	struct ring_buffer_per_cpu *cpu_buffer;
3407
3408	cpu_buffer = iter->cpu_buffer;
3409
3410	return iter->head_page == cpu_buffer->commit_page &&
3411		iter->head == rb_commit_index(cpu_buffer);
3412}
3413EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3414
3415static void
3416rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3417		     struct ring_buffer_event *event)
3418{
3419	u64 delta;
3420
3421	switch (event->type_len) {
3422	case RINGBUF_TYPE_PADDING:
3423		return;
3424
3425	case RINGBUF_TYPE_TIME_EXTEND:
3426		delta = event->array[0];
3427		delta <<= TS_SHIFT;
3428		delta += event->time_delta;
3429		cpu_buffer->read_stamp += delta;
3430		return;
3431
3432	case RINGBUF_TYPE_TIME_STAMP:
3433		/* FIXME: not implemented */
3434		return;
3435
3436	case RINGBUF_TYPE_DATA:
3437		cpu_buffer->read_stamp += event->time_delta;
3438		return;
3439
3440	default:
3441		BUG();
3442	}
3443	return;
3444}
3445
3446static void
3447rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3448			  struct ring_buffer_event *event)
3449{
3450	u64 delta;
3451
3452	switch (event->type_len) {
3453	case RINGBUF_TYPE_PADDING:
3454		return;
3455
3456	case RINGBUF_TYPE_TIME_EXTEND:
3457		delta = event->array[0];
3458		delta <<= TS_SHIFT;
3459		delta += event->time_delta;
3460		iter->read_stamp += delta;
3461		return;
3462
3463	case RINGBUF_TYPE_TIME_STAMP:
3464		/* FIXME: not implemented */
3465		return;
3466
3467	case RINGBUF_TYPE_DATA:
3468		iter->read_stamp += event->time_delta;
3469		return;
3470
3471	default:
3472		BUG();
3473	}
3474	return;
3475}
3476
3477static struct buffer_page *
3478rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3479{
3480	struct buffer_page *reader = NULL;
3481	unsigned long overwrite;
3482	unsigned long flags;
3483	int nr_loops = 0;
3484	int ret;
3485
3486	local_irq_save(flags);
3487	arch_spin_lock(&cpu_buffer->lock);
3488
3489 again:
3490	/*
3491	 * This should normally only loop twice. But because the
3492	 * start of the reader inserts an empty page, it causes
3493	 * a case where we will loop three times. There should be no
3494	 * reason to loop four times (that I know of).
3495	 */
3496	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3497		reader = NULL;
3498		goto out;
3499	}
3500
3501	reader = cpu_buffer->reader_page;
3502
3503	/* If there's more to read, return this page */
3504	if (cpu_buffer->reader_page->read < rb_page_size(reader))
3505		goto out;
3506
3507	/* Never should we have an index greater than the size */
3508	if (RB_WARN_ON(cpu_buffer,
3509		       cpu_buffer->reader_page->read > rb_page_size(reader)))
3510		goto out;
3511
3512	/* check if we caught up to the tail */
3513	reader = NULL;
3514	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3515		goto out;
3516
3517	/* Don't bother swapping if the ring buffer is empty */
3518	if (rb_num_of_entries(cpu_buffer) == 0)
3519		goto out;
3520
3521	/*
3522	 * Reset the reader page to size zero.
3523	 */
3524	local_set(&cpu_buffer->reader_page->write, 0);
3525	local_set(&cpu_buffer->reader_page->entries, 0);
3526	local_set(&cpu_buffer->reader_page->page->commit, 0);
3527	cpu_buffer->reader_page->real_end = 0;
3528
3529 spin:
3530	/*
3531	 * Splice the empty reader page into the list around the head.
3532	 */
3533	reader = rb_set_head_page(cpu_buffer);
3534	if (!reader)
3535		goto out;
3536	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3537	cpu_buffer->reader_page->list.prev = reader->list.prev;
3538
3539	/*
3540	 * cpu_buffer->pages just needs to point to the buffer, it
3541	 *  has no specific buffer page to point to. Lets move it out
3542	 *  of our way so we don't accidentally swap it.
3543	 */
3544	cpu_buffer->pages = reader->list.prev;
3545
3546	/* The reader page will be pointing to the new head */
3547	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3548
3549	/*
3550	 * We want to make sure we read the overruns after we set up our
3551	 * pointers to the next object. The writer side does a
3552	 * cmpxchg to cross pages which acts as the mb on the writer
3553	 * side. Note, the reader will constantly fail the swap
3554	 * while the writer is updating the pointers, so this
3555	 * guarantees that the overwrite recorded here is the one we
3556	 * want to compare with the last_overrun.
3557	 */
3558	smp_mb();
3559	overwrite = local_read(&(cpu_buffer->overrun));
3560
3561	/*
3562	 * Here's the tricky part.
3563	 *
3564	 * We need to move the pointer past the header page.
3565	 * But we can only do that if a writer is not currently
3566	 * moving it. The page before the header page has the
3567	 * flag bit '1' set if it is pointing to the page we want.
3568	 * but if the writer is in the process of moving it
3569	 * than it will be '2' or already moved '0'.
3570	 */
3571
3572	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3573
3574	/*
3575	 * If we did not convert it, then we must try again.
3576	 */
3577	if (!ret)
3578		goto spin;
3579
3580	/*
3581	 * Yeah! We succeeded in replacing the page.
3582	 *
3583	 * Now make the new head point back to the reader page.
3584	 */
3585	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3586	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3587
3588	/* Finally update the reader page to the new head */
3589	cpu_buffer->reader_page = reader;
3590	cpu_buffer->reader_page->read = 0;
3591
3592	if (overwrite != cpu_buffer->last_overrun) {
3593		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3594		cpu_buffer->last_overrun = overwrite;
3595	}
3596
3597	goto again;
3598
3599 out:
3600	/* Update the read_stamp on the first event */
3601	if (reader && reader->read == 0)
3602		cpu_buffer->read_stamp = reader->page->time_stamp;
3603
3604	arch_spin_unlock(&cpu_buffer->lock);
3605	local_irq_restore(flags);
3606
3607	return reader;
3608}
3609
3610static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3611{
3612	struct ring_buffer_event *event;
3613	struct buffer_page *reader;
3614	unsigned length;
3615
3616	reader = rb_get_reader_page(cpu_buffer);
3617
3618	/* This function should not be called when buffer is empty */
3619	if (RB_WARN_ON(cpu_buffer, !reader))
3620		return;
3621
3622	event = rb_reader_event(cpu_buffer);
3623
3624	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3625		cpu_buffer->read++;
3626
3627	rb_update_read_stamp(cpu_buffer, event);
3628
3629	length = rb_event_length(event);
3630	cpu_buffer->reader_page->read += length;
3631}
3632
3633static void rb_advance_iter(struct ring_buffer_iter *iter)
3634{
3635	struct ring_buffer_per_cpu *cpu_buffer;
3636	struct ring_buffer_event *event;
3637	unsigned length;
3638
3639	cpu_buffer = iter->cpu_buffer;
3640
3641	/*
3642	 * Check if we are at the end of the buffer.
3643	 */
3644	if (iter->head >= rb_page_size(iter->head_page)) {
3645		/* discarded commits can make the page empty */
3646		if (iter->head_page == cpu_buffer->commit_page)
3647			return;
3648		rb_inc_iter(iter);
3649		return;
3650	}
3651
3652	event = rb_iter_head_event(iter);
3653
3654	length = rb_event_length(event);
3655
3656	/*
3657	 * This should not be called to advance the header if we are
3658	 * at the tail of the buffer.
3659	 */
3660	if (RB_WARN_ON(cpu_buffer,
3661		       (iter->head_page == cpu_buffer->commit_page) &&
3662		       (iter->head + length > rb_commit_index(cpu_buffer))))
3663		return;
3664
3665	rb_update_iter_read_stamp(iter, event);
3666
3667	iter->head += length;
3668
3669	/* check for end of page padding */
3670	if ((iter->head >= rb_page_size(iter->head_page)) &&
3671	    (iter->head_page != cpu_buffer->commit_page))
3672		rb_inc_iter(iter);
3673}
3674
3675static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3676{
3677	return cpu_buffer->lost_events;
3678}
3679
3680static struct ring_buffer_event *
3681rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3682	       unsigned long *lost_events)
3683{
3684	struct ring_buffer_event *event;
3685	struct buffer_page *reader;
3686	int nr_loops = 0;
3687
3688 again:
3689	/*
3690	 * We repeat when a time extend is encountered.
3691	 * Since the time extend is always attached to a data event,
3692	 * we should never loop more than once.
3693	 * (We never hit the following condition more than twice).
3694	 */
3695	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3696		return NULL;
3697
3698	reader = rb_get_reader_page(cpu_buffer);
3699	if (!reader)
3700		return NULL;
3701
3702	event = rb_reader_event(cpu_buffer);
3703
3704	switch (event->type_len) {
3705	case RINGBUF_TYPE_PADDING:
3706		if (rb_null_event(event))
3707			RB_WARN_ON(cpu_buffer, 1);
3708		/*
3709		 * Because the writer could be discarding every
3710		 * event it creates (which would probably be bad)
3711		 * if we were to go back to "again" then we may never
3712		 * catch up, and will trigger the warn on, or lock
3713		 * the box. Return the padding, and we will release
3714		 * the current locks, and try again.
3715		 */
3716		return event;
3717
3718	case RINGBUF_TYPE_TIME_EXTEND:
3719		/* Internal data, OK to advance */
3720		rb_advance_reader(cpu_buffer);
3721		goto again;
3722
3723	case RINGBUF_TYPE_TIME_STAMP:
3724		/* FIXME: not implemented */
3725		rb_advance_reader(cpu_buffer);
3726		goto again;
3727
3728	case RINGBUF_TYPE_DATA:
3729		if (ts) {
3730			*ts = cpu_buffer->read_stamp + event->time_delta;
3731			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3732							 cpu_buffer->cpu, ts);
3733		}
3734		if (lost_events)
3735			*lost_events = rb_lost_events(cpu_buffer);
3736		return event;
3737
3738	default:
3739		BUG();
3740	}
3741
3742	return NULL;
3743}
3744EXPORT_SYMBOL_GPL(ring_buffer_peek);
3745
3746static struct ring_buffer_event *
3747rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3748{
3749	struct ring_buffer *buffer;
3750	struct ring_buffer_per_cpu *cpu_buffer;
3751	struct ring_buffer_event *event;
3752	int nr_loops = 0;
3753
3754	cpu_buffer = iter->cpu_buffer;
3755	buffer = cpu_buffer->buffer;
3756
3757	/*
3758	 * Check if someone performed a consuming read to
3759	 * the buffer. A consuming read invalidates the iterator
3760	 * and we need to reset the iterator in this case.
3761	 */
3762	if (unlikely(iter->cache_read != cpu_buffer->read ||
3763		     iter->cache_reader_page != cpu_buffer->reader_page))
3764		rb_iter_reset(iter);
3765
3766 again:
3767	if (ring_buffer_iter_empty(iter))
3768		return NULL;
3769
3770	/*
3771	 * We repeat when a time extend is encountered or we hit
3772	 * the end of the page. Since the time extend is always attached
3773	 * to a data event, we should never loop more than three times.
3774	 * Once for going to next page, once on time extend, and
3775	 * finally once to get the event.
3776	 * (We never hit the following condition more than thrice).
3777	 */
3778	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3779		return NULL;
3780
3781	if (rb_per_cpu_empty(cpu_buffer))
3782		return NULL;
3783
3784	if (iter->head >= rb_page_size(iter->head_page)) {
3785		rb_inc_iter(iter);
3786		goto again;
3787	}
3788
3789	event = rb_iter_head_event(iter);
3790
3791	switch (event->type_len) {
3792	case RINGBUF_TYPE_PADDING:
3793		if (rb_null_event(event)) {
3794			rb_inc_iter(iter);
3795			goto again;
3796		}
3797		rb_advance_iter(iter);
3798		return event;
3799
3800	case RINGBUF_TYPE_TIME_EXTEND:
3801		/* Internal data, OK to advance */
3802		rb_advance_iter(iter);
3803		goto again;
3804
3805	case RINGBUF_TYPE_TIME_STAMP:
3806		/* FIXME: not implemented */
3807		rb_advance_iter(iter);
3808		goto again;
3809
3810	case RINGBUF_TYPE_DATA:
3811		if (ts) {
3812			*ts = iter->read_stamp + event->time_delta;
3813			ring_buffer_normalize_time_stamp(buffer,
3814							 cpu_buffer->cpu, ts);
3815		}
3816		return event;
3817
3818	default:
3819		BUG();
3820	}
3821
3822	return NULL;
3823}
3824EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3825
3826static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3827{
3828	if (likely(!in_nmi())) {
3829		raw_spin_lock(&cpu_buffer->reader_lock);
3830		return true;
3831	}
3832
3833	/*
3834	 * If an NMI die dumps out the content of the ring buffer
3835	 * trylock must be used to prevent a deadlock if the NMI
3836	 * preempted a task that holds the ring buffer locks. If
3837	 * we get the lock then all is fine, if not, then continue
3838	 * to do the read, but this can corrupt the ring buffer,
3839	 * so it must be permanently disabled from future writes.
3840	 * Reading from NMI is a oneshot deal.
3841	 */
3842	if (raw_spin_trylock(&cpu_buffer->reader_lock))
3843		return true;
3844
3845	/* Continue without locking, but disable the ring buffer */
3846	atomic_inc(&cpu_buffer->record_disabled);
3847	return false;
3848}
3849
3850static inline void
3851rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3852{
3853	if (likely(locked))
3854		raw_spin_unlock(&cpu_buffer->reader_lock);
3855	return;
3856}
3857
3858/**
3859 * ring_buffer_peek - peek at the next event to be read
3860 * @buffer: The ring buffer to read
3861 * @cpu: The cpu to peak at
3862 * @ts: The timestamp counter of this event.
3863 * @lost_events: a variable to store if events were lost (may be NULL)
3864 *
3865 * This will return the event that will be read next, but does
3866 * not consume the data.
3867 */
3868struct ring_buffer_event *
3869ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3870		 unsigned long *lost_events)
3871{
3872	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3873	struct ring_buffer_event *event;
3874	unsigned long flags;
3875	bool dolock;
3876
3877	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3878		return NULL;
3879
 
3880 again:
3881	local_irq_save(flags);
3882	dolock = rb_reader_lock(cpu_buffer);
 
3883	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3884	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3885		rb_advance_reader(cpu_buffer);
3886	rb_reader_unlock(cpu_buffer, dolock);
 
3887	local_irq_restore(flags);
3888
3889	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3890		goto again;
3891
3892	return event;
3893}
3894
3895/**
3896 * ring_buffer_iter_peek - peek at the next event to be read
3897 * @iter: The ring buffer iterator
3898 * @ts: The timestamp counter of this event.
3899 *
3900 * This will return the event that will be read next, but does
3901 * not increment the iterator.
3902 */
3903struct ring_buffer_event *
3904ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3905{
3906	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3907	struct ring_buffer_event *event;
3908	unsigned long flags;
3909
3910 again:
3911	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3912	event = rb_iter_peek(iter, ts);
3913	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3914
3915	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3916		goto again;
3917
3918	return event;
3919}
3920
3921/**
3922 * ring_buffer_consume - return an event and consume it
3923 * @buffer: The ring buffer to get the next event from
3924 * @cpu: the cpu to read the buffer from
3925 * @ts: a variable to store the timestamp (may be NULL)
3926 * @lost_events: a variable to store if events were lost (may be NULL)
3927 *
3928 * Returns the next event in the ring buffer, and that event is consumed.
3929 * Meaning, that sequential reads will keep returning a different event,
3930 * and eventually empty the ring buffer if the producer is slower.
3931 */
3932struct ring_buffer_event *
3933ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3934		    unsigned long *lost_events)
3935{
3936	struct ring_buffer_per_cpu *cpu_buffer;
3937	struct ring_buffer_event *event = NULL;
3938	unsigned long flags;
3939	bool dolock;
 
 
3940
3941 again:
3942	/* might be called in atomic */
3943	preempt_disable();
3944
3945	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3946		goto out;
3947
3948	cpu_buffer = buffer->buffers[cpu];
3949	local_irq_save(flags);
3950	dolock = rb_reader_lock(cpu_buffer);
 
3951
3952	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3953	if (event) {
3954		cpu_buffer->lost_events = 0;
3955		rb_advance_reader(cpu_buffer);
3956	}
3957
3958	rb_reader_unlock(cpu_buffer, dolock);
 
3959	local_irq_restore(flags);
3960
3961 out:
3962	preempt_enable();
3963
3964	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3965		goto again;
3966
3967	return event;
3968}
3969EXPORT_SYMBOL_GPL(ring_buffer_consume);
3970
3971/**
3972 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3973 * @buffer: The ring buffer to read from
3974 * @cpu: The cpu buffer to iterate over
3975 *
3976 * This performs the initial preparations necessary to iterate
3977 * through the buffer.  Memory is allocated, buffer recording
3978 * is disabled, and the iterator pointer is returned to the caller.
3979 *
3980 * Disabling buffer recordng prevents the reading from being
3981 * corrupted. This is not a consuming read, so a producer is not
3982 * expected.
3983 *
3984 * After a sequence of ring_buffer_read_prepare calls, the user is
3985 * expected to make at least one call to ring_buffer_read_prepare_sync.
3986 * Afterwards, ring_buffer_read_start is invoked to get things going
3987 * for real.
3988 *
3989 * This overall must be paired with ring_buffer_read_finish.
3990 */
3991struct ring_buffer_iter *
3992ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3993{
3994	struct ring_buffer_per_cpu *cpu_buffer;
3995	struct ring_buffer_iter *iter;
3996
3997	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3998		return NULL;
3999
4000	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4001	if (!iter)
4002		return NULL;
4003
4004	cpu_buffer = buffer->buffers[cpu];
4005
4006	iter->cpu_buffer = cpu_buffer;
4007
4008	atomic_inc(&buffer->resize_disabled);
4009	atomic_inc(&cpu_buffer->record_disabled);
4010
4011	return iter;
4012}
4013EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4014
4015/**
4016 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4017 *
4018 * All previously invoked ring_buffer_read_prepare calls to prepare
4019 * iterators will be synchronized.  Afterwards, read_buffer_read_start
4020 * calls on those iterators are allowed.
4021 */
4022void
4023ring_buffer_read_prepare_sync(void)
4024{
4025	synchronize_sched();
4026}
4027EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4028
4029/**
4030 * ring_buffer_read_start - start a non consuming read of the buffer
4031 * @iter: The iterator returned by ring_buffer_read_prepare
4032 *
4033 * This finalizes the startup of an iteration through the buffer.
4034 * The iterator comes from a call to ring_buffer_read_prepare and
4035 * an intervening ring_buffer_read_prepare_sync must have been
4036 * performed.
4037 *
4038 * Must be paired with ring_buffer_read_finish.
4039 */
4040void
4041ring_buffer_read_start(struct ring_buffer_iter *iter)
4042{
4043	struct ring_buffer_per_cpu *cpu_buffer;
4044	unsigned long flags;
4045
4046	if (!iter)
4047		return;
4048
4049	cpu_buffer = iter->cpu_buffer;
4050
4051	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4052	arch_spin_lock(&cpu_buffer->lock);
4053	rb_iter_reset(iter);
4054	arch_spin_unlock(&cpu_buffer->lock);
4055	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4056}
4057EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4058
4059/**
4060 * ring_buffer_read_finish - finish reading the iterator of the buffer
4061 * @iter: The iterator retrieved by ring_buffer_start
4062 *
4063 * This re-enables the recording to the buffer, and frees the
4064 * iterator.
4065 */
4066void
4067ring_buffer_read_finish(struct ring_buffer_iter *iter)
4068{
4069	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4070	unsigned long flags;
4071
4072	/*
4073	 * Ring buffer is disabled from recording, here's a good place
4074	 * to check the integrity of the ring buffer.
4075	 * Must prevent readers from trying to read, as the check
4076	 * clears the HEAD page and readers require it.
4077	 */
4078	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4079	rb_check_pages(cpu_buffer);
4080	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4081
4082	atomic_dec(&cpu_buffer->record_disabled);
4083	atomic_dec(&cpu_buffer->buffer->resize_disabled);
4084	kfree(iter);
4085}
4086EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4087
4088/**
4089 * ring_buffer_read - read the next item in the ring buffer by the iterator
4090 * @iter: The ring buffer iterator
4091 * @ts: The time stamp of the event read.
4092 *
4093 * This reads the next event in the ring buffer and increments the iterator.
4094 */
4095struct ring_buffer_event *
4096ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4097{
4098	struct ring_buffer_event *event;
4099	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4100	unsigned long flags;
4101
4102	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4103 again:
4104	event = rb_iter_peek(iter, ts);
4105	if (!event)
4106		goto out;
4107
4108	if (event->type_len == RINGBUF_TYPE_PADDING)
4109		goto again;
4110
4111	rb_advance_iter(iter);
4112 out:
4113	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4114
4115	return event;
4116}
4117EXPORT_SYMBOL_GPL(ring_buffer_read);
4118
4119/**
4120 * ring_buffer_size - return the size of the ring buffer (in bytes)
4121 * @buffer: The ring buffer.
4122 */
4123unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4124{
4125	/*
4126	 * Earlier, this method returned
4127	 *	BUF_PAGE_SIZE * buffer->nr_pages
4128	 * Since the nr_pages field is now removed, we have converted this to
4129	 * return the per cpu buffer value.
4130	 */
4131	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4132		return 0;
4133
4134	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4135}
4136EXPORT_SYMBOL_GPL(ring_buffer_size);
4137
4138static void
4139rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4140{
4141	rb_head_page_deactivate(cpu_buffer);
4142
4143	cpu_buffer->head_page
4144		= list_entry(cpu_buffer->pages, struct buffer_page, list);
4145	local_set(&cpu_buffer->head_page->write, 0);
4146	local_set(&cpu_buffer->head_page->entries, 0);
4147	local_set(&cpu_buffer->head_page->page->commit, 0);
4148
4149	cpu_buffer->head_page->read = 0;
4150
4151	cpu_buffer->tail_page = cpu_buffer->head_page;
4152	cpu_buffer->commit_page = cpu_buffer->head_page;
4153
4154	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4155	INIT_LIST_HEAD(&cpu_buffer->new_pages);
4156	local_set(&cpu_buffer->reader_page->write, 0);
4157	local_set(&cpu_buffer->reader_page->entries, 0);
4158	local_set(&cpu_buffer->reader_page->page->commit, 0);
4159	cpu_buffer->reader_page->read = 0;
4160
4161	local_set(&cpu_buffer->entries_bytes, 0);
4162	local_set(&cpu_buffer->overrun, 0);
4163	local_set(&cpu_buffer->commit_overrun, 0);
4164	local_set(&cpu_buffer->dropped_events, 0);
4165	local_set(&cpu_buffer->entries, 0);
4166	local_set(&cpu_buffer->committing, 0);
4167	local_set(&cpu_buffer->commits, 0);
4168	cpu_buffer->read = 0;
4169	cpu_buffer->read_bytes = 0;
4170
4171	cpu_buffer->write_stamp = 0;
4172	cpu_buffer->read_stamp = 0;
4173
4174	cpu_buffer->lost_events = 0;
4175	cpu_buffer->last_overrun = 0;
4176
4177	rb_head_page_activate(cpu_buffer);
4178}
4179
4180/**
4181 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4182 * @buffer: The ring buffer to reset a per cpu buffer of
4183 * @cpu: The CPU buffer to be reset
4184 */
4185void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4186{
4187	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4188	unsigned long flags;
4189
4190	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4191		return;
4192
4193	atomic_inc(&buffer->resize_disabled);
4194	atomic_inc(&cpu_buffer->record_disabled);
4195
4196	/* Make sure all commits have finished */
4197	synchronize_sched();
4198
4199	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4200
4201	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4202		goto out;
4203
4204	arch_spin_lock(&cpu_buffer->lock);
4205
4206	rb_reset_cpu(cpu_buffer);
4207
4208	arch_spin_unlock(&cpu_buffer->lock);
4209
4210 out:
4211	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4212
4213	atomic_dec(&cpu_buffer->record_disabled);
4214	atomic_dec(&buffer->resize_disabled);
4215}
4216EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4217
4218/**
4219 * ring_buffer_reset - reset a ring buffer
4220 * @buffer: The ring buffer to reset all cpu buffers
4221 */
4222void ring_buffer_reset(struct ring_buffer *buffer)
4223{
4224	int cpu;
4225
4226	for_each_buffer_cpu(buffer, cpu)
4227		ring_buffer_reset_cpu(buffer, cpu);
4228}
4229EXPORT_SYMBOL_GPL(ring_buffer_reset);
4230
4231/**
4232 * rind_buffer_empty - is the ring buffer empty?
4233 * @buffer: The ring buffer to test
4234 */
4235bool ring_buffer_empty(struct ring_buffer *buffer)
4236{
4237	struct ring_buffer_per_cpu *cpu_buffer;
4238	unsigned long flags;
4239	bool dolock;
4240	int cpu;
4241	int ret;
4242
 
 
4243	/* yes this is racy, but if you don't like the race, lock the buffer */
4244	for_each_buffer_cpu(buffer, cpu) {
4245		cpu_buffer = buffer->buffers[cpu];
4246		local_irq_save(flags);
4247		dolock = rb_reader_lock(cpu_buffer);
 
4248		ret = rb_per_cpu_empty(cpu_buffer);
4249		rb_reader_unlock(cpu_buffer, dolock);
 
4250		local_irq_restore(flags);
4251
4252		if (!ret)
4253			return false;
4254	}
4255
4256	return true;
4257}
4258EXPORT_SYMBOL_GPL(ring_buffer_empty);
4259
4260/**
4261 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4262 * @buffer: The ring buffer
4263 * @cpu: The CPU buffer to test
4264 */
4265bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4266{
4267	struct ring_buffer_per_cpu *cpu_buffer;
4268	unsigned long flags;
4269	bool dolock;
4270	int ret;
4271
4272	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4273		return true;
 
 
4274
4275	cpu_buffer = buffer->buffers[cpu];
4276	local_irq_save(flags);
4277	dolock = rb_reader_lock(cpu_buffer);
 
4278	ret = rb_per_cpu_empty(cpu_buffer);
4279	rb_reader_unlock(cpu_buffer, dolock);
 
4280	local_irq_restore(flags);
4281
4282	return ret;
4283}
4284EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4285
4286#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4287/**
4288 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4289 * @buffer_a: One buffer to swap with
4290 * @buffer_b: The other buffer to swap with
4291 *
4292 * This function is useful for tracers that want to take a "snapshot"
4293 * of a CPU buffer and has another back up buffer lying around.
4294 * it is expected that the tracer handles the cpu buffer not being
4295 * used at the moment.
4296 */
4297int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4298			 struct ring_buffer *buffer_b, int cpu)
4299{
4300	struct ring_buffer_per_cpu *cpu_buffer_a;
4301	struct ring_buffer_per_cpu *cpu_buffer_b;
4302	int ret = -EINVAL;
4303
4304	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4305	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
4306		goto out;
4307
4308	cpu_buffer_a = buffer_a->buffers[cpu];
4309	cpu_buffer_b = buffer_b->buffers[cpu];
4310
4311	/* At least make sure the two buffers are somewhat the same */
4312	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4313		goto out;
4314
4315	ret = -EAGAIN;
4316
 
 
 
4317	if (atomic_read(&buffer_a->record_disabled))
4318		goto out;
4319
4320	if (atomic_read(&buffer_b->record_disabled))
4321		goto out;
4322
4323	if (atomic_read(&cpu_buffer_a->record_disabled))
4324		goto out;
4325
4326	if (atomic_read(&cpu_buffer_b->record_disabled))
4327		goto out;
4328
4329	/*
4330	 * We can't do a synchronize_sched here because this
4331	 * function can be called in atomic context.
4332	 * Normally this will be called from the same CPU as cpu.
4333	 * If not it's up to the caller to protect this.
4334	 */
4335	atomic_inc(&cpu_buffer_a->record_disabled);
4336	atomic_inc(&cpu_buffer_b->record_disabled);
4337
4338	ret = -EBUSY;
4339	if (local_read(&cpu_buffer_a->committing))
4340		goto out_dec;
4341	if (local_read(&cpu_buffer_b->committing))
4342		goto out_dec;
4343
4344	buffer_a->buffers[cpu] = cpu_buffer_b;
4345	buffer_b->buffers[cpu] = cpu_buffer_a;
4346
4347	cpu_buffer_b->buffer = buffer_a;
4348	cpu_buffer_a->buffer = buffer_b;
4349
4350	ret = 0;
4351
4352out_dec:
4353	atomic_dec(&cpu_buffer_a->record_disabled);
4354	atomic_dec(&cpu_buffer_b->record_disabled);
4355out:
4356	return ret;
4357}
4358EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4359#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4360
4361/**
4362 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4363 * @buffer: the buffer to allocate for.
4364 * @cpu: the cpu buffer to allocate.
4365 *
4366 * This function is used in conjunction with ring_buffer_read_page.
4367 * When reading a full page from the ring buffer, these functions
4368 * can be used to speed up the process. The calling function should
4369 * allocate a few pages first with this function. Then when it
4370 * needs to get pages from the ring buffer, it passes the result
4371 * of this function into ring_buffer_read_page, which will swap
4372 * the page that was allocated, with the read page of the buffer.
4373 *
4374 * Returns:
4375 *  The page allocated, or NULL on error.
4376 */
4377void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4378{
4379	struct buffer_data_page *bpage;
4380	struct page *page;
4381
4382	page = alloc_pages_node(cpu_to_node(cpu),
4383				GFP_KERNEL | __GFP_NORETRY, 0);
4384	if (!page)
4385		return NULL;
4386
4387	bpage = page_address(page);
4388
4389	rb_init_page(bpage);
4390
4391	return bpage;
4392}
4393EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4394
4395/**
4396 * ring_buffer_free_read_page - free an allocated read page
4397 * @buffer: the buffer the page was allocate for
4398 * @data: the page to free
4399 *
4400 * Free a page allocated from ring_buffer_alloc_read_page.
4401 */
4402void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4403{
4404	free_page((unsigned long)data);
4405}
4406EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4407
4408/**
4409 * ring_buffer_read_page - extract a page from the ring buffer
4410 * @buffer: buffer to extract from
4411 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4412 * @len: amount to extract
4413 * @cpu: the cpu of the buffer to extract
4414 * @full: should the extraction only happen when the page is full.
4415 *
4416 * This function will pull out a page from the ring buffer and consume it.
4417 * @data_page must be the address of the variable that was returned
4418 * from ring_buffer_alloc_read_page. This is because the page might be used
4419 * to swap with a page in the ring buffer.
4420 *
4421 * for example:
4422 *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
4423 *	if (!rpage)
4424 *		return error;
4425 *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4426 *	if (ret >= 0)
4427 *		process_page(rpage, ret);
4428 *
4429 * When @full is set, the function will not return true unless
4430 * the writer is off the reader page.
4431 *
4432 * Note: it is up to the calling functions to handle sleeps and wakeups.
4433 *  The ring buffer can be used anywhere in the kernel and can not
4434 *  blindly call wake_up. The layer that uses the ring buffer must be
4435 *  responsible for that.
4436 *
4437 * Returns:
4438 *  >=0 if data has been transferred, returns the offset of consumed data.
4439 *  <0 if no data has been transferred.
4440 */
4441int ring_buffer_read_page(struct ring_buffer *buffer,
4442			  void **data_page, size_t len, int cpu, int full)
4443{
4444	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4445	struct ring_buffer_event *event;
4446	struct buffer_data_page *bpage;
4447	struct buffer_page *reader;
4448	unsigned long missed_events;
4449	unsigned long flags;
4450	unsigned int commit;
4451	unsigned int read;
4452	u64 save_timestamp;
4453	int ret = -1;
4454
4455	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4456		goto out;
4457
4458	/*
4459	 * If len is not big enough to hold the page header, then
4460	 * we can not copy anything.
4461	 */
4462	if (len <= BUF_PAGE_HDR_SIZE)
4463		goto out;
4464
4465	len -= BUF_PAGE_HDR_SIZE;
4466
4467	if (!data_page)
4468		goto out;
4469
4470	bpage = *data_page;
4471	if (!bpage)
4472		goto out;
4473
4474	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4475
4476	reader = rb_get_reader_page(cpu_buffer);
4477	if (!reader)
4478		goto out_unlock;
4479
4480	event = rb_reader_event(cpu_buffer);
4481
4482	read = reader->read;
4483	commit = rb_page_commit(reader);
4484
4485	/* Check if any events were dropped */
4486	missed_events = cpu_buffer->lost_events;
4487
4488	/*
4489	 * If this page has been partially read or
4490	 * if len is not big enough to read the rest of the page or
4491	 * a writer is still on the page, then
4492	 * we must copy the data from the page to the buffer.
4493	 * Otherwise, we can simply swap the page with the one passed in.
4494	 */
4495	if (read || (len < (commit - read)) ||
4496	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
4497		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4498		unsigned int rpos = read;
4499		unsigned int pos = 0;
4500		unsigned int size;
4501
4502		if (full)
4503			goto out_unlock;
4504
4505		if (len > (commit - read))
4506			len = (commit - read);
4507
4508		/* Always keep the time extend and data together */
4509		size = rb_event_ts_length(event);
4510
4511		if (len < size)
4512			goto out_unlock;
4513
4514		/* save the current timestamp, since the user will need it */
4515		save_timestamp = cpu_buffer->read_stamp;
4516
4517		/* Need to copy one event at a time */
4518		do {
4519			/* We need the size of one event, because
4520			 * rb_advance_reader only advances by one event,
4521			 * whereas rb_event_ts_length may include the size of
4522			 * one or two events.
4523			 * We have already ensured there's enough space if this
4524			 * is a time extend. */
4525			size = rb_event_length(event);
4526			memcpy(bpage->data + pos, rpage->data + rpos, size);
4527
4528			len -= size;
4529
4530			rb_advance_reader(cpu_buffer);
4531			rpos = reader->read;
4532			pos += size;
4533
4534			if (rpos >= commit)
4535				break;
4536
4537			event = rb_reader_event(cpu_buffer);
4538			/* Always keep the time extend and data together */
4539			size = rb_event_ts_length(event);
4540		} while (len >= size);
4541
4542		/* update bpage */
4543		local_set(&bpage->commit, pos);
4544		bpage->time_stamp = save_timestamp;
4545
4546		/* we copied everything to the beginning */
4547		read = 0;
4548	} else {
4549		/* update the entry counter */
4550		cpu_buffer->read += rb_page_entries(reader);
4551		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4552
4553		/* swap the pages */
4554		rb_init_page(bpage);
4555		bpage = reader->page;
4556		reader->page = *data_page;
4557		local_set(&reader->write, 0);
4558		local_set(&reader->entries, 0);
4559		reader->read = 0;
4560		*data_page = bpage;
4561
4562		/*
4563		 * Use the real_end for the data size,
4564		 * This gives us a chance to store the lost events
4565		 * on the page.
4566		 */
4567		if (reader->real_end)
4568			local_set(&bpage->commit, reader->real_end);
4569	}
4570	ret = read;
4571
4572	cpu_buffer->lost_events = 0;
4573
4574	commit = local_read(&bpage->commit);
4575	/*
4576	 * Set a flag in the commit field if we lost events
4577	 */
4578	if (missed_events) {
4579		/* If there is room at the end of the page to save the
4580		 * missed events, then record it there.
4581		 */
4582		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4583			memcpy(&bpage->data[commit], &missed_events,
4584			       sizeof(missed_events));
4585			local_add(RB_MISSED_STORED, &bpage->commit);
4586			commit += sizeof(missed_events);
4587		}
4588		local_add(RB_MISSED_EVENTS, &bpage->commit);
4589	}
4590
4591	/*
4592	 * This page may be off to user land. Zero it out here.
4593	 */
4594	if (commit < BUF_PAGE_SIZE)
4595		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4596
4597 out_unlock:
4598	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4599
4600 out:
4601	return ret;
4602}
4603EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4604
4605/*
4606 * We only allocate new buffers, never free them if the CPU goes down.
4607 * If we were to free the buffer, then the user would lose any trace that was in
4608 * the buffer.
4609 */
4610int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4611{
4612	struct ring_buffer *buffer;
4613	long nr_pages_same;
4614	int cpu_i;
4615	unsigned long nr_pages;
4616
4617	buffer = container_of(node, struct ring_buffer, node);
4618	if (cpumask_test_cpu(cpu, buffer->cpumask))
4619		return 0;
4620
4621	nr_pages = 0;
4622	nr_pages_same = 1;
4623	/* check if all cpu sizes are same */
4624	for_each_buffer_cpu(buffer, cpu_i) {
4625		/* fill in the size from first enabled cpu */
4626		if (nr_pages == 0)
4627			nr_pages = buffer->buffers[cpu_i]->nr_pages;
4628		if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4629			nr_pages_same = 0;
4630			break;
 
4631		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4632	}
4633	/* allocate minimum pages, user can later expand it */
4634	if (!nr_pages_same)
4635		nr_pages = 2;
4636	buffer->buffers[cpu] =
4637		rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4638	if (!buffer->buffers[cpu]) {
4639		WARN(1, "failed to allocate ring buffer on CPU %u\n",
4640		     cpu);
4641		return -ENOMEM;
4642	}
4643	smp_wmb();
4644	cpumask_set_cpu(cpu, buffer->cpumask);
4645	return 0;
4646}
 
4647
4648#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4649/*
4650 * This is a basic integrity check of the ring buffer.
4651 * Late in the boot cycle this test will run when configured in.
4652 * It will kick off a thread per CPU that will go into a loop
4653 * writing to the per cpu ring buffer various sizes of data.
4654 * Some of the data will be large items, some small.
4655 *
4656 * Another thread is created that goes into a spin, sending out
4657 * IPIs to the other CPUs to also write into the ring buffer.
4658 * this is to test the nesting ability of the buffer.
4659 *
4660 * Basic stats are recorded and reported. If something in the
4661 * ring buffer should happen that's not expected, a big warning
4662 * is displayed and all ring buffers are disabled.
4663 */
4664static struct task_struct *rb_threads[NR_CPUS] __initdata;
4665
4666struct rb_test_data {
4667	struct ring_buffer	*buffer;
4668	unsigned long		events;
4669	unsigned long		bytes_written;
4670	unsigned long		bytes_alloc;
4671	unsigned long		bytes_dropped;
4672	unsigned long		events_nested;
4673	unsigned long		bytes_written_nested;
4674	unsigned long		bytes_alloc_nested;
4675	unsigned long		bytes_dropped_nested;
4676	int			min_size_nested;
4677	int			max_size_nested;
4678	int			max_size;
4679	int			min_size;
4680	int			cpu;
4681	int			cnt;
4682};
4683
4684static struct rb_test_data rb_data[NR_CPUS] __initdata;
4685
4686/* 1 meg per cpu */
4687#define RB_TEST_BUFFER_SIZE	1048576
4688
4689static char rb_string[] __initdata =
4690	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4691	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4692	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4693
4694static bool rb_test_started __initdata;
4695
4696struct rb_item {
4697	int size;
4698	char str[];
4699};
4700
4701static __init int rb_write_something(struct rb_test_data *data, bool nested)
4702{
4703	struct ring_buffer_event *event;
4704	struct rb_item *item;
4705	bool started;
4706	int event_len;
4707	int size;
4708	int len;
4709	int cnt;
4710
4711	/* Have nested writes different that what is written */
4712	cnt = data->cnt + (nested ? 27 : 0);
4713
4714	/* Multiply cnt by ~e, to make some unique increment */
4715	size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4716
4717	len = size + sizeof(struct rb_item);
4718
4719	started = rb_test_started;
4720	/* read rb_test_started before checking buffer enabled */
4721	smp_rmb();
4722
4723	event = ring_buffer_lock_reserve(data->buffer, len);
4724	if (!event) {
4725		/* Ignore dropped events before test starts. */
4726		if (started) {
4727			if (nested)
4728				data->bytes_dropped += len;
4729			else
4730				data->bytes_dropped_nested += len;
4731		}
4732		return len;
4733	}
4734
4735	event_len = ring_buffer_event_length(event);
4736
4737	if (RB_WARN_ON(data->buffer, event_len < len))
4738		goto out;
4739
4740	item = ring_buffer_event_data(event);
4741	item->size = size;
4742	memcpy(item->str, rb_string, size);
4743
4744	if (nested) {
4745		data->bytes_alloc_nested += event_len;
4746		data->bytes_written_nested += len;
4747		data->events_nested++;
4748		if (!data->min_size_nested || len < data->min_size_nested)
4749			data->min_size_nested = len;
4750		if (len > data->max_size_nested)
4751			data->max_size_nested = len;
4752	} else {
4753		data->bytes_alloc += event_len;
4754		data->bytes_written += len;
4755		data->events++;
4756		if (!data->min_size || len < data->min_size)
4757			data->max_size = len;
4758		if (len > data->max_size)
4759			data->max_size = len;
4760	}
4761
4762 out:
4763	ring_buffer_unlock_commit(data->buffer, event);
4764
4765	return 0;
4766}
4767
4768static __init int rb_test(void *arg)
4769{
4770	struct rb_test_data *data = arg;
4771
4772	while (!kthread_should_stop()) {
4773		rb_write_something(data, false);
4774		data->cnt++;
4775
4776		set_current_state(TASK_INTERRUPTIBLE);
4777		/* Now sleep between a min of 100-300us and a max of 1ms */
4778		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4779	}
4780
4781	return 0;
4782}
4783
4784static __init void rb_ipi(void *ignore)
4785{
4786	struct rb_test_data *data;
4787	int cpu = smp_processor_id();
4788
4789	data = &rb_data[cpu];
4790	rb_write_something(data, true);
4791}
4792
4793static __init int rb_hammer_test(void *arg)
4794{
4795	while (!kthread_should_stop()) {
4796
4797		/* Send an IPI to all cpus to write data! */
4798		smp_call_function(rb_ipi, NULL, 1);
4799		/* No sleep, but for non preempt, let others run */
4800		schedule();
4801	}
4802
4803	return 0;
4804}
4805
4806static __init int test_ringbuffer(void)
4807{
4808	struct task_struct *rb_hammer;
4809	struct ring_buffer *buffer;
4810	int cpu;
4811	int ret = 0;
4812
4813	pr_info("Running ring buffer tests...\n");
4814
4815	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4816	if (WARN_ON(!buffer))
4817		return 0;
4818
4819	/* Disable buffer so that threads can't write to it yet */
4820	ring_buffer_record_off(buffer);
4821
4822	for_each_online_cpu(cpu) {
4823		rb_data[cpu].buffer = buffer;
4824		rb_data[cpu].cpu = cpu;
4825		rb_data[cpu].cnt = cpu;
4826		rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4827						 "rbtester/%d", cpu);
4828		if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4829			pr_cont("FAILED\n");
4830			ret = PTR_ERR(rb_threads[cpu]);
4831			goto out_free;
4832		}
4833
4834		kthread_bind(rb_threads[cpu], cpu);
4835 		wake_up_process(rb_threads[cpu]);
4836	}
4837
4838	/* Now create the rb hammer! */
4839	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4840	if (WARN_ON(IS_ERR(rb_hammer))) {
4841		pr_cont("FAILED\n");
4842		ret = PTR_ERR(rb_hammer);
4843		goto out_free;
4844	}
4845
4846	ring_buffer_record_on(buffer);
4847	/*
4848	 * Show buffer is enabled before setting rb_test_started.
4849	 * Yes there's a small race window where events could be
4850	 * dropped and the thread wont catch it. But when a ring
4851	 * buffer gets enabled, there will always be some kind of
4852	 * delay before other CPUs see it. Thus, we don't care about
4853	 * those dropped events. We care about events dropped after
4854	 * the threads see that the buffer is active.
4855	 */
4856	smp_wmb();
4857	rb_test_started = true;
4858
4859	set_current_state(TASK_INTERRUPTIBLE);
4860	/* Just run for 10 seconds */;
4861	schedule_timeout(10 * HZ);
4862
4863	kthread_stop(rb_hammer);
4864
4865 out_free:
4866	for_each_online_cpu(cpu) {
4867		if (!rb_threads[cpu])
4868			break;
4869		kthread_stop(rb_threads[cpu]);
4870	}
4871	if (ret) {
4872		ring_buffer_free(buffer);
4873		return ret;
4874	}
4875
4876	/* Report! */
4877	pr_info("finished\n");
4878	for_each_online_cpu(cpu) {
4879		struct ring_buffer_event *event;
4880		struct rb_test_data *data = &rb_data[cpu];
4881		struct rb_item *item;
4882		unsigned long total_events;
4883		unsigned long total_dropped;
4884		unsigned long total_written;
4885		unsigned long total_alloc;
4886		unsigned long total_read = 0;
4887		unsigned long total_size = 0;
4888		unsigned long total_len = 0;
4889		unsigned long total_lost = 0;
4890		unsigned long lost;
4891		int big_event_size;
4892		int small_event_size;
4893
4894		ret = -1;
4895
4896		total_events = data->events + data->events_nested;
4897		total_written = data->bytes_written + data->bytes_written_nested;
4898		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4899		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4900
4901		big_event_size = data->max_size + data->max_size_nested;
4902		small_event_size = data->min_size + data->min_size_nested;
4903
4904		pr_info("CPU %d:\n", cpu);
4905		pr_info("              events:    %ld\n", total_events);
4906		pr_info("       dropped bytes:    %ld\n", total_dropped);
4907		pr_info("       alloced bytes:    %ld\n", total_alloc);
4908		pr_info("       written bytes:    %ld\n", total_written);
4909		pr_info("       biggest event:    %d\n", big_event_size);
4910		pr_info("      smallest event:    %d\n", small_event_size);
4911
4912		if (RB_WARN_ON(buffer, total_dropped))
4913			break;
4914
4915		ret = 0;
4916
4917		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4918			total_lost += lost;
4919			item = ring_buffer_event_data(event);
4920			total_len += ring_buffer_event_length(event);
4921			total_size += item->size + sizeof(struct rb_item);
4922			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4923				pr_info("FAILED!\n");
4924				pr_info("buffer had: %.*s\n", item->size, item->str);
4925				pr_info("expected:   %.*s\n", item->size, rb_string);
4926				RB_WARN_ON(buffer, 1);
4927				ret = -1;
4928				break;
4929			}
4930			total_read++;
4931		}
4932		if (ret)
4933			break;
4934
4935		ret = -1;
4936
4937		pr_info("         read events:   %ld\n", total_read);
4938		pr_info("         lost events:   %ld\n", total_lost);
4939		pr_info("        total events:   %ld\n", total_lost + total_read);
4940		pr_info("  recorded len bytes:   %ld\n", total_len);
4941		pr_info(" recorded size bytes:   %ld\n", total_size);
4942		if (total_lost)
4943			pr_info(" With dropped events, record len and size may not match\n"
4944				" alloced and written from above\n");
4945		if (!total_lost) {
4946			if (RB_WARN_ON(buffer, total_len != total_alloc ||
4947				       total_size != total_written))
4948				break;
4949		}
4950		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4951			break;
4952
4953		ret = 0;
4954	}
4955	if (!ret)
4956		pr_info("Ring buffer PASSED!\n");
4957
4958	ring_buffer_free(buffer);
4959	return 0;
4960}
4961
4962late_initcall(test_ringbuffer);
4963#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
v3.15
   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 */