Linux Audio

Check our new training course

Loading...
v4.10.11
 
   1/*
   2 * Generic ring buffer
   3 *
   4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
   5 */
   6#include <linux/trace_events.h>
   7#include <linux/ring_buffer.h>
   8#include <linux/trace_clock.h>
 
   9#include <linux/trace_seq.h>
  10#include <linux/spinlock.h>
  11#include <linux/irq_work.h>
 
  12#include <linux/uaccess.h>
  13#include <linux/hardirq.h>
  14#include <linux/kthread.h>	/* for self test */
  15#include <linux/kmemcheck.h>
  16#include <linux/module.h>
  17#include <linux/percpu.h>
  18#include <linux/mutex.h>
  19#include <linux/delay.h>
  20#include <linux/slab.h>
  21#include <linux/init.h>
  22#include <linux/hash.h>
  23#include <linux/list.h>
  24#include <linux/cpu.h>
 
  25
  26#include <asm/local.h>
  27
  28static void update_pages_handler(struct work_struct *work);
  29
  30/*
  31 * The ring buffer header is special. We must manually up keep it.
  32 */
  33int ring_buffer_print_entry_header(struct trace_seq *s)
  34{
  35	trace_seq_puts(s, "# compressed entry header\n");
  36	trace_seq_puts(s, "\ttype_len    :    5 bits\n");
  37	trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
  38	trace_seq_puts(s, "\tarray       :   32 bits\n");
  39	trace_seq_putc(s, '\n');
  40	trace_seq_printf(s, "\tpadding     : type == %d\n",
  41			 RINGBUF_TYPE_PADDING);
  42	trace_seq_printf(s, "\ttime_extend : type == %d\n",
  43			 RINGBUF_TYPE_TIME_EXTEND);
 
 
  44	trace_seq_printf(s, "\tdata max type_len  == %d\n",
  45			 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
  46
  47	return !trace_seq_has_overflowed(s);
  48}
  49
  50/*
  51 * The ring buffer is made up of a list of pages. A separate list of pages is
  52 * allocated for each CPU. A writer may only write to a buffer that is
  53 * associated with the CPU it is currently executing on.  A reader may read
  54 * from any per cpu buffer.
  55 *
  56 * The reader is special. For each per cpu buffer, the reader has its own
  57 * reader page. When a reader has read the entire reader page, this reader
  58 * page is swapped with another page in the ring buffer.
  59 *
  60 * Now, as long as the writer is off the reader page, the reader can do what
  61 * ever it wants with that page. The writer will never write to that page
  62 * again (as long as it is out of the ring buffer).
  63 *
  64 * Here's some silly ASCII art.
  65 *
  66 *   +------+
  67 *   |reader|          RING BUFFER
  68 *   |page  |
  69 *   +------+        +---+   +---+   +---+
  70 *                   |   |-->|   |-->|   |
  71 *                   +---+   +---+   +---+
  72 *                     ^               |
  73 *                     |               |
  74 *                     +---------------+
  75 *
  76 *
  77 *   +------+
  78 *   |reader|          RING BUFFER
  79 *   |page  |------------------v
  80 *   +------+        +---+   +---+   +---+
  81 *                   |   |-->|   |-->|   |
  82 *                   +---+   +---+   +---+
  83 *                     ^               |
  84 *                     |               |
  85 *                     +---------------+
  86 *
  87 *
  88 *   +------+
  89 *   |reader|          RING BUFFER
  90 *   |page  |------------------v
  91 *   +------+        +---+   +---+   +---+
  92 *      ^            |   |-->|   |-->|   |
  93 *      |            +---+   +---+   +---+
  94 *      |                              |
  95 *      |                              |
  96 *      +------------------------------+
  97 *
  98 *
  99 *   +------+
 100 *   |buffer|          RING BUFFER
 101 *   |page  |------------------v
 102 *   +------+        +---+   +---+   +---+
 103 *      ^            |   |   |   |-->|   |
 104 *      |   New      +---+   +---+   +---+
 105 *      |  Reader------^               |
 106 *      |   page                       |
 107 *      +------------------------------+
 108 *
 109 *
 110 * After we make this swap, the reader can hand this page off to the splice
 111 * code and be done with it. It can even allocate a new page if it needs to
 112 * and swap that into the ring buffer.
 113 *
 114 * We will be using cmpxchg soon to make all this lockless.
 115 *
 116 */
 117
 118/* Used for individual buffers (after the counter) */
 119#define RB_BUFFER_OFF		(1 << 20)
 120
 121#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
 122
 123#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
 124#define RB_ALIGNMENT		4U
 125#define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 126#define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
 127
 128#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
 129# define RB_FORCE_8BYTE_ALIGNMENT	0
 130# define RB_ARCH_ALIGNMENT		RB_ALIGNMENT
 131#else
 132# define RB_FORCE_8BYTE_ALIGNMENT	1
 133# define RB_ARCH_ALIGNMENT		8U
 134#endif
 135
 136#define RB_ALIGN_DATA		__aligned(RB_ARCH_ALIGNMENT)
 137
 138/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
 139#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
 140
 141enum {
 142	RB_LEN_TIME_EXTEND = 8,
 143	RB_LEN_TIME_STAMP = 16,
 144};
 145
 146#define skip_time_extend(event) \
 147	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
 148
 
 
 
 149static inline int rb_null_event(struct ring_buffer_event *event)
 150{
 151	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
 152}
 153
 154static void rb_event_set_padding(struct ring_buffer_event *event)
 155{
 156	/* padding has a NULL time_delta */
 157	event->type_len = RINGBUF_TYPE_PADDING;
 158	event->time_delta = 0;
 159}
 160
 161static unsigned
 162rb_event_data_length(struct ring_buffer_event *event)
 163{
 164	unsigned length;
 165
 166	if (event->type_len)
 167		length = event->type_len * RB_ALIGNMENT;
 168	else
 169		length = event->array[0];
 170	return length + RB_EVNT_HDR_SIZE;
 171}
 172
 173/*
 174 * Return the length of the given event. Will return
 175 * the length of the time extend if the event is a
 176 * time extend.
 177 */
 178static inline unsigned
 179rb_event_length(struct ring_buffer_event *event)
 180{
 181	switch (event->type_len) {
 182	case RINGBUF_TYPE_PADDING:
 183		if (rb_null_event(event))
 184			/* undefined */
 185			return -1;
 186		return  event->array[0] + RB_EVNT_HDR_SIZE;
 187
 188	case RINGBUF_TYPE_TIME_EXTEND:
 189		return RB_LEN_TIME_EXTEND;
 190
 191	case RINGBUF_TYPE_TIME_STAMP:
 192		return RB_LEN_TIME_STAMP;
 193
 194	case RINGBUF_TYPE_DATA:
 195		return rb_event_data_length(event);
 196	default:
 197		BUG();
 198	}
 199	/* not hit */
 200	return 0;
 201}
 202
 203/*
 204 * Return total length of time extend and data,
 205 *   or just the event length for all other events.
 206 */
 207static inline unsigned
 208rb_event_ts_length(struct ring_buffer_event *event)
 209{
 210	unsigned len = 0;
 211
 212	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
 213		/* time extends include the data event after it */
 214		len = RB_LEN_TIME_EXTEND;
 215		event = skip_time_extend(event);
 216	}
 217	return len + rb_event_length(event);
 218}
 219
 220/**
 221 * ring_buffer_event_length - return the length of the event
 222 * @event: the event to get the length of
 223 *
 224 * Returns the size of the data load of a data event.
 225 * If the event is something other than a data event, it
 226 * returns the size of the event itself. With the exception
 227 * of a TIME EXTEND, where it still returns the size of the
 228 * data load of the data event after it.
 229 */
 230unsigned ring_buffer_event_length(struct ring_buffer_event *event)
 231{
 232	unsigned length;
 233
 234	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
 235		event = skip_time_extend(event);
 236
 237	length = rb_event_length(event);
 238	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 239		return length;
 240	length -= RB_EVNT_HDR_SIZE;
 241	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
 242                length -= sizeof(event->array[0]);
 243	return length;
 244}
 245EXPORT_SYMBOL_GPL(ring_buffer_event_length);
 246
 247/* inline for ring buffer fast paths */
 248static __always_inline void *
 249rb_event_data(struct ring_buffer_event *event)
 250{
 251	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
 252		event = skip_time_extend(event);
 253	BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 254	/* If length is in len field, then array[0] has the data */
 255	if (event->type_len)
 256		return (void *)&event->array[0];
 257	/* Otherwise length is in array[0] and array[1] has the data */
 258	return (void *)&event->array[1];
 259}
 260
 261/**
 262 * ring_buffer_event_data - return the data of the event
 263 * @event: the event to get the data from
 264 */
 265void *ring_buffer_event_data(struct ring_buffer_event *event)
 266{
 267	return rb_event_data(event);
 268}
 269EXPORT_SYMBOL_GPL(ring_buffer_event_data);
 270
 271#define for_each_buffer_cpu(buffer, cpu)		\
 272	for_each_cpu(cpu, buffer->cpumask)
 273
 
 
 
 274#define TS_SHIFT	27
 275#define TS_MASK		((1ULL << TS_SHIFT) - 1)
 276#define TS_DELTA_TEST	(~TS_MASK)
 277
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 278/* Flag when events were overwritten */
 279#define RB_MISSED_EVENTS	(1 << 31)
 280/* Missed count stored at end */
 281#define RB_MISSED_STORED	(1 << 30)
 282
 283struct buffer_data_page {
 284	u64		 time_stamp;	/* page time stamp */
 285	local_t		 commit;	/* write committed index */
 286	unsigned char	 data[] RB_ALIGN_DATA;	/* data of buffer page */
 287};
 288
 289/*
 290 * Note, the buffer_page list must be first. The buffer pages
 291 * are allocated in cache lines, which means that each buffer
 292 * page will be at the beginning of a cache line, and thus
 293 * the least significant bits will be zero. We use this to
 294 * add flags in the list struct pointers, to make the ring buffer
 295 * lockless.
 296 */
 297struct buffer_page {
 298	struct list_head list;		/* list of buffer pages */
 299	local_t		 write;		/* index for next write */
 300	unsigned	 read;		/* index for next read */
 301	local_t		 entries;	/* entries on this page */
 302	unsigned long	 real_end;	/* real end of data */
 303	struct buffer_data_page *page;	/* Actual data page */
 304};
 305
 306/*
 307 * The buffer page counters, write and entries, must be reset
 308 * atomically when crossing page boundaries. To synchronize this
 309 * update, two counters are inserted into the number. One is
 310 * the actual counter for the write position or count on the page.
 311 *
 312 * The other is a counter of updaters. Before an update happens
 313 * the update partition of the counter is incremented. This will
 314 * allow the updater to update the counter atomically.
 315 *
 316 * The counter is 20 bits, and the state data is 12.
 317 */
 318#define RB_WRITE_MASK		0xfffff
 319#define RB_WRITE_INTCNT		(1 << 20)
 320
 321static void rb_init_page(struct buffer_data_page *bpage)
 322{
 323	local_set(&bpage->commit, 0);
 324}
 325
 326/**
 327 * ring_buffer_page_len - the size of data on the page.
 328 * @page: The page to read
 329 *
 330 * Returns the amount of data on the page, including buffer page header.
 331 */
 332size_t ring_buffer_page_len(void *page)
 333{
 334	return local_read(&((struct buffer_data_page *)page)->commit)
 335		+ BUF_PAGE_HDR_SIZE;
 336}
 337
 338/*
 339 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
 340 * this issue out.
 341 */
 342static void free_buffer_page(struct buffer_page *bpage)
 343{
 344	free_page((unsigned long)bpage->page);
 345	kfree(bpage);
 346}
 347
 348/*
 349 * We need to fit the time_stamp delta into 27 bits.
 350 */
 351static inline int test_time_stamp(u64 delta)
 352{
 353	if (delta & TS_DELTA_TEST)
 354		return 1;
 355	return 0;
 356}
 357
 358#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
 359
 360/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
 361#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
 362
 363int ring_buffer_print_page_header(struct trace_seq *s)
 364{
 365	struct buffer_data_page field;
 366
 367	trace_seq_printf(s, "\tfield: u64 timestamp;\t"
 368			 "offset:0;\tsize:%u;\tsigned:%u;\n",
 369			 (unsigned int)sizeof(field.time_stamp),
 370			 (unsigned int)is_signed_type(u64));
 371
 372	trace_seq_printf(s, "\tfield: local_t commit;\t"
 373			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 374			 (unsigned int)offsetof(typeof(field), commit),
 375			 (unsigned int)sizeof(field.commit),
 376			 (unsigned int)is_signed_type(long));
 377
 378	trace_seq_printf(s, "\tfield: int overwrite;\t"
 379			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 380			 (unsigned int)offsetof(typeof(field), commit),
 381			 1,
 382			 (unsigned int)is_signed_type(long));
 383
 384	trace_seq_printf(s, "\tfield: char data;\t"
 385			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 386			 (unsigned int)offsetof(typeof(field), data),
 387			 (unsigned int)BUF_PAGE_SIZE,
 388			 (unsigned int)is_signed_type(char));
 389
 390	return !trace_seq_has_overflowed(s);
 391}
 392
 393struct rb_irq_work {
 394	struct irq_work			work;
 395	wait_queue_head_t		waiters;
 396	wait_queue_head_t		full_waiters;
 397	bool				waiters_pending;
 398	bool				full_waiters_pending;
 399	bool				wakeup_full;
 400};
 401
 402/*
 403 * Structure to hold event state and handle nested events.
 404 */
 405struct rb_event_info {
 406	u64			ts;
 407	u64			delta;
 
 
 408	unsigned long		length;
 409	struct buffer_page	*tail_page;
 410	int			add_timestamp;
 411};
 412
 413/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 414 * Used for which event context the event is in.
 415 *  NMI     = 0
 416 *  IRQ     = 1
 417 *  SOFTIRQ = 2
 418 *  NORMAL  = 3
 419 *
 420 * See trace_recursive_lock() comment below for more details.
 421 */
 422enum {
 423	RB_CTX_NMI,
 424	RB_CTX_IRQ,
 425	RB_CTX_SOFTIRQ,
 426	RB_CTX_NORMAL,
 427	RB_CTX_MAX
 428};
 429
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 430/*
 431 * head_page == tail_page && head == tail then buffer is empty.
 432 */
 433struct ring_buffer_per_cpu {
 434	int				cpu;
 435	atomic_t			record_disabled;
 436	struct ring_buffer		*buffer;
 
 437	raw_spinlock_t			reader_lock;	/* serialize readers */
 438	arch_spinlock_t			lock;
 439	struct lock_class_key		lock_key;
 
 440	unsigned long			nr_pages;
 441	unsigned int			current_context;
 442	struct list_head		*pages;
 443	struct buffer_page		*head_page;	/* read from head */
 444	struct buffer_page		*tail_page;	/* write to tail */
 445	struct buffer_page		*commit_page;	/* committed pages */
 446	struct buffer_page		*reader_page;
 447	unsigned long			lost_events;
 448	unsigned long			last_overrun;
 
 449	local_t				entries_bytes;
 450	local_t				entries;
 451	local_t				overrun;
 452	local_t				commit_overrun;
 453	local_t				dropped_events;
 454	local_t				committing;
 455	local_t				commits;
 
 
 
 
 456	unsigned long			read;
 457	unsigned long			read_bytes;
 458	u64				write_stamp;
 
 459	u64				read_stamp;
 460	/* ring buffer pages to update, > 0 to add, < 0 to remove */
 461	long				nr_pages_to_update;
 462	struct list_head		new_pages; /* new pages to add */
 463	struct work_struct		update_pages_work;
 464	struct completion		update_done;
 465
 466	struct rb_irq_work		irq_work;
 467};
 468
 469struct ring_buffer {
 470	unsigned			flags;
 471	int				cpus;
 472	atomic_t			record_disabled;
 473	atomic_t			resize_disabled;
 474	cpumask_var_t			cpumask;
 475
 476	struct lock_class_key		*reader_lock_key;
 477
 478	struct mutex			mutex;
 479
 480	struct ring_buffer_per_cpu	**buffers;
 481
 482	struct hlist_node		node;
 483	u64				(*clock)(void);
 484
 485	struct rb_irq_work		irq_work;
 
 486};
 487
 488struct ring_buffer_iter {
 489	struct ring_buffer_per_cpu	*cpu_buffer;
 490	unsigned long			head;
 
 491	struct buffer_page		*head_page;
 492	struct buffer_page		*cache_reader_page;
 493	unsigned long			cache_read;
 494	u64				read_stamp;
 
 
 
 495};
 496
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 497/*
 498 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
 499 *
 500 * Schedules a delayed work to wake up any task that is blocked on the
 501 * ring buffer waiters queue.
 502 */
 503static void rb_wake_up_waiters(struct irq_work *work)
 504{
 505	struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
 506
 507	wake_up_all(&rbwork->waiters);
 508	if (rbwork->wakeup_full) {
 509		rbwork->wakeup_full = false;
 510		wake_up_all(&rbwork->full_waiters);
 511	}
 512}
 513
 514/**
 515 * ring_buffer_wait - wait for input to the ring buffer
 516 * @buffer: buffer to wait on
 517 * @cpu: the cpu buffer to wait on
 518 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
 519 *
 520 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 521 * as data is added to any of the @buffer's cpu buffers. Otherwise
 522 * it will wait for data to be added to a specific cpu buffer.
 523 */
 524int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
 525{
 526	struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
 527	DEFINE_WAIT(wait);
 528	struct rb_irq_work *work;
 529	int ret = 0;
 530
 531	/*
 532	 * Depending on what the caller is waiting for, either any
 533	 * data in any cpu buffer, or a specific buffer, put the
 534	 * caller on the appropriate wait queue.
 535	 */
 536	if (cpu == RING_BUFFER_ALL_CPUS) {
 537		work = &buffer->irq_work;
 538		/* Full only makes sense on per cpu reads */
 539		full = false;
 540	} else {
 541		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 542			return -ENODEV;
 543		cpu_buffer = buffer->buffers[cpu];
 544		work = &cpu_buffer->irq_work;
 545	}
 546
 547
 548	while (true) {
 549		if (full)
 550			prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
 551		else
 552			prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
 553
 554		/*
 555		 * The events can happen in critical sections where
 556		 * checking a work queue can cause deadlocks.
 557		 * After adding a task to the queue, this flag is set
 558		 * only to notify events to try to wake up the queue
 559		 * using irq_work.
 560		 *
 561		 * We don't clear it even if the buffer is no longer
 562		 * empty. The flag only causes the next event to run
 563		 * irq_work to do the work queue wake up. The worse
 564		 * that can happen if we race with !trace_empty() is that
 565		 * an event will cause an irq_work to try to wake up
 566		 * an empty queue.
 567		 *
 568		 * There's no reason to protect this flag either, as
 569		 * the work queue and irq_work logic will do the necessary
 570		 * synchronization for the wake ups. The only thing
 571		 * that is necessary is that the wake up happens after
 572		 * a task has been queued. It's OK for spurious wake ups.
 573		 */
 574		if (full)
 575			work->full_waiters_pending = true;
 576		else
 577			work->waiters_pending = true;
 578
 579		if (signal_pending(current)) {
 580			ret = -EINTR;
 581			break;
 582		}
 583
 584		if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
 585			break;
 586
 587		if (cpu != RING_BUFFER_ALL_CPUS &&
 588		    !ring_buffer_empty_cpu(buffer, cpu)) {
 589			unsigned long flags;
 590			bool pagebusy;
 
 
 591
 592			if (!full)
 593				break;
 594
 595			raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
 596			pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 
 
 
 
 
 597			raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
 598
 599			if (!pagebusy)
 600				break;
 601		}
 602
 603		schedule();
 604	}
 605
 606	if (full)
 607		finish_wait(&work->full_waiters, &wait);
 608	else
 609		finish_wait(&work->waiters, &wait);
 610
 611	return ret;
 612}
 613
 614/**
 615 * ring_buffer_poll_wait - poll on buffer input
 616 * @buffer: buffer to wait on
 617 * @cpu: the cpu buffer to wait on
 618 * @filp: the file descriptor
 619 * @poll_table: The poll descriptor
 620 *
 621 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 622 * as data is added to any of the @buffer's cpu buffers. Otherwise
 623 * it will wait for data to be added to a specific cpu buffer.
 624 *
 625 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
 626 * zero otherwise.
 627 */
 628int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
 629			  struct file *filp, poll_table *poll_table)
 630{
 631	struct ring_buffer_per_cpu *cpu_buffer;
 632	struct rb_irq_work *work;
 633
 634	if (cpu == RING_BUFFER_ALL_CPUS)
 635		work = &buffer->irq_work;
 636	else {
 637		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 638			return -EINVAL;
 639
 640		cpu_buffer = buffer->buffers[cpu];
 641		work = &cpu_buffer->irq_work;
 642	}
 643
 644	poll_wait(filp, &work->waiters, poll_table);
 645	work->waiters_pending = true;
 646	/*
 647	 * There's a tight race between setting the waiters_pending and
 648	 * checking if the ring buffer is empty.  Once the waiters_pending bit
 649	 * is set, the next event will wake the task up, but we can get stuck
 650	 * if there's only a single event in.
 651	 *
 652	 * FIXME: Ideally, we need a memory barrier on the writer side as well,
 653	 * but adding a memory barrier to all events will cause too much of a
 654	 * performance hit in the fast path.  We only need a memory barrier when
 655	 * the buffer goes from empty to having content.  But as this race is
 656	 * extremely small, and it's not a problem if another event comes in, we
 657	 * will fix it later.
 658	 */
 659	smp_mb();
 660
 661	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
 662	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
 663		return POLLIN | POLLRDNORM;
 664	return 0;
 665}
 666
 667/* buffer may be either ring_buffer or ring_buffer_per_cpu */
 668#define RB_WARN_ON(b, cond)						\
 669	({								\
 670		int _____ret = unlikely(cond);				\
 671		if (_____ret) {						\
 672			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
 673				struct ring_buffer_per_cpu *__b =	\
 674					(void *)b;			\
 675				atomic_inc(&__b->buffer->record_disabled); \
 676			} else						\
 677				atomic_inc(&b->record_disabled);	\
 678			WARN_ON(1);					\
 679		}							\
 680		_____ret;						\
 681	})
 682
 683/* Up this if you want to test the TIME_EXTENTS and normalization */
 684#define DEBUG_SHIFT 0
 685
 686static inline u64 rb_time_stamp(struct ring_buffer *buffer)
 687{
 
 
 
 
 
 
 
 
 688	/* shift to debug/test normalization and TIME_EXTENTS */
 689	return buffer->clock() << DEBUG_SHIFT;
 690}
 691
 692u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
 693{
 694	u64 time;
 695
 696	preempt_disable_notrace();
 697	time = rb_time_stamp(buffer);
 698	preempt_enable_no_resched_notrace();
 699
 700	return time;
 701}
 702EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
 703
 704void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
 705				      int cpu, u64 *ts)
 706{
 707	/* Just stupid testing the normalize function and deltas */
 708	*ts >>= DEBUG_SHIFT;
 709}
 710EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
 711
 712/*
 713 * Making the ring buffer lockless makes things tricky.
 714 * Although writes only happen on the CPU that they are on,
 715 * and they only need to worry about interrupts. Reads can
 716 * happen on any CPU.
 717 *
 718 * The reader page is always off the ring buffer, but when the
 719 * reader finishes with a page, it needs to swap its page with
 720 * a new one from the buffer. The reader needs to take from
 721 * the head (writes go to the tail). But if a writer is in overwrite
 722 * mode and wraps, it must push the head page forward.
 723 *
 724 * Here lies the problem.
 725 *
 726 * The reader must be careful to replace only the head page, and
 727 * not another one. As described at the top of the file in the
 728 * ASCII art, the reader sets its old page to point to the next
 729 * page after head. It then sets the page after head to point to
 730 * the old reader page. But if the writer moves the head page
 731 * during this operation, the reader could end up with the tail.
 732 *
 733 * We use cmpxchg to help prevent this race. We also do something
 734 * special with the page before head. We set the LSB to 1.
 735 *
 736 * When the writer must push the page forward, it will clear the
 737 * bit that points to the head page, move the head, and then set
 738 * the bit that points to the new head page.
 739 *
 740 * We also don't want an interrupt coming in and moving the head
 741 * page on another writer. Thus we use the second LSB to catch
 742 * that too. Thus:
 743 *
 744 * head->list->prev->next        bit 1          bit 0
 745 *                              -------        -------
 746 * Normal page                     0              0
 747 * Points to head page             0              1
 748 * New head page                   1              0
 749 *
 750 * Note we can not trust the prev pointer of the head page, because:
 751 *
 752 * +----+       +-----+        +-----+
 753 * |    |------>|  T  |---X--->|  N  |
 754 * |    |<------|     |        |     |
 755 * +----+       +-----+        +-----+
 756 *   ^                           ^ |
 757 *   |          +-----+          | |
 758 *   +----------|  R  |----------+ |
 759 *              |     |<-----------+
 760 *              +-----+
 761 *
 762 * Key:  ---X-->  HEAD flag set in pointer
 763 *         T      Tail page
 764 *         R      Reader page
 765 *         N      Next page
 766 *
 767 * (see __rb_reserve_next() to see where this happens)
 768 *
 769 *  What the above shows is that the reader just swapped out
 770 *  the reader page with a page in the buffer, but before it
 771 *  could make the new header point back to the new page added
 772 *  it was preempted by a writer. The writer moved forward onto
 773 *  the new page added by the reader and is about to move forward
 774 *  again.
 775 *
 776 *  You can see, it is legitimate for the previous pointer of
 777 *  the head (or any page) not to point back to itself. But only
 778 *  temporarially.
 779 */
 780
 781#define RB_PAGE_NORMAL		0UL
 782#define RB_PAGE_HEAD		1UL
 783#define RB_PAGE_UPDATE		2UL
 784
 785
 786#define RB_FLAG_MASK		3UL
 787
 788/* PAGE_MOVED is not part of the mask */
 789#define RB_PAGE_MOVED		4UL
 790
 791/*
 792 * rb_list_head - remove any bit
 793 */
 794static struct list_head *rb_list_head(struct list_head *list)
 795{
 796	unsigned long val = (unsigned long)list;
 797
 798	return (struct list_head *)(val & ~RB_FLAG_MASK);
 799}
 800
 801/*
 802 * rb_is_head_page - test if the given page is the head page
 803 *
 804 * Because the reader may move the head_page pointer, we can
 805 * not trust what the head page is (it may be pointing to
 806 * the reader page). But if the next page is a header page,
 807 * its flags will be non zero.
 808 */
 809static inline int
 810rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
 811		struct buffer_page *page, struct list_head *list)
 812{
 813	unsigned long val;
 814
 815	val = (unsigned long)list->next;
 816
 817	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
 818		return RB_PAGE_MOVED;
 819
 820	return val & RB_FLAG_MASK;
 821}
 822
 823/*
 824 * rb_is_reader_page
 825 *
 826 * The unique thing about the reader page, is that, if the
 827 * writer is ever on it, the previous pointer never points
 828 * back to the reader page.
 829 */
 830static bool rb_is_reader_page(struct buffer_page *page)
 831{
 832	struct list_head *list = page->list.prev;
 833
 834	return rb_list_head(list->next) != &page->list;
 835}
 836
 837/*
 838 * rb_set_list_to_head - set a list_head to be pointing to head.
 839 */
 840static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
 841				struct list_head *list)
 842{
 843	unsigned long *ptr;
 844
 845	ptr = (unsigned long *)&list->next;
 846	*ptr |= RB_PAGE_HEAD;
 847	*ptr &= ~RB_PAGE_UPDATE;
 848}
 849
 850/*
 851 * rb_head_page_activate - sets up head page
 852 */
 853static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
 854{
 855	struct buffer_page *head;
 856
 857	head = cpu_buffer->head_page;
 858	if (!head)
 859		return;
 860
 861	/*
 862	 * Set the previous list pointer to have the HEAD flag.
 863	 */
 864	rb_set_list_to_head(cpu_buffer, head->list.prev);
 865}
 866
 867static void rb_list_head_clear(struct list_head *list)
 868{
 869	unsigned long *ptr = (unsigned long *)&list->next;
 870
 871	*ptr &= ~RB_FLAG_MASK;
 872}
 873
 874/*
 875 * rb_head_page_dactivate - clears head page ptr (for free list)
 876 */
 877static void
 878rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
 879{
 880	struct list_head *hd;
 881
 882	/* Go through the whole list and clear any pointers found. */
 883	rb_list_head_clear(cpu_buffer->pages);
 884
 885	list_for_each(hd, cpu_buffer->pages)
 886		rb_list_head_clear(hd);
 887}
 888
 889static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
 890			    struct buffer_page *head,
 891			    struct buffer_page *prev,
 892			    int old_flag, int new_flag)
 893{
 894	struct list_head *list;
 895	unsigned long val = (unsigned long)&head->list;
 896	unsigned long ret;
 897
 898	list = &prev->list;
 899
 900	val &= ~RB_FLAG_MASK;
 901
 902	ret = cmpxchg((unsigned long *)&list->next,
 903		      val | old_flag, val | new_flag);
 904
 905	/* check if the reader took the page */
 906	if ((ret & ~RB_FLAG_MASK) != val)
 907		return RB_PAGE_MOVED;
 908
 909	return ret & RB_FLAG_MASK;
 910}
 911
 912static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
 913				   struct buffer_page *head,
 914				   struct buffer_page *prev,
 915				   int old_flag)
 916{
 917	return rb_head_page_set(cpu_buffer, head, prev,
 918				old_flag, RB_PAGE_UPDATE);
 919}
 920
 921static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
 922				 struct buffer_page *head,
 923				 struct buffer_page *prev,
 924				 int old_flag)
 925{
 926	return rb_head_page_set(cpu_buffer, head, prev,
 927				old_flag, RB_PAGE_HEAD);
 928}
 929
 930static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
 931				   struct buffer_page *head,
 932				   struct buffer_page *prev,
 933				   int old_flag)
 934{
 935	return rb_head_page_set(cpu_buffer, head, prev,
 936				old_flag, RB_PAGE_NORMAL);
 937}
 938
 939static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
 940			       struct buffer_page **bpage)
 941{
 942	struct list_head *p = rb_list_head((*bpage)->list.next);
 943
 944	*bpage = list_entry(p, struct buffer_page, list);
 945}
 946
 947static struct buffer_page *
 948rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
 949{
 950	struct buffer_page *head;
 951	struct buffer_page *page;
 952	struct list_head *list;
 953	int i;
 954
 955	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
 956		return NULL;
 957
 958	/* sanity check */
 959	list = cpu_buffer->pages;
 960	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
 961		return NULL;
 962
 963	page = head = cpu_buffer->head_page;
 964	/*
 965	 * It is possible that the writer moves the header behind
 966	 * where we started, and we miss in one loop.
 967	 * A second loop should grab the header, but we'll do
 968	 * three loops just because I'm paranoid.
 969	 */
 970	for (i = 0; i < 3; i++) {
 971		do {
 972			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
 973				cpu_buffer->head_page = page;
 974				return page;
 975			}
 976			rb_inc_page(cpu_buffer, &page);
 977		} while (page != head);
 978	}
 979
 980	RB_WARN_ON(cpu_buffer, 1);
 981
 982	return NULL;
 983}
 984
 985static int rb_head_page_replace(struct buffer_page *old,
 986				struct buffer_page *new)
 987{
 988	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
 989	unsigned long val;
 990	unsigned long ret;
 991
 992	val = *ptr & ~RB_FLAG_MASK;
 993	val |= RB_PAGE_HEAD;
 994
 995	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
 996
 997	return ret == val;
 998}
 999
1000/*
1001 * rb_tail_page_update - move the tail page forward
1002 */
1003static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1004			       struct buffer_page *tail_page,
1005			       struct buffer_page *next_page)
1006{
1007	unsigned long old_entries;
1008	unsigned long old_write;
1009
1010	/*
1011	 * The tail page now needs to be moved forward.
1012	 *
1013	 * We need to reset the tail page, but without messing
1014	 * with possible erasing of data brought in by interrupts
1015	 * that have moved the tail page and are currently on it.
1016	 *
1017	 * We add a counter to the write field to denote this.
1018	 */
1019	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1020	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1021
 
1022	/*
1023	 * Just make sure we have seen our old_write and synchronize
1024	 * with any interrupts that come in.
1025	 */
1026	barrier();
1027
1028	/*
1029	 * If the tail page is still the same as what we think
1030	 * it is, then it is up to us to update the tail
1031	 * pointer.
1032	 */
1033	if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1034		/* Zero the write counter */
1035		unsigned long val = old_write & ~RB_WRITE_MASK;
1036		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1037
1038		/*
1039		 * This will only succeed if an interrupt did
1040		 * not come in and change it. In which case, we
1041		 * do not want to modify it.
1042		 *
1043		 * We add (void) to let the compiler know that we do not care
1044		 * about the return value of these functions. We use the
1045		 * cmpxchg to only update if an interrupt did not already
1046		 * do it for us. If the cmpxchg fails, we don't care.
1047		 */
1048		(void)local_cmpxchg(&next_page->write, old_write, val);
1049		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1050
1051		/*
1052		 * No need to worry about races with clearing out the commit.
1053		 * it only can increment when a commit takes place. But that
1054		 * only happens in the outer most nested commit.
1055		 */
1056		local_set(&next_page->page->commit, 0);
1057
1058		/* Again, either we update tail_page or an interrupt does */
1059		(void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1060	}
1061}
1062
1063static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1064			  struct buffer_page *bpage)
1065{
1066	unsigned long val = (unsigned long)bpage;
1067
1068	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1069		return 1;
1070
1071	return 0;
1072}
1073
1074/**
1075 * rb_check_list - make sure a pointer to a list has the last bits zero
1076 */
1077static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1078			 struct list_head *list)
1079{
1080	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1081		return 1;
1082	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1083		return 1;
1084	return 0;
1085}
1086
1087/**
1088 * rb_check_pages - integrity check of buffer pages
1089 * @cpu_buffer: CPU buffer with pages to test
1090 *
1091 * As a safety measure we check to make sure the data pages have not
1092 * been corrupted.
1093 */
1094static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1095{
1096	struct list_head *head = cpu_buffer->pages;
1097	struct buffer_page *bpage, *tmp;
1098
1099	/* Reset the head page if it exists */
1100	if (cpu_buffer->head_page)
1101		rb_set_head_page(cpu_buffer);
1102
1103	rb_head_page_deactivate(cpu_buffer);
1104
1105	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1106		return -1;
1107	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1108		return -1;
1109
1110	if (rb_check_list(cpu_buffer, head))
1111		return -1;
1112
1113	list_for_each_entry_safe(bpage, tmp, head, list) {
1114		if (RB_WARN_ON(cpu_buffer,
1115			       bpage->list.next->prev != &bpage->list))
1116			return -1;
1117		if (RB_WARN_ON(cpu_buffer,
1118			       bpage->list.prev->next != &bpage->list))
1119			return -1;
1120		if (rb_check_list(cpu_buffer, &bpage->list))
1121			return -1;
1122	}
1123
1124	rb_head_page_activate(cpu_buffer);
1125
1126	return 0;
1127}
1128
1129static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1130{
1131	struct buffer_page *bpage, *tmp;
 
 
1132	long i;
1133
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1134	for (i = 0; i < nr_pages; i++) {
1135		struct page *page;
1136		/*
1137		 * __GFP_NORETRY flag makes sure that the allocation fails
1138		 * gracefully without invoking oom-killer and the system is
1139		 * not destabilized.
1140		 */
1141		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1142				    GFP_KERNEL | __GFP_NORETRY,
1143				    cpu_to_node(cpu));
1144		if (!bpage)
1145			goto free_pages;
1146
1147		list_add(&bpage->list, pages);
1148
1149		page = alloc_pages_node(cpu_to_node(cpu),
1150					GFP_KERNEL | __GFP_NORETRY, 0);
1151		if (!page)
1152			goto free_pages;
1153		bpage->page = page_address(page);
1154		rb_init_page(bpage->page);
 
 
 
1155	}
 
 
1156
1157	return 0;
1158
1159free_pages:
1160	list_for_each_entry_safe(bpage, tmp, pages, list) {
1161		list_del_init(&bpage->list);
1162		free_buffer_page(bpage);
1163	}
 
 
1164
1165	return -ENOMEM;
1166}
1167
1168static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1169			     unsigned long nr_pages)
1170{
1171	LIST_HEAD(pages);
1172
1173	WARN_ON(!nr_pages);
1174
1175	if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1176		return -ENOMEM;
1177
1178	/*
1179	 * The ring buffer page list is a circular list that does not
1180	 * start and end with a list head. All page list items point to
1181	 * other pages.
1182	 */
1183	cpu_buffer->pages = pages.next;
1184	list_del(&pages);
1185
1186	cpu_buffer->nr_pages = nr_pages;
1187
1188	rb_check_pages(cpu_buffer);
1189
1190	return 0;
1191}
1192
1193static struct ring_buffer_per_cpu *
1194rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1195{
1196	struct ring_buffer_per_cpu *cpu_buffer;
1197	struct buffer_page *bpage;
1198	struct page *page;
1199	int ret;
1200
1201	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1202				  GFP_KERNEL, cpu_to_node(cpu));
1203	if (!cpu_buffer)
1204		return NULL;
1205
1206	cpu_buffer->cpu = cpu;
1207	cpu_buffer->buffer = buffer;
1208	raw_spin_lock_init(&cpu_buffer->reader_lock);
1209	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1210	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1211	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1212	init_completion(&cpu_buffer->update_done);
1213	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1214	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1215	init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1216
1217	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1218			    GFP_KERNEL, cpu_to_node(cpu));
1219	if (!bpage)
1220		goto fail_free_buffer;
1221
1222	rb_check_bpage(cpu_buffer, bpage);
1223
1224	cpu_buffer->reader_page = bpage;
1225	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1226	if (!page)
1227		goto fail_free_reader;
1228	bpage->page = page_address(page);
1229	rb_init_page(bpage->page);
1230
1231	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1232	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1233
1234	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1235	if (ret < 0)
1236		goto fail_free_reader;
1237
1238	cpu_buffer->head_page
1239		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1240	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1241
1242	rb_head_page_activate(cpu_buffer);
1243
1244	return cpu_buffer;
1245
1246 fail_free_reader:
1247	free_buffer_page(cpu_buffer->reader_page);
1248
1249 fail_free_buffer:
1250	kfree(cpu_buffer);
1251	return NULL;
1252}
1253
1254static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1255{
1256	struct list_head *head = cpu_buffer->pages;
1257	struct buffer_page *bpage, *tmp;
1258
1259	free_buffer_page(cpu_buffer->reader_page);
1260
1261	rb_head_page_deactivate(cpu_buffer);
1262
1263	if (head) {
1264		list_for_each_entry_safe(bpage, tmp, head, list) {
1265			list_del_init(&bpage->list);
1266			free_buffer_page(bpage);
1267		}
1268		bpage = list_entry(head, struct buffer_page, list);
1269		free_buffer_page(bpage);
1270	}
1271
1272	kfree(cpu_buffer);
1273}
1274
1275/**
1276 * __ring_buffer_alloc - allocate a new ring_buffer
1277 * @size: the size in bytes per cpu that is needed.
1278 * @flags: attributes to set for the ring buffer.
 
1279 *
1280 * Currently the only flag that is available is the RB_FL_OVERWRITE
1281 * flag. This flag means that the buffer will overwrite old data
1282 * when the buffer wraps. If this flag is not set, the buffer will
1283 * drop data when the tail hits the head.
1284 */
1285struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1286					struct lock_class_key *key)
1287{
1288	struct ring_buffer *buffer;
1289	long nr_pages;
1290	int bsize;
1291	int cpu;
1292	int ret;
1293
1294	/* keep it in its own cache line */
1295	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1296			 GFP_KERNEL);
1297	if (!buffer)
1298		return NULL;
1299
1300	if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1301		goto fail_free_buffer;
1302
1303	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1304	buffer->flags = flags;
1305	buffer->clock = trace_clock_local;
1306	buffer->reader_lock_key = key;
1307
1308	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1309	init_waitqueue_head(&buffer->irq_work.waiters);
1310
1311	/* need at least two pages */
1312	if (nr_pages < 2)
1313		nr_pages = 2;
1314
1315	buffer->cpus = nr_cpu_ids;
1316
1317	bsize = sizeof(void *) * nr_cpu_ids;
1318	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1319				  GFP_KERNEL);
1320	if (!buffer->buffers)
1321		goto fail_free_cpumask;
1322
1323	cpu = raw_smp_processor_id();
1324	cpumask_set_cpu(cpu, buffer->cpumask);
1325	buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1326	if (!buffer->buffers[cpu])
1327		goto fail_free_buffers;
1328
1329	ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1330	if (ret < 0)
1331		goto fail_free_buffers;
1332
1333	mutex_init(&buffer->mutex);
1334
1335	return buffer;
1336
1337 fail_free_buffers:
1338	for_each_buffer_cpu(buffer, cpu) {
1339		if (buffer->buffers[cpu])
1340			rb_free_cpu_buffer(buffer->buffers[cpu]);
1341	}
1342	kfree(buffer->buffers);
1343
1344 fail_free_cpumask:
1345	free_cpumask_var(buffer->cpumask);
1346
1347 fail_free_buffer:
1348	kfree(buffer);
1349	return NULL;
1350}
1351EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1352
1353/**
1354 * ring_buffer_free - free a ring buffer.
1355 * @buffer: the buffer to free.
1356 */
1357void
1358ring_buffer_free(struct ring_buffer *buffer)
1359{
1360	int cpu;
1361
1362	cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1363
1364	for_each_buffer_cpu(buffer, cpu)
1365		rb_free_cpu_buffer(buffer->buffers[cpu]);
1366
1367	kfree(buffer->buffers);
1368	free_cpumask_var(buffer->cpumask);
1369
1370	kfree(buffer);
1371}
1372EXPORT_SYMBOL_GPL(ring_buffer_free);
1373
1374void ring_buffer_set_clock(struct ring_buffer *buffer,
1375			   u64 (*clock)(void))
1376{
1377	buffer->clock = clock;
1378}
1379
 
 
 
 
 
 
 
 
 
 
1380static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1381
1382static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1383{
1384	return local_read(&bpage->entries) & RB_WRITE_MASK;
1385}
1386
1387static inline unsigned long rb_page_write(struct buffer_page *bpage)
1388{
1389	return local_read(&bpage->write) & RB_WRITE_MASK;
1390}
1391
1392static int
1393rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1394{
1395	struct list_head *tail_page, *to_remove, *next_page;
1396	struct buffer_page *to_remove_page, *tmp_iter_page;
1397	struct buffer_page *last_page, *first_page;
1398	unsigned long nr_removed;
1399	unsigned long head_bit;
1400	int page_entries;
1401
1402	head_bit = 0;
1403
1404	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1405	atomic_inc(&cpu_buffer->record_disabled);
1406	/*
1407	 * We don't race with the readers since we have acquired the reader
1408	 * lock. We also don't race with writers after disabling recording.
1409	 * This makes it easy to figure out the first and the last page to be
1410	 * removed from the list. We unlink all the pages in between including
1411	 * the first and last pages. This is done in a busy loop so that we
1412	 * lose the least number of traces.
1413	 * The pages are freed after we restart recording and unlock readers.
1414	 */
1415	tail_page = &cpu_buffer->tail_page->list;
1416
1417	/*
1418	 * tail page might be on reader page, we remove the next page
1419	 * from the ring buffer
1420	 */
1421	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1422		tail_page = rb_list_head(tail_page->next);
1423	to_remove = tail_page;
1424
1425	/* start of pages to remove */
1426	first_page = list_entry(rb_list_head(to_remove->next),
1427				struct buffer_page, list);
1428
1429	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1430		to_remove = rb_list_head(to_remove)->next;
1431		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1432	}
1433
1434	next_page = rb_list_head(to_remove)->next;
1435
1436	/*
1437	 * Now we remove all pages between tail_page and next_page.
1438	 * Make sure that we have head_bit value preserved for the
1439	 * next page
1440	 */
1441	tail_page->next = (struct list_head *)((unsigned long)next_page |
1442						head_bit);
1443	next_page = rb_list_head(next_page);
1444	next_page->prev = tail_page;
1445
1446	/* make sure pages points to a valid page in the ring buffer */
1447	cpu_buffer->pages = next_page;
1448
1449	/* update head page */
1450	if (head_bit)
1451		cpu_buffer->head_page = list_entry(next_page,
1452						struct buffer_page, list);
1453
1454	/*
1455	 * change read pointer to make sure any read iterators reset
1456	 * themselves
1457	 */
1458	cpu_buffer->read = 0;
1459
1460	/* pages are removed, resume tracing and then free the pages */
1461	atomic_dec(&cpu_buffer->record_disabled);
1462	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1463
1464	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1465
1466	/* last buffer page to remove */
1467	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1468				list);
1469	tmp_iter_page = first_page;
1470
1471	do {
 
 
1472		to_remove_page = tmp_iter_page;
1473		rb_inc_page(cpu_buffer, &tmp_iter_page);
1474
1475		/* update the counters */
1476		page_entries = rb_page_entries(to_remove_page);
1477		if (page_entries) {
1478			/*
1479			 * If something was added to this page, it was full
1480			 * since it is not the tail page. So we deduct the
1481			 * bytes consumed in ring buffer from here.
1482			 * Increment overrun to account for the lost events.
1483			 */
1484			local_add(page_entries, &cpu_buffer->overrun);
1485			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1486		}
1487
1488		/*
1489		 * We have already removed references to this list item, just
1490		 * free up the buffer_page and its page
1491		 */
1492		free_buffer_page(to_remove_page);
1493		nr_removed--;
1494
1495	} while (to_remove_page != last_page);
1496
1497	RB_WARN_ON(cpu_buffer, nr_removed);
1498
1499	return nr_removed == 0;
1500}
1501
1502static int
1503rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1504{
1505	struct list_head *pages = &cpu_buffer->new_pages;
1506	int retries, success;
1507
1508	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1509	/*
1510	 * We are holding the reader lock, so the reader page won't be swapped
1511	 * in the ring buffer. Now we are racing with the writer trying to
1512	 * move head page and the tail page.
1513	 * We are going to adapt the reader page update process where:
1514	 * 1. We first splice the start and end of list of new pages between
1515	 *    the head page and its previous page.
1516	 * 2. We cmpxchg the prev_page->next to point from head page to the
1517	 *    start of new pages list.
1518	 * 3. Finally, we update the head->prev to the end of new list.
1519	 *
1520	 * We will try this process 10 times, to make sure that we don't keep
1521	 * spinning.
1522	 */
1523	retries = 10;
1524	success = 0;
1525	while (retries--) {
1526		struct list_head *head_page, *prev_page, *r;
1527		struct list_head *last_page, *first_page;
1528		struct list_head *head_page_with_bit;
1529
1530		head_page = &rb_set_head_page(cpu_buffer)->list;
1531		if (!head_page)
1532			break;
1533		prev_page = head_page->prev;
1534
1535		first_page = pages->next;
1536		last_page  = pages->prev;
1537
1538		head_page_with_bit = (struct list_head *)
1539				     ((unsigned long)head_page | RB_PAGE_HEAD);
1540
1541		last_page->next = head_page_with_bit;
1542		first_page->prev = prev_page;
1543
1544		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1545
1546		if (r == head_page_with_bit) {
1547			/*
1548			 * yay, we replaced the page pointer to our new list,
1549			 * now, we just have to update to head page's prev
1550			 * pointer to point to end of list
1551			 */
1552			head_page->prev = last_page;
1553			success = 1;
1554			break;
1555		}
1556	}
1557
1558	if (success)
1559		INIT_LIST_HEAD(pages);
1560	/*
1561	 * If we weren't successful in adding in new pages, warn and stop
1562	 * tracing
1563	 */
1564	RB_WARN_ON(cpu_buffer, !success);
1565	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1566
1567	/* free pages if they weren't inserted */
1568	if (!success) {
1569		struct buffer_page *bpage, *tmp;
1570		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1571					 list) {
1572			list_del_init(&bpage->list);
1573			free_buffer_page(bpage);
1574		}
1575	}
1576	return success;
1577}
1578
1579static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1580{
1581	int success;
1582
1583	if (cpu_buffer->nr_pages_to_update > 0)
1584		success = rb_insert_pages(cpu_buffer);
1585	else
1586		success = rb_remove_pages(cpu_buffer,
1587					-cpu_buffer->nr_pages_to_update);
1588
1589	if (success)
1590		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1591}
1592
1593static void update_pages_handler(struct work_struct *work)
1594{
1595	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1596			struct ring_buffer_per_cpu, update_pages_work);
1597	rb_update_pages(cpu_buffer);
1598	complete(&cpu_buffer->update_done);
1599}
1600
1601/**
1602 * ring_buffer_resize - resize the ring buffer
1603 * @buffer: the buffer to resize.
1604 * @size: the new size.
1605 * @cpu_id: the cpu buffer to resize
1606 *
1607 * Minimum size is 2 * BUF_PAGE_SIZE.
1608 *
1609 * Returns 0 on success and < 0 on failure.
1610 */
1611int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1612			int cpu_id)
1613{
1614	struct ring_buffer_per_cpu *cpu_buffer;
1615	unsigned long nr_pages;
1616	int cpu, err = 0;
1617
1618	/*
1619	 * Always succeed at resizing a non-existent buffer:
1620	 */
1621	if (!buffer)
1622		return size;
1623
1624	/* Make sure the requested buffer exists */
1625	if (cpu_id != RING_BUFFER_ALL_CPUS &&
1626	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
1627		return size;
1628
1629	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1630
1631	/* we need a minimum of two pages */
1632	if (nr_pages < 2)
1633		nr_pages = 2;
1634
1635	size = nr_pages * BUF_PAGE_SIZE;
1636
1637	/*
1638	 * Don't succeed if resizing is disabled, as a reader might be
1639	 * manipulating the ring buffer and is expecting a sane state while
1640	 * this is true.
1641	 */
1642	if (atomic_read(&buffer->resize_disabled))
1643		return -EBUSY;
1644
1645	/* prevent another thread from changing buffer sizes */
1646	mutex_lock(&buffer->mutex);
1647
 
1648	if (cpu_id == RING_BUFFER_ALL_CPUS) {
 
 
 
 
 
 
 
 
 
 
 
 
 
1649		/* calculate the pages to update */
1650		for_each_buffer_cpu(buffer, cpu) {
1651			cpu_buffer = buffer->buffers[cpu];
1652
1653			cpu_buffer->nr_pages_to_update = nr_pages -
1654							cpu_buffer->nr_pages;
1655			/*
1656			 * nothing more to do for removing pages or no update
1657			 */
1658			if (cpu_buffer->nr_pages_to_update <= 0)
1659				continue;
1660			/*
1661			 * to add pages, make sure all new pages can be
1662			 * allocated without receiving ENOMEM
1663			 */
1664			INIT_LIST_HEAD(&cpu_buffer->new_pages);
1665			if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1666						&cpu_buffer->new_pages, cpu)) {
1667				/* not enough memory for new pages */
1668				err = -ENOMEM;
1669				goto out_err;
1670			}
1671		}
1672
1673		get_online_cpus();
1674		/*
1675		 * Fire off all the required work handlers
1676		 * We can't schedule on offline CPUs, but it's not necessary
1677		 * since we can change their buffer sizes without any race.
1678		 */
1679		for_each_buffer_cpu(buffer, cpu) {
1680			cpu_buffer = buffer->buffers[cpu];
1681			if (!cpu_buffer->nr_pages_to_update)
1682				continue;
1683
1684			/* Can't run something on an offline CPU. */
1685			if (!cpu_online(cpu)) {
1686				rb_update_pages(cpu_buffer);
1687				cpu_buffer->nr_pages_to_update = 0;
1688			} else {
1689				schedule_work_on(cpu,
1690						&cpu_buffer->update_pages_work);
1691			}
1692		}
1693
1694		/* wait for all the updates to complete */
1695		for_each_buffer_cpu(buffer, cpu) {
1696			cpu_buffer = buffer->buffers[cpu];
1697			if (!cpu_buffer->nr_pages_to_update)
1698				continue;
1699
1700			if (cpu_online(cpu))
1701				wait_for_completion(&cpu_buffer->update_done);
1702			cpu_buffer->nr_pages_to_update = 0;
1703		}
1704
1705		put_online_cpus();
1706	} else {
1707		/* Make sure this CPU has been intitialized */
1708		if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1709			goto out;
1710
1711		cpu_buffer = buffer->buffers[cpu_id];
1712
1713		if (nr_pages == cpu_buffer->nr_pages)
1714			goto out;
1715
 
 
 
 
 
 
 
 
 
 
1716		cpu_buffer->nr_pages_to_update = nr_pages -
1717						cpu_buffer->nr_pages;
1718
1719		INIT_LIST_HEAD(&cpu_buffer->new_pages);
1720		if (cpu_buffer->nr_pages_to_update > 0 &&
1721			__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1722					    &cpu_buffer->new_pages, cpu_id)) {
1723			err = -ENOMEM;
1724			goto out_err;
1725		}
1726
1727		get_online_cpus();
1728
1729		/* Can't run something on an offline CPU. */
1730		if (!cpu_online(cpu_id))
1731			rb_update_pages(cpu_buffer);
1732		else {
1733			schedule_work_on(cpu_id,
1734					 &cpu_buffer->update_pages_work);
1735			wait_for_completion(&cpu_buffer->update_done);
1736		}
1737
1738		cpu_buffer->nr_pages_to_update = 0;
1739		put_online_cpus();
1740	}
1741
1742 out:
1743	/*
1744	 * The ring buffer resize can happen with the ring buffer
1745	 * enabled, so that the update disturbs the tracing as little
1746	 * as possible. But if the buffer is disabled, we do not need
1747	 * to worry about that, and we can take the time to verify
1748	 * that the buffer is not corrupt.
1749	 */
1750	if (atomic_read(&buffer->record_disabled)) {
1751		atomic_inc(&buffer->record_disabled);
1752		/*
1753		 * Even though the buffer was disabled, we must make sure
1754		 * that it is truly disabled before calling rb_check_pages.
1755		 * There could have been a race between checking
1756		 * record_disable and incrementing it.
1757		 */
1758		synchronize_sched();
1759		for_each_buffer_cpu(buffer, cpu) {
1760			cpu_buffer = buffer->buffers[cpu];
1761			rb_check_pages(cpu_buffer);
1762		}
1763		atomic_dec(&buffer->record_disabled);
1764	}
1765
1766	mutex_unlock(&buffer->mutex);
1767	return size;
1768
1769 out_err:
1770	for_each_buffer_cpu(buffer, cpu) {
1771		struct buffer_page *bpage, *tmp;
1772
1773		cpu_buffer = buffer->buffers[cpu];
1774		cpu_buffer->nr_pages_to_update = 0;
1775
1776		if (list_empty(&cpu_buffer->new_pages))
1777			continue;
1778
1779		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1780					list) {
1781			list_del_init(&bpage->list);
1782			free_buffer_page(bpage);
1783		}
1784	}
 
1785	mutex_unlock(&buffer->mutex);
1786	return err;
1787}
1788EXPORT_SYMBOL_GPL(ring_buffer_resize);
1789
1790void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1791{
1792	mutex_lock(&buffer->mutex);
1793	if (val)
1794		buffer->flags |= RB_FL_OVERWRITE;
1795	else
1796		buffer->flags &= ~RB_FL_OVERWRITE;
1797	mutex_unlock(&buffer->mutex);
1798}
1799EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1800
1801static __always_inline void *
1802__rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1803{
1804	return bpage->data + index;
1805}
1806
1807static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1808{
1809	return bpage->page->data + index;
1810}
1811
1812static __always_inline struct ring_buffer_event *
1813rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1814{
1815	return __rb_page_index(cpu_buffer->reader_page,
1816			       cpu_buffer->reader_page->read);
1817}
1818
1819static __always_inline struct ring_buffer_event *
1820rb_iter_head_event(struct ring_buffer_iter *iter)
1821{
1822	return __rb_page_index(iter->head_page, iter->head);
1823}
1824
1825static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
 
1826{
1827	return local_read(&bpage->page->commit);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1828}
1829
1830/* Size is determined by what has been committed */
1831static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1832{
1833	return rb_page_commit(bpage);
1834}
1835
1836static __always_inline unsigned
1837rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1838{
1839	return rb_page_commit(cpu_buffer->commit_page);
1840}
1841
1842static __always_inline unsigned
1843rb_event_index(struct ring_buffer_event *event)
1844{
1845	unsigned long addr = (unsigned long)event;
1846
1847	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1848}
1849
1850static void rb_inc_iter(struct ring_buffer_iter *iter)
1851{
1852	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1853
1854	/*
1855	 * The iterator could be on the reader page (it starts there).
1856	 * But the head could have moved, since the reader was
1857	 * found. Check for this case and assign the iterator
1858	 * to the head page instead of next.
1859	 */
1860	if (iter->head_page == cpu_buffer->reader_page)
1861		iter->head_page = rb_set_head_page(cpu_buffer);
1862	else
1863		rb_inc_page(cpu_buffer, &iter->head_page);
1864
1865	iter->read_stamp = iter->head_page->page->time_stamp;
1866	iter->head = 0;
 
1867}
1868
1869/*
1870 * rb_handle_head_page - writer hit the head page
1871 *
1872 * Returns: +1 to retry page
1873 *           0 to continue
1874 *          -1 on error
1875 */
1876static int
1877rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1878		    struct buffer_page *tail_page,
1879		    struct buffer_page *next_page)
1880{
1881	struct buffer_page *new_head;
1882	int entries;
1883	int type;
1884	int ret;
1885
1886	entries = rb_page_entries(next_page);
1887
1888	/*
1889	 * The hard part is here. We need to move the head
1890	 * forward, and protect against both readers on
1891	 * other CPUs and writers coming in via interrupts.
1892	 */
1893	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1894				       RB_PAGE_HEAD);
1895
1896	/*
1897	 * type can be one of four:
1898	 *  NORMAL - an interrupt already moved it for us
1899	 *  HEAD   - we are the first to get here.
1900	 *  UPDATE - we are the interrupt interrupting
1901	 *           a current move.
1902	 *  MOVED  - a reader on another CPU moved the next
1903	 *           pointer to its reader page. Give up
1904	 *           and try again.
1905	 */
1906
1907	switch (type) {
1908	case RB_PAGE_HEAD:
1909		/*
1910		 * We changed the head to UPDATE, thus
1911		 * it is our responsibility to update
1912		 * the counters.
1913		 */
1914		local_add(entries, &cpu_buffer->overrun);
1915		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1916
1917		/*
1918		 * The entries will be zeroed out when we move the
1919		 * tail page.
1920		 */
1921
1922		/* still more to do */
1923		break;
1924
1925	case RB_PAGE_UPDATE:
1926		/*
1927		 * This is an interrupt that interrupt the
1928		 * previous update. Still more to do.
1929		 */
1930		break;
1931	case RB_PAGE_NORMAL:
1932		/*
1933		 * An interrupt came in before the update
1934		 * and processed this for us.
1935		 * Nothing left to do.
1936		 */
1937		return 1;
1938	case RB_PAGE_MOVED:
1939		/*
1940		 * The reader is on another CPU and just did
1941		 * a swap with our next_page.
1942		 * Try again.
1943		 */
1944		return 1;
1945	default:
1946		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1947		return -1;
1948	}
1949
1950	/*
1951	 * Now that we are here, the old head pointer is
1952	 * set to UPDATE. This will keep the reader from
1953	 * swapping the head page with the reader page.
1954	 * The reader (on another CPU) will spin till
1955	 * we are finished.
1956	 *
1957	 * We just need to protect against interrupts
1958	 * doing the job. We will set the next pointer
1959	 * to HEAD. After that, we set the old pointer
1960	 * to NORMAL, but only if it was HEAD before.
1961	 * otherwise we are an interrupt, and only
1962	 * want the outer most commit to reset it.
1963	 */
1964	new_head = next_page;
1965	rb_inc_page(cpu_buffer, &new_head);
1966
1967	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1968				    RB_PAGE_NORMAL);
1969
1970	/*
1971	 * Valid returns are:
1972	 *  HEAD   - an interrupt came in and already set it.
1973	 *  NORMAL - One of two things:
1974	 *            1) We really set it.
1975	 *            2) A bunch of interrupts came in and moved
1976	 *               the page forward again.
1977	 */
1978	switch (ret) {
1979	case RB_PAGE_HEAD:
1980	case RB_PAGE_NORMAL:
1981		/* OK */
1982		break;
1983	default:
1984		RB_WARN_ON(cpu_buffer, 1);
1985		return -1;
1986	}
1987
1988	/*
1989	 * It is possible that an interrupt came in,
1990	 * set the head up, then more interrupts came in
1991	 * and moved it again. When we get back here,
1992	 * the page would have been set to NORMAL but we
1993	 * just set it back to HEAD.
1994	 *
1995	 * How do you detect this? Well, if that happened
1996	 * the tail page would have moved.
1997	 */
1998	if (ret == RB_PAGE_NORMAL) {
1999		struct buffer_page *buffer_tail_page;
2000
2001		buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2002		/*
2003		 * If the tail had moved passed next, then we need
2004		 * to reset the pointer.
2005		 */
2006		if (buffer_tail_page != tail_page &&
2007		    buffer_tail_page != next_page)
2008			rb_head_page_set_normal(cpu_buffer, new_head,
2009						next_page,
2010						RB_PAGE_HEAD);
2011	}
2012
2013	/*
2014	 * If this was the outer most commit (the one that
2015	 * changed the original pointer from HEAD to UPDATE),
2016	 * then it is up to us to reset it to NORMAL.
2017	 */
2018	if (type == RB_PAGE_HEAD) {
2019		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2020					      tail_page,
2021					      RB_PAGE_UPDATE);
2022		if (RB_WARN_ON(cpu_buffer,
2023			       ret != RB_PAGE_UPDATE))
2024			return -1;
2025	}
2026
2027	return 0;
2028}
2029
2030static inline void
2031rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2032	      unsigned long tail, struct rb_event_info *info)
2033{
2034	struct buffer_page *tail_page = info->tail_page;
2035	struct ring_buffer_event *event;
2036	unsigned long length = info->length;
2037
2038	/*
2039	 * Only the event that crossed the page boundary
2040	 * must fill the old tail_page with padding.
2041	 */
2042	if (tail >= BUF_PAGE_SIZE) {
2043		/*
2044		 * If the page was filled, then we still need
2045		 * to update the real_end. Reset it to zero
2046		 * and the reader will ignore it.
2047		 */
2048		if (tail == BUF_PAGE_SIZE)
2049			tail_page->real_end = 0;
2050
2051		local_sub(length, &tail_page->write);
2052		return;
2053	}
2054
2055	event = __rb_page_index(tail_page, tail);
2056	kmemcheck_annotate_bitfield(event, bitfield);
2057
2058	/* account for padding bytes */
2059	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2060
2061	/*
2062	 * Save the original length to the meta data.
2063	 * This will be used by the reader to add lost event
2064	 * counter.
2065	 */
2066	tail_page->real_end = tail;
2067
2068	/*
2069	 * If this event is bigger than the minimum size, then
2070	 * we need to be careful that we don't subtract the
2071	 * write counter enough to allow another writer to slip
2072	 * in on this page.
2073	 * We put in a discarded commit instead, to make sure
2074	 * that this space is not used again.
2075	 *
2076	 * If we are less than the minimum size, we don't need to
2077	 * worry about it.
2078	 */
2079	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2080		/* No room for any events */
2081
2082		/* Mark the rest of the page with padding */
2083		rb_event_set_padding(event);
2084
2085		/* Set the write back to the previous setting */
2086		local_sub(length, &tail_page->write);
2087		return;
2088	}
2089
2090	/* Put in a discarded event */
2091	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2092	event->type_len = RINGBUF_TYPE_PADDING;
2093	/* time delta must be non zero */
2094	event->time_delta = 1;
2095
2096	/* Set write to end of buffer */
2097	length = (tail + length) - BUF_PAGE_SIZE;
2098	local_sub(length, &tail_page->write);
2099}
2100
2101static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2102
2103/*
2104 * This is the slow path, force gcc not to inline it.
2105 */
2106static noinline struct ring_buffer_event *
2107rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2108	     unsigned long tail, struct rb_event_info *info)
2109{
2110	struct buffer_page *tail_page = info->tail_page;
2111	struct buffer_page *commit_page = cpu_buffer->commit_page;
2112	struct ring_buffer *buffer = cpu_buffer->buffer;
2113	struct buffer_page *next_page;
2114	int ret;
2115
2116	next_page = tail_page;
2117
2118	rb_inc_page(cpu_buffer, &next_page);
2119
2120	/*
2121	 * If for some reason, we had an interrupt storm that made
2122	 * it all the way around the buffer, bail, and warn
2123	 * about it.
2124	 */
2125	if (unlikely(next_page == commit_page)) {
2126		local_inc(&cpu_buffer->commit_overrun);
2127		goto out_reset;
2128	}
2129
2130	/*
2131	 * This is where the fun begins!
2132	 *
2133	 * We are fighting against races between a reader that
2134	 * could be on another CPU trying to swap its reader
2135	 * page with the buffer head.
2136	 *
2137	 * We are also fighting against interrupts coming in and
2138	 * moving the head or tail on us as well.
2139	 *
2140	 * If the next page is the head page then we have filled
2141	 * the buffer, unless the commit page is still on the
2142	 * reader page.
2143	 */
2144	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2145
2146		/*
2147		 * If the commit is not on the reader page, then
2148		 * move the header page.
2149		 */
2150		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2151			/*
2152			 * If we are not in overwrite mode,
2153			 * this is easy, just stop here.
2154			 */
2155			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2156				local_inc(&cpu_buffer->dropped_events);
2157				goto out_reset;
2158			}
2159
2160			ret = rb_handle_head_page(cpu_buffer,
2161						  tail_page,
2162						  next_page);
2163			if (ret < 0)
2164				goto out_reset;
2165			if (ret)
2166				goto out_again;
2167		} else {
2168			/*
2169			 * We need to be careful here too. The
2170			 * commit page could still be on the reader
2171			 * page. We could have a small buffer, and
2172			 * have filled up the buffer with events
2173			 * from interrupts and such, and wrapped.
2174			 *
2175			 * Note, if the tail page is also the on the
2176			 * reader_page, we let it move out.
2177			 */
2178			if (unlikely((cpu_buffer->commit_page !=
2179				      cpu_buffer->tail_page) &&
2180				     (cpu_buffer->commit_page ==
2181				      cpu_buffer->reader_page))) {
2182				local_inc(&cpu_buffer->commit_overrun);
2183				goto out_reset;
2184			}
2185		}
2186	}
2187
2188	rb_tail_page_update(cpu_buffer, tail_page, next_page);
2189
2190 out_again:
2191
2192	rb_reset_tail(cpu_buffer, tail, info);
2193
2194	/* Commit what we have for now. */
2195	rb_end_commit(cpu_buffer);
2196	/* rb_end_commit() decs committing */
2197	local_inc(&cpu_buffer->committing);
2198
2199	/* fail and let the caller try again */
2200	return ERR_PTR(-EAGAIN);
2201
2202 out_reset:
2203	/* reset write */
2204	rb_reset_tail(cpu_buffer, tail, info);
2205
2206	return NULL;
2207}
2208
2209/* Slow path, do not inline */
2210static noinline struct ring_buffer_event *
2211rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2212{
2213	event->type_len = RINGBUF_TYPE_TIME_EXTEND;
 
 
 
2214
2215	/* Not the first event on the page? */
2216	if (rb_event_index(event)) {
2217		event->time_delta = delta & TS_MASK;
2218		event->array[0] = delta >> TS_SHIFT;
2219	} else {
2220		/* nope, just zero it */
2221		event->time_delta = 0;
2222		event->array[0] = 0;
2223	}
2224
2225	return skip_time_extend(event);
2226}
2227
2228static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2229				     struct ring_buffer_event *event);
2230
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2231/**
2232 * rb_update_event - update event type and data
 
2233 * @event: the event to update
2234 * @type: the type of event
2235 * @length: the size of the event field in the ring buffer
2236 *
2237 * Update the type and data fields of the event. The length
2238 * is the actual size that is written to the ring buffer,
2239 * and with this, we can determine what to place into the
2240 * data field.
2241 */
2242static void
2243rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2244		struct ring_buffer_event *event,
2245		struct rb_event_info *info)
2246{
2247	unsigned length = info->length;
2248	u64 delta = info->delta;
2249
2250	/* Only a commit updates the timestamp */
2251	if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2252		delta = 0;
2253
2254	/*
2255	 * If we need to add a timestamp, then we
2256	 * add it to the start of the resevered space.
2257	 */
2258	if (unlikely(info->add_timestamp)) {
2259		event = rb_add_time_stamp(event, delta);
2260		length -= RB_LEN_TIME_EXTEND;
2261		delta = 0;
2262	}
2263
2264	event->time_delta = delta;
2265	length -= RB_EVNT_HDR_SIZE;
2266	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2267		event->type_len = 0;
2268		event->array[0] = length;
2269	} else
2270		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2271}
2272
2273static unsigned rb_calculate_event_length(unsigned length)
2274{
2275	struct ring_buffer_event event; /* Used only for sizeof array */
2276
2277	/* zero length can cause confusions */
2278	if (!length)
2279		length++;
2280
2281	if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2282		length += sizeof(event.array[0]);
2283
2284	length += RB_EVNT_HDR_SIZE;
2285	length = ALIGN(length, RB_ARCH_ALIGNMENT);
2286
2287	/*
2288	 * In case the time delta is larger than the 27 bits for it
2289	 * in the header, we need to add a timestamp. If another
2290	 * event comes in when trying to discard this one to increase
2291	 * the length, then the timestamp will be added in the allocated
2292	 * space of this event. If length is bigger than the size needed
2293	 * for the TIME_EXTEND, then padding has to be used. The events
2294	 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2295	 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2296	 * As length is a multiple of 4, we only need to worry if it
2297	 * is 12 (RB_LEN_TIME_EXTEND + 4).
2298	 */
2299	if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2300		length += RB_ALIGNMENT;
2301
2302	return length;
2303}
2304
2305#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2306static inline bool sched_clock_stable(void)
 
2307{
2308	return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2309}
2310#endif
2311
2312static inline int
2313rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2314		  struct ring_buffer_event *event)
2315{
2316	unsigned long new_index, old_index;
2317	struct buffer_page *bpage;
2318	unsigned long index;
2319	unsigned long addr;
 
 
2320
2321	new_index = rb_event_index(event);
2322	old_index = new_index + rb_event_ts_length(event);
2323	addr = (unsigned long)event;
2324	addr &= PAGE_MASK;
2325
2326	bpage = READ_ONCE(cpu_buffer->tail_page);
2327
 
 
 
 
 
 
 
 
2328	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2329		unsigned long write_mask =
2330			local_read(&bpage->write) & ~RB_WRITE_MASK;
2331		unsigned long event_length = rb_event_length(event);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2332		/*
2333		 * This is on the tail page. It is possible that
2334		 * a write could come in and move the tail page
2335		 * and write to the next page. That is fine
2336		 * because we just shorten what is on this page.
2337		 */
2338		old_index += write_mask;
2339		new_index += write_mask;
2340		index = local_cmpxchg(&bpage->write, old_index, new_index);
2341		if (index == old_index) {
2342			/* update counters */
2343			local_sub(event_length, &cpu_buffer->entries_bytes);
2344			return 1;
2345		}
2346	}
2347
2348	/* could not discard */
2349	return 0;
2350}
2351
2352static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2353{
2354	local_inc(&cpu_buffer->committing);
2355	local_inc(&cpu_buffer->commits);
2356}
2357
2358static __always_inline void
2359rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2360{
2361	unsigned long max_count;
2362
2363	/*
2364	 * We only race with interrupts and NMIs on this CPU.
2365	 * If we own the commit event, then we can commit
2366	 * all others that interrupted us, since the interruptions
2367	 * are in stack format (they finish before they come
2368	 * back to us). This allows us to do a simple loop to
2369	 * assign the commit to the tail.
2370	 */
2371 again:
2372	max_count = cpu_buffer->nr_pages * 100;
2373
2374	while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2375		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2376			return;
2377		if (RB_WARN_ON(cpu_buffer,
2378			       rb_is_reader_page(cpu_buffer->tail_page)))
2379			return;
2380		local_set(&cpu_buffer->commit_page->page->commit,
2381			  rb_page_write(cpu_buffer->commit_page));
2382		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2383		/* Only update the write stamp if the page has an event */
2384		if (rb_page_write(cpu_buffer->commit_page))
2385			cpu_buffer->write_stamp =
2386				cpu_buffer->commit_page->page->time_stamp;
2387		/* add barrier to keep gcc from optimizing too much */
2388		barrier();
2389	}
2390	while (rb_commit_index(cpu_buffer) !=
2391	       rb_page_write(cpu_buffer->commit_page)) {
2392
2393		local_set(&cpu_buffer->commit_page->page->commit,
2394			  rb_page_write(cpu_buffer->commit_page));
2395		RB_WARN_ON(cpu_buffer,
2396			   local_read(&cpu_buffer->commit_page->page->commit) &
2397			   ~RB_WRITE_MASK);
2398		barrier();
2399	}
2400
2401	/* again, keep gcc from optimizing */
2402	barrier();
2403
2404	/*
2405	 * If an interrupt came in just after the first while loop
2406	 * and pushed the tail page forward, we will be left with
2407	 * a dangling commit that will never go forward.
2408	 */
2409	if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2410		goto again;
2411}
2412
2413static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2414{
2415	unsigned long commits;
2416
2417	if (RB_WARN_ON(cpu_buffer,
2418		       !local_read(&cpu_buffer->committing)))
2419		return;
2420
2421 again:
2422	commits = local_read(&cpu_buffer->commits);
2423	/* synchronize with interrupts */
2424	barrier();
2425	if (local_read(&cpu_buffer->committing) == 1)
2426		rb_set_commit_to_write(cpu_buffer);
2427
2428	local_dec(&cpu_buffer->committing);
2429
2430	/* synchronize with interrupts */
2431	barrier();
2432
2433	/*
2434	 * Need to account for interrupts coming in between the
2435	 * updating of the commit page and the clearing of the
2436	 * committing counter.
2437	 */
2438	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2439	    !local_read(&cpu_buffer->committing)) {
2440		local_inc(&cpu_buffer->committing);
2441		goto again;
2442	}
2443}
2444
2445static inline void rb_event_discard(struct ring_buffer_event *event)
2446{
2447	if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2448		event = skip_time_extend(event);
2449
2450	/* array[0] holds the actual length for the discarded event */
2451	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2452	event->type_len = RINGBUF_TYPE_PADDING;
2453	/* time delta must be non zero */
2454	if (!event->time_delta)
2455		event->time_delta = 1;
2456}
2457
2458static __always_inline bool
2459rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2460		   struct ring_buffer_event *event)
2461{
2462	unsigned long addr = (unsigned long)event;
2463	unsigned long index;
2464
2465	index = rb_event_index(event);
2466	addr &= PAGE_MASK;
2467
2468	return cpu_buffer->commit_page->page == (void *)addr &&
2469		rb_commit_index(cpu_buffer) == index;
2470}
2471
2472static __always_inline void
2473rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2474		      struct ring_buffer_event *event)
2475{
2476	u64 delta;
2477
2478	/*
2479	 * The event first in the commit queue updates the
2480	 * time stamp.
2481	 */
2482	if (rb_event_is_commit(cpu_buffer, event)) {
2483		/*
2484		 * A commit event that is first on a page
2485		 * updates the write timestamp with the page stamp
2486		 */
2487		if (!rb_event_index(event))
2488			cpu_buffer->write_stamp =
2489				cpu_buffer->commit_page->page->time_stamp;
2490		else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2491			delta = event->array[0];
2492			delta <<= TS_SHIFT;
2493			delta += event->time_delta;
2494			cpu_buffer->write_stamp += delta;
2495		} else
2496			cpu_buffer->write_stamp += event->time_delta;
2497	}
2498}
2499
2500static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2501		      struct ring_buffer_event *event)
2502{
2503	local_inc(&cpu_buffer->entries);
2504	rb_update_write_stamp(cpu_buffer, event);
2505	rb_end_commit(cpu_buffer);
2506}
2507
2508static __always_inline void
2509rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2510{
2511	bool pagebusy;
 
 
2512
2513	if (buffer->irq_work.waiters_pending) {
2514		buffer->irq_work.waiters_pending = false;
2515		/* irq_work_queue() supplies it's own memory barriers */
2516		irq_work_queue(&buffer->irq_work.work);
2517	}
2518
2519	if (cpu_buffer->irq_work.waiters_pending) {
2520		cpu_buffer->irq_work.waiters_pending = false;
2521		/* irq_work_queue() supplies it's own memory barriers */
2522		irq_work_queue(&cpu_buffer->irq_work.work);
2523	}
2524
2525	pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 
2526
2527	if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2528		cpu_buffer->irq_work.wakeup_full = true;
2529		cpu_buffer->irq_work.full_waiters_pending = false;
2530		/* irq_work_queue() supplies it's own memory barriers */
2531		irq_work_queue(&cpu_buffer->irq_work.work);
2532	}
 
 
 
 
 
 
 
 
 
 
 
 
2533}
2534
2535/*
2536 * The lock and unlock are done within a preempt disable section.
2537 * The current_context per_cpu variable can only be modified
2538 * by the current task between lock and unlock. But it can
2539 * be modified more than once via an interrupt. To pass this
2540 * information from the lock to the unlock without having to
2541 * access the 'in_interrupt()' functions again (which do show
2542 * a bit of overhead in something as critical as function tracing,
2543 * we use a bitmask trick.
2544 *
2545 *  bit 0 =  NMI context
2546 *  bit 1 =  IRQ context
2547 *  bit 2 =  SoftIRQ context
2548 *  bit 3 =  normal context.
2549 *
2550 * This works because this is the order of contexts that can
2551 * preempt other contexts. A SoftIRQ never preempts an IRQ
2552 * context.
2553 *
2554 * When the context is determined, the corresponding bit is
2555 * checked and set (if it was set, then a recursion of that context
2556 * happened).
2557 *
2558 * On unlock, we need to clear this bit. To do so, just subtract
2559 * 1 from the current_context and AND it to itself.
2560 *
2561 * (binary)
2562 *  101 - 1 = 100
2563 *  101 & 100 = 100 (clearing bit zero)
2564 *
2565 *  1010 - 1 = 1001
2566 *  1010 & 1001 = 1000 (clearing bit 1)
2567 *
2568 * The least significant bit can be cleared this way, and it
2569 * just so happens that it is the same bit corresponding to
2570 * the current context.
2571 */
2572
2573static __always_inline int
2574trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2575{
2576	unsigned int val = cpu_buffer->current_context;
 
2577	int bit;
2578
2579	if (in_interrupt()) {
2580		if (in_nmi())
2581			bit = RB_CTX_NMI;
2582		else if (in_irq())
2583			bit = RB_CTX_IRQ;
2584		else
2585			bit = RB_CTX_SOFTIRQ;
2586	} else
2587		bit = RB_CTX_NORMAL;
 
 
 
2588
2589	if (unlikely(val & (1 << bit)))
2590		return 1;
2591
2592	val |= (1 << bit);
2593	cpu_buffer->current_context = val;
2594
2595	return 0;
2596}
2597
2598static __always_inline void
2599trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2600{
2601	cpu_buffer->current_context &= cpu_buffer->current_context - 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2602}
2603
2604/**
2605 * ring_buffer_unlock_commit - commit a reserved
2606 * @buffer: The buffer to commit to
2607 * @event: The event pointer to commit.
2608 *
2609 * This commits the data to the ring buffer, and releases any locks held.
2610 *
2611 * Must be paired with ring_buffer_lock_reserve.
2612 */
2613int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2614			      struct ring_buffer_event *event)
2615{
2616	struct ring_buffer_per_cpu *cpu_buffer;
2617	int cpu = raw_smp_processor_id();
2618
2619	cpu_buffer = buffer->buffers[cpu];
2620
2621	rb_commit(cpu_buffer, event);
2622
2623	rb_wakeups(buffer, cpu_buffer);
2624
2625	trace_recursive_unlock(cpu_buffer);
2626
2627	preempt_enable_notrace();
2628
2629	return 0;
2630}
2631EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2632
2633static noinline void
2634rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2635		    struct rb_event_info *info)
2636{
2637	WARN_ONCE(info->delta > (1ULL << 59),
2638		  KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2639		  (unsigned long long)info->delta,
2640		  (unsigned long long)info->ts,
2641		  (unsigned long long)cpu_buffer->write_stamp,
2642		  sched_clock_stable() ? "" :
2643		  "If you just came from a suspend/resume,\n"
2644		  "please switch to the trace global clock:\n"
2645		  "  echo global > /sys/kernel/debug/tracing/trace_clock\n");
2646	info->add_timestamp = 1;
2647}
2648
2649static struct ring_buffer_event *
2650__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2651		  struct rb_event_info *info)
2652{
2653	struct ring_buffer_event *event;
2654	struct buffer_page *tail_page;
2655	unsigned long tail, write;
2656
2657	/*
2658	 * If the time delta since the last event is too big to
2659	 * hold in the time field of the event, then we append a
2660	 * TIME EXTEND event ahead of the data event.
2661	 */
2662	if (unlikely(info->add_timestamp))
2663		info->length += RB_LEN_TIME_EXTEND;
2664
2665	/* Don't let the compiler play games with cpu_buffer->tail_page */
2666	tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2667	write = local_add_return(info->length, &tail_page->write);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2668
2669	/* set write to only the index of the write */
2670	write &= RB_WRITE_MASK;
 
2671	tail = write - info->length;
2672
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2673	/*
2674	 * If this is the first commit on the page, then it has the same
2675	 * timestamp as the page itself.
2676	 */
2677	if (!tail)
 
2678		info->delta = 0;
2679
2680	/* See if we shot pass the end of this buffer page */
2681	if (unlikely(write > BUF_PAGE_SIZE))
2682		return rb_move_tail(cpu_buffer, tail, info);
2683
2684	/* We reserved something on the buffer */
2685
2686	event = __rb_page_index(tail_page, tail);
2687	kmemcheck_annotate_bitfield(event, bitfield);
2688	rb_update_event(cpu_buffer, event, info);
2689
2690	local_inc(&tail_page->entries);
2691
2692	/*
2693	 * If this is the first commit on the page, then update
2694	 * its timestamp.
2695	 */
2696	if (!tail)
2697		tail_page->page->time_stamp = info->ts;
2698
2699	/* account for these added bytes */
2700	local_add(info->length, &cpu_buffer->entries_bytes);
2701
2702	return event;
2703}
2704
2705static __always_inline struct ring_buffer_event *
2706rb_reserve_next_event(struct ring_buffer *buffer,
2707		      struct ring_buffer_per_cpu *cpu_buffer,
2708		      unsigned long length)
2709{
2710	struct ring_buffer_event *event;
2711	struct rb_event_info info;
2712	int nr_loops = 0;
2713	u64 diff;
2714
2715	rb_start_commit(cpu_buffer);
 
2716
2717#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2718	/*
2719	 * Due to the ability to swap a cpu buffer from a buffer
2720	 * it is possible it was swapped before we committed.
2721	 * (committing stops a swap). We check for it here and
2722	 * if it happened, we have to fail the write.
2723	 */
2724	barrier();
2725	if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2726		local_dec(&cpu_buffer->committing);
2727		local_dec(&cpu_buffer->commits);
2728		return NULL;
2729	}
2730#endif
2731
2732	info.length = rb_calculate_event_length(length);
 
 
 
 
 
 
 
 
2733 again:
2734	info.add_timestamp = 0;
2735	info.delta = 0;
2736
2737	/*
2738	 * We allow for interrupts to reenter here and do a trace.
2739	 * If one does, it will cause this original code to loop
2740	 * back here. Even with heavy interrupts happening, this
2741	 * should only happen a few times in a row. If this happens
2742	 * 1000 times in a row, there must be either an interrupt
2743	 * storm or we have something buggy.
2744	 * Bail!
2745	 */
2746	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2747		goto out_fail;
2748
2749	info.ts = rb_time_stamp(cpu_buffer->buffer);
2750	diff = info.ts - cpu_buffer->write_stamp;
2751
2752	/* make sure this diff is calculated here */
2753	barrier();
2754
2755	/* Did the write stamp get updated already? */
2756	if (likely(info.ts >= cpu_buffer->write_stamp)) {
2757		info.delta = diff;
2758		if (unlikely(test_time_stamp(info.delta)))
2759			rb_handle_timestamp(cpu_buffer, &info);
2760	}
2761
2762	event = __rb_reserve_next(cpu_buffer, &info);
2763
2764	if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2765		if (info.add_timestamp)
2766			info.length -= RB_LEN_TIME_EXTEND;
2767		goto again;
2768	}
2769
2770	if (!event)
2771		goto out_fail;
2772
2773	return event;
2774
2775 out_fail:
2776	rb_end_commit(cpu_buffer);
2777	return NULL;
2778}
2779
2780/**
2781 * ring_buffer_lock_reserve - reserve a part of the buffer
2782 * @buffer: the ring buffer to reserve from
2783 * @length: the length of the data to reserve (excluding event header)
2784 *
2785 * Returns a reseverd event on the ring buffer to copy directly to.
2786 * The user of this interface will need to get the body to write into
2787 * and can use the ring_buffer_event_data() interface.
2788 *
2789 * The length is the length of the data needed, not the event length
2790 * which also includes the event header.
2791 *
2792 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2793 * If NULL is returned, then nothing has been allocated or locked.
2794 */
2795struct ring_buffer_event *
2796ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2797{
2798	struct ring_buffer_per_cpu *cpu_buffer;
2799	struct ring_buffer_event *event;
2800	int cpu;
2801
2802	/* If we are tracing schedule, we don't want to recurse */
2803	preempt_disable_notrace();
2804
2805	if (unlikely(atomic_read(&buffer->record_disabled)))
2806		goto out;
2807
2808	cpu = raw_smp_processor_id();
2809
2810	if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2811		goto out;
2812
2813	cpu_buffer = buffer->buffers[cpu];
2814
2815	if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2816		goto out;
2817
2818	if (unlikely(length > BUF_MAX_DATA_SIZE))
2819		goto out;
2820
2821	if (unlikely(trace_recursive_lock(cpu_buffer)))
2822		goto out;
2823
2824	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2825	if (!event)
2826		goto out_unlock;
2827
2828	return event;
2829
2830 out_unlock:
2831	trace_recursive_unlock(cpu_buffer);
2832 out:
2833	preempt_enable_notrace();
2834	return NULL;
2835}
2836EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2837
2838/*
2839 * Decrement the entries to the page that an event is on.
2840 * The event does not even need to exist, only the pointer
2841 * to the page it is on. This may only be called before the commit
2842 * takes place.
2843 */
2844static inline void
2845rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2846		   struct ring_buffer_event *event)
2847{
2848	unsigned long addr = (unsigned long)event;
2849	struct buffer_page *bpage = cpu_buffer->commit_page;
2850	struct buffer_page *start;
2851
2852	addr &= PAGE_MASK;
2853
2854	/* Do the likely case first */
2855	if (likely(bpage->page == (void *)addr)) {
2856		local_dec(&bpage->entries);
2857		return;
2858	}
2859
2860	/*
2861	 * Because the commit page may be on the reader page we
2862	 * start with the next page and check the end loop there.
2863	 */
2864	rb_inc_page(cpu_buffer, &bpage);
2865	start = bpage;
2866	do {
2867		if (bpage->page == (void *)addr) {
2868			local_dec(&bpage->entries);
2869			return;
2870		}
2871		rb_inc_page(cpu_buffer, &bpage);
2872	} while (bpage != start);
2873
2874	/* commit not part of this buffer?? */
2875	RB_WARN_ON(cpu_buffer, 1);
2876}
2877
2878/**
2879 * ring_buffer_commit_discard - discard an event that has not been committed
2880 * @buffer: the ring buffer
2881 * @event: non committed event to discard
2882 *
2883 * Sometimes an event that is in the ring buffer needs to be ignored.
2884 * This function lets the user discard an event in the ring buffer
2885 * and then that event will not be read later.
2886 *
2887 * This function only works if it is called before the the item has been
2888 * committed. It will try to free the event from the ring buffer
2889 * if another event has not been added behind it.
2890 *
2891 * If another event has been added behind it, it will set the event
2892 * up as discarded, and perform the commit.
2893 *
2894 * If this function is called, do not call ring_buffer_unlock_commit on
2895 * the event.
2896 */
2897void ring_buffer_discard_commit(struct ring_buffer *buffer,
2898				struct ring_buffer_event *event)
2899{
2900	struct ring_buffer_per_cpu *cpu_buffer;
2901	int cpu;
2902
2903	/* The event is discarded regardless */
2904	rb_event_discard(event);
2905
2906	cpu = smp_processor_id();
2907	cpu_buffer = buffer->buffers[cpu];
2908
2909	/*
2910	 * This must only be called if the event has not been
2911	 * committed yet. Thus we can assume that preemption
2912	 * is still disabled.
2913	 */
2914	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2915
2916	rb_decrement_entry(cpu_buffer, event);
2917	if (rb_try_to_discard(cpu_buffer, event))
2918		goto out;
2919
2920	/*
2921	 * The commit is still visible by the reader, so we
2922	 * must still update the timestamp.
2923	 */
2924	rb_update_write_stamp(cpu_buffer, event);
2925 out:
2926	rb_end_commit(cpu_buffer);
2927
2928	trace_recursive_unlock(cpu_buffer);
2929
2930	preempt_enable_notrace();
2931
2932}
2933EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2934
2935/**
2936 * ring_buffer_write - write data to the buffer without reserving
2937 * @buffer: The ring buffer to write to.
2938 * @length: The length of the data being written (excluding the event header)
2939 * @data: The data to write to the buffer.
2940 *
2941 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2942 * one function. If you already have the data to write to the buffer, it
2943 * may be easier to simply call this function.
2944 *
2945 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2946 * and not the length of the event which would hold the header.
2947 */
2948int ring_buffer_write(struct ring_buffer *buffer,
2949		      unsigned long length,
2950		      void *data)
2951{
2952	struct ring_buffer_per_cpu *cpu_buffer;
2953	struct ring_buffer_event *event;
2954	void *body;
2955	int ret = -EBUSY;
2956	int cpu;
2957
2958	preempt_disable_notrace();
2959
2960	if (atomic_read(&buffer->record_disabled))
2961		goto out;
2962
2963	cpu = raw_smp_processor_id();
2964
2965	if (!cpumask_test_cpu(cpu, buffer->cpumask))
2966		goto out;
2967
2968	cpu_buffer = buffer->buffers[cpu];
2969
2970	if (atomic_read(&cpu_buffer->record_disabled))
2971		goto out;
2972
2973	if (length > BUF_MAX_DATA_SIZE)
2974		goto out;
2975
2976	if (unlikely(trace_recursive_lock(cpu_buffer)))
2977		goto out;
2978
2979	event = rb_reserve_next_event(buffer, cpu_buffer, length);
2980	if (!event)
2981		goto out_unlock;
2982
2983	body = rb_event_data(event);
2984
2985	memcpy(body, data, length);
2986
2987	rb_commit(cpu_buffer, event);
2988
2989	rb_wakeups(buffer, cpu_buffer);
2990
2991	ret = 0;
2992
2993 out_unlock:
2994	trace_recursive_unlock(cpu_buffer);
2995
2996 out:
2997	preempt_enable_notrace();
2998
2999	return ret;
3000}
3001EXPORT_SYMBOL_GPL(ring_buffer_write);
3002
3003static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3004{
3005	struct buffer_page *reader = cpu_buffer->reader_page;
3006	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3007	struct buffer_page *commit = cpu_buffer->commit_page;
3008
3009	/* In case of error, head will be NULL */
3010	if (unlikely(!head))
3011		return true;
3012
3013	return reader->read == rb_page_commit(reader) &&
3014		(commit == reader ||
3015		 (commit == head &&
3016		  head->read == rb_page_commit(commit)));
3017}
3018
3019/**
3020 * ring_buffer_record_disable - stop all writes into the buffer
3021 * @buffer: The ring buffer to stop writes to.
3022 *
3023 * This prevents all writes to the buffer. Any attempt to write
3024 * to the buffer after this will fail and return NULL.
3025 *
3026 * The caller should call synchronize_sched() after this.
3027 */
3028void ring_buffer_record_disable(struct ring_buffer *buffer)
3029{
3030	atomic_inc(&buffer->record_disabled);
3031}
3032EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3033
3034/**
3035 * ring_buffer_record_enable - enable writes to the buffer
3036 * @buffer: The ring buffer to enable writes
3037 *
3038 * Note, multiple disables will need the same number of enables
3039 * to truly enable the writing (much like preempt_disable).
3040 */
3041void ring_buffer_record_enable(struct ring_buffer *buffer)
3042{
3043	atomic_dec(&buffer->record_disabled);
3044}
3045EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3046
3047/**
3048 * ring_buffer_record_off - stop all writes into the buffer
3049 * @buffer: The ring buffer to stop writes to.
3050 *
3051 * This prevents all writes to the buffer. Any attempt to write
3052 * to the buffer after this will fail and return NULL.
3053 *
3054 * This is different than ring_buffer_record_disable() as
3055 * it works like an on/off switch, where as the disable() version
3056 * must be paired with a enable().
3057 */
3058void ring_buffer_record_off(struct ring_buffer *buffer)
3059{
3060	unsigned int rd;
3061	unsigned int new_rd;
3062
3063	do {
3064		rd = atomic_read(&buffer->record_disabled);
3065		new_rd = rd | RB_BUFFER_OFF;
3066	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3067}
3068EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3069
3070/**
3071 * ring_buffer_record_on - restart writes into the buffer
3072 * @buffer: The ring buffer to start writes to.
3073 *
3074 * This enables all writes to the buffer that was disabled by
3075 * ring_buffer_record_off().
3076 *
3077 * This is different than ring_buffer_record_enable() as
3078 * it works like an on/off switch, where as the enable() version
3079 * must be paired with a disable().
3080 */
3081void ring_buffer_record_on(struct ring_buffer *buffer)
3082{
3083	unsigned int rd;
3084	unsigned int new_rd;
3085
3086	do {
3087		rd = atomic_read(&buffer->record_disabled);
3088		new_rd = rd & ~RB_BUFFER_OFF;
3089	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3090}
3091EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3092
3093/**
3094 * ring_buffer_record_is_on - return true if the ring buffer can write
3095 * @buffer: The ring buffer to see if write is enabled
3096 *
3097 * Returns true if the ring buffer is in a state that it accepts writes.
3098 */
3099int ring_buffer_record_is_on(struct ring_buffer *buffer)
3100{
3101	return !atomic_read(&buffer->record_disabled);
3102}
3103
3104/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3105 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3106 * @buffer: The ring buffer to stop writes to.
3107 * @cpu: The CPU buffer to stop
3108 *
3109 * This prevents all writes to the buffer. Any attempt to write
3110 * to the buffer after this will fail and return NULL.
3111 *
3112 * The caller should call synchronize_sched() after this.
3113 */
3114void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3115{
3116	struct ring_buffer_per_cpu *cpu_buffer;
3117
3118	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3119		return;
3120
3121	cpu_buffer = buffer->buffers[cpu];
3122	atomic_inc(&cpu_buffer->record_disabled);
3123}
3124EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3125
3126/**
3127 * ring_buffer_record_enable_cpu - enable writes to the buffer
3128 * @buffer: The ring buffer to enable writes
3129 * @cpu: The CPU to enable.
3130 *
3131 * Note, multiple disables will need the same number of enables
3132 * to truly enable the writing (much like preempt_disable).
3133 */
3134void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3135{
3136	struct ring_buffer_per_cpu *cpu_buffer;
3137
3138	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3139		return;
3140
3141	cpu_buffer = buffer->buffers[cpu];
3142	atomic_dec(&cpu_buffer->record_disabled);
3143}
3144EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3145
3146/*
3147 * The total entries in the ring buffer is the running counter
3148 * of entries entered into the ring buffer, minus the sum of
3149 * the entries read from the ring buffer and the number of
3150 * entries that were overwritten.
3151 */
3152static inline unsigned long
3153rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3154{
3155	return local_read(&cpu_buffer->entries) -
3156		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3157}
3158
3159/**
3160 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3161 * @buffer: The ring buffer
3162 * @cpu: The per CPU buffer to read from.
3163 */
3164u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3165{
3166	unsigned long flags;
3167	struct ring_buffer_per_cpu *cpu_buffer;
3168	struct buffer_page *bpage;
3169	u64 ret = 0;
3170
3171	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3172		return 0;
3173
3174	cpu_buffer = buffer->buffers[cpu];
3175	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3176	/*
3177	 * if the tail is on reader_page, oldest time stamp is on the reader
3178	 * page
3179	 */
3180	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3181		bpage = cpu_buffer->reader_page;
3182	else
3183		bpage = rb_set_head_page(cpu_buffer);
3184	if (bpage)
3185		ret = bpage->page->time_stamp;
3186	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3187
3188	return ret;
3189}
3190EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3191
3192/**
3193 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3194 * @buffer: The ring buffer
3195 * @cpu: The per CPU buffer to read from.
3196 */
3197unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3198{
3199	struct ring_buffer_per_cpu *cpu_buffer;
3200	unsigned long ret;
3201
3202	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3203		return 0;
3204
3205	cpu_buffer = buffer->buffers[cpu];
3206	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3207
3208	return ret;
3209}
3210EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3211
3212/**
3213 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3214 * @buffer: The ring buffer
3215 * @cpu: The per CPU buffer to get the entries from.
3216 */
3217unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3218{
3219	struct ring_buffer_per_cpu *cpu_buffer;
3220
3221	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3222		return 0;
3223
3224	cpu_buffer = buffer->buffers[cpu];
3225
3226	return rb_num_of_entries(cpu_buffer);
3227}
3228EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3229
3230/**
3231 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3232 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3233 * @buffer: The ring buffer
3234 * @cpu: The per CPU buffer to get the number of overruns from
3235 */
3236unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3237{
3238	struct ring_buffer_per_cpu *cpu_buffer;
3239	unsigned long ret;
3240
3241	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3242		return 0;
3243
3244	cpu_buffer = buffer->buffers[cpu];
3245	ret = local_read(&cpu_buffer->overrun);
3246
3247	return ret;
3248}
3249EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3250
3251/**
3252 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3253 * commits failing due to the buffer wrapping around while there are uncommitted
3254 * events, such as during an interrupt storm.
3255 * @buffer: The ring buffer
3256 * @cpu: The per CPU buffer to get the number of overruns from
3257 */
3258unsigned long
3259ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3260{
3261	struct ring_buffer_per_cpu *cpu_buffer;
3262	unsigned long ret;
3263
3264	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3265		return 0;
3266
3267	cpu_buffer = buffer->buffers[cpu];
3268	ret = local_read(&cpu_buffer->commit_overrun);
3269
3270	return ret;
3271}
3272EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3273
3274/**
3275 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3276 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3277 * @buffer: The ring buffer
3278 * @cpu: The per CPU buffer to get the number of overruns from
3279 */
3280unsigned long
3281ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3282{
3283	struct ring_buffer_per_cpu *cpu_buffer;
3284	unsigned long ret;
3285
3286	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3287		return 0;
3288
3289	cpu_buffer = buffer->buffers[cpu];
3290	ret = local_read(&cpu_buffer->dropped_events);
3291
3292	return ret;
3293}
3294EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3295
3296/**
3297 * ring_buffer_read_events_cpu - get the number of events successfully read
3298 * @buffer: The ring buffer
3299 * @cpu: The per CPU buffer to get the number of events read
3300 */
3301unsigned long
3302ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3303{
3304	struct ring_buffer_per_cpu *cpu_buffer;
3305
3306	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3307		return 0;
3308
3309	cpu_buffer = buffer->buffers[cpu];
3310	return cpu_buffer->read;
3311}
3312EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3313
3314/**
3315 * ring_buffer_entries - get the number of entries in a buffer
3316 * @buffer: The ring buffer
3317 *
3318 * Returns the total number of entries in the ring buffer
3319 * (all CPU entries)
3320 */
3321unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3322{
3323	struct ring_buffer_per_cpu *cpu_buffer;
3324	unsigned long entries = 0;
3325	int cpu;
3326
3327	/* if you care about this being correct, lock the buffer */
3328	for_each_buffer_cpu(buffer, cpu) {
3329		cpu_buffer = buffer->buffers[cpu];
3330		entries += rb_num_of_entries(cpu_buffer);
3331	}
3332
3333	return entries;
3334}
3335EXPORT_SYMBOL_GPL(ring_buffer_entries);
3336
3337/**
3338 * ring_buffer_overruns - get the number of overruns in buffer
3339 * @buffer: The ring buffer
3340 *
3341 * Returns the total number of overruns in the ring buffer
3342 * (all CPU entries)
3343 */
3344unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3345{
3346	struct ring_buffer_per_cpu *cpu_buffer;
3347	unsigned long overruns = 0;
3348	int cpu;
3349
3350	/* if you care about this being correct, lock the buffer */
3351	for_each_buffer_cpu(buffer, cpu) {
3352		cpu_buffer = buffer->buffers[cpu];
3353		overruns += local_read(&cpu_buffer->overrun);
3354	}
3355
3356	return overruns;
3357}
3358EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3359
3360static void rb_iter_reset(struct ring_buffer_iter *iter)
3361{
3362	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3363
3364	/* Iterator usage is expected to have record disabled */
3365	iter->head_page = cpu_buffer->reader_page;
3366	iter->head = cpu_buffer->reader_page->read;
 
3367
3368	iter->cache_reader_page = iter->head_page;
3369	iter->cache_read = cpu_buffer->read;
3370
3371	if (iter->head)
3372		iter->read_stamp = cpu_buffer->read_stamp;
3373	else
 
3374		iter->read_stamp = iter->head_page->page->time_stamp;
 
 
3375}
3376
3377/**
3378 * ring_buffer_iter_reset - reset an iterator
3379 * @iter: The iterator to reset
3380 *
3381 * Resets the iterator, so that it will start from the beginning
3382 * again.
3383 */
3384void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3385{
3386	struct ring_buffer_per_cpu *cpu_buffer;
3387	unsigned long flags;
3388
3389	if (!iter)
3390		return;
3391
3392	cpu_buffer = iter->cpu_buffer;
3393
3394	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3395	rb_iter_reset(iter);
3396	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3397}
3398EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3399
3400/**
3401 * ring_buffer_iter_empty - check if an iterator has no more to read
3402 * @iter: The iterator to check
3403 */
3404int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3405{
3406	struct ring_buffer_per_cpu *cpu_buffer;
 
 
 
 
 
 
 
3407
3408	cpu_buffer = iter->cpu_buffer;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3409
3410	return iter->head_page == cpu_buffer->commit_page &&
3411		iter->head == rb_commit_index(cpu_buffer);
 
 
 
3412}
3413EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3414
3415static void
3416rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3417		     struct ring_buffer_event *event)
3418{
3419	u64 delta;
3420
3421	switch (event->type_len) {
3422	case RINGBUF_TYPE_PADDING:
3423		return;
3424
3425	case RINGBUF_TYPE_TIME_EXTEND:
3426		delta = event->array[0];
3427		delta <<= TS_SHIFT;
3428		delta += event->time_delta;
3429		cpu_buffer->read_stamp += delta;
3430		return;
3431
3432	case RINGBUF_TYPE_TIME_STAMP:
3433		/* FIXME: not implemented */
 
3434		return;
3435
3436	case RINGBUF_TYPE_DATA:
3437		cpu_buffer->read_stamp += event->time_delta;
3438		return;
3439
3440	default:
3441		BUG();
3442	}
3443	return;
3444}
3445
3446static void
3447rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3448			  struct ring_buffer_event *event)
3449{
3450	u64 delta;
3451
3452	switch (event->type_len) {
3453	case RINGBUF_TYPE_PADDING:
3454		return;
3455
3456	case RINGBUF_TYPE_TIME_EXTEND:
3457		delta = event->array[0];
3458		delta <<= TS_SHIFT;
3459		delta += event->time_delta;
3460		iter->read_stamp += delta;
3461		return;
3462
3463	case RINGBUF_TYPE_TIME_STAMP:
3464		/* FIXME: not implemented */
 
3465		return;
3466
3467	case RINGBUF_TYPE_DATA:
3468		iter->read_stamp += event->time_delta;
3469		return;
3470
3471	default:
3472		BUG();
3473	}
3474	return;
3475}
3476
3477static struct buffer_page *
3478rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3479{
3480	struct buffer_page *reader = NULL;
3481	unsigned long overwrite;
3482	unsigned long flags;
3483	int nr_loops = 0;
3484	int ret;
3485
3486	local_irq_save(flags);
3487	arch_spin_lock(&cpu_buffer->lock);
3488
3489 again:
3490	/*
3491	 * This should normally only loop twice. But because the
3492	 * start of the reader inserts an empty page, it causes
3493	 * a case where we will loop three times. There should be no
3494	 * reason to loop four times (that I know of).
3495	 */
3496	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3497		reader = NULL;
3498		goto out;
3499	}
3500
3501	reader = cpu_buffer->reader_page;
3502
3503	/* If there's more to read, return this page */
3504	if (cpu_buffer->reader_page->read < rb_page_size(reader))
3505		goto out;
3506
3507	/* Never should we have an index greater than the size */
3508	if (RB_WARN_ON(cpu_buffer,
3509		       cpu_buffer->reader_page->read > rb_page_size(reader)))
3510		goto out;
3511
3512	/* check if we caught up to the tail */
3513	reader = NULL;
3514	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3515		goto out;
3516
3517	/* Don't bother swapping if the ring buffer is empty */
3518	if (rb_num_of_entries(cpu_buffer) == 0)
3519		goto out;
3520
3521	/*
3522	 * Reset the reader page to size zero.
3523	 */
3524	local_set(&cpu_buffer->reader_page->write, 0);
3525	local_set(&cpu_buffer->reader_page->entries, 0);
3526	local_set(&cpu_buffer->reader_page->page->commit, 0);
3527	cpu_buffer->reader_page->real_end = 0;
3528
3529 spin:
3530	/*
3531	 * Splice the empty reader page into the list around the head.
3532	 */
3533	reader = rb_set_head_page(cpu_buffer);
3534	if (!reader)
3535		goto out;
3536	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3537	cpu_buffer->reader_page->list.prev = reader->list.prev;
3538
3539	/*
3540	 * cpu_buffer->pages just needs to point to the buffer, it
3541	 *  has no specific buffer page to point to. Lets move it out
3542	 *  of our way so we don't accidentally swap it.
3543	 */
3544	cpu_buffer->pages = reader->list.prev;
3545
3546	/* The reader page will be pointing to the new head */
3547	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3548
3549	/*
3550	 * We want to make sure we read the overruns after we set up our
3551	 * pointers to the next object. The writer side does a
3552	 * cmpxchg to cross pages which acts as the mb on the writer
3553	 * side. Note, the reader will constantly fail the swap
3554	 * while the writer is updating the pointers, so this
3555	 * guarantees that the overwrite recorded here is the one we
3556	 * want to compare with the last_overrun.
3557	 */
3558	smp_mb();
3559	overwrite = local_read(&(cpu_buffer->overrun));
3560
3561	/*
3562	 * Here's the tricky part.
3563	 *
3564	 * We need to move the pointer past the header page.
3565	 * But we can only do that if a writer is not currently
3566	 * moving it. The page before the header page has the
3567	 * flag bit '1' set if it is pointing to the page we want.
3568	 * but if the writer is in the process of moving it
3569	 * than it will be '2' or already moved '0'.
3570	 */
3571
3572	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3573
3574	/*
3575	 * If we did not convert it, then we must try again.
3576	 */
3577	if (!ret)
3578		goto spin;
3579
3580	/*
3581	 * Yeah! We succeeded in replacing the page.
3582	 *
3583	 * Now make the new head point back to the reader page.
3584	 */
3585	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3586	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3587
 
 
3588	/* Finally update the reader page to the new head */
3589	cpu_buffer->reader_page = reader;
3590	cpu_buffer->reader_page->read = 0;
3591
3592	if (overwrite != cpu_buffer->last_overrun) {
3593		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3594		cpu_buffer->last_overrun = overwrite;
3595	}
3596
3597	goto again;
3598
3599 out:
3600	/* Update the read_stamp on the first event */
3601	if (reader && reader->read == 0)
3602		cpu_buffer->read_stamp = reader->page->time_stamp;
3603
3604	arch_spin_unlock(&cpu_buffer->lock);
3605	local_irq_restore(flags);
3606
3607	return reader;
3608}
3609
3610static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3611{
3612	struct ring_buffer_event *event;
3613	struct buffer_page *reader;
3614	unsigned length;
3615
3616	reader = rb_get_reader_page(cpu_buffer);
3617
3618	/* This function should not be called when buffer is empty */
3619	if (RB_WARN_ON(cpu_buffer, !reader))
3620		return;
3621
3622	event = rb_reader_event(cpu_buffer);
3623
3624	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3625		cpu_buffer->read++;
3626
3627	rb_update_read_stamp(cpu_buffer, event);
3628
3629	length = rb_event_length(event);
3630	cpu_buffer->reader_page->read += length;
3631}
3632
3633static void rb_advance_iter(struct ring_buffer_iter *iter)
3634{
3635	struct ring_buffer_per_cpu *cpu_buffer;
3636	struct ring_buffer_event *event;
3637	unsigned length;
3638
3639	cpu_buffer = iter->cpu_buffer;
3640
 
 
 
 
 
 
 
 
 
3641	/*
3642	 * Check if we are at the end of the buffer.
3643	 */
3644	if (iter->head >= rb_page_size(iter->head_page)) {
3645		/* discarded commits can make the page empty */
3646		if (iter->head_page == cpu_buffer->commit_page)
3647			return;
3648		rb_inc_iter(iter);
3649		return;
3650	}
3651
3652	event = rb_iter_head_event(iter);
3653
3654	length = rb_event_length(event);
3655
3656	/*
3657	 * This should not be called to advance the header if we are
3658	 * at the tail of the buffer.
3659	 */
3660	if (RB_WARN_ON(cpu_buffer,
3661		       (iter->head_page == cpu_buffer->commit_page) &&
3662		       (iter->head + length > rb_commit_index(cpu_buffer))))
3663		return;
3664
3665	rb_update_iter_read_stamp(iter, event);
3666
3667	iter->head += length;
3668
3669	/* check for end of page padding */
3670	if ((iter->head >= rb_page_size(iter->head_page)) &&
3671	    (iter->head_page != cpu_buffer->commit_page))
3672		rb_inc_iter(iter);
3673}
3674
3675static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3676{
3677	return cpu_buffer->lost_events;
3678}
3679
3680static struct ring_buffer_event *
3681rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3682	       unsigned long *lost_events)
3683{
3684	struct ring_buffer_event *event;
3685	struct buffer_page *reader;
3686	int nr_loops = 0;
3687
 
 
3688 again:
3689	/*
3690	 * We repeat when a time extend is encountered.
3691	 * Since the time extend is always attached to a data event,
3692	 * we should never loop more than once.
3693	 * (We never hit the following condition more than twice).
3694	 */
3695	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3696		return NULL;
3697
3698	reader = rb_get_reader_page(cpu_buffer);
3699	if (!reader)
3700		return NULL;
3701
3702	event = rb_reader_event(cpu_buffer);
3703
3704	switch (event->type_len) {
3705	case RINGBUF_TYPE_PADDING:
3706		if (rb_null_event(event))
3707			RB_WARN_ON(cpu_buffer, 1);
3708		/*
3709		 * Because the writer could be discarding every
3710		 * event it creates (which would probably be bad)
3711		 * if we were to go back to "again" then we may never
3712		 * catch up, and will trigger the warn on, or lock
3713		 * the box. Return the padding, and we will release
3714		 * the current locks, and try again.
3715		 */
3716		return event;
3717
3718	case RINGBUF_TYPE_TIME_EXTEND:
3719		/* Internal data, OK to advance */
3720		rb_advance_reader(cpu_buffer);
3721		goto again;
3722
3723	case RINGBUF_TYPE_TIME_STAMP:
3724		/* FIXME: not implemented */
 
 
 
 
 
3725		rb_advance_reader(cpu_buffer);
3726		goto again;
3727
3728	case RINGBUF_TYPE_DATA:
3729		if (ts) {
3730			*ts = cpu_buffer->read_stamp + event->time_delta;
3731			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3732							 cpu_buffer->cpu, ts);
3733		}
3734		if (lost_events)
3735			*lost_events = rb_lost_events(cpu_buffer);
3736		return event;
3737
3738	default:
3739		BUG();
3740	}
3741
3742	return NULL;
3743}
3744EXPORT_SYMBOL_GPL(ring_buffer_peek);
3745
3746static struct ring_buffer_event *
3747rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3748{
3749	struct ring_buffer *buffer;
3750	struct ring_buffer_per_cpu *cpu_buffer;
3751	struct ring_buffer_event *event;
3752	int nr_loops = 0;
3753
 
 
 
3754	cpu_buffer = iter->cpu_buffer;
3755	buffer = cpu_buffer->buffer;
3756
3757	/*
3758	 * Check if someone performed a consuming read to
3759	 * the buffer. A consuming read invalidates the iterator
3760	 * and we need to reset the iterator in this case.
3761	 */
3762	if (unlikely(iter->cache_read != cpu_buffer->read ||
3763		     iter->cache_reader_page != cpu_buffer->reader_page))
3764		rb_iter_reset(iter);
3765
3766 again:
3767	if (ring_buffer_iter_empty(iter))
3768		return NULL;
3769
3770	/*
3771	 * We repeat when a time extend is encountered or we hit
3772	 * the end of the page. Since the time extend is always attached
3773	 * to a data event, we should never loop more than three times.
3774	 * Once for going to next page, once on time extend, and
3775	 * finally once to get the event.
3776	 * (We never hit the following condition more than thrice).
3777	 */
3778	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3779		return NULL;
3780
3781	if (rb_per_cpu_empty(cpu_buffer))
3782		return NULL;
3783
3784	if (iter->head >= rb_page_size(iter->head_page)) {
3785		rb_inc_iter(iter);
3786		goto again;
3787	}
3788
3789	event = rb_iter_head_event(iter);
 
 
3790
3791	switch (event->type_len) {
3792	case RINGBUF_TYPE_PADDING:
3793		if (rb_null_event(event)) {
3794			rb_inc_iter(iter);
3795			goto again;
3796		}
3797		rb_advance_iter(iter);
3798		return event;
3799
3800	case RINGBUF_TYPE_TIME_EXTEND:
3801		/* Internal data, OK to advance */
3802		rb_advance_iter(iter);
3803		goto again;
3804
3805	case RINGBUF_TYPE_TIME_STAMP:
3806		/* FIXME: not implemented */
 
 
 
 
 
3807		rb_advance_iter(iter);
3808		goto again;
3809
3810	case RINGBUF_TYPE_DATA:
3811		if (ts) {
3812			*ts = iter->read_stamp + event->time_delta;
3813			ring_buffer_normalize_time_stamp(buffer,
3814							 cpu_buffer->cpu, ts);
3815		}
3816		return event;
3817
3818	default:
3819		BUG();
3820	}
3821
3822	return NULL;
3823}
3824EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3825
3826static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3827{
3828	if (likely(!in_nmi())) {
3829		raw_spin_lock(&cpu_buffer->reader_lock);
3830		return true;
3831	}
3832
3833	/*
3834	 * If an NMI die dumps out the content of the ring buffer
3835	 * trylock must be used to prevent a deadlock if the NMI
3836	 * preempted a task that holds the ring buffer locks. If
3837	 * we get the lock then all is fine, if not, then continue
3838	 * to do the read, but this can corrupt the ring buffer,
3839	 * so it must be permanently disabled from future writes.
3840	 * Reading from NMI is a oneshot deal.
3841	 */
3842	if (raw_spin_trylock(&cpu_buffer->reader_lock))
3843		return true;
3844
3845	/* Continue without locking, but disable the ring buffer */
3846	atomic_inc(&cpu_buffer->record_disabled);
3847	return false;
3848}
3849
3850static inline void
3851rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3852{
3853	if (likely(locked))
3854		raw_spin_unlock(&cpu_buffer->reader_lock);
3855	return;
3856}
3857
3858/**
3859 * ring_buffer_peek - peek at the next event to be read
3860 * @buffer: The ring buffer to read
3861 * @cpu: The cpu to peak at
3862 * @ts: The timestamp counter of this event.
3863 * @lost_events: a variable to store if events were lost (may be NULL)
3864 *
3865 * This will return the event that will be read next, but does
3866 * not consume the data.
3867 */
3868struct ring_buffer_event *
3869ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3870		 unsigned long *lost_events)
3871{
3872	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3873	struct ring_buffer_event *event;
3874	unsigned long flags;
3875	bool dolock;
3876
3877	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3878		return NULL;
3879
3880 again:
3881	local_irq_save(flags);
3882	dolock = rb_reader_lock(cpu_buffer);
3883	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3884	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3885		rb_advance_reader(cpu_buffer);
3886	rb_reader_unlock(cpu_buffer, dolock);
3887	local_irq_restore(flags);
3888
3889	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3890		goto again;
3891
3892	return event;
3893}
3894
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3895/**
3896 * ring_buffer_iter_peek - peek at the next event to be read
3897 * @iter: The ring buffer iterator
3898 * @ts: The timestamp counter of this event.
3899 *
3900 * This will return the event that will be read next, but does
3901 * not increment the iterator.
3902 */
3903struct ring_buffer_event *
3904ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3905{
3906	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3907	struct ring_buffer_event *event;
3908	unsigned long flags;
3909
3910 again:
3911	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3912	event = rb_iter_peek(iter, ts);
3913	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3914
3915	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3916		goto again;
3917
3918	return event;
3919}
3920
3921/**
3922 * ring_buffer_consume - return an event and consume it
3923 * @buffer: The ring buffer to get the next event from
3924 * @cpu: the cpu to read the buffer from
3925 * @ts: a variable to store the timestamp (may be NULL)
3926 * @lost_events: a variable to store if events were lost (may be NULL)
3927 *
3928 * Returns the next event in the ring buffer, and that event is consumed.
3929 * Meaning, that sequential reads will keep returning a different event,
3930 * and eventually empty the ring buffer if the producer is slower.
3931 */
3932struct ring_buffer_event *
3933ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3934		    unsigned long *lost_events)
3935{
3936	struct ring_buffer_per_cpu *cpu_buffer;
3937	struct ring_buffer_event *event = NULL;
3938	unsigned long flags;
3939	bool dolock;
3940
3941 again:
3942	/* might be called in atomic */
3943	preempt_disable();
3944
3945	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3946		goto out;
3947
3948	cpu_buffer = buffer->buffers[cpu];
3949	local_irq_save(flags);
3950	dolock = rb_reader_lock(cpu_buffer);
3951
3952	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3953	if (event) {
3954		cpu_buffer->lost_events = 0;
3955		rb_advance_reader(cpu_buffer);
3956	}
3957
3958	rb_reader_unlock(cpu_buffer, dolock);
3959	local_irq_restore(flags);
3960
3961 out:
3962	preempt_enable();
3963
3964	if (event && event->type_len == RINGBUF_TYPE_PADDING)
3965		goto again;
3966
3967	return event;
3968}
3969EXPORT_SYMBOL_GPL(ring_buffer_consume);
3970
3971/**
3972 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3973 * @buffer: The ring buffer to read from
3974 * @cpu: The cpu buffer to iterate over
 
3975 *
3976 * This performs the initial preparations necessary to iterate
3977 * through the buffer.  Memory is allocated, buffer recording
3978 * is disabled, and the iterator pointer is returned to the caller.
3979 *
3980 * Disabling buffer recordng prevents the reading from being
3981 * corrupted. This is not a consuming read, so a producer is not
3982 * expected.
3983 *
3984 * After a sequence of ring_buffer_read_prepare calls, the user is
3985 * expected to make at least one call to ring_buffer_read_prepare_sync.
3986 * Afterwards, ring_buffer_read_start is invoked to get things going
3987 * for real.
3988 *
3989 * This overall must be paired with ring_buffer_read_finish.
3990 */
3991struct ring_buffer_iter *
3992ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3993{
3994	struct ring_buffer_per_cpu *cpu_buffer;
3995	struct ring_buffer_iter *iter;
3996
3997	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3998		return NULL;
3999
4000	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4001	if (!iter)
4002		return NULL;
4003
 
 
 
 
 
 
4004	cpu_buffer = buffer->buffers[cpu];
4005
4006	iter->cpu_buffer = cpu_buffer;
4007
4008	atomic_inc(&buffer->resize_disabled);
4009	atomic_inc(&cpu_buffer->record_disabled);
4010
4011	return iter;
4012}
4013EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4014
4015/**
4016 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4017 *
4018 * All previously invoked ring_buffer_read_prepare calls to prepare
4019 * iterators will be synchronized.  Afterwards, read_buffer_read_start
4020 * calls on those iterators are allowed.
4021 */
4022void
4023ring_buffer_read_prepare_sync(void)
4024{
4025	synchronize_sched();
4026}
4027EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4028
4029/**
4030 * ring_buffer_read_start - start a non consuming read of the buffer
4031 * @iter: The iterator returned by ring_buffer_read_prepare
4032 *
4033 * This finalizes the startup of an iteration through the buffer.
4034 * The iterator comes from a call to ring_buffer_read_prepare and
4035 * an intervening ring_buffer_read_prepare_sync must have been
4036 * performed.
4037 *
4038 * Must be paired with ring_buffer_read_finish.
4039 */
4040void
4041ring_buffer_read_start(struct ring_buffer_iter *iter)
4042{
4043	struct ring_buffer_per_cpu *cpu_buffer;
4044	unsigned long flags;
4045
4046	if (!iter)
4047		return;
4048
4049	cpu_buffer = iter->cpu_buffer;
4050
4051	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4052	arch_spin_lock(&cpu_buffer->lock);
4053	rb_iter_reset(iter);
4054	arch_spin_unlock(&cpu_buffer->lock);
4055	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4056}
4057EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4058
4059/**
4060 * ring_buffer_read_finish - finish reading the iterator of the buffer
4061 * @iter: The iterator retrieved by ring_buffer_start
4062 *
4063 * This re-enables the recording to the buffer, and frees the
4064 * iterator.
4065 */
4066void
4067ring_buffer_read_finish(struct ring_buffer_iter *iter)
4068{
4069	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4070	unsigned long flags;
4071
4072	/*
4073	 * Ring buffer is disabled from recording, here's a good place
4074	 * to check the integrity of the ring buffer.
4075	 * Must prevent readers from trying to read, as the check
4076	 * clears the HEAD page and readers require it.
4077	 */
4078	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4079	rb_check_pages(cpu_buffer);
4080	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4081
4082	atomic_dec(&cpu_buffer->record_disabled);
4083	atomic_dec(&cpu_buffer->buffer->resize_disabled);
4084	kfree(iter);
4085}
4086EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4087
4088/**
4089 * ring_buffer_read - read the next item in the ring buffer by the iterator
4090 * @iter: The ring buffer iterator
4091 * @ts: The time stamp of the event read.
4092 *
4093 * This reads the next event in the ring buffer and increments the iterator.
 
4094 */
4095struct ring_buffer_event *
4096ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4097{
4098	struct ring_buffer_event *event;
4099	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4100	unsigned long flags;
4101
4102	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4103 again:
4104	event = rb_iter_peek(iter, ts);
4105	if (!event)
4106		goto out;
4107
4108	if (event->type_len == RINGBUF_TYPE_PADDING)
4109		goto again;
4110
4111	rb_advance_iter(iter);
4112 out:
4113	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4114
4115	return event;
4116}
4117EXPORT_SYMBOL_GPL(ring_buffer_read);
4118
4119/**
4120 * ring_buffer_size - return the size of the ring buffer (in bytes)
4121 * @buffer: The ring buffer.
 
4122 */
4123unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4124{
4125	/*
4126	 * Earlier, this method returned
4127	 *	BUF_PAGE_SIZE * buffer->nr_pages
4128	 * Since the nr_pages field is now removed, we have converted this to
4129	 * return the per cpu buffer value.
4130	 */
4131	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4132		return 0;
4133
4134	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4135}
4136EXPORT_SYMBOL_GPL(ring_buffer_size);
4137
4138static void
4139rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4140{
4141	rb_head_page_deactivate(cpu_buffer);
4142
4143	cpu_buffer->head_page
4144		= list_entry(cpu_buffer->pages, struct buffer_page, list);
4145	local_set(&cpu_buffer->head_page->write, 0);
4146	local_set(&cpu_buffer->head_page->entries, 0);
4147	local_set(&cpu_buffer->head_page->page->commit, 0);
4148
4149	cpu_buffer->head_page->read = 0;
4150
4151	cpu_buffer->tail_page = cpu_buffer->head_page;
4152	cpu_buffer->commit_page = cpu_buffer->head_page;
4153
4154	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4155	INIT_LIST_HEAD(&cpu_buffer->new_pages);
4156	local_set(&cpu_buffer->reader_page->write, 0);
4157	local_set(&cpu_buffer->reader_page->entries, 0);
4158	local_set(&cpu_buffer->reader_page->page->commit, 0);
4159	cpu_buffer->reader_page->read = 0;
4160
4161	local_set(&cpu_buffer->entries_bytes, 0);
4162	local_set(&cpu_buffer->overrun, 0);
4163	local_set(&cpu_buffer->commit_overrun, 0);
4164	local_set(&cpu_buffer->dropped_events, 0);
4165	local_set(&cpu_buffer->entries, 0);
4166	local_set(&cpu_buffer->committing, 0);
4167	local_set(&cpu_buffer->commits, 0);
 
 
 
 
4168	cpu_buffer->read = 0;
4169	cpu_buffer->read_bytes = 0;
4170
4171	cpu_buffer->write_stamp = 0;
4172	cpu_buffer->read_stamp = 0;
4173
4174	cpu_buffer->lost_events = 0;
4175	cpu_buffer->last_overrun = 0;
4176
4177	rb_head_page_activate(cpu_buffer);
4178}
4179
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4180/**
4181 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4182 * @buffer: The ring buffer to reset a per cpu buffer of
4183 * @cpu: The CPU buffer to be reset
4184 */
4185void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4186{
4187	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4188	unsigned long flags;
4189
4190	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4191		return;
4192
4193	atomic_inc(&buffer->resize_disabled);
4194	atomic_inc(&cpu_buffer->record_disabled);
4195
4196	/* Make sure all commits have finished */
4197	synchronize_sched();
4198
4199	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4200
4201	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4202		goto out;
 
 
4203
4204	arch_spin_lock(&cpu_buffer->lock);
 
 
 
 
 
 
 
 
4205
4206	rb_reset_cpu(cpu_buffer);
 
4207
4208	arch_spin_unlock(&cpu_buffer->lock);
 
 
4209
4210 out:
4211	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4212
4213	atomic_dec(&cpu_buffer->record_disabled);
4214	atomic_dec(&buffer->resize_disabled);
 
 
 
 
 
 
4215}
4216EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4217
4218/**
4219 * ring_buffer_reset - reset a ring buffer
4220 * @buffer: The ring buffer to reset all cpu buffers
4221 */
4222void ring_buffer_reset(struct ring_buffer *buffer)
4223{
 
4224	int cpu;
4225
4226	for_each_buffer_cpu(buffer, cpu)
4227		ring_buffer_reset_cpu(buffer, cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4228}
4229EXPORT_SYMBOL_GPL(ring_buffer_reset);
4230
4231/**
4232 * rind_buffer_empty - is the ring buffer empty?
4233 * @buffer: The ring buffer to test
4234 */
4235bool ring_buffer_empty(struct ring_buffer *buffer)
4236{
4237	struct ring_buffer_per_cpu *cpu_buffer;
4238	unsigned long flags;
4239	bool dolock;
4240	int cpu;
4241	int ret;
4242
4243	/* yes this is racy, but if you don't like the race, lock the buffer */
4244	for_each_buffer_cpu(buffer, cpu) {
4245		cpu_buffer = buffer->buffers[cpu];
4246		local_irq_save(flags);
4247		dolock = rb_reader_lock(cpu_buffer);
4248		ret = rb_per_cpu_empty(cpu_buffer);
4249		rb_reader_unlock(cpu_buffer, dolock);
4250		local_irq_restore(flags);
4251
4252		if (!ret)
4253			return false;
4254	}
4255
4256	return true;
4257}
4258EXPORT_SYMBOL_GPL(ring_buffer_empty);
4259
4260/**
4261 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4262 * @buffer: The ring buffer
4263 * @cpu: The CPU buffer to test
4264 */
4265bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4266{
4267	struct ring_buffer_per_cpu *cpu_buffer;
4268	unsigned long flags;
4269	bool dolock;
4270	int ret;
4271
4272	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4273		return true;
4274
4275	cpu_buffer = buffer->buffers[cpu];
4276	local_irq_save(flags);
4277	dolock = rb_reader_lock(cpu_buffer);
4278	ret = rb_per_cpu_empty(cpu_buffer);
4279	rb_reader_unlock(cpu_buffer, dolock);
4280	local_irq_restore(flags);
4281
4282	return ret;
4283}
4284EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4285
4286#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4287/**
4288 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4289 * @buffer_a: One buffer to swap with
4290 * @buffer_b: The other buffer to swap with
 
4291 *
4292 * This function is useful for tracers that want to take a "snapshot"
4293 * of a CPU buffer and has another back up buffer lying around.
4294 * it is expected that the tracer handles the cpu buffer not being
4295 * used at the moment.
4296 */
4297int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4298			 struct ring_buffer *buffer_b, int cpu)
4299{
4300	struct ring_buffer_per_cpu *cpu_buffer_a;
4301	struct ring_buffer_per_cpu *cpu_buffer_b;
4302	int ret = -EINVAL;
4303
4304	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4305	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
4306		goto out;
4307
4308	cpu_buffer_a = buffer_a->buffers[cpu];
4309	cpu_buffer_b = buffer_b->buffers[cpu];
4310
4311	/* At least make sure the two buffers are somewhat the same */
4312	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4313		goto out;
4314
4315	ret = -EAGAIN;
4316
4317	if (atomic_read(&buffer_a->record_disabled))
4318		goto out;
4319
4320	if (atomic_read(&buffer_b->record_disabled))
4321		goto out;
4322
4323	if (atomic_read(&cpu_buffer_a->record_disabled))
4324		goto out;
4325
4326	if (atomic_read(&cpu_buffer_b->record_disabled))
4327		goto out;
4328
4329	/*
4330	 * We can't do a synchronize_sched here because this
4331	 * function can be called in atomic context.
4332	 * Normally this will be called from the same CPU as cpu.
4333	 * If not it's up to the caller to protect this.
4334	 */
4335	atomic_inc(&cpu_buffer_a->record_disabled);
4336	atomic_inc(&cpu_buffer_b->record_disabled);
4337
4338	ret = -EBUSY;
4339	if (local_read(&cpu_buffer_a->committing))
4340		goto out_dec;
4341	if (local_read(&cpu_buffer_b->committing))
4342		goto out_dec;
4343
4344	buffer_a->buffers[cpu] = cpu_buffer_b;
4345	buffer_b->buffers[cpu] = cpu_buffer_a;
4346
4347	cpu_buffer_b->buffer = buffer_a;
4348	cpu_buffer_a->buffer = buffer_b;
4349
4350	ret = 0;
4351
4352out_dec:
4353	atomic_dec(&cpu_buffer_a->record_disabled);
4354	atomic_dec(&cpu_buffer_b->record_disabled);
4355out:
4356	return ret;
4357}
4358EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4359#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4360
4361/**
4362 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4363 * @buffer: the buffer to allocate for.
4364 * @cpu: the cpu buffer to allocate.
4365 *
4366 * This function is used in conjunction with ring_buffer_read_page.
4367 * When reading a full page from the ring buffer, these functions
4368 * can be used to speed up the process. The calling function should
4369 * allocate a few pages first with this function. Then when it
4370 * needs to get pages from the ring buffer, it passes the result
4371 * of this function into ring_buffer_read_page, which will swap
4372 * the page that was allocated, with the read page of the buffer.
4373 *
4374 * Returns:
4375 *  The page allocated, or NULL on error.
4376 */
4377void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4378{
4379	struct buffer_data_page *bpage;
 
 
4380	struct page *page;
4381
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4382	page = alloc_pages_node(cpu_to_node(cpu),
4383				GFP_KERNEL | __GFP_NORETRY, 0);
4384	if (!page)
4385		return NULL;
4386
4387	bpage = page_address(page);
4388
 
4389	rb_init_page(bpage);
4390
4391	return bpage;
4392}
4393EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4394
4395/**
4396 * ring_buffer_free_read_page - free an allocated read page
4397 * @buffer: the buffer the page was allocate for
 
4398 * @data: the page to free
4399 *
4400 * Free a page allocated from ring_buffer_alloc_read_page.
4401 */
4402void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4403{
4404	free_page((unsigned long)data);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4405}
4406EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4407
4408/**
4409 * ring_buffer_read_page - extract a page from the ring buffer
4410 * @buffer: buffer to extract from
4411 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4412 * @len: amount to extract
4413 * @cpu: the cpu of the buffer to extract
4414 * @full: should the extraction only happen when the page is full.
4415 *
4416 * This function will pull out a page from the ring buffer and consume it.
4417 * @data_page must be the address of the variable that was returned
4418 * from ring_buffer_alloc_read_page. This is because the page might be used
4419 * to swap with a page in the ring buffer.
4420 *
4421 * for example:
4422 *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
4423 *	if (!rpage)
4424 *		return error;
4425 *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4426 *	if (ret >= 0)
4427 *		process_page(rpage, ret);
4428 *
4429 * When @full is set, the function will not return true unless
4430 * the writer is off the reader page.
4431 *
4432 * Note: it is up to the calling functions to handle sleeps and wakeups.
4433 *  The ring buffer can be used anywhere in the kernel and can not
4434 *  blindly call wake_up. The layer that uses the ring buffer must be
4435 *  responsible for that.
4436 *
4437 * Returns:
4438 *  >=0 if data has been transferred, returns the offset of consumed data.
4439 *  <0 if no data has been transferred.
4440 */
4441int ring_buffer_read_page(struct ring_buffer *buffer,
4442			  void **data_page, size_t len, int cpu, int full)
4443{
4444	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4445	struct ring_buffer_event *event;
4446	struct buffer_data_page *bpage;
4447	struct buffer_page *reader;
4448	unsigned long missed_events;
4449	unsigned long flags;
4450	unsigned int commit;
4451	unsigned int read;
4452	u64 save_timestamp;
4453	int ret = -1;
4454
4455	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4456		goto out;
4457
4458	/*
4459	 * If len is not big enough to hold the page header, then
4460	 * we can not copy anything.
4461	 */
4462	if (len <= BUF_PAGE_HDR_SIZE)
4463		goto out;
4464
4465	len -= BUF_PAGE_HDR_SIZE;
4466
4467	if (!data_page)
4468		goto out;
4469
4470	bpage = *data_page;
4471	if (!bpage)
4472		goto out;
4473
4474	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4475
4476	reader = rb_get_reader_page(cpu_buffer);
4477	if (!reader)
4478		goto out_unlock;
4479
4480	event = rb_reader_event(cpu_buffer);
4481
4482	read = reader->read;
4483	commit = rb_page_commit(reader);
4484
4485	/* Check if any events were dropped */
4486	missed_events = cpu_buffer->lost_events;
4487
4488	/*
4489	 * If this page has been partially read or
4490	 * if len is not big enough to read the rest of the page or
4491	 * a writer is still on the page, then
4492	 * we must copy the data from the page to the buffer.
4493	 * Otherwise, we can simply swap the page with the one passed in.
4494	 */
4495	if (read || (len < (commit - read)) ||
4496	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
4497		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4498		unsigned int rpos = read;
4499		unsigned int pos = 0;
4500		unsigned int size;
4501
4502		if (full)
4503			goto out_unlock;
4504
4505		if (len > (commit - read))
4506			len = (commit - read);
4507
4508		/* Always keep the time extend and data together */
4509		size = rb_event_ts_length(event);
4510
4511		if (len < size)
4512			goto out_unlock;
4513
4514		/* save the current timestamp, since the user will need it */
4515		save_timestamp = cpu_buffer->read_stamp;
4516
4517		/* Need to copy one event at a time */
4518		do {
4519			/* We need the size of one event, because
4520			 * rb_advance_reader only advances by one event,
4521			 * whereas rb_event_ts_length may include the size of
4522			 * one or two events.
4523			 * We have already ensured there's enough space if this
4524			 * is a time extend. */
4525			size = rb_event_length(event);
4526			memcpy(bpage->data + pos, rpage->data + rpos, size);
4527
4528			len -= size;
4529
4530			rb_advance_reader(cpu_buffer);
4531			rpos = reader->read;
4532			pos += size;
4533
4534			if (rpos >= commit)
4535				break;
4536
4537			event = rb_reader_event(cpu_buffer);
4538			/* Always keep the time extend and data together */
4539			size = rb_event_ts_length(event);
4540		} while (len >= size);
4541
4542		/* update bpage */
4543		local_set(&bpage->commit, pos);
4544		bpage->time_stamp = save_timestamp;
4545
4546		/* we copied everything to the beginning */
4547		read = 0;
4548	} else {
4549		/* update the entry counter */
4550		cpu_buffer->read += rb_page_entries(reader);
4551		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4552
4553		/* swap the pages */
4554		rb_init_page(bpage);
4555		bpage = reader->page;
4556		reader->page = *data_page;
4557		local_set(&reader->write, 0);
4558		local_set(&reader->entries, 0);
4559		reader->read = 0;
4560		*data_page = bpage;
4561
4562		/*
4563		 * Use the real_end for the data size,
4564		 * This gives us a chance to store the lost events
4565		 * on the page.
4566		 */
4567		if (reader->real_end)
4568			local_set(&bpage->commit, reader->real_end);
4569	}
4570	ret = read;
4571
4572	cpu_buffer->lost_events = 0;
4573
4574	commit = local_read(&bpage->commit);
4575	/*
4576	 * Set a flag in the commit field if we lost events
4577	 */
4578	if (missed_events) {
4579		/* If there is room at the end of the page to save the
4580		 * missed events, then record it there.
4581		 */
4582		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4583			memcpy(&bpage->data[commit], &missed_events,
4584			       sizeof(missed_events));
4585			local_add(RB_MISSED_STORED, &bpage->commit);
4586			commit += sizeof(missed_events);
4587		}
4588		local_add(RB_MISSED_EVENTS, &bpage->commit);
4589	}
4590
4591	/*
4592	 * This page may be off to user land. Zero it out here.
4593	 */
4594	if (commit < BUF_PAGE_SIZE)
4595		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4596
4597 out_unlock:
4598	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4599
4600 out:
4601	return ret;
4602}
4603EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4604
4605/*
4606 * We only allocate new buffers, never free them if the CPU goes down.
4607 * If we were to free the buffer, then the user would lose any trace that was in
4608 * the buffer.
4609 */
4610int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4611{
4612	struct ring_buffer *buffer;
4613	long nr_pages_same;
4614	int cpu_i;
4615	unsigned long nr_pages;
4616
4617	buffer = container_of(node, struct ring_buffer, node);
4618	if (cpumask_test_cpu(cpu, buffer->cpumask))
4619		return 0;
4620
4621	nr_pages = 0;
4622	nr_pages_same = 1;
4623	/* check if all cpu sizes are same */
4624	for_each_buffer_cpu(buffer, cpu_i) {
4625		/* fill in the size from first enabled cpu */
4626		if (nr_pages == 0)
4627			nr_pages = buffer->buffers[cpu_i]->nr_pages;
4628		if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4629			nr_pages_same = 0;
4630			break;
4631		}
4632	}
4633	/* allocate minimum pages, user can later expand it */
4634	if (!nr_pages_same)
4635		nr_pages = 2;
4636	buffer->buffers[cpu] =
4637		rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4638	if (!buffer->buffers[cpu]) {
4639		WARN(1, "failed to allocate ring buffer on CPU %u\n",
4640		     cpu);
4641		return -ENOMEM;
4642	}
4643	smp_wmb();
4644	cpumask_set_cpu(cpu, buffer->cpumask);
4645	return 0;
4646}
4647
4648#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4649/*
4650 * This is a basic integrity check of the ring buffer.
4651 * Late in the boot cycle this test will run when configured in.
4652 * It will kick off a thread per CPU that will go into a loop
4653 * writing to the per cpu ring buffer various sizes of data.
4654 * Some of the data will be large items, some small.
4655 *
4656 * Another thread is created that goes into a spin, sending out
4657 * IPIs to the other CPUs to also write into the ring buffer.
4658 * this is to test the nesting ability of the buffer.
4659 *
4660 * Basic stats are recorded and reported. If something in the
4661 * ring buffer should happen that's not expected, a big warning
4662 * is displayed and all ring buffers are disabled.
4663 */
4664static struct task_struct *rb_threads[NR_CPUS] __initdata;
4665
4666struct rb_test_data {
4667	struct ring_buffer	*buffer;
4668	unsigned long		events;
4669	unsigned long		bytes_written;
4670	unsigned long		bytes_alloc;
4671	unsigned long		bytes_dropped;
4672	unsigned long		events_nested;
4673	unsigned long		bytes_written_nested;
4674	unsigned long		bytes_alloc_nested;
4675	unsigned long		bytes_dropped_nested;
4676	int			min_size_nested;
4677	int			max_size_nested;
4678	int			max_size;
4679	int			min_size;
4680	int			cpu;
4681	int			cnt;
4682};
4683
4684static struct rb_test_data rb_data[NR_CPUS] __initdata;
4685
4686/* 1 meg per cpu */
4687#define RB_TEST_BUFFER_SIZE	1048576
4688
4689static char rb_string[] __initdata =
4690	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4691	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4692	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4693
4694static bool rb_test_started __initdata;
4695
4696struct rb_item {
4697	int size;
4698	char str[];
4699};
4700
4701static __init int rb_write_something(struct rb_test_data *data, bool nested)
4702{
4703	struct ring_buffer_event *event;
4704	struct rb_item *item;
4705	bool started;
4706	int event_len;
4707	int size;
4708	int len;
4709	int cnt;
4710
4711	/* Have nested writes different that what is written */
4712	cnt = data->cnt + (nested ? 27 : 0);
4713
4714	/* Multiply cnt by ~e, to make some unique increment */
4715	size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4716
4717	len = size + sizeof(struct rb_item);
4718
4719	started = rb_test_started;
4720	/* read rb_test_started before checking buffer enabled */
4721	smp_rmb();
4722
4723	event = ring_buffer_lock_reserve(data->buffer, len);
4724	if (!event) {
4725		/* Ignore dropped events before test starts. */
4726		if (started) {
4727			if (nested)
4728				data->bytes_dropped += len;
4729			else
4730				data->bytes_dropped_nested += len;
4731		}
4732		return len;
4733	}
4734
4735	event_len = ring_buffer_event_length(event);
4736
4737	if (RB_WARN_ON(data->buffer, event_len < len))
4738		goto out;
4739
4740	item = ring_buffer_event_data(event);
4741	item->size = size;
4742	memcpy(item->str, rb_string, size);
4743
4744	if (nested) {
4745		data->bytes_alloc_nested += event_len;
4746		data->bytes_written_nested += len;
4747		data->events_nested++;
4748		if (!data->min_size_nested || len < data->min_size_nested)
4749			data->min_size_nested = len;
4750		if (len > data->max_size_nested)
4751			data->max_size_nested = len;
4752	} else {
4753		data->bytes_alloc += event_len;
4754		data->bytes_written += len;
4755		data->events++;
4756		if (!data->min_size || len < data->min_size)
4757			data->max_size = len;
4758		if (len > data->max_size)
4759			data->max_size = len;
4760	}
4761
4762 out:
4763	ring_buffer_unlock_commit(data->buffer, event);
4764
4765	return 0;
4766}
4767
4768static __init int rb_test(void *arg)
4769{
4770	struct rb_test_data *data = arg;
4771
4772	while (!kthread_should_stop()) {
4773		rb_write_something(data, false);
4774		data->cnt++;
4775
4776		set_current_state(TASK_INTERRUPTIBLE);
4777		/* Now sleep between a min of 100-300us and a max of 1ms */
4778		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4779	}
4780
4781	return 0;
4782}
4783
4784static __init void rb_ipi(void *ignore)
4785{
4786	struct rb_test_data *data;
4787	int cpu = smp_processor_id();
4788
4789	data = &rb_data[cpu];
4790	rb_write_something(data, true);
4791}
4792
4793static __init int rb_hammer_test(void *arg)
4794{
4795	while (!kthread_should_stop()) {
4796
4797		/* Send an IPI to all cpus to write data! */
4798		smp_call_function(rb_ipi, NULL, 1);
4799		/* No sleep, but for non preempt, let others run */
4800		schedule();
4801	}
4802
4803	return 0;
4804}
4805
4806static __init int test_ringbuffer(void)
4807{
4808	struct task_struct *rb_hammer;
4809	struct ring_buffer *buffer;
4810	int cpu;
4811	int ret = 0;
 
 
 
 
 
4812
4813	pr_info("Running ring buffer tests...\n");
4814
4815	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4816	if (WARN_ON(!buffer))
4817		return 0;
4818
4819	/* Disable buffer so that threads can't write to it yet */
4820	ring_buffer_record_off(buffer);
4821
4822	for_each_online_cpu(cpu) {
4823		rb_data[cpu].buffer = buffer;
4824		rb_data[cpu].cpu = cpu;
4825		rb_data[cpu].cnt = cpu;
4826		rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4827						 "rbtester/%d", cpu);
4828		if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4829			pr_cont("FAILED\n");
4830			ret = PTR_ERR(rb_threads[cpu]);
4831			goto out_free;
4832		}
4833
4834		kthread_bind(rb_threads[cpu], cpu);
4835 		wake_up_process(rb_threads[cpu]);
4836	}
4837
4838	/* Now create the rb hammer! */
4839	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4840	if (WARN_ON(IS_ERR(rb_hammer))) {
4841		pr_cont("FAILED\n");
4842		ret = PTR_ERR(rb_hammer);
4843		goto out_free;
4844	}
4845
4846	ring_buffer_record_on(buffer);
4847	/*
4848	 * Show buffer is enabled before setting rb_test_started.
4849	 * Yes there's a small race window where events could be
4850	 * dropped and the thread wont catch it. But when a ring
4851	 * buffer gets enabled, there will always be some kind of
4852	 * delay before other CPUs see it. Thus, we don't care about
4853	 * those dropped events. We care about events dropped after
4854	 * the threads see that the buffer is active.
4855	 */
4856	smp_wmb();
4857	rb_test_started = true;
4858
4859	set_current_state(TASK_INTERRUPTIBLE);
4860	/* Just run for 10 seconds */;
4861	schedule_timeout(10 * HZ);
4862
4863	kthread_stop(rb_hammer);
4864
4865 out_free:
4866	for_each_online_cpu(cpu) {
4867		if (!rb_threads[cpu])
4868			break;
4869		kthread_stop(rb_threads[cpu]);
4870	}
4871	if (ret) {
4872		ring_buffer_free(buffer);
4873		return ret;
4874	}
4875
4876	/* Report! */
4877	pr_info("finished\n");
4878	for_each_online_cpu(cpu) {
4879		struct ring_buffer_event *event;
4880		struct rb_test_data *data = &rb_data[cpu];
4881		struct rb_item *item;
4882		unsigned long total_events;
4883		unsigned long total_dropped;
4884		unsigned long total_written;
4885		unsigned long total_alloc;
4886		unsigned long total_read = 0;
4887		unsigned long total_size = 0;
4888		unsigned long total_len = 0;
4889		unsigned long total_lost = 0;
4890		unsigned long lost;
4891		int big_event_size;
4892		int small_event_size;
4893
4894		ret = -1;
4895
4896		total_events = data->events + data->events_nested;
4897		total_written = data->bytes_written + data->bytes_written_nested;
4898		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4899		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4900
4901		big_event_size = data->max_size + data->max_size_nested;
4902		small_event_size = data->min_size + data->min_size_nested;
4903
4904		pr_info("CPU %d:\n", cpu);
4905		pr_info("              events:    %ld\n", total_events);
4906		pr_info("       dropped bytes:    %ld\n", total_dropped);
4907		pr_info("       alloced bytes:    %ld\n", total_alloc);
4908		pr_info("       written bytes:    %ld\n", total_written);
4909		pr_info("       biggest event:    %d\n", big_event_size);
4910		pr_info("      smallest event:    %d\n", small_event_size);
4911
4912		if (RB_WARN_ON(buffer, total_dropped))
4913			break;
4914
4915		ret = 0;
4916
4917		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4918			total_lost += lost;
4919			item = ring_buffer_event_data(event);
4920			total_len += ring_buffer_event_length(event);
4921			total_size += item->size + sizeof(struct rb_item);
4922			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4923				pr_info("FAILED!\n");
4924				pr_info("buffer had: %.*s\n", item->size, item->str);
4925				pr_info("expected:   %.*s\n", item->size, rb_string);
4926				RB_WARN_ON(buffer, 1);
4927				ret = -1;
4928				break;
4929			}
4930			total_read++;
4931		}
4932		if (ret)
4933			break;
4934
4935		ret = -1;
4936
4937		pr_info("         read events:   %ld\n", total_read);
4938		pr_info("         lost events:   %ld\n", total_lost);
4939		pr_info("        total events:   %ld\n", total_lost + total_read);
4940		pr_info("  recorded len bytes:   %ld\n", total_len);
4941		pr_info(" recorded size bytes:   %ld\n", total_size);
4942		if (total_lost)
4943			pr_info(" With dropped events, record len and size may not match\n"
4944				" alloced and written from above\n");
4945		if (!total_lost) {
4946			if (RB_WARN_ON(buffer, total_len != total_alloc ||
4947				       total_size != total_written))
4948				break;
4949		}
4950		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4951			break;
4952
4953		ret = 0;
4954	}
4955	if (!ret)
4956		pr_info("Ring buffer PASSED!\n");
4957
4958	ring_buffer_free(buffer);
4959	return 0;
4960}
4961
4962late_initcall(test_ringbuffer);
4963#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
v5.9
   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/security.h>
  15#include <linux/uaccess.h>
  16#include <linux/hardirq.h>
  17#include <linux/kthread.h>	/* for self test */
 
  18#include <linux/module.h>
  19#include <linux/percpu.h>
  20#include <linux/mutex.h>
  21#include <linux/delay.h>
  22#include <linux/slab.h>
  23#include <linux/init.h>
  24#include <linux/hash.h>
  25#include <linux/list.h>
  26#include <linux/cpu.h>
  27#include <linux/oom.h>
  28
  29#include <asm/local.h>
  30
  31static void update_pages_handler(struct work_struct *work);
  32
  33/*
  34 * The ring buffer header is special. We must manually up keep it.
  35 */
  36int ring_buffer_print_entry_header(struct trace_seq *s)
  37{
  38	trace_seq_puts(s, "# compressed entry header\n");
  39	trace_seq_puts(s, "\ttype_len    :    5 bits\n");
  40	trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
  41	trace_seq_puts(s, "\tarray       :   32 bits\n");
  42	trace_seq_putc(s, '\n');
  43	trace_seq_printf(s, "\tpadding     : type == %d\n",
  44			 RINGBUF_TYPE_PADDING);
  45	trace_seq_printf(s, "\ttime_extend : type == %d\n",
  46			 RINGBUF_TYPE_TIME_EXTEND);
  47	trace_seq_printf(s, "\ttime_stamp : type == %d\n",
  48			 RINGBUF_TYPE_TIME_STAMP);
  49	trace_seq_printf(s, "\tdata max type_len  == %d\n",
  50			 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
  51
  52	return !trace_seq_has_overflowed(s);
  53}
  54
  55/*
  56 * The ring buffer is made up of a list of pages. A separate list of pages is
  57 * allocated for each CPU. A writer may only write to a buffer that is
  58 * associated with the CPU it is currently executing on.  A reader may read
  59 * from any per cpu buffer.
  60 *
  61 * The reader is special. For each per cpu buffer, the reader has its own
  62 * reader page. When a reader has read the entire reader page, this reader
  63 * page is swapped with another page in the ring buffer.
  64 *
  65 * Now, as long as the writer is off the reader page, the reader can do what
  66 * ever it wants with that page. The writer will never write to that page
  67 * again (as long as it is out of the ring buffer).
  68 *
  69 * Here's some silly ASCII art.
  70 *
  71 *   +------+
  72 *   |reader|          RING BUFFER
  73 *   |page  |
  74 *   +------+        +---+   +---+   +---+
  75 *                   |   |-->|   |-->|   |
  76 *                   +---+   +---+   +---+
  77 *                     ^               |
  78 *                     |               |
  79 *                     +---------------+
  80 *
  81 *
  82 *   +------+
  83 *   |reader|          RING BUFFER
  84 *   |page  |------------------v
  85 *   +------+        +---+   +---+   +---+
  86 *                   |   |-->|   |-->|   |
  87 *                   +---+   +---+   +---+
  88 *                     ^               |
  89 *                     |               |
  90 *                     +---------------+
  91 *
  92 *
  93 *   +------+
  94 *   |reader|          RING BUFFER
  95 *   |page  |------------------v
  96 *   +------+        +---+   +---+   +---+
  97 *      ^            |   |-->|   |-->|   |
  98 *      |            +---+   +---+   +---+
  99 *      |                              |
 100 *      |                              |
 101 *      +------------------------------+
 102 *
 103 *
 104 *   +------+
 105 *   |buffer|          RING BUFFER
 106 *   |page  |------------------v
 107 *   +------+        +---+   +---+   +---+
 108 *      ^            |   |   |   |-->|   |
 109 *      |   New      +---+   +---+   +---+
 110 *      |  Reader------^               |
 111 *      |   page                       |
 112 *      +------------------------------+
 113 *
 114 *
 115 * After we make this swap, the reader can hand this page off to the splice
 116 * code and be done with it. It can even allocate a new page if it needs to
 117 * and swap that into the ring buffer.
 118 *
 119 * We will be using cmpxchg soon to make all this lockless.
 120 *
 121 */
 122
 123/* Used for individual buffers (after the counter) */
 124#define RB_BUFFER_OFF		(1 << 20)
 125
 126#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
 127
 128#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
 129#define RB_ALIGNMENT		4U
 130#define RB_MAX_SMALL_DATA	(RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 131#define RB_EVNT_MIN_SIZE	8U	/* two 32bit words */
 132#define RB_ALIGN_DATA		__aligned(RB_ALIGNMENT)
 
 
 
 
 
 
 
 
 
 133
 134/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
 135#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
 136
 137enum {
 138	RB_LEN_TIME_EXTEND = 8,
 139	RB_LEN_TIME_STAMP =  8,
 140};
 141
 142#define skip_time_extend(event) \
 143	((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
 144
 145#define extended_time(event) \
 146	(event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
 147
 148static inline int rb_null_event(struct ring_buffer_event *event)
 149{
 150	return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
 151}
 152
 153static void rb_event_set_padding(struct ring_buffer_event *event)
 154{
 155	/* padding has a NULL time_delta */
 156	event->type_len = RINGBUF_TYPE_PADDING;
 157	event->time_delta = 0;
 158}
 159
 160static unsigned
 161rb_event_data_length(struct ring_buffer_event *event)
 162{
 163	unsigned length;
 164
 165	if (event->type_len)
 166		length = event->type_len * RB_ALIGNMENT;
 167	else
 168		length = event->array[0];
 169	return length + RB_EVNT_HDR_SIZE;
 170}
 171
 172/*
 173 * Return the length of the given event. Will return
 174 * the length of the time extend if the event is a
 175 * time extend.
 176 */
 177static inline unsigned
 178rb_event_length(struct ring_buffer_event *event)
 179{
 180	switch (event->type_len) {
 181	case RINGBUF_TYPE_PADDING:
 182		if (rb_null_event(event))
 183			/* undefined */
 184			return -1;
 185		return  event->array[0] + RB_EVNT_HDR_SIZE;
 186
 187	case RINGBUF_TYPE_TIME_EXTEND:
 188		return RB_LEN_TIME_EXTEND;
 189
 190	case RINGBUF_TYPE_TIME_STAMP:
 191		return RB_LEN_TIME_STAMP;
 192
 193	case RINGBUF_TYPE_DATA:
 194		return rb_event_data_length(event);
 195	default:
 196		WARN_ON_ONCE(1);
 197	}
 198	/* not hit */
 199	return 0;
 200}
 201
 202/*
 203 * Return total length of time extend and data,
 204 *   or just the event length for all other events.
 205 */
 206static inline unsigned
 207rb_event_ts_length(struct ring_buffer_event *event)
 208{
 209	unsigned len = 0;
 210
 211	if (extended_time(event)) {
 212		/* time extends include the data event after it */
 213		len = RB_LEN_TIME_EXTEND;
 214		event = skip_time_extend(event);
 215	}
 216	return len + rb_event_length(event);
 217}
 218
 219/**
 220 * ring_buffer_event_length - return the length of the event
 221 * @event: the event to get the length of
 222 *
 223 * Returns the size of the data load of a data event.
 224 * If the event is something other than a data event, it
 225 * returns the size of the event itself. With the exception
 226 * of a TIME EXTEND, where it still returns the size of the
 227 * data load of the data event after it.
 228 */
 229unsigned ring_buffer_event_length(struct ring_buffer_event *event)
 230{
 231	unsigned length;
 232
 233	if (extended_time(event))
 234		event = skip_time_extend(event);
 235
 236	length = rb_event_length(event);
 237	if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 238		return length;
 239	length -= RB_EVNT_HDR_SIZE;
 240	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
 241                length -= sizeof(event->array[0]);
 242	return length;
 243}
 244EXPORT_SYMBOL_GPL(ring_buffer_event_length);
 245
 246/* inline for ring buffer fast paths */
 247static __always_inline void *
 248rb_event_data(struct ring_buffer_event *event)
 249{
 250	if (extended_time(event))
 251		event = skip_time_extend(event);
 252	WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 253	/* If length is in len field, then array[0] has the data */
 254	if (event->type_len)
 255		return (void *)&event->array[0];
 256	/* Otherwise length is in array[0] and array[1] has the data */
 257	return (void *)&event->array[1];
 258}
 259
 260/**
 261 * ring_buffer_event_data - return the data of the event
 262 * @event: the event to get the data from
 263 */
 264void *ring_buffer_event_data(struct ring_buffer_event *event)
 265{
 266	return rb_event_data(event);
 267}
 268EXPORT_SYMBOL_GPL(ring_buffer_event_data);
 269
 270#define for_each_buffer_cpu(buffer, cpu)		\
 271	for_each_cpu(cpu, buffer->cpumask)
 272
 273#define for_each_online_buffer_cpu(buffer, cpu)		\
 274	for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
 275
 276#define TS_SHIFT	27
 277#define TS_MASK		((1ULL << TS_SHIFT) - 1)
 278#define TS_DELTA_TEST	(~TS_MASK)
 279
 280/**
 281 * ring_buffer_event_time_stamp - return the event's extended timestamp
 282 * @event: the event to get the timestamp of
 283 *
 284 * Returns the extended timestamp associated with a data event.
 285 * An extended time_stamp is a 64-bit timestamp represented
 286 * internally in a special way that makes the best use of space
 287 * contained within a ring buffer event.  This function decodes
 288 * it and maps it to a straight u64 value.
 289 */
 290u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
 291{
 292	u64 ts;
 293
 294	ts = event->array[0];
 295	ts <<= TS_SHIFT;
 296	ts += event->time_delta;
 297
 298	return ts;
 299}
 300
 301/* Flag when events were overwritten */
 302#define RB_MISSED_EVENTS	(1 << 31)
 303/* Missed count stored at end */
 304#define RB_MISSED_STORED	(1 << 30)
 305
 306struct buffer_data_page {
 307	u64		 time_stamp;	/* page time stamp */
 308	local_t		 commit;	/* write committed index */
 309	unsigned char	 data[] RB_ALIGN_DATA;	/* data of buffer page */
 310};
 311
 312/*
 313 * Note, the buffer_page list must be first. The buffer pages
 314 * are allocated in cache lines, which means that each buffer
 315 * page will be at the beginning of a cache line, and thus
 316 * the least significant bits will be zero. We use this to
 317 * add flags in the list struct pointers, to make the ring buffer
 318 * lockless.
 319 */
 320struct buffer_page {
 321	struct list_head list;		/* list of buffer pages */
 322	local_t		 write;		/* index for next write */
 323	unsigned	 read;		/* index for next read */
 324	local_t		 entries;	/* entries on this page */
 325	unsigned long	 real_end;	/* real end of data */
 326	struct buffer_data_page *page;	/* Actual data page */
 327};
 328
 329/*
 330 * The buffer page counters, write and entries, must be reset
 331 * atomically when crossing page boundaries. To synchronize this
 332 * update, two counters are inserted into the number. One is
 333 * the actual counter for the write position or count on the page.
 334 *
 335 * The other is a counter of updaters. Before an update happens
 336 * the update partition of the counter is incremented. This will
 337 * allow the updater to update the counter atomically.
 338 *
 339 * The counter is 20 bits, and the state data is 12.
 340 */
 341#define RB_WRITE_MASK		0xfffff
 342#define RB_WRITE_INTCNT		(1 << 20)
 343
 344static void rb_init_page(struct buffer_data_page *bpage)
 345{
 346	local_set(&bpage->commit, 0);
 347}
 348
 
 
 
 
 
 
 
 
 
 
 
 
 349/*
 350 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
 351 * this issue out.
 352 */
 353static void free_buffer_page(struct buffer_page *bpage)
 354{
 355	free_page((unsigned long)bpage->page);
 356	kfree(bpage);
 357}
 358
 359/*
 360 * We need to fit the time_stamp delta into 27 bits.
 361 */
 362static inline int test_time_stamp(u64 delta)
 363{
 364	if (delta & TS_DELTA_TEST)
 365		return 1;
 366	return 0;
 367}
 368
 369#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
 370
 371/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
 372#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
 373
 374int ring_buffer_print_page_header(struct trace_seq *s)
 375{
 376	struct buffer_data_page field;
 377
 378	trace_seq_printf(s, "\tfield: u64 timestamp;\t"
 379			 "offset:0;\tsize:%u;\tsigned:%u;\n",
 380			 (unsigned int)sizeof(field.time_stamp),
 381			 (unsigned int)is_signed_type(u64));
 382
 383	trace_seq_printf(s, "\tfield: local_t commit;\t"
 384			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 385			 (unsigned int)offsetof(typeof(field), commit),
 386			 (unsigned int)sizeof(field.commit),
 387			 (unsigned int)is_signed_type(long));
 388
 389	trace_seq_printf(s, "\tfield: int overwrite;\t"
 390			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 391			 (unsigned int)offsetof(typeof(field), commit),
 392			 1,
 393			 (unsigned int)is_signed_type(long));
 394
 395	trace_seq_printf(s, "\tfield: char data;\t"
 396			 "offset:%u;\tsize:%u;\tsigned:%u;\n",
 397			 (unsigned int)offsetof(typeof(field), data),
 398			 (unsigned int)BUF_PAGE_SIZE,
 399			 (unsigned int)is_signed_type(char));
 400
 401	return !trace_seq_has_overflowed(s);
 402}
 403
 404struct rb_irq_work {
 405	struct irq_work			work;
 406	wait_queue_head_t		waiters;
 407	wait_queue_head_t		full_waiters;
 408	bool				waiters_pending;
 409	bool				full_waiters_pending;
 410	bool				wakeup_full;
 411};
 412
 413/*
 414 * Structure to hold event state and handle nested events.
 415 */
 416struct rb_event_info {
 417	u64			ts;
 418	u64			delta;
 419	u64			before;
 420	u64			after;
 421	unsigned long		length;
 422	struct buffer_page	*tail_page;
 423	int			add_timestamp;
 424};
 425
 426/*
 427 * Used for the add_timestamp
 428 *  NONE
 429 *  EXTEND - wants a time extend
 430 *  ABSOLUTE - the buffer requests all events to have absolute time stamps
 431 *  FORCE - force a full time stamp.
 432 */
 433enum {
 434	RB_ADD_STAMP_NONE		= 0,
 435	RB_ADD_STAMP_EXTEND		= BIT(1),
 436	RB_ADD_STAMP_ABSOLUTE		= BIT(2),
 437	RB_ADD_STAMP_FORCE		= BIT(3)
 438};
 439/*
 440 * Used for which event context the event is in.
 441 *  NMI     = 0
 442 *  IRQ     = 1
 443 *  SOFTIRQ = 2
 444 *  NORMAL  = 3
 445 *
 446 * See trace_recursive_lock() comment below for more details.
 447 */
 448enum {
 449	RB_CTX_NMI,
 450	RB_CTX_IRQ,
 451	RB_CTX_SOFTIRQ,
 452	RB_CTX_NORMAL,
 453	RB_CTX_MAX
 454};
 455
 456#if BITS_PER_LONG == 32
 457#define RB_TIME_32
 458#endif
 459
 460/* To test on 64 bit machines */
 461//#define RB_TIME_32
 462
 463#ifdef RB_TIME_32
 464
 465struct rb_time_struct {
 466	local_t		cnt;
 467	local_t		top;
 468	local_t		bottom;
 469};
 470#else
 471#include <asm/local64.h>
 472struct rb_time_struct {
 473	local64_t	time;
 474};
 475#endif
 476typedef struct rb_time_struct rb_time_t;
 477
 478/*
 479 * head_page == tail_page && head == tail then buffer is empty.
 480 */
 481struct ring_buffer_per_cpu {
 482	int				cpu;
 483	atomic_t			record_disabled;
 484	atomic_t			resize_disabled;
 485	struct trace_buffer	*buffer;
 486	raw_spinlock_t			reader_lock;	/* serialize readers */
 487	arch_spinlock_t			lock;
 488	struct lock_class_key		lock_key;
 489	struct buffer_data_page		*free_page;
 490	unsigned long			nr_pages;
 491	unsigned int			current_context;
 492	struct list_head		*pages;
 493	struct buffer_page		*head_page;	/* read from head */
 494	struct buffer_page		*tail_page;	/* write to tail */
 495	struct buffer_page		*commit_page;	/* committed pages */
 496	struct buffer_page		*reader_page;
 497	unsigned long			lost_events;
 498	unsigned long			last_overrun;
 499	unsigned long			nest;
 500	local_t				entries_bytes;
 501	local_t				entries;
 502	local_t				overrun;
 503	local_t				commit_overrun;
 504	local_t				dropped_events;
 505	local_t				committing;
 506	local_t				commits;
 507	local_t				pages_touched;
 508	local_t				pages_read;
 509	long				last_pages_touch;
 510	size_t				shortest_full;
 511	unsigned long			read;
 512	unsigned long			read_bytes;
 513	rb_time_t			write_stamp;
 514	rb_time_t			before_stamp;
 515	u64				read_stamp;
 516	/* ring buffer pages to update, > 0 to add, < 0 to remove */
 517	long				nr_pages_to_update;
 518	struct list_head		new_pages; /* new pages to add */
 519	struct work_struct		update_pages_work;
 520	struct completion		update_done;
 521
 522	struct rb_irq_work		irq_work;
 523};
 524
 525struct trace_buffer {
 526	unsigned			flags;
 527	int				cpus;
 528	atomic_t			record_disabled;
 
 529	cpumask_var_t			cpumask;
 530
 531	struct lock_class_key		*reader_lock_key;
 532
 533	struct mutex			mutex;
 534
 535	struct ring_buffer_per_cpu	**buffers;
 536
 537	struct hlist_node		node;
 538	u64				(*clock)(void);
 539
 540	struct rb_irq_work		irq_work;
 541	bool				time_stamp_abs;
 542};
 543
 544struct ring_buffer_iter {
 545	struct ring_buffer_per_cpu	*cpu_buffer;
 546	unsigned long			head;
 547	unsigned long			next_event;
 548	struct buffer_page		*head_page;
 549	struct buffer_page		*cache_reader_page;
 550	unsigned long			cache_read;
 551	u64				read_stamp;
 552	u64				page_stamp;
 553	struct ring_buffer_event	*event;
 554	int				missed_events;
 555};
 556
 557#ifdef RB_TIME_32
 558
 559/*
 560 * On 32 bit machines, local64_t is very expensive. As the ring
 561 * buffer doesn't need all the features of a true 64 bit atomic,
 562 * on 32 bit, it uses these functions (64 still uses local64_t).
 563 *
 564 * For the ring buffer, 64 bit required operations for the time is
 565 * the following:
 566 *
 567 *  - Only need 59 bits (uses 60 to make it even).
 568 *  - Reads may fail if it interrupted a modification of the time stamp.
 569 *      It will succeed if it did not interrupt another write even if
 570 *      the read itself is interrupted by a write.
 571 *      It returns whether it was successful or not.
 572 *
 573 *  - Writes always succeed and will overwrite other writes and writes
 574 *      that were done by events interrupting the current write.
 575 *
 576 *  - A write followed by a read of the same time stamp will always succeed,
 577 *      but may not contain the same value.
 578 *
 579 *  - A cmpxchg will fail if it interrupted another write or cmpxchg.
 580 *      Other than that, it acts like a normal cmpxchg.
 581 *
 582 * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
 583 *  (bottom being the least significant 30 bits of the 60 bit time stamp).
 584 *
 585 * The two most significant bits of each half holds a 2 bit counter (0-3).
 586 * Each update will increment this counter by one.
 587 * When reading the top and bottom, if the two counter bits match then the
 588 *  top and bottom together make a valid 60 bit number.
 589 */
 590#define RB_TIME_SHIFT	30
 591#define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
 592
 593static inline int rb_time_cnt(unsigned long val)
 594{
 595	return (val >> RB_TIME_SHIFT) & 3;
 596}
 597
 598static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
 599{
 600	u64 val;
 601
 602	val = top & RB_TIME_VAL_MASK;
 603	val <<= RB_TIME_SHIFT;
 604	val |= bottom & RB_TIME_VAL_MASK;
 605
 606	return val;
 607}
 608
 609static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
 610{
 611	unsigned long top, bottom;
 612	unsigned long c;
 613
 614	/*
 615	 * If the read is interrupted by a write, then the cnt will
 616	 * be different. Loop until both top and bottom have been read
 617	 * without interruption.
 618	 */
 619	do {
 620		c = local_read(&t->cnt);
 621		top = local_read(&t->top);
 622		bottom = local_read(&t->bottom);
 623	} while (c != local_read(&t->cnt));
 624
 625	*cnt = rb_time_cnt(top);
 626
 627	/* If top and bottom counts don't match, this interrupted a write */
 628	if (*cnt != rb_time_cnt(bottom))
 629		return false;
 630
 631	*ret = rb_time_val(top, bottom);
 632	return true;
 633}
 634
 635static bool rb_time_read(rb_time_t *t, u64 *ret)
 636{
 637	unsigned long cnt;
 638
 639	return __rb_time_read(t, ret, &cnt);
 640}
 641
 642static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
 643{
 644	return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
 645}
 646
 647static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom)
 648{
 649	*top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
 650	*bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
 651}
 652
 653static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
 654{
 655	val = rb_time_val_cnt(val, cnt);
 656	local_set(t, val);
 657}
 658
 659static void rb_time_set(rb_time_t *t, u64 val)
 660{
 661	unsigned long cnt, top, bottom;
 662
 663	rb_time_split(val, &top, &bottom);
 664
 665	/* Writes always succeed with a valid number even if it gets interrupted. */
 666	do {
 667		cnt = local_inc_return(&t->cnt);
 668		rb_time_val_set(&t->top, top, cnt);
 669		rb_time_val_set(&t->bottom, bottom, cnt);
 670	} while (cnt != local_read(&t->cnt));
 671}
 672
 673static inline bool
 674rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
 675{
 676	unsigned long ret;
 677
 678	ret = local_cmpxchg(l, expect, set);
 679	return ret == expect;
 680}
 681
 682static int rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
 683{
 684	unsigned long cnt, top, bottom;
 685	unsigned long cnt2, top2, bottom2;
 686	u64 val;
 687
 688	/* The cmpxchg always fails if it interrupted an update */
 689	 if (!__rb_time_read(t, &val, &cnt2))
 690		 return false;
 691
 692	 if (val != expect)
 693		 return false;
 694
 695	 cnt = local_read(&t->cnt);
 696	 if ((cnt & 3) != cnt2)
 697		 return false;
 698
 699	 cnt2 = cnt + 1;
 700
 701	 rb_time_split(val, &top, &bottom);
 702	 top = rb_time_val_cnt(top, cnt);
 703	 bottom = rb_time_val_cnt(bottom, cnt);
 704
 705	 rb_time_split(set, &top2, &bottom2);
 706	 top2 = rb_time_val_cnt(top2, cnt2);
 707	 bottom2 = rb_time_val_cnt(bottom2, cnt2);
 708
 709	if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
 710		return false;
 711	if (!rb_time_read_cmpxchg(&t->top, top, top2))
 712		return false;
 713	if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
 714		return false;
 715	return true;
 716}
 717
 718#else /* 64 bits */
 719
 720/* local64_t always succeeds */
 721
 722static inline bool rb_time_read(rb_time_t *t, u64 *ret)
 723{
 724	*ret = local64_read(&t->time);
 725	return true;
 726}
 727static void rb_time_set(rb_time_t *t, u64 val)
 728{
 729	local64_set(&t->time, val);
 730}
 731
 732static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
 733{
 734	u64 val;
 735	val = local64_cmpxchg(&t->time, expect, set);
 736	return val == expect;
 737}
 738#endif
 739
 740/**
 741 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
 742 * @buffer: The ring_buffer to get the number of pages from
 743 * @cpu: The cpu of the ring_buffer to get the number of pages from
 744 *
 745 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
 746 */
 747size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
 748{
 749	return buffer->buffers[cpu]->nr_pages;
 750}
 751
 752/**
 753 * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
 754 * @buffer: The ring_buffer to get the number of pages from
 755 * @cpu: The cpu of the ring_buffer to get the number of pages from
 756 *
 757 * Returns the number of pages that have content in the ring buffer.
 758 */
 759size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
 760{
 761	size_t read;
 762	size_t cnt;
 763
 764	read = local_read(&buffer->buffers[cpu]->pages_read);
 765	cnt = local_read(&buffer->buffers[cpu]->pages_touched);
 766	/* The reader can read an empty page, but not more than that */
 767	if (cnt < read) {
 768		WARN_ON_ONCE(read > cnt + 1);
 769		return 0;
 770	}
 771
 772	return cnt - read;
 773}
 774
 775/*
 776 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
 777 *
 778 * Schedules a delayed work to wake up any task that is blocked on the
 779 * ring buffer waiters queue.
 780 */
 781static void rb_wake_up_waiters(struct irq_work *work)
 782{
 783	struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
 784
 785	wake_up_all(&rbwork->waiters);
 786	if (rbwork->wakeup_full) {
 787		rbwork->wakeup_full = false;
 788		wake_up_all(&rbwork->full_waiters);
 789	}
 790}
 791
 792/**
 793 * ring_buffer_wait - wait for input to the ring buffer
 794 * @buffer: buffer to wait on
 795 * @cpu: the cpu buffer to wait on
 796 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
 797 *
 798 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 799 * as data is added to any of the @buffer's cpu buffers. Otherwise
 800 * it will wait for data to be added to a specific cpu buffer.
 801 */
 802int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
 803{
 804	struct ring_buffer_per_cpu *cpu_buffer;
 805	DEFINE_WAIT(wait);
 806	struct rb_irq_work *work;
 807	int ret = 0;
 808
 809	/*
 810	 * Depending on what the caller is waiting for, either any
 811	 * data in any cpu buffer, or a specific buffer, put the
 812	 * caller on the appropriate wait queue.
 813	 */
 814	if (cpu == RING_BUFFER_ALL_CPUS) {
 815		work = &buffer->irq_work;
 816		/* Full only makes sense on per cpu reads */
 817		full = 0;
 818	} else {
 819		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 820			return -ENODEV;
 821		cpu_buffer = buffer->buffers[cpu];
 822		work = &cpu_buffer->irq_work;
 823	}
 824
 825
 826	while (true) {
 827		if (full)
 828			prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
 829		else
 830			prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
 831
 832		/*
 833		 * The events can happen in critical sections where
 834		 * checking a work queue can cause deadlocks.
 835		 * After adding a task to the queue, this flag is set
 836		 * only to notify events to try to wake up the queue
 837		 * using irq_work.
 838		 *
 839		 * We don't clear it even if the buffer is no longer
 840		 * empty. The flag only causes the next event to run
 841		 * irq_work to do the work queue wake up. The worse
 842		 * that can happen if we race with !trace_empty() is that
 843		 * an event will cause an irq_work to try to wake up
 844		 * an empty queue.
 845		 *
 846		 * There's no reason to protect this flag either, as
 847		 * the work queue and irq_work logic will do the necessary
 848		 * synchronization for the wake ups. The only thing
 849		 * that is necessary is that the wake up happens after
 850		 * a task has been queued. It's OK for spurious wake ups.
 851		 */
 852		if (full)
 853			work->full_waiters_pending = true;
 854		else
 855			work->waiters_pending = true;
 856
 857		if (signal_pending(current)) {
 858			ret = -EINTR;
 859			break;
 860		}
 861
 862		if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
 863			break;
 864
 865		if (cpu != RING_BUFFER_ALL_CPUS &&
 866		    !ring_buffer_empty_cpu(buffer, cpu)) {
 867			unsigned long flags;
 868			bool pagebusy;
 869			size_t nr_pages;
 870			size_t dirty;
 871
 872			if (!full)
 873				break;
 874
 875			raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
 876			pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 877			nr_pages = cpu_buffer->nr_pages;
 878			dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
 879			if (!cpu_buffer->shortest_full ||
 880			    cpu_buffer->shortest_full < full)
 881				cpu_buffer->shortest_full = full;
 882			raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
 883			if (!pagebusy &&
 884			    (!nr_pages || (dirty * 100) > full * nr_pages))
 885				break;
 886		}
 887
 888		schedule();
 889	}
 890
 891	if (full)
 892		finish_wait(&work->full_waiters, &wait);
 893	else
 894		finish_wait(&work->waiters, &wait);
 895
 896	return ret;
 897}
 898
 899/**
 900 * ring_buffer_poll_wait - poll on buffer input
 901 * @buffer: buffer to wait on
 902 * @cpu: the cpu buffer to wait on
 903 * @filp: the file descriptor
 904 * @poll_table: The poll descriptor
 905 *
 906 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 907 * as data is added to any of the @buffer's cpu buffers. Otherwise
 908 * it will wait for data to be added to a specific cpu buffer.
 909 *
 910 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
 911 * zero otherwise.
 912 */
 913__poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
 914			  struct file *filp, poll_table *poll_table)
 915{
 916	struct ring_buffer_per_cpu *cpu_buffer;
 917	struct rb_irq_work *work;
 918
 919	if (cpu == RING_BUFFER_ALL_CPUS)
 920		work = &buffer->irq_work;
 921	else {
 922		if (!cpumask_test_cpu(cpu, buffer->cpumask))
 923			return -EINVAL;
 924
 925		cpu_buffer = buffer->buffers[cpu];
 926		work = &cpu_buffer->irq_work;
 927	}
 928
 929	poll_wait(filp, &work->waiters, poll_table);
 930	work->waiters_pending = true;
 931	/*
 932	 * There's a tight race between setting the waiters_pending and
 933	 * checking if the ring buffer is empty.  Once the waiters_pending bit
 934	 * is set, the next event will wake the task up, but we can get stuck
 935	 * if there's only a single event in.
 936	 *
 937	 * FIXME: Ideally, we need a memory barrier on the writer side as well,
 938	 * but adding a memory barrier to all events will cause too much of a
 939	 * performance hit in the fast path.  We only need a memory barrier when
 940	 * the buffer goes from empty to having content.  But as this race is
 941	 * extremely small, and it's not a problem if another event comes in, we
 942	 * will fix it later.
 943	 */
 944	smp_mb();
 945
 946	if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
 947	    (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
 948		return EPOLLIN | EPOLLRDNORM;
 949	return 0;
 950}
 951
 952/* buffer may be either ring_buffer or ring_buffer_per_cpu */
 953#define RB_WARN_ON(b, cond)						\
 954	({								\
 955		int _____ret = unlikely(cond);				\
 956		if (_____ret) {						\
 957			if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
 958				struct ring_buffer_per_cpu *__b =	\
 959					(void *)b;			\
 960				atomic_inc(&__b->buffer->record_disabled); \
 961			} else						\
 962				atomic_inc(&b->record_disabled);	\
 963			WARN_ON(1);					\
 964		}							\
 965		_____ret;						\
 966	})
 967
 968/* Up this if you want to test the TIME_EXTENTS and normalization */
 969#define DEBUG_SHIFT 0
 970
 971static inline u64 rb_time_stamp(struct trace_buffer *buffer)
 972{
 973	u64 ts;
 974
 975	/* Skip retpolines :-( */
 976	if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
 977		ts = trace_clock_local();
 978	else
 979		ts = buffer->clock();
 980
 981	/* shift to debug/test normalization and TIME_EXTENTS */
 982	return ts << DEBUG_SHIFT;
 983}
 984
 985u64 ring_buffer_time_stamp(struct trace_buffer *buffer, int cpu)
 986{
 987	u64 time;
 988
 989	preempt_disable_notrace();
 990	time = rb_time_stamp(buffer);
 991	preempt_enable_notrace();
 992
 993	return time;
 994}
 995EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
 996
 997void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
 998				      int cpu, u64 *ts)
 999{
1000	/* Just stupid testing the normalize function and deltas */
1001	*ts >>= DEBUG_SHIFT;
1002}
1003EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1004
1005/*
1006 * Making the ring buffer lockless makes things tricky.
1007 * Although writes only happen on the CPU that they are on,
1008 * and they only need to worry about interrupts. Reads can
1009 * happen on any CPU.
1010 *
1011 * The reader page is always off the ring buffer, but when the
1012 * reader finishes with a page, it needs to swap its page with
1013 * a new one from the buffer. The reader needs to take from
1014 * the head (writes go to the tail). But if a writer is in overwrite
1015 * mode and wraps, it must push the head page forward.
1016 *
1017 * Here lies the problem.
1018 *
1019 * The reader must be careful to replace only the head page, and
1020 * not another one. As described at the top of the file in the
1021 * ASCII art, the reader sets its old page to point to the next
1022 * page after head. It then sets the page after head to point to
1023 * the old reader page. But if the writer moves the head page
1024 * during this operation, the reader could end up with the tail.
1025 *
1026 * We use cmpxchg to help prevent this race. We also do something
1027 * special with the page before head. We set the LSB to 1.
1028 *
1029 * When the writer must push the page forward, it will clear the
1030 * bit that points to the head page, move the head, and then set
1031 * the bit that points to the new head page.
1032 *
1033 * We also don't want an interrupt coming in and moving the head
1034 * page on another writer. Thus we use the second LSB to catch
1035 * that too. Thus:
1036 *
1037 * head->list->prev->next        bit 1          bit 0
1038 *                              -------        -------
1039 * Normal page                     0              0
1040 * Points to head page             0              1
1041 * New head page                   1              0
1042 *
1043 * Note we can not trust the prev pointer of the head page, because:
1044 *
1045 * +----+       +-----+        +-----+
1046 * |    |------>|  T  |---X--->|  N  |
1047 * |    |<------|     |        |     |
1048 * +----+       +-----+        +-----+
1049 *   ^                           ^ |
1050 *   |          +-----+          | |
1051 *   +----------|  R  |----------+ |
1052 *              |     |<-----------+
1053 *              +-----+
1054 *
1055 * Key:  ---X-->  HEAD flag set in pointer
1056 *         T      Tail page
1057 *         R      Reader page
1058 *         N      Next page
1059 *
1060 * (see __rb_reserve_next() to see where this happens)
1061 *
1062 *  What the above shows is that the reader just swapped out
1063 *  the reader page with a page in the buffer, but before it
1064 *  could make the new header point back to the new page added
1065 *  it was preempted by a writer. The writer moved forward onto
1066 *  the new page added by the reader and is about to move forward
1067 *  again.
1068 *
1069 *  You can see, it is legitimate for the previous pointer of
1070 *  the head (or any page) not to point back to itself. But only
1071 *  temporarily.
1072 */
1073
1074#define RB_PAGE_NORMAL		0UL
1075#define RB_PAGE_HEAD		1UL
1076#define RB_PAGE_UPDATE		2UL
1077
1078
1079#define RB_FLAG_MASK		3UL
1080
1081/* PAGE_MOVED is not part of the mask */
1082#define RB_PAGE_MOVED		4UL
1083
1084/*
1085 * rb_list_head - remove any bit
1086 */
1087static struct list_head *rb_list_head(struct list_head *list)
1088{
1089	unsigned long val = (unsigned long)list;
1090
1091	return (struct list_head *)(val & ~RB_FLAG_MASK);
1092}
1093
1094/*
1095 * rb_is_head_page - test if the given page is the head page
1096 *
1097 * Because the reader may move the head_page pointer, we can
1098 * not trust what the head page is (it may be pointing to
1099 * the reader page). But if the next page is a header page,
1100 * its flags will be non zero.
1101 */
1102static inline int
1103rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1104		struct buffer_page *page, struct list_head *list)
1105{
1106	unsigned long val;
1107
1108	val = (unsigned long)list->next;
1109
1110	if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1111		return RB_PAGE_MOVED;
1112
1113	return val & RB_FLAG_MASK;
1114}
1115
1116/*
1117 * rb_is_reader_page
1118 *
1119 * The unique thing about the reader page, is that, if the
1120 * writer is ever on it, the previous pointer never points
1121 * back to the reader page.
1122 */
1123static bool rb_is_reader_page(struct buffer_page *page)
1124{
1125	struct list_head *list = page->list.prev;
1126
1127	return rb_list_head(list->next) != &page->list;
1128}
1129
1130/*
1131 * rb_set_list_to_head - set a list_head to be pointing to head.
1132 */
1133static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
1134				struct list_head *list)
1135{
1136	unsigned long *ptr;
1137
1138	ptr = (unsigned long *)&list->next;
1139	*ptr |= RB_PAGE_HEAD;
1140	*ptr &= ~RB_PAGE_UPDATE;
1141}
1142
1143/*
1144 * rb_head_page_activate - sets up head page
1145 */
1146static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1147{
1148	struct buffer_page *head;
1149
1150	head = cpu_buffer->head_page;
1151	if (!head)
1152		return;
1153
1154	/*
1155	 * Set the previous list pointer to have the HEAD flag.
1156	 */
1157	rb_set_list_to_head(cpu_buffer, head->list.prev);
1158}
1159
1160static void rb_list_head_clear(struct list_head *list)
1161{
1162	unsigned long *ptr = (unsigned long *)&list->next;
1163
1164	*ptr &= ~RB_FLAG_MASK;
1165}
1166
1167/*
1168 * rb_head_page_deactivate - clears head page ptr (for free list)
1169 */
1170static void
1171rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1172{
1173	struct list_head *hd;
1174
1175	/* Go through the whole list and clear any pointers found. */
1176	rb_list_head_clear(cpu_buffer->pages);
1177
1178	list_for_each(hd, cpu_buffer->pages)
1179		rb_list_head_clear(hd);
1180}
1181
1182static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1183			    struct buffer_page *head,
1184			    struct buffer_page *prev,
1185			    int old_flag, int new_flag)
1186{
1187	struct list_head *list;
1188	unsigned long val = (unsigned long)&head->list;
1189	unsigned long ret;
1190
1191	list = &prev->list;
1192
1193	val &= ~RB_FLAG_MASK;
1194
1195	ret = cmpxchg((unsigned long *)&list->next,
1196		      val | old_flag, val | new_flag);
1197
1198	/* check if the reader took the page */
1199	if ((ret & ~RB_FLAG_MASK) != val)
1200		return RB_PAGE_MOVED;
1201
1202	return ret & RB_FLAG_MASK;
1203}
1204
1205static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1206				   struct buffer_page *head,
1207				   struct buffer_page *prev,
1208				   int old_flag)
1209{
1210	return rb_head_page_set(cpu_buffer, head, prev,
1211				old_flag, RB_PAGE_UPDATE);
1212}
1213
1214static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1215				 struct buffer_page *head,
1216				 struct buffer_page *prev,
1217				 int old_flag)
1218{
1219	return rb_head_page_set(cpu_buffer, head, prev,
1220				old_flag, RB_PAGE_HEAD);
1221}
1222
1223static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1224				   struct buffer_page *head,
1225				   struct buffer_page *prev,
1226				   int old_flag)
1227{
1228	return rb_head_page_set(cpu_buffer, head, prev,
1229				old_flag, RB_PAGE_NORMAL);
1230}
1231
1232static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
1233			       struct buffer_page **bpage)
1234{
1235	struct list_head *p = rb_list_head((*bpage)->list.next);
1236
1237	*bpage = list_entry(p, struct buffer_page, list);
1238}
1239
1240static struct buffer_page *
1241rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1242{
1243	struct buffer_page *head;
1244	struct buffer_page *page;
1245	struct list_head *list;
1246	int i;
1247
1248	if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1249		return NULL;
1250
1251	/* sanity check */
1252	list = cpu_buffer->pages;
1253	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1254		return NULL;
1255
1256	page = head = cpu_buffer->head_page;
1257	/*
1258	 * It is possible that the writer moves the header behind
1259	 * where we started, and we miss in one loop.
1260	 * A second loop should grab the header, but we'll do
1261	 * three loops just because I'm paranoid.
1262	 */
1263	for (i = 0; i < 3; i++) {
1264		do {
1265			if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1266				cpu_buffer->head_page = page;
1267				return page;
1268			}
1269			rb_inc_page(cpu_buffer, &page);
1270		} while (page != head);
1271	}
1272
1273	RB_WARN_ON(cpu_buffer, 1);
1274
1275	return NULL;
1276}
1277
1278static int rb_head_page_replace(struct buffer_page *old,
1279				struct buffer_page *new)
1280{
1281	unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1282	unsigned long val;
1283	unsigned long ret;
1284
1285	val = *ptr & ~RB_FLAG_MASK;
1286	val |= RB_PAGE_HEAD;
1287
1288	ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1289
1290	return ret == val;
1291}
1292
1293/*
1294 * rb_tail_page_update - move the tail page forward
1295 */
1296static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1297			       struct buffer_page *tail_page,
1298			       struct buffer_page *next_page)
1299{
1300	unsigned long old_entries;
1301	unsigned long old_write;
1302
1303	/*
1304	 * The tail page now needs to be moved forward.
1305	 *
1306	 * We need to reset the tail page, but without messing
1307	 * with possible erasing of data brought in by interrupts
1308	 * that have moved the tail page and are currently on it.
1309	 *
1310	 * We add a counter to the write field to denote this.
1311	 */
1312	old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1313	old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1314
1315	local_inc(&cpu_buffer->pages_touched);
1316	/*
1317	 * Just make sure we have seen our old_write and synchronize
1318	 * with any interrupts that come in.
1319	 */
1320	barrier();
1321
1322	/*
1323	 * If the tail page is still the same as what we think
1324	 * it is, then it is up to us to update the tail
1325	 * pointer.
1326	 */
1327	if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1328		/* Zero the write counter */
1329		unsigned long val = old_write & ~RB_WRITE_MASK;
1330		unsigned long eval = old_entries & ~RB_WRITE_MASK;
1331
1332		/*
1333		 * This will only succeed if an interrupt did
1334		 * not come in and change it. In which case, we
1335		 * do not want to modify it.
1336		 *
1337		 * We add (void) to let the compiler know that we do not care
1338		 * about the return value of these functions. We use the
1339		 * cmpxchg to only update if an interrupt did not already
1340		 * do it for us. If the cmpxchg fails, we don't care.
1341		 */
1342		(void)local_cmpxchg(&next_page->write, old_write, val);
1343		(void)local_cmpxchg(&next_page->entries, old_entries, eval);
1344
1345		/*
1346		 * No need to worry about races with clearing out the commit.
1347		 * it only can increment when a commit takes place. But that
1348		 * only happens in the outer most nested commit.
1349		 */
1350		local_set(&next_page->page->commit, 0);
1351
1352		/* Again, either we update tail_page or an interrupt does */
1353		(void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1354	}
1355}
1356
1357static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1358			  struct buffer_page *bpage)
1359{
1360	unsigned long val = (unsigned long)bpage;
1361
1362	if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1363		return 1;
1364
1365	return 0;
1366}
1367
1368/**
1369 * rb_check_list - make sure a pointer to a list has the last bits zero
1370 */
1371static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1372			 struct list_head *list)
1373{
1374	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1375		return 1;
1376	if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1377		return 1;
1378	return 0;
1379}
1380
1381/**
1382 * rb_check_pages - integrity check of buffer pages
1383 * @cpu_buffer: CPU buffer with pages to test
1384 *
1385 * As a safety measure we check to make sure the data pages have not
1386 * been corrupted.
1387 */
1388static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1389{
1390	struct list_head *head = cpu_buffer->pages;
1391	struct buffer_page *bpage, *tmp;
1392
1393	/* Reset the head page if it exists */
1394	if (cpu_buffer->head_page)
1395		rb_set_head_page(cpu_buffer);
1396
1397	rb_head_page_deactivate(cpu_buffer);
1398
1399	if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1400		return -1;
1401	if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1402		return -1;
1403
1404	if (rb_check_list(cpu_buffer, head))
1405		return -1;
1406
1407	list_for_each_entry_safe(bpage, tmp, head, list) {
1408		if (RB_WARN_ON(cpu_buffer,
1409			       bpage->list.next->prev != &bpage->list))
1410			return -1;
1411		if (RB_WARN_ON(cpu_buffer,
1412			       bpage->list.prev->next != &bpage->list))
1413			return -1;
1414		if (rb_check_list(cpu_buffer, &bpage->list))
1415			return -1;
1416	}
1417
1418	rb_head_page_activate(cpu_buffer);
1419
1420	return 0;
1421}
1422
1423static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1424{
1425	struct buffer_page *bpage, *tmp;
1426	bool user_thread = current->mm != NULL;
1427	gfp_t mflags;
1428	long i;
1429
1430	/*
1431	 * Check if the available memory is there first.
1432	 * Note, si_mem_available() only gives us a rough estimate of available
1433	 * memory. It may not be accurate. But we don't care, we just want
1434	 * to prevent doing any allocation when it is obvious that it is
1435	 * not going to succeed.
1436	 */
1437	i = si_mem_available();
1438	if (i < nr_pages)
1439		return -ENOMEM;
1440
1441	/*
1442	 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1443	 * gracefully without invoking oom-killer and the system is not
1444	 * destabilized.
1445	 */
1446	mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1447
1448	/*
1449	 * If a user thread allocates too much, and si_mem_available()
1450	 * reports there's enough memory, even though there is not.
1451	 * Make sure the OOM killer kills this thread. This can happen
1452	 * even with RETRY_MAYFAIL because another task may be doing
1453	 * an allocation after this task has taken all memory.
1454	 * This is the task the OOM killer needs to take out during this
1455	 * loop, even if it was triggered by an allocation somewhere else.
1456	 */
1457	if (user_thread)
1458		set_current_oom_origin();
1459	for (i = 0; i < nr_pages; i++) {
1460		struct page *page;
1461
 
 
 
 
1462		bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1463				    mflags, cpu_to_node(cpu));
 
1464		if (!bpage)
1465			goto free_pages;
1466
1467		list_add(&bpage->list, pages);
1468
1469		page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
 
1470		if (!page)
1471			goto free_pages;
1472		bpage->page = page_address(page);
1473		rb_init_page(bpage->page);
1474
1475		if (user_thread && fatal_signal_pending(current))
1476			goto free_pages;
1477	}
1478	if (user_thread)
1479		clear_current_oom_origin();
1480
1481	return 0;
1482
1483free_pages:
1484	list_for_each_entry_safe(bpage, tmp, pages, list) {
1485		list_del_init(&bpage->list);
1486		free_buffer_page(bpage);
1487	}
1488	if (user_thread)
1489		clear_current_oom_origin();
1490
1491	return -ENOMEM;
1492}
1493
1494static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1495			     unsigned long nr_pages)
1496{
1497	LIST_HEAD(pages);
1498
1499	WARN_ON(!nr_pages);
1500
1501	if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1502		return -ENOMEM;
1503
1504	/*
1505	 * The ring buffer page list is a circular list that does not
1506	 * start and end with a list head. All page list items point to
1507	 * other pages.
1508	 */
1509	cpu_buffer->pages = pages.next;
1510	list_del(&pages);
1511
1512	cpu_buffer->nr_pages = nr_pages;
1513
1514	rb_check_pages(cpu_buffer);
1515
1516	return 0;
1517}
1518
1519static struct ring_buffer_per_cpu *
1520rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1521{
1522	struct ring_buffer_per_cpu *cpu_buffer;
1523	struct buffer_page *bpage;
1524	struct page *page;
1525	int ret;
1526
1527	cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1528				  GFP_KERNEL, cpu_to_node(cpu));
1529	if (!cpu_buffer)
1530		return NULL;
1531
1532	cpu_buffer->cpu = cpu;
1533	cpu_buffer->buffer = buffer;
1534	raw_spin_lock_init(&cpu_buffer->reader_lock);
1535	lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1536	cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1537	INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1538	init_completion(&cpu_buffer->update_done);
1539	init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1540	init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1541	init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1542
1543	bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1544			    GFP_KERNEL, cpu_to_node(cpu));
1545	if (!bpage)
1546		goto fail_free_buffer;
1547
1548	rb_check_bpage(cpu_buffer, bpage);
1549
1550	cpu_buffer->reader_page = bpage;
1551	page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1552	if (!page)
1553		goto fail_free_reader;
1554	bpage->page = page_address(page);
1555	rb_init_page(bpage->page);
1556
1557	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1558	INIT_LIST_HEAD(&cpu_buffer->new_pages);
1559
1560	ret = rb_allocate_pages(cpu_buffer, nr_pages);
1561	if (ret < 0)
1562		goto fail_free_reader;
1563
1564	cpu_buffer->head_page
1565		= list_entry(cpu_buffer->pages, struct buffer_page, list);
1566	cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1567
1568	rb_head_page_activate(cpu_buffer);
1569
1570	return cpu_buffer;
1571
1572 fail_free_reader:
1573	free_buffer_page(cpu_buffer->reader_page);
1574
1575 fail_free_buffer:
1576	kfree(cpu_buffer);
1577	return NULL;
1578}
1579
1580static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1581{
1582	struct list_head *head = cpu_buffer->pages;
1583	struct buffer_page *bpage, *tmp;
1584
1585	free_buffer_page(cpu_buffer->reader_page);
1586
1587	rb_head_page_deactivate(cpu_buffer);
1588
1589	if (head) {
1590		list_for_each_entry_safe(bpage, tmp, head, list) {
1591			list_del_init(&bpage->list);
1592			free_buffer_page(bpage);
1593		}
1594		bpage = list_entry(head, struct buffer_page, list);
1595		free_buffer_page(bpage);
1596	}
1597
1598	kfree(cpu_buffer);
1599}
1600
1601/**
1602 * __ring_buffer_alloc - allocate a new ring_buffer
1603 * @size: the size in bytes per cpu that is needed.
1604 * @flags: attributes to set for the ring buffer.
1605 * @key: ring buffer reader_lock_key.
1606 *
1607 * Currently the only flag that is available is the RB_FL_OVERWRITE
1608 * flag. This flag means that the buffer will overwrite old data
1609 * when the buffer wraps. If this flag is not set, the buffer will
1610 * drop data when the tail hits the head.
1611 */
1612struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1613					struct lock_class_key *key)
1614{
1615	struct trace_buffer *buffer;
1616	long nr_pages;
1617	int bsize;
1618	int cpu;
1619	int ret;
1620
1621	/* keep it in its own cache line */
1622	buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1623			 GFP_KERNEL);
1624	if (!buffer)
1625		return NULL;
1626
1627	if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1628		goto fail_free_buffer;
1629
1630	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1631	buffer->flags = flags;
1632	buffer->clock = trace_clock_local;
1633	buffer->reader_lock_key = key;
1634
1635	init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1636	init_waitqueue_head(&buffer->irq_work.waiters);
1637
1638	/* need at least two pages */
1639	if (nr_pages < 2)
1640		nr_pages = 2;
1641
1642	buffer->cpus = nr_cpu_ids;
1643
1644	bsize = sizeof(void *) * nr_cpu_ids;
1645	buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1646				  GFP_KERNEL);
1647	if (!buffer->buffers)
1648		goto fail_free_cpumask;
1649
1650	cpu = raw_smp_processor_id();
1651	cpumask_set_cpu(cpu, buffer->cpumask);
1652	buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1653	if (!buffer->buffers[cpu])
1654		goto fail_free_buffers;
1655
1656	ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1657	if (ret < 0)
1658		goto fail_free_buffers;
1659
1660	mutex_init(&buffer->mutex);
1661
1662	return buffer;
1663
1664 fail_free_buffers:
1665	for_each_buffer_cpu(buffer, cpu) {
1666		if (buffer->buffers[cpu])
1667			rb_free_cpu_buffer(buffer->buffers[cpu]);
1668	}
1669	kfree(buffer->buffers);
1670
1671 fail_free_cpumask:
1672	free_cpumask_var(buffer->cpumask);
1673
1674 fail_free_buffer:
1675	kfree(buffer);
1676	return NULL;
1677}
1678EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1679
1680/**
1681 * ring_buffer_free - free a ring buffer.
1682 * @buffer: the buffer to free.
1683 */
1684void
1685ring_buffer_free(struct trace_buffer *buffer)
1686{
1687	int cpu;
1688
1689	cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1690
1691	for_each_buffer_cpu(buffer, cpu)
1692		rb_free_cpu_buffer(buffer->buffers[cpu]);
1693
1694	kfree(buffer->buffers);
1695	free_cpumask_var(buffer->cpumask);
1696
1697	kfree(buffer);
1698}
1699EXPORT_SYMBOL_GPL(ring_buffer_free);
1700
1701void ring_buffer_set_clock(struct trace_buffer *buffer,
1702			   u64 (*clock)(void))
1703{
1704	buffer->clock = clock;
1705}
1706
1707void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1708{
1709	buffer->time_stamp_abs = abs;
1710}
1711
1712bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1713{
1714	return buffer->time_stamp_abs;
1715}
1716
1717static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1718
1719static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1720{
1721	return local_read(&bpage->entries) & RB_WRITE_MASK;
1722}
1723
1724static inline unsigned long rb_page_write(struct buffer_page *bpage)
1725{
1726	return local_read(&bpage->write) & RB_WRITE_MASK;
1727}
1728
1729static int
1730rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1731{
1732	struct list_head *tail_page, *to_remove, *next_page;
1733	struct buffer_page *to_remove_page, *tmp_iter_page;
1734	struct buffer_page *last_page, *first_page;
1735	unsigned long nr_removed;
1736	unsigned long head_bit;
1737	int page_entries;
1738
1739	head_bit = 0;
1740
1741	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1742	atomic_inc(&cpu_buffer->record_disabled);
1743	/*
1744	 * We don't race with the readers since we have acquired the reader
1745	 * lock. We also don't race with writers after disabling recording.
1746	 * This makes it easy to figure out the first and the last page to be
1747	 * removed from the list. We unlink all the pages in between including
1748	 * the first and last pages. This is done in a busy loop so that we
1749	 * lose the least number of traces.
1750	 * The pages are freed after we restart recording and unlock readers.
1751	 */
1752	tail_page = &cpu_buffer->tail_page->list;
1753
1754	/*
1755	 * tail page might be on reader page, we remove the next page
1756	 * from the ring buffer
1757	 */
1758	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1759		tail_page = rb_list_head(tail_page->next);
1760	to_remove = tail_page;
1761
1762	/* start of pages to remove */
1763	first_page = list_entry(rb_list_head(to_remove->next),
1764				struct buffer_page, list);
1765
1766	for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1767		to_remove = rb_list_head(to_remove)->next;
1768		head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1769	}
1770
1771	next_page = rb_list_head(to_remove)->next;
1772
1773	/*
1774	 * Now we remove all pages between tail_page and next_page.
1775	 * Make sure that we have head_bit value preserved for the
1776	 * next page
1777	 */
1778	tail_page->next = (struct list_head *)((unsigned long)next_page |
1779						head_bit);
1780	next_page = rb_list_head(next_page);
1781	next_page->prev = tail_page;
1782
1783	/* make sure pages points to a valid page in the ring buffer */
1784	cpu_buffer->pages = next_page;
1785
1786	/* update head page */
1787	if (head_bit)
1788		cpu_buffer->head_page = list_entry(next_page,
1789						struct buffer_page, list);
1790
1791	/*
1792	 * change read pointer to make sure any read iterators reset
1793	 * themselves
1794	 */
1795	cpu_buffer->read = 0;
1796
1797	/* pages are removed, resume tracing and then free the pages */
1798	atomic_dec(&cpu_buffer->record_disabled);
1799	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1800
1801	RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1802
1803	/* last buffer page to remove */
1804	last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1805				list);
1806	tmp_iter_page = first_page;
1807
1808	do {
1809		cond_resched();
1810
1811		to_remove_page = tmp_iter_page;
1812		rb_inc_page(cpu_buffer, &tmp_iter_page);
1813
1814		/* update the counters */
1815		page_entries = rb_page_entries(to_remove_page);
1816		if (page_entries) {
1817			/*
1818			 * If something was added to this page, it was full
1819			 * since it is not the tail page. So we deduct the
1820			 * bytes consumed in ring buffer from here.
1821			 * Increment overrun to account for the lost events.
1822			 */
1823			local_add(page_entries, &cpu_buffer->overrun);
1824			local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1825		}
1826
1827		/*
1828		 * We have already removed references to this list item, just
1829		 * free up the buffer_page and its page
1830		 */
1831		free_buffer_page(to_remove_page);
1832		nr_removed--;
1833
1834	} while (to_remove_page != last_page);
1835
1836	RB_WARN_ON(cpu_buffer, nr_removed);
1837
1838	return nr_removed == 0;
1839}
1840
1841static int
1842rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1843{
1844	struct list_head *pages = &cpu_buffer->new_pages;
1845	int retries, success;
1846
1847	raw_spin_lock_irq(&cpu_buffer->reader_lock);
1848	/*
1849	 * We are holding the reader lock, so the reader page won't be swapped
1850	 * in the ring buffer. Now we are racing with the writer trying to
1851	 * move head page and the tail page.
1852	 * We are going to adapt the reader page update process where:
1853	 * 1. We first splice the start and end of list of new pages between
1854	 *    the head page and its previous page.
1855	 * 2. We cmpxchg the prev_page->next to point from head page to the
1856	 *    start of new pages list.
1857	 * 3. Finally, we update the head->prev to the end of new list.
1858	 *
1859	 * We will try this process 10 times, to make sure that we don't keep
1860	 * spinning.
1861	 */
1862	retries = 10;
1863	success = 0;
1864	while (retries--) {
1865		struct list_head *head_page, *prev_page, *r;
1866		struct list_head *last_page, *first_page;
1867		struct list_head *head_page_with_bit;
1868
1869		head_page = &rb_set_head_page(cpu_buffer)->list;
1870		if (!head_page)
1871			break;
1872		prev_page = head_page->prev;
1873
1874		first_page = pages->next;
1875		last_page  = pages->prev;
1876
1877		head_page_with_bit = (struct list_head *)
1878				     ((unsigned long)head_page | RB_PAGE_HEAD);
1879
1880		last_page->next = head_page_with_bit;
1881		first_page->prev = prev_page;
1882
1883		r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1884
1885		if (r == head_page_with_bit) {
1886			/*
1887			 * yay, we replaced the page pointer to our new list,
1888			 * now, we just have to update to head page's prev
1889			 * pointer to point to end of list
1890			 */
1891			head_page->prev = last_page;
1892			success = 1;
1893			break;
1894		}
1895	}
1896
1897	if (success)
1898		INIT_LIST_HEAD(pages);
1899	/*
1900	 * If we weren't successful in adding in new pages, warn and stop
1901	 * tracing
1902	 */
1903	RB_WARN_ON(cpu_buffer, !success);
1904	raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1905
1906	/* free pages if they weren't inserted */
1907	if (!success) {
1908		struct buffer_page *bpage, *tmp;
1909		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1910					 list) {
1911			list_del_init(&bpage->list);
1912			free_buffer_page(bpage);
1913		}
1914	}
1915	return success;
1916}
1917
1918static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1919{
1920	int success;
1921
1922	if (cpu_buffer->nr_pages_to_update > 0)
1923		success = rb_insert_pages(cpu_buffer);
1924	else
1925		success = rb_remove_pages(cpu_buffer,
1926					-cpu_buffer->nr_pages_to_update);
1927
1928	if (success)
1929		cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1930}
1931
1932static void update_pages_handler(struct work_struct *work)
1933{
1934	struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1935			struct ring_buffer_per_cpu, update_pages_work);
1936	rb_update_pages(cpu_buffer);
1937	complete(&cpu_buffer->update_done);
1938}
1939
1940/**
1941 * ring_buffer_resize - resize the ring buffer
1942 * @buffer: the buffer to resize.
1943 * @size: the new size.
1944 * @cpu_id: the cpu buffer to resize
1945 *
1946 * Minimum size is 2 * BUF_PAGE_SIZE.
1947 *
1948 * Returns 0 on success and < 0 on failure.
1949 */
1950int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
1951			int cpu_id)
1952{
1953	struct ring_buffer_per_cpu *cpu_buffer;
1954	unsigned long nr_pages;
1955	int cpu, err = 0;
1956
1957	/*
1958	 * Always succeed at resizing a non-existent buffer:
1959	 */
1960	if (!buffer)
1961		return size;
1962
1963	/* Make sure the requested buffer exists */
1964	if (cpu_id != RING_BUFFER_ALL_CPUS &&
1965	    !cpumask_test_cpu(cpu_id, buffer->cpumask))
1966		return size;
1967
1968	nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1969
1970	/* we need a minimum of two pages */
1971	if (nr_pages < 2)
1972		nr_pages = 2;
1973
1974	size = nr_pages * BUF_PAGE_SIZE;
1975
 
 
 
 
 
 
 
 
1976	/* prevent another thread from changing buffer sizes */
1977	mutex_lock(&buffer->mutex);
1978
1979
1980	if (cpu_id == RING_BUFFER_ALL_CPUS) {
1981		/*
1982		 * Don't succeed if resizing is disabled, as a reader might be
1983		 * manipulating the ring buffer and is expecting a sane state while
1984		 * this is true.
1985		 */
1986		for_each_buffer_cpu(buffer, cpu) {
1987			cpu_buffer = buffer->buffers[cpu];
1988			if (atomic_read(&cpu_buffer->resize_disabled)) {
1989				err = -EBUSY;
1990				goto out_err_unlock;
1991			}
1992		}
1993
1994		/* calculate the pages to update */
1995		for_each_buffer_cpu(buffer, cpu) {
1996			cpu_buffer = buffer->buffers[cpu];
1997
1998			cpu_buffer->nr_pages_to_update = nr_pages -
1999							cpu_buffer->nr_pages;
2000			/*
2001			 * nothing more to do for removing pages or no update
2002			 */
2003			if (cpu_buffer->nr_pages_to_update <= 0)
2004				continue;
2005			/*
2006			 * to add pages, make sure all new pages can be
2007			 * allocated without receiving ENOMEM
2008			 */
2009			INIT_LIST_HEAD(&cpu_buffer->new_pages);
2010			if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
2011						&cpu_buffer->new_pages, cpu)) {
2012				/* not enough memory for new pages */
2013				err = -ENOMEM;
2014				goto out_err;
2015			}
2016		}
2017
2018		get_online_cpus();
2019		/*
2020		 * Fire off all the required work handlers
2021		 * We can't schedule on offline CPUs, but it's not necessary
2022		 * since we can change their buffer sizes without any race.
2023		 */
2024		for_each_buffer_cpu(buffer, cpu) {
2025			cpu_buffer = buffer->buffers[cpu];
2026			if (!cpu_buffer->nr_pages_to_update)
2027				continue;
2028
2029			/* Can't run something on an offline CPU. */
2030			if (!cpu_online(cpu)) {
2031				rb_update_pages(cpu_buffer);
2032				cpu_buffer->nr_pages_to_update = 0;
2033			} else {
2034				schedule_work_on(cpu,
2035						&cpu_buffer->update_pages_work);
2036			}
2037		}
2038
2039		/* wait for all the updates to complete */
2040		for_each_buffer_cpu(buffer, cpu) {
2041			cpu_buffer = buffer->buffers[cpu];
2042			if (!cpu_buffer->nr_pages_to_update)
2043				continue;
2044
2045			if (cpu_online(cpu))
2046				wait_for_completion(&cpu_buffer->update_done);
2047			cpu_buffer->nr_pages_to_update = 0;
2048		}
2049
2050		put_online_cpus();
2051	} else {
2052		/* Make sure this CPU has been initialized */
2053		if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
2054			goto out;
2055
2056		cpu_buffer = buffer->buffers[cpu_id];
2057
2058		if (nr_pages == cpu_buffer->nr_pages)
2059			goto out;
2060
2061		/*
2062		 * Don't succeed if resizing is disabled, as a reader might be
2063		 * manipulating the ring buffer and is expecting a sane state while
2064		 * this is true.
2065		 */
2066		if (atomic_read(&cpu_buffer->resize_disabled)) {
2067			err = -EBUSY;
2068			goto out_err_unlock;
2069		}
2070
2071		cpu_buffer->nr_pages_to_update = nr_pages -
2072						cpu_buffer->nr_pages;
2073
2074		INIT_LIST_HEAD(&cpu_buffer->new_pages);
2075		if (cpu_buffer->nr_pages_to_update > 0 &&
2076			__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
2077					    &cpu_buffer->new_pages, cpu_id)) {
2078			err = -ENOMEM;
2079			goto out_err;
2080		}
2081
2082		get_online_cpus();
2083
2084		/* Can't run something on an offline CPU. */
2085		if (!cpu_online(cpu_id))
2086			rb_update_pages(cpu_buffer);
2087		else {
2088			schedule_work_on(cpu_id,
2089					 &cpu_buffer->update_pages_work);
2090			wait_for_completion(&cpu_buffer->update_done);
2091		}
2092
2093		cpu_buffer->nr_pages_to_update = 0;
2094		put_online_cpus();
2095	}
2096
2097 out:
2098	/*
2099	 * The ring buffer resize can happen with the ring buffer
2100	 * enabled, so that the update disturbs the tracing as little
2101	 * as possible. But if the buffer is disabled, we do not need
2102	 * to worry about that, and we can take the time to verify
2103	 * that the buffer is not corrupt.
2104	 */
2105	if (atomic_read(&buffer->record_disabled)) {
2106		atomic_inc(&buffer->record_disabled);
2107		/*
2108		 * Even though the buffer was disabled, we must make sure
2109		 * that it is truly disabled before calling rb_check_pages.
2110		 * There could have been a race between checking
2111		 * record_disable and incrementing it.
2112		 */
2113		synchronize_rcu();
2114		for_each_buffer_cpu(buffer, cpu) {
2115			cpu_buffer = buffer->buffers[cpu];
2116			rb_check_pages(cpu_buffer);
2117		}
2118		atomic_dec(&buffer->record_disabled);
2119	}
2120
2121	mutex_unlock(&buffer->mutex);
2122	return size;
2123
2124 out_err:
2125	for_each_buffer_cpu(buffer, cpu) {
2126		struct buffer_page *bpage, *tmp;
2127
2128		cpu_buffer = buffer->buffers[cpu];
2129		cpu_buffer->nr_pages_to_update = 0;
2130
2131		if (list_empty(&cpu_buffer->new_pages))
2132			continue;
2133
2134		list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2135					list) {
2136			list_del_init(&bpage->list);
2137			free_buffer_page(bpage);
2138		}
2139	}
2140 out_err_unlock:
2141	mutex_unlock(&buffer->mutex);
2142	return err;
2143}
2144EXPORT_SYMBOL_GPL(ring_buffer_resize);
2145
2146void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2147{
2148	mutex_lock(&buffer->mutex);
2149	if (val)
2150		buffer->flags |= RB_FL_OVERWRITE;
2151	else
2152		buffer->flags &= ~RB_FL_OVERWRITE;
2153	mutex_unlock(&buffer->mutex);
2154}
2155EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2156
 
 
 
 
 
 
2157static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2158{
2159	return bpage->page->data + index;
2160}
2161
2162static __always_inline struct ring_buffer_event *
2163rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2164{
2165	return __rb_page_index(cpu_buffer->reader_page,
2166			       cpu_buffer->reader_page->read);
2167}
2168
2169static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
 
2170{
2171	return local_read(&bpage->page->commit);
2172}
2173
2174static struct ring_buffer_event *
2175rb_iter_head_event(struct ring_buffer_iter *iter)
2176{
2177	struct ring_buffer_event *event;
2178	struct buffer_page *iter_head_page = iter->head_page;
2179	unsigned long commit;
2180	unsigned length;
2181
2182	if (iter->head != iter->next_event)
2183		return iter->event;
2184
2185	/*
2186	 * When the writer goes across pages, it issues a cmpxchg which
2187	 * is a mb(), which will synchronize with the rmb here.
2188	 * (see rb_tail_page_update() and __rb_reserve_next())
2189	 */
2190	commit = rb_page_commit(iter_head_page);
2191	smp_rmb();
2192	event = __rb_page_index(iter_head_page, iter->head);
2193	length = rb_event_length(event);
2194
2195	/*
2196	 * READ_ONCE() doesn't work on functions and we don't want the
2197	 * compiler doing any crazy optimizations with length.
2198	 */
2199	barrier();
2200
2201	if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
2202		/* Writer corrupted the read? */
2203		goto reset;
2204
2205	memcpy(iter->event, event, length);
2206	/*
2207	 * If the page stamp is still the same after this rmb() then the
2208	 * event was safely copied without the writer entering the page.
2209	 */
2210	smp_rmb();
2211
2212	/* Make sure the page didn't change since we read this */
2213	if (iter->page_stamp != iter_head_page->page->time_stamp ||
2214	    commit > rb_page_commit(iter_head_page))
2215		goto reset;
2216
2217	iter->next_event = iter->head + length;
2218	return iter->event;
2219 reset:
2220	/* Reset to the beginning */
2221	iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2222	iter->head = 0;
2223	iter->next_event = 0;
2224	iter->missed_events = 1;
2225	return NULL;
2226}
2227
2228/* Size is determined by what has been committed */
2229static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2230{
2231	return rb_page_commit(bpage);
2232}
2233
2234static __always_inline unsigned
2235rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2236{
2237	return rb_page_commit(cpu_buffer->commit_page);
2238}
2239
2240static __always_inline unsigned
2241rb_event_index(struct ring_buffer_event *event)
2242{
2243	unsigned long addr = (unsigned long)event;
2244
2245	return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2246}
2247
2248static void rb_inc_iter(struct ring_buffer_iter *iter)
2249{
2250	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2251
2252	/*
2253	 * The iterator could be on the reader page (it starts there).
2254	 * But the head could have moved, since the reader was
2255	 * found. Check for this case and assign the iterator
2256	 * to the head page instead of next.
2257	 */
2258	if (iter->head_page == cpu_buffer->reader_page)
2259		iter->head_page = rb_set_head_page(cpu_buffer);
2260	else
2261		rb_inc_page(cpu_buffer, &iter->head_page);
2262
2263	iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2264	iter->head = 0;
2265	iter->next_event = 0;
2266}
2267
2268/*
2269 * rb_handle_head_page - writer hit the head page
2270 *
2271 * Returns: +1 to retry page
2272 *           0 to continue
2273 *          -1 on error
2274 */
2275static int
2276rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2277		    struct buffer_page *tail_page,
2278		    struct buffer_page *next_page)
2279{
2280	struct buffer_page *new_head;
2281	int entries;
2282	int type;
2283	int ret;
2284
2285	entries = rb_page_entries(next_page);
2286
2287	/*
2288	 * The hard part is here. We need to move the head
2289	 * forward, and protect against both readers on
2290	 * other CPUs and writers coming in via interrupts.
2291	 */
2292	type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2293				       RB_PAGE_HEAD);
2294
2295	/*
2296	 * type can be one of four:
2297	 *  NORMAL - an interrupt already moved it for us
2298	 *  HEAD   - we are the first to get here.
2299	 *  UPDATE - we are the interrupt interrupting
2300	 *           a current move.
2301	 *  MOVED  - a reader on another CPU moved the next
2302	 *           pointer to its reader page. Give up
2303	 *           and try again.
2304	 */
2305
2306	switch (type) {
2307	case RB_PAGE_HEAD:
2308		/*
2309		 * We changed the head to UPDATE, thus
2310		 * it is our responsibility to update
2311		 * the counters.
2312		 */
2313		local_add(entries, &cpu_buffer->overrun);
2314		local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2315
2316		/*
2317		 * The entries will be zeroed out when we move the
2318		 * tail page.
2319		 */
2320
2321		/* still more to do */
2322		break;
2323
2324	case RB_PAGE_UPDATE:
2325		/*
2326		 * This is an interrupt that interrupt the
2327		 * previous update. Still more to do.
2328		 */
2329		break;
2330	case RB_PAGE_NORMAL:
2331		/*
2332		 * An interrupt came in before the update
2333		 * and processed this for us.
2334		 * Nothing left to do.
2335		 */
2336		return 1;
2337	case RB_PAGE_MOVED:
2338		/*
2339		 * The reader is on another CPU and just did
2340		 * a swap with our next_page.
2341		 * Try again.
2342		 */
2343		return 1;
2344	default:
2345		RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2346		return -1;
2347	}
2348
2349	/*
2350	 * Now that we are here, the old head pointer is
2351	 * set to UPDATE. This will keep the reader from
2352	 * swapping the head page with the reader page.
2353	 * The reader (on another CPU) will spin till
2354	 * we are finished.
2355	 *
2356	 * We just need to protect against interrupts
2357	 * doing the job. We will set the next pointer
2358	 * to HEAD. After that, we set the old pointer
2359	 * to NORMAL, but only if it was HEAD before.
2360	 * otherwise we are an interrupt, and only
2361	 * want the outer most commit to reset it.
2362	 */
2363	new_head = next_page;
2364	rb_inc_page(cpu_buffer, &new_head);
2365
2366	ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2367				    RB_PAGE_NORMAL);
2368
2369	/*
2370	 * Valid returns are:
2371	 *  HEAD   - an interrupt came in and already set it.
2372	 *  NORMAL - One of two things:
2373	 *            1) We really set it.
2374	 *            2) A bunch of interrupts came in and moved
2375	 *               the page forward again.
2376	 */
2377	switch (ret) {
2378	case RB_PAGE_HEAD:
2379	case RB_PAGE_NORMAL:
2380		/* OK */
2381		break;
2382	default:
2383		RB_WARN_ON(cpu_buffer, 1);
2384		return -1;
2385	}
2386
2387	/*
2388	 * It is possible that an interrupt came in,
2389	 * set the head up, then more interrupts came in
2390	 * and moved it again. When we get back here,
2391	 * the page would have been set to NORMAL but we
2392	 * just set it back to HEAD.
2393	 *
2394	 * How do you detect this? Well, if that happened
2395	 * the tail page would have moved.
2396	 */
2397	if (ret == RB_PAGE_NORMAL) {
2398		struct buffer_page *buffer_tail_page;
2399
2400		buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2401		/*
2402		 * If the tail had moved passed next, then we need
2403		 * to reset the pointer.
2404		 */
2405		if (buffer_tail_page != tail_page &&
2406		    buffer_tail_page != next_page)
2407			rb_head_page_set_normal(cpu_buffer, new_head,
2408						next_page,
2409						RB_PAGE_HEAD);
2410	}
2411
2412	/*
2413	 * If this was the outer most commit (the one that
2414	 * changed the original pointer from HEAD to UPDATE),
2415	 * then it is up to us to reset it to NORMAL.
2416	 */
2417	if (type == RB_PAGE_HEAD) {
2418		ret = rb_head_page_set_normal(cpu_buffer, next_page,
2419					      tail_page,
2420					      RB_PAGE_UPDATE);
2421		if (RB_WARN_ON(cpu_buffer,
2422			       ret != RB_PAGE_UPDATE))
2423			return -1;
2424	}
2425
2426	return 0;
2427}
2428
2429static inline void
2430rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2431	      unsigned long tail, struct rb_event_info *info)
2432{
2433	struct buffer_page *tail_page = info->tail_page;
2434	struct ring_buffer_event *event;
2435	unsigned long length = info->length;
2436
2437	/*
2438	 * Only the event that crossed the page boundary
2439	 * must fill the old tail_page with padding.
2440	 */
2441	if (tail >= BUF_PAGE_SIZE) {
2442		/*
2443		 * If the page was filled, then we still need
2444		 * to update the real_end. Reset it to zero
2445		 * and the reader will ignore it.
2446		 */
2447		if (tail == BUF_PAGE_SIZE)
2448			tail_page->real_end = 0;
2449
2450		local_sub(length, &tail_page->write);
2451		return;
2452	}
2453
2454	event = __rb_page_index(tail_page, tail);
 
2455
2456	/* account for padding bytes */
2457	local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2458
2459	/*
2460	 * Save the original length to the meta data.
2461	 * This will be used by the reader to add lost event
2462	 * counter.
2463	 */
2464	tail_page->real_end = tail;
2465
2466	/*
2467	 * If this event is bigger than the minimum size, then
2468	 * we need to be careful that we don't subtract the
2469	 * write counter enough to allow another writer to slip
2470	 * in on this page.
2471	 * We put in a discarded commit instead, to make sure
2472	 * that this space is not used again.
2473	 *
2474	 * If we are less than the minimum size, we don't need to
2475	 * worry about it.
2476	 */
2477	if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2478		/* No room for any events */
2479
2480		/* Mark the rest of the page with padding */
2481		rb_event_set_padding(event);
2482
2483		/* Set the write back to the previous setting */
2484		local_sub(length, &tail_page->write);
2485		return;
2486	}
2487
2488	/* Put in a discarded event */
2489	event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2490	event->type_len = RINGBUF_TYPE_PADDING;
2491	/* time delta must be non zero */
2492	event->time_delta = 1;
2493
2494	/* Set write to end of buffer */
2495	length = (tail + length) - BUF_PAGE_SIZE;
2496	local_sub(length, &tail_page->write);
2497}
2498
2499static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2500
2501/*
2502 * This is the slow path, force gcc not to inline it.
2503 */
2504static noinline struct ring_buffer_event *
2505rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2506	     unsigned long tail, struct rb_event_info *info)
2507{
2508	struct buffer_page *tail_page = info->tail_page;
2509	struct buffer_page *commit_page = cpu_buffer->commit_page;
2510	struct trace_buffer *buffer = cpu_buffer->buffer;
2511	struct buffer_page *next_page;
2512	int ret;
2513
2514	next_page = tail_page;
2515
2516	rb_inc_page(cpu_buffer, &next_page);
2517
2518	/*
2519	 * If for some reason, we had an interrupt storm that made
2520	 * it all the way around the buffer, bail, and warn
2521	 * about it.
2522	 */
2523	if (unlikely(next_page == commit_page)) {
2524		local_inc(&cpu_buffer->commit_overrun);
2525		goto out_reset;
2526	}
2527
2528	/*
2529	 * This is where the fun begins!
2530	 *
2531	 * We are fighting against races between a reader that
2532	 * could be on another CPU trying to swap its reader
2533	 * page with the buffer head.
2534	 *
2535	 * We are also fighting against interrupts coming in and
2536	 * moving the head or tail on us as well.
2537	 *
2538	 * If the next page is the head page then we have filled
2539	 * the buffer, unless the commit page is still on the
2540	 * reader page.
2541	 */
2542	if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2543
2544		/*
2545		 * If the commit is not on the reader page, then
2546		 * move the header page.
2547		 */
2548		if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2549			/*
2550			 * If we are not in overwrite mode,
2551			 * this is easy, just stop here.
2552			 */
2553			if (!(buffer->flags & RB_FL_OVERWRITE)) {
2554				local_inc(&cpu_buffer->dropped_events);
2555				goto out_reset;
2556			}
2557
2558			ret = rb_handle_head_page(cpu_buffer,
2559						  tail_page,
2560						  next_page);
2561			if (ret < 0)
2562				goto out_reset;
2563			if (ret)
2564				goto out_again;
2565		} else {
2566			/*
2567			 * We need to be careful here too. The
2568			 * commit page could still be on the reader
2569			 * page. We could have a small buffer, and
2570			 * have filled up the buffer with events
2571			 * from interrupts and such, and wrapped.
2572			 *
2573			 * Note, if the tail page is also the on the
2574			 * reader_page, we let it move out.
2575			 */
2576			if (unlikely((cpu_buffer->commit_page !=
2577				      cpu_buffer->tail_page) &&
2578				     (cpu_buffer->commit_page ==
2579				      cpu_buffer->reader_page))) {
2580				local_inc(&cpu_buffer->commit_overrun);
2581				goto out_reset;
2582			}
2583		}
2584	}
2585
2586	rb_tail_page_update(cpu_buffer, tail_page, next_page);
2587
2588 out_again:
2589
2590	rb_reset_tail(cpu_buffer, tail, info);
2591
2592	/* Commit what we have for now. */
2593	rb_end_commit(cpu_buffer);
2594	/* rb_end_commit() decs committing */
2595	local_inc(&cpu_buffer->committing);
2596
2597	/* fail and let the caller try again */
2598	return ERR_PTR(-EAGAIN);
2599
2600 out_reset:
2601	/* reset write */
2602	rb_reset_tail(cpu_buffer, tail, info);
2603
2604	return NULL;
2605}
2606
2607/* Slow path */
2608static struct ring_buffer_event *
2609rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2610{
2611	if (abs)
2612		event->type_len = RINGBUF_TYPE_TIME_STAMP;
2613	else
2614		event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2615
2616	/* Not the first event on the page, or not delta? */
2617	if (abs || rb_event_index(event)) {
2618		event->time_delta = delta & TS_MASK;
2619		event->array[0] = delta >> TS_SHIFT;
2620	} else {
2621		/* nope, just zero it */
2622		event->time_delta = 0;
2623		event->array[0] = 0;
2624	}
2625
2626	return skip_time_extend(event);
2627}
2628
2629static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2630				     struct ring_buffer_event *event);
2631
2632#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2633static inline bool sched_clock_stable(void)
2634{
2635	return true;
2636}
2637#endif
2638
2639static void
2640rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2641		   struct rb_event_info *info)
2642{
2643	u64 write_stamp;
2644
2645	WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2646		  (unsigned long long)info->delta,
2647		  (unsigned long long)info->ts,
2648		  (unsigned long long)info->before,
2649		  (unsigned long long)info->after,
2650		  (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2651		  sched_clock_stable() ? "" :
2652		  "If you just came from a suspend/resume,\n"
2653		  "please switch to the trace global clock:\n"
2654		  "  echo global > /sys/kernel/debug/tracing/trace_clock\n"
2655		  "or add trace_clock=global to the kernel command line\n");
2656}
2657
2658static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2659				      struct ring_buffer_event **event,
2660				      struct rb_event_info *info,
2661				      u64 *delta,
2662				      unsigned int *length)
2663{
2664	bool abs = info->add_timestamp &
2665		(RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2666
2667	if (unlikely(info->delta > (1ULL << 59))) {
2668		/* did the clock go backwards */
2669		if (info->before == info->after && info->before > info->ts) {
2670			/* not interrupted */
2671			static int once;
2672
2673			/*
2674			 * This is possible with a recalibrating of the TSC.
2675			 * Do not produce a call stack, but just report it.
2676			 */
2677			if (!once) {
2678				once++;
2679				pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2680					info->before, info->ts);
2681			}
2682		} else
2683			rb_check_timestamp(cpu_buffer, info);
2684		if (!abs)
2685			info->delta = 0;
2686	}
2687	*event = rb_add_time_stamp(*event, info->delta, abs);
2688	*length -= RB_LEN_TIME_EXTEND;
2689	*delta = 0;
2690}
2691
2692/**
2693 * rb_update_event - update event type and data
2694 * @cpu_buffer: The per cpu buffer of the @event
2695 * @event: the event to update
2696 * @info: The info to update the @event with (contains length and delta)
 
2697 *
2698 * Update the type and data fields of the @event. The length
2699 * is the actual size that is written to the ring buffer,
2700 * and with this, we can determine what to place into the
2701 * data field.
2702 */
2703static void
2704rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2705		struct ring_buffer_event *event,
2706		struct rb_event_info *info)
2707{
2708	unsigned length = info->length;
2709	u64 delta = info->delta;
2710
 
 
 
 
2711	/*
2712	 * If we need to add a timestamp, then we
2713	 * add it to the start of the reserved space.
2714	 */
2715	if (unlikely(info->add_timestamp))
2716		rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
 
 
 
2717
2718	event->time_delta = delta;
2719	length -= RB_EVNT_HDR_SIZE;
2720	if (length > RB_MAX_SMALL_DATA) {
2721		event->type_len = 0;
2722		event->array[0] = length;
2723	} else
2724		event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2725}
2726
2727static unsigned rb_calculate_event_length(unsigned length)
2728{
2729	struct ring_buffer_event event; /* Used only for sizeof array */
2730
2731	/* zero length can cause confusions */
2732	if (!length)
2733		length++;
2734
2735	if (length > RB_MAX_SMALL_DATA)
2736		length += sizeof(event.array[0]);
2737
2738	length += RB_EVNT_HDR_SIZE;
2739	length = ALIGN(length, RB_ALIGNMENT);
2740
2741	/*
2742	 * In case the time delta is larger than the 27 bits for it
2743	 * in the header, we need to add a timestamp. If another
2744	 * event comes in when trying to discard this one to increase
2745	 * the length, then the timestamp will be added in the allocated
2746	 * space of this event. If length is bigger than the size needed
2747	 * for the TIME_EXTEND, then padding has to be used. The events
2748	 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2749	 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2750	 * As length is a multiple of 4, we only need to worry if it
2751	 * is 12 (RB_LEN_TIME_EXTEND + 4).
2752	 */
2753	if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2754		length += RB_ALIGNMENT;
2755
2756	return length;
2757}
2758
2759static __always_inline bool
2760rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2761		   struct ring_buffer_event *event)
2762{
2763	unsigned long addr = (unsigned long)event;
2764	unsigned long index;
2765
2766	index = rb_event_index(event);
2767	addr &= PAGE_MASK;
2768
2769	return cpu_buffer->commit_page->page == (void *)addr &&
2770		rb_commit_index(cpu_buffer) == index;
2771}
2772
2773static u64 rb_time_delta(struct ring_buffer_event *event)
2774{
2775	switch (event->type_len) {
2776	case RINGBUF_TYPE_PADDING:
2777		return 0;
2778
2779	case RINGBUF_TYPE_TIME_EXTEND:
2780		return ring_buffer_event_time_stamp(event);
2781
2782	case RINGBUF_TYPE_TIME_STAMP:
2783		return 0;
2784
2785	case RINGBUF_TYPE_DATA:
2786		return event->time_delta;
2787	default:
2788		return 0;
2789	}
2790}
 
2791
2792static inline int
2793rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2794		  struct ring_buffer_event *event)
2795{
2796	unsigned long new_index, old_index;
2797	struct buffer_page *bpage;
2798	unsigned long index;
2799	unsigned long addr;
2800	u64 write_stamp;
2801	u64 delta;
2802
2803	new_index = rb_event_index(event);
2804	old_index = new_index + rb_event_ts_length(event);
2805	addr = (unsigned long)event;
2806	addr &= PAGE_MASK;
2807
2808	bpage = READ_ONCE(cpu_buffer->tail_page);
2809
2810	delta = rb_time_delta(event);
2811
2812	if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
2813		return 0;
2814
2815	/* Make sure the write stamp is read before testing the location */
2816	barrier();
2817
2818	if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2819		unsigned long write_mask =
2820			local_read(&bpage->write) & ~RB_WRITE_MASK;
2821		unsigned long event_length = rb_event_length(event);
2822
2823		/* Something came in, can't discard */
2824		if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
2825				       write_stamp, write_stamp - delta))
2826			return 0;
2827
2828		/*
2829		 * If an event were to come in now, it would see that the
2830		 * write_stamp and the before_stamp are different, and assume
2831		 * that this event just added itself before updating
2832		 * the write stamp. The interrupting event will fix the
2833		 * write stamp for us, and use the before stamp as its delta.
2834		 */
2835
2836		/*
2837		 * This is on the tail page. It is possible that
2838		 * a write could come in and move the tail page
2839		 * and write to the next page. That is fine
2840		 * because we just shorten what is on this page.
2841		 */
2842		old_index += write_mask;
2843		new_index += write_mask;
2844		index = local_cmpxchg(&bpage->write, old_index, new_index);
2845		if (index == old_index) {
2846			/* update counters */
2847			local_sub(event_length, &cpu_buffer->entries_bytes);
2848			return 1;
2849		}
2850	}
2851
2852	/* could not discard */
2853	return 0;
2854}
2855
2856static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2857{
2858	local_inc(&cpu_buffer->committing);
2859	local_inc(&cpu_buffer->commits);
2860}
2861
2862static __always_inline void
2863rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2864{
2865	unsigned long max_count;
2866
2867	/*
2868	 * We only race with interrupts and NMIs on this CPU.
2869	 * If we own the commit event, then we can commit
2870	 * all others that interrupted us, since the interruptions
2871	 * are in stack format (they finish before they come
2872	 * back to us). This allows us to do a simple loop to
2873	 * assign the commit to the tail.
2874	 */
2875 again:
2876	max_count = cpu_buffer->nr_pages * 100;
2877
2878	while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2879		if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2880			return;
2881		if (RB_WARN_ON(cpu_buffer,
2882			       rb_is_reader_page(cpu_buffer->tail_page)))
2883			return;
2884		local_set(&cpu_buffer->commit_page->page->commit,
2885			  rb_page_write(cpu_buffer->commit_page));
2886		rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
 
 
 
 
2887		/* add barrier to keep gcc from optimizing too much */
2888		barrier();
2889	}
2890	while (rb_commit_index(cpu_buffer) !=
2891	       rb_page_write(cpu_buffer->commit_page)) {
2892
2893		local_set(&cpu_buffer->commit_page->page->commit,
2894			  rb_page_write(cpu_buffer->commit_page));
2895		RB_WARN_ON(cpu_buffer,
2896			   local_read(&cpu_buffer->commit_page->page->commit) &
2897			   ~RB_WRITE_MASK);
2898		barrier();
2899	}
2900
2901	/* again, keep gcc from optimizing */
2902	barrier();
2903
2904	/*
2905	 * If an interrupt came in just after the first while loop
2906	 * and pushed the tail page forward, we will be left with
2907	 * a dangling commit that will never go forward.
2908	 */
2909	if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2910		goto again;
2911}
2912
2913static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2914{
2915	unsigned long commits;
2916
2917	if (RB_WARN_ON(cpu_buffer,
2918		       !local_read(&cpu_buffer->committing)))
2919		return;
2920
2921 again:
2922	commits = local_read(&cpu_buffer->commits);
2923	/* synchronize with interrupts */
2924	barrier();
2925	if (local_read(&cpu_buffer->committing) == 1)
2926		rb_set_commit_to_write(cpu_buffer);
2927
2928	local_dec(&cpu_buffer->committing);
2929
2930	/* synchronize with interrupts */
2931	barrier();
2932
2933	/*
2934	 * Need to account for interrupts coming in between the
2935	 * updating of the commit page and the clearing of the
2936	 * committing counter.
2937	 */
2938	if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2939	    !local_read(&cpu_buffer->committing)) {
2940		local_inc(&cpu_buffer->committing);
2941		goto again;
2942	}
2943}
2944
2945static inline void rb_event_discard(struct ring_buffer_event *event)
2946{
2947	if (extended_time(event))
2948		event = skip_time_extend(event);
2949
2950	/* array[0] holds the actual length for the discarded event */
2951	event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2952	event->type_len = RINGBUF_TYPE_PADDING;
2953	/* time delta must be non zero */
2954	if (!event->time_delta)
2955		event->time_delta = 1;
2956}
2957
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2958static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2959		      struct ring_buffer_event *event)
2960{
2961	local_inc(&cpu_buffer->entries);
 
2962	rb_end_commit(cpu_buffer);
2963}
2964
2965static __always_inline void
2966rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2967{
2968	size_t nr_pages;
2969	size_t dirty;
2970	size_t full;
2971
2972	if (buffer->irq_work.waiters_pending) {
2973		buffer->irq_work.waiters_pending = false;
2974		/* irq_work_queue() supplies it's own memory barriers */
2975		irq_work_queue(&buffer->irq_work.work);
2976	}
2977
2978	if (cpu_buffer->irq_work.waiters_pending) {
2979		cpu_buffer->irq_work.waiters_pending = false;
2980		/* irq_work_queue() supplies it's own memory barriers */
2981		irq_work_queue(&cpu_buffer->irq_work.work);
2982	}
2983
2984	if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
2985		return;
2986
2987	if (cpu_buffer->reader_page == cpu_buffer->commit_page)
2988		return;
2989
2990	if (!cpu_buffer->irq_work.full_waiters_pending)
2991		return;
2992
2993	cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
2994
2995	full = cpu_buffer->shortest_full;
2996	nr_pages = cpu_buffer->nr_pages;
2997	dirty = ring_buffer_nr_dirty_pages(buffer, cpu_buffer->cpu);
2998	if (full && nr_pages && (dirty * 100) <= full * nr_pages)
2999		return;
3000
3001	cpu_buffer->irq_work.wakeup_full = true;
3002	cpu_buffer->irq_work.full_waiters_pending = false;
3003	/* irq_work_queue() supplies it's own memory barriers */
3004	irq_work_queue(&cpu_buffer->irq_work.work);
3005}
3006
3007/*
3008 * The lock and unlock are done within a preempt disable section.
3009 * The current_context per_cpu variable can only be modified
3010 * by the current task between lock and unlock. But it can
3011 * be modified more than once via an interrupt. To pass this
3012 * information from the lock to the unlock without having to
3013 * access the 'in_interrupt()' functions again (which do show
3014 * a bit of overhead in something as critical as function tracing,
3015 * we use a bitmask trick.
3016 *
3017 *  bit 0 =  NMI context
3018 *  bit 1 =  IRQ context
3019 *  bit 2 =  SoftIRQ context
3020 *  bit 3 =  normal context.
3021 *
3022 * This works because this is the order of contexts that can
3023 * preempt other contexts. A SoftIRQ never preempts an IRQ
3024 * context.
3025 *
3026 * When the context is determined, the corresponding bit is
3027 * checked and set (if it was set, then a recursion of that context
3028 * happened).
3029 *
3030 * On unlock, we need to clear this bit. To do so, just subtract
3031 * 1 from the current_context and AND it to itself.
3032 *
3033 * (binary)
3034 *  101 - 1 = 100
3035 *  101 & 100 = 100 (clearing bit zero)
3036 *
3037 *  1010 - 1 = 1001
3038 *  1010 & 1001 = 1000 (clearing bit 1)
3039 *
3040 * The least significant bit can be cleared this way, and it
3041 * just so happens that it is the same bit corresponding to
3042 * the current context.
3043 */
3044
3045static __always_inline int
3046trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3047{
3048	unsigned int val = cpu_buffer->current_context;
3049	unsigned long pc = preempt_count();
3050	int bit;
3051
3052	if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
 
 
 
 
 
 
 
3053		bit = RB_CTX_NORMAL;
3054	else
3055		bit = pc & NMI_MASK ? RB_CTX_NMI :
3056			pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
3057
3058	if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
3059		return 1;
3060
3061	val |= (1 << (bit + cpu_buffer->nest));
3062	cpu_buffer->current_context = val;
3063
3064	return 0;
3065}
3066
3067static __always_inline void
3068trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3069{
3070	cpu_buffer->current_context &=
3071		cpu_buffer->current_context - (1 << cpu_buffer->nest);
3072}
3073
3074/* The recursive locking above uses 4 bits */
3075#define NESTED_BITS 4
3076
3077/**
3078 * ring_buffer_nest_start - Allow to trace while nested
3079 * @buffer: The ring buffer to modify
3080 *
3081 * The ring buffer has a safety mechanism to prevent recursion.
3082 * But there may be a case where a trace needs to be done while
3083 * tracing something else. In this case, calling this function
3084 * will allow this function to nest within a currently active
3085 * ring_buffer_lock_reserve().
3086 *
3087 * Call this function before calling another ring_buffer_lock_reserve() and
3088 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3089 */
3090void ring_buffer_nest_start(struct trace_buffer *buffer)
3091{
3092	struct ring_buffer_per_cpu *cpu_buffer;
3093	int cpu;
3094
3095	/* Enabled by ring_buffer_nest_end() */
3096	preempt_disable_notrace();
3097	cpu = raw_smp_processor_id();
3098	cpu_buffer = buffer->buffers[cpu];
3099	/* This is the shift value for the above recursive locking */
3100	cpu_buffer->nest += NESTED_BITS;
3101}
3102
3103/**
3104 * ring_buffer_nest_end - Allow to trace while nested
3105 * @buffer: The ring buffer to modify
3106 *
3107 * Must be called after ring_buffer_nest_start() and after the
3108 * ring_buffer_unlock_commit().
3109 */
3110void ring_buffer_nest_end(struct trace_buffer *buffer)
3111{
3112	struct ring_buffer_per_cpu *cpu_buffer;
3113	int cpu;
3114
3115	/* disabled by ring_buffer_nest_start() */
3116	cpu = raw_smp_processor_id();
3117	cpu_buffer = buffer->buffers[cpu];
3118	/* This is the shift value for the above recursive locking */
3119	cpu_buffer->nest -= NESTED_BITS;
3120	preempt_enable_notrace();
3121}
3122
3123/**
3124 * ring_buffer_unlock_commit - commit a reserved
3125 * @buffer: The buffer to commit to
3126 * @event: The event pointer to commit.
3127 *
3128 * This commits the data to the ring buffer, and releases any locks held.
3129 *
3130 * Must be paired with ring_buffer_lock_reserve.
3131 */
3132int ring_buffer_unlock_commit(struct trace_buffer *buffer,
3133			      struct ring_buffer_event *event)
3134{
3135	struct ring_buffer_per_cpu *cpu_buffer;
3136	int cpu = raw_smp_processor_id();
3137
3138	cpu_buffer = buffer->buffers[cpu];
3139
3140	rb_commit(cpu_buffer, event);
3141
3142	rb_wakeups(buffer, cpu_buffer);
3143
3144	trace_recursive_unlock(cpu_buffer);
3145
3146	preempt_enable_notrace();
3147
3148	return 0;
3149}
3150EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3151
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3152static struct ring_buffer_event *
3153__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3154		  struct rb_event_info *info)
3155{
3156	struct ring_buffer_event *event;
3157	struct buffer_page *tail_page;
3158	unsigned long tail, write, w;
3159	bool a_ok;
3160	bool b_ok;
 
 
 
 
 
 
3161
3162	/* Don't let the compiler play games with cpu_buffer->tail_page */
3163	tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3164
3165 /*A*/	w = local_read(&tail_page->write) & RB_WRITE_MASK;
3166	barrier();
3167	b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3168	a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3169	barrier();
3170	info->ts = rb_time_stamp(cpu_buffer->buffer);
3171
3172	if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3173		info->delta = info->ts;
3174	} else {
3175		/*
3176		 * If interrupting an event time update, we may need an
3177		 * absolute timestamp.
3178		 * Don't bother if this is the start of a new page (w == 0).
3179		 */
3180		if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
3181			info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3182			info->length += RB_LEN_TIME_EXTEND;
3183		} else {
3184			info->delta = info->ts - info->after;
3185			if (unlikely(test_time_stamp(info->delta))) {
3186				info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3187				info->length += RB_LEN_TIME_EXTEND;
3188			}
3189		}
3190	}
3191
3192 /*B*/	rb_time_set(&cpu_buffer->before_stamp, info->ts);
3193
3194 /*C*/	write = local_add_return(info->length, &tail_page->write);
3195
3196	/* set write to only the index of the write */
3197	write &= RB_WRITE_MASK;
3198
3199	tail = write - info->length;
3200
3201	/* See if we shot pass the end of this buffer page */
3202	if (unlikely(write > BUF_PAGE_SIZE)) {
3203		if (tail != w) {
3204			/* before and after may now different, fix it up*/
3205			b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3206			a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3207			if (a_ok && b_ok && info->before != info->after)
3208				(void)rb_time_cmpxchg(&cpu_buffer->before_stamp,
3209						      info->before, info->after);
3210		}
3211		return rb_move_tail(cpu_buffer, tail, info);
3212	}
3213
3214	if (likely(tail == w)) {
3215		u64 save_before;
3216		bool s_ok;
3217
3218		/* Nothing interrupted us between A and C */
3219 /*D*/		rb_time_set(&cpu_buffer->write_stamp, info->ts);
3220		barrier();
3221 /*E*/		s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
3222		RB_WARN_ON(cpu_buffer, !s_ok);
3223		if (likely(!(info->add_timestamp &
3224			     (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3225			/* This did not interrupt any time update */
3226			info->delta = info->ts - info->after;
3227		else
3228			/* Just use full timestamp for inerrupting event */
3229			info->delta = info->ts;
3230		barrier();
3231		if (unlikely(info->ts != save_before)) {
3232			/* SLOW PATH - Interrupted between C and E */
3233
3234			a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3235			RB_WARN_ON(cpu_buffer, !a_ok);
3236
3237			/* Write stamp must only go forward */
3238			if (save_before > info->after) {
3239				/*
3240				 * We do not care about the result, only that
3241				 * it gets updated atomically.
3242				 */
3243				(void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3244						      info->after, save_before);
3245			}
3246		}
3247	} else {
3248		u64 ts;
3249		/* SLOW PATH - Interrupted between A and C */
3250		a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3251		/* Was interrupted before here, write_stamp must be valid */
3252		RB_WARN_ON(cpu_buffer, !a_ok);
3253		ts = rb_time_stamp(cpu_buffer->buffer);
3254		barrier();
3255 /*E*/		if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3256		    info->after < ts) {
3257			/* Nothing came after this event between C and E */
3258			info->delta = ts - info->after;
3259			(void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3260					      info->after, info->ts);
3261			info->ts = ts;
3262		} else {
3263			/*
3264			 * Interrupted beween C and E:
3265			 * Lost the previous events time stamp. Just set the
3266			 * delta to zero, and this will be the same time as
3267			 * the event this event interrupted. And the events that
3268			 * came after this will still be correct (as they would
3269			 * have built their delta on the previous event.
3270			 */
3271			info->delta = 0;
3272		}
3273		info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3274	}
3275
3276	/*
3277	 * If this is the first commit on the page, then it has the same
3278	 * timestamp as the page itself.
3279	 */
3280	if (unlikely(!tail && !(info->add_timestamp &
3281				(RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3282		info->delta = 0;
3283
 
 
 
 
3284	/* We reserved something on the buffer */
3285
3286	event = __rb_page_index(tail_page, tail);
 
3287	rb_update_event(cpu_buffer, event, info);
3288
3289	local_inc(&tail_page->entries);
3290
3291	/*
3292	 * If this is the first commit on the page, then update
3293	 * its timestamp.
3294	 */
3295	if (unlikely(!tail))
3296		tail_page->page->time_stamp = info->ts;
3297
3298	/* account for these added bytes */
3299	local_add(info->length, &cpu_buffer->entries_bytes);
3300
3301	return event;
3302}
3303
3304static __always_inline struct ring_buffer_event *
3305rb_reserve_next_event(struct trace_buffer *buffer,
3306		      struct ring_buffer_per_cpu *cpu_buffer,
3307		      unsigned long length)
3308{
3309	struct ring_buffer_event *event;
3310	struct rb_event_info info;
3311	int nr_loops = 0;
3312	int add_ts_default;
3313
3314	rb_start_commit(cpu_buffer);
3315	/* The commit page can not change after this */
3316
3317#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3318	/*
3319	 * Due to the ability to swap a cpu buffer from a buffer
3320	 * it is possible it was swapped before we committed.
3321	 * (committing stops a swap). We check for it here and
3322	 * if it happened, we have to fail the write.
3323	 */
3324	barrier();
3325	if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3326		local_dec(&cpu_buffer->committing);
3327		local_dec(&cpu_buffer->commits);
3328		return NULL;
3329	}
3330#endif
3331
3332	info.length = rb_calculate_event_length(length);
3333
3334	if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3335		add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3336		info.length += RB_LEN_TIME_EXTEND;
3337	} else {
3338		add_ts_default = RB_ADD_STAMP_NONE;
3339	}
3340
3341 again:
3342	info.add_timestamp = add_ts_default;
3343	info.delta = 0;
3344
3345	/*
3346	 * We allow for interrupts to reenter here and do a trace.
3347	 * If one does, it will cause this original code to loop
3348	 * back here. Even with heavy interrupts happening, this
3349	 * should only happen a few times in a row. If this happens
3350	 * 1000 times in a row, there must be either an interrupt
3351	 * storm or we have something buggy.
3352	 * Bail!
3353	 */
3354	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3355		goto out_fail;
3356
 
 
 
 
 
 
 
 
 
 
 
 
 
3357	event = __rb_reserve_next(cpu_buffer, &info);
3358
3359	if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3360		if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3361			info.length -= RB_LEN_TIME_EXTEND;
3362		goto again;
3363	}
3364
3365	if (likely(event))
3366		return event;
 
 
 
3367 out_fail:
3368	rb_end_commit(cpu_buffer);
3369	return NULL;
3370}
3371
3372/**
3373 * ring_buffer_lock_reserve - reserve a part of the buffer
3374 * @buffer: the ring buffer to reserve from
3375 * @length: the length of the data to reserve (excluding event header)
3376 *
3377 * Returns a reserved event on the ring buffer to copy directly to.
3378 * The user of this interface will need to get the body to write into
3379 * and can use the ring_buffer_event_data() interface.
3380 *
3381 * The length is the length of the data needed, not the event length
3382 * which also includes the event header.
3383 *
3384 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3385 * If NULL is returned, then nothing has been allocated or locked.
3386 */
3387struct ring_buffer_event *
3388ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3389{
3390	struct ring_buffer_per_cpu *cpu_buffer;
3391	struct ring_buffer_event *event;
3392	int cpu;
3393
3394	/* If we are tracing schedule, we don't want to recurse */
3395	preempt_disable_notrace();
3396
3397	if (unlikely(atomic_read(&buffer->record_disabled)))
3398		goto out;
3399
3400	cpu = raw_smp_processor_id();
3401
3402	if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3403		goto out;
3404
3405	cpu_buffer = buffer->buffers[cpu];
3406
3407	if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3408		goto out;
3409
3410	if (unlikely(length > BUF_MAX_DATA_SIZE))
3411		goto out;
3412
3413	if (unlikely(trace_recursive_lock(cpu_buffer)))
3414		goto out;
3415
3416	event = rb_reserve_next_event(buffer, cpu_buffer, length);
3417	if (!event)
3418		goto out_unlock;
3419
3420	return event;
3421
3422 out_unlock:
3423	trace_recursive_unlock(cpu_buffer);
3424 out:
3425	preempt_enable_notrace();
3426	return NULL;
3427}
3428EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3429
3430/*
3431 * Decrement the entries to the page that an event is on.
3432 * The event does not even need to exist, only the pointer
3433 * to the page it is on. This may only be called before the commit
3434 * takes place.
3435 */
3436static inline void
3437rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3438		   struct ring_buffer_event *event)
3439{
3440	unsigned long addr = (unsigned long)event;
3441	struct buffer_page *bpage = cpu_buffer->commit_page;
3442	struct buffer_page *start;
3443
3444	addr &= PAGE_MASK;
3445
3446	/* Do the likely case first */
3447	if (likely(bpage->page == (void *)addr)) {
3448		local_dec(&bpage->entries);
3449		return;
3450	}
3451
3452	/*
3453	 * Because the commit page may be on the reader page we
3454	 * start with the next page and check the end loop there.
3455	 */
3456	rb_inc_page(cpu_buffer, &bpage);
3457	start = bpage;
3458	do {
3459		if (bpage->page == (void *)addr) {
3460			local_dec(&bpage->entries);
3461			return;
3462		}
3463		rb_inc_page(cpu_buffer, &bpage);
3464	} while (bpage != start);
3465
3466	/* commit not part of this buffer?? */
3467	RB_WARN_ON(cpu_buffer, 1);
3468}
3469
3470/**
3471 * ring_buffer_commit_discard - discard an event that has not been committed
3472 * @buffer: the ring buffer
3473 * @event: non committed event to discard
3474 *
3475 * Sometimes an event that is in the ring buffer needs to be ignored.
3476 * This function lets the user discard an event in the ring buffer
3477 * and then that event will not be read later.
3478 *
3479 * This function only works if it is called before the item has been
3480 * committed. It will try to free the event from the ring buffer
3481 * if another event has not been added behind it.
3482 *
3483 * If another event has been added behind it, it will set the event
3484 * up as discarded, and perform the commit.
3485 *
3486 * If this function is called, do not call ring_buffer_unlock_commit on
3487 * the event.
3488 */
3489void ring_buffer_discard_commit(struct trace_buffer *buffer,
3490				struct ring_buffer_event *event)
3491{
3492	struct ring_buffer_per_cpu *cpu_buffer;
3493	int cpu;
3494
3495	/* The event is discarded regardless */
3496	rb_event_discard(event);
3497
3498	cpu = smp_processor_id();
3499	cpu_buffer = buffer->buffers[cpu];
3500
3501	/*
3502	 * This must only be called if the event has not been
3503	 * committed yet. Thus we can assume that preemption
3504	 * is still disabled.
3505	 */
3506	RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3507
3508	rb_decrement_entry(cpu_buffer, event);
3509	if (rb_try_to_discard(cpu_buffer, event))
3510		goto out;
3511
 
 
 
 
 
3512 out:
3513	rb_end_commit(cpu_buffer);
3514
3515	trace_recursive_unlock(cpu_buffer);
3516
3517	preempt_enable_notrace();
3518
3519}
3520EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3521
3522/**
3523 * ring_buffer_write - write data to the buffer without reserving
3524 * @buffer: The ring buffer to write to.
3525 * @length: The length of the data being written (excluding the event header)
3526 * @data: The data to write to the buffer.
3527 *
3528 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3529 * one function. If you already have the data to write to the buffer, it
3530 * may be easier to simply call this function.
3531 *
3532 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3533 * and not the length of the event which would hold the header.
3534 */
3535int ring_buffer_write(struct trace_buffer *buffer,
3536		      unsigned long length,
3537		      void *data)
3538{
3539	struct ring_buffer_per_cpu *cpu_buffer;
3540	struct ring_buffer_event *event;
3541	void *body;
3542	int ret = -EBUSY;
3543	int cpu;
3544
3545	preempt_disable_notrace();
3546
3547	if (atomic_read(&buffer->record_disabled))
3548		goto out;
3549
3550	cpu = raw_smp_processor_id();
3551
3552	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3553		goto out;
3554
3555	cpu_buffer = buffer->buffers[cpu];
3556
3557	if (atomic_read(&cpu_buffer->record_disabled))
3558		goto out;
3559
3560	if (length > BUF_MAX_DATA_SIZE)
3561		goto out;
3562
3563	if (unlikely(trace_recursive_lock(cpu_buffer)))
3564		goto out;
3565
3566	event = rb_reserve_next_event(buffer, cpu_buffer, length);
3567	if (!event)
3568		goto out_unlock;
3569
3570	body = rb_event_data(event);
3571
3572	memcpy(body, data, length);
3573
3574	rb_commit(cpu_buffer, event);
3575
3576	rb_wakeups(buffer, cpu_buffer);
3577
3578	ret = 0;
3579
3580 out_unlock:
3581	trace_recursive_unlock(cpu_buffer);
3582
3583 out:
3584	preempt_enable_notrace();
3585
3586	return ret;
3587}
3588EXPORT_SYMBOL_GPL(ring_buffer_write);
3589
3590static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3591{
3592	struct buffer_page *reader = cpu_buffer->reader_page;
3593	struct buffer_page *head = rb_set_head_page(cpu_buffer);
3594	struct buffer_page *commit = cpu_buffer->commit_page;
3595
3596	/* In case of error, head will be NULL */
3597	if (unlikely(!head))
3598		return true;
3599
3600	return reader->read == rb_page_commit(reader) &&
3601		(commit == reader ||
3602		 (commit == head &&
3603		  head->read == rb_page_commit(commit)));
3604}
3605
3606/**
3607 * ring_buffer_record_disable - stop all writes into the buffer
3608 * @buffer: The ring buffer to stop writes to.
3609 *
3610 * This prevents all writes to the buffer. Any attempt to write
3611 * to the buffer after this will fail and return NULL.
3612 *
3613 * The caller should call synchronize_rcu() after this.
3614 */
3615void ring_buffer_record_disable(struct trace_buffer *buffer)
3616{
3617	atomic_inc(&buffer->record_disabled);
3618}
3619EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3620
3621/**
3622 * ring_buffer_record_enable - enable writes to the buffer
3623 * @buffer: The ring buffer to enable writes
3624 *
3625 * Note, multiple disables will need the same number of enables
3626 * to truly enable the writing (much like preempt_disable).
3627 */
3628void ring_buffer_record_enable(struct trace_buffer *buffer)
3629{
3630	atomic_dec(&buffer->record_disabled);
3631}
3632EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3633
3634/**
3635 * ring_buffer_record_off - stop all writes into the buffer
3636 * @buffer: The ring buffer to stop writes to.
3637 *
3638 * This prevents all writes to the buffer. Any attempt to write
3639 * to the buffer after this will fail and return NULL.
3640 *
3641 * This is different than ring_buffer_record_disable() as
3642 * it works like an on/off switch, where as the disable() version
3643 * must be paired with a enable().
3644 */
3645void ring_buffer_record_off(struct trace_buffer *buffer)
3646{
3647	unsigned int rd;
3648	unsigned int new_rd;
3649
3650	do {
3651		rd = atomic_read(&buffer->record_disabled);
3652		new_rd = rd | RB_BUFFER_OFF;
3653	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3654}
3655EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3656
3657/**
3658 * ring_buffer_record_on - restart writes into the buffer
3659 * @buffer: The ring buffer to start writes to.
3660 *
3661 * This enables all writes to the buffer that was disabled by
3662 * ring_buffer_record_off().
3663 *
3664 * This is different than ring_buffer_record_enable() as
3665 * it works like an on/off switch, where as the enable() version
3666 * must be paired with a disable().
3667 */
3668void ring_buffer_record_on(struct trace_buffer *buffer)
3669{
3670	unsigned int rd;
3671	unsigned int new_rd;
3672
3673	do {
3674		rd = atomic_read(&buffer->record_disabled);
3675		new_rd = rd & ~RB_BUFFER_OFF;
3676	} while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3677}
3678EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3679
3680/**
3681 * ring_buffer_record_is_on - return true if the ring buffer can write
3682 * @buffer: The ring buffer to see if write is enabled
3683 *
3684 * Returns true if the ring buffer is in a state that it accepts writes.
3685 */
3686bool ring_buffer_record_is_on(struct trace_buffer *buffer)
3687{
3688	return !atomic_read(&buffer->record_disabled);
3689}
3690
3691/**
3692 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3693 * @buffer: The ring buffer to see if write is set enabled
3694 *
3695 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3696 * Note that this does NOT mean it is in a writable state.
3697 *
3698 * It may return true when the ring buffer has been disabled by
3699 * ring_buffer_record_disable(), as that is a temporary disabling of
3700 * the ring buffer.
3701 */
3702bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
3703{
3704	return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3705}
3706
3707/**
3708 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3709 * @buffer: The ring buffer to stop writes to.
3710 * @cpu: The CPU buffer to stop
3711 *
3712 * This prevents all writes to the buffer. Any attempt to write
3713 * to the buffer after this will fail and return NULL.
3714 *
3715 * The caller should call synchronize_rcu() after this.
3716 */
3717void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
3718{
3719	struct ring_buffer_per_cpu *cpu_buffer;
3720
3721	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3722		return;
3723
3724	cpu_buffer = buffer->buffers[cpu];
3725	atomic_inc(&cpu_buffer->record_disabled);
3726}
3727EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3728
3729/**
3730 * ring_buffer_record_enable_cpu - enable writes to the buffer
3731 * @buffer: The ring buffer to enable writes
3732 * @cpu: The CPU to enable.
3733 *
3734 * Note, multiple disables will need the same number of enables
3735 * to truly enable the writing (much like preempt_disable).
3736 */
3737void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
3738{
3739	struct ring_buffer_per_cpu *cpu_buffer;
3740
3741	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3742		return;
3743
3744	cpu_buffer = buffer->buffers[cpu];
3745	atomic_dec(&cpu_buffer->record_disabled);
3746}
3747EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3748
3749/*
3750 * The total entries in the ring buffer is the running counter
3751 * of entries entered into the ring buffer, minus the sum of
3752 * the entries read from the ring buffer and the number of
3753 * entries that were overwritten.
3754 */
3755static inline unsigned long
3756rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3757{
3758	return local_read(&cpu_buffer->entries) -
3759		(local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3760}
3761
3762/**
3763 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3764 * @buffer: The ring buffer
3765 * @cpu: The per CPU buffer to read from.
3766 */
3767u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
3768{
3769	unsigned long flags;
3770	struct ring_buffer_per_cpu *cpu_buffer;
3771	struct buffer_page *bpage;
3772	u64 ret = 0;
3773
3774	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3775		return 0;
3776
3777	cpu_buffer = buffer->buffers[cpu];
3778	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3779	/*
3780	 * if the tail is on reader_page, oldest time stamp is on the reader
3781	 * page
3782	 */
3783	if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3784		bpage = cpu_buffer->reader_page;
3785	else
3786		bpage = rb_set_head_page(cpu_buffer);
3787	if (bpage)
3788		ret = bpage->page->time_stamp;
3789	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3790
3791	return ret;
3792}
3793EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3794
3795/**
3796 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3797 * @buffer: The ring buffer
3798 * @cpu: The per CPU buffer to read from.
3799 */
3800unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
3801{
3802	struct ring_buffer_per_cpu *cpu_buffer;
3803	unsigned long ret;
3804
3805	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3806		return 0;
3807
3808	cpu_buffer = buffer->buffers[cpu];
3809	ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3810
3811	return ret;
3812}
3813EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3814
3815/**
3816 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3817 * @buffer: The ring buffer
3818 * @cpu: The per CPU buffer to get the entries from.
3819 */
3820unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
3821{
3822	struct ring_buffer_per_cpu *cpu_buffer;
3823
3824	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3825		return 0;
3826
3827	cpu_buffer = buffer->buffers[cpu];
3828
3829	return rb_num_of_entries(cpu_buffer);
3830}
3831EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3832
3833/**
3834 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3835 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3836 * @buffer: The ring buffer
3837 * @cpu: The per CPU buffer to get the number of overruns from
3838 */
3839unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
3840{
3841	struct ring_buffer_per_cpu *cpu_buffer;
3842	unsigned long ret;
3843
3844	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3845		return 0;
3846
3847	cpu_buffer = buffer->buffers[cpu];
3848	ret = local_read(&cpu_buffer->overrun);
3849
3850	return ret;
3851}
3852EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3853
3854/**
3855 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3856 * commits failing due to the buffer wrapping around while there are uncommitted
3857 * events, such as during an interrupt storm.
3858 * @buffer: The ring buffer
3859 * @cpu: The per CPU buffer to get the number of overruns from
3860 */
3861unsigned long
3862ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
3863{
3864	struct ring_buffer_per_cpu *cpu_buffer;
3865	unsigned long ret;
3866
3867	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3868		return 0;
3869
3870	cpu_buffer = buffer->buffers[cpu];
3871	ret = local_read(&cpu_buffer->commit_overrun);
3872
3873	return ret;
3874}
3875EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3876
3877/**
3878 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3879 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3880 * @buffer: The ring buffer
3881 * @cpu: The per CPU buffer to get the number of overruns from
3882 */
3883unsigned long
3884ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
3885{
3886	struct ring_buffer_per_cpu *cpu_buffer;
3887	unsigned long ret;
3888
3889	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3890		return 0;
3891
3892	cpu_buffer = buffer->buffers[cpu];
3893	ret = local_read(&cpu_buffer->dropped_events);
3894
3895	return ret;
3896}
3897EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3898
3899/**
3900 * ring_buffer_read_events_cpu - get the number of events successfully read
3901 * @buffer: The ring buffer
3902 * @cpu: The per CPU buffer to get the number of events read
3903 */
3904unsigned long
3905ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
3906{
3907	struct ring_buffer_per_cpu *cpu_buffer;
3908
3909	if (!cpumask_test_cpu(cpu, buffer->cpumask))
3910		return 0;
3911
3912	cpu_buffer = buffer->buffers[cpu];
3913	return cpu_buffer->read;
3914}
3915EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3916
3917/**
3918 * ring_buffer_entries - get the number of entries in a buffer
3919 * @buffer: The ring buffer
3920 *
3921 * Returns the total number of entries in the ring buffer
3922 * (all CPU entries)
3923 */
3924unsigned long ring_buffer_entries(struct trace_buffer *buffer)
3925{
3926	struct ring_buffer_per_cpu *cpu_buffer;
3927	unsigned long entries = 0;
3928	int cpu;
3929
3930	/* if you care about this being correct, lock the buffer */
3931	for_each_buffer_cpu(buffer, cpu) {
3932		cpu_buffer = buffer->buffers[cpu];
3933		entries += rb_num_of_entries(cpu_buffer);
3934	}
3935
3936	return entries;
3937}
3938EXPORT_SYMBOL_GPL(ring_buffer_entries);
3939
3940/**
3941 * ring_buffer_overruns - get the number of overruns in buffer
3942 * @buffer: The ring buffer
3943 *
3944 * Returns the total number of overruns in the ring buffer
3945 * (all CPU entries)
3946 */
3947unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
3948{
3949	struct ring_buffer_per_cpu *cpu_buffer;
3950	unsigned long overruns = 0;
3951	int cpu;
3952
3953	/* if you care about this being correct, lock the buffer */
3954	for_each_buffer_cpu(buffer, cpu) {
3955		cpu_buffer = buffer->buffers[cpu];
3956		overruns += local_read(&cpu_buffer->overrun);
3957	}
3958
3959	return overruns;
3960}
3961EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3962
3963static void rb_iter_reset(struct ring_buffer_iter *iter)
3964{
3965	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3966
3967	/* Iterator usage is expected to have record disabled */
3968	iter->head_page = cpu_buffer->reader_page;
3969	iter->head = cpu_buffer->reader_page->read;
3970	iter->next_event = iter->head;
3971
3972	iter->cache_reader_page = iter->head_page;
3973	iter->cache_read = cpu_buffer->read;
3974
3975	if (iter->head) {
3976		iter->read_stamp = cpu_buffer->read_stamp;
3977		iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
3978	} else {
3979		iter->read_stamp = iter->head_page->page->time_stamp;
3980		iter->page_stamp = iter->read_stamp;
3981	}
3982}
3983
3984/**
3985 * ring_buffer_iter_reset - reset an iterator
3986 * @iter: The iterator to reset
3987 *
3988 * Resets the iterator, so that it will start from the beginning
3989 * again.
3990 */
3991void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3992{
3993	struct ring_buffer_per_cpu *cpu_buffer;
3994	unsigned long flags;
3995
3996	if (!iter)
3997		return;
3998
3999	cpu_buffer = iter->cpu_buffer;
4000
4001	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4002	rb_iter_reset(iter);
4003	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4004}
4005EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4006
4007/**
4008 * ring_buffer_iter_empty - check if an iterator has no more to read
4009 * @iter: The iterator to check
4010 */
4011int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4012{
4013	struct ring_buffer_per_cpu *cpu_buffer;
4014	struct buffer_page *reader;
4015	struct buffer_page *head_page;
4016	struct buffer_page *commit_page;
4017	struct buffer_page *curr_commit_page;
4018	unsigned commit;
4019	u64 curr_commit_ts;
4020	u64 commit_ts;
4021
4022	cpu_buffer = iter->cpu_buffer;
4023	reader = cpu_buffer->reader_page;
4024	head_page = cpu_buffer->head_page;
4025	commit_page = cpu_buffer->commit_page;
4026	commit_ts = commit_page->page->time_stamp;
4027
4028	/*
4029	 * When the writer goes across pages, it issues a cmpxchg which
4030	 * is a mb(), which will synchronize with the rmb here.
4031	 * (see rb_tail_page_update())
4032	 */
4033	smp_rmb();
4034	commit = rb_page_commit(commit_page);
4035	/* We want to make sure that the commit page doesn't change */
4036	smp_rmb();
4037
4038	/* Make sure commit page didn't change */
4039	curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4040	curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4041
4042	/* If the commit page changed, then there's more data */
4043	if (curr_commit_page != commit_page ||
4044	    curr_commit_ts != commit_ts)
4045		return 0;
4046
4047	/* Still racy, as it may return a false positive, but that's OK */
4048	return ((iter->head_page == commit_page && iter->head >= commit) ||
4049		(iter->head_page == reader && commit_page == head_page &&
4050		 head_page->read == commit &&
4051		 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4052}
4053EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4054
4055static void
4056rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4057		     struct ring_buffer_event *event)
4058{
4059	u64 delta;
4060
4061	switch (event->type_len) {
4062	case RINGBUF_TYPE_PADDING:
4063		return;
4064
4065	case RINGBUF_TYPE_TIME_EXTEND:
4066		delta = ring_buffer_event_time_stamp(event);
 
 
4067		cpu_buffer->read_stamp += delta;
4068		return;
4069
4070	case RINGBUF_TYPE_TIME_STAMP:
4071		delta = ring_buffer_event_time_stamp(event);
4072		cpu_buffer->read_stamp = delta;
4073		return;
4074
4075	case RINGBUF_TYPE_DATA:
4076		cpu_buffer->read_stamp += event->time_delta;
4077		return;
4078
4079	default:
4080		RB_WARN_ON(cpu_buffer, 1);
4081	}
4082	return;
4083}
4084
4085static void
4086rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4087			  struct ring_buffer_event *event)
4088{
4089	u64 delta;
4090
4091	switch (event->type_len) {
4092	case RINGBUF_TYPE_PADDING:
4093		return;
4094
4095	case RINGBUF_TYPE_TIME_EXTEND:
4096		delta = ring_buffer_event_time_stamp(event);
 
 
4097		iter->read_stamp += delta;
4098		return;
4099
4100	case RINGBUF_TYPE_TIME_STAMP:
4101		delta = ring_buffer_event_time_stamp(event);
4102		iter->read_stamp = delta;
4103		return;
4104
4105	case RINGBUF_TYPE_DATA:
4106		iter->read_stamp += event->time_delta;
4107		return;
4108
4109	default:
4110		RB_WARN_ON(iter->cpu_buffer, 1);
4111	}
4112	return;
4113}
4114
4115static struct buffer_page *
4116rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4117{
4118	struct buffer_page *reader = NULL;
4119	unsigned long overwrite;
4120	unsigned long flags;
4121	int nr_loops = 0;
4122	int ret;
4123
4124	local_irq_save(flags);
4125	arch_spin_lock(&cpu_buffer->lock);
4126
4127 again:
4128	/*
4129	 * This should normally only loop twice. But because the
4130	 * start of the reader inserts an empty page, it causes
4131	 * a case where we will loop three times. There should be no
4132	 * reason to loop four times (that I know of).
4133	 */
4134	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4135		reader = NULL;
4136		goto out;
4137	}
4138
4139	reader = cpu_buffer->reader_page;
4140
4141	/* If there's more to read, return this page */
4142	if (cpu_buffer->reader_page->read < rb_page_size(reader))
4143		goto out;
4144
4145	/* Never should we have an index greater than the size */
4146	if (RB_WARN_ON(cpu_buffer,
4147		       cpu_buffer->reader_page->read > rb_page_size(reader)))
4148		goto out;
4149
4150	/* check if we caught up to the tail */
4151	reader = NULL;
4152	if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4153		goto out;
4154
4155	/* Don't bother swapping if the ring buffer is empty */
4156	if (rb_num_of_entries(cpu_buffer) == 0)
4157		goto out;
4158
4159	/*
4160	 * Reset the reader page to size zero.
4161	 */
4162	local_set(&cpu_buffer->reader_page->write, 0);
4163	local_set(&cpu_buffer->reader_page->entries, 0);
4164	local_set(&cpu_buffer->reader_page->page->commit, 0);
4165	cpu_buffer->reader_page->real_end = 0;
4166
4167 spin:
4168	/*
4169	 * Splice the empty reader page into the list around the head.
4170	 */
4171	reader = rb_set_head_page(cpu_buffer);
4172	if (!reader)
4173		goto out;
4174	cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4175	cpu_buffer->reader_page->list.prev = reader->list.prev;
4176
4177	/*
4178	 * cpu_buffer->pages just needs to point to the buffer, it
4179	 *  has no specific buffer page to point to. Lets move it out
4180	 *  of our way so we don't accidentally swap it.
4181	 */
4182	cpu_buffer->pages = reader->list.prev;
4183
4184	/* The reader page will be pointing to the new head */
4185	rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
4186
4187	/*
4188	 * We want to make sure we read the overruns after we set up our
4189	 * pointers to the next object. The writer side does a
4190	 * cmpxchg to cross pages which acts as the mb on the writer
4191	 * side. Note, the reader will constantly fail the swap
4192	 * while the writer is updating the pointers, so this
4193	 * guarantees that the overwrite recorded here is the one we
4194	 * want to compare with the last_overrun.
4195	 */
4196	smp_mb();
4197	overwrite = local_read(&(cpu_buffer->overrun));
4198
4199	/*
4200	 * Here's the tricky part.
4201	 *
4202	 * We need to move the pointer past the header page.
4203	 * But we can only do that if a writer is not currently
4204	 * moving it. The page before the header page has the
4205	 * flag bit '1' set if it is pointing to the page we want.
4206	 * but if the writer is in the process of moving it
4207	 * than it will be '2' or already moved '0'.
4208	 */
4209
4210	ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4211
4212	/*
4213	 * If we did not convert it, then we must try again.
4214	 */
4215	if (!ret)
4216		goto spin;
4217
4218	/*
4219	 * Yay! We succeeded in replacing the page.
4220	 *
4221	 * Now make the new head point back to the reader page.
4222	 */
4223	rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4224	rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
4225
4226	local_inc(&cpu_buffer->pages_read);
4227
4228	/* Finally update the reader page to the new head */
4229	cpu_buffer->reader_page = reader;
4230	cpu_buffer->reader_page->read = 0;
4231
4232	if (overwrite != cpu_buffer->last_overrun) {
4233		cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4234		cpu_buffer->last_overrun = overwrite;
4235	}
4236
4237	goto again;
4238
4239 out:
4240	/* Update the read_stamp on the first event */
4241	if (reader && reader->read == 0)
4242		cpu_buffer->read_stamp = reader->page->time_stamp;
4243
4244	arch_spin_unlock(&cpu_buffer->lock);
4245	local_irq_restore(flags);
4246
4247	return reader;
4248}
4249
4250static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4251{
4252	struct ring_buffer_event *event;
4253	struct buffer_page *reader;
4254	unsigned length;
4255
4256	reader = rb_get_reader_page(cpu_buffer);
4257
4258	/* This function should not be called when buffer is empty */
4259	if (RB_WARN_ON(cpu_buffer, !reader))
4260		return;
4261
4262	event = rb_reader_event(cpu_buffer);
4263
4264	if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4265		cpu_buffer->read++;
4266
4267	rb_update_read_stamp(cpu_buffer, event);
4268
4269	length = rb_event_length(event);
4270	cpu_buffer->reader_page->read += length;
4271}
4272
4273static void rb_advance_iter(struct ring_buffer_iter *iter)
4274{
4275	struct ring_buffer_per_cpu *cpu_buffer;
 
 
4276
4277	cpu_buffer = iter->cpu_buffer;
4278
4279	/* If head == next_event then we need to jump to the next event */
4280	if (iter->head == iter->next_event) {
4281		/* If the event gets overwritten again, there's nothing to do */
4282		if (rb_iter_head_event(iter) == NULL)
4283			return;
4284	}
4285
4286	iter->head = iter->next_event;
4287
4288	/*
4289	 * Check if we are at the end of the buffer.
4290	 */
4291	if (iter->next_event >= rb_page_size(iter->head_page)) {
4292		/* discarded commits can make the page empty */
4293		if (iter->head_page == cpu_buffer->commit_page)
4294			return;
4295		rb_inc_iter(iter);
4296		return;
4297	}
4298
4299	rb_update_iter_read_stamp(iter, iter->event);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4300}
4301
4302static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4303{
4304	return cpu_buffer->lost_events;
4305}
4306
4307static struct ring_buffer_event *
4308rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4309	       unsigned long *lost_events)
4310{
4311	struct ring_buffer_event *event;
4312	struct buffer_page *reader;
4313	int nr_loops = 0;
4314
4315	if (ts)
4316		*ts = 0;
4317 again:
4318	/*
4319	 * We repeat when a time extend is encountered.
4320	 * Since the time extend is always attached to a data event,
4321	 * we should never loop more than once.
4322	 * (We never hit the following condition more than twice).
4323	 */
4324	if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4325		return NULL;
4326
4327	reader = rb_get_reader_page(cpu_buffer);
4328	if (!reader)
4329		return NULL;
4330
4331	event = rb_reader_event(cpu_buffer);
4332
4333	switch (event->type_len) {
4334	case RINGBUF_TYPE_PADDING:
4335		if (rb_null_event(event))
4336			RB_WARN_ON(cpu_buffer, 1);
4337		/*
4338		 * Because the writer could be discarding every
4339		 * event it creates (which would probably be bad)
4340		 * if we were to go back to "again" then we may never
4341		 * catch up, and will trigger the warn on, or lock
4342		 * the box. Return the padding, and we will release
4343		 * the current locks, and try again.
4344		 */
4345		return event;
4346
4347	case RINGBUF_TYPE_TIME_EXTEND:
4348		/* Internal data, OK to advance */
4349		rb_advance_reader(cpu_buffer);
4350		goto again;
4351
4352	case RINGBUF_TYPE_TIME_STAMP:
4353		if (ts) {
4354			*ts = ring_buffer_event_time_stamp(event);
4355			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4356							 cpu_buffer->cpu, ts);
4357		}
4358		/* Internal data, OK to advance */
4359		rb_advance_reader(cpu_buffer);
4360		goto again;
4361
4362	case RINGBUF_TYPE_DATA:
4363		if (ts && !(*ts)) {
4364			*ts = cpu_buffer->read_stamp + event->time_delta;
4365			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4366							 cpu_buffer->cpu, ts);
4367		}
4368		if (lost_events)
4369			*lost_events = rb_lost_events(cpu_buffer);
4370		return event;
4371
4372	default:
4373		RB_WARN_ON(cpu_buffer, 1);
4374	}
4375
4376	return NULL;
4377}
4378EXPORT_SYMBOL_GPL(ring_buffer_peek);
4379
4380static struct ring_buffer_event *
4381rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4382{
4383	struct trace_buffer *buffer;
4384	struct ring_buffer_per_cpu *cpu_buffer;
4385	struct ring_buffer_event *event;
4386	int nr_loops = 0;
4387
4388	if (ts)
4389		*ts = 0;
4390
4391	cpu_buffer = iter->cpu_buffer;
4392	buffer = cpu_buffer->buffer;
4393
4394	/*
4395	 * Check if someone performed a consuming read to
4396	 * the buffer. A consuming read invalidates the iterator
4397	 * and we need to reset the iterator in this case.
4398	 */
4399	if (unlikely(iter->cache_read != cpu_buffer->read ||
4400		     iter->cache_reader_page != cpu_buffer->reader_page))
4401		rb_iter_reset(iter);
4402
4403 again:
4404	if (ring_buffer_iter_empty(iter))
4405		return NULL;
4406
4407	/*
4408	 * As the writer can mess with what the iterator is trying
4409	 * to read, just give up if we fail to get an event after
4410	 * three tries. The iterator is not as reliable when reading
4411	 * the ring buffer with an active write as the consumer is.
4412	 * Do not warn if the three failures is reached.
 
4413	 */
4414	if (++nr_loops > 3)
4415		return NULL;
4416
4417	if (rb_per_cpu_empty(cpu_buffer))
4418		return NULL;
4419
4420	if (iter->head >= rb_page_size(iter->head_page)) {
4421		rb_inc_iter(iter);
4422		goto again;
4423	}
4424
4425	event = rb_iter_head_event(iter);
4426	if (!event)
4427		goto again;
4428
4429	switch (event->type_len) {
4430	case RINGBUF_TYPE_PADDING:
4431		if (rb_null_event(event)) {
4432			rb_inc_iter(iter);
4433			goto again;
4434		}
4435		rb_advance_iter(iter);
4436		return event;
4437
4438	case RINGBUF_TYPE_TIME_EXTEND:
4439		/* Internal data, OK to advance */
4440		rb_advance_iter(iter);
4441		goto again;
4442
4443	case RINGBUF_TYPE_TIME_STAMP:
4444		if (ts) {
4445			*ts = ring_buffer_event_time_stamp(event);
4446			ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4447							 cpu_buffer->cpu, ts);
4448		}
4449		/* Internal data, OK to advance */
4450		rb_advance_iter(iter);
4451		goto again;
4452
4453	case RINGBUF_TYPE_DATA:
4454		if (ts && !(*ts)) {
4455			*ts = iter->read_stamp + event->time_delta;
4456			ring_buffer_normalize_time_stamp(buffer,
4457							 cpu_buffer->cpu, ts);
4458		}
4459		return event;
4460
4461	default:
4462		RB_WARN_ON(cpu_buffer, 1);
4463	}
4464
4465	return NULL;
4466}
4467EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4468
4469static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4470{
4471	if (likely(!in_nmi())) {
4472		raw_spin_lock(&cpu_buffer->reader_lock);
4473		return true;
4474	}
4475
4476	/*
4477	 * If an NMI die dumps out the content of the ring buffer
4478	 * trylock must be used to prevent a deadlock if the NMI
4479	 * preempted a task that holds the ring buffer locks. If
4480	 * we get the lock then all is fine, if not, then continue
4481	 * to do the read, but this can corrupt the ring buffer,
4482	 * so it must be permanently disabled from future writes.
4483	 * Reading from NMI is a oneshot deal.
4484	 */
4485	if (raw_spin_trylock(&cpu_buffer->reader_lock))
4486		return true;
4487
4488	/* Continue without locking, but disable the ring buffer */
4489	atomic_inc(&cpu_buffer->record_disabled);
4490	return false;
4491}
4492
4493static inline void
4494rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4495{
4496	if (likely(locked))
4497		raw_spin_unlock(&cpu_buffer->reader_lock);
4498	return;
4499}
4500
4501/**
4502 * ring_buffer_peek - peek at the next event to be read
4503 * @buffer: The ring buffer to read
4504 * @cpu: The cpu to peak at
4505 * @ts: The timestamp counter of this event.
4506 * @lost_events: a variable to store if events were lost (may be NULL)
4507 *
4508 * This will return the event that will be read next, but does
4509 * not consume the data.
4510 */
4511struct ring_buffer_event *
4512ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4513		 unsigned long *lost_events)
4514{
4515	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4516	struct ring_buffer_event *event;
4517	unsigned long flags;
4518	bool dolock;
4519
4520	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4521		return NULL;
4522
4523 again:
4524	local_irq_save(flags);
4525	dolock = rb_reader_lock(cpu_buffer);
4526	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4527	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4528		rb_advance_reader(cpu_buffer);
4529	rb_reader_unlock(cpu_buffer, dolock);
4530	local_irq_restore(flags);
4531
4532	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4533		goto again;
4534
4535	return event;
4536}
4537
4538/** ring_buffer_iter_dropped - report if there are dropped events
4539 * @iter: The ring buffer iterator
4540 *
4541 * Returns true if there was dropped events since the last peek.
4542 */
4543bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4544{
4545	bool ret = iter->missed_events != 0;
4546
4547	iter->missed_events = 0;
4548	return ret;
4549}
4550EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4551
4552/**
4553 * ring_buffer_iter_peek - peek at the next event to be read
4554 * @iter: The ring buffer iterator
4555 * @ts: The timestamp counter of this event.
4556 *
4557 * This will return the event that will be read next, but does
4558 * not increment the iterator.
4559 */
4560struct ring_buffer_event *
4561ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4562{
4563	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4564	struct ring_buffer_event *event;
4565	unsigned long flags;
4566
4567 again:
4568	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4569	event = rb_iter_peek(iter, ts);
4570	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4571
4572	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4573		goto again;
4574
4575	return event;
4576}
4577
4578/**
4579 * ring_buffer_consume - return an event and consume it
4580 * @buffer: The ring buffer to get the next event from
4581 * @cpu: the cpu to read the buffer from
4582 * @ts: a variable to store the timestamp (may be NULL)
4583 * @lost_events: a variable to store if events were lost (may be NULL)
4584 *
4585 * Returns the next event in the ring buffer, and that event is consumed.
4586 * Meaning, that sequential reads will keep returning a different event,
4587 * and eventually empty the ring buffer if the producer is slower.
4588 */
4589struct ring_buffer_event *
4590ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4591		    unsigned long *lost_events)
4592{
4593	struct ring_buffer_per_cpu *cpu_buffer;
4594	struct ring_buffer_event *event = NULL;
4595	unsigned long flags;
4596	bool dolock;
4597
4598 again:
4599	/* might be called in atomic */
4600	preempt_disable();
4601
4602	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4603		goto out;
4604
4605	cpu_buffer = buffer->buffers[cpu];
4606	local_irq_save(flags);
4607	dolock = rb_reader_lock(cpu_buffer);
4608
4609	event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4610	if (event) {
4611		cpu_buffer->lost_events = 0;
4612		rb_advance_reader(cpu_buffer);
4613	}
4614
4615	rb_reader_unlock(cpu_buffer, dolock);
4616	local_irq_restore(flags);
4617
4618 out:
4619	preempt_enable();
4620
4621	if (event && event->type_len == RINGBUF_TYPE_PADDING)
4622		goto again;
4623
4624	return event;
4625}
4626EXPORT_SYMBOL_GPL(ring_buffer_consume);
4627
4628/**
4629 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4630 * @buffer: The ring buffer to read from
4631 * @cpu: The cpu buffer to iterate over
4632 * @flags: gfp flags to use for memory allocation
4633 *
4634 * This performs the initial preparations necessary to iterate
4635 * through the buffer.  Memory is allocated, buffer recording
4636 * is disabled, and the iterator pointer is returned to the caller.
4637 *
4638 * Disabling buffer recording prevents the reading from being
4639 * corrupted. This is not a consuming read, so a producer is not
4640 * expected.
4641 *
4642 * After a sequence of ring_buffer_read_prepare calls, the user is
4643 * expected to make at least one call to ring_buffer_read_prepare_sync.
4644 * Afterwards, ring_buffer_read_start is invoked to get things going
4645 * for real.
4646 *
4647 * This overall must be paired with ring_buffer_read_finish.
4648 */
4649struct ring_buffer_iter *
4650ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
4651{
4652	struct ring_buffer_per_cpu *cpu_buffer;
4653	struct ring_buffer_iter *iter;
4654
4655	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4656		return NULL;
4657
4658	iter = kzalloc(sizeof(*iter), flags);
4659	if (!iter)
4660		return NULL;
4661
4662	iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
4663	if (!iter->event) {
4664		kfree(iter);
4665		return NULL;
4666	}
4667
4668	cpu_buffer = buffer->buffers[cpu];
4669
4670	iter->cpu_buffer = cpu_buffer;
4671
4672	atomic_inc(&cpu_buffer->resize_disabled);
 
4673
4674	return iter;
4675}
4676EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4677
4678/**
4679 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4680 *
4681 * All previously invoked ring_buffer_read_prepare calls to prepare
4682 * iterators will be synchronized.  Afterwards, read_buffer_read_start
4683 * calls on those iterators are allowed.
4684 */
4685void
4686ring_buffer_read_prepare_sync(void)
4687{
4688	synchronize_rcu();
4689}
4690EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4691
4692/**
4693 * ring_buffer_read_start - start a non consuming read of the buffer
4694 * @iter: The iterator returned by ring_buffer_read_prepare
4695 *
4696 * This finalizes the startup of an iteration through the buffer.
4697 * The iterator comes from a call to ring_buffer_read_prepare and
4698 * an intervening ring_buffer_read_prepare_sync must have been
4699 * performed.
4700 *
4701 * Must be paired with ring_buffer_read_finish.
4702 */
4703void
4704ring_buffer_read_start(struct ring_buffer_iter *iter)
4705{
4706	struct ring_buffer_per_cpu *cpu_buffer;
4707	unsigned long flags;
4708
4709	if (!iter)
4710		return;
4711
4712	cpu_buffer = iter->cpu_buffer;
4713
4714	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4715	arch_spin_lock(&cpu_buffer->lock);
4716	rb_iter_reset(iter);
4717	arch_spin_unlock(&cpu_buffer->lock);
4718	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4719}
4720EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4721
4722/**
4723 * ring_buffer_read_finish - finish reading the iterator of the buffer
4724 * @iter: The iterator retrieved by ring_buffer_start
4725 *
4726 * This re-enables the recording to the buffer, and frees the
4727 * iterator.
4728 */
4729void
4730ring_buffer_read_finish(struct ring_buffer_iter *iter)
4731{
4732	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4733	unsigned long flags;
4734
4735	/*
4736	 * Ring buffer is disabled from recording, here's a good place
4737	 * to check the integrity of the ring buffer.
4738	 * Must prevent readers from trying to read, as the check
4739	 * clears the HEAD page and readers require it.
4740	 */
4741	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4742	rb_check_pages(cpu_buffer);
4743	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4744
4745	atomic_dec(&cpu_buffer->resize_disabled);
4746	kfree(iter->event);
4747	kfree(iter);
4748}
4749EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4750
4751/**
4752 * ring_buffer_iter_advance - advance the iterator to the next location
4753 * @iter: The ring buffer iterator
 
4754 *
4755 * Move the location of the iterator such that the next read will
4756 * be the next location of the iterator.
4757 */
4758void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
 
4759{
 
4760	struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4761	unsigned long flags;
4762
4763	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
 
 
 
 
 
 
 
4764
4765	rb_advance_iter(iter);
 
 
4766
4767	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4768}
4769EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
4770
4771/**
4772 * ring_buffer_size - return the size of the ring buffer (in bytes)
4773 * @buffer: The ring buffer.
4774 * @cpu: The CPU to get ring buffer size from.
4775 */
4776unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
4777{
4778	/*
4779	 * Earlier, this method returned
4780	 *	BUF_PAGE_SIZE * buffer->nr_pages
4781	 * Since the nr_pages field is now removed, we have converted this to
4782	 * return the per cpu buffer value.
4783	 */
4784	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4785		return 0;
4786
4787	return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4788}
4789EXPORT_SYMBOL_GPL(ring_buffer_size);
4790
4791static void
4792rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4793{
4794	rb_head_page_deactivate(cpu_buffer);
4795
4796	cpu_buffer->head_page
4797		= list_entry(cpu_buffer->pages, struct buffer_page, list);
4798	local_set(&cpu_buffer->head_page->write, 0);
4799	local_set(&cpu_buffer->head_page->entries, 0);
4800	local_set(&cpu_buffer->head_page->page->commit, 0);
4801
4802	cpu_buffer->head_page->read = 0;
4803
4804	cpu_buffer->tail_page = cpu_buffer->head_page;
4805	cpu_buffer->commit_page = cpu_buffer->head_page;
4806
4807	INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4808	INIT_LIST_HEAD(&cpu_buffer->new_pages);
4809	local_set(&cpu_buffer->reader_page->write, 0);
4810	local_set(&cpu_buffer->reader_page->entries, 0);
4811	local_set(&cpu_buffer->reader_page->page->commit, 0);
4812	cpu_buffer->reader_page->read = 0;
4813
4814	local_set(&cpu_buffer->entries_bytes, 0);
4815	local_set(&cpu_buffer->overrun, 0);
4816	local_set(&cpu_buffer->commit_overrun, 0);
4817	local_set(&cpu_buffer->dropped_events, 0);
4818	local_set(&cpu_buffer->entries, 0);
4819	local_set(&cpu_buffer->committing, 0);
4820	local_set(&cpu_buffer->commits, 0);
4821	local_set(&cpu_buffer->pages_touched, 0);
4822	local_set(&cpu_buffer->pages_read, 0);
4823	cpu_buffer->last_pages_touch = 0;
4824	cpu_buffer->shortest_full = 0;
4825	cpu_buffer->read = 0;
4826	cpu_buffer->read_bytes = 0;
4827
4828	rb_time_set(&cpu_buffer->write_stamp, 0);
4829	rb_time_set(&cpu_buffer->before_stamp, 0);
4830
4831	cpu_buffer->lost_events = 0;
4832	cpu_buffer->last_overrun = 0;
4833
4834	rb_head_page_activate(cpu_buffer);
4835}
4836
4837/* Must have disabled the cpu buffer then done a synchronize_rcu */
4838static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
4839{
4840	unsigned long flags;
4841
4842	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4843
4844	if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4845		goto out;
4846
4847	arch_spin_lock(&cpu_buffer->lock);
4848
4849	rb_reset_cpu(cpu_buffer);
4850
4851	arch_spin_unlock(&cpu_buffer->lock);
4852
4853 out:
4854	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4855}
4856
4857/**
4858 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4859 * @buffer: The ring buffer to reset a per cpu buffer of
4860 * @cpu: The CPU buffer to be reset
4861 */
4862void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
4863{
4864	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
 
4865
4866	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4867		return;
4868
4869	atomic_inc(&cpu_buffer->resize_disabled);
4870	atomic_inc(&cpu_buffer->record_disabled);
4871
4872	/* Make sure all commits have finished */
4873	synchronize_rcu();
4874
4875	reset_disabled_cpu_buffer(cpu_buffer);
4876
4877	atomic_dec(&cpu_buffer->record_disabled);
4878	atomic_dec(&cpu_buffer->resize_disabled);
4879}
4880EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4881
4882/**
4883 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4884 * @buffer: The ring buffer to reset a per cpu buffer of
4885 * @cpu: The CPU buffer to be reset
4886 */
4887void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
4888{
4889	struct ring_buffer_per_cpu *cpu_buffer;
4890	int cpu;
4891
4892	for_each_online_buffer_cpu(buffer, cpu) {
4893		cpu_buffer = buffer->buffers[cpu];
4894
4895		atomic_inc(&cpu_buffer->resize_disabled);
4896		atomic_inc(&cpu_buffer->record_disabled);
4897	}
4898
4899	/* Make sure all commits have finished */
4900	synchronize_rcu();
4901
4902	for_each_online_buffer_cpu(buffer, cpu) {
4903		cpu_buffer = buffer->buffers[cpu];
4904
4905		reset_disabled_cpu_buffer(cpu_buffer);
4906
4907		atomic_dec(&cpu_buffer->record_disabled);
4908		atomic_dec(&cpu_buffer->resize_disabled);
4909	}
4910}
 
4911
4912/**
4913 * ring_buffer_reset - reset a ring buffer
4914 * @buffer: The ring buffer to reset all cpu buffers
4915 */
4916void ring_buffer_reset(struct trace_buffer *buffer)
4917{
4918	struct ring_buffer_per_cpu *cpu_buffer;
4919	int cpu;
4920
4921	for_each_buffer_cpu(buffer, cpu) {
4922		cpu_buffer = buffer->buffers[cpu];
4923
4924		atomic_inc(&cpu_buffer->resize_disabled);
4925		atomic_inc(&cpu_buffer->record_disabled);
4926	}
4927
4928	/* Make sure all commits have finished */
4929	synchronize_rcu();
4930
4931	for_each_buffer_cpu(buffer, cpu) {
4932		cpu_buffer = buffer->buffers[cpu];
4933
4934		reset_disabled_cpu_buffer(cpu_buffer);
4935
4936		atomic_dec(&cpu_buffer->record_disabled);
4937		atomic_dec(&cpu_buffer->resize_disabled);
4938	}
4939}
4940EXPORT_SYMBOL_GPL(ring_buffer_reset);
4941
4942/**
4943 * rind_buffer_empty - is the ring buffer empty?
4944 * @buffer: The ring buffer to test
4945 */
4946bool ring_buffer_empty(struct trace_buffer *buffer)
4947{
4948	struct ring_buffer_per_cpu *cpu_buffer;
4949	unsigned long flags;
4950	bool dolock;
4951	int cpu;
4952	int ret;
4953
4954	/* yes this is racy, but if you don't like the race, lock the buffer */
4955	for_each_buffer_cpu(buffer, cpu) {
4956		cpu_buffer = buffer->buffers[cpu];
4957		local_irq_save(flags);
4958		dolock = rb_reader_lock(cpu_buffer);
4959		ret = rb_per_cpu_empty(cpu_buffer);
4960		rb_reader_unlock(cpu_buffer, dolock);
4961		local_irq_restore(flags);
4962
4963		if (!ret)
4964			return false;
4965	}
4966
4967	return true;
4968}
4969EXPORT_SYMBOL_GPL(ring_buffer_empty);
4970
4971/**
4972 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4973 * @buffer: The ring buffer
4974 * @cpu: The CPU buffer to test
4975 */
4976bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
4977{
4978	struct ring_buffer_per_cpu *cpu_buffer;
4979	unsigned long flags;
4980	bool dolock;
4981	int ret;
4982
4983	if (!cpumask_test_cpu(cpu, buffer->cpumask))
4984		return true;
4985
4986	cpu_buffer = buffer->buffers[cpu];
4987	local_irq_save(flags);
4988	dolock = rb_reader_lock(cpu_buffer);
4989	ret = rb_per_cpu_empty(cpu_buffer);
4990	rb_reader_unlock(cpu_buffer, dolock);
4991	local_irq_restore(flags);
4992
4993	return ret;
4994}
4995EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4996
4997#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4998/**
4999 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5000 * @buffer_a: One buffer to swap with
5001 * @buffer_b: The other buffer to swap with
5002 * @cpu: the CPU of the buffers to swap
5003 *
5004 * This function is useful for tracers that want to take a "snapshot"
5005 * of a CPU buffer and has another back up buffer lying around.
5006 * it is expected that the tracer handles the cpu buffer not being
5007 * used at the moment.
5008 */
5009int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5010			 struct trace_buffer *buffer_b, int cpu)
5011{
5012	struct ring_buffer_per_cpu *cpu_buffer_a;
5013	struct ring_buffer_per_cpu *cpu_buffer_b;
5014	int ret = -EINVAL;
5015
5016	if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5017	    !cpumask_test_cpu(cpu, buffer_b->cpumask))
5018		goto out;
5019
5020	cpu_buffer_a = buffer_a->buffers[cpu];
5021	cpu_buffer_b = buffer_b->buffers[cpu];
5022
5023	/* At least make sure the two buffers are somewhat the same */
5024	if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5025		goto out;
5026
5027	ret = -EAGAIN;
5028
5029	if (atomic_read(&buffer_a->record_disabled))
5030		goto out;
5031
5032	if (atomic_read(&buffer_b->record_disabled))
5033		goto out;
5034
5035	if (atomic_read(&cpu_buffer_a->record_disabled))
5036		goto out;
5037
5038	if (atomic_read(&cpu_buffer_b->record_disabled))
5039		goto out;
5040
5041	/*
5042	 * We can't do a synchronize_rcu here because this
5043	 * function can be called in atomic context.
5044	 * Normally this will be called from the same CPU as cpu.
5045	 * If not it's up to the caller to protect this.
5046	 */
5047	atomic_inc(&cpu_buffer_a->record_disabled);
5048	atomic_inc(&cpu_buffer_b->record_disabled);
5049
5050	ret = -EBUSY;
5051	if (local_read(&cpu_buffer_a->committing))
5052		goto out_dec;
5053	if (local_read(&cpu_buffer_b->committing))
5054		goto out_dec;
5055
5056	buffer_a->buffers[cpu] = cpu_buffer_b;
5057	buffer_b->buffers[cpu] = cpu_buffer_a;
5058
5059	cpu_buffer_b->buffer = buffer_a;
5060	cpu_buffer_a->buffer = buffer_b;
5061
5062	ret = 0;
5063
5064out_dec:
5065	atomic_dec(&cpu_buffer_a->record_disabled);
5066	atomic_dec(&cpu_buffer_b->record_disabled);
5067out:
5068	return ret;
5069}
5070EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5071#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5072
5073/**
5074 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5075 * @buffer: the buffer to allocate for.
5076 * @cpu: the cpu buffer to allocate.
5077 *
5078 * This function is used in conjunction with ring_buffer_read_page.
5079 * When reading a full page from the ring buffer, these functions
5080 * can be used to speed up the process. The calling function should
5081 * allocate a few pages first with this function. Then when it
5082 * needs to get pages from the ring buffer, it passes the result
5083 * of this function into ring_buffer_read_page, which will swap
5084 * the page that was allocated, with the read page of the buffer.
5085 *
5086 * Returns:
5087 *  The page allocated, or ERR_PTR
5088 */
5089void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5090{
5091	struct ring_buffer_per_cpu *cpu_buffer;
5092	struct buffer_data_page *bpage = NULL;
5093	unsigned long flags;
5094	struct page *page;
5095
5096	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5097		return ERR_PTR(-ENODEV);
5098
5099	cpu_buffer = buffer->buffers[cpu];
5100	local_irq_save(flags);
5101	arch_spin_lock(&cpu_buffer->lock);
5102
5103	if (cpu_buffer->free_page) {
5104		bpage = cpu_buffer->free_page;
5105		cpu_buffer->free_page = NULL;
5106	}
5107
5108	arch_spin_unlock(&cpu_buffer->lock);
5109	local_irq_restore(flags);
5110
5111	if (bpage)
5112		goto out;
5113
5114	page = alloc_pages_node(cpu_to_node(cpu),
5115				GFP_KERNEL | __GFP_NORETRY, 0);
5116	if (!page)
5117		return ERR_PTR(-ENOMEM);
5118
5119	bpage = page_address(page);
5120
5121 out:
5122	rb_init_page(bpage);
5123
5124	return bpage;
5125}
5126EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5127
5128/**
5129 * ring_buffer_free_read_page - free an allocated read page
5130 * @buffer: the buffer the page was allocate for
5131 * @cpu: the cpu buffer the page came from
5132 * @data: the page to free
5133 *
5134 * Free a page allocated from ring_buffer_alloc_read_page.
5135 */
5136void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5137{
5138	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5139	struct buffer_data_page *bpage = data;
5140	struct page *page = virt_to_page(bpage);
5141	unsigned long flags;
5142
5143	/* If the page is still in use someplace else, we can't reuse it */
5144	if (page_ref_count(page) > 1)
5145		goto out;
5146
5147	local_irq_save(flags);
5148	arch_spin_lock(&cpu_buffer->lock);
5149
5150	if (!cpu_buffer->free_page) {
5151		cpu_buffer->free_page = bpage;
5152		bpage = NULL;
5153	}
5154
5155	arch_spin_unlock(&cpu_buffer->lock);
5156	local_irq_restore(flags);
5157
5158 out:
5159	free_page((unsigned long)bpage);
5160}
5161EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5162
5163/**
5164 * ring_buffer_read_page - extract a page from the ring buffer
5165 * @buffer: buffer to extract from
5166 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5167 * @len: amount to extract
5168 * @cpu: the cpu of the buffer to extract
5169 * @full: should the extraction only happen when the page is full.
5170 *
5171 * This function will pull out a page from the ring buffer and consume it.
5172 * @data_page must be the address of the variable that was returned
5173 * from ring_buffer_alloc_read_page. This is because the page might be used
5174 * to swap with a page in the ring buffer.
5175 *
5176 * for example:
5177 *	rpage = ring_buffer_alloc_read_page(buffer, cpu);
5178 *	if (IS_ERR(rpage))
5179 *		return PTR_ERR(rpage);
5180 *	ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5181 *	if (ret >= 0)
5182 *		process_page(rpage, ret);
5183 *
5184 * When @full is set, the function will not return true unless
5185 * the writer is off the reader page.
5186 *
5187 * Note: it is up to the calling functions to handle sleeps and wakeups.
5188 *  The ring buffer can be used anywhere in the kernel and can not
5189 *  blindly call wake_up. The layer that uses the ring buffer must be
5190 *  responsible for that.
5191 *
5192 * Returns:
5193 *  >=0 if data has been transferred, returns the offset of consumed data.
5194 *  <0 if no data has been transferred.
5195 */
5196int ring_buffer_read_page(struct trace_buffer *buffer,
5197			  void **data_page, size_t len, int cpu, int full)
5198{
5199	struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5200	struct ring_buffer_event *event;
5201	struct buffer_data_page *bpage;
5202	struct buffer_page *reader;
5203	unsigned long missed_events;
5204	unsigned long flags;
5205	unsigned int commit;
5206	unsigned int read;
5207	u64 save_timestamp;
5208	int ret = -1;
5209
5210	if (!cpumask_test_cpu(cpu, buffer->cpumask))
5211		goto out;
5212
5213	/*
5214	 * If len is not big enough to hold the page header, then
5215	 * we can not copy anything.
5216	 */
5217	if (len <= BUF_PAGE_HDR_SIZE)
5218		goto out;
5219
5220	len -= BUF_PAGE_HDR_SIZE;
5221
5222	if (!data_page)
5223		goto out;
5224
5225	bpage = *data_page;
5226	if (!bpage)
5227		goto out;
5228
5229	raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5230
5231	reader = rb_get_reader_page(cpu_buffer);
5232	if (!reader)
5233		goto out_unlock;
5234
5235	event = rb_reader_event(cpu_buffer);
5236
5237	read = reader->read;
5238	commit = rb_page_commit(reader);
5239
5240	/* Check if any events were dropped */
5241	missed_events = cpu_buffer->lost_events;
5242
5243	/*
5244	 * If this page has been partially read or
5245	 * if len is not big enough to read the rest of the page or
5246	 * a writer is still on the page, then
5247	 * we must copy the data from the page to the buffer.
5248	 * Otherwise, we can simply swap the page with the one passed in.
5249	 */
5250	if (read || (len < (commit - read)) ||
5251	    cpu_buffer->reader_page == cpu_buffer->commit_page) {
5252		struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5253		unsigned int rpos = read;
5254		unsigned int pos = 0;
5255		unsigned int size;
5256
5257		if (full)
5258			goto out_unlock;
5259
5260		if (len > (commit - read))
5261			len = (commit - read);
5262
5263		/* Always keep the time extend and data together */
5264		size = rb_event_ts_length(event);
5265
5266		if (len < size)
5267			goto out_unlock;
5268
5269		/* save the current timestamp, since the user will need it */
5270		save_timestamp = cpu_buffer->read_stamp;
5271
5272		/* Need to copy one event at a time */
5273		do {
5274			/* We need the size of one event, because
5275			 * rb_advance_reader only advances by one event,
5276			 * whereas rb_event_ts_length may include the size of
5277			 * one or two events.
5278			 * We have already ensured there's enough space if this
5279			 * is a time extend. */
5280			size = rb_event_length(event);
5281			memcpy(bpage->data + pos, rpage->data + rpos, size);
5282
5283			len -= size;
5284
5285			rb_advance_reader(cpu_buffer);
5286			rpos = reader->read;
5287			pos += size;
5288
5289			if (rpos >= commit)
5290				break;
5291
5292			event = rb_reader_event(cpu_buffer);
5293			/* Always keep the time extend and data together */
5294			size = rb_event_ts_length(event);
5295		} while (len >= size);
5296
5297		/* update bpage */
5298		local_set(&bpage->commit, pos);
5299		bpage->time_stamp = save_timestamp;
5300
5301		/* we copied everything to the beginning */
5302		read = 0;
5303	} else {
5304		/* update the entry counter */
5305		cpu_buffer->read += rb_page_entries(reader);
5306		cpu_buffer->read_bytes += BUF_PAGE_SIZE;
5307
5308		/* swap the pages */
5309		rb_init_page(bpage);
5310		bpage = reader->page;
5311		reader->page = *data_page;
5312		local_set(&reader->write, 0);
5313		local_set(&reader->entries, 0);
5314		reader->read = 0;
5315		*data_page = bpage;
5316
5317		/*
5318		 * Use the real_end for the data size,
5319		 * This gives us a chance to store the lost events
5320		 * on the page.
5321		 */
5322		if (reader->real_end)
5323			local_set(&bpage->commit, reader->real_end);
5324	}
5325	ret = read;
5326
5327	cpu_buffer->lost_events = 0;
5328
5329	commit = local_read(&bpage->commit);
5330	/*
5331	 * Set a flag in the commit field if we lost events
5332	 */
5333	if (missed_events) {
5334		/* If there is room at the end of the page to save the
5335		 * missed events, then record it there.
5336		 */
5337		if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5338			memcpy(&bpage->data[commit], &missed_events,
5339			       sizeof(missed_events));
5340			local_add(RB_MISSED_STORED, &bpage->commit);
5341			commit += sizeof(missed_events);
5342		}
5343		local_add(RB_MISSED_EVENTS, &bpage->commit);
5344	}
5345
5346	/*
5347	 * This page may be off to user land. Zero it out here.
5348	 */
5349	if (commit < BUF_PAGE_SIZE)
5350		memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5351
5352 out_unlock:
5353	raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5354
5355 out:
5356	return ret;
5357}
5358EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5359
5360/*
5361 * We only allocate new buffers, never free them if the CPU goes down.
5362 * If we were to free the buffer, then the user would lose any trace that was in
5363 * the buffer.
5364 */
5365int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5366{
5367	struct trace_buffer *buffer;
5368	long nr_pages_same;
5369	int cpu_i;
5370	unsigned long nr_pages;
5371
5372	buffer = container_of(node, struct trace_buffer, node);
5373	if (cpumask_test_cpu(cpu, buffer->cpumask))
5374		return 0;
5375
5376	nr_pages = 0;
5377	nr_pages_same = 1;
5378	/* check if all cpu sizes are same */
5379	for_each_buffer_cpu(buffer, cpu_i) {
5380		/* fill in the size from first enabled cpu */
5381		if (nr_pages == 0)
5382			nr_pages = buffer->buffers[cpu_i]->nr_pages;
5383		if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5384			nr_pages_same = 0;
5385			break;
5386		}
5387	}
5388	/* allocate minimum pages, user can later expand it */
5389	if (!nr_pages_same)
5390		nr_pages = 2;
5391	buffer->buffers[cpu] =
5392		rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5393	if (!buffer->buffers[cpu]) {
5394		WARN(1, "failed to allocate ring buffer on CPU %u\n",
5395		     cpu);
5396		return -ENOMEM;
5397	}
5398	smp_wmb();
5399	cpumask_set_cpu(cpu, buffer->cpumask);
5400	return 0;
5401}
5402
5403#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5404/*
5405 * This is a basic integrity check of the ring buffer.
5406 * Late in the boot cycle this test will run when configured in.
5407 * It will kick off a thread per CPU that will go into a loop
5408 * writing to the per cpu ring buffer various sizes of data.
5409 * Some of the data will be large items, some small.
5410 *
5411 * Another thread is created that goes into a spin, sending out
5412 * IPIs to the other CPUs to also write into the ring buffer.
5413 * this is to test the nesting ability of the buffer.
5414 *
5415 * Basic stats are recorded and reported. If something in the
5416 * ring buffer should happen that's not expected, a big warning
5417 * is displayed and all ring buffers are disabled.
5418 */
5419static struct task_struct *rb_threads[NR_CPUS] __initdata;
5420
5421struct rb_test_data {
5422	struct trace_buffer *buffer;
5423	unsigned long		events;
5424	unsigned long		bytes_written;
5425	unsigned long		bytes_alloc;
5426	unsigned long		bytes_dropped;
5427	unsigned long		events_nested;
5428	unsigned long		bytes_written_nested;
5429	unsigned long		bytes_alloc_nested;
5430	unsigned long		bytes_dropped_nested;
5431	int			min_size_nested;
5432	int			max_size_nested;
5433	int			max_size;
5434	int			min_size;
5435	int			cpu;
5436	int			cnt;
5437};
5438
5439static struct rb_test_data rb_data[NR_CPUS] __initdata;
5440
5441/* 1 meg per cpu */
5442#define RB_TEST_BUFFER_SIZE	1048576
5443
5444static char rb_string[] __initdata =
5445	"abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5446	"?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5447	"!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5448
5449static bool rb_test_started __initdata;
5450
5451struct rb_item {
5452	int size;
5453	char str[];
5454};
5455
5456static __init int rb_write_something(struct rb_test_data *data, bool nested)
5457{
5458	struct ring_buffer_event *event;
5459	struct rb_item *item;
5460	bool started;
5461	int event_len;
5462	int size;
5463	int len;
5464	int cnt;
5465
5466	/* Have nested writes different that what is written */
5467	cnt = data->cnt + (nested ? 27 : 0);
5468
5469	/* Multiply cnt by ~e, to make some unique increment */
5470	size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5471
5472	len = size + sizeof(struct rb_item);
5473
5474	started = rb_test_started;
5475	/* read rb_test_started before checking buffer enabled */
5476	smp_rmb();
5477
5478	event = ring_buffer_lock_reserve(data->buffer, len);
5479	if (!event) {
5480		/* Ignore dropped events before test starts. */
5481		if (started) {
5482			if (nested)
5483				data->bytes_dropped += len;
5484			else
5485				data->bytes_dropped_nested += len;
5486		}
5487		return len;
5488	}
5489
5490	event_len = ring_buffer_event_length(event);
5491
5492	if (RB_WARN_ON(data->buffer, event_len < len))
5493		goto out;
5494
5495	item = ring_buffer_event_data(event);
5496	item->size = size;
5497	memcpy(item->str, rb_string, size);
5498
5499	if (nested) {
5500		data->bytes_alloc_nested += event_len;
5501		data->bytes_written_nested += len;
5502		data->events_nested++;
5503		if (!data->min_size_nested || len < data->min_size_nested)
5504			data->min_size_nested = len;
5505		if (len > data->max_size_nested)
5506			data->max_size_nested = len;
5507	} else {
5508		data->bytes_alloc += event_len;
5509		data->bytes_written += len;
5510		data->events++;
5511		if (!data->min_size || len < data->min_size)
5512			data->max_size = len;
5513		if (len > data->max_size)
5514			data->max_size = len;
5515	}
5516
5517 out:
5518	ring_buffer_unlock_commit(data->buffer, event);
5519
5520	return 0;
5521}
5522
5523static __init int rb_test(void *arg)
5524{
5525	struct rb_test_data *data = arg;
5526
5527	while (!kthread_should_stop()) {
5528		rb_write_something(data, false);
5529		data->cnt++;
5530
5531		set_current_state(TASK_INTERRUPTIBLE);
5532		/* Now sleep between a min of 100-300us and a max of 1ms */
5533		usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5534	}
5535
5536	return 0;
5537}
5538
5539static __init void rb_ipi(void *ignore)
5540{
5541	struct rb_test_data *data;
5542	int cpu = smp_processor_id();
5543
5544	data = &rb_data[cpu];
5545	rb_write_something(data, true);
5546}
5547
5548static __init int rb_hammer_test(void *arg)
5549{
5550	while (!kthread_should_stop()) {
5551
5552		/* Send an IPI to all cpus to write data! */
5553		smp_call_function(rb_ipi, NULL, 1);
5554		/* No sleep, but for non preempt, let others run */
5555		schedule();
5556	}
5557
5558	return 0;
5559}
5560
5561static __init int test_ringbuffer(void)
5562{
5563	struct task_struct *rb_hammer;
5564	struct trace_buffer *buffer;
5565	int cpu;
5566	int ret = 0;
5567
5568	if (security_locked_down(LOCKDOWN_TRACEFS)) {
5569		pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
5570		return 0;
5571	}
5572
5573	pr_info("Running ring buffer tests...\n");
5574
5575	buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5576	if (WARN_ON(!buffer))
5577		return 0;
5578
5579	/* Disable buffer so that threads can't write to it yet */
5580	ring_buffer_record_off(buffer);
5581
5582	for_each_online_cpu(cpu) {
5583		rb_data[cpu].buffer = buffer;
5584		rb_data[cpu].cpu = cpu;
5585		rb_data[cpu].cnt = cpu;
5586		rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5587						 "rbtester/%d", cpu);
5588		if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5589			pr_cont("FAILED\n");
5590			ret = PTR_ERR(rb_threads[cpu]);
5591			goto out_free;
5592		}
5593
5594		kthread_bind(rb_threads[cpu], cpu);
5595 		wake_up_process(rb_threads[cpu]);
5596	}
5597
5598	/* Now create the rb hammer! */
5599	rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5600	if (WARN_ON(IS_ERR(rb_hammer))) {
5601		pr_cont("FAILED\n");
5602		ret = PTR_ERR(rb_hammer);
5603		goto out_free;
5604	}
5605
5606	ring_buffer_record_on(buffer);
5607	/*
5608	 * Show buffer is enabled before setting rb_test_started.
5609	 * Yes there's a small race window where events could be
5610	 * dropped and the thread wont catch it. But when a ring
5611	 * buffer gets enabled, there will always be some kind of
5612	 * delay before other CPUs see it. Thus, we don't care about
5613	 * those dropped events. We care about events dropped after
5614	 * the threads see that the buffer is active.
5615	 */
5616	smp_wmb();
5617	rb_test_started = true;
5618
5619	set_current_state(TASK_INTERRUPTIBLE);
5620	/* Just run for 10 seconds */;
5621	schedule_timeout(10 * HZ);
5622
5623	kthread_stop(rb_hammer);
5624
5625 out_free:
5626	for_each_online_cpu(cpu) {
5627		if (!rb_threads[cpu])
5628			break;
5629		kthread_stop(rb_threads[cpu]);
5630	}
5631	if (ret) {
5632		ring_buffer_free(buffer);
5633		return ret;
5634	}
5635
5636	/* Report! */
5637	pr_info("finished\n");
5638	for_each_online_cpu(cpu) {
5639		struct ring_buffer_event *event;
5640		struct rb_test_data *data = &rb_data[cpu];
5641		struct rb_item *item;
5642		unsigned long total_events;
5643		unsigned long total_dropped;
5644		unsigned long total_written;
5645		unsigned long total_alloc;
5646		unsigned long total_read = 0;
5647		unsigned long total_size = 0;
5648		unsigned long total_len = 0;
5649		unsigned long total_lost = 0;
5650		unsigned long lost;
5651		int big_event_size;
5652		int small_event_size;
5653
5654		ret = -1;
5655
5656		total_events = data->events + data->events_nested;
5657		total_written = data->bytes_written + data->bytes_written_nested;
5658		total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5659		total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5660
5661		big_event_size = data->max_size + data->max_size_nested;
5662		small_event_size = data->min_size + data->min_size_nested;
5663
5664		pr_info("CPU %d:\n", cpu);
5665		pr_info("              events:    %ld\n", total_events);
5666		pr_info("       dropped bytes:    %ld\n", total_dropped);
5667		pr_info("       alloced bytes:    %ld\n", total_alloc);
5668		pr_info("       written bytes:    %ld\n", total_written);
5669		pr_info("       biggest event:    %d\n", big_event_size);
5670		pr_info("      smallest event:    %d\n", small_event_size);
5671
5672		if (RB_WARN_ON(buffer, total_dropped))
5673			break;
5674
5675		ret = 0;
5676
5677		while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5678			total_lost += lost;
5679			item = ring_buffer_event_data(event);
5680			total_len += ring_buffer_event_length(event);
5681			total_size += item->size + sizeof(struct rb_item);
5682			if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5683				pr_info("FAILED!\n");
5684				pr_info("buffer had: %.*s\n", item->size, item->str);
5685				pr_info("expected:   %.*s\n", item->size, rb_string);
5686				RB_WARN_ON(buffer, 1);
5687				ret = -1;
5688				break;
5689			}
5690			total_read++;
5691		}
5692		if (ret)
5693			break;
5694
5695		ret = -1;
5696
5697		pr_info("         read events:   %ld\n", total_read);
5698		pr_info("         lost events:   %ld\n", total_lost);
5699		pr_info("        total events:   %ld\n", total_lost + total_read);
5700		pr_info("  recorded len bytes:   %ld\n", total_len);
5701		pr_info(" recorded size bytes:   %ld\n", total_size);
5702		if (total_lost)
5703			pr_info(" With dropped events, record len and size may not match\n"
5704				" alloced and written from above\n");
5705		if (!total_lost) {
5706			if (RB_WARN_ON(buffer, total_len != total_alloc ||
5707				       total_size != total_written))
5708				break;
5709		}
5710		if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5711			break;
5712
5713		ret = 0;
5714	}
5715	if (!ret)
5716		pr_info("Ring buffer PASSED!\n");
5717
5718	ring_buffer_free(buffer);
5719	return 0;
5720}
5721
5722late_initcall(test_ringbuffer);
5723#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */