Linux Audio

Check our new training course

Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Generic ring buffer
   4 *
   5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
   6 */
   7#include <linux/trace_events.h>
   8#include <linux/ring_buffer.h>
   9#include <linux/trace_clock.h>
  10#include <linux/sched/clock.h>
  11#include <linux/trace_seq.h>
  12#include <linux/spinlock.h>
  13#include <linux/irq_work.h>
  14#include <linux/uaccess.h>
  15#include <linux/hardirq.h>
  16#include <linux/kthread.h>	/* for self test */
 
  17#include <linux/module.h>
  18#include <linux/percpu.h>
  19#include <linux/mutex.h>
  20#include <linux/delay.h>
  21#include <linux/slab.h>
  22#include <linux/init.h>
  23#include <linux/hash.h>
  24#include <linux/list.h>
  25#include <linux/cpu.h>
  26#include <linux/oom.h>
  27
  28#include <asm/local.h>
  29
  30static void update_pages_handler(struct work_struct *work);
  31
  32/*
  33 * The ring buffer header is special. We must manually up keep it.
  34 */
  35int ring_buffer_print_entry_header(struct trace_seq *s)
  36{
  37	trace_seq_puts(s, "# compressed entry header\n");
  38	trace_seq_puts(s, "\ttype_len    :    5 bits\n");
  39	trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
  40	trace_seq_puts(s, "\tarray       :   32 bits\n");
  41	trace_seq_putc(s, '\n');
  42	trace_seq_printf(s, "\tpadding     : type == %d\n",
  43			 RINGBUF_TYPE_PADDING);
  44	trace_seq_printf(s, "\ttime_extend : type == %d\n",
  45			 RINGBUF_TYPE_TIME_EXTEND);
  46	trace_seq_printf(s, "\ttime_stamp : type == %d\n",
  47			 RINGBUF_TYPE_TIME_STAMP);
  48	trace_seq_printf(s, "\tdata max type_len  == %d\n",
  49			 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
  50
  51	return !trace_seq_has_overflowed(s);
  52}
  53
  54/*
  55 * The ring buffer is made up of a list of pages. A separate list of pages is
  56 * allocated for each CPU. A writer may only write to a buffer that is
  57 * associated with the CPU it is currently executing on.  A reader may read
  58 * from any per cpu buffer.
  59 *
  60 * The reader is special. For each per cpu buffer, the reader has its own
  61 * reader page. When a reader has read the entire reader page, this reader
  62 * page is swapped with another page in the ring buffer.
  63 *
  64 * Now, as long as the writer is off the reader page, the reader can do what
  65 * ever it wants with that page. The writer will never write to that page
  66 * again (as long as it is out of the ring buffer).
  67 *
  68 * Here's some silly ASCII art.
  69 *
  70 *   +------+
  71 *   |reader|          RING BUFFER
  72 *   |page  |
  73 *   +------+        +---+   +---+   +---+
  74 *                   |   |-->|   |-->|   |
  75 *                   +---+   +---+   +---+
  76 *                     ^               |
  77 *                     |               |
  78 *                     +---------------+
  79 *
  80 *
  81 *   +------+
  82 *   |reader|          RING BUFFER
  83 *   |page  |------------------v
  84 *   +------+        +---+   +---+   +---+
  85 *                   |   |-->|   |-->|   |
  86 *                   +---+   +---+   +---+
  87 *                     ^               |
  88 *                     |               |
  89 *                     +---------------+
  90 *
  91 *
  92 *   +------+
  93 *   |reader|          RING BUFFER
  94 *   |page  |------------------v
  95 *   +------+        +---+   +---+   +---+
  96 *      ^            |   |-->|   |-->|   |
  97 *      |            +---+   +---+   +---+
  98 *      |                              |
  99 *      |                              |
 100 *      +------------------------------+
 101 *
 102 *
 103 *   +------+
 104 *   |buffer|          RING BUFFER
 105 *   |page  |------------------v
 106 *   +------+        +---+   +---+   +---+
 107 *      ^            |   |   |   |-->|   |
 108 *      |   New      +---+   +---+   +---+
 109 *      |  Reader------^               |
 110 *      |   page                       |
 111 *      +------------------------------+
 112 *
 113 *
 114 * After we make this swap, the reader can hand this page off to the splice
 115 * code and be done with it. It can even allocate a new page if it needs to
 116 * and swap that into the ring buffer.
 117 *
 118 * We will be using cmpxchg soon to make all this lockless.
 119 *
 120 */
 121
 122/* Used for individual buffers (after the counter) */
 123#define RB_BUFFER_OFF		(1 << 20)
 124
 125#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
 126
 127#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
 128#define RB_ALIGNMENT		4U
 129#define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 130#define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
 131#define RB_ALIGN_DATA		__aligned(RB_ALIGNMENT)
 
 
 
 
 
 
 
 
 
 132
 133/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
 134#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
 135
 136enum {
 137	RB_LEN_TIME_EXTEND = 8,
 138	RB_LEN_TIME_STAMP =  8,
 139};
 140
 141#define skip_time_extend(event) \
 142	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
 143
 144#define extended_time(event) \
 145	(event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
 146
 147static inline int rb_null_event(struct ring_buffer_event *event)
 148{
 149	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
 150}
 151
 152static void rb_event_set_padding(struct ring_buffer_event *event)
 153{
 154	/* padding has a NULL time_delta */
 155	event->type_len = RINGBUF_TYPE_PADDING;
 156	event->time_delta = 0;
 157}
 158
 159static unsigned
 160rb_event_data_length(struct ring_buffer_event *event)
 161{
 162	unsigned length;
 163
 164	if (event->type_len)
 165		length = event->type_len * RB_ALIGNMENT;
 166	else
 167		length = event->array[0];
 168	return length + RB_EVNT_HDR_SIZE;
 169}
 170
 171/*
 172 * Return the length of the given event. Will return
 173 * the length of the time extend if the event is a
 174 * time extend.
 175 */
 176static inline unsigned
 177rb_event_length(struct ring_buffer_event *event)
 178{
 179	switch (event->type_len) {
 180	case RINGBUF_TYPE_PADDING:
 181		if (rb_null_event(event))
 182			/* undefined */
 183			return -1;
 184		return  event->array[0] + RB_EVNT_HDR_SIZE;
 185
 186	case RINGBUF_TYPE_TIME_EXTEND:
 187		return RB_LEN_TIME_EXTEND;
 188
 189	case RINGBUF_TYPE_TIME_STAMP:
 190		return RB_LEN_TIME_STAMP;
 191
 192	case RINGBUF_TYPE_DATA:
 193		return rb_event_data_length(event);
 194	default:
 195		BUG();
 196	}
 197	/* not hit */
 198	return 0;
 199}
 200
 201/*
 202 * Return total length of time extend and data,
 203 *   or just the event length for all other events.
 204 */
 205static inline unsigned
 206rb_event_ts_length(struct ring_buffer_event *event)
 207{
 208	unsigned len = 0;
 209
 210	if (extended_time(event)) {
 211		/* time extends include the data event after it */
 212		len = RB_LEN_TIME_EXTEND;
 213		event = skip_time_extend(event);
 214	}
 215	return len + rb_event_length(event);
 216}
 217
 218/**
 219 * ring_buffer_event_length - return the length of the event
 220 * @event: the event to get the length of
 221 *
 222 * Returns the size of the data load of a data event.
 223 * If the event is something other than a data event, it
 224 * returns the size of the event itself. With the exception
 225 * of a TIME EXTEND, where it still returns the size of the
 226 * data load of the data event after it.
 227 */
 228unsigned ring_buffer_event_length(struct ring_buffer_event *event)
 229{
 230	unsigned length;
 231
 232	if (extended_time(event))
 233		event = skip_time_extend(event);
 234
 235	length = rb_event_length(event);
 236	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 237		return length;
 238	length -= RB_EVNT_HDR_SIZE;
 239	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
 240                length -= sizeof(event->array[0]);
 241	return length;
 242}
 243EXPORT_SYMBOL_GPL(ring_buffer_event_length);
 244
 245/* inline for ring buffer fast paths */
 246static __always_inline void *
 247rb_event_data(struct ring_buffer_event *event)
 248{
 249	if (extended_time(event))
 250		event = skip_time_extend(event);
 251	BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 252	/* If length is in len field, then array[0] has the data */
 253	if (event->type_len)
 254		return (void *)&event->array[0];
 255	/* Otherwise length is in array[0] and array[1] has the data */
 256	return (void *)&event->array[1];
 257}
 258
 259/**
 260 * ring_buffer_event_data - return the data of the event
 261 * @event: the event to get the data from
 262 */
 263void *ring_buffer_event_data(struct ring_buffer_event *event)
 264{
 265	return rb_event_data(event);
 266}
 267EXPORT_SYMBOL_GPL(ring_buffer_event_data);
 268
 269#define for_each_buffer_cpu(buffer, cpu)		\
 270	for_each_cpu(cpu, buffer->cpumask)
 271
 272#define TS_SHIFT	27
 273#define TS_MASK		((1ULL << TS_SHIFT) - 1)
 274#define TS_DELTA_TEST	(~TS_MASK)
 275
 276/**
 277 * ring_buffer_event_time_stamp - return the event's extended timestamp
 278 * @event: the event to get the timestamp of
 279 *
 280 * Returns the extended timestamp associated with a data event.
 281 * An extended time_stamp is a 64-bit timestamp represented
 282 * internally in a special way that makes the best use of space
 283 * contained within a ring buffer event.  This function decodes
 284 * it and maps it to a straight u64 value.
 285 */
 286u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
 287{
 288	u64 ts;
 289
 290	ts = event->array[0];
 291	ts <<= TS_SHIFT;
 292	ts += event->time_delta;
 293
 294	return ts;
 295}
 296
 297/* Flag when events were overwritten */
 298#define RB_MISSED_EVENTS	(1 << 31)
 299/* Missed count stored at end */
 300#define RB_MISSED_STORED	(1 << 30)
 301
 302#define RB_MISSED_FLAGS		(RB_MISSED_EVENTS|RB_MISSED_STORED)
 303
 304struct buffer_data_page {
 305	u64		 time_stamp;	/* page time stamp */
 306	local_t		 commit;	/* write committed index */
 307	unsigned char	 data[] RB_ALIGN_DATA;	/* data of buffer page */
 308};
 309
 310/*
 311 * Note, the buffer_page list must be first. The buffer pages
 312 * are allocated in cache lines, which means that each buffer
 313 * page will be at the beginning of a cache line, and thus
 314 * the least significant bits will be zero. We use this to
 315 * add flags in the list struct pointers, to make the ring buffer
 316 * lockless.
 317 */
 318struct buffer_page {
 319	struct list_head list;		/* list of buffer pages */
 320	local_t		 write;		/* index for next write */
 321	unsigned	 read;		/* index for next read */
 322	local_t		 entries;	/* entries on this page */
 323	unsigned long	 real_end;	/* real end of data */
 324	struct buffer_data_page *page;	/* Actual data page */
 325};
 326
 327/*
 328 * The buffer page counters, write and entries, must be reset
 329 * atomically when crossing page boundaries. To synchronize this
 330 * update, two counters are inserted into the number. One is
 331 * the actual counter for the write position or count on the page.
 332 *
 333 * The other is a counter of updaters. Before an update happens
 334 * the update partition of the counter is incremented. This will
 335 * allow the updater to update the counter atomically.
 336 *
 337 * The counter is 20 bits, and the state data is 12.
 338 */
 339#define RB_WRITE_MASK		0xfffff
 340#define RB_WRITE_INTCNT		(1 << 20)
 341
 342static void rb_init_page(struct buffer_data_page *bpage)
 343{
 344	local_set(&bpage->commit, 0);
 345}
 346
 
 
 
 
 
 
 
 
 
 
 
 
 347/*
 348 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
 349 * this issue out.
 350 */
 351static void free_buffer_page(struct buffer_page *bpage)
 352{
 353	free_page((unsigned long)bpage->page);
 354	kfree(bpage);
 355}
 356
 357/*
 358 * We need to fit the time_stamp delta into 27 bits.
 359 */
 360static inline int test_time_stamp(u64 delta)
 361{
 362	if (delta & TS_DELTA_TEST)
 363		return 1;
 364	return 0;
 365}
 366
 367#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
 368
 369/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
 370#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
 371
 372int ring_buffer_print_page_header(struct trace_seq *s)
 373{
 374	struct buffer_data_page field;
 375
 376	trace_seq_printf(s, "\tfield: u64 timestamp;\t"
 377			 "offset:0;\tsize:%u;\tsigned:%u;\n",
 378			 (unsigned int)sizeof(field.time_stamp),
 379			 (unsigned int)is_signed_type(u64));
 380
 381	trace_seq_printf(s, "\tfield: local_t commit;\t"
 382			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 383			 (unsigned int)offsetof(typeof(field), commit),
 384			 (unsigned int)sizeof(field.commit),
 385			 (unsigned int)is_signed_type(long));
 386
 387	trace_seq_printf(s, "\tfield: int overwrite;\t"
 388			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 389			 (unsigned int)offsetof(typeof(field), commit),
 390			 1,
 391			 (unsigned int)is_signed_type(long));
 392
 393	trace_seq_printf(s, "\tfield: char data;\t"
 394			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 395			 (unsigned int)offsetof(typeof(field), data),
 396			 (unsigned int)BUF_PAGE_SIZE,
 397			 (unsigned int)is_signed_type(char));
 398
 399	return !trace_seq_has_overflowed(s);
 400}
 401
 402struct rb_irq_work {
 403	struct irq_work			work;
 404	wait_queue_head_t		waiters;
 405	wait_queue_head_t		full_waiters;
 406	bool				waiters_pending;
 407	bool				full_waiters_pending;
 408	bool				wakeup_full;
 409};
 410
 411/*
 412 * Structure to hold event state and handle nested events.
 413 */
 414struct rb_event_info {
 415	u64			ts;
 416	u64			delta;
 417	unsigned long		length;
 418	struct buffer_page	*tail_page;
 419	int			add_timestamp;
 420};
 421
 422/*
 423 * Used for which event context the event is in.
 424 *  NMI     = 0
 425 *  IRQ     = 1
 426 *  SOFTIRQ = 2
 427 *  NORMAL  = 3
 428 *
 429 * See trace_recursive_lock() comment below for more details.
 430 */
 431enum {
 432	RB_CTX_NMI,
 433	RB_CTX_IRQ,
 434	RB_CTX_SOFTIRQ,
 435	RB_CTX_NORMAL,
 436	RB_CTX_MAX
 437};
 438
 439/*
 440 * head_page == tail_page && head == tail then buffer is empty.
 441 */
 442struct ring_buffer_per_cpu {
 443	int				cpu;
 444	atomic_t			record_disabled;
 445	struct ring_buffer		*buffer;
 446	raw_spinlock_t			reader_lock;	/* serialize readers */
 447	arch_spinlock_t			lock;
 448	struct lock_class_key		lock_key;
 449	struct buffer_data_page		*free_page;
 450	unsigned long			nr_pages;
 451	unsigned int			current_context;
 452	struct list_head		*pages;
 453	struct buffer_page		*head_page;	/* read from head */
 454	struct buffer_page		*tail_page;	/* write to tail */
 455	struct buffer_page		*commit_page;	/* committed pages */
 456	struct buffer_page		*reader_page;
 457	unsigned long			lost_events;
 458	unsigned long			last_overrun;
 459	unsigned long			nest;
 460	local_t				entries_bytes;
 461	local_t				entries;
 462	local_t				overrun;
 463	local_t				commit_overrun;
 464	local_t				dropped_events;
 465	local_t				committing;
 466	local_t				commits;
 467	local_t				pages_touched;
 468	local_t				pages_read;
 469	long				last_pages_touch;
 470	size_t				shortest_full;
 471	unsigned long			read;
 472	unsigned long			read_bytes;
 473	u64				write_stamp;
 474	u64				read_stamp;
 475	/* ring buffer pages to update, > 0 to add, < 0 to remove */
 476	long				nr_pages_to_update;
 477	struct list_head		new_pages; /* new pages to add */
 478	struct work_struct		update_pages_work;
 479	struct completion		update_done;
 480
 481	struct rb_irq_work		irq_work;
 482};
 483
 484struct ring_buffer {
 485	unsigned			flags;
 486	int				cpus;
 487	atomic_t			record_disabled;
 488	atomic_t			resize_disabled;
 489	cpumask_var_t			cpumask;
 490
 491	struct lock_class_key		*reader_lock_key;
 492
 493	struct mutex			mutex;
 494
 495	struct ring_buffer_per_cpu	**buffers;
 496
 497	struct hlist_node		node;
 498	u64				(*clock)(void);
 499
 500	struct rb_irq_work		irq_work;
 501	bool				time_stamp_abs;
 502};
 503
 504struct ring_buffer_iter {
 505	struct ring_buffer_per_cpu	*cpu_buffer;
 506	unsigned long			head;
 507	struct buffer_page		*head_page;
 508	struct buffer_page		*cache_reader_page;
 509	unsigned long			cache_read;
 510	u64				read_stamp;
 511};
 512
 513/**
 514 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
 515 * @buffer: The ring_buffer to get the number of pages from
 516 * @cpu: The cpu of the ring_buffer to get the number of pages from
 517 *
 518 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
 519 */
 520size_t ring_buffer_nr_pages(struct ring_buffer *buffer, int cpu)
 521{
 522	return buffer->buffers[cpu]->nr_pages;
 523}
 524
 525/**
 526 * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
 527 * @buffer: The ring_buffer to get the number of pages from
 528 * @cpu: The cpu of the ring_buffer to get the number of pages from
 529 *
 530 * Returns the number of pages that have content in the ring buffer.
 531 */
 532size_t ring_buffer_nr_dirty_pages(struct ring_buffer *buffer, int cpu)
 533{
 534	size_t read;
 535	size_t cnt;
 536
 537	read = local_read(&buffer->buffers[cpu]->pages_read);
 538	cnt = local_read(&buffer->buffers[cpu]->pages_touched);
 539	/* The reader can read an empty page, but not more than that */
 540	if (cnt < read) {
 541		WARN_ON_ONCE(read > cnt + 1);
 542		return 0;
 543	}
 544
 545	return cnt - read;
 546}
 547
 548/*
 549 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
 550 *
 551 * Schedules a delayed work to wake up any task that is blocked on the
 552 * ring buffer waiters queue.
 553 */
 554static void rb_wake_up_waiters(struct irq_work *work)
 555{
 556	struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
 557
 558	wake_up_all(&rbwork->waiters);
 559	if (rbwork->wakeup_full) {
 560		rbwork->wakeup_full = false;
 561		wake_up_all(&rbwork->full_waiters);
 562	}
 563}
 564
 565/**
 566 * ring_buffer_wait - wait for input to the ring buffer
 567 * @buffer: buffer to wait on
 568 * @cpu: the cpu buffer to wait on
 569 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
 570 *
 571 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 572 * as data is added to any of the @buffer's cpu buffers. Otherwise
 573 * it will wait for data to be added to a specific cpu buffer.
 574 */
 575int ring_buffer_wait(struct ring_buffer *buffer, int cpu, int full)
 576{
 577	struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
 578	DEFINE_WAIT(wait);
 579	struct rb_irq_work *work;
 580	int ret = 0;
 581
 582	/*
 583	 * Depending on what the caller is waiting for, either any
 584	 * data in any cpu buffer, or a specific buffer, put the
 585	 * caller on the appropriate wait queue.
 586	 */
 587	if (cpu == RING_BUFFER_ALL_CPUS) {
 588		work = &buffer->irq_work;
 589		/* Full only makes sense on per cpu reads */
 590		full = 0;
 591	} else {
 592		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 593			return -ENODEV;
 594		cpu_buffer = buffer->buffers[cpu];
 595		work = &cpu_buffer->irq_work;
 596	}
 597
 598
 599	while (true) {
 600		if (full)
 601			prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
 602		else
 603			prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
 604
 605		/*
 606		 * The events can happen in critical sections where
 607		 * checking a work queue can cause deadlocks.
 608		 * After adding a task to the queue, this flag is set
 609		 * only to notify events to try to wake up the queue
 610		 * using irq_work.
 611		 *
 612		 * We don't clear it even if the buffer is no longer
 613		 * empty. The flag only causes the next event to run
 614		 * irq_work to do the work queue wake up. The worse
 615		 * that can happen if we race with !trace_empty() is that
 616		 * an event will cause an irq_work to try to wake up
 617		 * an empty queue.
 618		 *
 619		 * There's no reason to protect this flag either, as
 620		 * the work queue and irq_work logic will do the necessary
 621		 * synchronization for the wake ups. The only thing
 622		 * that is necessary is that the wake up happens after
 623		 * a task has been queued. It's OK for spurious wake ups.
 624		 */
 625		if (full)
 626			work->full_waiters_pending = true;
 627		else
 628			work->waiters_pending = true;
 629
 630		if (signal_pending(current)) {
 631			ret = -EINTR;
 632			break;
 633		}
 634
 635		if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
 636			break;
 637
 638		if (cpu != RING_BUFFER_ALL_CPUS &&
 639		    !ring_buffer_empty_cpu(buffer, cpu)) {
 640			unsigned long flags;
 641			bool pagebusy;
 642			size_t nr_pages;
 643			size_t dirty;
 644
 645			if (!full)
 646				break;
 647
 648			raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
 649			pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 650			nr_pages = cpu_buffer->nr_pages;
 651			dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
 652			if (!cpu_buffer->shortest_full ||
 653			    cpu_buffer->shortest_full < full)
 654				cpu_buffer->shortest_full = full;
 655			raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
 656			if (!pagebusy &&
 657			    (!nr_pages || (dirty * 100) > full * nr_pages))
 658				break;
 659		}
 660
 661		schedule();
 662	}
 663
 664	if (full)
 665		finish_wait(&work->full_waiters, &wait);
 666	else
 667		finish_wait(&work->waiters, &wait);
 668
 669	return ret;
 670}
 671
 672/**
 673 * ring_buffer_poll_wait - poll on buffer input
 674 * @buffer: buffer to wait on
 675 * @cpu: the cpu buffer to wait on
 676 * @filp: the file descriptor
 677 * @poll_table: The poll descriptor
 678 *
 679 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 680 * as data is added to any of the @buffer's cpu buffers. Otherwise
 681 * it will wait for data to be added to a specific cpu buffer.
 682 *
 683 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
 684 * zero otherwise.
 685 */
 686__poll_t ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
 687			  struct file *filp, poll_table *poll_table)
 688{
 689	struct ring_buffer_per_cpu *cpu_buffer;
 690	struct rb_irq_work *work;
 691
 692	if (cpu == RING_BUFFER_ALL_CPUS)
 693		work = &buffer->irq_work;
 694	else {
 695		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 696			return -EINVAL;
 697
 698		cpu_buffer = buffer->buffers[cpu];
 699		work = &cpu_buffer->irq_work;
 700	}
 701
 702	poll_wait(filp, &work->waiters, poll_table);
 703	work->waiters_pending = true;
 704	/*
 705	 * There's a tight race between setting the waiters_pending and
 706	 * checking if the ring buffer is empty.  Once the waiters_pending bit
 707	 * is set, the next event will wake the task up, but we can get stuck
 708	 * if there's only a single event in.
 709	 *
 710	 * FIXME: Ideally, we need a memory barrier on the writer side as well,
 711	 * but adding a memory barrier to all events will cause too much of a
 712	 * performance hit in the fast path.  We only need a memory barrier when
 713	 * the buffer goes from empty to having content.  But as this race is
 714	 * extremely small, and it's not a problem if another event comes in, we
 715	 * will fix it later.
 716	 */
 717	smp_mb();
 718
 719	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
 720	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
 721		return EPOLLIN | EPOLLRDNORM;
 722	return 0;
 723}
 724
 725/* buffer may be either ring_buffer or ring_buffer_per_cpu */
 726#define RB_WARN_ON(b, cond)						\
 727	({								\
 728		int _____ret = unlikely(cond);				\
 729		if (_____ret) {						\
 730			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
 731				struct ring_buffer_per_cpu *__b =	\
 732					(void *)b;			\
 733				atomic_inc(&__b->buffer->record_disabled); \
 734			} else						\
 735				atomic_inc(&b->record_disabled);	\
 736			WARN_ON(1);					\
 737		}							\
 738		_____ret;						\
 739	})
 740
 741/* Up this if you want to test the TIME_EXTENTS and normalization */
 742#define DEBUG_SHIFT 0
 743
 744static inline u64 rb_time_stamp(struct ring_buffer *buffer)
 745{
 746	/* shift to debug/test normalization and TIME_EXTENTS */
 747	return buffer->clock() << DEBUG_SHIFT;
 748}
 749
 750u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
 751{
 752	u64 time;
 753
 754	preempt_disable_notrace();
 755	time = rb_time_stamp(buffer);
 756	preempt_enable_notrace();
 757
 758	return time;
 759}
 760EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
 761
 762void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
 763				      int cpu, u64 *ts)
 764{
 765	/* Just stupid testing the normalize function and deltas */
 766	*ts >>= DEBUG_SHIFT;
 767}
 768EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
 769
 770/*
 771 * Making the ring buffer lockless makes things tricky.
 772 * Although writes only happen on the CPU that they are on,
 773 * and they only need to worry about interrupts. Reads can
 774 * happen on any CPU.
 775 *
 776 * The reader page is always off the ring buffer, but when the
 777 * reader finishes with a page, it needs to swap its page with
 778 * a new one from the buffer. The reader needs to take from
 779 * the head (writes go to the tail). But if a writer is in overwrite
 780 * mode and wraps, it must push the head page forward.
 781 *
 782 * Here lies the problem.
 783 *
 784 * The reader must be careful to replace only the head page, and
 785 * not another one. As described at the top of the file in the
 786 * ASCII art, the reader sets its old page to point to the next
 787 * page after head. It then sets the page after head to point to
 788 * the old reader page. But if the writer moves the head page
 789 * during this operation, the reader could end up with the tail.
 790 *
 791 * We use cmpxchg to help prevent this race. We also do something
 792 * special with the page before head. We set the LSB to 1.
 793 *
 794 * When the writer must push the page forward, it will clear the
 795 * bit that points to the head page, move the head, and then set
 796 * the bit that points to the new head page.
 797 *
 798 * We also don't want an interrupt coming in and moving the head
 799 * page on another writer. Thus we use the second LSB to catch
 800 * that too. Thus:
 801 *
 802 * head->list->prev->next        bit 1          bit 0
 803 *                              -------        -------
 804 * Normal page                     0              0
 805 * Points to head page             0              1
 806 * New head page                   1              0
 807 *
 808 * Note we can not trust the prev pointer of the head page, because:
 809 *
 810 * +----+       +-----+        +-----+
 811 * |    |------>|  T  |---X--->|  N  |
 812 * |    |<------|     |        |     |
 813 * +----+       +-----+        +-----+
 814 *   ^                           ^ |
 815 *   |          +-----+          | |
 816 *   +----------|  R  |----------+ |
 817 *              |     |<-----------+
 818 *              +-----+
 819 *
 820 * Key:  ---X-->  HEAD flag set in pointer
 821 *         T      Tail page
 822 *         R      Reader page
 823 *         N      Next page
 824 *
 825 * (see __rb_reserve_next() to see where this happens)
 826 *
 827 *  What the above shows is that the reader just swapped out
 828 *  the reader page with a page in the buffer, but before it
 829 *  could make the new header point back to the new page added
 830 *  it was preempted by a writer. The writer moved forward onto
 831 *  the new page added by the reader and is about to move forward
 832 *  again.
 833 *
 834 *  You can see, it is legitimate for the previous pointer of
 835 *  the head (or any page) not to point back to itself. But only
 836 *  temporarily.
 837 */
 838
 839#define RB_PAGE_NORMAL		0UL
 840#define RB_PAGE_HEAD		1UL
 841#define RB_PAGE_UPDATE		2UL
 842
 843
 844#define RB_FLAG_MASK		3UL
 845
 846/* PAGE_MOVED is not part of the mask */
 847#define RB_PAGE_MOVED		4UL
 848
 849/*
 850 * rb_list_head - remove any bit
 851 */
 852static struct list_head *rb_list_head(struct list_head *list)
 853{
 854	unsigned long val = (unsigned long)list;
 855
 856	return (struct list_head *)(val & ~RB_FLAG_MASK);
 857}
 858
 859/*
 860 * rb_is_head_page - test if the given page is the head page
 861 *
 862 * Because the reader may move the head_page pointer, we can
 863 * not trust what the head page is (it may be pointing to
 864 * the reader page). But if the next page is a header page,
 865 * its flags will be non zero.
 866 */
 867static inline int
 868rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
 869		struct buffer_page *page, struct list_head *list)
 870{
 871	unsigned long val;
 872
 873	val = (unsigned long)list->next;
 874
 875	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
 876		return RB_PAGE_MOVED;
 877
 878	return val & RB_FLAG_MASK;
 879}
 880
 881/*
 882 * rb_is_reader_page
 883 *
 884 * The unique thing about the reader page, is that, if the
 885 * writer is ever on it, the previous pointer never points
 886 * back to the reader page.
 887 */
 888static bool rb_is_reader_page(struct buffer_page *page)
 889{
 890	struct list_head *list = page->list.prev;
 891
 892	return rb_list_head(list->next) != &page->list;
 893}
 894
 895/*
 896 * rb_set_list_to_head - set a list_head to be pointing to head.
 897 */
 898static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
 899				struct list_head *list)
 900{
 901	unsigned long *ptr;
 902
 903	ptr = (unsigned long *)&list->next;
 904	*ptr |= RB_PAGE_HEAD;
 905	*ptr &= ~RB_PAGE_UPDATE;
 906}
 907
 908/*
 909 * rb_head_page_activate - sets up head page
 910 */
 911static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
 912{
 913	struct buffer_page *head;
 914
 915	head = cpu_buffer->head_page;
 916	if (!head)
 917		return;
 918
 919	/*
 920	 * Set the previous list pointer to have the HEAD flag.
 921	 */
 922	rb_set_list_to_head(cpu_buffer, head->list.prev);
 923}
 924
 925static void rb_list_head_clear(struct list_head *list)
 926{
 927	unsigned long *ptr = (unsigned long *)&list->next;
 928
 929	*ptr &= ~RB_FLAG_MASK;
 930}
 931
 932/*
 933 * rb_head_page_deactivate - clears head page ptr (for free list)
 934 */
 935static void
 936rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
 937{
 938	struct list_head *hd;
 939
 940	/* Go through the whole list and clear any pointers found. */
 941	rb_list_head_clear(cpu_buffer->pages);
 942
 943	list_for_each(hd, cpu_buffer->pages)
 944		rb_list_head_clear(hd);
 945}
 946
 947static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
 948			    struct buffer_page *head,
 949			    struct buffer_page *prev,
 950			    int old_flag, int new_flag)
 951{
 952	struct list_head *list;
 953	unsigned long val = (unsigned long)&head->list;
 954	unsigned long ret;
 955
 956	list = &prev->list;
 957
 958	val &= ~RB_FLAG_MASK;
 959
 960	ret = cmpxchg((unsigned long *)&list->next,
 961		      val | old_flag, val | new_flag);
 962
 963	/* check if the reader took the page */
 964	if ((ret & ~RB_FLAG_MASK) != val)
 965		return RB_PAGE_MOVED;
 966
 967	return ret & RB_FLAG_MASK;
 968}
 969
 970static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
 971				   struct buffer_page *head,
 972				   struct buffer_page *prev,
 973				   int old_flag)
 974{
 975	return rb_head_page_set(cpu_buffer, head, prev,
 976				old_flag, RB_PAGE_UPDATE);
 977}
 978
 979static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
 980				 struct buffer_page *head,
 981				 struct buffer_page *prev,
 982				 int old_flag)
 983{
 984	return rb_head_page_set(cpu_buffer, head, prev,
 985				old_flag, RB_PAGE_HEAD);
 986}
 987
 988static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
 989				   struct buffer_page *head,
 990				   struct buffer_page *prev,
 991				   int old_flag)
 992{
 993	return rb_head_page_set(cpu_buffer, head, prev,
 994				old_flag, RB_PAGE_NORMAL);
 995}
 996
 997static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
 998			       struct buffer_page **bpage)
 999{
1000	struct list_head *p = rb_list_head((*bpage)->list.next);
1001
1002	*bpage = list_entry(p, struct buffer_page, list);
1003}
1004
1005static struct buffer_page *
1006rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1007{
1008	struct buffer_page *head;
1009	struct buffer_page *page;
1010	struct list_head *list;
1011	int i;
1012
1013	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1014		return NULL;
1015
1016	/* sanity check */
1017	list = cpu_buffer->pages;
1018	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1019		return NULL;
1020
1021	page = head = cpu_buffer->head_page;
1022	/*
1023	 * It is possible that the writer moves the header behind
1024	 * where we started, and we miss in one loop.
1025	 * A second loop should grab the header, but we'll do
1026	 * three loops just because I'm paranoid.
1027	 */
1028	for (i = 0; i < 3; i++) {
1029		do {
1030			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1031				cpu_buffer->head_page = page;
1032				return page;
1033			}
1034			rb_inc_page(cpu_buffer, &page);
1035		} while (page != head);
1036	}
1037
1038	RB_WARN_ON(cpu_buffer, 1);
1039
1040	return NULL;
1041}
1042
1043static int rb_head_page_replace(struct buffer_page *old,
1044				struct buffer_page *new)
1045{
1046	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1047	unsigned long val;
1048	unsigned long ret;
1049
1050	val = *ptr & ~RB_FLAG_MASK;
1051	val |= RB_PAGE_HEAD;
1052
1053	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1054
1055	return ret == val;
1056}
1057
1058/*
1059 * rb_tail_page_update - move the tail page forward
1060 */
1061static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1062			       struct buffer_page *tail_page,
1063			       struct buffer_page *next_page)
1064{
1065	unsigned long old_entries;
1066	unsigned long old_write;
1067
1068	/*
1069	 * The tail page now needs to be moved forward.
1070	 *
1071	 * We need to reset the tail page, but without messing
1072	 * with possible erasing of data brought in by interrupts
1073	 * that have moved the tail page and are currently on it.
1074	 *
1075	 * We add a counter to the write field to denote this.
1076	 */
1077	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1078	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1079
1080	local_inc(&cpu_buffer->pages_touched);
1081	/*
1082	 * Just make sure we have seen our old_write and synchronize
1083	 * with any interrupts that come in.
1084	 */
1085	barrier();
1086
1087	/*
1088	 * If the tail page is still the same as what we think
1089	 * it is, then it is up to us to update the tail
1090	 * pointer.
1091	 */
1092	if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1093		/* Zero the write counter */
1094		unsigned long val = old_write & ~RB_WRITE_MASK;
1095		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1096
1097		/*
1098		 * This will only succeed if an interrupt did
1099		 * not come in and change it. In which case, we
1100		 * do not want to modify it.
1101		 *
1102		 * We add (void) to let the compiler know that we do not care
1103		 * about the return value of these functions. We use the
1104		 * cmpxchg to only update if an interrupt did not already
1105		 * do it for us. If the cmpxchg fails, we don't care.
1106		 */
1107		(void)local_cmpxchg(&next_page->write, old_write, val);
1108		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1109
1110		/*
1111		 * No need to worry about races with clearing out the commit.
1112		 * it only can increment when a commit takes place. But that
1113		 * only happens in the outer most nested commit.
1114		 */
1115		local_set(&next_page->page->commit, 0);
1116
1117		/* Again, either we update tail_page or an interrupt does */
1118		(void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1119	}
1120}
1121
1122static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1123			  struct buffer_page *bpage)
1124{
1125	unsigned long val = (unsigned long)bpage;
1126
1127	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1128		return 1;
1129
1130	return 0;
1131}
1132
1133/**
1134 * rb_check_list - make sure a pointer to a list has the last bits zero
1135 */
1136static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1137			 struct list_head *list)
1138{
1139	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1140		return 1;
1141	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1142		return 1;
1143	return 0;
1144}
1145
1146/**
1147 * rb_check_pages - integrity check of buffer pages
1148 * @cpu_buffer: CPU buffer with pages to test
1149 *
1150 * As a safety measure we check to make sure the data pages have not
1151 * been corrupted.
1152 */
1153static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1154{
1155	struct list_head *head = cpu_buffer->pages;
1156	struct buffer_page *bpage, *tmp;
1157
1158	/* Reset the head page if it exists */
1159	if (cpu_buffer->head_page)
1160		rb_set_head_page(cpu_buffer);
1161
1162	rb_head_page_deactivate(cpu_buffer);
1163
1164	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1165		return -1;
1166	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1167		return -1;
1168
1169	if (rb_check_list(cpu_buffer, head))
1170		return -1;
1171
1172	list_for_each_entry_safe(bpage, tmp, head, list) {
1173		if (RB_WARN_ON(cpu_buffer,
1174			       bpage->list.next->prev != &bpage->list))
1175			return -1;
1176		if (RB_WARN_ON(cpu_buffer,
1177			       bpage->list.prev->next != &bpage->list))
1178			return -1;
1179		if (rb_check_list(cpu_buffer, &bpage->list))
1180			return -1;
1181	}
1182
1183	rb_head_page_activate(cpu_buffer);
1184
1185	return 0;
1186}
1187
1188static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1189{
1190	struct buffer_page *bpage, *tmp;
1191	bool user_thread = current->mm != NULL;
1192	gfp_t mflags;
1193	long i;
1194
1195	/*
1196	 * Check if the available memory is there first.
1197	 * Note, si_mem_available() only gives us a rough estimate of available
1198	 * memory. It may not be accurate. But we don't care, we just want
1199	 * to prevent doing any allocation when it is obvious that it is
1200	 * not going to succeed.
1201	 */
1202	i = si_mem_available();
1203	if (i < nr_pages)
1204		return -ENOMEM;
1205
1206	/*
1207	 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1208	 * gracefully without invoking oom-killer and the system is not
1209	 * destabilized.
1210	 */
1211	mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1212
1213	/*
1214	 * If a user thread allocates too much, and si_mem_available()
1215	 * reports there's enough memory, even though there is not.
1216	 * Make sure the OOM killer kills this thread. This can happen
1217	 * even with RETRY_MAYFAIL because another task may be doing
1218	 * an allocation after this task has taken all memory.
1219	 * This is the task the OOM killer needs to take out during this
1220	 * loop, even if it was triggered by an allocation somewhere else.
1221	 */
1222	if (user_thread)
1223		set_current_oom_origin();
1224	for (i = 0; i < nr_pages; i++) {
1225		struct page *page;
1226
 
 
 
 
1227		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1228				    mflags, cpu_to_node(cpu));
 
1229		if (!bpage)
1230			goto free_pages;
1231
1232		list_add(&bpage->list, pages);
1233
1234		page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
 
1235		if (!page)
1236			goto free_pages;
1237		bpage->page = page_address(page);
1238		rb_init_page(bpage->page);
1239
1240		if (user_thread && fatal_signal_pending(current))
1241			goto free_pages;
1242	}
1243	if (user_thread)
1244		clear_current_oom_origin();
1245
1246	return 0;
1247
1248free_pages:
1249	list_for_each_entry_safe(bpage, tmp, pages, list) {
1250		list_del_init(&bpage->list);
1251		free_buffer_page(bpage);
1252	}
1253	if (user_thread)
1254		clear_current_oom_origin();
1255
1256	return -ENOMEM;
1257}
1258
1259static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1260			     unsigned long nr_pages)
1261{
1262	LIST_HEAD(pages);
1263
1264	WARN_ON(!nr_pages);
1265
1266	if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1267		return -ENOMEM;
1268
1269	/*
1270	 * The ring buffer page list is a circular list that does not
1271	 * start and end with a list head. All page list items point to
1272	 * other pages.
1273	 */
1274	cpu_buffer->pages = pages.next;
1275	list_del(&pages);
1276
1277	cpu_buffer->nr_pages = nr_pages;
1278
1279	rb_check_pages(cpu_buffer);
1280
1281	return 0;
1282}
1283
1284static struct ring_buffer_per_cpu *
1285rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1286{
1287	struct ring_buffer_per_cpu *cpu_buffer;
1288	struct buffer_page *bpage;
1289	struct page *page;
1290	int ret;
1291
1292	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1293				  GFP_KERNEL, cpu_to_node(cpu));
1294	if (!cpu_buffer)
1295		return NULL;
1296
1297	cpu_buffer->cpu = cpu;
1298	cpu_buffer->buffer = buffer;
1299	raw_spin_lock_init(&cpu_buffer->reader_lock);
1300	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1301	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1302	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1303	init_completion(&cpu_buffer->update_done);
1304	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1305	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1306	init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1307
1308	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1309			    GFP_KERNEL, cpu_to_node(cpu));
1310	if (!bpage)
1311		goto fail_free_buffer;
1312
1313	rb_check_bpage(cpu_buffer, bpage);
1314
1315	cpu_buffer->reader_page = bpage;
1316	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1317	if (!page)
1318		goto fail_free_reader;
1319	bpage->page = page_address(page);
1320	rb_init_page(bpage->page);
1321
1322	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1323	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1324
1325	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1326	if (ret < 0)
1327		goto fail_free_reader;
1328
1329	cpu_buffer->head_page
1330		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1331	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1332
1333	rb_head_page_activate(cpu_buffer);
1334
1335	return cpu_buffer;
1336
1337 fail_free_reader:
1338	free_buffer_page(cpu_buffer->reader_page);
1339
1340 fail_free_buffer:
1341	kfree(cpu_buffer);
1342	return NULL;
1343}
1344
1345static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1346{
1347	struct list_head *head = cpu_buffer->pages;
1348	struct buffer_page *bpage, *tmp;
1349
1350	free_buffer_page(cpu_buffer->reader_page);
1351
1352	rb_head_page_deactivate(cpu_buffer);
1353
1354	if (head) {
1355		list_for_each_entry_safe(bpage, tmp, head, list) {
1356			list_del_init(&bpage->list);
1357			free_buffer_page(bpage);
1358		}
1359		bpage = list_entry(head, struct buffer_page, list);
1360		free_buffer_page(bpage);
1361	}
1362
1363	kfree(cpu_buffer);
1364}
1365
1366/**
1367 * __ring_buffer_alloc - allocate a new ring_buffer
1368 * @size: the size in bytes per cpu that is needed.
1369 * @flags: attributes to set for the ring buffer.
1370 *
1371 * Currently the only flag that is available is the RB_FL_OVERWRITE
1372 * flag. This flag means that the buffer will overwrite old data
1373 * when the buffer wraps. If this flag is not set, the buffer will
1374 * drop data when the tail hits the head.
1375 */
1376struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1377					struct lock_class_key *key)
1378{
1379	struct ring_buffer *buffer;
1380	long nr_pages;
1381	int bsize;
1382	int cpu;
1383	int ret;
1384
1385	/* keep it in its own cache line */
1386	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1387			 GFP_KERNEL);
1388	if (!buffer)
1389		return NULL;
1390
1391	if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1392		goto fail_free_buffer;
1393
1394	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1395	buffer->flags = flags;
1396	buffer->clock = trace_clock_local;
1397	buffer->reader_lock_key = key;
1398
1399	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1400	init_waitqueue_head(&buffer->irq_work.waiters);
1401
1402	/* need at least two pages */
1403	if (nr_pages < 2)
1404		nr_pages = 2;
1405
1406	buffer->cpus = nr_cpu_ids;
1407
1408	bsize = sizeof(void *) * nr_cpu_ids;
1409	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1410				  GFP_KERNEL);
1411	if (!buffer->buffers)
1412		goto fail_free_cpumask;
1413
1414	cpu = raw_smp_processor_id();
1415	cpumask_set_cpu(cpu, buffer->cpumask);
1416	buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1417	if (!buffer->buffers[cpu])
1418		goto fail_free_buffers;
1419
1420	ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1421	if (ret < 0)
1422		goto fail_free_buffers;
1423
1424	mutex_init(&buffer->mutex);
1425
1426	return buffer;
1427
1428 fail_free_buffers:
1429	for_each_buffer_cpu(buffer, cpu) {
1430		if (buffer->buffers[cpu])
1431			rb_free_cpu_buffer(buffer->buffers[cpu]);
1432	}
1433	kfree(buffer->buffers);
1434
1435 fail_free_cpumask:
1436	free_cpumask_var(buffer->cpumask);
1437
1438 fail_free_buffer:
1439	kfree(buffer);
1440	return NULL;
1441}
1442EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1443
1444/**
1445 * ring_buffer_free - free a ring buffer.
1446 * @buffer: the buffer to free.
1447 */
1448void
1449ring_buffer_free(struct ring_buffer *buffer)
1450{
1451	int cpu;
1452
1453	cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1454
1455	for_each_buffer_cpu(buffer, cpu)
1456		rb_free_cpu_buffer(buffer->buffers[cpu]);
1457
1458	kfree(buffer->buffers);
1459	free_cpumask_var(buffer->cpumask);
1460
1461	kfree(buffer);
1462}
1463EXPORT_SYMBOL_GPL(ring_buffer_free);
1464
1465void ring_buffer_set_clock(struct ring_buffer *buffer,
1466			   u64 (*clock)(void))
1467{
1468	buffer->clock = clock;
1469}
1470
1471void ring_buffer_set_time_stamp_abs(struct ring_buffer *buffer, bool abs)
1472{
1473	buffer->time_stamp_abs = abs;
1474}
1475
1476bool ring_buffer_time_stamp_abs(struct ring_buffer *buffer)
1477{
1478	return buffer->time_stamp_abs;
1479}
1480
1481static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1482
1483static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1484{
1485	return local_read(&bpage->entries) & RB_WRITE_MASK;
1486}
1487
1488static inline unsigned long rb_page_write(struct buffer_page *bpage)
1489{
1490	return local_read(&bpage->write) & RB_WRITE_MASK;
1491}
1492
1493static int
1494rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1495{
1496	struct list_head *tail_page, *to_remove, *next_page;
1497	struct buffer_page *to_remove_page, *tmp_iter_page;
1498	struct buffer_page *last_page, *first_page;
1499	unsigned long nr_removed;
1500	unsigned long head_bit;
1501	int page_entries;
1502
1503	head_bit = 0;
1504
1505	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1506	atomic_inc(&cpu_buffer->record_disabled);
1507	/*
1508	 * We don't race with the readers since we have acquired the reader
1509	 * lock. We also don't race with writers after disabling recording.
1510	 * This makes it easy to figure out the first and the last page to be
1511	 * removed from the list. We unlink all the pages in between including
1512	 * the first and last pages. This is done in a busy loop so that we
1513	 * lose the least number of traces.
1514	 * The pages are freed after we restart recording and unlock readers.
1515	 */
1516	tail_page = &cpu_buffer->tail_page->list;
1517
1518	/*
1519	 * tail page might be on reader page, we remove the next page
1520	 * from the ring buffer
1521	 */
1522	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1523		tail_page = rb_list_head(tail_page->next);
1524	to_remove = tail_page;
1525
1526	/* start of pages to remove */
1527	first_page = list_entry(rb_list_head(to_remove->next),
1528				struct buffer_page, list);
1529
1530	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1531		to_remove = rb_list_head(to_remove)->next;
1532		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1533	}
1534
1535	next_page = rb_list_head(to_remove)->next;
1536
1537	/*
1538	 * Now we remove all pages between tail_page and next_page.
1539	 * Make sure that we have head_bit value preserved for the
1540	 * next page
1541	 */
1542	tail_page->next = (struct list_head *)((unsigned long)next_page |
1543						head_bit);
1544	next_page = rb_list_head(next_page);
1545	next_page->prev = tail_page;
1546
1547	/* make sure pages points to a valid page in the ring buffer */
1548	cpu_buffer->pages = next_page;
1549
1550	/* update head page */
1551	if (head_bit)
1552		cpu_buffer->head_page = list_entry(next_page,
1553						struct buffer_page, list);
1554
1555	/*
1556	 * change read pointer to make sure any read iterators reset
1557	 * themselves
1558	 */
1559	cpu_buffer->read = 0;
1560
1561	/* pages are removed, resume tracing and then free the pages */
1562	atomic_dec(&cpu_buffer->record_disabled);
1563	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1564
1565	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1566
1567	/* last buffer page to remove */
1568	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1569				list);
1570	tmp_iter_page = first_page;
1571
1572	do {
1573		cond_resched();
1574
1575		to_remove_page = tmp_iter_page;
1576		rb_inc_page(cpu_buffer, &tmp_iter_page);
1577
1578		/* update the counters */
1579		page_entries = rb_page_entries(to_remove_page);
1580		if (page_entries) {
1581			/*
1582			 * If something was added to this page, it was full
1583			 * since it is not the tail page. So we deduct the
1584			 * bytes consumed in ring buffer from here.
1585			 * Increment overrun to account for the lost events.
1586			 */
1587			local_add(page_entries, &cpu_buffer->overrun);
1588			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1589		}
1590
1591		/*
1592		 * We have already removed references to this list item, just
1593		 * free up the buffer_page and its page
1594		 */
1595		free_buffer_page(to_remove_page);
1596		nr_removed--;
1597
1598	} while (to_remove_page != last_page);
1599
1600	RB_WARN_ON(cpu_buffer, nr_removed);
1601
1602	return nr_removed == 0;
1603}
1604
1605static int
1606rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1607{
1608	struct list_head *pages = &cpu_buffer->new_pages;
1609	int retries, success;
1610
1611	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1612	/*
1613	 * We are holding the reader lock, so the reader page won't be swapped
1614	 * in the ring buffer. Now we are racing with the writer trying to
1615	 * move head page and the tail page.
1616	 * We are going to adapt the reader page update process where:
1617	 * 1. We first splice the start and end of list of new pages between
1618	 *    the head page and its previous page.
1619	 * 2. We cmpxchg the prev_page->next to point from head page to the
1620	 *    start of new pages list.
1621	 * 3. Finally, we update the head->prev to the end of new list.
1622	 *
1623	 * We will try this process 10 times, to make sure that we don't keep
1624	 * spinning.
1625	 */
1626	retries = 10;
1627	success = 0;
1628	while (retries--) {
1629		struct list_head *head_page, *prev_page, *r;
1630		struct list_head *last_page, *first_page;
1631		struct list_head *head_page_with_bit;
1632
1633		head_page = &rb_set_head_page(cpu_buffer)->list;
1634		if (!head_page)
1635			break;
1636		prev_page = head_page->prev;
1637
1638		first_page = pages->next;
1639		last_page  = pages->prev;
1640
1641		head_page_with_bit = (struct list_head *)
1642				     ((unsigned long)head_page | RB_PAGE_HEAD);
1643
1644		last_page->next = head_page_with_bit;
1645		first_page->prev = prev_page;
1646
1647		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1648
1649		if (r == head_page_with_bit) {
1650			/*
1651			 * yay, we replaced the page pointer to our new list,
1652			 * now, we just have to update to head page's prev
1653			 * pointer to point to end of list
1654			 */
1655			head_page->prev = last_page;
1656			success = 1;
1657			break;
1658		}
1659	}
1660
1661	if (success)
1662		INIT_LIST_HEAD(pages);
1663	/*
1664	 * If we weren't successful in adding in new pages, warn and stop
1665	 * tracing
1666	 */
1667	RB_WARN_ON(cpu_buffer, !success);
1668	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1669
1670	/* free pages if they weren't inserted */
1671	if (!success) {
1672		struct buffer_page *bpage, *tmp;
1673		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1674					 list) {
1675			list_del_init(&bpage->list);
1676			free_buffer_page(bpage);
1677		}
1678	}
1679	return success;
1680}
1681
1682static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1683{
1684	int success;
1685
1686	if (cpu_buffer->nr_pages_to_update > 0)
1687		success = rb_insert_pages(cpu_buffer);
1688	else
1689		success = rb_remove_pages(cpu_buffer,
1690					-cpu_buffer->nr_pages_to_update);
1691
1692	if (success)
1693		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1694}
1695
1696static void update_pages_handler(struct work_struct *work)
1697{
1698	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1699			struct ring_buffer_per_cpu, update_pages_work);
1700	rb_update_pages(cpu_buffer);
1701	complete(&cpu_buffer->update_done);
1702}
1703
1704/**
1705 * ring_buffer_resize - resize the ring buffer
1706 * @buffer: the buffer to resize.
1707 * @size: the new size.
1708 * @cpu_id: the cpu buffer to resize
1709 *
1710 * Minimum size is 2 * BUF_PAGE_SIZE.
1711 *
1712 * Returns 0 on success and < 0 on failure.
1713 */
1714int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1715			int cpu_id)
1716{
1717	struct ring_buffer_per_cpu *cpu_buffer;
1718	unsigned long nr_pages;
1719	int cpu, err = 0;
1720
1721	/*
1722	 * Always succeed at resizing a non-existent buffer:
1723	 */
1724	if (!buffer)
1725		return size;
1726
1727	/* Make sure the requested buffer exists */
1728	if (cpu_id != RING_BUFFER_ALL_CPUS &&
1729	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
1730		return size;
1731
1732	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1733
1734	/* we need a minimum of two pages */
1735	if (nr_pages < 2)
1736		nr_pages = 2;
1737
1738	size = nr_pages * BUF_PAGE_SIZE;
1739
1740	/*
1741	 * Don't succeed if resizing is disabled, as a reader might be
1742	 * manipulating the ring buffer and is expecting a sane state while
1743	 * this is true.
1744	 */
1745	if (atomic_read(&buffer->resize_disabled))
1746		return -EBUSY;
1747
1748	/* prevent another thread from changing buffer sizes */
1749	mutex_lock(&buffer->mutex);
1750
1751	if (cpu_id == RING_BUFFER_ALL_CPUS) {
1752		/* calculate the pages to update */
1753		for_each_buffer_cpu(buffer, cpu) {
1754			cpu_buffer = buffer->buffers[cpu];
1755
1756			cpu_buffer->nr_pages_to_update = nr_pages -
1757							cpu_buffer->nr_pages;
1758			/*
1759			 * nothing more to do for removing pages or no update
1760			 */
1761			if (cpu_buffer->nr_pages_to_update <= 0)
1762				continue;
1763			/*
1764			 * to add pages, make sure all new pages can be
1765			 * allocated without receiving ENOMEM
1766			 */
1767			INIT_LIST_HEAD(&cpu_buffer->new_pages);
1768			if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1769						&cpu_buffer->new_pages, cpu)) {
1770				/* not enough memory for new pages */
1771				err = -ENOMEM;
1772				goto out_err;
1773			}
1774		}
1775
1776		get_online_cpus();
1777		/*
1778		 * Fire off all the required work handlers
1779		 * We can't schedule on offline CPUs, but it's not necessary
1780		 * since we can change their buffer sizes without any race.
1781		 */
1782		for_each_buffer_cpu(buffer, cpu) {
1783			cpu_buffer = buffer->buffers[cpu];
1784			if (!cpu_buffer->nr_pages_to_update)
1785				continue;
1786
1787			/* Can't run something on an offline CPU. */
1788			if (!cpu_online(cpu)) {
1789				rb_update_pages(cpu_buffer);
1790				cpu_buffer->nr_pages_to_update = 0;
1791			} else {
1792				schedule_work_on(cpu,
1793						&cpu_buffer->update_pages_work);
1794			}
1795		}
1796
1797		/* wait for all the updates to complete */
1798		for_each_buffer_cpu(buffer, cpu) {
1799			cpu_buffer = buffer->buffers[cpu];
1800			if (!cpu_buffer->nr_pages_to_update)
1801				continue;
1802
1803			if (cpu_online(cpu))
1804				wait_for_completion(&cpu_buffer->update_done);
1805			cpu_buffer->nr_pages_to_update = 0;
1806		}
1807
1808		put_online_cpus();
1809	} else {
1810		/* Make sure this CPU has been initialized */
1811		if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1812			goto out;
1813
1814		cpu_buffer = buffer->buffers[cpu_id];
1815
1816		if (nr_pages == cpu_buffer->nr_pages)
1817			goto out;
1818
1819		cpu_buffer->nr_pages_to_update = nr_pages -
1820						cpu_buffer->nr_pages;
1821
1822		INIT_LIST_HEAD(&cpu_buffer->new_pages);
1823		if (cpu_buffer->nr_pages_to_update > 0 &&
1824			__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1825					    &cpu_buffer->new_pages, cpu_id)) {
1826			err = -ENOMEM;
1827			goto out_err;
1828		}
1829
1830		get_online_cpus();
1831
1832		/* Can't run something on an offline CPU. */
1833		if (!cpu_online(cpu_id))
1834			rb_update_pages(cpu_buffer);
1835		else {
1836			schedule_work_on(cpu_id,
1837					 &cpu_buffer->update_pages_work);
1838			wait_for_completion(&cpu_buffer->update_done);
1839		}
1840
1841		cpu_buffer->nr_pages_to_update = 0;
1842		put_online_cpus();
1843	}
1844
1845 out:
1846	/*
1847	 * The ring buffer resize can happen with the ring buffer
1848	 * enabled, so that the update disturbs the tracing as little
1849	 * as possible. But if the buffer is disabled, we do not need
1850	 * to worry about that, and we can take the time to verify
1851	 * that the buffer is not corrupt.
1852	 */
1853	if (atomic_read(&buffer->record_disabled)) {
1854		atomic_inc(&buffer->record_disabled);
1855		/*
1856		 * Even though the buffer was disabled, we must make sure
1857		 * that it is truly disabled before calling rb_check_pages.
1858		 * There could have been a race between checking
1859		 * record_disable and incrementing it.
1860		 */
1861		synchronize_rcu();
1862		for_each_buffer_cpu(buffer, cpu) {
1863			cpu_buffer = buffer->buffers[cpu];
1864			rb_check_pages(cpu_buffer);
1865		}
1866		atomic_dec(&buffer->record_disabled);
1867	}
1868
1869	mutex_unlock(&buffer->mutex);
1870	return size;
1871
1872 out_err:
1873	for_each_buffer_cpu(buffer, cpu) {
1874		struct buffer_page *bpage, *tmp;
1875
1876		cpu_buffer = buffer->buffers[cpu];
1877		cpu_buffer->nr_pages_to_update = 0;
1878
1879		if (list_empty(&cpu_buffer->new_pages))
1880			continue;
1881
1882		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1883					list) {
1884			list_del_init(&bpage->list);
1885			free_buffer_page(bpage);
1886		}
1887	}
1888	mutex_unlock(&buffer->mutex);
1889	return err;
1890}
1891EXPORT_SYMBOL_GPL(ring_buffer_resize);
1892
1893void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1894{
1895	mutex_lock(&buffer->mutex);
1896	if (val)
1897		buffer->flags |= RB_FL_OVERWRITE;
1898	else
1899		buffer->flags &= ~RB_FL_OVERWRITE;
1900	mutex_unlock(&buffer->mutex);
1901}
1902EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1903
 
 
 
 
 
 
1904static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1905{
1906	return bpage->page->data + index;
1907}
1908
1909static __always_inline struct ring_buffer_event *
1910rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1911{
1912	return __rb_page_index(cpu_buffer->reader_page,
1913			       cpu_buffer->reader_page->read);
1914}
1915
1916static __always_inline struct ring_buffer_event *
1917rb_iter_head_event(struct ring_buffer_iter *iter)
1918{
1919	return __rb_page_index(iter->head_page, iter->head);
1920}
1921
1922static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1923{
1924	return local_read(&bpage->page->commit);
1925}
1926
1927/* Size is determined by what has been committed */
1928static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1929{
1930	return rb_page_commit(bpage);
1931}
1932
1933static __always_inline unsigned
1934rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1935{
1936	return rb_page_commit(cpu_buffer->commit_page);
1937}
1938
1939static __always_inline unsigned
1940rb_event_index(struct ring_buffer_event *event)
1941{
1942	unsigned long addr = (unsigned long)event;
1943
1944	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1945}
1946
1947static void rb_inc_iter(struct ring_buffer_iter *iter)
1948{
1949	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1950
1951	/*
1952	 * The iterator could be on the reader page (it starts there).
1953	 * But the head could have moved, since the reader was
1954	 * found. Check for this case and assign the iterator
1955	 * to the head page instead of next.
1956	 */
1957	if (iter->head_page == cpu_buffer->reader_page)
1958		iter->head_page = rb_set_head_page(cpu_buffer);
1959	else
1960		rb_inc_page(cpu_buffer, &iter->head_page);
1961
1962	iter->read_stamp = iter->head_page->page->time_stamp;
1963	iter->head = 0;
1964}
1965
1966/*
1967 * rb_handle_head_page - writer hit the head page
1968 *
1969 * Returns: +1 to retry page
1970 *           0 to continue
1971 *          -1 on error
1972 */
1973static int
1974rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1975		    struct buffer_page *tail_page,
1976		    struct buffer_page *next_page)
1977{
1978	struct buffer_page *new_head;
1979	int entries;
1980	int type;
1981	int ret;
1982
1983	entries = rb_page_entries(next_page);
1984
1985	/*
1986	 * The hard part is here. We need to move the head
1987	 * forward, and protect against both readers on
1988	 * other CPUs and writers coming in via interrupts.
1989	 */
1990	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1991				       RB_PAGE_HEAD);
1992
1993	/*
1994	 * type can be one of four:
1995	 *  NORMAL - an interrupt already moved it for us
1996	 *  HEAD   - we are the first to get here.
1997	 *  UPDATE - we are the interrupt interrupting
1998	 *           a current move.
1999	 *  MOVED  - a reader on another CPU moved the next
2000	 *           pointer to its reader page. Give up
2001	 *           and try again.
2002	 */
2003
2004	switch (type) {
2005	case RB_PAGE_HEAD:
2006		/*
2007		 * We changed the head to UPDATE, thus
2008		 * it is our responsibility to update
2009		 * the counters.
2010		 */
2011		local_add(entries, &cpu_buffer->overrun);
2012		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2013
2014		/*
2015		 * The entries will be zeroed out when we move the
2016		 * tail page.
2017		 */
2018
2019		/* still more to do */
2020		break;
2021
2022	case RB_PAGE_UPDATE:
2023		/*
2024		 * This is an interrupt that interrupt the
2025		 * previous update. Still more to do.
2026		 */
2027		break;
2028	case RB_PAGE_NORMAL:
2029		/*
2030		 * An interrupt came in before the update
2031		 * and processed this for us.
2032		 * Nothing left to do.
2033		 */
2034		return 1;
2035	case RB_PAGE_MOVED:
2036		/*
2037		 * The reader is on another CPU and just did
2038		 * a swap with our next_page.
2039		 * Try again.
2040		 */
2041		return 1;
2042	default:
2043		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2044		return -1;
2045	}
2046
2047	/*
2048	 * Now that we are here, the old head pointer is
2049	 * set to UPDATE. This will keep the reader from
2050	 * swapping the head page with the reader page.
2051	 * The reader (on another CPU) will spin till
2052	 * we are finished.
2053	 *
2054	 * We just need to protect against interrupts
2055	 * doing the job. We will set the next pointer
2056	 * to HEAD. After that, we set the old pointer
2057	 * to NORMAL, but only if it was HEAD before.
2058	 * otherwise we are an interrupt, and only
2059	 * want the outer most commit to reset it.
2060	 */
2061	new_head = next_page;
2062	rb_inc_page(cpu_buffer, &new_head);
2063
2064	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2065				    RB_PAGE_NORMAL);
2066
2067	/*
2068	 * Valid returns are:
2069	 *  HEAD   - an interrupt came in and already set it.
2070	 *  NORMAL - One of two things:
2071	 *            1) We really set it.
2072	 *            2) A bunch of interrupts came in and moved
2073	 *               the page forward again.
2074	 */
2075	switch (ret) {
2076	case RB_PAGE_HEAD:
2077	case RB_PAGE_NORMAL:
2078		/* OK */
2079		break;
2080	default:
2081		RB_WARN_ON(cpu_buffer, 1);
2082		return -1;
2083	}
2084
2085	/*
2086	 * It is possible that an interrupt came in,
2087	 * set the head up, then more interrupts came in
2088	 * and moved it again. When we get back here,
2089	 * the page would have been set to NORMAL but we
2090	 * just set it back to HEAD.
2091	 *
2092	 * How do you detect this? Well, if that happened
2093	 * the tail page would have moved.
2094	 */
2095	if (ret == RB_PAGE_NORMAL) {
2096		struct buffer_page *buffer_tail_page;
2097
2098		buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2099		/*
2100		 * If the tail had moved passed next, then we need
2101		 * to reset the pointer.
2102		 */
2103		if (buffer_tail_page != tail_page &&
2104		    buffer_tail_page != next_page)
2105			rb_head_page_set_normal(cpu_buffer, new_head,
2106						next_page,
2107						RB_PAGE_HEAD);
2108	}
2109
2110	/*
2111	 * If this was the outer most commit (the one that
2112	 * changed the original pointer from HEAD to UPDATE),
2113	 * then it is up to us to reset it to NORMAL.
2114	 */
2115	if (type == RB_PAGE_HEAD) {
2116		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2117					      tail_page,
2118					      RB_PAGE_UPDATE);
2119		if (RB_WARN_ON(cpu_buffer,
2120			       ret != RB_PAGE_UPDATE))
2121			return -1;
2122	}
2123
2124	return 0;
2125}
2126
2127static inline void
2128rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2129	      unsigned long tail, struct rb_event_info *info)
2130{
2131	struct buffer_page *tail_page = info->tail_page;
2132	struct ring_buffer_event *event;
2133	unsigned long length = info->length;
2134
2135	/*
2136	 * Only the event that crossed the page boundary
2137	 * must fill the old tail_page with padding.
2138	 */
2139	if (tail >= BUF_PAGE_SIZE) {
2140		/*
2141		 * If the page was filled, then we still need
2142		 * to update the real_end. Reset it to zero
2143		 * and the reader will ignore it.
2144		 */
2145		if (tail == BUF_PAGE_SIZE)
2146			tail_page->real_end = 0;
2147
2148		local_sub(length, &tail_page->write);
2149		return;
2150	}
2151
2152	event = __rb_page_index(tail_page, tail);
 
2153
2154	/* account for padding bytes */
2155	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2156
2157	/*
2158	 * Save the original length to the meta data.
2159	 * This will be used by the reader to add lost event
2160	 * counter.
2161	 */
2162	tail_page->real_end = tail;
2163
2164	/*
2165	 * If this event is bigger than the minimum size, then
2166	 * we need to be careful that we don't subtract the
2167	 * write counter enough to allow another writer to slip
2168	 * in on this page.
2169	 * We put in a discarded commit instead, to make sure
2170	 * that this space is not used again.
2171	 *
2172	 * If we are less than the minimum size, we don't need to
2173	 * worry about it.
2174	 */
2175	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2176		/* No room for any events */
2177
2178		/* Mark the rest of the page with padding */
2179		rb_event_set_padding(event);
2180
2181		/* Set the write back to the previous setting */
2182		local_sub(length, &tail_page->write);
2183		return;
2184	}
2185
2186	/* Put in a discarded event */
2187	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2188	event->type_len = RINGBUF_TYPE_PADDING;
2189	/* time delta must be non zero */
2190	event->time_delta = 1;
2191
2192	/* Set write to end of buffer */
2193	length = (tail + length) - BUF_PAGE_SIZE;
2194	local_sub(length, &tail_page->write);
2195}
2196
2197static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2198
2199/*
2200 * This is the slow path, force gcc not to inline it.
2201 */
2202static noinline struct ring_buffer_event *
2203rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2204	     unsigned long tail, struct rb_event_info *info)
2205{
2206	struct buffer_page *tail_page = info->tail_page;
2207	struct buffer_page *commit_page = cpu_buffer->commit_page;
2208	struct ring_buffer *buffer = cpu_buffer->buffer;
2209	struct buffer_page *next_page;
2210	int ret;
2211
2212	next_page = tail_page;
2213
2214	rb_inc_page(cpu_buffer, &next_page);
2215
2216	/*
2217	 * If for some reason, we had an interrupt storm that made
2218	 * it all the way around the buffer, bail, and warn
2219	 * about it.
2220	 */
2221	if (unlikely(next_page == commit_page)) {
2222		local_inc(&cpu_buffer->commit_overrun);
2223		goto out_reset;
2224	}
2225
2226	/*
2227	 * This is where the fun begins!
2228	 *
2229	 * We are fighting against races between a reader that
2230	 * could be on another CPU trying to swap its reader
2231	 * page with the buffer head.
2232	 *
2233	 * We are also fighting against interrupts coming in and
2234	 * moving the head or tail on us as well.
2235	 *
2236	 * If the next page is the head page then we have filled
2237	 * the buffer, unless the commit page is still on the
2238	 * reader page.
2239	 */
2240	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2241
2242		/*
2243		 * If the commit is not on the reader page, then
2244		 * move the header page.
2245		 */
2246		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2247			/*
2248			 * If we are not in overwrite mode,
2249			 * this is easy, just stop here.
2250			 */
2251			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2252				local_inc(&cpu_buffer->dropped_events);
2253				goto out_reset;
2254			}
2255
2256			ret = rb_handle_head_page(cpu_buffer,
2257						  tail_page,
2258						  next_page);
2259			if (ret < 0)
2260				goto out_reset;
2261			if (ret)
2262				goto out_again;
2263		} else {
2264			/*
2265			 * We need to be careful here too. The
2266			 * commit page could still be on the reader
2267			 * page. We could have a small buffer, and
2268			 * have filled up the buffer with events
2269			 * from interrupts and such, and wrapped.
2270			 *
2271			 * Note, if the tail page is also the on the
2272			 * reader_page, we let it move out.
2273			 */
2274			if (unlikely((cpu_buffer->commit_page !=
2275				      cpu_buffer->tail_page) &&
2276				     (cpu_buffer->commit_page ==
2277				      cpu_buffer->reader_page))) {
2278				local_inc(&cpu_buffer->commit_overrun);
2279				goto out_reset;
2280			}
2281		}
2282	}
2283
2284	rb_tail_page_update(cpu_buffer, tail_page, next_page);
2285
2286 out_again:
2287
2288	rb_reset_tail(cpu_buffer, tail, info);
2289
2290	/* Commit what we have for now. */
2291	rb_end_commit(cpu_buffer);
2292	/* rb_end_commit() decs committing */
2293	local_inc(&cpu_buffer->committing);
2294
2295	/* fail and let the caller try again */
2296	return ERR_PTR(-EAGAIN);
2297
2298 out_reset:
2299	/* reset write */
2300	rb_reset_tail(cpu_buffer, tail, info);
2301
2302	return NULL;
2303}
2304
2305/* Slow path, do not inline */
2306static noinline struct ring_buffer_event *
2307rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2308{
2309	if (abs)
2310		event->type_len = RINGBUF_TYPE_TIME_STAMP;
2311	else
2312		event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2313
2314	/* Not the first event on the page, or not delta? */
2315	if (abs || rb_event_index(event)) {
2316		event->time_delta = delta & TS_MASK;
2317		event->array[0] = delta >> TS_SHIFT;
2318	} else {
2319		/* nope, just zero it */
2320		event->time_delta = 0;
2321		event->array[0] = 0;
2322	}
2323
2324	return skip_time_extend(event);
2325}
2326
2327static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2328				     struct ring_buffer_event *event);
2329
2330/**
2331 * rb_update_event - update event type and data
2332 * @event: the event to update
2333 * @type: the type of event
2334 * @length: the size of the event field in the ring buffer
2335 *
2336 * Update the type and data fields of the event. The length
2337 * is the actual size that is written to the ring buffer,
2338 * and with this, we can determine what to place into the
2339 * data field.
2340 */
2341static void
2342rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2343		struct ring_buffer_event *event,
2344		struct rb_event_info *info)
2345{
2346	unsigned length = info->length;
2347	u64 delta = info->delta;
2348
2349	/* Only a commit updates the timestamp */
2350	if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2351		delta = 0;
2352
2353	/*
2354	 * If we need to add a timestamp, then we
2355	 * add it to the start of the reserved space.
2356	 */
2357	if (unlikely(info->add_timestamp)) {
2358		bool abs = ring_buffer_time_stamp_abs(cpu_buffer->buffer);
2359
2360		event = rb_add_time_stamp(event, info->delta, abs);
2361		length -= RB_LEN_TIME_EXTEND;
2362		delta = 0;
2363	}
2364
2365	event->time_delta = delta;
2366	length -= RB_EVNT_HDR_SIZE;
2367	if (length > RB_MAX_SMALL_DATA) {
2368		event->type_len = 0;
2369		event->array[0] = length;
2370	} else
2371		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2372}
2373
2374static unsigned rb_calculate_event_length(unsigned length)
2375{
2376	struct ring_buffer_event event; /* Used only for sizeof array */
2377
2378	/* zero length can cause confusions */
2379	if (!length)
2380		length++;
2381
2382	if (length > RB_MAX_SMALL_DATA)
2383		length += sizeof(event.array[0]);
2384
2385	length += RB_EVNT_HDR_SIZE;
2386	length = ALIGN(length, RB_ALIGNMENT);
2387
2388	/*
2389	 * In case the time delta is larger than the 27 bits for it
2390	 * in the header, we need to add a timestamp. If another
2391	 * event comes in when trying to discard this one to increase
2392	 * the length, then the timestamp will be added in the allocated
2393	 * space of this event. If length is bigger than the size needed
2394	 * for the TIME_EXTEND, then padding has to be used. The events
2395	 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2396	 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2397	 * As length is a multiple of 4, we only need to worry if it
2398	 * is 12 (RB_LEN_TIME_EXTEND + 4).
2399	 */
2400	if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2401		length += RB_ALIGNMENT;
2402
2403	return length;
2404}
2405
2406#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2407static inline bool sched_clock_stable(void)
2408{
2409	return true;
2410}
2411#endif
2412
2413static inline int
2414rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2415		  struct ring_buffer_event *event)
2416{
2417	unsigned long new_index, old_index;
2418	struct buffer_page *bpage;
2419	unsigned long index;
2420	unsigned long addr;
2421
2422	new_index = rb_event_index(event);
2423	old_index = new_index + rb_event_ts_length(event);
2424	addr = (unsigned long)event;
2425	addr &= PAGE_MASK;
2426
2427	bpage = READ_ONCE(cpu_buffer->tail_page);
2428
2429	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2430		unsigned long write_mask =
2431			local_read(&bpage->write) & ~RB_WRITE_MASK;
2432		unsigned long event_length = rb_event_length(event);
2433		/*
2434		 * This is on the tail page. It is possible that
2435		 * a write could come in and move the tail page
2436		 * and write to the next page. That is fine
2437		 * because we just shorten what is on this page.
2438		 */
2439		old_index += write_mask;
2440		new_index += write_mask;
2441		index = local_cmpxchg(&bpage->write, old_index, new_index);
2442		if (index == old_index) {
2443			/* update counters */
2444			local_sub(event_length, &cpu_buffer->entries_bytes);
2445			return 1;
2446		}
2447	}
2448
2449	/* could not discard */
2450	return 0;
2451}
2452
2453static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2454{
2455	local_inc(&cpu_buffer->committing);
2456	local_inc(&cpu_buffer->commits);
2457}
2458
2459static __always_inline void
2460rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2461{
2462	unsigned long max_count;
2463
2464	/*
2465	 * We only race with interrupts and NMIs on this CPU.
2466	 * If we own the commit event, then we can commit
2467	 * all others that interrupted us, since the interruptions
2468	 * are in stack format (they finish before they come
2469	 * back to us). This allows us to do a simple loop to
2470	 * assign the commit to the tail.
2471	 */
2472 again:
2473	max_count = cpu_buffer->nr_pages * 100;
2474
2475	while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2476		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2477			return;
2478		if (RB_WARN_ON(cpu_buffer,
2479			       rb_is_reader_page(cpu_buffer->tail_page)))
2480			return;
2481		local_set(&cpu_buffer->commit_page->page->commit,
2482			  rb_page_write(cpu_buffer->commit_page));
2483		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2484		/* Only update the write stamp if the page has an event */
2485		if (rb_page_write(cpu_buffer->commit_page))
2486			cpu_buffer->write_stamp =
2487				cpu_buffer->commit_page->page->time_stamp;
2488		/* add barrier to keep gcc from optimizing too much */
2489		barrier();
2490	}
2491	while (rb_commit_index(cpu_buffer) !=
2492	       rb_page_write(cpu_buffer->commit_page)) {
2493
2494		local_set(&cpu_buffer->commit_page->page->commit,
2495			  rb_page_write(cpu_buffer->commit_page));
2496		RB_WARN_ON(cpu_buffer,
2497			   local_read(&cpu_buffer->commit_page->page->commit) &
2498			   ~RB_WRITE_MASK);
2499		barrier();
2500	}
2501
2502	/* again, keep gcc from optimizing */
2503	barrier();
2504
2505	/*
2506	 * If an interrupt came in just after the first while loop
2507	 * and pushed the tail page forward, we will be left with
2508	 * a dangling commit that will never go forward.
2509	 */
2510	if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2511		goto again;
2512}
2513
2514static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2515{
2516	unsigned long commits;
2517
2518	if (RB_WARN_ON(cpu_buffer,
2519		       !local_read(&cpu_buffer->committing)))
2520		return;
2521
2522 again:
2523	commits = local_read(&cpu_buffer->commits);
2524	/* synchronize with interrupts */
2525	barrier();
2526	if (local_read(&cpu_buffer->committing) == 1)
2527		rb_set_commit_to_write(cpu_buffer);
2528
2529	local_dec(&cpu_buffer->committing);
2530
2531	/* synchronize with interrupts */
2532	barrier();
2533
2534	/*
2535	 * Need to account for interrupts coming in between the
2536	 * updating of the commit page and the clearing of the
2537	 * committing counter.
2538	 */
2539	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2540	    !local_read(&cpu_buffer->committing)) {
2541		local_inc(&cpu_buffer->committing);
2542		goto again;
2543	}
2544}
2545
2546static inline void rb_event_discard(struct ring_buffer_event *event)
2547{
2548	if (extended_time(event))
2549		event = skip_time_extend(event);
2550
2551	/* array[0] holds the actual length for the discarded event */
2552	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2553	event->type_len = RINGBUF_TYPE_PADDING;
2554	/* time delta must be non zero */
2555	if (!event->time_delta)
2556		event->time_delta = 1;
2557}
2558
2559static __always_inline bool
2560rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2561		   struct ring_buffer_event *event)
2562{
2563	unsigned long addr = (unsigned long)event;
2564	unsigned long index;
2565
2566	index = rb_event_index(event);
2567	addr &= PAGE_MASK;
2568
2569	return cpu_buffer->commit_page->page == (void *)addr &&
2570		rb_commit_index(cpu_buffer) == index;
2571}
2572
2573static __always_inline void
2574rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2575		      struct ring_buffer_event *event)
2576{
2577	u64 delta;
2578
2579	/*
2580	 * The event first in the commit queue updates the
2581	 * time stamp.
2582	 */
2583	if (rb_event_is_commit(cpu_buffer, event)) {
2584		/*
2585		 * A commit event that is first on a page
2586		 * updates the write timestamp with the page stamp
2587		 */
2588		if (!rb_event_index(event))
2589			cpu_buffer->write_stamp =
2590				cpu_buffer->commit_page->page->time_stamp;
2591		else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2592			delta = ring_buffer_event_time_stamp(event);
 
 
2593			cpu_buffer->write_stamp += delta;
2594		} else if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
2595			delta = ring_buffer_event_time_stamp(event);
2596			cpu_buffer->write_stamp = delta;
2597		} else
2598			cpu_buffer->write_stamp += event->time_delta;
2599	}
2600}
2601
2602static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2603		      struct ring_buffer_event *event)
2604{
2605	local_inc(&cpu_buffer->entries);
2606	rb_update_write_stamp(cpu_buffer, event);
2607	rb_end_commit(cpu_buffer);
2608}
2609
2610static __always_inline void
2611rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2612{
2613	size_t nr_pages;
2614	size_t dirty;
2615	size_t full;
2616
2617	if (buffer->irq_work.waiters_pending) {
2618		buffer->irq_work.waiters_pending = false;
2619		/* irq_work_queue() supplies it's own memory barriers */
2620		irq_work_queue(&buffer->irq_work.work);
2621	}
2622
2623	if (cpu_buffer->irq_work.waiters_pending) {
2624		cpu_buffer->irq_work.waiters_pending = false;
2625		/* irq_work_queue() supplies it's own memory barriers */
2626		irq_work_queue(&cpu_buffer->irq_work.work);
2627	}
2628
2629	if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
2630		return;
2631
2632	if (cpu_buffer->reader_page == cpu_buffer->commit_page)
2633		return;
2634
2635	if (!cpu_buffer->irq_work.full_waiters_pending)
2636		return;
2637
2638	cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
2639
2640	full = cpu_buffer->shortest_full;
2641	nr_pages = cpu_buffer->nr_pages;
2642	dirty = ring_buffer_nr_dirty_pages(buffer, cpu_buffer->cpu);
2643	if (full && nr_pages && (dirty * 100) <= full * nr_pages)
2644		return;
2645
2646	cpu_buffer->irq_work.wakeup_full = true;
2647	cpu_buffer->irq_work.full_waiters_pending = false;
2648	/* irq_work_queue() supplies it's own memory barriers */
2649	irq_work_queue(&cpu_buffer->irq_work.work);
2650}
2651
2652/*
2653 * The lock and unlock are done within a preempt disable section.
2654 * The current_context per_cpu variable can only be modified
2655 * by the current task between lock and unlock. But it can
2656 * be modified more than once via an interrupt. To pass this
2657 * information from the lock to the unlock without having to
2658 * access the 'in_interrupt()' functions again (which do show
2659 * a bit of overhead in something as critical as function tracing,
2660 * we use a bitmask trick.
2661 *
2662 *  bit 0 =  NMI context
2663 *  bit 1 =  IRQ context
2664 *  bit 2 =  SoftIRQ context
2665 *  bit 3 =  normal context.
2666 *
2667 * This works because this is the order of contexts that can
2668 * preempt other contexts. A SoftIRQ never preempts an IRQ
2669 * context.
2670 *
2671 * When the context is determined, the corresponding bit is
2672 * checked and set (if it was set, then a recursion of that context
2673 * happened).
2674 *
2675 * On unlock, we need to clear this bit. To do so, just subtract
2676 * 1 from the current_context and AND it to itself.
2677 *
2678 * (binary)
2679 *  101 - 1 = 100
2680 *  101 & 100 = 100 (clearing bit zero)
2681 *
2682 *  1010 - 1 = 1001
2683 *  1010 & 1001 = 1000 (clearing bit 1)
2684 *
2685 * The least significant bit can be cleared this way, and it
2686 * just so happens that it is the same bit corresponding to
2687 * the current context.
2688 */
2689
2690static __always_inline int
2691trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2692{
2693	unsigned int val = cpu_buffer->current_context;
2694	unsigned long pc = preempt_count();
2695	int bit;
2696
2697	if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
 
 
 
 
 
 
 
2698		bit = RB_CTX_NORMAL;
2699	else
2700		bit = pc & NMI_MASK ? RB_CTX_NMI :
2701			pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
2702
2703	if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
2704		return 1;
2705
2706	val |= (1 << (bit + cpu_buffer->nest));
2707	cpu_buffer->current_context = val;
2708
2709	return 0;
2710}
2711
2712static __always_inline void
2713trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2714{
2715	cpu_buffer->current_context &=
2716		cpu_buffer->current_context - (1 << cpu_buffer->nest);
2717}
2718
2719/* The recursive locking above uses 4 bits */
2720#define NESTED_BITS 4
2721
2722/**
2723 * ring_buffer_nest_start - Allow to trace while nested
2724 * @buffer: The ring buffer to modify
2725 *
2726 * The ring buffer has a safety mechanism to prevent recursion.
2727 * But there may be a case where a trace needs to be done while
2728 * tracing something else. In this case, calling this function
2729 * will allow this function to nest within a currently active
2730 * ring_buffer_lock_reserve().
2731 *
2732 * Call this function before calling another ring_buffer_lock_reserve() and
2733 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
2734 */
2735void ring_buffer_nest_start(struct ring_buffer *buffer)
2736{
2737	struct ring_buffer_per_cpu *cpu_buffer;
2738	int cpu;
2739
2740	/* Enabled by ring_buffer_nest_end() */
2741	preempt_disable_notrace();
2742	cpu = raw_smp_processor_id();
2743	cpu_buffer = buffer->buffers[cpu];
2744	/* This is the shift value for the above recursive locking */
2745	cpu_buffer->nest += NESTED_BITS;
2746}
2747
2748/**
2749 * ring_buffer_nest_end - Allow to trace while nested
2750 * @buffer: The ring buffer to modify
2751 *
2752 * Must be called after ring_buffer_nest_start() and after the
2753 * ring_buffer_unlock_commit().
2754 */
2755void ring_buffer_nest_end(struct ring_buffer *buffer)
2756{
2757	struct ring_buffer_per_cpu *cpu_buffer;
2758	int cpu;
2759
2760	/* disabled by ring_buffer_nest_start() */
2761	cpu = raw_smp_processor_id();
2762	cpu_buffer = buffer->buffers[cpu];
2763	/* This is the shift value for the above recursive locking */
2764	cpu_buffer->nest -= NESTED_BITS;
2765	preempt_enable_notrace();
2766}
2767
2768/**
2769 * ring_buffer_unlock_commit - commit a reserved
2770 * @buffer: The buffer to commit to
2771 * @event: The event pointer to commit.
2772 *
2773 * This commits the data to the ring buffer, and releases any locks held.
2774 *
2775 * Must be paired with ring_buffer_lock_reserve.
2776 */
2777int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2778			      struct ring_buffer_event *event)
2779{
2780	struct ring_buffer_per_cpu *cpu_buffer;
2781	int cpu = raw_smp_processor_id();
2782
2783	cpu_buffer = buffer->buffers[cpu];
2784
2785	rb_commit(cpu_buffer, event);
2786
2787	rb_wakeups(buffer, cpu_buffer);
2788
2789	trace_recursive_unlock(cpu_buffer);
2790
2791	preempt_enable_notrace();
2792
2793	return 0;
2794}
2795EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2796
2797static noinline void
2798rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2799		    struct rb_event_info *info)
2800{
2801	WARN_ONCE(info->delta > (1ULL << 59),
2802		  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2803		  (unsigned long long)info->delta,
2804		  (unsigned long long)info->ts,
2805		  (unsigned long long)cpu_buffer->write_stamp,
2806		  sched_clock_stable() ? "" :
2807		  "If you just came from a suspend/resume,\n"
2808		  "please switch to the trace global clock:\n"
2809		  "  echo global > /sys/kernel/debug/tracing/trace_clock\n"
2810		  "or add trace_clock=global to the kernel command line\n");
2811	info->add_timestamp = 1;
2812}
2813
2814static struct ring_buffer_event *
2815__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2816		  struct rb_event_info *info)
2817{
2818	struct ring_buffer_event *event;
2819	struct buffer_page *tail_page;
2820	unsigned long tail, write;
2821
2822	/*
2823	 * If the time delta since the last event is too big to
2824	 * hold in the time field of the event, then we append a
2825	 * TIME EXTEND event ahead of the data event.
2826	 */
2827	if (unlikely(info->add_timestamp))
2828		info->length += RB_LEN_TIME_EXTEND;
2829
2830	/* Don't let the compiler play games with cpu_buffer->tail_page */
2831	tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2832	write = local_add_return(info->length, &tail_page->write);
2833
2834	/* set write to only the index of the write */
2835	write &= RB_WRITE_MASK;
2836	tail = write - info->length;
2837
2838	/*
2839	 * If this is the first commit on the page, then it has the same
2840	 * timestamp as the page itself.
2841	 */
2842	if (!tail && !ring_buffer_time_stamp_abs(cpu_buffer->buffer))
2843		info->delta = 0;
2844
2845	/* See if we shot pass the end of this buffer page */
2846	if (unlikely(write > BUF_PAGE_SIZE))
2847		return rb_move_tail(cpu_buffer, tail, info);
2848
2849	/* We reserved something on the buffer */
2850
2851	event = __rb_page_index(tail_page, tail);
 
2852	rb_update_event(cpu_buffer, event, info);
2853
2854	local_inc(&tail_page->entries);
2855
2856	/*
2857	 * If this is the first commit on the page, then update
2858	 * its timestamp.
2859	 */
2860	if (!tail)
2861		tail_page->page->time_stamp = info->ts;
2862
2863	/* account for these added bytes */
2864	local_add(info->length, &cpu_buffer->entries_bytes);
2865
2866	return event;
2867}
2868
2869static __always_inline struct ring_buffer_event *
2870rb_reserve_next_event(struct ring_buffer *buffer,
2871		      struct ring_buffer_per_cpu *cpu_buffer,
2872		      unsigned long length)
2873{
2874	struct ring_buffer_event *event;
2875	struct rb_event_info info;
2876	int nr_loops = 0;
2877	u64 diff;
2878
2879	rb_start_commit(cpu_buffer);
2880
2881#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2882	/*
2883	 * Due to the ability to swap a cpu buffer from a buffer
2884	 * it is possible it was swapped before we committed.
2885	 * (committing stops a swap). We check for it here and
2886	 * if it happened, we have to fail the write.
2887	 */
2888	barrier();
2889	if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2890		local_dec(&cpu_buffer->committing);
2891		local_dec(&cpu_buffer->commits);
2892		return NULL;
2893	}
2894#endif
2895
2896	info.length = rb_calculate_event_length(length);
2897 again:
2898	info.add_timestamp = 0;
2899	info.delta = 0;
2900
2901	/*
2902	 * We allow for interrupts to reenter here and do a trace.
2903	 * If one does, it will cause this original code to loop
2904	 * back here. Even with heavy interrupts happening, this
2905	 * should only happen a few times in a row. If this happens
2906	 * 1000 times in a row, there must be either an interrupt
2907	 * storm or we have something buggy.
2908	 * Bail!
2909	 */
2910	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2911		goto out_fail;
2912
2913	info.ts = rb_time_stamp(cpu_buffer->buffer);
2914	diff = info.ts - cpu_buffer->write_stamp;
2915
2916	/* make sure this diff is calculated here */
2917	barrier();
2918
2919	if (ring_buffer_time_stamp_abs(buffer)) {
2920		info.delta = info.ts;
2921		rb_handle_timestamp(cpu_buffer, &info);
2922	} else /* Did the write stamp get updated already? */
2923		if (likely(info.ts >= cpu_buffer->write_stamp)) {
2924		info.delta = diff;
2925		if (unlikely(test_time_stamp(info.delta)))
2926			rb_handle_timestamp(cpu_buffer, &info);
2927	}
2928
2929	event = __rb_reserve_next(cpu_buffer, &info);
2930
2931	if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2932		if (info.add_timestamp)
2933			info.length -= RB_LEN_TIME_EXTEND;
2934		goto again;
2935	}
2936
2937	if (!event)
2938		goto out_fail;
2939
2940	return event;
2941
2942 out_fail:
2943	rb_end_commit(cpu_buffer);
2944	return NULL;
2945}
2946
2947/**
2948 * ring_buffer_lock_reserve - reserve a part of the buffer
2949 * @buffer: the ring buffer to reserve from
2950 * @length: the length of the data to reserve (excluding event header)
2951 *
2952 * Returns a reserved event on the ring buffer to copy directly to.
2953 * The user of this interface will need to get the body to write into
2954 * and can use the ring_buffer_event_data() interface.
2955 *
2956 * The length is the length of the data needed, not the event length
2957 * which also includes the event header.
2958 *
2959 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2960 * If NULL is returned, then nothing has been allocated or locked.
2961 */
2962struct ring_buffer_event *
2963ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2964{
2965	struct ring_buffer_per_cpu *cpu_buffer;
2966	struct ring_buffer_event *event;
2967	int cpu;
2968
2969	/* If we are tracing schedule, we don't want to recurse */
2970	preempt_disable_notrace();
2971
2972	if (unlikely(atomic_read(&buffer->record_disabled)))
2973		goto out;
2974
2975	cpu = raw_smp_processor_id();
2976
2977	if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2978		goto out;
2979
2980	cpu_buffer = buffer->buffers[cpu];
2981
2982	if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2983		goto out;
2984
2985	if (unlikely(length > BUF_MAX_DATA_SIZE))
2986		goto out;
2987
2988	if (unlikely(trace_recursive_lock(cpu_buffer)))
2989		goto out;
2990
2991	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2992	if (!event)
2993		goto out_unlock;
2994
2995	return event;
2996
2997 out_unlock:
2998	trace_recursive_unlock(cpu_buffer);
2999 out:
3000	preempt_enable_notrace();
3001	return NULL;
3002}
3003EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3004
3005/*
3006 * Decrement the entries to the page that an event is on.
3007 * The event does not even need to exist, only the pointer
3008 * to the page it is on. This may only be called before the commit
3009 * takes place.
3010 */
3011static inline void
3012rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3013		   struct ring_buffer_event *event)
3014{
3015	unsigned long addr = (unsigned long)event;
3016	struct buffer_page *bpage = cpu_buffer->commit_page;
3017	struct buffer_page *start;
3018
3019	addr &= PAGE_MASK;
3020
3021	/* Do the likely case first */
3022	if (likely(bpage->page == (void *)addr)) {
3023		local_dec(&bpage->entries);
3024		return;
3025	}
3026
3027	/*
3028	 * Because the commit page may be on the reader page we
3029	 * start with the next page and check the end loop there.
3030	 */
3031	rb_inc_page(cpu_buffer, &bpage);
3032	start = bpage;
3033	do {
3034		if (bpage->page == (void *)addr) {
3035			local_dec(&bpage->entries);
3036			return;
3037		}
3038		rb_inc_page(cpu_buffer, &bpage);
3039	} while (bpage != start);
3040
3041	/* commit not part of this buffer?? */
3042	RB_WARN_ON(cpu_buffer, 1);
3043}
3044
3045/**
3046 * ring_buffer_commit_discard - discard an event that has not been committed
3047 * @buffer: the ring buffer
3048 * @event: non committed event to discard
3049 *
3050 * Sometimes an event that is in the ring buffer needs to be ignored.
3051 * This function lets the user discard an event in the ring buffer
3052 * and then that event will not be read later.
3053 *
3054 * This function only works if it is called before the item has been
3055 * committed. It will try to free the event from the ring buffer
3056 * if another event has not been added behind it.
3057 *
3058 * If another event has been added behind it, it will set the event
3059 * up as discarded, and perform the commit.
3060 *
3061 * If this function is called, do not call ring_buffer_unlock_commit on
3062 * the event.
3063 */
3064void ring_buffer_discard_commit(struct ring_buffer *buffer,
3065				struct ring_buffer_event *event)
3066{
3067	struct ring_buffer_per_cpu *cpu_buffer;
3068	int cpu;
3069
3070	/* The event is discarded regardless */
3071	rb_event_discard(event);
3072
3073	cpu = smp_processor_id();
3074	cpu_buffer = buffer->buffers[cpu];
3075
3076	/*
3077	 * This must only be called if the event has not been
3078	 * committed yet. Thus we can assume that preemption
3079	 * is still disabled.
3080	 */
3081	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3082
3083	rb_decrement_entry(cpu_buffer, event);
3084	if (rb_try_to_discard(cpu_buffer, event))
3085		goto out;
3086
3087	/*
3088	 * The commit is still visible by the reader, so we
3089	 * must still update the timestamp.
3090	 */
3091	rb_update_write_stamp(cpu_buffer, event);
3092 out:
3093	rb_end_commit(cpu_buffer);
3094
3095	trace_recursive_unlock(cpu_buffer);
3096
3097	preempt_enable_notrace();
3098
3099}
3100EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3101
3102/**
3103 * ring_buffer_write - write data to the buffer without reserving
3104 * @buffer: The ring buffer to write to.
3105 * @length: The length of the data being written (excluding the event header)
3106 * @data: The data to write to the buffer.
3107 *
3108 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3109 * one function. If you already have the data to write to the buffer, it
3110 * may be easier to simply call this function.
3111 *
3112 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3113 * and not the length of the event which would hold the header.
3114 */
3115int ring_buffer_write(struct ring_buffer *buffer,
3116		      unsigned long length,
3117		      void *data)
3118{
3119	struct ring_buffer_per_cpu *cpu_buffer;
3120	struct ring_buffer_event *event;
3121	void *body;
3122	int ret = -EBUSY;
3123	int cpu;
3124
3125	preempt_disable_notrace();
3126
3127	if (atomic_read(&buffer->record_disabled))
3128		goto out;
3129
3130	cpu = raw_smp_processor_id();
3131
3132	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3133		goto out;
3134
3135	cpu_buffer = buffer->buffers[cpu];
3136
3137	if (atomic_read(&cpu_buffer->record_disabled))
3138		goto out;
3139
3140	if (length > BUF_MAX_DATA_SIZE)
3141		goto out;
3142
3143	if (unlikely(trace_recursive_lock(cpu_buffer)))
3144		goto out;
3145
3146	event = rb_reserve_next_event(buffer, cpu_buffer, length);
3147	if (!event)
3148		goto out_unlock;
3149
3150	body = rb_event_data(event);
3151
3152	memcpy(body, data, length);
3153
3154	rb_commit(cpu_buffer, event);
3155
3156	rb_wakeups(buffer, cpu_buffer);
3157
3158	ret = 0;
3159
3160 out_unlock:
3161	trace_recursive_unlock(cpu_buffer);
3162
3163 out:
3164	preempt_enable_notrace();
3165
3166	return ret;
3167}
3168EXPORT_SYMBOL_GPL(ring_buffer_write);
3169
3170static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3171{
3172	struct buffer_page *reader = cpu_buffer->reader_page;
3173	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3174	struct buffer_page *commit = cpu_buffer->commit_page;
3175
3176	/* In case of error, head will be NULL */
3177	if (unlikely(!head))
3178		return true;
3179
3180	return reader->read == rb_page_commit(reader) &&
3181		(commit == reader ||
3182		 (commit == head &&
3183		  head->read == rb_page_commit(commit)));
3184}
3185
3186/**
3187 * ring_buffer_record_disable - stop all writes into the buffer
3188 * @buffer: The ring buffer to stop writes to.
3189 *
3190 * This prevents all writes to the buffer. Any attempt to write
3191 * to the buffer after this will fail and return NULL.
3192 *
3193 * The caller should call synchronize_rcu() after this.
3194 */
3195void ring_buffer_record_disable(struct ring_buffer *buffer)
3196{
3197	atomic_inc(&buffer->record_disabled);
3198}
3199EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3200
3201/**
3202 * ring_buffer_record_enable - enable writes to the buffer
3203 * @buffer: The ring buffer to enable writes
3204 *
3205 * Note, multiple disables will need the same number of enables
3206 * to truly enable the writing (much like preempt_disable).
3207 */
3208void ring_buffer_record_enable(struct ring_buffer *buffer)
3209{
3210	atomic_dec(&buffer->record_disabled);
3211}
3212EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3213
3214/**
3215 * ring_buffer_record_off - stop all writes into the buffer
3216 * @buffer: The ring buffer to stop writes to.
3217 *
3218 * This prevents all writes to the buffer. Any attempt to write
3219 * to the buffer after this will fail and return NULL.
3220 *
3221 * This is different than ring_buffer_record_disable() as
3222 * it works like an on/off switch, where as the disable() version
3223 * must be paired with a enable().
3224 */
3225void ring_buffer_record_off(struct ring_buffer *buffer)
3226{
3227	unsigned int rd;
3228	unsigned int new_rd;
3229
3230	do {
3231		rd = atomic_read(&buffer->record_disabled);
3232		new_rd = rd | RB_BUFFER_OFF;
3233	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3234}
3235EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3236
3237/**
3238 * ring_buffer_record_on - restart writes into the buffer
3239 * @buffer: The ring buffer to start writes to.
3240 *
3241 * This enables all writes to the buffer that was disabled by
3242 * ring_buffer_record_off().
3243 *
3244 * This is different than ring_buffer_record_enable() as
3245 * it works like an on/off switch, where as the enable() version
3246 * must be paired with a disable().
3247 */
3248void ring_buffer_record_on(struct ring_buffer *buffer)
3249{
3250	unsigned int rd;
3251	unsigned int new_rd;
3252
3253	do {
3254		rd = atomic_read(&buffer->record_disabled);
3255		new_rd = rd & ~RB_BUFFER_OFF;
3256	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3257}
3258EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3259
3260/**
3261 * ring_buffer_record_is_on - return true if the ring buffer can write
3262 * @buffer: The ring buffer to see if write is enabled
3263 *
3264 * Returns true if the ring buffer is in a state that it accepts writes.
3265 */
3266bool ring_buffer_record_is_on(struct ring_buffer *buffer)
3267{
3268	return !atomic_read(&buffer->record_disabled);
3269}
3270
3271/**
3272 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3273 * @buffer: The ring buffer to see if write is set enabled
3274 *
3275 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3276 * Note that this does NOT mean it is in a writable state.
3277 *
3278 * It may return true when the ring buffer has been disabled by
3279 * ring_buffer_record_disable(), as that is a temporary disabling of
3280 * the ring buffer.
3281 */
3282bool ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3283{
3284	return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3285}
3286
3287/**
3288 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3289 * @buffer: The ring buffer to stop writes to.
3290 * @cpu: The CPU buffer to stop
3291 *
3292 * This prevents all writes to the buffer. Any attempt to write
3293 * to the buffer after this will fail and return NULL.
3294 *
3295 * The caller should call synchronize_rcu() after this.
3296 */
3297void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3298{
3299	struct ring_buffer_per_cpu *cpu_buffer;
3300
3301	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3302		return;
3303
3304	cpu_buffer = buffer->buffers[cpu];
3305	atomic_inc(&cpu_buffer->record_disabled);
3306}
3307EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3308
3309/**
3310 * ring_buffer_record_enable_cpu - enable writes to the buffer
3311 * @buffer: The ring buffer to enable writes
3312 * @cpu: The CPU to enable.
3313 *
3314 * Note, multiple disables will need the same number of enables
3315 * to truly enable the writing (much like preempt_disable).
3316 */
3317void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3318{
3319	struct ring_buffer_per_cpu *cpu_buffer;
3320
3321	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3322		return;
3323
3324	cpu_buffer = buffer->buffers[cpu];
3325	atomic_dec(&cpu_buffer->record_disabled);
3326}
3327EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3328
3329/*
3330 * The total entries in the ring buffer is the running counter
3331 * of entries entered into the ring buffer, minus the sum of
3332 * the entries read from the ring buffer and the number of
3333 * entries that were overwritten.
3334 */
3335static inline unsigned long
3336rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3337{
3338	return local_read(&cpu_buffer->entries) -
3339		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3340}
3341
3342/**
3343 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3344 * @buffer: The ring buffer
3345 * @cpu: The per CPU buffer to read from.
3346 */
3347u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3348{
3349	unsigned long flags;
3350	struct ring_buffer_per_cpu *cpu_buffer;
3351	struct buffer_page *bpage;
3352	u64 ret = 0;
3353
3354	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3355		return 0;
3356
3357	cpu_buffer = buffer->buffers[cpu];
3358	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3359	/*
3360	 * if the tail is on reader_page, oldest time stamp is on the reader
3361	 * page
3362	 */
3363	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3364		bpage = cpu_buffer->reader_page;
3365	else
3366		bpage = rb_set_head_page(cpu_buffer);
3367	if (bpage)
3368		ret = bpage->page->time_stamp;
3369	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3370
3371	return ret;
3372}
3373EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3374
3375/**
3376 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3377 * @buffer: The ring buffer
3378 * @cpu: The per CPU buffer to read from.
3379 */
3380unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3381{
3382	struct ring_buffer_per_cpu *cpu_buffer;
3383	unsigned long ret;
3384
3385	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3386		return 0;
3387
3388	cpu_buffer = buffer->buffers[cpu];
3389	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3390
3391	return ret;
3392}
3393EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3394
3395/**
3396 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3397 * @buffer: The ring buffer
3398 * @cpu: The per CPU buffer to get the entries from.
3399 */
3400unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3401{
3402	struct ring_buffer_per_cpu *cpu_buffer;
3403
3404	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3405		return 0;
3406
3407	cpu_buffer = buffer->buffers[cpu];
3408
3409	return rb_num_of_entries(cpu_buffer);
3410}
3411EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3412
3413/**
3414 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3415 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3416 * @buffer: The ring buffer
3417 * @cpu: The per CPU buffer to get the number of overruns from
3418 */
3419unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3420{
3421	struct ring_buffer_per_cpu *cpu_buffer;
3422	unsigned long ret;
3423
3424	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3425		return 0;
3426
3427	cpu_buffer = buffer->buffers[cpu];
3428	ret = local_read(&cpu_buffer->overrun);
3429
3430	return ret;
3431}
3432EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3433
3434/**
3435 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3436 * commits failing due to the buffer wrapping around while there are uncommitted
3437 * events, such as during an interrupt storm.
3438 * @buffer: The ring buffer
3439 * @cpu: The per CPU buffer to get the number of overruns from
3440 */
3441unsigned long
3442ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3443{
3444	struct ring_buffer_per_cpu *cpu_buffer;
3445	unsigned long ret;
3446
3447	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3448		return 0;
3449
3450	cpu_buffer = buffer->buffers[cpu];
3451	ret = local_read(&cpu_buffer->commit_overrun);
3452
3453	return ret;
3454}
3455EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3456
3457/**
3458 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3459 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3460 * @buffer: The ring buffer
3461 * @cpu: The per CPU buffer to get the number of overruns from
3462 */
3463unsigned long
3464ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3465{
3466	struct ring_buffer_per_cpu *cpu_buffer;
3467	unsigned long ret;
3468
3469	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3470		return 0;
3471
3472	cpu_buffer = buffer->buffers[cpu];
3473	ret = local_read(&cpu_buffer->dropped_events);
3474
3475	return ret;
3476}
3477EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3478
3479/**
3480 * ring_buffer_read_events_cpu - get the number of events successfully read
3481 * @buffer: The ring buffer
3482 * @cpu: The per CPU buffer to get the number of events read
3483 */
3484unsigned long
3485ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3486{
3487	struct ring_buffer_per_cpu *cpu_buffer;
3488
3489	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3490		return 0;
3491
3492	cpu_buffer = buffer->buffers[cpu];
3493	return cpu_buffer->read;
3494}
3495EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3496
3497/**
3498 * ring_buffer_entries - get the number of entries in a buffer
3499 * @buffer: The ring buffer
3500 *
3501 * Returns the total number of entries in the ring buffer
3502 * (all CPU entries)
3503 */
3504unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3505{
3506	struct ring_buffer_per_cpu *cpu_buffer;
3507	unsigned long entries = 0;
3508	int cpu;
3509
3510	/* if you care about this being correct, lock the buffer */
3511	for_each_buffer_cpu(buffer, cpu) {
3512		cpu_buffer = buffer->buffers[cpu];
3513		entries += rb_num_of_entries(cpu_buffer);
3514	}
3515
3516	return entries;
3517}
3518EXPORT_SYMBOL_GPL(ring_buffer_entries);
3519
3520/**
3521 * ring_buffer_overruns - get the number of overruns in buffer
3522 * @buffer: The ring buffer
3523 *
3524 * Returns the total number of overruns in the ring buffer
3525 * (all CPU entries)
3526 */
3527unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3528{
3529	struct ring_buffer_per_cpu *cpu_buffer;
3530	unsigned long overruns = 0;
3531	int cpu;
3532
3533	/* if you care about this being correct, lock the buffer */
3534	for_each_buffer_cpu(buffer, cpu) {
3535		cpu_buffer = buffer->buffers[cpu];
3536		overruns += local_read(&cpu_buffer->overrun);
3537	}
3538
3539	return overruns;
3540}
3541EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3542
3543static void rb_iter_reset(struct ring_buffer_iter *iter)
3544{
3545	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3546
3547	/* Iterator usage is expected to have record disabled */
3548	iter->head_page = cpu_buffer->reader_page;
3549	iter->head = cpu_buffer->reader_page->read;
3550
3551	iter->cache_reader_page = iter->head_page;
3552	iter->cache_read = cpu_buffer->read;
3553
3554	if (iter->head)
3555		iter->read_stamp = cpu_buffer->read_stamp;
3556	else
3557		iter->read_stamp = iter->head_page->page->time_stamp;
3558}
3559
3560/**
3561 * ring_buffer_iter_reset - reset an iterator
3562 * @iter: The iterator to reset
3563 *
3564 * Resets the iterator, so that it will start from the beginning
3565 * again.
3566 */
3567void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3568{
3569	struct ring_buffer_per_cpu *cpu_buffer;
3570	unsigned long flags;
3571
3572	if (!iter)
3573		return;
3574
3575	cpu_buffer = iter->cpu_buffer;
3576
3577	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3578	rb_iter_reset(iter);
3579	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3580}
3581EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3582
3583/**
3584 * ring_buffer_iter_empty - check if an iterator has no more to read
3585 * @iter: The iterator to check
3586 */
3587int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3588{
3589	struct ring_buffer_per_cpu *cpu_buffer;
3590	struct buffer_page *reader;
3591	struct buffer_page *head_page;
3592	struct buffer_page *commit_page;
3593	unsigned commit;
3594
3595	cpu_buffer = iter->cpu_buffer;
3596
3597	/* Remember, trace recording is off when iterator is in use */
3598	reader = cpu_buffer->reader_page;
3599	head_page = cpu_buffer->head_page;
3600	commit_page = cpu_buffer->commit_page;
3601	commit = rb_page_commit(commit_page);
3602
3603	return ((iter->head_page == commit_page && iter->head == commit) ||
3604		(iter->head_page == reader && commit_page == head_page &&
3605		 head_page->read == commit &&
3606		 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3607}
3608EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3609
3610static void
3611rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3612		     struct ring_buffer_event *event)
3613{
3614	u64 delta;
3615
3616	switch (event->type_len) {
3617	case RINGBUF_TYPE_PADDING:
3618		return;
3619
3620	case RINGBUF_TYPE_TIME_EXTEND:
3621		delta = ring_buffer_event_time_stamp(event);
 
 
3622		cpu_buffer->read_stamp += delta;
3623		return;
3624
3625	case RINGBUF_TYPE_TIME_STAMP:
3626		delta = ring_buffer_event_time_stamp(event);
3627		cpu_buffer->read_stamp = delta;
3628		return;
3629
3630	case RINGBUF_TYPE_DATA:
3631		cpu_buffer->read_stamp += event->time_delta;
3632		return;
3633
3634	default:
3635		BUG();
3636	}
3637	return;
3638}
3639
3640static void
3641rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3642			  struct ring_buffer_event *event)
3643{
3644	u64 delta;
3645
3646	switch (event->type_len) {
3647	case RINGBUF_TYPE_PADDING:
3648		return;
3649
3650	case RINGBUF_TYPE_TIME_EXTEND:
3651		delta = ring_buffer_event_time_stamp(event);
 
 
3652		iter->read_stamp += delta;
3653		return;
3654
3655	case RINGBUF_TYPE_TIME_STAMP:
3656		delta = ring_buffer_event_time_stamp(event);
3657		iter->read_stamp = delta;
3658		return;
3659
3660	case RINGBUF_TYPE_DATA:
3661		iter->read_stamp += event->time_delta;
3662		return;
3663
3664	default:
3665		BUG();
3666	}
3667	return;
3668}
3669
3670static struct buffer_page *
3671rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3672{
3673	struct buffer_page *reader = NULL;
3674	unsigned long overwrite;
3675	unsigned long flags;
3676	int nr_loops = 0;
3677	int ret;
3678
3679	local_irq_save(flags);
3680	arch_spin_lock(&cpu_buffer->lock);
3681
3682 again:
3683	/*
3684	 * This should normally only loop twice. But because the
3685	 * start of the reader inserts an empty page, it causes
3686	 * a case where we will loop three times. There should be no
3687	 * reason to loop four times (that I know of).
3688	 */
3689	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3690		reader = NULL;
3691		goto out;
3692	}
3693
3694	reader = cpu_buffer->reader_page;
3695
3696	/* If there's more to read, return this page */
3697	if (cpu_buffer->reader_page->read < rb_page_size(reader))
3698		goto out;
3699
3700	/* Never should we have an index greater than the size */
3701	if (RB_WARN_ON(cpu_buffer,
3702		       cpu_buffer->reader_page->read > rb_page_size(reader)))
3703		goto out;
3704
3705	/* check if we caught up to the tail */
3706	reader = NULL;
3707	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3708		goto out;
3709
3710	/* Don't bother swapping if the ring buffer is empty */
3711	if (rb_num_of_entries(cpu_buffer) == 0)
3712		goto out;
3713
3714	/*
3715	 * Reset the reader page to size zero.
3716	 */
3717	local_set(&cpu_buffer->reader_page->write, 0);
3718	local_set(&cpu_buffer->reader_page->entries, 0);
3719	local_set(&cpu_buffer->reader_page->page->commit, 0);
3720	cpu_buffer->reader_page->real_end = 0;
3721
3722 spin:
3723	/*
3724	 * Splice the empty reader page into the list around the head.
3725	 */
3726	reader = rb_set_head_page(cpu_buffer);
3727	if (!reader)
3728		goto out;
3729	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3730	cpu_buffer->reader_page->list.prev = reader->list.prev;
3731
3732	/*
3733	 * cpu_buffer->pages just needs to point to the buffer, it
3734	 *  has no specific buffer page to point to. Lets move it out
3735	 *  of our way so we don't accidentally swap it.
3736	 */
3737	cpu_buffer->pages = reader->list.prev;
3738
3739	/* The reader page will be pointing to the new head */
3740	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3741
3742	/*
3743	 * We want to make sure we read the overruns after we set up our
3744	 * pointers to the next object. The writer side does a
3745	 * cmpxchg to cross pages which acts as the mb on the writer
3746	 * side. Note, the reader will constantly fail the swap
3747	 * while the writer is updating the pointers, so this
3748	 * guarantees that the overwrite recorded here is the one we
3749	 * want to compare with the last_overrun.
3750	 */
3751	smp_mb();
3752	overwrite = local_read(&(cpu_buffer->overrun));
3753
3754	/*
3755	 * Here's the tricky part.
3756	 *
3757	 * We need to move the pointer past the header page.
3758	 * But we can only do that if a writer is not currently
3759	 * moving it. The page before the header page has the
3760	 * flag bit '1' set if it is pointing to the page we want.
3761	 * but if the writer is in the process of moving it
3762	 * than it will be '2' or already moved '0'.
3763	 */
3764
3765	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3766
3767	/*
3768	 * If we did not convert it, then we must try again.
3769	 */
3770	if (!ret)
3771		goto spin;
3772
3773	/*
3774	 * Yay! We succeeded in replacing the page.
3775	 *
3776	 * Now make the new head point back to the reader page.
3777	 */
3778	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3779	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3780
3781	local_inc(&cpu_buffer->pages_read);
3782
3783	/* Finally update the reader page to the new head */
3784	cpu_buffer->reader_page = reader;
3785	cpu_buffer->reader_page->read = 0;
3786
3787	if (overwrite != cpu_buffer->last_overrun) {
3788		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3789		cpu_buffer->last_overrun = overwrite;
3790	}
3791
3792	goto again;
3793
3794 out:
3795	/* Update the read_stamp on the first event */
3796	if (reader && reader->read == 0)
3797		cpu_buffer->read_stamp = reader->page->time_stamp;
3798
3799	arch_spin_unlock(&cpu_buffer->lock);
3800	local_irq_restore(flags);
3801
3802	return reader;
3803}
3804
3805static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3806{
3807	struct ring_buffer_event *event;
3808	struct buffer_page *reader;
3809	unsigned length;
3810
3811	reader = rb_get_reader_page(cpu_buffer);
3812
3813	/* This function should not be called when buffer is empty */
3814	if (RB_WARN_ON(cpu_buffer, !reader))
3815		return;
3816
3817	event = rb_reader_event(cpu_buffer);
3818
3819	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3820		cpu_buffer->read++;
3821
3822	rb_update_read_stamp(cpu_buffer, event);
3823
3824	length = rb_event_length(event);
3825	cpu_buffer->reader_page->read += length;
3826}
3827
3828static void rb_advance_iter(struct ring_buffer_iter *iter)
3829{
3830	struct ring_buffer_per_cpu *cpu_buffer;
3831	struct ring_buffer_event *event;
3832	unsigned length;
3833
3834	cpu_buffer = iter->cpu_buffer;
3835
3836	/*
3837	 * Check if we are at the end of the buffer.
3838	 */
3839	if (iter->head >= rb_page_size(iter->head_page)) {
3840		/* discarded commits can make the page empty */
3841		if (iter->head_page == cpu_buffer->commit_page)
3842			return;
3843		rb_inc_iter(iter);
3844		return;
3845	}
3846
3847	event = rb_iter_head_event(iter);
3848
3849	length = rb_event_length(event);
3850
3851	/*
3852	 * This should not be called to advance the header if we are
3853	 * at the tail of the buffer.
3854	 */
3855	if (RB_WARN_ON(cpu_buffer,
3856		       (iter->head_page == cpu_buffer->commit_page) &&
3857		       (iter->head + length > rb_commit_index(cpu_buffer))))
3858		return;
3859
3860	rb_update_iter_read_stamp(iter, event);
3861
3862	iter->head += length;
3863
3864	/* check for end of page padding */
3865	if ((iter->head >= rb_page_size(iter->head_page)) &&
3866	    (iter->head_page != cpu_buffer->commit_page))
3867		rb_inc_iter(iter);
3868}
3869
3870static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3871{
3872	return cpu_buffer->lost_events;
3873}
3874
3875static struct ring_buffer_event *
3876rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3877	       unsigned long *lost_events)
3878{
3879	struct ring_buffer_event *event;
3880	struct buffer_page *reader;
3881	int nr_loops = 0;
3882
3883	if (ts)
3884		*ts = 0;
3885 again:
3886	/*
3887	 * We repeat when a time extend is encountered.
3888	 * Since the time extend is always attached to a data event,
3889	 * we should never loop more than once.
3890	 * (We never hit the following condition more than twice).
3891	 */
3892	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3893		return NULL;
3894
3895	reader = rb_get_reader_page(cpu_buffer);
3896	if (!reader)
3897		return NULL;
3898
3899	event = rb_reader_event(cpu_buffer);
3900
3901	switch (event->type_len) {
3902	case RINGBUF_TYPE_PADDING:
3903		if (rb_null_event(event))
3904			RB_WARN_ON(cpu_buffer, 1);
3905		/*
3906		 * Because the writer could be discarding every
3907		 * event it creates (which would probably be bad)
3908		 * if we were to go back to "again" then we may never
3909		 * catch up, and will trigger the warn on, or lock
3910		 * the box. Return the padding, and we will release
3911		 * the current locks, and try again.
3912		 */
3913		return event;
3914
3915	case RINGBUF_TYPE_TIME_EXTEND:
3916		/* Internal data, OK to advance */
3917		rb_advance_reader(cpu_buffer);
3918		goto again;
3919
3920	case RINGBUF_TYPE_TIME_STAMP:
3921		if (ts) {
3922			*ts = ring_buffer_event_time_stamp(event);
3923			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3924							 cpu_buffer->cpu, ts);
3925		}
3926		/* Internal data, OK to advance */
3927		rb_advance_reader(cpu_buffer);
3928		goto again;
3929
3930	case RINGBUF_TYPE_DATA:
3931		if (ts && !(*ts)) {
3932			*ts = cpu_buffer->read_stamp + event->time_delta;
3933			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3934							 cpu_buffer->cpu, ts);
3935		}
3936		if (lost_events)
3937			*lost_events = rb_lost_events(cpu_buffer);
3938		return event;
3939
3940	default:
3941		BUG();
3942	}
3943
3944	return NULL;
3945}
3946EXPORT_SYMBOL_GPL(ring_buffer_peek);
3947
3948static struct ring_buffer_event *
3949rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3950{
3951	struct ring_buffer *buffer;
3952	struct ring_buffer_per_cpu *cpu_buffer;
3953	struct ring_buffer_event *event;
3954	int nr_loops = 0;
3955
3956	if (ts)
3957		*ts = 0;
3958
3959	cpu_buffer = iter->cpu_buffer;
3960	buffer = cpu_buffer->buffer;
3961
3962	/*
3963	 * Check if someone performed a consuming read to
3964	 * the buffer. A consuming read invalidates the iterator
3965	 * and we need to reset the iterator in this case.
3966	 */
3967	if (unlikely(iter->cache_read != cpu_buffer->read ||
3968		     iter->cache_reader_page != cpu_buffer->reader_page))
3969		rb_iter_reset(iter);
3970
3971 again:
3972	if (ring_buffer_iter_empty(iter))
3973		return NULL;
3974
3975	/*
3976	 * We repeat when a time extend is encountered or we hit
3977	 * the end of the page. Since the time extend is always attached
3978	 * to a data event, we should never loop more than three times.
3979	 * Once for going to next page, once on time extend, and
3980	 * finally once to get the event.
3981	 * (We never hit the following condition more than thrice).
3982	 */
3983	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3984		return NULL;
3985
3986	if (rb_per_cpu_empty(cpu_buffer))
3987		return NULL;
3988
3989	if (iter->head >= rb_page_size(iter->head_page)) {
3990		rb_inc_iter(iter);
3991		goto again;
3992	}
3993
3994	event = rb_iter_head_event(iter);
3995
3996	switch (event->type_len) {
3997	case RINGBUF_TYPE_PADDING:
3998		if (rb_null_event(event)) {
3999			rb_inc_iter(iter);
4000			goto again;
4001		}
4002		rb_advance_iter(iter);
4003		return event;
4004
4005	case RINGBUF_TYPE_TIME_EXTEND:
4006		/* Internal data, OK to advance */
4007		rb_advance_iter(iter);
4008		goto again;
4009
4010	case RINGBUF_TYPE_TIME_STAMP:
4011		if (ts) {
4012			*ts = ring_buffer_event_time_stamp(event);
4013			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4014							 cpu_buffer->cpu, ts);
4015		}
4016		/* Internal data, OK to advance */
4017		rb_advance_iter(iter);
4018		goto again;
4019
4020	case RINGBUF_TYPE_DATA:
4021		if (ts && !(*ts)) {
4022			*ts = iter->read_stamp + event->time_delta;
4023			ring_buffer_normalize_time_stamp(buffer,
4024							 cpu_buffer->cpu, ts);
4025		}
4026		return event;
4027
4028	default:
4029		BUG();
4030	}
4031
4032	return NULL;
4033}
4034EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4035
4036static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4037{
4038	if (likely(!in_nmi())) {
4039		raw_spin_lock(&cpu_buffer->reader_lock);
4040		return true;
4041	}
4042
4043	/*
4044	 * If an NMI die dumps out the content of the ring buffer
4045	 * trylock must be used to prevent a deadlock if the NMI
4046	 * preempted a task that holds the ring buffer locks. If
4047	 * we get the lock then all is fine, if not, then continue
4048	 * to do the read, but this can corrupt the ring buffer,
4049	 * so it must be permanently disabled from future writes.
4050	 * Reading from NMI is a oneshot deal.
4051	 */
4052	if (raw_spin_trylock(&cpu_buffer->reader_lock))
4053		return true;
4054
4055	/* Continue without locking, but disable the ring buffer */
4056	atomic_inc(&cpu_buffer->record_disabled);
4057	return false;
4058}
4059
4060static inline void
4061rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4062{
4063	if (likely(locked))
4064		raw_spin_unlock(&cpu_buffer->reader_lock);
4065	return;
4066}
4067
4068/**
4069 * ring_buffer_peek - peek at the next event to be read
4070 * @buffer: The ring buffer to read
4071 * @cpu: The cpu to peak at
4072 * @ts: The timestamp counter of this event.
4073 * @lost_events: a variable to store if events were lost (may be NULL)
4074 *
4075 * This will return the event that will be read next, but does
4076 * not consume the data.
4077 */
4078struct ring_buffer_event *
4079ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
4080		 unsigned long *lost_events)
4081{
4082	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4083	struct ring_buffer_event *event;
4084	unsigned long flags;
4085	bool dolock;
4086
4087	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4088		return NULL;
4089
4090 again:
4091	local_irq_save(flags);
4092	dolock = rb_reader_lock(cpu_buffer);
4093	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4094	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4095		rb_advance_reader(cpu_buffer);
4096	rb_reader_unlock(cpu_buffer, dolock);
4097	local_irq_restore(flags);
4098
4099	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4100		goto again;
4101
4102	return event;
4103}
4104
4105/**
4106 * ring_buffer_iter_peek - peek at the next event to be read
4107 * @iter: The ring buffer iterator
4108 * @ts: The timestamp counter of this event.
4109 *
4110 * This will return the event that will be read next, but does
4111 * not increment the iterator.
4112 */
4113struct ring_buffer_event *
4114ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4115{
4116	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4117	struct ring_buffer_event *event;
4118	unsigned long flags;
4119
4120 again:
4121	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4122	event = rb_iter_peek(iter, ts);
4123	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4124
4125	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4126		goto again;
4127
4128	return event;
4129}
4130
4131/**
4132 * ring_buffer_consume - return an event and consume it
4133 * @buffer: The ring buffer to get the next event from
4134 * @cpu: the cpu to read the buffer from
4135 * @ts: a variable to store the timestamp (may be NULL)
4136 * @lost_events: a variable to store if events were lost (may be NULL)
4137 *
4138 * Returns the next event in the ring buffer, and that event is consumed.
4139 * Meaning, that sequential reads will keep returning a different event,
4140 * and eventually empty the ring buffer if the producer is slower.
4141 */
4142struct ring_buffer_event *
4143ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4144		    unsigned long *lost_events)
4145{
4146	struct ring_buffer_per_cpu *cpu_buffer;
4147	struct ring_buffer_event *event = NULL;
4148	unsigned long flags;
4149	bool dolock;
4150
4151 again:
4152	/* might be called in atomic */
4153	preempt_disable();
4154
4155	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4156		goto out;
4157
4158	cpu_buffer = buffer->buffers[cpu];
4159	local_irq_save(flags);
4160	dolock = rb_reader_lock(cpu_buffer);
4161
4162	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4163	if (event) {
4164		cpu_buffer->lost_events = 0;
4165		rb_advance_reader(cpu_buffer);
4166	}
4167
4168	rb_reader_unlock(cpu_buffer, dolock);
4169	local_irq_restore(flags);
4170
4171 out:
4172	preempt_enable();
4173
4174	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4175		goto again;
4176
4177	return event;
4178}
4179EXPORT_SYMBOL_GPL(ring_buffer_consume);
4180
4181/**
4182 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4183 * @buffer: The ring buffer to read from
4184 * @cpu: The cpu buffer to iterate over
4185 * @flags: gfp flags to use for memory allocation
4186 *
4187 * This performs the initial preparations necessary to iterate
4188 * through the buffer.  Memory is allocated, buffer recording
4189 * is disabled, and the iterator pointer is returned to the caller.
4190 *
4191 * Disabling buffer recording prevents the reading from being
4192 * corrupted. This is not a consuming read, so a producer is not
4193 * expected.
4194 *
4195 * After a sequence of ring_buffer_read_prepare calls, the user is
4196 * expected to make at least one call to ring_buffer_read_prepare_sync.
4197 * Afterwards, ring_buffer_read_start is invoked to get things going
4198 * for real.
4199 *
4200 * This overall must be paired with ring_buffer_read_finish.
4201 */
4202struct ring_buffer_iter *
4203ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu, gfp_t flags)
4204{
4205	struct ring_buffer_per_cpu *cpu_buffer;
4206	struct ring_buffer_iter *iter;
4207
4208	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4209		return NULL;
4210
4211	iter = kmalloc(sizeof(*iter), flags);
4212	if (!iter)
4213		return NULL;
4214
4215	cpu_buffer = buffer->buffers[cpu];
4216
4217	iter->cpu_buffer = cpu_buffer;
4218
4219	atomic_inc(&buffer->resize_disabled);
4220	atomic_inc(&cpu_buffer->record_disabled);
4221
4222	return iter;
4223}
4224EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4225
4226/**
4227 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4228 *
4229 * All previously invoked ring_buffer_read_prepare calls to prepare
4230 * iterators will be synchronized.  Afterwards, read_buffer_read_start
4231 * calls on those iterators are allowed.
4232 */
4233void
4234ring_buffer_read_prepare_sync(void)
4235{
4236	synchronize_rcu();
4237}
4238EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4239
4240/**
4241 * ring_buffer_read_start - start a non consuming read of the buffer
4242 * @iter: The iterator returned by ring_buffer_read_prepare
4243 *
4244 * This finalizes the startup of an iteration through the buffer.
4245 * The iterator comes from a call to ring_buffer_read_prepare and
4246 * an intervening ring_buffer_read_prepare_sync must have been
4247 * performed.
4248 *
4249 * Must be paired with ring_buffer_read_finish.
4250 */
4251void
4252ring_buffer_read_start(struct ring_buffer_iter *iter)
4253{
4254	struct ring_buffer_per_cpu *cpu_buffer;
4255	unsigned long flags;
4256
4257	if (!iter)
4258		return;
4259
4260	cpu_buffer = iter->cpu_buffer;
4261
4262	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4263	arch_spin_lock(&cpu_buffer->lock);
4264	rb_iter_reset(iter);
4265	arch_spin_unlock(&cpu_buffer->lock);
4266	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4267}
4268EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4269
4270/**
4271 * ring_buffer_read_finish - finish reading the iterator of the buffer
4272 * @iter: The iterator retrieved by ring_buffer_start
4273 *
4274 * This re-enables the recording to the buffer, and frees the
4275 * iterator.
4276 */
4277void
4278ring_buffer_read_finish(struct ring_buffer_iter *iter)
4279{
4280	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4281	unsigned long flags;
4282
4283	/*
4284	 * Ring buffer is disabled from recording, here's a good place
4285	 * to check the integrity of the ring buffer.
4286	 * Must prevent readers from trying to read, as the check
4287	 * clears the HEAD page and readers require it.
4288	 */
4289	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4290	rb_check_pages(cpu_buffer);
4291	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4292
4293	atomic_dec(&cpu_buffer->record_disabled);
4294	atomic_dec(&cpu_buffer->buffer->resize_disabled);
4295	kfree(iter);
4296}
4297EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4298
4299/**
4300 * ring_buffer_read - read the next item in the ring buffer by the iterator
4301 * @iter: The ring buffer iterator
4302 * @ts: The time stamp of the event read.
4303 *
4304 * This reads the next event in the ring buffer and increments the iterator.
4305 */
4306struct ring_buffer_event *
4307ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4308{
4309	struct ring_buffer_event *event;
4310	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4311	unsigned long flags;
4312
4313	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4314 again:
4315	event = rb_iter_peek(iter, ts);
4316	if (!event)
4317		goto out;
4318
4319	if (event->type_len == RINGBUF_TYPE_PADDING)
4320		goto again;
4321
4322	rb_advance_iter(iter);
4323 out:
4324	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4325
4326	return event;
4327}
4328EXPORT_SYMBOL_GPL(ring_buffer_read);
4329
4330/**
4331 * ring_buffer_size - return the size of the ring buffer (in bytes)
4332 * @buffer: The ring buffer.
4333 */
4334unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4335{
4336	/*
4337	 * Earlier, this method returned
4338	 *	BUF_PAGE_SIZE * buffer->nr_pages
4339	 * Since the nr_pages field is now removed, we have converted this to
4340	 * return the per cpu buffer value.
4341	 */
4342	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4343		return 0;
4344
4345	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4346}
4347EXPORT_SYMBOL_GPL(ring_buffer_size);
4348
4349static void
4350rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4351{
4352	rb_head_page_deactivate(cpu_buffer);
4353
4354	cpu_buffer->head_page
4355		= list_entry(cpu_buffer->pages, struct buffer_page, list);
4356	local_set(&cpu_buffer->head_page->write, 0);
4357	local_set(&cpu_buffer->head_page->entries, 0);
4358	local_set(&cpu_buffer->head_page->page->commit, 0);
4359
4360	cpu_buffer->head_page->read = 0;
4361
4362	cpu_buffer->tail_page = cpu_buffer->head_page;
4363	cpu_buffer->commit_page = cpu_buffer->head_page;
4364
4365	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4366	INIT_LIST_HEAD(&cpu_buffer->new_pages);
4367	local_set(&cpu_buffer->reader_page->write, 0);
4368	local_set(&cpu_buffer->reader_page->entries, 0);
4369	local_set(&cpu_buffer->reader_page->page->commit, 0);
4370	cpu_buffer->reader_page->read = 0;
4371
4372	local_set(&cpu_buffer->entries_bytes, 0);
4373	local_set(&cpu_buffer->overrun, 0);
4374	local_set(&cpu_buffer->commit_overrun, 0);
4375	local_set(&cpu_buffer->dropped_events, 0);
4376	local_set(&cpu_buffer->entries, 0);
4377	local_set(&cpu_buffer->committing, 0);
4378	local_set(&cpu_buffer->commits, 0);
4379	local_set(&cpu_buffer->pages_touched, 0);
4380	local_set(&cpu_buffer->pages_read, 0);
4381	cpu_buffer->last_pages_touch = 0;
4382	cpu_buffer->shortest_full = 0;
4383	cpu_buffer->read = 0;
4384	cpu_buffer->read_bytes = 0;
4385
4386	cpu_buffer->write_stamp = 0;
4387	cpu_buffer->read_stamp = 0;
4388
4389	cpu_buffer->lost_events = 0;
4390	cpu_buffer->last_overrun = 0;
4391
4392	rb_head_page_activate(cpu_buffer);
4393}
4394
4395/**
4396 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4397 * @buffer: The ring buffer to reset a per cpu buffer of
4398 * @cpu: The CPU buffer to be reset
4399 */
4400void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4401{
4402	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4403	unsigned long flags;
4404
4405	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4406		return;
4407
4408	atomic_inc(&buffer->resize_disabled);
4409	atomic_inc(&cpu_buffer->record_disabled);
4410
4411	/* Make sure all commits have finished */
4412	synchronize_rcu();
4413
4414	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4415
4416	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4417		goto out;
4418
4419	arch_spin_lock(&cpu_buffer->lock);
4420
4421	rb_reset_cpu(cpu_buffer);
4422
4423	arch_spin_unlock(&cpu_buffer->lock);
4424
4425 out:
4426	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4427
4428	atomic_dec(&cpu_buffer->record_disabled);
4429	atomic_dec(&buffer->resize_disabled);
4430}
4431EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4432
4433/**
4434 * ring_buffer_reset - reset a ring buffer
4435 * @buffer: The ring buffer to reset all cpu buffers
4436 */
4437void ring_buffer_reset(struct ring_buffer *buffer)
4438{
4439	int cpu;
4440
4441	for_each_buffer_cpu(buffer, cpu)
4442		ring_buffer_reset_cpu(buffer, cpu);
4443}
4444EXPORT_SYMBOL_GPL(ring_buffer_reset);
4445
4446/**
4447 * rind_buffer_empty - is the ring buffer empty?
4448 * @buffer: The ring buffer to test
4449 */
4450bool ring_buffer_empty(struct ring_buffer *buffer)
4451{
4452	struct ring_buffer_per_cpu *cpu_buffer;
4453	unsigned long flags;
4454	bool dolock;
4455	int cpu;
4456	int ret;
4457
4458	/* yes this is racy, but if you don't like the race, lock the buffer */
4459	for_each_buffer_cpu(buffer, cpu) {
4460		cpu_buffer = buffer->buffers[cpu];
4461		local_irq_save(flags);
4462		dolock = rb_reader_lock(cpu_buffer);
4463		ret = rb_per_cpu_empty(cpu_buffer);
4464		rb_reader_unlock(cpu_buffer, dolock);
4465		local_irq_restore(flags);
4466
4467		if (!ret)
4468			return false;
4469	}
4470
4471	return true;
4472}
4473EXPORT_SYMBOL_GPL(ring_buffer_empty);
4474
4475/**
4476 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4477 * @buffer: The ring buffer
4478 * @cpu: The CPU buffer to test
4479 */
4480bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4481{
4482	struct ring_buffer_per_cpu *cpu_buffer;
4483	unsigned long flags;
4484	bool dolock;
4485	int ret;
4486
4487	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4488		return true;
4489
4490	cpu_buffer = buffer->buffers[cpu];
4491	local_irq_save(flags);
4492	dolock = rb_reader_lock(cpu_buffer);
4493	ret = rb_per_cpu_empty(cpu_buffer);
4494	rb_reader_unlock(cpu_buffer, dolock);
4495	local_irq_restore(flags);
4496
4497	return ret;
4498}
4499EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4500
4501#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4502/**
4503 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4504 * @buffer_a: One buffer to swap with
4505 * @buffer_b: The other buffer to swap with
4506 *
4507 * This function is useful for tracers that want to take a "snapshot"
4508 * of a CPU buffer and has another back up buffer lying around.
4509 * it is expected that the tracer handles the cpu buffer not being
4510 * used at the moment.
4511 */
4512int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4513			 struct ring_buffer *buffer_b, int cpu)
4514{
4515	struct ring_buffer_per_cpu *cpu_buffer_a;
4516	struct ring_buffer_per_cpu *cpu_buffer_b;
4517	int ret = -EINVAL;
4518
4519	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4520	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
4521		goto out;
4522
4523	cpu_buffer_a = buffer_a->buffers[cpu];
4524	cpu_buffer_b = buffer_b->buffers[cpu];
4525
4526	/* At least make sure the two buffers are somewhat the same */
4527	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4528		goto out;
4529
4530	ret = -EAGAIN;
4531
4532	if (atomic_read(&buffer_a->record_disabled))
4533		goto out;
4534
4535	if (atomic_read(&buffer_b->record_disabled))
4536		goto out;
4537
4538	if (atomic_read(&cpu_buffer_a->record_disabled))
4539		goto out;
4540
4541	if (atomic_read(&cpu_buffer_b->record_disabled))
4542		goto out;
4543
4544	/*
4545	 * We can't do a synchronize_rcu here because this
4546	 * function can be called in atomic context.
4547	 * Normally this will be called from the same CPU as cpu.
4548	 * If not it's up to the caller to protect this.
4549	 */
4550	atomic_inc(&cpu_buffer_a->record_disabled);
4551	atomic_inc(&cpu_buffer_b->record_disabled);
4552
4553	ret = -EBUSY;
4554	if (local_read(&cpu_buffer_a->committing))
4555		goto out_dec;
4556	if (local_read(&cpu_buffer_b->committing))
4557		goto out_dec;
4558
4559	buffer_a->buffers[cpu] = cpu_buffer_b;
4560	buffer_b->buffers[cpu] = cpu_buffer_a;
4561
4562	cpu_buffer_b->buffer = buffer_a;
4563	cpu_buffer_a->buffer = buffer_b;
4564
4565	ret = 0;
4566
4567out_dec:
4568	atomic_dec(&cpu_buffer_a->record_disabled);
4569	atomic_dec(&cpu_buffer_b->record_disabled);
4570out:
4571	return ret;
4572}
4573EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4574#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4575
4576/**
4577 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4578 * @buffer: the buffer to allocate for.
4579 * @cpu: the cpu buffer to allocate.
4580 *
4581 * This function is used in conjunction with ring_buffer_read_page.
4582 * When reading a full page from the ring buffer, these functions
4583 * can be used to speed up the process. The calling function should
4584 * allocate a few pages first with this function. Then when it
4585 * needs to get pages from the ring buffer, it passes the result
4586 * of this function into ring_buffer_read_page, which will swap
4587 * the page that was allocated, with the read page of the buffer.
4588 *
4589 * Returns:
4590 *  The page allocated, or ERR_PTR
4591 */
4592void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4593{
4594	struct ring_buffer_per_cpu *cpu_buffer;
4595	struct buffer_data_page *bpage = NULL;
4596	unsigned long flags;
4597	struct page *page;
4598
4599	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4600		return ERR_PTR(-ENODEV);
4601
4602	cpu_buffer = buffer->buffers[cpu];
4603	local_irq_save(flags);
4604	arch_spin_lock(&cpu_buffer->lock);
4605
4606	if (cpu_buffer->free_page) {
4607		bpage = cpu_buffer->free_page;
4608		cpu_buffer->free_page = NULL;
4609	}
4610
4611	arch_spin_unlock(&cpu_buffer->lock);
4612	local_irq_restore(flags);
4613
4614	if (bpage)
4615		goto out;
4616
4617	page = alloc_pages_node(cpu_to_node(cpu),
4618				GFP_KERNEL | __GFP_NORETRY, 0);
4619	if (!page)
4620		return ERR_PTR(-ENOMEM);
4621
4622	bpage = page_address(page);
4623
4624 out:
4625	rb_init_page(bpage);
4626
4627	return bpage;
4628}
4629EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4630
4631/**
4632 * ring_buffer_free_read_page - free an allocated read page
4633 * @buffer: the buffer the page was allocate for
4634 * @cpu: the cpu buffer the page came from
4635 * @data: the page to free
4636 *
4637 * Free a page allocated from ring_buffer_alloc_read_page.
4638 */
4639void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4640{
4641	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4642	struct buffer_data_page *bpage = data;
4643	struct page *page = virt_to_page(bpage);
4644	unsigned long flags;
4645
4646	/* If the page is still in use someplace else, we can't reuse it */
4647	if (page_ref_count(page) > 1)
4648		goto out;
4649
4650	local_irq_save(flags);
4651	arch_spin_lock(&cpu_buffer->lock);
4652
4653	if (!cpu_buffer->free_page) {
4654		cpu_buffer->free_page = bpage;
4655		bpage = NULL;
4656	}
4657
4658	arch_spin_unlock(&cpu_buffer->lock);
4659	local_irq_restore(flags);
4660
4661 out:
4662	free_page((unsigned long)bpage);
4663}
4664EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4665
4666/**
4667 * ring_buffer_read_page - extract a page from the ring buffer
4668 * @buffer: buffer to extract from
4669 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4670 * @len: amount to extract
4671 * @cpu: the cpu of the buffer to extract
4672 * @full: should the extraction only happen when the page is full.
4673 *
4674 * This function will pull out a page from the ring buffer and consume it.
4675 * @data_page must be the address of the variable that was returned
4676 * from ring_buffer_alloc_read_page. This is because the page might be used
4677 * to swap with a page in the ring buffer.
4678 *
4679 * for example:
4680 *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
4681 *	if (IS_ERR(rpage))
4682 *		return PTR_ERR(rpage);
4683 *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4684 *	if (ret >= 0)
4685 *		process_page(rpage, ret);
4686 *
4687 * When @full is set, the function will not return true unless
4688 * the writer is off the reader page.
4689 *
4690 * Note: it is up to the calling functions to handle sleeps and wakeups.
4691 *  The ring buffer can be used anywhere in the kernel and can not
4692 *  blindly call wake_up. The layer that uses the ring buffer must be
4693 *  responsible for that.
4694 *
4695 * Returns:
4696 *  >=0 if data has been transferred, returns the offset of consumed data.
4697 *  <0 if no data has been transferred.
4698 */
4699int ring_buffer_read_page(struct ring_buffer *buffer,
4700			  void **data_page, size_t len, int cpu, int full)
4701{
4702	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4703	struct ring_buffer_event *event;
4704	struct buffer_data_page *bpage;
4705	struct buffer_page *reader;
4706	unsigned long missed_events;
4707	unsigned long flags;
4708	unsigned int commit;
4709	unsigned int read;
4710	u64 save_timestamp;
4711	int ret = -1;
4712
4713	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4714		goto out;
4715
4716	/*
4717	 * If len is not big enough to hold the page header, then
4718	 * we can not copy anything.
4719	 */
4720	if (len <= BUF_PAGE_HDR_SIZE)
4721		goto out;
4722
4723	len -= BUF_PAGE_HDR_SIZE;
4724
4725	if (!data_page)
4726		goto out;
4727
4728	bpage = *data_page;
4729	if (!bpage)
4730		goto out;
4731
4732	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4733
4734	reader = rb_get_reader_page(cpu_buffer);
4735	if (!reader)
4736		goto out_unlock;
4737
4738	event = rb_reader_event(cpu_buffer);
4739
4740	read = reader->read;
4741	commit = rb_page_commit(reader);
4742
4743	/* Check if any events were dropped */
4744	missed_events = cpu_buffer->lost_events;
4745
4746	/*
4747	 * If this page has been partially read or
4748	 * if len is not big enough to read the rest of the page or
4749	 * a writer is still on the page, then
4750	 * we must copy the data from the page to the buffer.
4751	 * Otherwise, we can simply swap the page with the one passed in.
4752	 */
4753	if (read || (len < (commit - read)) ||
4754	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
4755		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4756		unsigned int rpos = read;
4757		unsigned int pos = 0;
4758		unsigned int size;
4759
4760		if (full)
4761			goto out_unlock;
4762
4763		if (len > (commit - read))
4764			len = (commit - read);
4765
4766		/* Always keep the time extend and data together */
4767		size = rb_event_ts_length(event);
4768
4769		if (len < size)
4770			goto out_unlock;
4771
4772		/* save the current timestamp, since the user will need it */
4773		save_timestamp = cpu_buffer->read_stamp;
4774
4775		/* Need to copy one event at a time */
4776		do {
4777			/* We need the size of one event, because
4778			 * rb_advance_reader only advances by one event,
4779			 * whereas rb_event_ts_length may include the size of
4780			 * one or two events.
4781			 * We have already ensured there's enough space if this
4782			 * is a time extend. */
4783			size = rb_event_length(event);
4784			memcpy(bpage->data + pos, rpage->data + rpos, size);
4785
4786			len -= size;
4787
4788			rb_advance_reader(cpu_buffer);
4789			rpos = reader->read;
4790			pos += size;
4791
4792			if (rpos >= commit)
4793				break;
4794
4795			event = rb_reader_event(cpu_buffer);
4796			/* Always keep the time extend and data together */
4797			size = rb_event_ts_length(event);
4798		} while (len >= size);
4799
4800		/* update bpage */
4801		local_set(&bpage->commit, pos);
4802		bpage->time_stamp = save_timestamp;
4803
4804		/* we copied everything to the beginning */
4805		read = 0;
4806	} else {
4807		/* update the entry counter */
4808		cpu_buffer->read += rb_page_entries(reader);
4809		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4810
4811		/* swap the pages */
4812		rb_init_page(bpage);
4813		bpage = reader->page;
4814		reader->page = *data_page;
4815		local_set(&reader->write, 0);
4816		local_set(&reader->entries, 0);
4817		reader->read = 0;
4818		*data_page = bpage;
4819
4820		/*
4821		 * Use the real_end for the data size,
4822		 * This gives us a chance to store the lost events
4823		 * on the page.
4824		 */
4825		if (reader->real_end)
4826			local_set(&bpage->commit, reader->real_end);
4827	}
4828	ret = read;
4829
4830	cpu_buffer->lost_events = 0;
4831
4832	commit = local_read(&bpage->commit);
4833	/*
4834	 * Set a flag in the commit field if we lost events
4835	 */
4836	if (missed_events) {
4837		/* If there is room at the end of the page to save the
4838		 * missed events, then record it there.
4839		 */
4840		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4841			memcpy(&bpage->data[commit], &missed_events,
4842			       sizeof(missed_events));
4843			local_add(RB_MISSED_STORED, &bpage->commit);
4844			commit += sizeof(missed_events);
4845		}
4846		local_add(RB_MISSED_EVENTS, &bpage->commit);
4847	}
4848
4849	/*
4850	 * This page may be off to user land. Zero it out here.
4851	 */
4852	if (commit < BUF_PAGE_SIZE)
4853		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4854
4855 out_unlock:
4856	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4857
4858 out:
4859	return ret;
4860}
4861EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4862
4863/*
4864 * We only allocate new buffers, never free them if the CPU goes down.
4865 * If we were to free the buffer, then the user would lose any trace that was in
4866 * the buffer.
4867 */
4868int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4869{
4870	struct ring_buffer *buffer;
4871	long nr_pages_same;
4872	int cpu_i;
4873	unsigned long nr_pages;
4874
4875	buffer = container_of(node, struct ring_buffer, node);
4876	if (cpumask_test_cpu(cpu, buffer->cpumask))
4877		return 0;
4878
4879	nr_pages = 0;
4880	nr_pages_same = 1;
4881	/* check if all cpu sizes are same */
4882	for_each_buffer_cpu(buffer, cpu_i) {
4883		/* fill in the size from first enabled cpu */
4884		if (nr_pages == 0)
4885			nr_pages = buffer->buffers[cpu_i]->nr_pages;
4886		if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4887			nr_pages_same = 0;
4888			break;
4889		}
4890	}
4891	/* allocate minimum pages, user can later expand it */
4892	if (!nr_pages_same)
4893		nr_pages = 2;
4894	buffer->buffers[cpu] =
4895		rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4896	if (!buffer->buffers[cpu]) {
4897		WARN(1, "failed to allocate ring buffer on CPU %u\n",
4898		     cpu);
4899		return -ENOMEM;
4900	}
4901	smp_wmb();
4902	cpumask_set_cpu(cpu, buffer->cpumask);
4903	return 0;
4904}
4905
4906#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4907/*
4908 * This is a basic integrity check of the ring buffer.
4909 * Late in the boot cycle this test will run when configured in.
4910 * It will kick off a thread per CPU that will go into a loop
4911 * writing to the per cpu ring buffer various sizes of data.
4912 * Some of the data will be large items, some small.
4913 *
4914 * Another thread is created that goes into a spin, sending out
4915 * IPIs to the other CPUs to also write into the ring buffer.
4916 * this is to test the nesting ability of the buffer.
4917 *
4918 * Basic stats are recorded and reported. If something in the
4919 * ring buffer should happen that's not expected, a big warning
4920 * is displayed and all ring buffers are disabled.
4921 */
4922static struct task_struct *rb_threads[NR_CPUS] __initdata;
4923
4924struct rb_test_data {
4925	struct ring_buffer	*buffer;
4926	unsigned long		events;
4927	unsigned long		bytes_written;
4928	unsigned long		bytes_alloc;
4929	unsigned long		bytes_dropped;
4930	unsigned long		events_nested;
4931	unsigned long		bytes_written_nested;
4932	unsigned long		bytes_alloc_nested;
4933	unsigned long		bytes_dropped_nested;
4934	int			min_size_nested;
4935	int			max_size_nested;
4936	int			max_size;
4937	int			min_size;
4938	int			cpu;
4939	int			cnt;
4940};
4941
4942static struct rb_test_data rb_data[NR_CPUS] __initdata;
4943
4944/* 1 meg per cpu */
4945#define RB_TEST_BUFFER_SIZE	1048576
4946
4947static char rb_string[] __initdata =
4948	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4949	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4950	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4951
4952static bool rb_test_started __initdata;
4953
4954struct rb_item {
4955	int size;
4956	char str[];
4957};
4958
4959static __init int rb_write_something(struct rb_test_data *data, bool nested)
4960{
4961	struct ring_buffer_event *event;
4962	struct rb_item *item;
4963	bool started;
4964	int event_len;
4965	int size;
4966	int len;
4967	int cnt;
4968
4969	/* Have nested writes different that what is written */
4970	cnt = data->cnt + (nested ? 27 : 0);
4971
4972	/* Multiply cnt by ~e, to make some unique increment */
4973	size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
4974
4975	len = size + sizeof(struct rb_item);
4976
4977	started = rb_test_started;
4978	/* read rb_test_started before checking buffer enabled */
4979	smp_rmb();
4980
4981	event = ring_buffer_lock_reserve(data->buffer, len);
4982	if (!event) {
4983		/* Ignore dropped events before test starts. */
4984		if (started) {
4985			if (nested)
4986				data->bytes_dropped += len;
4987			else
4988				data->bytes_dropped_nested += len;
4989		}
4990		return len;
4991	}
4992
4993	event_len = ring_buffer_event_length(event);
4994
4995	if (RB_WARN_ON(data->buffer, event_len < len))
4996		goto out;
4997
4998	item = ring_buffer_event_data(event);
4999	item->size = size;
5000	memcpy(item->str, rb_string, size);
5001
5002	if (nested) {
5003		data->bytes_alloc_nested += event_len;
5004		data->bytes_written_nested += len;
5005		data->events_nested++;
5006		if (!data->min_size_nested || len < data->min_size_nested)
5007			data->min_size_nested = len;
5008		if (len > data->max_size_nested)
5009			data->max_size_nested = len;
5010	} else {
5011		data->bytes_alloc += event_len;
5012		data->bytes_written += len;
5013		data->events++;
5014		if (!data->min_size || len < data->min_size)
5015			data->max_size = len;
5016		if (len > data->max_size)
5017			data->max_size = len;
5018	}
5019
5020 out:
5021	ring_buffer_unlock_commit(data->buffer, event);
5022
5023	return 0;
5024}
5025
5026static __init int rb_test(void *arg)
5027{
5028	struct rb_test_data *data = arg;
5029
5030	while (!kthread_should_stop()) {
5031		rb_write_something(data, false);
5032		data->cnt++;
5033
5034		set_current_state(TASK_INTERRUPTIBLE);
5035		/* Now sleep between a min of 100-300us and a max of 1ms */
5036		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5037	}
5038
5039	return 0;
5040}
5041
5042static __init void rb_ipi(void *ignore)
5043{
5044	struct rb_test_data *data;
5045	int cpu = smp_processor_id();
5046
5047	data = &rb_data[cpu];
5048	rb_write_something(data, true);
5049}
5050
5051static __init int rb_hammer_test(void *arg)
5052{
5053	while (!kthread_should_stop()) {
5054
5055		/* Send an IPI to all cpus to write data! */
5056		smp_call_function(rb_ipi, NULL, 1);
5057		/* No sleep, but for non preempt, let others run */
5058		schedule();
5059	}
5060
5061	return 0;
5062}
5063
5064static __init int test_ringbuffer(void)
5065{
5066	struct task_struct *rb_hammer;
5067	struct ring_buffer *buffer;
5068	int cpu;
5069	int ret = 0;
5070
5071	pr_info("Running ring buffer tests...\n");
5072
5073	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5074	if (WARN_ON(!buffer))
5075		return 0;
5076
5077	/* Disable buffer so that threads can't write to it yet */
5078	ring_buffer_record_off(buffer);
5079
5080	for_each_online_cpu(cpu) {
5081		rb_data[cpu].buffer = buffer;
5082		rb_data[cpu].cpu = cpu;
5083		rb_data[cpu].cnt = cpu;
5084		rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5085						 "rbtester/%d", cpu);
5086		if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5087			pr_cont("FAILED\n");
5088			ret = PTR_ERR(rb_threads[cpu]);
5089			goto out_free;
5090		}
5091
5092		kthread_bind(rb_threads[cpu], cpu);
5093 		wake_up_process(rb_threads[cpu]);
5094	}
5095
5096	/* Now create the rb hammer! */
5097	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5098	if (WARN_ON(IS_ERR(rb_hammer))) {
5099		pr_cont("FAILED\n");
5100		ret = PTR_ERR(rb_hammer);
5101		goto out_free;
5102	}
5103
5104	ring_buffer_record_on(buffer);
5105	/*
5106	 * Show buffer is enabled before setting rb_test_started.
5107	 * Yes there's a small race window where events could be
5108	 * dropped and the thread wont catch it. But when a ring
5109	 * buffer gets enabled, there will always be some kind of
5110	 * delay before other CPUs see it. Thus, we don't care about
5111	 * those dropped events. We care about events dropped after
5112	 * the threads see that the buffer is active.
5113	 */
5114	smp_wmb();
5115	rb_test_started = true;
5116
5117	set_current_state(TASK_INTERRUPTIBLE);
5118	/* Just run for 10 seconds */;
5119	schedule_timeout(10 * HZ);
5120
5121	kthread_stop(rb_hammer);
5122
5123 out_free:
5124	for_each_online_cpu(cpu) {
5125		if (!rb_threads[cpu])
5126			break;
5127		kthread_stop(rb_threads[cpu]);
5128	}
5129	if (ret) {
5130		ring_buffer_free(buffer);
5131		return ret;
5132	}
5133
5134	/* Report! */
5135	pr_info("finished\n");
5136	for_each_online_cpu(cpu) {
5137		struct ring_buffer_event *event;
5138		struct rb_test_data *data = &rb_data[cpu];
5139		struct rb_item *item;
5140		unsigned long total_events;
5141		unsigned long total_dropped;
5142		unsigned long total_written;
5143		unsigned long total_alloc;
5144		unsigned long total_read = 0;
5145		unsigned long total_size = 0;
5146		unsigned long total_len = 0;
5147		unsigned long total_lost = 0;
5148		unsigned long lost;
5149		int big_event_size;
5150		int small_event_size;
5151
5152		ret = -1;
5153
5154		total_events = data->events + data->events_nested;
5155		total_written = data->bytes_written + data->bytes_written_nested;
5156		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5157		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5158
5159		big_event_size = data->max_size + data->max_size_nested;
5160		small_event_size = data->min_size + data->min_size_nested;
5161
5162		pr_info("CPU %d:\n", cpu);
5163		pr_info("              events:    %ld\n", total_events);
5164		pr_info("       dropped bytes:    %ld\n", total_dropped);
5165		pr_info("       alloced bytes:    %ld\n", total_alloc);
5166		pr_info("       written bytes:    %ld\n", total_written);
5167		pr_info("       biggest event:    %d\n", big_event_size);
5168		pr_info("      smallest event:    %d\n", small_event_size);
5169
5170		if (RB_WARN_ON(buffer, total_dropped))
5171			break;
5172
5173		ret = 0;
5174
5175		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5176			total_lost += lost;
5177			item = ring_buffer_event_data(event);
5178			total_len += ring_buffer_event_length(event);
5179			total_size += item->size + sizeof(struct rb_item);
5180			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5181				pr_info("FAILED!\n");
5182				pr_info("buffer had: %.*s\n", item->size, item->str);
5183				pr_info("expected:   %.*s\n", item->size, rb_string);
5184				RB_WARN_ON(buffer, 1);
5185				ret = -1;
5186				break;
5187			}
5188			total_read++;
5189		}
5190		if (ret)
5191			break;
5192
5193		ret = -1;
5194
5195		pr_info("         read events:   %ld\n", total_read);
5196		pr_info("         lost events:   %ld\n", total_lost);
5197		pr_info("        total events:   %ld\n", total_lost + total_read);
5198		pr_info("  recorded len bytes:   %ld\n", total_len);
5199		pr_info(" recorded size bytes:   %ld\n", total_size);
5200		if (total_lost)
5201			pr_info(" With dropped events, record len and size may not match\n"
5202				" alloced and written from above\n");
5203		if (!total_lost) {
5204			if (RB_WARN_ON(buffer, total_len != total_alloc ||
5205				       total_size != total_written))
5206				break;
5207		}
5208		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5209			break;
5210
5211		ret = 0;
5212	}
5213	if (!ret)
5214		pr_info("Ring buffer PASSED!\n");
5215
5216	ring_buffer_free(buffer);
5217	return 0;
5218}
5219
5220late_initcall(test_ringbuffer);
5221#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
v4.10.11
 
   1/*
   2 * Generic ring buffer
   3 *
   4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
   5 */
   6#include <linux/trace_events.h>
   7#include <linux/ring_buffer.h>
   8#include <linux/trace_clock.h>
 
   9#include <linux/trace_seq.h>
  10#include <linux/spinlock.h>
  11#include <linux/irq_work.h>
  12#include <linux/uaccess.h>
  13#include <linux/hardirq.h>
  14#include <linux/kthread.h>	/* for self test */
  15#include <linux/kmemcheck.h>
  16#include <linux/module.h>
  17#include <linux/percpu.h>
  18#include <linux/mutex.h>
  19#include <linux/delay.h>
  20#include <linux/slab.h>
  21#include <linux/init.h>
  22#include <linux/hash.h>
  23#include <linux/list.h>
  24#include <linux/cpu.h>
 
  25
  26#include <asm/local.h>
  27
  28static void update_pages_handler(struct work_struct *work);
  29
  30/*
  31 * The ring buffer header is special. We must manually up keep it.
  32 */
  33int ring_buffer_print_entry_header(struct trace_seq *s)
  34{
  35	trace_seq_puts(s, "# compressed entry header\n");
  36	trace_seq_puts(s, "\ttype_len    :    5 bits\n");
  37	trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
  38	trace_seq_puts(s, "\tarray       :   32 bits\n");
  39	trace_seq_putc(s, '\n');
  40	trace_seq_printf(s, "\tpadding     : type == %d\n",
  41			 RINGBUF_TYPE_PADDING);
  42	trace_seq_printf(s, "\ttime_extend : type == %d\n",
  43			 RINGBUF_TYPE_TIME_EXTEND);
 
 
  44	trace_seq_printf(s, "\tdata max type_len  == %d\n",
  45			 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
  46
  47	return !trace_seq_has_overflowed(s);
  48}
  49
  50/*
  51 * The ring buffer is made up of a list of pages. A separate list of pages is
  52 * allocated for each CPU. A writer may only write to a buffer that is
  53 * associated with the CPU it is currently executing on.  A reader may read
  54 * from any per cpu buffer.
  55 *
  56 * The reader is special. For each per cpu buffer, the reader has its own
  57 * reader page. When a reader has read the entire reader page, this reader
  58 * page is swapped with another page in the ring buffer.
  59 *
  60 * Now, as long as the writer is off the reader page, the reader can do what
  61 * ever it wants with that page. The writer will never write to that page
  62 * again (as long as it is out of the ring buffer).
  63 *
  64 * Here's some silly ASCII art.
  65 *
  66 *   +------+
  67 *   |reader|          RING BUFFER
  68 *   |page  |
  69 *   +------+        +---+   +---+   +---+
  70 *                   |   |-->|   |-->|   |
  71 *                   +---+   +---+   +---+
  72 *                     ^               |
  73 *                     |               |
  74 *                     +---------------+
  75 *
  76 *
  77 *   +------+
  78 *   |reader|          RING BUFFER
  79 *   |page  |------------------v
  80 *   +------+        +---+   +---+   +---+
  81 *                   |   |-->|   |-->|   |
  82 *                   +---+   +---+   +---+
  83 *                     ^               |
  84 *                     |               |
  85 *                     +---------------+
  86 *
  87 *
  88 *   +------+
  89 *   |reader|          RING BUFFER
  90 *   |page  |------------------v
  91 *   +------+        +---+   +---+   +---+
  92 *      ^            |   |-->|   |-->|   |
  93 *      |            +---+   +---+   +---+
  94 *      |                              |
  95 *      |                              |
  96 *      +------------------------------+
  97 *
  98 *
  99 *   +------+
 100 *   |buffer|          RING BUFFER
 101 *   |page  |------------------v
 102 *   +------+        +---+   +---+   +---+
 103 *      ^            |   |   |   |-->|   |
 104 *      |   New      +---+   +---+   +---+
 105 *      |  Reader------^               |
 106 *      |   page                       |
 107 *      +------------------------------+
 108 *
 109 *
 110 * After we make this swap, the reader can hand this page off to the splice
 111 * code and be done with it. It can even allocate a new page if it needs to
 112 * and swap that into the ring buffer.
 113 *
 114 * We will be using cmpxchg soon to make all this lockless.
 115 *
 116 */
 117
 118/* Used for individual buffers (after the counter) */
 119#define RB_BUFFER_OFF		(1 << 20)
 120
 121#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
 122
 123#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
 124#define RB_ALIGNMENT		4U
 125#define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 126#define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
 127
 128#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
 129# define RB_FORCE_8BYTE_ALIGNMENT	0
 130# define RB_ARCH_ALIGNMENT		RB_ALIGNMENT
 131#else
 132# define RB_FORCE_8BYTE_ALIGNMENT	1
 133# define RB_ARCH_ALIGNMENT		8U
 134#endif
 135
 136#define RB_ALIGN_DATA		__aligned(RB_ARCH_ALIGNMENT)
 137
 138/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
 139#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
 140
 141enum {
 142	RB_LEN_TIME_EXTEND = 8,
 143	RB_LEN_TIME_STAMP = 16,
 144};
 145
 146#define skip_time_extend(event) \
 147	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
 148
 
 
 
 149static inline int rb_null_event(struct ring_buffer_event *event)
 150{
 151	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
 152}
 153
 154static void rb_event_set_padding(struct ring_buffer_event *event)
 155{
 156	/* padding has a NULL time_delta */
 157	event->type_len = RINGBUF_TYPE_PADDING;
 158	event->time_delta = 0;
 159}
 160
 161static unsigned
 162rb_event_data_length(struct ring_buffer_event *event)
 163{
 164	unsigned length;
 165
 166	if (event->type_len)
 167		length = event->type_len * RB_ALIGNMENT;
 168	else
 169		length = event->array[0];
 170	return length + RB_EVNT_HDR_SIZE;
 171}
 172
 173/*
 174 * Return the length of the given event. Will return
 175 * the length of the time extend if the event is a
 176 * time extend.
 177 */
 178static inline unsigned
 179rb_event_length(struct ring_buffer_event *event)
 180{
 181	switch (event->type_len) {
 182	case RINGBUF_TYPE_PADDING:
 183		if (rb_null_event(event))
 184			/* undefined */
 185			return -1;
 186		return  event->array[0] + RB_EVNT_HDR_SIZE;
 187
 188	case RINGBUF_TYPE_TIME_EXTEND:
 189		return RB_LEN_TIME_EXTEND;
 190
 191	case RINGBUF_TYPE_TIME_STAMP:
 192		return RB_LEN_TIME_STAMP;
 193
 194	case RINGBUF_TYPE_DATA:
 195		return rb_event_data_length(event);
 196	default:
 197		BUG();
 198	}
 199	/* not hit */
 200	return 0;
 201}
 202
 203/*
 204 * Return total length of time extend and data,
 205 *   or just the event length for all other events.
 206 */
 207static inline unsigned
 208rb_event_ts_length(struct ring_buffer_event *event)
 209{
 210	unsigned len = 0;
 211
 212	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
 213		/* time extends include the data event after it */
 214		len = RB_LEN_TIME_EXTEND;
 215		event = skip_time_extend(event);
 216	}
 217	return len + rb_event_length(event);
 218}
 219
 220/**
 221 * ring_buffer_event_length - return the length of the event
 222 * @event: the event to get the length of
 223 *
 224 * Returns the size of the data load of a data event.
 225 * If the event is something other than a data event, it
 226 * returns the size of the event itself. With the exception
 227 * of a TIME EXTEND, where it still returns the size of the
 228 * data load of the data event after it.
 229 */
 230unsigned ring_buffer_event_length(struct ring_buffer_event *event)
 231{
 232	unsigned length;
 233
 234	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
 235		event = skip_time_extend(event);
 236
 237	length = rb_event_length(event);
 238	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 239		return length;
 240	length -= RB_EVNT_HDR_SIZE;
 241	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
 242                length -= sizeof(event->array[0]);
 243	return length;
 244}
 245EXPORT_SYMBOL_GPL(ring_buffer_event_length);
 246
 247/* inline for ring buffer fast paths */
 248static __always_inline void *
 249rb_event_data(struct ring_buffer_event *event)
 250{
 251	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
 252		event = skip_time_extend(event);
 253	BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 254	/* If length is in len field, then array[0] has the data */
 255	if (event->type_len)
 256		return (void *)&event->array[0];
 257	/* Otherwise length is in array[0] and array[1] has the data */
 258	return (void *)&event->array[1];
 259}
 260
 261/**
 262 * ring_buffer_event_data - return the data of the event
 263 * @event: the event to get the data from
 264 */
 265void *ring_buffer_event_data(struct ring_buffer_event *event)
 266{
 267	return rb_event_data(event);
 268}
 269EXPORT_SYMBOL_GPL(ring_buffer_event_data);
 270
 271#define for_each_buffer_cpu(buffer, cpu)		\
 272	for_each_cpu(cpu, buffer->cpumask)
 273
 274#define TS_SHIFT	27
 275#define TS_MASK		((1ULL << TS_SHIFT) - 1)
 276#define TS_DELTA_TEST	(~TS_MASK)
 277
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 278/* Flag when events were overwritten */
 279#define RB_MISSED_EVENTS	(1 << 31)
 280/* Missed count stored at end */
 281#define RB_MISSED_STORED	(1 << 30)
 282
 
 
 283struct buffer_data_page {
 284	u64		 time_stamp;	/* page time stamp */
 285	local_t		 commit;	/* write committed index */
 286	unsigned char	 data[] RB_ALIGN_DATA;	/* data of buffer page */
 287};
 288
 289/*
 290 * Note, the buffer_page list must be first. The buffer pages
 291 * are allocated in cache lines, which means that each buffer
 292 * page will be at the beginning of a cache line, and thus
 293 * the least significant bits will be zero. We use this to
 294 * add flags in the list struct pointers, to make the ring buffer
 295 * lockless.
 296 */
 297struct buffer_page {
 298	struct list_head list;		/* list of buffer pages */
 299	local_t		 write;		/* index for next write */
 300	unsigned	 read;		/* index for next read */
 301	local_t		 entries;	/* entries on this page */
 302	unsigned long	 real_end;	/* real end of data */
 303	struct buffer_data_page *page;	/* Actual data page */
 304};
 305
 306/*
 307 * The buffer page counters, write and entries, must be reset
 308 * atomically when crossing page boundaries. To synchronize this
 309 * update, two counters are inserted into the number. One is
 310 * the actual counter for the write position or count on the page.
 311 *
 312 * The other is a counter of updaters. Before an update happens
 313 * the update partition of the counter is incremented. This will
 314 * allow the updater to update the counter atomically.
 315 *
 316 * The counter is 20 bits, and the state data is 12.
 317 */
 318#define RB_WRITE_MASK		0xfffff
 319#define RB_WRITE_INTCNT		(1 << 20)
 320
 321static void rb_init_page(struct buffer_data_page *bpage)
 322{
 323	local_set(&bpage->commit, 0);
 324}
 325
 326/**
 327 * ring_buffer_page_len - the size of data on the page.
 328 * @page: The page to read
 329 *
 330 * Returns the amount of data on the page, including buffer page header.
 331 */
 332size_t ring_buffer_page_len(void *page)
 333{
 334	return local_read(&((struct buffer_data_page *)page)->commit)
 335		+ BUF_PAGE_HDR_SIZE;
 336}
 337
 338/*
 339 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
 340 * this issue out.
 341 */
 342static void free_buffer_page(struct buffer_page *bpage)
 343{
 344	free_page((unsigned long)bpage->page);
 345	kfree(bpage);
 346}
 347
 348/*
 349 * We need to fit the time_stamp delta into 27 bits.
 350 */
 351static inline int test_time_stamp(u64 delta)
 352{
 353	if (delta & TS_DELTA_TEST)
 354		return 1;
 355	return 0;
 356}
 357
 358#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
 359
 360/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
 361#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
 362
 363int ring_buffer_print_page_header(struct trace_seq *s)
 364{
 365	struct buffer_data_page field;
 366
 367	trace_seq_printf(s, "\tfield: u64 timestamp;\t"
 368			 "offset:0;\tsize:%u;\tsigned:%u;\n",
 369			 (unsigned int)sizeof(field.time_stamp),
 370			 (unsigned int)is_signed_type(u64));
 371
 372	trace_seq_printf(s, "\tfield: local_t commit;\t"
 373			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 374			 (unsigned int)offsetof(typeof(field), commit),
 375			 (unsigned int)sizeof(field.commit),
 376			 (unsigned int)is_signed_type(long));
 377
 378	trace_seq_printf(s, "\tfield: int overwrite;\t"
 379			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 380			 (unsigned int)offsetof(typeof(field), commit),
 381			 1,
 382			 (unsigned int)is_signed_type(long));
 383
 384	trace_seq_printf(s, "\tfield: char data;\t"
 385			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 386			 (unsigned int)offsetof(typeof(field), data),
 387			 (unsigned int)BUF_PAGE_SIZE,
 388			 (unsigned int)is_signed_type(char));
 389
 390	return !trace_seq_has_overflowed(s);
 391}
 392
 393struct rb_irq_work {
 394	struct irq_work			work;
 395	wait_queue_head_t		waiters;
 396	wait_queue_head_t		full_waiters;
 397	bool				waiters_pending;
 398	bool				full_waiters_pending;
 399	bool				wakeup_full;
 400};
 401
 402/*
 403 * Structure to hold event state and handle nested events.
 404 */
 405struct rb_event_info {
 406	u64			ts;
 407	u64			delta;
 408	unsigned long		length;
 409	struct buffer_page	*tail_page;
 410	int			add_timestamp;
 411};
 412
 413/*
 414 * Used for which event context the event is in.
 415 *  NMI     = 0
 416 *  IRQ     = 1
 417 *  SOFTIRQ = 2
 418 *  NORMAL  = 3
 419 *
 420 * See trace_recursive_lock() comment below for more details.
 421 */
 422enum {
 423	RB_CTX_NMI,
 424	RB_CTX_IRQ,
 425	RB_CTX_SOFTIRQ,
 426	RB_CTX_NORMAL,
 427	RB_CTX_MAX
 428};
 429
 430/*
 431 * head_page == tail_page && head == tail then buffer is empty.
 432 */
 433struct ring_buffer_per_cpu {
 434	int				cpu;
 435	atomic_t			record_disabled;
 436	struct ring_buffer		*buffer;
 437	raw_spinlock_t			reader_lock;	/* serialize readers */
 438	arch_spinlock_t			lock;
 439	struct lock_class_key		lock_key;
 
 440	unsigned long			nr_pages;
 441	unsigned int			current_context;
 442	struct list_head		*pages;
 443	struct buffer_page		*head_page;	/* read from head */
 444	struct buffer_page		*tail_page;	/* write to tail */
 445	struct buffer_page		*commit_page;	/* committed pages */
 446	struct buffer_page		*reader_page;
 447	unsigned long			lost_events;
 448	unsigned long			last_overrun;
 
 449	local_t				entries_bytes;
 450	local_t				entries;
 451	local_t				overrun;
 452	local_t				commit_overrun;
 453	local_t				dropped_events;
 454	local_t				committing;
 455	local_t				commits;
 
 
 
 
 456	unsigned long			read;
 457	unsigned long			read_bytes;
 458	u64				write_stamp;
 459	u64				read_stamp;
 460	/* ring buffer pages to update, > 0 to add, < 0 to remove */
 461	long				nr_pages_to_update;
 462	struct list_head		new_pages; /* new pages to add */
 463	struct work_struct		update_pages_work;
 464	struct completion		update_done;
 465
 466	struct rb_irq_work		irq_work;
 467};
 468
 469struct ring_buffer {
 470	unsigned			flags;
 471	int				cpus;
 472	atomic_t			record_disabled;
 473	atomic_t			resize_disabled;
 474	cpumask_var_t			cpumask;
 475
 476	struct lock_class_key		*reader_lock_key;
 477
 478	struct mutex			mutex;
 479
 480	struct ring_buffer_per_cpu	**buffers;
 481
 482	struct hlist_node		node;
 483	u64				(*clock)(void);
 484
 485	struct rb_irq_work		irq_work;
 
 486};
 487
 488struct ring_buffer_iter {
 489	struct ring_buffer_per_cpu	*cpu_buffer;
 490	unsigned long			head;
 491	struct buffer_page		*head_page;
 492	struct buffer_page		*cache_reader_page;
 493	unsigned long			cache_read;
 494	u64				read_stamp;
 495};
 496
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 497/*
 498 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
 499 *
 500 * Schedules a delayed work to wake up any task that is blocked on the
 501 * ring buffer waiters queue.
 502 */
 503static void rb_wake_up_waiters(struct irq_work *work)
 504{
 505	struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
 506
 507	wake_up_all(&rbwork->waiters);
 508	if (rbwork->wakeup_full) {
 509		rbwork->wakeup_full = false;
 510		wake_up_all(&rbwork->full_waiters);
 511	}
 512}
 513
 514/**
 515 * ring_buffer_wait - wait for input to the ring buffer
 516 * @buffer: buffer to wait on
 517 * @cpu: the cpu buffer to wait on
 518 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
 519 *
 520 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 521 * as data is added to any of the @buffer's cpu buffers. Otherwise
 522 * it will wait for data to be added to a specific cpu buffer.
 523 */
 524int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
 525{
 526	struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
 527	DEFINE_WAIT(wait);
 528	struct rb_irq_work *work;
 529	int ret = 0;
 530
 531	/*
 532	 * Depending on what the caller is waiting for, either any
 533	 * data in any cpu buffer, or a specific buffer, put the
 534	 * caller on the appropriate wait queue.
 535	 */
 536	if (cpu == RING_BUFFER_ALL_CPUS) {
 537		work = &buffer->irq_work;
 538		/* Full only makes sense on per cpu reads */
 539		full = false;
 540	} else {
 541		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 542			return -ENODEV;
 543		cpu_buffer = buffer->buffers[cpu];
 544		work = &cpu_buffer->irq_work;
 545	}
 546
 547
 548	while (true) {
 549		if (full)
 550			prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
 551		else
 552			prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
 553
 554		/*
 555		 * The events can happen in critical sections where
 556		 * checking a work queue can cause deadlocks.
 557		 * After adding a task to the queue, this flag is set
 558		 * only to notify events to try to wake up the queue
 559		 * using irq_work.
 560		 *
 561		 * We don't clear it even if the buffer is no longer
 562		 * empty. The flag only causes the next event to run
 563		 * irq_work to do the work queue wake up. The worse
 564		 * that can happen if we race with !trace_empty() is that
 565		 * an event will cause an irq_work to try to wake up
 566		 * an empty queue.
 567		 *
 568		 * There's no reason to protect this flag either, as
 569		 * the work queue and irq_work logic will do the necessary
 570		 * synchronization for the wake ups. The only thing
 571		 * that is necessary is that the wake up happens after
 572		 * a task has been queued. It's OK for spurious wake ups.
 573		 */
 574		if (full)
 575			work->full_waiters_pending = true;
 576		else
 577			work->waiters_pending = true;
 578
 579		if (signal_pending(current)) {
 580			ret = -EINTR;
 581			break;
 582		}
 583
 584		if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
 585			break;
 586
 587		if (cpu != RING_BUFFER_ALL_CPUS &&
 588		    !ring_buffer_empty_cpu(buffer, cpu)) {
 589			unsigned long flags;
 590			bool pagebusy;
 
 
 591
 592			if (!full)
 593				break;
 594
 595			raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
 596			pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 
 
 
 
 
 597			raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
 598
 599			if (!pagebusy)
 600				break;
 601		}
 602
 603		schedule();
 604	}
 605
 606	if (full)
 607		finish_wait(&work->full_waiters, &wait);
 608	else
 609		finish_wait(&work->waiters, &wait);
 610
 611	return ret;
 612}
 613
 614/**
 615 * ring_buffer_poll_wait - poll on buffer input
 616 * @buffer: buffer to wait on
 617 * @cpu: the cpu buffer to wait on
 618 * @filp: the file descriptor
 619 * @poll_table: The poll descriptor
 620 *
 621 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 622 * as data is added to any of the @buffer's cpu buffers. Otherwise
 623 * it will wait for data to be added to a specific cpu buffer.
 624 *
 625 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
 626 * zero otherwise.
 627 */
 628int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
 629			  struct file *filp, poll_table *poll_table)
 630{
 631	struct ring_buffer_per_cpu *cpu_buffer;
 632	struct rb_irq_work *work;
 633
 634	if (cpu == RING_BUFFER_ALL_CPUS)
 635		work = &buffer->irq_work;
 636	else {
 637		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 638			return -EINVAL;
 639
 640		cpu_buffer = buffer->buffers[cpu];
 641		work = &cpu_buffer->irq_work;
 642	}
 643
 644	poll_wait(filp, &work->waiters, poll_table);
 645	work->waiters_pending = true;
 646	/*
 647	 * There's a tight race between setting the waiters_pending and
 648	 * checking if the ring buffer is empty.  Once the waiters_pending bit
 649	 * is set, the next event will wake the task up, but we can get stuck
 650	 * if there's only a single event in.
 651	 *
 652	 * FIXME: Ideally, we need a memory barrier on the writer side as well,
 653	 * but adding a memory barrier to all events will cause too much of a
 654	 * performance hit in the fast path.  We only need a memory barrier when
 655	 * the buffer goes from empty to having content.  But as this race is
 656	 * extremely small, and it's not a problem if another event comes in, we
 657	 * will fix it later.
 658	 */
 659	smp_mb();
 660
 661	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
 662	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
 663		return POLLIN | POLLRDNORM;
 664	return 0;
 665}
 666
 667/* buffer may be either ring_buffer or ring_buffer_per_cpu */
 668#define RB_WARN_ON(b, cond)						\
 669	({								\
 670		int _____ret = unlikely(cond);				\
 671		if (_____ret) {						\
 672			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
 673				struct ring_buffer_per_cpu *__b =	\
 674					(void *)b;			\
 675				atomic_inc(&__b->buffer->record_disabled); \
 676			} else						\
 677				atomic_inc(&b->record_disabled);	\
 678			WARN_ON(1);					\
 679		}							\
 680		_____ret;						\
 681	})
 682
 683/* Up this if you want to test the TIME_EXTENTS and normalization */
 684#define DEBUG_SHIFT 0
 685
 686static inline u64 rb_time_stamp(struct ring_buffer *buffer)
 687{
 688	/* shift to debug/test normalization and TIME_EXTENTS */
 689	return buffer->clock() << DEBUG_SHIFT;
 690}
 691
 692u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
 693{
 694	u64 time;
 695
 696	preempt_disable_notrace();
 697	time = rb_time_stamp(buffer);
 698	preempt_enable_no_resched_notrace();
 699
 700	return time;
 701}
 702EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
 703
 704void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
 705				      int cpu, u64 *ts)
 706{
 707	/* Just stupid testing the normalize function and deltas */
 708	*ts >>= DEBUG_SHIFT;
 709}
 710EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
 711
 712/*
 713 * Making the ring buffer lockless makes things tricky.
 714 * Although writes only happen on the CPU that they are on,
 715 * and they only need to worry about interrupts. Reads can
 716 * happen on any CPU.
 717 *
 718 * The reader page is always off the ring buffer, but when the
 719 * reader finishes with a page, it needs to swap its page with
 720 * a new one from the buffer. The reader needs to take from
 721 * the head (writes go to the tail). But if a writer is in overwrite
 722 * mode and wraps, it must push the head page forward.
 723 *
 724 * Here lies the problem.
 725 *
 726 * The reader must be careful to replace only the head page, and
 727 * not another one. As described at the top of the file in the
 728 * ASCII art, the reader sets its old page to point to the next
 729 * page after head. It then sets the page after head to point to
 730 * the old reader page. But if the writer moves the head page
 731 * during this operation, the reader could end up with the tail.
 732 *
 733 * We use cmpxchg to help prevent this race. We also do something
 734 * special with the page before head. We set the LSB to 1.
 735 *
 736 * When the writer must push the page forward, it will clear the
 737 * bit that points to the head page, move the head, and then set
 738 * the bit that points to the new head page.
 739 *
 740 * We also don't want an interrupt coming in and moving the head
 741 * page on another writer. Thus we use the second LSB to catch
 742 * that too. Thus:
 743 *
 744 * head->list->prev->next        bit 1          bit 0
 745 *                              -------        -------
 746 * Normal page                     0              0
 747 * Points to head page             0              1
 748 * New head page                   1              0
 749 *
 750 * Note we can not trust the prev pointer of the head page, because:
 751 *
 752 * +----+       +-----+        +-----+
 753 * |    |------>|  T  |---X--->|  N  |
 754 * |    |<------|     |        |     |
 755 * +----+       +-----+        +-----+
 756 *   ^                           ^ |
 757 *   |          +-----+          | |
 758 *   +----------|  R  |----------+ |
 759 *              |     |<-----------+
 760 *              +-----+
 761 *
 762 * Key:  ---X-->  HEAD flag set in pointer
 763 *         T      Tail page
 764 *         R      Reader page
 765 *         N      Next page
 766 *
 767 * (see __rb_reserve_next() to see where this happens)
 768 *
 769 *  What the above shows is that the reader just swapped out
 770 *  the reader page with a page in the buffer, but before it
 771 *  could make the new header point back to the new page added
 772 *  it was preempted by a writer. The writer moved forward onto
 773 *  the new page added by the reader and is about to move forward
 774 *  again.
 775 *
 776 *  You can see, it is legitimate for the previous pointer of
 777 *  the head (or any page) not to point back to itself. But only
 778 *  temporarially.
 779 */
 780
 781#define RB_PAGE_NORMAL		0UL
 782#define RB_PAGE_HEAD		1UL
 783#define RB_PAGE_UPDATE		2UL
 784
 785
 786#define RB_FLAG_MASK		3UL
 787
 788/* PAGE_MOVED is not part of the mask */
 789#define RB_PAGE_MOVED		4UL
 790
 791/*
 792 * rb_list_head - remove any bit
 793 */
 794static struct list_head *rb_list_head(struct list_head *list)
 795{
 796	unsigned long val = (unsigned long)list;
 797
 798	return (struct list_head *)(val & ~RB_FLAG_MASK);
 799}
 800
 801/*
 802 * rb_is_head_page - test if the given page is the head page
 803 *
 804 * Because the reader may move the head_page pointer, we can
 805 * not trust what the head page is (it may be pointing to
 806 * the reader page). But if the next page is a header page,
 807 * its flags will be non zero.
 808 */
 809static inline int
 810rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
 811		struct buffer_page *page, struct list_head *list)
 812{
 813	unsigned long val;
 814
 815	val = (unsigned long)list->next;
 816
 817	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
 818		return RB_PAGE_MOVED;
 819
 820	return val & RB_FLAG_MASK;
 821}
 822
 823/*
 824 * rb_is_reader_page
 825 *
 826 * The unique thing about the reader page, is that, if the
 827 * writer is ever on it, the previous pointer never points
 828 * back to the reader page.
 829 */
 830static bool rb_is_reader_page(struct buffer_page *page)
 831{
 832	struct list_head *list = page->list.prev;
 833
 834	return rb_list_head(list->next) != &page->list;
 835}
 836
 837/*
 838 * rb_set_list_to_head - set a list_head to be pointing to head.
 839 */
 840static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
 841				struct list_head *list)
 842{
 843	unsigned long *ptr;
 844
 845	ptr = (unsigned long *)&list->next;
 846	*ptr |= RB_PAGE_HEAD;
 847	*ptr &= ~RB_PAGE_UPDATE;
 848}
 849
 850/*
 851 * rb_head_page_activate - sets up head page
 852 */
 853static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
 854{
 855	struct buffer_page *head;
 856
 857	head = cpu_buffer->head_page;
 858	if (!head)
 859		return;
 860
 861	/*
 862	 * Set the previous list pointer to have the HEAD flag.
 863	 */
 864	rb_set_list_to_head(cpu_buffer, head->list.prev);
 865}
 866
 867static void rb_list_head_clear(struct list_head *list)
 868{
 869	unsigned long *ptr = (unsigned long *)&list->next;
 870
 871	*ptr &= ~RB_FLAG_MASK;
 872}
 873
 874/*
 875 * rb_head_page_dactivate - clears head page ptr (for free list)
 876 */
 877static void
 878rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
 879{
 880	struct list_head *hd;
 881
 882	/* Go through the whole list and clear any pointers found. */
 883	rb_list_head_clear(cpu_buffer->pages);
 884
 885	list_for_each(hd, cpu_buffer->pages)
 886		rb_list_head_clear(hd);
 887}
 888
 889static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
 890			    struct buffer_page *head,
 891			    struct buffer_page *prev,
 892			    int old_flag, int new_flag)
 893{
 894	struct list_head *list;
 895	unsigned long val = (unsigned long)&head->list;
 896	unsigned long ret;
 897
 898	list = &prev->list;
 899
 900	val &= ~RB_FLAG_MASK;
 901
 902	ret = cmpxchg((unsigned long *)&list->next,
 903		      val | old_flag, val | new_flag);
 904
 905	/* check if the reader took the page */
 906	if ((ret & ~RB_FLAG_MASK) != val)
 907		return RB_PAGE_MOVED;
 908
 909	return ret & RB_FLAG_MASK;
 910}
 911
 912static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
 913				   struct buffer_page *head,
 914				   struct buffer_page *prev,
 915				   int old_flag)
 916{
 917	return rb_head_page_set(cpu_buffer, head, prev,
 918				old_flag, RB_PAGE_UPDATE);
 919}
 920
 921static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
 922				 struct buffer_page *head,
 923				 struct buffer_page *prev,
 924				 int old_flag)
 925{
 926	return rb_head_page_set(cpu_buffer, head, prev,
 927				old_flag, RB_PAGE_HEAD);
 928}
 929
 930static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
 931				   struct buffer_page *head,
 932				   struct buffer_page *prev,
 933				   int old_flag)
 934{
 935	return rb_head_page_set(cpu_buffer, head, prev,
 936				old_flag, RB_PAGE_NORMAL);
 937}
 938
 939static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
 940			       struct buffer_page **bpage)
 941{
 942	struct list_head *p = rb_list_head((*bpage)->list.next);
 943
 944	*bpage = list_entry(p, struct buffer_page, list);
 945}
 946
 947static struct buffer_page *
 948rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
 949{
 950	struct buffer_page *head;
 951	struct buffer_page *page;
 952	struct list_head *list;
 953	int i;
 954
 955	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
 956		return NULL;
 957
 958	/* sanity check */
 959	list = cpu_buffer->pages;
 960	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
 961		return NULL;
 962
 963	page = head = cpu_buffer->head_page;
 964	/*
 965	 * It is possible that the writer moves the header behind
 966	 * where we started, and we miss in one loop.
 967	 * A second loop should grab the header, but we'll do
 968	 * three loops just because I'm paranoid.
 969	 */
 970	for (i = 0; i < 3; i++) {
 971		do {
 972			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
 973				cpu_buffer->head_page = page;
 974				return page;
 975			}
 976			rb_inc_page(cpu_buffer, &page);
 977		} while (page != head);
 978	}
 979
 980	RB_WARN_ON(cpu_buffer, 1);
 981
 982	return NULL;
 983}
 984
 985static int rb_head_page_replace(struct buffer_page *old,
 986				struct buffer_page *new)
 987{
 988	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
 989	unsigned long val;
 990	unsigned long ret;
 991
 992	val = *ptr & ~RB_FLAG_MASK;
 993	val |= RB_PAGE_HEAD;
 994
 995	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
 996
 997	return ret == val;
 998}
 999
1000/*
1001 * rb_tail_page_update - move the tail page forward
1002 */
1003static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1004			       struct buffer_page *tail_page,
1005			       struct buffer_page *next_page)
1006{
1007	unsigned long old_entries;
1008	unsigned long old_write;
1009
1010	/*
1011	 * The tail page now needs to be moved forward.
1012	 *
1013	 * We need to reset the tail page, but without messing
1014	 * with possible erasing of data brought in by interrupts
1015	 * that have moved the tail page and are currently on it.
1016	 *
1017	 * We add a counter to the write field to denote this.
1018	 */
1019	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1020	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1021
 
1022	/*
1023	 * Just make sure we have seen our old_write and synchronize
1024	 * with any interrupts that come in.
1025	 */
1026	barrier();
1027
1028	/*
1029	 * If the tail page is still the same as what we think
1030	 * it is, then it is up to us to update the tail
1031	 * pointer.
1032	 */
1033	if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1034		/* Zero the write counter */
1035		unsigned long val = old_write & ~RB_WRITE_MASK;
1036		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1037
1038		/*
1039		 * This will only succeed if an interrupt did
1040		 * not come in and change it. In which case, we
1041		 * do not want to modify it.
1042		 *
1043		 * We add (void) to let the compiler know that we do not care
1044		 * about the return value of these functions. We use the
1045		 * cmpxchg to only update if an interrupt did not already
1046		 * do it for us. If the cmpxchg fails, we don't care.
1047		 */
1048		(void)local_cmpxchg(&next_page->write, old_write, val);
1049		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1050
1051		/*
1052		 * No need to worry about races with clearing out the commit.
1053		 * it only can increment when a commit takes place. But that
1054		 * only happens in the outer most nested commit.
1055		 */
1056		local_set(&next_page->page->commit, 0);
1057
1058		/* Again, either we update tail_page or an interrupt does */
1059		(void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1060	}
1061}
1062
1063static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1064			  struct buffer_page *bpage)
1065{
1066	unsigned long val = (unsigned long)bpage;
1067
1068	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1069		return 1;
1070
1071	return 0;
1072}
1073
1074/**
1075 * rb_check_list - make sure a pointer to a list has the last bits zero
1076 */
1077static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1078			 struct list_head *list)
1079{
1080	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1081		return 1;
1082	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1083		return 1;
1084	return 0;
1085}
1086
1087/**
1088 * rb_check_pages - integrity check of buffer pages
1089 * @cpu_buffer: CPU buffer with pages to test
1090 *
1091 * As a safety measure we check to make sure the data pages have not
1092 * been corrupted.
1093 */
1094static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1095{
1096	struct list_head *head = cpu_buffer->pages;
1097	struct buffer_page *bpage, *tmp;
1098
1099	/* Reset the head page if it exists */
1100	if (cpu_buffer->head_page)
1101		rb_set_head_page(cpu_buffer);
1102
1103	rb_head_page_deactivate(cpu_buffer);
1104
1105	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1106		return -1;
1107	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1108		return -1;
1109
1110	if (rb_check_list(cpu_buffer, head))
1111		return -1;
1112
1113	list_for_each_entry_safe(bpage, tmp, head, list) {
1114		if (RB_WARN_ON(cpu_buffer,
1115			       bpage->list.next->prev != &bpage->list))
1116			return -1;
1117		if (RB_WARN_ON(cpu_buffer,
1118			       bpage->list.prev->next != &bpage->list))
1119			return -1;
1120		if (rb_check_list(cpu_buffer, &bpage->list))
1121			return -1;
1122	}
1123
1124	rb_head_page_activate(cpu_buffer);
1125
1126	return 0;
1127}
1128
1129static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1130{
1131	struct buffer_page *bpage, *tmp;
 
 
1132	long i;
1133
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1134	for (i = 0; i < nr_pages; i++) {
1135		struct page *page;
1136		/*
1137		 * __GFP_NORETRY flag makes sure that the allocation fails
1138		 * gracefully without invoking oom-killer and the system is
1139		 * not destabilized.
1140		 */
1141		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1142				    GFP_KERNEL | __GFP_NORETRY,
1143				    cpu_to_node(cpu));
1144		if (!bpage)
1145			goto free_pages;
1146
1147		list_add(&bpage->list, pages);
1148
1149		page = alloc_pages_node(cpu_to_node(cpu),
1150					GFP_KERNEL | __GFP_NORETRY, 0);
1151		if (!page)
1152			goto free_pages;
1153		bpage->page = page_address(page);
1154		rb_init_page(bpage->page);
 
 
 
1155	}
 
 
1156
1157	return 0;
1158
1159free_pages:
1160	list_for_each_entry_safe(bpage, tmp, pages, list) {
1161		list_del_init(&bpage->list);
1162		free_buffer_page(bpage);
1163	}
 
 
1164
1165	return -ENOMEM;
1166}
1167
1168static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1169			     unsigned long nr_pages)
1170{
1171	LIST_HEAD(pages);
1172
1173	WARN_ON(!nr_pages);
1174
1175	if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1176		return -ENOMEM;
1177
1178	/*
1179	 * The ring buffer page list is a circular list that does not
1180	 * start and end with a list head. All page list items point to
1181	 * other pages.
1182	 */
1183	cpu_buffer->pages = pages.next;
1184	list_del(&pages);
1185
1186	cpu_buffer->nr_pages = nr_pages;
1187
1188	rb_check_pages(cpu_buffer);
1189
1190	return 0;
1191}
1192
1193static struct ring_buffer_per_cpu *
1194rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1195{
1196	struct ring_buffer_per_cpu *cpu_buffer;
1197	struct buffer_page *bpage;
1198	struct page *page;
1199	int ret;
1200
1201	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1202				  GFP_KERNEL, cpu_to_node(cpu));
1203	if (!cpu_buffer)
1204		return NULL;
1205
1206	cpu_buffer->cpu = cpu;
1207	cpu_buffer->buffer = buffer;
1208	raw_spin_lock_init(&cpu_buffer->reader_lock);
1209	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1210	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1211	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1212	init_completion(&cpu_buffer->update_done);
1213	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1214	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1215	init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1216
1217	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1218			    GFP_KERNEL, cpu_to_node(cpu));
1219	if (!bpage)
1220		goto fail_free_buffer;
1221
1222	rb_check_bpage(cpu_buffer, bpage);
1223
1224	cpu_buffer->reader_page = bpage;
1225	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1226	if (!page)
1227		goto fail_free_reader;
1228	bpage->page = page_address(page);
1229	rb_init_page(bpage->page);
1230
1231	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1232	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1233
1234	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1235	if (ret < 0)
1236		goto fail_free_reader;
1237
1238	cpu_buffer->head_page
1239		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1240	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1241
1242	rb_head_page_activate(cpu_buffer);
1243
1244	return cpu_buffer;
1245
1246 fail_free_reader:
1247	free_buffer_page(cpu_buffer->reader_page);
1248
1249 fail_free_buffer:
1250	kfree(cpu_buffer);
1251	return NULL;
1252}
1253
1254static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1255{
1256	struct list_head *head = cpu_buffer->pages;
1257	struct buffer_page *bpage, *tmp;
1258
1259	free_buffer_page(cpu_buffer->reader_page);
1260
1261	rb_head_page_deactivate(cpu_buffer);
1262
1263	if (head) {
1264		list_for_each_entry_safe(bpage, tmp, head, list) {
1265			list_del_init(&bpage->list);
1266			free_buffer_page(bpage);
1267		}
1268		bpage = list_entry(head, struct buffer_page, list);
1269		free_buffer_page(bpage);
1270	}
1271
1272	kfree(cpu_buffer);
1273}
1274
1275/**
1276 * __ring_buffer_alloc - allocate a new ring_buffer
1277 * @size: the size in bytes per cpu that is needed.
1278 * @flags: attributes to set for the ring buffer.
1279 *
1280 * Currently the only flag that is available is the RB_FL_OVERWRITE
1281 * flag. This flag means that the buffer will overwrite old data
1282 * when the buffer wraps. If this flag is not set, the buffer will
1283 * drop data when the tail hits the head.
1284 */
1285struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1286					struct lock_class_key *key)
1287{
1288	struct ring_buffer *buffer;
1289	long nr_pages;
1290	int bsize;
1291	int cpu;
1292	int ret;
1293
1294	/* keep it in its own cache line */
1295	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1296			 GFP_KERNEL);
1297	if (!buffer)
1298		return NULL;
1299
1300	if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1301		goto fail_free_buffer;
1302
1303	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1304	buffer->flags = flags;
1305	buffer->clock = trace_clock_local;
1306	buffer->reader_lock_key = key;
1307
1308	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1309	init_waitqueue_head(&buffer->irq_work.waiters);
1310
1311	/* need at least two pages */
1312	if (nr_pages < 2)
1313		nr_pages = 2;
1314
1315	buffer->cpus = nr_cpu_ids;
1316
1317	bsize = sizeof(void *) * nr_cpu_ids;
1318	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1319				  GFP_KERNEL);
1320	if (!buffer->buffers)
1321		goto fail_free_cpumask;
1322
1323	cpu = raw_smp_processor_id();
1324	cpumask_set_cpu(cpu, buffer->cpumask);
1325	buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1326	if (!buffer->buffers[cpu])
1327		goto fail_free_buffers;
1328
1329	ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1330	if (ret < 0)
1331		goto fail_free_buffers;
1332
1333	mutex_init(&buffer->mutex);
1334
1335	return buffer;
1336
1337 fail_free_buffers:
1338	for_each_buffer_cpu(buffer, cpu) {
1339		if (buffer->buffers[cpu])
1340			rb_free_cpu_buffer(buffer->buffers[cpu]);
1341	}
1342	kfree(buffer->buffers);
1343
1344 fail_free_cpumask:
1345	free_cpumask_var(buffer->cpumask);
1346
1347 fail_free_buffer:
1348	kfree(buffer);
1349	return NULL;
1350}
1351EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1352
1353/**
1354 * ring_buffer_free - free a ring buffer.
1355 * @buffer: the buffer to free.
1356 */
1357void
1358ring_buffer_free(struct ring_buffer *buffer)
1359{
1360	int cpu;
1361
1362	cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1363
1364	for_each_buffer_cpu(buffer, cpu)
1365		rb_free_cpu_buffer(buffer->buffers[cpu]);
1366
1367	kfree(buffer->buffers);
1368	free_cpumask_var(buffer->cpumask);
1369
1370	kfree(buffer);
1371}
1372EXPORT_SYMBOL_GPL(ring_buffer_free);
1373
1374void ring_buffer_set_clock(struct ring_buffer *buffer,
1375			   u64 (*clock)(void))
1376{
1377	buffer->clock = clock;
1378}
1379
 
 
 
 
 
 
 
 
 
 
1380static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1381
1382static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1383{
1384	return local_read(&bpage->entries) & RB_WRITE_MASK;
1385}
1386
1387static inline unsigned long rb_page_write(struct buffer_page *bpage)
1388{
1389	return local_read(&bpage->write) & RB_WRITE_MASK;
1390}
1391
1392static int
1393rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1394{
1395	struct list_head *tail_page, *to_remove, *next_page;
1396	struct buffer_page *to_remove_page, *tmp_iter_page;
1397	struct buffer_page *last_page, *first_page;
1398	unsigned long nr_removed;
1399	unsigned long head_bit;
1400	int page_entries;
1401
1402	head_bit = 0;
1403
1404	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1405	atomic_inc(&cpu_buffer->record_disabled);
1406	/*
1407	 * We don't race with the readers since we have acquired the reader
1408	 * lock. We also don't race with writers after disabling recording.
1409	 * This makes it easy to figure out the first and the last page to be
1410	 * removed from the list. We unlink all the pages in between including
1411	 * the first and last pages. This is done in a busy loop so that we
1412	 * lose the least number of traces.
1413	 * The pages are freed after we restart recording and unlock readers.
1414	 */
1415	tail_page = &cpu_buffer->tail_page->list;
1416
1417	/*
1418	 * tail page might be on reader page, we remove the next page
1419	 * from the ring buffer
1420	 */
1421	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1422		tail_page = rb_list_head(tail_page->next);
1423	to_remove = tail_page;
1424
1425	/* start of pages to remove */
1426	first_page = list_entry(rb_list_head(to_remove->next),
1427				struct buffer_page, list);
1428
1429	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1430		to_remove = rb_list_head(to_remove)->next;
1431		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1432	}
1433
1434	next_page = rb_list_head(to_remove)->next;
1435
1436	/*
1437	 * Now we remove all pages between tail_page and next_page.
1438	 * Make sure that we have head_bit value preserved for the
1439	 * next page
1440	 */
1441	tail_page->next = (struct list_head *)((unsigned long)next_page |
1442						head_bit);
1443	next_page = rb_list_head(next_page);
1444	next_page->prev = tail_page;
1445
1446	/* make sure pages points to a valid page in the ring buffer */
1447	cpu_buffer->pages = next_page;
1448
1449	/* update head page */
1450	if (head_bit)
1451		cpu_buffer->head_page = list_entry(next_page,
1452						struct buffer_page, list);
1453
1454	/*
1455	 * change read pointer to make sure any read iterators reset
1456	 * themselves
1457	 */
1458	cpu_buffer->read = 0;
1459
1460	/* pages are removed, resume tracing and then free the pages */
1461	atomic_dec(&cpu_buffer->record_disabled);
1462	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1463
1464	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1465
1466	/* last buffer page to remove */
1467	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1468				list);
1469	tmp_iter_page = first_page;
1470
1471	do {
 
 
1472		to_remove_page = tmp_iter_page;
1473		rb_inc_page(cpu_buffer, &tmp_iter_page);
1474
1475		/* update the counters */
1476		page_entries = rb_page_entries(to_remove_page);
1477		if (page_entries) {
1478			/*
1479			 * If something was added to this page, it was full
1480			 * since it is not the tail page. So we deduct the
1481			 * bytes consumed in ring buffer from here.
1482			 * Increment overrun to account for the lost events.
1483			 */
1484			local_add(page_entries, &cpu_buffer->overrun);
1485			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1486		}
1487
1488		/*
1489		 * We have already removed references to this list item, just
1490		 * free up the buffer_page and its page
1491		 */
1492		free_buffer_page(to_remove_page);
1493		nr_removed--;
1494
1495	} while (to_remove_page != last_page);
1496
1497	RB_WARN_ON(cpu_buffer, nr_removed);
1498
1499	return nr_removed == 0;
1500}
1501
1502static int
1503rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1504{
1505	struct list_head *pages = &cpu_buffer->new_pages;
1506	int retries, success;
1507
1508	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1509	/*
1510	 * We are holding the reader lock, so the reader page won't be swapped
1511	 * in the ring buffer. Now we are racing with the writer trying to
1512	 * move head page and the tail page.
1513	 * We are going to adapt the reader page update process where:
1514	 * 1. We first splice the start and end of list of new pages between
1515	 *    the head page and its previous page.
1516	 * 2. We cmpxchg the prev_page->next to point from head page to the
1517	 *    start of new pages list.
1518	 * 3. Finally, we update the head->prev to the end of new list.
1519	 *
1520	 * We will try this process 10 times, to make sure that we don't keep
1521	 * spinning.
1522	 */
1523	retries = 10;
1524	success = 0;
1525	while (retries--) {
1526		struct list_head *head_page, *prev_page, *r;
1527		struct list_head *last_page, *first_page;
1528		struct list_head *head_page_with_bit;
1529
1530		head_page = &rb_set_head_page(cpu_buffer)->list;
1531		if (!head_page)
1532			break;
1533		prev_page = head_page->prev;
1534
1535		first_page = pages->next;
1536		last_page  = pages->prev;
1537
1538		head_page_with_bit = (struct list_head *)
1539				     ((unsigned long)head_page | RB_PAGE_HEAD);
1540
1541		last_page->next = head_page_with_bit;
1542		first_page->prev = prev_page;
1543
1544		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1545
1546		if (r == head_page_with_bit) {
1547			/*
1548			 * yay, we replaced the page pointer to our new list,
1549			 * now, we just have to update to head page's prev
1550			 * pointer to point to end of list
1551			 */
1552			head_page->prev = last_page;
1553			success = 1;
1554			break;
1555		}
1556	}
1557
1558	if (success)
1559		INIT_LIST_HEAD(pages);
1560	/*
1561	 * If we weren't successful in adding in new pages, warn and stop
1562	 * tracing
1563	 */
1564	RB_WARN_ON(cpu_buffer, !success);
1565	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1566
1567	/* free pages if they weren't inserted */
1568	if (!success) {
1569		struct buffer_page *bpage, *tmp;
1570		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1571					 list) {
1572			list_del_init(&bpage->list);
1573			free_buffer_page(bpage);
1574		}
1575	}
1576	return success;
1577}
1578
1579static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1580{
1581	int success;
1582
1583	if (cpu_buffer->nr_pages_to_update > 0)
1584		success = rb_insert_pages(cpu_buffer);
1585	else
1586		success = rb_remove_pages(cpu_buffer,
1587					-cpu_buffer->nr_pages_to_update);
1588
1589	if (success)
1590		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1591}
1592
1593static void update_pages_handler(struct work_struct *work)
1594{
1595	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1596			struct ring_buffer_per_cpu, update_pages_work);
1597	rb_update_pages(cpu_buffer);
1598	complete(&cpu_buffer->update_done);
1599}
1600
1601/**
1602 * ring_buffer_resize - resize the ring buffer
1603 * @buffer: the buffer to resize.
1604 * @size: the new size.
1605 * @cpu_id: the cpu buffer to resize
1606 *
1607 * Minimum size is 2 * BUF_PAGE_SIZE.
1608 *
1609 * Returns 0 on success and < 0 on failure.
1610 */
1611int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1612			int cpu_id)
1613{
1614	struct ring_buffer_per_cpu *cpu_buffer;
1615	unsigned long nr_pages;
1616	int cpu, err = 0;
1617
1618	/*
1619	 * Always succeed at resizing a non-existent buffer:
1620	 */
1621	if (!buffer)
1622		return size;
1623
1624	/* Make sure the requested buffer exists */
1625	if (cpu_id != RING_BUFFER_ALL_CPUS &&
1626	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
1627		return size;
1628
1629	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1630
1631	/* we need a minimum of two pages */
1632	if (nr_pages < 2)
1633		nr_pages = 2;
1634
1635	size = nr_pages * BUF_PAGE_SIZE;
1636
1637	/*
1638	 * Don't succeed if resizing is disabled, as a reader might be
1639	 * manipulating the ring buffer and is expecting a sane state while
1640	 * this is true.
1641	 */
1642	if (atomic_read(&buffer->resize_disabled))
1643		return -EBUSY;
1644
1645	/* prevent another thread from changing buffer sizes */
1646	mutex_lock(&buffer->mutex);
1647
1648	if (cpu_id == RING_BUFFER_ALL_CPUS) {
1649		/* calculate the pages to update */
1650		for_each_buffer_cpu(buffer, cpu) {
1651			cpu_buffer = buffer->buffers[cpu];
1652
1653			cpu_buffer->nr_pages_to_update = nr_pages -
1654							cpu_buffer->nr_pages;
1655			/*
1656			 * nothing more to do for removing pages or no update
1657			 */
1658			if (cpu_buffer->nr_pages_to_update <= 0)
1659				continue;
1660			/*
1661			 * to add pages, make sure all new pages can be
1662			 * allocated without receiving ENOMEM
1663			 */
1664			INIT_LIST_HEAD(&cpu_buffer->new_pages);
1665			if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1666						&cpu_buffer->new_pages, cpu)) {
1667				/* not enough memory for new pages */
1668				err = -ENOMEM;
1669				goto out_err;
1670			}
1671		}
1672
1673		get_online_cpus();
1674		/*
1675		 * Fire off all the required work handlers
1676		 * We can't schedule on offline CPUs, but it's not necessary
1677		 * since we can change their buffer sizes without any race.
1678		 */
1679		for_each_buffer_cpu(buffer, cpu) {
1680			cpu_buffer = buffer->buffers[cpu];
1681			if (!cpu_buffer->nr_pages_to_update)
1682				continue;
1683
1684			/* Can't run something on an offline CPU. */
1685			if (!cpu_online(cpu)) {
1686				rb_update_pages(cpu_buffer);
1687				cpu_buffer->nr_pages_to_update = 0;
1688			} else {
1689				schedule_work_on(cpu,
1690						&cpu_buffer->update_pages_work);
1691			}
1692		}
1693
1694		/* wait for all the updates to complete */
1695		for_each_buffer_cpu(buffer, cpu) {
1696			cpu_buffer = buffer->buffers[cpu];
1697			if (!cpu_buffer->nr_pages_to_update)
1698				continue;
1699
1700			if (cpu_online(cpu))
1701				wait_for_completion(&cpu_buffer->update_done);
1702			cpu_buffer->nr_pages_to_update = 0;
1703		}
1704
1705		put_online_cpus();
1706	} else {
1707		/* Make sure this CPU has been intitialized */
1708		if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1709			goto out;
1710
1711		cpu_buffer = buffer->buffers[cpu_id];
1712
1713		if (nr_pages == cpu_buffer->nr_pages)
1714			goto out;
1715
1716		cpu_buffer->nr_pages_to_update = nr_pages -
1717						cpu_buffer->nr_pages;
1718
1719		INIT_LIST_HEAD(&cpu_buffer->new_pages);
1720		if (cpu_buffer->nr_pages_to_update > 0 &&
1721			__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1722					    &cpu_buffer->new_pages, cpu_id)) {
1723			err = -ENOMEM;
1724			goto out_err;
1725		}
1726
1727		get_online_cpus();
1728
1729		/* Can't run something on an offline CPU. */
1730		if (!cpu_online(cpu_id))
1731			rb_update_pages(cpu_buffer);
1732		else {
1733			schedule_work_on(cpu_id,
1734					 &cpu_buffer->update_pages_work);
1735			wait_for_completion(&cpu_buffer->update_done);
1736		}
1737
1738		cpu_buffer->nr_pages_to_update = 0;
1739		put_online_cpus();
1740	}
1741
1742 out:
1743	/*
1744	 * The ring buffer resize can happen with the ring buffer
1745	 * enabled, so that the update disturbs the tracing as little
1746	 * as possible. But if the buffer is disabled, we do not need
1747	 * to worry about that, and we can take the time to verify
1748	 * that the buffer is not corrupt.
1749	 */
1750	if (atomic_read(&buffer->record_disabled)) {
1751		atomic_inc(&buffer->record_disabled);
1752		/*
1753		 * Even though the buffer was disabled, we must make sure
1754		 * that it is truly disabled before calling rb_check_pages.
1755		 * There could have been a race between checking
1756		 * record_disable and incrementing it.
1757		 */
1758		synchronize_sched();
1759		for_each_buffer_cpu(buffer, cpu) {
1760			cpu_buffer = buffer->buffers[cpu];
1761			rb_check_pages(cpu_buffer);
1762		}
1763		atomic_dec(&buffer->record_disabled);
1764	}
1765
1766	mutex_unlock(&buffer->mutex);
1767	return size;
1768
1769 out_err:
1770	for_each_buffer_cpu(buffer, cpu) {
1771		struct buffer_page *bpage, *tmp;
1772
1773		cpu_buffer = buffer->buffers[cpu];
1774		cpu_buffer->nr_pages_to_update = 0;
1775
1776		if (list_empty(&cpu_buffer->new_pages))
1777			continue;
1778
1779		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1780					list) {
1781			list_del_init(&bpage->list);
1782			free_buffer_page(bpage);
1783		}
1784	}
1785	mutex_unlock(&buffer->mutex);
1786	return err;
1787}
1788EXPORT_SYMBOL_GPL(ring_buffer_resize);
1789
1790void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1791{
1792	mutex_lock(&buffer->mutex);
1793	if (val)
1794		buffer->flags |= RB_FL_OVERWRITE;
1795	else
1796		buffer->flags &= ~RB_FL_OVERWRITE;
1797	mutex_unlock(&buffer->mutex);
1798}
1799EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1800
1801static __always_inline void *
1802__rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1803{
1804	return bpage->data + index;
1805}
1806
1807static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1808{
1809	return bpage->page->data + index;
1810}
1811
1812static __always_inline struct ring_buffer_event *
1813rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1814{
1815	return __rb_page_index(cpu_buffer->reader_page,
1816			       cpu_buffer->reader_page->read);
1817}
1818
1819static __always_inline struct ring_buffer_event *
1820rb_iter_head_event(struct ring_buffer_iter *iter)
1821{
1822	return __rb_page_index(iter->head_page, iter->head);
1823}
1824
1825static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1826{
1827	return local_read(&bpage->page->commit);
1828}
1829
1830/* Size is determined by what has been committed */
1831static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1832{
1833	return rb_page_commit(bpage);
1834}
1835
1836static __always_inline unsigned
1837rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1838{
1839	return rb_page_commit(cpu_buffer->commit_page);
1840}
1841
1842static __always_inline unsigned
1843rb_event_index(struct ring_buffer_event *event)
1844{
1845	unsigned long addr = (unsigned long)event;
1846
1847	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1848}
1849
1850static void rb_inc_iter(struct ring_buffer_iter *iter)
1851{
1852	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1853
1854	/*
1855	 * The iterator could be on the reader page (it starts there).
1856	 * But the head could have moved, since the reader was
1857	 * found. Check for this case and assign the iterator
1858	 * to the head page instead of next.
1859	 */
1860	if (iter->head_page == cpu_buffer->reader_page)
1861		iter->head_page = rb_set_head_page(cpu_buffer);
1862	else
1863		rb_inc_page(cpu_buffer, &iter->head_page);
1864
1865	iter->read_stamp = iter->head_page->page->time_stamp;
1866	iter->head = 0;
1867}
1868
1869/*
1870 * rb_handle_head_page - writer hit the head page
1871 *
1872 * Returns: +1 to retry page
1873 *           0 to continue
1874 *          -1 on error
1875 */
1876static int
1877rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1878		    struct buffer_page *tail_page,
1879		    struct buffer_page *next_page)
1880{
1881	struct buffer_page *new_head;
1882	int entries;
1883	int type;
1884	int ret;
1885
1886	entries = rb_page_entries(next_page);
1887
1888	/*
1889	 * The hard part is here. We need to move the head
1890	 * forward, and protect against both readers on
1891	 * other CPUs and writers coming in via interrupts.
1892	 */
1893	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1894				       RB_PAGE_HEAD);
1895
1896	/*
1897	 * type can be one of four:
1898	 *  NORMAL - an interrupt already moved it for us
1899	 *  HEAD   - we are the first to get here.
1900	 *  UPDATE - we are the interrupt interrupting
1901	 *           a current move.
1902	 *  MOVED  - a reader on another CPU moved the next
1903	 *           pointer to its reader page. Give up
1904	 *           and try again.
1905	 */
1906
1907	switch (type) {
1908	case RB_PAGE_HEAD:
1909		/*
1910		 * We changed the head to UPDATE, thus
1911		 * it is our responsibility to update
1912		 * the counters.
1913		 */
1914		local_add(entries, &cpu_buffer->overrun);
1915		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1916
1917		/*
1918		 * The entries will be zeroed out when we move the
1919		 * tail page.
1920		 */
1921
1922		/* still more to do */
1923		break;
1924
1925	case RB_PAGE_UPDATE:
1926		/*
1927		 * This is an interrupt that interrupt the
1928		 * previous update. Still more to do.
1929		 */
1930		break;
1931	case RB_PAGE_NORMAL:
1932		/*
1933		 * An interrupt came in before the update
1934		 * and processed this for us.
1935		 * Nothing left to do.
1936		 */
1937		return 1;
1938	case RB_PAGE_MOVED:
1939		/*
1940		 * The reader is on another CPU and just did
1941		 * a swap with our next_page.
1942		 * Try again.
1943		 */
1944		return 1;
1945	default:
1946		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1947		return -1;
1948	}
1949
1950	/*
1951	 * Now that we are here, the old head pointer is
1952	 * set to UPDATE. This will keep the reader from
1953	 * swapping the head page with the reader page.
1954	 * The reader (on another CPU) will spin till
1955	 * we are finished.
1956	 *
1957	 * We just need to protect against interrupts
1958	 * doing the job. We will set the next pointer
1959	 * to HEAD. After that, we set the old pointer
1960	 * to NORMAL, but only if it was HEAD before.
1961	 * otherwise we are an interrupt, and only
1962	 * want the outer most commit to reset it.
1963	 */
1964	new_head = next_page;
1965	rb_inc_page(cpu_buffer, &new_head);
1966
1967	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1968				    RB_PAGE_NORMAL);
1969
1970	/*
1971	 * Valid returns are:
1972	 *  HEAD   - an interrupt came in and already set it.
1973	 *  NORMAL - One of two things:
1974	 *            1) We really set it.
1975	 *            2) A bunch of interrupts came in and moved
1976	 *               the page forward again.
1977	 */
1978	switch (ret) {
1979	case RB_PAGE_HEAD:
1980	case RB_PAGE_NORMAL:
1981		/* OK */
1982		break;
1983	default:
1984		RB_WARN_ON(cpu_buffer, 1);
1985		return -1;
1986	}
1987
1988	/*
1989	 * It is possible that an interrupt came in,
1990	 * set the head up, then more interrupts came in
1991	 * and moved it again. When we get back here,
1992	 * the page would have been set to NORMAL but we
1993	 * just set it back to HEAD.
1994	 *
1995	 * How do you detect this? Well, if that happened
1996	 * the tail page would have moved.
1997	 */
1998	if (ret == RB_PAGE_NORMAL) {
1999		struct buffer_page *buffer_tail_page;
2000
2001		buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2002		/*
2003		 * If the tail had moved passed next, then we need
2004		 * to reset the pointer.
2005		 */
2006		if (buffer_tail_page != tail_page &&
2007		    buffer_tail_page != next_page)
2008			rb_head_page_set_normal(cpu_buffer, new_head,
2009						next_page,
2010						RB_PAGE_HEAD);
2011	}
2012
2013	/*
2014	 * If this was the outer most commit (the one that
2015	 * changed the original pointer from HEAD to UPDATE),
2016	 * then it is up to us to reset it to NORMAL.
2017	 */
2018	if (type == RB_PAGE_HEAD) {
2019		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2020					      tail_page,
2021					      RB_PAGE_UPDATE);
2022		if (RB_WARN_ON(cpu_buffer,
2023			       ret != RB_PAGE_UPDATE))
2024			return -1;
2025	}
2026
2027	return 0;
2028}
2029
2030static inline void
2031rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2032	      unsigned long tail, struct rb_event_info *info)
2033{
2034	struct buffer_page *tail_page = info->tail_page;
2035	struct ring_buffer_event *event;
2036	unsigned long length = info->length;
2037
2038	/*
2039	 * Only the event that crossed the page boundary
2040	 * must fill the old tail_page with padding.
2041	 */
2042	if (tail >= BUF_PAGE_SIZE) {
2043		/*
2044		 * If the page was filled, then we still need
2045		 * to update the real_end. Reset it to zero
2046		 * and the reader will ignore it.
2047		 */
2048		if (tail == BUF_PAGE_SIZE)
2049			tail_page->real_end = 0;
2050
2051		local_sub(length, &tail_page->write);
2052		return;
2053	}
2054
2055	event = __rb_page_index(tail_page, tail);
2056	kmemcheck_annotate_bitfield(event, bitfield);
2057
2058	/* account for padding bytes */
2059	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2060
2061	/*
2062	 * Save the original length to the meta data.
2063	 * This will be used by the reader to add lost event
2064	 * counter.
2065	 */
2066	tail_page->real_end = tail;
2067
2068	/*
2069	 * If this event is bigger than the minimum size, then
2070	 * we need to be careful that we don't subtract the
2071	 * write counter enough to allow another writer to slip
2072	 * in on this page.
2073	 * We put in a discarded commit instead, to make sure
2074	 * that this space is not used again.
2075	 *
2076	 * If we are less than the minimum size, we don't need to
2077	 * worry about it.
2078	 */
2079	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2080		/* No room for any events */
2081
2082		/* Mark the rest of the page with padding */
2083		rb_event_set_padding(event);
2084
2085		/* Set the write back to the previous setting */
2086		local_sub(length, &tail_page->write);
2087		return;
2088	}
2089
2090	/* Put in a discarded event */
2091	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2092	event->type_len = RINGBUF_TYPE_PADDING;
2093	/* time delta must be non zero */
2094	event->time_delta = 1;
2095
2096	/* Set write to end of buffer */
2097	length = (tail + length) - BUF_PAGE_SIZE;
2098	local_sub(length, &tail_page->write);
2099}
2100
2101static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2102
2103/*
2104 * This is the slow path, force gcc not to inline it.
2105 */
2106static noinline struct ring_buffer_event *
2107rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2108	     unsigned long tail, struct rb_event_info *info)
2109{
2110	struct buffer_page *tail_page = info->tail_page;
2111	struct buffer_page *commit_page = cpu_buffer->commit_page;
2112	struct ring_buffer *buffer = cpu_buffer->buffer;
2113	struct buffer_page *next_page;
2114	int ret;
2115
2116	next_page = tail_page;
2117
2118	rb_inc_page(cpu_buffer, &next_page);
2119
2120	/*
2121	 * If for some reason, we had an interrupt storm that made
2122	 * it all the way around the buffer, bail, and warn
2123	 * about it.
2124	 */
2125	if (unlikely(next_page == commit_page)) {
2126		local_inc(&cpu_buffer->commit_overrun);
2127		goto out_reset;
2128	}
2129
2130	/*
2131	 * This is where the fun begins!
2132	 *
2133	 * We are fighting against races between a reader that
2134	 * could be on another CPU trying to swap its reader
2135	 * page with the buffer head.
2136	 *
2137	 * We are also fighting against interrupts coming in and
2138	 * moving the head or tail on us as well.
2139	 *
2140	 * If the next page is the head page then we have filled
2141	 * the buffer, unless the commit page is still on the
2142	 * reader page.
2143	 */
2144	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2145
2146		/*
2147		 * If the commit is not on the reader page, then
2148		 * move the header page.
2149		 */
2150		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2151			/*
2152			 * If we are not in overwrite mode,
2153			 * this is easy, just stop here.
2154			 */
2155			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2156				local_inc(&cpu_buffer->dropped_events);
2157				goto out_reset;
2158			}
2159
2160			ret = rb_handle_head_page(cpu_buffer,
2161						  tail_page,
2162						  next_page);
2163			if (ret < 0)
2164				goto out_reset;
2165			if (ret)
2166				goto out_again;
2167		} else {
2168			/*
2169			 * We need to be careful here too. The
2170			 * commit page could still be on the reader
2171			 * page. We could have a small buffer, and
2172			 * have filled up the buffer with events
2173			 * from interrupts and such, and wrapped.
2174			 *
2175			 * Note, if the tail page is also the on the
2176			 * reader_page, we let it move out.
2177			 */
2178			if (unlikely((cpu_buffer->commit_page !=
2179				      cpu_buffer->tail_page) &&
2180				     (cpu_buffer->commit_page ==
2181				      cpu_buffer->reader_page))) {
2182				local_inc(&cpu_buffer->commit_overrun);
2183				goto out_reset;
2184			}
2185		}
2186	}
2187
2188	rb_tail_page_update(cpu_buffer, tail_page, next_page);
2189
2190 out_again:
2191
2192	rb_reset_tail(cpu_buffer, tail, info);
2193
2194	/* Commit what we have for now. */
2195	rb_end_commit(cpu_buffer);
2196	/* rb_end_commit() decs committing */
2197	local_inc(&cpu_buffer->committing);
2198
2199	/* fail and let the caller try again */
2200	return ERR_PTR(-EAGAIN);
2201
2202 out_reset:
2203	/* reset write */
2204	rb_reset_tail(cpu_buffer, tail, info);
2205
2206	return NULL;
2207}
2208
2209/* Slow path, do not inline */
2210static noinline struct ring_buffer_event *
2211rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2212{
2213	event->type_len = RINGBUF_TYPE_TIME_EXTEND;
 
 
 
2214
2215	/* Not the first event on the page? */
2216	if (rb_event_index(event)) {
2217		event->time_delta = delta & TS_MASK;
2218		event->array[0] = delta >> TS_SHIFT;
2219	} else {
2220		/* nope, just zero it */
2221		event->time_delta = 0;
2222		event->array[0] = 0;
2223	}
2224
2225	return skip_time_extend(event);
2226}
2227
2228static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2229				     struct ring_buffer_event *event);
2230
2231/**
2232 * rb_update_event - update event type and data
2233 * @event: the event to update
2234 * @type: the type of event
2235 * @length: the size of the event field in the ring buffer
2236 *
2237 * Update the type and data fields of the event. The length
2238 * is the actual size that is written to the ring buffer,
2239 * and with this, we can determine what to place into the
2240 * data field.
2241 */
2242static void
2243rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2244		struct ring_buffer_event *event,
2245		struct rb_event_info *info)
2246{
2247	unsigned length = info->length;
2248	u64 delta = info->delta;
2249
2250	/* Only a commit updates the timestamp */
2251	if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2252		delta = 0;
2253
2254	/*
2255	 * If we need to add a timestamp, then we
2256	 * add it to the start of the resevered space.
2257	 */
2258	if (unlikely(info->add_timestamp)) {
2259		event = rb_add_time_stamp(event, delta);
 
 
2260		length -= RB_LEN_TIME_EXTEND;
2261		delta = 0;
2262	}
2263
2264	event->time_delta = delta;
2265	length -= RB_EVNT_HDR_SIZE;
2266	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2267		event->type_len = 0;
2268		event->array[0] = length;
2269	} else
2270		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2271}
2272
2273static unsigned rb_calculate_event_length(unsigned length)
2274{
2275	struct ring_buffer_event event; /* Used only for sizeof array */
2276
2277	/* zero length can cause confusions */
2278	if (!length)
2279		length++;
2280
2281	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2282		length += sizeof(event.array[0]);
2283
2284	length += RB_EVNT_HDR_SIZE;
2285	length = ALIGN(length, RB_ARCH_ALIGNMENT);
2286
2287	/*
2288	 * In case the time delta is larger than the 27 bits for it
2289	 * in the header, we need to add a timestamp. If another
2290	 * event comes in when trying to discard this one to increase
2291	 * the length, then the timestamp will be added in the allocated
2292	 * space of this event. If length is bigger than the size needed
2293	 * for the TIME_EXTEND, then padding has to be used. The events
2294	 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2295	 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2296	 * As length is a multiple of 4, we only need to worry if it
2297	 * is 12 (RB_LEN_TIME_EXTEND + 4).
2298	 */
2299	if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2300		length += RB_ALIGNMENT;
2301
2302	return length;
2303}
2304
2305#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2306static inline bool sched_clock_stable(void)
2307{
2308	return true;
2309}
2310#endif
2311
2312static inline int
2313rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2314		  struct ring_buffer_event *event)
2315{
2316	unsigned long new_index, old_index;
2317	struct buffer_page *bpage;
2318	unsigned long index;
2319	unsigned long addr;
2320
2321	new_index = rb_event_index(event);
2322	old_index = new_index + rb_event_ts_length(event);
2323	addr = (unsigned long)event;
2324	addr &= PAGE_MASK;
2325
2326	bpage = READ_ONCE(cpu_buffer->tail_page);
2327
2328	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2329		unsigned long write_mask =
2330			local_read(&bpage->write) & ~RB_WRITE_MASK;
2331		unsigned long event_length = rb_event_length(event);
2332		/*
2333		 * This is on the tail page. It is possible that
2334		 * a write could come in and move the tail page
2335		 * and write to the next page. That is fine
2336		 * because we just shorten what is on this page.
2337		 */
2338		old_index += write_mask;
2339		new_index += write_mask;
2340		index = local_cmpxchg(&bpage->write, old_index, new_index);
2341		if (index == old_index) {
2342			/* update counters */
2343			local_sub(event_length, &cpu_buffer->entries_bytes);
2344			return 1;
2345		}
2346	}
2347
2348	/* could not discard */
2349	return 0;
2350}
2351
2352static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2353{
2354	local_inc(&cpu_buffer->committing);
2355	local_inc(&cpu_buffer->commits);
2356}
2357
2358static __always_inline void
2359rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2360{
2361	unsigned long max_count;
2362
2363	/*
2364	 * We only race with interrupts and NMIs on this CPU.
2365	 * If we own the commit event, then we can commit
2366	 * all others that interrupted us, since the interruptions
2367	 * are in stack format (they finish before they come
2368	 * back to us). This allows us to do a simple loop to
2369	 * assign the commit to the tail.
2370	 */
2371 again:
2372	max_count = cpu_buffer->nr_pages * 100;
2373
2374	while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2375		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2376			return;
2377		if (RB_WARN_ON(cpu_buffer,
2378			       rb_is_reader_page(cpu_buffer->tail_page)))
2379			return;
2380		local_set(&cpu_buffer->commit_page->page->commit,
2381			  rb_page_write(cpu_buffer->commit_page));
2382		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2383		/* Only update the write stamp if the page has an event */
2384		if (rb_page_write(cpu_buffer->commit_page))
2385			cpu_buffer->write_stamp =
2386				cpu_buffer->commit_page->page->time_stamp;
2387		/* add barrier to keep gcc from optimizing too much */
2388		barrier();
2389	}
2390	while (rb_commit_index(cpu_buffer) !=
2391	       rb_page_write(cpu_buffer->commit_page)) {
2392
2393		local_set(&cpu_buffer->commit_page->page->commit,
2394			  rb_page_write(cpu_buffer->commit_page));
2395		RB_WARN_ON(cpu_buffer,
2396			   local_read(&cpu_buffer->commit_page->page->commit) &
2397			   ~RB_WRITE_MASK);
2398		barrier();
2399	}
2400
2401	/* again, keep gcc from optimizing */
2402	barrier();
2403
2404	/*
2405	 * If an interrupt came in just after the first while loop
2406	 * and pushed the tail page forward, we will be left with
2407	 * a dangling commit that will never go forward.
2408	 */
2409	if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2410		goto again;
2411}
2412
2413static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2414{
2415	unsigned long commits;
2416
2417	if (RB_WARN_ON(cpu_buffer,
2418		       !local_read(&cpu_buffer->committing)))
2419		return;
2420
2421 again:
2422	commits = local_read(&cpu_buffer->commits);
2423	/* synchronize with interrupts */
2424	barrier();
2425	if (local_read(&cpu_buffer->committing) == 1)
2426		rb_set_commit_to_write(cpu_buffer);
2427
2428	local_dec(&cpu_buffer->committing);
2429
2430	/* synchronize with interrupts */
2431	barrier();
2432
2433	/*
2434	 * Need to account for interrupts coming in between the
2435	 * updating of the commit page and the clearing of the
2436	 * committing counter.
2437	 */
2438	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2439	    !local_read(&cpu_buffer->committing)) {
2440		local_inc(&cpu_buffer->committing);
2441		goto again;
2442	}
2443}
2444
2445static inline void rb_event_discard(struct ring_buffer_event *event)
2446{
2447	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2448		event = skip_time_extend(event);
2449
2450	/* array[0] holds the actual length for the discarded event */
2451	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2452	event->type_len = RINGBUF_TYPE_PADDING;
2453	/* time delta must be non zero */
2454	if (!event->time_delta)
2455		event->time_delta = 1;
2456}
2457
2458static __always_inline bool
2459rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2460		   struct ring_buffer_event *event)
2461{
2462	unsigned long addr = (unsigned long)event;
2463	unsigned long index;
2464
2465	index = rb_event_index(event);
2466	addr &= PAGE_MASK;
2467
2468	return cpu_buffer->commit_page->page == (void *)addr &&
2469		rb_commit_index(cpu_buffer) == index;
2470}
2471
2472static __always_inline void
2473rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2474		      struct ring_buffer_event *event)
2475{
2476	u64 delta;
2477
2478	/*
2479	 * The event first in the commit queue updates the
2480	 * time stamp.
2481	 */
2482	if (rb_event_is_commit(cpu_buffer, event)) {
2483		/*
2484		 * A commit event that is first on a page
2485		 * updates the write timestamp with the page stamp
2486		 */
2487		if (!rb_event_index(event))
2488			cpu_buffer->write_stamp =
2489				cpu_buffer->commit_page->page->time_stamp;
2490		else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2491			delta = event->array[0];
2492			delta <<= TS_SHIFT;
2493			delta += event->time_delta;
2494			cpu_buffer->write_stamp += delta;
 
 
 
2495		} else
2496			cpu_buffer->write_stamp += event->time_delta;
2497	}
2498}
2499
2500static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2501		      struct ring_buffer_event *event)
2502{
2503	local_inc(&cpu_buffer->entries);
2504	rb_update_write_stamp(cpu_buffer, event);
2505	rb_end_commit(cpu_buffer);
2506}
2507
2508static __always_inline void
2509rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2510{
2511	bool pagebusy;
 
 
2512
2513	if (buffer->irq_work.waiters_pending) {
2514		buffer->irq_work.waiters_pending = false;
2515		/* irq_work_queue() supplies it's own memory barriers */
2516		irq_work_queue(&buffer->irq_work.work);
2517	}
2518
2519	if (cpu_buffer->irq_work.waiters_pending) {
2520		cpu_buffer->irq_work.waiters_pending = false;
2521		/* irq_work_queue() supplies it's own memory barriers */
2522		irq_work_queue(&cpu_buffer->irq_work.work);
2523	}
2524
2525	pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 
 
 
 
 
 
 
 
 
2526
2527	if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2528		cpu_buffer->irq_work.wakeup_full = true;
2529		cpu_buffer->irq_work.full_waiters_pending = false;
2530		/* irq_work_queue() supplies it's own memory barriers */
2531		irq_work_queue(&cpu_buffer->irq_work.work);
2532	}
 
 
 
 
2533}
2534
2535/*
2536 * The lock and unlock are done within a preempt disable section.
2537 * The current_context per_cpu variable can only be modified
2538 * by the current task between lock and unlock. But it can
2539 * be modified more than once via an interrupt. To pass this
2540 * information from the lock to the unlock without having to
2541 * access the 'in_interrupt()' functions again (which do show
2542 * a bit of overhead in something as critical as function tracing,
2543 * we use a bitmask trick.
2544 *
2545 *  bit 0 =  NMI context
2546 *  bit 1 =  IRQ context
2547 *  bit 2 =  SoftIRQ context
2548 *  bit 3 =  normal context.
2549 *
2550 * This works because this is the order of contexts that can
2551 * preempt other contexts. A SoftIRQ never preempts an IRQ
2552 * context.
2553 *
2554 * When the context is determined, the corresponding bit is
2555 * checked and set (if it was set, then a recursion of that context
2556 * happened).
2557 *
2558 * On unlock, we need to clear this bit. To do so, just subtract
2559 * 1 from the current_context and AND it to itself.
2560 *
2561 * (binary)
2562 *  101 - 1 = 100
2563 *  101 & 100 = 100 (clearing bit zero)
2564 *
2565 *  1010 - 1 = 1001
2566 *  1010 & 1001 = 1000 (clearing bit 1)
2567 *
2568 * The least significant bit can be cleared this way, and it
2569 * just so happens that it is the same bit corresponding to
2570 * the current context.
2571 */
2572
2573static __always_inline int
2574trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2575{
2576	unsigned int val = cpu_buffer->current_context;
 
2577	int bit;
2578
2579	if (in_interrupt()) {
2580		if (in_nmi())
2581			bit = RB_CTX_NMI;
2582		else if (in_irq())
2583			bit = RB_CTX_IRQ;
2584		else
2585			bit = RB_CTX_SOFTIRQ;
2586	} else
2587		bit = RB_CTX_NORMAL;
 
 
 
2588
2589	if (unlikely(val & (1 << bit)))
2590		return 1;
2591
2592	val |= (1 << bit);
2593	cpu_buffer->current_context = val;
2594
2595	return 0;
2596}
2597
2598static __always_inline void
2599trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2600{
2601	cpu_buffer->current_context &= cpu_buffer->current_context - 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2602}
2603
2604/**
2605 * ring_buffer_unlock_commit - commit a reserved
2606 * @buffer: The buffer to commit to
2607 * @event: The event pointer to commit.
2608 *
2609 * This commits the data to the ring buffer, and releases any locks held.
2610 *
2611 * Must be paired with ring_buffer_lock_reserve.
2612 */
2613int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2614			      struct ring_buffer_event *event)
2615{
2616	struct ring_buffer_per_cpu *cpu_buffer;
2617	int cpu = raw_smp_processor_id();
2618
2619	cpu_buffer = buffer->buffers[cpu];
2620
2621	rb_commit(cpu_buffer, event);
2622
2623	rb_wakeups(buffer, cpu_buffer);
2624
2625	trace_recursive_unlock(cpu_buffer);
2626
2627	preempt_enable_notrace();
2628
2629	return 0;
2630}
2631EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2632
2633static noinline void
2634rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2635		    struct rb_event_info *info)
2636{
2637	WARN_ONCE(info->delta > (1ULL << 59),
2638		  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2639		  (unsigned long long)info->delta,
2640		  (unsigned long long)info->ts,
2641		  (unsigned long long)cpu_buffer->write_stamp,
2642		  sched_clock_stable() ? "" :
2643		  "If you just came from a suspend/resume,\n"
2644		  "please switch to the trace global clock:\n"
2645		  "  echo global > /sys/kernel/debug/tracing/trace_clock\n");
 
2646	info->add_timestamp = 1;
2647}
2648
2649static struct ring_buffer_event *
2650__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2651		  struct rb_event_info *info)
2652{
2653	struct ring_buffer_event *event;
2654	struct buffer_page *tail_page;
2655	unsigned long tail, write;
2656
2657	/*
2658	 * If the time delta since the last event is too big to
2659	 * hold in the time field of the event, then we append a
2660	 * TIME EXTEND event ahead of the data event.
2661	 */
2662	if (unlikely(info->add_timestamp))
2663		info->length += RB_LEN_TIME_EXTEND;
2664
2665	/* Don't let the compiler play games with cpu_buffer->tail_page */
2666	tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2667	write = local_add_return(info->length, &tail_page->write);
2668
2669	/* set write to only the index of the write */
2670	write &= RB_WRITE_MASK;
2671	tail = write - info->length;
2672
2673	/*
2674	 * If this is the first commit on the page, then it has the same
2675	 * timestamp as the page itself.
2676	 */
2677	if (!tail)
2678		info->delta = 0;
2679
2680	/* See if we shot pass the end of this buffer page */
2681	if (unlikely(write > BUF_PAGE_SIZE))
2682		return rb_move_tail(cpu_buffer, tail, info);
2683
2684	/* We reserved something on the buffer */
2685
2686	event = __rb_page_index(tail_page, tail);
2687	kmemcheck_annotate_bitfield(event, bitfield);
2688	rb_update_event(cpu_buffer, event, info);
2689
2690	local_inc(&tail_page->entries);
2691
2692	/*
2693	 * If this is the first commit on the page, then update
2694	 * its timestamp.
2695	 */
2696	if (!tail)
2697		tail_page->page->time_stamp = info->ts;
2698
2699	/* account for these added bytes */
2700	local_add(info->length, &cpu_buffer->entries_bytes);
2701
2702	return event;
2703}
2704
2705static __always_inline struct ring_buffer_event *
2706rb_reserve_next_event(struct ring_buffer *buffer,
2707		      struct ring_buffer_per_cpu *cpu_buffer,
2708		      unsigned long length)
2709{
2710	struct ring_buffer_event *event;
2711	struct rb_event_info info;
2712	int nr_loops = 0;
2713	u64 diff;
2714
2715	rb_start_commit(cpu_buffer);
2716
2717#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2718	/*
2719	 * Due to the ability to swap a cpu buffer from a buffer
2720	 * it is possible it was swapped before we committed.
2721	 * (committing stops a swap). We check for it here and
2722	 * if it happened, we have to fail the write.
2723	 */
2724	barrier();
2725	if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2726		local_dec(&cpu_buffer->committing);
2727		local_dec(&cpu_buffer->commits);
2728		return NULL;
2729	}
2730#endif
2731
2732	info.length = rb_calculate_event_length(length);
2733 again:
2734	info.add_timestamp = 0;
2735	info.delta = 0;
2736
2737	/*
2738	 * We allow for interrupts to reenter here and do a trace.
2739	 * If one does, it will cause this original code to loop
2740	 * back here. Even with heavy interrupts happening, this
2741	 * should only happen a few times in a row. If this happens
2742	 * 1000 times in a row, there must be either an interrupt
2743	 * storm or we have something buggy.
2744	 * Bail!
2745	 */
2746	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2747		goto out_fail;
2748
2749	info.ts = rb_time_stamp(cpu_buffer->buffer);
2750	diff = info.ts - cpu_buffer->write_stamp;
2751
2752	/* make sure this diff is calculated here */
2753	barrier();
2754
2755	/* Did the write stamp get updated already? */
2756	if (likely(info.ts >= cpu_buffer->write_stamp)) {
 
 
 
2757		info.delta = diff;
2758		if (unlikely(test_time_stamp(info.delta)))
2759			rb_handle_timestamp(cpu_buffer, &info);
2760	}
2761
2762	event = __rb_reserve_next(cpu_buffer, &info);
2763
2764	if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2765		if (info.add_timestamp)
2766			info.length -= RB_LEN_TIME_EXTEND;
2767		goto again;
2768	}
2769
2770	if (!event)
2771		goto out_fail;
2772
2773	return event;
2774
2775 out_fail:
2776	rb_end_commit(cpu_buffer);
2777	return NULL;
2778}
2779
2780/**
2781 * ring_buffer_lock_reserve - reserve a part of the buffer
2782 * @buffer: the ring buffer to reserve from
2783 * @length: the length of the data to reserve (excluding event header)
2784 *
2785 * Returns a reseverd event on the ring buffer to copy directly to.
2786 * The user of this interface will need to get the body to write into
2787 * and can use the ring_buffer_event_data() interface.
2788 *
2789 * The length is the length of the data needed, not the event length
2790 * which also includes the event header.
2791 *
2792 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2793 * If NULL is returned, then nothing has been allocated or locked.
2794 */
2795struct ring_buffer_event *
2796ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2797{
2798	struct ring_buffer_per_cpu *cpu_buffer;
2799	struct ring_buffer_event *event;
2800	int cpu;
2801
2802	/* If we are tracing schedule, we don't want to recurse */
2803	preempt_disable_notrace();
2804
2805	if (unlikely(atomic_read(&buffer->record_disabled)))
2806		goto out;
2807
2808	cpu = raw_smp_processor_id();
2809
2810	if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2811		goto out;
2812
2813	cpu_buffer = buffer->buffers[cpu];
2814
2815	if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2816		goto out;
2817
2818	if (unlikely(length > BUF_MAX_DATA_SIZE))
2819		goto out;
2820
2821	if (unlikely(trace_recursive_lock(cpu_buffer)))
2822		goto out;
2823
2824	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2825	if (!event)
2826		goto out_unlock;
2827
2828	return event;
2829
2830 out_unlock:
2831	trace_recursive_unlock(cpu_buffer);
2832 out:
2833	preempt_enable_notrace();
2834	return NULL;
2835}
2836EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2837
2838/*
2839 * Decrement the entries to the page that an event is on.
2840 * The event does not even need to exist, only the pointer
2841 * to the page it is on. This may only be called before the commit
2842 * takes place.
2843 */
2844static inline void
2845rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2846		   struct ring_buffer_event *event)
2847{
2848	unsigned long addr = (unsigned long)event;
2849	struct buffer_page *bpage = cpu_buffer->commit_page;
2850	struct buffer_page *start;
2851
2852	addr &= PAGE_MASK;
2853
2854	/* Do the likely case first */
2855	if (likely(bpage->page == (void *)addr)) {
2856		local_dec(&bpage->entries);
2857		return;
2858	}
2859
2860	/*
2861	 * Because the commit page may be on the reader page we
2862	 * start with the next page and check the end loop there.
2863	 */
2864	rb_inc_page(cpu_buffer, &bpage);
2865	start = bpage;
2866	do {
2867		if (bpage->page == (void *)addr) {
2868			local_dec(&bpage->entries);
2869			return;
2870		}
2871		rb_inc_page(cpu_buffer, &bpage);
2872	} while (bpage != start);
2873
2874	/* commit not part of this buffer?? */
2875	RB_WARN_ON(cpu_buffer, 1);
2876}
2877
2878/**
2879 * ring_buffer_commit_discard - discard an event that has not been committed
2880 * @buffer: the ring buffer
2881 * @event: non committed event to discard
2882 *
2883 * Sometimes an event that is in the ring buffer needs to be ignored.
2884 * This function lets the user discard an event in the ring buffer
2885 * and then that event will not be read later.
2886 *
2887 * This function only works if it is called before the the item has been
2888 * committed. It will try to free the event from the ring buffer
2889 * if another event has not been added behind it.
2890 *
2891 * If another event has been added behind it, it will set the event
2892 * up as discarded, and perform the commit.
2893 *
2894 * If this function is called, do not call ring_buffer_unlock_commit on
2895 * the event.
2896 */
2897void ring_buffer_discard_commit(struct ring_buffer *buffer,
2898				struct ring_buffer_event *event)
2899{
2900	struct ring_buffer_per_cpu *cpu_buffer;
2901	int cpu;
2902
2903	/* The event is discarded regardless */
2904	rb_event_discard(event);
2905
2906	cpu = smp_processor_id();
2907	cpu_buffer = buffer->buffers[cpu];
2908
2909	/*
2910	 * This must only be called if the event has not been
2911	 * committed yet. Thus we can assume that preemption
2912	 * is still disabled.
2913	 */
2914	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2915
2916	rb_decrement_entry(cpu_buffer, event);
2917	if (rb_try_to_discard(cpu_buffer, event))
2918		goto out;
2919
2920	/*
2921	 * The commit is still visible by the reader, so we
2922	 * must still update the timestamp.
2923	 */
2924	rb_update_write_stamp(cpu_buffer, event);
2925 out:
2926	rb_end_commit(cpu_buffer);
2927
2928	trace_recursive_unlock(cpu_buffer);
2929
2930	preempt_enable_notrace();
2931
2932}
2933EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2934
2935/**
2936 * ring_buffer_write - write data to the buffer without reserving
2937 * @buffer: The ring buffer to write to.
2938 * @length: The length of the data being written (excluding the event header)
2939 * @data: The data to write to the buffer.
2940 *
2941 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2942 * one function. If you already have the data to write to the buffer, it
2943 * may be easier to simply call this function.
2944 *
2945 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2946 * and not the length of the event which would hold the header.
2947 */
2948int ring_buffer_write(struct ring_buffer *buffer,
2949		      unsigned long length,
2950		      void *data)
2951{
2952	struct ring_buffer_per_cpu *cpu_buffer;
2953	struct ring_buffer_event *event;
2954	void *body;
2955	int ret = -EBUSY;
2956	int cpu;
2957
2958	preempt_disable_notrace();
2959
2960	if (atomic_read(&buffer->record_disabled))
2961		goto out;
2962
2963	cpu = raw_smp_processor_id();
2964
2965	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2966		goto out;
2967
2968	cpu_buffer = buffer->buffers[cpu];
2969
2970	if (atomic_read(&cpu_buffer->record_disabled))
2971		goto out;
2972
2973	if (length > BUF_MAX_DATA_SIZE)
2974		goto out;
2975
2976	if (unlikely(trace_recursive_lock(cpu_buffer)))
2977		goto out;
2978
2979	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2980	if (!event)
2981		goto out_unlock;
2982
2983	body = rb_event_data(event);
2984
2985	memcpy(body, data, length);
2986
2987	rb_commit(cpu_buffer, event);
2988
2989	rb_wakeups(buffer, cpu_buffer);
2990
2991	ret = 0;
2992
2993 out_unlock:
2994	trace_recursive_unlock(cpu_buffer);
2995
2996 out:
2997	preempt_enable_notrace();
2998
2999	return ret;
3000}
3001EXPORT_SYMBOL_GPL(ring_buffer_write);
3002
3003static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3004{
3005	struct buffer_page *reader = cpu_buffer->reader_page;
3006	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3007	struct buffer_page *commit = cpu_buffer->commit_page;
3008
3009	/* In case of error, head will be NULL */
3010	if (unlikely(!head))
3011		return true;
3012
3013	return reader->read == rb_page_commit(reader) &&
3014		(commit == reader ||
3015		 (commit == head &&
3016		  head->read == rb_page_commit(commit)));
3017}
3018
3019/**
3020 * ring_buffer_record_disable - stop all writes into the buffer
3021 * @buffer: The ring buffer to stop writes to.
3022 *
3023 * This prevents all writes to the buffer. Any attempt to write
3024 * to the buffer after this will fail and return NULL.
3025 *
3026 * The caller should call synchronize_sched() after this.
3027 */
3028void ring_buffer_record_disable(struct ring_buffer *buffer)
3029{
3030	atomic_inc(&buffer->record_disabled);
3031}
3032EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3033
3034/**
3035 * ring_buffer_record_enable - enable writes to the buffer
3036 * @buffer: The ring buffer to enable writes
3037 *
3038 * Note, multiple disables will need the same number of enables
3039 * to truly enable the writing (much like preempt_disable).
3040 */
3041void ring_buffer_record_enable(struct ring_buffer *buffer)
3042{
3043	atomic_dec(&buffer->record_disabled);
3044}
3045EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3046
3047/**
3048 * ring_buffer_record_off - stop all writes into the buffer
3049 * @buffer: The ring buffer to stop writes to.
3050 *
3051 * This prevents all writes to the buffer. Any attempt to write
3052 * to the buffer after this will fail and return NULL.
3053 *
3054 * This is different than ring_buffer_record_disable() as
3055 * it works like an on/off switch, where as the disable() version
3056 * must be paired with a enable().
3057 */
3058void ring_buffer_record_off(struct ring_buffer *buffer)
3059{
3060	unsigned int rd;
3061	unsigned int new_rd;
3062
3063	do {
3064		rd = atomic_read(&buffer->record_disabled);
3065		new_rd = rd | RB_BUFFER_OFF;
3066	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3067}
3068EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3069
3070/**
3071 * ring_buffer_record_on - restart writes into the buffer
3072 * @buffer: The ring buffer to start writes to.
3073 *
3074 * This enables all writes to the buffer that was disabled by
3075 * ring_buffer_record_off().
3076 *
3077 * This is different than ring_buffer_record_enable() as
3078 * it works like an on/off switch, where as the enable() version
3079 * must be paired with a disable().
3080 */
3081void ring_buffer_record_on(struct ring_buffer *buffer)
3082{
3083	unsigned int rd;
3084	unsigned int new_rd;
3085
3086	do {
3087		rd = atomic_read(&buffer->record_disabled);
3088		new_rd = rd & ~RB_BUFFER_OFF;
3089	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3090}
3091EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3092
3093/**
3094 * ring_buffer_record_is_on - return true if the ring buffer can write
3095 * @buffer: The ring buffer to see if write is enabled
3096 *
3097 * Returns true if the ring buffer is in a state that it accepts writes.
3098 */
3099int ring_buffer_record_is_on(struct ring_buffer *buffer)
3100{
3101	return !atomic_read(&buffer->record_disabled);
3102}
3103
3104/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3105 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3106 * @buffer: The ring buffer to stop writes to.
3107 * @cpu: The CPU buffer to stop
3108 *
3109 * This prevents all writes to the buffer. Any attempt to write
3110 * to the buffer after this will fail and return NULL.
3111 *
3112 * The caller should call synchronize_sched() after this.
3113 */
3114void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3115{
3116	struct ring_buffer_per_cpu *cpu_buffer;
3117
3118	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3119		return;
3120
3121	cpu_buffer = buffer->buffers[cpu];
3122	atomic_inc(&cpu_buffer->record_disabled);
3123}
3124EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3125
3126/**
3127 * ring_buffer_record_enable_cpu - enable writes to the buffer
3128 * @buffer: The ring buffer to enable writes
3129 * @cpu: The CPU to enable.
3130 *
3131 * Note, multiple disables will need the same number of enables
3132 * to truly enable the writing (much like preempt_disable).
3133 */
3134void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3135{
3136	struct ring_buffer_per_cpu *cpu_buffer;
3137
3138	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3139		return;
3140
3141	cpu_buffer = buffer->buffers[cpu];
3142	atomic_dec(&cpu_buffer->record_disabled);
3143}
3144EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3145
3146/*
3147 * The total entries in the ring buffer is the running counter
3148 * of entries entered into the ring buffer, minus the sum of
3149 * the entries read from the ring buffer and the number of
3150 * entries that were overwritten.
3151 */
3152static inline unsigned long
3153rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3154{
3155	return local_read(&cpu_buffer->entries) -
3156		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3157}
3158
3159/**
3160 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3161 * @buffer: The ring buffer
3162 * @cpu: The per CPU buffer to read from.
3163 */
3164u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3165{
3166	unsigned long flags;
3167	struct ring_buffer_per_cpu *cpu_buffer;
3168	struct buffer_page *bpage;
3169	u64 ret = 0;
3170
3171	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3172		return 0;
3173
3174	cpu_buffer = buffer->buffers[cpu];
3175	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3176	/*
3177	 * if the tail is on reader_page, oldest time stamp is on the reader
3178	 * page
3179	 */
3180	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3181		bpage = cpu_buffer->reader_page;
3182	else
3183		bpage = rb_set_head_page(cpu_buffer);
3184	if (bpage)
3185		ret = bpage->page->time_stamp;
3186	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3187
3188	return ret;
3189}
3190EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3191
3192/**
3193 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3194 * @buffer: The ring buffer
3195 * @cpu: The per CPU buffer to read from.
3196 */
3197unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3198{
3199	struct ring_buffer_per_cpu *cpu_buffer;
3200	unsigned long ret;
3201
3202	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3203		return 0;
3204
3205	cpu_buffer = buffer->buffers[cpu];
3206	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3207
3208	return ret;
3209}
3210EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3211
3212/**
3213 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3214 * @buffer: The ring buffer
3215 * @cpu: The per CPU buffer to get the entries from.
3216 */
3217unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3218{
3219	struct ring_buffer_per_cpu *cpu_buffer;
3220
3221	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3222		return 0;
3223
3224	cpu_buffer = buffer->buffers[cpu];
3225
3226	return rb_num_of_entries(cpu_buffer);
3227}
3228EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3229
3230/**
3231 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3232 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3233 * @buffer: The ring buffer
3234 * @cpu: The per CPU buffer to get the number of overruns from
3235 */
3236unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3237{
3238	struct ring_buffer_per_cpu *cpu_buffer;
3239	unsigned long ret;
3240
3241	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3242		return 0;
3243
3244	cpu_buffer = buffer->buffers[cpu];
3245	ret = local_read(&cpu_buffer->overrun);
3246
3247	return ret;
3248}
3249EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3250
3251/**
3252 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3253 * commits failing due to the buffer wrapping around while there are uncommitted
3254 * events, such as during an interrupt storm.
3255 * @buffer: The ring buffer
3256 * @cpu: The per CPU buffer to get the number of overruns from
3257 */
3258unsigned long
3259ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3260{
3261	struct ring_buffer_per_cpu *cpu_buffer;
3262	unsigned long ret;
3263
3264	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3265		return 0;
3266
3267	cpu_buffer = buffer->buffers[cpu];
3268	ret = local_read(&cpu_buffer->commit_overrun);
3269
3270	return ret;
3271}
3272EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3273
3274/**
3275 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3276 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3277 * @buffer: The ring buffer
3278 * @cpu: The per CPU buffer to get the number of overruns from
3279 */
3280unsigned long
3281ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3282{
3283	struct ring_buffer_per_cpu *cpu_buffer;
3284	unsigned long ret;
3285
3286	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3287		return 0;
3288
3289	cpu_buffer = buffer->buffers[cpu];
3290	ret = local_read(&cpu_buffer->dropped_events);
3291
3292	return ret;
3293}
3294EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3295
3296/**
3297 * ring_buffer_read_events_cpu - get the number of events successfully read
3298 * @buffer: The ring buffer
3299 * @cpu: The per CPU buffer to get the number of events read
3300 */
3301unsigned long
3302ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3303{
3304	struct ring_buffer_per_cpu *cpu_buffer;
3305
3306	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3307		return 0;
3308
3309	cpu_buffer = buffer->buffers[cpu];
3310	return cpu_buffer->read;
3311}
3312EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3313
3314/**
3315 * ring_buffer_entries - get the number of entries in a buffer
3316 * @buffer: The ring buffer
3317 *
3318 * Returns the total number of entries in the ring buffer
3319 * (all CPU entries)
3320 */
3321unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3322{
3323	struct ring_buffer_per_cpu *cpu_buffer;
3324	unsigned long entries = 0;
3325	int cpu;
3326
3327	/* if you care about this being correct, lock the buffer */
3328	for_each_buffer_cpu(buffer, cpu) {
3329		cpu_buffer = buffer->buffers[cpu];
3330		entries += rb_num_of_entries(cpu_buffer);
3331	}
3332
3333	return entries;
3334}
3335EXPORT_SYMBOL_GPL(ring_buffer_entries);
3336
3337/**
3338 * ring_buffer_overruns - get the number of overruns in buffer
3339 * @buffer: The ring buffer
3340 *
3341 * Returns the total number of overruns in the ring buffer
3342 * (all CPU entries)
3343 */
3344unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3345{
3346	struct ring_buffer_per_cpu *cpu_buffer;
3347	unsigned long overruns = 0;
3348	int cpu;
3349
3350	/* if you care about this being correct, lock the buffer */
3351	for_each_buffer_cpu(buffer, cpu) {
3352		cpu_buffer = buffer->buffers[cpu];
3353		overruns += local_read(&cpu_buffer->overrun);
3354	}
3355
3356	return overruns;
3357}
3358EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3359
3360static void rb_iter_reset(struct ring_buffer_iter *iter)
3361{
3362	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3363
3364	/* Iterator usage is expected to have record disabled */
3365	iter->head_page = cpu_buffer->reader_page;
3366	iter->head = cpu_buffer->reader_page->read;
3367
3368	iter->cache_reader_page = iter->head_page;
3369	iter->cache_read = cpu_buffer->read;
3370
3371	if (iter->head)
3372		iter->read_stamp = cpu_buffer->read_stamp;
3373	else
3374		iter->read_stamp = iter->head_page->page->time_stamp;
3375}
3376
3377/**
3378 * ring_buffer_iter_reset - reset an iterator
3379 * @iter: The iterator to reset
3380 *
3381 * Resets the iterator, so that it will start from the beginning
3382 * again.
3383 */
3384void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3385{
3386	struct ring_buffer_per_cpu *cpu_buffer;
3387	unsigned long flags;
3388
3389	if (!iter)
3390		return;
3391
3392	cpu_buffer = iter->cpu_buffer;
3393
3394	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3395	rb_iter_reset(iter);
3396	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3397}
3398EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3399
3400/**
3401 * ring_buffer_iter_empty - check if an iterator has no more to read
3402 * @iter: The iterator to check
3403 */
3404int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3405{
3406	struct ring_buffer_per_cpu *cpu_buffer;
 
 
 
 
3407
3408	cpu_buffer = iter->cpu_buffer;
3409
3410	return iter->head_page == cpu_buffer->commit_page &&
3411		iter->head == rb_commit_index(cpu_buffer);
 
 
 
 
 
 
 
 
3412}
3413EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3414
3415static void
3416rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3417		     struct ring_buffer_event *event)
3418{
3419	u64 delta;
3420
3421	switch (event->type_len) {
3422	case RINGBUF_TYPE_PADDING:
3423		return;
3424
3425	case RINGBUF_TYPE_TIME_EXTEND:
3426		delta = event->array[0];
3427		delta <<= TS_SHIFT;
3428		delta += event->time_delta;
3429		cpu_buffer->read_stamp += delta;
3430		return;
3431
3432	case RINGBUF_TYPE_TIME_STAMP:
3433		/* FIXME: not implemented */
 
3434		return;
3435
3436	case RINGBUF_TYPE_DATA:
3437		cpu_buffer->read_stamp += event->time_delta;
3438		return;
3439
3440	default:
3441		BUG();
3442	}
3443	return;
3444}
3445
3446static void
3447rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3448			  struct ring_buffer_event *event)
3449{
3450	u64 delta;
3451
3452	switch (event->type_len) {
3453	case RINGBUF_TYPE_PADDING:
3454		return;
3455
3456	case RINGBUF_TYPE_TIME_EXTEND:
3457		delta = event->array[0];
3458		delta <<= TS_SHIFT;
3459		delta += event->time_delta;
3460		iter->read_stamp += delta;
3461		return;
3462
3463	case RINGBUF_TYPE_TIME_STAMP:
3464		/* FIXME: not implemented */
 
3465		return;
3466
3467	case RINGBUF_TYPE_DATA:
3468		iter->read_stamp += event->time_delta;
3469		return;
3470
3471	default:
3472		BUG();
3473	}
3474	return;
3475}
3476
3477static struct buffer_page *
3478rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3479{
3480	struct buffer_page *reader = NULL;
3481	unsigned long overwrite;
3482	unsigned long flags;
3483	int nr_loops = 0;
3484	int ret;
3485
3486	local_irq_save(flags);
3487	arch_spin_lock(&cpu_buffer->lock);
3488
3489 again:
3490	/*
3491	 * This should normally only loop twice. But because the
3492	 * start of the reader inserts an empty page, it causes
3493	 * a case where we will loop three times. There should be no
3494	 * reason to loop four times (that I know of).
3495	 */
3496	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3497		reader = NULL;
3498		goto out;
3499	}
3500
3501	reader = cpu_buffer->reader_page;
3502
3503	/* If there's more to read, return this page */
3504	if (cpu_buffer->reader_page->read < rb_page_size(reader))
3505		goto out;
3506
3507	/* Never should we have an index greater than the size */
3508	if (RB_WARN_ON(cpu_buffer,
3509		       cpu_buffer->reader_page->read > rb_page_size(reader)))
3510		goto out;
3511
3512	/* check if we caught up to the tail */
3513	reader = NULL;
3514	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3515		goto out;
3516
3517	/* Don't bother swapping if the ring buffer is empty */
3518	if (rb_num_of_entries(cpu_buffer) == 0)
3519		goto out;
3520
3521	/*
3522	 * Reset the reader page to size zero.
3523	 */
3524	local_set(&cpu_buffer->reader_page->write, 0);
3525	local_set(&cpu_buffer->reader_page->entries, 0);
3526	local_set(&cpu_buffer->reader_page->page->commit, 0);
3527	cpu_buffer->reader_page->real_end = 0;
3528
3529 spin:
3530	/*
3531	 * Splice the empty reader page into the list around the head.
3532	 */
3533	reader = rb_set_head_page(cpu_buffer);
3534	if (!reader)
3535		goto out;
3536	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3537	cpu_buffer->reader_page->list.prev = reader->list.prev;
3538
3539	/*
3540	 * cpu_buffer->pages just needs to point to the buffer, it
3541	 *  has no specific buffer page to point to. Lets move it out
3542	 *  of our way so we don't accidentally swap it.
3543	 */
3544	cpu_buffer->pages = reader->list.prev;
3545
3546	/* The reader page will be pointing to the new head */
3547	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3548
3549	/*
3550	 * We want to make sure we read the overruns after we set up our
3551	 * pointers to the next object. The writer side does a
3552	 * cmpxchg to cross pages which acts as the mb on the writer
3553	 * side. Note, the reader will constantly fail the swap
3554	 * while the writer is updating the pointers, so this
3555	 * guarantees that the overwrite recorded here is the one we
3556	 * want to compare with the last_overrun.
3557	 */
3558	smp_mb();
3559	overwrite = local_read(&(cpu_buffer->overrun));
3560
3561	/*
3562	 * Here's the tricky part.
3563	 *
3564	 * We need to move the pointer past the header page.
3565	 * But we can only do that if a writer is not currently
3566	 * moving it. The page before the header page has the
3567	 * flag bit '1' set if it is pointing to the page we want.
3568	 * but if the writer is in the process of moving it
3569	 * than it will be '2' or already moved '0'.
3570	 */
3571
3572	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3573
3574	/*
3575	 * If we did not convert it, then we must try again.
3576	 */
3577	if (!ret)
3578		goto spin;
3579
3580	/*
3581	 * Yeah! We succeeded in replacing the page.
3582	 *
3583	 * Now make the new head point back to the reader page.
3584	 */
3585	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3586	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3587
 
 
3588	/* Finally update the reader page to the new head */
3589	cpu_buffer->reader_page = reader;
3590	cpu_buffer->reader_page->read = 0;
3591
3592	if (overwrite != cpu_buffer->last_overrun) {
3593		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3594		cpu_buffer->last_overrun = overwrite;
3595	}
3596
3597	goto again;
3598
3599 out:
3600	/* Update the read_stamp on the first event */
3601	if (reader && reader->read == 0)
3602		cpu_buffer->read_stamp = reader->page->time_stamp;
3603
3604	arch_spin_unlock(&cpu_buffer->lock);
3605	local_irq_restore(flags);
3606
3607	return reader;
3608}
3609
3610static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3611{
3612	struct ring_buffer_event *event;
3613	struct buffer_page *reader;
3614	unsigned length;
3615
3616	reader = rb_get_reader_page(cpu_buffer);
3617
3618	/* This function should not be called when buffer is empty */
3619	if (RB_WARN_ON(cpu_buffer, !reader))
3620		return;
3621
3622	event = rb_reader_event(cpu_buffer);
3623
3624	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3625		cpu_buffer->read++;
3626
3627	rb_update_read_stamp(cpu_buffer, event);
3628
3629	length = rb_event_length(event);
3630	cpu_buffer->reader_page->read += length;
3631}
3632
3633static void rb_advance_iter(struct ring_buffer_iter *iter)
3634{
3635	struct ring_buffer_per_cpu *cpu_buffer;
3636	struct ring_buffer_event *event;
3637	unsigned length;
3638
3639	cpu_buffer = iter->cpu_buffer;
3640
3641	/*
3642	 * Check if we are at the end of the buffer.
3643	 */
3644	if (iter->head >= rb_page_size(iter->head_page)) {
3645		/* discarded commits can make the page empty */
3646		if (iter->head_page == cpu_buffer->commit_page)
3647			return;
3648		rb_inc_iter(iter);
3649		return;
3650	}
3651
3652	event = rb_iter_head_event(iter);
3653
3654	length = rb_event_length(event);
3655
3656	/*
3657	 * This should not be called to advance the header if we are
3658	 * at the tail of the buffer.
3659	 */
3660	if (RB_WARN_ON(cpu_buffer,
3661		       (iter->head_page == cpu_buffer->commit_page) &&
3662		       (iter->head + length > rb_commit_index(cpu_buffer))))
3663		return;
3664
3665	rb_update_iter_read_stamp(iter, event);
3666
3667	iter->head += length;
3668
3669	/* check for end of page padding */
3670	if ((iter->head >= rb_page_size(iter->head_page)) &&
3671	    (iter->head_page != cpu_buffer->commit_page))
3672		rb_inc_iter(iter);
3673}
3674
3675static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3676{
3677	return cpu_buffer->lost_events;
3678}
3679
3680static struct ring_buffer_event *
3681rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3682	       unsigned long *lost_events)
3683{
3684	struct ring_buffer_event *event;
3685	struct buffer_page *reader;
3686	int nr_loops = 0;
3687
 
 
3688 again:
3689	/*
3690	 * We repeat when a time extend is encountered.
3691	 * Since the time extend is always attached to a data event,
3692	 * we should never loop more than once.
3693	 * (We never hit the following condition more than twice).
3694	 */
3695	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3696		return NULL;
3697
3698	reader = rb_get_reader_page(cpu_buffer);
3699	if (!reader)
3700		return NULL;
3701
3702	event = rb_reader_event(cpu_buffer);
3703
3704	switch (event->type_len) {
3705	case RINGBUF_TYPE_PADDING:
3706		if (rb_null_event(event))
3707			RB_WARN_ON(cpu_buffer, 1);
3708		/*
3709		 * Because the writer could be discarding every
3710		 * event it creates (which would probably be bad)
3711		 * if we were to go back to "again" then we may never
3712		 * catch up, and will trigger the warn on, or lock
3713		 * the box. Return the padding, and we will release
3714		 * the current locks, and try again.
3715		 */
3716		return event;
3717
3718	case RINGBUF_TYPE_TIME_EXTEND:
3719		/* Internal data, OK to advance */
3720		rb_advance_reader(cpu_buffer);
3721		goto again;
3722
3723	case RINGBUF_TYPE_TIME_STAMP:
3724		/* FIXME: not implemented */
 
 
 
 
 
3725		rb_advance_reader(cpu_buffer);
3726		goto again;
3727
3728	case RINGBUF_TYPE_DATA:
3729		if (ts) {
3730			*ts = cpu_buffer->read_stamp + event->time_delta;
3731			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3732							 cpu_buffer->cpu, ts);
3733		}
3734		if (lost_events)
3735			*lost_events = rb_lost_events(cpu_buffer);
3736		return event;
3737
3738	default:
3739		BUG();
3740	}
3741
3742	return NULL;
3743}
3744EXPORT_SYMBOL_GPL(ring_buffer_peek);
3745
3746static struct ring_buffer_event *
3747rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3748{
3749	struct ring_buffer *buffer;
3750	struct ring_buffer_per_cpu *cpu_buffer;
3751	struct ring_buffer_event *event;
3752	int nr_loops = 0;
3753
 
 
 
3754	cpu_buffer = iter->cpu_buffer;
3755	buffer = cpu_buffer->buffer;
3756
3757	/*
3758	 * Check if someone performed a consuming read to
3759	 * the buffer. A consuming read invalidates the iterator
3760	 * and we need to reset the iterator in this case.
3761	 */
3762	if (unlikely(iter->cache_read != cpu_buffer->read ||
3763		     iter->cache_reader_page != cpu_buffer->reader_page))
3764		rb_iter_reset(iter);
3765
3766 again:
3767	if (ring_buffer_iter_empty(iter))
3768		return NULL;
3769
3770	/*
3771	 * We repeat when a time extend is encountered or we hit
3772	 * the end of the page. Since the time extend is always attached
3773	 * to a data event, we should never loop more than three times.
3774	 * Once for going to next page, once on time extend, and
3775	 * finally once to get the event.
3776	 * (We never hit the following condition more than thrice).
3777	 */
3778	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3779		return NULL;
3780
3781	if (rb_per_cpu_empty(cpu_buffer))
3782		return NULL;
3783
3784	if (iter->head >= rb_page_size(iter->head_page)) {
3785		rb_inc_iter(iter);
3786		goto again;
3787	}
3788
3789	event = rb_iter_head_event(iter);
3790
3791	switch (event->type_len) {
3792	case RINGBUF_TYPE_PADDING:
3793		if (rb_null_event(event)) {
3794			rb_inc_iter(iter);
3795			goto again;
3796		}
3797		rb_advance_iter(iter);
3798		return event;
3799
3800	case RINGBUF_TYPE_TIME_EXTEND:
3801		/* Internal data, OK to advance */
3802		rb_advance_iter(iter);
3803		goto again;
3804
3805	case RINGBUF_TYPE_TIME_STAMP:
3806		/* FIXME: not implemented */
 
 
 
 
 
3807		rb_advance_iter(iter);
3808		goto again;
3809
3810	case RINGBUF_TYPE_DATA:
3811		if (ts) {
3812			*ts = iter->read_stamp + event->time_delta;
3813			ring_buffer_normalize_time_stamp(buffer,
3814							 cpu_buffer->cpu, ts);
3815		}
3816		return event;
3817
3818	default:
3819		BUG();
3820	}
3821
3822	return NULL;
3823}
3824EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3825
3826static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3827{
3828	if (likely(!in_nmi())) {
3829		raw_spin_lock(&cpu_buffer->reader_lock);
3830		return true;
3831	}
3832
3833	/*
3834	 * If an NMI die dumps out the content of the ring buffer
3835	 * trylock must be used to prevent a deadlock if the NMI
3836	 * preempted a task that holds the ring buffer locks. If
3837	 * we get the lock then all is fine, if not, then continue
3838	 * to do the read, but this can corrupt the ring buffer,
3839	 * so it must be permanently disabled from future writes.
3840	 * Reading from NMI is a oneshot deal.
3841	 */
3842	if (raw_spin_trylock(&cpu_buffer->reader_lock))
3843		return true;
3844
3845	/* Continue without locking, but disable the ring buffer */
3846	atomic_inc(&cpu_buffer->record_disabled);
3847	return false;
3848}
3849
3850static inline void
3851rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3852{
3853	if (likely(locked))
3854		raw_spin_unlock(&cpu_buffer->reader_lock);
3855	return;
3856}
3857
3858/**
3859 * ring_buffer_peek - peek at the next event to be read
3860 * @buffer: The ring buffer to read
3861 * @cpu: The cpu to peak at
3862 * @ts: The timestamp counter of this event.
3863 * @lost_events: a variable to store if events were lost (may be NULL)
3864 *
3865 * This will return the event that will be read next, but does
3866 * not consume the data.
3867 */
3868struct ring_buffer_event *
3869ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3870		 unsigned long *lost_events)
3871{
3872	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3873	struct ring_buffer_event *event;
3874	unsigned long flags;
3875	bool dolock;
3876
3877	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3878		return NULL;
3879
3880 again:
3881	local_irq_save(flags);
3882	dolock = rb_reader_lock(cpu_buffer);
3883	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3884	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3885		rb_advance_reader(cpu_buffer);
3886	rb_reader_unlock(cpu_buffer, dolock);
3887	local_irq_restore(flags);
3888
3889	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3890		goto again;
3891
3892	return event;
3893}
3894
3895/**
3896 * ring_buffer_iter_peek - peek at the next event to be read
3897 * @iter: The ring buffer iterator
3898 * @ts: The timestamp counter of this event.
3899 *
3900 * This will return the event that will be read next, but does
3901 * not increment the iterator.
3902 */
3903struct ring_buffer_event *
3904ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3905{
3906	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3907	struct ring_buffer_event *event;
3908	unsigned long flags;
3909
3910 again:
3911	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3912	event = rb_iter_peek(iter, ts);
3913	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3914
3915	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3916		goto again;
3917
3918	return event;
3919}
3920
3921/**
3922 * ring_buffer_consume - return an event and consume it
3923 * @buffer: The ring buffer to get the next event from
3924 * @cpu: the cpu to read the buffer from
3925 * @ts: a variable to store the timestamp (may be NULL)
3926 * @lost_events: a variable to store if events were lost (may be NULL)
3927 *
3928 * Returns the next event in the ring buffer, and that event is consumed.
3929 * Meaning, that sequential reads will keep returning a different event,
3930 * and eventually empty the ring buffer if the producer is slower.
3931 */
3932struct ring_buffer_event *
3933ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3934		    unsigned long *lost_events)
3935{
3936	struct ring_buffer_per_cpu *cpu_buffer;
3937	struct ring_buffer_event *event = NULL;
3938	unsigned long flags;
3939	bool dolock;
3940
3941 again:
3942	/* might be called in atomic */
3943	preempt_disable();
3944
3945	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3946		goto out;
3947
3948	cpu_buffer = buffer->buffers[cpu];
3949	local_irq_save(flags);
3950	dolock = rb_reader_lock(cpu_buffer);
3951
3952	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3953	if (event) {
3954		cpu_buffer->lost_events = 0;
3955		rb_advance_reader(cpu_buffer);
3956	}
3957
3958	rb_reader_unlock(cpu_buffer, dolock);
3959	local_irq_restore(flags);
3960
3961 out:
3962	preempt_enable();
3963
3964	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3965		goto again;
3966
3967	return event;
3968}
3969EXPORT_SYMBOL_GPL(ring_buffer_consume);
3970
3971/**
3972 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3973 * @buffer: The ring buffer to read from
3974 * @cpu: The cpu buffer to iterate over
 
3975 *
3976 * This performs the initial preparations necessary to iterate
3977 * through the buffer.  Memory is allocated, buffer recording
3978 * is disabled, and the iterator pointer is returned to the caller.
3979 *
3980 * Disabling buffer recordng prevents the reading from being
3981 * corrupted. This is not a consuming read, so a producer is not
3982 * expected.
3983 *
3984 * After a sequence of ring_buffer_read_prepare calls, the user is
3985 * expected to make at least one call to ring_buffer_read_prepare_sync.
3986 * Afterwards, ring_buffer_read_start is invoked to get things going
3987 * for real.
3988 *
3989 * This overall must be paired with ring_buffer_read_finish.
3990 */
3991struct ring_buffer_iter *
3992ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3993{
3994	struct ring_buffer_per_cpu *cpu_buffer;
3995	struct ring_buffer_iter *iter;
3996
3997	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3998		return NULL;
3999
4000	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4001	if (!iter)
4002		return NULL;
4003
4004	cpu_buffer = buffer->buffers[cpu];
4005
4006	iter->cpu_buffer = cpu_buffer;
4007
4008	atomic_inc(&buffer->resize_disabled);
4009	atomic_inc(&cpu_buffer->record_disabled);
4010
4011	return iter;
4012}
4013EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4014
4015/**
4016 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4017 *
4018 * All previously invoked ring_buffer_read_prepare calls to prepare
4019 * iterators will be synchronized.  Afterwards, read_buffer_read_start
4020 * calls on those iterators are allowed.
4021 */
4022void
4023ring_buffer_read_prepare_sync(void)
4024{
4025	synchronize_sched();
4026}
4027EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4028
4029/**
4030 * ring_buffer_read_start - start a non consuming read of the buffer
4031 * @iter: The iterator returned by ring_buffer_read_prepare
4032 *
4033 * This finalizes the startup of an iteration through the buffer.
4034 * The iterator comes from a call to ring_buffer_read_prepare and
4035 * an intervening ring_buffer_read_prepare_sync must have been
4036 * performed.
4037 *
4038 * Must be paired with ring_buffer_read_finish.
4039 */
4040void
4041ring_buffer_read_start(struct ring_buffer_iter *iter)
4042{
4043	struct ring_buffer_per_cpu *cpu_buffer;
4044	unsigned long flags;
4045
4046	if (!iter)
4047		return;
4048
4049	cpu_buffer = iter->cpu_buffer;
4050
4051	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4052	arch_spin_lock(&cpu_buffer->lock);
4053	rb_iter_reset(iter);
4054	arch_spin_unlock(&cpu_buffer->lock);
4055	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4056}
4057EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4058
4059/**
4060 * ring_buffer_read_finish - finish reading the iterator of the buffer
4061 * @iter: The iterator retrieved by ring_buffer_start
4062 *
4063 * This re-enables the recording to the buffer, and frees the
4064 * iterator.
4065 */
4066void
4067ring_buffer_read_finish(struct ring_buffer_iter *iter)
4068{
4069	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4070	unsigned long flags;
4071
4072	/*
4073	 * Ring buffer is disabled from recording, here's a good place
4074	 * to check the integrity of the ring buffer.
4075	 * Must prevent readers from trying to read, as the check
4076	 * clears the HEAD page and readers require it.
4077	 */
4078	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4079	rb_check_pages(cpu_buffer);
4080	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4081
4082	atomic_dec(&cpu_buffer->record_disabled);
4083	atomic_dec(&cpu_buffer->buffer->resize_disabled);
4084	kfree(iter);
4085}
4086EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4087
4088/**
4089 * ring_buffer_read - read the next item in the ring buffer by the iterator
4090 * @iter: The ring buffer iterator
4091 * @ts: The time stamp of the event read.
4092 *
4093 * This reads the next event in the ring buffer and increments the iterator.
4094 */
4095struct ring_buffer_event *
4096ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4097{
4098	struct ring_buffer_event *event;
4099	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4100	unsigned long flags;
4101
4102	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4103 again:
4104	event = rb_iter_peek(iter, ts);
4105	if (!event)
4106		goto out;
4107
4108	if (event->type_len == RINGBUF_TYPE_PADDING)
4109		goto again;
4110
4111	rb_advance_iter(iter);
4112 out:
4113	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4114
4115	return event;
4116}
4117EXPORT_SYMBOL_GPL(ring_buffer_read);
4118
4119/**
4120 * ring_buffer_size - return the size of the ring buffer (in bytes)
4121 * @buffer: The ring buffer.
4122 */
4123unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4124{
4125	/*
4126	 * Earlier, this method returned
4127	 *	BUF_PAGE_SIZE * buffer->nr_pages
4128	 * Since the nr_pages field is now removed, we have converted this to
4129	 * return the per cpu buffer value.
4130	 */
4131	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4132		return 0;
4133
4134	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4135}
4136EXPORT_SYMBOL_GPL(ring_buffer_size);
4137
4138static void
4139rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4140{
4141	rb_head_page_deactivate(cpu_buffer);
4142
4143	cpu_buffer->head_page
4144		= list_entry(cpu_buffer->pages, struct buffer_page, list);
4145	local_set(&cpu_buffer->head_page->write, 0);
4146	local_set(&cpu_buffer->head_page->entries, 0);
4147	local_set(&cpu_buffer->head_page->page->commit, 0);
4148
4149	cpu_buffer->head_page->read = 0;
4150
4151	cpu_buffer->tail_page = cpu_buffer->head_page;
4152	cpu_buffer->commit_page = cpu_buffer->head_page;
4153
4154	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4155	INIT_LIST_HEAD(&cpu_buffer->new_pages);
4156	local_set(&cpu_buffer->reader_page->write, 0);
4157	local_set(&cpu_buffer->reader_page->entries, 0);
4158	local_set(&cpu_buffer->reader_page->page->commit, 0);
4159	cpu_buffer->reader_page->read = 0;
4160
4161	local_set(&cpu_buffer->entries_bytes, 0);
4162	local_set(&cpu_buffer->overrun, 0);
4163	local_set(&cpu_buffer->commit_overrun, 0);
4164	local_set(&cpu_buffer->dropped_events, 0);
4165	local_set(&cpu_buffer->entries, 0);
4166	local_set(&cpu_buffer->committing, 0);
4167	local_set(&cpu_buffer->commits, 0);
 
 
 
 
4168	cpu_buffer->read = 0;
4169	cpu_buffer->read_bytes = 0;
4170
4171	cpu_buffer->write_stamp = 0;
4172	cpu_buffer->read_stamp = 0;
4173
4174	cpu_buffer->lost_events = 0;
4175	cpu_buffer->last_overrun = 0;
4176
4177	rb_head_page_activate(cpu_buffer);
4178}
4179
4180/**
4181 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4182 * @buffer: The ring buffer to reset a per cpu buffer of
4183 * @cpu: The CPU buffer to be reset
4184 */
4185void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4186{
4187	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4188	unsigned long flags;
4189
4190	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4191		return;
4192
4193	atomic_inc(&buffer->resize_disabled);
4194	atomic_inc(&cpu_buffer->record_disabled);
4195
4196	/* Make sure all commits have finished */
4197	synchronize_sched();
4198
4199	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4200
4201	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4202		goto out;
4203
4204	arch_spin_lock(&cpu_buffer->lock);
4205
4206	rb_reset_cpu(cpu_buffer);
4207
4208	arch_spin_unlock(&cpu_buffer->lock);
4209
4210 out:
4211	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4212
4213	atomic_dec(&cpu_buffer->record_disabled);
4214	atomic_dec(&buffer->resize_disabled);
4215}
4216EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4217
4218/**
4219 * ring_buffer_reset - reset a ring buffer
4220 * @buffer: The ring buffer to reset all cpu buffers
4221 */
4222void ring_buffer_reset(struct ring_buffer *buffer)
4223{
4224	int cpu;
4225
4226	for_each_buffer_cpu(buffer, cpu)
4227		ring_buffer_reset_cpu(buffer, cpu);
4228}
4229EXPORT_SYMBOL_GPL(ring_buffer_reset);
4230
4231/**
4232 * rind_buffer_empty - is the ring buffer empty?
4233 * @buffer: The ring buffer to test
4234 */
4235bool ring_buffer_empty(struct ring_buffer *buffer)
4236{
4237	struct ring_buffer_per_cpu *cpu_buffer;
4238	unsigned long flags;
4239	bool dolock;
4240	int cpu;
4241	int ret;
4242
4243	/* yes this is racy, but if you don't like the race, lock the buffer */
4244	for_each_buffer_cpu(buffer, cpu) {
4245		cpu_buffer = buffer->buffers[cpu];
4246		local_irq_save(flags);
4247		dolock = rb_reader_lock(cpu_buffer);
4248		ret = rb_per_cpu_empty(cpu_buffer);
4249		rb_reader_unlock(cpu_buffer, dolock);
4250		local_irq_restore(flags);
4251
4252		if (!ret)
4253			return false;
4254	}
4255
4256	return true;
4257}
4258EXPORT_SYMBOL_GPL(ring_buffer_empty);
4259
4260/**
4261 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4262 * @buffer: The ring buffer
4263 * @cpu: The CPU buffer to test
4264 */
4265bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4266{
4267	struct ring_buffer_per_cpu *cpu_buffer;
4268	unsigned long flags;
4269	bool dolock;
4270	int ret;
4271
4272	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4273		return true;
4274
4275	cpu_buffer = buffer->buffers[cpu];
4276	local_irq_save(flags);
4277	dolock = rb_reader_lock(cpu_buffer);
4278	ret = rb_per_cpu_empty(cpu_buffer);
4279	rb_reader_unlock(cpu_buffer, dolock);
4280	local_irq_restore(flags);
4281
4282	return ret;
4283}
4284EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4285
4286#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4287/**
4288 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4289 * @buffer_a: One buffer to swap with
4290 * @buffer_b: The other buffer to swap with
4291 *
4292 * This function is useful for tracers that want to take a "snapshot"
4293 * of a CPU buffer and has another back up buffer lying around.
4294 * it is expected that the tracer handles the cpu buffer not being
4295 * used at the moment.
4296 */
4297int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4298			 struct ring_buffer *buffer_b, int cpu)
4299{
4300	struct ring_buffer_per_cpu *cpu_buffer_a;
4301	struct ring_buffer_per_cpu *cpu_buffer_b;
4302	int ret = -EINVAL;
4303
4304	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4305	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
4306		goto out;
4307
4308	cpu_buffer_a = buffer_a->buffers[cpu];
4309	cpu_buffer_b = buffer_b->buffers[cpu];
4310
4311	/* At least make sure the two buffers are somewhat the same */
4312	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4313		goto out;
4314
4315	ret = -EAGAIN;
4316
4317	if (atomic_read(&buffer_a->record_disabled))
4318		goto out;
4319
4320	if (atomic_read(&buffer_b->record_disabled))
4321		goto out;
4322
4323	if (atomic_read(&cpu_buffer_a->record_disabled))
4324		goto out;
4325
4326	if (atomic_read(&cpu_buffer_b->record_disabled))
4327		goto out;
4328
4329	/*
4330	 * We can't do a synchronize_sched here because this
4331	 * function can be called in atomic context.
4332	 * Normally this will be called from the same CPU as cpu.
4333	 * If not it's up to the caller to protect this.
4334	 */
4335	atomic_inc(&cpu_buffer_a->record_disabled);
4336	atomic_inc(&cpu_buffer_b->record_disabled);
4337
4338	ret = -EBUSY;
4339	if (local_read(&cpu_buffer_a->committing))
4340		goto out_dec;
4341	if (local_read(&cpu_buffer_b->committing))
4342		goto out_dec;
4343
4344	buffer_a->buffers[cpu] = cpu_buffer_b;
4345	buffer_b->buffers[cpu] = cpu_buffer_a;
4346
4347	cpu_buffer_b->buffer = buffer_a;
4348	cpu_buffer_a->buffer = buffer_b;
4349
4350	ret = 0;
4351
4352out_dec:
4353	atomic_dec(&cpu_buffer_a->record_disabled);
4354	atomic_dec(&cpu_buffer_b->record_disabled);
4355out:
4356	return ret;
4357}
4358EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4359#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4360
4361/**
4362 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4363 * @buffer: the buffer to allocate for.
4364 * @cpu: the cpu buffer to allocate.
4365 *
4366 * This function is used in conjunction with ring_buffer_read_page.
4367 * When reading a full page from the ring buffer, these functions
4368 * can be used to speed up the process. The calling function should
4369 * allocate a few pages first with this function. Then when it
4370 * needs to get pages from the ring buffer, it passes the result
4371 * of this function into ring_buffer_read_page, which will swap
4372 * the page that was allocated, with the read page of the buffer.
4373 *
4374 * Returns:
4375 *  The page allocated, or NULL on error.
4376 */
4377void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4378{
4379	struct buffer_data_page *bpage;
 
 
4380	struct page *page;
4381
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4382	page = alloc_pages_node(cpu_to_node(cpu),
4383				GFP_KERNEL | __GFP_NORETRY, 0);
4384	if (!page)
4385		return NULL;
4386
4387	bpage = page_address(page);
4388
 
4389	rb_init_page(bpage);
4390
4391	return bpage;
4392}
4393EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4394
4395/**
4396 * ring_buffer_free_read_page - free an allocated read page
4397 * @buffer: the buffer the page was allocate for
 
4398 * @data: the page to free
4399 *
4400 * Free a page allocated from ring_buffer_alloc_read_page.
4401 */
4402void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4403{
4404	free_page((unsigned long)data);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4405}
4406EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4407
4408/**
4409 * ring_buffer_read_page - extract a page from the ring buffer
4410 * @buffer: buffer to extract from
4411 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4412 * @len: amount to extract
4413 * @cpu: the cpu of the buffer to extract
4414 * @full: should the extraction only happen when the page is full.
4415 *
4416 * This function will pull out a page from the ring buffer and consume it.
4417 * @data_page must be the address of the variable that was returned
4418 * from ring_buffer_alloc_read_page. This is because the page might be used
4419 * to swap with a page in the ring buffer.
4420 *
4421 * for example:
4422 *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
4423 *	if (!rpage)
4424 *		return error;
4425 *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4426 *	if (ret >= 0)
4427 *		process_page(rpage, ret);
4428 *
4429 * When @full is set, the function will not return true unless
4430 * the writer is off the reader page.
4431 *
4432 * Note: it is up to the calling functions to handle sleeps and wakeups.
4433 *  The ring buffer can be used anywhere in the kernel and can not
4434 *  blindly call wake_up. The layer that uses the ring buffer must be
4435 *  responsible for that.
4436 *
4437 * Returns:
4438 *  >=0 if data has been transferred, returns the offset of consumed data.
4439 *  <0 if no data has been transferred.
4440 */
4441int ring_buffer_read_page(struct ring_buffer *buffer,
4442			  void **data_page, size_t len, int cpu, int full)
4443{
4444	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4445	struct ring_buffer_event *event;
4446	struct buffer_data_page *bpage;
4447	struct buffer_page *reader;
4448	unsigned long missed_events;
4449	unsigned long flags;
4450	unsigned int commit;
4451	unsigned int read;
4452	u64 save_timestamp;
4453	int ret = -1;
4454
4455	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4456		goto out;
4457
4458	/*
4459	 * If len is not big enough to hold the page header, then
4460	 * we can not copy anything.
4461	 */
4462	if (len <= BUF_PAGE_HDR_SIZE)
4463		goto out;
4464
4465	len -= BUF_PAGE_HDR_SIZE;
4466
4467	if (!data_page)
4468		goto out;
4469
4470	bpage = *data_page;
4471	if (!bpage)
4472		goto out;
4473
4474	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4475
4476	reader = rb_get_reader_page(cpu_buffer);
4477	if (!reader)
4478		goto out_unlock;
4479
4480	event = rb_reader_event(cpu_buffer);
4481
4482	read = reader->read;
4483	commit = rb_page_commit(reader);
4484
4485	/* Check if any events were dropped */
4486	missed_events = cpu_buffer->lost_events;
4487
4488	/*
4489	 * If this page has been partially read or
4490	 * if len is not big enough to read the rest of the page or
4491	 * a writer is still on the page, then
4492	 * we must copy the data from the page to the buffer.
4493	 * Otherwise, we can simply swap the page with the one passed in.
4494	 */
4495	if (read || (len < (commit - read)) ||
4496	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
4497		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4498		unsigned int rpos = read;
4499		unsigned int pos = 0;
4500		unsigned int size;
4501
4502		if (full)
4503			goto out_unlock;
4504
4505		if (len > (commit - read))
4506			len = (commit - read);
4507
4508		/* Always keep the time extend and data together */
4509		size = rb_event_ts_length(event);
4510
4511		if (len < size)
4512			goto out_unlock;
4513
4514		/* save the current timestamp, since the user will need it */
4515		save_timestamp = cpu_buffer->read_stamp;
4516
4517		/* Need to copy one event at a time */
4518		do {
4519			/* We need the size of one event, because
4520			 * rb_advance_reader only advances by one event,
4521			 * whereas rb_event_ts_length may include the size of
4522			 * one or two events.
4523			 * We have already ensured there's enough space if this
4524			 * is a time extend. */
4525			size = rb_event_length(event);
4526			memcpy(bpage->data + pos, rpage->data + rpos, size);
4527
4528			len -= size;
4529
4530			rb_advance_reader(cpu_buffer);
4531			rpos = reader->read;
4532			pos += size;
4533
4534			if (rpos >= commit)
4535				break;
4536
4537			event = rb_reader_event(cpu_buffer);
4538			/* Always keep the time extend and data together */
4539			size = rb_event_ts_length(event);
4540		} while (len >= size);
4541
4542		/* update bpage */
4543		local_set(&bpage->commit, pos);
4544		bpage->time_stamp = save_timestamp;
4545
4546		/* we copied everything to the beginning */
4547		read = 0;
4548	} else {
4549		/* update the entry counter */
4550		cpu_buffer->read += rb_page_entries(reader);
4551		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4552
4553		/* swap the pages */
4554		rb_init_page(bpage);
4555		bpage = reader->page;
4556		reader->page = *data_page;
4557		local_set(&reader->write, 0);
4558		local_set(&reader->entries, 0);
4559		reader->read = 0;
4560		*data_page = bpage;
4561
4562		/*
4563		 * Use the real_end for the data size,
4564		 * This gives us a chance to store the lost events
4565		 * on the page.
4566		 */
4567		if (reader->real_end)
4568			local_set(&bpage->commit, reader->real_end);
4569	}
4570	ret = read;
4571
4572	cpu_buffer->lost_events = 0;
4573
4574	commit = local_read(&bpage->commit);
4575	/*
4576	 * Set a flag in the commit field if we lost events
4577	 */
4578	if (missed_events) {
4579		/* If there is room at the end of the page to save the
4580		 * missed events, then record it there.
4581		 */
4582		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4583			memcpy(&bpage->data[commit], &missed_events,
4584			       sizeof(missed_events));
4585			local_add(RB_MISSED_STORED, &bpage->commit);
4586			commit += sizeof(missed_events);
4587		}
4588		local_add(RB_MISSED_EVENTS, &bpage->commit);
4589	}
4590
4591	/*
4592	 * This page may be off to user land. Zero it out here.
4593	 */
4594	if (commit < BUF_PAGE_SIZE)
4595		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4596
4597 out_unlock:
4598	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4599
4600 out:
4601	return ret;
4602}
4603EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4604
4605/*
4606 * We only allocate new buffers, never free them if the CPU goes down.
4607 * If we were to free the buffer, then the user would lose any trace that was in
4608 * the buffer.
4609 */
4610int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4611{
4612	struct ring_buffer *buffer;
4613	long nr_pages_same;
4614	int cpu_i;
4615	unsigned long nr_pages;
4616
4617	buffer = container_of(node, struct ring_buffer, node);
4618	if (cpumask_test_cpu(cpu, buffer->cpumask))
4619		return 0;
4620
4621	nr_pages = 0;
4622	nr_pages_same = 1;
4623	/* check if all cpu sizes are same */
4624	for_each_buffer_cpu(buffer, cpu_i) {
4625		/* fill in the size from first enabled cpu */
4626		if (nr_pages == 0)
4627			nr_pages = buffer->buffers[cpu_i]->nr_pages;
4628		if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4629			nr_pages_same = 0;
4630			break;
4631		}
4632	}
4633	/* allocate minimum pages, user can later expand it */
4634	if (!nr_pages_same)
4635		nr_pages = 2;
4636	buffer->buffers[cpu] =
4637		rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4638	if (!buffer->buffers[cpu]) {
4639		WARN(1, "failed to allocate ring buffer on CPU %u\n",
4640		     cpu);
4641		return -ENOMEM;
4642	}
4643	smp_wmb();
4644	cpumask_set_cpu(cpu, buffer->cpumask);
4645	return 0;
4646}
4647
4648#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4649/*
4650 * This is a basic integrity check of the ring buffer.
4651 * Late in the boot cycle this test will run when configured in.
4652 * It will kick off a thread per CPU that will go into a loop
4653 * writing to the per cpu ring buffer various sizes of data.
4654 * Some of the data will be large items, some small.
4655 *
4656 * Another thread is created that goes into a spin, sending out
4657 * IPIs to the other CPUs to also write into the ring buffer.
4658 * this is to test the nesting ability of the buffer.
4659 *
4660 * Basic stats are recorded and reported. If something in the
4661 * ring buffer should happen that's not expected, a big warning
4662 * is displayed and all ring buffers are disabled.
4663 */
4664static struct task_struct *rb_threads[NR_CPUS] __initdata;
4665
4666struct rb_test_data {
4667	struct ring_buffer	*buffer;
4668	unsigned long		events;
4669	unsigned long		bytes_written;
4670	unsigned long		bytes_alloc;
4671	unsigned long		bytes_dropped;
4672	unsigned long		events_nested;
4673	unsigned long		bytes_written_nested;
4674	unsigned long		bytes_alloc_nested;
4675	unsigned long		bytes_dropped_nested;
4676	int			min_size_nested;
4677	int			max_size_nested;
4678	int			max_size;
4679	int			min_size;
4680	int			cpu;
4681	int			cnt;
4682};
4683
4684static struct rb_test_data rb_data[NR_CPUS] __initdata;
4685
4686/* 1 meg per cpu */
4687#define RB_TEST_BUFFER_SIZE	1048576
4688
4689static char rb_string[] __initdata =
4690	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4691	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4692	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4693
4694static bool rb_test_started __initdata;
4695
4696struct rb_item {
4697	int size;
4698	char str[];
4699};
4700
4701static __init int rb_write_something(struct rb_test_data *data, bool nested)
4702{
4703	struct ring_buffer_event *event;
4704	struct rb_item *item;
4705	bool started;
4706	int event_len;
4707	int size;
4708	int len;
4709	int cnt;
4710
4711	/* Have nested writes different that what is written */
4712	cnt = data->cnt + (nested ? 27 : 0);
4713
4714	/* Multiply cnt by ~e, to make some unique increment */
4715	size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4716
4717	len = size + sizeof(struct rb_item);
4718
4719	started = rb_test_started;
4720	/* read rb_test_started before checking buffer enabled */
4721	smp_rmb();
4722
4723	event = ring_buffer_lock_reserve(data->buffer, len);
4724	if (!event) {
4725		/* Ignore dropped events before test starts. */
4726		if (started) {
4727			if (nested)
4728				data->bytes_dropped += len;
4729			else
4730				data->bytes_dropped_nested += len;
4731		}
4732		return len;
4733	}
4734
4735	event_len = ring_buffer_event_length(event);
4736
4737	if (RB_WARN_ON(data->buffer, event_len < len))
4738		goto out;
4739
4740	item = ring_buffer_event_data(event);
4741	item->size = size;
4742	memcpy(item->str, rb_string, size);
4743
4744	if (nested) {
4745		data->bytes_alloc_nested += event_len;
4746		data->bytes_written_nested += len;
4747		data->events_nested++;
4748		if (!data->min_size_nested || len < data->min_size_nested)
4749			data->min_size_nested = len;
4750		if (len > data->max_size_nested)
4751			data->max_size_nested = len;
4752	} else {
4753		data->bytes_alloc += event_len;
4754		data->bytes_written += len;
4755		data->events++;
4756		if (!data->min_size || len < data->min_size)
4757			data->max_size = len;
4758		if (len > data->max_size)
4759			data->max_size = len;
4760	}
4761
4762 out:
4763	ring_buffer_unlock_commit(data->buffer, event);
4764
4765	return 0;
4766}
4767
4768static __init int rb_test(void *arg)
4769{
4770	struct rb_test_data *data = arg;
4771
4772	while (!kthread_should_stop()) {
4773		rb_write_something(data, false);
4774		data->cnt++;
4775
4776		set_current_state(TASK_INTERRUPTIBLE);
4777		/* Now sleep between a min of 100-300us and a max of 1ms */
4778		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4779	}
4780
4781	return 0;
4782}
4783
4784static __init void rb_ipi(void *ignore)
4785{
4786	struct rb_test_data *data;
4787	int cpu = smp_processor_id();
4788
4789	data = &rb_data[cpu];
4790	rb_write_something(data, true);
4791}
4792
4793static __init int rb_hammer_test(void *arg)
4794{
4795	while (!kthread_should_stop()) {
4796
4797		/* Send an IPI to all cpus to write data! */
4798		smp_call_function(rb_ipi, NULL, 1);
4799		/* No sleep, but for non preempt, let others run */
4800		schedule();
4801	}
4802
4803	return 0;
4804}
4805
4806static __init int test_ringbuffer(void)
4807{
4808	struct task_struct *rb_hammer;
4809	struct ring_buffer *buffer;
4810	int cpu;
4811	int ret = 0;
4812
4813	pr_info("Running ring buffer tests...\n");
4814
4815	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4816	if (WARN_ON(!buffer))
4817		return 0;
4818
4819	/* Disable buffer so that threads can't write to it yet */
4820	ring_buffer_record_off(buffer);
4821
4822	for_each_online_cpu(cpu) {
4823		rb_data[cpu].buffer = buffer;
4824		rb_data[cpu].cpu = cpu;
4825		rb_data[cpu].cnt = cpu;
4826		rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4827						 "rbtester/%d", cpu);
4828		if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4829			pr_cont("FAILED\n");
4830			ret = PTR_ERR(rb_threads[cpu]);
4831			goto out_free;
4832		}
4833
4834		kthread_bind(rb_threads[cpu], cpu);
4835 		wake_up_process(rb_threads[cpu]);
4836	}
4837
4838	/* Now create the rb hammer! */
4839	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4840	if (WARN_ON(IS_ERR(rb_hammer))) {
4841		pr_cont("FAILED\n");
4842		ret = PTR_ERR(rb_hammer);
4843		goto out_free;
4844	}
4845
4846	ring_buffer_record_on(buffer);
4847	/*
4848	 * Show buffer is enabled before setting rb_test_started.
4849	 * Yes there's a small race window where events could be
4850	 * dropped and the thread wont catch it. But when a ring
4851	 * buffer gets enabled, there will always be some kind of
4852	 * delay before other CPUs see it. Thus, we don't care about
4853	 * those dropped events. We care about events dropped after
4854	 * the threads see that the buffer is active.
4855	 */
4856	smp_wmb();
4857	rb_test_started = true;
4858
4859	set_current_state(TASK_INTERRUPTIBLE);
4860	/* Just run for 10 seconds */;
4861	schedule_timeout(10 * HZ);
4862
4863	kthread_stop(rb_hammer);
4864
4865 out_free:
4866	for_each_online_cpu(cpu) {
4867		if (!rb_threads[cpu])
4868			break;
4869		kthread_stop(rb_threads[cpu]);
4870	}
4871	if (ret) {
4872		ring_buffer_free(buffer);
4873		return ret;
4874	}
4875
4876	/* Report! */
4877	pr_info("finished\n");
4878	for_each_online_cpu(cpu) {
4879		struct ring_buffer_event *event;
4880		struct rb_test_data *data = &rb_data[cpu];
4881		struct rb_item *item;
4882		unsigned long total_events;
4883		unsigned long total_dropped;
4884		unsigned long total_written;
4885		unsigned long total_alloc;
4886		unsigned long total_read = 0;
4887		unsigned long total_size = 0;
4888		unsigned long total_len = 0;
4889		unsigned long total_lost = 0;
4890		unsigned long lost;
4891		int big_event_size;
4892		int small_event_size;
4893
4894		ret = -1;
4895
4896		total_events = data->events + data->events_nested;
4897		total_written = data->bytes_written + data->bytes_written_nested;
4898		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4899		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4900
4901		big_event_size = data->max_size + data->max_size_nested;
4902		small_event_size = data->min_size + data->min_size_nested;
4903
4904		pr_info("CPU %d:\n", cpu);
4905		pr_info("              events:    %ld\n", total_events);
4906		pr_info("       dropped bytes:    %ld\n", total_dropped);
4907		pr_info("       alloced bytes:    %ld\n", total_alloc);
4908		pr_info("       written bytes:    %ld\n", total_written);
4909		pr_info("       biggest event:    %d\n", big_event_size);
4910		pr_info("      smallest event:    %d\n", small_event_size);
4911
4912		if (RB_WARN_ON(buffer, total_dropped))
4913			break;
4914
4915		ret = 0;
4916
4917		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4918			total_lost += lost;
4919			item = ring_buffer_event_data(event);
4920			total_len += ring_buffer_event_length(event);
4921			total_size += item->size + sizeof(struct rb_item);
4922			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4923				pr_info("FAILED!\n");
4924				pr_info("buffer had: %.*s\n", item->size, item->str);
4925				pr_info("expected:   %.*s\n", item->size, rb_string);
4926				RB_WARN_ON(buffer, 1);
4927				ret = -1;
4928				break;
4929			}
4930			total_read++;
4931		}
4932		if (ret)
4933			break;
4934
4935		ret = -1;
4936
4937		pr_info("         read events:   %ld\n", total_read);
4938		pr_info("         lost events:   %ld\n", total_lost);
4939		pr_info("        total events:   %ld\n", total_lost + total_read);
4940		pr_info("  recorded len bytes:   %ld\n", total_len);
4941		pr_info(" recorded size bytes:   %ld\n", total_size);
4942		if (total_lost)
4943			pr_info(" With dropped events, record len and size may not match\n"
4944				" alloced and written from above\n");
4945		if (!total_lost) {
4946			if (RB_WARN_ON(buffer, total_len != total_alloc ||
4947				       total_size != total_written))
4948				break;
4949		}
4950		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4951			break;
4952
4953		ret = 0;
4954	}
4955	if (!ret)
4956		pr_info("Ring buffer PASSED!\n");
4957
4958	ring_buffer_free(buffer);
4959	return 0;
4960}
4961
4962late_initcall(test_ringbuffer);
4963#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */