kernel/trace/ring_buffer.c

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