linux/kernel/trace/ring_buffer.c
<<
>>
Prefs
   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_recursion.h>
   8#include <linux/trace_events.h>
   9#include <linux/ring_buffer.h>
  10#include <linux/trace_clock.h>
  11#include <linux/sched/clock.h>
  12#include <linux/trace_seq.h>
  13#include <linux/spinlock.h>
  14#include <linux/irq_work.h>
  15#include <linux/security.h>
  16#include <linux/uaccess.h>
  17#include <linux/hardirq.h>
  18#include <linux/kthread.h>      /* for self test */
  19#include <linux/module.h>
  20#include <linux/percpu.h>
  21#include <linux/mutex.h>
  22#include <linux/delay.h>
  23#include <linux/slab.h>
  24#include <linux/init.h>
  25#include <linux/hash.h>
  26#include <linux/list.h>
  27#include <linux/cpu.h>
  28#include <linux/oom.h>
  29
  30#include <asm/local.h>
  31
  32static void update_pages_handler(struct work_struct *work);
  33
  34/*
  35 * The ring buffer header is special. We must manually up keep it.
  36 */
  37int ring_buffer_print_entry_header(struct trace_seq *s)
  38{
  39        trace_seq_puts(s, "# compressed entry header\n");
  40        trace_seq_puts(s, "\ttype_len    :    5 bits\n");
  41        trace_seq_puts(s, "\ttime_delta  :   27 bits\n");
  42        trace_seq_puts(s, "\tarray       :   32 bits\n");
  43        trace_seq_putc(s, '\n');
  44        trace_seq_printf(s, "\tpadding     : type == %d\n",
  45                         RINGBUF_TYPE_PADDING);
  46        trace_seq_printf(s, "\ttime_extend : type == %d\n",
  47                         RINGBUF_TYPE_TIME_EXTEND);
  48        trace_seq_printf(s, "\ttime_stamp : type == %d\n",
  49                         RINGBUF_TYPE_TIME_STAMP);
  50        trace_seq_printf(s, "\tdata max type_len  == %d\n",
  51                         RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
  52
  53        return !trace_seq_has_overflowed(s);
  54}
  55
  56/*
  57 * The ring buffer is made up of a list of pages. A separate list of pages is
  58 * allocated for each CPU. A writer may only write to a buffer that is
  59 * associated with the CPU it is currently executing on.  A reader may read
  60 * from any per cpu buffer.
  61 *
  62 * The reader is special. For each per cpu buffer, the reader has its own
  63 * reader page. When a reader has read the entire reader page, this reader
  64 * page is swapped with another page in the ring buffer.
  65 *
  66 * Now, as long as the writer is off the reader page, the reader can do what
  67 * ever it wants with that page. The writer will never write to that page
  68 * again (as long as it is out of the ring buffer).
  69 *
  70 * Here's some silly ASCII art.
  71 *
  72 *   +------+
  73 *   |reader|          RING BUFFER
  74 *   |page  |
  75 *   +------+        +---+   +---+   +---+
  76 *                   |   |-->|   |-->|   |
  77 *                   +---+   +---+   +---+
  78 *                     ^               |
  79 *                     |               |
  80 *                     +---------------+
  81 *
  82 *
  83 *   +------+
  84 *   |reader|          RING BUFFER
  85 *   |page  |------------------v
  86 *   +------+        +---+   +---+   +---+
  87 *                   |   |-->|   |-->|   |
  88 *                   +---+   +---+   +---+
  89 *                     ^               |
  90 *                     |               |
  91 *                     +---------------+
  92 *
  93 *
  94 *   +------+
  95 *   |reader|          RING BUFFER
  96 *   |page  |------------------v
  97 *   +------+        +---+   +---+   +---+
  98 *      ^            |   |-->|   |-->|   |
  99 *      |            +---+   +---+   +---+
 100 *      |                              |
 101 *      |                              |
 102 *      +------------------------------+
 103 *
 104 *
 105 *   +------+
 106 *   |buffer|          RING BUFFER
 107 *   |page  |------------------v
 108 *   +------+        +---+   +---+   +---+
 109 *      ^            |   |   |   |-->|   |
 110 *      |   New      +---+   +---+   +---+
 111 *      |  Reader------^               |
 112 *      |   page                       |
 113 *      +------------------------------+
 114 *
 115 *
 116 * After we make this swap, the reader can hand this page off to the splice
 117 * code and be done with it. It can even allocate a new page if it needs to
 118 * and swap that into the ring buffer.
 119 *
 120 * We will be using cmpxchg soon to make all this lockless.
 121 *
 122 */
 123
 124/* Used for individual buffers (after the counter) */
 125#define RB_BUFFER_OFF           (1 << 20)
 126
 127#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
 128
 129#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
 130#define RB_ALIGNMENT            4U
 131#define RB_MAX_SMALL_DATA       (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 132#define RB_EVNT_MIN_SIZE        8U      /* two 32bit words */
 133
 134#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
 135# define RB_FORCE_8BYTE_ALIGNMENT       0
 136# define RB_ARCH_ALIGNMENT              RB_ALIGNMENT
 137#else
 138# define RB_FORCE_8BYTE_ALIGNMENT       1
 139# define RB_ARCH_ALIGNMENT              8U
 140#endif
 141
 142#define RB_ALIGN_DATA           __aligned(RB_ARCH_ALIGNMENT)
 143
 144/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
 145#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
 146
 147enum {
 148        RB_LEN_TIME_EXTEND = 8,
 149        RB_LEN_TIME_STAMP =  8,
 150};
 151
 152#define skip_time_extend(event) \
 153        ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
 154
 155#define extended_time(event) \
 156        (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
 157
 158static inline int rb_null_event(struct ring_buffer_event *event)
 159{
 160        return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
 161}
 162
 163static void rb_event_set_padding(struct ring_buffer_event *event)
 164{
 165        /* padding has a NULL time_delta */
 166        event->type_len = RINGBUF_TYPE_PADDING;
 167        event->time_delta = 0;
 168}
 169
 170static unsigned
 171rb_event_data_length(struct ring_buffer_event *event)
 172{
 173        unsigned length;
 174
 175        if (event->type_len)
 176                length = event->type_len * RB_ALIGNMENT;
 177        else
 178                length = event->array[0];
 179        return length + RB_EVNT_HDR_SIZE;
 180}
 181
 182/*
 183 * Return the length of the given event. Will return
 184 * the length of the time extend if the event is a
 185 * time extend.
 186 */
 187static inline unsigned
 188rb_event_length(struct ring_buffer_event *event)
 189{
 190        switch (event->type_len) {
 191        case RINGBUF_TYPE_PADDING:
 192                if (rb_null_event(event))
 193                        /* undefined */
 194                        return -1;
 195                return  event->array[0] + RB_EVNT_HDR_SIZE;
 196
 197        case RINGBUF_TYPE_TIME_EXTEND:
 198                return RB_LEN_TIME_EXTEND;
 199
 200        case RINGBUF_TYPE_TIME_STAMP:
 201                return RB_LEN_TIME_STAMP;
 202
 203        case RINGBUF_TYPE_DATA:
 204                return rb_event_data_length(event);
 205        default:
 206                WARN_ON_ONCE(1);
 207        }
 208        /* not hit */
 209        return 0;
 210}
 211
 212/*
 213 * Return total length of time extend and data,
 214 *   or just the event length for all other events.
 215 */
 216static inline unsigned
 217rb_event_ts_length(struct ring_buffer_event *event)
 218{
 219        unsigned len = 0;
 220
 221        if (extended_time(event)) {
 222                /* time extends include the data event after it */
 223                len = RB_LEN_TIME_EXTEND;
 224                event = skip_time_extend(event);
 225        }
 226        return len + rb_event_length(event);
 227}
 228
 229/**
 230 * ring_buffer_event_length - return the length of the event
 231 * @event: the event to get the length of
 232 *
 233 * Returns the size of the data load of a data event.
 234 * If the event is something other than a data event, it
 235 * returns the size of the event itself. With the exception
 236 * of a TIME EXTEND, where it still returns the size of the
 237 * data load of the data event after it.
 238 */
 239unsigned ring_buffer_event_length(struct ring_buffer_event *event)
 240{
 241        unsigned length;
 242
 243        if (extended_time(event))
 244                event = skip_time_extend(event);
 245
 246        length = rb_event_length(event);
 247        if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
 248                return length;
 249        length -= RB_EVNT_HDR_SIZE;
 250        if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
 251                length -= sizeof(event->array[0]);
 252        return length;
 253}
 254EXPORT_SYMBOL_GPL(ring_buffer_event_length);
 255
 256/* inline for ring buffer fast paths */
 257static __always_inline void *
 258rb_event_data(struct ring_buffer_event *event)
 259{
 260        if (extended_time(event))
 261                event = skip_time_extend(event);
 262        WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
 263        /* If length is in len field, then array[0] has the data */
 264        if (event->type_len)
 265                return (void *)&event->array[0];
 266        /* Otherwise length is in array[0] and array[1] has the data */
 267        return (void *)&event->array[1];
 268}
 269
 270/**
 271 * ring_buffer_event_data - return the data of the event
 272 * @event: the event to get the data from
 273 */
 274void *ring_buffer_event_data(struct ring_buffer_event *event)
 275{
 276        return rb_event_data(event);
 277}
 278EXPORT_SYMBOL_GPL(ring_buffer_event_data);
 279
 280#define for_each_buffer_cpu(buffer, cpu)                \
 281        for_each_cpu(cpu, buffer->cpumask)
 282
 283#define for_each_online_buffer_cpu(buffer, cpu)         \
 284        for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
 285
 286#define TS_SHIFT        27
 287#define TS_MASK         ((1ULL << TS_SHIFT) - 1)
 288#define TS_DELTA_TEST   (~TS_MASK)
 289
 290static u64 rb_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 *  TRANSITION = 0
 442 *  NMI     = 1
 443 *  IRQ     = 2
 444 *  SOFTIRQ = 3
 445 *  NORMAL  = 4
 446 *
 447 * See trace_recursive_lock() comment below for more details.
 448 */
 449enum {
 450        RB_CTX_TRANSITION,
 451        RB_CTX_NMI,
 452        RB_CTX_IRQ,
 453        RB_CTX_SOFTIRQ,
 454        RB_CTX_NORMAL,
 455        RB_CTX_MAX
 456};
 457
 458#if BITS_PER_LONG == 32
 459#define RB_TIME_32
 460#endif
 461
 462/* To test on 64 bit machines */
 463//#define RB_TIME_32
 464
 465#ifdef RB_TIME_32
 466
 467struct rb_time_struct {
 468        local_t         cnt;
 469        local_t         top;
 470        local_t         bottom;
 471};
 472#else
 473#include <asm/local64.h>
 474struct rb_time_struct {
 475        local64_t       time;
 476};
 477#endif
 478typedef struct rb_time_struct rb_time_t;
 479
 480#define MAX_NEST        5
 481
 482/*
 483 * head_page == tail_page && head == tail then buffer is empty.
 484 */
 485struct ring_buffer_per_cpu {
 486        int                             cpu;
 487        atomic_t                        record_disabled;
 488        atomic_t                        resize_disabled;
 489        struct trace_buffer     *buffer;
 490        raw_spinlock_t                  reader_lock;    /* serialize readers */
 491        arch_spinlock_t                 lock;
 492        struct lock_class_key           lock_key;
 493        struct buffer_data_page         *free_page;
 494        unsigned long                   nr_pages;
 495        unsigned int                    current_context;
 496        struct list_head                *pages;
 497        struct buffer_page              *head_page;     /* read from head */
 498        struct buffer_page              *tail_page;     /* write to tail */
 499        struct buffer_page              *commit_page;   /* committed pages */
 500        struct buffer_page              *reader_page;
 501        unsigned long                   lost_events;
 502        unsigned long                   last_overrun;
 503        unsigned long                   nest;
 504        local_t                         entries_bytes;
 505        local_t                         entries;
 506        local_t                         overrun;
 507        local_t                         commit_overrun;
 508        local_t                         dropped_events;
 509        local_t                         committing;
 510        local_t                         commits;
 511        local_t                         pages_touched;
 512        local_t                         pages_read;
 513        long                            last_pages_touch;
 514        size_t                          shortest_full;
 515        unsigned long                   read;
 516        unsigned long                   read_bytes;
 517        rb_time_t                       write_stamp;
 518        rb_time_t                       before_stamp;
 519        u64                             event_stamp[MAX_NEST];
 520        u64                             read_stamp;
 521        /* ring buffer pages to update, > 0 to add, < 0 to remove */
 522        long                            nr_pages_to_update;
 523        struct list_head                new_pages; /* new pages to add */
 524        struct work_struct              update_pages_work;
 525        struct completion               update_done;
 526
 527        struct rb_irq_work              irq_work;
 528};
 529
 530struct trace_buffer {
 531        unsigned                        flags;
 532        int                             cpus;
 533        atomic_t                        record_disabled;
 534        cpumask_var_t                   cpumask;
 535
 536        struct lock_class_key           *reader_lock_key;
 537
 538        struct mutex                    mutex;
 539
 540        struct ring_buffer_per_cpu      **buffers;
 541
 542        struct hlist_node               node;
 543        u64                             (*clock)(void);
 544
 545        struct rb_irq_work              irq_work;
 546        bool                            time_stamp_abs;
 547};
 548
 549struct ring_buffer_iter {
 550        struct ring_buffer_per_cpu      *cpu_buffer;
 551        unsigned long                   head;
 552        unsigned long                   next_event;
 553        struct buffer_page              *head_page;
 554        struct buffer_page              *cache_reader_page;
 555        unsigned long                   cache_read;
 556        u64                             read_stamp;
 557        u64                             page_stamp;
 558        struct ring_buffer_event        *event;
 559        int                             missed_events;
 560};
 561
 562#ifdef RB_TIME_32
 563
 564/*
 565 * On 32 bit machines, local64_t is very expensive. As the ring
 566 * buffer doesn't need all the features of a true 64 bit atomic,
 567 * on 32 bit, it uses these functions (64 still uses local64_t).
 568 *
 569 * For the ring buffer, 64 bit required operations for the time is
 570 * the following:
 571 *
 572 *  - Only need 59 bits (uses 60 to make it even).
 573 *  - Reads may fail if it interrupted a modification of the time stamp.
 574 *      It will succeed if it did not interrupt another write even if
 575 *      the read itself is interrupted by a write.
 576 *      It returns whether it was successful or not.
 577 *
 578 *  - Writes always succeed and will overwrite other writes and writes
 579 *      that were done by events interrupting the current write.
 580 *
 581 *  - A write followed by a read of the same time stamp will always succeed,
 582 *      but may not contain the same value.
 583 *
 584 *  - A cmpxchg will fail if it interrupted another write or cmpxchg.
 585 *      Other than that, it acts like a normal cmpxchg.
 586 *
 587 * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
 588 *  (bottom being the least significant 30 bits of the 60 bit time stamp).
 589 *
 590 * The two most significant bits of each half holds a 2 bit counter (0-3).
 591 * Each update will increment this counter by one.
 592 * When reading the top and bottom, if the two counter bits match then the
 593 *  top and bottom together make a valid 60 bit number.
 594 */
 595#define RB_TIME_SHIFT   30
 596#define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
 597
 598static inline int rb_time_cnt(unsigned long val)
 599{
 600        return (val >> RB_TIME_SHIFT) & 3;
 601}
 602
 603static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
 604{
 605        u64 val;
 606
 607        val = top & RB_TIME_VAL_MASK;
 608        val <<= RB_TIME_SHIFT;
 609        val |= bottom & RB_TIME_VAL_MASK;
 610
 611        return val;
 612}
 613
 614static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
 615{
 616        unsigned long top, bottom;
 617        unsigned long c;
 618
 619        /*
 620         * If the read is interrupted by a write, then the cnt will
 621         * be different. Loop until both top and bottom have been read
 622         * without interruption.
 623         */
 624        do {
 625                c = local_read(&t->cnt);
 626                top = local_read(&t->top);
 627                bottom = local_read(&t->bottom);
 628        } while (c != local_read(&t->cnt));
 629
 630        *cnt = rb_time_cnt(top);
 631
 632        /* If top and bottom counts don't match, this interrupted a write */
 633        if (*cnt != rb_time_cnt(bottom))
 634                return false;
 635
 636        *ret = rb_time_val(top, bottom);
 637        return true;
 638}
 639
 640static bool rb_time_read(rb_time_t *t, u64 *ret)
 641{
 642        unsigned long cnt;
 643
 644        return __rb_time_read(t, ret, &cnt);
 645}
 646
 647static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
 648{
 649        return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
 650}
 651
 652static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom)
 653{
 654        *top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
 655        *bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
 656}
 657
 658static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
 659{
 660        val = rb_time_val_cnt(val, cnt);
 661        local_set(t, val);
 662}
 663
 664static void rb_time_set(rb_time_t *t, u64 val)
 665{
 666        unsigned long cnt, top, bottom;
 667
 668        rb_time_split(val, &top, &bottom);
 669
 670        /* Writes always succeed with a valid number even if it gets interrupted. */
 671        do {
 672                cnt = local_inc_return(&t->cnt);
 673                rb_time_val_set(&t->top, top, cnt);
 674                rb_time_val_set(&t->bottom, bottom, cnt);
 675        } while (cnt != local_read(&t->cnt));
 676}
 677
 678static inline bool
 679rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
 680{
 681        unsigned long ret;
 682
 683        ret = local_cmpxchg(l, expect, set);
 684        return ret == expect;
 685}
 686
 687static int rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
 688{
 689        unsigned long cnt, top, bottom;
 690        unsigned long cnt2, top2, bottom2;
 691        u64 val;
 692
 693        /* The cmpxchg always fails if it interrupted an update */
 694         if (!__rb_time_read(t, &val, &cnt2))
 695                 return false;
 696
 697         if (val != expect)
 698                 return false;
 699
 700         cnt = local_read(&t->cnt);
 701         if ((cnt & 3) != cnt2)
 702                 return false;
 703
 704         cnt2 = cnt + 1;
 705
 706         rb_time_split(val, &top, &bottom);
 707         top = rb_time_val_cnt(top, cnt);
 708         bottom = rb_time_val_cnt(bottom, cnt);
 709
 710         rb_time_split(set, &top2, &bottom2);
 711         top2 = rb_time_val_cnt(top2, cnt2);
 712         bottom2 = rb_time_val_cnt(bottom2, cnt2);
 713
 714        if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
 715                return false;
 716        if (!rb_time_read_cmpxchg(&t->top, top, top2))
 717                return false;
 718        if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
 719                return false;
 720        return true;
 721}
 722
 723#else /* 64 bits */
 724
 725/* local64_t always succeeds */
 726
 727static inline bool rb_time_read(rb_time_t *t, u64 *ret)
 728{
 729        *ret = local64_read(&t->time);
 730        return true;
 731}
 732static void rb_time_set(rb_time_t *t, u64 val)
 733{
 734        local64_set(&t->time, val);
 735}
 736
 737static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
 738{
 739        u64 val;
 740        val = local64_cmpxchg(&t->time, expect, set);
 741        return val == expect;
 742}
 743#endif
 744
 745/*
 746 * Enable this to make sure that the event passed to
 747 * ring_buffer_event_time_stamp() is not committed and also
 748 * is on the buffer that it passed in.
 749 */
 750//#define RB_VERIFY_EVENT
 751#ifdef RB_VERIFY_EVENT
 752static struct list_head *rb_list_head(struct list_head *list);
 753static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
 754                         void *event)
 755{
 756        struct buffer_page *page = cpu_buffer->commit_page;
 757        struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
 758        struct list_head *next;
 759        long commit, write;
 760        unsigned long addr = (unsigned long)event;
 761        bool done = false;
 762        int stop = 0;
 763
 764        /* Make sure the event exists and is not committed yet */
 765        do {
 766                if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
 767                        done = true;
 768                commit = local_read(&page->page->commit);
 769                write = local_read(&page->write);
 770                if (addr >= (unsigned long)&page->page->data[commit] &&
 771                    addr < (unsigned long)&page->page->data[write])
 772                        return;
 773
 774                next = rb_list_head(page->list.next);
 775                page = list_entry(next, struct buffer_page, list);
 776        } while (!done);
 777        WARN_ON_ONCE(1);
 778}
 779#else
 780static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
 781                         void *event)
 782{
 783}
 784#endif
 785
 786
 787static inline u64 rb_time_stamp(struct trace_buffer *buffer);
 788
 789/**
 790 * ring_buffer_event_time_stamp - return the event's current time stamp
 791 * @buffer: The buffer that the event is on
 792 * @event: the event to get the time stamp of
 793 *
 794 * Note, this must be called after @event is reserved, and before it is
 795 * committed to the ring buffer. And must be called from the same
 796 * context where the event was reserved (normal, softirq, irq, etc).
 797 *
 798 * Returns the time stamp associated with the current event.
 799 * If the event has an extended time stamp, then that is used as
 800 * the time stamp to return.
 801 * In the highly unlikely case that the event was nested more than
 802 * the max nesting, then the write_stamp of the buffer is returned,
 803 * otherwise  current time is returned, but that really neither of
 804 * the last two cases should ever happen.
 805 */
 806u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
 807                                 struct ring_buffer_event *event)
 808{
 809        struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
 810        unsigned int nest;
 811        u64 ts;
 812
 813        /* If the event includes an absolute time, then just use that */
 814        if (event->type_len == RINGBUF_TYPE_TIME_STAMP)
 815                return rb_event_time_stamp(event);
 816
 817        nest = local_read(&cpu_buffer->committing);
 818        verify_event(cpu_buffer, event);
 819        if (WARN_ON_ONCE(!nest))
 820                goto fail;
 821
 822        /* Read the current saved nesting level time stamp */
 823        if (likely(--nest < MAX_NEST))
 824                return cpu_buffer->event_stamp[nest];
 825
 826        /* Shouldn't happen, warn if it does */
 827        WARN_ONCE(1, "nest (%d) greater than max", nest);
 828
 829 fail:
 830        /* Can only fail on 32 bit */
 831        if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
 832                /* Screw it, just read the current time */
 833                ts = rb_time_stamp(cpu_buffer->buffer);
 834
 835        return ts;
 836}
 837
 838/**
 839 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
 840 * @buffer: The ring_buffer to get the number of pages from
 841 * @cpu: The cpu of the ring_buffer to get the number of pages from
 842 *
 843 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
 844 */
 845size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
 846{
 847        return buffer->buffers[cpu]->nr_pages;
 848}
 849
 850/**
 851 * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
 852 * @buffer: The ring_buffer to get the number of pages from
 853 * @cpu: The cpu of the ring_buffer to get the number of pages from
 854 *
 855 * Returns the number of pages that have content in the ring buffer.
 856 */
 857size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
 858{
 859        size_t read;
 860        size_t cnt;
 861
 862        read = local_read(&buffer->buffers[cpu]->pages_read);
 863        cnt = local_read(&buffer->buffers[cpu]->pages_touched);
 864        /* The reader can read an empty page, but not more than that */
 865        if (cnt < read) {
 866                WARN_ON_ONCE(read > cnt + 1);
 867                return 0;
 868        }
 869
 870        return cnt - read;
 871}
 872
 873/*
 874 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
 875 *
 876 * Schedules a delayed work to wake up any task that is blocked on the
 877 * ring buffer waiters queue.
 878 */
 879static void rb_wake_up_waiters(struct irq_work *work)
 880{
 881        struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
 882
 883        wake_up_all(&rbwork->waiters);
 884        if (rbwork->wakeup_full) {
 885                rbwork->wakeup_full = false;
 886                wake_up_all(&rbwork->full_waiters);
 887        }
 888}
 889
 890/**
 891 * ring_buffer_wait - wait for input to the ring buffer
 892 * @buffer: buffer to wait on
 893 * @cpu: the cpu buffer to wait on
 894 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
 895 *
 896 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
 897 * as data is added to any of the @buffer's cpu buffers. Otherwise
 898 * it will wait for data to be added to a specific cpu buffer.
 899 */
 900int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
 901{
 902        struct ring_buffer_per_cpu *cpu_buffer;
 903        DEFINE_WAIT(wait);
 904        struct rb_irq_work *work;
 905        int ret = 0;
 906
 907        /*
 908         * Depending on what the caller is waiting for, either any
 909         * data in any cpu buffer, or a specific buffer, put the
 910         * caller on the appropriate wait queue.
 911         */
 912        if (cpu == RING_BUFFER_ALL_CPUS) {
 913                work = &buffer->irq_work;
 914                /* Full only makes sense on per cpu reads */
 915                full = 0;
 916        } else {
 917                if (!cpumask_test_cpu(cpu, buffer->cpumask))
 918                        return -ENODEV;
 919                cpu_buffer = buffer->buffers[cpu];
 920                work = &cpu_buffer->irq_work;
 921        }
 922
 923
 924        while (true) {
 925                if (full)
 926                        prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
 927                else
 928                        prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
 929
 930                /*
 931                 * The events can happen in critical sections where
 932                 * checking a work queue can cause deadlocks.
 933                 * After adding a task to the queue, this flag is set
 934                 * only to notify events to try to wake up the queue
 935                 * using irq_work.
 936                 *
 937                 * We don't clear it even if the buffer is no longer
 938                 * empty. The flag only causes the next event to run
 939                 * irq_work to do the work queue wake up. The worse
 940                 * that can happen if we race with !trace_empty() is that
 941                 * an event will cause an irq_work to try to wake up
 942                 * an empty queue.
 943                 *
 944                 * There's no reason to protect this flag either, as
 945                 * the work queue and irq_work logic will do the necessary
 946                 * synchronization for the wake ups. The only thing
 947                 * that is necessary is that the wake up happens after
 948                 * a task has been queued. It's OK for spurious wake ups.
 949                 */
 950                if (full)
 951                        work->full_waiters_pending = true;
 952                else
 953                        work->waiters_pending = true;
 954
 955                if (signal_pending(current)) {
 956                        ret = -EINTR;
 957                        break;
 958                }
 959
 960                if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
 961                        break;
 962
 963                if (cpu != RING_BUFFER_ALL_CPUS &&
 964                    !ring_buffer_empty_cpu(buffer, cpu)) {
 965                        unsigned long flags;
 966                        bool pagebusy;
 967                        size_t nr_pages;
 968                        size_t dirty;
 969
 970                        if (!full)
 971                                break;
 972
 973                        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
 974                        pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
 975                        nr_pages = cpu_buffer->nr_pages;
 976                        dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
 977                        if (!cpu_buffer->shortest_full ||
 978                            cpu_buffer->shortest_full < full)
 979                                cpu_buffer->shortest_full = full;
 980                        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
 981                        if (!pagebusy &&
 982                            (!nr_pages || (dirty * 100) > full * nr_pages))
 983                                break;
 984                }
 985
 986                schedule();
 987        }
 988
 989        if (full)
 990                finish_wait(&work->full_waiters, &wait);
 991        else
 992                finish_wait(&work->waiters, &wait);
 993
 994        return ret;
 995}
 996
 997/**
 998 * ring_buffer_poll_wait - poll on buffer input
 999 * @buffer: buffer to wait on
1000 * @cpu: the cpu buffer to wait on
1001 * @filp: the file descriptor
1002 * @poll_table: The poll descriptor
1003 *
1004 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1005 * as data is added to any of the @buffer's cpu buffers. Otherwise
1006 * it will wait for data to be added to a specific cpu buffer.
1007 *
1008 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1009 * zero otherwise.
1010 */
1011__poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1012                          struct file *filp, poll_table *poll_table)
1013{
1014        struct ring_buffer_per_cpu *cpu_buffer;
1015        struct rb_irq_work *work;
1016
1017        if (cpu == RING_BUFFER_ALL_CPUS)
1018                work = &buffer->irq_work;
1019        else {
1020                if (!cpumask_test_cpu(cpu, buffer->cpumask))
1021                        return -EINVAL;
1022
1023                cpu_buffer = buffer->buffers[cpu];
1024                work = &cpu_buffer->irq_work;
1025        }
1026
1027        poll_wait(filp, &work->waiters, poll_table);
1028        work->waiters_pending = true;
1029        /*
1030         * There's a tight race between setting the waiters_pending and
1031         * checking if the ring buffer is empty.  Once the waiters_pending bit
1032         * is set, the next event will wake the task up, but we can get stuck
1033         * if there's only a single event in.
1034         *
1035         * FIXME: Ideally, we need a memory barrier on the writer side as well,
1036         * but adding a memory barrier to all events will cause too much of a
1037         * performance hit in the fast path.  We only need a memory barrier when
1038         * the buffer goes from empty to having content.  But as this race is
1039         * extremely small, and it's not a problem if another event comes in, we
1040         * will fix it later.
1041         */
1042        smp_mb();
1043
1044        if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1045            (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1046                return EPOLLIN | EPOLLRDNORM;
1047        return 0;
1048}
1049
1050/* buffer may be either ring_buffer or ring_buffer_per_cpu */
1051#define RB_WARN_ON(b, cond)                                             \
1052        ({                                                              \
1053                int _____ret = unlikely(cond);                          \
1054                if (_____ret) {                                         \
1055                        if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1056                                struct ring_buffer_per_cpu *__b =       \
1057                                        (void *)b;                      \
1058                                atomic_inc(&__b->buffer->record_disabled); \
1059                        } else                                          \
1060                                atomic_inc(&b->record_disabled);        \
1061                        WARN_ON(1);                                     \
1062                }                                                       \
1063                _____ret;                                               \
1064        })
1065
1066/* Up this if you want to test the TIME_EXTENTS and normalization */
1067#define DEBUG_SHIFT 0
1068
1069static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1070{
1071        u64 ts;
1072
1073        /* Skip retpolines :-( */
1074        if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1075                ts = trace_clock_local();
1076        else
1077                ts = buffer->clock();
1078
1079        /* shift to debug/test normalization and TIME_EXTENTS */
1080        return ts << DEBUG_SHIFT;
1081}
1082
1083u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1084{
1085        u64 time;
1086
1087        preempt_disable_notrace();
1088        time = rb_time_stamp(buffer);
1089        preempt_enable_notrace();
1090
1091        return time;
1092}
1093EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1094
1095void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1096                                      int cpu, u64 *ts)
1097{
1098        /* Just stupid testing the normalize function and deltas */
1099        *ts >>= DEBUG_SHIFT;
1100}
1101EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1102
1103/*
1104 * Making the ring buffer lockless makes things tricky.
1105 * Although writes only happen on the CPU that they are on,
1106 * and they only need to worry about interrupts. Reads can
1107 * happen on any CPU.
1108 *
1109 * The reader page is always off the ring buffer, but when the
1110 * reader finishes with a page, it needs to swap its page with
1111 * a new one from the buffer. The reader needs to take from
1112 * the head (writes go to the tail). But if a writer is in overwrite
1113 * mode and wraps, it must push the head page forward.
1114 *
1115 * Here lies the problem.
1116 *
1117 * The reader must be careful to replace only the head page, and
1118 * not another one. As described at the top of the file in the
1119 * ASCII art, the reader sets its old page to point to the next
1120 * page after head. It then sets the page after head to point to
1121 * the old reader page. But if the writer moves the head page
1122 * during this operation, the reader could end up with the tail.
1123 *
1124 * We use cmpxchg to help prevent this race. We also do something
1125 * special with the page before head. We set the LSB to 1.
1126 *
1127 * When the writer must push the page forward, it will clear the
1128 * bit that points to the head page, move the head, and then set
1129 * the bit that points to the new head page.
1130 *
1131 * We also don't want an interrupt coming in and moving the head
1132 * page on another writer. Thus we use the second LSB to catch
1133 * that too. Thus:
1134 *
1135 * head->list->prev->next        bit 1          bit 0
1136 *                              -------        -------
1137 * Normal page                     0              0
1138 * Points to head page             0              1
1139 * New head page                   1              0
1140 *
1141 * Note we can not trust the prev pointer of the head page, because:
1142 *
1143 * +----+       +-----+        +-----+
1144 * |    |------>|  T  |---X--->|  N  |
1145 * |    |<------|     |        |     |
1146 * +----+       +-----+        +-----+
1147 *   ^                           ^ |
1148 *   |          +-----+          | |
1149 *   +----------|  R  |----------+ |
1150 *              |     |<-----------+
1151 *              +-----+
1152 *
1153 * Key:  ---X-->  HEAD flag set in pointer
1154 *         T      Tail page
1155 *         R      Reader page
1156 *         N      Next page
1157 *
1158 * (see __rb_reserve_next() to see where this happens)
1159 *
1160 *  What the above shows is that the reader just swapped out
1161 *  the reader page with a page in the buffer, but before it
1162 *  could make the new header point back to the new page added
1163 *  it was preempted by a writer. The writer moved forward onto
1164 *  the new page added by the reader and is about to move forward
1165 *  again.
1166 *
1167 *  You can see, it is legitimate for the previous pointer of
1168 *  the head (or any page) not to point back to itself. But only
1169 *  temporarily.
1170 */
1171
1172#define RB_PAGE_NORMAL          0UL
1173#define RB_PAGE_HEAD            1UL
1174#define RB_PAGE_UPDATE          2UL
1175
1176
1177#define RB_FLAG_MASK            3UL
1178
1179/* PAGE_MOVED is not part of the mask */
1180#define RB_PAGE_MOVED           4UL
1181
1182/*
1183 * rb_list_head - remove any bit
1184 */
1185static struct list_head *rb_list_head(struct list_head *list)
1186{
1187        unsigned long val = (unsigned long)list;
1188
1189        return (struct list_head *)(val & ~RB_FLAG_MASK);
1190}
1191
1192/*
1193 * rb_is_head_page - test if the given page is the head page
1194 *
1195 * Because the reader may move the head_page pointer, we can
1196 * not trust what the head page is (it may be pointing to
1197 * the reader page). But if the next page is a header page,
1198 * its flags will be non zero.
1199 */
1200static inline int
1201rb_is_head_page(struct buffer_page *page, struct list_head *list)
1202{
1203        unsigned long val;
1204
1205        val = (unsigned long)list->next;
1206
1207        if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1208                return RB_PAGE_MOVED;
1209
1210        return val & RB_FLAG_MASK;
1211}
1212
1213/*
1214 * rb_is_reader_page
1215 *
1216 * The unique thing about the reader page, is that, if the
1217 * writer is ever on it, the previous pointer never points
1218 * back to the reader page.
1219 */
1220static bool rb_is_reader_page(struct buffer_page *page)
1221{
1222        struct list_head *list = page->list.prev;
1223
1224        return rb_list_head(list->next) != &page->list;
1225}
1226
1227/*
1228 * rb_set_list_to_head - set a list_head to be pointing to head.
1229 */
1230static void rb_set_list_to_head(struct list_head *list)
1231{
1232        unsigned long *ptr;
1233
1234        ptr = (unsigned long *)&list->next;
1235        *ptr |= RB_PAGE_HEAD;
1236        *ptr &= ~RB_PAGE_UPDATE;
1237}
1238
1239/*
1240 * rb_head_page_activate - sets up head page
1241 */
1242static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1243{
1244        struct buffer_page *head;
1245
1246        head = cpu_buffer->head_page;
1247        if (!head)
1248                return;
1249
1250        /*
1251         * Set the previous list pointer to have the HEAD flag.
1252         */
1253        rb_set_list_to_head(head->list.prev);
1254}
1255
1256static void rb_list_head_clear(struct list_head *list)
1257{
1258        unsigned long *ptr = (unsigned long *)&list->next;
1259
1260        *ptr &= ~RB_FLAG_MASK;
1261}
1262
1263/*
1264 * rb_head_page_deactivate - clears head page ptr (for free list)
1265 */
1266static void
1267rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1268{
1269        struct list_head *hd;
1270
1271        /* Go through the whole list and clear any pointers found. */
1272        rb_list_head_clear(cpu_buffer->pages);
1273
1274        list_for_each(hd, cpu_buffer->pages)
1275                rb_list_head_clear(hd);
1276}
1277
1278static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1279                            struct buffer_page *head,
1280                            struct buffer_page *prev,
1281                            int old_flag, int new_flag)
1282{
1283        struct list_head *list;
1284        unsigned long val = (unsigned long)&head->list;
1285        unsigned long ret;
1286
1287        list = &prev->list;
1288
1289        val &= ~RB_FLAG_MASK;
1290
1291        ret = cmpxchg((unsigned long *)&list->next,
1292                      val | old_flag, val | new_flag);
1293
1294        /* check if the reader took the page */
1295        if ((ret & ~RB_FLAG_MASK) != val)
1296                return RB_PAGE_MOVED;
1297
1298        return ret & RB_FLAG_MASK;
1299}
1300
1301static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1302                                   struct buffer_page *head,
1303                                   struct buffer_page *prev,
1304                                   int old_flag)
1305{
1306        return rb_head_page_set(cpu_buffer, head, prev,
1307                                old_flag, RB_PAGE_UPDATE);
1308}
1309
1310static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1311                                 struct buffer_page *head,
1312                                 struct buffer_page *prev,
1313                                 int old_flag)
1314{
1315        return rb_head_page_set(cpu_buffer, head, prev,
1316                                old_flag, RB_PAGE_HEAD);
1317}
1318
1319static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1320                                   struct buffer_page *head,
1321                                   struct buffer_page *prev,
1322                                   int old_flag)
1323{
1324        return rb_head_page_set(cpu_buffer, head, prev,
1325                                old_flag, RB_PAGE_NORMAL);
1326}
1327
1328static inline void rb_inc_page(struct buffer_page **bpage)
1329{
1330        struct list_head *p = rb_list_head((*bpage)->list.next);
1331
1332        *bpage = list_entry(p, struct buffer_page, list);
1333}
1334
1335static struct buffer_page *
1336rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1337{
1338        struct buffer_page *head;
1339        struct buffer_page *page;
1340        struct list_head *list;
1341        int i;
1342
1343        if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1344                return NULL;
1345
1346        /* sanity check */
1347        list = cpu_buffer->pages;
1348        if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1349                return NULL;
1350
1351        page = head = cpu_buffer->head_page;
1352        /*
1353         * It is possible that the writer moves the header behind
1354         * where we started, and we miss in one loop.
1355         * A second loop should grab the header, but we'll do
1356         * three loops just because I'm paranoid.
1357         */
1358        for (i = 0; i < 3; i++) {
1359                do {
1360                        if (rb_is_head_page(page, page->list.prev)) {
1361                                cpu_buffer->head_page = page;
1362                                return page;
1363                        }
1364                        rb_inc_page(&page);
1365                } while (page != head);
1366        }
1367
1368        RB_WARN_ON(cpu_buffer, 1);
1369
1370        return NULL;
1371}
1372
1373static int rb_head_page_replace(struct buffer_page *old,
1374                                struct buffer_page *new)
1375{
1376        unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1377        unsigned long val;
1378        unsigned long ret;
1379
1380        val = *ptr & ~RB_FLAG_MASK;
1381        val |= RB_PAGE_HEAD;
1382
1383        ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1384
1385        return ret == val;
1386}
1387
1388/*
1389 * rb_tail_page_update - move the tail page forward
1390 */
1391static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1392                               struct buffer_page *tail_page,
1393                               struct buffer_page *next_page)
1394{
1395        unsigned long old_entries;
1396        unsigned long old_write;
1397
1398        /*
1399         * The tail page now needs to be moved forward.
1400         *
1401         * We need to reset the tail page, but without messing
1402         * with possible erasing of data brought in by interrupts
1403         * that have moved the tail page and are currently on it.
1404         *
1405         * We add a counter to the write field to denote this.
1406         */
1407        old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1408        old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1409
1410        local_inc(&cpu_buffer->pages_touched);
1411        /*
1412         * Just make sure we have seen our old_write and synchronize
1413         * with any interrupts that come in.
1414         */
1415        barrier();
1416
1417        /*
1418         * If the tail page is still the same as what we think
1419         * it is, then it is up to us to update the tail
1420         * pointer.
1421         */
1422        if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1423                /* Zero the write counter */
1424                unsigned long val = old_write & ~RB_WRITE_MASK;
1425                unsigned long eval = old_entries & ~RB_WRITE_MASK;
1426
1427                /*
1428                 * This will only succeed if an interrupt did
1429                 * not come in and change it. In which case, we
1430                 * do not want to modify it.
1431                 *
1432                 * We add (void) to let the compiler know that we do not care
1433                 * about the return value of these functions. We use the
1434                 * cmpxchg to only update if an interrupt did not already
1435                 * do it for us. If the cmpxchg fails, we don't care.
1436                 */
1437                (void)local_cmpxchg(&next_page->write, old_write, val);
1438                (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1439
1440                /*
1441                 * No need to worry about races with clearing out the commit.
1442                 * it only can increment when a commit takes place. But that
1443                 * only happens in the outer most nested commit.
1444                 */
1445                local_set(&next_page->page->commit, 0);
1446
1447                /* Again, either we update tail_page or an interrupt does */
1448                (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1449        }
1450}
1451
1452static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1453                          struct buffer_page *bpage)
1454{
1455        unsigned long val = (unsigned long)bpage;
1456
1457        if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1458                return 1;
1459
1460        return 0;
1461}
1462
1463/**
1464 * rb_check_list - make sure a pointer to a list has the last bits zero
1465 */
1466static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1467                         struct list_head *list)
1468{
1469        if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1470                return 1;
1471        if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1472                return 1;
1473        return 0;
1474}
1475
1476/**
1477 * rb_check_pages - integrity check of buffer pages
1478 * @cpu_buffer: CPU buffer with pages to test
1479 *
1480 * As a safety measure we check to make sure the data pages have not
1481 * been corrupted.
1482 */
1483static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1484{
1485        struct list_head *head = cpu_buffer->pages;
1486        struct buffer_page *bpage, *tmp;
1487
1488        /* Reset the head page if it exists */
1489        if (cpu_buffer->head_page)
1490                rb_set_head_page(cpu_buffer);
1491
1492        rb_head_page_deactivate(cpu_buffer);
1493
1494        if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1495                return -1;
1496        if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1497                return -1;
1498
1499        if (rb_check_list(cpu_buffer, head))
1500                return -1;
1501
1502        list_for_each_entry_safe(bpage, tmp, head, list) {
1503                if (RB_WARN_ON(cpu_buffer,
1504                               bpage->list.next->prev != &bpage->list))
1505                        return -1;
1506                if (RB_WARN_ON(cpu_buffer,
1507                               bpage->list.prev->next != &bpage->list))
1508                        return -1;
1509                if (rb_check_list(cpu_buffer, &bpage->list))
1510                        return -1;
1511        }
1512
1513        rb_head_page_activate(cpu_buffer);
1514
1515        return 0;
1516}
1517
1518static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1519                long nr_pages, struct list_head *pages)
1520{
1521        struct buffer_page *bpage, *tmp;
1522        bool user_thread = current->mm != NULL;
1523        gfp_t mflags;
1524        long i;
1525
1526        /*
1527         * Check if the available memory is there first.
1528         * Note, si_mem_available() only gives us a rough estimate of available
1529         * memory. It may not be accurate. But we don't care, we just want
1530         * to prevent doing any allocation when it is obvious that it is
1531         * not going to succeed.
1532         */
1533        i = si_mem_available();
1534        if (i < nr_pages)
1535                return -ENOMEM;
1536
1537        /*
1538         * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1539         * gracefully without invoking oom-killer and the system is not
1540         * destabilized.
1541         */
1542        mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1543
1544        /*
1545         * If a user thread allocates too much, and si_mem_available()
1546         * reports there's enough memory, even though there is not.
1547         * Make sure the OOM killer kills this thread. This can happen
1548         * even with RETRY_MAYFAIL because another task may be doing
1549         * an allocation after this task has taken all memory.
1550         * This is the task the OOM killer needs to take out during this
1551         * loop, even if it was triggered by an allocation somewhere else.
1552         */
1553        if (user_thread)
1554                set_current_oom_origin();
1555        for (i = 0; i < nr_pages; i++) {
1556                struct page *page;
1557
1558                bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1559                                    mflags, cpu_to_node(cpu_buffer->cpu));
1560                if (!bpage)
1561                        goto free_pages;
1562
1563                rb_check_bpage(cpu_buffer, bpage);
1564
1565                list_add(&bpage->list, pages);
1566
1567                page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
1568                if (!page)
1569                        goto free_pages;
1570                bpage->page = page_address(page);
1571                rb_init_page(bpage->page);
1572
1573                if (user_thread && fatal_signal_pending(current))
1574                        goto free_pages;
1575        }
1576        if (user_thread)
1577                clear_current_oom_origin();
1578
1579        return 0;
1580
1581free_pages:
1582        list_for_each_entry_safe(bpage, tmp, pages, list) {
1583                list_del_init(&bpage->list);
1584                free_buffer_page(bpage);
1585        }
1586        if (user_thread)
1587                clear_current_oom_origin();
1588
1589        return -ENOMEM;
1590}
1591
1592static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1593                             unsigned long nr_pages)
1594{
1595        LIST_HEAD(pages);
1596
1597        WARN_ON(!nr_pages);
1598
1599        if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1600                return -ENOMEM;
1601
1602        /*
1603         * The ring buffer page list is a circular list that does not
1604         * start and end with a list head. All page list items point to
1605         * other pages.
1606         */
1607        cpu_buffer->pages = pages.next;
1608        list_del(&pages);
1609
1610        cpu_buffer->nr_pages = nr_pages;
1611
1612        rb_check_pages(cpu_buffer);
1613
1614        return 0;
1615}
1616
1617static struct ring_buffer_per_cpu *
1618rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1619{
1620        struct ring_buffer_per_cpu *cpu_buffer;
1621        struct buffer_page *bpage;
1622        struct page *page;
1623        int ret;
1624
1625        cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1626                                  GFP_KERNEL, cpu_to_node(cpu));
1627        if (!cpu_buffer)
1628                return NULL;
1629
1630        cpu_buffer->cpu = cpu;
1631        cpu_buffer->buffer = buffer;
1632        raw_spin_lock_init(&cpu_buffer->reader_lock);
1633        lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1634        cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1635        INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1636        init_completion(&cpu_buffer->update_done);
1637        init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1638        init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1639        init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1640
1641        bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1642                            GFP_KERNEL, cpu_to_node(cpu));
1643        if (!bpage)
1644                goto fail_free_buffer;
1645
1646        rb_check_bpage(cpu_buffer, bpage);
1647
1648        cpu_buffer->reader_page = bpage;
1649        page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1650        if (!page)
1651                goto fail_free_reader;
1652        bpage->page = page_address(page);
1653        rb_init_page(bpage->page);
1654
1655        INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1656        INIT_LIST_HEAD(&cpu_buffer->new_pages);
1657
1658        ret = rb_allocate_pages(cpu_buffer, nr_pages);
1659        if (ret < 0)
1660                goto fail_free_reader;
1661
1662        cpu_buffer->head_page
1663                = list_entry(cpu_buffer->pages, struct buffer_page, list);
1664        cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1665
1666        rb_head_page_activate(cpu_buffer);
1667
1668        return cpu_buffer;
1669
1670 fail_free_reader:
1671        free_buffer_page(cpu_buffer->reader_page);
1672
1673 fail_free_buffer:
1674        kfree(cpu_buffer);
1675        return NULL;
1676}
1677
1678static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1679{
1680        struct list_head *head = cpu_buffer->pages;
1681        struct buffer_page *bpage, *tmp;
1682
1683        free_buffer_page(cpu_buffer->reader_page);
1684
1685        rb_head_page_deactivate(cpu_buffer);
1686
1687        if (head) {
1688                list_for_each_entry_safe(bpage, tmp, head, list) {
1689                        list_del_init(&bpage->list);
1690                        free_buffer_page(bpage);
1691                }
1692                bpage = list_entry(head, struct buffer_page, list);
1693                free_buffer_page(bpage);
1694        }
1695
1696        kfree(cpu_buffer);
1697}
1698
1699/**
1700 * __ring_buffer_alloc - allocate a new ring_buffer
1701 * @size: the size in bytes per cpu that is needed.
1702 * @flags: attributes to set for the ring buffer.
1703 * @key: ring buffer reader_lock_key.
1704 *
1705 * Currently the only flag that is available is the RB_FL_OVERWRITE
1706 * flag. This flag means that the buffer will overwrite old data
1707 * when the buffer wraps. If this flag is not set, the buffer will
1708 * drop data when the tail hits the head.
1709 */
1710struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1711                                        struct lock_class_key *key)
1712{
1713        struct trace_buffer *buffer;
1714        long nr_pages;
1715        int bsize;
1716        int cpu;
1717        int ret;
1718
1719        /* keep it in its own cache line */
1720        buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1721                         GFP_KERNEL);
1722        if (!buffer)
1723                return NULL;
1724
1725        if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1726                goto fail_free_buffer;
1727
1728        nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1729        buffer->flags = flags;
1730        buffer->clock = trace_clock_local;
1731        buffer->reader_lock_key = key;
1732
1733        init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1734        init_waitqueue_head(&buffer->irq_work.waiters);
1735
1736        /* need at least two pages */
1737        if (nr_pages < 2)
1738                nr_pages = 2;
1739
1740        buffer->cpus = nr_cpu_ids;
1741
1742        bsize = sizeof(void *) * nr_cpu_ids;
1743        buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1744                                  GFP_KERNEL);
1745        if (!buffer->buffers)
1746                goto fail_free_cpumask;
1747
1748        cpu = raw_smp_processor_id();
1749        cpumask_set_cpu(cpu, buffer->cpumask);
1750        buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1751        if (!buffer->buffers[cpu])
1752                goto fail_free_buffers;
1753
1754        ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1755        if (ret < 0)
1756                goto fail_free_buffers;
1757
1758        mutex_init(&buffer->mutex);
1759
1760        return buffer;
1761
1762 fail_free_buffers:
1763        for_each_buffer_cpu(buffer, cpu) {
1764                if (buffer->buffers[cpu])
1765                        rb_free_cpu_buffer(buffer->buffers[cpu]);
1766        }
1767        kfree(buffer->buffers);
1768
1769 fail_free_cpumask:
1770        free_cpumask_var(buffer->cpumask);
1771
1772 fail_free_buffer:
1773        kfree(buffer);
1774        return NULL;
1775}
1776EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1777
1778/**
1779 * ring_buffer_free - free a ring buffer.
1780 * @buffer: the buffer to free.
1781 */
1782void
1783ring_buffer_free(struct trace_buffer *buffer)
1784{
1785        int cpu;
1786
1787        cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1788
1789        for_each_buffer_cpu(buffer, cpu)
1790                rb_free_cpu_buffer(buffer->buffers[cpu]);
1791
1792        kfree(buffer->buffers);
1793        free_cpumask_var(buffer->cpumask);
1794
1795        kfree(buffer);
1796}
1797EXPORT_SYMBOL_GPL(ring_buffer_free);
1798
1799void ring_buffer_set_clock(struct trace_buffer *buffer,
1800                           u64 (*clock)(void))
1801{
1802        buffer->clock = clock;
1803}
1804
1805void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1806{
1807        buffer->time_stamp_abs = abs;
1808}
1809
1810bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1811{
1812        return buffer->time_stamp_abs;
1813}
1814
1815static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1816
1817static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1818{
1819        return local_read(&bpage->entries) & RB_WRITE_MASK;
1820}
1821
1822static inline unsigned long rb_page_write(struct buffer_page *bpage)
1823{
1824        return local_read(&bpage->write) & RB_WRITE_MASK;
1825}
1826
1827static int
1828rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1829{
1830        struct list_head *tail_page, *to_remove, *next_page;
1831        struct buffer_page *to_remove_page, *tmp_iter_page;
1832        struct buffer_page *last_page, *first_page;
1833        unsigned long nr_removed;
1834        unsigned long head_bit;
1835        int page_entries;
1836
1837        head_bit = 0;
1838
1839        raw_spin_lock_irq(&cpu_buffer->reader_lock);
1840        atomic_inc(&cpu_buffer->record_disabled);
1841        /*
1842         * We don't race with the readers since we have acquired the reader
1843         * lock. We also don't race with writers after disabling recording.
1844         * This makes it easy to figure out the first and the last page to be
1845         * removed from the list. We unlink all the pages in between including
1846         * the first and last pages. This is done in a busy loop so that we
1847         * lose the least number of traces.
1848         * The pages are freed after we restart recording and unlock readers.
1849         */
1850        tail_page = &cpu_buffer->tail_page->list;
1851
1852        /*
1853         * tail page might be on reader page, we remove the next page
1854         * from the ring buffer
1855         */
1856        if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1857                tail_page = rb_list_head(tail_page->next);
1858        to_remove = tail_page;
1859
1860        /* start of pages to remove */
1861        first_page = list_entry(rb_list_head(to_remove->next),
1862                                struct buffer_page, list);
1863
1864        for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1865                to_remove = rb_list_head(to_remove)->next;
1866                head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1867        }
1868
1869        next_page = rb_list_head(to_remove)->next;
1870
1871        /*
1872         * Now we remove all pages between tail_page and next_page.
1873         * Make sure that we have head_bit value preserved for the
1874         * next page
1875         */
1876        tail_page->next = (struct list_head *)((unsigned long)next_page |
1877                                                head_bit);
1878        next_page = rb_list_head(next_page);
1879        next_page->prev = tail_page;
1880
1881        /* make sure pages points to a valid page in the ring buffer */
1882        cpu_buffer->pages = next_page;
1883
1884        /* update head page */
1885        if (head_bit)
1886                cpu_buffer->head_page = list_entry(next_page,
1887                                                struct buffer_page, list);
1888
1889        /*
1890         * change read pointer to make sure any read iterators reset
1891         * themselves
1892         */
1893        cpu_buffer->read = 0;
1894
1895        /* pages are removed, resume tracing and then free the pages */
1896        atomic_dec(&cpu_buffer->record_disabled);
1897        raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1898
1899        RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1900
1901        /* last buffer page to remove */
1902        last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1903                                list);
1904        tmp_iter_page = first_page;
1905
1906        do {
1907                cond_resched();
1908
1909                to_remove_page = tmp_iter_page;
1910                rb_inc_page(&tmp_iter_page);
1911
1912                /* update the counters */
1913                page_entries = rb_page_entries(to_remove_page);
1914                if (page_entries) {
1915                        /*
1916                         * If something was added to this page, it was full
1917                         * since it is not the tail page. So we deduct the
1918                         * bytes consumed in ring buffer from here.
1919                         * Increment overrun to account for the lost events.
1920                         */
1921                        local_add(page_entries, &cpu_buffer->overrun);
1922                        local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1923                }
1924
1925                /*
1926                 * We have already removed references to this list item, just
1927                 * free up the buffer_page and its page
1928                 */
1929                free_buffer_page(to_remove_page);
1930                nr_removed--;
1931
1932        } while (to_remove_page != last_page);
1933
1934        RB_WARN_ON(cpu_buffer, nr_removed);
1935
1936        return nr_removed == 0;
1937}
1938
1939static int
1940rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1941{
1942        struct list_head *pages = &cpu_buffer->new_pages;
1943        int retries, success;
1944
1945        raw_spin_lock_irq(&cpu_buffer->reader_lock);
1946        /*
1947         * We are holding the reader lock, so the reader page won't be swapped
1948         * in the ring buffer. Now we are racing with the writer trying to
1949         * move head page and the tail page.
1950         * We are going to adapt the reader page update process where:
1951         * 1. We first splice the start and end of list of new pages between
1952         *    the head page and its previous page.
1953         * 2. We cmpxchg the prev_page->next to point from head page to the
1954         *    start of new pages list.
1955         * 3. Finally, we update the head->prev to the end of new list.
1956         *
1957         * We will try this process 10 times, to make sure that we don't keep
1958         * spinning.
1959         */
1960        retries = 10;
1961        success = 0;
1962        while (retries--) {
1963                struct list_head *head_page, *prev_page, *r;
1964                struct list_head *last_page, *first_page;
1965                struct list_head *head_page_with_bit;
1966
1967                head_page = &rb_set_head_page(cpu_buffer)->list;
1968                if (!head_page)
1969                        break;
1970                prev_page = head_page->prev;
1971
1972                first_page = pages->next;
1973                last_page  = pages->prev;
1974
1975                head_page_with_bit = (struct list_head *)
1976                                     ((unsigned long)head_page | RB_PAGE_HEAD);
1977
1978                last_page->next = head_page_with_bit;
1979                first_page->prev = prev_page;
1980
1981                r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1982
1983                if (r == head_page_with_bit) {
1984                        /*
1985                         * yay, we replaced the page pointer to our new list,
1986                         * now, we just have to update to head page's prev
1987                         * pointer to point to end of list
1988                         */
1989                        head_page->prev = last_page;
1990                        success = 1;
1991                        break;
1992                }
1993        }
1994
1995        if (success)
1996                INIT_LIST_HEAD(pages);
1997        /*
1998         * If we weren't successful in adding in new pages, warn and stop
1999         * tracing
2000         */
2001        RB_WARN_ON(cpu_buffer, !success);
2002        raw_spin_unlock_irq(&cpu_buffer->reader_lock);
2003
2004        /* free pages if they weren't inserted */
2005        if (!success) {
2006                struct buffer_page *bpage, *tmp;
2007                list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2008                                         list) {
2009                        list_del_init(&bpage->list);
2010                        free_buffer_page(bpage);
2011                }
2012        }
2013        return success;
2014}
2015
2016static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2017{
2018        int success;
2019
2020        if (cpu_buffer->nr_pages_to_update > 0)
2021                success = rb_insert_pages(cpu_buffer);
2022        else
2023                success = rb_remove_pages(cpu_buffer,
2024                                        -cpu_buffer->nr_pages_to_update);
2025
2026        if (success)
2027                cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2028}
2029
2030static void update_pages_handler(struct work_struct *work)
2031{
2032        struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2033                        struct ring_buffer_per_cpu, update_pages_work);
2034        rb_update_pages(cpu_buffer);
2035        complete(&cpu_buffer->update_done);
2036}
2037
2038/**
2039 * ring_buffer_resize - resize the ring buffer
2040 * @buffer: the buffer to resize.
2041 * @size: the new size.
2042 * @cpu_id: the cpu buffer to resize
2043 *
2044 * Minimum size is 2 * BUF_PAGE_SIZE.
2045 *
2046 * Returns 0 on success and < 0 on failure.
2047 */
2048int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2049                        int cpu_id)
2050{
2051        struct ring_buffer_per_cpu *cpu_buffer;
2052        unsigned long nr_pages;
2053        int cpu, err;
2054
2055        /*
2056         * Always succeed at resizing a non-existent buffer:
2057         */
2058        if (!buffer)
2059                return 0;
2060
2061        /* Make sure the requested buffer exists */
2062        if (cpu_id != RING_BUFFER_ALL_CPUS &&
2063            !cpumask_test_cpu(cpu_id, buffer->cpumask))
2064                return 0;
2065
2066        nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2067
2068        /* we need a minimum of two pages */
2069        if (nr_pages < 2)
2070                nr_pages = 2;
2071
2072        /* prevent another thread from changing buffer sizes */
2073        mutex_lock(&buffer->mutex);
2074
2075
2076        if (cpu_id == RING_BUFFER_ALL_CPUS) {
2077                /*
2078                 * Don't succeed if resizing is disabled, as a reader might be
2079                 * manipulating the ring buffer and is expecting a sane state while
2080                 * this is true.
2081                 */
2082                for_each_buffer_cpu(buffer, cpu) {
2083                        cpu_buffer = buffer->buffers[cpu];
2084                        if (atomic_read(&cpu_buffer->resize_disabled)) {
2085                                err = -EBUSY;
2086                                goto out_err_unlock;
2087                        }
2088                }
2089
2090                /* calculate the pages to update */
2091                for_each_buffer_cpu(buffer, cpu) {
2092                        cpu_buffer = buffer->buffers[cpu];
2093
2094                        cpu_buffer->nr_pages_to_update = nr_pages -
2095                                                        cpu_buffer->nr_pages;
2096                        /*
2097                         * nothing more to do for removing pages or no update
2098                         */
2099                        if (cpu_buffer->nr_pages_to_update <= 0)
2100                                continue;
2101                        /*
2102                         * to add pages, make sure all new pages can be
2103                         * allocated without receiving ENOMEM
2104                         */
2105                        INIT_LIST_HEAD(&cpu_buffer->new_pages);
2106                        if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2107                                                &cpu_buffer->new_pages)) {
2108                                /* not enough memory for new pages */
2109                                err = -ENOMEM;
2110                                goto out_err;
2111                        }
2112                }
2113
2114                cpus_read_lock();
2115                /*
2116                 * Fire off all the required work handlers
2117                 * We can't schedule on offline CPUs, but it's not necessary
2118                 * since we can change their buffer sizes without any race.
2119                 */
2120                for_each_buffer_cpu(buffer, cpu) {
2121                        cpu_buffer = buffer->buffers[cpu];
2122                        if (!cpu_buffer->nr_pages_to_update)
2123                                continue;
2124
2125                        /* Can't run something on an offline CPU. */
2126                        if (!cpu_online(cpu)) {
2127                                rb_update_pages(cpu_buffer);
2128                                cpu_buffer->nr_pages_to_update = 0;
2129                        } else {
2130                                schedule_work_on(cpu,
2131                                                &cpu_buffer->update_pages_work);
2132                        }
2133                }
2134
2135                /* wait for all the updates to complete */
2136                for_each_buffer_cpu(buffer, cpu) {
2137                        cpu_buffer = buffer->buffers[cpu];
2138                        if (!cpu_buffer->nr_pages_to_update)
2139                                continue;
2140
2141                        if (cpu_online(cpu))
2142                                wait_for_completion(&cpu_buffer->update_done);
2143                        cpu_buffer->nr_pages_to_update = 0;
2144                }
2145
2146                cpus_read_unlock();
2147        } else {
2148                cpu_buffer = buffer->buffers[cpu_id];
2149
2150                if (nr_pages == cpu_buffer->nr_pages)
2151                        goto out;
2152
2153                /*
2154                 * Don't succeed if resizing is disabled, as a reader might be
2155                 * manipulating the ring buffer and is expecting a sane state while
2156                 * this is true.
2157                 */
2158                if (atomic_read(&cpu_buffer->resize_disabled)) {
2159                        err = -EBUSY;
2160                        goto out_err_unlock;
2161                }
2162
2163                cpu_buffer->nr_pages_to_update = nr_pages -
2164                                                cpu_buffer->nr_pages;
2165
2166                INIT_LIST_HEAD(&cpu_buffer->new_pages);
2167                if (cpu_buffer->nr_pages_to_update > 0 &&
2168                        __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2169                                            &cpu_buffer->new_pages)) {
2170                        err = -ENOMEM;
2171                        goto out_err;
2172                }
2173
2174                cpus_read_lock();
2175
2176                /* Can't run something on an offline CPU. */
2177                if (!cpu_online(cpu_id))
2178                        rb_update_pages(cpu_buffer);
2179                else {
2180                        schedule_work_on(cpu_id,
2181                                         &cpu_buffer->update_pages_work);
2182                        wait_for_completion(&cpu_buffer->update_done);
2183                }
2184
2185                cpu_buffer->nr_pages_to_update = 0;
2186                cpus_read_unlock();
2187        }
2188
2189 out:
2190        /*
2191         * The ring buffer resize can happen with the ring buffer
2192         * enabled, so that the update disturbs the tracing as little
2193         * as possible. But if the buffer is disabled, we do not need
2194         * to worry about that, and we can take the time to verify
2195         * that the buffer is not corrupt.
2196         */
2197        if (atomic_read(&buffer->record_disabled)) {
2198                atomic_inc(&buffer->record_disabled);
2199                /*
2200                 * Even though the buffer was disabled, we must make sure
2201                 * that it is truly disabled before calling rb_check_pages.
2202                 * There could have been a race between checking
2203                 * record_disable and incrementing it.
2204                 */
2205                synchronize_rcu();
2206                for_each_buffer_cpu(buffer, cpu) {
2207                        cpu_buffer = buffer->buffers[cpu];
2208                        rb_check_pages(cpu_buffer);
2209                }
2210                atomic_dec(&buffer->record_disabled);
2211        }
2212
2213        mutex_unlock(&buffer->mutex);
2214        return 0;
2215
2216 out_err:
2217        for_each_buffer_cpu(buffer, cpu) {
2218                struct buffer_page *bpage, *tmp;
2219
2220                cpu_buffer = buffer->buffers[cpu];
2221                cpu_buffer->nr_pages_to_update = 0;
2222
2223                if (list_empty(&cpu_buffer->new_pages))
2224                        continue;
2225
2226                list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2227                                        list) {
2228                        list_del_init(&bpage->list);
2229                        free_buffer_page(bpage);
2230                }
2231        }
2232 out_err_unlock:
2233        mutex_unlock(&buffer->mutex);
2234        return err;
2235}
2236EXPORT_SYMBOL_GPL(ring_buffer_resize);
2237
2238void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2239{
2240        mutex_lock(&buffer->mutex);
2241        if (val)
2242                buffer->flags |= RB_FL_OVERWRITE;
2243        else
2244                buffer->flags &= ~RB_FL_OVERWRITE;
2245        mutex_unlock(&buffer->mutex);
2246}
2247EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2248
2249static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2250{
2251        return bpage->page->data + index;
2252}
2253
2254static __always_inline struct ring_buffer_event *
2255rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2256{
2257        return __rb_page_index(cpu_buffer->reader_page,
2258                               cpu_buffer->reader_page->read);
2259}
2260
2261static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
2262{
2263        return local_read(&bpage->page->commit);
2264}
2265
2266static struct ring_buffer_event *
2267rb_iter_head_event(struct ring_buffer_iter *iter)
2268{
2269        struct ring_buffer_event *event;
2270        struct buffer_page *iter_head_page = iter->head_page;
2271        unsigned long commit;
2272        unsigned length;
2273
2274        if (iter->head != iter->next_event)
2275                return iter->event;
2276
2277        /*
2278         * When the writer goes across pages, it issues a cmpxchg which
2279         * is a mb(), which will synchronize with the rmb here.
2280         * (see rb_tail_page_update() and __rb_reserve_next())
2281         */
2282        commit = rb_page_commit(iter_head_page);
2283        smp_rmb();
2284        event = __rb_page_index(iter_head_page, iter->head);
2285        length = rb_event_length(event);
2286
2287        /*
2288         * READ_ONCE() doesn't work on functions and we don't want the
2289         * compiler doing any crazy optimizations with length.
2290         */
2291        barrier();
2292
2293        if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
2294                /* Writer corrupted the read? */
2295                goto reset;
2296
2297        memcpy(iter->event, event, length);
2298        /*
2299         * If the page stamp is still the same after this rmb() then the
2300         * event was safely copied without the writer entering the page.
2301         */
2302        smp_rmb();
2303
2304        /* Make sure the page didn't change since we read this */
2305        if (iter->page_stamp != iter_head_page->page->time_stamp ||
2306            commit > rb_page_commit(iter_head_page))
2307                goto reset;
2308
2309        iter->next_event = iter->head + length;
2310        return iter->event;
2311 reset:
2312        /* Reset to the beginning */
2313        iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2314        iter->head = 0;
2315        iter->next_event = 0;
2316        iter->missed_events = 1;
2317        return NULL;
2318}
2319
2320/* Size is determined by what has been committed */
2321static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2322{
2323        return rb_page_commit(bpage);
2324}
2325
2326static __always_inline unsigned
2327rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2328{
2329        return rb_page_commit(cpu_buffer->commit_page);
2330}
2331
2332static __always_inline unsigned
2333rb_event_index(struct ring_buffer_event *event)
2334{
2335        unsigned long addr = (unsigned long)event;
2336
2337        return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2338}
2339
2340static void rb_inc_iter(struct ring_buffer_iter *iter)
2341{
2342        struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2343
2344        /*
2345         * The iterator could be on the reader page (it starts there).
2346         * But the head could have moved, since the reader was
2347         * found. Check for this case and assign the iterator
2348         * to the head page instead of next.
2349         */
2350        if (iter->head_page == cpu_buffer->reader_page)
2351                iter->head_page = rb_set_head_page(cpu_buffer);
2352        else
2353                rb_inc_page(&iter->head_page);
2354
2355        iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2356        iter->head = 0;
2357        iter->next_event = 0;
2358}
2359
2360/*
2361 * rb_handle_head_page - writer hit the head page
2362 *
2363 * Returns: +1 to retry page
2364 *           0 to continue
2365 *          -1 on error
2366 */
2367static int
2368rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2369                    struct buffer_page *tail_page,
2370                    struct buffer_page *next_page)
2371{
2372        struct buffer_page *new_head;
2373        int entries;
2374        int type;
2375        int ret;
2376
2377        entries = rb_page_entries(next_page);
2378
2379        /*
2380         * The hard part is here. We need to move the head
2381         * forward, and protect against both readers on
2382         * other CPUs and writers coming in via interrupts.
2383         */
2384        type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2385                                       RB_PAGE_HEAD);
2386
2387        /*
2388         * type can be one of four:
2389         *  NORMAL - an interrupt already moved it for us
2390         *  HEAD   - we are the first to get here.
2391         *  UPDATE - we are the interrupt interrupting
2392         *           a current move.
2393         *  MOVED  - a reader on another CPU moved the next
2394         *           pointer to its reader page. Give up
2395         *           and try again.
2396         */
2397
2398        switch (type) {
2399        case RB_PAGE_HEAD:
2400                /*
2401                 * We changed the head to UPDATE, thus
2402                 * it is our responsibility to update
2403                 * the counters.
2404                 */
2405                local_add(entries, &cpu_buffer->overrun);
2406                local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2407
2408                /*
2409                 * The entries will be zeroed out when we move the
2410                 * tail page.
2411                 */
2412
2413                /* still more to do */
2414                break;
2415
2416        case RB_PAGE_UPDATE:
2417                /*
2418                 * This is an interrupt that interrupt the
2419                 * previous update. Still more to do.
2420                 */
2421                break;
2422        case RB_PAGE_NORMAL:
2423                /*
2424                 * An interrupt came in before the update
2425                 * and processed this for us.
2426                 * Nothing left to do.
2427                 */
2428                return 1;
2429        case RB_PAGE_MOVED:
2430                /*
2431                 * The reader is on another CPU and just did
2432                 * a swap with our next_page.
2433                 * Try again.
2434                 */
2435                return 1;
2436        default:
2437                RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2438                return -1;
2439        }
2440
2441        /*
2442         * Now that we are here, the old head pointer is
2443         * set to UPDATE. This will keep the reader from
2444         * swapping the head page with the reader page.
2445         * The reader (on another CPU) will spin till
2446         * we are finished.
2447         *
2448         * We just need to protect against interrupts
2449         * doing the job. We will set the next pointer
2450         * to HEAD. After that, we set the old pointer
2451         * to NORMAL, but only if it was HEAD before.
2452         * otherwise we are an interrupt, and only
2453         * want the outer most commit to reset it.
2454         */
2455        new_head = next_page;
2456        rb_inc_page(&new_head);
2457
2458        ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2459                                    RB_PAGE_NORMAL);
2460
2461        /*
2462         * Valid returns are:
2463         *  HEAD   - an interrupt came in and already set it.
2464         *  NORMAL - One of two things:
2465         *            1) We really set it.
2466         *            2) A bunch of interrupts came in and moved
2467         *               the page forward again.
2468         */
2469        switch (ret) {
2470        case RB_PAGE_HEAD:
2471        case RB_PAGE_NORMAL:
2472                /* OK */
2473                break;
2474        default:
2475                RB_WARN_ON(cpu_buffer, 1);
2476                return -1;
2477        }
2478
2479        /*
2480         * It is possible that an interrupt came in,
2481         * set the head up, then more interrupts came in
2482         * and moved it again. When we get back here,
2483         * the page would have been set to NORMAL but we
2484         * just set it back to HEAD.
2485         *
2486         * How do you detect this? Well, if that happened
2487         * the tail page would have moved.
2488         */
2489        if (ret == RB_PAGE_NORMAL) {
2490                struct buffer_page *buffer_tail_page;
2491
2492                buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2493                /*
2494                 * If the tail had moved passed next, then we need
2495                 * to reset the pointer.
2496                 */
2497                if (buffer_tail_page != tail_page &&
2498                    buffer_tail_page != next_page)
2499                        rb_head_page_set_normal(cpu_buffer, new_head,
2500                                                next_page,
2501                                                RB_PAGE_HEAD);
2502        }
2503
2504        /*
2505         * If this was the outer most commit (the one that
2506         * changed the original pointer from HEAD to UPDATE),
2507         * then it is up to us to reset it to NORMAL.
2508         */
2509        if (type == RB_PAGE_HEAD) {
2510                ret = rb_head_page_set_normal(cpu_buffer, next_page,
2511                                              tail_page,
2512                                              RB_PAGE_UPDATE);
2513                if (RB_WARN_ON(cpu_buffer,
2514                               ret != RB_PAGE_UPDATE))
2515                        return -1;
2516        }
2517
2518        return 0;
2519}
2520
2521static inline void
2522rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2523              unsigned long tail, struct rb_event_info *info)
2524{
2525        struct buffer_page *tail_page = info->tail_page;
2526        struct ring_buffer_event *event;
2527        unsigned long length = info->length;
2528
2529        /*
2530         * Only the event that crossed the page boundary
2531         * must fill the old tail_page with padding.
2532         */
2533        if (tail >= BUF_PAGE_SIZE) {
2534                /*
2535                 * If the page was filled, then we still need
2536                 * to update the real_end. Reset it to zero
2537                 * and the reader will ignore it.
2538                 */
2539                if (tail == BUF_PAGE_SIZE)
2540                        tail_page->real_end = 0;
2541
2542                local_sub(length, &tail_page->write);
2543                return;
2544        }
2545
2546        event = __rb_page_index(tail_page, tail);
2547
2548        /* account for padding bytes */
2549        local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2550
2551        /*
2552         * Save the original length to the meta data.
2553         * This will be used by the reader to add lost event
2554         * counter.
2555         */
2556        tail_page->real_end = tail;
2557
2558        /*
2559         * If this event is bigger than the minimum size, then
2560         * we need to be careful that we don't subtract the
2561         * write counter enough to allow another writer to slip
2562         * in on this page.
2563         * We put in a discarded commit instead, to make sure
2564         * that this space is not used again.
2565         *
2566         * If we are less than the minimum size, we don't need to
2567         * worry about it.
2568         */
2569        if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2570                /* No room for any events */
2571
2572                /* Mark the rest of the page with padding */
2573                rb_event_set_padding(event);
2574
2575                /* Set the write back to the previous setting */
2576                local_sub(length, &tail_page->write);
2577                return;
2578        }
2579
2580        /* Put in a discarded event */
2581        event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2582        event->type_len = RINGBUF_TYPE_PADDING;
2583        /* time delta must be non zero */
2584        event->time_delta = 1;
2585
2586        /* Set write to end of buffer */
2587        length = (tail + length) - BUF_PAGE_SIZE;
2588        local_sub(length, &tail_page->write);
2589}
2590
2591static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2592
2593/*
2594 * This is the slow path, force gcc not to inline it.
2595 */
2596static noinline struct ring_buffer_event *
2597rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2598             unsigned long tail, struct rb_event_info *info)
2599{
2600        struct buffer_page *tail_page = info->tail_page;
2601        struct buffer_page *commit_page = cpu_buffer->commit_page;
2602        struct trace_buffer *buffer = cpu_buffer->buffer;
2603        struct buffer_page *next_page;
2604        int ret;
2605
2606        next_page = tail_page;
2607
2608        rb_inc_page(&next_page);
2609
2610        /*
2611         * If for some reason, we had an interrupt storm that made
2612         * it all the way around the buffer, bail, and warn
2613         * about it.
2614         */
2615        if (unlikely(next_page == commit_page)) {
2616                local_inc(&cpu_buffer->commit_overrun);
2617                goto out_reset;
2618        }
2619
2620        /*
2621         * This is where the fun begins!
2622         *
2623         * We are fighting against races between a reader that
2624         * could be on another CPU trying to swap its reader
2625         * page with the buffer head.
2626         *
2627         * We are also fighting against interrupts coming in and
2628         * moving the head or tail on us as well.
2629         *
2630         * If the next page is the head page then we have filled
2631         * the buffer, unless the commit page is still on the
2632         * reader page.
2633         */
2634        if (rb_is_head_page(next_page, &tail_page->list)) {
2635
2636                /*
2637                 * If the commit is not on the reader page, then
2638                 * move the header page.
2639                 */
2640                if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2641                        /*
2642                         * If we are not in overwrite mode,
2643                         * this is easy, just stop here.
2644                         */
2645                        if (!(buffer->flags & RB_FL_OVERWRITE)) {
2646                                local_inc(&cpu_buffer->dropped_events);
2647                                goto out_reset;
2648                        }
2649
2650                        ret = rb_handle_head_page(cpu_buffer,
2651                                                  tail_page,
2652                                                  next_page);
2653                        if (ret < 0)
2654                                goto out_reset;
2655                        if (ret)
2656                                goto out_again;
2657                } else {
2658                        /*
2659                         * We need to be careful here too. The
2660                         * commit page could still be on the reader
2661                         * page. We could have a small buffer, and
2662                         * have filled up the buffer with events
2663                         * from interrupts and such, and wrapped.
2664                         *
2665                         * Note, if the tail page is also on the
2666                         * reader_page, we let it move out.
2667                         */
2668                        if (unlikely((cpu_buffer->commit_page !=
2669                                      cpu_buffer->tail_page) &&
2670                                     (cpu_buffer->commit_page ==
2671                                      cpu_buffer->reader_page))) {
2672                                local_inc(&cpu_buffer->commit_overrun);
2673                                goto out_reset;
2674                        }
2675                }
2676        }
2677
2678        rb_tail_page_update(cpu_buffer, tail_page, next_page);
2679
2680 out_again:
2681
2682        rb_reset_tail(cpu_buffer, tail, info);
2683
2684        /* Commit what we have for now. */
2685        rb_end_commit(cpu_buffer);
2686        /* rb_end_commit() decs committing */
2687        local_inc(&cpu_buffer->committing);
2688
2689        /* fail and let the caller try again */
2690        return ERR_PTR(-EAGAIN);
2691
2692 out_reset:
2693        /* reset write */
2694        rb_reset_tail(cpu_buffer, tail, info);
2695
2696        return NULL;
2697}
2698
2699/* Slow path */
2700static struct ring_buffer_event *
2701rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2702{
2703        if (abs)
2704                event->type_len = RINGBUF_TYPE_TIME_STAMP;
2705        else
2706                event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2707
2708        /* Not the first event on the page, or not delta? */
2709        if (abs || rb_event_index(event)) {
2710                event->time_delta = delta & TS_MASK;
2711                event->array[0] = delta >> TS_SHIFT;
2712        } else {
2713                /* nope, just zero it */
2714                event->time_delta = 0;
2715                event->array[0] = 0;
2716        }
2717
2718        return skip_time_extend(event);
2719}
2720
2721#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2722static inline bool sched_clock_stable(void)
2723{
2724        return true;
2725}
2726#endif
2727
2728static void
2729rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2730                   struct rb_event_info *info)
2731{
2732        u64 write_stamp;
2733
2734        WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2735                  (unsigned long long)info->delta,
2736                  (unsigned long long)info->ts,
2737                  (unsigned long long)info->before,
2738                  (unsigned long long)info->after,
2739                  (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2740                  sched_clock_stable() ? "" :
2741                  "If you just came from a suspend/resume,\n"
2742                  "please switch to the trace global clock:\n"
2743                  "  echo global > /sys/kernel/debug/tracing/trace_clock\n"
2744                  "or add trace_clock=global to the kernel command line\n");
2745}
2746
2747static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2748                                      struct ring_buffer_event **event,
2749                                      struct rb_event_info *info,
2750                                      u64 *delta,
2751                                      unsigned int *length)
2752{
2753        bool abs = info->add_timestamp &
2754                (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2755
2756        if (unlikely(info->delta > (1ULL << 59))) {
2757                /* did the clock go backwards */
2758                if (info->before == info->after && info->before > info->ts) {
2759                        /* not interrupted */
2760                        static int once;
2761
2762                        /*
2763                         * This is possible with a recalibrating of the TSC.
2764                         * Do not produce a call stack, but just report it.
2765                         */
2766                        if (!once) {
2767                                once++;
2768                                pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2769                                        info->before, info->ts);
2770                        }
2771                } else
2772                        rb_check_timestamp(cpu_buffer, info);
2773                if (!abs)
2774                        info->delta = 0;
2775        }
2776        *event = rb_add_time_stamp(*event, info->delta, abs);
2777        *length -= RB_LEN_TIME_EXTEND;
2778        *delta = 0;
2779}
2780
2781/**
2782 * rb_update_event - update event type and data
2783 * @cpu_buffer: The per cpu buffer of the @event
2784 * @event: the event to update
2785 * @info: The info to update the @event with (contains length and delta)
2786 *
2787 * Update the type and data fields of the @event. The length
2788 * is the actual size that is written to the ring buffer,
2789 * and with this, we can determine what to place into the
2790 * data field.
2791 */
2792static void
2793rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2794                struct ring_buffer_event *event,
2795                struct rb_event_info *info)
2796{
2797        unsigned length = info->length;
2798        u64 delta = info->delta;
2799        unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2800
2801        if (!WARN_ON_ONCE(nest >= MAX_NEST))
2802                cpu_buffer->event_stamp[nest] = info->ts;
2803
2804        /*
2805         * If we need to add a timestamp, then we
2806         * add it to the start of the reserved space.
2807         */
2808        if (unlikely(info->add_timestamp))
2809                rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2810
2811        event->time_delta = delta;
2812        length -= RB_EVNT_HDR_SIZE;
2813        if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2814                event->type_len = 0;
2815                event->array[0] = length;
2816        } else
2817                event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2818}
2819
2820static unsigned rb_calculate_event_length(unsigned length)
2821{
2822        struct ring_buffer_event event; /* Used only for sizeof array */
2823
2824        /* zero length can cause confusions */
2825        if (!length)
2826                length++;
2827
2828        if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2829                length += sizeof(event.array[0]);
2830
2831        length += RB_EVNT_HDR_SIZE;
2832        length = ALIGN(length, RB_ARCH_ALIGNMENT);
2833
2834        /*
2835         * In case the time delta is larger than the 27 bits for it
2836         * in the header, we need to add a timestamp. If another
2837         * event comes in when trying to discard this one to increase
2838         * the length, then the timestamp will be added in the allocated
2839         * space of this event. If length is bigger than the size needed
2840         * for the TIME_EXTEND, then padding has to be used. The events
2841         * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2842         * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2843         * As length is a multiple of 4, we only need to worry if it
2844         * is 12 (RB_LEN_TIME_EXTEND + 4).
2845         */
2846        if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2847                length += RB_ALIGNMENT;
2848
2849        return length;
2850}
2851
2852static u64 rb_time_delta(struct ring_buffer_event *event)
2853{
2854        switch (event->type_len) {
2855        case RINGBUF_TYPE_PADDING:
2856                return 0;
2857
2858        case RINGBUF_TYPE_TIME_EXTEND:
2859                return rb_event_time_stamp(event);
2860
2861        case RINGBUF_TYPE_TIME_STAMP:
2862                return 0;
2863
2864        case RINGBUF_TYPE_DATA:
2865                return event->time_delta;
2866        default:
2867                return 0;
2868        }
2869}
2870
2871static inline int
2872rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2873                  struct ring_buffer_event *event)
2874{
2875        unsigned long new_index, old_index;
2876        struct buffer_page *bpage;
2877        unsigned long index;
2878        unsigned long addr;
2879        u64 write_stamp;
2880        u64 delta;
2881
2882        new_index = rb_event_index(event);
2883        old_index = new_index + rb_event_ts_length(event);
2884        addr = (unsigned long)event;
2885        addr &= PAGE_MASK;
2886
2887        bpage = READ_ONCE(cpu_buffer->tail_page);
2888
2889        delta = rb_time_delta(event);
2890
2891        if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
2892                return 0;
2893
2894        /* Make sure the write stamp is read before testing the location */
2895        barrier();
2896
2897        if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2898                unsigned long write_mask =
2899                        local_read(&bpage->write) & ~RB_WRITE_MASK;
2900                unsigned long event_length = rb_event_length(event);
2901
2902                /* Something came in, can't discard */
2903                if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
2904                                       write_stamp, write_stamp - delta))
2905                        return 0;
2906
2907                /*
2908                 * It's possible that the event time delta is zero
2909                 * (has the same time stamp as the previous event)
2910                 * in which case write_stamp and before_stamp could
2911                 * be the same. In such a case, force before_stamp
2912                 * to be different than write_stamp. It doesn't
2913                 * matter what it is, as long as its different.
2914                 */
2915                if (!delta)
2916                        rb_time_set(&cpu_buffer->before_stamp, 0);
2917
2918                /*
2919                 * If an event were to come in now, it would see that the
2920                 * write_stamp and the before_stamp are different, and assume
2921                 * that this event just added itself before updating
2922                 * the write stamp. The interrupting event will fix the
2923                 * write stamp for us, and use the before stamp as its delta.
2924                 */
2925
2926                /*
2927                 * This is on the tail page. It is possible that
2928                 * a write could come in and move the tail page
2929                 * and write to the next page. That is fine
2930                 * because we just shorten what is on this page.
2931                 */
2932                old_index += write_mask;
2933                new_index += write_mask;
2934                index = local_cmpxchg(&bpage->write, old_index, new_index);
2935                if (index == old_index) {
2936                        /* update counters */
2937                        local_sub(event_length, &cpu_buffer->entries_bytes);
2938                        return 1;
2939                }
2940        }
2941
2942        /* could not discard */
2943        return 0;
2944}
2945
2946static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2947{
2948        local_inc(&cpu_buffer->committing);
2949        local_inc(&cpu_buffer->commits);
2950}
2951
2952static __always_inline void
2953rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2954{
2955        unsigned long max_count;
2956
2957        /*
2958         * We only race with interrupts and NMIs on this CPU.
2959         * If we own the commit event, then we can commit
2960         * all others that interrupted us, since the interruptions
2961         * are in stack format (they finish before they come
2962         * back to us). This allows us to do a simple loop to
2963         * assign the commit to the tail.
2964         */
2965 again:
2966        max_count = cpu_buffer->nr_pages * 100;
2967
2968        while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2969                if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2970                        return;
2971                if (RB_WARN_ON(cpu_buffer,
2972                               rb_is_reader_page(cpu_buffer->tail_page)))
2973                        return;
2974                local_set(&cpu_buffer->commit_page->page->commit,
2975                          rb_page_write(cpu_buffer->commit_page));
2976                rb_inc_page(&cpu_buffer->commit_page);
2977                /* add barrier to keep gcc from optimizing too much */
2978                barrier();
2979        }
2980        while (rb_commit_index(cpu_buffer) !=
2981               rb_page_write(cpu_buffer->commit_page)) {
2982
2983                local_set(&cpu_buffer->commit_page->page->commit,
2984                          rb_page_write(cpu_buffer->commit_page));
2985                RB_WARN_ON(cpu_buffer,
2986                           local_read(&cpu_buffer->commit_page->page->commit) &
2987                           ~RB_WRITE_MASK);
2988                barrier();
2989        }
2990
2991        /* again, keep gcc from optimizing */
2992        barrier();
2993
2994        /*
2995         * If an interrupt came in just after the first while loop
2996         * and pushed the tail page forward, we will be left with
2997         * a dangling commit that will never go forward.
2998         */
2999        if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3000                goto again;
3001}
3002
3003static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3004{
3005        unsigned long commits;
3006
3007        if (RB_WARN_ON(cpu_buffer,
3008                       !local_read(&cpu_buffer->committing)))
3009                return;
3010
3011 again:
3012        commits = local_read(&cpu_buffer->commits);
3013        /* synchronize with interrupts */
3014        barrier();
3015        if (local_read(&cpu_buffer->committing) == 1)
3016                rb_set_commit_to_write(cpu_buffer);
3017
3018        local_dec(&cpu_buffer->committing);
3019
3020        /* synchronize with interrupts */
3021        barrier();
3022
3023        /*
3024         * Need to account for interrupts coming in between the
3025         * updating of the commit page and the clearing of the
3026         * committing counter.
3027         */
3028        if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3029            !local_read(&cpu_buffer->committing)) {
3030                local_inc(&cpu_buffer->committing);
3031                goto again;
3032        }
3033}
3034
3035static inline void rb_event_discard(struct ring_buffer_event *event)
3036{
3037        if (extended_time(event))
3038                event = skip_time_extend(event);
3039
3040        /* array[0] holds the actual length for the discarded event */
3041        event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3042        event->type_len = RINGBUF_TYPE_PADDING;
3043        /* time delta must be non zero */
3044        if (!event->time_delta)
3045                event->time_delta = 1;
3046}
3047
3048static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
3049                      struct ring_buffer_event *event)
3050{
3051        local_inc(&cpu_buffer->entries);
3052        rb_end_commit(cpu_buffer);
3053}
3054
3055static __always_inline void
3056rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3057{
3058        size_t nr_pages;
3059        size_t dirty;
3060        size_t full;
3061
3062        if (buffer->irq_work.waiters_pending) {
3063                buffer->irq_work.waiters_pending = false;
3064                /* irq_work_queue() supplies it's own memory barriers */
3065                irq_work_queue(&buffer->irq_work.work);
3066        }
3067
3068        if (cpu_buffer->irq_work.waiters_pending) {
3069                cpu_buffer->irq_work.waiters_pending = false;
3070                /* irq_work_queue() supplies it's own memory barriers */
3071                irq_work_queue(&cpu_buffer->irq_work.work);
3072        }
3073
3074        if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3075                return;
3076
3077        if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3078                return;
3079
3080        if (!cpu_buffer->irq_work.full_waiters_pending)
3081                return;
3082
3083        cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3084
3085        full = cpu_buffer->shortest_full;
3086        nr_pages = cpu_buffer->nr_pages;
3087        dirty = ring_buffer_nr_dirty_pages(buffer, cpu_buffer->cpu);
3088        if (full && nr_pages && (dirty * 100) <= full * nr_pages)
3089                return;
3090
3091        cpu_buffer->irq_work.wakeup_full = true;
3092        cpu_buffer->irq_work.full_waiters_pending = false;
3093        /* irq_work_queue() supplies it's own memory barriers */
3094        irq_work_queue(&cpu_buffer->irq_work.work);
3095}
3096
3097#ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3098# define do_ring_buffer_record_recursion()      \
3099        do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3100#else
3101# define do_ring_buffer_record_recursion() do { } while (0)
3102#endif
3103
3104/*
3105 * The lock and unlock are done within a preempt disable section.
3106 * The current_context per_cpu variable can only be modified
3107 * by the current task between lock and unlock. But it can
3108 * be modified more than once via an interrupt. To pass this
3109 * information from the lock to the unlock without having to
3110 * access the 'in_interrupt()' functions again (which do show
3111 * a bit of overhead in something as critical as function tracing,
3112 * we use a bitmask trick.
3113 *
3114 *  bit 1 =  NMI context
3115 *  bit 2 =  IRQ context
3116 *  bit 3 =  SoftIRQ context
3117 *  bit 4 =  normal context.
3118 *
3119 * This works because this is the order of contexts that can
3120 * preempt other contexts. A SoftIRQ never preempts an IRQ
3121 * context.
3122 *
3123 * When the context is determined, the corresponding bit is
3124 * checked and set (if it was set, then a recursion of that context
3125 * happened).
3126 *
3127 * On unlock, we need to clear this bit. To do so, just subtract
3128 * 1 from the current_context and AND it to itself.
3129 *
3130 * (binary)
3131 *  101 - 1 = 100
3132 *  101 & 100 = 100 (clearing bit zero)
3133 *
3134 *  1010 - 1 = 1001
3135 *  1010 & 1001 = 1000 (clearing bit 1)
3136 *
3137 * The least significant bit can be cleared this way, and it
3138 * just so happens that it is the same bit corresponding to
3139 * the current context.
3140 *
3141 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3142 * is set when a recursion is detected at the current context, and if
3143 * the TRANSITION bit is already set, it will fail the recursion.
3144 * This is needed because there's a lag between the changing of
3145 * interrupt context and updating the preempt count. In this case,
3146 * a false positive will be found. To handle this, one extra recursion
3147 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3148 * bit is already set, then it is considered a recursion and the function
3149 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3150 *
3151 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3152 * to be cleared. Even if it wasn't the context that set it. That is,
3153 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3154 * is called before preempt_count() is updated, since the check will
3155 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3156 * NMI then comes in, it will set the NMI bit, but when the NMI code
3157 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3158 * and leave the NMI bit set. But this is fine, because the interrupt
3159 * code that set the TRANSITION bit will then clear the NMI bit when it
3160 * calls trace_recursive_unlock(). If another NMI comes in, it will
3161 * set the TRANSITION bit and continue.
3162 *
3163 * Note: The TRANSITION bit only handles a single transition between context.
3164 */
3165
3166static __always_inline int
3167trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3168{
3169        unsigned int val = cpu_buffer->current_context;
3170        int bit = interrupt_context_level();
3171
3172        bit = RB_CTX_NORMAL - bit;
3173
3174        if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3175                /*
3176                 * It is possible that this was called by transitioning
3177                 * between interrupt context, and preempt_count() has not
3178                 * been updated yet. In this case, use the TRANSITION bit.
3179                 */
3180                bit = RB_CTX_TRANSITION;
3181                if (val & (1 << (bit + cpu_buffer->nest))) {
3182                        do_ring_buffer_record_recursion();
3183                        return 1;
3184                }
3185        }
3186
3187        val |= (1 << (bit + cpu_buffer->nest));
3188        cpu_buffer->current_context = val;
3189
3190        return 0;
3191}
3192
3193static __always_inline void
3194trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3195{
3196        cpu_buffer->current_context &=
3197                cpu_buffer->current_context - (1 << cpu_buffer->nest);
3198}
3199
3200/* The recursive locking above uses 5 bits */
3201#define NESTED_BITS 5
3202
3203/**
3204 * ring_buffer_nest_start - Allow to trace while nested
3205 * @buffer: The ring buffer to modify
3206 *
3207 * The ring buffer has a safety mechanism to prevent recursion.
3208 * But there may be a case where a trace needs to be done while
3209 * tracing something else. In this case, calling this function
3210 * will allow this function to nest within a currently active
3211 * ring_buffer_lock_reserve().
3212 *
3213 * Call this function before calling another ring_buffer_lock_reserve() and
3214 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3215 */
3216void ring_buffer_nest_start(struct trace_buffer *buffer)
3217{
3218        struct ring_buffer_per_cpu *cpu_buffer;
3219        int cpu;
3220
3221        /* Enabled by ring_buffer_nest_end() */
3222        preempt_disable_notrace();
3223        cpu = raw_smp_processor_id();
3224        cpu_buffer = buffer->buffers[cpu];
3225        /* This is the shift value for the above recursive locking */
3226        cpu_buffer->nest += NESTED_BITS;
3227}
3228
3229/**
3230 * ring_buffer_nest_end - Allow to trace while nested
3231 * @buffer: The ring buffer to modify
3232 *
3233 * Must be called after ring_buffer_nest_start() and after the
3234 * ring_buffer_unlock_commit().
3235 */
3236void ring_buffer_nest_end(struct trace_buffer *buffer)
3237{
3238        struct ring_buffer_per_cpu *cpu_buffer;
3239        int cpu;
3240
3241        /* disabled by ring_buffer_nest_start() */
3242        cpu = raw_smp_processor_id();
3243        cpu_buffer = buffer->buffers[cpu];
3244        /* This is the shift value for the above recursive locking */
3245        cpu_buffer->nest -= NESTED_BITS;
3246        preempt_enable_notrace();
3247}
3248
3249/**
3250 * ring_buffer_unlock_commit - commit a reserved
3251 * @buffer: The buffer to commit to
3252 * @event: The event pointer to commit.
3253 *
3254 * This commits the data to the ring buffer, and releases any locks held.
3255 *
3256 * Must be paired with ring_buffer_lock_reserve.
3257 */
3258int ring_buffer_unlock_commit(struct trace_buffer *buffer,
3259                              struct ring_buffer_event *event)
3260{
3261        struct ring_buffer_per_cpu *cpu_buffer;
3262        int cpu = raw_smp_processor_id();
3263
3264        cpu_buffer = buffer->buffers[cpu];
3265
3266        rb_commit(cpu_buffer, event);
3267
3268        rb_wakeups(buffer, cpu_buffer);
3269
3270        trace_recursive_unlock(cpu_buffer);
3271
3272        preempt_enable_notrace();
3273
3274        return 0;
3275}
3276EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3277
3278/* Special value to validate all deltas on a page. */
3279#define CHECK_FULL_PAGE         1L
3280
3281#ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3282static void dump_buffer_page(struct buffer_data_page *bpage,
3283                             struct rb_event_info *info,
3284                             unsigned long tail)
3285{
3286        struct ring_buffer_event *event;
3287        u64 ts, delta;
3288        int e;
3289
3290        ts = bpage->time_stamp;
3291        pr_warn("  [%lld] PAGE TIME STAMP\n", ts);
3292
3293        for (e = 0; e < tail; e += rb_event_length(event)) {
3294
3295                event = (struct ring_buffer_event *)(bpage->data + e);
3296
3297                switch (event->type_len) {
3298
3299                case RINGBUF_TYPE_TIME_EXTEND:
3300                        delta = rb_event_time_stamp(event);
3301                        ts += delta;
3302                        pr_warn("  [%lld] delta:%lld TIME EXTEND\n", ts, delta);
3303                        break;
3304
3305                case RINGBUF_TYPE_TIME_STAMP:
3306                        delta = rb_event_time_stamp(event);
3307                        ts = delta;
3308                        pr_warn("  [%lld] absolute:%lld TIME STAMP\n", ts, delta);
3309                        break;
3310
3311                case RINGBUF_TYPE_PADDING:
3312                        ts += event->time_delta;
3313                        pr_warn("  [%lld] delta:%d PADDING\n", ts, event->time_delta);
3314                        break;
3315
3316                case RINGBUF_TYPE_DATA:
3317                        ts += event->time_delta;
3318                        pr_warn("  [%lld] delta:%d\n", ts, event->time_delta);
3319                        break;
3320
3321                default:
3322                        break;
3323                }
3324        }
3325}
3326
3327static DEFINE_PER_CPU(atomic_t, checking);
3328static atomic_t ts_dump;
3329
3330/*
3331 * Check if the current event time stamp matches the deltas on
3332 * the buffer page.
3333 */
3334static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3335                         struct rb_event_info *info,
3336                         unsigned long tail)
3337{
3338        struct ring_buffer_event *event;
3339        struct buffer_data_page *bpage;
3340        u64 ts, delta;
3341        bool full = false;
3342        int e;
3343
3344        bpage = info->tail_page->page;
3345
3346        if (tail == CHECK_FULL_PAGE) {
3347                full = true;
3348                tail = local_read(&bpage->commit);
3349        } else if (info->add_timestamp &
3350                   (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3351                /* Ignore events with absolute time stamps */
3352                return;
3353        }
3354
3355        /*
3356         * Do not check the first event (skip possible extends too).
3357         * Also do not check if previous events have not been committed.
3358         */
3359        if (tail <= 8 || tail > local_read(&bpage->commit))
3360                return;
3361
3362        /*
3363         * If this interrupted another event, 
3364         */
3365        if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3366                goto out;
3367
3368        ts = bpage->time_stamp;
3369
3370        for (e = 0; e < tail; e += rb_event_length(event)) {
3371
3372                event = (struct ring_buffer_event *)(bpage->data + e);
3373
3374                switch (event->type_len) {
3375
3376                case RINGBUF_TYPE_TIME_EXTEND:
3377                        delta = rb_event_time_stamp(event);
3378                        ts += delta;
3379                        break;
3380
3381                case RINGBUF_TYPE_TIME_STAMP:
3382                        delta = rb_event_time_stamp(event);
3383                        ts = delta;
3384                        break;
3385
3386                case RINGBUF_TYPE_PADDING:
3387                        if (event->time_delta == 1)
3388                                break;
3389                        fallthrough;
3390                case RINGBUF_TYPE_DATA:
3391                        ts += event->time_delta;
3392                        break;
3393
3394                default:
3395                        RB_WARN_ON(cpu_buffer, 1);
3396                }
3397        }
3398        if ((full && ts > info->ts) ||
3399            (!full && ts + info->delta != info->ts)) {
3400                /* If another report is happening, ignore this one */
3401                if (atomic_inc_return(&ts_dump) != 1) {
3402                        atomic_dec(&ts_dump);
3403                        goto out;
3404                }
3405                atomic_inc(&cpu_buffer->record_disabled);
3406                /* There's some cases in boot up that this can happen */
3407                WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
3408                pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
3409                        cpu_buffer->cpu,
3410                        ts + info->delta, info->ts, info->delta,
3411                        info->before, info->after,
3412                        full ? " (full)" : "");
3413                dump_buffer_page(bpage, info, tail);
3414                atomic_dec(&ts_dump);
3415                /* Do not re-enable checking */
3416                return;
3417        }
3418out:
3419        atomic_dec(this_cpu_ptr(&checking));
3420}
3421#else
3422static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3423                         struct rb_event_info *info,
3424                         unsigned long tail)
3425{
3426}
3427#endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3428
3429static struct ring_buffer_event *
3430__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3431                  struct rb_event_info *info)
3432{
3433        struct ring_buffer_event *event;
3434        struct buffer_page *tail_page;
3435        unsigned long tail, write, w;
3436        bool a_ok;
3437        bool b_ok;
3438
3439        /* Don't let the compiler play games with cpu_buffer->tail_page */
3440        tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3441
3442 /*A*/  w = local_read(&tail_page->write) & RB_WRITE_MASK;
3443        barrier();
3444        b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3445        a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3446        barrier();
3447        info->ts = rb_time_stamp(cpu_buffer->buffer);
3448
3449        if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3450                info->delta = info->ts;
3451        } else {
3452                /*
3453                 * If interrupting an event time update, we may need an
3454                 * absolute timestamp.
3455                 * Don't bother if this is the start of a new page (w == 0).
3456                 */
3457                if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
3458                        info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3459                        info->length += RB_LEN_TIME_EXTEND;
3460                } else {
3461                        info->delta = info->ts - info->after;
3462                        if (unlikely(test_time_stamp(info->delta))) {
3463                                info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3464                                info->length += RB_LEN_TIME_EXTEND;
3465                        }
3466                }
3467        }
3468
3469 /*B*/  rb_time_set(&cpu_buffer->before_stamp, info->ts);
3470
3471 /*C*/  write = local_add_return(info->length, &tail_page->write);
3472
3473        /* set write to only the index of the write */
3474        write &= RB_WRITE_MASK;
3475
3476        tail = write - info->length;
3477
3478        /* See if we shot pass the end of this buffer page */
3479        if (unlikely(write > BUF_PAGE_SIZE)) {
3480                /* before and after may now different, fix it up*/
3481                b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3482                a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3483                if (a_ok && b_ok && info->before != info->after)
3484                        (void)rb_time_cmpxchg(&cpu_buffer->before_stamp,
3485                                              info->before, info->after);
3486                if (a_ok && b_ok)
3487                        check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3488                return rb_move_tail(cpu_buffer, tail, info);
3489        }
3490
3491        if (likely(tail == w)) {
3492                u64 save_before;
3493                bool s_ok;
3494
3495                /* Nothing interrupted us between A and C */
3496 /*D*/          rb_time_set(&cpu_buffer->write_stamp, info->ts);
3497                barrier();
3498 /*E*/          s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
3499                RB_WARN_ON(cpu_buffer, !s_ok);
3500                if (likely(!(info->add_timestamp &
3501                             (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3502                        /* This did not interrupt any time update */
3503                        info->delta = info->ts - info->after;
3504                else
3505                        /* Just use full timestamp for interrupting event */
3506                        info->delta = info->ts;
3507                barrier();
3508                check_buffer(cpu_buffer, info, tail);
3509                if (unlikely(info->ts != save_before)) {
3510                        /* SLOW PATH - Interrupted between C and E */
3511
3512                        a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3513                        RB_WARN_ON(cpu_buffer, !a_ok);
3514
3515                        /* Write stamp must only go forward */
3516                        if (save_before > info->after) {
3517                                /*
3518                                 * We do not care about the result, only that
3519                                 * it gets updated atomically.
3520                                 */
3521                                (void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3522                                                      info->after, save_before);
3523                        }
3524                }
3525        } else {
3526                u64 ts;
3527                /* SLOW PATH - Interrupted between A and C */
3528                a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3529                /* Was interrupted before here, write_stamp must be valid */
3530                RB_WARN_ON(cpu_buffer, !a_ok);
3531                ts = rb_time_stamp(cpu_buffer->buffer);
3532                barrier();
3533 /*E*/          if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3534                    info->after < ts &&
3535                    rb_time_cmpxchg(&cpu_buffer->write_stamp,
3536                                    info->after, ts)) {
3537                        /* Nothing came after this event between C and E */
3538                        info->delta = ts - info->after;
3539                } else {
3540                        /*
3541                         * Interrupted between C and E:
3542                         * Lost the previous events time stamp. Just set the
3543                         * delta to zero, and this will be the same time as
3544                         * the event this event interrupted. And the events that
3545                         * came after this will still be correct (as they would
3546                         * have built their delta on the previous event.
3547                         */
3548                        info->delta = 0;
3549                }
3550                info->ts = ts;
3551                info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3552        }
3553
3554        /*
3555         * If this is the first commit on the page, then it has the same
3556         * timestamp as the page itself.
3557         */
3558        if (unlikely(!tail && !(info->add_timestamp &
3559                                (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3560                info->delta = 0;
3561
3562        /* We reserved something on the buffer */
3563
3564        event = __rb_page_index(tail_page, tail);
3565        rb_update_event(cpu_buffer, event, info);
3566
3567        local_inc(&tail_page->entries);
3568
3569        /*
3570         * If this is the first commit on the page, then update
3571         * its timestamp.
3572         */
3573        if (unlikely(!tail))
3574                tail_page->page->time_stamp = info->ts;
3575
3576        /* account for these added bytes */
3577        local_add(info->length, &cpu_buffer->entries_bytes);
3578
3579        return event;
3580}
3581
3582static __always_inline struct ring_buffer_event *
3583rb_reserve_next_event(struct trace_buffer *buffer,
3584                      struct ring_buffer_per_cpu *cpu_buffer,
3585                      unsigned long length)
3586{
3587        struct ring_buffer_event *event;
3588        struct rb_event_info info;
3589        int nr_loops = 0;
3590        int add_ts_default;
3591
3592        rb_start_commit(cpu_buffer);
3593        /* The commit page can not change after this */
3594
3595#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3596        /*
3597         * Due to the ability to swap a cpu buffer from a buffer
3598         * it is possible it was swapped before we committed.
3599         * (committing stops a swap). We check for it here and
3600         * if it happened, we have to fail the write.
3601         */
3602        barrier();
3603        if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3604                local_dec(&cpu_buffer->committing);
3605                local_dec(&cpu_buffer->commits);
3606                return NULL;
3607        }
3608#endif
3609
3610        info.length = rb_calculate_event_length(length);
3611
3612        if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3613                add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3614                info.length += RB_LEN_TIME_EXTEND;
3615        } else {
3616                add_ts_default = RB_ADD_STAMP_NONE;
3617        }
3618
3619 again:
3620        info.add_timestamp = add_ts_default;
3621        info.delta = 0;
3622
3623        /*
3624         * We allow for interrupts to reenter here and do a trace.
3625         * If one does, it will cause this original code to loop
3626         * back here. Even with heavy interrupts happening, this
3627         * should only happen a few times in a row. If this happens
3628         * 1000 times in a row, there must be either an interrupt
3629         * storm or we have something buggy.
3630         * Bail!
3631         */
3632        if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3633                goto out_fail;
3634
3635        event = __rb_reserve_next(cpu_buffer, &info);
3636
3637        if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3638                if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3639                        info.length -= RB_LEN_TIME_EXTEND;
3640                goto again;
3641        }
3642
3643        if (likely(event))
3644                return event;
3645 out_fail:
3646        rb_end_commit(cpu_buffer);
3647        return NULL;
3648}
3649
3650/**
3651 * ring_buffer_lock_reserve - reserve a part of the buffer
3652 * @buffer: the ring buffer to reserve from
3653 * @length: the length of the data to reserve (excluding event header)
3654 *
3655 * Returns a reserved event on the ring buffer to copy directly to.
3656 * The user of this interface will need to get the body to write into
3657 * and can use the ring_buffer_event_data() interface.
3658 *
3659 * The length is the length of the data needed, not the event length
3660 * which also includes the event header.
3661 *
3662 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3663 * If NULL is returned, then nothing has been allocated or locked.
3664 */
3665struct ring_buffer_event *
3666ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3667{
3668        struct ring_buffer_per_cpu *cpu_buffer;
3669        struct ring_buffer_event *event;
3670        int cpu;
3671
3672        /* If we are tracing schedule, we don't want to recurse */
3673        preempt_disable_notrace();
3674
3675        if (unlikely(atomic_read(&buffer->record_disabled)))
3676                goto out;
3677
3678        cpu = raw_smp_processor_id();
3679
3680        if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3681                goto out;
3682
3683        cpu_buffer = buffer->buffers[cpu];
3684
3685        if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3686                goto out;
3687
3688        if (unlikely(length > BUF_MAX_DATA_SIZE))
3689                goto out;
3690
3691        if (unlikely(trace_recursive_lock(cpu_buffer)))
3692                goto out;
3693
3694        event = rb_reserve_next_event(buffer, cpu_buffer, length);
3695        if (!event)
3696                goto out_unlock;
3697
3698        return event;
3699
3700 out_unlock:
3701        trace_recursive_unlock(cpu_buffer);
3702 out:
3703        preempt_enable_notrace();
3704        return NULL;
3705}
3706EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3707
3708/*
3709 * Decrement the entries to the page that an event is on.
3710 * The event does not even need to exist, only the pointer
3711 * to the page it is on. This may only be called before the commit
3712 * takes place.
3713 */
3714static inline void
3715rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3716                   struct ring_buffer_event *event)
3717{
3718        unsigned long addr = (unsigned long)event;
3719        struct buffer_page *bpage = cpu_buffer->commit_page;
3720        struct buffer_page *start;
3721
3722        addr &= PAGE_MASK;
3723
3724        /* Do the likely case first */
3725        if (likely(bpage->page == (void *)addr)) {
3726                local_dec(&bpage->entries);
3727                return;
3728        }
3729
3730        /*
3731         * Because the commit page may be on the reader page we
3732         * start with the next page and check the end loop there.
3733         */
3734        rb_inc_page(&bpage);
3735        start = bpage;
3736        do {
3737                if (bpage->page == (void *)addr) {
3738                        local_dec(&bpage->entries);
3739                        return;
3740                }
3741                rb_inc_page(&bpage);
3742        } while (bpage != start);
3743
3744        /* commit not part of this buffer?? */
3745        RB_WARN_ON(cpu_buffer, 1);
3746}
3747
3748/**
3749 * ring_buffer_discard_commit - discard an event that has not been committed
3750 * @buffer: the ring buffer
3751 * @event: non committed event to discard
3752 *
3753 * Sometimes an event that is in the ring buffer needs to be ignored.
3754 * This function lets the user discard an event in the ring buffer
3755 * and then that event will not be read later.
3756 *
3757 * This function only works if it is called before the item has been
3758 * committed. It will try to free the event from the ring buffer
3759 * if another event has not been added behind it.
3760 *
3761 * If another event has been added behind it, it will set the event
3762 * up as discarded, and perform the commit.
3763 *
3764 * If this function is called, do not call ring_buffer_unlock_commit on
3765 * the event.
3766 */
3767void ring_buffer_discard_commit(struct trace_buffer *buffer,
3768                                struct ring_buffer_event *event)
3769{
3770        struct ring_buffer_per_cpu *cpu_buffer;
3771        int cpu;
3772
3773        /* The event is discarded regardless */
3774        rb_event_discard(event);
3775
3776        cpu = smp_processor_id();
3777        cpu_buffer = buffer->buffers[cpu];
3778
3779        /*
3780         * This must only be called if the event has not been
3781         * committed yet. Thus we can assume that preemption
3782         * is still disabled.
3783         */
3784        RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3785
3786        rb_decrement_entry(cpu_buffer, event);
3787        if (rb_try_to_discard(cpu_buffer, event))
3788                goto out;
3789
3790 out:
3791        rb_end_commit(cpu_buffer);
3792
3793        trace_recursive_unlock(cpu_buffer);
3794
3795        preempt_enable_notrace();
3796
3797}
3798EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3799
3800/**
3801 * ring_buffer_write - write data to the buffer without reserving
3802 * @buffer: The ring buffer to write to.
3803 * @length: The length of the data being written (excluding the event header)
3804 * @data: The data to write to the buffer.
3805 *
3806 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3807 * one function. If you already have the data to write to the buffer, it
3808 * may be easier to simply call this function.
3809 *
3810 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3811 * and not the length of the event which would hold the header.
3812 */
3813int ring_buffer_write(struct trace_buffer *buffer,
3814                      unsigned long length,
3815                      void *data)
3816{
3817        struct ring_buffer_per_cpu *cpu_buffer;
3818        struct ring_buffer_event *event;
3819        void *body;
3820        int ret = -EBUSY;
3821        int cpu;
3822
3823        preempt_disable_notrace();
3824
3825        if (atomic_read(&buffer->record_disabled))
3826                goto out;
3827
3828        cpu = raw_smp_processor_id();
3829
3830        if (!cpumask_test_cpu(cpu, buffer->cpumask))
3831                goto out;
3832
3833        cpu_buffer = buffer->buffers[cpu];
3834
3835        if (atomic_read(&cpu_buffer->record_disabled))
3836                goto out;
3837
3838        if (length > BUF_MAX_DATA_SIZE)
3839                goto out;
3840
3841        if (unlikely(trace_recursive_lock(cpu_buffer)))
3842                goto out;
3843
3844        event = rb_reserve_next_event(buffer, cpu_buffer, length);
3845        if (!event)
3846                goto out_unlock;
3847
3848        body = rb_event_data(event);
3849
3850        memcpy(body, data, length);
3851
3852        rb_commit(cpu_buffer, event);
3853
3854        rb_wakeups(buffer, cpu_buffer);
3855
3856        ret = 0;
3857
3858 out_unlock:
3859        trace_recursive_unlock(cpu_buffer);
3860
3861 out:
3862        preempt_enable_notrace();
3863
3864        return ret;
3865}
3866EXPORT_SYMBOL_GPL(ring_buffer_write);
3867
3868static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3869{
3870        struct buffer_page *reader = cpu_buffer->reader_page;
3871        struct buffer_page *head = rb_set_head_page(cpu_buffer);
3872        struct buffer_page *commit = cpu_buffer->commit_page;
3873
3874        /* In case of error, head will be NULL */
3875        if (unlikely(!head))
3876                return true;
3877
3878        /* Reader should exhaust content in reader page */
3879        if (reader->read != rb_page_commit(reader))
3880                return false;
3881
3882        /*
3883         * If writers are committing on the reader page, knowing all
3884         * committed content has been read, the ring buffer is empty.
3885         */
3886        if (commit == reader)
3887                return true;
3888
3889        /*
3890         * If writers are committing on a page other than reader page
3891         * and head page, there should always be content to read.
3892         */
3893        if (commit != head)
3894                return false;
3895
3896        /*
3897         * Writers are committing on the head page, we just need
3898         * to care about there're committed data, and the reader will
3899         * swap reader page with head page when it is to read data.
3900         */
3901        return rb_page_commit(commit) == 0;
3902}
3903
3904/**
3905 * ring_buffer_record_disable - stop all writes into the buffer
3906 * @buffer: The ring buffer to stop writes to.
3907 *
3908 * This prevents all writes to the buffer. Any attempt to write
3909 * to the buffer after this will fail and return NULL.
3910 *
3911 * The caller should call synchronize_rcu() after this.
3912 */
3913void ring_buffer_record_disable(struct trace_buffer *buffer)
3914{
3915        atomic_inc(&buffer->record_disabled);
3916}
3917EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3918
3919/**
3920 * ring_buffer_record_enable - enable writes to the buffer
3921 * @buffer: The ring buffer to enable writes
3922 *
3923 * Note, multiple disables will need the same number of enables
3924 * to truly enable the writing (much like preempt_disable).
3925 */
3926void ring_buffer_record_enable(struct trace_buffer *buffer)
3927{
3928        atomic_dec(&buffer->record_disabled);
3929}
3930EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3931
3932/**
3933 * ring_buffer_record_off - stop all writes into the buffer
3934 * @buffer: The ring buffer to stop writes to.
3935 *
3936 * This prevents all writes to the buffer. Any attempt to write
3937 * to the buffer after this will fail and return NULL.
3938 *
3939 * This is different than ring_buffer_record_disable() as
3940 * it works like an on/off switch, where as the disable() version
3941 * must be paired with a enable().
3942 */
3943void ring_buffer_record_off(struct trace_buffer *buffer)
3944{
3945        unsigned int rd;
3946        unsigned int new_rd;
3947
3948        do {
3949                rd = atomic_read(&buffer->record_disabled);
3950                new_rd = rd | RB_BUFFER_OFF;
3951        } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3952}
3953EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3954
3955/**
3956 * ring_buffer_record_on - restart writes into the buffer
3957 * @buffer: The ring buffer to start writes to.
3958 *
3959 * This enables all writes to the buffer that was disabled by
3960 * ring_buffer_record_off().
3961 *
3962 * This is different than ring_buffer_record_enable() as
3963 * it works like an on/off switch, where as the enable() version
3964 * must be paired with a disable().
3965 */
3966void ring_buffer_record_on(struct trace_buffer *buffer)
3967{
3968        unsigned int rd;
3969        unsigned int new_rd;
3970
3971        do {
3972                rd = atomic_read(&buffer->record_disabled);
3973                new_rd = rd & ~RB_BUFFER_OFF;
3974        } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3975}
3976EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3977
3978/**
3979 * ring_buffer_record_is_on - return true if the ring buffer can write
3980 * @buffer: The ring buffer to see if write is enabled
3981 *
3982 * Returns true if the ring buffer is in a state that it accepts writes.
3983 */
3984bool ring_buffer_record_is_on(struct trace_buffer *buffer)
3985{
3986        return !atomic_read(&buffer->record_disabled);
3987}
3988
3989/**
3990 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3991 * @buffer: The ring buffer to see if write is set enabled
3992 *
3993 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3994 * Note that this does NOT mean it is in a writable state.
3995 *
3996 * It may return true when the ring buffer has been disabled by
3997 * ring_buffer_record_disable(), as that is a temporary disabling of
3998 * the ring buffer.
3999 */
4000bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4001{
4002        return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4003}
4004
4005/**
4006 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4007 * @buffer: The ring buffer to stop writes to.
4008 * @cpu: The CPU buffer to stop
4009 *
4010 * This prevents all writes to the buffer. Any attempt to write
4011 * to the buffer after this will fail and return NULL.
4012 *
4013 * The caller should call synchronize_rcu() after this.
4014 */
4015void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4016{
4017        struct ring_buffer_per_cpu *cpu_buffer;
4018
4019        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4020                return;
4021
4022        cpu_buffer = buffer->buffers[cpu];
4023        atomic_inc(&cpu_buffer->record_disabled);
4024}
4025EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4026
4027/**
4028 * ring_buffer_record_enable_cpu - enable writes to the buffer
4029 * @buffer: The ring buffer to enable writes
4030 * @cpu: The CPU to enable.
4031 *
4032 * Note, multiple disables will need the same number of enables
4033 * to truly enable the writing (much like preempt_disable).
4034 */
4035void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4036{
4037        struct ring_buffer_per_cpu *cpu_buffer;
4038
4039        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4040                return;
4041
4042        cpu_buffer = buffer->buffers[cpu];
4043        atomic_dec(&cpu_buffer->record_disabled);
4044}
4045EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4046
4047/*
4048 * The total entries in the ring buffer is the running counter
4049 * of entries entered into the ring buffer, minus the sum of
4050 * the entries read from the ring buffer and the number of
4051 * entries that were overwritten.
4052 */
4053static inline unsigned long
4054rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4055{
4056        return local_read(&cpu_buffer->entries) -
4057                (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4058}
4059
4060/**
4061 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4062 * @buffer: The ring buffer
4063 * @cpu: The per CPU buffer to read from.
4064 */
4065u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4066{
4067        unsigned long flags;
4068        struct ring_buffer_per_cpu *cpu_buffer;
4069        struct buffer_page *bpage;
4070        u64 ret = 0;
4071
4072        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4073                return 0;
4074
4075        cpu_buffer = buffer->buffers[cpu];
4076        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4077        /*
4078         * if the tail is on reader_page, oldest time stamp is on the reader
4079         * page
4080         */
4081        if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4082                bpage = cpu_buffer->reader_page;
4083        else
4084                bpage = rb_set_head_page(cpu_buffer);
4085        if (bpage)
4086                ret = bpage->page->time_stamp;
4087        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4088
4089        return ret;
4090}
4091EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4092
4093/**
4094 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
4095 * @buffer: The ring buffer
4096 * @cpu: The per CPU buffer to read from.
4097 */
4098unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4099{
4100        struct ring_buffer_per_cpu *cpu_buffer;
4101        unsigned long ret;
4102
4103        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4104                return 0;
4105
4106        cpu_buffer = buffer->buffers[cpu];
4107        ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4108
4109        return ret;
4110}
4111EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4112
4113/**
4114 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4115 * @buffer: The ring buffer
4116 * @cpu: The per CPU buffer to get the entries from.
4117 */
4118unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4119{
4120        struct ring_buffer_per_cpu *cpu_buffer;
4121
4122        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4123                return 0;
4124
4125        cpu_buffer = buffer->buffers[cpu];
4126
4127        return rb_num_of_entries(cpu_buffer);
4128}
4129EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4130
4131/**
4132 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4133 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4134 * @buffer: The ring buffer
4135 * @cpu: The per CPU buffer to get the number of overruns from
4136 */
4137unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4138{
4139        struct ring_buffer_per_cpu *cpu_buffer;
4140        unsigned long ret;
4141
4142        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4143                return 0;
4144
4145        cpu_buffer = buffer->buffers[cpu];
4146        ret = local_read(&cpu_buffer->overrun);
4147
4148        return ret;
4149}
4150EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4151
4152/**
4153 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4154 * commits failing due to the buffer wrapping around while there are uncommitted
4155 * events, such as during an interrupt storm.
4156 * @buffer: The ring buffer
4157 * @cpu: The per CPU buffer to get the number of overruns from
4158 */
4159unsigned long
4160ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4161{
4162        struct ring_buffer_per_cpu *cpu_buffer;
4163        unsigned long ret;
4164
4165        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4166                return 0;
4167
4168        cpu_buffer = buffer->buffers[cpu];
4169        ret = local_read(&cpu_buffer->commit_overrun);
4170
4171        return ret;
4172}
4173EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4174
4175/**
4176 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4177 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4178 * @buffer: The ring buffer
4179 * @cpu: The per CPU buffer to get the number of overruns from
4180 */
4181unsigned long
4182ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4183{
4184        struct ring_buffer_per_cpu *cpu_buffer;
4185        unsigned long ret;
4186
4187        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4188                return 0;
4189
4190        cpu_buffer = buffer->buffers[cpu];
4191        ret = local_read(&cpu_buffer->dropped_events);
4192
4193        return ret;
4194}
4195EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4196
4197/**
4198 * ring_buffer_read_events_cpu - get the number of events successfully read
4199 * @buffer: The ring buffer
4200 * @cpu: The per CPU buffer to get the number of events read
4201 */
4202unsigned long
4203ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4204{
4205        struct ring_buffer_per_cpu *cpu_buffer;
4206
4207        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4208                return 0;
4209
4210        cpu_buffer = buffer->buffers[cpu];
4211        return cpu_buffer->read;
4212}
4213EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4214
4215/**
4216 * ring_buffer_entries - get the number of entries in a buffer
4217 * @buffer: The ring buffer
4218 *
4219 * Returns the total number of entries in the ring buffer
4220 * (all CPU entries)
4221 */
4222unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4223{
4224        struct ring_buffer_per_cpu *cpu_buffer;
4225        unsigned long entries = 0;
4226        int cpu;
4227
4228        /* if you care about this being correct, lock the buffer */
4229        for_each_buffer_cpu(buffer, cpu) {
4230                cpu_buffer = buffer->buffers[cpu];
4231                entries += rb_num_of_entries(cpu_buffer);
4232        }
4233
4234        return entries;
4235}
4236EXPORT_SYMBOL_GPL(ring_buffer_entries);
4237
4238/**
4239 * ring_buffer_overruns - get the number of overruns in buffer
4240 * @buffer: The ring buffer
4241 *
4242 * Returns the total number of overruns in the ring buffer
4243 * (all CPU entries)
4244 */
4245unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4246{
4247        struct ring_buffer_per_cpu *cpu_buffer;
4248        unsigned long overruns = 0;
4249        int cpu;
4250
4251        /* if you care about this being correct, lock the buffer */
4252        for_each_buffer_cpu(buffer, cpu) {
4253                cpu_buffer = buffer->buffers[cpu];
4254                overruns += local_read(&cpu_buffer->overrun);
4255        }
4256
4257        return overruns;
4258}
4259EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4260
4261static void rb_iter_reset(struct ring_buffer_iter *iter)
4262{
4263        struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4264
4265        /* Iterator usage is expected to have record disabled */
4266        iter->head_page = cpu_buffer->reader_page;
4267        iter->head = cpu_buffer->reader_page->read;
4268        iter->next_event = iter->head;
4269
4270        iter->cache_reader_page = iter->head_page;
4271        iter->cache_read = cpu_buffer->read;
4272
4273        if (iter->head) {
4274                iter->read_stamp = cpu_buffer->read_stamp;
4275                iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4276        } else {
4277                iter->read_stamp = iter->head_page->page->time_stamp;
4278                iter->page_stamp = iter->read_stamp;
4279        }
4280}
4281
4282/**
4283 * ring_buffer_iter_reset - reset an iterator
4284 * @iter: The iterator to reset
4285 *
4286 * Resets the iterator, so that it will start from the beginning
4287 * again.
4288 */
4289void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4290{
4291        struct ring_buffer_per_cpu *cpu_buffer;
4292        unsigned long flags;
4293
4294        if (!iter)
4295                return;
4296
4297        cpu_buffer = iter->cpu_buffer;
4298
4299        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4300        rb_iter_reset(iter);
4301        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4302}
4303EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4304
4305/**
4306 * ring_buffer_iter_empty - check if an iterator has no more to read
4307 * @iter: The iterator to check
4308 */
4309int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4310{
4311        struct ring_buffer_per_cpu *cpu_buffer;
4312        struct buffer_page *reader;
4313        struct buffer_page *head_page;
4314        struct buffer_page *commit_page;
4315        struct buffer_page *curr_commit_page;
4316        unsigned commit;
4317        u64 curr_commit_ts;
4318        u64 commit_ts;
4319
4320        cpu_buffer = iter->cpu_buffer;
4321        reader = cpu_buffer->reader_page;
4322        head_page = cpu_buffer->head_page;
4323        commit_page = cpu_buffer->commit_page;
4324        commit_ts = commit_page->page->time_stamp;
4325
4326        /*
4327         * When the writer goes across pages, it issues a cmpxchg which
4328         * is a mb(), which will synchronize with the rmb here.
4329         * (see rb_tail_page_update())
4330         */
4331        smp_rmb();
4332        commit = rb_page_commit(commit_page);
4333        /* We want to make sure that the commit page doesn't change */
4334        smp_rmb();
4335
4336        /* Make sure commit page didn't change */
4337        curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4338        curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4339
4340        /* If the commit page changed, then there's more data */
4341        if (curr_commit_page != commit_page ||
4342            curr_commit_ts != commit_ts)
4343                return 0;
4344
4345        /* Still racy, as it may return a false positive, but that's OK */
4346        return ((iter->head_page == commit_page && iter->head >= commit) ||
4347                (iter->head_page == reader && commit_page == head_page &&
4348                 head_page->read == commit &&
4349                 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4350}
4351EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4352
4353static void
4354rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4355                     struct ring_buffer_event *event)
4356{
4357        u64 delta;
4358
4359        switch (event->type_len) {
4360        case RINGBUF_TYPE_PADDING:
4361                return;
4362
4363        case RINGBUF_TYPE_TIME_EXTEND:
4364                delta = rb_event_time_stamp(event);
4365                cpu_buffer->read_stamp += delta;
4366                return;
4367
4368        case RINGBUF_TYPE_TIME_STAMP:
4369                delta = rb_event_time_stamp(event);
4370                cpu_buffer->read_stamp = delta;
4371                return;
4372
4373        case RINGBUF_TYPE_DATA:
4374                cpu_buffer->read_stamp += event->time_delta;
4375                return;
4376
4377        default:
4378                RB_WARN_ON(cpu_buffer, 1);
4379        }
4380        return;
4381}
4382
4383static void
4384rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4385                          struct ring_buffer_event *event)
4386{
4387        u64 delta;
4388
4389        switch (event->type_len) {
4390        case RINGBUF_TYPE_PADDING:
4391                return;
4392
4393        case RINGBUF_TYPE_TIME_EXTEND:
4394                delta = rb_event_time_stamp(event);
4395                iter->read_stamp += delta;
4396                return;
4397
4398        case RINGBUF_TYPE_TIME_STAMP:
4399                delta = rb_event_time_stamp(event);
4400                iter->read_stamp = delta;
4401                return;
4402
4403        case RINGBUF_TYPE_DATA:
4404                iter->read_stamp += event->time_delta;
4405                return;
4406
4407        default:
4408                RB_WARN_ON(iter->cpu_buffer, 1);
4409        }
4410        return;
4411}
4412
4413static struct buffer_page *
4414rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4415{
4416        struct buffer_page *reader = NULL;
4417        unsigned long overwrite;
4418        unsigned long flags;
4419        int nr_loops = 0;
4420        int ret;
4421
4422        local_irq_save(flags);
4423        arch_spin_lock(&cpu_buffer->lock);
4424
4425 again:
4426        /*
4427         * This should normally only loop twice. But because the
4428         * start of the reader inserts an empty page, it causes
4429         * a case where we will loop three times. There should be no
4430         * reason to loop four times (that I know of).
4431         */
4432        if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4433                reader = NULL;
4434                goto out;
4435        }
4436
4437        reader = cpu_buffer->reader_page;
4438
4439        /* If there's more to read, return this page */
4440        if (cpu_buffer->reader_page->read < rb_page_size(reader))
4441                goto out;
4442
4443        /* Never should we have an index greater than the size */
4444        if (RB_WARN_ON(cpu_buffer,
4445                       cpu_buffer->reader_page->read > rb_page_size(reader)))
4446                goto out;
4447
4448        /* check if we caught up to the tail */
4449        reader = NULL;
4450        if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4451                goto out;
4452
4453        /* Don't bother swapping if the ring buffer is empty */
4454        if (rb_num_of_entries(cpu_buffer) == 0)
4455                goto out;
4456
4457        /*
4458         * Reset the reader page to size zero.
4459         */
4460        local_set(&cpu_buffer->reader_page->write, 0);
4461        local_set(&cpu_buffer->reader_page->entries, 0);
4462        local_set(&cpu_buffer->reader_page->page->commit, 0);
4463        cpu_buffer->reader_page->real_end = 0;
4464
4465 spin:
4466        /*
4467         * Splice the empty reader page into the list around the head.
4468         */
4469        reader = rb_set_head_page(cpu_buffer);
4470        if (!reader)
4471                goto out;
4472        cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4473        cpu_buffer->reader_page->list.prev = reader->list.prev;
4474
4475        /*
4476         * cpu_buffer->pages just needs to point to the buffer, it
4477         *  has no specific buffer page to point to. Lets move it out
4478         *  of our way so we don't accidentally swap it.
4479         */
4480        cpu_buffer->pages = reader->list.prev;
4481
4482        /* The reader page will be pointing to the new head */
4483        rb_set_list_to_head(&cpu_buffer->reader_page->list);
4484
4485        /*
4486         * We want to make sure we read the overruns after we set up our
4487         * pointers to the next object. The writer side does a
4488         * cmpxchg to cross pages which acts as the mb on the writer
4489         * side. Note, the reader will constantly fail the swap
4490         * while the writer is updating the pointers, so this
4491         * guarantees that the overwrite recorded here is the one we
4492         * want to compare with the last_overrun.
4493         */
4494        smp_mb();
4495        overwrite = local_read(&(cpu_buffer->overrun));
4496
4497        /*
4498         * Here's the tricky part.
4499         *
4500         * We need to move the pointer past the header page.
4501         * But we can only do that if a writer is not currently
4502         * moving it. The page before the header page has the
4503         * flag bit '1' set if it is pointing to the page we want.
4504         * but if the writer is in the process of moving it
4505         * than it will be '2' or already moved '0'.
4506         */
4507
4508        ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4509
4510        /*
4511         * If we did not convert it, then we must try again.
4512         */
4513        if (!ret)
4514                goto spin;
4515
4516        /*
4517         * Yay! We succeeded in replacing the page.
4518         *
4519         * Now make the new head point back to the reader page.
4520         */
4521        rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4522        rb_inc_page(&cpu_buffer->head_page);
4523
4524        local_inc(&cpu_buffer->pages_read);
4525
4526        /* Finally update the reader page to the new head */
4527        cpu_buffer->reader_page = reader;
4528        cpu_buffer->reader_page->read = 0;
4529
4530        if (overwrite != cpu_buffer->last_overrun) {
4531                cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4532                cpu_buffer->last_overrun = overwrite;
4533        }
4534
4535        goto again;
4536
4537 out:
4538        /* Update the read_stamp on the first event */
4539        if (reader && reader->read == 0)
4540                cpu_buffer->read_stamp = reader->page->time_stamp;
4541
4542        arch_spin_unlock(&cpu_buffer->lock);
4543        local_irq_restore(flags);
4544
4545        return reader;
4546}
4547
4548static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4549{
4550        struct ring_buffer_event *event;
4551        struct buffer_page *reader;
4552        unsigned length;
4553
4554        reader = rb_get_reader_page(cpu_buffer);
4555
4556        /* This function should not be called when buffer is empty */
4557        if (RB_WARN_ON(cpu_buffer, !reader))
4558                return;
4559
4560        event = rb_reader_event(cpu_buffer);
4561
4562        if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4563                cpu_buffer->read++;
4564
4565        rb_update_read_stamp(cpu_buffer, event);
4566
4567        length = rb_event_length(event);
4568        cpu_buffer->reader_page->read += length;
4569}
4570
4571static void rb_advance_iter(struct ring_buffer_iter *iter)
4572{
4573        struct ring_buffer_per_cpu *cpu_buffer;
4574
4575        cpu_buffer = iter->cpu_buffer;
4576
4577        /* If head == next_event then we need to jump to the next event */
4578        if (iter->head == iter->next_event) {
4579                /* If the event gets overwritten again, there's nothing to do */
4580                if (rb_iter_head_event(iter) == NULL)
4581                        return;
4582        }
4583
4584        iter->head = iter->next_event;
4585
4586        /*
4587         * Check if we are at the end of the buffer.
4588         */
4589        if (iter->next_event >= rb_page_size(iter->head_page)) {
4590                /* discarded commits can make the page empty */
4591                if (iter->head_page == cpu_buffer->commit_page)
4592                        return;
4593                rb_inc_iter(iter);
4594                return;
4595        }
4596
4597        rb_update_iter_read_stamp(iter, iter->event);
4598}
4599
4600static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4601{
4602        return cpu_buffer->lost_events;
4603}
4604
4605static struct ring_buffer_event *
4606rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4607               unsigned long *lost_events)
4608{
4609        struct ring_buffer_event *event;
4610        struct buffer_page *reader;
4611        int nr_loops = 0;
4612
4613        if (ts)
4614                *ts = 0;
4615 again:
4616        /*
4617         * We repeat when a time extend is encountered.
4618         * Since the time extend is always attached to a data event,
4619         * we should never loop more than once.
4620         * (We never hit the following condition more than twice).
4621         */
4622        if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4623                return NULL;
4624
4625        reader = rb_get_reader_page(cpu_buffer);
4626        if (!reader)
4627                return NULL;
4628
4629        event = rb_reader_event(cpu_buffer);
4630
4631        switch (event->type_len) {
4632        case RINGBUF_TYPE_PADDING:
4633                if (rb_null_event(event))
4634                        RB_WARN_ON(cpu_buffer, 1);
4635                /*
4636                 * Because the writer could be discarding every
4637                 * event it creates (which would probably be bad)
4638                 * if we were to go back to "again" then we may never
4639                 * catch up, and will trigger the warn on, or lock
4640                 * the box. Return the padding, and we will release
4641                 * the current locks, and try again.
4642                 */
4643                return event;
4644
4645        case RINGBUF_TYPE_TIME_EXTEND:
4646                /* Internal data, OK to advance */
4647                rb_advance_reader(cpu_buffer);
4648                goto again;
4649
4650        case RINGBUF_TYPE_TIME_STAMP:
4651                if (ts) {
4652                        *ts = rb_event_time_stamp(event);
4653                        ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4654                                                         cpu_buffer->cpu, ts);
4655                }
4656                /* Internal data, OK to advance */
4657                rb_advance_reader(cpu_buffer);
4658                goto again;
4659
4660        case RINGBUF_TYPE_DATA:
4661                if (ts && !(*ts)) {
4662                        *ts = cpu_buffer->read_stamp + event->time_delta;
4663                        ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4664                                                         cpu_buffer->cpu, ts);
4665                }
4666                if (lost_events)
4667                        *lost_events = rb_lost_events(cpu_buffer);
4668                return event;
4669
4670        default:
4671                RB_WARN_ON(cpu_buffer, 1);
4672        }
4673
4674        return NULL;
4675}
4676EXPORT_SYMBOL_GPL(ring_buffer_peek);
4677
4678static struct ring_buffer_event *
4679rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4680{
4681        struct trace_buffer *buffer;
4682        struct ring_buffer_per_cpu *cpu_buffer;
4683        struct ring_buffer_event *event;
4684        int nr_loops = 0;
4685
4686        if (ts)
4687                *ts = 0;
4688
4689        cpu_buffer = iter->cpu_buffer;
4690        buffer = cpu_buffer->buffer;
4691
4692        /*
4693         * Check if someone performed a consuming read to
4694         * the buffer. A consuming read invalidates the iterator
4695         * and we need to reset the iterator in this case.
4696         */
4697        if (unlikely(iter->cache_read != cpu_buffer->read ||
4698                     iter->cache_reader_page != cpu_buffer->reader_page))
4699                rb_iter_reset(iter);
4700
4701 again:
4702        if (ring_buffer_iter_empty(iter))
4703                return NULL;
4704
4705        /*
4706         * As the writer can mess with what the iterator is trying
4707         * to read, just give up if we fail to get an event after
4708         * three tries. The iterator is not as reliable when reading
4709         * the ring buffer with an active write as the consumer is.
4710         * Do not warn if the three failures is reached.
4711         */
4712        if (++nr_loops > 3)
4713                return NULL;
4714
4715        if (rb_per_cpu_empty(cpu_buffer))
4716                return NULL;
4717
4718        if (iter->head >= rb_page_size(iter->head_page)) {
4719                rb_inc_iter(iter);
4720                goto again;
4721        }
4722
4723        event = rb_iter_head_event(iter);
4724        if (!event)
4725                goto again;
4726
4727        switch (event->type_len) {
4728        case RINGBUF_TYPE_PADDING:
4729                if (rb_null_event(event)) {
4730                        rb_inc_iter(iter);
4731                        goto again;
4732                }
4733                rb_advance_iter(iter);
4734                return event;
4735
4736        case RINGBUF_TYPE_TIME_EXTEND:
4737                /* Internal data, OK to advance */
4738                rb_advance_iter(iter);
4739                goto again;
4740
4741        case RINGBUF_TYPE_TIME_STAMP:
4742                if (ts) {
4743                        *ts = rb_event_time_stamp(event);
4744                        ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4745                                                         cpu_buffer->cpu, ts);
4746                }
4747                /* Internal data, OK to advance */
4748                rb_advance_iter(iter);
4749                goto again;
4750
4751        case RINGBUF_TYPE_DATA:
4752                if (ts && !(*ts)) {
4753                        *ts = iter->read_stamp + event->time_delta;
4754                        ring_buffer_normalize_time_stamp(buffer,
4755                                                         cpu_buffer->cpu, ts);
4756                }
4757                return event;
4758
4759        default:
4760                RB_WARN_ON(cpu_buffer, 1);
4761        }
4762
4763        return NULL;
4764}
4765EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4766
4767static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4768{
4769        if (likely(!in_nmi())) {
4770                raw_spin_lock(&cpu_buffer->reader_lock);
4771                return true;
4772        }
4773
4774        /*
4775         * If an NMI die dumps out the content of the ring buffer
4776         * trylock must be used to prevent a deadlock if the NMI
4777         * preempted a task that holds the ring buffer locks. If
4778         * we get the lock then all is fine, if not, then continue
4779         * to do the read, but this can corrupt the ring buffer,
4780         * so it must be permanently disabled from future writes.
4781         * Reading from NMI is a oneshot deal.
4782         */
4783        if (raw_spin_trylock(&cpu_buffer->reader_lock))
4784                return true;
4785
4786        /* Continue without locking, but disable the ring buffer */
4787        atomic_inc(&cpu_buffer->record_disabled);
4788        return false;
4789}
4790
4791static inline void
4792rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4793{
4794        if (likely(locked))
4795                raw_spin_unlock(&cpu_buffer->reader_lock);
4796        return;
4797}
4798
4799/**
4800 * ring_buffer_peek - peek at the next event to be read
4801 * @buffer: The ring buffer to read
4802 * @cpu: The cpu to peak at
4803 * @ts: The timestamp counter of this event.
4804 * @lost_events: a variable to store if events were lost (may be NULL)
4805 *
4806 * This will return the event that will be read next, but does
4807 * not consume the data.
4808 */
4809struct ring_buffer_event *
4810ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4811                 unsigned long *lost_events)
4812{
4813        struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4814        struct ring_buffer_event *event;
4815        unsigned long flags;
4816        bool dolock;
4817
4818        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4819                return NULL;
4820
4821 again:
4822        local_irq_save(flags);
4823        dolock = rb_reader_lock(cpu_buffer);
4824        event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4825        if (event && event->type_len == RINGBUF_TYPE_PADDING)
4826                rb_advance_reader(cpu_buffer);
4827        rb_reader_unlock(cpu_buffer, dolock);
4828        local_irq_restore(flags);
4829
4830        if (event && event->type_len == RINGBUF_TYPE_PADDING)
4831                goto again;
4832
4833        return event;
4834}
4835
4836/** ring_buffer_iter_dropped - report if there are dropped events
4837 * @iter: The ring buffer iterator
4838 *
4839 * Returns true if there was dropped events since the last peek.
4840 */
4841bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4842{
4843        bool ret = iter->missed_events != 0;
4844
4845        iter->missed_events = 0;
4846        return ret;
4847}
4848EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4849
4850/**
4851 * ring_buffer_iter_peek - peek at the next event to be read
4852 * @iter: The ring buffer iterator
4853 * @ts: The timestamp counter of this event.
4854 *
4855 * This will return the event that will be read next, but does
4856 * not increment the iterator.
4857 */
4858struct ring_buffer_event *
4859ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4860{
4861        struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4862        struct ring_buffer_event *event;
4863        unsigned long flags;
4864
4865 again:
4866        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4867        event = rb_iter_peek(iter, ts);
4868        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4869
4870        if (event && event->type_len == RINGBUF_TYPE_PADDING)
4871                goto again;
4872
4873        return event;
4874}
4875
4876/**
4877 * ring_buffer_consume - return an event and consume it
4878 * @buffer: The ring buffer to get the next event from
4879 * @cpu: the cpu to read the buffer from
4880 * @ts: a variable to store the timestamp (may be NULL)
4881 * @lost_events: a variable to store if events were lost (may be NULL)
4882 *
4883 * Returns the next event in the ring buffer, and that event is consumed.
4884 * Meaning, that sequential reads will keep returning a different event,
4885 * and eventually empty the ring buffer if the producer is slower.
4886 */
4887struct ring_buffer_event *
4888ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4889                    unsigned long *lost_events)
4890{
4891        struct ring_buffer_per_cpu *cpu_buffer;
4892        struct ring_buffer_event *event = NULL;
4893        unsigned long flags;
4894        bool dolock;
4895
4896 again:
4897        /* might be called in atomic */
4898        preempt_disable();
4899
4900        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4901                goto out;
4902
4903        cpu_buffer = buffer->buffers[cpu];
4904        local_irq_save(flags);
4905        dolock = rb_reader_lock(cpu_buffer);
4906
4907        event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4908        if (event) {
4909                cpu_buffer->lost_events = 0;
4910                rb_advance_reader(cpu_buffer);
4911        }
4912
4913        rb_reader_unlock(cpu_buffer, dolock);
4914        local_irq_restore(flags);
4915
4916 out:
4917        preempt_enable();
4918
4919        if (event && event->type_len == RINGBUF_TYPE_PADDING)
4920                goto again;
4921
4922        return event;
4923}
4924EXPORT_SYMBOL_GPL(ring_buffer_consume);
4925
4926/**
4927 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4928 * @buffer: The ring buffer to read from
4929 * @cpu: The cpu buffer to iterate over
4930 * @flags: gfp flags to use for memory allocation
4931 *
4932 * This performs the initial preparations necessary to iterate
4933 * through the buffer.  Memory is allocated, buffer recording
4934 * is disabled, and the iterator pointer is returned to the caller.
4935 *
4936 * Disabling buffer recording prevents the reading from being
4937 * corrupted. This is not a consuming read, so a producer is not
4938 * expected.
4939 *
4940 * After a sequence of ring_buffer_read_prepare calls, the user is
4941 * expected to make at least one call to ring_buffer_read_prepare_sync.
4942 * Afterwards, ring_buffer_read_start is invoked to get things going
4943 * for real.
4944 *
4945 * This overall must be paired with ring_buffer_read_finish.
4946 */
4947struct ring_buffer_iter *
4948ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
4949{
4950        struct ring_buffer_per_cpu *cpu_buffer;
4951        struct ring_buffer_iter *iter;
4952
4953        if (!cpumask_test_cpu(cpu, buffer->cpumask))
4954                return NULL;
4955
4956        iter = kzalloc(sizeof(*iter), flags);
4957        if (!iter)
4958                return NULL;
4959
4960        iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
4961        if (!iter->event) {
4962                kfree(iter);
4963                return NULL;
4964        }
4965
4966        cpu_buffer = buffer->buffers[cpu];
4967
4968        iter->cpu_buffer = cpu_buffer;
4969
4970        atomic_inc(&cpu_buffer->resize_disabled);
4971
4972        return iter;
4973}
4974EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4975
4976/**
4977 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4978 *
4979 * All previously invoked ring_buffer_read_prepare calls to prepare
4980 * iterators will be synchronized.  Afterwards, read_buffer_read_start
4981 * calls on those iterators are allowed.
4982 */
4983void
4984ring_buffer_read_prepare_sync(void)
4985{
4986        synchronize_rcu();
4987}
4988EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4989
4990/**
4991 * ring_buffer_read_start - start a non consuming read of the buffer
4992 * @iter: The iterator returned by ring_buffer_read_prepare
4993 *
4994 * This finalizes the startup of an iteration through the buffer.
4995 * The iterator comes from a call to ring_buffer_read_prepare and
4996 * an intervening ring_buffer_read_prepare_sync must have been
4997 * performed.
4998 *
4999 * Must be paired with ring_buffer_read_finish.
5000 */
5001void
5002ring_buffer_read_start(struct ring_buffer_iter *iter)
5003{
5004        struct ring_buffer_per_cpu *cpu_buffer;
5005        unsigned long flags;
5006
5007        if (!iter)
5008                return;
5009
5010        cpu_buffer = iter->cpu_buffer;
5011
5012        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5013        arch_spin_lock(&cpu_buffer->lock);
5014        rb_iter_reset(iter);
5015        arch_spin_unlock(&cpu_buffer->lock);
5016        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5017}
5018EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5019
5020/**
5021 * ring_buffer_read_finish - finish reading the iterator of the buffer
5022 * @iter: The iterator retrieved by ring_buffer_start
5023 *
5024 * This re-enables the recording to the buffer, and frees the
5025 * iterator.
5026 */
5027void
5028ring_buffer_read_finish(struct ring_buffer_iter *iter)
5029{
5030        struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5031        unsigned long flags;
5032
5033        /*
5034         * Ring buffer is disabled from recording, here's a good place
5035         * to check the integrity of the ring buffer.
5036         * Must prevent readers from trying to read, as the check
5037         * clears the HEAD page and readers require it.
5038         */
5039        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5040        rb_check_pages(cpu_buffer);
5041        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5042
5043        atomic_dec(&cpu_buffer->resize_disabled);
5044        kfree(iter->event);
5045        kfree(iter);
5046}
5047EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5048
5049/**
5050 * ring_buffer_iter_advance - advance the iterator to the next location
5051 * @iter: The ring buffer iterator
5052 *
5053 * Move the location of the iterator such that the next read will
5054 * be the next location of the iterator.
5055 */
5056void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5057{
5058        struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5059        unsigned long flags;
5060
5061        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5062
5063        rb_advance_iter(iter);
5064
5065        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5066}
5067EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5068
5069/**
5070 * ring_buffer_size - return the size of the ring buffer (in bytes)
5071 * @buffer: The ring buffer.
5072 * @cpu: The CPU to get ring buffer size from.
5073 */
5074unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5075{
5076        /*
5077         * Earlier, this method returned
5078         *      BUF_PAGE_SIZE * buffer->nr_pages
5079         * Since the nr_pages field is now removed, we have converted this to
5080         * return the per cpu buffer value.
5081         */
5082        if (!cpumask_test_cpu(cpu, buffer->cpumask))
5083                return 0;
5084
5085        return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
5086}
5087EXPORT_SYMBOL_GPL(ring_buffer_size);
5088
5089static void
5090rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5091{
5092        rb_head_page_deactivate(cpu_buffer);
5093
5094        cpu_buffer->head_page
5095                = list_entry(cpu_buffer->pages, struct buffer_page, list);
5096        local_set(&cpu_buffer->head_page->write, 0);
5097        local_set(&cpu_buffer->head_page->entries, 0);
5098        local_set(&cpu_buffer->head_page->page->commit, 0);
5099
5100        cpu_buffer->head_page->read = 0;
5101
5102        cpu_buffer->tail_page = cpu_buffer->head_page;
5103        cpu_buffer->commit_page = cpu_buffer->head_page;
5104
5105        INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5106        INIT_LIST_HEAD(&cpu_buffer->new_pages);
5107        local_set(&cpu_buffer->reader_page->write, 0);
5108        local_set(&cpu_buffer->reader_page->entries, 0);
5109        local_set(&cpu_buffer->reader_page->page->commit, 0);
5110        cpu_buffer->reader_page->read = 0;
5111
5112        local_set(&cpu_buffer->entries_bytes, 0);
5113        local_set(&cpu_buffer->overrun, 0);
5114        local_set(&cpu_buffer->commit_overrun, 0);
5115        local_set(&cpu_buffer->dropped_events, 0);
5116        local_set(&cpu_buffer->entries, 0);
5117        local_set(&cpu_buffer->committing, 0);
5118        local_set(&cpu_buffer->commits, 0);
5119        local_set(&cpu_buffer->pages_touched, 0);
5120        local_set(&cpu_buffer->pages_read, 0);
5121        cpu_buffer->last_pages_touch = 0;
5122        cpu_buffer->shortest_full = 0;
5123        cpu_buffer->read = 0;
5124        cpu_buffer->read_bytes = 0;
5125
5126        rb_time_set(&cpu_buffer->write_stamp, 0);
5127        rb_time_set(&cpu_buffer->before_stamp, 0);
5128
5129        memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5130
5131        cpu_buffer->lost_events = 0;
5132        cpu_buffer->last_overrun = 0;
5133
5134        rb_head_page_activate(cpu_buffer);
5135}
5136
5137/* Must have disabled the cpu buffer then done a synchronize_rcu */
5138static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5139{
5140        unsigned long flags;
5141
5142        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5143
5144        if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5145                goto out;
5146
5147        arch_spin_lock(&cpu_buffer->lock);
5148
5149        rb_reset_cpu(cpu_buffer);
5150
5151        arch_spin_unlock(&cpu_buffer->lock);
5152
5153 out:
5154        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5155}
5156
5157/**
5158 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5159 * @buffer: The ring buffer to reset a per cpu buffer of
5160 * @cpu: The CPU buffer to be reset
5161 */
5162void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5163{
5164        struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5165
5166        if (!cpumask_test_cpu(cpu, buffer->cpumask))
5167                return;
5168
5169        /* prevent another thread from changing buffer sizes */
5170        mutex_lock(&buffer->mutex);
5171
5172        atomic_inc(&cpu_buffer->resize_disabled);
5173        atomic_inc(&cpu_buffer->record_disabled);
5174
5175        /* Make sure all commits have finished */
5176        synchronize_rcu();
5177
5178        reset_disabled_cpu_buffer(cpu_buffer);
5179
5180        atomic_dec(&cpu_buffer->record_disabled);
5181        atomic_dec(&cpu_buffer->resize_disabled);
5182
5183        mutex_unlock(&buffer->mutex);
5184}
5185EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5186
5187/**
5188 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5189 * @buffer: The ring buffer to reset a per cpu buffer of
5190 * @cpu: The CPU buffer to be reset
5191 */
5192void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5193{
5194        struct ring_buffer_per_cpu *cpu_buffer;
5195        int cpu;
5196
5197        /* prevent another thread from changing buffer sizes */
5198        mutex_lock(&buffer->mutex);
5199
5200        for_each_online_buffer_cpu(buffer, cpu) {
5201                cpu_buffer = buffer->buffers[cpu];
5202
5203                atomic_inc(&cpu_buffer->resize_disabled);
5204                atomic_inc(&cpu_buffer->record_disabled);
5205        }
5206
5207        /* Make sure all commits have finished */
5208        synchronize_rcu();
5209
5210        for_each_online_buffer_cpu(buffer, cpu) {
5211                cpu_buffer = buffer->buffers[cpu];
5212
5213                reset_disabled_cpu_buffer(cpu_buffer);
5214
5215                atomic_dec(&cpu_buffer->record_disabled);
5216                atomic_dec(&cpu_buffer->resize_disabled);
5217        }
5218
5219        mutex_unlock(&buffer->mutex);
5220}
5221
5222/**
5223 * ring_buffer_reset - reset a ring buffer
5224 * @buffer: The ring buffer to reset all cpu buffers
5225 */
5226void ring_buffer_reset(struct trace_buffer *buffer)
5227{
5228        struct ring_buffer_per_cpu *cpu_buffer;
5229        int cpu;
5230
5231        /* prevent another thread from changing buffer sizes */
5232        mutex_lock(&buffer->mutex);
5233
5234        for_each_buffer_cpu(buffer, cpu) {
5235                cpu_buffer = buffer->buffers[cpu];
5236
5237                atomic_inc(&cpu_buffer->resize_disabled);
5238                atomic_inc(&cpu_buffer->record_disabled);
5239        }
5240
5241        /* Make sure all commits have finished */
5242        synchronize_rcu();
5243
5244        for_each_buffer_cpu(buffer, cpu) {
5245                cpu_buffer = buffer->buffers[cpu];
5246
5247                reset_disabled_cpu_buffer(cpu_buffer);
5248
5249                atomic_dec(&cpu_buffer->record_disabled);
5250                atomic_dec(&cpu_buffer->resize_disabled);
5251        }
5252
5253        mutex_unlock(&buffer->mutex);
5254}
5255EXPORT_SYMBOL_GPL(ring_buffer_reset);
5256
5257/**
5258 * rind_buffer_empty - is the ring buffer empty?
5259 * @buffer: The ring buffer to test
5260 */
5261bool ring_buffer_empty(struct trace_buffer *buffer)
5262{
5263        struct ring_buffer_per_cpu *cpu_buffer;
5264        unsigned long flags;
5265        bool dolock;
5266        int cpu;
5267        int ret;
5268
5269        /* yes this is racy, but if you don't like the race, lock the buffer */
5270        for_each_buffer_cpu(buffer, cpu) {
5271                cpu_buffer = buffer->buffers[cpu];
5272                local_irq_save(flags);
5273                dolock = rb_reader_lock(cpu_buffer);
5274                ret = rb_per_cpu_empty(cpu_buffer);
5275                rb_reader_unlock(cpu_buffer, dolock);
5276                local_irq_restore(flags);
5277
5278                if (!ret)
5279                        return false;
5280        }
5281
5282        return true;
5283}
5284EXPORT_SYMBOL_GPL(ring_buffer_empty);
5285
5286/**
5287 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5288 * @buffer: The ring buffer
5289 * @cpu: The CPU buffer to test
5290 */
5291bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5292{
5293        struct ring_buffer_per_cpu *cpu_buffer;
5294        unsigned long flags;
5295        bool dolock;
5296        int ret;
5297
5298        if (!cpumask_test_cpu(cpu, buffer->cpumask))
5299                return true;
5300
5301        cpu_buffer = buffer->buffers[cpu];
5302        local_irq_save(flags);
5303        dolock = rb_reader_lock(cpu_buffer);
5304        ret = rb_per_cpu_empty(cpu_buffer);
5305        rb_reader_unlock(cpu_buffer, dolock);
5306        local_irq_restore(flags);
5307
5308        return ret;
5309}
5310EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5311
5312#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5313/**
5314 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5315 * @buffer_a: One buffer to swap with
5316 * @buffer_b: The other buffer to swap with
5317 * @cpu: the CPU of the buffers to swap
5318 *
5319 * This function is useful for tracers that want to take a "snapshot"
5320 * of a CPU buffer and has another back up buffer lying around.
5321 * it is expected that the tracer handles the cpu buffer not being
5322 * used at the moment.
5323 */
5324int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5325                         struct trace_buffer *buffer_b, int cpu)
5326{
5327        struct ring_buffer_per_cpu *cpu_buffer_a;
5328        struct ring_buffer_per_cpu *cpu_buffer_b;
5329        int ret = -EINVAL;
5330
5331        if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5332            !cpumask_test_cpu(cpu, buffer_b->cpumask))
5333                goto out;
5334
5335        cpu_buffer_a = buffer_a->buffers[cpu];
5336        cpu_buffer_b = buffer_b->buffers[cpu];
5337
5338        /* At least make sure the two buffers are somewhat the same */
5339        if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5340                goto out;
5341
5342        ret = -EAGAIN;
5343
5344        if (atomic_read(&buffer_a->record_disabled))
5345                goto out;
5346
5347        if (atomic_read(&buffer_b->record_disabled))
5348                goto out;
5349
5350        if (atomic_read(&cpu_buffer_a->record_disabled))
5351                goto out;
5352
5353        if (atomic_read(&cpu_buffer_b->record_disabled))
5354                goto out;
5355
5356        /*
5357         * We can't do a synchronize_rcu here because this
5358         * function can be called in atomic context.
5359         * Normally this will be called from the same CPU as cpu.
5360         * If not it's up to the caller to protect this.
5361         */
5362        atomic_inc(&cpu_buffer_a->record_disabled);
5363        atomic_inc(&cpu_buffer_b->record_disabled);
5364
5365        ret = -EBUSY;
5366        if (local_read(&cpu_buffer_a->committing))
5367                goto out_dec;
5368        if (local_read(&cpu_buffer_b->committing))
5369                goto out_dec;
5370
5371        buffer_a->buffers[cpu] = cpu_buffer_b;
5372        buffer_b->buffers[cpu] = cpu_buffer_a;
5373
5374        cpu_buffer_b->buffer = buffer_a;
5375        cpu_buffer_a->buffer = buffer_b;
5376
5377        ret = 0;
5378
5379out_dec:
5380        atomic_dec(&cpu_buffer_a->record_disabled);
5381        atomic_dec(&cpu_buffer_b->record_disabled);
5382out:
5383        return ret;
5384}
5385EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5386#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5387
5388/**
5389 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5390 * @buffer: the buffer to allocate for.
5391 * @cpu: the cpu buffer to allocate.
5392 *
5393 * This function is used in conjunction with ring_buffer_read_page.
5394 * When reading a full page from the ring buffer, these functions
5395 * can be used to speed up the process. The calling function should
5396 * allocate a few pages first with this function. Then when it
5397 * needs to get pages from the ring buffer, it passes the result
5398 * of this function into ring_buffer_read_page, which will swap
5399 * the page that was allocated, with the read page of the buffer.
5400 *
5401 * Returns:
5402 *  The page allocated, or ERR_PTR
5403 */
5404void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5405{
5406        struct ring_buffer_per_cpu *cpu_buffer;
5407        struct buffer_data_page *bpage = NULL;
5408        unsigned long flags;
5409        struct page *page;
5410
5411        if (!cpumask_test_cpu(cpu, buffer->cpumask))
5412                return ERR_PTR(-ENODEV);
5413
5414        cpu_buffer = buffer->buffers[cpu];
5415        local_irq_save(flags);
5416        arch_spin_lock(&cpu_buffer->lock);
5417
5418        if (cpu_buffer->free_page) {
5419                bpage = cpu_buffer->free_page;
5420                cpu_buffer->free_page = NULL;
5421        }
5422
5423        arch_spin_unlock(&cpu_buffer->lock);
5424        local_irq_restore(flags);
5425
5426        if (bpage)
5427                goto out;
5428
5429        page = alloc_pages_node(cpu_to_node(cpu),
5430                                GFP_KERNEL | __GFP_NORETRY, 0);
5431        if (!page)
5432                return ERR_PTR(-ENOMEM);
5433
5434        bpage = page_address(page);
5435
5436 out:
5437        rb_init_page(bpage);
5438
5439        return bpage;
5440}
5441EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5442
5443/**
5444 * ring_buffer_free_read_page - free an allocated read page
5445 * @buffer: the buffer the page was allocate for
5446 * @cpu: the cpu buffer the page came from
5447 * @data: the page to free
5448 *
5449 * Free a page allocated from ring_buffer_alloc_read_page.
5450 */
5451void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5452{
5453        struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5454        struct buffer_data_page *bpage = data;
5455        struct page *page = virt_to_page(bpage);
5456        unsigned long flags;
5457
5458        /* If the page is still in use someplace else, we can't reuse it */
5459        if (page_ref_count(page) > 1)
5460                goto out;
5461
5462        local_irq_save(flags);
5463        arch_spin_lock(&cpu_buffer->lock);
5464
5465        if (!cpu_buffer->free_page) {
5466                cpu_buffer->free_page = bpage;
5467                bpage = NULL;
5468        }
5469
5470        arch_spin_unlock(&cpu_buffer->lock);
5471        local_irq_restore(flags);
5472
5473 out:
5474        free_page((unsigned long)bpage);
5475}
5476EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5477
5478/**
5479 * ring_buffer_read_page - extract a page from the ring buffer
5480 * @buffer: buffer to extract from
5481 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5482 * @len: amount to extract
5483 * @cpu: the cpu of the buffer to extract
5484 * @full: should the extraction only happen when the page is full.
5485 *
5486 * This function will pull out a page from the ring buffer and consume it.
5487 * @data_page must be the address of the variable that was returned
5488 * from ring_buffer_alloc_read_page. This is because the page might be used
5489 * to swap with a page in the ring buffer.
5490 *
5491 * for example:
5492 *      rpage = ring_buffer_alloc_read_page(buffer, cpu);
5493 *      if (IS_ERR(rpage))
5494 *              return PTR_ERR(rpage);
5495 *      ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5496 *      if (ret >= 0)
5497 *              process_page(rpage, ret);
5498 *
5499 * When @full is set, the function will not return true unless
5500 * the writer is off the reader page.
5501 *
5502 * Note: it is up to the calling functions to handle sleeps and wakeups.
5503 *  The ring buffer can be used anywhere in the kernel and can not
5504 *  blindly call wake_up. The layer that uses the ring buffer must be
5505 *  responsible for that.
5506 *
5507 * Returns:
5508 *  >=0 if data has been transferred, returns the offset of consumed data.
5509 *  <0 if no data has been transferred.
5510 */
5511int ring_buffer_read_page(struct trace_buffer *buffer,
5512                          void **data_page, size_t len, int cpu, int full)
5513{
5514        struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5515        struct ring_buffer_event *event;
5516        struct buffer_data_page *bpage;
5517        struct buffer_page *reader;
5518        unsigned long missed_events;
5519        unsigned long flags;
5520        unsigned int commit;
5521        unsigned int read;
5522        u64 save_timestamp;
5523        int ret = -1;
5524
5525        if (!cpumask_test_cpu(cpu, buffer->cpumask))
5526                goto out;
5527
5528        /*
5529         * If len is not big enough to hold the page header, then
5530         * we can not copy anything.
5531         */
5532        if (len <= BUF_PAGE_HDR_SIZE)
5533                goto out;
5534
5535        len -= BUF_PAGE_HDR_SIZE;
5536
5537        if (!data_page)
5538                goto out;
5539
5540        bpage = *data_page;
5541        if (!bpage)
5542                goto out;
5543
5544        raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5545
5546        reader = rb_get_reader_page(cpu_buffer);
5547        if (!reader)
5548                goto out_unlock;
5549
5550        event = rb_reader_event(cpu_buffer);
5551
5552        read = reader->read;
5553        commit = rb_page_commit(reader);
5554
5555        /* Check if any events were dropped */
5556        missed_events = cpu_buffer->lost_events;
5557
5558        /*
5559         * If this page has been partially read or
5560         * if len is not big enough to read the rest of the page or
5561         * a writer is still on the page, then
5562         * we must copy the data from the page to the buffer.
5563         * Otherwise, we can simply swap the page with the one passed in.
5564         */
5565        if (read || (len < (commit - read)) ||
5566            cpu_buffer->reader_page == cpu_buffer->commit_page) {
5567                struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5568                unsigned int rpos = read;
5569                unsigned int pos = 0;
5570                unsigned int size;
5571
5572                if (full)
5573                        goto out_unlock;
5574
5575                if (len > (commit - read))
5576                        len = (commit - read);
5577
5578                /* Always keep the time extend and data together */
5579                size = rb_event_ts_length(event);
5580
5581                if (len < size)
5582                        goto out_unlock;
5583
5584                /* save the current timestamp, since the user will need it */
5585                save_timestamp = cpu_buffer->read_stamp;
5586
5587                /* Need to copy one event at a time */
5588                do {
5589                        /* We need the size of one event, because
5590                         * rb_advance_reader only advances by one event,
5591                         * whereas rb_event_ts_length may include the size of
5592                         * one or two events.
5593                         * We have already ensured there's enough space if this
5594                         * is a time extend. */
5595                        size = rb_event_length(event);
5596                        memcpy(bpage->data + pos, rpage->data + rpos, size);
5597
5598                        len -= size;
5599
5600                        rb_advance_reader(cpu_buffer);
5601                        rpos = reader->read;
5602                        pos += size;
5603
5604                        if (rpos >= commit)
5605                                break;
5606
5607                        event = rb_reader_event(cpu_buffer);
5608                        /* Always keep the time extend and data together */
5609                        size = rb_event_ts_length(event);
5610                } while (len >= size);
5611
5612                /* update bpage */
5613                local_set(&bpage->commit, pos);
5614                bpage->time_stamp = save_timestamp;
5615
5616                /* we copied everything to the beginning */
5617                read = 0;
5618        } else {
5619                /* update the entry counter */
5620                cpu_buffer->read += rb_page_entries(reader);
5621                cpu_buffer->read_bytes += BUF_PAGE_SIZE;
5622
5623                /* swap the pages */
5624                rb_init_page(bpage);
5625                bpage = reader->page;
5626                reader->page = *data_page;
5627                local_set(&reader->write, 0);
5628                local_set(&reader->entries, 0);
5629                reader->read = 0;
5630                *data_page = bpage;
5631
5632                /*
5633                 * Use the real_end for the data size,
5634                 * This gives us a chance to store the lost events
5635                 * on the page.
5636                 */
5637                if (reader->real_end)
5638                        local_set(&bpage->commit, reader->real_end);
5639        }
5640        ret = read;
5641
5642        cpu_buffer->lost_events = 0;
5643
5644        commit = local_read(&bpage->commit);
5645        /*
5646         * Set a flag in the commit field if we lost events
5647         */
5648        if (missed_events) {
5649                /* If there is room at the end of the page to save the
5650                 * missed events, then record it there.
5651                 */
5652                if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5653                        memcpy(&bpage->data[commit], &missed_events,
5654                               sizeof(missed_events));
5655                        local_add(RB_MISSED_STORED, &bpage->commit);
5656                        commit += sizeof(missed_events);
5657                }
5658                local_add(RB_MISSED_EVENTS, &bpage->commit);
5659        }
5660
5661        /*
5662         * This page may be off to user land. Zero it out here.
5663         */
5664        if (commit < BUF_PAGE_SIZE)
5665                memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5666
5667 out_unlock:
5668        raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5669
5670 out:
5671        return ret;
5672}
5673EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5674
5675/*
5676 * We only allocate new buffers, never free them if the CPU goes down.
5677 * If we were to free the buffer, then the user would lose any trace that was in
5678 * the buffer.
5679 */
5680int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5681{
5682        struct trace_buffer *buffer;
5683        long nr_pages_same;
5684        int cpu_i;
5685        unsigned long nr_pages;
5686
5687        buffer = container_of(node, struct trace_buffer, node);
5688        if (cpumask_test_cpu(cpu, buffer->cpumask))
5689                return 0;
5690
5691        nr_pages = 0;
5692        nr_pages_same = 1;
5693        /* check if all cpu sizes are same */
5694        for_each_buffer_cpu(buffer, cpu_i) {
5695                /* fill in the size from first enabled cpu */
5696                if (nr_pages == 0)
5697                        nr_pages = buffer->buffers[cpu_i]->nr_pages;
5698                if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5699                        nr_pages_same = 0;
5700                        break;
5701                }
5702        }
5703        /* allocate minimum pages, user can later expand it */
5704        if (!nr_pages_same)
5705                nr_pages = 2;
5706        buffer->buffers[cpu] =
5707                rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5708        if (!buffer->buffers[cpu]) {
5709                WARN(1, "failed to allocate ring buffer on CPU %u\n",
5710                     cpu);
5711                return -ENOMEM;
5712        }
5713        smp_wmb();
5714        cpumask_set_cpu(cpu, buffer->cpumask);
5715        return 0;
5716}
5717
5718#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5719/*
5720 * This is a basic integrity check of the ring buffer.
5721 * Late in the boot cycle this test will run when configured in.
5722 * It will kick off a thread per CPU that will go into a loop
5723 * writing to the per cpu ring buffer various sizes of data.
5724 * Some of the data will be large items, some small.
5725 *
5726 * Another thread is created that goes into a spin, sending out
5727 * IPIs to the other CPUs to also write into the ring buffer.
5728 * this is to test the nesting ability of the buffer.
5729 *
5730 * Basic stats are recorded and reported. If something in the
5731 * ring buffer should happen that's not expected, a big warning
5732 * is displayed and all ring buffers are disabled.
5733 */
5734static struct task_struct *rb_threads[NR_CPUS] __initdata;
5735
5736struct rb_test_data {
5737        struct trace_buffer *buffer;
5738        unsigned long           events;
5739        unsigned long           bytes_written;
5740        unsigned long           bytes_alloc;
5741        unsigned long           bytes_dropped;
5742        unsigned long           events_nested;
5743        unsigned long           bytes_written_nested;
5744        unsigned long           bytes_alloc_nested;
5745        unsigned long           bytes_dropped_nested;
5746        int                     min_size_nested;
5747        int                     max_size_nested;
5748        int                     max_size;
5749        int                     min_size;
5750        int                     cpu;
5751        int                     cnt;
5752};
5753
5754static struct rb_test_data rb_data[NR_CPUS] __initdata;
5755
5756/* 1 meg per cpu */
5757#define RB_TEST_BUFFER_SIZE     1048576
5758
5759static char rb_string[] __initdata =
5760        "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5761        "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5762        "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5763
5764static bool rb_test_started __initdata;
5765
5766struct rb_item {
5767        int size;
5768        char str[];
5769};
5770
5771static __init int rb_write_something(struct rb_test_data *data, bool nested)
5772{
5773        struct ring_buffer_event *event;
5774        struct rb_item *item;
5775        bool started;
5776        int event_len;
5777        int size;
5778        int len;
5779        int cnt;
5780
5781        /* Have nested writes different that what is written */
5782        cnt = data->cnt + (nested ? 27 : 0);
5783
5784        /* Multiply cnt by ~e, to make some unique increment */
5785        size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5786
5787        len = size + sizeof(struct rb_item);
5788
5789        started = rb_test_started;
5790        /* read rb_test_started before checking buffer enabled */
5791        smp_rmb();
5792
5793        event = ring_buffer_lock_reserve(data->buffer, len);
5794        if (!event) {
5795                /* Ignore dropped events before test starts. */
5796                if (started) {
5797                        if (nested)
5798                                data->bytes_dropped += len;
5799                        else
5800                                data->bytes_dropped_nested += len;
5801                }
5802                return len;
5803        }
5804
5805        event_len = ring_buffer_event_length(event);
5806
5807        if (RB_WARN_ON(data->buffer, event_len < len))
5808                goto out;
5809
5810        item = ring_buffer_event_data(event);
5811        item->size = size;
5812        memcpy(item->str, rb_string, size);
5813
5814        if (nested) {
5815                data->bytes_alloc_nested += event_len;
5816                data->bytes_written_nested += len;
5817                data->events_nested++;
5818                if (!data->min_size_nested || len < data->min_size_nested)
5819                        data->min_size_nested = len;
5820                if (len > data->max_size_nested)
5821                        data->max_size_nested = len;
5822        } else {
5823                data->bytes_alloc += event_len;
5824                data->bytes_written += len;
5825                data->events++;
5826                if (!data->min_size || len < data->min_size)
5827                        data->max_size = len;
5828                if (len > data->max_size)
5829                        data->max_size = len;
5830        }
5831
5832 out:
5833        ring_buffer_unlock_commit(data->buffer, event);
5834
5835        return 0;
5836}
5837
5838static __init int rb_test(void *arg)
5839{
5840        struct rb_test_data *data = arg;
5841
5842        while (!kthread_should_stop()) {
5843                rb_write_something(data, false);
5844                data->cnt++;
5845
5846                set_current_state(TASK_INTERRUPTIBLE);
5847                /* Now sleep between a min of 100-300us and a max of 1ms */
5848                usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5849        }
5850
5851        return 0;
5852}
5853
5854static __init void rb_ipi(void *ignore)
5855{
5856        struct rb_test_data *data;
5857        int cpu = smp_processor_id();
5858
5859        data = &rb_data[cpu];
5860        rb_write_something(data, true);
5861}
5862
5863static __init int rb_hammer_test(void *arg)
5864{
5865        while (!kthread_should_stop()) {
5866
5867                /* Send an IPI to all cpus to write data! */
5868                smp_call_function(rb_ipi, NULL, 1);
5869                /* No sleep, but for non preempt, let others run */
5870                schedule();
5871        }
5872
5873        return 0;
5874}
5875
5876static __init int test_ringbuffer(void)
5877{
5878        struct task_struct *rb_hammer;
5879        struct trace_buffer *buffer;
5880        int cpu;
5881        int ret = 0;
5882
5883        if (security_locked_down(LOCKDOWN_TRACEFS)) {
5884                pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
5885                return 0;
5886        }
5887
5888        pr_info("Running ring buffer tests...\n");
5889
5890        buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5891        if (WARN_ON(!buffer))
5892                return 0;
5893
5894        /* Disable buffer so that threads can't write to it yet */
5895        ring_buffer_record_off(buffer);
5896
5897        for_each_online_cpu(cpu) {
5898                rb_data[cpu].buffer = buffer;
5899                rb_data[cpu].cpu = cpu;
5900                rb_data[cpu].cnt = cpu;
5901                rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
5902                                                     cpu, "rbtester/%u");
5903                if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5904                        pr_cont("FAILED\n");
5905                        ret = PTR_ERR(rb_threads[cpu]);
5906                        goto out_free;
5907                }
5908        }
5909
5910        /* Now create the rb hammer! */
5911        rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5912        if (WARN_ON(IS_ERR(rb_hammer))) {
5913                pr_cont("FAILED\n");
5914                ret = PTR_ERR(rb_hammer);
5915                goto out_free;
5916        }
5917
5918        ring_buffer_record_on(buffer);
5919        /*
5920         * Show buffer is enabled before setting rb_test_started.
5921         * Yes there's a small race window where events could be
5922         * dropped and the thread wont catch it. But when a ring
5923         * buffer gets enabled, there will always be some kind of
5924         * delay before other CPUs see it. Thus, we don't care about
5925         * those dropped events. We care about events dropped after
5926         * the threads see that the buffer is active.
5927         */
5928        smp_wmb();
5929        rb_test_started = true;
5930
5931        set_current_state(TASK_INTERRUPTIBLE);
5932        /* Just run for 10 seconds */;
5933        schedule_timeout(10 * HZ);
5934
5935        kthread_stop(rb_hammer);
5936
5937 out_free:
5938        for_each_online_cpu(cpu) {
5939                if (!rb_threads[cpu])
5940                        break;
5941                kthread_stop(rb_threads[cpu]);
5942        }
5943        if (ret) {
5944                ring_buffer_free(buffer);
5945                return ret;
5946        }
5947
5948        /* Report! */
5949        pr_info("finished\n");
5950        for_each_online_cpu(cpu) {
5951                struct ring_buffer_event *event;
5952                struct rb_test_data *data = &rb_data[cpu];
5953                struct rb_item *item;
5954                unsigned long total_events;
5955                unsigned long total_dropped;
5956                unsigned long total_written;
5957                unsigned long total_alloc;
5958                unsigned long total_read = 0;
5959                unsigned long total_size = 0;
5960                unsigned long total_len = 0;
5961                unsigned long total_lost = 0;
5962                unsigned long lost;
5963                int big_event_size;
5964                int small_event_size;
5965
5966                ret = -1;
5967
5968                total_events = data->events + data->events_nested;
5969                total_written = data->bytes_written + data->bytes_written_nested;
5970                total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5971                total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5972
5973                big_event_size = data->max_size + data->max_size_nested;
5974                small_event_size = data->min_size + data->min_size_nested;
5975
5976                pr_info("CPU %d:\n", cpu);
5977                pr_info("              events:    %ld\n", total_events);
5978                pr_info("       dropped bytes:    %ld\n", total_dropped);
5979                pr_info("       alloced bytes:    %ld\n", total_alloc);
5980                pr_info("       written bytes:    %ld\n", total_written);
5981                pr_info("       biggest event:    %d\n", big_event_size);
5982                pr_info("      smallest event:    %d\n", small_event_size);
5983
5984                if (RB_WARN_ON(buffer, total_dropped))
5985                        break;
5986
5987                ret = 0;
5988
5989                while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5990                        total_lost += lost;
5991                        item = ring_buffer_event_data(event);
5992                        total_len += ring_buffer_event_length(event);
5993                        total_size += item->size + sizeof(struct rb_item);
5994                        if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5995                                pr_info("FAILED!\n");
5996                                pr_info("buffer had: %.*s\n", item->size, item->str);
5997                                pr_info("expected:   %.*s\n", item->size, rb_string);
5998                                RB_WARN_ON(buffer, 1);
5999                                ret = -1;
6000                                break;
6001                        }
6002                        total_read++;
6003                }
6004                if (ret)
6005                        break;
6006
6007                ret = -1;
6008
6009                pr_info("         read events:   %ld\n", total_read);
6010                pr_info("         lost events:   %ld\n", total_lost);
6011                pr_info("        total events:   %ld\n", total_lost + total_read);
6012                pr_info("  recorded len bytes:   %ld\n", total_len);
6013                pr_info(" recorded size bytes:   %ld\n", total_size);
6014                if (total_lost)
6015                        pr_info(" With dropped events, record len and size may not match\n"
6016                                " alloced and written from above\n");
6017                if (!total_lost) {
6018                        if (RB_WARN_ON(buffer, total_len != total_alloc ||
6019                                       total_size != total_written))
6020                                break;
6021                }
6022                if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6023                        break;
6024
6025                ret = 0;
6026        }
6027        if (!ret)
6028                pr_info("Ring buffer PASSED!\n");
6029
6030        ring_buffer_free(buffer);
6031        return 0;
6032}
6033
6034late_initcall(test_ringbuffer);
6035#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
6036