kernel/trace/ring_buffer.c

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