Linux Audio

Check our new training course

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 */
v4.17
 
   1/*
   2 * Generic ring buffer
   3 *
   4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
   5 */
   6#include <linux/trace_events.h>
   7#include <linux/ring_buffer.h>
   8#include <linux/trace_clock.h>
   9#include <linux/sched/clock.h>
  10#include <linux/trace_seq.h>
  11#include <linux/spinlock.h>
  12#include <linux/irq_work.h>
  13#include <linux/uaccess.h>
  14#include <linux/hardirq.h>
  15#include <linux/kthread.h>	/* for self test */
  16#include <linux/module.h>
  17#include <linux/percpu.h>
  18#include <linux/mutex.h>
  19#include <linux/delay.h>
  20#include <linux/slab.h>
  21#include <linux/init.h>
  22#include <linux/hash.h>
  23#include <linux/list.h>
  24#include <linux/cpu.h>
  25#include <linux/oom.h>
  26
  27#include <asm/local.h>
  28
  29static void update_pages_handler(struct work_struct *work);
  30
  31/*
  32 * The ring buffer header is special. We must manually up keep it.
  33 */
  34int ring_buffer_print_entry_header(struct trace_seq *s)
  35{
  36	trace_seq_puts(s, "# compressed entry header\n");
  37	trace_seq_puts(s, "\ttype_len    :    5 bits\n");
  38	trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
  39	trace_seq_puts(s, "\tarray       :   32 bits\n");
  40	trace_seq_putc(s, '\n');
  41	trace_seq_printf(s, "\tpadding     : type == %d\n",
  42			 RINGBUF_TYPE_PADDING);
  43	trace_seq_printf(s, "\ttime_extend : type == %d\n",
  44			 RINGBUF_TYPE_TIME_EXTEND);
  45	trace_seq_printf(s, "\ttime_stamp : type == %d\n",
  46			 RINGBUF_TYPE_TIME_STAMP);
  47	trace_seq_printf(s, "\tdata max type_len  == %d\n",
  48			 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
  49
  50	return !trace_seq_has_overflowed(s);
  51}
  52
  53/*
  54 * The ring buffer is made up of a list of pages. A separate list of pages is
  55 * allocated for each CPU. A writer may only write to a buffer that is
  56 * associated with the CPU it is currently executing on.  A reader may read
  57 * from any per cpu buffer.
  58 *
  59 * The reader is special. For each per cpu buffer, the reader has its own
  60 * reader page. When a reader has read the entire reader page, this reader
  61 * page is swapped with another page in the ring buffer.
  62 *
  63 * Now, as long as the writer is off the reader page, the reader can do what
  64 * ever it wants with that page. The writer will never write to that page
  65 * again (as long as it is out of the ring buffer).
  66 *
  67 * Here's some silly ASCII art.
  68 *
  69 *   +------+
  70 *   |reader|          RING BUFFER
  71 *   |page  |
  72 *   +------+        +---+   +---+   +---+
  73 *                   |   |-->|   |-->|   |
  74 *                   +---+   +---+   +---+
  75 *                     ^               |
  76 *                     |               |
  77 *                     +---------------+
  78 *
  79 *
  80 *   +------+
  81 *   |reader|          RING BUFFER
  82 *   |page  |------------------v
  83 *   +------+        +---+   +---+   +---+
  84 *                   |   |-->|   |-->|   |
  85 *                   +---+   +---+   +---+
  86 *                     ^               |
  87 *                     |               |
  88 *                     +---------------+
  89 *
  90 *
  91 *   +------+
  92 *   |reader|          RING BUFFER
  93 *   |page  |------------------v
  94 *   +------+        +---+   +---+   +---+
  95 *      ^            |   |-->|   |-->|   |
  96 *      |            +---+   +---+   +---+
  97 *      |                              |
  98 *      |                              |
  99 *      +------------------------------+
 100 *
 101 *
 102 *   +------+
 103 *   |buffer|          RING BUFFER
 104 *   |page  |------------------v
 105 *   +------+        +---+   +---+   +---+
 106 *      ^            |   |   |   |-->|   |
 107 *      |   New      +---+   +---+   +---+
 108 *      |  Reader------^               |
 109 *      |   page                       |
 110 *      +------------------------------+
 111 *
 112 *
 113 * After we make this swap, the reader can hand this page off to the splice
 114 * code and be done with it. It can even allocate a new page if it needs to
 115 * and swap that into the ring buffer.
 116 *
 117 * We will be using cmpxchg soon to make all this lockless.
 118 *
 119 */
 120
 121/* Used for individual buffers (after the counter) */
 122#define RB_BUFFER_OFF		(1 << 20)
 123
 124#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
 125
 126#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
 127#define RB_ALIGNMENT		4U
 128#define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 129#define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
 130
 131#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
 132# define RB_FORCE_8BYTE_ALIGNMENT	0
 133# define RB_ARCH_ALIGNMENT		RB_ALIGNMENT
 134#else
 135# define RB_FORCE_8BYTE_ALIGNMENT	1
 136# define RB_ARCH_ALIGNMENT		8U
 137#endif
 138
 139#define RB_ALIGN_DATA		__aligned(RB_ARCH_ALIGNMENT)
 140
 141/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
 142#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
 143
 144enum {
 145	RB_LEN_TIME_EXTEND = 8,
 146	RB_LEN_TIME_STAMP =  8,
 147};
 148
 149#define skip_time_extend(event) \
 150	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
 151
 152#define extended_time(event) \
 153	(event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
 154
 155static inline int rb_null_event(struct ring_buffer_event *event)
 156{
 157	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
 158}
 159
 160static void rb_event_set_padding(struct ring_buffer_event *event)
 161{
 162	/* padding has a NULL time_delta */
 163	event->type_len = RINGBUF_TYPE_PADDING;
 164	event->time_delta = 0;
 165}
 166
 167static unsigned
 168rb_event_data_length(struct ring_buffer_event *event)
 169{
 170	unsigned length;
 171
 172	if (event->type_len)
 173		length = event->type_len * RB_ALIGNMENT;
 174	else
 175		length = event->array[0];
 176	return length + RB_EVNT_HDR_SIZE;
 177}
 178
 179/*
 180 * Return the length of the given event. Will return
 181 * the length of the time extend if the event is a
 182 * time extend.
 183 */
 184static inline unsigned
 185rb_event_length(struct ring_buffer_event *event)
 186{
 187	switch (event->type_len) {
 188	case RINGBUF_TYPE_PADDING:
 189		if (rb_null_event(event))
 190			/* undefined */
 191			return -1;
 192		return  event->array[0] + RB_EVNT_HDR_SIZE;
 193
 194	case RINGBUF_TYPE_TIME_EXTEND:
 195		return RB_LEN_TIME_EXTEND;
 196
 197	case RINGBUF_TYPE_TIME_STAMP:
 198		return RB_LEN_TIME_STAMP;
 199
 200	case RINGBUF_TYPE_DATA:
 201		return rb_event_data_length(event);
 202	default:
 203		BUG();
 204	}
 205	/* not hit */
 206	return 0;
 207}
 208
 209/*
 210 * Return total length of time extend and data,
 211 *   or just the event length for all other events.
 212 */
 213static inline unsigned
 214rb_event_ts_length(struct ring_buffer_event *event)
 215{
 216	unsigned len = 0;
 217
 218	if (extended_time(event)) {
 219		/* time extends include the data event after it */
 220		len = RB_LEN_TIME_EXTEND;
 221		event = skip_time_extend(event);
 222	}
 223	return len + rb_event_length(event);
 224}
 225
 226/**
 227 * ring_buffer_event_length - return the length of the event
 228 * @event: the event to get the length of
 229 *
 230 * Returns the size of the data load of a data event.
 231 * If the event is something other than a data event, it
 232 * returns the size of the event itself. With the exception
 233 * of a TIME EXTEND, where it still returns the size of the
 234 * data load of the data event after it.
 235 */
 236unsigned ring_buffer_event_length(struct ring_buffer_event *event)
 237{
 238	unsigned length;
 239
 240	if (extended_time(event))
 241		event = skip_time_extend(event);
 242
 243	length = rb_event_length(event);
 244	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 245		return length;
 246	length -= RB_EVNT_HDR_SIZE;
 247	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
 248                length -= sizeof(event->array[0]);
 249	return length;
 250}
 251EXPORT_SYMBOL_GPL(ring_buffer_event_length);
 252
 253/* inline for ring buffer fast paths */
 254static __always_inline void *
 255rb_event_data(struct ring_buffer_event *event)
 256{
 257	if (extended_time(event))
 258		event = skip_time_extend(event);
 259	BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 260	/* If length is in len field, then array[0] has the data */
 261	if (event->type_len)
 262		return (void *)&event->array[0];
 263	/* Otherwise length is in array[0] and array[1] has the data */
 264	return (void *)&event->array[1];
 265}
 266
 267/**
 268 * ring_buffer_event_data - return the data of the event
 269 * @event: the event to get the data from
 270 */
 271void *ring_buffer_event_data(struct ring_buffer_event *event)
 272{
 273	return rb_event_data(event);
 274}
 275EXPORT_SYMBOL_GPL(ring_buffer_event_data);
 276
 277#define for_each_buffer_cpu(buffer, cpu)		\
 278	for_each_cpu(cpu, buffer->cpumask)
 279
 280#define TS_SHIFT	27
 281#define TS_MASK		((1ULL << TS_SHIFT) - 1)
 282#define TS_DELTA_TEST	(~TS_MASK)
 283
 284/**
 285 * ring_buffer_event_time_stamp - return the event's extended timestamp
 286 * @event: the event to get the timestamp of
 287 *
 288 * Returns the extended timestamp associated with a data event.
 289 * An extended time_stamp is a 64-bit timestamp represented
 290 * internally in a special way that makes the best use of space
 291 * contained within a ring buffer event.  This function decodes
 292 * it and maps it to a straight u64 value.
 293 */
 294u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
 295{
 296	u64 ts;
 297
 298	ts = event->array[0];
 299	ts <<= TS_SHIFT;
 300	ts += event->time_delta;
 301
 302	return ts;
 303}
 304
 305/* Flag when events were overwritten */
 306#define RB_MISSED_EVENTS	(1 << 31)
 307/* Missed count stored at end */
 308#define RB_MISSED_STORED	(1 << 30)
 309
 310#define RB_MISSED_FLAGS		(RB_MISSED_EVENTS|RB_MISSED_STORED)
 311
 312struct buffer_data_page {
 313	u64		 time_stamp;	/* page time stamp */
 314	local_t		 commit;	/* write committed index */
 315	unsigned char	 data[] RB_ALIGN_DATA;	/* data of buffer page */
 316};
 317
 318/*
 319 * Note, the buffer_page list must be first. The buffer pages
 320 * are allocated in cache lines, which means that each buffer
 321 * page will be at the beginning of a cache line, and thus
 322 * the least significant bits will be zero. We use this to
 323 * add flags in the list struct pointers, to make the ring buffer
 324 * lockless.
 325 */
 326struct buffer_page {
 327	struct list_head list;		/* list of buffer pages */
 328	local_t		 write;		/* index for next write */
 329	unsigned	 read;		/* index for next read */
 330	local_t		 entries;	/* entries on this page */
 331	unsigned long	 real_end;	/* real end of data */
 332	struct buffer_data_page *page;	/* Actual data page */
 333};
 334
 335/*
 336 * The buffer page counters, write and entries, must be reset
 337 * atomically when crossing page boundaries. To synchronize this
 338 * update, two counters are inserted into the number. One is
 339 * the actual counter for the write position or count on the page.
 340 *
 341 * The other is a counter of updaters. Before an update happens
 342 * the update partition of the counter is incremented. This will
 343 * allow the updater to update the counter atomically.
 344 *
 345 * The counter is 20 bits, and the state data is 12.
 346 */
 347#define RB_WRITE_MASK		0xfffff
 348#define RB_WRITE_INTCNT		(1 << 20)
 349
 350static void rb_init_page(struct buffer_data_page *bpage)
 351{
 352	local_set(&bpage->commit, 0);
 353}
 354
 355/**
 356 * ring_buffer_page_len - the size of data on the page.
 357 * @page: The page to read
 358 *
 359 * Returns the amount of data on the page, including buffer page header.
 360 */
 361size_t ring_buffer_page_len(void *page)
 362{
 363	struct buffer_data_page *bpage = page;
 364
 365	return (local_read(&bpage->commit) & ~RB_MISSED_FLAGS)
 366		+ BUF_PAGE_HDR_SIZE;
 367}
 368
 369/*
 370 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
 371 * this issue out.
 372 */
 373static void free_buffer_page(struct buffer_page *bpage)
 374{
 375	free_page((unsigned long)bpage->page);
 376	kfree(bpage);
 377}
 378
 379/*
 380 * We need to fit the time_stamp delta into 27 bits.
 381 */
 382static inline int test_time_stamp(u64 delta)
 383{
 384	if (delta & TS_DELTA_TEST)
 385		return 1;
 386	return 0;
 387}
 388
 389#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
 390
 391/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
 392#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
 393
 394int ring_buffer_print_page_header(struct trace_seq *s)
 395{
 396	struct buffer_data_page field;
 397
 398	trace_seq_printf(s, "\tfield: u64 timestamp;\t"
 399			 "offset:0;\tsize:%u;\tsigned:%u;\n",
 400			 (unsigned int)sizeof(field.time_stamp),
 401			 (unsigned int)is_signed_type(u64));
 402
 403	trace_seq_printf(s, "\tfield: local_t commit;\t"
 404			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 405			 (unsigned int)offsetof(typeof(field), commit),
 406			 (unsigned int)sizeof(field.commit),
 407			 (unsigned int)is_signed_type(long));
 408
 409	trace_seq_printf(s, "\tfield: int overwrite;\t"
 410			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 411			 (unsigned int)offsetof(typeof(field), commit),
 412			 1,
 413			 (unsigned int)is_signed_type(long));
 414
 415	trace_seq_printf(s, "\tfield: char data;\t"
 416			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 417			 (unsigned int)offsetof(typeof(field), data),
 418			 (unsigned int)BUF_PAGE_SIZE,
 419			 (unsigned int)is_signed_type(char));
 420
 421	return !trace_seq_has_overflowed(s);
 422}
 423
 424struct rb_irq_work {
 425	struct irq_work			work;
 426	wait_queue_head_t		waiters;
 427	wait_queue_head_t		full_waiters;
 428	bool				waiters_pending;
 429	bool				full_waiters_pending;
 430	bool				wakeup_full;
 431};
 432
 433/*
 434 * Structure to hold event state and handle nested events.
 435 */
 436struct rb_event_info {
 437	u64			ts;
 438	u64			delta;
 439	unsigned long		length;
 440	struct buffer_page	*tail_page;
 441	int			add_timestamp;
 442};
 443
 444/*
 445 * Used for which event context the event is in.
 446 *  NMI     = 0
 447 *  IRQ     = 1
 448 *  SOFTIRQ = 2
 449 *  NORMAL  = 3
 450 *
 451 * See trace_recursive_lock() comment below for more details.
 452 */
 453enum {
 454	RB_CTX_NMI,
 455	RB_CTX_IRQ,
 456	RB_CTX_SOFTIRQ,
 457	RB_CTX_NORMAL,
 458	RB_CTX_MAX
 459};
 460
 461/*
 462 * head_page == tail_page && head == tail then buffer is empty.
 463 */
 464struct ring_buffer_per_cpu {
 465	int				cpu;
 466	atomic_t			record_disabled;
 467	struct ring_buffer		*buffer;
 468	raw_spinlock_t			reader_lock;	/* serialize readers */
 469	arch_spinlock_t			lock;
 470	struct lock_class_key		lock_key;
 471	struct buffer_data_page		*free_page;
 472	unsigned long			nr_pages;
 473	unsigned int			current_context;
 474	struct list_head		*pages;
 475	struct buffer_page		*head_page;	/* read from head */
 476	struct buffer_page		*tail_page;	/* write to tail */
 477	struct buffer_page		*commit_page;	/* committed pages */
 478	struct buffer_page		*reader_page;
 479	unsigned long			lost_events;
 480	unsigned long			last_overrun;
 481	unsigned long			nest;
 482	local_t				entries_bytes;
 483	local_t				entries;
 484	local_t				overrun;
 485	local_t				commit_overrun;
 486	local_t				dropped_events;
 487	local_t				committing;
 488	local_t				commits;
 
 
 
 
 489	unsigned long			read;
 490	unsigned long			read_bytes;
 491	u64				write_stamp;
 492	u64				read_stamp;
 493	/* ring buffer pages to update, > 0 to add, < 0 to remove */
 494	long				nr_pages_to_update;
 495	struct list_head		new_pages; /* new pages to add */
 496	struct work_struct		update_pages_work;
 497	struct completion		update_done;
 498
 499	struct rb_irq_work		irq_work;
 500};
 501
 502struct ring_buffer {
 503	unsigned			flags;
 504	int				cpus;
 505	atomic_t			record_disabled;
 506	atomic_t			resize_disabled;
 507	cpumask_var_t			cpumask;
 508
 509	struct lock_class_key		*reader_lock_key;
 510
 511	struct mutex			mutex;
 512
 513	struct ring_buffer_per_cpu	**buffers;
 514
 515	struct hlist_node		node;
 516	u64				(*clock)(void);
 517
 518	struct rb_irq_work		irq_work;
 519	bool				time_stamp_abs;
 520};
 521
 522struct ring_buffer_iter {
 523	struct ring_buffer_per_cpu	*cpu_buffer;
 524	unsigned long			head;
 525	struct buffer_page		*head_page;
 526	struct buffer_page		*cache_reader_page;
 527	unsigned long			cache_read;
 528	u64				read_stamp;
 529};
 530
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 531/*
 532 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
 533 *
 534 * Schedules a delayed work to wake up any task that is blocked on the
 535 * ring buffer waiters queue.
 536 */
 537static void rb_wake_up_waiters(struct irq_work *work)
 538{
 539	struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
 540
 541	wake_up_all(&rbwork->waiters);
 542	if (rbwork->wakeup_full) {
 543		rbwork->wakeup_full = false;
 544		wake_up_all(&rbwork->full_waiters);
 545	}
 546}
 547
 548/**
 549 * ring_buffer_wait - wait for input to the ring buffer
 550 * @buffer: buffer to wait on
 551 * @cpu: the cpu buffer to wait on
 552 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
 553 *
 554 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 555 * as data is added to any of the @buffer's cpu buffers. Otherwise
 556 * it will wait for data to be added to a specific cpu buffer.
 557 */
 558int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
 559{
 560	struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
 561	DEFINE_WAIT(wait);
 562	struct rb_irq_work *work;
 563	int ret = 0;
 564
 565	/*
 566	 * Depending on what the caller is waiting for, either any
 567	 * data in any cpu buffer, or a specific buffer, put the
 568	 * caller on the appropriate wait queue.
 569	 */
 570	if (cpu == RING_BUFFER_ALL_CPUS) {
 571		work = &buffer->irq_work;
 572		/* Full only makes sense on per cpu reads */
 573		full = false;
 574	} else {
 575		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 576			return -ENODEV;
 577		cpu_buffer = buffer->buffers[cpu];
 578		work = &cpu_buffer->irq_work;
 579	}
 580
 581
 582	while (true) {
 583		if (full)
 584			prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
 585		else
 586			prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
 587
 588		/*
 589		 * The events can happen in critical sections where
 590		 * checking a work queue can cause deadlocks.
 591		 * After adding a task to the queue, this flag is set
 592		 * only to notify events to try to wake up the queue
 593		 * using irq_work.
 594		 *
 595		 * We don't clear it even if the buffer is no longer
 596		 * empty. The flag only causes the next event to run
 597		 * irq_work to do the work queue wake up. The worse
 598		 * that can happen if we race with !trace_empty() is that
 599		 * an event will cause an irq_work to try to wake up
 600		 * an empty queue.
 601		 *
 602		 * There's no reason to protect this flag either, as
 603		 * the work queue and irq_work logic will do the necessary
 604		 * synchronization for the wake ups. The only thing
 605		 * that is necessary is that the wake up happens after
 606		 * a task has been queued. It's OK for spurious wake ups.
 607		 */
 608		if (full)
 609			work->full_waiters_pending = true;
 610		else
 611			work->waiters_pending = true;
 612
 613		if (signal_pending(current)) {
 614			ret = -EINTR;
 615			break;
 616		}
 617
 618		if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
 619			break;
 620
 621		if (cpu != RING_BUFFER_ALL_CPUS &&
 622		    !ring_buffer_empty_cpu(buffer, cpu)) {
 623			unsigned long flags;
 624			bool pagebusy;
 
 
 625
 626			if (!full)
 627				break;
 628
 629			raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
 630			pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 
 
 
 
 
 631			raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
 632
 633			if (!pagebusy)
 634				break;
 635		}
 636
 637		schedule();
 638	}
 639
 640	if (full)
 641		finish_wait(&work->full_waiters, &wait);
 642	else
 643		finish_wait(&work->waiters, &wait);
 644
 645	return ret;
 646}
 647
 648/**
 649 * ring_buffer_poll_wait - poll on buffer input
 650 * @buffer: buffer to wait on
 651 * @cpu: the cpu buffer to wait on
 652 * @filp: the file descriptor
 653 * @poll_table: The poll descriptor
 654 *
 655 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 656 * as data is added to any of the @buffer's cpu buffers. Otherwise
 657 * it will wait for data to be added to a specific cpu buffer.
 658 *
 659 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
 660 * zero otherwise.
 661 */
 662__poll_t ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
 663			  struct file *filp, poll_table *poll_table)
 664{
 665	struct ring_buffer_per_cpu *cpu_buffer;
 666	struct rb_irq_work *work;
 667
 668	if (cpu == RING_BUFFER_ALL_CPUS)
 669		work = &buffer->irq_work;
 670	else {
 671		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 672			return -EINVAL;
 673
 674		cpu_buffer = buffer->buffers[cpu];
 675		work = &cpu_buffer->irq_work;
 676	}
 677
 678	poll_wait(filp, &work->waiters, poll_table);
 679	work->waiters_pending = true;
 680	/*
 681	 * There's a tight race between setting the waiters_pending and
 682	 * checking if the ring buffer is empty.  Once the waiters_pending bit
 683	 * is set, the next event will wake the task up, but we can get stuck
 684	 * if there's only a single event in.
 685	 *
 686	 * FIXME: Ideally, we need a memory barrier on the writer side as well,
 687	 * but adding a memory barrier to all events will cause too much of a
 688	 * performance hit in the fast path.  We only need a memory barrier when
 689	 * the buffer goes from empty to having content.  But as this race is
 690	 * extremely small, and it's not a problem if another event comes in, we
 691	 * will fix it later.
 692	 */
 693	smp_mb();
 694
 695	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
 696	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
 697		return EPOLLIN | EPOLLRDNORM;
 698	return 0;
 699}
 700
 701/* buffer may be either ring_buffer or ring_buffer_per_cpu */
 702#define RB_WARN_ON(b, cond)						\
 703	({								\
 704		int _____ret = unlikely(cond);				\
 705		if (_____ret) {						\
 706			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
 707				struct ring_buffer_per_cpu *__b =	\
 708					(void *)b;			\
 709				atomic_inc(&__b->buffer->record_disabled); \
 710			} else						\
 711				atomic_inc(&b->record_disabled);	\
 712			WARN_ON(1);					\
 713		}							\
 714		_____ret;						\
 715	})
 716
 717/* Up this if you want to test the TIME_EXTENTS and normalization */
 718#define DEBUG_SHIFT 0
 719
 720static inline u64 rb_time_stamp(struct ring_buffer *buffer)
 721{
 722	/* shift to debug/test normalization and TIME_EXTENTS */
 723	return buffer->clock() << DEBUG_SHIFT;
 724}
 725
 726u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
 727{
 728	u64 time;
 729
 730	preempt_disable_notrace();
 731	time = rb_time_stamp(buffer);
 732	preempt_enable_no_resched_notrace();
 733
 734	return time;
 735}
 736EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
 737
 738void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
 739				      int cpu, u64 *ts)
 740{
 741	/* Just stupid testing the normalize function and deltas */
 742	*ts >>= DEBUG_SHIFT;
 743}
 744EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
 745
 746/*
 747 * Making the ring buffer lockless makes things tricky.
 748 * Although writes only happen on the CPU that they are on,
 749 * and they only need to worry about interrupts. Reads can
 750 * happen on any CPU.
 751 *
 752 * The reader page is always off the ring buffer, but when the
 753 * reader finishes with a page, it needs to swap its page with
 754 * a new one from the buffer. The reader needs to take from
 755 * the head (writes go to the tail). But if a writer is in overwrite
 756 * mode and wraps, it must push the head page forward.
 757 *
 758 * Here lies the problem.
 759 *
 760 * The reader must be careful to replace only the head page, and
 761 * not another one. As described at the top of the file in the
 762 * ASCII art, the reader sets its old page to point to the next
 763 * page after head. It then sets the page after head to point to
 764 * the old reader page. But if the writer moves the head page
 765 * during this operation, the reader could end up with the tail.
 766 *
 767 * We use cmpxchg to help prevent this race. We also do something
 768 * special with the page before head. We set the LSB to 1.
 769 *
 770 * When the writer must push the page forward, it will clear the
 771 * bit that points to the head page, move the head, and then set
 772 * the bit that points to the new head page.
 773 *
 774 * We also don't want an interrupt coming in and moving the head
 775 * page on another writer. Thus we use the second LSB to catch
 776 * that too. Thus:
 777 *
 778 * head->list->prev->next        bit 1          bit 0
 779 *                              -------        -------
 780 * Normal page                     0              0
 781 * Points to head page             0              1
 782 * New head page                   1              0
 783 *
 784 * Note we can not trust the prev pointer of the head page, because:
 785 *
 786 * +----+       +-----+        +-----+
 787 * |    |------>|  T  |---X--->|  N  |
 788 * |    |<------|     |        |     |
 789 * +----+       +-----+        +-----+
 790 *   ^                           ^ |
 791 *   |          +-----+          | |
 792 *   +----------|  R  |----------+ |
 793 *              |     |<-----------+
 794 *              +-----+
 795 *
 796 * Key:  ---X-->  HEAD flag set in pointer
 797 *         T      Tail page
 798 *         R      Reader page
 799 *         N      Next page
 800 *
 801 * (see __rb_reserve_next() to see where this happens)
 802 *
 803 *  What the above shows is that the reader just swapped out
 804 *  the reader page with a page in the buffer, but before it
 805 *  could make the new header point back to the new page added
 806 *  it was preempted by a writer. The writer moved forward onto
 807 *  the new page added by the reader and is about to move forward
 808 *  again.
 809 *
 810 *  You can see, it is legitimate for the previous pointer of
 811 *  the head (or any page) not to point back to itself. But only
 812 *  temporarially.
 813 */
 814
 815#define RB_PAGE_NORMAL		0UL
 816#define RB_PAGE_HEAD		1UL
 817#define RB_PAGE_UPDATE		2UL
 818
 819
 820#define RB_FLAG_MASK		3UL
 821
 822/* PAGE_MOVED is not part of the mask */
 823#define RB_PAGE_MOVED		4UL
 824
 825/*
 826 * rb_list_head - remove any bit
 827 */
 828static struct list_head *rb_list_head(struct list_head *list)
 829{
 830	unsigned long val = (unsigned long)list;
 831
 832	return (struct list_head *)(val & ~RB_FLAG_MASK);
 833}
 834
 835/*
 836 * rb_is_head_page - test if the given page is the head page
 837 *
 838 * Because the reader may move the head_page pointer, we can
 839 * not trust what the head page is (it may be pointing to
 840 * the reader page). But if the next page is a header page,
 841 * its flags will be non zero.
 842 */
 843static inline int
 844rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
 845		struct buffer_page *page, struct list_head *list)
 846{
 847	unsigned long val;
 848
 849	val = (unsigned long)list->next;
 850
 851	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
 852		return RB_PAGE_MOVED;
 853
 854	return val & RB_FLAG_MASK;
 855}
 856
 857/*
 858 * rb_is_reader_page
 859 *
 860 * The unique thing about the reader page, is that, if the
 861 * writer is ever on it, the previous pointer never points
 862 * back to the reader page.
 863 */
 864static bool rb_is_reader_page(struct buffer_page *page)
 865{
 866	struct list_head *list = page->list.prev;
 867
 868	return rb_list_head(list->next) != &page->list;
 869}
 870
 871/*
 872 * rb_set_list_to_head - set a list_head to be pointing to head.
 873 */
 874static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
 875				struct list_head *list)
 876{
 877	unsigned long *ptr;
 878
 879	ptr = (unsigned long *)&list->next;
 880	*ptr |= RB_PAGE_HEAD;
 881	*ptr &= ~RB_PAGE_UPDATE;
 882}
 883
 884/*
 885 * rb_head_page_activate - sets up head page
 886 */
 887static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
 888{
 889	struct buffer_page *head;
 890
 891	head = cpu_buffer->head_page;
 892	if (!head)
 893		return;
 894
 895	/*
 896	 * Set the previous list pointer to have the HEAD flag.
 897	 */
 898	rb_set_list_to_head(cpu_buffer, head->list.prev);
 899}
 900
 901static void rb_list_head_clear(struct list_head *list)
 902{
 903	unsigned long *ptr = (unsigned long *)&list->next;
 904
 905	*ptr &= ~RB_FLAG_MASK;
 906}
 907
 908/*
 909 * rb_head_page_dactivate - clears head page ptr (for free list)
 910 */
 911static void
 912rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
 913{
 914	struct list_head *hd;
 915
 916	/* Go through the whole list and clear any pointers found. */
 917	rb_list_head_clear(cpu_buffer->pages);
 918
 919	list_for_each(hd, cpu_buffer->pages)
 920		rb_list_head_clear(hd);
 921}
 922
 923static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
 924			    struct buffer_page *head,
 925			    struct buffer_page *prev,
 926			    int old_flag, int new_flag)
 927{
 928	struct list_head *list;
 929	unsigned long val = (unsigned long)&head->list;
 930	unsigned long ret;
 931
 932	list = &prev->list;
 933
 934	val &= ~RB_FLAG_MASK;
 935
 936	ret = cmpxchg((unsigned long *)&list->next,
 937		      val | old_flag, val | new_flag);
 938
 939	/* check if the reader took the page */
 940	if ((ret & ~RB_FLAG_MASK) != val)
 941		return RB_PAGE_MOVED;
 942
 943	return ret & RB_FLAG_MASK;
 944}
 945
 946static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
 947				   struct buffer_page *head,
 948				   struct buffer_page *prev,
 949				   int old_flag)
 950{
 951	return rb_head_page_set(cpu_buffer, head, prev,
 952				old_flag, RB_PAGE_UPDATE);
 953}
 954
 955static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
 956				 struct buffer_page *head,
 957				 struct buffer_page *prev,
 958				 int old_flag)
 959{
 960	return rb_head_page_set(cpu_buffer, head, prev,
 961				old_flag, RB_PAGE_HEAD);
 962}
 963
 964static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
 965				   struct buffer_page *head,
 966				   struct buffer_page *prev,
 967				   int old_flag)
 968{
 969	return rb_head_page_set(cpu_buffer, head, prev,
 970				old_flag, RB_PAGE_NORMAL);
 971}
 972
 973static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
 974			       struct buffer_page **bpage)
 975{
 976	struct list_head *p = rb_list_head((*bpage)->list.next);
 977
 978	*bpage = list_entry(p, struct buffer_page, list);
 979}
 980
 981static struct buffer_page *
 982rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
 983{
 984	struct buffer_page *head;
 985	struct buffer_page *page;
 986	struct list_head *list;
 987	int i;
 988
 989	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
 990		return NULL;
 991
 992	/* sanity check */
 993	list = cpu_buffer->pages;
 994	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
 995		return NULL;
 996
 997	page = head = cpu_buffer->head_page;
 998	/*
 999	 * It is possible that the writer moves the header behind
1000	 * where we started, and we miss in one loop.
1001	 * A second loop should grab the header, but we'll do
1002	 * three loops just because I'm paranoid.
1003	 */
1004	for (i = 0; i < 3; i++) {
1005		do {
1006			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1007				cpu_buffer->head_page = page;
1008				return page;
1009			}
1010			rb_inc_page(cpu_buffer, &page);
1011		} while (page != head);
1012	}
1013
1014	RB_WARN_ON(cpu_buffer, 1);
1015
1016	return NULL;
1017}
1018
1019static int rb_head_page_replace(struct buffer_page *old,
1020				struct buffer_page *new)
1021{
1022	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1023	unsigned long val;
1024	unsigned long ret;
1025
1026	val = *ptr & ~RB_FLAG_MASK;
1027	val |= RB_PAGE_HEAD;
1028
1029	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1030
1031	return ret == val;
1032}
1033
1034/*
1035 * rb_tail_page_update - move the tail page forward
1036 */
1037static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1038			       struct buffer_page *tail_page,
1039			       struct buffer_page *next_page)
1040{
1041	unsigned long old_entries;
1042	unsigned long old_write;
1043
1044	/*
1045	 * The tail page now needs to be moved forward.
1046	 *
1047	 * We need to reset the tail page, but without messing
1048	 * with possible erasing of data brought in by interrupts
1049	 * that have moved the tail page and are currently on it.
1050	 *
1051	 * We add a counter to the write field to denote this.
1052	 */
1053	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1054	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1055
 
1056	/*
1057	 * Just make sure we have seen our old_write and synchronize
1058	 * with any interrupts that come in.
1059	 */
1060	barrier();
1061
1062	/*
1063	 * If the tail page is still the same as what we think
1064	 * it is, then it is up to us to update the tail
1065	 * pointer.
1066	 */
1067	if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1068		/* Zero the write counter */
1069		unsigned long val = old_write & ~RB_WRITE_MASK;
1070		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1071
1072		/*
1073		 * This will only succeed if an interrupt did
1074		 * not come in and change it. In which case, we
1075		 * do not want to modify it.
1076		 *
1077		 * We add (void) to let the compiler know that we do not care
1078		 * about the return value of these functions. We use the
1079		 * cmpxchg to only update if an interrupt did not already
1080		 * do it for us. If the cmpxchg fails, we don't care.
1081		 */
1082		(void)local_cmpxchg(&next_page->write, old_write, val);
1083		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1084
1085		/*
1086		 * No need to worry about races with clearing out the commit.
1087		 * it only can increment when a commit takes place. But that
1088		 * only happens in the outer most nested commit.
1089		 */
1090		local_set(&next_page->page->commit, 0);
1091
1092		/* Again, either we update tail_page or an interrupt does */
1093		(void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1094	}
1095}
1096
1097static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1098			  struct buffer_page *bpage)
1099{
1100	unsigned long val = (unsigned long)bpage;
1101
1102	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1103		return 1;
1104
1105	return 0;
1106}
1107
1108/**
1109 * rb_check_list - make sure a pointer to a list has the last bits zero
1110 */
1111static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1112			 struct list_head *list)
1113{
1114	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1115		return 1;
1116	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1117		return 1;
1118	return 0;
1119}
1120
1121/**
1122 * rb_check_pages - integrity check of buffer pages
1123 * @cpu_buffer: CPU buffer with pages to test
1124 *
1125 * As a safety measure we check to make sure the data pages have not
1126 * been corrupted.
1127 */
1128static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1129{
1130	struct list_head *head = cpu_buffer->pages;
1131	struct buffer_page *bpage, *tmp;
1132
1133	/* Reset the head page if it exists */
1134	if (cpu_buffer->head_page)
1135		rb_set_head_page(cpu_buffer);
1136
1137	rb_head_page_deactivate(cpu_buffer);
1138
1139	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1140		return -1;
1141	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1142		return -1;
1143
1144	if (rb_check_list(cpu_buffer, head))
1145		return -1;
1146
1147	list_for_each_entry_safe(bpage, tmp, head, list) {
1148		if (RB_WARN_ON(cpu_buffer,
1149			       bpage->list.next->prev != &bpage->list))
1150			return -1;
1151		if (RB_WARN_ON(cpu_buffer,
1152			       bpage->list.prev->next != &bpage->list))
1153			return -1;
1154		if (rb_check_list(cpu_buffer, &bpage->list))
1155			return -1;
1156	}
1157
1158	rb_head_page_activate(cpu_buffer);
1159
1160	return 0;
1161}
1162
1163static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1164{
1165	struct buffer_page *bpage, *tmp;
1166	bool user_thread = current->mm != NULL;
1167	gfp_t mflags;
1168	long i;
1169
1170	/*
1171	 * Check if the available memory is there first.
1172	 * Note, si_mem_available() only gives us a rough estimate of available
1173	 * memory. It may not be accurate. But we don't care, we just want
1174	 * to prevent doing any allocation when it is obvious that it is
1175	 * not going to succeed.
1176	 */
1177	i = si_mem_available();
1178	if (i < nr_pages)
1179		return -ENOMEM;
1180
1181	/*
1182	 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1183	 * gracefully without invoking oom-killer and the system is not
1184	 * destabilized.
1185	 */
1186	mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1187
1188	/*
1189	 * If a user thread allocates too much, and si_mem_available()
1190	 * reports there's enough memory, even though there is not.
1191	 * Make sure the OOM killer kills this thread. This can happen
1192	 * even with RETRY_MAYFAIL because another task may be doing
1193	 * an allocation after this task has taken all memory.
1194	 * This is the task the OOM killer needs to take out during this
1195	 * loop, even if it was triggered by an allocation somewhere else.
1196	 */
1197	if (user_thread)
1198		set_current_oom_origin();
1199	for (i = 0; i < nr_pages; i++) {
1200		struct page *page;
1201
1202		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1203				    mflags, cpu_to_node(cpu));
1204		if (!bpage)
1205			goto free_pages;
1206
1207		list_add(&bpage->list, pages);
1208
1209		page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
1210		if (!page)
1211			goto free_pages;
1212		bpage->page = page_address(page);
1213		rb_init_page(bpage->page);
1214
1215		if (user_thread && fatal_signal_pending(current))
1216			goto free_pages;
1217	}
1218	if (user_thread)
1219		clear_current_oom_origin();
1220
1221	return 0;
1222
1223free_pages:
1224	list_for_each_entry_safe(bpage, tmp, pages, list) {
1225		list_del_init(&bpage->list);
1226		free_buffer_page(bpage);
1227	}
1228	if (user_thread)
1229		clear_current_oom_origin();
1230
1231	return -ENOMEM;
1232}
1233
1234static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1235			     unsigned long nr_pages)
1236{
1237	LIST_HEAD(pages);
1238
1239	WARN_ON(!nr_pages);
1240
1241	if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1242		return -ENOMEM;
1243
1244	/*
1245	 * The ring buffer page list is a circular list that does not
1246	 * start and end with a list head. All page list items point to
1247	 * other pages.
1248	 */
1249	cpu_buffer->pages = pages.next;
1250	list_del(&pages);
1251
1252	cpu_buffer->nr_pages = nr_pages;
1253
1254	rb_check_pages(cpu_buffer);
1255
1256	return 0;
1257}
1258
1259static struct ring_buffer_per_cpu *
1260rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1261{
1262	struct ring_buffer_per_cpu *cpu_buffer;
1263	struct buffer_page *bpage;
1264	struct page *page;
1265	int ret;
1266
1267	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1268				  GFP_KERNEL, cpu_to_node(cpu));
1269	if (!cpu_buffer)
1270		return NULL;
1271
1272	cpu_buffer->cpu = cpu;
1273	cpu_buffer->buffer = buffer;
1274	raw_spin_lock_init(&cpu_buffer->reader_lock);
1275	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1276	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1277	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1278	init_completion(&cpu_buffer->update_done);
1279	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1280	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1281	init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1282
1283	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1284			    GFP_KERNEL, cpu_to_node(cpu));
1285	if (!bpage)
1286		goto fail_free_buffer;
1287
1288	rb_check_bpage(cpu_buffer, bpage);
1289
1290	cpu_buffer->reader_page = bpage;
1291	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1292	if (!page)
1293		goto fail_free_reader;
1294	bpage->page = page_address(page);
1295	rb_init_page(bpage->page);
1296
1297	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1298	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1299
1300	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1301	if (ret < 0)
1302		goto fail_free_reader;
1303
1304	cpu_buffer->head_page
1305		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1306	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1307
1308	rb_head_page_activate(cpu_buffer);
1309
1310	return cpu_buffer;
1311
1312 fail_free_reader:
1313	free_buffer_page(cpu_buffer->reader_page);
1314
1315 fail_free_buffer:
1316	kfree(cpu_buffer);
1317	return NULL;
1318}
1319
1320static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1321{
1322	struct list_head *head = cpu_buffer->pages;
1323	struct buffer_page *bpage, *tmp;
1324
1325	free_buffer_page(cpu_buffer->reader_page);
1326
1327	rb_head_page_deactivate(cpu_buffer);
1328
1329	if (head) {
1330		list_for_each_entry_safe(bpage, tmp, head, list) {
1331			list_del_init(&bpage->list);
1332			free_buffer_page(bpage);
1333		}
1334		bpage = list_entry(head, struct buffer_page, list);
1335		free_buffer_page(bpage);
1336	}
1337
1338	kfree(cpu_buffer);
1339}
1340
1341/**
1342 * __ring_buffer_alloc - allocate a new ring_buffer
1343 * @size: the size in bytes per cpu that is needed.
1344 * @flags: attributes to set for the ring buffer.
1345 *
1346 * Currently the only flag that is available is the RB_FL_OVERWRITE
1347 * flag. This flag means that the buffer will overwrite old data
1348 * when the buffer wraps. If this flag is not set, the buffer will
1349 * drop data when the tail hits the head.
1350 */
1351struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1352					struct lock_class_key *key)
1353{
1354	struct ring_buffer *buffer;
1355	long nr_pages;
1356	int bsize;
1357	int cpu;
1358	int ret;
1359
1360	/* keep it in its own cache line */
1361	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1362			 GFP_KERNEL);
1363	if (!buffer)
1364		return NULL;
1365
1366	if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1367		goto fail_free_buffer;
1368
1369	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1370	buffer->flags = flags;
1371	buffer->clock = trace_clock_local;
1372	buffer->reader_lock_key = key;
1373
1374	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1375	init_waitqueue_head(&buffer->irq_work.waiters);
1376
1377	/* need at least two pages */
1378	if (nr_pages < 2)
1379		nr_pages = 2;
1380
1381	buffer->cpus = nr_cpu_ids;
1382
1383	bsize = sizeof(void *) * nr_cpu_ids;
1384	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1385				  GFP_KERNEL);
1386	if (!buffer->buffers)
1387		goto fail_free_cpumask;
1388
1389	cpu = raw_smp_processor_id();
1390	cpumask_set_cpu(cpu, buffer->cpumask);
1391	buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1392	if (!buffer->buffers[cpu])
1393		goto fail_free_buffers;
1394
1395	ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1396	if (ret < 0)
1397		goto fail_free_buffers;
1398
1399	mutex_init(&buffer->mutex);
1400
1401	return buffer;
1402
1403 fail_free_buffers:
1404	for_each_buffer_cpu(buffer, cpu) {
1405		if (buffer->buffers[cpu])
1406			rb_free_cpu_buffer(buffer->buffers[cpu]);
1407	}
1408	kfree(buffer->buffers);
1409
1410 fail_free_cpumask:
1411	free_cpumask_var(buffer->cpumask);
1412
1413 fail_free_buffer:
1414	kfree(buffer);
1415	return NULL;
1416}
1417EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1418
1419/**
1420 * ring_buffer_free - free a ring buffer.
1421 * @buffer: the buffer to free.
1422 */
1423void
1424ring_buffer_free(struct ring_buffer *buffer)
1425{
1426	int cpu;
1427
1428	cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1429
1430	for_each_buffer_cpu(buffer, cpu)
1431		rb_free_cpu_buffer(buffer->buffers[cpu]);
1432
1433	kfree(buffer->buffers);
1434	free_cpumask_var(buffer->cpumask);
1435
1436	kfree(buffer);
1437}
1438EXPORT_SYMBOL_GPL(ring_buffer_free);
1439
1440void ring_buffer_set_clock(struct ring_buffer *buffer,
1441			   u64 (*clock)(void))
1442{
1443	buffer->clock = clock;
1444}
1445
1446void ring_buffer_set_time_stamp_abs(struct ring_buffer *buffer, bool abs)
1447{
1448	buffer->time_stamp_abs = abs;
1449}
1450
1451bool ring_buffer_time_stamp_abs(struct ring_buffer *buffer)
1452{
1453	return buffer->time_stamp_abs;
1454}
1455
1456static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1457
1458static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1459{
1460	return local_read(&bpage->entries) & RB_WRITE_MASK;
1461}
1462
1463static inline unsigned long rb_page_write(struct buffer_page *bpage)
1464{
1465	return local_read(&bpage->write) & RB_WRITE_MASK;
1466}
1467
1468static int
1469rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1470{
1471	struct list_head *tail_page, *to_remove, *next_page;
1472	struct buffer_page *to_remove_page, *tmp_iter_page;
1473	struct buffer_page *last_page, *first_page;
1474	unsigned long nr_removed;
1475	unsigned long head_bit;
1476	int page_entries;
1477
1478	head_bit = 0;
1479
1480	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1481	atomic_inc(&cpu_buffer->record_disabled);
1482	/*
1483	 * We don't race with the readers since we have acquired the reader
1484	 * lock. We also don't race with writers after disabling recording.
1485	 * This makes it easy to figure out the first and the last page to be
1486	 * removed from the list. We unlink all the pages in between including
1487	 * the first and last pages. This is done in a busy loop so that we
1488	 * lose the least number of traces.
1489	 * The pages are freed after we restart recording and unlock readers.
1490	 */
1491	tail_page = &cpu_buffer->tail_page->list;
1492
1493	/*
1494	 * tail page might be on reader page, we remove the next page
1495	 * from the ring buffer
1496	 */
1497	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1498		tail_page = rb_list_head(tail_page->next);
1499	to_remove = tail_page;
1500
1501	/* start of pages to remove */
1502	first_page = list_entry(rb_list_head(to_remove->next),
1503				struct buffer_page, list);
1504
1505	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1506		to_remove = rb_list_head(to_remove)->next;
1507		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1508	}
1509
1510	next_page = rb_list_head(to_remove)->next;
1511
1512	/*
1513	 * Now we remove all pages between tail_page and next_page.
1514	 * Make sure that we have head_bit value preserved for the
1515	 * next page
1516	 */
1517	tail_page->next = (struct list_head *)((unsigned long)next_page |
1518						head_bit);
1519	next_page = rb_list_head(next_page);
1520	next_page->prev = tail_page;
1521
1522	/* make sure pages points to a valid page in the ring buffer */
1523	cpu_buffer->pages = next_page;
1524
1525	/* update head page */
1526	if (head_bit)
1527		cpu_buffer->head_page = list_entry(next_page,
1528						struct buffer_page, list);
1529
1530	/*
1531	 * change read pointer to make sure any read iterators reset
1532	 * themselves
1533	 */
1534	cpu_buffer->read = 0;
1535
1536	/* pages are removed, resume tracing and then free the pages */
1537	atomic_dec(&cpu_buffer->record_disabled);
1538	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1539
1540	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1541
1542	/* last buffer page to remove */
1543	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1544				list);
1545	tmp_iter_page = first_page;
1546
1547	do {
 
 
1548		to_remove_page = tmp_iter_page;
1549		rb_inc_page(cpu_buffer, &tmp_iter_page);
1550
1551		/* update the counters */
1552		page_entries = rb_page_entries(to_remove_page);
1553		if (page_entries) {
1554			/*
1555			 * If something was added to this page, it was full
1556			 * since it is not the tail page. So we deduct the
1557			 * bytes consumed in ring buffer from here.
1558			 * Increment overrun to account for the lost events.
1559			 */
1560			local_add(page_entries, &cpu_buffer->overrun);
1561			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1562		}
1563
1564		/*
1565		 * We have already removed references to this list item, just
1566		 * free up the buffer_page and its page
1567		 */
1568		free_buffer_page(to_remove_page);
1569		nr_removed--;
1570
1571	} while (to_remove_page != last_page);
1572
1573	RB_WARN_ON(cpu_buffer, nr_removed);
1574
1575	return nr_removed == 0;
1576}
1577
1578static int
1579rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1580{
1581	struct list_head *pages = &cpu_buffer->new_pages;
1582	int retries, success;
1583
1584	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1585	/*
1586	 * We are holding the reader lock, so the reader page won't be swapped
1587	 * in the ring buffer. Now we are racing with the writer trying to
1588	 * move head page and the tail page.
1589	 * We are going to adapt the reader page update process where:
1590	 * 1. We first splice the start and end of list of new pages between
1591	 *    the head page and its previous page.
1592	 * 2. We cmpxchg the prev_page->next to point from head page to the
1593	 *    start of new pages list.
1594	 * 3. Finally, we update the head->prev to the end of new list.
1595	 *
1596	 * We will try this process 10 times, to make sure that we don't keep
1597	 * spinning.
1598	 */
1599	retries = 10;
1600	success = 0;
1601	while (retries--) {
1602		struct list_head *head_page, *prev_page, *r;
1603		struct list_head *last_page, *first_page;
1604		struct list_head *head_page_with_bit;
1605
1606		head_page = &rb_set_head_page(cpu_buffer)->list;
1607		if (!head_page)
1608			break;
1609		prev_page = head_page->prev;
1610
1611		first_page = pages->next;
1612		last_page  = pages->prev;
1613
1614		head_page_with_bit = (struct list_head *)
1615				     ((unsigned long)head_page | RB_PAGE_HEAD);
1616
1617		last_page->next = head_page_with_bit;
1618		first_page->prev = prev_page;
1619
1620		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1621
1622		if (r == head_page_with_bit) {
1623			/*
1624			 * yay, we replaced the page pointer to our new list,
1625			 * now, we just have to update to head page's prev
1626			 * pointer to point to end of list
1627			 */
1628			head_page->prev = last_page;
1629			success = 1;
1630			break;
1631		}
1632	}
1633
1634	if (success)
1635		INIT_LIST_HEAD(pages);
1636	/*
1637	 * If we weren't successful in adding in new pages, warn and stop
1638	 * tracing
1639	 */
1640	RB_WARN_ON(cpu_buffer, !success);
1641	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1642
1643	/* free pages if they weren't inserted */
1644	if (!success) {
1645		struct buffer_page *bpage, *tmp;
1646		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1647					 list) {
1648			list_del_init(&bpage->list);
1649			free_buffer_page(bpage);
1650		}
1651	}
1652	return success;
1653}
1654
1655static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1656{
1657	int success;
1658
1659	if (cpu_buffer->nr_pages_to_update > 0)
1660		success = rb_insert_pages(cpu_buffer);
1661	else
1662		success = rb_remove_pages(cpu_buffer,
1663					-cpu_buffer->nr_pages_to_update);
1664
1665	if (success)
1666		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1667}
1668
1669static void update_pages_handler(struct work_struct *work)
1670{
1671	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1672			struct ring_buffer_per_cpu, update_pages_work);
1673	rb_update_pages(cpu_buffer);
1674	complete(&cpu_buffer->update_done);
1675}
1676
1677/**
1678 * ring_buffer_resize - resize the ring buffer
1679 * @buffer: the buffer to resize.
1680 * @size: the new size.
1681 * @cpu_id: the cpu buffer to resize
1682 *
1683 * Minimum size is 2 * BUF_PAGE_SIZE.
1684 *
1685 * Returns 0 on success and < 0 on failure.
1686 */
1687int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1688			int cpu_id)
1689{
1690	struct ring_buffer_per_cpu *cpu_buffer;
1691	unsigned long nr_pages;
1692	int cpu, err = 0;
1693
1694	/*
1695	 * Always succeed at resizing a non-existent buffer:
1696	 */
1697	if (!buffer)
1698		return size;
1699
1700	/* Make sure the requested buffer exists */
1701	if (cpu_id != RING_BUFFER_ALL_CPUS &&
1702	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
1703		return size;
1704
1705	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1706
1707	/* we need a minimum of two pages */
1708	if (nr_pages < 2)
1709		nr_pages = 2;
1710
1711	size = nr_pages * BUF_PAGE_SIZE;
1712
1713	/*
1714	 * Don't succeed if resizing is disabled, as a reader might be
1715	 * manipulating the ring buffer and is expecting a sane state while
1716	 * this is true.
1717	 */
1718	if (atomic_read(&buffer->resize_disabled))
1719		return -EBUSY;
1720
1721	/* prevent another thread from changing buffer sizes */
1722	mutex_lock(&buffer->mutex);
1723
1724	if (cpu_id == RING_BUFFER_ALL_CPUS) {
1725		/* calculate the pages to update */
1726		for_each_buffer_cpu(buffer, cpu) {
1727			cpu_buffer = buffer->buffers[cpu];
1728
1729			cpu_buffer->nr_pages_to_update = nr_pages -
1730							cpu_buffer->nr_pages;
1731			/*
1732			 * nothing more to do for removing pages or no update
1733			 */
1734			if (cpu_buffer->nr_pages_to_update <= 0)
1735				continue;
1736			/*
1737			 * to add pages, make sure all new pages can be
1738			 * allocated without receiving ENOMEM
1739			 */
1740			INIT_LIST_HEAD(&cpu_buffer->new_pages);
1741			if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1742						&cpu_buffer->new_pages, cpu)) {
1743				/* not enough memory for new pages */
1744				err = -ENOMEM;
1745				goto out_err;
1746			}
1747		}
1748
1749		get_online_cpus();
1750		/*
1751		 * Fire off all the required work handlers
1752		 * We can't schedule on offline CPUs, but it's not necessary
1753		 * since we can change their buffer sizes without any race.
1754		 */
1755		for_each_buffer_cpu(buffer, cpu) {
1756			cpu_buffer = buffer->buffers[cpu];
1757			if (!cpu_buffer->nr_pages_to_update)
1758				continue;
1759
1760			/* Can't run something on an offline CPU. */
1761			if (!cpu_online(cpu)) {
1762				rb_update_pages(cpu_buffer);
1763				cpu_buffer->nr_pages_to_update = 0;
1764			} else {
1765				schedule_work_on(cpu,
1766						&cpu_buffer->update_pages_work);
1767			}
1768		}
1769
1770		/* wait for all the updates to complete */
1771		for_each_buffer_cpu(buffer, cpu) {
1772			cpu_buffer = buffer->buffers[cpu];
1773			if (!cpu_buffer->nr_pages_to_update)
1774				continue;
1775
1776			if (cpu_online(cpu))
1777				wait_for_completion(&cpu_buffer->update_done);
1778			cpu_buffer->nr_pages_to_update = 0;
1779		}
1780
1781		put_online_cpus();
1782	} else {
1783		/* Make sure this CPU has been intitialized */
1784		if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1785			goto out;
1786
1787		cpu_buffer = buffer->buffers[cpu_id];
1788
1789		if (nr_pages == cpu_buffer->nr_pages)
1790			goto out;
1791
1792		cpu_buffer->nr_pages_to_update = nr_pages -
1793						cpu_buffer->nr_pages;
1794
1795		INIT_LIST_HEAD(&cpu_buffer->new_pages);
1796		if (cpu_buffer->nr_pages_to_update > 0 &&
1797			__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1798					    &cpu_buffer->new_pages, cpu_id)) {
1799			err = -ENOMEM;
1800			goto out_err;
1801		}
1802
1803		get_online_cpus();
1804
1805		/* Can't run something on an offline CPU. */
1806		if (!cpu_online(cpu_id))
1807			rb_update_pages(cpu_buffer);
1808		else {
1809			schedule_work_on(cpu_id,
1810					 &cpu_buffer->update_pages_work);
1811			wait_for_completion(&cpu_buffer->update_done);
1812		}
1813
1814		cpu_buffer->nr_pages_to_update = 0;
1815		put_online_cpus();
1816	}
1817
1818 out:
1819	/*
1820	 * The ring buffer resize can happen with the ring buffer
1821	 * enabled, so that the update disturbs the tracing as little
1822	 * as possible. But if the buffer is disabled, we do not need
1823	 * to worry about that, and we can take the time to verify
1824	 * that the buffer is not corrupt.
1825	 */
1826	if (atomic_read(&buffer->record_disabled)) {
1827		atomic_inc(&buffer->record_disabled);
1828		/*
1829		 * Even though the buffer was disabled, we must make sure
1830		 * that it is truly disabled before calling rb_check_pages.
1831		 * There could have been a race between checking
1832		 * record_disable and incrementing it.
1833		 */
1834		synchronize_sched();
1835		for_each_buffer_cpu(buffer, cpu) {
1836			cpu_buffer = buffer->buffers[cpu];
1837			rb_check_pages(cpu_buffer);
1838		}
1839		atomic_dec(&buffer->record_disabled);
1840	}
1841
1842	mutex_unlock(&buffer->mutex);
1843	return size;
1844
1845 out_err:
1846	for_each_buffer_cpu(buffer, cpu) {
1847		struct buffer_page *bpage, *tmp;
1848
1849		cpu_buffer = buffer->buffers[cpu];
1850		cpu_buffer->nr_pages_to_update = 0;
1851
1852		if (list_empty(&cpu_buffer->new_pages))
1853			continue;
1854
1855		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1856					list) {
1857			list_del_init(&bpage->list);
1858			free_buffer_page(bpage);
1859		}
1860	}
1861	mutex_unlock(&buffer->mutex);
1862	return err;
1863}
1864EXPORT_SYMBOL_GPL(ring_buffer_resize);
1865
1866void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1867{
1868	mutex_lock(&buffer->mutex);
1869	if (val)
1870		buffer->flags |= RB_FL_OVERWRITE;
1871	else
1872		buffer->flags &= ~RB_FL_OVERWRITE;
1873	mutex_unlock(&buffer->mutex);
1874}
1875EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1876
1877static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1878{
1879	return bpage->page->data + index;
1880}
1881
1882static __always_inline struct ring_buffer_event *
1883rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1884{
1885	return __rb_page_index(cpu_buffer->reader_page,
1886			       cpu_buffer->reader_page->read);
1887}
1888
1889static __always_inline struct ring_buffer_event *
1890rb_iter_head_event(struct ring_buffer_iter *iter)
1891{
1892	return __rb_page_index(iter->head_page, iter->head);
1893}
1894
1895static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1896{
1897	return local_read(&bpage->page->commit);
1898}
1899
1900/* Size is determined by what has been committed */
1901static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1902{
1903	return rb_page_commit(bpage);
1904}
1905
1906static __always_inline unsigned
1907rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1908{
1909	return rb_page_commit(cpu_buffer->commit_page);
1910}
1911
1912static __always_inline unsigned
1913rb_event_index(struct ring_buffer_event *event)
1914{
1915	unsigned long addr = (unsigned long)event;
1916
1917	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1918}
1919
1920static void rb_inc_iter(struct ring_buffer_iter *iter)
1921{
1922	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1923
1924	/*
1925	 * The iterator could be on the reader page (it starts there).
1926	 * But the head could have moved, since the reader was
1927	 * found. Check for this case and assign the iterator
1928	 * to the head page instead of next.
1929	 */
1930	if (iter->head_page == cpu_buffer->reader_page)
1931		iter->head_page = rb_set_head_page(cpu_buffer);
1932	else
1933		rb_inc_page(cpu_buffer, &iter->head_page);
1934
1935	iter->read_stamp = iter->head_page->page->time_stamp;
1936	iter->head = 0;
1937}
1938
1939/*
1940 * rb_handle_head_page - writer hit the head page
1941 *
1942 * Returns: +1 to retry page
1943 *           0 to continue
1944 *          -1 on error
1945 */
1946static int
1947rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1948		    struct buffer_page *tail_page,
1949		    struct buffer_page *next_page)
1950{
1951	struct buffer_page *new_head;
1952	int entries;
1953	int type;
1954	int ret;
1955
1956	entries = rb_page_entries(next_page);
1957
1958	/*
1959	 * The hard part is here. We need to move the head
1960	 * forward, and protect against both readers on
1961	 * other CPUs and writers coming in via interrupts.
1962	 */
1963	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1964				       RB_PAGE_HEAD);
1965
1966	/*
1967	 * type can be one of four:
1968	 *  NORMAL - an interrupt already moved it for us
1969	 *  HEAD   - we are the first to get here.
1970	 *  UPDATE - we are the interrupt interrupting
1971	 *           a current move.
1972	 *  MOVED  - a reader on another CPU moved the next
1973	 *           pointer to its reader page. Give up
1974	 *           and try again.
1975	 */
1976
1977	switch (type) {
1978	case RB_PAGE_HEAD:
1979		/*
1980		 * We changed the head to UPDATE, thus
1981		 * it is our responsibility to update
1982		 * the counters.
1983		 */
1984		local_add(entries, &cpu_buffer->overrun);
1985		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1986
1987		/*
1988		 * The entries will be zeroed out when we move the
1989		 * tail page.
1990		 */
1991
1992		/* still more to do */
1993		break;
1994
1995	case RB_PAGE_UPDATE:
1996		/*
1997		 * This is an interrupt that interrupt the
1998		 * previous update. Still more to do.
1999		 */
2000		break;
2001	case RB_PAGE_NORMAL:
2002		/*
2003		 * An interrupt came in before the update
2004		 * and processed this for us.
2005		 * Nothing left to do.
2006		 */
2007		return 1;
2008	case RB_PAGE_MOVED:
2009		/*
2010		 * The reader is on another CPU and just did
2011		 * a swap with our next_page.
2012		 * Try again.
2013		 */
2014		return 1;
2015	default:
2016		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2017		return -1;
2018	}
2019
2020	/*
2021	 * Now that we are here, the old head pointer is
2022	 * set to UPDATE. This will keep the reader from
2023	 * swapping the head page with the reader page.
2024	 * The reader (on another CPU) will spin till
2025	 * we are finished.
2026	 *
2027	 * We just need to protect against interrupts
2028	 * doing the job. We will set the next pointer
2029	 * to HEAD. After that, we set the old pointer
2030	 * to NORMAL, but only if it was HEAD before.
2031	 * otherwise we are an interrupt, and only
2032	 * want the outer most commit to reset it.
2033	 */
2034	new_head = next_page;
2035	rb_inc_page(cpu_buffer, &new_head);
2036
2037	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2038				    RB_PAGE_NORMAL);
2039
2040	/*
2041	 * Valid returns are:
2042	 *  HEAD   - an interrupt came in and already set it.
2043	 *  NORMAL - One of two things:
2044	 *            1) We really set it.
2045	 *            2) A bunch of interrupts came in and moved
2046	 *               the page forward again.
2047	 */
2048	switch (ret) {
2049	case RB_PAGE_HEAD:
2050	case RB_PAGE_NORMAL:
2051		/* OK */
2052		break;
2053	default:
2054		RB_WARN_ON(cpu_buffer, 1);
2055		return -1;
2056	}
2057
2058	/*
2059	 * It is possible that an interrupt came in,
2060	 * set the head up, then more interrupts came in
2061	 * and moved it again. When we get back here,
2062	 * the page would have been set to NORMAL but we
2063	 * just set it back to HEAD.
2064	 *
2065	 * How do you detect this? Well, if that happened
2066	 * the tail page would have moved.
2067	 */
2068	if (ret == RB_PAGE_NORMAL) {
2069		struct buffer_page *buffer_tail_page;
2070
2071		buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2072		/*
2073		 * If the tail had moved passed next, then we need
2074		 * to reset the pointer.
2075		 */
2076		if (buffer_tail_page != tail_page &&
2077		    buffer_tail_page != next_page)
2078			rb_head_page_set_normal(cpu_buffer, new_head,
2079						next_page,
2080						RB_PAGE_HEAD);
2081	}
2082
2083	/*
2084	 * If this was the outer most commit (the one that
2085	 * changed the original pointer from HEAD to UPDATE),
2086	 * then it is up to us to reset it to NORMAL.
2087	 */
2088	if (type == RB_PAGE_HEAD) {
2089		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2090					      tail_page,
2091					      RB_PAGE_UPDATE);
2092		if (RB_WARN_ON(cpu_buffer,
2093			       ret != RB_PAGE_UPDATE))
2094			return -1;
2095	}
2096
2097	return 0;
2098}
2099
2100static inline void
2101rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2102	      unsigned long tail, struct rb_event_info *info)
2103{
2104	struct buffer_page *tail_page = info->tail_page;
2105	struct ring_buffer_event *event;
2106	unsigned long length = info->length;
2107
2108	/*
2109	 * Only the event that crossed the page boundary
2110	 * must fill the old tail_page with padding.
2111	 */
2112	if (tail >= BUF_PAGE_SIZE) {
2113		/*
2114		 * If the page was filled, then we still need
2115		 * to update the real_end. Reset it to zero
2116		 * and the reader will ignore it.
2117		 */
2118		if (tail == BUF_PAGE_SIZE)
2119			tail_page->real_end = 0;
2120
2121		local_sub(length, &tail_page->write);
2122		return;
2123	}
2124
2125	event = __rb_page_index(tail_page, tail);
2126
2127	/* account for padding bytes */
2128	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2129
2130	/*
2131	 * Save the original length to the meta data.
2132	 * This will be used by the reader to add lost event
2133	 * counter.
2134	 */
2135	tail_page->real_end = tail;
2136
2137	/*
2138	 * If this event is bigger than the minimum size, then
2139	 * we need to be careful that we don't subtract the
2140	 * write counter enough to allow another writer to slip
2141	 * in on this page.
2142	 * We put in a discarded commit instead, to make sure
2143	 * that this space is not used again.
2144	 *
2145	 * If we are less than the minimum size, we don't need to
2146	 * worry about it.
2147	 */
2148	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2149		/* No room for any events */
2150
2151		/* Mark the rest of the page with padding */
2152		rb_event_set_padding(event);
2153
2154		/* Set the write back to the previous setting */
2155		local_sub(length, &tail_page->write);
2156		return;
2157	}
2158
2159	/* Put in a discarded event */
2160	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2161	event->type_len = RINGBUF_TYPE_PADDING;
2162	/* time delta must be non zero */
2163	event->time_delta = 1;
2164
2165	/* Set write to end of buffer */
2166	length = (tail + length) - BUF_PAGE_SIZE;
2167	local_sub(length, &tail_page->write);
2168}
2169
2170static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2171
2172/*
2173 * This is the slow path, force gcc not to inline it.
2174 */
2175static noinline struct ring_buffer_event *
2176rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2177	     unsigned long tail, struct rb_event_info *info)
2178{
2179	struct buffer_page *tail_page = info->tail_page;
2180	struct buffer_page *commit_page = cpu_buffer->commit_page;
2181	struct ring_buffer *buffer = cpu_buffer->buffer;
2182	struct buffer_page *next_page;
2183	int ret;
2184
2185	next_page = tail_page;
2186
2187	rb_inc_page(cpu_buffer, &next_page);
2188
2189	/*
2190	 * If for some reason, we had an interrupt storm that made
2191	 * it all the way around the buffer, bail, and warn
2192	 * about it.
2193	 */
2194	if (unlikely(next_page == commit_page)) {
2195		local_inc(&cpu_buffer->commit_overrun);
2196		goto out_reset;
2197	}
2198
2199	/*
2200	 * This is where the fun begins!
2201	 *
2202	 * We are fighting against races between a reader that
2203	 * could be on another CPU trying to swap its reader
2204	 * page with the buffer head.
2205	 *
2206	 * We are also fighting against interrupts coming in and
2207	 * moving the head or tail on us as well.
2208	 *
2209	 * If the next page is the head page then we have filled
2210	 * the buffer, unless the commit page is still on the
2211	 * reader page.
2212	 */
2213	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2214
2215		/*
2216		 * If the commit is not on the reader page, then
2217		 * move the header page.
2218		 */
2219		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2220			/*
2221			 * If we are not in overwrite mode,
2222			 * this is easy, just stop here.
2223			 */
2224			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2225				local_inc(&cpu_buffer->dropped_events);
2226				goto out_reset;
2227			}
2228
2229			ret = rb_handle_head_page(cpu_buffer,
2230						  tail_page,
2231						  next_page);
2232			if (ret < 0)
2233				goto out_reset;
2234			if (ret)
2235				goto out_again;
2236		} else {
2237			/*
2238			 * We need to be careful here too. The
2239			 * commit page could still be on the reader
2240			 * page. We could have a small buffer, and
2241			 * have filled up the buffer with events
2242			 * from interrupts and such, and wrapped.
2243			 *
2244			 * Note, if the tail page is also the on the
2245			 * reader_page, we let it move out.
2246			 */
2247			if (unlikely((cpu_buffer->commit_page !=
2248				      cpu_buffer->tail_page) &&
2249				     (cpu_buffer->commit_page ==
2250				      cpu_buffer->reader_page))) {
2251				local_inc(&cpu_buffer->commit_overrun);
2252				goto out_reset;
2253			}
2254		}
2255	}
2256
2257	rb_tail_page_update(cpu_buffer, tail_page, next_page);
2258
2259 out_again:
2260
2261	rb_reset_tail(cpu_buffer, tail, info);
2262
2263	/* Commit what we have for now. */
2264	rb_end_commit(cpu_buffer);
2265	/* rb_end_commit() decs committing */
2266	local_inc(&cpu_buffer->committing);
2267
2268	/* fail and let the caller try again */
2269	return ERR_PTR(-EAGAIN);
2270
2271 out_reset:
2272	/* reset write */
2273	rb_reset_tail(cpu_buffer, tail, info);
2274
2275	return NULL;
2276}
2277
2278/* Slow path, do not inline */
2279static noinline struct ring_buffer_event *
2280rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2281{
2282	if (abs)
2283		event->type_len = RINGBUF_TYPE_TIME_STAMP;
2284	else
2285		event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2286
2287	/* Not the first event on the page, or not delta? */
2288	if (abs || rb_event_index(event)) {
2289		event->time_delta = delta & TS_MASK;
2290		event->array[0] = delta >> TS_SHIFT;
2291	} else {
2292		/* nope, just zero it */
2293		event->time_delta = 0;
2294		event->array[0] = 0;
2295	}
2296
2297	return skip_time_extend(event);
2298}
2299
2300static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2301				     struct ring_buffer_event *event);
2302
2303/**
2304 * rb_update_event - update event type and data
2305 * @event: the event to update
2306 * @type: the type of event
2307 * @length: the size of the event field in the ring buffer
2308 *
2309 * Update the type and data fields of the event. The length
2310 * is the actual size that is written to the ring buffer,
2311 * and with this, we can determine what to place into the
2312 * data field.
2313 */
2314static void
2315rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2316		struct ring_buffer_event *event,
2317		struct rb_event_info *info)
2318{
2319	unsigned length = info->length;
2320	u64 delta = info->delta;
2321
2322	/* Only a commit updates the timestamp */
2323	if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2324		delta = 0;
2325
2326	/*
2327	 * If we need to add a timestamp, then we
2328	 * add it to the start of the resevered space.
2329	 */
2330	if (unlikely(info->add_timestamp)) {
2331		bool abs = ring_buffer_time_stamp_abs(cpu_buffer->buffer);
2332
2333		event = rb_add_time_stamp(event, info->delta, abs);
2334		length -= RB_LEN_TIME_EXTEND;
2335		delta = 0;
2336	}
2337
2338	event->time_delta = delta;
2339	length -= RB_EVNT_HDR_SIZE;
2340	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2341		event->type_len = 0;
2342		event->array[0] = length;
2343	} else
2344		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2345}
2346
2347static unsigned rb_calculate_event_length(unsigned length)
2348{
2349	struct ring_buffer_event event; /* Used only for sizeof array */
2350
2351	/* zero length can cause confusions */
2352	if (!length)
2353		length++;
2354
2355	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2356		length += sizeof(event.array[0]);
2357
2358	length += RB_EVNT_HDR_SIZE;
2359	length = ALIGN(length, RB_ARCH_ALIGNMENT);
2360
2361	/*
2362	 * In case the time delta is larger than the 27 bits for it
2363	 * in the header, we need to add a timestamp. If another
2364	 * event comes in when trying to discard this one to increase
2365	 * the length, then the timestamp will be added in the allocated
2366	 * space of this event. If length is bigger than the size needed
2367	 * for the TIME_EXTEND, then padding has to be used. The events
2368	 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2369	 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2370	 * As length is a multiple of 4, we only need to worry if it
2371	 * is 12 (RB_LEN_TIME_EXTEND + 4).
2372	 */
2373	if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2374		length += RB_ALIGNMENT;
2375
2376	return length;
2377}
2378
2379#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2380static inline bool sched_clock_stable(void)
2381{
2382	return true;
2383}
2384#endif
2385
2386static inline int
2387rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2388		  struct ring_buffer_event *event)
2389{
2390	unsigned long new_index, old_index;
2391	struct buffer_page *bpage;
2392	unsigned long index;
2393	unsigned long addr;
2394
2395	new_index = rb_event_index(event);
2396	old_index = new_index + rb_event_ts_length(event);
2397	addr = (unsigned long)event;
2398	addr &= PAGE_MASK;
2399
2400	bpage = READ_ONCE(cpu_buffer->tail_page);
2401
2402	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2403		unsigned long write_mask =
2404			local_read(&bpage->write) & ~RB_WRITE_MASK;
2405		unsigned long event_length = rb_event_length(event);
2406		/*
2407		 * This is on the tail page. It is possible that
2408		 * a write could come in and move the tail page
2409		 * and write to the next page. That is fine
2410		 * because we just shorten what is on this page.
2411		 */
2412		old_index += write_mask;
2413		new_index += write_mask;
2414		index = local_cmpxchg(&bpage->write, old_index, new_index);
2415		if (index == old_index) {
2416			/* update counters */
2417			local_sub(event_length, &cpu_buffer->entries_bytes);
2418			return 1;
2419		}
2420	}
2421
2422	/* could not discard */
2423	return 0;
2424}
2425
2426static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2427{
2428	local_inc(&cpu_buffer->committing);
2429	local_inc(&cpu_buffer->commits);
2430}
2431
2432static __always_inline void
2433rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2434{
2435	unsigned long max_count;
2436
2437	/*
2438	 * We only race with interrupts and NMIs on this CPU.
2439	 * If we own the commit event, then we can commit
2440	 * all others that interrupted us, since the interruptions
2441	 * are in stack format (they finish before they come
2442	 * back to us). This allows us to do a simple loop to
2443	 * assign the commit to the tail.
2444	 */
2445 again:
2446	max_count = cpu_buffer->nr_pages * 100;
2447
2448	while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2449		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2450			return;
2451		if (RB_WARN_ON(cpu_buffer,
2452			       rb_is_reader_page(cpu_buffer->tail_page)))
2453			return;
2454		local_set(&cpu_buffer->commit_page->page->commit,
2455			  rb_page_write(cpu_buffer->commit_page));
2456		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2457		/* Only update the write stamp if the page has an event */
2458		if (rb_page_write(cpu_buffer->commit_page))
2459			cpu_buffer->write_stamp =
2460				cpu_buffer->commit_page->page->time_stamp;
2461		/* add barrier to keep gcc from optimizing too much */
2462		barrier();
2463	}
2464	while (rb_commit_index(cpu_buffer) !=
2465	       rb_page_write(cpu_buffer->commit_page)) {
2466
2467		local_set(&cpu_buffer->commit_page->page->commit,
2468			  rb_page_write(cpu_buffer->commit_page));
2469		RB_WARN_ON(cpu_buffer,
2470			   local_read(&cpu_buffer->commit_page->page->commit) &
2471			   ~RB_WRITE_MASK);
2472		barrier();
2473	}
2474
2475	/* again, keep gcc from optimizing */
2476	barrier();
2477
2478	/*
2479	 * If an interrupt came in just after the first while loop
2480	 * and pushed the tail page forward, we will be left with
2481	 * a dangling commit that will never go forward.
2482	 */
2483	if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2484		goto again;
2485}
2486
2487static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2488{
2489	unsigned long commits;
2490
2491	if (RB_WARN_ON(cpu_buffer,
2492		       !local_read(&cpu_buffer->committing)))
2493		return;
2494
2495 again:
2496	commits = local_read(&cpu_buffer->commits);
2497	/* synchronize with interrupts */
2498	barrier();
2499	if (local_read(&cpu_buffer->committing) == 1)
2500		rb_set_commit_to_write(cpu_buffer);
2501
2502	local_dec(&cpu_buffer->committing);
2503
2504	/* synchronize with interrupts */
2505	barrier();
2506
2507	/*
2508	 * Need to account for interrupts coming in between the
2509	 * updating of the commit page and the clearing of the
2510	 * committing counter.
2511	 */
2512	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2513	    !local_read(&cpu_buffer->committing)) {
2514		local_inc(&cpu_buffer->committing);
2515		goto again;
2516	}
2517}
2518
2519static inline void rb_event_discard(struct ring_buffer_event *event)
2520{
2521	if (extended_time(event))
2522		event = skip_time_extend(event);
2523
2524	/* array[0] holds the actual length for the discarded event */
2525	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2526	event->type_len = RINGBUF_TYPE_PADDING;
2527	/* time delta must be non zero */
2528	if (!event->time_delta)
2529		event->time_delta = 1;
2530}
2531
2532static __always_inline bool
2533rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2534		   struct ring_buffer_event *event)
2535{
2536	unsigned long addr = (unsigned long)event;
2537	unsigned long index;
2538
2539	index = rb_event_index(event);
2540	addr &= PAGE_MASK;
2541
2542	return cpu_buffer->commit_page->page == (void *)addr &&
2543		rb_commit_index(cpu_buffer) == index;
2544}
2545
2546static __always_inline void
2547rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2548		      struct ring_buffer_event *event)
2549{
2550	u64 delta;
2551
2552	/*
2553	 * The event first in the commit queue updates the
2554	 * time stamp.
2555	 */
2556	if (rb_event_is_commit(cpu_buffer, event)) {
2557		/*
2558		 * A commit event that is first on a page
2559		 * updates the write timestamp with the page stamp
2560		 */
2561		if (!rb_event_index(event))
2562			cpu_buffer->write_stamp =
2563				cpu_buffer->commit_page->page->time_stamp;
2564		else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2565			delta = ring_buffer_event_time_stamp(event);
2566			cpu_buffer->write_stamp += delta;
2567		} else if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
2568			delta = ring_buffer_event_time_stamp(event);
2569			cpu_buffer->write_stamp = delta;
2570		} else
2571			cpu_buffer->write_stamp += event->time_delta;
2572	}
2573}
2574
2575static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2576		      struct ring_buffer_event *event)
2577{
2578	local_inc(&cpu_buffer->entries);
2579	rb_update_write_stamp(cpu_buffer, event);
2580	rb_end_commit(cpu_buffer);
2581}
2582
2583static __always_inline void
2584rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2585{
2586	bool pagebusy;
 
 
2587
2588	if (buffer->irq_work.waiters_pending) {
2589		buffer->irq_work.waiters_pending = false;
2590		/* irq_work_queue() supplies it's own memory barriers */
2591		irq_work_queue(&buffer->irq_work.work);
2592	}
2593
2594	if (cpu_buffer->irq_work.waiters_pending) {
2595		cpu_buffer->irq_work.waiters_pending = false;
2596		/* irq_work_queue() supplies it's own memory barriers */
2597		irq_work_queue(&cpu_buffer->irq_work.work);
2598	}
2599
2600	pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2601
2602	if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2603		cpu_buffer->irq_work.wakeup_full = true;
2604		cpu_buffer->irq_work.full_waiters_pending = false;
2605		/* irq_work_queue() supplies it's own memory barriers */
2606		irq_work_queue(&cpu_buffer->irq_work.work);
2607	}
2608}
2609
2610/*
2611 * The lock and unlock are done within a preempt disable section.
2612 * The current_context per_cpu variable can only be modified
2613 * by the current task between lock and unlock. But it can
2614 * be modified more than once via an interrupt. To pass this
2615 * information from the lock to the unlock without having to
2616 * access the 'in_interrupt()' functions again (which do show
2617 * a bit of overhead in something as critical as function tracing,
2618 * we use a bitmask trick.
2619 *
2620 *  bit 0 =  NMI context
2621 *  bit 1 =  IRQ context
2622 *  bit 2 =  SoftIRQ context
2623 *  bit 3 =  normal context.
2624 *
2625 * This works because this is the order of contexts that can
2626 * preempt other contexts. A SoftIRQ never preempts an IRQ
2627 * context.
2628 *
2629 * When the context is determined, the corresponding bit is
2630 * checked and set (if it was set, then a recursion of that context
2631 * happened).
2632 *
2633 * On unlock, we need to clear this bit. To do so, just subtract
2634 * 1 from the current_context and AND it to itself.
2635 *
2636 * (binary)
2637 *  101 - 1 = 100
2638 *  101 & 100 = 100 (clearing bit zero)
2639 *
2640 *  1010 - 1 = 1001
2641 *  1010 & 1001 = 1000 (clearing bit 1)
2642 *
2643 * The least significant bit can be cleared this way, and it
2644 * just so happens that it is the same bit corresponding to
2645 * the current context.
2646 */
2647
2648static __always_inline int
2649trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2650{
2651	unsigned int val = cpu_buffer->current_context;
2652	unsigned long pc = preempt_count();
2653	int bit;
2654
2655	if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
2656		bit = RB_CTX_NORMAL;
2657	else
2658		bit = pc & NMI_MASK ? RB_CTX_NMI :
2659			pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
2660
2661	if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
2662		return 1;
2663
2664	val |= (1 << (bit + cpu_buffer->nest));
2665	cpu_buffer->current_context = val;
2666
2667	return 0;
2668}
2669
2670static __always_inline void
2671trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2672{
2673	cpu_buffer->current_context &=
2674		cpu_buffer->current_context - (1 << cpu_buffer->nest);
2675}
2676
2677/* The recursive locking above uses 4 bits */
2678#define NESTED_BITS 4
2679
2680/**
2681 * ring_buffer_nest_start - Allow to trace while nested
2682 * @buffer: The ring buffer to modify
2683 *
2684 * The ring buffer has a safty mechanism to prevent recursion.
2685 * But there may be a case where a trace needs to be done while
2686 * tracing something else. In this case, calling this function
2687 * will allow this function to nest within a currently active
2688 * ring_buffer_lock_reserve().
2689 *
2690 * Call this function before calling another ring_buffer_lock_reserve() and
2691 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
2692 */
2693void ring_buffer_nest_start(struct ring_buffer *buffer)
2694{
2695	struct ring_buffer_per_cpu *cpu_buffer;
2696	int cpu;
2697
2698	/* Enabled by ring_buffer_nest_end() */
2699	preempt_disable_notrace();
2700	cpu = raw_smp_processor_id();
2701	cpu_buffer = buffer->buffers[cpu];
2702	/* This is the shift value for the above recusive locking */
2703	cpu_buffer->nest += NESTED_BITS;
2704}
2705
2706/**
2707 * ring_buffer_nest_end - Allow to trace while nested
2708 * @buffer: The ring buffer to modify
2709 *
2710 * Must be called after ring_buffer_nest_start() and after the
2711 * ring_buffer_unlock_commit().
2712 */
2713void ring_buffer_nest_end(struct ring_buffer *buffer)
2714{
2715	struct ring_buffer_per_cpu *cpu_buffer;
2716	int cpu;
2717
2718	/* disabled by ring_buffer_nest_start() */
2719	cpu = raw_smp_processor_id();
2720	cpu_buffer = buffer->buffers[cpu];
2721	/* This is the shift value for the above recusive locking */
2722	cpu_buffer->nest -= NESTED_BITS;
2723	preempt_enable_notrace();
2724}
2725
2726/**
2727 * ring_buffer_unlock_commit - commit a reserved
2728 * @buffer: The buffer to commit to
2729 * @event: The event pointer to commit.
2730 *
2731 * This commits the data to the ring buffer, and releases any locks held.
2732 *
2733 * Must be paired with ring_buffer_lock_reserve.
2734 */
2735int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2736			      struct ring_buffer_event *event)
2737{
2738	struct ring_buffer_per_cpu *cpu_buffer;
2739	int cpu = raw_smp_processor_id();
2740
2741	cpu_buffer = buffer->buffers[cpu];
2742
2743	rb_commit(cpu_buffer, event);
2744
2745	rb_wakeups(buffer, cpu_buffer);
2746
2747	trace_recursive_unlock(cpu_buffer);
2748
2749	preempt_enable_notrace();
2750
2751	return 0;
2752}
2753EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2754
2755static noinline void
2756rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2757		    struct rb_event_info *info)
2758{
2759	WARN_ONCE(info->delta > (1ULL << 59),
2760		  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2761		  (unsigned long long)info->delta,
2762		  (unsigned long long)info->ts,
2763		  (unsigned long long)cpu_buffer->write_stamp,
2764		  sched_clock_stable() ? "" :
2765		  "If you just came from a suspend/resume,\n"
2766		  "please switch to the trace global clock:\n"
2767		  "  echo global > /sys/kernel/debug/tracing/trace_clock\n"
2768		  "or add trace_clock=global to the kernel command line\n");
2769	info->add_timestamp = 1;
2770}
2771
2772static struct ring_buffer_event *
2773__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2774		  struct rb_event_info *info)
2775{
2776	struct ring_buffer_event *event;
2777	struct buffer_page *tail_page;
2778	unsigned long tail, write;
2779
2780	/*
2781	 * If the time delta since the last event is too big to
2782	 * hold in the time field of the event, then we append a
2783	 * TIME EXTEND event ahead of the data event.
2784	 */
2785	if (unlikely(info->add_timestamp))
2786		info->length += RB_LEN_TIME_EXTEND;
2787
2788	/* Don't let the compiler play games with cpu_buffer->tail_page */
2789	tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2790	write = local_add_return(info->length, &tail_page->write);
2791
2792	/* set write to only the index of the write */
2793	write &= RB_WRITE_MASK;
2794	tail = write - info->length;
2795
2796	/*
2797	 * If this is the first commit on the page, then it has the same
2798	 * timestamp as the page itself.
2799	 */
2800	if (!tail && !ring_buffer_time_stamp_abs(cpu_buffer->buffer))
2801		info->delta = 0;
2802
2803	/* See if we shot pass the end of this buffer page */
2804	if (unlikely(write > BUF_PAGE_SIZE))
2805		return rb_move_tail(cpu_buffer, tail, info);
2806
2807	/* We reserved something on the buffer */
2808
2809	event = __rb_page_index(tail_page, tail);
2810	rb_update_event(cpu_buffer, event, info);
2811
2812	local_inc(&tail_page->entries);
2813
2814	/*
2815	 * If this is the first commit on the page, then update
2816	 * its timestamp.
2817	 */
2818	if (!tail)
2819		tail_page->page->time_stamp = info->ts;
2820
2821	/* account for these added bytes */
2822	local_add(info->length, &cpu_buffer->entries_bytes);
2823
2824	return event;
2825}
2826
2827static __always_inline struct ring_buffer_event *
2828rb_reserve_next_event(struct ring_buffer *buffer,
2829		      struct ring_buffer_per_cpu *cpu_buffer,
2830		      unsigned long length)
2831{
2832	struct ring_buffer_event *event;
2833	struct rb_event_info info;
2834	int nr_loops = 0;
2835	u64 diff;
2836
2837	rb_start_commit(cpu_buffer);
2838
2839#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2840	/*
2841	 * Due to the ability to swap a cpu buffer from a buffer
2842	 * it is possible it was swapped before we committed.
2843	 * (committing stops a swap). We check for it here and
2844	 * if it happened, we have to fail the write.
2845	 */
2846	barrier();
2847	if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2848		local_dec(&cpu_buffer->committing);
2849		local_dec(&cpu_buffer->commits);
2850		return NULL;
2851	}
2852#endif
2853
2854	info.length = rb_calculate_event_length(length);
2855 again:
2856	info.add_timestamp = 0;
2857	info.delta = 0;
2858
2859	/*
2860	 * We allow for interrupts to reenter here and do a trace.
2861	 * If one does, it will cause this original code to loop
2862	 * back here. Even with heavy interrupts happening, this
2863	 * should only happen a few times in a row. If this happens
2864	 * 1000 times in a row, there must be either an interrupt
2865	 * storm or we have something buggy.
2866	 * Bail!
2867	 */
2868	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2869		goto out_fail;
2870
2871	info.ts = rb_time_stamp(cpu_buffer->buffer);
2872	diff = info.ts - cpu_buffer->write_stamp;
2873
2874	/* make sure this diff is calculated here */
2875	barrier();
2876
2877	if (ring_buffer_time_stamp_abs(buffer)) {
2878		info.delta = info.ts;
2879		rb_handle_timestamp(cpu_buffer, &info);
2880	} else /* Did the write stamp get updated already? */
2881		if (likely(info.ts >= cpu_buffer->write_stamp)) {
2882		info.delta = diff;
2883		if (unlikely(test_time_stamp(info.delta)))
2884			rb_handle_timestamp(cpu_buffer, &info);
2885	}
2886
2887	event = __rb_reserve_next(cpu_buffer, &info);
2888
2889	if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2890		if (info.add_timestamp)
2891			info.length -= RB_LEN_TIME_EXTEND;
2892		goto again;
2893	}
2894
2895	if (!event)
2896		goto out_fail;
2897
2898	return event;
2899
2900 out_fail:
2901	rb_end_commit(cpu_buffer);
2902	return NULL;
2903}
2904
2905/**
2906 * ring_buffer_lock_reserve - reserve a part of the buffer
2907 * @buffer: the ring buffer to reserve from
2908 * @length: the length of the data to reserve (excluding event header)
2909 *
2910 * Returns a reseverd event on the ring buffer to copy directly to.
2911 * The user of this interface will need to get the body to write into
2912 * and can use the ring_buffer_event_data() interface.
2913 *
2914 * The length is the length of the data needed, not the event length
2915 * which also includes the event header.
2916 *
2917 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2918 * If NULL is returned, then nothing has been allocated or locked.
2919 */
2920struct ring_buffer_event *
2921ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2922{
2923	struct ring_buffer_per_cpu *cpu_buffer;
2924	struct ring_buffer_event *event;
2925	int cpu;
2926
2927	/* If we are tracing schedule, we don't want to recurse */
2928	preempt_disable_notrace();
2929
2930	if (unlikely(atomic_read(&buffer->record_disabled)))
2931		goto out;
2932
2933	cpu = raw_smp_processor_id();
2934
2935	if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2936		goto out;
2937
2938	cpu_buffer = buffer->buffers[cpu];
2939
2940	if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2941		goto out;
2942
2943	if (unlikely(length > BUF_MAX_DATA_SIZE))
2944		goto out;
2945
2946	if (unlikely(trace_recursive_lock(cpu_buffer)))
2947		goto out;
2948
2949	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2950	if (!event)
2951		goto out_unlock;
2952
2953	return event;
2954
2955 out_unlock:
2956	trace_recursive_unlock(cpu_buffer);
2957 out:
2958	preempt_enable_notrace();
2959	return NULL;
2960}
2961EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2962
2963/*
2964 * Decrement the entries to the page that an event is on.
2965 * The event does not even need to exist, only the pointer
2966 * to the page it is on. This may only be called before the commit
2967 * takes place.
2968 */
2969static inline void
2970rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2971		   struct ring_buffer_event *event)
2972{
2973	unsigned long addr = (unsigned long)event;
2974	struct buffer_page *bpage = cpu_buffer->commit_page;
2975	struct buffer_page *start;
2976
2977	addr &= PAGE_MASK;
2978
2979	/* Do the likely case first */
2980	if (likely(bpage->page == (void *)addr)) {
2981		local_dec(&bpage->entries);
2982		return;
2983	}
2984
2985	/*
2986	 * Because the commit page may be on the reader page we
2987	 * start with the next page and check the end loop there.
2988	 */
2989	rb_inc_page(cpu_buffer, &bpage);
2990	start = bpage;
2991	do {
2992		if (bpage->page == (void *)addr) {
2993			local_dec(&bpage->entries);
2994			return;
2995		}
2996		rb_inc_page(cpu_buffer, &bpage);
2997	} while (bpage != start);
2998
2999	/* commit not part of this buffer?? */
3000	RB_WARN_ON(cpu_buffer, 1);
3001}
3002
3003/**
3004 * ring_buffer_commit_discard - discard an event that has not been committed
3005 * @buffer: the ring buffer
3006 * @event: non committed event to discard
3007 *
3008 * Sometimes an event that is in the ring buffer needs to be ignored.
3009 * This function lets the user discard an event in the ring buffer
3010 * and then that event will not be read later.
3011 *
3012 * This function only works if it is called before the the item has been
3013 * committed. It will try to free the event from the ring buffer
3014 * if another event has not been added behind it.
3015 *
3016 * If another event has been added behind it, it will set the event
3017 * up as discarded, and perform the commit.
3018 *
3019 * If this function is called, do not call ring_buffer_unlock_commit on
3020 * the event.
3021 */
3022void ring_buffer_discard_commit(struct ring_buffer *buffer,
3023				struct ring_buffer_event *event)
3024{
3025	struct ring_buffer_per_cpu *cpu_buffer;
3026	int cpu;
3027
3028	/* The event is discarded regardless */
3029	rb_event_discard(event);
3030
3031	cpu = smp_processor_id();
3032	cpu_buffer = buffer->buffers[cpu];
3033
3034	/*
3035	 * This must only be called if the event has not been
3036	 * committed yet. Thus we can assume that preemption
3037	 * is still disabled.
3038	 */
3039	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3040
3041	rb_decrement_entry(cpu_buffer, event);
3042	if (rb_try_to_discard(cpu_buffer, event))
3043		goto out;
3044
3045	/*
3046	 * The commit is still visible by the reader, so we
3047	 * must still update the timestamp.
3048	 */
3049	rb_update_write_stamp(cpu_buffer, event);
3050 out:
3051	rb_end_commit(cpu_buffer);
3052
3053	trace_recursive_unlock(cpu_buffer);
3054
3055	preempt_enable_notrace();
3056
3057}
3058EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3059
3060/**
3061 * ring_buffer_write - write data to the buffer without reserving
3062 * @buffer: The ring buffer to write to.
3063 * @length: The length of the data being written (excluding the event header)
3064 * @data: The data to write to the buffer.
3065 *
3066 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3067 * one function. If you already have the data to write to the buffer, it
3068 * may be easier to simply call this function.
3069 *
3070 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3071 * and not the length of the event which would hold the header.
3072 */
3073int ring_buffer_write(struct ring_buffer *buffer,
3074		      unsigned long length,
3075		      void *data)
3076{
3077	struct ring_buffer_per_cpu *cpu_buffer;
3078	struct ring_buffer_event *event;
3079	void *body;
3080	int ret = -EBUSY;
3081	int cpu;
3082
3083	preempt_disable_notrace();
3084
3085	if (atomic_read(&buffer->record_disabled))
3086		goto out;
3087
3088	cpu = raw_smp_processor_id();
3089
3090	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3091		goto out;
3092
3093	cpu_buffer = buffer->buffers[cpu];
3094
3095	if (atomic_read(&cpu_buffer->record_disabled))
3096		goto out;
3097
3098	if (length > BUF_MAX_DATA_SIZE)
3099		goto out;
3100
3101	if (unlikely(trace_recursive_lock(cpu_buffer)))
3102		goto out;
3103
3104	event = rb_reserve_next_event(buffer, cpu_buffer, length);
3105	if (!event)
3106		goto out_unlock;
3107
3108	body = rb_event_data(event);
3109
3110	memcpy(body, data, length);
3111
3112	rb_commit(cpu_buffer, event);
3113
3114	rb_wakeups(buffer, cpu_buffer);
3115
3116	ret = 0;
3117
3118 out_unlock:
3119	trace_recursive_unlock(cpu_buffer);
3120
3121 out:
3122	preempt_enable_notrace();
3123
3124	return ret;
3125}
3126EXPORT_SYMBOL_GPL(ring_buffer_write);
3127
3128static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3129{
3130	struct buffer_page *reader = cpu_buffer->reader_page;
3131	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3132	struct buffer_page *commit = cpu_buffer->commit_page;
3133
3134	/* In case of error, head will be NULL */
3135	if (unlikely(!head))
3136		return true;
3137
3138	return reader->read == rb_page_commit(reader) &&
3139		(commit == reader ||
3140		 (commit == head &&
3141		  head->read == rb_page_commit(commit)));
3142}
3143
3144/**
3145 * ring_buffer_record_disable - stop all writes into the buffer
3146 * @buffer: The ring buffer to stop writes to.
3147 *
3148 * This prevents all writes to the buffer. Any attempt to write
3149 * to the buffer after this will fail and return NULL.
3150 *
3151 * The caller should call synchronize_sched() after this.
3152 */
3153void ring_buffer_record_disable(struct ring_buffer *buffer)
3154{
3155	atomic_inc(&buffer->record_disabled);
3156}
3157EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3158
3159/**
3160 * ring_buffer_record_enable - enable writes to the buffer
3161 * @buffer: The ring buffer to enable writes
3162 *
3163 * Note, multiple disables will need the same number of enables
3164 * to truly enable the writing (much like preempt_disable).
3165 */
3166void ring_buffer_record_enable(struct ring_buffer *buffer)
3167{
3168	atomic_dec(&buffer->record_disabled);
3169}
3170EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3171
3172/**
3173 * ring_buffer_record_off - stop all writes into the buffer
3174 * @buffer: The ring buffer to stop writes to.
3175 *
3176 * This prevents all writes to the buffer. Any attempt to write
3177 * to the buffer after this will fail and return NULL.
3178 *
3179 * This is different than ring_buffer_record_disable() as
3180 * it works like an on/off switch, where as the disable() version
3181 * must be paired with a enable().
3182 */
3183void ring_buffer_record_off(struct ring_buffer *buffer)
3184{
3185	unsigned int rd;
3186	unsigned int new_rd;
3187
3188	do {
3189		rd = atomic_read(&buffer->record_disabled);
3190		new_rd = rd | RB_BUFFER_OFF;
3191	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3192}
3193EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3194
3195/**
3196 * ring_buffer_record_on - restart writes into the buffer
3197 * @buffer: The ring buffer to start writes to.
3198 *
3199 * This enables all writes to the buffer that was disabled by
3200 * ring_buffer_record_off().
3201 *
3202 * This is different than ring_buffer_record_enable() as
3203 * it works like an on/off switch, where as the enable() version
3204 * must be paired with a disable().
3205 */
3206void ring_buffer_record_on(struct ring_buffer *buffer)
3207{
3208	unsigned int rd;
3209	unsigned int new_rd;
3210
3211	do {
3212		rd = atomic_read(&buffer->record_disabled);
3213		new_rd = rd & ~RB_BUFFER_OFF;
3214	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3215}
3216EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3217
3218/**
3219 * ring_buffer_record_is_on - return true if the ring buffer can write
3220 * @buffer: The ring buffer to see if write is enabled
3221 *
3222 * Returns true if the ring buffer is in a state that it accepts writes.
3223 */
3224int ring_buffer_record_is_on(struct ring_buffer *buffer)
3225{
3226	return !atomic_read(&buffer->record_disabled);
3227}
3228
3229/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3230 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3231 * @buffer: The ring buffer to stop writes to.
3232 * @cpu: The CPU buffer to stop
3233 *
3234 * This prevents all writes to the buffer. Any attempt to write
3235 * to the buffer after this will fail and return NULL.
3236 *
3237 * The caller should call synchronize_sched() after this.
3238 */
3239void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3240{
3241	struct ring_buffer_per_cpu *cpu_buffer;
3242
3243	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3244		return;
3245
3246	cpu_buffer = buffer->buffers[cpu];
3247	atomic_inc(&cpu_buffer->record_disabled);
3248}
3249EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3250
3251/**
3252 * ring_buffer_record_enable_cpu - enable writes to the buffer
3253 * @buffer: The ring buffer to enable writes
3254 * @cpu: The CPU to enable.
3255 *
3256 * Note, multiple disables will need the same number of enables
3257 * to truly enable the writing (much like preempt_disable).
3258 */
3259void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3260{
3261	struct ring_buffer_per_cpu *cpu_buffer;
3262
3263	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3264		return;
3265
3266	cpu_buffer = buffer->buffers[cpu];
3267	atomic_dec(&cpu_buffer->record_disabled);
3268}
3269EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3270
3271/*
3272 * The total entries in the ring buffer is the running counter
3273 * of entries entered into the ring buffer, minus the sum of
3274 * the entries read from the ring buffer and the number of
3275 * entries that were overwritten.
3276 */
3277static inline unsigned long
3278rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3279{
3280	return local_read(&cpu_buffer->entries) -
3281		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3282}
3283
3284/**
3285 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3286 * @buffer: The ring buffer
3287 * @cpu: The per CPU buffer to read from.
3288 */
3289u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3290{
3291	unsigned long flags;
3292	struct ring_buffer_per_cpu *cpu_buffer;
3293	struct buffer_page *bpage;
3294	u64 ret = 0;
3295
3296	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3297		return 0;
3298
3299	cpu_buffer = buffer->buffers[cpu];
3300	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3301	/*
3302	 * if the tail is on reader_page, oldest time stamp is on the reader
3303	 * page
3304	 */
3305	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3306		bpage = cpu_buffer->reader_page;
3307	else
3308		bpage = rb_set_head_page(cpu_buffer);
3309	if (bpage)
3310		ret = bpage->page->time_stamp;
3311	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3312
3313	return ret;
3314}
3315EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3316
3317/**
3318 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3319 * @buffer: The ring buffer
3320 * @cpu: The per CPU buffer to read from.
3321 */
3322unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3323{
3324	struct ring_buffer_per_cpu *cpu_buffer;
3325	unsigned long ret;
3326
3327	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3328		return 0;
3329
3330	cpu_buffer = buffer->buffers[cpu];
3331	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3332
3333	return ret;
3334}
3335EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3336
3337/**
3338 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3339 * @buffer: The ring buffer
3340 * @cpu: The per CPU buffer to get the entries from.
3341 */
3342unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3343{
3344	struct ring_buffer_per_cpu *cpu_buffer;
3345
3346	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3347		return 0;
3348
3349	cpu_buffer = buffer->buffers[cpu];
3350
3351	return rb_num_of_entries(cpu_buffer);
3352}
3353EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3354
3355/**
3356 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3357 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3358 * @buffer: The ring buffer
3359 * @cpu: The per CPU buffer to get the number of overruns from
3360 */
3361unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3362{
3363	struct ring_buffer_per_cpu *cpu_buffer;
3364	unsigned long ret;
3365
3366	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3367		return 0;
3368
3369	cpu_buffer = buffer->buffers[cpu];
3370	ret = local_read(&cpu_buffer->overrun);
3371
3372	return ret;
3373}
3374EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3375
3376/**
3377 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3378 * commits failing due to the buffer wrapping around while there are uncommitted
3379 * events, such as during an interrupt storm.
3380 * @buffer: The ring buffer
3381 * @cpu: The per CPU buffer to get the number of overruns from
3382 */
3383unsigned long
3384ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3385{
3386	struct ring_buffer_per_cpu *cpu_buffer;
3387	unsigned long ret;
3388
3389	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3390		return 0;
3391
3392	cpu_buffer = buffer->buffers[cpu];
3393	ret = local_read(&cpu_buffer->commit_overrun);
3394
3395	return ret;
3396}
3397EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3398
3399/**
3400 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3401 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3402 * @buffer: The ring buffer
3403 * @cpu: The per CPU buffer to get the number of overruns from
3404 */
3405unsigned long
3406ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3407{
3408	struct ring_buffer_per_cpu *cpu_buffer;
3409	unsigned long ret;
3410
3411	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3412		return 0;
3413
3414	cpu_buffer = buffer->buffers[cpu];
3415	ret = local_read(&cpu_buffer->dropped_events);
3416
3417	return ret;
3418}
3419EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3420
3421/**
3422 * ring_buffer_read_events_cpu - get the number of events successfully read
3423 * @buffer: The ring buffer
3424 * @cpu: The per CPU buffer to get the number of events read
3425 */
3426unsigned long
3427ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3428{
3429	struct ring_buffer_per_cpu *cpu_buffer;
3430
3431	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3432		return 0;
3433
3434	cpu_buffer = buffer->buffers[cpu];
3435	return cpu_buffer->read;
3436}
3437EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3438
3439/**
3440 * ring_buffer_entries - get the number of entries in a buffer
3441 * @buffer: The ring buffer
3442 *
3443 * Returns the total number of entries in the ring buffer
3444 * (all CPU entries)
3445 */
3446unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3447{
3448	struct ring_buffer_per_cpu *cpu_buffer;
3449	unsigned long entries = 0;
3450	int cpu;
3451
3452	/* if you care about this being correct, lock the buffer */
3453	for_each_buffer_cpu(buffer, cpu) {
3454		cpu_buffer = buffer->buffers[cpu];
3455		entries += rb_num_of_entries(cpu_buffer);
3456	}
3457
3458	return entries;
3459}
3460EXPORT_SYMBOL_GPL(ring_buffer_entries);
3461
3462/**
3463 * ring_buffer_overruns - get the number of overruns in buffer
3464 * @buffer: The ring buffer
3465 *
3466 * Returns the total number of overruns in the ring buffer
3467 * (all CPU entries)
3468 */
3469unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3470{
3471	struct ring_buffer_per_cpu *cpu_buffer;
3472	unsigned long overruns = 0;
3473	int cpu;
3474
3475	/* if you care about this being correct, lock the buffer */
3476	for_each_buffer_cpu(buffer, cpu) {
3477		cpu_buffer = buffer->buffers[cpu];
3478		overruns += local_read(&cpu_buffer->overrun);
3479	}
3480
3481	return overruns;
3482}
3483EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3484
3485static void rb_iter_reset(struct ring_buffer_iter *iter)
3486{
3487	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3488
3489	/* Iterator usage is expected to have record disabled */
3490	iter->head_page = cpu_buffer->reader_page;
3491	iter->head = cpu_buffer->reader_page->read;
3492
3493	iter->cache_reader_page = iter->head_page;
3494	iter->cache_read = cpu_buffer->read;
3495
3496	if (iter->head)
3497		iter->read_stamp = cpu_buffer->read_stamp;
3498	else
3499		iter->read_stamp = iter->head_page->page->time_stamp;
3500}
3501
3502/**
3503 * ring_buffer_iter_reset - reset an iterator
3504 * @iter: The iterator to reset
3505 *
3506 * Resets the iterator, so that it will start from the beginning
3507 * again.
3508 */
3509void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3510{
3511	struct ring_buffer_per_cpu *cpu_buffer;
3512	unsigned long flags;
3513
3514	if (!iter)
3515		return;
3516
3517	cpu_buffer = iter->cpu_buffer;
3518
3519	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3520	rb_iter_reset(iter);
3521	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3522}
3523EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3524
3525/**
3526 * ring_buffer_iter_empty - check if an iterator has no more to read
3527 * @iter: The iterator to check
3528 */
3529int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3530{
3531	struct ring_buffer_per_cpu *cpu_buffer;
3532	struct buffer_page *reader;
3533	struct buffer_page *head_page;
3534	struct buffer_page *commit_page;
3535	unsigned commit;
3536
3537	cpu_buffer = iter->cpu_buffer;
3538
3539	/* Remember, trace recording is off when iterator is in use */
3540	reader = cpu_buffer->reader_page;
3541	head_page = cpu_buffer->head_page;
3542	commit_page = cpu_buffer->commit_page;
3543	commit = rb_page_commit(commit_page);
3544
3545	return ((iter->head_page == commit_page && iter->head == commit) ||
3546		(iter->head_page == reader && commit_page == head_page &&
3547		 head_page->read == commit &&
3548		 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3549}
3550EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3551
3552static void
3553rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3554		     struct ring_buffer_event *event)
3555{
3556	u64 delta;
3557
3558	switch (event->type_len) {
3559	case RINGBUF_TYPE_PADDING:
3560		return;
3561
3562	case RINGBUF_TYPE_TIME_EXTEND:
3563		delta = ring_buffer_event_time_stamp(event);
3564		cpu_buffer->read_stamp += delta;
3565		return;
3566
3567	case RINGBUF_TYPE_TIME_STAMP:
3568		delta = ring_buffer_event_time_stamp(event);
3569		cpu_buffer->read_stamp = delta;
3570		return;
3571
3572	case RINGBUF_TYPE_DATA:
3573		cpu_buffer->read_stamp += event->time_delta;
3574		return;
3575
3576	default:
3577		BUG();
3578	}
3579	return;
3580}
3581
3582static void
3583rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3584			  struct ring_buffer_event *event)
3585{
3586	u64 delta;
3587
3588	switch (event->type_len) {
3589	case RINGBUF_TYPE_PADDING:
3590		return;
3591
3592	case RINGBUF_TYPE_TIME_EXTEND:
3593		delta = ring_buffer_event_time_stamp(event);
3594		iter->read_stamp += delta;
3595		return;
3596
3597	case RINGBUF_TYPE_TIME_STAMP:
3598		delta = ring_buffer_event_time_stamp(event);
3599		iter->read_stamp = delta;
3600		return;
3601
3602	case RINGBUF_TYPE_DATA:
3603		iter->read_stamp += event->time_delta;
3604		return;
3605
3606	default:
3607		BUG();
3608	}
3609	return;
3610}
3611
3612static struct buffer_page *
3613rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3614{
3615	struct buffer_page *reader = NULL;
3616	unsigned long overwrite;
3617	unsigned long flags;
3618	int nr_loops = 0;
3619	int ret;
3620
3621	local_irq_save(flags);
3622	arch_spin_lock(&cpu_buffer->lock);
3623
3624 again:
3625	/*
3626	 * This should normally only loop twice. But because the
3627	 * start of the reader inserts an empty page, it causes
3628	 * a case where we will loop three times. There should be no
3629	 * reason to loop four times (that I know of).
3630	 */
3631	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3632		reader = NULL;
3633		goto out;
3634	}
3635
3636	reader = cpu_buffer->reader_page;
3637
3638	/* If there's more to read, return this page */
3639	if (cpu_buffer->reader_page->read < rb_page_size(reader))
3640		goto out;
3641
3642	/* Never should we have an index greater than the size */
3643	if (RB_WARN_ON(cpu_buffer,
3644		       cpu_buffer->reader_page->read > rb_page_size(reader)))
3645		goto out;
3646
3647	/* check if we caught up to the tail */
3648	reader = NULL;
3649	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3650		goto out;
3651
3652	/* Don't bother swapping if the ring buffer is empty */
3653	if (rb_num_of_entries(cpu_buffer) == 0)
3654		goto out;
3655
3656	/*
3657	 * Reset the reader page to size zero.
3658	 */
3659	local_set(&cpu_buffer->reader_page->write, 0);
3660	local_set(&cpu_buffer->reader_page->entries, 0);
3661	local_set(&cpu_buffer->reader_page->page->commit, 0);
3662	cpu_buffer->reader_page->real_end = 0;
3663
3664 spin:
3665	/*
3666	 * Splice the empty reader page into the list around the head.
3667	 */
3668	reader = rb_set_head_page(cpu_buffer);
3669	if (!reader)
3670		goto out;
3671	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3672	cpu_buffer->reader_page->list.prev = reader->list.prev;
3673
3674	/*
3675	 * cpu_buffer->pages just needs to point to the buffer, it
3676	 *  has no specific buffer page to point to. Lets move it out
3677	 *  of our way so we don't accidentally swap it.
3678	 */
3679	cpu_buffer->pages = reader->list.prev;
3680
3681	/* The reader page will be pointing to the new head */
3682	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3683
3684	/*
3685	 * We want to make sure we read the overruns after we set up our
3686	 * pointers to the next object. The writer side does a
3687	 * cmpxchg to cross pages which acts as the mb on the writer
3688	 * side. Note, the reader will constantly fail the swap
3689	 * while the writer is updating the pointers, so this
3690	 * guarantees that the overwrite recorded here is the one we
3691	 * want to compare with the last_overrun.
3692	 */
3693	smp_mb();
3694	overwrite = local_read(&(cpu_buffer->overrun));
3695
3696	/*
3697	 * Here's the tricky part.
3698	 *
3699	 * We need to move the pointer past the header page.
3700	 * But we can only do that if a writer is not currently
3701	 * moving it. The page before the header page has the
3702	 * flag bit '1' set if it is pointing to the page we want.
3703	 * but if the writer is in the process of moving it
3704	 * than it will be '2' or already moved '0'.
3705	 */
3706
3707	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3708
3709	/*
3710	 * If we did not convert it, then we must try again.
3711	 */
3712	if (!ret)
3713		goto spin;
3714
3715	/*
3716	 * Yeah! We succeeded in replacing the page.
3717	 *
3718	 * Now make the new head point back to the reader page.
3719	 */
3720	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3721	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3722
 
 
3723	/* Finally update the reader page to the new head */
3724	cpu_buffer->reader_page = reader;
3725	cpu_buffer->reader_page->read = 0;
3726
3727	if (overwrite != cpu_buffer->last_overrun) {
3728		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3729		cpu_buffer->last_overrun = overwrite;
3730	}
3731
3732	goto again;
3733
3734 out:
3735	/* Update the read_stamp on the first event */
3736	if (reader && reader->read == 0)
3737		cpu_buffer->read_stamp = reader->page->time_stamp;
3738
3739	arch_spin_unlock(&cpu_buffer->lock);
3740	local_irq_restore(flags);
3741
3742	return reader;
3743}
3744
3745static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3746{
3747	struct ring_buffer_event *event;
3748	struct buffer_page *reader;
3749	unsigned length;
3750
3751	reader = rb_get_reader_page(cpu_buffer);
3752
3753	/* This function should not be called when buffer is empty */
3754	if (RB_WARN_ON(cpu_buffer, !reader))
3755		return;
3756
3757	event = rb_reader_event(cpu_buffer);
3758
3759	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3760		cpu_buffer->read++;
3761
3762	rb_update_read_stamp(cpu_buffer, event);
3763
3764	length = rb_event_length(event);
3765	cpu_buffer->reader_page->read += length;
3766}
3767
3768static void rb_advance_iter(struct ring_buffer_iter *iter)
3769{
3770	struct ring_buffer_per_cpu *cpu_buffer;
3771	struct ring_buffer_event *event;
3772	unsigned length;
3773
3774	cpu_buffer = iter->cpu_buffer;
3775
3776	/*
3777	 * Check if we are at the end of the buffer.
3778	 */
3779	if (iter->head >= rb_page_size(iter->head_page)) {
3780		/* discarded commits can make the page empty */
3781		if (iter->head_page == cpu_buffer->commit_page)
3782			return;
3783		rb_inc_iter(iter);
3784		return;
3785	}
3786
3787	event = rb_iter_head_event(iter);
3788
3789	length = rb_event_length(event);
3790
3791	/*
3792	 * This should not be called to advance the header if we are
3793	 * at the tail of the buffer.
3794	 */
3795	if (RB_WARN_ON(cpu_buffer,
3796		       (iter->head_page == cpu_buffer->commit_page) &&
3797		       (iter->head + length > rb_commit_index(cpu_buffer))))
3798		return;
3799
3800	rb_update_iter_read_stamp(iter, event);
3801
3802	iter->head += length;
3803
3804	/* check for end of page padding */
3805	if ((iter->head >= rb_page_size(iter->head_page)) &&
3806	    (iter->head_page != cpu_buffer->commit_page))
3807		rb_inc_iter(iter);
3808}
3809
3810static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3811{
3812	return cpu_buffer->lost_events;
3813}
3814
3815static struct ring_buffer_event *
3816rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3817	       unsigned long *lost_events)
3818{
3819	struct ring_buffer_event *event;
3820	struct buffer_page *reader;
3821	int nr_loops = 0;
3822
3823	if (ts)
3824		*ts = 0;
3825 again:
3826	/*
3827	 * We repeat when a time extend is encountered.
3828	 * Since the time extend is always attached to a data event,
3829	 * we should never loop more than once.
3830	 * (We never hit the following condition more than twice).
3831	 */
3832	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3833		return NULL;
3834
3835	reader = rb_get_reader_page(cpu_buffer);
3836	if (!reader)
3837		return NULL;
3838
3839	event = rb_reader_event(cpu_buffer);
3840
3841	switch (event->type_len) {
3842	case RINGBUF_TYPE_PADDING:
3843		if (rb_null_event(event))
3844			RB_WARN_ON(cpu_buffer, 1);
3845		/*
3846		 * Because the writer could be discarding every
3847		 * event it creates (which would probably be bad)
3848		 * if we were to go back to "again" then we may never
3849		 * catch up, and will trigger the warn on, or lock
3850		 * the box. Return the padding, and we will release
3851		 * the current locks, and try again.
3852		 */
3853		return event;
3854
3855	case RINGBUF_TYPE_TIME_EXTEND:
3856		/* Internal data, OK to advance */
3857		rb_advance_reader(cpu_buffer);
3858		goto again;
3859
3860	case RINGBUF_TYPE_TIME_STAMP:
3861		if (ts) {
3862			*ts = ring_buffer_event_time_stamp(event);
3863			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3864							 cpu_buffer->cpu, ts);
3865		}
3866		/* Internal data, OK to advance */
3867		rb_advance_reader(cpu_buffer);
3868		goto again;
3869
3870	case RINGBUF_TYPE_DATA:
3871		if (ts && !(*ts)) {
3872			*ts = cpu_buffer->read_stamp + event->time_delta;
3873			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3874							 cpu_buffer->cpu, ts);
3875		}
3876		if (lost_events)
3877			*lost_events = rb_lost_events(cpu_buffer);
3878		return event;
3879
3880	default:
3881		BUG();
3882	}
3883
3884	return NULL;
3885}
3886EXPORT_SYMBOL_GPL(ring_buffer_peek);
3887
3888static struct ring_buffer_event *
3889rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3890{
3891	struct ring_buffer *buffer;
3892	struct ring_buffer_per_cpu *cpu_buffer;
3893	struct ring_buffer_event *event;
3894	int nr_loops = 0;
3895
3896	if (ts)
3897		*ts = 0;
3898
3899	cpu_buffer = iter->cpu_buffer;
3900	buffer = cpu_buffer->buffer;
3901
3902	/*
3903	 * Check if someone performed a consuming read to
3904	 * the buffer. A consuming read invalidates the iterator
3905	 * and we need to reset the iterator in this case.
3906	 */
3907	if (unlikely(iter->cache_read != cpu_buffer->read ||
3908		     iter->cache_reader_page != cpu_buffer->reader_page))
3909		rb_iter_reset(iter);
3910
3911 again:
3912	if (ring_buffer_iter_empty(iter))
3913		return NULL;
3914
3915	/*
3916	 * We repeat when a time extend is encountered or we hit
3917	 * the end of the page. Since the time extend is always attached
3918	 * to a data event, we should never loop more than three times.
3919	 * Once for going to next page, once on time extend, and
3920	 * finally once to get the event.
3921	 * (We never hit the following condition more than thrice).
3922	 */
3923	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3924		return NULL;
3925
3926	if (rb_per_cpu_empty(cpu_buffer))
3927		return NULL;
3928
3929	if (iter->head >= rb_page_size(iter->head_page)) {
3930		rb_inc_iter(iter);
3931		goto again;
3932	}
3933
3934	event = rb_iter_head_event(iter);
3935
3936	switch (event->type_len) {
3937	case RINGBUF_TYPE_PADDING:
3938		if (rb_null_event(event)) {
3939			rb_inc_iter(iter);
3940			goto again;
3941		}
3942		rb_advance_iter(iter);
3943		return event;
3944
3945	case RINGBUF_TYPE_TIME_EXTEND:
3946		/* Internal data, OK to advance */
3947		rb_advance_iter(iter);
3948		goto again;
3949
3950	case RINGBUF_TYPE_TIME_STAMP:
3951		if (ts) {
3952			*ts = ring_buffer_event_time_stamp(event);
3953			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3954							 cpu_buffer->cpu, ts);
3955		}
3956		/* Internal data, OK to advance */
3957		rb_advance_iter(iter);
3958		goto again;
3959
3960	case RINGBUF_TYPE_DATA:
3961		if (ts && !(*ts)) {
3962			*ts = iter->read_stamp + event->time_delta;
3963			ring_buffer_normalize_time_stamp(buffer,
3964							 cpu_buffer->cpu, ts);
3965		}
3966		return event;
3967
3968	default:
3969		BUG();
3970	}
3971
3972	return NULL;
3973}
3974EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3975
3976static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3977{
3978	if (likely(!in_nmi())) {
3979		raw_spin_lock(&cpu_buffer->reader_lock);
3980		return true;
3981	}
3982
3983	/*
3984	 * If an NMI die dumps out the content of the ring buffer
3985	 * trylock must be used to prevent a deadlock if the NMI
3986	 * preempted a task that holds the ring buffer locks. If
3987	 * we get the lock then all is fine, if not, then continue
3988	 * to do the read, but this can corrupt the ring buffer,
3989	 * so it must be permanently disabled from future writes.
3990	 * Reading from NMI is a oneshot deal.
3991	 */
3992	if (raw_spin_trylock(&cpu_buffer->reader_lock))
3993		return true;
3994
3995	/* Continue without locking, but disable the ring buffer */
3996	atomic_inc(&cpu_buffer->record_disabled);
3997	return false;
3998}
3999
4000static inline void
4001rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4002{
4003	if (likely(locked))
4004		raw_spin_unlock(&cpu_buffer->reader_lock);
4005	return;
4006}
4007
4008/**
4009 * ring_buffer_peek - peek at the next event to be read
4010 * @buffer: The ring buffer to read
4011 * @cpu: The cpu to peak at
4012 * @ts: The timestamp counter of this event.
4013 * @lost_events: a variable to store if events were lost (may be NULL)
4014 *
4015 * This will return the event that will be read next, but does
4016 * not consume the data.
4017 */
4018struct ring_buffer_event *
4019ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
4020		 unsigned long *lost_events)
4021{
4022	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4023	struct ring_buffer_event *event;
4024	unsigned long flags;
4025	bool dolock;
4026
4027	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4028		return NULL;
4029
4030 again:
4031	local_irq_save(flags);
4032	dolock = rb_reader_lock(cpu_buffer);
4033	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4034	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4035		rb_advance_reader(cpu_buffer);
4036	rb_reader_unlock(cpu_buffer, dolock);
4037	local_irq_restore(flags);
4038
4039	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4040		goto again;
4041
4042	return event;
4043}
4044
4045/**
4046 * ring_buffer_iter_peek - peek at the next event to be read
4047 * @iter: The ring buffer iterator
4048 * @ts: The timestamp counter of this event.
4049 *
4050 * This will return the event that will be read next, but does
4051 * not increment the iterator.
4052 */
4053struct ring_buffer_event *
4054ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4055{
4056	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4057	struct ring_buffer_event *event;
4058	unsigned long flags;
4059
4060 again:
4061	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4062	event = rb_iter_peek(iter, ts);
4063	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4064
4065	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4066		goto again;
4067
4068	return event;
4069}
4070
4071/**
4072 * ring_buffer_consume - return an event and consume it
4073 * @buffer: The ring buffer to get the next event from
4074 * @cpu: the cpu to read the buffer from
4075 * @ts: a variable to store the timestamp (may be NULL)
4076 * @lost_events: a variable to store if events were lost (may be NULL)
4077 *
4078 * Returns the next event in the ring buffer, and that event is consumed.
4079 * Meaning, that sequential reads will keep returning a different event,
4080 * and eventually empty the ring buffer if the producer is slower.
4081 */
4082struct ring_buffer_event *
4083ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4084		    unsigned long *lost_events)
4085{
4086	struct ring_buffer_per_cpu *cpu_buffer;
4087	struct ring_buffer_event *event = NULL;
4088	unsigned long flags;
4089	bool dolock;
4090
4091 again:
4092	/* might be called in atomic */
4093	preempt_disable();
4094
4095	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4096		goto out;
4097
4098	cpu_buffer = buffer->buffers[cpu];
4099	local_irq_save(flags);
4100	dolock = rb_reader_lock(cpu_buffer);
4101
4102	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4103	if (event) {
4104		cpu_buffer->lost_events = 0;
4105		rb_advance_reader(cpu_buffer);
4106	}
4107
4108	rb_reader_unlock(cpu_buffer, dolock);
4109	local_irq_restore(flags);
4110
4111 out:
4112	preempt_enable();
4113
4114	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4115		goto again;
4116
4117	return event;
4118}
4119EXPORT_SYMBOL_GPL(ring_buffer_consume);
4120
4121/**
4122 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4123 * @buffer: The ring buffer to read from
4124 * @cpu: The cpu buffer to iterate over
 
4125 *
4126 * This performs the initial preparations necessary to iterate
4127 * through the buffer.  Memory is allocated, buffer recording
4128 * is disabled, and the iterator pointer is returned to the caller.
4129 *
4130 * Disabling buffer recordng prevents the reading from being
4131 * corrupted. This is not a consuming read, so a producer is not
4132 * expected.
4133 *
4134 * After a sequence of ring_buffer_read_prepare calls, the user is
4135 * expected to make at least one call to ring_buffer_read_prepare_sync.
4136 * Afterwards, ring_buffer_read_start is invoked to get things going
4137 * for real.
4138 *
4139 * This overall must be paired with ring_buffer_read_finish.
4140 */
4141struct ring_buffer_iter *
4142ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4143{
4144	struct ring_buffer_per_cpu *cpu_buffer;
4145	struct ring_buffer_iter *iter;
4146
4147	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4148		return NULL;
4149
4150	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4151	if (!iter)
4152		return NULL;
4153
4154	cpu_buffer = buffer->buffers[cpu];
4155
4156	iter->cpu_buffer = cpu_buffer;
4157
4158	atomic_inc(&buffer->resize_disabled);
4159	atomic_inc(&cpu_buffer->record_disabled);
4160
4161	return iter;
4162}
4163EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4164
4165/**
4166 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4167 *
4168 * All previously invoked ring_buffer_read_prepare calls to prepare
4169 * iterators will be synchronized.  Afterwards, read_buffer_read_start
4170 * calls on those iterators are allowed.
4171 */
4172void
4173ring_buffer_read_prepare_sync(void)
4174{
4175	synchronize_sched();
4176}
4177EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4178
4179/**
4180 * ring_buffer_read_start - start a non consuming read of the buffer
4181 * @iter: The iterator returned by ring_buffer_read_prepare
4182 *
4183 * This finalizes the startup of an iteration through the buffer.
4184 * The iterator comes from a call to ring_buffer_read_prepare and
4185 * an intervening ring_buffer_read_prepare_sync must have been
4186 * performed.
4187 *
4188 * Must be paired with ring_buffer_read_finish.
4189 */
4190void
4191ring_buffer_read_start(struct ring_buffer_iter *iter)
4192{
4193	struct ring_buffer_per_cpu *cpu_buffer;
4194	unsigned long flags;
4195
4196	if (!iter)
4197		return;
4198
4199	cpu_buffer = iter->cpu_buffer;
4200
4201	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4202	arch_spin_lock(&cpu_buffer->lock);
4203	rb_iter_reset(iter);
4204	arch_spin_unlock(&cpu_buffer->lock);
4205	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4206}
4207EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4208
4209/**
4210 * ring_buffer_read_finish - finish reading the iterator of the buffer
4211 * @iter: The iterator retrieved by ring_buffer_start
4212 *
4213 * This re-enables the recording to the buffer, and frees the
4214 * iterator.
4215 */
4216void
4217ring_buffer_read_finish(struct ring_buffer_iter *iter)
4218{
4219	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4220	unsigned long flags;
4221
4222	/*
4223	 * Ring buffer is disabled from recording, here's a good place
4224	 * to check the integrity of the ring buffer.
4225	 * Must prevent readers from trying to read, as the check
4226	 * clears the HEAD page and readers require it.
4227	 */
4228	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4229	rb_check_pages(cpu_buffer);
4230	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4231
4232	atomic_dec(&cpu_buffer->record_disabled);
4233	atomic_dec(&cpu_buffer->buffer->resize_disabled);
4234	kfree(iter);
4235}
4236EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4237
4238/**
4239 * ring_buffer_read - read the next item in the ring buffer by the iterator
4240 * @iter: The ring buffer iterator
4241 * @ts: The time stamp of the event read.
4242 *
4243 * This reads the next event in the ring buffer and increments the iterator.
4244 */
4245struct ring_buffer_event *
4246ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4247{
4248	struct ring_buffer_event *event;
4249	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4250	unsigned long flags;
4251
4252	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4253 again:
4254	event = rb_iter_peek(iter, ts);
4255	if (!event)
4256		goto out;
4257
4258	if (event->type_len == RINGBUF_TYPE_PADDING)
4259		goto again;
4260
4261	rb_advance_iter(iter);
4262 out:
4263	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4264
4265	return event;
4266}
4267EXPORT_SYMBOL_GPL(ring_buffer_read);
4268
4269/**
4270 * ring_buffer_size - return the size of the ring buffer (in bytes)
4271 * @buffer: The ring buffer.
4272 */
4273unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4274{
4275	/*
4276	 * Earlier, this method returned
4277	 *	BUF_PAGE_SIZE * buffer->nr_pages
4278	 * Since the nr_pages field is now removed, we have converted this to
4279	 * return the per cpu buffer value.
4280	 */
4281	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4282		return 0;
4283
4284	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4285}
4286EXPORT_SYMBOL_GPL(ring_buffer_size);
4287
4288static void
4289rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4290{
4291	rb_head_page_deactivate(cpu_buffer);
4292
4293	cpu_buffer->head_page
4294		= list_entry(cpu_buffer->pages, struct buffer_page, list);
4295	local_set(&cpu_buffer->head_page->write, 0);
4296	local_set(&cpu_buffer->head_page->entries, 0);
4297	local_set(&cpu_buffer->head_page->page->commit, 0);
4298
4299	cpu_buffer->head_page->read = 0;
4300
4301	cpu_buffer->tail_page = cpu_buffer->head_page;
4302	cpu_buffer->commit_page = cpu_buffer->head_page;
4303
4304	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4305	INIT_LIST_HEAD(&cpu_buffer->new_pages);
4306	local_set(&cpu_buffer->reader_page->write, 0);
4307	local_set(&cpu_buffer->reader_page->entries, 0);
4308	local_set(&cpu_buffer->reader_page->page->commit, 0);
4309	cpu_buffer->reader_page->read = 0;
4310
4311	local_set(&cpu_buffer->entries_bytes, 0);
4312	local_set(&cpu_buffer->overrun, 0);
4313	local_set(&cpu_buffer->commit_overrun, 0);
4314	local_set(&cpu_buffer->dropped_events, 0);
4315	local_set(&cpu_buffer->entries, 0);
4316	local_set(&cpu_buffer->committing, 0);
4317	local_set(&cpu_buffer->commits, 0);
 
 
 
 
4318	cpu_buffer->read = 0;
4319	cpu_buffer->read_bytes = 0;
4320
4321	cpu_buffer->write_stamp = 0;
4322	cpu_buffer->read_stamp = 0;
4323
4324	cpu_buffer->lost_events = 0;
4325	cpu_buffer->last_overrun = 0;
4326
4327	rb_head_page_activate(cpu_buffer);
4328}
4329
4330/**
4331 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4332 * @buffer: The ring buffer to reset a per cpu buffer of
4333 * @cpu: The CPU buffer to be reset
4334 */
4335void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4336{
4337	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4338	unsigned long flags;
4339
4340	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4341		return;
4342
4343	atomic_inc(&buffer->resize_disabled);
4344	atomic_inc(&cpu_buffer->record_disabled);
4345
4346	/* Make sure all commits have finished */
4347	synchronize_sched();
4348
4349	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4350
4351	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4352		goto out;
4353
4354	arch_spin_lock(&cpu_buffer->lock);
4355
4356	rb_reset_cpu(cpu_buffer);
4357
4358	arch_spin_unlock(&cpu_buffer->lock);
4359
4360 out:
4361	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4362
4363	atomic_dec(&cpu_buffer->record_disabled);
4364	atomic_dec(&buffer->resize_disabled);
4365}
4366EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4367
4368/**
4369 * ring_buffer_reset - reset a ring buffer
4370 * @buffer: The ring buffer to reset all cpu buffers
4371 */
4372void ring_buffer_reset(struct ring_buffer *buffer)
4373{
4374	int cpu;
4375
4376	for_each_buffer_cpu(buffer, cpu)
4377		ring_buffer_reset_cpu(buffer, cpu);
4378}
4379EXPORT_SYMBOL_GPL(ring_buffer_reset);
4380
4381/**
4382 * rind_buffer_empty - is the ring buffer empty?
4383 * @buffer: The ring buffer to test
4384 */
4385bool ring_buffer_empty(struct ring_buffer *buffer)
4386{
4387	struct ring_buffer_per_cpu *cpu_buffer;
4388	unsigned long flags;
4389	bool dolock;
4390	int cpu;
4391	int ret;
4392
4393	/* yes this is racy, but if you don't like the race, lock the buffer */
4394	for_each_buffer_cpu(buffer, cpu) {
4395		cpu_buffer = buffer->buffers[cpu];
4396		local_irq_save(flags);
4397		dolock = rb_reader_lock(cpu_buffer);
4398		ret = rb_per_cpu_empty(cpu_buffer);
4399		rb_reader_unlock(cpu_buffer, dolock);
4400		local_irq_restore(flags);
4401
4402		if (!ret)
4403			return false;
4404	}
4405
4406	return true;
4407}
4408EXPORT_SYMBOL_GPL(ring_buffer_empty);
4409
4410/**
4411 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4412 * @buffer: The ring buffer
4413 * @cpu: The CPU buffer to test
4414 */
4415bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4416{
4417	struct ring_buffer_per_cpu *cpu_buffer;
4418	unsigned long flags;
4419	bool dolock;
4420	int ret;
4421
4422	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4423		return true;
4424
4425	cpu_buffer = buffer->buffers[cpu];
4426	local_irq_save(flags);
4427	dolock = rb_reader_lock(cpu_buffer);
4428	ret = rb_per_cpu_empty(cpu_buffer);
4429	rb_reader_unlock(cpu_buffer, dolock);
4430	local_irq_restore(flags);
4431
4432	return ret;
4433}
4434EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4435
4436#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4437/**
4438 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4439 * @buffer_a: One buffer to swap with
4440 * @buffer_b: The other buffer to swap with
4441 *
4442 * This function is useful for tracers that want to take a "snapshot"
4443 * of a CPU buffer and has another back up buffer lying around.
4444 * it is expected that the tracer handles the cpu buffer not being
4445 * used at the moment.
4446 */
4447int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4448			 struct ring_buffer *buffer_b, int cpu)
4449{
4450	struct ring_buffer_per_cpu *cpu_buffer_a;
4451	struct ring_buffer_per_cpu *cpu_buffer_b;
4452	int ret = -EINVAL;
4453
4454	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4455	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
4456		goto out;
4457
4458	cpu_buffer_a = buffer_a->buffers[cpu];
4459	cpu_buffer_b = buffer_b->buffers[cpu];
4460
4461	/* At least make sure the two buffers are somewhat the same */
4462	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4463		goto out;
4464
4465	ret = -EAGAIN;
4466
4467	if (atomic_read(&buffer_a->record_disabled))
4468		goto out;
4469
4470	if (atomic_read(&buffer_b->record_disabled))
4471		goto out;
4472
4473	if (atomic_read(&cpu_buffer_a->record_disabled))
4474		goto out;
4475
4476	if (atomic_read(&cpu_buffer_b->record_disabled))
4477		goto out;
4478
4479	/*
4480	 * We can't do a synchronize_sched here because this
4481	 * function can be called in atomic context.
4482	 * Normally this will be called from the same CPU as cpu.
4483	 * If not it's up to the caller to protect this.
4484	 */
4485	atomic_inc(&cpu_buffer_a->record_disabled);
4486	atomic_inc(&cpu_buffer_b->record_disabled);
4487
4488	ret = -EBUSY;
4489	if (local_read(&cpu_buffer_a->committing))
4490		goto out_dec;
4491	if (local_read(&cpu_buffer_b->committing))
4492		goto out_dec;
4493
4494	buffer_a->buffers[cpu] = cpu_buffer_b;
4495	buffer_b->buffers[cpu] = cpu_buffer_a;
4496
4497	cpu_buffer_b->buffer = buffer_a;
4498	cpu_buffer_a->buffer = buffer_b;
4499
4500	ret = 0;
4501
4502out_dec:
4503	atomic_dec(&cpu_buffer_a->record_disabled);
4504	atomic_dec(&cpu_buffer_b->record_disabled);
4505out:
4506	return ret;
4507}
4508EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4509#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4510
4511/**
4512 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4513 * @buffer: the buffer to allocate for.
4514 * @cpu: the cpu buffer to allocate.
4515 *
4516 * This function is used in conjunction with ring_buffer_read_page.
4517 * When reading a full page from the ring buffer, these functions
4518 * can be used to speed up the process. The calling function should
4519 * allocate a few pages first with this function. Then when it
4520 * needs to get pages from the ring buffer, it passes the result
4521 * of this function into ring_buffer_read_page, which will swap
4522 * the page that was allocated, with the read page of the buffer.
4523 *
4524 * Returns:
4525 *  The page allocated, or ERR_PTR
4526 */
4527void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4528{
4529	struct ring_buffer_per_cpu *cpu_buffer;
4530	struct buffer_data_page *bpage = NULL;
4531	unsigned long flags;
4532	struct page *page;
4533
4534	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4535		return ERR_PTR(-ENODEV);
4536
4537	cpu_buffer = buffer->buffers[cpu];
4538	local_irq_save(flags);
4539	arch_spin_lock(&cpu_buffer->lock);
4540
4541	if (cpu_buffer->free_page) {
4542		bpage = cpu_buffer->free_page;
4543		cpu_buffer->free_page = NULL;
4544	}
4545
4546	arch_spin_unlock(&cpu_buffer->lock);
4547	local_irq_restore(flags);
4548
4549	if (bpage)
4550		goto out;
4551
4552	page = alloc_pages_node(cpu_to_node(cpu),
4553				GFP_KERNEL | __GFP_NORETRY, 0);
4554	if (!page)
4555		return ERR_PTR(-ENOMEM);
4556
4557	bpage = page_address(page);
4558
4559 out:
4560	rb_init_page(bpage);
4561
4562	return bpage;
4563}
4564EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4565
4566/**
4567 * ring_buffer_free_read_page - free an allocated read page
4568 * @buffer: the buffer the page was allocate for
4569 * @cpu: the cpu buffer the page came from
4570 * @data: the page to free
4571 *
4572 * Free a page allocated from ring_buffer_alloc_read_page.
4573 */
4574void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4575{
4576	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4577	struct buffer_data_page *bpage = data;
4578	struct page *page = virt_to_page(bpage);
4579	unsigned long flags;
4580
4581	/* If the page is still in use someplace else, we can't reuse it */
4582	if (page_ref_count(page) > 1)
4583		goto out;
4584
4585	local_irq_save(flags);
4586	arch_spin_lock(&cpu_buffer->lock);
4587
4588	if (!cpu_buffer->free_page) {
4589		cpu_buffer->free_page = bpage;
4590		bpage = NULL;
4591	}
4592
4593	arch_spin_unlock(&cpu_buffer->lock);
4594	local_irq_restore(flags);
4595
4596 out:
4597	free_page((unsigned long)bpage);
4598}
4599EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4600
4601/**
4602 * ring_buffer_read_page - extract a page from the ring buffer
4603 * @buffer: buffer to extract from
4604 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4605 * @len: amount to extract
4606 * @cpu: the cpu of the buffer to extract
4607 * @full: should the extraction only happen when the page is full.
4608 *
4609 * This function will pull out a page from the ring buffer and consume it.
4610 * @data_page must be the address of the variable that was returned
4611 * from ring_buffer_alloc_read_page. This is because the page might be used
4612 * to swap with a page in the ring buffer.
4613 *
4614 * for example:
4615 *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
4616 *	if (IS_ERR(rpage))
4617 *		return PTR_ERR(rpage);
4618 *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4619 *	if (ret >= 0)
4620 *		process_page(rpage, ret);
4621 *
4622 * When @full is set, the function will not return true unless
4623 * the writer is off the reader page.
4624 *
4625 * Note: it is up to the calling functions to handle sleeps and wakeups.
4626 *  The ring buffer can be used anywhere in the kernel and can not
4627 *  blindly call wake_up. The layer that uses the ring buffer must be
4628 *  responsible for that.
4629 *
4630 * Returns:
4631 *  >=0 if data has been transferred, returns the offset of consumed data.
4632 *  <0 if no data has been transferred.
4633 */
4634int ring_buffer_read_page(struct ring_buffer *buffer,
4635			  void **data_page, size_t len, int cpu, int full)
4636{
4637	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4638	struct ring_buffer_event *event;
4639	struct buffer_data_page *bpage;
4640	struct buffer_page *reader;
4641	unsigned long missed_events;
4642	unsigned long flags;
4643	unsigned int commit;
4644	unsigned int read;
4645	u64 save_timestamp;
4646	int ret = -1;
4647
4648	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4649		goto out;
4650
4651	/*
4652	 * If len is not big enough to hold the page header, then
4653	 * we can not copy anything.
4654	 */
4655	if (len <= BUF_PAGE_HDR_SIZE)
4656		goto out;
4657
4658	len -= BUF_PAGE_HDR_SIZE;
4659
4660	if (!data_page)
4661		goto out;
4662
4663	bpage = *data_page;
4664	if (!bpage)
4665		goto out;
4666
4667	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4668
4669	reader = rb_get_reader_page(cpu_buffer);
4670	if (!reader)
4671		goto out_unlock;
4672
4673	event = rb_reader_event(cpu_buffer);
4674
4675	read = reader->read;
4676	commit = rb_page_commit(reader);
4677
4678	/* Check if any events were dropped */
4679	missed_events = cpu_buffer->lost_events;
4680
4681	/*
4682	 * If this page has been partially read or
4683	 * if len is not big enough to read the rest of the page or
4684	 * a writer is still on the page, then
4685	 * we must copy the data from the page to the buffer.
4686	 * Otherwise, we can simply swap the page with the one passed in.
4687	 */
4688	if (read || (len < (commit - read)) ||
4689	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
4690		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4691		unsigned int rpos = read;
4692		unsigned int pos = 0;
4693		unsigned int size;
4694
4695		if (full)
4696			goto out_unlock;
4697
4698		if (len > (commit - read))
4699			len = (commit - read);
4700
4701		/* Always keep the time extend and data together */
4702		size = rb_event_ts_length(event);
4703
4704		if (len < size)
4705			goto out_unlock;
4706
4707		/* save the current timestamp, since the user will need it */
4708		save_timestamp = cpu_buffer->read_stamp;
4709
4710		/* Need to copy one event at a time */
4711		do {
4712			/* We need the size of one event, because
4713			 * rb_advance_reader only advances by one event,
4714			 * whereas rb_event_ts_length may include the size of
4715			 * one or two events.
4716			 * We have already ensured there's enough space if this
4717			 * is a time extend. */
4718			size = rb_event_length(event);
4719			memcpy(bpage->data + pos, rpage->data + rpos, size);
4720
4721			len -= size;
4722
4723			rb_advance_reader(cpu_buffer);
4724			rpos = reader->read;
4725			pos += size;
4726
4727			if (rpos >= commit)
4728				break;
4729
4730			event = rb_reader_event(cpu_buffer);
4731			/* Always keep the time extend and data together */
4732			size = rb_event_ts_length(event);
4733		} while (len >= size);
4734
4735		/* update bpage */
4736		local_set(&bpage->commit, pos);
4737		bpage->time_stamp = save_timestamp;
4738
4739		/* we copied everything to the beginning */
4740		read = 0;
4741	} else {
4742		/* update the entry counter */
4743		cpu_buffer->read += rb_page_entries(reader);
4744		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4745
4746		/* swap the pages */
4747		rb_init_page(bpage);
4748		bpage = reader->page;
4749		reader->page = *data_page;
4750		local_set(&reader->write, 0);
4751		local_set(&reader->entries, 0);
4752		reader->read = 0;
4753		*data_page = bpage;
4754
4755		/*
4756		 * Use the real_end for the data size,
4757		 * This gives us a chance to store the lost events
4758		 * on the page.
4759		 */
4760		if (reader->real_end)
4761			local_set(&bpage->commit, reader->real_end);
4762	}
4763	ret = read;
4764
4765	cpu_buffer->lost_events = 0;
4766
4767	commit = local_read(&bpage->commit);
4768	/*
4769	 * Set a flag in the commit field if we lost events
4770	 */
4771	if (missed_events) {
4772		/* If there is room at the end of the page to save the
4773		 * missed events, then record it there.
4774		 */
4775		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4776			memcpy(&bpage->data[commit], &missed_events,
4777			       sizeof(missed_events));
4778			local_add(RB_MISSED_STORED, &bpage->commit);
4779			commit += sizeof(missed_events);
4780		}
4781		local_add(RB_MISSED_EVENTS, &bpage->commit);
4782	}
4783
4784	/*
4785	 * This page may be off to user land. Zero it out here.
4786	 */
4787	if (commit < BUF_PAGE_SIZE)
4788		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4789
4790 out_unlock:
4791	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4792
4793 out:
4794	return ret;
4795}
4796EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4797
4798/*
4799 * We only allocate new buffers, never free them if the CPU goes down.
4800 * If we were to free the buffer, then the user would lose any trace that was in
4801 * the buffer.
4802 */
4803int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4804{
4805	struct ring_buffer *buffer;
4806	long nr_pages_same;
4807	int cpu_i;
4808	unsigned long nr_pages;
4809
4810	buffer = container_of(node, struct ring_buffer, node);
4811	if (cpumask_test_cpu(cpu, buffer->cpumask))
4812		return 0;
4813
4814	nr_pages = 0;
4815	nr_pages_same = 1;
4816	/* check if all cpu sizes are same */
4817	for_each_buffer_cpu(buffer, cpu_i) {
4818		/* fill in the size from first enabled cpu */
4819		if (nr_pages == 0)
4820			nr_pages = buffer->buffers[cpu_i]->nr_pages;
4821		if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4822			nr_pages_same = 0;
4823			break;
4824		}
4825	}
4826	/* allocate minimum pages, user can later expand it */
4827	if (!nr_pages_same)
4828		nr_pages = 2;
4829	buffer->buffers[cpu] =
4830		rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4831	if (!buffer->buffers[cpu]) {
4832		WARN(1, "failed to allocate ring buffer on CPU %u\n",
4833		     cpu);
4834		return -ENOMEM;
4835	}
4836	smp_wmb();
4837	cpumask_set_cpu(cpu, buffer->cpumask);
4838	return 0;
4839}
4840
4841#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4842/*
4843 * This is a basic integrity check of the ring buffer.
4844 * Late in the boot cycle this test will run when configured in.
4845 * It will kick off a thread per CPU that will go into a loop
4846 * writing to the per cpu ring buffer various sizes of data.
4847 * Some of the data will be large items, some small.
4848 *
4849 * Another thread is created that goes into a spin, sending out
4850 * IPIs to the other CPUs to also write into the ring buffer.
4851 * this is to test the nesting ability of the buffer.
4852 *
4853 * Basic stats are recorded and reported. If something in the
4854 * ring buffer should happen that's not expected, a big warning
4855 * is displayed and all ring buffers are disabled.
4856 */
4857static struct task_struct *rb_threads[NR_CPUS] __initdata;
4858
4859struct rb_test_data {
4860	struct ring_buffer	*buffer;
4861	unsigned long		events;
4862	unsigned long		bytes_written;
4863	unsigned long		bytes_alloc;
4864	unsigned long		bytes_dropped;
4865	unsigned long		events_nested;
4866	unsigned long		bytes_written_nested;
4867	unsigned long		bytes_alloc_nested;
4868	unsigned long		bytes_dropped_nested;
4869	int			min_size_nested;
4870	int			max_size_nested;
4871	int			max_size;
4872	int			min_size;
4873	int			cpu;
4874	int			cnt;
4875};
4876
4877static struct rb_test_data rb_data[NR_CPUS] __initdata;
4878
4879/* 1 meg per cpu */
4880#define RB_TEST_BUFFER_SIZE	1048576
4881
4882static char rb_string[] __initdata =
4883	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4884	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4885	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4886
4887static bool rb_test_started __initdata;
4888
4889struct rb_item {
4890	int size;
4891	char str[];
4892};
4893
4894static __init int rb_write_something(struct rb_test_data *data, bool nested)
4895{
4896	struct ring_buffer_event *event;
4897	struct rb_item *item;
4898	bool started;
4899	int event_len;
4900	int size;
4901	int len;
4902	int cnt;
4903
4904	/* Have nested writes different that what is written */
4905	cnt = data->cnt + (nested ? 27 : 0);
4906
4907	/* Multiply cnt by ~e, to make some unique increment */
4908	size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4909
4910	len = size + sizeof(struct rb_item);
4911
4912	started = rb_test_started;
4913	/* read rb_test_started before checking buffer enabled */
4914	smp_rmb();
4915
4916	event = ring_buffer_lock_reserve(data->buffer, len);
4917	if (!event) {
4918		/* Ignore dropped events before test starts. */
4919		if (started) {
4920			if (nested)
4921				data->bytes_dropped += len;
4922			else
4923				data->bytes_dropped_nested += len;
4924		}
4925		return len;
4926	}
4927
4928	event_len = ring_buffer_event_length(event);
4929
4930	if (RB_WARN_ON(data->buffer, event_len < len))
4931		goto out;
4932
4933	item = ring_buffer_event_data(event);
4934	item->size = size;
4935	memcpy(item->str, rb_string, size);
4936
4937	if (nested) {
4938		data->bytes_alloc_nested += event_len;
4939		data->bytes_written_nested += len;
4940		data->events_nested++;
4941		if (!data->min_size_nested || len < data->min_size_nested)
4942			data->min_size_nested = len;
4943		if (len > data->max_size_nested)
4944			data->max_size_nested = len;
4945	} else {
4946		data->bytes_alloc += event_len;
4947		data->bytes_written += len;
4948		data->events++;
4949		if (!data->min_size || len < data->min_size)
4950			data->max_size = len;
4951		if (len > data->max_size)
4952			data->max_size = len;
4953	}
4954
4955 out:
4956	ring_buffer_unlock_commit(data->buffer, event);
4957
4958	return 0;
4959}
4960
4961static __init int rb_test(void *arg)
4962{
4963	struct rb_test_data *data = arg;
4964
4965	while (!kthread_should_stop()) {
4966		rb_write_something(data, false);
4967		data->cnt++;
4968
4969		set_current_state(TASK_INTERRUPTIBLE);
4970		/* Now sleep between a min of 100-300us and a max of 1ms */
4971		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4972	}
4973
4974	return 0;
4975}
4976
4977static __init void rb_ipi(void *ignore)
4978{
4979	struct rb_test_data *data;
4980	int cpu = smp_processor_id();
4981
4982	data = &rb_data[cpu];
4983	rb_write_something(data, true);
4984}
4985
4986static __init int rb_hammer_test(void *arg)
4987{
4988	while (!kthread_should_stop()) {
4989
4990		/* Send an IPI to all cpus to write data! */
4991		smp_call_function(rb_ipi, NULL, 1);
4992		/* No sleep, but for non preempt, let others run */
4993		schedule();
4994	}
4995
4996	return 0;
4997}
4998
4999static __init int test_ringbuffer(void)
5000{
5001	struct task_struct *rb_hammer;
5002	struct ring_buffer *buffer;
5003	int cpu;
5004	int ret = 0;
5005
5006	pr_info("Running ring buffer tests...\n");
5007
5008	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5009	if (WARN_ON(!buffer))
5010		return 0;
5011
5012	/* Disable buffer so that threads can't write to it yet */
5013	ring_buffer_record_off(buffer);
5014
5015	for_each_online_cpu(cpu) {
5016		rb_data[cpu].buffer = buffer;
5017		rb_data[cpu].cpu = cpu;
5018		rb_data[cpu].cnt = cpu;
5019		rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5020						 "rbtester/%d", cpu);
5021		if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5022			pr_cont("FAILED\n");
5023			ret = PTR_ERR(rb_threads[cpu]);
5024			goto out_free;
5025		}
5026
5027		kthread_bind(rb_threads[cpu], cpu);
5028 		wake_up_process(rb_threads[cpu]);
5029	}
5030
5031	/* Now create the rb hammer! */
5032	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5033	if (WARN_ON(IS_ERR(rb_hammer))) {
5034		pr_cont("FAILED\n");
5035		ret = PTR_ERR(rb_hammer);
5036		goto out_free;
5037	}
5038
5039	ring_buffer_record_on(buffer);
5040	/*
5041	 * Show buffer is enabled before setting rb_test_started.
5042	 * Yes there's a small race window where events could be
5043	 * dropped and the thread wont catch it. But when a ring
5044	 * buffer gets enabled, there will always be some kind of
5045	 * delay before other CPUs see it. Thus, we don't care about
5046	 * those dropped events. We care about events dropped after
5047	 * the threads see that the buffer is active.
5048	 */
5049	smp_wmb();
5050	rb_test_started = true;
5051
5052	set_current_state(TASK_INTERRUPTIBLE);
5053	/* Just run for 10 seconds */;
5054	schedule_timeout(10 * HZ);
5055
5056	kthread_stop(rb_hammer);
5057
5058 out_free:
5059	for_each_online_cpu(cpu) {
5060		if (!rb_threads[cpu])
5061			break;
5062		kthread_stop(rb_threads[cpu]);
5063	}
5064	if (ret) {
5065		ring_buffer_free(buffer);
5066		return ret;
5067	}
5068
5069	/* Report! */
5070	pr_info("finished\n");
5071	for_each_online_cpu(cpu) {
5072		struct ring_buffer_event *event;
5073		struct rb_test_data *data = &rb_data[cpu];
5074		struct rb_item *item;
5075		unsigned long total_events;
5076		unsigned long total_dropped;
5077		unsigned long total_written;
5078		unsigned long total_alloc;
5079		unsigned long total_read = 0;
5080		unsigned long total_size = 0;
5081		unsigned long total_len = 0;
5082		unsigned long total_lost = 0;
5083		unsigned long lost;
5084		int big_event_size;
5085		int small_event_size;
5086
5087		ret = -1;
5088
5089		total_events = data->events + data->events_nested;
5090		total_written = data->bytes_written + data->bytes_written_nested;
5091		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5092		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5093
5094		big_event_size = data->max_size + data->max_size_nested;
5095		small_event_size = data->min_size + data->min_size_nested;
5096
5097		pr_info("CPU %d:\n", cpu);
5098		pr_info("              events:    %ld\n", total_events);
5099		pr_info("       dropped bytes:    %ld\n", total_dropped);
5100		pr_info("       alloced bytes:    %ld\n", total_alloc);
5101		pr_info("       written bytes:    %ld\n", total_written);
5102		pr_info("       biggest event:    %d\n", big_event_size);
5103		pr_info("      smallest event:    %d\n", small_event_size);
5104
5105		if (RB_WARN_ON(buffer, total_dropped))
5106			break;
5107
5108		ret = 0;
5109
5110		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5111			total_lost += lost;
5112			item = ring_buffer_event_data(event);
5113			total_len += ring_buffer_event_length(event);
5114			total_size += item->size + sizeof(struct rb_item);
5115			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5116				pr_info("FAILED!\n");
5117				pr_info("buffer had: %.*s\n", item->size, item->str);
5118				pr_info("expected:   %.*s\n", item->size, rb_string);
5119				RB_WARN_ON(buffer, 1);
5120				ret = -1;
5121				break;
5122			}
5123			total_read++;
5124		}
5125		if (ret)
5126			break;
5127
5128		ret = -1;
5129
5130		pr_info("         read events:   %ld\n", total_read);
5131		pr_info("         lost events:   %ld\n", total_lost);
5132		pr_info("        total events:   %ld\n", total_lost + total_read);
5133		pr_info("  recorded len bytes:   %ld\n", total_len);
5134		pr_info(" recorded size bytes:   %ld\n", total_size);
5135		if (total_lost)
5136			pr_info(" With dropped events, record len and size may not match\n"
5137				" alloced and written from above\n");
5138		if (!total_lost) {
5139			if (RB_WARN_ON(buffer, total_len != total_alloc ||
5140				       total_size != total_written))
5141				break;
5142		}
5143		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5144			break;
5145
5146		ret = 0;
5147	}
5148	if (!ret)
5149		pr_info("Ring buffer PASSED!\n");
5150
5151	ring_buffer_free(buffer);
5152	return 0;
5153}
5154
5155late_initcall(test_ringbuffer);
5156#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */