Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Apr 14-17, 2025
Register
Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Generic ring buffer
   4 *
   5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
   6 */
   7#include <linux/trace_events.h>
   8#include <linux/ring_buffer.h>
   9#include <linux/trace_clock.h>
  10#include <linux/sched/clock.h>
  11#include <linux/trace_seq.h>
  12#include <linux/spinlock.h>
  13#include <linux/irq_work.h>
 
  14#include <linux/uaccess.h>
  15#include <linux/hardirq.h>
  16#include <linux/kthread.h>	/* for self test */
 
  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/oom.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	trace_seq_puts(s, "# compressed entry header\n");
  38	trace_seq_puts(s, "\ttype_len    :    5 bits\n");
  39	trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
  40	trace_seq_puts(s, "\tarray       :   32 bits\n");
  41	trace_seq_putc(s, '\n');
  42	trace_seq_printf(s, "\tpadding     : type == %d\n",
  43			 RINGBUF_TYPE_PADDING);
  44	trace_seq_printf(s, "\ttime_extend : type == %d\n",
  45			 RINGBUF_TYPE_TIME_EXTEND);
  46	trace_seq_printf(s, "\ttime_stamp : type == %d\n",
  47			 RINGBUF_TYPE_TIME_STAMP);
  48	trace_seq_printf(s, "\tdata max type_len  == %d\n",
  49			 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
  50
  51	return !trace_seq_has_overflowed(s);
  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/* Used for individual buffers (after the counter) */
 123#define RB_BUFFER_OFF		(1 << 20)
 124
 125#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
 126
 
 
 
 
 
 
 
 
 
 
 
 127#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
 128#define RB_ALIGNMENT		4U
 129#define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 130#define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
 131#define RB_ALIGN_DATA		__aligned(RB_ALIGNMENT)
 
 
 
 
 
 
 
 
 
 132
 133/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
 134#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
 135
 136enum {
 137	RB_LEN_TIME_EXTEND = 8,
 138	RB_LEN_TIME_STAMP =  8,
 139};
 140
 141#define skip_time_extend(event) \
 142	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
 143
 144#define extended_time(event) \
 145	(event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
 146
 147static inline int rb_null_event(struct ring_buffer_event *event)
 148{
 149	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
 150}
 151
 152static void rb_event_set_padding(struct ring_buffer_event *event)
 153{
 154	/* padding has a NULL time_delta */
 155	event->type_len = RINGBUF_TYPE_PADDING;
 156	event->time_delta = 0;
 157}
 158
 159static unsigned
 160rb_event_data_length(struct ring_buffer_event *event)
 161{
 162	unsigned length;
 163
 164	if (event->type_len)
 165		length = event->type_len * RB_ALIGNMENT;
 166	else
 167		length = event->array[0];
 168	return length + RB_EVNT_HDR_SIZE;
 169}
 170
 171/*
 172 * Return the length of the given event. Will return
 173 * the length of the time extend if the event is a
 174 * time extend.
 175 */
 176static inline unsigned
 177rb_event_length(struct ring_buffer_event *event)
 178{
 179	switch (event->type_len) {
 180	case RINGBUF_TYPE_PADDING:
 181		if (rb_null_event(event))
 182			/* undefined */
 183			return -1;
 184		return  event->array[0] + RB_EVNT_HDR_SIZE;
 185
 186	case RINGBUF_TYPE_TIME_EXTEND:
 187		return RB_LEN_TIME_EXTEND;
 188
 189	case RINGBUF_TYPE_TIME_STAMP:
 190		return RB_LEN_TIME_STAMP;
 191
 192	case RINGBUF_TYPE_DATA:
 193		return rb_event_data_length(event);
 194	default:
 195		BUG();
 196	}
 197	/* not hit */
 198	return 0;
 199}
 200
 201/*
 202 * Return total length of time extend and data,
 203 *   or just the event length for all other events.
 204 */
 205static inline unsigned
 206rb_event_ts_length(struct ring_buffer_event *event)
 207{
 208	unsigned len = 0;
 209
 210	if (extended_time(event)) {
 211		/* time extends include the data event after it */
 212		len = RB_LEN_TIME_EXTEND;
 213		event = skip_time_extend(event);
 214	}
 215	return len + rb_event_length(event);
 216}
 217
 218/**
 219 * ring_buffer_event_length - return the length of the event
 220 * @event: the event to get the length of
 221 *
 222 * Returns the size of the data load of a data event.
 223 * If the event is something other than a data event, it
 224 * returns the size of the event itself. With the exception
 225 * of a TIME EXTEND, where it still returns the size of the
 226 * data load of the data event after it.
 227 */
 228unsigned ring_buffer_event_length(struct ring_buffer_event *event)
 229{
 230	unsigned length;
 231
 232	if (extended_time(event))
 233		event = skip_time_extend(event);
 234
 235	length = rb_event_length(event);
 236	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 237		return length;
 238	length -= RB_EVNT_HDR_SIZE;
 239	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
 240                length -= sizeof(event->array[0]);
 241	return length;
 242}
 243EXPORT_SYMBOL_GPL(ring_buffer_event_length);
 244
 245/* inline for ring buffer fast paths */
 246static __always_inline void *
 247rb_event_data(struct ring_buffer_event *event)
 248{
 249	if (extended_time(event))
 250		event = skip_time_extend(event);
 251	BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 252	/* If length is in len field, then array[0] has the data */
 253	if (event->type_len)
 254		return (void *)&event->array[0];
 255	/* Otherwise length is in array[0] and array[1] has the data */
 256	return (void *)&event->array[1];
 257}
 258
 259/**
 260 * ring_buffer_event_data - return the data of the event
 261 * @event: the event to get the data from
 262 */
 263void *ring_buffer_event_data(struct ring_buffer_event *event)
 264{
 265	return rb_event_data(event);
 266}
 267EXPORT_SYMBOL_GPL(ring_buffer_event_data);
 268
 269#define for_each_buffer_cpu(buffer, cpu)		\
 270	for_each_cpu(cpu, buffer->cpumask)
 271
 272#define TS_SHIFT	27
 273#define TS_MASK		((1ULL << TS_SHIFT) - 1)
 274#define TS_DELTA_TEST	(~TS_MASK)
 275
 276/**
 277 * ring_buffer_event_time_stamp - return the event's extended timestamp
 278 * @event: the event to get the timestamp of
 279 *
 280 * Returns the extended timestamp associated with a data event.
 281 * An extended time_stamp is a 64-bit timestamp represented
 282 * internally in a special way that makes the best use of space
 283 * contained within a ring buffer event.  This function decodes
 284 * it and maps it to a straight u64 value.
 285 */
 286u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
 287{
 288	u64 ts;
 289
 290	ts = event->array[0];
 291	ts <<= TS_SHIFT;
 292	ts += event->time_delta;
 293
 294	return ts;
 295}
 296
 297/* Flag when events were overwritten */
 298#define RB_MISSED_EVENTS	(1 << 31)
 299/* Missed count stored at end */
 300#define RB_MISSED_STORED	(1 << 30)
 301
 302#define RB_MISSED_FLAGS		(RB_MISSED_EVENTS|RB_MISSED_STORED)
 303
 304struct buffer_data_page {
 305	u64		 time_stamp;	/* page time stamp */
 306	local_t		 commit;	/* write committed index */
 307	unsigned char	 data[] RB_ALIGN_DATA;	/* data of buffer page */
 308};
 309
 310/*
 311 * Note, the buffer_page list must be first. The buffer pages
 312 * are allocated in cache lines, which means that each buffer
 313 * page will be at the beginning of a cache line, and thus
 314 * the least significant bits will be zero. We use this to
 315 * add flags in the list struct pointers, to make the ring buffer
 316 * lockless.
 317 */
 318struct buffer_page {
 319	struct list_head list;		/* list of buffer pages */
 320	local_t		 write;		/* index for next write */
 321	unsigned	 read;		/* index for next read */
 322	local_t		 entries;	/* entries on this page */
 323	unsigned long	 real_end;	/* real end of data */
 324	struct buffer_data_page *page;	/* Actual data page */
 325};
 326
 327/*
 328 * The buffer page counters, write and entries, must be reset
 329 * atomically when crossing page boundaries. To synchronize this
 330 * update, two counters are inserted into the number. One is
 331 * the actual counter for the write position or count on the page.
 332 *
 333 * The other is a counter of updaters. Before an update happens
 334 * the update partition of the counter is incremented. This will
 335 * allow the updater to update the counter atomically.
 336 *
 337 * The counter is 20 bits, and the state data is 12.
 338 */
 339#define RB_WRITE_MASK		0xfffff
 340#define RB_WRITE_INTCNT		(1 << 20)
 341
 342static void rb_init_page(struct buffer_data_page *bpage)
 343{
 344	local_set(&bpage->commit, 0);
 345}
 346
 
 
 
 
 
 
 
 
 
 
 
 
 347/*
 348 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
 349 * this issue out.
 350 */
 351static void free_buffer_page(struct buffer_page *bpage)
 352{
 353	free_page((unsigned long)bpage->page);
 354	kfree(bpage);
 355}
 356
 357/*
 358 * We need to fit the time_stamp delta into 27 bits.
 359 */
 360static inline int test_time_stamp(u64 delta)
 361{
 362	if (delta & TS_DELTA_TEST)
 363		return 1;
 364	return 0;
 365}
 366
 367#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
 368
 369/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
 370#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
 371
 372int ring_buffer_print_page_header(struct trace_seq *s)
 373{
 374	struct buffer_data_page field;
 
 375
 376	trace_seq_printf(s, "\tfield: u64 timestamp;\t"
 377			 "offset:0;\tsize:%u;\tsigned:%u;\n",
 378			 (unsigned int)sizeof(field.time_stamp),
 379			 (unsigned int)is_signed_type(u64));
 380
 381	trace_seq_printf(s, "\tfield: local_t commit;\t"
 382			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 383			 (unsigned int)offsetof(typeof(field), commit),
 384			 (unsigned int)sizeof(field.commit),
 385			 (unsigned int)is_signed_type(long));
 386
 387	trace_seq_printf(s, "\tfield: int overwrite;\t"
 388			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 389			 (unsigned int)offsetof(typeof(field), commit),
 390			 1,
 391			 (unsigned int)is_signed_type(long));
 392
 393	trace_seq_printf(s, "\tfield: char data;\t"
 394			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 395			 (unsigned int)offsetof(typeof(field), data),
 396			 (unsigned int)BUF_PAGE_SIZE,
 397			 (unsigned int)is_signed_type(char));
 398
 399	return !trace_seq_has_overflowed(s);
 400}
 401
 402struct rb_irq_work {
 403	struct irq_work			work;
 404	wait_queue_head_t		waiters;
 405	wait_queue_head_t		full_waiters;
 406	bool				waiters_pending;
 407	bool				full_waiters_pending;
 408	bool				wakeup_full;
 409};
 410
 411/*
 412 * Structure to hold event state and handle nested events.
 413 */
 414struct rb_event_info {
 415	u64			ts;
 416	u64			delta;
 417	unsigned long		length;
 418	struct buffer_page	*tail_page;
 419	int			add_timestamp;
 420};
 421
 422/*
 423 * Used for which event context the event is in.
 424 *  NMI     = 0
 425 *  IRQ     = 1
 426 *  SOFTIRQ = 2
 427 *  NORMAL  = 3
 428 *
 429 * See trace_recursive_lock() comment below for more details.
 430 */
 431enum {
 432	RB_CTX_NMI,
 433	RB_CTX_IRQ,
 434	RB_CTX_SOFTIRQ,
 435	RB_CTX_NORMAL,
 436	RB_CTX_MAX
 437};
 438
 439/*
 440 * head_page == tail_page && head == tail then buffer is empty.
 441 */
 442struct ring_buffer_per_cpu {
 443	int				cpu;
 444	atomic_t			record_disabled;
 445	struct ring_buffer		*buffer;
 446	raw_spinlock_t			reader_lock;	/* serialize readers */
 447	arch_spinlock_t			lock;
 448	struct lock_class_key		lock_key;
 449	struct buffer_data_page		*free_page;
 450	unsigned long			nr_pages;
 451	unsigned int			current_context;
 452	struct list_head		*pages;
 453	struct buffer_page		*head_page;	/* read from head */
 454	struct buffer_page		*tail_page;	/* write to tail */
 455	struct buffer_page		*commit_page;	/* committed pages */
 456	struct buffer_page		*reader_page;
 457	unsigned long			lost_events;
 458	unsigned long			last_overrun;
 459	unsigned long			nest;
 460	local_t				entries_bytes;
 461	local_t				entries;
 462	local_t				overrun;
 463	local_t				commit_overrun;
 464	local_t				dropped_events;
 465	local_t				committing;
 466	local_t				commits;
 467	local_t				pages_touched;
 468	local_t				pages_read;
 469	long				last_pages_touch;
 470	size_t				shortest_full;
 471	unsigned long			read;
 472	unsigned long			read_bytes;
 473	u64				write_stamp;
 474	u64				read_stamp;
 475	/* ring buffer pages to update, > 0 to add, < 0 to remove */
 476	long				nr_pages_to_update;
 477	struct list_head		new_pages; /* new pages to add */
 478	struct work_struct		update_pages_work;
 479	struct completion		update_done;
 480
 481	struct rb_irq_work		irq_work;
 482};
 483
 484struct ring_buffer {
 485	unsigned			flags;
 486	int				cpus;
 487	atomic_t			record_disabled;
 488	atomic_t			resize_disabled;
 489	cpumask_var_t			cpumask;
 490
 491	struct lock_class_key		*reader_lock_key;
 492
 493	struct mutex			mutex;
 494
 495	struct ring_buffer_per_cpu	**buffers;
 496
 497	struct hlist_node		node;
 
 
 498	u64				(*clock)(void);
 499
 500	struct rb_irq_work		irq_work;
 501	bool				time_stamp_abs;
 502};
 503
 504struct ring_buffer_iter {
 505	struct ring_buffer_per_cpu	*cpu_buffer;
 506	unsigned long			head;
 507	struct buffer_page		*head_page;
 508	struct buffer_page		*cache_reader_page;
 509	unsigned long			cache_read;
 510	u64				read_stamp;
 511};
 512
 513/**
 514 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
 515 * @buffer: The ring_buffer to get the number of pages from
 516 * @cpu: The cpu of the ring_buffer to get the number of pages from
 517 *
 518 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
 519 */
 520size_t ring_buffer_nr_pages(struct ring_buffer *buffer, int cpu)
 521{
 522	return buffer->buffers[cpu]->nr_pages;
 523}
 524
 525/**
 526 * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
 527 * @buffer: The ring_buffer to get the number of pages from
 528 * @cpu: The cpu of the ring_buffer to get the number of pages from
 529 *
 530 * Returns the number of pages that have content in the ring buffer.
 531 */
 532size_t ring_buffer_nr_dirty_pages(struct ring_buffer *buffer, int cpu)
 533{
 534	size_t read;
 535	size_t cnt;
 536
 537	read = local_read(&buffer->buffers[cpu]->pages_read);
 538	cnt = local_read(&buffer->buffers[cpu]->pages_touched);
 539	/* The reader can read an empty page, but not more than that */
 540	if (cnt < read) {
 541		WARN_ON_ONCE(read > cnt + 1);
 542		return 0;
 543	}
 544
 545	return cnt - read;
 546}
 547
 548/*
 549 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
 550 *
 551 * Schedules a delayed work to wake up any task that is blocked on the
 552 * ring buffer waiters queue.
 553 */
 554static void rb_wake_up_waiters(struct irq_work *work)
 555{
 556	struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
 557
 558	wake_up_all(&rbwork->waiters);
 559	if (rbwork->wakeup_full) {
 560		rbwork->wakeup_full = false;
 561		wake_up_all(&rbwork->full_waiters);
 562	}
 563}
 564
 565/**
 566 * ring_buffer_wait - wait for input to the ring buffer
 567 * @buffer: buffer to wait on
 568 * @cpu: the cpu buffer to wait on
 569 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
 570 *
 571 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 572 * as data is added to any of the @buffer's cpu buffers. Otherwise
 573 * it will wait for data to be added to a specific cpu buffer.
 574 */
 575int ring_buffer_wait(struct ring_buffer *buffer, int cpu, int full)
 576{
 577	struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
 578	DEFINE_WAIT(wait);
 579	struct rb_irq_work *work;
 580	int ret = 0;
 581
 582	/*
 583	 * Depending on what the caller is waiting for, either any
 584	 * data in any cpu buffer, or a specific buffer, put the
 585	 * caller on the appropriate wait queue.
 586	 */
 587	if (cpu == RING_BUFFER_ALL_CPUS) {
 588		work = &buffer->irq_work;
 589		/* Full only makes sense on per cpu reads */
 590		full = 0;
 591	} else {
 592		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 593			return -ENODEV;
 594		cpu_buffer = buffer->buffers[cpu];
 595		work = &cpu_buffer->irq_work;
 596	}
 597
 598
 599	while (true) {
 600		if (full)
 601			prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
 602		else
 603			prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
 604
 605		/*
 606		 * The events can happen in critical sections where
 607		 * checking a work queue can cause deadlocks.
 608		 * After adding a task to the queue, this flag is set
 609		 * only to notify events to try to wake up the queue
 610		 * using irq_work.
 611		 *
 612		 * We don't clear it even if the buffer is no longer
 613		 * empty. The flag only causes the next event to run
 614		 * irq_work to do the work queue wake up. The worse
 615		 * that can happen if we race with !trace_empty() is that
 616		 * an event will cause an irq_work to try to wake up
 617		 * an empty queue.
 618		 *
 619		 * There's no reason to protect this flag either, as
 620		 * the work queue and irq_work logic will do the necessary
 621		 * synchronization for the wake ups. The only thing
 622		 * that is necessary is that the wake up happens after
 623		 * a task has been queued. It's OK for spurious wake ups.
 624		 */
 625		if (full)
 626			work->full_waiters_pending = true;
 627		else
 628			work->waiters_pending = true;
 629
 630		if (signal_pending(current)) {
 631			ret = -EINTR;
 632			break;
 633		}
 634
 635		if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
 636			break;
 637
 638		if (cpu != RING_BUFFER_ALL_CPUS &&
 639		    !ring_buffer_empty_cpu(buffer, cpu)) {
 640			unsigned long flags;
 641			bool pagebusy;
 642			size_t nr_pages;
 643			size_t dirty;
 644
 645			if (!full)
 646				break;
 647
 648			raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
 649			pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 650			nr_pages = cpu_buffer->nr_pages;
 651			dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
 652			if (!cpu_buffer->shortest_full ||
 653			    cpu_buffer->shortest_full < full)
 654				cpu_buffer->shortest_full = full;
 655			raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
 656			if (!pagebusy &&
 657			    (!nr_pages || (dirty * 100) > full * nr_pages))
 658				break;
 659		}
 
 
 
 
 
 
 660
 
 
 661		schedule();
 662	}
 663
 664	if (full)
 665		finish_wait(&work->full_waiters, &wait);
 666	else
 667		finish_wait(&work->waiters, &wait);
 668
 669	return ret;
 670}
 671
 672/**
 673 * ring_buffer_poll_wait - poll on buffer input
 674 * @buffer: buffer to wait on
 675 * @cpu: the cpu buffer to wait on
 676 * @filp: the file descriptor
 677 * @poll_table: The poll descriptor
 678 *
 679 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 680 * as data is added to any of the @buffer's cpu buffers. Otherwise
 681 * it will wait for data to be added to a specific cpu buffer.
 682 *
 683 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
 684 * zero otherwise.
 685 */
 686__poll_t ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
 687			  struct file *filp, poll_table *poll_table)
 688{
 689	struct ring_buffer_per_cpu *cpu_buffer;
 690	struct rb_irq_work *work;
 691
 
 
 
 
 692	if (cpu == RING_BUFFER_ALL_CPUS)
 693		work = &buffer->irq_work;
 694	else {
 695		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 696			return -EINVAL;
 697
 698		cpu_buffer = buffer->buffers[cpu];
 699		work = &cpu_buffer->irq_work;
 700	}
 701
 702	poll_wait(filp, &work->waiters, poll_table);
 703	work->waiters_pending = true;
 704	/*
 705	 * There's a tight race between setting the waiters_pending and
 706	 * checking if the ring buffer is empty.  Once the waiters_pending bit
 707	 * is set, the next event will wake the task up, but we can get stuck
 708	 * if there's only a single event in.
 709	 *
 710	 * FIXME: Ideally, we need a memory barrier on the writer side as well,
 711	 * but adding a memory barrier to all events will cause too much of a
 712	 * performance hit in the fast path.  We only need a memory barrier when
 713	 * the buffer goes from empty to having content.  But as this race is
 714	 * extremely small, and it's not a problem if another event comes in, we
 715	 * will fix it later.
 716	 */
 717	smp_mb();
 718
 719	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
 720	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
 721		return EPOLLIN | EPOLLRDNORM;
 722	return 0;
 723}
 724
 725/* buffer may be either ring_buffer or ring_buffer_per_cpu */
 726#define RB_WARN_ON(b, cond)						\
 727	({								\
 728		int _____ret = unlikely(cond);				\
 729		if (_____ret) {						\
 730			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
 731				struct ring_buffer_per_cpu *__b =	\
 732					(void *)b;			\
 733				atomic_inc(&__b->buffer->record_disabled); \
 734			} else						\
 735				atomic_inc(&b->record_disabled);	\
 736			WARN_ON(1);					\
 737		}							\
 738		_____ret;						\
 739	})
 740
 741/* Up this if you want to test the TIME_EXTENTS and normalization */
 742#define DEBUG_SHIFT 0
 743
 744static inline u64 rb_time_stamp(struct ring_buffer *buffer)
 745{
 746	/* shift to debug/test normalization and TIME_EXTENTS */
 747	return buffer->clock() << DEBUG_SHIFT;
 748}
 749
 750u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
 751{
 752	u64 time;
 753
 754	preempt_disable_notrace();
 755	time = rb_time_stamp(buffer);
 756	preempt_enable_notrace();
 757
 758	return time;
 759}
 760EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
 761
 762void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
 763				      int cpu, u64 *ts)
 764{
 765	/* Just stupid testing the normalize function and deltas */
 766	*ts >>= DEBUG_SHIFT;
 767}
 768EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
 769
 770/*
 771 * Making the ring buffer lockless makes things tricky.
 772 * Although writes only happen on the CPU that they are on,
 773 * and they only need to worry about interrupts. Reads can
 774 * happen on any CPU.
 775 *
 776 * The reader page is always off the ring buffer, but when the
 777 * reader finishes with a page, it needs to swap its page with
 778 * a new one from the buffer. The reader needs to take from
 779 * the head (writes go to the tail). But if a writer is in overwrite
 780 * mode and wraps, it must push the head page forward.
 781 *
 782 * Here lies the problem.
 783 *
 784 * The reader must be careful to replace only the head page, and
 785 * not another one. As described at the top of the file in the
 786 * ASCII art, the reader sets its old page to point to the next
 787 * page after head. It then sets the page after head to point to
 788 * the old reader page. But if the writer moves the head page
 789 * during this operation, the reader could end up with the tail.
 790 *
 791 * We use cmpxchg to help prevent this race. We also do something
 792 * special with the page before head. We set the LSB to 1.
 793 *
 794 * When the writer must push the page forward, it will clear the
 795 * bit that points to the head page, move the head, and then set
 796 * the bit that points to the new head page.
 797 *
 798 * We also don't want an interrupt coming in and moving the head
 799 * page on another writer. Thus we use the second LSB to catch
 800 * that too. Thus:
 801 *
 802 * head->list->prev->next        bit 1          bit 0
 803 *                              -------        -------
 804 * Normal page                     0              0
 805 * Points to head page             0              1
 806 * New head page                   1              0
 807 *
 808 * Note we can not trust the prev pointer of the head page, because:
 809 *
 810 * +----+       +-----+        +-----+
 811 * |    |------>|  T  |---X--->|  N  |
 812 * |    |<------|     |        |     |
 813 * +----+       +-----+        +-----+
 814 *   ^                           ^ |
 815 *   |          +-----+          | |
 816 *   +----------|  R  |----------+ |
 817 *              |     |<-----------+
 818 *              +-----+
 819 *
 820 * Key:  ---X-->  HEAD flag set in pointer
 821 *         T      Tail page
 822 *         R      Reader page
 823 *         N      Next page
 824 *
 825 * (see __rb_reserve_next() to see where this happens)
 826 *
 827 *  What the above shows is that the reader just swapped out
 828 *  the reader page with a page in the buffer, but before it
 829 *  could make the new header point back to the new page added
 830 *  it was preempted by a writer. The writer moved forward onto
 831 *  the new page added by the reader and is about to move forward
 832 *  again.
 833 *
 834 *  You can see, it is legitimate for the previous pointer of
 835 *  the head (or any page) not to point back to itself. But only
 836 *  temporarily.
 837 */
 838
 839#define RB_PAGE_NORMAL		0UL
 840#define RB_PAGE_HEAD		1UL
 841#define RB_PAGE_UPDATE		2UL
 842
 843
 844#define RB_FLAG_MASK		3UL
 845
 846/* PAGE_MOVED is not part of the mask */
 847#define RB_PAGE_MOVED		4UL
 848
 849/*
 850 * rb_list_head - remove any bit
 851 */
 852static struct list_head *rb_list_head(struct list_head *list)
 853{
 854	unsigned long val = (unsigned long)list;
 855
 856	return (struct list_head *)(val & ~RB_FLAG_MASK);
 857}
 858
 859/*
 860 * rb_is_head_page - test if the given page is the head page
 861 *
 862 * Because the reader may move the head_page pointer, we can
 863 * not trust what the head page is (it may be pointing to
 864 * the reader page). But if the next page is a header page,
 865 * its flags will be non zero.
 866 */
 867static inline int
 868rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
 869		struct buffer_page *page, struct list_head *list)
 870{
 871	unsigned long val;
 872
 873	val = (unsigned long)list->next;
 874
 875	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
 876		return RB_PAGE_MOVED;
 877
 878	return val & RB_FLAG_MASK;
 879}
 880
 881/*
 882 * rb_is_reader_page
 883 *
 884 * The unique thing about the reader page, is that, if the
 885 * writer is ever on it, the previous pointer never points
 886 * back to the reader page.
 887 */
 888static bool rb_is_reader_page(struct buffer_page *page)
 889{
 890	struct list_head *list = page->list.prev;
 891
 892	return rb_list_head(list->next) != &page->list;
 893}
 894
 895/*
 896 * rb_set_list_to_head - set a list_head to be pointing to head.
 897 */
 898static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
 899				struct list_head *list)
 900{
 901	unsigned long *ptr;
 902
 903	ptr = (unsigned long *)&list->next;
 904	*ptr |= RB_PAGE_HEAD;
 905	*ptr &= ~RB_PAGE_UPDATE;
 906}
 907
 908/*
 909 * rb_head_page_activate - sets up head page
 910 */
 911static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
 912{
 913	struct buffer_page *head;
 914
 915	head = cpu_buffer->head_page;
 916	if (!head)
 917		return;
 918
 919	/*
 920	 * Set the previous list pointer to have the HEAD flag.
 921	 */
 922	rb_set_list_to_head(cpu_buffer, head->list.prev);
 923}
 924
 925static void rb_list_head_clear(struct list_head *list)
 926{
 927	unsigned long *ptr = (unsigned long *)&list->next;
 928
 929	*ptr &= ~RB_FLAG_MASK;
 930}
 931
 932/*
 933 * rb_head_page_deactivate - clears head page ptr (for free list)
 934 */
 935static void
 936rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
 937{
 938	struct list_head *hd;
 939
 940	/* Go through the whole list and clear any pointers found. */
 941	rb_list_head_clear(cpu_buffer->pages);
 942
 943	list_for_each(hd, cpu_buffer->pages)
 944		rb_list_head_clear(hd);
 945}
 946
 947static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
 948			    struct buffer_page *head,
 949			    struct buffer_page *prev,
 950			    int old_flag, int new_flag)
 951{
 952	struct list_head *list;
 953	unsigned long val = (unsigned long)&head->list;
 954	unsigned long ret;
 955
 956	list = &prev->list;
 957
 958	val &= ~RB_FLAG_MASK;
 959
 960	ret = cmpxchg((unsigned long *)&list->next,
 961		      val | old_flag, val | new_flag);
 962
 963	/* check if the reader took the page */
 964	if ((ret & ~RB_FLAG_MASK) != val)
 965		return RB_PAGE_MOVED;
 966
 967	return ret & RB_FLAG_MASK;
 968}
 969
 970static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
 971				   struct buffer_page *head,
 972				   struct buffer_page *prev,
 973				   int old_flag)
 974{
 975	return rb_head_page_set(cpu_buffer, head, prev,
 976				old_flag, RB_PAGE_UPDATE);
 977}
 978
 979static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
 980				 struct buffer_page *head,
 981				 struct buffer_page *prev,
 982				 int old_flag)
 983{
 984	return rb_head_page_set(cpu_buffer, head, prev,
 985				old_flag, RB_PAGE_HEAD);
 986}
 987
 988static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
 989				   struct buffer_page *head,
 990				   struct buffer_page *prev,
 991				   int old_flag)
 992{
 993	return rb_head_page_set(cpu_buffer, head, prev,
 994				old_flag, RB_PAGE_NORMAL);
 995}
 996
 997static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
 998			       struct buffer_page **bpage)
 999{
1000	struct list_head *p = rb_list_head((*bpage)->list.next);
1001
1002	*bpage = list_entry(p, struct buffer_page, list);
1003}
1004
1005static struct buffer_page *
1006rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1007{
1008	struct buffer_page *head;
1009	struct buffer_page *page;
1010	struct list_head *list;
1011	int i;
1012
1013	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1014		return NULL;
1015
1016	/* sanity check */
1017	list = cpu_buffer->pages;
1018	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1019		return NULL;
1020
1021	page = head = cpu_buffer->head_page;
1022	/*
1023	 * It is possible that the writer moves the header behind
1024	 * where we started, and we miss in one loop.
1025	 * A second loop should grab the header, but we'll do
1026	 * three loops just because I'm paranoid.
1027	 */
1028	for (i = 0; i < 3; i++) {
1029		do {
1030			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1031				cpu_buffer->head_page = page;
1032				return page;
1033			}
1034			rb_inc_page(cpu_buffer, &page);
1035		} while (page != head);
1036	}
1037
1038	RB_WARN_ON(cpu_buffer, 1);
1039
1040	return NULL;
1041}
1042
1043static int rb_head_page_replace(struct buffer_page *old,
1044				struct buffer_page *new)
1045{
1046	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1047	unsigned long val;
1048	unsigned long ret;
1049
1050	val = *ptr & ~RB_FLAG_MASK;
1051	val |= RB_PAGE_HEAD;
1052
1053	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1054
1055	return ret == val;
1056}
1057
1058/*
1059 * rb_tail_page_update - move the tail page forward
 
 
1060 */
1061static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1062			       struct buffer_page *tail_page,
1063			       struct buffer_page *next_page)
1064{
 
1065	unsigned long old_entries;
1066	unsigned long old_write;
 
1067
1068	/*
1069	 * The tail page now needs to be moved forward.
1070	 *
1071	 * We need to reset the tail page, but without messing
1072	 * with possible erasing of data brought in by interrupts
1073	 * that have moved the tail page and are currently on it.
1074	 *
1075	 * We add a counter to the write field to denote this.
1076	 */
1077	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1078	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1079
1080	local_inc(&cpu_buffer->pages_touched);
1081	/*
1082	 * Just make sure we have seen our old_write and synchronize
1083	 * with any interrupts that come in.
1084	 */
1085	barrier();
1086
1087	/*
1088	 * If the tail page is still the same as what we think
1089	 * it is, then it is up to us to update the tail
1090	 * pointer.
1091	 */
1092	if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1093		/* Zero the write counter */
1094		unsigned long val = old_write & ~RB_WRITE_MASK;
1095		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1096
1097		/*
1098		 * This will only succeed if an interrupt did
1099		 * not come in and change it. In which case, we
1100		 * do not want to modify it.
1101		 *
1102		 * We add (void) to let the compiler know that we do not care
1103		 * about the return value of these functions. We use the
1104		 * cmpxchg to only update if an interrupt did not already
1105		 * do it for us. If the cmpxchg fails, we don't care.
1106		 */
1107		(void)local_cmpxchg(&next_page->write, old_write, val);
1108		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1109
1110		/*
1111		 * No need to worry about races with clearing out the commit.
1112		 * it only can increment when a commit takes place. But that
1113		 * only happens in the outer most nested commit.
1114		 */
1115		local_set(&next_page->page->commit, 0);
1116
1117		/* Again, either we update tail_page or an interrupt does */
1118		(void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
 
 
 
1119	}
 
 
1120}
1121
1122static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1123			  struct buffer_page *bpage)
1124{
1125	unsigned long val = (unsigned long)bpage;
1126
1127	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1128		return 1;
1129
1130	return 0;
1131}
1132
1133/**
1134 * rb_check_list - make sure a pointer to a list has the last bits zero
1135 */
1136static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1137			 struct list_head *list)
1138{
1139	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1140		return 1;
1141	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1142		return 1;
1143	return 0;
1144}
1145
1146/**
1147 * rb_check_pages - integrity check of buffer pages
1148 * @cpu_buffer: CPU buffer with pages to test
1149 *
1150 * As a safety measure we check to make sure the data pages have not
1151 * been corrupted.
1152 */
1153static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1154{
1155	struct list_head *head = cpu_buffer->pages;
1156	struct buffer_page *bpage, *tmp;
1157
1158	/* Reset the head page if it exists */
1159	if (cpu_buffer->head_page)
1160		rb_set_head_page(cpu_buffer);
1161
1162	rb_head_page_deactivate(cpu_buffer);
1163
1164	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1165		return -1;
1166	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1167		return -1;
1168
1169	if (rb_check_list(cpu_buffer, head))
1170		return -1;
1171
1172	list_for_each_entry_safe(bpage, tmp, head, list) {
1173		if (RB_WARN_ON(cpu_buffer,
1174			       bpage->list.next->prev != &bpage->list))
1175			return -1;
1176		if (RB_WARN_ON(cpu_buffer,
1177			       bpage->list.prev->next != &bpage->list))
1178			return -1;
1179		if (rb_check_list(cpu_buffer, &bpage->list))
1180			return -1;
1181	}
1182
1183	rb_head_page_activate(cpu_buffer);
1184
1185	return 0;
1186}
1187
1188static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1189{
 
1190	struct buffer_page *bpage, *tmp;
1191	bool user_thread = current->mm != NULL;
1192	gfp_t mflags;
1193	long i;
1194
1195	/*
1196	 * Check if the available memory is there first.
1197	 * Note, si_mem_available() only gives us a rough estimate of available
1198	 * memory. It may not be accurate. But we don't care, we just want
1199	 * to prevent doing any allocation when it is obvious that it is
1200	 * not going to succeed.
1201	 */
1202	i = si_mem_available();
1203	if (i < nr_pages)
1204		return -ENOMEM;
1205
1206	/*
1207	 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1208	 * gracefully without invoking oom-killer and the system is not
1209	 * destabilized.
1210	 */
1211	mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1212
1213	/*
1214	 * If a user thread allocates too much, and si_mem_available()
1215	 * reports there's enough memory, even though there is not.
1216	 * Make sure the OOM killer kills this thread. This can happen
1217	 * even with RETRY_MAYFAIL because another task may be doing
1218	 * an allocation after this task has taken all memory.
1219	 * This is the task the OOM killer needs to take out during this
1220	 * loop, even if it was triggered by an allocation somewhere else.
1221	 */
1222	if (user_thread)
1223		set_current_oom_origin();
1224	for (i = 0; i < nr_pages; i++) {
1225		struct page *page;
1226
 
 
 
 
1227		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1228				    mflags, cpu_to_node(cpu));
 
1229		if (!bpage)
1230			goto free_pages;
1231
1232		list_add(&bpage->list, pages);
1233
1234		page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
 
1235		if (!page)
1236			goto free_pages;
1237		bpage->page = page_address(page);
1238		rb_init_page(bpage->page);
1239
1240		if (user_thread && fatal_signal_pending(current))
1241			goto free_pages;
1242	}
1243	if (user_thread)
1244		clear_current_oom_origin();
1245
1246	return 0;
1247
1248free_pages:
1249	list_for_each_entry_safe(bpage, tmp, pages, list) {
1250		list_del_init(&bpage->list);
1251		free_buffer_page(bpage);
1252	}
1253	if (user_thread)
1254		clear_current_oom_origin();
1255
1256	return -ENOMEM;
1257}
1258
1259static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1260			     unsigned long nr_pages)
1261{
1262	LIST_HEAD(pages);
1263
1264	WARN_ON(!nr_pages);
1265
1266	if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1267		return -ENOMEM;
1268
1269	/*
1270	 * The ring buffer page list is a circular list that does not
1271	 * start and end with a list head. All page list items point to
1272	 * other pages.
1273	 */
1274	cpu_buffer->pages = pages.next;
1275	list_del(&pages);
1276
1277	cpu_buffer->nr_pages = nr_pages;
1278
1279	rb_check_pages(cpu_buffer);
1280
1281	return 0;
1282}
1283
1284static struct ring_buffer_per_cpu *
1285rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1286{
1287	struct ring_buffer_per_cpu *cpu_buffer;
1288	struct buffer_page *bpage;
1289	struct page *page;
1290	int ret;
1291
1292	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1293				  GFP_KERNEL, cpu_to_node(cpu));
1294	if (!cpu_buffer)
1295		return NULL;
1296
1297	cpu_buffer->cpu = cpu;
1298	cpu_buffer->buffer = buffer;
1299	raw_spin_lock_init(&cpu_buffer->reader_lock);
1300	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1301	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1302	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1303	init_completion(&cpu_buffer->update_done);
1304	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1305	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1306	init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1307
1308	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1309			    GFP_KERNEL, cpu_to_node(cpu));
1310	if (!bpage)
1311		goto fail_free_buffer;
1312
1313	rb_check_bpage(cpu_buffer, bpage);
1314
1315	cpu_buffer->reader_page = bpage;
1316	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1317	if (!page)
1318		goto fail_free_reader;
1319	bpage->page = page_address(page);
1320	rb_init_page(bpage->page);
1321
1322	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1323	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1324
1325	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1326	if (ret < 0)
1327		goto fail_free_reader;
1328
1329	cpu_buffer->head_page
1330		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1331	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1332
1333	rb_head_page_activate(cpu_buffer);
1334
1335	return cpu_buffer;
1336
1337 fail_free_reader:
1338	free_buffer_page(cpu_buffer->reader_page);
1339
1340 fail_free_buffer:
1341	kfree(cpu_buffer);
1342	return NULL;
1343}
1344
1345static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1346{
1347	struct list_head *head = cpu_buffer->pages;
1348	struct buffer_page *bpage, *tmp;
1349
1350	free_buffer_page(cpu_buffer->reader_page);
1351
1352	rb_head_page_deactivate(cpu_buffer);
1353
1354	if (head) {
1355		list_for_each_entry_safe(bpage, tmp, head, list) {
1356			list_del_init(&bpage->list);
1357			free_buffer_page(bpage);
1358		}
1359		bpage = list_entry(head, struct buffer_page, list);
1360		free_buffer_page(bpage);
1361	}
1362
1363	kfree(cpu_buffer);
1364}
1365
 
 
 
 
 
1366/**
1367 * __ring_buffer_alloc - allocate a new ring_buffer
1368 * @size: the size in bytes per cpu that is needed.
1369 * @flags: attributes to set for the ring buffer.
1370 *
1371 * Currently the only flag that is available is the RB_FL_OVERWRITE
1372 * flag. This flag means that the buffer will overwrite old data
1373 * when the buffer wraps. If this flag is not set, the buffer will
1374 * drop data when the tail hits the head.
1375 */
1376struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1377					struct lock_class_key *key)
1378{
1379	struct ring_buffer *buffer;
1380	long nr_pages;
1381	int bsize;
1382	int cpu;
1383	int ret;
1384
1385	/* keep it in its own cache line */
1386	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1387			 GFP_KERNEL);
1388	if (!buffer)
1389		return NULL;
1390
1391	if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1392		goto fail_free_buffer;
1393
1394	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1395	buffer->flags = flags;
1396	buffer->clock = trace_clock_local;
1397	buffer->reader_lock_key = key;
1398
1399	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1400	init_waitqueue_head(&buffer->irq_work.waiters);
1401
1402	/* need at least two pages */
1403	if (nr_pages < 2)
1404		nr_pages = 2;
1405
 
 
 
 
 
 
 
 
 
 
 
1406	buffer->cpus = nr_cpu_ids;
1407
1408	bsize = sizeof(void *) * nr_cpu_ids;
1409	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1410				  GFP_KERNEL);
1411	if (!buffer->buffers)
1412		goto fail_free_cpumask;
1413
1414	cpu = raw_smp_processor_id();
1415	cpumask_set_cpu(cpu, buffer->cpumask);
1416	buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1417	if (!buffer->buffers[cpu])
1418		goto fail_free_buffers;
 
1419
1420	ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1421	if (ret < 0)
1422		goto fail_free_buffers;
 
 
 
1423
1424	mutex_init(&buffer->mutex);
1425
1426	return buffer;
1427
1428 fail_free_buffers:
1429	for_each_buffer_cpu(buffer, cpu) {
1430		if (buffer->buffers[cpu])
1431			rb_free_cpu_buffer(buffer->buffers[cpu]);
1432	}
1433	kfree(buffer->buffers);
1434
1435 fail_free_cpumask:
1436	free_cpumask_var(buffer->cpumask);
 
 
 
1437
1438 fail_free_buffer:
1439	kfree(buffer);
1440	return NULL;
1441}
1442EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1443
1444/**
1445 * ring_buffer_free - free a ring buffer.
1446 * @buffer: the buffer to free.
1447 */
1448void
1449ring_buffer_free(struct ring_buffer *buffer)
1450{
1451	int cpu;
1452
1453	cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
 
 
 
1454
1455	for_each_buffer_cpu(buffer, cpu)
1456		rb_free_cpu_buffer(buffer->buffers[cpu]);
1457
 
 
 
 
1458	kfree(buffer->buffers);
1459	free_cpumask_var(buffer->cpumask);
1460
1461	kfree(buffer);
1462}
1463EXPORT_SYMBOL_GPL(ring_buffer_free);
1464
1465void ring_buffer_set_clock(struct ring_buffer *buffer,
1466			   u64 (*clock)(void))
1467{
1468	buffer->clock = clock;
1469}
1470
1471void ring_buffer_set_time_stamp_abs(struct ring_buffer *buffer, bool abs)
1472{
1473	buffer->time_stamp_abs = abs;
1474}
1475
1476bool ring_buffer_time_stamp_abs(struct ring_buffer *buffer)
1477{
1478	return buffer->time_stamp_abs;
1479}
1480
1481static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1482
1483static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1484{
1485	return local_read(&bpage->entries) & RB_WRITE_MASK;
1486}
1487
1488static inline unsigned long rb_page_write(struct buffer_page *bpage)
1489{
1490	return local_read(&bpage->write) & RB_WRITE_MASK;
1491}
1492
1493static int
1494rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1495{
1496	struct list_head *tail_page, *to_remove, *next_page;
1497	struct buffer_page *to_remove_page, *tmp_iter_page;
1498	struct buffer_page *last_page, *first_page;
1499	unsigned long nr_removed;
1500	unsigned long head_bit;
1501	int page_entries;
1502
1503	head_bit = 0;
1504
1505	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1506	atomic_inc(&cpu_buffer->record_disabled);
1507	/*
1508	 * We don't race with the readers since we have acquired the reader
1509	 * lock. We also don't race with writers after disabling recording.
1510	 * This makes it easy to figure out the first and the last page to be
1511	 * removed from the list. We unlink all the pages in between including
1512	 * the first and last pages. This is done in a busy loop so that we
1513	 * lose the least number of traces.
1514	 * The pages are freed after we restart recording and unlock readers.
1515	 */
1516	tail_page = &cpu_buffer->tail_page->list;
1517
1518	/*
1519	 * tail page might be on reader page, we remove the next page
1520	 * from the ring buffer
1521	 */
1522	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1523		tail_page = rb_list_head(tail_page->next);
1524	to_remove = tail_page;
1525
1526	/* start of pages to remove */
1527	first_page = list_entry(rb_list_head(to_remove->next),
1528				struct buffer_page, list);
1529
1530	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1531		to_remove = rb_list_head(to_remove)->next;
1532		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1533	}
1534
1535	next_page = rb_list_head(to_remove)->next;
1536
1537	/*
1538	 * Now we remove all pages between tail_page and next_page.
1539	 * Make sure that we have head_bit value preserved for the
1540	 * next page
1541	 */
1542	tail_page->next = (struct list_head *)((unsigned long)next_page |
1543						head_bit);
1544	next_page = rb_list_head(next_page);
1545	next_page->prev = tail_page;
1546
1547	/* make sure pages points to a valid page in the ring buffer */
1548	cpu_buffer->pages = next_page;
1549
1550	/* update head page */
1551	if (head_bit)
1552		cpu_buffer->head_page = list_entry(next_page,
1553						struct buffer_page, list);
1554
1555	/*
1556	 * change read pointer to make sure any read iterators reset
1557	 * themselves
1558	 */
1559	cpu_buffer->read = 0;
1560
1561	/* pages are removed, resume tracing and then free the pages */
1562	atomic_dec(&cpu_buffer->record_disabled);
1563	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1564
1565	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1566
1567	/* last buffer page to remove */
1568	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1569				list);
1570	tmp_iter_page = first_page;
1571
1572	do {
1573		cond_resched();
1574
1575		to_remove_page = tmp_iter_page;
1576		rb_inc_page(cpu_buffer, &tmp_iter_page);
1577
1578		/* update the counters */
1579		page_entries = rb_page_entries(to_remove_page);
1580		if (page_entries) {
1581			/*
1582			 * If something was added to this page, it was full
1583			 * since it is not the tail page. So we deduct the
1584			 * bytes consumed in ring buffer from here.
1585			 * Increment overrun to account for the lost events.
1586			 */
1587			local_add(page_entries, &cpu_buffer->overrun);
1588			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1589		}
1590
1591		/*
1592		 * We have already removed references to this list item, just
1593		 * free up the buffer_page and its page
1594		 */
1595		free_buffer_page(to_remove_page);
1596		nr_removed--;
1597
1598	} while (to_remove_page != last_page);
1599
1600	RB_WARN_ON(cpu_buffer, nr_removed);
1601
1602	return nr_removed == 0;
1603}
1604
1605static int
1606rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1607{
1608	struct list_head *pages = &cpu_buffer->new_pages;
1609	int retries, success;
1610
1611	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1612	/*
1613	 * We are holding the reader lock, so the reader page won't be swapped
1614	 * in the ring buffer. Now we are racing with the writer trying to
1615	 * move head page and the tail page.
1616	 * We are going to adapt the reader page update process where:
1617	 * 1. We first splice the start and end of list of new pages between
1618	 *    the head page and its previous page.
1619	 * 2. We cmpxchg the prev_page->next to point from head page to the
1620	 *    start of new pages list.
1621	 * 3. Finally, we update the head->prev to the end of new list.
1622	 *
1623	 * We will try this process 10 times, to make sure that we don't keep
1624	 * spinning.
1625	 */
1626	retries = 10;
1627	success = 0;
1628	while (retries--) {
1629		struct list_head *head_page, *prev_page, *r;
1630		struct list_head *last_page, *first_page;
1631		struct list_head *head_page_with_bit;
1632
1633		head_page = &rb_set_head_page(cpu_buffer)->list;
1634		if (!head_page)
1635			break;
1636		prev_page = head_page->prev;
1637
1638		first_page = pages->next;
1639		last_page  = pages->prev;
1640
1641		head_page_with_bit = (struct list_head *)
1642				     ((unsigned long)head_page | RB_PAGE_HEAD);
1643
1644		last_page->next = head_page_with_bit;
1645		first_page->prev = prev_page;
1646
1647		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1648
1649		if (r == head_page_with_bit) {
1650			/*
1651			 * yay, we replaced the page pointer to our new list,
1652			 * now, we just have to update to head page's prev
1653			 * pointer to point to end of list
1654			 */
1655			head_page->prev = last_page;
1656			success = 1;
1657			break;
1658		}
1659	}
1660
1661	if (success)
1662		INIT_LIST_HEAD(pages);
1663	/*
1664	 * If we weren't successful in adding in new pages, warn and stop
1665	 * tracing
1666	 */
1667	RB_WARN_ON(cpu_buffer, !success);
1668	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1669
1670	/* free pages if they weren't inserted */
1671	if (!success) {
1672		struct buffer_page *bpage, *tmp;
1673		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1674					 list) {
1675			list_del_init(&bpage->list);
1676			free_buffer_page(bpage);
1677		}
1678	}
1679	return success;
1680}
1681
1682static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1683{
1684	int success;
1685
1686	if (cpu_buffer->nr_pages_to_update > 0)
1687		success = rb_insert_pages(cpu_buffer);
1688	else
1689		success = rb_remove_pages(cpu_buffer,
1690					-cpu_buffer->nr_pages_to_update);
1691
1692	if (success)
1693		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1694}
1695
1696static void update_pages_handler(struct work_struct *work)
1697{
1698	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1699			struct ring_buffer_per_cpu, update_pages_work);
1700	rb_update_pages(cpu_buffer);
1701	complete(&cpu_buffer->update_done);
1702}
1703
1704/**
1705 * ring_buffer_resize - resize the ring buffer
1706 * @buffer: the buffer to resize.
1707 * @size: the new size.
1708 * @cpu_id: the cpu buffer to resize
1709 *
1710 * Minimum size is 2 * BUF_PAGE_SIZE.
1711 *
1712 * Returns 0 on success and < 0 on failure.
1713 */
1714int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1715			int cpu_id)
1716{
1717	struct ring_buffer_per_cpu *cpu_buffer;
1718	unsigned long nr_pages;
1719	int cpu, err = 0;
1720
1721	/*
1722	 * Always succeed at resizing a non-existent buffer:
1723	 */
1724	if (!buffer)
1725		return size;
1726
1727	/* Make sure the requested buffer exists */
1728	if (cpu_id != RING_BUFFER_ALL_CPUS &&
1729	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
1730		return size;
1731
1732	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
 
1733
1734	/* we need a minimum of two pages */
1735	if (nr_pages < 2)
1736		nr_pages = 2;
1737
1738	size = nr_pages * BUF_PAGE_SIZE;
1739
1740	/*
1741	 * Don't succeed if resizing is disabled, as a reader might be
1742	 * manipulating the ring buffer and is expecting a sane state while
1743	 * this is true.
1744	 */
1745	if (atomic_read(&buffer->resize_disabled))
1746		return -EBUSY;
1747
1748	/* prevent another thread from changing buffer sizes */
1749	mutex_lock(&buffer->mutex);
1750
1751	if (cpu_id == RING_BUFFER_ALL_CPUS) {
1752		/* calculate the pages to update */
1753		for_each_buffer_cpu(buffer, cpu) {
1754			cpu_buffer = buffer->buffers[cpu];
1755
1756			cpu_buffer->nr_pages_to_update = nr_pages -
1757							cpu_buffer->nr_pages;
1758			/*
1759			 * nothing more to do for removing pages or no update
1760			 */
1761			if (cpu_buffer->nr_pages_to_update <= 0)
1762				continue;
1763			/*
1764			 * to add pages, make sure all new pages can be
1765			 * allocated without receiving ENOMEM
1766			 */
1767			INIT_LIST_HEAD(&cpu_buffer->new_pages);
1768			if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1769						&cpu_buffer->new_pages, cpu)) {
1770				/* not enough memory for new pages */
1771				err = -ENOMEM;
1772				goto out_err;
1773			}
1774		}
1775
1776		get_online_cpus();
1777		/*
1778		 * Fire off all the required work handlers
1779		 * We can't schedule on offline CPUs, but it's not necessary
1780		 * since we can change their buffer sizes without any race.
1781		 */
1782		for_each_buffer_cpu(buffer, cpu) {
1783			cpu_buffer = buffer->buffers[cpu];
1784			if (!cpu_buffer->nr_pages_to_update)
1785				continue;
1786
1787			/* Can't run something on an offline CPU. */
1788			if (!cpu_online(cpu)) {
 
1789				rb_update_pages(cpu_buffer);
1790				cpu_buffer->nr_pages_to_update = 0;
1791			} else {
 
 
 
 
 
1792				schedule_work_on(cpu,
1793						&cpu_buffer->update_pages_work);
 
1794			}
 
1795		}
1796
1797		/* wait for all the updates to complete */
1798		for_each_buffer_cpu(buffer, cpu) {
1799			cpu_buffer = buffer->buffers[cpu];
1800			if (!cpu_buffer->nr_pages_to_update)
1801				continue;
1802
1803			if (cpu_online(cpu))
1804				wait_for_completion(&cpu_buffer->update_done);
1805			cpu_buffer->nr_pages_to_update = 0;
1806		}
1807
1808		put_online_cpus();
1809	} else {
1810		/* Make sure this CPU has been initialized */
1811		if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1812			goto out;
1813
1814		cpu_buffer = buffer->buffers[cpu_id];
1815
1816		if (nr_pages == cpu_buffer->nr_pages)
1817			goto out;
1818
1819		cpu_buffer->nr_pages_to_update = nr_pages -
1820						cpu_buffer->nr_pages;
1821
1822		INIT_LIST_HEAD(&cpu_buffer->new_pages);
1823		if (cpu_buffer->nr_pages_to_update > 0 &&
1824			__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1825					    &cpu_buffer->new_pages, cpu_id)) {
1826			err = -ENOMEM;
1827			goto out_err;
1828		}
1829
1830		get_online_cpus();
1831
1832		/* Can't run something on an offline CPU. */
1833		if (!cpu_online(cpu_id))
 
1834			rb_update_pages(cpu_buffer);
1835		else {
 
 
 
 
 
1836			schedule_work_on(cpu_id,
1837					 &cpu_buffer->update_pages_work);
1838			wait_for_completion(&cpu_buffer->update_done);
 
1839		}
 
1840
1841		cpu_buffer->nr_pages_to_update = 0;
1842		put_online_cpus();
1843	}
1844
1845 out:
1846	/*
1847	 * The ring buffer resize can happen with the ring buffer
1848	 * enabled, so that the update disturbs the tracing as little
1849	 * as possible. But if the buffer is disabled, we do not need
1850	 * to worry about that, and we can take the time to verify
1851	 * that the buffer is not corrupt.
1852	 */
1853	if (atomic_read(&buffer->record_disabled)) {
1854		atomic_inc(&buffer->record_disabled);
1855		/*
1856		 * Even though the buffer was disabled, we must make sure
1857		 * that it is truly disabled before calling rb_check_pages.
1858		 * There could have been a race between checking
1859		 * record_disable and incrementing it.
1860		 */
1861		synchronize_rcu();
1862		for_each_buffer_cpu(buffer, cpu) {
1863			cpu_buffer = buffer->buffers[cpu];
1864			rb_check_pages(cpu_buffer);
1865		}
1866		atomic_dec(&buffer->record_disabled);
1867	}
1868
1869	mutex_unlock(&buffer->mutex);
1870	return size;
1871
1872 out_err:
1873	for_each_buffer_cpu(buffer, cpu) {
1874		struct buffer_page *bpage, *tmp;
1875
1876		cpu_buffer = buffer->buffers[cpu];
1877		cpu_buffer->nr_pages_to_update = 0;
1878
1879		if (list_empty(&cpu_buffer->new_pages))
1880			continue;
1881
1882		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1883					list) {
1884			list_del_init(&bpage->list);
1885			free_buffer_page(bpage);
1886		}
1887	}
1888	mutex_unlock(&buffer->mutex);
1889	return err;
1890}
1891EXPORT_SYMBOL_GPL(ring_buffer_resize);
1892
1893void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1894{
1895	mutex_lock(&buffer->mutex);
1896	if (val)
1897		buffer->flags |= RB_FL_OVERWRITE;
1898	else
1899		buffer->flags &= ~RB_FL_OVERWRITE;
1900	mutex_unlock(&buffer->mutex);
1901}
1902EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1903
1904static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
 
 
 
 
 
 
1905{
1906	return bpage->page->data + index;
1907}
1908
1909static __always_inline struct ring_buffer_event *
1910rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1911{
1912	return __rb_page_index(cpu_buffer->reader_page,
1913			       cpu_buffer->reader_page->read);
1914}
1915
1916static __always_inline struct ring_buffer_event *
1917rb_iter_head_event(struct ring_buffer_iter *iter)
1918{
1919	return __rb_page_index(iter->head_page, iter->head);
1920}
1921
1922static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1923{
1924	return local_read(&bpage->page->commit);
1925}
1926
1927/* Size is determined by what has been committed */
1928static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1929{
1930	return rb_page_commit(bpage);
1931}
1932
1933static __always_inline unsigned
1934rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1935{
1936	return rb_page_commit(cpu_buffer->commit_page);
1937}
1938
1939static __always_inline unsigned
1940rb_event_index(struct ring_buffer_event *event)
1941{
1942	unsigned long addr = (unsigned long)event;
1943
1944	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1945}
1946
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1947static void rb_inc_iter(struct ring_buffer_iter *iter)
1948{
1949	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1950
1951	/*
1952	 * The iterator could be on the reader page (it starts there).
1953	 * But the head could have moved, since the reader was
1954	 * found. Check for this case and assign the iterator
1955	 * to the head page instead of next.
1956	 */
1957	if (iter->head_page == cpu_buffer->reader_page)
1958		iter->head_page = rb_set_head_page(cpu_buffer);
1959	else
1960		rb_inc_page(cpu_buffer, &iter->head_page);
1961
1962	iter->read_stamp = iter->head_page->page->time_stamp;
1963	iter->head = 0;
1964}
1965
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1966/*
1967 * rb_handle_head_page - writer hit the head page
1968 *
1969 * Returns: +1 to retry page
1970 *           0 to continue
1971 *          -1 on error
1972 */
1973static int
1974rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1975		    struct buffer_page *tail_page,
1976		    struct buffer_page *next_page)
1977{
1978	struct buffer_page *new_head;
1979	int entries;
1980	int type;
1981	int ret;
1982
1983	entries = rb_page_entries(next_page);
1984
1985	/*
1986	 * The hard part is here. We need to move the head
1987	 * forward, and protect against both readers on
1988	 * other CPUs and writers coming in via interrupts.
1989	 */
1990	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1991				       RB_PAGE_HEAD);
1992
1993	/*
1994	 * type can be one of four:
1995	 *  NORMAL - an interrupt already moved it for us
1996	 *  HEAD   - we are the first to get here.
1997	 *  UPDATE - we are the interrupt interrupting
1998	 *           a current move.
1999	 *  MOVED  - a reader on another CPU moved the next
2000	 *           pointer to its reader page. Give up
2001	 *           and try again.
2002	 */
2003
2004	switch (type) {
2005	case RB_PAGE_HEAD:
2006		/*
2007		 * We changed the head to UPDATE, thus
2008		 * it is our responsibility to update
2009		 * the counters.
2010		 */
2011		local_add(entries, &cpu_buffer->overrun);
2012		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2013
2014		/*
2015		 * The entries will be zeroed out when we move the
2016		 * tail page.
2017		 */
2018
2019		/* still more to do */
2020		break;
2021
2022	case RB_PAGE_UPDATE:
2023		/*
2024		 * This is an interrupt that interrupt the
2025		 * previous update. Still more to do.
2026		 */
2027		break;
2028	case RB_PAGE_NORMAL:
2029		/*
2030		 * An interrupt came in before the update
2031		 * and processed this for us.
2032		 * Nothing left to do.
2033		 */
2034		return 1;
2035	case RB_PAGE_MOVED:
2036		/*
2037		 * The reader is on another CPU and just did
2038		 * a swap with our next_page.
2039		 * Try again.
2040		 */
2041		return 1;
2042	default:
2043		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2044		return -1;
2045	}
2046
2047	/*
2048	 * Now that we are here, the old head pointer is
2049	 * set to UPDATE. This will keep the reader from
2050	 * swapping the head page with the reader page.
2051	 * The reader (on another CPU) will spin till
2052	 * we are finished.
2053	 *
2054	 * We just need to protect against interrupts
2055	 * doing the job. We will set the next pointer
2056	 * to HEAD. After that, we set the old pointer
2057	 * to NORMAL, but only if it was HEAD before.
2058	 * otherwise we are an interrupt, and only
2059	 * want the outer most commit to reset it.
2060	 */
2061	new_head = next_page;
2062	rb_inc_page(cpu_buffer, &new_head);
2063
2064	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2065				    RB_PAGE_NORMAL);
2066
2067	/*
2068	 * Valid returns are:
2069	 *  HEAD   - an interrupt came in and already set it.
2070	 *  NORMAL - One of two things:
2071	 *            1) We really set it.
2072	 *            2) A bunch of interrupts came in and moved
2073	 *               the page forward again.
2074	 */
2075	switch (ret) {
2076	case RB_PAGE_HEAD:
2077	case RB_PAGE_NORMAL:
2078		/* OK */
2079		break;
2080	default:
2081		RB_WARN_ON(cpu_buffer, 1);
2082		return -1;
2083	}
2084
2085	/*
2086	 * It is possible that an interrupt came in,
2087	 * set the head up, then more interrupts came in
2088	 * and moved it again. When we get back here,
2089	 * the page would have been set to NORMAL but we
2090	 * just set it back to HEAD.
2091	 *
2092	 * How do you detect this? Well, if that happened
2093	 * the tail page would have moved.
2094	 */
2095	if (ret == RB_PAGE_NORMAL) {
2096		struct buffer_page *buffer_tail_page;
2097
2098		buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2099		/*
2100		 * If the tail had moved passed next, then we need
2101		 * to reset the pointer.
2102		 */
2103		if (buffer_tail_page != tail_page &&
2104		    buffer_tail_page != next_page)
2105			rb_head_page_set_normal(cpu_buffer, new_head,
2106						next_page,
2107						RB_PAGE_HEAD);
2108	}
2109
2110	/*
2111	 * If this was the outer most commit (the one that
2112	 * changed the original pointer from HEAD to UPDATE),
2113	 * then it is up to us to reset it to NORMAL.
2114	 */
2115	if (type == RB_PAGE_HEAD) {
2116		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2117					      tail_page,
2118					      RB_PAGE_UPDATE);
2119		if (RB_WARN_ON(cpu_buffer,
2120			       ret != RB_PAGE_UPDATE))
2121			return -1;
2122	}
2123
2124	return 0;
2125}
2126
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2127static inline void
2128rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2129	      unsigned long tail, struct rb_event_info *info)
 
2130{
2131	struct buffer_page *tail_page = info->tail_page;
2132	struct ring_buffer_event *event;
2133	unsigned long length = info->length;
2134
2135	/*
2136	 * Only the event that crossed the page boundary
2137	 * must fill the old tail_page with padding.
2138	 */
2139	if (tail >= BUF_PAGE_SIZE) {
2140		/*
2141		 * If the page was filled, then we still need
2142		 * to update the real_end. Reset it to zero
2143		 * and the reader will ignore it.
2144		 */
2145		if (tail == BUF_PAGE_SIZE)
2146			tail_page->real_end = 0;
2147
2148		local_sub(length, &tail_page->write);
2149		return;
2150	}
2151
2152	event = __rb_page_index(tail_page, tail);
 
2153
2154	/* account for padding bytes */
2155	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2156
2157	/*
2158	 * Save the original length to the meta data.
2159	 * This will be used by the reader to add lost event
2160	 * counter.
2161	 */
2162	tail_page->real_end = tail;
2163
2164	/*
2165	 * If this event is bigger than the minimum size, then
2166	 * we need to be careful that we don't subtract the
2167	 * write counter enough to allow another writer to slip
2168	 * in on this page.
2169	 * We put in a discarded commit instead, to make sure
2170	 * that this space is not used again.
2171	 *
2172	 * If we are less than the minimum size, we don't need to
2173	 * worry about it.
2174	 */
2175	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2176		/* No room for any events */
2177
2178		/* Mark the rest of the page with padding */
2179		rb_event_set_padding(event);
2180
2181		/* Set the write back to the previous setting */
2182		local_sub(length, &tail_page->write);
2183		return;
2184	}
2185
2186	/* Put in a discarded event */
2187	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2188	event->type_len = RINGBUF_TYPE_PADDING;
2189	/* time delta must be non zero */
2190	event->time_delta = 1;
2191
2192	/* Set write to end of buffer */
2193	length = (tail + length) - BUF_PAGE_SIZE;
2194	local_sub(length, &tail_page->write);
2195}
2196
2197static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2198
2199/*
2200 * This is the slow path, force gcc not to inline it.
2201 */
2202static noinline struct ring_buffer_event *
2203rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2204	     unsigned long tail, struct rb_event_info *info)
 
2205{
2206	struct buffer_page *tail_page = info->tail_page;
2207	struct buffer_page *commit_page = cpu_buffer->commit_page;
2208	struct ring_buffer *buffer = cpu_buffer->buffer;
2209	struct buffer_page *next_page;
2210	int ret;
2211
2212	next_page = tail_page;
2213
2214	rb_inc_page(cpu_buffer, &next_page);
2215
2216	/*
2217	 * If for some reason, we had an interrupt storm that made
2218	 * it all the way around the buffer, bail, and warn
2219	 * about it.
2220	 */
2221	if (unlikely(next_page == commit_page)) {
2222		local_inc(&cpu_buffer->commit_overrun);
2223		goto out_reset;
2224	}
2225
2226	/*
2227	 * This is where the fun begins!
2228	 *
2229	 * We are fighting against races between a reader that
2230	 * could be on another CPU trying to swap its reader
2231	 * page with the buffer head.
2232	 *
2233	 * We are also fighting against interrupts coming in and
2234	 * moving the head or tail on us as well.
2235	 *
2236	 * If the next page is the head page then we have filled
2237	 * the buffer, unless the commit page is still on the
2238	 * reader page.
2239	 */
2240	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2241
2242		/*
2243		 * If the commit is not on the reader page, then
2244		 * move the header page.
2245		 */
2246		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2247			/*
2248			 * If we are not in overwrite mode,
2249			 * this is easy, just stop here.
2250			 */
2251			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2252				local_inc(&cpu_buffer->dropped_events);
2253				goto out_reset;
2254			}
2255
2256			ret = rb_handle_head_page(cpu_buffer,
2257						  tail_page,
2258						  next_page);
2259			if (ret < 0)
2260				goto out_reset;
2261			if (ret)
2262				goto out_again;
2263		} else {
2264			/*
2265			 * We need to be careful here too. The
2266			 * commit page could still be on the reader
2267			 * page. We could have a small buffer, and
2268			 * have filled up the buffer with events
2269			 * from interrupts and such, and wrapped.
2270			 *
2271			 * Note, if the tail page is also the on the
2272			 * reader_page, we let it move out.
2273			 */
2274			if (unlikely((cpu_buffer->commit_page !=
2275				      cpu_buffer->tail_page) &&
2276				     (cpu_buffer->commit_page ==
2277				      cpu_buffer->reader_page))) {
2278				local_inc(&cpu_buffer->commit_overrun);
2279				goto out_reset;
2280			}
2281		}
2282	}
2283
2284	rb_tail_page_update(cpu_buffer, tail_page, next_page);
 
 
 
 
 
 
 
 
2285
2286 out_again:
2287
2288	rb_reset_tail(cpu_buffer, tail, info);
2289
2290	/* Commit what we have for now. */
2291	rb_end_commit(cpu_buffer);
2292	/* rb_end_commit() decs committing */
2293	local_inc(&cpu_buffer->committing);
2294
2295	/* fail and let the caller try again */
2296	return ERR_PTR(-EAGAIN);
2297
2298 out_reset:
2299	/* reset write */
2300	rb_reset_tail(cpu_buffer, tail, info);
2301
2302	return NULL;
2303}
2304
2305/* Slow path, do not inline */
2306static noinline struct ring_buffer_event *
2307rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
 
2308{
2309	if (abs)
2310		event->type_len = RINGBUF_TYPE_TIME_STAMP;
2311	else
2312		event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2313
2314	/* Not the first event on the page, or not delta? */
2315	if (abs || rb_event_index(event)) {
2316		event->time_delta = delta & TS_MASK;
2317		event->array[0] = delta >> TS_SHIFT;
2318	} else {
2319		/* nope, just zero it */
2320		event->time_delta = 0;
2321		event->array[0] = 0;
2322	}
2323
2324	return skip_time_extend(event);
2325}
2326
2327static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2328				     struct ring_buffer_event *event);
 
 
 
 
 
2329
2330/**
2331 * rb_update_event - update event type and data
2332 * @event: the event to update
2333 * @type: the type of event
2334 * @length: the size of the event field in the ring buffer
2335 *
2336 * Update the type and data fields of the event. The length
2337 * is the actual size that is written to the ring buffer,
2338 * and with this, we can determine what to place into the
2339 * data field.
2340 */
2341static void
2342rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2343		struct ring_buffer_event *event,
2344		struct rb_event_info *info)
2345{
2346	unsigned length = info->length;
2347	u64 delta = info->delta;
2348
2349	/* Only a commit updates the timestamp */
2350	if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2351		delta = 0;
2352
2353	/*
2354	 * If we need to add a timestamp, then we
2355	 * add it to the start of the reserved space.
2356	 */
2357	if (unlikely(info->add_timestamp)) {
2358		bool abs = ring_buffer_time_stamp_abs(cpu_buffer->buffer);
2359
2360		event = rb_add_time_stamp(event, info->delta, abs);
2361		length -= RB_LEN_TIME_EXTEND;
2362		delta = 0;
2363	}
2364
2365	event->time_delta = delta;
2366	length -= RB_EVNT_HDR_SIZE;
2367	if (length > RB_MAX_SMALL_DATA) {
2368		event->type_len = 0;
2369		event->array[0] = length;
2370	} else
2371		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2372}
2373
2374static unsigned rb_calculate_event_length(unsigned length)
2375{
2376	struct ring_buffer_event event; /* Used only for sizeof array */
2377
2378	/* zero length can cause confusions */
2379	if (!length)
2380		length++;
2381
2382	if (length > RB_MAX_SMALL_DATA)
2383		length += sizeof(event.array[0]);
 
2384
2385	length += RB_EVNT_HDR_SIZE;
2386	length = ALIGN(length, RB_ALIGNMENT);
2387
2388	/*
2389	 * In case the time delta is larger than the 27 bits for it
2390	 * in the header, we need to add a timestamp. If another
2391	 * event comes in when trying to discard this one to increase
2392	 * the length, then the timestamp will be added in the allocated
2393	 * space of this event. If length is bigger than the size needed
2394	 * for the TIME_EXTEND, then padding has to be used. The events
2395	 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2396	 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2397	 * As length is a multiple of 4, we only need to worry if it
2398	 * is 12 (RB_LEN_TIME_EXTEND + 4).
2399	 */
2400	if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2401		length += RB_ALIGNMENT;
2402
2403	return length;
2404}
2405
2406#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2407static inline bool sched_clock_stable(void)
2408{
2409	return true;
2410}
2411#endif
2412
2413static inline int
2414rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2415		  struct ring_buffer_event *event)
2416{
2417	unsigned long new_index, old_index;
2418	struct buffer_page *bpage;
2419	unsigned long index;
2420	unsigned long addr;
2421
2422	new_index = rb_event_index(event);
2423	old_index = new_index + rb_event_ts_length(event);
2424	addr = (unsigned long)event;
2425	addr &= PAGE_MASK;
2426
2427	bpage = READ_ONCE(cpu_buffer->tail_page);
2428
2429	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2430		unsigned long write_mask =
2431			local_read(&bpage->write) & ~RB_WRITE_MASK;
2432		unsigned long event_length = rb_event_length(event);
2433		/*
2434		 * This is on the tail page. It is possible that
2435		 * a write could come in and move the tail page
2436		 * and write to the next page. That is fine
2437		 * because we just shorten what is on this page.
2438		 */
2439		old_index += write_mask;
2440		new_index += write_mask;
2441		index = local_cmpxchg(&bpage->write, old_index, new_index);
2442		if (index == old_index) {
2443			/* update counters */
2444			local_sub(event_length, &cpu_buffer->entries_bytes);
2445			return 1;
2446		}
2447	}
2448
2449	/* could not discard */
2450	return 0;
2451}
2452
2453static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2454{
2455	local_inc(&cpu_buffer->committing);
2456	local_inc(&cpu_buffer->commits);
2457}
2458
2459static __always_inline void
2460rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2461{
2462	unsigned long max_count;
2463
2464	/*
2465	 * We only race with interrupts and NMIs on this CPU.
2466	 * If we own the commit event, then we can commit
2467	 * all others that interrupted us, since the interruptions
2468	 * are in stack format (they finish before they come
2469	 * back to us). This allows us to do a simple loop to
2470	 * assign the commit to the tail.
2471	 */
2472 again:
2473	max_count = cpu_buffer->nr_pages * 100;
2474
2475	while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2476		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2477			return;
2478		if (RB_WARN_ON(cpu_buffer,
2479			       rb_is_reader_page(cpu_buffer->tail_page)))
2480			return;
2481		local_set(&cpu_buffer->commit_page->page->commit,
2482			  rb_page_write(cpu_buffer->commit_page));
2483		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2484		/* Only update the write stamp if the page has an event */
2485		if (rb_page_write(cpu_buffer->commit_page))
2486			cpu_buffer->write_stamp =
2487				cpu_buffer->commit_page->page->time_stamp;
2488		/* add barrier to keep gcc from optimizing too much */
2489		barrier();
2490	}
2491	while (rb_commit_index(cpu_buffer) !=
2492	       rb_page_write(cpu_buffer->commit_page)) {
2493
2494		local_set(&cpu_buffer->commit_page->page->commit,
2495			  rb_page_write(cpu_buffer->commit_page));
2496		RB_WARN_ON(cpu_buffer,
2497			   local_read(&cpu_buffer->commit_page->page->commit) &
2498			   ~RB_WRITE_MASK);
2499		barrier();
2500	}
2501
2502	/* again, keep gcc from optimizing */
2503	barrier();
2504
2505	/*
2506	 * If an interrupt came in just after the first while loop
2507	 * and pushed the tail page forward, we will be left with
2508	 * a dangling commit that will never go forward.
2509	 */
2510	if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2511		goto again;
2512}
2513
2514static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2515{
2516	unsigned long commits;
2517
2518	if (RB_WARN_ON(cpu_buffer,
2519		       !local_read(&cpu_buffer->committing)))
2520		return;
2521
2522 again:
2523	commits = local_read(&cpu_buffer->commits);
2524	/* synchronize with interrupts */
2525	barrier();
2526	if (local_read(&cpu_buffer->committing) == 1)
2527		rb_set_commit_to_write(cpu_buffer);
2528
2529	local_dec(&cpu_buffer->committing);
2530
2531	/* synchronize with interrupts */
2532	barrier();
2533
2534	/*
2535	 * Need to account for interrupts coming in between the
2536	 * updating of the commit page and the clearing of the
2537	 * committing counter.
2538	 */
2539	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2540	    !local_read(&cpu_buffer->committing)) {
2541		local_inc(&cpu_buffer->committing);
2542		goto again;
2543	}
2544}
2545
2546static inline void rb_event_discard(struct ring_buffer_event *event)
2547{
2548	if (extended_time(event))
2549		event = skip_time_extend(event);
2550
2551	/* array[0] holds the actual length for the discarded event */
2552	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2553	event->type_len = RINGBUF_TYPE_PADDING;
2554	/* time delta must be non zero */
2555	if (!event->time_delta)
2556		event->time_delta = 1;
2557}
2558
2559static __always_inline bool
2560rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2561		   struct ring_buffer_event *event)
2562{
2563	unsigned long addr = (unsigned long)event;
2564	unsigned long index;
2565
2566	index = rb_event_index(event);
2567	addr &= PAGE_MASK;
2568
2569	return cpu_buffer->commit_page->page == (void *)addr &&
2570		rb_commit_index(cpu_buffer) == index;
2571}
2572
2573static __always_inline void
2574rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2575		      struct ring_buffer_event *event)
2576{
2577	u64 delta;
2578
 
2579	/*
2580	 * The event first in the commit queue updates the
2581	 * time stamp.
 
 
2582	 */
2583	if (rb_event_is_commit(cpu_buffer, event)) {
2584		/*
2585		 * A commit event that is first on a page
2586		 * updates the write timestamp with the page stamp
2587		 */
2588		if (!rb_event_index(event))
2589			cpu_buffer->write_stamp =
2590				cpu_buffer->commit_page->page->time_stamp;
2591		else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2592			delta = ring_buffer_event_time_stamp(event);
2593			cpu_buffer->write_stamp += delta;
2594		} else if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
2595			delta = ring_buffer_event_time_stamp(event);
2596			cpu_buffer->write_stamp = delta;
2597		} else
2598			cpu_buffer->write_stamp += event->time_delta;
2599	}
2600}
2601
2602static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2603		      struct ring_buffer_event *event)
2604{
2605	local_inc(&cpu_buffer->entries);
2606	rb_update_write_stamp(cpu_buffer, event);
2607	rb_end_commit(cpu_buffer);
2608}
2609
2610static __always_inline void
2611rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2612{
2613	size_t nr_pages;
2614	size_t dirty;
2615	size_t full;
2616
2617	if (buffer->irq_work.waiters_pending) {
2618		buffer->irq_work.waiters_pending = false;
2619		/* irq_work_queue() supplies it's own memory barriers */
2620		irq_work_queue(&buffer->irq_work.work);
2621	}
 
 
 
 
 
 
2622
2623	if (cpu_buffer->irq_work.waiters_pending) {
2624		cpu_buffer->irq_work.waiters_pending = false;
2625		/* irq_work_queue() supplies it's own memory barriers */
2626		irq_work_queue(&cpu_buffer->irq_work.work);
2627	}
2628
2629	if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
2630		return;
2631
2632	if (cpu_buffer->reader_page == cpu_buffer->commit_page)
2633		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2634
2635	if (!cpu_buffer->irq_work.full_waiters_pending)
2636		return;
 
 
2637
2638	cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
 
2639
2640	full = cpu_buffer->shortest_full;
2641	nr_pages = cpu_buffer->nr_pages;
2642	dirty = ring_buffer_nr_dirty_pages(buffer, cpu_buffer->cpu);
2643	if (full && nr_pages && (dirty * 100) <= full * nr_pages)
2644		return;
2645
2646	cpu_buffer->irq_work.wakeup_full = true;
2647	cpu_buffer->irq_work.full_waiters_pending = false;
2648	/* irq_work_queue() supplies it's own memory barriers */
2649	irq_work_queue(&cpu_buffer->irq_work.work);
2650}
2651
 
 
2652/*
2653 * The lock and unlock are done within a preempt disable section.
2654 * The current_context per_cpu variable can only be modified
2655 * by the current task between lock and unlock. But it can
2656 * be modified more than once via an interrupt. To pass this
2657 * information from the lock to the unlock without having to
2658 * access the 'in_interrupt()' functions again (which do show
2659 * a bit of overhead in something as critical as function tracing,
2660 * we use a bitmask trick.
2661 *
2662 *  bit 0 =  NMI context
2663 *  bit 1 =  IRQ context
2664 *  bit 2 =  SoftIRQ context
2665 *  bit 3 =  normal context.
2666 *
2667 * This works because this is the order of contexts that can
2668 * preempt other contexts. A SoftIRQ never preempts an IRQ
2669 * context.
2670 *
2671 * When the context is determined, the corresponding bit is
2672 * checked and set (if it was set, then a recursion of that context
2673 * happened).
2674 *
2675 * On unlock, we need to clear this bit. To do so, just subtract
2676 * 1 from the current_context and AND it to itself.
2677 *
2678 * (binary)
2679 *  101 - 1 = 100
2680 *  101 & 100 = 100 (clearing bit zero)
2681 *
2682 *  1010 - 1 = 1001
2683 *  1010 & 1001 = 1000 (clearing bit 1)
2684 *
2685 * The least significant bit can be cleared this way, and it
2686 * just so happens that it is the same bit corresponding to
2687 * the current context.
2688 */
 
2689
2690static __always_inline int
2691trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2692{
2693	unsigned int val = cpu_buffer->current_context;
2694	unsigned long pc = preempt_count();
2695	int bit;
2696
2697	if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
2698		bit = RB_CTX_NORMAL;
2699	else
2700		bit = pc & NMI_MASK ? RB_CTX_NMI :
2701			pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
 
 
 
 
2702
2703	if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
2704		return 1;
2705
2706	val |= (1 << (bit + cpu_buffer->nest));
2707	cpu_buffer->current_context = val;
2708
2709	return 0;
2710}
2711
2712static __always_inline void
2713trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2714{
2715	cpu_buffer->current_context &=
2716		cpu_buffer->current_context - (1 << cpu_buffer->nest);
 
 
 
2717}
2718
2719/* The recursive locking above uses 4 bits */
2720#define NESTED_BITS 4
2721
2722/**
2723 * ring_buffer_nest_start - Allow to trace while nested
2724 * @buffer: The ring buffer to modify
2725 *
2726 * The ring buffer has a safety mechanism to prevent recursion.
2727 * But there may be a case where a trace needs to be done while
2728 * tracing something else. In this case, calling this function
2729 * will allow this function to nest within a currently active
2730 * ring_buffer_lock_reserve().
2731 *
2732 * Call this function before calling another ring_buffer_lock_reserve() and
2733 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
2734 */
2735void ring_buffer_nest_start(struct ring_buffer *buffer)
2736{
2737	struct ring_buffer_per_cpu *cpu_buffer;
2738	int cpu;
2739
2740	/* Enabled by ring_buffer_nest_end() */
2741	preempt_disable_notrace();
2742	cpu = raw_smp_processor_id();
2743	cpu_buffer = buffer->buffers[cpu];
2744	/* This is the shift value for the above recursive locking */
2745	cpu_buffer->nest += NESTED_BITS;
2746}
2747
2748/**
2749 * ring_buffer_nest_end - Allow to trace while nested
2750 * @buffer: The ring buffer to modify
 
2751 *
2752 * Must be called after ring_buffer_nest_start() and after the
2753 * ring_buffer_unlock_commit().
2754 */
2755void ring_buffer_nest_end(struct ring_buffer *buffer)
2756{
2757	struct ring_buffer_per_cpu *cpu_buffer;
2758	int cpu;
2759
2760	/* disabled by ring_buffer_nest_start() */
2761	cpu = raw_smp_processor_id();
2762	cpu_buffer = buffer->buffers[cpu];
2763	/* This is the shift value for the above recursive locking */
2764	cpu_buffer->nest -= NESTED_BITS;
2765	preempt_enable_notrace();
2766}
2767
2768/**
2769 * ring_buffer_unlock_commit - commit a reserved
2770 * @buffer: The buffer to commit to
2771 * @event: The event pointer to commit.
2772 *
2773 * This commits the data to the ring buffer, and releases any locks held.
 
2774 *
2775 * Must be paired with ring_buffer_lock_reserve.
 
2776 */
2777int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2778			      struct ring_buffer_event *event)
2779{
2780	struct ring_buffer_per_cpu *cpu_buffer;
2781	int cpu = raw_smp_processor_id();
2782
2783	cpu_buffer = buffer->buffers[cpu];
2784
2785	rb_commit(cpu_buffer, event);
2786
2787	rb_wakeups(buffer, cpu_buffer);
2788
2789	trace_recursive_unlock(cpu_buffer);
2790
2791	preempt_enable_notrace();
2792
2793	return 0;
2794}
2795EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2796
2797static noinline void
2798rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2799		    struct rb_event_info *info)
2800{
2801	WARN_ONCE(info->delta > (1ULL << 59),
2802		  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2803		  (unsigned long long)info->delta,
2804		  (unsigned long long)info->ts,
2805		  (unsigned long long)cpu_buffer->write_stamp,
2806		  sched_clock_stable() ? "" :
2807		  "If you just came from a suspend/resume,\n"
2808		  "please switch to the trace global clock:\n"
2809		  "  echo global > /sys/kernel/debug/tracing/trace_clock\n"
2810		  "or add trace_clock=global to the kernel command line\n");
2811	info->add_timestamp = 1;
2812}
2813
2814static struct ring_buffer_event *
2815__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2816		  struct rb_event_info *info)
2817{
2818	struct ring_buffer_event *event;
2819	struct buffer_page *tail_page;
2820	unsigned long tail, write;
2821
2822	/*
2823	 * If the time delta since the last event is too big to
2824	 * hold in the time field of the event, then we append a
2825	 * TIME EXTEND event ahead of the data event.
2826	 */
2827	if (unlikely(info->add_timestamp))
2828		info->length += RB_LEN_TIME_EXTEND;
2829
2830	/* Don't let the compiler play games with cpu_buffer->tail_page */
2831	tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2832	write = local_add_return(info->length, &tail_page->write);
2833
2834	/* set write to only the index of the write */
2835	write &= RB_WRITE_MASK;
2836	tail = write - info->length;
2837
2838	/*
2839	 * If this is the first commit on the page, then it has the same
2840	 * timestamp as the page itself.
2841	 */
2842	if (!tail && !ring_buffer_time_stamp_abs(cpu_buffer->buffer))
2843		info->delta = 0;
2844
2845	/* See if we shot pass the end of this buffer page */
2846	if (unlikely(write > BUF_PAGE_SIZE))
2847		return rb_move_tail(cpu_buffer, tail, info);
2848
2849	/* We reserved something on the buffer */
 
2850
2851	event = __rb_page_index(tail_page, tail);
2852	rb_update_event(cpu_buffer, event, info);
2853
2854	local_inc(&tail_page->entries);
 
2855
2856	/*
2857	 * If this is the first commit on the page, then update
2858	 * its timestamp.
2859	 */
2860	if (!tail)
2861		tail_page->page->time_stamp = info->ts;
2862
2863	/* account for these added bytes */
2864	local_add(info->length, &cpu_buffer->entries_bytes);
 
2865
2866	return event;
 
 
 
 
 
 
 
2867}
 
2868
2869static __always_inline struct ring_buffer_event *
2870rb_reserve_next_event(struct ring_buffer *buffer,
2871		      struct ring_buffer_per_cpu *cpu_buffer,
2872		      unsigned long length)
2873{
2874	struct ring_buffer_event *event;
2875	struct rb_event_info info;
2876	int nr_loops = 0;
2877	u64 diff;
2878
2879	rb_start_commit(cpu_buffer);
2880
2881#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2882	/*
2883	 * Due to the ability to swap a cpu buffer from a buffer
2884	 * it is possible it was swapped before we committed.
2885	 * (committing stops a swap). We check for it here and
2886	 * if it happened, we have to fail the write.
2887	 */
2888	barrier();
2889	if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2890		local_dec(&cpu_buffer->committing);
2891		local_dec(&cpu_buffer->commits);
2892		return NULL;
 
 
 
 
 
 
 
 
 
 
2893	}
2894#endif
2895
2896	info.length = rb_calculate_event_length(length);
2897 again:
2898	info.add_timestamp = 0;
2899	info.delta = 0;
2900
2901	/*
2902	 * We allow for interrupts to reenter here and do a trace.
2903	 * If one does, it will cause this original code to loop
2904	 * back here. Even with heavy interrupts happening, this
2905	 * should only happen a few times in a row. If this happens
2906	 * 1000 times in a row, there must be either an interrupt
2907	 * storm or we have something buggy.
2908	 * Bail!
2909	 */
2910	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2911		goto out_fail;
2912
2913	info.ts = rb_time_stamp(cpu_buffer->buffer);
2914	diff = info.ts - cpu_buffer->write_stamp;
2915
2916	/* make sure this diff is calculated here */
2917	barrier();
 
 
 
 
 
2918
2919	if (ring_buffer_time_stamp_abs(buffer)) {
2920		info.delta = info.ts;
2921		rb_handle_timestamp(cpu_buffer, &info);
2922	} else /* Did the write stamp get updated already? */
2923		if (likely(info.ts >= cpu_buffer->write_stamp)) {
2924		info.delta = diff;
2925		if (unlikely(test_time_stamp(info.delta)))
2926			rb_handle_timestamp(cpu_buffer, &info);
2927	}
2928
2929	event = __rb_reserve_next(cpu_buffer, &info);
2930
2931	if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2932		if (info.add_timestamp)
2933			info.length -= RB_LEN_TIME_EXTEND;
2934		goto again;
2935	}
2936
2937	if (!event)
2938		goto out_fail;
2939
2940	return event;
2941
2942 out_fail:
2943	rb_end_commit(cpu_buffer);
2944	return NULL;
2945}
2946
2947/**
2948 * ring_buffer_lock_reserve - reserve a part of the buffer
2949 * @buffer: the ring buffer to reserve from
2950 * @length: the length of the data to reserve (excluding event header)
2951 *
2952 * Returns a reserved event on the ring buffer to copy directly to.
2953 * The user of this interface will need to get the body to write into
2954 * and can use the ring_buffer_event_data() interface.
2955 *
2956 * The length is the length of the data needed, not the event length
2957 * which also includes the event header.
2958 *
2959 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2960 * If NULL is returned, then nothing has been allocated or locked.
2961 */
2962struct ring_buffer_event *
2963ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2964{
2965	struct ring_buffer_per_cpu *cpu_buffer;
2966	struct ring_buffer_event *event;
2967	int cpu;
2968
2969	/* If we are tracing schedule, we don't want to recurse */
2970	preempt_disable_notrace();
2971
2972	if (unlikely(atomic_read(&buffer->record_disabled)))
2973		goto out;
2974
2975	cpu = raw_smp_processor_id();
2976
2977	if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2978		goto out;
2979
2980	cpu_buffer = buffer->buffers[cpu];
2981
2982	if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2983		goto out;
2984
2985	if (unlikely(length > BUF_MAX_DATA_SIZE))
2986		goto out;
2987
2988	if (unlikely(trace_recursive_lock(cpu_buffer)))
2989		goto out;
2990
2991	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2992	if (!event)
2993		goto out_unlock;
2994
2995	return event;
 
 
2996
2997 out_unlock:
2998	trace_recursive_unlock(cpu_buffer);
2999 out:
3000	preempt_enable_notrace();
3001	return NULL;
 
 
 
 
 
 
3002}
3003EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3004
3005/*
3006 * Decrement the entries to the page that an event is on.
3007 * The event does not even need to exist, only the pointer
3008 * to the page it is on. This may only be called before the commit
3009 * takes place.
3010 */
3011static inline void
3012rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3013		   struct ring_buffer_event *event)
3014{
3015	unsigned long addr = (unsigned long)event;
3016	struct buffer_page *bpage = cpu_buffer->commit_page;
3017	struct buffer_page *start;
3018
3019	addr &= PAGE_MASK;
3020
3021	/* Do the likely case first */
3022	if (likely(bpage->page == (void *)addr)) {
3023		local_dec(&bpage->entries);
3024		return;
3025	}
3026
3027	/*
3028	 * Because the commit page may be on the reader page we
3029	 * start with the next page and check the end loop there.
3030	 */
3031	rb_inc_page(cpu_buffer, &bpage);
3032	start = bpage;
3033	do {
3034		if (bpage->page == (void *)addr) {
3035			local_dec(&bpage->entries);
3036			return;
3037		}
3038		rb_inc_page(cpu_buffer, &bpage);
3039	} while (bpage != start);
3040
3041	/* commit not part of this buffer?? */
3042	RB_WARN_ON(cpu_buffer, 1);
3043}
3044
3045/**
3046 * ring_buffer_commit_discard - discard an event that has not been committed
3047 * @buffer: the ring buffer
3048 * @event: non committed event to discard
3049 *
3050 * Sometimes an event that is in the ring buffer needs to be ignored.
3051 * This function lets the user discard an event in the ring buffer
3052 * and then that event will not be read later.
3053 *
3054 * This function only works if it is called before the item has been
3055 * committed. It will try to free the event from the ring buffer
3056 * if another event has not been added behind it.
3057 *
3058 * If another event has been added behind it, it will set the event
3059 * up as discarded, and perform the commit.
3060 *
3061 * If this function is called, do not call ring_buffer_unlock_commit on
3062 * the event.
3063 */
3064void ring_buffer_discard_commit(struct ring_buffer *buffer,
3065				struct ring_buffer_event *event)
3066{
3067	struct ring_buffer_per_cpu *cpu_buffer;
3068	int cpu;
3069
3070	/* The event is discarded regardless */
3071	rb_event_discard(event);
3072
3073	cpu = smp_processor_id();
3074	cpu_buffer = buffer->buffers[cpu];
3075
3076	/*
3077	 * This must only be called if the event has not been
3078	 * committed yet. Thus we can assume that preemption
3079	 * is still disabled.
3080	 */
3081	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3082
3083	rb_decrement_entry(cpu_buffer, event);
3084	if (rb_try_to_discard(cpu_buffer, event))
3085		goto out;
3086
3087	/*
3088	 * The commit is still visible by the reader, so we
3089	 * must still update the timestamp.
3090	 */
3091	rb_update_write_stamp(cpu_buffer, event);
3092 out:
3093	rb_end_commit(cpu_buffer);
3094
3095	trace_recursive_unlock(cpu_buffer);
3096
3097	preempt_enable_notrace();
3098
3099}
3100EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3101
3102/**
3103 * ring_buffer_write - write data to the buffer without reserving
3104 * @buffer: The ring buffer to write to.
3105 * @length: The length of the data being written (excluding the event header)
3106 * @data: The data to write to the buffer.
3107 *
3108 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3109 * one function. If you already have the data to write to the buffer, it
3110 * may be easier to simply call this function.
3111 *
3112 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3113 * and not the length of the event which would hold the header.
3114 */
3115int ring_buffer_write(struct ring_buffer *buffer,
3116		      unsigned long length,
3117		      void *data)
3118{
3119	struct ring_buffer_per_cpu *cpu_buffer;
3120	struct ring_buffer_event *event;
3121	void *body;
3122	int ret = -EBUSY;
3123	int cpu;
3124
 
 
 
3125	preempt_disable_notrace();
3126
3127	if (atomic_read(&buffer->record_disabled))
3128		goto out;
3129
3130	cpu = raw_smp_processor_id();
3131
3132	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3133		goto out;
3134
3135	cpu_buffer = buffer->buffers[cpu];
3136
3137	if (atomic_read(&cpu_buffer->record_disabled))
3138		goto out;
3139
3140	if (length > BUF_MAX_DATA_SIZE)
3141		goto out;
3142
3143	if (unlikely(trace_recursive_lock(cpu_buffer)))
3144		goto out;
3145
3146	event = rb_reserve_next_event(buffer, cpu_buffer, length);
3147	if (!event)
3148		goto out_unlock;
3149
3150	body = rb_event_data(event);
3151
3152	memcpy(body, data, length);
3153
3154	rb_commit(cpu_buffer, event);
3155
3156	rb_wakeups(buffer, cpu_buffer);
3157
3158	ret = 0;
3159
3160 out_unlock:
3161	trace_recursive_unlock(cpu_buffer);
3162
3163 out:
3164	preempt_enable_notrace();
3165
3166	return ret;
3167}
3168EXPORT_SYMBOL_GPL(ring_buffer_write);
3169
3170static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3171{
3172	struct buffer_page *reader = cpu_buffer->reader_page;
3173	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3174	struct buffer_page *commit = cpu_buffer->commit_page;
3175
3176	/* In case of error, head will be NULL */
3177	if (unlikely(!head))
3178		return true;
3179
3180	return reader->read == rb_page_commit(reader) &&
3181		(commit == reader ||
3182		 (commit == head &&
3183		  head->read == rb_page_commit(commit)));
3184}
3185
3186/**
3187 * ring_buffer_record_disable - stop all writes into the buffer
3188 * @buffer: The ring buffer to stop writes to.
3189 *
3190 * This prevents all writes to the buffer. Any attempt to write
3191 * to the buffer after this will fail and return NULL.
3192 *
3193 * The caller should call synchronize_rcu() after this.
3194 */
3195void ring_buffer_record_disable(struct ring_buffer *buffer)
3196{
3197	atomic_inc(&buffer->record_disabled);
3198}
3199EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3200
3201/**
3202 * ring_buffer_record_enable - enable writes to the buffer
3203 * @buffer: The ring buffer to enable writes
3204 *
3205 * Note, multiple disables will need the same number of enables
3206 * to truly enable the writing (much like preempt_disable).
3207 */
3208void ring_buffer_record_enable(struct ring_buffer *buffer)
3209{
3210	atomic_dec(&buffer->record_disabled);
3211}
3212EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3213
3214/**
3215 * ring_buffer_record_off - stop all writes into the buffer
3216 * @buffer: The ring buffer to stop writes to.
3217 *
3218 * This prevents all writes to the buffer. Any attempt to write
3219 * to the buffer after this will fail and return NULL.
3220 *
3221 * This is different than ring_buffer_record_disable() as
3222 * it works like an on/off switch, where as the disable() version
3223 * must be paired with a enable().
3224 */
3225void ring_buffer_record_off(struct ring_buffer *buffer)
3226{
3227	unsigned int rd;
3228	unsigned int new_rd;
3229
3230	do {
3231		rd = atomic_read(&buffer->record_disabled);
3232		new_rd = rd | RB_BUFFER_OFF;
3233	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3234}
3235EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3236
3237/**
3238 * ring_buffer_record_on - restart writes into the buffer
3239 * @buffer: The ring buffer to start writes to.
3240 *
3241 * This enables all writes to the buffer that was disabled by
3242 * ring_buffer_record_off().
3243 *
3244 * This is different than ring_buffer_record_enable() as
3245 * it works like an on/off switch, where as the enable() version
3246 * must be paired with a disable().
3247 */
3248void ring_buffer_record_on(struct ring_buffer *buffer)
3249{
3250	unsigned int rd;
3251	unsigned int new_rd;
3252
3253	do {
3254		rd = atomic_read(&buffer->record_disabled);
3255		new_rd = rd & ~RB_BUFFER_OFF;
3256	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3257}
3258EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3259
3260/**
3261 * ring_buffer_record_is_on - return true if the ring buffer can write
3262 * @buffer: The ring buffer to see if write is enabled
3263 *
3264 * Returns true if the ring buffer is in a state that it accepts writes.
3265 */
3266bool ring_buffer_record_is_on(struct ring_buffer *buffer)
3267{
3268	return !atomic_read(&buffer->record_disabled);
3269}
3270
3271/**
3272 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3273 * @buffer: The ring buffer to see if write is set enabled
3274 *
3275 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3276 * Note that this does NOT mean it is in a writable state.
3277 *
3278 * It may return true when the ring buffer has been disabled by
3279 * ring_buffer_record_disable(), as that is a temporary disabling of
3280 * the ring buffer.
3281 */
3282bool ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3283{
3284	return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3285}
3286
3287/**
3288 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3289 * @buffer: The ring buffer to stop writes to.
3290 * @cpu: The CPU buffer to stop
3291 *
3292 * This prevents all writes to the buffer. Any attempt to write
3293 * to the buffer after this will fail and return NULL.
3294 *
3295 * The caller should call synchronize_rcu() after this.
3296 */
3297void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3298{
3299	struct ring_buffer_per_cpu *cpu_buffer;
3300
3301	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3302		return;
3303
3304	cpu_buffer = buffer->buffers[cpu];
3305	atomic_inc(&cpu_buffer->record_disabled);
3306}
3307EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3308
3309/**
3310 * ring_buffer_record_enable_cpu - enable writes to the buffer
3311 * @buffer: The ring buffer to enable writes
3312 * @cpu: The CPU to enable.
3313 *
3314 * Note, multiple disables will need the same number of enables
3315 * to truly enable the writing (much like preempt_disable).
3316 */
3317void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3318{
3319	struct ring_buffer_per_cpu *cpu_buffer;
3320
3321	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3322		return;
3323
3324	cpu_buffer = buffer->buffers[cpu];
3325	atomic_dec(&cpu_buffer->record_disabled);
3326}
3327EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3328
3329/*
3330 * The total entries in the ring buffer is the running counter
3331 * of entries entered into the ring buffer, minus the sum of
3332 * the entries read from the ring buffer and the number of
3333 * entries that were overwritten.
3334 */
3335static inline unsigned long
3336rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3337{
3338	return local_read(&cpu_buffer->entries) -
3339		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3340}
3341
3342/**
3343 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3344 * @buffer: The ring buffer
3345 * @cpu: The per CPU buffer to read from.
3346 */
3347u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3348{
3349	unsigned long flags;
3350	struct ring_buffer_per_cpu *cpu_buffer;
3351	struct buffer_page *bpage;
3352	u64 ret = 0;
3353
3354	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3355		return 0;
3356
3357	cpu_buffer = buffer->buffers[cpu];
3358	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3359	/*
3360	 * if the tail is on reader_page, oldest time stamp is on the reader
3361	 * page
3362	 */
3363	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3364		bpage = cpu_buffer->reader_page;
3365	else
3366		bpage = rb_set_head_page(cpu_buffer);
3367	if (bpage)
3368		ret = bpage->page->time_stamp;
3369	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3370
3371	return ret;
3372}
3373EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3374
3375/**
3376 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3377 * @buffer: The ring buffer
3378 * @cpu: The per CPU buffer to read from.
3379 */
3380unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3381{
3382	struct ring_buffer_per_cpu *cpu_buffer;
3383	unsigned long ret;
3384
3385	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3386		return 0;
3387
3388	cpu_buffer = buffer->buffers[cpu];
3389	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3390
3391	return ret;
3392}
3393EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3394
3395/**
3396 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3397 * @buffer: The ring buffer
3398 * @cpu: The per CPU buffer to get the entries from.
3399 */
3400unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3401{
3402	struct ring_buffer_per_cpu *cpu_buffer;
3403
3404	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3405		return 0;
3406
3407	cpu_buffer = buffer->buffers[cpu];
3408
3409	return rb_num_of_entries(cpu_buffer);
3410}
3411EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3412
3413/**
3414 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3415 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3416 * @buffer: The ring buffer
3417 * @cpu: The per CPU buffer to get the number of overruns from
3418 */
3419unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3420{
3421	struct ring_buffer_per_cpu *cpu_buffer;
3422	unsigned long ret;
3423
3424	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3425		return 0;
3426
3427	cpu_buffer = buffer->buffers[cpu];
3428	ret = local_read(&cpu_buffer->overrun);
3429
3430	return ret;
3431}
3432EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3433
3434/**
3435 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3436 * commits failing due to the buffer wrapping around while there are uncommitted
3437 * events, such as during an interrupt storm.
3438 * @buffer: The ring buffer
3439 * @cpu: The per CPU buffer to get the number of overruns from
3440 */
3441unsigned long
3442ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3443{
3444	struct ring_buffer_per_cpu *cpu_buffer;
3445	unsigned long ret;
3446
3447	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3448		return 0;
3449
3450	cpu_buffer = buffer->buffers[cpu];
3451	ret = local_read(&cpu_buffer->commit_overrun);
3452
3453	return ret;
3454}
3455EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3456
3457/**
3458 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3459 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3460 * @buffer: The ring buffer
3461 * @cpu: The per CPU buffer to get the number of overruns from
3462 */
3463unsigned long
3464ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3465{
3466	struct ring_buffer_per_cpu *cpu_buffer;
3467	unsigned long ret;
3468
3469	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3470		return 0;
3471
3472	cpu_buffer = buffer->buffers[cpu];
3473	ret = local_read(&cpu_buffer->dropped_events);
3474
3475	return ret;
3476}
3477EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3478
3479/**
3480 * ring_buffer_read_events_cpu - get the number of events successfully read
3481 * @buffer: The ring buffer
3482 * @cpu: The per CPU buffer to get the number of events read
3483 */
3484unsigned long
3485ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3486{
3487	struct ring_buffer_per_cpu *cpu_buffer;
3488
3489	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3490		return 0;
3491
3492	cpu_buffer = buffer->buffers[cpu];
3493	return cpu_buffer->read;
3494}
3495EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3496
3497/**
3498 * ring_buffer_entries - get the number of entries in a buffer
3499 * @buffer: The ring buffer
3500 *
3501 * Returns the total number of entries in the ring buffer
3502 * (all CPU entries)
3503 */
3504unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3505{
3506	struct ring_buffer_per_cpu *cpu_buffer;
3507	unsigned long entries = 0;
3508	int cpu;
3509
3510	/* if you care about this being correct, lock the buffer */
3511	for_each_buffer_cpu(buffer, cpu) {
3512		cpu_buffer = buffer->buffers[cpu];
3513		entries += rb_num_of_entries(cpu_buffer);
3514	}
3515
3516	return entries;
3517}
3518EXPORT_SYMBOL_GPL(ring_buffer_entries);
3519
3520/**
3521 * ring_buffer_overruns - get the number of overruns in buffer
3522 * @buffer: The ring buffer
3523 *
3524 * Returns the total number of overruns in the ring buffer
3525 * (all CPU entries)
3526 */
3527unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3528{
3529	struct ring_buffer_per_cpu *cpu_buffer;
3530	unsigned long overruns = 0;
3531	int cpu;
3532
3533	/* if you care about this being correct, lock the buffer */
3534	for_each_buffer_cpu(buffer, cpu) {
3535		cpu_buffer = buffer->buffers[cpu];
3536		overruns += local_read(&cpu_buffer->overrun);
3537	}
3538
3539	return overruns;
3540}
3541EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3542
3543static void rb_iter_reset(struct ring_buffer_iter *iter)
3544{
3545	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3546
3547	/* Iterator usage is expected to have record disabled */
3548	iter->head_page = cpu_buffer->reader_page;
3549	iter->head = cpu_buffer->reader_page->read;
3550
3551	iter->cache_reader_page = iter->head_page;
3552	iter->cache_read = cpu_buffer->read;
3553
 
 
 
3554	if (iter->head)
3555		iter->read_stamp = cpu_buffer->read_stamp;
3556	else
3557		iter->read_stamp = iter->head_page->page->time_stamp;
 
 
3558}
3559
3560/**
3561 * ring_buffer_iter_reset - reset an iterator
3562 * @iter: The iterator to reset
3563 *
3564 * Resets the iterator, so that it will start from the beginning
3565 * again.
3566 */
3567void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3568{
3569	struct ring_buffer_per_cpu *cpu_buffer;
3570	unsigned long flags;
3571
3572	if (!iter)
3573		return;
3574
3575	cpu_buffer = iter->cpu_buffer;
3576
3577	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3578	rb_iter_reset(iter);
3579	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3580}
3581EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3582
3583/**
3584 * ring_buffer_iter_empty - check if an iterator has no more to read
3585 * @iter: The iterator to check
3586 */
3587int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3588{
3589	struct ring_buffer_per_cpu *cpu_buffer;
3590	struct buffer_page *reader;
3591	struct buffer_page *head_page;
3592	struct buffer_page *commit_page;
3593	unsigned commit;
3594
3595	cpu_buffer = iter->cpu_buffer;
3596
3597	/* Remember, trace recording is off when iterator is in use */
3598	reader = cpu_buffer->reader_page;
3599	head_page = cpu_buffer->head_page;
3600	commit_page = cpu_buffer->commit_page;
3601	commit = rb_page_commit(commit_page);
3602
3603	return ((iter->head_page == commit_page && iter->head == commit) ||
3604		(iter->head_page == reader && commit_page == head_page &&
3605		 head_page->read == commit &&
3606		 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3607}
3608EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3609
3610static void
3611rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3612		     struct ring_buffer_event *event)
3613{
3614	u64 delta;
3615
3616	switch (event->type_len) {
3617	case RINGBUF_TYPE_PADDING:
3618		return;
3619
3620	case RINGBUF_TYPE_TIME_EXTEND:
3621		delta = ring_buffer_event_time_stamp(event);
 
 
3622		cpu_buffer->read_stamp += delta;
3623		return;
3624
3625	case RINGBUF_TYPE_TIME_STAMP:
3626		delta = ring_buffer_event_time_stamp(event);
3627		cpu_buffer->read_stamp = delta;
3628		return;
3629
3630	case RINGBUF_TYPE_DATA:
3631		cpu_buffer->read_stamp += event->time_delta;
3632		return;
3633
3634	default:
3635		BUG();
3636	}
3637	return;
3638}
3639
3640static void
3641rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3642			  struct ring_buffer_event *event)
3643{
3644	u64 delta;
3645
3646	switch (event->type_len) {
3647	case RINGBUF_TYPE_PADDING:
3648		return;
3649
3650	case RINGBUF_TYPE_TIME_EXTEND:
3651		delta = ring_buffer_event_time_stamp(event);
 
 
3652		iter->read_stamp += delta;
3653		return;
3654
3655	case RINGBUF_TYPE_TIME_STAMP:
3656		delta = ring_buffer_event_time_stamp(event);
3657		iter->read_stamp = delta;
3658		return;
3659
3660	case RINGBUF_TYPE_DATA:
3661		iter->read_stamp += event->time_delta;
3662		return;
3663
3664	default:
3665		BUG();
3666	}
3667	return;
3668}
3669
3670static struct buffer_page *
3671rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3672{
3673	struct buffer_page *reader = NULL;
3674	unsigned long overwrite;
3675	unsigned long flags;
3676	int nr_loops = 0;
3677	int ret;
3678
3679	local_irq_save(flags);
3680	arch_spin_lock(&cpu_buffer->lock);
3681
3682 again:
3683	/*
3684	 * This should normally only loop twice. But because the
3685	 * start of the reader inserts an empty page, it causes
3686	 * a case where we will loop three times. There should be no
3687	 * reason to loop four times (that I know of).
3688	 */
3689	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3690		reader = NULL;
3691		goto out;
3692	}
3693
3694	reader = cpu_buffer->reader_page;
3695
3696	/* If there's more to read, return this page */
3697	if (cpu_buffer->reader_page->read < rb_page_size(reader))
3698		goto out;
3699
3700	/* Never should we have an index greater than the size */
3701	if (RB_WARN_ON(cpu_buffer,
3702		       cpu_buffer->reader_page->read > rb_page_size(reader)))
3703		goto out;
3704
3705	/* check if we caught up to the tail */
3706	reader = NULL;
3707	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3708		goto out;
3709
3710	/* Don't bother swapping if the ring buffer is empty */
3711	if (rb_num_of_entries(cpu_buffer) == 0)
3712		goto out;
3713
3714	/*
3715	 * Reset the reader page to size zero.
3716	 */
3717	local_set(&cpu_buffer->reader_page->write, 0);
3718	local_set(&cpu_buffer->reader_page->entries, 0);
3719	local_set(&cpu_buffer->reader_page->page->commit, 0);
3720	cpu_buffer->reader_page->real_end = 0;
3721
3722 spin:
3723	/*
3724	 * Splice the empty reader page into the list around the head.
3725	 */
3726	reader = rb_set_head_page(cpu_buffer);
3727	if (!reader)
3728		goto out;
3729	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3730	cpu_buffer->reader_page->list.prev = reader->list.prev;
3731
3732	/*
3733	 * cpu_buffer->pages just needs to point to the buffer, it
3734	 *  has no specific buffer page to point to. Lets move it out
3735	 *  of our way so we don't accidentally swap it.
3736	 */
3737	cpu_buffer->pages = reader->list.prev;
3738
3739	/* The reader page will be pointing to the new head */
3740	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3741
3742	/*
3743	 * We want to make sure we read the overruns after we set up our
3744	 * pointers to the next object. The writer side does a
3745	 * cmpxchg to cross pages which acts as the mb on the writer
3746	 * side. Note, the reader will constantly fail the swap
3747	 * while the writer is updating the pointers, so this
3748	 * guarantees that the overwrite recorded here is the one we
3749	 * want to compare with the last_overrun.
3750	 */
3751	smp_mb();
3752	overwrite = local_read(&(cpu_buffer->overrun));
3753
3754	/*
3755	 * Here's the tricky part.
3756	 *
3757	 * We need to move the pointer past the header page.
3758	 * But we can only do that if a writer is not currently
3759	 * moving it. The page before the header page has the
3760	 * flag bit '1' set if it is pointing to the page we want.
3761	 * but if the writer is in the process of moving it
3762	 * than it will be '2' or already moved '0'.
3763	 */
3764
3765	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3766
3767	/*
3768	 * If we did not convert it, then we must try again.
3769	 */
3770	if (!ret)
3771		goto spin;
3772
3773	/*
3774	 * Yay! We succeeded in replacing the page.
3775	 *
3776	 * Now make the new head point back to the reader page.
3777	 */
3778	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3779	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3780
3781	local_inc(&cpu_buffer->pages_read);
3782
3783	/* Finally update the reader page to the new head */
3784	cpu_buffer->reader_page = reader;
3785	cpu_buffer->reader_page->read = 0;
3786
3787	if (overwrite != cpu_buffer->last_overrun) {
3788		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3789		cpu_buffer->last_overrun = overwrite;
3790	}
3791
3792	goto again;
3793
3794 out:
3795	/* Update the read_stamp on the first event */
3796	if (reader && reader->read == 0)
3797		cpu_buffer->read_stamp = reader->page->time_stamp;
3798
3799	arch_spin_unlock(&cpu_buffer->lock);
3800	local_irq_restore(flags);
3801
3802	return reader;
3803}
3804
3805static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3806{
3807	struct ring_buffer_event *event;
3808	struct buffer_page *reader;
3809	unsigned length;
3810
3811	reader = rb_get_reader_page(cpu_buffer);
3812
3813	/* This function should not be called when buffer is empty */
3814	if (RB_WARN_ON(cpu_buffer, !reader))
3815		return;
3816
3817	event = rb_reader_event(cpu_buffer);
3818
3819	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3820		cpu_buffer->read++;
3821
3822	rb_update_read_stamp(cpu_buffer, event);
3823
3824	length = rb_event_length(event);
3825	cpu_buffer->reader_page->read += length;
3826}
3827
3828static void rb_advance_iter(struct ring_buffer_iter *iter)
3829{
3830	struct ring_buffer_per_cpu *cpu_buffer;
3831	struct ring_buffer_event *event;
3832	unsigned length;
3833
3834	cpu_buffer = iter->cpu_buffer;
3835
3836	/*
3837	 * Check if we are at the end of the buffer.
3838	 */
3839	if (iter->head >= rb_page_size(iter->head_page)) {
3840		/* discarded commits can make the page empty */
3841		if (iter->head_page == cpu_buffer->commit_page)
3842			return;
3843		rb_inc_iter(iter);
3844		return;
3845	}
3846
3847	event = rb_iter_head_event(iter);
3848
3849	length = rb_event_length(event);
3850
3851	/*
3852	 * This should not be called to advance the header if we are
3853	 * at the tail of the buffer.
3854	 */
3855	if (RB_WARN_ON(cpu_buffer,
3856		       (iter->head_page == cpu_buffer->commit_page) &&
3857		       (iter->head + length > rb_commit_index(cpu_buffer))))
3858		return;
3859
3860	rb_update_iter_read_stamp(iter, event);
3861
3862	iter->head += length;
3863
3864	/* check for end of page padding */
3865	if ((iter->head >= rb_page_size(iter->head_page)) &&
3866	    (iter->head_page != cpu_buffer->commit_page))
3867		rb_inc_iter(iter);
3868}
3869
3870static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3871{
3872	return cpu_buffer->lost_events;
3873}
3874
3875static struct ring_buffer_event *
3876rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3877	       unsigned long *lost_events)
3878{
3879	struct ring_buffer_event *event;
3880	struct buffer_page *reader;
3881	int nr_loops = 0;
3882
3883	if (ts)
3884		*ts = 0;
3885 again:
3886	/*
3887	 * We repeat when a time extend is encountered.
3888	 * Since the time extend is always attached to a data event,
3889	 * we should never loop more than once.
3890	 * (We never hit the following condition more than twice).
3891	 */
3892	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3893		return NULL;
3894
3895	reader = rb_get_reader_page(cpu_buffer);
3896	if (!reader)
3897		return NULL;
3898
3899	event = rb_reader_event(cpu_buffer);
3900
3901	switch (event->type_len) {
3902	case RINGBUF_TYPE_PADDING:
3903		if (rb_null_event(event))
3904			RB_WARN_ON(cpu_buffer, 1);
3905		/*
3906		 * Because the writer could be discarding every
3907		 * event it creates (which would probably be bad)
3908		 * if we were to go back to "again" then we may never
3909		 * catch up, and will trigger the warn on, or lock
3910		 * the box. Return the padding, and we will release
3911		 * the current locks, and try again.
3912		 */
3913		return event;
3914
3915	case RINGBUF_TYPE_TIME_EXTEND:
3916		/* Internal data, OK to advance */
3917		rb_advance_reader(cpu_buffer);
3918		goto again;
3919
3920	case RINGBUF_TYPE_TIME_STAMP:
3921		if (ts) {
3922			*ts = ring_buffer_event_time_stamp(event);
3923			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3924							 cpu_buffer->cpu, ts);
3925		}
3926		/* Internal data, OK to advance */
3927		rb_advance_reader(cpu_buffer);
3928		goto again;
3929
3930	case RINGBUF_TYPE_DATA:
3931		if (ts && !(*ts)) {
3932			*ts = cpu_buffer->read_stamp + event->time_delta;
3933			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3934							 cpu_buffer->cpu, ts);
3935		}
3936		if (lost_events)
3937			*lost_events = rb_lost_events(cpu_buffer);
3938		return event;
3939
3940	default:
3941		BUG();
3942	}
3943
3944	return NULL;
3945}
3946EXPORT_SYMBOL_GPL(ring_buffer_peek);
3947
3948static struct ring_buffer_event *
3949rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3950{
3951	struct ring_buffer *buffer;
3952	struct ring_buffer_per_cpu *cpu_buffer;
3953	struct ring_buffer_event *event;
3954	int nr_loops = 0;
3955
3956	if (ts)
3957		*ts = 0;
3958
3959	cpu_buffer = iter->cpu_buffer;
3960	buffer = cpu_buffer->buffer;
3961
3962	/*
3963	 * Check if someone performed a consuming read to
3964	 * the buffer. A consuming read invalidates the iterator
3965	 * and we need to reset the iterator in this case.
3966	 */
3967	if (unlikely(iter->cache_read != cpu_buffer->read ||
3968		     iter->cache_reader_page != cpu_buffer->reader_page))
3969		rb_iter_reset(iter);
3970
3971 again:
3972	if (ring_buffer_iter_empty(iter))
3973		return NULL;
3974
3975	/*
3976	 * We repeat when a time extend is encountered or we hit
3977	 * the end of the page. Since the time extend is always attached
3978	 * to a data event, we should never loop more than three times.
3979	 * Once for going to next page, once on time extend, and
3980	 * finally once to get the event.
3981	 * (We never hit the following condition more than thrice).
3982	 */
3983	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3984		return NULL;
3985
3986	if (rb_per_cpu_empty(cpu_buffer))
3987		return NULL;
3988
3989	if (iter->head >= rb_page_size(iter->head_page)) {
3990		rb_inc_iter(iter);
3991		goto again;
3992	}
3993
3994	event = rb_iter_head_event(iter);
3995
3996	switch (event->type_len) {
3997	case RINGBUF_TYPE_PADDING:
3998		if (rb_null_event(event)) {
3999			rb_inc_iter(iter);
4000			goto again;
4001		}
4002		rb_advance_iter(iter);
4003		return event;
4004
4005	case RINGBUF_TYPE_TIME_EXTEND:
4006		/* Internal data, OK to advance */
4007		rb_advance_iter(iter);
4008		goto again;
4009
4010	case RINGBUF_TYPE_TIME_STAMP:
4011		if (ts) {
4012			*ts = ring_buffer_event_time_stamp(event);
4013			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4014							 cpu_buffer->cpu, ts);
4015		}
4016		/* Internal data, OK to advance */
4017		rb_advance_iter(iter);
4018		goto again;
4019
4020	case RINGBUF_TYPE_DATA:
4021		if (ts && !(*ts)) {
4022			*ts = iter->read_stamp + event->time_delta;
4023			ring_buffer_normalize_time_stamp(buffer,
4024							 cpu_buffer->cpu, ts);
4025		}
4026		return event;
4027
4028	default:
4029		BUG();
4030	}
4031
4032	return NULL;
4033}
4034EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4035
4036static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4037{
4038	if (likely(!in_nmi())) {
4039		raw_spin_lock(&cpu_buffer->reader_lock);
4040		return true;
4041	}
4042
4043	/*
4044	 * If an NMI die dumps out the content of the ring buffer
4045	 * trylock must be used to prevent a deadlock if the NMI
4046	 * preempted a task that holds the ring buffer locks. If
4047	 * we get the lock then all is fine, if not, then continue
4048	 * to do the read, but this can corrupt the ring buffer,
4049	 * so it must be permanently disabled from future writes.
4050	 * Reading from NMI is a oneshot deal.
4051	 */
4052	if (raw_spin_trylock(&cpu_buffer->reader_lock))
4053		return true;
4054
4055	/* Continue without locking, but disable the ring buffer */
4056	atomic_inc(&cpu_buffer->record_disabled);
4057	return false;
4058}
4059
4060static inline void
4061rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4062{
4063	if (likely(locked))
4064		raw_spin_unlock(&cpu_buffer->reader_lock);
4065	return;
4066}
4067
4068/**
4069 * ring_buffer_peek - peek at the next event to be read
4070 * @buffer: The ring buffer to read
4071 * @cpu: The cpu to peak at
4072 * @ts: The timestamp counter of this event.
4073 * @lost_events: a variable to store if events were lost (may be NULL)
4074 *
4075 * This will return the event that will be read next, but does
4076 * not consume the data.
4077 */
4078struct ring_buffer_event *
4079ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
4080		 unsigned long *lost_events)
4081{
4082	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4083	struct ring_buffer_event *event;
4084	unsigned long flags;
4085	bool dolock;
4086
4087	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4088		return NULL;
4089
 
4090 again:
4091	local_irq_save(flags);
4092	dolock = rb_reader_lock(cpu_buffer);
 
4093	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4094	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4095		rb_advance_reader(cpu_buffer);
4096	rb_reader_unlock(cpu_buffer, dolock);
 
4097	local_irq_restore(flags);
4098
4099	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4100		goto again;
4101
4102	return event;
4103}
4104
4105/**
4106 * ring_buffer_iter_peek - peek at the next event to be read
4107 * @iter: The ring buffer iterator
4108 * @ts: The timestamp counter of this event.
4109 *
4110 * This will return the event that will be read next, but does
4111 * not increment the iterator.
4112 */
4113struct ring_buffer_event *
4114ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4115{
4116	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4117	struct ring_buffer_event *event;
4118	unsigned long flags;
4119
4120 again:
4121	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4122	event = rb_iter_peek(iter, ts);
4123	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4124
4125	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4126		goto again;
4127
4128	return event;
4129}
4130
4131/**
4132 * ring_buffer_consume - return an event and consume it
4133 * @buffer: The ring buffer to get the next event from
4134 * @cpu: the cpu to read the buffer from
4135 * @ts: a variable to store the timestamp (may be NULL)
4136 * @lost_events: a variable to store if events were lost (may be NULL)
4137 *
4138 * Returns the next event in the ring buffer, and that event is consumed.
4139 * Meaning, that sequential reads will keep returning a different event,
4140 * and eventually empty the ring buffer if the producer is slower.
4141 */
4142struct ring_buffer_event *
4143ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4144		    unsigned long *lost_events)
4145{
4146	struct ring_buffer_per_cpu *cpu_buffer;
4147	struct ring_buffer_event *event = NULL;
4148	unsigned long flags;
4149	bool dolock;
 
 
4150
4151 again:
4152	/* might be called in atomic */
4153	preempt_disable();
4154
4155	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4156		goto out;
4157
4158	cpu_buffer = buffer->buffers[cpu];
4159	local_irq_save(flags);
4160	dolock = rb_reader_lock(cpu_buffer);
 
4161
4162	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4163	if (event) {
4164		cpu_buffer->lost_events = 0;
4165		rb_advance_reader(cpu_buffer);
4166	}
4167
4168	rb_reader_unlock(cpu_buffer, dolock);
 
4169	local_irq_restore(flags);
4170
4171 out:
4172	preempt_enable();
4173
4174	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4175		goto again;
4176
4177	return event;
4178}
4179EXPORT_SYMBOL_GPL(ring_buffer_consume);
4180
4181/**
4182 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4183 * @buffer: The ring buffer to read from
4184 * @cpu: The cpu buffer to iterate over
4185 * @flags: gfp flags to use for memory allocation
4186 *
4187 * This performs the initial preparations necessary to iterate
4188 * through the buffer.  Memory is allocated, buffer recording
4189 * is disabled, and the iterator pointer is returned to the caller.
4190 *
4191 * Disabling buffer recording prevents the reading from being
4192 * corrupted. This is not a consuming read, so a producer is not
4193 * expected.
4194 *
4195 * After a sequence of ring_buffer_read_prepare calls, the user is
4196 * expected to make at least one call to ring_buffer_read_prepare_sync.
4197 * Afterwards, ring_buffer_read_start is invoked to get things going
4198 * for real.
4199 *
4200 * This overall must be paired with ring_buffer_read_finish.
4201 */
4202struct ring_buffer_iter *
4203ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu, gfp_t flags)
4204{
4205	struct ring_buffer_per_cpu *cpu_buffer;
4206	struct ring_buffer_iter *iter;
4207
4208	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4209		return NULL;
4210
4211	iter = kmalloc(sizeof(*iter), flags);
4212	if (!iter)
4213		return NULL;
4214
4215	cpu_buffer = buffer->buffers[cpu];
4216
4217	iter->cpu_buffer = cpu_buffer;
4218
4219	atomic_inc(&buffer->resize_disabled);
4220	atomic_inc(&cpu_buffer->record_disabled);
4221
4222	return iter;
4223}
4224EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4225
4226/**
4227 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4228 *
4229 * All previously invoked ring_buffer_read_prepare calls to prepare
4230 * iterators will be synchronized.  Afterwards, read_buffer_read_start
4231 * calls on those iterators are allowed.
4232 */
4233void
4234ring_buffer_read_prepare_sync(void)
4235{
4236	synchronize_rcu();
4237}
4238EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4239
4240/**
4241 * ring_buffer_read_start - start a non consuming read of the buffer
4242 * @iter: The iterator returned by ring_buffer_read_prepare
4243 *
4244 * This finalizes the startup of an iteration through the buffer.
4245 * The iterator comes from a call to ring_buffer_read_prepare and
4246 * an intervening ring_buffer_read_prepare_sync must have been
4247 * performed.
4248 *
4249 * Must be paired with ring_buffer_read_finish.
4250 */
4251void
4252ring_buffer_read_start(struct ring_buffer_iter *iter)
4253{
4254	struct ring_buffer_per_cpu *cpu_buffer;
4255	unsigned long flags;
4256
4257	if (!iter)
4258		return;
4259
4260	cpu_buffer = iter->cpu_buffer;
4261
4262	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4263	arch_spin_lock(&cpu_buffer->lock);
4264	rb_iter_reset(iter);
4265	arch_spin_unlock(&cpu_buffer->lock);
4266	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4267}
4268EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4269
4270/**
4271 * ring_buffer_read_finish - finish reading the iterator of the buffer
4272 * @iter: The iterator retrieved by ring_buffer_start
4273 *
4274 * This re-enables the recording to the buffer, and frees the
4275 * iterator.
4276 */
4277void
4278ring_buffer_read_finish(struct ring_buffer_iter *iter)
4279{
4280	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4281	unsigned long flags;
4282
4283	/*
4284	 * Ring buffer is disabled from recording, here's a good place
4285	 * to check the integrity of the ring buffer.
4286	 * Must prevent readers from trying to read, as the check
4287	 * clears the HEAD page and readers require it.
4288	 */
4289	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4290	rb_check_pages(cpu_buffer);
4291	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4292
4293	atomic_dec(&cpu_buffer->record_disabled);
4294	atomic_dec(&cpu_buffer->buffer->resize_disabled);
4295	kfree(iter);
4296}
4297EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4298
4299/**
4300 * ring_buffer_read - read the next item in the ring buffer by the iterator
4301 * @iter: The ring buffer iterator
4302 * @ts: The time stamp of the event read.
4303 *
4304 * This reads the next event in the ring buffer and increments the iterator.
4305 */
4306struct ring_buffer_event *
4307ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4308{
4309	struct ring_buffer_event *event;
4310	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4311	unsigned long flags;
4312
4313	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4314 again:
4315	event = rb_iter_peek(iter, ts);
4316	if (!event)
4317		goto out;
4318
4319	if (event->type_len == RINGBUF_TYPE_PADDING)
4320		goto again;
4321
4322	rb_advance_iter(iter);
4323 out:
4324	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4325
4326	return event;
4327}
4328EXPORT_SYMBOL_GPL(ring_buffer_read);
4329
4330/**
4331 * ring_buffer_size - return the size of the ring buffer (in bytes)
4332 * @buffer: The ring buffer.
4333 */
4334unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4335{
4336	/*
4337	 * Earlier, this method returned
4338	 *	BUF_PAGE_SIZE * buffer->nr_pages
4339	 * Since the nr_pages field is now removed, we have converted this to
4340	 * return the per cpu buffer value.
4341	 */
4342	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4343		return 0;
4344
4345	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4346}
4347EXPORT_SYMBOL_GPL(ring_buffer_size);
4348
4349static void
4350rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4351{
4352	rb_head_page_deactivate(cpu_buffer);
4353
4354	cpu_buffer->head_page
4355		= list_entry(cpu_buffer->pages, struct buffer_page, list);
4356	local_set(&cpu_buffer->head_page->write, 0);
4357	local_set(&cpu_buffer->head_page->entries, 0);
4358	local_set(&cpu_buffer->head_page->page->commit, 0);
4359
4360	cpu_buffer->head_page->read = 0;
4361
4362	cpu_buffer->tail_page = cpu_buffer->head_page;
4363	cpu_buffer->commit_page = cpu_buffer->head_page;
4364
4365	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4366	INIT_LIST_HEAD(&cpu_buffer->new_pages);
4367	local_set(&cpu_buffer->reader_page->write, 0);
4368	local_set(&cpu_buffer->reader_page->entries, 0);
4369	local_set(&cpu_buffer->reader_page->page->commit, 0);
4370	cpu_buffer->reader_page->read = 0;
4371
4372	local_set(&cpu_buffer->entries_bytes, 0);
4373	local_set(&cpu_buffer->overrun, 0);
4374	local_set(&cpu_buffer->commit_overrun, 0);
4375	local_set(&cpu_buffer->dropped_events, 0);
4376	local_set(&cpu_buffer->entries, 0);
4377	local_set(&cpu_buffer->committing, 0);
4378	local_set(&cpu_buffer->commits, 0);
4379	local_set(&cpu_buffer->pages_touched, 0);
4380	local_set(&cpu_buffer->pages_read, 0);
4381	cpu_buffer->last_pages_touch = 0;
4382	cpu_buffer->shortest_full = 0;
4383	cpu_buffer->read = 0;
4384	cpu_buffer->read_bytes = 0;
4385
4386	cpu_buffer->write_stamp = 0;
4387	cpu_buffer->read_stamp = 0;
4388
4389	cpu_buffer->lost_events = 0;
4390	cpu_buffer->last_overrun = 0;
4391
4392	rb_head_page_activate(cpu_buffer);
4393}
4394
4395/**
4396 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4397 * @buffer: The ring buffer to reset a per cpu buffer of
4398 * @cpu: The CPU buffer to be reset
4399 */
4400void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4401{
4402	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4403	unsigned long flags;
4404
4405	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4406		return;
4407
4408	atomic_inc(&buffer->resize_disabled);
4409	atomic_inc(&cpu_buffer->record_disabled);
4410
4411	/* Make sure all commits have finished */
4412	synchronize_rcu();
4413
4414	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4415
4416	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4417		goto out;
4418
4419	arch_spin_lock(&cpu_buffer->lock);
4420
4421	rb_reset_cpu(cpu_buffer);
4422
4423	arch_spin_unlock(&cpu_buffer->lock);
4424
4425 out:
4426	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4427
4428	atomic_dec(&cpu_buffer->record_disabled);
4429	atomic_dec(&buffer->resize_disabled);
4430}
4431EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4432
4433/**
4434 * ring_buffer_reset - reset a ring buffer
4435 * @buffer: The ring buffer to reset all cpu buffers
4436 */
4437void ring_buffer_reset(struct ring_buffer *buffer)
4438{
4439	int cpu;
4440
4441	for_each_buffer_cpu(buffer, cpu)
4442		ring_buffer_reset_cpu(buffer, cpu);
4443}
4444EXPORT_SYMBOL_GPL(ring_buffer_reset);
4445
4446/**
4447 * rind_buffer_empty - is the ring buffer empty?
4448 * @buffer: The ring buffer to test
4449 */
4450bool ring_buffer_empty(struct ring_buffer *buffer)
4451{
4452	struct ring_buffer_per_cpu *cpu_buffer;
4453	unsigned long flags;
4454	bool dolock;
4455	int cpu;
4456	int ret;
4457
 
 
4458	/* yes this is racy, but if you don't like the race, lock the buffer */
4459	for_each_buffer_cpu(buffer, cpu) {
4460		cpu_buffer = buffer->buffers[cpu];
4461		local_irq_save(flags);
4462		dolock = rb_reader_lock(cpu_buffer);
 
4463		ret = rb_per_cpu_empty(cpu_buffer);
4464		rb_reader_unlock(cpu_buffer, dolock);
 
4465		local_irq_restore(flags);
4466
4467		if (!ret)
4468			return false;
4469	}
4470
4471	return true;
4472}
4473EXPORT_SYMBOL_GPL(ring_buffer_empty);
4474
4475/**
4476 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4477 * @buffer: The ring buffer
4478 * @cpu: The CPU buffer to test
4479 */
4480bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4481{
4482	struct ring_buffer_per_cpu *cpu_buffer;
4483	unsigned long flags;
4484	bool dolock;
4485	int ret;
4486
4487	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4488		return true;
 
 
4489
4490	cpu_buffer = buffer->buffers[cpu];
4491	local_irq_save(flags);
4492	dolock = rb_reader_lock(cpu_buffer);
 
4493	ret = rb_per_cpu_empty(cpu_buffer);
4494	rb_reader_unlock(cpu_buffer, dolock);
 
4495	local_irq_restore(flags);
4496
4497	return ret;
4498}
4499EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4500
4501#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4502/**
4503 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4504 * @buffer_a: One buffer to swap with
4505 * @buffer_b: The other buffer to swap with
4506 *
4507 * This function is useful for tracers that want to take a "snapshot"
4508 * of a CPU buffer and has another back up buffer lying around.
4509 * it is expected that the tracer handles the cpu buffer not being
4510 * used at the moment.
4511 */
4512int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4513			 struct ring_buffer *buffer_b, int cpu)
4514{
4515	struct ring_buffer_per_cpu *cpu_buffer_a;
4516	struct ring_buffer_per_cpu *cpu_buffer_b;
4517	int ret = -EINVAL;
4518
4519	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4520	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
4521		goto out;
4522
4523	cpu_buffer_a = buffer_a->buffers[cpu];
4524	cpu_buffer_b = buffer_b->buffers[cpu];
4525
4526	/* At least make sure the two buffers are somewhat the same */
4527	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4528		goto out;
4529
4530	ret = -EAGAIN;
4531
 
 
 
4532	if (atomic_read(&buffer_a->record_disabled))
4533		goto out;
4534
4535	if (atomic_read(&buffer_b->record_disabled))
4536		goto out;
4537
4538	if (atomic_read(&cpu_buffer_a->record_disabled))
4539		goto out;
4540
4541	if (atomic_read(&cpu_buffer_b->record_disabled))
4542		goto out;
4543
4544	/*
4545	 * We can't do a synchronize_rcu here because this
4546	 * function can be called in atomic context.
4547	 * Normally this will be called from the same CPU as cpu.
4548	 * If not it's up to the caller to protect this.
4549	 */
4550	atomic_inc(&cpu_buffer_a->record_disabled);
4551	atomic_inc(&cpu_buffer_b->record_disabled);
4552
4553	ret = -EBUSY;
4554	if (local_read(&cpu_buffer_a->committing))
4555		goto out_dec;
4556	if (local_read(&cpu_buffer_b->committing))
4557		goto out_dec;
4558
4559	buffer_a->buffers[cpu] = cpu_buffer_b;
4560	buffer_b->buffers[cpu] = cpu_buffer_a;
4561
4562	cpu_buffer_b->buffer = buffer_a;
4563	cpu_buffer_a->buffer = buffer_b;
4564
4565	ret = 0;
4566
4567out_dec:
4568	atomic_dec(&cpu_buffer_a->record_disabled);
4569	atomic_dec(&cpu_buffer_b->record_disabled);
4570out:
4571	return ret;
4572}
4573EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4574#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4575
4576/**
4577 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4578 * @buffer: the buffer to allocate for.
4579 * @cpu: the cpu buffer to allocate.
4580 *
4581 * This function is used in conjunction with ring_buffer_read_page.
4582 * When reading a full page from the ring buffer, these functions
4583 * can be used to speed up the process. The calling function should
4584 * allocate a few pages first with this function. Then when it
4585 * needs to get pages from the ring buffer, it passes the result
4586 * of this function into ring_buffer_read_page, which will swap
4587 * the page that was allocated, with the read page of the buffer.
4588 *
4589 * Returns:
4590 *  The page allocated, or ERR_PTR
4591 */
4592void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4593{
4594	struct ring_buffer_per_cpu *cpu_buffer;
4595	struct buffer_data_page *bpage = NULL;
4596	unsigned long flags;
4597	struct page *page;
4598
4599	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4600		return ERR_PTR(-ENODEV);
4601
4602	cpu_buffer = buffer->buffers[cpu];
4603	local_irq_save(flags);
4604	arch_spin_lock(&cpu_buffer->lock);
4605
4606	if (cpu_buffer->free_page) {
4607		bpage = cpu_buffer->free_page;
4608		cpu_buffer->free_page = NULL;
4609	}
4610
4611	arch_spin_unlock(&cpu_buffer->lock);
4612	local_irq_restore(flags);
4613
4614	if (bpage)
4615		goto out;
4616
4617	page = alloc_pages_node(cpu_to_node(cpu),
4618				GFP_KERNEL | __GFP_NORETRY, 0);
4619	if (!page)
4620		return ERR_PTR(-ENOMEM);
4621
4622	bpage = page_address(page);
4623
4624 out:
4625	rb_init_page(bpage);
4626
4627	return bpage;
4628}
4629EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4630
4631/**
4632 * ring_buffer_free_read_page - free an allocated read page
4633 * @buffer: the buffer the page was allocate for
4634 * @cpu: the cpu buffer the page came from
4635 * @data: the page to free
4636 *
4637 * Free a page allocated from ring_buffer_alloc_read_page.
4638 */
4639void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4640{
4641	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4642	struct buffer_data_page *bpage = data;
4643	struct page *page = virt_to_page(bpage);
4644	unsigned long flags;
4645
4646	/* If the page is still in use someplace else, we can't reuse it */
4647	if (page_ref_count(page) > 1)
4648		goto out;
4649
4650	local_irq_save(flags);
4651	arch_spin_lock(&cpu_buffer->lock);
4652
4653	if (!cpu_buffer->free_page) {
4654		cpu_buffer->free_page = bpage;
4655		bpage = NULL;
4656	}
4657
4658	arch_spin_unlock(&cpu_buffer->lock);
4659	local_irq_restore(flags);
4660
4661 out:
4662	free_page((unsigned long)bpage);
4663}
4664EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4665
4666/**
4667 * ring_buffer_read_page - extract a page from the ring buffer
4668 * @buffer: buffer to extract from
4669 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4670 * @len: amount to extract
4671 * @cpu: the cpu of the buffer to extract
4672 * @full: should the extraction only happen when the page is full.
4673 *
4674 * This function will pull out a page from the ring buffer and consume it.
4675 * @data_page must be the address of the variable that was returned
4676 * from ring_buffer_alloc_read_page. This is because the page might be used
4677 * to swap with a page in the ring buffer.
4678 *
4679 * for example:
4680 *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
4681 *	if (IS_ERR(rpage))
4682 *		return PTR_ERR(rpage);
4683 *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4684 *	if (ret >= 0)
4685 *		process_page(rpage, ret);
4686 *
4687 * When @full is set, the function will not return true unless
4688 * the writer is off the reader page.
4689 *
4690 * Note: it is up to the calling functions to handle sleeps and wakeups.
4691 *  The ring buffer can be used anywhere in the kernel and can not
4692 *  blindly call wake_up. The layer that uses the ring buffer must be
4693 *  responsible for that.
4694 *
4695 * Returns:
4696 *  >=0 if data has been transferred, returns the offset of consumed data.
4697 *  <0 if no data has been transferred.
4698 */
4699int ring_buffer_read_page(struct ring_buffer *buffer,
4700			  void **data_page, size_t len, int cpu, int full)
4701{
4702	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4703	struct ring_buffer_event *event;
4704	struct buffer_data_page *bpage;
4705	struct buffer_page *reader;
4706	unsigned long missed_events;
4707	unsigned long flags;
4708	unsigned int commit;
4709	unsigned int read;
4710	u64 save_timestamp;
4711	int ret = -1;
4712
4713	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4714		goto out;
4715
4716	/*
4717	 * If len is not big enough to hold the page header, then
4718	 * we can not copy anything.
4719	 */
4720	if (len <= BUF_PAGE_HDR_SIZE)
4721		goto out;
4722
4723	len -= BUF_PAGE_HDR_SIZE;
4724
4725	if (!data_page)
4726		goto out;
4727
4728	bpage = *data_page;
4729	if (!bpage)
4730		goto out;
4731
4732	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4733
4734	reader = rb_get_reader_page(cpu_buffer);
4735	if (!reader)
4736		goto out_unlock;
4737
4738	event = rb_reader_event(cpu_buffer);
4739
4740	read = reader->read;
4741	commit = rb_page_commit(reader);
4742
4743	/* Check if any events were dropped */
4744	missed_events = cpu_buffer->lost_events;
4745
4746	/*
4747	 * If this page has been partially read or
4748	 * if len is not big enough to read the rest of the page or
4749	 * a writer is still on the page, then
4750	 * we must copy the data from the page to the buffer.
4751	 * Otherwise, we can simply swap the page with the one passed in.
4752	 */
4753	if (read || (len < (commit - read)) ||
4754	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
4755		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4756		unsigned int rpos = read;
4757		unsigned int pos = 0;
4758		unsigned int size;
4759
4760		if (full)
4761			goto out_unlock;
4762
4763		if (len > (commit - read))
4764			len = (commit - read);
4765
4766		/* Always keep the time extend and data together */
4767		size = rb_event_ts_length(event);
4768
4769		if (len < size)
4770			goto out_unlock;
4771
4772		/* save the current timestamp, since the user will need it */
4773		save_timestamp = cpu_buffer->read_stamp;
4774
4775		/* Need to copy one event at a time */
4776		do {
4777			/* We need the size of one event, because
4778			 * rb_advance_reader only advances by one event,
4779			 * whereas rb_event_ts_length may include the size of
4780			 * one or two events.
4781			 * We have already ensured there's enough space if this
4782			 * is a time extend. */
4783			size = rb_event_length(event);
4784			memcpy(bpage->data + pos, rpage->data + rpos, size);
4785
4786			len -= size;
4787
4788			rb_advance_reader(cpu_buffer);
4789			rpos = reader->read;
4790			pos += size;
4791
4792			if (rpos >= commit)
4793				break;
4794
4795			event = rb_reader_event(cpu_buffer);
4796			/* Always keep the time extend and data together */
4797			size = rb_event_ts_length(event);
4798		} while (len >= size);
4799
4800		/* update bpage */
4801		local_set(&bpage->commit, pos);
4802		bpage->time_stamp = save_timestamp;
4803
4804		/* we copied everything to the beginning */
4805		read = 0;
4806	} else {
4807		/* update the entry counter */
4808		cpu_buffer->read += rb_page_entries(reader);
4809		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4810
4811		/* swap the pages */
4812		rb_init_page(bpage);
4813		bpage = reader->page;
4814		reader->page = *data_page;
4815		local_set(&reader->write, 0);
4816		local_set(&reader->entries, 0);
4817		reader->read = 0;
4818		*data_page = bpage;
4819
4820		/*
4821		 * Use the real_end for the data size,
4822		 * This gives us a chance to store the lost events
4823		 * on the page.
4824		 */
4825		if (reader->real_end)
4826			local_set(&bpage->commit, reader->real_end);
4827	}
4828	ret = read;
4829
4830	cpu_buffer->lost_events = 0;
4831
4832	commit = local_read(&bpage->commit);
4833	/*
4834	 * Set a flag in the commit field if we lost events
4835	 */
4836	if (missed_events) {
4837		/* If there is room at the end of the page to save the
4838		 * missed events, then record it there.
4839		 */
4840		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4841			memcpy(&bpage->data[commit], &missed_events,
4842			       sizeof(missed_events));
4843			local_add(RB_MISSED_STORED, &bpage->commit);
4844			commit += sizeof(missed_events);
4845		}
4846		local_add(RB_MISSED_EVENTS, &bpage->commit);
4847	}
4848
4849	/*
4850	 * This page may be off to user land. Zero it out here.
4851	 */
4852	if (commit < BUF_PAGE_SIZE)
4853		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4854
4855 out_unlock:
4856	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4857
4858 out:
4859	return ret;
4860}
4861EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4862
4863/*
4864 * We only allocate new buffers, never free them if the CPU goes down.
4865 * If we were to free the buffer, then the user would lose any trace that was in
4866 * the buffer.
4867 */
4868int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4869{
4870	struct ring_buffer *buffer;
4871	long nr_pages_same;
4872	int cpu_i;
4873	unsigned long nr_pages;
4874
4875	buffer = container_of(node, struct ring_buffer, node);
4876	if (cpumask_test_cpu(cpu, buffer->cpumask))
4877		return 0;
4878
4879	nr_pages = 0;
4880	nr_pages_same = 1;
4881	/* check if all cpu sizes are same */
4882	for_each_buffer_cpu(buffer, cpu_i) {
4883		/* fill in the size from first enabled cpu */
4884		if (nr_pages == 0)
4885			nr_pages = buffer->buffers[cpu_i]->nr_pages;
4886		if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4887			nr_pages_same = 0;
4888			break;
 
4889		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4890	}
4891	/* allocate minimum pages, user can later expand it */
4892	if (!nr_pages_same)
4893		nr_pages = 2;
4894	buffer->buffers[cpu] =
4895		rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4896	if (!buffer->buffers[cpu]) {
4897		WARN(1, "failed to allocate ring buffer on CPU %u\n",
4898		     cpu);
4899		return -ENOMEM;
4900	}
4901	smp_wmb();
4902	cpumask_set_cpu(cpu, buffer->cpumask);
4903	return 0;
4904}
 
4905
4906#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4907/*
4908 * This is a basic integrity check of the ring buffer.
4909 * Late in the boot cycle this test will run when configured in.
4910 * It will kick off a thread per CPU that will go into a loop
4911 * writing to the per cpu ring buffer various sizes of data.
4912 * Some of the data will be large items, some small.
4913 *
4914 * Another thread is created that goes into a spin, sending out
4915 * IPIs to the other CPUs to also write into the ring buffer.
4916 * this is to test the nesting ability of the buffer.
4917 *
4918 * Basic stats are recorded and reported. If something in the
4919 * ring buffer should happen that's not expected, a big warning
4920 * is displayed and all ring buffers are disabled.
4921 */
4922static struct task_struct *rb_threads[NR_CPUS] __initdata;
4923
4924struct rb_test_data {
4925	struct ring_buffer	*buffer;
4926	unsigned long		events;
4927	unsigned long		bytes_written;
4928	unsigned long		bytes_alloc;
4929	unsigned long		bytes_dropped;
4930	unsigned long		events_nested;
4931	unsigned long		bytes_written_nested;
4932	unsigned long		bytes_alloc_nested;
4933	unsigned long		bytes_dropped_nested;
4934	int			min_size_nested;
4935	int			max_size_nested;
4936	int			max_size;
4937	int			min_size;
4938	int			cpu;
4939	int			cnt;
4940};
4941
4942static struct rb_test_data rb_data[NR_CPUS] __initdata;
4943
4944/* 1 meg per cpu */
4945#define RB_TEST_BUFFER_SIZE	1048576
4946
4947static char rb_string[] __initdata =
4948	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4949	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4950	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4951
4952static bool rb_test_started __initdata;
4953
4954struct rb_item {
4955	int size;
4956	char str[];
4957};
4958
4959static __init int rb_write_something(struct rb_test_data *data, bool nested)
4960{
4961	struct ring_buffer_event *event;
4962	struct rb_item *item;
4963	bool started;
4964	int event_len;
4965	int size;
4966	int len;
4967	int cnt;
4968
4969	/* Have nested writes different that what is written */
4970	cnt = data->cnt + (nested ? 27 : 0);
4971
4972	/* Multiply cnt by ~e, to make some unique increment */
4973	size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
4974
4975	len = size + sizeof(struct rb_item);
4976
4977	started = rb_test_started;
4978	/* read rb_test_started before checking buffer enabled */
4979	smp_rmb();
4980
4981	event = ring_buffer_lock_reserve(data->buffer, len);
4982	if (!event) {
4983		/* Ignore dropped events before test starts. */
4984		if (started) {
4985			if (nested)
4986				data->bytes_dropped += len;
4987			else
4988				data->bytes_dropped_nested += len;
4989		}
4990		return len;
4991	}
4992
4993	event_len = ring_buffer_event_length(event);
4994
4995	if (RB_WARN_ON(data->buffer, event_len < len))
4996		goto out;
4997
4998	item = ring_buffer_event_data(event);
4999	item->size = size;
5000	memcpy(item->str, rb_string, size);
5001
5002	if (nested) {
5003		data->bytes_alloc_nested += event_len;
5004		data->bytes_written_nested += len;
5005		data->events_nested++;
5006		if (!data->min_size_nested || len < data->min_size_nested)
5007			data->min_size_nested = len;
5008		if (len > data->max_size_nested)
5009			data->max_size_nested = len;
5010	} else {
5011		data->bytes_alloc += event_len;
5012		data->bytes_written += len;
5013		data->events++;
5014		if (!data->min_size || len < data->min_size)
5015			data->max_size = len;
5016		if (len > data->max_size)
5017			data->max_size = len;
5018	}
5019
5020 out:
5021	ring_buffer_unlock_commit(data->buffer, event);
5022
5023	return 0;
5024}
5025
5026static __init int rb_test(void *arg)
5027{
5028	struct rb_test_data *data = arg;
5029
5030	while (!kthread_should_stop()) {
5031		rb_write_something(data, false);
5032		data->cnt++;
5033
5034		set_current_state(TASK_INTERRUPTIBLE);
5035		/* Now sleep between a min of 100-300us and a max of 1ms */
5036		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5037	}
5038
5039	return 0;
5040}
5041
5042static __init void rb_ipi(void *ignore)
5043{
5044	struct rb_test_data *data;
5045	int cpu = smp_processor_id();
5046
5047	data = &rb_data[cpu];
5048	rb_write_something(data, true);
5049}
5050
5051static __init int rb_hammer_test(void *arg)
5052{
5053	while (!kthread_should_stop()) {
5054
5055		/* Send an IPI to all cpus to write data! */
5056		smp_call_function(rb_ipi, NULL, 1);
5057		/* No sleep, but for non preempt, let others run */
5058		schedule();
5059	}
5060
5061	return 0;
5062}
5063
5064static __init int test_ringbuffer(void)
5065{
5066	struct task_struct *rb_hammer;
5067	struct ring_buffer *buffer;
5068	int cpu;
5069	int ret = 0;
5070
5071	pr_info("Running ring buffer tests...\n");
5072
5073	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5074	if (WARN_ON(!buffer))
5075		return 0;
5076
5077	/* Disable buffer so that threads can't write to it yet */
5078	ring_buffer_record_off(buffer);
5079
5080	for_each_online_cpu(cpu) {
5081		rb_data[cpu].buffer = buffer;
5082		rb_data[cpu].cpu = cpu;
5083		rb_data[cpu].cnt = cpu;
5084		rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5085						 "rbtester/%d", cpu);
5086		if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5087			pr_cont("FAILED\n");
5088			ret = PTR_ERR(rb_threads[cpu]);
5089			goto out_free;
5090		}
5091
5092		kthread_bind(rb_threads[cpu], cpu);
5093 		wake_up_process(rb_threads[cpu]);
5094	}
5095
5096	/* Now create the rb hammer! */
5097	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5098	if (WARN_ON(IS_ERR(rb_hammer))) {
5099		pr_cont("FAILED\n");
5100		ret = PTR_ERR(rb_hammer);
5101		goto out_free;
5102	}
5103
5104	ring_buffer_record_on(buffer);
5105	/*
5106	 * Show buffer is enabled before setting rb_test_started.
5107	 * Yes there's a small race window where events could be
5108	 * dropped and the thread wont catch it. But when a ring
5109	 * buffer gets enabled, there will always be some kind of
5110	 * delay before other CPUs see it. Thus, we don't care about
5111	 * those dropped events. We care about events dropped after
5112	 * the threads see that the buffer is active.
5113	 */
5114	smp_wmb();
5115	rb_test_started = true;
5116
5117	set_current_state(TASK_INTERRUPTIBLE);
5118	/* Just run for 10 seconds */;
5119	schedule_timeout(10 * HZ);
5120
5121	kthread_stop(rb_hammer);
5122
5123 out_free:
5124	for_each_online_cpu(cpu) {
5125		if (!rb_threads[cpu])
5126			break;
5127		kthread_stop(rb_threads[cpu]);
5128	}
5129	if (ret) {
5130		ring_buffer_free(buffer);
5131		return ret;
5132	}
5133
5134	/* Report! */
5135	pr_info("finished\n");
5136	for_each_online_cpu(cpu) {
5137		struct ring_buffer_event *event;
5138		struct rb_test_data *data = &rb_data[cpu];
5139		struct rb_item *item;
5140		unsigned long total_events;
5141		unsigned long total_dropped;
5142		unsigned long total_written;
5143		unsigned long total_alloc;
5144		unsigned long total_read = 0;
5145		unsigned long total_size = 0;
5146		unsigned long total_len = 0;
5147		unsigned long total_lost = 0;
5148		unsigned long lost;
5149		int big_event_size;
5150		int small_event_size;
5151
5152		ret = -1;
5153
5154		total_events = data->events + data->events_nested;
5155		total_written = data->bytes_written + data->bytes_written_nested;
5156		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5157		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5158
5159		big_event_size = data->max_size + data->max_size_nested;
5160		small_event_size = data->min_size + data->min_size_nested;
5161
5162		pr_info("CPU %d:\n", cpu);
5163		pr_info("              events:    %ld\n", total_events);
5164		pr_info("       dropped bytes:    %ld\n", total_dropped);
5165		pr_info("       alloced bytes:    %ld\n", total_alloc);
5166		pr_info("       written bytes:    %ld\n", total_written);
5167		pr_info("       biggest event:    %d\n", big_event_size);
5168		pr_info("      smallest event:    %d\n", small_event_size);
5169
5170		if (RB_WARN_ON(buffer, total_dropped))
5171			break;
5172
5173		ret = 0;
5174
5175		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5176			total_lost += lost;
5177			item = ring_buffer_event_data(event);
5178			total_len += ring_buffer_event_length(event);
5179			total_size += item->size + sizeof(struct rb_item);
5180			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5181				pr_info("FAILED!\n");
5182				pr_info("buffer had: %.*s\n", item->size, item->str);
5183				pr_info("expected:   %.*s\n", item->size, rb_string);
5184				RB_WARN_ON(buffer, 1);
5185				ret = -1;
5186				break;
5187			}
5188			total_read++;
5189		}
5190		if (ret)
5191			break;
5192
5193		ret = -1;
5194
5195		pr_info("         read events:   %ld\n", total_read);
5196		pr_info("         lost events:   %ld\n", total_lost);
5197		pr_info("        total events:   %ld\n", total_lost + total_read);
5198		pr_info("  recorded len bytes:   %ld\n", total_len);
5199		pr_info(" recorded size bytes:   %ld\n", total_size);
5200		if (total_lost)
5201			pr_info(" With dropped events, record len and size may not match\n"
5202				" alloced and written from above\n");
5203		if (!total_lost) {
5204			if (RB_WARN_ON(buffer, total_len != total_alloc ||
5205				       total_size != total_written))
5206				break;
5207		}
5208		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5209			break;
5210
5211		ret = 0;
5212	}
5213	if (!ret)
5214		pr_info("Ring buffer PASSED!\n");
5215
5216	ring_buffer_free(buffer);
5217	return 0;
5218}
5219
5220late_initcall(test_ringbuffer);
5221#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 */