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 */
v3.1
 
   1/*
   2 * Generic ring buffer
   3 *
   4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
   5 */
 
   6#include <linux/ring_buffer.h>
   7#include <linux/trace_clock.h>
 
 
   8#include <linux/spinlock.h>
   9#include <linux/debugfs.h>
  10#include <linux/uaccess.h>
  11#include <linux/hardirq.h>
  12#include <linux/kmemcheck.h>
  13#include <linux/module.h>
  14#include <linux/percpu.h>
  15#include <linux/mutex.h>
 
  16#include <linux/slab.h>
  17#include <linux/init.h>
  18#include <linux/hash.h>
  19#include <linux/list.h>
  20#include <linux/cpu.h>
  21#include <linux/fs.h>
  22
  23#include <asm/local.h>
  24#include "trace.h"
 
  25
  26/*
  27 * The ring buffer header is special. We must manually up keep it.
  28 */
  29int ring_buffer_print_entry_header(struct trace_seq *s)
  30{
  31	int ret;
  32
  33	ret = trace_seq_printf(s, "# compressed entry header\n");
  34	ret = trace_seq_printf(s, "\ttype_len    :    5 bits\n");
  35	ret = trace_seq_printf(s, "\ttime_delta  :   27 bits\n");
  36	ret = trace_seq_printf(s, "\tarray       :   32 bits\n");
  37	ret = trace_seq_printf(s, "\n");
  38	ret = trace_seq_printf(s, "\tpadding     : type == %d\n",
  39			       RINGBUF_TYPE_PADDING);
  40	ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
  41			       RINGBUF_TYPE_TIME_EXTEND);
  42	ret = trace_seq_printf(s, "\tdata max type_len  == %d\n",
  43			       RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
  44
  45	return ret;
  46}
  47
  48/*
  49 * The ring buffer is made up of a list of pages. A separate list of pages is
  50 * allocated for each CPU. A writer may only write to a buffer that is
  51 * associated with the CPU it is currently executing on.  A reader may read
  52 * from any per cpu buffer.
  53 *
  54 * The reader is special. For each per cpu buffer, the reader has its own
  55 * reader page. When a reader has read the entire reader page, this reader
  56 * page is swapped with another page in the ring buffer.
  57 *
  58 * Now, as long as the writer is off the reader page, the reader can do what
  59 * ever it wants with that page. The writer will never write to that page
  60 * again (as long as it is out of the ring buffer).
  61 *
  62 * Here's some silly ASCII art.
  63 *
  64 *   +------+
  65 *   |reader|          RING BUFFER
  66 *   |page  |
  67 *   +------+        +---+   +---+   +---+
  68 *                   |   |-->|   |-->|   |
  69 *                   +---+   +---+   +---+
  70 *                     ^               |
  71 *                     |               |
  72 *                     +---------------+
  73 *
  74 *
  75 *   +------+
  76 *   |reader|          RING BUFFER
  77 *   |page  |------------------v
  78 *   +------+        +---+   +---+   +---+
  79 *                   |   |-->|   |-->|   |
  80 *                   +---+   +---+   +---+
  81 *                     ^               |
  82 *                     |               |
  83 *                     +---------------+
  84 *
  85 *
  86 *   +------+
  87 *   |reader|          RING BUFFER
  88 *   |page  |------------------v
  89 *   +------+        +---+   +---+   +---+
  90 *      ^            |   |-->|   |-->|   |
  91 *      |            +---+   +---+   +---+
  92 *      |                              |
  93 *      |                              |
  94 *      +------------------------------+
  95 *
  96 *
  97 *   +------+
  98 *   |buffer|          RING BUFFER
  99 *   |page  |------------------v
 100 *   +------+        +---+   +---+   +---+
 101 *      ^            |   |   |   |-->|   |
 102 *      |   New      +---+   +---+   +---+
 103 *      |  Reader------^               |
 104 *      |   page                       |
 105 *      +------------------------------+
 106 *
 107 *
 108 * After we make this swap, the reader can hand this page off to the splice
 109 * code and be done with it. It can even allocate a new page if it needs to
 110 * and swap that into the ring buffer.
 111 *
 112 * We will be using cmpxchg soon to make all this lockless.
 113 *
 114 */
 115
 116/*
 117 * A fast way to enable or disable all ring buffers is to
 118 * call tracing_on or tracing_off. Turning off the ring buffers
 119 * prevents all ring buffers from being recorded to.
 120 * Turning this switch on, makes it OK to write to the
 121 * ring buffer, if the ring buffer is enabled itself.
 122 *
 123 * There's three layers that must be on in order to write
 124 * to the ring buffer.
 125 *
 126 * 1) This global flag must be set.
 127 * 2) The ring buffer must be enabled for recording.
 128 * 3) The per cpu buffer must be enabled for recording.
 129 *
 130 * In case of an anomaly, this global flag has a bit set that
 131 * will permantly disable all ring buffers.
 132 */
 133
 134/*
 135 * Global flag to disable all recording to ring buffers
 136 *  This has two bits: ON, DISABLED
 137 *
 138 *  ON   DISABLED
 139 * ---- ----------
 140 *   0      0        : ring buffers are off
 141 *   1      0        : ring buffers are on
 142 *   X      1        : ring buffers are permanently disabled
 143 */
 144
 145enum {
 146	RB_BUFFERS_ON_BIT	= 0,
 147	RB_BUFFERS_DISABLED_BIT	= 1,
 148};
 149
 150enum {
 151	RB_BUFFERS_ON		= 1 << RB_BUFFERS_ON_BIT,
 152	RB_BUFFERS_DISABLED	= 1 << RB_BUFFERS_DISABLED_BIT,
 153};
 154
 155static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
 156
 157#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
 158
 159/**
 160 * tracing_on - enable all tracing buffers
 161 *
 162 * This function enables all tracing buffers that may have been
 163 * disabled with tracing_off.
 164 */
 165void tracing_on(void)
 166{
 167	set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
 168}
 169EXPORT_SYMBOL_GPL(tracing_on);
 170
 171/**
 172 * tracing_off - turn off all tracing buffers
 173 *
 174 * This function stops all tracing buffers from recording data.
 175 * It does not disable any overhead the tracers themselves may
 176 * be causing. This function simply causes all recording to
 177 * the ring buffers to fail.
 178 */
 179void tracing_off(void)
 180{
 181	clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
 182}
 183EXPORT_SYMBOL_GPL(tracing_off);
 184
 185/**
 186 * tracing_off_permanent - permanently disable ring buffers
 187 *
 188 * This function, once called, will disable all ring buffers
 189 * permanently.
 190 */
 191void tracing_off_permanent(void)
 192{
 193	set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
 194}
 195
 196/**
 197 * tracing_is_on - show state of ring buffers enabled
 198 */
 199int tracing_is_on(void)
 200{
 201	return ring_buffer_flags == RB_BUFFERS_ON;
 202}
 203EXPORT_SYMBOL_GPL(tracing_is_on);
 204
 205#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
 206#define RB_ALIGNMENT		4U
 207#define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 208#define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
 209
 210#if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
 211# define RB_FORCE_8BYTE_ALIGNMENT	0
 212# define RB_ARCH_ALIGNMENT		RB_ALIGNMENT
 213#else
 214# define RB_FORCE_8BYTE_ALIGNMENT	1
 215# define RB_ARCH_ALIGNMENT		8U
 216#endif
 217
 218/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
 219#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
 220
 221enum {
 222	RB_LEN_TIME_EXTEND = 8,
 223	RB_LEN_TIME_STAMP = 16,
 224};
 225
 226#define skip_time_extend(event) \
 227	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
 228
 
 
 
 229static inline int rb_null_event(struct ring_buffer_event *event)
 230{
 231	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
 232}
 233
 234static void rb_event_set_padding(struct ring_buffer_event *event)
 235{
 236	/* padding has a NULL time_delta */
 237	event->type_len = RINGBUF_TYPE_PADDING;
 238	event->time_delta = 0;
 239}
 240
 241static unsigned
 242rb_event_data_length(struct ring_buffer_event *event)
 243{
 244	unsigned length;
 245
 246	if (event->type_len)
 247		length = event->type_len * RB_ALIGNMENT;
 248	else
 249		length = event->array[0];
 250	return length + RB_EVNT_HDR_SIZE;
 251}
 252
 253/*
 254 * Return the length of the given event. Will return
 255 * the length of the time extend if the event is a
 256 * time extend.
 257 */
 258static inline unsigned
 259rb_event_length(struct ring_buffer_event *event)
 260{
 261	switch (event->type_len) {
 262	case RINGBUF_TYPE_PADDING:
 263		if (rb_null_event(event))
 264			/* undefined */
 265			return -1;
 266		return  event->array[0] + RB_EVNT_HDR_SIZE;
 267
 268	case RINGBUF_TYPE_TIME_EXTEND:
 269		return RB_LEN_TIME_EXTEND;
 270
 271	case RINGBUF_TYPE_TIME_STAMP:
 272		return RB_LEN_TIME_STAMP;
 273
 274	case RINGBUF_TYPE_DATA:
 275		return rb_event_data_length(event);
 276	default:
 277		BUG();
 278	}
 279	/* not hit */
 280	return 0;
 281}
 282
 283/*
 284 * Return total length of time extend and data,
 285 *   or just the event length for all other events.
 286 */
 287static inline unsigned
 288rb_event_ts_length(struct ring_buffer_event *event)
 289{
 290	unsigned len = 0;
 291
 292	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
 293		/* time extends include the data event after it */
 294		len = RB_LEN_TIME_EXTEND;
 295		event = skip_time_extend(event);
 296	}
 297	return len + rb_event_length(event);
 298}
 299
 300/**
 301 * ring_buffer_event_length - return the length of the event
 302 * @event: the event to get the length of
 303 *
 304 * Returns the size of the data load of a data event.
 305 * If the event is something other than a data event, it
 306 * returns the size of the event itself. With the exception
 307 * of a TIME EXTEND, where it still returns the size of the
 308 * data load of the data event after it.
 309 */
 310unsigned ring_buffer_event_length(struct ring_buffer_event *event)
 311{
 312	unsigned length;
 313
 314	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
 315		event = skip_time_extend(event);
 316
 317	length = rb_event_length(event);
 318	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 319		return length;
 320	length -= RB_EVNT_HDR_SIZE;
 321	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
 322                length -= sizeof(event->array[0]);
 323	return length;
 324}
 325EXPORT_SYMBOL_GPL(ring_buffer_event_length);
 326
 327/* inline for ring buffer fast paths */
 328static void *
 329rb_event_data(struct ring_buffer_event *event)
 330{
 331	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
 332		event = skip_time_extend(event);
 333	BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 334	/* If length is in len field, then array[0] has the data */
 335	if (event->type_len)
 336		return (void *)&event->array[0];
 337	/* Otherwise length is in array[0] and array[1] has the data */
 338	return (void *)&event->array[1];
 339}
 340
 341/**
 342 * ring_buffer_event_data - return the data of the event
 343 * @event: the event to get the data from
 344 */
 345void *ring_buffer_event_data(struct ring_buffer_event *event)
 346{
 347	return rb_event_data(event);
 348}
 349EXPORT_SYMBOL_GPL(ring_buffer_event_data);
 350
 351#define for_each_buffer_cpu(buffer, cpu)		\
 352	for_each_cpu(cpu, buffer->cpumask)
 353
 354#define TS_SHIFT	27
 355#define TS_MASK		((1ULL << TS_SHIFT) - 1)
 356#define TS_DELTA_TEST	(~TS_MASK)
 357
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 358/* Flag when events were overwritten */
 359#define RB_MISSED_EVENTS	(1 << 31)
 360/* Missed count stored at end */
 361#define RB_MISSED_STORED	(1 << 30)
 362
 
 
 363struct buffer_data_page {
 364	u64		 time_stamp;	/* page time stamp */
 365	local_t		 commit;	/* write committed index */
 366	unsigned char	 data[];	/* data of buffer page */
 367};
 368
 369/*
 370 * Note, the buffer_page list must be first. The buffer pages
 371 * are allocated in cache lines, which means that each buffer
 372 * page will be at the beginning of a cache line, and thus
 373 * the least significant bits will be zero. We use this to
 374 * add flags in the list struct pointers, to make the ring buffer
 375 * lockless.
 376 */
 377struct buffer_page {
 378	struct list_head list;		/* list of buffer pages */
 379	local_t		 write;		/* index for next write */
 380	unsigned	 read;		/* index for next read */
 381	local_t		 entries;	/* entries on this page */
 382	unsigned long	 real_end;	/* real end of data */
 383	struct buffer_data_page *page;	/* Actual data page */
 384};
 385
 386/*
 387 * The buffer page counters, write and entries, must be reset
 388 * atomically when crossing page boundaries. To synchronize this
 389 * update, two counters are inserted into the number. One is
 390 * the actual counter for the write position or count on the page.
 391 *
 392 * The other is a counter of updaters. Before an update happens
 393 * the update partition of the counter is incremented. This will
 394 * allow the updater to update the counter atomically.
 395 *
 396 * The counter is 20 bits, and the state data is 12.
 397 */
 398#define RB_WRITE_MASK		0xfffff
 399#define RB_WRITE_INTCNT		(1 << 20)
 400
 401static void rb_init_page(struct buffer_data_page *bpage)
 402{
 403	local_set(&bpage->commit, 0);
 404}
 405
 406/**
 407 * ring_buffer_page_len - the size of data on the page.
 408 * @page: The page to read
 409 *
 410 * Returns the amount of data on the page, including buffer page header.
 411 */
 412size_t ring_buffer_page_len(void *page)
 413{
 414	return local_read(&((struct buffer_data_page *)page)->commit)
 415		+ BUF_PAGE_HDR_SIZE;
 416}
 417
 418/*
 419 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
 420 * this issue out.
 421 */
 422static void free_buffer_page(struct buffer_page *bpage)
 423{
 424	free_page((unsigned long)bpage->page);
 425	kfree(bpage);
 426}
 427
 428/*
 429 * We need to fit the time_stamp delta into 27 bits.
 430 */
 431static inline int test_time_stamp(u64 delta)
 432{
 433	if (delta & TS_DELTA_TEST)
 434		return 1;
 435	return 0;
 436}
 437
 438#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
 439
 440/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
 441#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
 442
 443int ring_buffer_print_page_header(struct trace_seq *s)
 444{
 445	struct buffer_data_page field;
 446	int ret;
 447
 448	ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
 449			       "offset:0;\tsize:%u;\tsigned:%u;\n",
 450			       (unsigned int)sizeof(field.time_stamp),
 451			       (unsigned int)is_signed_type(u64));
 452
 453	ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
 454			       "offset:%u;\tsize:%u;\tsigned:%u;\n",
 455			       (unsigned int)offsetof(typeof(field), commit),
 456			       (unsigned int)sizeof(field.commit),
 457			       (unsigned int)is_signed_type(long));
 458
 459	ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
 460			       "offset:%u;\tsize:%u;\tsigned:%u;\n",
 461			       (unsigned int)offsetof(typeof(field), commit),
 462			       1,
 463			       (unsigned int)is_signed_type(long));
 464
 465	ret = trace_seq_printf(s, "\tfield: char data;\t"
 466			       "offset:%u;\tsize:%u;\tsigned:%u;\n",
 467			       (unsigned int)offsetof(typeof(field), data),
 468			       (unsigned int)BUF_PAGE_SIZE,
 469			       (unsigned int)is_signed_type(char));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 470
 471	return ret;
 472}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 473
 474/*
 475 * head_page == tail_page && head == tail then buffer is empty.
 476 */
 477struct ring_buffer_per_cpu {
 478	int				cpu;
 479	atomic_t			record_disabled;
 480	struct ring_buffer		*buffer;
 481	spinlock_t			reader_lock;	/* serialize readers */
 482	arch_spinlock_t			lock;
 483	struct lock_class_key		lock_key;
 
 
 
 484	struct list_head		*pages;
 485	struct buffer_page		*head_page;	/* read from head */
 486	struct buffer_page		*tail_page;	/* write to tail */
 487	struct buffer_page		*commit_page;	/* committed pages */
 488	struct buffer_page		*reader_page;
 489	unsigned long			lost_events;
 490	unsigned long			last_overrun;
 
 
 
 
 491	local_t				commit_overrun;
 492	local_t				overrun;
 493	local_t				entries;
 494	local_t				committing;
 495	local_t				commits;
 
 
 
 
 496	unsigned long			read;
 
 497	u64				write_stamp;
 498	u64				read_stamp;
 
 
 
 
 
 
 
 499};
 500
 501struct ring_buffer {
 502	unsigned			pages;
 503	unsigned			flags;
 504	int				cpus;
 505	atomic_t			record_disabled;
 
 506	cpumask_var_t			cpumask;
 507
 508	struct lock_class_key		*reader_lock_key;
 509
 510	struct mutex			mutex;
 511
 512	struct ring_buffer_per_cpu	**buffers;
 513
 514#ifdef CONFIG_HOTPLUG_CPU
 515	struct notifier_block		cpu_notify;
 516#endif
 517	u64				(*clock)(void);
 
 
 
 518};
 519
 520struct ring_buffer_iter {
 521	struct ring_buffer_per_cpu	*cpu_buffer;
 522	unsigned long			head;
 523	struct buffer_page		*head_page;
 524	struct buffer_page		*cache_reader_page;
 525	unsigned long			cache_read;
 526	u64				read_stamp;
 527};
 528
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 529/* buffer may be either ring_buffer or ring_buffer_per_cpu */
 530#define RB_WARN_ON(b, cond)						\
 531	({								\
 532		int _____ret = unlikely(cond);				\
 533		if (_____ret) {						\
 534			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
 535				struct ring_buffer_per_cpu *__b =	\
 536					(void *)b;			\
 537				atomic_inc(&__b->buffer->record_disabled); \
 538			} else						\
 539				atomic_inc(&b->record_disabled);	\
 540			WARN_ON(1);					\
 541		}							\
 542		_____ret;						\
 543	})
 544
 545/* Up this if you want to test the TIME_EXTENTS and normalization */
 546#define DEBUG_SHIFT 0
 547
 548static inline u64 rb_time_stamp(struct ring_buffer *buffer)
 549{
 550	/* shift to debug/test normalization and TIME_EXTENTS */
 551	return buffer->clock() << DEBUG_SHIFT;
 552}
 553
 554u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
 555{
 556	u64 time;
 557
 558	preempt_disable_notrace();
 559	time = rb_time_stamp(buffer);
 560	preempt_enable_no_resched_notrace();
 561
 562	return time;
 563}
 564EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
 565
 566void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
 567				      int cpu, u64 *ts)
 568{
 569	/* Just stupid testing the normalize function and deltas */
 570	*ts >>= DEBUG_SHIFT;
 571}
 572EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
 573
 574/*
 575 * Making the ring buffer lockless makes things tricky.
 576 * Although writes only happen on the CPU that they are on,
 577 * and they only need to worry about interrupts. Reads can
 578 * happen on any CPU.
 579 *
 580 * The reader page is always off the ring buffer, but when the
 581 * reader finishes with a page, it needs to swap its page with
 582 * a new one from the buffer. The reader needs to take from
 583 * the head (writes go to the tail). But if a writer is in overwrite
 584 * mode and wraps, it must push the head page forward.
 585 *
 586 * Here lies the problem.
 587 *
 588 * The reader must be careful to replace only the head page, and
 589 * not another one. As described at the top of the file in the
 590 * ASCII art, the reader sets its old page to point to the next
 591 * page after head. It then sets the page after head to point to
 592 * the old reader page. But if the writer moves the head page
 593 * during this operation, the reader could end up with the tail.
 594 *
 595 * We use cmpxchg to help prevent this race. We also do something
 596 * special with the page before head. We set the LSB to 1.
 597 *
 598 * When the writer must push the page forward, it will clear the
 599 * bit that points to the head page, move the head, and then set
 600 * the bit that points to the new head page.
 601 *
 602 * We also don't want an interrupt coming in and moving the head
 603 * page on another writer. Thus we use the second LSB to catch
 604 * that too. Thus:
 605 *
 606 * head->list->prev->next        bit 1          bit 0
 607 *                              -------        -------
 608 * Normal page                     0              0
 609 * Points to head page             0              1
 610 * New head page                   1              0
 611 *
 612 * Note we can not trust the prev pointer of the head page, because:
 613 *
 614 * +----+       +-----+        +-----+
 615 * |    |------>|  T  |---X--->|  N  |
 616 * |    |<------|     |        |     |
 617 * +----+       +-----+        +-----+
 618 *   ^                           ^ |
 619 *   |          +-----+          | |
 620 *   +----------|  R  |----------+ |
 621 *              |     |<-----------+
 622 *              +-----+
 623 *
 624 * Key:  ---X-->  HEAD flag set in pointer
 625 *         T      Tail page
 626 *         R      Reader page
 627 *         N      Next page
 628 *
 629 * (see __rb_reserve_next() to see where this happens)
 630 *
 631 *  What the above shows is that the reader just swapped out
 632 *  the reader page with a page in the buffer, but before it
 633 *  could make the new header point back to the new page added
 634 *  it was preempted by a writer. The writer moved forward onto
 635 *  the new page added by the reader and is about to move forward
 636 *  again.
 637 *
 638 *  You can see, it is legitimate for the previous pointer of
 639 *  the head (or any page) not to point back to itself. But only
 640 *  temporarially.
 641 */
 642
 643#define RB_PAGE_NORMAL		0UL
 644#define RB_PAGE_HEAD		1UL
 645#define RB_PAGE_UPDATE		2UL
 646
 647
 648#define RB_FLAG_MASK		3UL
 649
 650/* PAGE_MOVED is not part of the mask */
 651#define RB_PAGE_MOVED		4UL
 652
 653/*
 654 * rb_list_head - remove any bit
 655 */
 656static struct list_head *rb_list_head(struct list_head *list)
 657{
 658	unsigned long val = (unsigned long)list;
 659
 660	return (struct list_head *)(val & ~RB_FLAG_MASK);
 661}
 662
 663/*
 664 * rb_is_head_page - test if the given page is the head page
 665 *
 666 * Because the reader may move the head_page pointer, we can
 667 * not trust what the head page is (it may be pointing to
 668 * the reader page). But if the next page is a header page,
 669 * its flags will be non zero.
 670 */
 671static inline int
 672rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
 673		struct buffer_page *page, struct list_head *list)
 674{
 675	unsigned long val;
 676
 677	val = (unsigned long)list->next;
 678
 679	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
 680		return RB_PAGE_MOVED;
 681
 682	return val & RB_FLAG_MASK;
 683}
 684
 685/*
 686 * rb_is_reader_page
 687 *
 688 * The unique thing about the reader page, is that, if the
 689 * writer is ever on it, the previous pointer never points
 690 * back to the reader page.
 691 */
 692static int rb_is_reader_page(struct buffer_page *page)
 693{
 694	struct list_head *list = page->list.prev;
 695
 696	return rb_list_head(list->next) != &page->list;
 697}
 698
 699/*
 700 * rb_set_list_to_head - set a list_head to be pointing to head.
 701 */
 702static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
 703				struct list_head *list)
 704{
 705	unsigned long *ptr;
 706
 707	ptr = (unsigned long *)&list->next;
 708	*ptr |= RB_PAGE_HEAD;
 709	*ptr &= ~RB_PAGE_UPDATE;
 710}
 711
 712/*
 713 * rb_head_page_activate - sets up head page
 714 */
 715static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
 716{
 717	struct buffer_page *head;
 718
 719	head = cpu_buffer->head_page;
 720	if (!head)
 721		return;
 722
 723	/*
 724	 * Set the previous list pointer to have the HEAD flag.
 725	 */
 726	rb_set_list_to_head(cpu_buffer, head->list.prev);
 727}
 728
 729static void rb_list_head_clear(struct list_head *list)
 730{
 731	unsigned long *ptr = (unsigned long *)&list->next;
 732
 733	*ptr &= ~RB_FLAG_MASK;
 734}
 735
 736/*
 737 * rb_head_page_dactivate - clears head page ptr (for free list)
 738 */
 739static void
 740rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
 741{
 742	struct list_head *hd;
 743
 744	/* Go through the whole list and clear any pointers found. */
 745	rb_list_head_clear(cpu_buffer->pages);
 746
 747	list_for_each(hd, cpu_buffer->pages)
 748		rb_list_head_clear(hd);
 749}
 750
 751static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
 752			    struct buffer_page *head,
 753			    struct buffer_page *prev,
 754			    int old_flag, int new_flag)
 755{
 756	struct list_head *list;
 757	unsigned long val = (unsigned long)&head->list;
 758	unsigned long ret;
 759
 760	list = &prev->list;
 761
 762	val &= ~RB_FLAG_MASK;
 763
 764	ret = cmpxchg((unsigned long *)&list->next,
 765		      val | old_flag, val | new_flag);
 766
 767	/* check if the reader took the page */
 768	if ((ret & ~RB_FLAG_MASK) != val)
 769		return RB_PAGE_MOVED;
 770
 771	return ret & RB_FLAG_MASK;
 772}
 773
 774static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
 775				   struct buffer_page *head,
 776				   struct buffer_page *prev,
 777				   int old_flag)
 778{
 779	return rb_head_page_set(cpu_buffer, head, prev,
 780				old_flag, RB_PAGE_UPDATE);
 781}
 782
 783static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
 784				 struct buffer_page *head,
 785				 struct buffer_page *prev,
 786				 int old_flag)
 787{
 788	return rb_head_page_set(cpu_buffer, head, prev,
 789				old_flag, RB_PAGE_HEAD);
 790}
 791
 792static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
 793				   struct buffer_page *head,
 794				   struct buffer_page *prev,
 795				   int old_flag)
 796{
 797	return rb_head_page_set(cpu_buffer, head, prev,
 798				old_flag, RB_PAGE_NORMAL);
 799}
 800
 801static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
 802			       struct buffer_page **bpage)
 803{
 804	struct list_head *p = rb_list_head((*bpage)->list.next);
 805
 806	*bpage = list_entry(p, struct buffer_page, list);
 807}
 808
 809static struct buffer_page *
 810rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
 811{
 812	struct buffer_page *head;
 813	struct buffer_page *page;
 814	struct list_head *list;
 815	int i;
 816
 817	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
 818		return NULL;
 819
 820	/* sanity check */
 821	list = cpu_buffer->pages;
 822	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
 823		return NULL;
 824
 825	page = head = cpu_buffer->head_page;
 826	/*
 827	 * It is possible that the writer moves the header behind
 828	 * where we started, and we miss in one loop.
 829	 * A second loop should grab the header, but we'll do
 830	 * three loops just because I'm paranoid.
 831	 */
 832	for (i = 0; i < 3; i++) {
 833		do {
 834			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
 835				cpu_buffer->head_page = page;
 836				return page;
 837			}
 838			rb_inc_page(cpu_buffer, &page);
 839		} while (page != head);
 840	}
 841
 842	RB_WARN_ON(cpu_buffer, 1);
 843
 844	return NULL;
 845}
 846
 847static int rb_head_page_replace(struct buffer_page *old,
 848				struct buffer_page *new)
 849{
 850	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
 851	unsigned long val;
 852	unsigned long ret;
 853
 854	val = *ptr & ~RB_FLAG_MASK;
 855	val |= RB_PAGE_HEAD;
 856
 857	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
 858
 859	return ret == val;
 860}
 861
 862/*
 863 * rb_tail_page_update - move the tail page forward
 864 *
 865 * Returns 1 if moved tail page, 0 if someone else did.
 866 */
 867static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
 868			       struct buffer_page *tail_page,
 869			       struct buffer_page *next_page)
 870{
 871	struct buffer_page *old_tail;
 872	unsigned long old_entries;
 873	unsigned long old_write;
 874	int ret = 0;
 875
 876	/*
 877	 * The tail page now needs to be moved forward.
 878	 *
 879	 * We need to reset the tail page, but without messing
 880	 * with possible erasing of data brought in by interrupts
 881	 * that have moved the tail page and are currently on it.
 882	 *
 883	 * We add a counter to the write field to denote this.
 884	 */
 885	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
 886	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
 887
 
 888	/*
 889	 * Just make sure we have seen our old_write and synchronize
 890	 * with any interrupts that come in.
 891	 */
 892	barrier();
 893
 894	/*
 895	 * If the tail page is still the same as what we think
 896	 * it is, then it is up to us to update the tail
 897	 * pointer.
 898	 */
 899	if (tail_page == cpu_buffer->tail_page) {
 900		/* Zero the write counter */
 901		unsigned long val = old_write & ~RB_WRITE_MASK;
 902		unsigned long eval = old_entries & ~RB_WRITE_MASK;
 903
 904		/*
 905		 * This will only succeed if an interrupt did
 906		 * not come in and change it. In which case, we
 907		 * do not want to modify it.
 908		 *
 909		 * We add (void) to let the compiler know that we do not care
 910		 * about the return value of these functions. We use the
 911		 * cmpxchg to only update if an interrupt did not already
 912		 * do it for us. If the cmpxchg fails, we don't care.
 913		 */
 914		(void)local_cmpxchg(&next_page->write, old_write, val);
 915		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
 916
 917		/*
 918		 * No need to worry about races with clearing out the commit.
 919		 * it only can increment when a commit takes place. But that
 920		 * only happens in the outer most nested commit.
 921		 */
 922		local_set(&next_page->page->commit, 0);
 923
 924		old_tail = cmpxchg(&cpu_buffer->tail_page,
 925				   tail_page, next_page);
 926
 927		if (old_tail == tail_page)
 928			ret = 1;
 929	}
 930
 931	return ret;
 932}
 933
 934static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
 935			  struct buffer_page *bpage)
 936{
 937	unsigned long val = (unsigned long)bpage;
 938
 939	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
 940		return 1;
 941
 942	return 0;
 943}
 944
 945/**
 946 * rb_check_list - make sure a pointer to a list has the last bits zero
 947 */
 948static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
 949			 struct list_head *list)
 950{
 951	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
 952		return 1;
 953	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
 954		return 1;
 955	return 0;
 956}
 957
 958/**
 959 * check_pages - integrity check of buffer pages
 960 * @cpu_buffer: CPU buffer with pages to test
 961 *
 962 * As a safety measure we check to make sure the data pages have not
 963 * been corrupted.
 964 */
 965static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
 966{
 967	struct list_head *head = cpu_buffer->pages;
 968	struct buffer_page *bpage, *tmp;
 969
 
 
 
 
 970	rb_head_page_deactivate(cpu_buffer);
 971
 972	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
 973		return -1;
 974	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
 975		return -1;
 976
 977	if (rb_check_list(cpu_buffer, head))
 978		return -1;
 979
 980	list_for_each_entry_safe(bpage, tmp, head, list) {
 981		if (RB_WARN_ON(cpu_buffer,
 982			       bpage->list.next->prev != &bpage->list))
 983			return -1;
 984		if (RB_WARN_ON(cpu_buffer,
 985			       bpage->list.prev->next != &bpage->list))
 986			return -1;
 987		if (rb_check_list(cpu_buffer, &bpage->list))
 988			return -1;
 989	}
 990
 991	rb_head_page_activate(cpu_buffer);
 992
 993	return 0;
 994}
 995
 996static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
 997			     unsigned nr_pages)
 998{
 999	struct buffer_page *bpage, *tmp;
1000	LIST_HEAD(pages);
1001	unsigned i;
 
 
 
 
 
 
 
 
 
 
 
 
1002
1003	WARN_ON(!nr_pages);
 
 
 
 
 
1004
 
 
 
 
 
 
 
 
 
 
 
1005	for (i = 0; i < nr_pages; i++) {
1006		struct page *page;
1007		/*
1008		 * __GFP_NORETRY flag makes sure that the allocation fails
1009		 * gracefully without invoking oom-killer and the system is
1010		 * not destabilized.
1011		 */
1012		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1013				    GFP_KERNEL | __GFP_NORETRY,
1014				    cpu_to_node(cpu_buffer->cpu));
1015		if (!bpage)
1016			goto free_pages;
1017
1018		rb_check_bpage(cpu_buffer, bpage);
1019
1020		list_add(&bpage->list, &pages);
1021
1022		page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu),
1023					GFP_KERNEL | __GFP_NORETRY, 0);
1024		if (!page)
1025			goto free_pages;
1026		bpage->page = page_address(page);
1027		rb_init_page(bpage->page);
 
 
 
 
 
 
 
 
 
 
 
 
 
1028	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1029
1030	/*
1031	 * The ring buffer page list is a circular list that does not
1032	 * start and end with a list head. All page list items point to
1033	 * other pages.
1034	 */
1035	cpu_buffer->pages = pages.next;
1036	list_del(&pages);
1037
 
 
1038	rb_check_pages(cpu_buffer);
1039
1040	return 0;
1041
1042 free_pages:
1043	list_for_each_entry_safe(bpage, tmp, &pages, list) {
1044		list_del_init(&bpage->list);
1045		free_buffer_page(bpage);
1046	}
1047	return -ENOMEM;
1048}
1049
1050static struct ring_buffer_per_cpu *
1051rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
1052{
1053	struct ring_buffer_per_cpu *cpu_buffer;
1054	struct buffer_page *bpage;
1055	struct page *page;
1056	int ret;
1057
1058	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1059				  GFP_KERNEL, cpu_to_node(cpu));
1060	if (!cpu_buffer)
1061		return NULL;
1062
1063	cpu_buffer->cpu = cpu;
1064	cpu_buffer->buffer = buffer;
1065	spin_lock_init(&cpu_buffer->reader_lock);
1066	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1067	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
 
 
 
 
 
1068
1069	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1070			    GFP_KERNEL, cpu_to_node(cpu));
1071	if (!bpage)
1072		goto fail_free_buffer;
1073
1074	rb_check_bpage(cpu_buffer, bpage);
1075
1076	cpu_buffer->reader_page = bpage;
1077	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1078	if (!page)
1079		goto fail_free_reader;
1080	bpage->page = page_address(page);
1081	rb_init_page(bpage->page);
1082
1083	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
 
1084
1085	ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1086	if (ret < 0)
1087		goto fail_free_reader;
1088
1089	cpu_buffer->head_page
1090		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1091	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1092
1093	rb_head_page_activate(cpu_buffer);
1094
1095	return cpu_buffer;
1096
1097 fail_free_reader:
1098	free_buffer_page(cpu_buffer->reader_page);
1099
1100 fail_free_buffer:
1101	kfree(cpu_buffer);
1102	return NULL;
1103}
1104
1105static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1106{
1107	struct list_head *head = cpu_buffer->pages;
1108	struct buffer_page *bpage, *tmp;
1109
1110	free_buffer_page(cpu_buffer->reader_page);
1111
1112	rb_head_page_deactivate(cpu_buffer);
1113
1114	if (head) {
1115		list_for_each_entry_safe(bpage, tmp, head, list) {
1116			list_del_init(&bpage->list);
1117			free_buffer_page(bpage);
1118		}
1119		bpage = list_entry(head, struct buffer_page, list);
1120		free_buffer_page(bpage);
1121	}
1122
1123	kfree(cpu_buffer);
1124}
1125
1126#ifdef CONFIG_HOTPLUG_CPU
1127static int rb_cpu_notify(struct notifier_block *self,
1128			 unsigned long action, void *hcpu);
1129#endif
1130
1131/**
1132 * ring_buffer_alloc - allocate a new ring_buffer
1133 * @size: the size in bytes per cpu that is needed.
1134 * @flags: attributes to set for the ring buffer.
1135 *
1136 * Currently the only flag that is available is the RB_FL_OVERWRITE
1137 * flag. This flag means that the buffer will overwrite old data
1138 * when the buffer wraps. If this flag is not set, the buffer will
1139 * drop data when the tail hits the head.
1140 */
1141struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1142					struct lock_class_key *key)
1143{
1144	struct ring_buffer *buffer;
 
1145	int bsize;
1146	int cpu;
 
1147
1148	/* keep it in its own cache line */
1149	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1150			 GFP_KERNEL);
1151	if (!buffer)
1152		return NULL;
1153
1154	if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1155		goto fail_free_buffer;
1156
1157	buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1158	buffer->flags = flags;
1159	buffer->clock = trace_clock_local;
1160	buffer->reader_lock_key = key;
1161
 
 
 
1162	/* need at least two pages */
1163	if (buffer->pages < 2)
1164		buffer->pages = 2;
1165
1166	/*
1167	 * In case of non-hotplug cpu, if the ring-buffer is allocated
1168	 * in early initcall, it will not be notified of secondary cpus.
1169	 * In that off case, we need to allocate for all possible cpus.
1170	 */
1171#ifdef CONFIG_HOTPLUG_CPU
1172	get_online_cpus();
1173	cpumask_copy(buffer->cpumask, cpu_online_mask);
1174#else
1175	cpumask_copy(buffer->cpumask, cpu_possible_mask);
1176#endif
1177	buffer->cpus = nr_cpu_ids;
1178
1179	bsize = sizeof(void *) * nr_cpu_ids;
1180	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1181				  GFP_KERNEL);
1182	if (!buffer->buffers)
1183		goto fail_free_cpumask;
1184
1185	for_each_buffer_cpu(buffer, cpu) {
1186		buffer->buffers[cpu] =
1187			rb_allocate_cpu_buffer(buffer, cpu);
1188		if (!buffer->buffers[cpu])
1189			goto fail_free_buffers;
1190	}
1191
1192#ifdef CONFIG_HOTPLUG_CPU
1193	buffer->cpu_notify.notifier_call = rb_cpu_notify;
1194	buffer->cpu_notify.priority = 0;
1195	register_cpu_notifier(&buffer->cpu_notify);
1196#endif
1197
1198	put_online_cpus();
1199	mutex_init(&buffer->mutex);
1200
1201	return buffer;
1202
1203 fail_free_buffers:
1204	for_each_buffer_cpu(buffer, cpu) {
1205		if (buffer->buffers[cpu])
1206			rb_free_cpu_buffer(buffer->buffers[cpu]);
1207	}
1208	kfree(buffer->buffers);
1209
1210 fail_free_cpumask:
1211	free_cpumask_var(buffer->cpumask);
1212	put_online_cpus();
1213
1214 fail_free_buffer:
1215	kfree(buffer);
1216	return NULL;
1217}
1218EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1219
1220/**
1221 * ring_buffer_free - free a ring buffer.
1222 * @buffer: the buffer to free.
1223 */
1224void
1225ring_buffer_free(struct ring_buffer *buffer)
1226{
1227	int cpu;
1228
1229	get_online_cpus();
1230
1231#ifdef CONFIG_HOTPLUG_CPU
1232	unregister_cpu_notifier(&buffer->cpu_notify);
1233#endif
1234
1235	for_each_buffer_cpu(buffer, cpu)
1236		rb_free_cpu_buffer(buffer->buffers[cpu]);
1237
1238	put_online_cpus();
1239
1240	kfree(buffer->buffers);
1241	free_cpumask_var(buffer->cpumask);
1242
1243	kfree(buffer);
1244}
1245EXPORT_SYMBOL_GPL(ring_buffer_free);
1246
1247void ring_buffer_set_clock(struct ring_buffer *buffer,
1248			   u64 (*clock)(void))
1249{
1250	buffer->clock = clock;
1251}
1252
 
 
 
 
 
 
 
 
 
 
1253static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1254
1255static void
1256rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
 
 
 
 
 
 
 
 
 
 
1257{
1258	struct buffer_page *bpage;
1259	struct list_head *p;
1260	unsigned i;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1261
1262	spin_lock_irq(&cpu_buffer->reader_lock);
1263	rb_head_page_deactivate(cpu_buffer);
 
1264
1265	for (i = 0; i < nr_pages; i++) {
1266		if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1267			goto out;
1268		p = cpu_buffer->pages->next;
1269		bpage = list_entry(p, struct buffer_page, list);
1270		list_del_init(&bpage->list);
1271		free_buffer_page(bpage);
1272	}
1273	if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1274		goto out;
1275
1276	rb_reset_cpu(cpu_buffer);
1277	rb_check_pages(cpu_buffer);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1278
1279out:
1280	spin_unlock_irq(&cpu_buffer->reader_lock);
1281}
1282
1283static void
1284rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1285		struct list_head *pages, unsigned nr_pages)
1286{
1287	struct buffer_page *bpage;
1288	struct list_head *p;
1289	unsigned i;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1290
1291	spin_lock_irq(&cpu_buffer->reader_lock);
1292	rb_head_page_deactivate(cpu_buffer);
 
 
 
 
 
 
1293
1294	for (i = 0; i < nr_pages; i++) {
1295		if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1296			goto out;
1297		p = pages->next;
1298		bpage = list_entry(p, struct buffer_page, list);
1299		list_del_init(&bpage->list);
1300		list_add_tail(&bpage->list, cpu_buffer->pages);
 
1301	}
1302	rb_reset_cpu(cpu_buffer);
1303	rb_check_pages(cpu_buffer);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1304
1305out:
1306	spin_unlock_irq(&cpu_buffer->reader_lock);
 
 
 
 
1307}
1308
1309/**
1310 * ring_buffer_resize - resize the ring buffer
1311 * @buffer: the buffer to resize.
1312 * @size: the new size.
 
1313 *
1314 * Minimum size is 2 * BUF_PAGE_SIZE.
1315 *
1316 * Returns -1 on failure.
1317 */
1318int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
 
1319{
1320	struct ring_buffer_per_cpu *cpu_buffer;
1321	unsigned nr_pages, rm_pages, new_pages;
1322	struct buffer_page *bpage, *tmp;
1323	unsigned long buffer_size;
1324	LIST_HEAD(pages);
1325	int i, cpu;
1326
1327	/*
1328	 * Always succeed at resizing a non-existent buffer:
1329	 */
1330	if (!buffer)
1331		return size;
1332
1333	size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1334	size *= BUF_PAGE_SIZE;
1335	buffer_size = buffer->pages * BUF_PAGE_SIZE;
 
 
 
1336
1337	/* we need a minimum of two pages */
1338	if (size < BUF_PAGE_SIZE * 2)
1339		size = BUF_PAGE_SIZE * 2;
1340
1341	if (size == buffer_size)
1342		return size;
1343
1344	atomic_inc(&buffer->record_disabled);
1345
1346	/* Make sure all writers are done with this buffer. */
1347	synchronize_sched();
 
 
 
1348
 
1349	mutex_lock(&buffer->mutex);
1350	get_online_cpus();
1351
1352	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
 
 
 
1353
1354	if (size < buffer_size) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1355
1356		/* easy case, just free pages */
1357		if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1358			goto out_fail;
 
 
 
 
 
 
 
1359
1360		rm_pages = buffer->pages - nr_pages;
 
 
 
 
 
 
 
 
1361
 
1362		for_each_buffer_cpu(buffer, cpu) {
1363			cpu_buffer = buffer->buffers[cpu];
1364			rb_remove_pages(cpu_buffer, rm_pages);
 
 
 
 
 
1365		}
1366		goto out;
1367	}
1368
1369	/*
1370	 * This is a bit more difficult. We only want to add pages
1371	 * when we can allocate enough for all CPUs. We do this
1372	 * by allocating all the pages and storing them on a local
1373	 * link list. If we succeed in our allocation, then we
1374	 * add these pages to the cpu_buffers. Otherwise we just free
1375	 * them all and return -ENOMEM;
1376	 */
1377	if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1378		goto out_fail;
 
 
 
 
 
 
 
 
 
 
 
1379
1380	new_pages = nr_pages - buffer->pages;
1381
1382	for_each_buffer_cpu(buffer, cpu) {
1383		for (i = 0; i < new_pages; i++) {
1384			struct page *page;
1385			/*
1386			 * __GFP_NORETRY flag makes sure that the allocation
1387			 * fails gracefully without invoking oom-killer and
1388			 * the system is not destabilized.
1389			 */
1390			bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1391						  cache_line_size()),
1392					    GFP_KERNEL | __GFP_NORETRY,
1393					    cpu_to_node(cpu));
1394			if (!bpage)
1395				goto free_pages;
1396			list_add(&bpage->list, &pages);
1397			page = alloc_pages_node(cpu_to_node(cpu),
1398						GFP_KERNEL | __GFP_NORETRY, 0);
1399			if (!page)
1400				goto free_pages;
1401			bpage->page = page_address(page);
1402			rb_init_page(bpage->page);
1403		}
 
 
 
1404	}
1405
1406	for_each_buffer_cpu(buffer, cpu) {
1407		cpu_buffer = buffer->buffers[cpu];
1408		rb_insert_pages(cpu_buffer, &pages, new_pages);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1409	}
1410
1411	if (RB_WARN_ON(buffer, !list_empty(&pages)))
1412		goto out_fail;
1413
1414 out:
1415	buffer->pages = nr_pages;
1416	put_online_cpus();
1417	mutex_unlock(&buffer->mutex);
1418
1419	atomic_dec(&buffer->record_disabled);
 
1420
1421	return size;
 
1422
1423 free_pages:
1424	list_for_each_entry_safe(bpage, tmp, &pages, list) {
1425		list_del_init(&bpage->list);
1426		free_buffer_page(bpage);
 
1427	}
1428	put_online_cpus();
1429	mutex_unlock(&buffer->mutex);
1430	atomic_dec(&buffer->record_disabled);
1431	return -ENOMEM;
1432
1433	/*
1434	 * Something went totally wrong, and we are too paranoid
1435	 * to even clean up the mess.
1436	 */
1437 out_fail:
1438	put_online_cpus();
1439	mutex_unlock(&buffer->mutex);
1440	atomic_dec(&buffer->record_disabled);
1441	return -1;
1442}
1443EXPORT_SYMBOL_GPL(ring_buffer_resize);
1444
1445void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1446{
1447	mutex_lock(&buffer->mutex);
1448	if (val)
1449		buffer->flags |= RB_FL_OVERWRITE;
1450	else
1451		buffer->flags &= ~RB_FL_OVERWRITE;
1452	mutex_unlock(&buffer->mutex);
1453}
1454EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1455
1456static inline void *
1457__rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1458{
1459	return bpage->data + index;
1460}
1461
1462static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1463{
1464	return bpage->page->data + index;
1465}
1466
1467static inline struct ring_buffer_event *
1468rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1469{
1470	return __rb_page_index(cpu_buffer->reader_page,
1471			       cpu_buffer->reader_page->read);
1472}
1473
1474static inline struct ring_buffer_event *
1475rb_iter_head_event(struct ring_buffer_iter *iter)
1476{
1477	return __rb_page_index(iter->head_page, iter->head);
1478}
1479
1480static inline unsigned long rb_page_write(struct buffer_page *bpage)
1481{
1482	return local_read(&bpage->write) & RB_WRITE_MASK;
1483}
1484
1485static inline unsigned rb_page_commit(struct buffer_page *bpage)
1486{
1487	return local_read(&bpage->page->commit);
1488}
1489
1490static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1491{
1492	return local_read(&bpage->entries) & RB_WRITE_MASK;
1493}
1494
1495/* Size is determined by what has been committed */
1496static inline unsigned rb_page_size(struct buffer_page *bpage)
1497{
1498	return rb_page_commit(bpage);
1499}
1500
1501static inline unsigned
1502rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1503{
1504	return rb_page_commit(cpu_buffer->commit_page);
1505}
1506
1507static inline unsigned
1508rb_event_index(struct ring_buffer_event *event)
1509{
1510	unsigned long addr = (unsigned long)event;
1511
1512	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1513}
1514
1515static inline int
1516rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1517		   struct ring_buffer_event *event)
1518{
1519	unsigned long addr = (unsigned long)event;
1520	unsigned long index;
1521
1522	index = rb_event_index(event);
1523	addr &= PAGE_MASK;
1524
1525	return cpu_buffer->commit_page->page == (void *)addr &&
1526		rb_commit_index(cpu_buffer) == index;
1527}
1528
1529static void
1530rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1531{
1532	unsigned long max_count;
1533
1534	/*
1535	 * We only race with interrupts and NMIs on this CPU.
1536	 * If we own the commit event, then we can commit
1537	 * all others that interrupted us, since the interruptions
1538	 * are in stack format (they finish before they come
1539	 * back to us). This allows us to do a simple loop to
1540	 * assign the commit to the tail.
1541	 */
1542 again:
1543	max_count = cpu_buffer->buffer->pages * 100;
1544
1545	while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1546		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1547			return;
1548		if (RB_WARN_ON(cpu_buffer,
1549			       rb_is_reader_page(cpu_buffer->tail_page)))
1550			return;
1551		local_set(&cpu_buffer->commit_page->page->commit,
1552			  rb_page_write(cpu_buffer->commit_page));
1553		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1554		cpu_buffer->write_stamp =
1555			cpu_buffer->commit_page->page->time_stamp;
1556		/* add barrier to keep gcc from optimizing too much */
1557		barrier();
1558	}
1559	while (rb_commit_index(cpu_buffer) !=
1560	       rb_page_write(cpu_buffer->commit_page)) {
1561
1562		local_set(&cpu_buffer->commit_page->page->commit,
1563			  rb_page_write(cpu_buffer->commit_page));
1564		RB_WARN_ON(cpu_buffer,
1565			   local_read(&cpu_buffer->commit_page->page->commit) &
1566			   ~RB_WRITE_MASK);
1567		barrier();
1568	}
1569
1570	/* again, keep gcc from optimizing */
1571	barrier();
1572
1573	/*
1574	 * If an interrupt came in just after the first while loop
1575	 * and pushed the tail page forward, we will be left with
1576	 * a dangling commit that will never go forward.
1577	 */
1578	if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1579		goto again;
1580}
1581
1582static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1583{
1584	cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1585	cpu_buffer->reader_page->read = 0;
1586}
1587
1588static void rb_inc_iter(struct ring_buffer_iter *iter)
1589{
1590	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1591
1592	/*
1593	 * The iterator could be on the reader page (it starts there).
1594	 * But the head could have moved, since the reader was
1595	 * found. Check for this case and assign the iterator
1596	 * to the head page instead of next.
1597	 */
1598	if (iter->head_page == cpu_buffer->reader_page)
1599		iter->head_page = rb_set_head_page(cpu_buffer);
1600	else
1601		rb_inc_page(cpu_buffer, &iter->head_page);
1602
1603	iter->read_stamp = iter->head_page->page->time_stamp;
1604	iter->head = 0;
1605}
1606
1607/* Slow path, do not inline */
1608static noinline struct ring_buffer_event *
1609rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1610{
1611	event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1612
1613	/* Not the first event on the page? */
1614	if (rb_event_index(event)) {
1615		event->time_delta = delta & TS_MASK;
1616		event->array[0] = delta >> TS_SHIFT;
1617	} else {
1618		/* nope, just zero it */
1619		event->time_delta = 0;
1620		event->array[0] = 0;
1621	}
1622
1623	return skip_time_extend(event);
1624}
1625
1626/**
1627 * ring_buffer_update_event - update event type and data
1628 * @event: the even to update
1629 * @type: the type of event
1630 * @length: the size of the event field in the ring buffer
1631 *
1632 * Update the type and data fields of the event. The length
1633 * is the actual size that is written to the ring buffer,
1634 * and with this, we can determine what to place into the
1635 * data field.
1636 */
1637static void
1638rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1639		struct ring_buffer_event *event, unsigned length,
1640		int add_timestamp, u64 delta)
1641{
1642	/* Only a commit updates the timestamp */
1643	if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1644		delta = 0;
1645
1646	/*
1647	 * If we need to add a timestamp, then we
1648	 * add it to the start of the resevered space.
1649	 */
1650	if (unlikely(add_timestamp)) {
1651		event = rb_add_time_stamp(event, delta);
1652		length -= RB_LEN_TIME_EXTEND;
1653		delta = 0;
1654	}
1655
1656	event->time_delta = delta;
1657	length -= RB_EVNT_HDR_SIZE;
1658	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1659		event->type_len = 0;
1660		event->array[0] = length;
1661	} else
1662		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1663}
1664
1665/*
1666 * rb_handle_head_page - writer hit the head page
1667 *
1668 * Returns: +1 to retry page
1669 *           0 to continue
1670 *          -1 on error
1671 */
1672static int
1673rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1674		    struct buffer_page *tail_page,
1675		    struct buffer_page *next_page)
1676{
1677	struct buffer_page *new_head;
1678	int entries;
1679	int type;
1680	int ret;
1681
1682	entries = rb_page_entries(next_page);
1683
1684	/*
1685	 * The hard part is here. We need to move the head
1686	 * forward, and protect against both readers on
1687	 * other CPUs and writers coming in via interrupts.
1688	 */
1689	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1690				       RB_PAGE_HEAD);
1691
1692	/*
1693	 * type can be one of four:
1694	 *  NORMAL - an interrupt already moved it for us
1695	 *  HEAD   - we are the first to get here.
1696	 *  UPDATE - we are the interrupt interrupting
1697	 *           a current move.
1698	 *  MOVED  - a reader on another CPU moved the next
1699	 *           pointer to its reader page. Give up
1700	 *           and try again.
1701	 */
1702
1703	switch (type) {
1704	case RB_PAGE_HEAD:
1705		/*
1706		 * We changed the head to UPDATE, thus
1707		 * it is our responsibility to update
1708		 * the counters.
1709		 */
1710		local_add(entries, &cpu_buffer->overrun);
 
1711
1712		/*
1713		 * The entries will be zeroed out when we move the
1714		 * tail page.
1715		 */
1716
1717		/* still more to do */
1718		break;
1719
1720	case RB_PAGE_UPDATE:
1721		/*
1722		 * This is an interrupt that interrupt the
1723		 * previous update. Still more to do.
1724		 */
1725		break;
1726	case RB_PAGE_NORMAL:
1727		/*
1728		 * An interrupt came in before the update
1729		 * and processed this for us.
1730		 * Nothing left to do.
1731		 */
1732		return 1;
1733	case RB_PAGE_MOVED:
1734		/*
1735		 * The reader is on another CPU and just did
1736		 * a swap with our next_page.
1737		 * Try again.
1738		 */
1739		return 1;
1740	default:
1741		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1742		return -1;
1743	}
1744
1745	/*
1746	 * Now that we are here, the old head pointer is
1747	 * set to UPDATE. This will keep the reader from
1748	 * swapping the head page with the reader page.
1749	 * The reader (on another CPU) will spin till
1750	 * we are finished.
1751	 *
1752	 * We just need to protect against interrupts
1753	 * doing the job. We will set the next pointer
1754	 * to HEAD. After that, we set the old pointer
1755	 * to NORMAL, but only if it was HEAD before.
1756	 * otherwise we are an interrupt, and only
1757	 * want the outer most commit to reset it.
1758	 */
1759	new_head = next_page;
1760	rb_inc_page(cpu_buffer, &new_head);
1761
1762	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1763				    RB_PAGE_NORMAL);
1764
1765	/*
1766	 * Valid returns are:
1767	 *  HEAD   - an interrupt came in and already set it.
1768	 *  NORMAL - One of two things:
1769	 *            1) We really set it.
1770	 *            2) A bunch of interrupts came in and moved
1771	 *               the page forward again.
1772	 */
1773	switch (ret) {
1774	case RB_PAGE_HEAD:
1775	case RB_PAGE_NORMAL:
1776		/* OK */
1777		break;
1778	default:
1779		RB_WARN_ON(cpu_buffer, 1);
1780		return -1;
1781	}
1782
1783	/*
1784	 * It is possible that an interrupt came in,
1785	 * set the head up, then more interrupts came in
1786	 * and moved it again. When we get back here,
1787	 * the page would have been set to NORMAL but we
1788	 * just set it back to HEAD.
1789	 *
1790	 * How do you detect this? Well, if that happened
1791	 * the tail page would have moved.
1792	 */
1793	if (ret == RB_PAGE_NORMAL) {
 
 
 
1794		/*
1795		 * If the tail had moved passed next, then we need
1796		 * to reset the pointer.
1797		 */
1798		if (cpu_buffer->tail_page != tail_page &&
1799		    cpu_buffer->tail_page != next_page)
1800			rb_head_page_set_normal(cpu_buffer, new_head,
1801						next_page,
1802						RB_PAGE_HEAD);
1803	}
1804
1805	/*
1806	 * If this was the outer most commit (the one that
1807	 * changed the original pointer from HEAD to UPDATE),
1808	 * then it is up to us to reset it to NORMAL.
1809	 */
1810	if (type == RB_PAGE_HEAD) {
1811		ret = rb_head_page_set_normal(cpu_buffer, next_page,
1812					      tail_page,
1813					      RB_PAGE_UPDATE);
1814		if (RB_WARN_ON(cpu_buffer,
1815			       ret != RB_PAGE_UPDATE))
1816			return -1;
1817	}
1818
1819	return 0;
1820}
1821
1822static unsigned rb_calculate_event_length(unsigned length)
1823{
1824	struct ring_buffer_event event; /* Used only for sizeof array */
1825
1826	/* zero length can cause confusions */
1827	if (!length)
1828		length = 1;
1829
1830	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
1831		length += sizeof(event.array[0]);
1832
1833	length += RB_EVNT_HDR_SIZE;
1834	length = ALIGN(length, RB_ARCH_ALIGNMENT);
1835
1836	return length;
1837}
1838
1839static inline void
1840rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1841	      struct buffer_page *tail_page,
1842	      unsigned long tail, unsigned long length)
1843{
 
1844	struct ring_buffer_event *event;
 
1845
1846	/*
1847	 * Only the event that crossed the page boundary
1848	 * must fill the old tail_page with padding.
1849	 */
1850	if (tail >= BUF_PAGE_SIZE) {
1851		/*
1852		 * If the page was filled, then we still need
1853		 * to update the real_end. Reset it to zero
1854		 * and the reader will ignore it.
1855		 */
1856		if (tail == BUF_PAGE_SIZE)
1857			tail_page->real_end = 0;
1858
1859		local_sub(length, &tail_page->write);
1860		return;
1861	}
1862
1863	event = __rb_page_index(tail_page, tail);
1864	kmemcheck_annotate_bitfield(event, bitfield);
 
 
1865
1866	/*
1867	 * Save the original length to the meta data.
1868	 * This will be used by the reader to add lost event
1869	 * counter.
1870	 */
1871	tail_page->real_end = tail;
1872
1873	/*
1874	 * If this event is bigger than the minimum size, then
1875	 * we need to be careful that we don't subtract the
1876	 * write counter enough to allow another writer to slip
1877	 * in on this page.
1878	 * We put in a discarded commit instead, to make sure
1879	 * that this space is not used again.
1880	 *
1881	 * If we are less than the minimum size, we don't need to
1882	 * worry about it.
1883	 */
1884	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1885		/* No room for any events */
1886
1887		/* Mark the rest of the page with padding */
1888		rb_event_set_padding(event);
1889
1890		/* Set the write back to the previous setting */
1891		local_sub(length, &tail_page->write);
1892		return;
1893	}
1894
1895	/* Put in a discarded event */
1896	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1897	event->type_len = RINGBUF_TYPE_PADDING;
1898	/* time delta must be non zero */
1899	event->time_delta = 1;
1900
1901	/* Set write to end of buffer */
1902	length = (tail + length) - BUF_PAGE_SIZE;
1903	local_sub(length, &tail_page->write);
1904}
1905
 
 
1906/*
1907 * This is the slow path, force gcc not to inline it.
1908 */
1909static noinline struct ring_buffer_event *
1910rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1911	     unsigned long length, unsigned long tail,
1912	     struct buffer_page *tail_page, u64 ts)
1913{
 
1914	struct buffer_page *commit_page = cpu_buffer->commit_page;
1915	struct ring_buffer *buffer = cpu_buffer->buffer;
1916	struct buffer_page *next_page;
1917	int ret;
1918
1919	next_page = tail_page;
1920
1921	rb_inc_page(cpu_buffer, &next_page);
1922
1923	/*
1924	 * If for some reason, we had an interrupt storm that made
1925	 * it all the way around the buffer, bail, and warn
1926	 * about it.
1927	 */
1928	if (unlikely(next_page == commit_page)) {
1929		local_inc(&cpu_buffer->commit_overrun);
1930		goto out_reset;
1931	}
1932
1933	/*
1934	 * This is where the fun begins!
1935	 *
1936	 * We are fighting against races between a reader that
1937	 * could be on another CPU trying to swap its reader
1938	 * page with the buffer head.
1939	 *
1940	 * We are also fighting against interrupts coming in and
1941	 * moving the head or tail on us as well.
1942	 *
1943	 * If the next page is the head page then we have filled
1944	 * the buffer, unless the commit page is still on the
1945	 * reader page.
1946	 */
1947	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1948
1949		/*
1950		 * If the commit is not on the reader page, then
1951		 * move the header page.
1952		 */
1953		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1954			/*
1955			 * If we are not in overwrite mode,
1956			 * this is easy, just stop here.
1957			 */
1958			if (!(buffer->flags & RB_FL_OVERWRITE))
 
1959				goto out_reset;
 
1960
1961			ret = rb_handle_head_page(cpu_buffer,
1962						  tail_page,
1963						  next_page);
1964			if (ret < 0)
1965				goto out_reset;
1966			if (ret)
1967				goto out_again;
1968		} else {
1969			/*
1970			 * We need to be careful here too. The
1971			 * commit page could still be on the reader
1972			 * page. We could have a small buffer, and
1973			 * have filled up the buffer with events
1974			 * from interrupts and such, and wrapped.
1975			 *
1976			 * Note, if the tail page is also the on the
1977			 * reader_page, we let it move out.
1978			 */
1979			if (unlikely((cpu_buffer->commit_page !=
1980				      cpu_buffer->tail_page) &&
1981				     (cpu_buffer->commit_page ==
1982				      cpu_buffer->reader_page))) {
1983				local_inc(&cpu_buffer->commit_overrun);
1984				goto out_reset;
1985			}
1986		}
1987	}
1988
1989	ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1990	if (ret) {
1991		/*
1992		 * Nested commits always have zero deltas, so
1993		 * just reread the time stamp
1994		 */
1995		ts = rb_time_stamp(buffer);
1996		next_page->page->time_stamp = ts;
1997	}
1998
1999 out_again:
2000
2001	rb_reset_tail(cpu_buffer, tail_page, tail, length);
 
 
 
 
 
2002
2003	/* fail and let the caller try again */
2004	return ERR_PTR(-EAGAIN);
2005
2006 out_reset:
2007	/* reset write */
2008	rb_reset_tail(cpu_buffer, tail_page, tail, length);
2009
2010	return NULL;
2011}
2012
2013static struct ring_buffer_event *
2014__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2015		  unsigned long length, u64 ts,
2016		  u64 delta, int add_timestamp)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2017{
2018	struct buffer_page *tail_page;
2019	struct ring_buffer_event *event;
2020	unsigned long tail, write;
 
 
 
2021
2022	/*
2023	 * If the time delta since the last event is too big to
2024	 * hold in the time field of the event, then we append a
2025	 * TIME EXTEND event ahead of the data event.
2026	 */
2027	if (unlikely(add_timestamp))
2028		length += RB_LEN_TIME_EXTEND;
2029
2030	tail_page = cpu_buffer->tail_page;
2031	write = local_add_return(length, &tail_page->write);
 
 
2032
2033	/* set write to only the index of the write */
2034	write &= RB_WRITE_MASK;
2035	tail = write - length;
 
 
 
 
 
2036
2037	/* See if we shot pass the end of this buffer page */
2038	if (unlikely(write > BUF_PAGE_SIZE))
2039		return rb_move_tail(cpu_buffer, length, tail,
2040				    tail_page, ts);
2041
2042	/* We reserved something on the buffer */
 
 
2043
2044	event = __rb_page_index(tail_page, tail);
2045	kmemcheck_annotate_bitfield(event, bitfield);
2046	rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2047
2048	local_inc(&tail_page->entries);
 
2049
2050	/*
2051	 * If this is the first commit on the page, then update
2052	 * its timestamp.
 
 
 
 
 
 
 
 
2053	 */
2054	if (!tail)
2055		tail_page->page->time_stamp = ts;
2056
2057	return event;
2058}
2059
 
 
 
 
 
 
 
2060static inline int
2061rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2062		  struct ring_buffer_event *event)
2063{
2064	unsigned long new_index, old_index;
2065	struct buffer_page *bpage;
2066	unsigned long index;
2067	unsigned long addr;
2068
2069	new_index = rb_event_index(event);
2070	old_index = new_index + rb_event_ts_length(event);
2071	addr = (unsigned long)event;
2072	addr &= PAGE_MASK;
2073
2074	bpage = cpu_buffer->tail_page;
2075
2076	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2077		unsigned long write_mask =
2078			local_read(&bpage->write) & ~RB_WRITE_MASK;
 
2079		/*
2080		 * This is on the tail page. It is possible that
2081		 * a write could come in and move the tail page
2082		 * and write to the next page. That is fine
2083		 * because we just shorten what is on this page.
2084		 */
2085		old_index += write_mask;
2086		new_index += write_mask;
2087		index = local_cmpxchg(&bpage->write, old_index, new_index);
2088		if (index == old_index)
 
 
2089			return 1;
 
2090	}
2091
2092	/* could not discard */
2093	return 0;
2094}
2095
2096static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2097{
2098	local_inc(&cpu_buffer->committing);
2099	local_inc(&cpu_buffer->commits);
2100}
2101
2102static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2103{
2104	unsigned long commits;
2105
2106	if (RB_WARN_ON(cpu_buffer,
2107		       !local_read(&cpu_buffer->committing)))
2108		return;
2109
2110 again:
2111	commits = local_read(&cpu_buffer->commits);
2112	/* synchronize with interrupts */
2113	barrier();
2114	if (local_read(&cpu_buffer->committing) == 1)
2115		rb_set_commit_to_write(cpu_buffer);
2116
2117	local_dec(&cpu_buffer->committing);
2118
2119	/* synchronize with interrupts */
2120	barrier();
2121
2122	/*
2123	 * Need to account for interrupts coming in between the
2124	 * updating of the commit page and the clearing of the
2125	 * committing counter.
2126	 */
2127	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2128	    !local_read(&cpu_buffer->committing)) {
2129		local_inc(&cpu_buffer->committing);
2130		goto again;
2131	}
2132}
2133
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2134static struct ring_buffer_event *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2135rb_reserve_next_event(struct ring_buffer *buffer,
2136		      struct ring_buffer_per_cpu *cpu_buffer,
2137		      unsigned long length)
2138{
2139	struct ring_buffer_event *event;
2140	u64 ts, delta;
2141	int nr_loops = 0;
2142	int add_timestamp;
2143	u64 diff;
2144
2145	rb_start_commit(cpu_buffer);
2146
2147#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2148	/*
2149	 * Due to the ability to swap a cpu buffer from a buffer
2150	 * it is possible it was swapped before we committed.
2151	 * (committing stops a swap). We check for it here and
2152	 * if it happened, we have to fail the write.
2153	 */
2154	barrier();
2155	if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2156		local_dec(&cpu_buffer->committing);
2157		local_dec(&cpu_buffer->commits);
2158		return NULL;
2159	}
2160#endif
2161
2162	length = rb_calculate_event_length(length);
2163 again:
2164	add_timestamp = 0;
2165	delta = 0;
2166
2167	/*
2168	 * We allow for interrupts to reenter here and do a trace.
2169	 * If one does, it will cause this original code to loop
2170	 * back here. Even with heavy interrupts happening, this
2171	 * should only happen a few times in a row. If this happens
2172	 * 1000 times in a row, there must be either an interrupt
2173	 * storm or we have something buggy.
2174	 * Bail!
2175	 */
2176	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2177		goto out_fail;
2178
2179	ts = rb_time_stamp(cpu_buffer->buffer);
2180	diff = ts - cpu_buffer->write_stamp;
2181
2182	/* make sure this diff is calculated here */
2183	barrier();
2184
2185	/* Did the write stamp get updated already? */
2186	if (likely(ts >= cpu_buffer->write_stamp)) {
2187		delta = diff;
2188		if (unlikely(test_time_stamp(delta))) {
2189			int local_clock_stable = 1;
2190#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2191			local_clock_stable = sched_clock_stable;
2192#endif
2193			WARN_ONCE(delta > (1ULL << 59),
2194				  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2195				  (unsigned long long)delta,
2196				  (unsigned long long)ts,
2197				  (unsigned long long)cpu_buffer->write_stamp,
2198				  local_clock_stable ? "" :
2199				  "If you just came from a suspend/resume,\n"
2200				  "please switch to the trace global clock:\n"
2201				  "  echo global > /sys/kernel/debug/tracing/trace_clock\n");
2202			add_timestamp = 1;
2203		}
2204	}
2205
2206	event = __rb_reserve_next(cpu_buffer, length, ts,
2207				  delta, add_timestamp);
2208	if (unlikely(PTR_ERR(event) == -EAGAIN))
 
 
2209		goto again;
 
2210
2211	if (!event)
2212		goto out_fail;
2213
2214	return event;
2215
2216 out_fail:
2217	rb_end_commit(cpu_buffer);
2218	return NULL;
2219}
2220
2221#ifdef CONFIG_TRACING
2222
2223#define TRACE_RECURSIVE_DEPTH 16
2224
2225/* Keep this code out of the fast path cache */
2226static noinline void trace_recursive_fail(void)
2227{
2228	/* Disable all tracing before we do anything else */
2229	tracing_off_permanent();
2230
2231	printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2232		    "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2233		    trace_recursion_buffer(),
2234		    hardirq_count() >> HARDIRQ_SHIFT,
2235		    softirq_count() >> SOFTIRQ_SHIFT,
2236		    in_nmi());
2237
2238	WARN_ON_ONCE(1);
2239}
2240
2241static inline int trace_recursive_lock(void)
2242{
2243	trace_recursion_inc();
2244
2245	if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH))
2246		return 0;
2247
2248	trace_recursive_fail();
2249
2250	return -1;
2251}
2252
2253static inline void trace_recursive_unlock(void)
2254{
2255	WARN_ON_ONCE(!trace_recursion_buffer());
2256
2257	trace_recursion_dec();
2258}
2259
2260#else
2261
2262#define trace_recursive_lock()		(0)
2263#define trace_recursive_unlock()	do { } while (0)
2264
2265#endif
2266
2267/**
2268 * ring_buffer_lock_reserve - reserve a part of the buffer
2269 * @buffer: the ring buffer to reserve from
2270 * @length: the length of the data to reserve (excluding event header)
2271 *
2272 * Returns a reseverd event on the ring buffer to copy directly to.
2273 * The user of this interface will need to get the body to write into
2274 * and can use the ring_buffer_event_data() interface.
2275 *
2276 * The length is the length of the data needed, not the event length
2277 * which also includes the event header.
2278 *
2279 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2280 * If NULL is returned, then nothing has been allocated or locked.
2281 */
2282struct ring_buffer_event *
2283ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2284{
2285	struct ring_buffer_per_cpu *cpu_buffer;
2286	struct ring_buffer_event *event;
2287	int cpu;
2288
2289	if (ring_buffer_flags != RB_BUFFERS_ON)
2290		return NULL;
2291
2292	/* If we are tracing schedule, we don't want to recurse */
2293	preempt_disable_notrace();
2294
2295	if (atomic_read(&buffer->record_disabled))
2296		goto out_nocheck;
2297
2298	if (trace_recursive_lock())
2299		goto out_nocheck;
2300
2301	cpu = raw_smp_processor_id();
2302
2303	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2304		goto out;
2305
2306	cpu_buffer = buffer->buffers[cpu];
2307
2308	if (atomic_read(&cpu_buffer->record_disabled))
 
 
 
2309		goto out;
2310
2311	if (length > BUF_MAX_DATA_SIZE)
2312		goto out;
2313
2314	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2315	if (!event)
2316		goto out;
2317
2318	return event;
2319
 
 
2320 out:
2321	trace_recursive_unlock();
2322
2323 out_nocheck:
2324	preempt_enable_notrace();
2325	return NULL;
2326}
2327EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2328
2329static void
2330rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2331		      struct ring_buffer_event *event)
2332{
2333	u64 delta;
2334
2335	/*
2336	 * The event first in the commit queue updates the
2337	 * time stamp.
2338	 */
2339	if (rb_event_is_commit(cpu_buffer, event)) {
2340		/*
2341		 * A commit event that is first on a page
2342		 * updates the write timestamp with the page stamp
2343		 */
2344		if (!rb_event_index(event))
2345			cpu_buffer->write_stamp =
2346				cpu_buffer->commit_page->page->time_stamp;
2347		else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2348			delta = event->array[0];
2349			delta <<= TS_SHIFT;
2350			delta += event->time_delta;
2351			cpu_buffer->write_stamp += delta;
2352		} else
2353			cpu_buffer->write_stamp += event->time_delta;
2354	}
2355}
2356
2357static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2358		      struct ring_buffer_event *event)
2359{
2360	local_inc(&cpu_buffer->entries);
2361	rb_update_write_stamp(cpu_buffer, event);
2362	rb_end_commit(cpu_buffer);
2363}
2364
2365/**
2366 * ring_buffer_unlock_commit - commit a reserved
2367 * @buffer: The buffer to commit to
2368 * @event: The event pointer to commit.
2369 *
2370 * This commits the data to the ring buffer, and releases any locks held.
2371 *
2372 * Must be paired with ring_buffer_lock_reserve.
2373 */
2374int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2375			      struct ring_buffer_event *event)
2376{
2377	struct ring_buffer_per_cpu *cpu_buffer;
2378	int cpu = raw_smp_processor_id();
2379
2380	cpu_buffer = buffer->buffers[cpu];
2381
2382	rb_commit(cpu_buffer, event);
2383
2384	trace_recursive_unlock();
2385
2386	preempt_enable_notrace();
2387
2388	return 0;
2389}
2390EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2391
2392static inline void rb_event_discard(struct ring_buffer_event *event)
2393{
2394	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2395		event = skip_time_extend(event);
2396
2397	/* array[0] holds the actual length for the discarded event */
2398	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2399	event->type_len = RINGBUF_TYPE_PADDING;
2400	/* time delta must be non zero */
2401	if (!event->time_delta)
2402		event->time_delta = 1;
2403}
2404
2405/*
2406 * Decrement the entries to the page that an event is on.
2407 * The event does not even need to exist, only the pointer
2408 * to the page it is on. This may only be called before the commit
2409 * takes place.
2410 */
2411static inline void
2412rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2413		   struct ring_buffer_event *event)
2414{
2415	unsigned long addr = (unsigned long)event;
2416	struct buffer_page *bpage = cpu_buffer->commit_page;
2417	struct buffer_page *start;
2418
2419	addr &= PAGE_MASK;
2420
2421	/* Do the likely case first */
2422	if (likely(bpage->page == (void *)addr)) {
2423		local_dec(&bpage->entries);
2424		return;
2425	}
2426
2427	/*
2428	 * Because the commit page may be on the reader page we
2429	 * start with the next page and check the end loop there.
2430	 */
2431	rb_inc_page(cpu_buffer, &bpage);
2432	start = bpage;
2433	do {
2434		if (bpage->page == (void *)addr) {
2435			local_dec(&bpage->entries);
2436			return;
2437		}
2438		rb_inc_page(cpu_buffer, &bpage);
2439	} while (bpage != start);
2440
2441	/* commit not part of this buffer?? */
2442	RB_WARN_ON(cpu_buffer, 1);
2443}
2444
2445/**
2446 * ring_buffer_commit_discard - discard an event that has not been committed
2447 * @buffer: the ring buffer
2448 * @event: non committed event to discard
2449 *
2450 * Sometimes an event that is in the ring buffer needs to be ignored.
2451 * This function lets the user discard an event in the ring buffer
2452 * and then that event will not be read later.
2453 *
2454 * This function only works if it is called before the the item has been
2455 * committed. It will try to free the event from the ring buffer
2456 * if another event has not been added behind it.
2457 *
2458 * If another event has been added behind it, it will set the event
2459 * up as discarded, and perform the commit.
2460 *
2461 * If this function is called, do not call ring_buffer_unlock_commit on
2462 * the event.
2463 */
2464void ring_buffer_discard_commit(struct ring_buffer *buffer,
2465				struct ring_buffer_event *event)
2466{
2467	struct ring_buffer_per_cpu *cpu_buffer;
2468	int cpu;
2469
2470	/* The event is discarded regardless */
2471	rb_event_discard(event);
2472
2473	cpu = smp_processor_id();
2474	cpu_buffer = buffer->buffers[cpu];
2475
2476	/*
2477	 * This must only be called if the event has not been
2478	 * committed yet. Thus we can assume that preemption
2479	 * is still disabled.
2480	 */
2481	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2482
2483	rb_decrement_entry(cpu_buffer, event);
2484	if (rb_try_to_discard(cpu_buffer, event))
2485		goto out;
2486
2487	/*
2488	 * The commit is still visible by the reader, so we
2489	 * must still update the timestamp.
2490	 */
2491	rb_update_write_stamp(cpu_buffer, event);
2492 out:
2493	rb_end_commit(cpu_buffer);
2494
2495	trace_recursive_unlock();
2496
2497	preempt_enable_notrace();
2498
2499}
2500EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2501
2502/**
2503 * ring_buffer_write - write data to the buffer without reserving
2504 * @buffer: The ring buffer to write to.
2505 * @length: The length of the data being written (excluding the event header)
2506 * @data: The data to write to the buffer.
2507 *
2508 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2509 * one function. If you already have the data to write to the buffer, it
2510 * may be easier to simply call this function.
2511 *
2512 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2513 * and not the length of the event which would hold the header.
2514 */
2515int ring_buffer_write(struct ring_buffer *buffer,
2516			unsigned long length,
2517			void *data)
2518{
2519	struct ring_buffer_per_cpu *cpu_buffer;
2520	struct ring_buffer_event *event;
2521	void *body;
2522	int ret = -EBUSY;
2523	int cpu;
2524
2525	if (ring_buffer_flags != RB_BUFFERS_ON)
2526		return -EBUSY;
2527
2528	preempt_disable_notrace();
2529
2530	if (atomic_read(&buffer->record_disabled))
2531		goto out;
2532
2533	cpu = raw_smp_processor_id();
2534
2535	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2536		goto out;
2537
2538	cpu_buffer = buffer->buffers[cpu];
2539
2540	if (atomic_read(&cpu_buffer->record_disabled))
2541		goto out;
2542
2543	if (length > BUF_MAX_DATA_SIZE)
2544		goto out;
2545
 
 
 
2546	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2547	if (!event)
2548		goto out;
2549
2550	body = rb_event_data(event);
2551
2552	memcpy(body, data, length);
2553
2554	rb_commit(cpu_buffer, event);
2555
 
 
2556	ret = 0;
 
 
 
 
2557 out:
2558	preempt_enable_notrace();
2559
2560	return ret;
2561}
2562EXPORT_SYMBOL_GPL(ring_buffer_write);
2563
2564static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2565{
2566	struct buffer_page *reader = cpu_buffer->reader_page;
2567	struct buffer_page *head = rb_set_head_page(cpu_buffer);
2568	struct buffer_page *commit = cpu_buffer->commit_page;
2569
2570	/* In case of error, head will be NULL */
2571	if (unlikely(!head))
2572		return 1;
2573
2574	return reader->read == rb_page_commit(reader) &&
2575		(commit == reader ||
2576		 (commit == head &&
2577		  head->read == rb_page_commit(commit)));
2578}
2579
2580/**
2581 * ring_buffer_record_disable - stop all writes into the buffer
2582 * @buffer: The ring buffer to stop writes to.
2583 *
2584 * This prevents all writes to the buffer. Any attempt to write
2585 * to the buffer after this will fail and return NULL.
2586 *
2587 * The caller should call synchronize_sched() after this.
2588 */
2589void ring_buffer_record_disable(struct ring_buffer *buffer)
2590{
2591	atomic_inc(&buffer->record_disabled);
2592}
2593EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2594
2595/**
2596 * ring_buffer_record_enable - enable writes to the buffer
2597 * @buffer: The ring buffer to enable writes
2598 *
2599 * Note, multiple disables will need the same number of enables
2600 * to truly enable the writing (much like preempt_disable).
2601 */
2602void ring_buffer_record_enable(struct ring_buffer *buffer)
2603{
2604	atomic_dec(&buffer->record_disabled);
2605}
2606EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2607
2608/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2609 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2610 * @buffer: The ring buffer to stop writes to.
2611 * @cpu: The CPU buffer to stop
2612 *
2613 * This prevents all writes to the buffer. Any attempt to write
2614 * to the buffer after this will fail and return NULL.
2615 *
2616 * The caller should call synchronize_sched() after this.
2617 */
2618void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2619{
2620	struct ring_buffer_per_cpu *cpu_buffer;
2621
2622	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2623		return;
2624
2625	cpu_buffer = buffer->buffers[cpu];
2626	atomic_inc(&cpu_buffer->record_disabled);
2627}
2628EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2629
2630/**
2631 * ring_buffer_record_enable_cpu - enable writes to the buffer
2632 * @buffer: The ring buffer to enable writes
2633 * @cpu: The CPU to enable.
2634 *
2635 * Note, multiple disables will need the same number of enables
2636 * to truly enable the writing (much like preempt_disable).
2637 */
2638void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2639{
2640	struct ring_buffer_per_cpu *cpu_buffer;
2641
2642	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2643		return;
2644
2645	cpu_buffer = buffer->buffers[cpu];
2646	atomic_dec(&cpu_buffer->record_disabled);
2647}
2648EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2649
2650/*
2651 * The total entries in the ring buffer is the running counter
2652 * of entries entered into the ring buffer, minus the sum of
2653 * the entries read from the ring buffer and the number of
2654 * entries that were overwritten.
2655 */
2656static inline unsigned long
2657rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2658{
2659	return local_read(&cpu_buffer->entries) -
2660		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2661}
2662
2663/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2664 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2665 * @buffer: The ring buffer
2666 * @cpu: The per CPU buffer to get the entries from.
2667 */
2668unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2669{
2670	struct ring_buffer_per_cpu *cpu_buffer;
2671
2672	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2673		return 0;
2674
2675	cpu_buffer = buffer->buffers[cpu];
2676
2677	return rb_num_of_entries(cpu_buffer);
2678}
2679EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2680
2681/**
2682 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
 
2683 * @buffer: The ring buffer
2684 * @cpu: The per CPU buffer to get the number of overruns from
2685 */
2686unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2687{
2688	struct ring_buffer_per_cpu *cpu_buffer;
2689	unsigned long ret;
2690
2691	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2692		return 0;
2693
2694	cpu_buffer = buffer->buffers[cpu];
2695	ret = local_read(&cpu_buffer->overrun);
2696
2697	return ret;
2698}
2699EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2700
2701/**
2702 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
 
 
2703 * @buffer: The ring buffer
2704 * @cpu: The per CPU buffer to get the number of overruns from
2705 */
2706unsigned long
2707ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2708{
2709	struct ring_buffer_per_cpu *cpu_buffer;
2710	unsigned long ret;
2711
2712	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2713		return 0;
2714
2715	cpu_buffer = buffer->buffers[cpu];
2716	ret = local_read(&cpu_buffer->commit_overrun);
2717
2718	return ret;
2719}
2720EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2721
2722/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2723 * ring_buffer_entries - get the number of entries in a buffer
2724 * @buffer: The ring buffer
2725 *
2726 * Returns the total number of entries in the ring buffer
2727 * (all CPU entries)
2728 */
2729unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2730{
2731	struct ring_buffer_per_cpu *cpu_buffer;
2732	unsigned long entries = 0;
2733	int cpu;
2734
2735	/* if you care about this being correct, lock the buffer */
2736	for_each_buffer_cpu(buffer, cpu) {
2737		cpu_buffer = buffer->buffers[cpu];
2738		entries += rb_num_of_entries(cpu_buffer);
2739	}
2740
2741	return entries;
2742}
2743EXPORT_SYMBOL_GPL(ring_buffer_entries);
2744
2745/**
2746 * ring_buffer_overruns - get the number of overruns in buffer
2747 * @buffer: The ring buffer
2748 *
2749 * Returns the total number of overruns in the ring buffer
2750 * (all CPU entries)
2751 */
2752unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2753{
2754	struct ring_buffer_per_cpu *cpu_buffer;
2755	unsigned long overruns = 0;
2756	int cpu;
2757
2758	/* if you care about this being correct, lock the buffer */
2759	for_each_buffer_cpu(buffer, cpu) {
2760		cpu_buffer = buffer->buffers[cpu];
2761		overruns += local_read(&cpu_buffer->overrun);
2762	}
2763
2764	return overruns;
2765}
2766EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2767
2768static void rb_iter_reset(struct ring_buffer_iter *iter)
2769{
2770	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2771
2772	/* Iterator usage is expected to have record disabled */
2773	if (list_empty(&cpu_buffer->reader_page->list)) {
2774		iter->head_page = rb_set_head_page(cpu_buffer);
2775		if (unlikely(!iter->head_page))
2776			return;
2777		iter->head = iter->head_page->read;
2778	} else {
2779		iter->head_page = cpu_buffer->reader_page;
2780		iter->head = cpu_buffer->reader_page->read;
2781	}
2782	if (iter->head)
2783		iter->read_stamp = cpu_buffer->read_stamp;
2784	else
2785		iter->read_stamp = iter->head_page->page->time_stamp;
2786	iter->cache_reader_page = cpu_buffer->reader_page;
2787	iter->cache_read = cpu_buffer->read;
2788}
2789
2790/**
2791 * ring_buffer_iter_reset - reset an iterator
2792 * @iter: The iterator to reset
2793 *
2794 * Resets the iterator, so that it will start from the beginning
2795 * again.
2796 */
2797void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2798{
2799	struct ring_buffer_per_cpu *cpu_buffer;
2800	unsigned long flags;
2801
2802	if (!iter)
2803		return;
2804
2805	cpu_buffer = iter->cpu_buffer;
2806
2807	spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2808	rb_iter_reset(iter);
2809	spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2810}
2811EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2812
2813/**
2814 * ring_buffer_iter_empty - check if an iterator has no more to read
2815 * @iter: The iterator to check
2816 */
2817int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2818{
2819	struct ring_buffer_per_cpu *cpu_buffer;
 
 
 
 
2820
2821	cpu_buffer = iter->cpu_buffer;
2822
2823	return iter->head_page == cpu_buffer->commit_page &&
2824		iter->head == rb_commit_index(cpu_buffer);
 
 
 
 
 
 
 
 
2825}
2826EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2827
2828static void
2829rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2830		     struct ring_buffer_event *event)
2831{
2832	u64 delta;
2833
2834	switch (event->type_len) {
2835	case RINGBUF_TYPE_PADDING:
2836		return;
2837
2838	case RINGBUF_TYPE_TIME_EXTEND:
2839		delta = event->array[0];
2840		delta <<= TS_SHIFT;
2841		delta += event->time_delta;
2842		cpu_buffer->read_stamp += delta;
2843		return;
2844
2845	case RINGBUF_TYPE_TIME_STAMP:
2846		/* FIXME: not implemented */
 
2847		return;
2848
2849	case RINGBUF_TYPE_DATA:
2850		cpu_buffer->read_stamp += event->time_delta;
2851		return;
2852
2853	default:
2854		BUG();
2855	}
2856	return;
2857}
2858
2859static void
2860rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2861			  struct ring_buffer_event *event)
2862{
2863	u64 delta;
2864
2865	switch (event->type_len) {
2866	case RINGBUF_TYPE_PADDING:
2867		return;
2868
2869	case RINGBUF_TYPE_TIME_EXTEND:
2870		delta = event->array[0];
2871		delta <<= TS_SHIFT;
2872		delta += event->time_delta;
2873		iter->read_stamp += delta;
2874		return;
2875
2876	case RINGBUF_TYPE_TIME_STAMP:
2877		/* FIXME: not implemented */
 
2878		return;
2879
2880	case RINGBUF_TYPE_DATA:
2881		iter->read_stamp += event->time_delta;
2882		return;
2883
2884	default:
2885		BUG();
2886	}
2887	return;
2888}
2889
2890static struct buffer_page *
2891rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2892{
2893	struct buffer_page *reader = NULL;
2894	unsigned long overwrite;
2895	unsigned long flags;
2896	int nr_loops = 0;
2897	int ret;
2898
2899	local_irq_save(flags);
2900	arch_spin_lock(&cpu_buffer->lock);
2901
2902 again:
2903	/*
2904	 * This should normally only loop twice. But because the
2905	 * start of the reader inserts an empty page, it causes
2906	 * a case where we will loop three times. There should be no
2907	 * reason to loop four times (that I know of).
2908	 */
2909	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2910		reader = NULL;
2911		goto out;
2912	}
2913
2914	reader = cpu_buffer->reader_page;
2915
2916	/* If there's more to read, return this page */
2917	if (cpu_buffer->reader_page->read < rb_page_size(reader))
2918		goto out;
2919
2920	/* Never should we have an index greater than the size */
2921	if (RB_WARN_ON(cpu_buffer,
2922		       cpu_buffer->reader_page->read > rb_page_size(reader)))
2923		goto out;
2924
2925	/* check if we caught up to the tail */
2926	reader = NULL;
2927	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2928		goto out;
2929
 
 
 
 
2930	/*
2931	 * Reset the reader page to size zero.
2932	 */
2933	local_set(&cpu_buffer->reader_page->write, 0);
2934	local_set(&cpu_buffer->reader_page->entries, 0);
2935	local_set(&cpu_buffer->reader_page->page->commit, 0);
2936	cpu_buffer->reader_page->real_end = 0;
2937
2938 spin:
2939	/*
2940	 * Splice the empty reader page into the list around the head.
2941	 */
2942	reader = rb_set_head_page(cpu_buffer);
 
 
2943	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
2944	cpu_buffer->reader_page->list.prev = reader->list.prev;
2945
2946	/*
2947	 * cpu_buffer->pages just needs to point to the buffer, it
2948	 *  has no specific buffer page to point to. Lets move it out
2949	 *  of our way so we don't accidentally swap it.
2950	 */
2951	cpu_buffer->pages = reader->list.prev;
2952
2953	/* The reader page will be pointing to the new head */
2954	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2955
2956	/*
2957	 * We want to make sure we read the overruns after we set up our
2958	 * pointers to the next object. The writer side does a
2959	 * cmpxchg to cross pages which acts as the mb on the writer
2960	 * side. Note, the reader will constantly fail the swap
2961	 * while the writer is updating the pointers, so this
2962	 * guarantees that the overwrite recorded here is the one we
2963	 * want to compare with the last_overrun.
2964	 */
2965	smp_mb();
2966	overwrite = local_read(&(cpu_buffer->overrun));
2967
2968	/*
2969	 * Here's the tricky part.
2970	 *
2971	 * We need to move the pointer past the header page.
2972	 * But we can only do that if a writer is not currently
2973	 * moving it. The page before the header page has the
2974	 * flag bit '1' set if it is pointing to the page we want.
2975	 * but if the writer is in the process of moving it
2976	 * than it will be '2' or already moved '0'.
2977	 */
2978
2979	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2980
2981	/*
2982	 * If we did not convert it, then we must try again.
2983	 */
2984	if (!ret)
2985		goto spin;
2986
2987	/*
2988	 * Yeah! We succeeded in replacing the page.
2989	 *
2990	 * Now make the new head point back to the reader page.
2991	 */
2992	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
2993	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2994
 
 
2995	/* Finally update the reader page to the new head */
2996	cpu_buffer->reader_page = reader;
2997	rb_reset_reader_page(cpu_buffer);
2998
2999	if (overwrite != cpu_buffer->last_overrun) {
3000		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3001		cpu_buffer->last_overrun = overwrite;
3002	}
3003
3004	goto again;
3005
3006 out:
 
 
 
 
3007	arch_spin_unlock(&cpu_buffer->lock);
3008	local_irq_restore(flags);
3009
3010	return reader;
3011}
3012
3013static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3014{
3015	struct ring_buffer_event *event;
3016	struct buffer_page *reader;
3017	unsigned length;
3018
3019	reader = rb_get_reader_page(cpu_buffer);
3020
3021	/* This function should not be called when buffer is empty */
3022	if (RB_WARN_ON(cpu_buffer, !reader))
3023		return;
3024
3025	event = rb_reader_event(cpu_buffer);
3026
3027	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3028		cpu_buffer->read++;
3029
3030	rb_update_read_stamp(cpu_buffer, event);
3031
3032	length = rb_event_length(event);
3033	cpu_buffer->reader_page->read += length;
3034}
3035
3036static void rb_advance_iter(struct ring_buffer_iter *iter)
3037{
3038	struct ring_buffer_per_cpu *cpu_buffer;
3039	struct ring_buffer_event *event;
3040	unsigned length;
3041
3042	cpu_buffer = iter->cpu_buffer;
3043
3044	/*
3045	 * Check if we are at the end of the buffer.
3046	 */
3047	if (iter->head >= rb_page_size(iter->head_page)) {
3048		/* discarded commits can make the page empty */
3049		if (iter->head_page == cpu_buffer->commit_page)
3050			return;
3051		rb_inc_iter(iter);
3052		return;
3053	}
3054
3055	event = rb_iter_head_event(iter);
3056
3057	length = rb_event_length(event);
3058
3059	/*
3060	 * This should not be called to advance the header if we are
3061	 * at the tail of the buffer.
3062	 */
3063	if (RB_WARN_ON(cpu_buffer,
3064		       (iter->head_page == cpu_buffer->commit_page) &&
3065		       (iter->head + length > rb_commit_index(cpu_buffer))))
3066		return;
3067
3068	rb_update_iter_read_stamp(iter, event);
3069
3070	iter->head += length;
3071
3072	/* check for end of page padding */
3073	if ((iter->head >= rb_page_size(iter->head_page)) &&
3074	    (iter->head_page != cpu_buffer->commit_page))
3075		rb_advance_iter(iter);
3076}
3077
3078static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3079{
3080	return cpu_buffer->lost_events;
3081}
3082
3083static struct ring_buffer_event *
3084rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3085	       unsigned long *lost_events)
3086{
3087	struct ring_buffer_event *event;
3088	struct buffer_page *reader;
3089	int nr_loops = 0;
3090
 
 
3091 again:
3092	/*
3093	 * We repeat when a time extend is encountered.
3094	 * Since the time extend is always attached to a data event,
3095	 * we should never loop more than once.
3096	 * (We never hit the following condition more than twice).
3097	 */
3098	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3099		return NULL;
3100
3101	reader = rb_get_reader_page(cpu_buffer);
3102	if (!reader)
3103		return NULL;
3104
3105	event = rb_reader_event(cpu_buffer);
3106
3107	switch (event->type_len) {
3108	case RINGBUF_TYPE_PADDING:
3109		if (rb_null_event(event))
3110			RB_WARN_ON(cpu_buffer, 1);
3111		/*
3112		 * Because the writer could be discarding every
3113		 * event it creates (which would probably be bad)
3114		 * if we were to go back to "again" then we may never
3115		 * catch up, and will trigger the warn on, or lock
3116		 * the box. Return the padding, and we will release
3117		 * the current locks, and try again.
3118		 */
3119		return event;
3120
3121	case RINGBUF_TYPE_TIME_EXTEND:
3122		/* Internal data, OK to advance */
3123		rb_advance_reader(cpu_buffer);
3124		goto again;
3125
3126	case RINGBUF_TYPE_TIME_STAMP:
3127		/* FIXME: not implemented */
 
 
 
 
 
3128		rb_advance_reader(cpu_buffer);
3129		goto again;
3130
3131	case RINGBUF_TYPE_DATA:
3132		if (ts) {
3133			*ts = cpu_buffer->read_stamp + event->time_delta;
3134			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3135							 cpu_buffer->cpu, ts);
3136		}
3137		if (lost_events)
3138			*lost_events = rb_lost_events(cpu_buffer);
3139		return event;
3140
3141	default:
3142		BUG();
3143	}
3144
3145	return NULL;
3146}
3147EXPORT_SYMBOL_GPL(ring_buffer_peek);
3148
3149static struct ring_buffer_event *
3150rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3151{
3152	struct ring_buffer *buffer;
3153	struct ring_buffer_per_cpu *cpu_buffer;
3154	struct ring_buffer_event *event;
3155	int nr_loops = 0;
3156
 
 
 
3157	cpu_buffer = iter->cpu_buffer;
3158	buffer = cpu_buffer->buffer;
3159
3160	/*
3161	 * Check if someone performed a consuming read to
3162	 * the buffer. A consuming read invalidates the iterator
3163	 * and we need to reset the iterator in this case.
3164	 */
3165	if (unlikely(iter->cache_read != cpu_buffer->read ||
3166		     iter->cache_reader_page != cpu_buffer->reader_page))
3167		rb_iter_reset(iter);
3168
3169 again:
3170	if (ring_buffer_iter_empty(iter))
3171		return NULL;
3172
3173	/*
3174	 * We repeat when a time extend is encountered.
3175	 * Since the time extend is always attached to a data event,
3176	 * we should never loop more than once.
3177	 * (We never hit the following condition more than twice).
 
 
3178	 */
3179	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3180		return NULL;
3181
3182	if (rb_per_cpu_empty(cpu_buffer))
3183		return NULL;
3184
3185	if (iter->head >= local_read(&iter->head_page->page->commit)) {
3186		rb_inc_iter(iter);
3187		goto again;
3188	}
3189
3190	event = rb_iter_head_event(iter);
3191
3192	switch (event->type_len) {
3193	case RINGBUF_TYPE_PADDING:
3194		if (rb_null_event(event)) {
3195			rb_inc_iter(iter);
3196			goto again;
3197		}
3198		rb_advance_iter(iter);
3199		return event;
3200
3201	case RINGBUF_TYPE_TIME_EXTEND:
3202		/* Internal data, OK to advance */
3203		rb_advance_iter(iter);
3204		goto again;
3205
3206	case RINGBUF_TYPE_TIME_STAMP:
3207		/* FIXME: not implemented */
 
 
 
 
 
3208		rb_advance_iter(iter);
3209		goto again;
3210
3211	case RINGBUF_TYPE_DATA:
3212		if (ts) {
3213			*ts = iter->read_stamp + event->time_delta;
3214			ring_buffer_normalize_time_stamp(buffer,
3215							 cpu_buffer->cpu, ts);
3216		}
3217		return event;
3218
3219	default:
3220		BUG();
3221	}
3222
3223	return NULL;
3224}
3225EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3226
3227static inline int rb_ok_to_lock(void)
3228{
 
 
 
 
 
3229	/*
3230	 * If an NMI die dumps out the content of the ring buffer
3231	 * do not grab locks. We also permanently disable the ring
3232	 * buffer too. A one time deal is all you get from reading
3233	 * the ring buffer from an NMI.
 
 
 
3234	 */
3235	if (likely(!in_nmi()))
3236		return 1;
 
 
 
 
 
3237
3238	tracing_off_permanent();
3239	return 0;
 
 
 
 
3240}
3241
3242/**
3243 * ring_buffer_peek - peek at the next event to be read
3244 * @buffer: The ring buffer to read
3245 * @cpu: The cpu to peak at
3246 * @ts: The timestamp counter of this event.
3247 * @lost_events: a variable to store if events were lost (may be NULL)
3248 *
3249 * This will return the event that will be read next, but does
3250 * not consume the data.
3251 */
3252struct ring_buffer_event *
3253ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3254		 unsigned long *lost_events)
3255{
3256	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3257	struct ring_buffer_event *event;
3258	unsigned long flags;
3259	int dolock;
3260
3261	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3262		return NULL;
3263
3264	dolock = rb_ok_to_lock();
3265 again:
3266	local_irq_save(flags);
3267	if (dolock)
3268		spin_lock(&cpu_buffer->reader_lock);
3269	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3270	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3271		rb_advance_reader(cpu_buffer);
3272	if (dolock)
3273		spin_unlock(&cpu_buffer->reader_lock);
3274	local_irq_restore(flags);
3275
3276	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3277		goto again;
3278
3279	return event;
3280}
3281
3282/**
3283 * ring_buffer_iter_peek - peek at the next event to be read
3284 * @iter: The ring buffer iterator
3285 * @ts: The timestamp counter of this event.
3286 *
3287 * This will return the event that will be read next, but does
3288 * not increment the iterator.
3289 */
3290struct ring_buffer_event *
3291ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3292{
3293	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3294	struct ring_buffer_event *event;
3295	unsigned long flags;
3296
3297 again:
3298	spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3299	event = rb_iter_peek(iter, ts);
3300	spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3301
3302	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3303		goto again;
3304
3305	return event;
3306}
3307
3308/**
3309 * ring_buffer_consume - return an event and consume it
3310 * @buffer: The ring buffer to get the next event from
3311 * @cpu: the cpu to read the buffer from
3312 * @ts: a variable to store the timestamp (may be NULL)
3313 * @lost_events: a variable to store if events were lost (may be NULL)
3314 *
3315 * Returns the next event in the ring buffer, and that event is consumed.
3316 * Meaning, that sequential reads will keep returning a different event,
3317 * and eventually empty the ring buffer if the producer is slower.
3318 */
3319struct ring_buffer_event *
3320ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3321		    unsigned long *lost_events)
3322{
3323	struct ring_buffer_per_cpu *cpu_buffer;
3324	struct ring_buffer_event *event = NULL;
3325	unsigned long flags;
3326	int dolock;
3327
3328	dolock = rb_ok_to_lock();
3329
3330 again:
3331	/* might be called in atomic */
3332	preempt_disable();
3333
3334	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3335		goto out;
3336
3337	cpu_buffer = buffer->buffers[cpu];
3338	local_irq_save(flags);
3339	if (dolock)
3340		spin_lock(&cpu_buffer->reader_lock);
3341
3342	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3343	if (event) {
3344		cpu_buffer->lost_events = 0;
3345		rb_advance_reader(cpu_buffer);
3346	}
3347
3348	if (dolock)
3349		spin_unlock(&cpu_buffer->reader_lock);
3350	local_irq_restore(flags);
3351
3352 out:
3353	preempt_enable();
3354
3355	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3356		goto again;
3357
3358	return event;
3359}
3360EXPORT_SYMBOL_GPL(ring_buffer_consume);
3361
3362/**
3363 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3364 * @buffer: The ring buffer to read from
3365 * @cpu: The cpu buffer to iterate over
 
3366 *
3367 * This performs the initial preparations necessary to iterate
3368 * through the buffer.  Memory is allocated, buffer recording
3369 * is disabled, and the iterator pointer is returned to the caller.
3370 *
3371 * Disabling buffer recordng prevents the reading from being
3372 * corrupted. This is not a consuming read, so a producer is not
3373 * expected.
3374 *
3375 * After a sequence of ring_buffer_read_prepare calls, the user is
3376 * expected to make at least one call to ring_buffer_prepare_sync.
3377 * Afterwards, ring_buffer_read_start is invoked to get things going
3378 * for real.
3379 *
3380 * This overall must be paired with ring_buffer_finish.
3381 */
3382struct ring_buffer_iter *
3383ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3384{
3385	struct ring_buffer_per_cpu *cpu_buffer;
3386	struct ring_buffer_iter *iter;
3387
3388	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3389		return NULL;
3390
3391	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3392	if (!iter)
3393		return NULL;
3394
3395	cpu_buffer = buffer->buffers[cpu];
3396
3397	iter->cpu_buffer = cpu_buffer;
3398
 
3399	atomic_inc(&cpu_buffer->record_disabled);
3400
3401	return iter;
3402}
3403EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3404
3405/**
3406 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3407 *
3408 * All previously invoked ring_buffer_read_prepare calls to prepare
3409 * iterators will be synchronized.  Afterwards, read_buffer_read_start
3410 * calls on those iterators are allowed.
3411 */
3412void
3413ring_buffer_read_prepare_sync(void)
3414{
3415	synchronize_sched();
3416}
3417EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3418
3419/**
3420 * ring_buffer_read_start - start a non consuming read of the buffer
3421 * @iter: The iterator returned by ring_buffer_read_prepare
3422 *
3423 * This finalizes the startup of an iteration through the buffer.
3424 * The iterator comes from a call to ring_buffer_read_prepare and
3425 * an intervening ring_buffer_read_prepare_sync must have been
3426 * performed.
3427 *
3428 * Must be paired with ring_buffer_finish.
3429 */
3430void
3431ring_buffer_read_start(struct ring_buffer_iter *iter)
3432{
3433	struct ring_buffer_per_cpu *cpu_buffer;
3434	unsigned long flags;
3435
3436	if (!iter)
3437		return;
3438
3439	cpu_buffer = iter->cpu_buffer;
3440
3441	spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3442	arch_spin_lock(&cpu_buffer->lock);
3443	rb_iter_reset(iter);
3444	arch_spin_unlock(&cpu_buffer->lock);
3445	spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3446}
3447EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3448
3449/**
3450 * ring_buffer_finish - finish reading the iterator of the buffer
3451 * @iter: The iterator retrieved by ring_buffer_start
3452 *
3453 * This re-enables the recording to the buffer, and frees the
3454 * iterator.
3455 */
3456void
3457ring_buffer_read_finish(struct ring_buffer_iter *iter)
3458{
3459	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
 
 
 
 
 
 
 
 
 
 
 
3460
3461	atomic_dec(&cpu_buffer->record_disabled);
 
3462	kfree(iter);
3463}
3464EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3465
3466/**
3467 * ring_buffer_read - read the next item in the ring buffer by the iterator
3468 * @iter: The ring buffer iterator
3469 * @ts: The time stamp of the event read.
3470 *
3471 * This reads the next event in the ring buffer and increments the iterator.
3472 */
3473struct ring_buffer_event *
3474ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3475{
3476	struct ring_buffer_event *event;
3477	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3478	unsigned long flags;
3479
3480	spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3481 again:
3482	event = rb_iter_peek(iter, ts);
3483	if (!event)
3484		goto out;
3485
3486	if (event->type_len == RINGBUF_TYPE_PADDING)
3487		goto again;
3488
3489	rb_advance_iter(iter);
3490 out:
3491	spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3492
3493	return event;
3494}
3495EXPORT_SYMBOL_GPL(ring_buffer_read);
3496
3497/**
3498 * ring_buffer_size - return the size of the ring buffer (in bytes)
3499 * @buffer: The ring buffer.
3500 */
3501unsigned long ring_buffer_size(struct ring_buffer *buffer)
3502{
3503	return BUF_PAGE_SIZE * buffer->pages;
 
 
 
 
 
 
 
 
 
3504}
3505EXPORT_SYMBOL_GPL(ring_buffer_size);
3506
3507static void
3508rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3509{
3510	rb_head_page_deactivate(cpu_buffer);
3511
3512	cpu_buffer->head_page
3513		= list_entry(cpu_buffer->pages, struct buffer_page, list);
3514	local_set(&cpu_buffer->head_page->write, 0);
3515	local_set(&cpu_buffer->head_page->entries, 0);
3516	local_set(&cpu_buffer->head_page->page->commit, 0);
3517
3518	cpu_buffer->head_page->read = 0;
3519
3520	cpu_buffer->tail_page = cpu_buffer->head_page;
3521	cpu_buffer->commit_page = cpu_buffer->head_page;
3522
3523	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
 
3524	local_set(&cpu_buffer->reader_page->write, 0);
3525	local_set(&cpu_buffer->reader_page->entries, 0);
3526	local_set(&cpu_buffer->reader_page->page->commit, 0);
3527	cpu_buffer->reader_page->read = 0;
3528
 
 
3529	local_set(&cpu_buffer->commit_overrun, 0);
3530	local_set(&cpu_buffer->overrun, 0);
3531	local_set(&cpu_buffer->entries, 0);
3532	local_set(&cpu_buffer->committing, 0);
3533	local_set(&cpu_buffer->commits, 0);
 
 
 
 
3534	cpu_buffer->read = 0;
 
3535
3536	cpu_buffer->write_stamp = 0;
3537	cpu_buffer->read_stamp = 0;
3538
3539	cpu_buffer->lost_events = 0;
3540	cpu_buffer->last_overrun = 0;
3541
3542	rb_head_page_activate(cpu_buffer);
3543}
3544
3545/**
3546 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3547 * @buffer: The ring buffer to reset a per cpu buffer of
3548 * @cpu: The CPU buffer to be reset
3549 */
3550void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3551{
3552	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3553	unsigned long flags;
3554
3555	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3556		return;
3557
 
3558	atomic_inc(&cpu_buffer->record_disabled);
3559
3560	spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
 
 
 
3561
3562	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3563		goto out;
3564
3565	arch_spin_lock(&cpu_buffer->lock);
3566
3567	rb_reset_cpu(cpu_buffer);
3568
3569	arch_spin_unlock(&cpu_buffer->lock);
3570
3571 out:
3572	spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3573
3574	atomic_dec(&cpu_buffer->record_disabled);
 
3575}
3576EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3577
3578/**
3579 * ring_buffer_reset - reset a ring buffer
3580 * @buffer: The ring buffer to reset all cpu buffers
3581 */
3582void ring_buffer_reset(struct ring_buffer *buffer)
3583{
3584	int cpu;
3585
3586	for_each_buffer_cpu(buffer, cpu)
3587		ring_buffer_reset_cpu(buffer, cpu);
3588}
3589EXPORT_SYMBOL_GPL(ring_buffer_reset);
3590
3591/**
3592 * rind_buffer_empty - is the ring buffer empty?
3593 * @buffer: The ring buffer to test
3594 */
3595int ring_buffer_empty(struct ring_buffer *buffer)
3596{
3597	struct ring_buffer_per_cpu *cpu_buffer;
3598	unsigned long flags;
3599	int dolock;
3600	int cpu;
3601	int ret;
3602
3603	dolock = rb_ok_to_lock();
3604
3605	/* yes this is racy, but if you don't like the race, lock the buffer */
3606	for_each_buffer_cpu(buffer, cpu) {
3607		cpu_buffer = buffer->buffers[cpu];
3608		local_irq_save(flags);
3609		if (dolock)
3610			spin_lock(&cpu_buffer->reader_lock);
3611		ret = rb_per_cpu_empty(cpu_buffer);
3612		if (dolock)
3613			spin_unlock(&cpu_buffer->reader_lock);
3614		local_irq_restore(flags);
3615
3616		if (!ret)
3617			return 0;
3618	}
3619
3620	return 1;
3621}
3622EXPORT_SYMBOL_GPL(ring_buffer_empty);
3623
3624/**
3625 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3626 * @buffer: The ring buffer
3627 * @cpu: The CPU buffer to test
3628 */
3629int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3630{
3631	struct ring_buffer_per_cpu *cpu_buffer;
3632	unsigned long flags;
3633	int dolock;
3634	int ret;
3635
3636	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3637		return 1;
3638
3639	dolock = rb_ok_to_lock();
3640
3641	cpu_buffer = buffer->buffers[cpu];
3642	local_irq_save(flags);
3643	if (dolock)
3644		spin_lock(&cpu_buffer->reader_lock);
3645	ret = rb_per_cpu_empty(cpu_buffer);
3646	if (dolock)
3647		spin_unlock(&cpu_buffer->reader_lock);
3648	local_irq_restore(flags);
3649
3650	return ret;
3651}
3652EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3653
3654#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3655/**
3656 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3657 * @buffer_a: One buffer to swap with
3658 * @buffer_b: The other buffer to swap with
3659 *
3660 * This function is useful for tracers that want to take a "snapshot"
3661 * of a CPU buffer and has another back up buffer lying around.
3662 * it is expected that the tracer handles the cpu buffer not being
3663 * used at the moment.
3664 */
3665int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3666			 struct ring_buffer *buffer_b, int cpu)
3667{
3668	struct ring_buffer_per_cpu *cpu_buffer_a;
3669	struct ring_buffer_per_cpu *cpu_buffer_b;
3670	int ret = -EINVAL;
3671
3672	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3673	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
3674		goto out;
3675
 
 
 
3676	/* At least make sure the two buffers are somewhat the same */
3677	if (buffer_a->pages != buffer_b->pages)
3678		goto out;
3679
3680	ret = -EAGAIN;
3681
3682	if (ring_buffer_flags != RB_BUFFERS_ON)
3683		goto out;
3684
3685	if (atomic_read(&buffer_a->record_disabled))
3686		goto out;
3687
3688	if (atomic_read(&buffer_b->record_disabled))
3689		goto out;
3690
3691	cpu_buffer_a = buffer_a->buffers[cpu];
3692	cpu_buffer_b = buffer_b->buffers[cpu];
3693
3694	if (atomic_read(&cpu_buffer_a->record_disabled))
3695		goto out;
3696
3697	if (atomic_read(&cpu_buffer_b->record_disabled))
3698		goto out;
3699
3700	/*
3701	 * We can't do a synchronize_sched here because this
3702	 * function can be called in atomic context.
3703	 * Normally this will be called from the same CPU as cpu.
3704	 * If not it's up to the caller to protect this.
3705	 */
3706	atomic_inc(&cpu_buffer_a->record_disabled);
3707	atomic_inc(&cpu_buffer_b->record_disabled);
3708
3709	ret = -EBUSY;
3710	if (local_read(&cpu_buffer_a->committing))
3711		goto out_dec;
3712	if (local_read(&cpu_buffer_b->committing))
3713		goto out_dec;
3714
3715	buffer_a->buffers[cpu] = cpu_buffer_b;
3716	buffer_b->buffers[cpu] = cpu_buffer_a;
3717
3718	cpu_buffer_b->buffer = buffer_a;
3719	cpu_buffer_a->buffer = buffer_b;
3720
3721	ret = 0;
3722
3723out_dec:
3724	atomic_dec(&cpu_buffer_a->record_disabled);
3725	atomic_dec(&cpu_buffer_b->record_disabled);
3726out:
3727	return ret;
3728}
3729EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3730#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
3731
3732/**
3733 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3734 * @buffer: the buffer to allocate for.
 
3735 *
3736 * This function is used in conjunction with ring_buffer_read_page.
3737 * When reading a full page from the ring buffer, these functions
3738 * can be used to speed up the process. The calling function should
3739 * allocate a few pages first with this function. Then when it
3740 * needs to get pages from the ring buffer, it passes the result
3741 * of this function into ring_buffer_read_page, which will swap
3742 * the page that was allocated, with the read page of the buffer.
3743 *
3744 * Returns:
3745 *  The page allocated, or NULL on error.
3746 */
3747void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
3748{
3749	struct buffer_data_page *bpage;
 
 
3750	struct page *page;
3751
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3752	page = alloc_pages_node(cpu_to_node(cpu),
3753				GFP_KERNEL | __GFP_NORETRY, 0);
3754	if (!page)
3755		return NULL;
3756
3757	bpage = page_address(page);
3758
 
3759	rb_init_page(bpage);
3760
3761	return bpage;
3762}
3763EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3764
3765/**
3766 * ring_buffer_free_read_page - free an allocated read page
3767 * @buffer: the buffer the page was allocate for
 
3768 * @data: the page to free
3769 *
3770 * Free a page allocated from ring_buffer_alloc_read_page.
3771 */
3772void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3773{
3774	free_page((unsigned long)data);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3775}
3776EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3777
3778/**
3779 * ring_buffer_read_page - extract a page from the ring buffer
3780 * @buffer: buffer to extract from
3781 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3782 * @len: amount to extract
3783 * @cpu: the cpu of the buffer to extract
3784 * @full: should the extraction only happen when the page is full.
3785 *
3786 * This function will pull out a page from the ring buffer and consume it.
3787 * @data_page must be the address of the variable that was returned
3788 * from ring_buffer_alloc_read_page. This is because the page might be used
3789 * to swap with a page in the ring buffer.
3790 *
3791 * for example:
3792 *	rpage = ring_buffer_alloc_read_page(buffer);
3793 *	if (!rpage)
3794 *		return error;
3795 *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3796 *	if (ret >= 0)
3797 *		process_page(rpage, ret);
3798 *
3799 * When @full is set, the function will not return true unless
3800 * the writer is off the reader page.
3801 *
3802 * Note: it is up to the calling functions to handle sleeps and wakeups.
3803 *  The ring buffer can be used anywhere in the kernel and can not
3804 *  blindly call wake_up. The layer that uses the ring buffer must be
3805 *  responsible for that.
3806 *
3807 * Returns:
3808 *  >=0 if data has been transferred, returns the offset of consumed data.
3809 *  <0 if no data has been transferred.
3810 */
3811int ring_buffer_read_page(struct ring_buffer *buffer,
3812			  void **data_page, size_t len, int cpu, int full)
3813{
3814	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3815	struct ring_buffer_event *event;
3816	struct buffer_data_page *bpage;
3817	struct buffer_page *reader;
3818	unsigned long missed_events;
3819	unsigned long flags;
3820	unsigned int commit;
3821	unsigned int read;
3822	u64 save_timestamp;
3823	int ret = -1;
3824
3825	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3826		goto out;
3827
3828	/*
3829	 * If len is not big enough to hold the page header, then
3830	 * we can not copy anything.
3831	 */
3832	if (len <= BUF_PAGE_HDR_SIZE)
3833		goto out;
3834
3835	len -= BUF_PAGE_HDR_SIZE;
3836
3837	if (!data_page)
3838		goto out;
3839
3840	bpage = *data_page;
3841	if (!bpage)
3842		goto out;
3843
3844	spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3845
3846	reader = rb_get_reader_page(cpu_buffer);
3847	if (!reader)
3848		goto out_unlock;
3849
3850	event = rb_reader_event(cpu_buffer);
3851
3852	read = reader->read;
3853	commit = rb_page_commit(reader);
3854
3855	/* Check if any events were dropped */
3856	missed_events = cpu_buffer->lost_events;
3857
3858	/*
3859	 * If this page has been partially read or
3860	 * if len is not big enough to read the rest of the page or
3861	 * a writer is still on the page, then
3862	 * we must copy the data from the page to the buffer.
3863	 * Otherwise, we can simply swap the page with the one passed in.
3864	 */
3865	if (read || (len < (commit - read)) ||
3866	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
3867		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3868		unsigned int rpos = read;
3869		unsigned int pos = 0;
3870		unsigned int size;
3871
3872		if (full)
3873			goto out_unlock;
3874
3875		if (len > (commit - read))
3876			len = (commit - read);
3877
3878		/* Always keep the time extend and data together */
3879		size = rb_event_ts_length(event);
3880
3881		if (len < size)
3882			goto out_unlock;
3883
3884		/* save the current timestamp, since the user will need it */
3885		save_timestamp = cpu_buffer->read_stamp;
3886
3887		/* Need to copy one event at a time */
3888		do {
3889			/* We need the size of one event, because
3890			 * rb_advance_reader only advances by one event,
3891			 * whereas rb_event_ts_length may include the size of
3892			 * one or two events.
3893			 * We have already ensured there's enough space if this
3894			 * is a time extend. */
3895			size = rb_event_length(event);
3896			memcpy(bpage->data + pos, rpage->data + rpos, size);
3897
3898			len -= size;
3899
3900			rb_advance_reader(cpu_buffer);
3901			rpos = reader->read;
3902			pos += size;
3903
3904			if (rpos >= commit)
3905				break;
3906
3907			event = rb_reader_event(cpu_buffer);
3908			/* Always keep the time extend and data together */
3909			size = rb_event_ts_length(event);
3910		} while (len >= size);
3911
3912		/* update bpage */
3913		local_set(&bpage->commit, pos);
3914		bpage->time_stamp = save_timestamp;
3915
3916		/* we copied everything to the beginning */
3917		read = 0;
3918	} else {
3919		/* update the entry counter */
3920		cpu_buffer->read += rb_page_entries(reader);
 
3921
3922		/* swap the pages */
3923		rb_init_page(bpage);
3924		bpage = reader->page;
3925		reader->page = *data_page;
3926		local_set(&reader->write, 0);
3927		local_set(&reader->entries, 0);
3928		reader->read = 0;
3929		*data_page = bpage;
3930
3931		/*
3932		 * Use the real_end for the data size,
3933		 * This gives us a chance to store the lost events
3934		 * on the page.
3935		 */
3936		if (reader->real_end)
3937			local_set(&bpage->commit, reader->real_end);
3938	}
3939	ret = read;
3940
3941	cpu_buffer->lost_events = 0;
3942
3943	commit = local_read(&bpage->commit);
3944	/*
3945	 * Set a flag in the commit field if we lost events
3946	 */
3947	if (missed_events) {
3948		/* If there is room at the end of the page to save the
3949		 * missed events, then record it there.
3950		 */
3951		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
3952			memcpy(&bpage->data[commit], &missed_events,
3953			       sizeof(missed_events));
3954			local_add(RB_MISSED_STORED, &bpage->commit);
3955			commit += sizeof(missed_events);
3956		}
3957		local_add(RB_MISSED_EVENTS, &bpage->commit);
3958	}
3959
3960	/*
3961	 * This page may be off to user land. Zero it out here.
3962	 */
3963	if (commit < BUF_PAGE_SIZE)
3964		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
3965
3966 out_unlock:
3967	spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3968
3969 out:
3970	return ret;
3971}
3972EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3973
3974#ifdef CONFIG_TRACING
3975static ssize_t
3976rb_simple_read(struct file *filp, char __user *ubuf,
3977	       size_t cnt, loff_t *ppos)
3978{
3979	unsigned long *p = filp->private_data;
3980	char buf[64];
3981	int r;
 
 
 
3982
3983	if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3984		r = sprintf(buf, "permanently disabled\n");
3985	else
3986		r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3987
3988	return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3989}
3990
3991static ssize_t
3992rb_simple_write(struct file *filp, const char __user *ubuf,
3993		size_t cnt, loff_t *ppos)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3994{
3995	unsigned long *p = filp->private_data;
3996	unsigned long val;
3997	int ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3998
3999	ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
4000	if (ret)
4001		return ret;
4002
4003	if (val)
4004		set_bit(RB_BUFFERS_ON_BIT, p);
4005	else
4006		clear_bit(RB_BUFFERS_ON_BIT, p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4007
4008	(*ppos)++;
 
4009
4010	return cnt;
4011}
4012
4013static const struct file_operations rb_simple_fops = {
4014	.open		= tracing_open_generic,
4015	.read		= rb_simple_read,
4016	.write		= rb_simple_write,
4017	.llseek		= default_llseek,
4018};
 
 
 
 
 
 
4019
 
 
4020
4021static __init int rb_init_debugfs(void)
4022{
4023	struct dentry *d_tracer;
 
 
 
 
 
4024
4025	d_tracer = tracing_init_dentry();
 
 
4026
4027	trace_create_file("tracing_on", 0644, d_tracer,
4028			    &ring_buffer_flags, &rb_simple_fops);
 
 
 
4029
4030	return 0;
4031}
4032
4033fs_initcall(rb_init_debugfs);
4034#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4035
4036#ifdef CONFIG_HOTPLUG_CPU
4037static int rb_cpu_notify(struct notifier_block *self,
4038			 unsigned long action, void *hcpu)
4039{
4040	struct ring_buffer *buffer =
4041		container_of(self, struct ring_buffer, cpu_notify);
4042	long cpu = (long)hcpu;
4043
4044	switch (action) {
4045	case CPU_UP_PREPARE:
4046	case CPU_UP_PREPARE_FROZEN:
4047		if (cpumask_test_cpu(cpu, buffer->cpumask))
4048			return NOTIFY_OK;
4049
4050		buffer->buffers[cpu] =
4051			rb_allocate_cpu_buffer(buffer, cpu);
4052		if (!buffer->buffers[cpu]) {
4053			WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4054			     cpu);
4055			return NOTIFY_OK;
4056		}
4057		smp_wmb();
4058		cpumask_set_cpu(cpu, buffer->cpumask);
4059		break;
4060	case CPU_DOWN_PREPARE:
4061	case CPU_DOWN_PREPARE_FROZEN:
4062		/*
4063		 * Do nothing.
4064		 *  If we were to free the buffer, then the user would
4065		 *  lose any trace that was in the buffer.
4066		 */
4067		break;
4068	default:
4069		break;
4070	}
4071	return NOTIFY_OK;
 
 
 
 
4072}
4073#endif