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;\tsigned:%u;\n", 401 (unsigned int)sizeof(field.time_stamp), 402 (unsigned int)is_signed_type(u64)); 403 404 ret = trace_seq_printf(s, "\tfield: local_t commit;\t" 405 "offset:%u;\tsize:%u;\tsigned:%u;\n", 406 (unsigned int)offsetof(typeof(field), commit), 407 (unsigned int)sizeof(field.commit), 408 (unsigned int)is_signed_type(long)); 409 410 ret = trace_seq_printf(s, "\tfield: char data;\t" 411 "offset:%u;\tsize:%u;\tsigned:%u;\n", 412 (unsigned int)offsetof(typeof(field), data), 413 (unsigned int)BUF_PAGE_SIZE, 414 (unsigned int)is_signed_type(char)); 415 416 return ret; 417 } 418 419 /* 420 * head_page == tail_page && head == tail then buffer is empty. 421 */ 422 struct ring_buffer_per_cpu { 423 int cpu; 424 struct ring_buffer *buffer; 425 spinlock_t reader_lock; /* serialize readers */ 426 arch_spinlock_t lock; 427 struct lock_class_key lock_key; 428 struct list_head *pages; 429 struct buffer_page *head_page; /* read from head */ 430 struct buffer_page *tail_page; /* write to tail */ 431 struct buffer_page *commit_page; /* committed pages */ 432 struct buffer_page *reader_page; 433 local_t commit_overrun; 434 local_t overrun; 435 local_t entries; 436 local_t committing; 437 local_t commits; 438 unsigned long read; 439 u64 write_stamp; 440 u64 read_stamp; 441 atomic_t record_disabled; 442 }; 443 444 struct ring_buffer { 445 unsigned pages; 446 unsigned flags; 447 int cpus; 448 atomic_t record_disabled; 449 cpumask_var_t cpumask; 450 451 struct lock_class_key *reader_lock_key; 452 453 struct mutex mutex; 454 455 struct ring_buffer_per_cpu **buffers; 456 457 #ifdef CONFIG_HOTPLUG_CPU 458 struct notifier_block cpu_notify; 459 #endif 460 u64 (*clock)(void); 461 }; 462 463 struct ring_buffer_iter { 464 struct ring_buffer_per_cpu *cpu_buffer; 465 unsigned long head; 466 struct buffer_page *head_page; 467 struct buffer_page *cache_reader_page; 468 unsigned long cache_read; 469 u64 read_stamp; 470 }; 471 472 /* buffer may be either ring_buffer or ring_buffer_per_cpu */ 473 #define RB_WARN_ON(b, cond) \ 474 ({ \ 475 int _____ret = unlikely(cond); \ 476 if (_____ret) { \ 477 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \ 478 struct ring_buffer_per_cpu *__b = \ 479 (void *)b; \ 480 atomic_inc(&__b->buffer->record_disabled); \ 481 } else \ 482 atomic_inc(&b->record_disabled); \ 483 WARN_ON(1); \ 484 } \ 485 _____ret; \ 486 }) 487 488 /* Up this if you want to test the TIME_EXTENTS and normalization */ 489 #define DEBUG_SHIFT 0 490 491 static inline u64 rb_time_stamp(struct ring_buffer *buffer) 492 { 493 /* shift to debug/test normalization and TIME_EXTENTS */ 494 return buffer->clock() << DEBUG_SHIFT; 495 } 496 497 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu) 498 { 499 u64 time; 500 501 preempt_disable_notrace(); 502 time = rb_time_stamp(buffer); 503 preempt_enable_no_resched_notrace(); 504 505 return time; 506 } 507 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp); 508 509 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer, 510 int cpu, u64 *ts) 511 { 512 /* Just stupid testing the normalize function and deltas */ 513 *ts >>= DEBUG_SHIFT; 514 } 515 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp); 516 517 /* 518 * Making the ring buffer lockless makes things tricky. 519 * Although writes only happen on the CPU that they are on, 520 * and they only need to worry about interrupts. Reads can 521 * happen on any CPU. 522 * 523 * The reader page is always off the ring buffer, but when the 524 * reader finishes with a page, it needs to swap its page with 525 * a new one from the buffer. The reader needs to take from 526 * the head (writes go to the tail). But if a writer is in overwrite 527 * mode and wraps, it must push the head page forward. 528 * 529 * Here lies the problem. 530 * 531 * The reader must be careful to replace only the head page, and 532 * not another one. As described at the top of the file in the 533 * ASCII art, the reader sets its old page to point to the next 534 * page after head. It then sets the page after head to point to 535 * the old reader page. But if the writer moves the head page 536 * during this operation, the reader could end up with the tail. 537 * 538 * We use cmpxchg to help prevent this race. We also do something 539 * special with the page before head. We set the LSB to 1. 540 * 541 * When the writer must push the page forward, it will clear the 542 * bit that points to the head page, move the head, and then set 543 * the bit that points to the new head page. 544 * 545 * We also don't want an interrupt coming in and moving the head 546 * page on another writer. Thus we use the second LSB to catch 547 * that too. Thus: 548 * 549 * head->list->prev->next bit 1 bit 0 550 * ------- ------- 551 * Normal page 0 0 552 * Points to head page 0 1 553 * New head page 1 0 554 * 555 * Note we can not trust the prev pointer of the head page, because: 556 * 557 * +----+ +-----+ +-----+ 558 * | |------>| T |---X--->| N | 559 * | |<------| | | | 560 * +----+ +-----+ +-----+ 561 * ^ ^ | 562 * | +-----+ | | 563 * +----------| R |----------+ | 564 * | |<-----------+ 565 * +-----+ 566 * 567 * Key: ---X--> HEAD flag set in pointer 568 * T Tail page 569 * R Reader page 570 * N Next page 571 * 572 * (see __rb_reserve_next() to see where this happens) 573 * 574 * What the above shows is that the reader just swapped out 575 * the reader page with a page in the buffer, but before it 576 * could make the new header point back to the new page added 577 * it was preempted by a writer. The writer moved forward onto 578 * the new page added by the reader and is about to move forward 579 * again. 580 * 581 * You can see, it is legitimate for the previous pointer of 582 * the head (or any page) not to point back to itself. But only 583 * temporarially. 584 */ 585 586 #define RB_PAGE_NORMAL 0UL 587 #define RB_PAGE_HEAD 1UL 588 #define RB_PAGE_UPDATE 2UL 589 590 591 #define RB_FLAG_MASK 3UL 592 593 /* PAGE_MOVED is not part of the mask */ 594 #define RB_PAGE_MOVED 4UL 595 596 /* 597 * rb_list_head - remove any bit 598 */ 599 static struct list_head *rb_list_head(struct list_head *list) 600 { 601 unsigned long val = (unsigned long)list; 602 603 return (struct list_head *)(val & ~RB_FLAG_MASK); 604 } 605 606 /* 607 * rb_is_head_page - test if the given page is the head page 608 * 609 * Because the reader may move the head_page pointer, we can 610 * not trust what the head page is (it may be pointing to 611 * the reader page). But if the next page is a header page, 612 * its flags will be non zero. 613 */ 614 static int inline 615 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer, 616 struct buffer_page *page, struct list_head *list) 617 { 618 unsigned long val; 619 620 val = (unsigned long)list->next; 621 622 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list) 623 return RB_PAGE_MOVED; 624 625 return val & RB_FLAG_MASK; 626 } 627 628 /* 629 * rb_is_reader_page 630 * 631 * The unique thing about the reader page, is that, if the 632 * writer is ever on it, the previous pointer never points 633 * back to the reader page. 634 */ 635 static int rb_is_reader_page(struct buffer_page *page) 636 { 637 struct list_head *list = page->list.prev; 638 639 return rb_list_head(list->next) != &page->list; 640 } 641 642 /* 643 * rb_set_list_to_head - set a list_head to be pointing to head. 644 */ 645 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer, 646 struct list_head *list) 647 { 648 unsigned long *ptr; 649 650 ptr = (unsigned long *)&list->next; 651 *ptr |= RB_PAGE_HEAD; 652 *ptr &= ~RB_PAGE_UPDATE; 653 } 654 655 /* 656 * rb_head_page_activate - sets up head page 657 */ 658 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer) 659 { 660 struct buffer_page *head; 661 662 head = cpu_buffer->head_page; 663 if (!head) 664 return; 665 666 /* 667 * Set the previous list pointer to have the HEAD flag. 668 */ 669 rb_set_list_to_head(cpu_buffer, head->list.prev); 670 } 671 672 static void rb_list_head_clear(struct list_head *list) 673 { 674 unsigned long *ptr = (unsigned long *)&list->next; 675 676 *ptr &= ~RB_FLAG_MASK; 677 } 678 679 /* 680 * rb_head_page_dactivate - clears head page ptr (for free list) 681 */ 682 static void 683 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer) 684 { 685 struct list_head *hd; 686 687 /* Go through the whole list and clear any pointers found. */ 688 rb_list_head_clear(cpu_buffer->pages); 689 690 list_for_each(hd, cpu_buffer->pages) 691 rb_list_head_clear(hd); 692 } 693 694 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer, 695 struct buffer_page *head, 696 struct buffer_page *prev, 697 int old_flag, int new_flag) 698 { 699 struct list_head *list; 700 unsigned long val = (unsigned long)&head->list; 701 unsigned long ret; 702 703 list = &prev->list; 704 705 val &= ~RB_FLAG_MASK; 706 707 ret = cmpxchg((unsigned long *)&list->next, 708 val | old_flag, val | new_flag); 709 710 /* check if the reader took the page */ 711 if ((ret & ~RB_FLAG_MASK) != val) 712 return RB_PAGE_MOVED; 713 714 return ret & RB_FLAG_MASK; 715 } 716 717 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer, 718 struct buffer_page *head, 719 struct buffer_page *prev, 720 int old_flag) 721 { 722 return rb_head_page_set(cpu_buffer, head, prev, 723 old_flag, RB_PAGE_UPDATE); 724 } 725 726 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer, 727 struct buffer_page *head, 728 struct buffer_page *prev, 729 int old_flag) 730 { 731 return rb_head_page_set(cpu_buffer, head, prev, 732 old_flag, RB_PAGE_HEAD); 733 } 734 735 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer, 736 struct buffer_page *head, 737 struct buffer_page *prev, 738 int old_flag) 739 { 740 return rb_head_page_set(cpu_buffer, head, prev, 741 old_flag, RB_PAGE_NORMAL); 742 } 743 744 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer, 745 struct buffer_page **bpage) 746 { 747 struct list_head *p = rb_list_head((*bpage)->list.next); 748 749 *bpage = list_entry(p, struct buffer_page, list); 750 } 751 752 static struct buffer_page * 753 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer) 754 { 755 struct buffer_page *head; 756 struct buffer_page *page; 757 struct list_head *list; 758 int i; 759 760 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page)) 761 return NULL; 762 763 /* sanity check */ 764 list = cpu_buffer->pages; 765 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list)) 766 return NULL; 767 768 page = head = cpu_buffer->head_page; 769 /* 770 * It is possible that the writer moves the header behind 771 * where we started, and we miss in one loop. 772 * A second loop should grab the header, but we'll do 773 * three loops just because I'm paranoid. 774 */ 775 for (i = 0; i < 3; i++) { 776 do { 777 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) { 778 cpu_buffer->head_page = page; 779 return page; 780 } 781 rb_inc_page(cpu_buffer, &page); 782 } while (page != head); 783 } 784 785 RB_WARN_ON(cpu_buffer, 1); 786 787 return NULL; 788 } 789 790 static int rb_head_page_replace(struct buffer_page *old, 791 struct buffer_page *new) 792 { 793 unsigned long *ptr = (unsigned long *)&old->list.prev->next; 794 unsigned long val; 795 unsigned long ret; 796 797 val = *ptr & ~RB_FLAG_MASK; 798 val |= RB_PAGE_HEAD; 799 800 ret = cmpxchg(ptr, val, (unsigned long)&new->list); 801 802 return ret == val; 803 } 804 805 /* 806 * rb_tail_page_update - move the tail page forward 807 * 808 * Returns 1 if moved tail page, 0 if someone else did. 809 */ 810 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer, 811 struct buffer_page *tail_page, 812 struct buffer_page *next_page) 813 { 814 struct buffer_page *old_tail; 815 unsigned long old_entries; 816 unsigned long old_write; 817 int ret = 0; 818 819 /* 820 * The tail page now needs to be moved forward. 821 * 822 * We need to reset the tail page, but without messing 823 * with possible erasing of data brought in by interrupts 824 * that have moved the tail page and are currently on it. 825 * 826 * We add a counter to the write field to denote this. 827 */ 828 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write); 829 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries); 830 831 /* 832 * Just make sure we have seen our old_write and synchronize 833 * with any interrupts that come in. 834 */ 835 barrier(); 836 837 /* 838 * If the tail page is still the same as what we think 839 * it is, then it is up to us to update the tail 840 * pointer. 841 */ 842 if (tail_page == cpu_buffer->tail_page) { 843 /* Zero the write counter */ 844 unsigned long val = old_write & ~RB_WRITE_MASK; 845 unsigned long eval = old_entries & ~RB_WRITE_MASK; 846 847 /* 848 * This will only succeed if an interrupt did 849 * not come in and change it. In which case, we 850 * do not want to modify it. 851 * 852 * We add (void) to let the compiler know that we do not care 853 * about the return value of these functions. We use the 854 * cmpxchg to only update if an interrupt did not already 855 * do it for us. If the cmpxchg fails, we don't care. 856 */ 857 (void)local_cmpxchg(&next_page->write, old_write, val); 858 (void)local_cmpxchg(&next_page->entries, old_entries, eval); 859 860 /* 861 * No need to worry about races with clearing out the commit. 862 * it only can increment when a commit takes place. But that 863 * only happens in the outer most nested commit. 864 */ 865 local_set(&next_page->page->commit, 0); 866 867 old_tail = cmpxchg(&cpu_buffer->tail_page, 868 tail_page, next_page); 869 870 if (old_tail == tail_page) 871 ret = 1; 872 } 873 874 return ret; 875 } 876 877 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer, 878 struct buffer_page *bpage) 879 { 880 unsigned long val = (unsigned long)bpage; 881 882 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK)) 883 return 1; 884 885 return 0; 886 } 887 888 /** 889 * rb_check_list - make sure a pointer to a list has the last bits zero 890 */ 891 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer, 892 struct list_head *list) 893 { 894 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev)) 895 return 1; 896 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next)) 897 return 1; 898 return 0; 899 } 900 901 /** 902 * check_pages - integrity check of buffer pages 903 * @cpu_buffer: CPU buffer with pages to test 904 * 905 * As a safety measure we check to make sure the data pages have not 906 * been corrupted. 907 */ 908 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer) 909 { 910 struct list_head *head = cpu_buffer->pages; 911 struct buffer_page *bpage, *tmp; 912 913 rb_head_page_deactivate(cpu_buffer); 914 915 if (RB_WARN_ON(cpu_buffer, head->next->prev != head)) 916 return -1; 917 if (RB_WARN_ON(cpu_buffer, head->prev->next != head)) 918 return -1; 919 920 if (rb_check_list(cpu_buffer, head)) 921 return -1; 922 923 list_for_each_entry_safe(bpage, tmp, head, list) { 924 if (RB_WARN_ON(cpu_buffer, 925 bpage->list.next->prev != &bpage->list)) 926 return -1; 927 if (RB_WARN_ON(cpu_buffer, 928 bpage->list.prev->next != &bpage->list)) 929 return -1; 930 if (rb_check_list(cpu_buffer, &bpage->list)) 931 return -1; 932 } 933 934 rb_head_page_activate(cpu_buffer); 935 936 return 0; 937 } 938 939 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer, 940 unsigned nr_pages) 941 { 942 struct buffer_page *bpage, *tmp; 943 unsigned long addr; 944 LIST_HEAD(pages); 945 unsigned i; 946 947 WARN_ON(!nr_pages); 948 949 for (i = 0; i < nr_pages; i++) { 950 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()), 951 GFP_KERNEL, cpu_to_node(cpu_buffer->cpu)); 952 if (!bpage) 953 goto free_pages; 954 955 rb_check_bpage(cpu_buffer, bpage); 956 957 list_add(&bpage->list, &pages); 958 959 addr = __get_free_page(GFP_KERNEL); 960 if (!addr) 961 goto free_pages; 962 bpage->page = (void *)addr; 963 rb_init_page(bpage->page); 964 } 965 966 /* 967 * The ring buffer page list is a circular list that does not 968 * start and end with a list head. All page list items point to 969 * other pages. 970 */ 971 cpu_buffer->pages = pages.next; 972 list_del(&pages); 973 974 rb_check_pages(cpu_buffer); 975 976 return 0; 977 978 free_pages: 979 list_for_each_entry_safe(bpage, tmp, &pages, list) { 980 list_del_init(&bpage->list); 981 free_buffer_page(bpage); 982 } 983 return -ENOMEM; 984 } 985 986 static struct ring_buffer_per_cpu * 987 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu) 988 { 989 struct ring_buffer_per_cpu *cpu_buffer; 990 struct buffer_page *bpage; 991 unsigned long addr; 992 int ret; 993 994 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()), 995 GFP_KERNEL, cpu_to_node(cpu)); 996 if (!cpu_buffer) 997 return NULL; 998 999 cpu_buffer->cpu = cpu; 1000 cpu_buffer->buffer = buffer; 1001 spin_lock_init(&cpu_buffer->reader_lock); 1002 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key); 1003 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED; 1004 1005 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()), 1006 GFP_KERNEL, cpu_to_node(cpu)); 1007 if (!bpage) 1008 goto fail_free_buffer; 1009 1010 rb_check_bpage(cpu_buffer, bpage); 1011 1012 cpu_buffer->reader_page = bpage; 1013 addr = __get_free_page(GFP_KERNEL); 1014 if (!addr) 1015 goto fail_free_reader; 1016 bpage->page = (void *)addr; 1017 rb_init_page(bpage->page); 1018 1019 INIT_LIST_HEAD(&cpu_buffer->reader_page->list); 1020 1021 ret = rb_allocate_pages(cpu_buffer, buffer->pages); 1022 if (ret < 0) 1023 goto fail_free_reader; 1024 1025 cpu_buffer->head_page 1026 = list_entry(cpu_buffer->pages, struct buffer_page, list); 1027 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page; 1028 1029 rb_head_page_activate(cpu_buffer); 1030 1031 return cpu_buffer; 1032 1033 fail_free_reader: 1034 free_buffer_page(cpu_buffer->reader_page); 1035 1036 fail_free_buffer: 1037 kfree(cpu_buffer); 1038 return NULL; 1039 } 1040 1041 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer) 1042 { 1043 struct list_head *head = cpu_buffer->pages; 1044 struct buffer_page *bpage, *tmp; 1045 1046 free_buffer_page(cpu_buffer->reader_page); 1047 1048 rb_head_page_deactivate(cpu_buffer); 1049 1050 if (head) { 1051 list_for_each_entry_safe(bpage, tmp, head, list) { 1052 list_del_init(&bpage->list); 1053 free_buffer_page(bpage); 1054 } 1055 bpage = list_entry(head, struct buffer_page, list); 1056 free_buffer_page(bpage); 1057 } 1058 1059 kfree(cpu_buffer); 1060 } 1061 1062 #ifdef CONFIG_HOTPLUG_CPU 1063 static int rb_cpu_notify(struct notifier_block *self, 1064 unsigned long action, void *hcpu); 1065 #endif 1066 1067 /** 1068 * ring_buffer_alloc - allocate a new ring_buffer 1069 * @size: the size in bytes per cpu that is needed. 1070 * @flags: attributes to set for the ring buffer. 1071 * 1072 * Currently the only flag that is available is the RB_FL_OVERWRITE 1073 * flag. This flag means that the buffer will overwrite old data 1074 * when the buffer wraps. If this flag is not set, the buffer will 1075 * drop data when the tail hits the head. 1076 */ 1077 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags, 1078 struct lock_class_key *key) 1079 { 1080 struct ring_buffer *buffer; 1081 int bsize; 1082 int cpu; 1083 1084 /* keep it in its own cache line */ 1085 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()), 1086 GFP_KERNEL); 1087 if (!buffer) 1088 return NULL; 1089 1090 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL)) 1091 goto fail_free_buffer; 1092 1093 buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE); 1094 buffer->flags = flags; 1095 buffer->clock = trace_clock_local; 1096 buffer->reader_lock_key = key; 1097 1098 /* need at least two pages */ 1099 if (buffer->pages < 2) 1100 buffer->pages = 2; 1101 1102 /* 1103 * In case of non-hotplug cpu, if the ring-buffer is allocated 1104 * in early initcall, it will not be notified of secondary cpus. 1105 * In that off case, we need to allocate for all possible cpus. 1106 */ 1107 #ifdef CONFIG_HOTPLUG_CPU 1108 get_online_cpus(); 1109 cpumask_copy(buffer->cpumask, cpu_online_mask); 1110 #else 1111 cpumask_copy(buffer->cpumask, cpu_possible_mask); 1112 #endif 1113 buffer->cpus = nr_cpu_ids; 1114 1115 bsize = sizeof(void *) * nr_cpu_ids; 1116 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()), 1117 GFP_KERNEL); 1118 if (!buffer->buffers) 1119 goto fail_free_cpumask; 1120 1121 for_each_buffer_cpu(buffer, cpu) { 1122 buffer->buffers[cpu] = 1123 rb_allocate_cpu_buffer(buffer, cpu); 1124 if (!buffer->buffers[cpu]) 1125 goto fail_free_buffers; 1126 } 1127 1128 #ifdef CONFIG_HOTPLUG_CPU 1129 buffer->cpu_notify.notifier_call = rb_cpu_notify; 1130 buffer->cpu_notify.priority = 0; 1131 register_cpu_notifier(&buffer->cpu_notify); 1132 #endif 1133 1134 put_online_cpus(); 1135 mutex_init(&buffer->mutex); 1136 1137 return buffer; 1138 1139 fail_free_buffers: 1140 for_each_buffer_cpu(buffer, cpu) { 1141 if (buffer->buffers[cpu]) 1142 rb_free_cpu_buffer(buffer->buffers[cpu]); 1143 } 1144 kfree(buffer->buffers); 1145 1146 fail_free_cpumask: 1147 free_cpumask_var(buffer->cpumask); 1148 put_online_cpus(); 1149 1150 fail_free_buffer: 1151 kfree(buffer); 1152 return NULL; 1153 } 1154 EXPORT_SYMBOL_GPL(__ring_buffer_alloc); 1155 1156 /** 1157 * ring_buffer_free - free a ring buffer. 1158 * @buffer: the buffer to free. 1159 */ 1160 void 1161 ring_buffer_free(struct ring_buffer *buffer) 1162 { 1163 int cpu; 1164 1165 get_online_cpus(); 1166 1167 #ifdef CONFIG_HOTPLUG_CPU 1168 unregister_cpu_notifier(&buffer->cpu_notify); 1169 #endif 1170 1171 for_each_buffer_cpu(buffer, cpu) 1172 rb_free_cpu_buffer(buffer->buffers[cpu]); 1173 1174 put_online_cpus(); 1175 1176 kfree(buffer->buffers); 1177 free_cpumask_var(buffer->cpumask); 1178 1179 kfree(buffer); 1180 } 1181 EXPORT_SYMBOL_GPL(ring_buffer_free); 1182 1183 void ring_buffer_set_clock(struct ring_buffer *buffer, 1184 u64 (*clock)(void)) 1185 { 1186 buffer->clock = clock; 1187 } 1188 1189 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer); 1190 1191 static void 1192 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages) 1193 { 1194 struct buffer_page *bpage; 1195 struct list_head *p; 1196 unsigned i; 1197 1198 spin_lock_irq(&cpu_buffer->reader_lock); 1199 rb_head_page_deactivate(cpu_buffer); 1200 1201 for (i = 0; i < nr_pages; i++) { 1202 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages))) 1203 return; 1204 p = cpu_buffer->pages->next; 1205 bpage = list_entry(p, struct buffer_page, list); 1206 list_del_init(&bpage->list); 1207 free_buffer_page(bpage); 1208 } 1209 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages))) 1210 return; 1211 1212 rb_reset_cpu(cpu_buffer); 1213 rb_check_pages(cpu_buffer); 1214 1215 spin_unlock_irq(&cpu_buffer->reader_lock); 1216 } 1217 1218 static void 1219 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer, 1220 struct list_head *pages, unsigned nr_pages) 1221 { 1222 struct buffer_page *bpage; 1223 struct list_head *p; 1224 unsigned i; 1225 1226 spin_lock_irq(&cpu_buffer->reader_lock); 1227 rb_head_page_deactivate(cpu_buffer); 1228 1229 for (i = 0; i < nr_pages; i++) { 1230 if (RB_WARN_ON(cpu_buffer, list_empty(pages))) 1231 return; 1232 p = pages->next; 1233 bpage = list_entry(p, struct buffer_page, list); 1234 list_del_init(&bpage->list); 1235 list_add_tail(&bpage->list, cpu_buffer->pages); 1236 } 1237 rb_reset_cpu(cpu_buffer); 1238 rb_check_pages(cpu_buffer); 1239 1240 spin_unlock_irq(&cpu_buffer->reader_lock); 1241 } 1242 1243 /** 1244 * ring_buffer_resize - resize the ring buffer 1245 * @buffer: the buffer to resize. 1246 * @size: the new size. 1247 * 1248 * Minimum size is 2 * BUF_PAGE_SIZE. 1249 * 1250 * Returns -1 on failure. 1251 */ 1252 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size) 1253 { 1254 struct ring_buffer_per_cpu *cpu_buffer; 1255 unsigned nr_pages, rm_pages, new_pages; 1256 struct buffer_page *bpage, *tmp; 1257 unsigned long buffer_size; 1258 unsigned long addr; 1259 LIST_HEAD(pages); 1260 int i, cpu; 1261 1262 /* 1263 * Always succeed at resizing a non-existent buffer: 1264 */ 1265 if (!buffer) 1266 return size; 1267 1268 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE); 1269 size *= BUF_PAGE_SIZE; 1270 buffer_size = buffer->pages * BUF_PAGE_SIZE; 1271 1272 /* we need a minimum of two pages */ 1273 if (size < BUF_PAGE_SIZE * 2) 1274 size = BUF_PAGE_SIZE * 2; 1275 1276 if (size == buffer_size) 1277 return size; 1278 1279 atomic_inc(&buffer->record_disabled); 1280 1281 /* Make sure all writers are done with this buffer. */ 1282 synchronize_sched(); 1283 1284 mutex_lock(&buffer->mutex); 1285 get_online_cpus(); 1286 1287 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE); 1288 1289 if (size < buffer_size) { 1290 1291 /* easy case, just free pages */ 1292 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages)) 1293 goto out_fail; 1294 1295 rm_pages = buffer->pages - nr_pages; 1296 1297 for_each_buffer_cpu(buffer, cpu) { 1298 cpu_buffer = buffer->buffers[cpu]; 1299 rb_remove_pages(cpu_buffer, rm_pages); 1300 } 1301 goto out; 1302 } 1303 1304 /* 1305 * This is a bit more difficult. We only want to add pages 1306 * when we can allocate enough for all CPUs. We do this 1307 * by allocating all the pages and storing them on a local 1308 * link list. If we succeed in our allocation, then we 1309 * add these pages to the cpu_buffers. Otherwise we just free 1310 * them all and return -ENOMEM; 1311 */ 1312 if (RB_WARN_ON(buffer, nr_pages <= buffer->pages)) 1313 goto out_fail; 1314 1315 new_pages = nr_pages - buffer->pages; 1316 1317 for_each_buffer_cpu(buffer, cpu) { 1318 for (i = 0; i < new_pages; i++) { 1319 bpage = kzalloc_node(ALIGN(sizeof(*bpage), 1320 cache_line_size()), 1321 GFP_KERNEL, cpu_to_node(cpu)); 1322 if (!bpage) 1323 goto free_pages; 1324 list_add(&bpage->list, &pages); 1325 addr = __get_free_page(GFP_KERNEL); 1326 if (!addr) 1327 goto free_pages; 1328 bpage->page = (void *)addr; 1329 rb_init_page(bpage->page); 1330 } 1331 } 1332 1333 for_each_buffer_cpu(buffer, cpu) { 1334 cpu_buffer = buffer->buffers[cpu]; 1335 rb_insert_pages(cpu_buffer, &pages, new_pages); 1336 } 1337 1338 if (RB_WARN_ON(buffer, !list_empty(&pages))) 1339 goto out_fail; 1340 1341 out: 1342 buffer->pages = nr_pages; 1343 put_online_cpus(); 1344 mutex_unlock(&buffer->mutex); 1345 1346 atomic_dec(&buffer->record_disabled); 1347 1348 return size; 1349 1350 free_pages: 1351 list_for_each_entry_safe(bpage, tmp, &pages, list) { 1352 list_del_init(&bpage->list); 1353 free_buffer_page(bpage); 1354 } 1355 put_online_cpus(); 1356 mutex_unlock(&buffer->mutex); 1357 atomic_dec(&buffer->record_disabled); 1358 return -ENOMEM; 1359 1360 /* 1361 * Something went totally wrong, and we are too paranoid 1362 * to even clean up the mess. 1363 */ 1364 out_fail: 1365 put_online_cpus(); 1366 mutex_unlock(&buffer->mutex); 1367 atomic_dec(&buffer->record_disabled); 1368 return -1; 1369 } 1370 EXPORT_SYMBOL_GPL(ring_buffer_resize); 1371 1372 static inline void * 1373 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index) 1374 { 1375 return bpage->data + index; 1376 } 1377 1378 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index) 1379 { 1380 return bpage->page->data + index; 1381 } 1382 1383 static inline struct ring_buffer_event * 1384 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer) 1385 { 1386 return __rb_page_index(cpu_buffer->reader_page, 1387 cpu_buffer->reader_page->read); 1388 } 1389 1390 static inline struct ring_buffer_event * 1391 rb_iter_head_event(struct ring_buffer_iter *iter) 1392 { 1393 return __rb_page_index(iter->head_page, iter->head); 1394 } 1395 1396 static inline unsigned long rb_page_write(struct buffer_page *bpage) 1397 { 1398 return local_read(&bpage->write) & RB_WRITE_MASK; 1399 } 1400 1401 static inline unsigned rb_page_commit(struct buffer_page *bpage) 1402 { 1403 return local_read(&bpage->page->commit); 1404 } 1405 1406 static inline unsigned long rb_page_entries(struct buffer_page *bpage) 1407 { 1408 return local_read(&bpage->entries) & RB_WRITE_MASK; 1409 } 1410 1411 /* Size is determined by what has been commited */ 1412 static inline unsigned rb_page_size(struct buffer_page *bpage) 1413 { 1414 return rb_page_commit(bpage); 1415 } 1416 1417 static inline unsigned 1418 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer) 1419 { 1420 return rb_page_commit(cpu_buffer->commit_page); 1421 } 1422 1423 static inline unsigned 1424 rb_event_index(struct ring_buffer_event *event) 1425 { 1426 unsigned long addr = (unsigned long)event; 1427 1428 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE; 1429 } 1430 1431 static inline int 1432 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer, 1433 struct ring_buffer_event *event) 1434 { 1435 unsigned long addr = (unsigned long)event; 1436 unsigned long index; 1437 1438 index = rb_event_index(event); 1439 addr &= PAGE_MASK; 1440 1441 return cpu_buffer->commit_page->page == (void *)addr && 1442 rb_commit_index(cpu_buffer) == index; 1443 } 1444 1445 static void 1446 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer) 1447 { 1448 unsigned long max_count; 1449 1450 /* 1451 * We only race with interrupts and NMIs on this CPU. 1452 * If we own the commit event, then we can commit 1453 * all others that interrupted us, since the interruptions 1454 * are in stack format (they finish before they come 1455 * back to us). This allows us to do a simple loop to 1456 * assign the commit to the tail. 1457 */ 1458 again: 1459 max_count = cpu_buffer->buffer->pages * 100; 1460 1461 while (cpu_buffer->commit_page != cpu_buffer->tail_page) { 1462 if (RB_WARN_ON(cpu_buffer, !(--max_count))) 1463 return; 1464 if (RB_WARN_ON(cpu_buffer, 1465 rb_is_reader_page(cpu_buffer->tail_page))) 1466 return; 1467 local_set(&cpu_buffer->commit_page->page->commit, 1468 rb_page_write(cpu_buffer->commit_page)); 1469 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page); 1470 cpu_buffer->write_stamp = 1471 cpu_buffer->commit_page->page->time_stamp; 1472 /* add barrier to keep gcc from optimizing too much */ 1473 barrier(); 1474 } 1475 while (rb_commit_index(cpu_buffer) != 1476 rb_page_write(cpu_buffer->commit_page)) { 1477 1478 local_set(&cpu_buffer->commit_page->page->commit, 1479 rb_page_write(cpu_buffer->commit_page)); 1480 RB_WARN_ON(cpu_buffer, 1481 local_read(&cpu_buffer->commit_page->page->commit) & 1482 ~RB_WRITE_MASK); 1483 barrier(); 1484 } 1485 1486 /* again, keep gcc from optimizing */ 1487 barrier(); 1488 1489 /* 1490 * If an interrupt came in just after the first while loop 1491 * and pushed the tail page forward, we will be left with 1492 * a dangling commit that will never go forward. 1493 */ 1494 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page)) 1495 goto again; 1496 } 1497 1498 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer) 1499 { 1500 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp; 1501 cpu_buffer->reader_page->read = 0; 1502 } 1503 1504 static void rb_inc_iter(struct ring_buffer_iter *iter) 1505 { 1506 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; 1507 1508 /* 1509 * The iterator could be on the reader page (it starts there). 1510 * But the head could have moved, since the reader was 1511 * found. Check for this case and assign the iterator 1512 * to the head page instead of next. 1513 */ 1514 if (iter->head_page == cpu_buffer->reader_page) 1515 iter->head_page = rb_set_head_page(cpu_buffer); 1516 else 1517 rb_inc_page(cpu_buffer, &iter->head_page); 1518 1519 iter->read_stamp = iter->head_page->page->time_stamp; 1520 iter->head = 0; 1521 } 1522 1523 /** 1524 * ring_buffer_update_event - update event type and data 1525 * @event: the even to update 1526 * @type: the type of event 1527 * @length: the size of the event field in the ring buffer 1528 * 1529 * Update the type and data fields of the event. The length 1530 * is the actual size that is written to the ring buffer, 1531 * and with this, we can determine what to place into the 1532 * data field. 1533 */ 1534 static void 1535 rb_update_event(struct ring_buffer_event *event, 1536 unsigned type, unsigned length) 1537 { 1538 event->type_len = type; 1539 1540 switch (type) { 1541 1542 case RINGBUF_TYPE_PADDING: 1543 case RINGBUF_TYPE_TIME_EXTEND: 1544 case RINGBUF_TYPE_TIME_STAMP: 1545 break; 1546 1547 case 0: 1548 length -= RB_EVNT_HDR_SIZE; 1549 if (length > RB_MAX_SMALL_DATA) 1550 event->array[0] = length; 1551 else 1552 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT); 1553 break; 1554 default: 1555 BUG(); 1556 } 1557 } 1558 1559 /* 1560 * rb_handle_head_page - writer hit the head page 1561 * 1562 * Returns: +1 to retry page 1563 * 0 to continue 1564 * -1 on error 1565 */ 1566 static int 1567 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer, 1568 struct buffer_page *tail_page, 1569 struct buffer_page *next_page) 1570 { 1571 struct buffer_page *new_head; 1572 int entries; 1573 int type; 1574 int ret; 1575 1576 entries = rb_page_entries(next_page); 1577 1578 /* 1579 * The hard part is here. We need to move the head 1580 * forward, and protect against both readers on 1581 * other CPUs and writers coming in via interrupts. 1582 */ 1583 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page, 1584 RB_PAGE_HEAD); 1585 1586 /* 1587 * type can be one of four: 1588 * NORMAL - an interrupt already moved it for us 1589 * HEAD - we are the first to get here. 1590 * UPDATE - we are the interrupt interrupting 1591 * a current move. 1592 * MOVED - a reader on another CPU moved the next 1593 * pointer to its reader page. Give up 1594 * and try again. 1595 */ 1596 1597 switch (type) { 1598 case RB_PAGE_HEAD: 1599 /* 1600 * We changed the head to UPDATE, thus 1601 * it is our responsibility to update 1602 * the counters. 1603 */ 1604 local_add(entries, &cpu_buffer->overrun); 1605 1606 /* 1607 * The entries will be zeroed out when we move the 1608 * tail page. 1609 */ 1610 1611 /* still more to do */ 1612 break; 1613 1614 case RB_PAGE_UPDATE: 1615 /* 1616 * This is an interrupt that interrupt the 1617 * previous update. Still more to do. 1618 */ 1619 break; 1620 case RB_PAGE_NORMAL: 1621 /* 1622 * An interrupt came in before the update 1623 * and processed this for us. 1624 * Nothing left to do. 1625 */ 1626 return 1; 1627 case RB_PAGE_MOVED: 1628 /* 1629 * The reader is on another CPU and just did 1630 * a swap with our next_page. 1631 * Try again. 1632 */ 1633 return 1; 1634 default: 1635 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */ 1636 return -1; 1637 } 1638 1639 /* 1640 * Now that we are here, the old head pointer is 1641 * set to UPDATE. This will keep the reader from 1642 * swapping the head page with the reader page. 1643 * The reader (on another CPU) will spin till 1644 * we are finished. 1645 * 1646 * We just need to protect against interrupts 1647 * doing the job. We will set the next pointer 1648 * to HEAD. After that, we set the old pointer 1649 * to NORMAL, but only if it was HEAD before. 1650 * otherwise we are an interrupt, and only 1651 * want the outer most commit to reset it. 1652 */ 1653 new_head = next_page; 1654 rb_inc_page(cpu_buffer, &new_head); 1655 1656 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page, 1657 RB_PAGE_NORMAL); 1658 1659 /* 1660 * Valid returns are: 1661 * HEAD - an interrupt came in and already set it. 1662 * NORMAL - One of two things: 1663 * 1) We really set it. 1664 * 2) A bunch of interrupts came in and moved 1665 * the page forward again. 1666 */ 1667 switch (ret) { 1668 case RB_PAGE_HEAD: 1669 case RB_PAGE_NORMAL: 1670 /* OK */ 1671 break; 1672 default: 1673 RB_WARN_ON(cpu_buffer, 1); 1674 return -1; 1675 } 1676 1677 /* 1678 * It is possible that an interrupt came in, 1679 * set the head up, then more interrupts came in 1680 * and moved it again. When we get back here, 1681 * the page would have been set to NORMAL but we 1682 * just set it back to HEAD. 1683 * 1684 * How do you detect this? Well, if that happened 1685 * the tail page would have moved. 1686 */ 1687 if (ret == RB_PAGE_NORMAL) { 1688 /* 1689 * If the tail had moved passed next, then we need 1690 * to reset the pointer. 1691 */ 1692 if (cpu_buffer->tail_page != tail_page && 1693 cpu_buffer->tail_page != next_page) 1694 rb_head_page_set_normal(cpu_buffer, new_head, 1695 next_page, 1696 RB_PAGE_HEAD); 1697 } 1698 1699 /* 1700 * If this was the outer most commit (the one that 1701 * changed the original pointer from HEAD to UPDATE), 1702 * then it is up to us to reset it to NORMAL. 1703 */ 1704 if (type == RB_PAGE_HEAD) { 1705 ret = rb_head_page_set_normal(cpu_buffer, next_page, 1706 tail_page, 1707 RB_PAGE_UPDATE); 1708 if (RB_WARN_ON(cpu_buffer, 1709 ret != RB_PAGE_UPDATE)) 1710 return -1; 1711 } 1712 1713 return 0; 1714 } 1715 1716 static unsigned rb_calculate_event_length(unsigned length) 1717 { 1718 struct ring_buffer_event event; /* Used only for sizeof array */ 1719 1720 /* zero length can cause confusions */ 1721 if (!length) 1722 length = 1; 1723 1724 if (length > RB_MAX_SMALL_DATA) 1725 length += sizeof(event.array[0]); 1726 1727 length += RB_EVNT_HDR_SIZE; 1728 length = ALIGN(length, RB_ALIGNMENT); 1729 1730 return length; 1731 } 1732 1733 static inline void 1734 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer, 1735 struct buffer_page *tail_page, 1736 unsigned long tail, unsigned long length) 1737 { 1738 struct ring_buffer_event *event; 1739 1740 /* 1741 * Only the event that crossed the page boundary 1742 * must fill the old tail_page with padding. 1743 */ 1744 if (tail >= BUF_PAGE_SIZE) { 1745 local_sub(length, &tail_page->write); 1746 return; 1747 } 1748 1749 event = __rb_page_index(tail_page, tail); 1750 kmemcheck_annotate_bitfield(event, bitfield); 1751 1752 /* 1753 * If this event is bigger than the minimum size, then 1754 * we need to be careful that we don't subtract the 1755 * write counter enough to allow another writer to slip 1756 * in on this page. 1757 * We put in a discarded commit instead, to make sure 1758 * that this space is not used again. 1759 * 1760 * If we are less than the minimum size, we don't need to 1761 * worry about it. 1762 */ 1763 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) { 1764 /* No room for any events */ 1765 1766 /* Mark the rest of the page with padding */ 1767 rb_event_set_padding(event); 1768 1769 /* Set the write back to the previous setting */ 1770 local_sub(length, &tail_page->write); 1771 return; 1772 } 1773 1774 /* Put in a discarded event */ 1775 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE; 1776 event->type_len = RINGBUF_TYPE_PADDING; 1777 /* time delta must be non zero */ 1778 event->time_delta = 1; 1779 1780 /* Set write to end of buffer */ 1781 length = (tail + length) - BUF_PAGE_SIZE; 1782 local_sub(length, &tail_page->write); 1783 } 1784 1785 static struct ring_buffer_event * 1786 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer, 1787 unsigned long length, unsigned long tail, 1788 struct buffer_page *tail_page, u64 *ts) 1789 { 1790 struct buffer_page *commit_page = cpu_buffer->commit_page; 1791 struct ring_buffer *buffer = cpu_buffer->buffer; 1792 struct buffer_page *next_page; 1793 int ret; 1794 1795 next_page = tail_page; 1796 1797 rb_inc_page(cpu_buffer, &next_page); 1798 1799 /* 1800 * If for some reason, we had an interrupt storm that made 1801 * it all the way around the buffer, bail, and warn 1802 * about it. 1803 */ 1804 if (unlikely(next_page == commit_page)) { 1805 local_inc(&cpu_buffer->commit_overrun); 1806 goto out_reset; 1807 } 1808 1809 /* 1810 * This is where the fun begins! 1811 * 1812 * We are fighting against races between a reader that 1813 * could be on another CPU trying to swap its reader 1814 * page with the buffer head. 1815 * 1816 * We are also fighting against interrupts coming in and 1817 * moving the head or tail on us as well. 1818 * 1819 * If the next page is the head page then we have filled 1820 * the buffer, unless the commit page is still on the 1821 * reader page. 1822 */ 1823 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) { 1824 1825 /* 1826 * If the commit is not on the reader page, then 1827 * move the header page. 1828 */ 1829 if (!rb_is_reader_page(cpu_buffer->commit_page)) { 1830 /* 1831 * If we are not in overwrite mode, 1832 * this is easy, just stop here. 1833 */ 1834 if (!(buffer->flags & RB_FL_OVERWRITE)) 1835 goto out_reset; 1836 1837 ret = rb_handle_head_page(cpu_buffer, 1838 tail_page, 1839 next_page); 1840 if (ret < 0) 1841 goto out_reset; 1842 if (ret) 1843 goto out_again; 1844 } else { 1845 /* 1846 * We need to be careful here too. The 1847 * commit page could still be on the reader 1848 * page. We could have a small buffer, and 1849 * have filled up the buffer with events 1850 * from interrupts and such, and wrapped. 1851 * 1852 * Note, if the tail page is also the on the 1853 * reader_page, we let it move out. 1854 */ 1855 if (unlikely((cpu_buffer->commit_page != 1856 cpu_buffer->tail_page) && 1857 (cpu_buffer->commit_page == 1858 cpu_buffer->reader_page))) { 1859 local_inc(&cpu_buffer->commit_overrun); 1860 goto out_reset; 1861 } 1862 } 1863 } 1864 1865 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page); 1866 if (ret) { 1867 /* 1868 * Nested commits always have zero deltas, so 1869 * just reread the time stamp 1870 */ 1871 *ts = rb_time_stamp(buffer); 1872 next_page->page->time_stamp = *ts; 1873 } 1874 1875 out_again: 1876 1877 rb_reset_tail(cpu_buffer, tail_page, tail, length); 1878 1879 /* fail and let the caller try again */ 1880 return ERR_PTR(-EAGAIN); 1881 1882 out_reset: 1883 /* reset write */ 1884 rb_reset_tail(cpu_buffer, tail_page, tail, length); 1885 1886 return NULL; 1887 } 1888 1889 static struct ring_buffer_event * 1890 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer, 1891 unsigned type, unsigned long length, u64 *ts) 1892 { 1893 struct buffer_page *tail_page; 1894 struct ring_buffer_event *event; 1895 unsigned long tail, write; 1896 1897 tail_page = cpu_buffer->tail_page; 1898 write = local_add_return(length, &tail_page->write); 1899 1900 /* set write to only the index of the write */ 1901 write &= RB_WRITE_MASK; 1902 tail = write - length; 1903 1904 /* See if we shot pass the end of this buffer page */ 1905 if (write > BUF_PAGE_SIZE) 1906 return rb_move_tail(cpu_buffer, length, tail, 1907 tail_page, ts); 1908 1909 /* We reserved something on the buffer */ 1910 1911 event = __rb_page_index(tail_page, tail); 1912 kmemcheck_annotate_bitfield(event, bitfield); 1913 rb_update_event(event, type, length); 1914 1915 /* The passed in type is zero for DATA */ 1916 if (likely(!type)) 1917 local_inc(&tail_page->entries); 1918 1919 /* 1920 * If this is the first commit on the page, then update 1921 * its timestamp. 1922 */ 1923 if (!tail) 1924 tail_page->page->time_stamp = *ts; 1925 1926 return event; 1927 } 1928 1929 static inline int 1930 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer, 1931 struct ring_buffer_event *event) 1932 { 1933 unsigned long new_index, old_index; 1934 struct buffer_page *bpage; 1935 unsigned long index; 1936 unsigned long addr; 1937 1938 new_index = rb_event_index(event); 1939 old_index = new_index + rb_event_length(event); 1940 addr = (unsigned long)event; 1941 addr &= PAGE_MASK; 1942 1943 bpage = cpu_buffer->tail_page; 1944 1945 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) { 1946 unsigned long write_mask = 1947 local_read(&bpage->write) & ~RB_WRITE_MASK; 1948 /* 1949 * This is on the tail page. It is possible that 1950 * a write could come in and move the tail page 1951 * and write to the next page. That is fine 1952 * because we just shorten what is on this page. 1953 */ 1954 old_index += write_mask; 1955 new_index += write_mask; 1956 index = local_cmpxchg(&bpage->write, old_index, new_index); 1957 if (index == old_index) 1958 return 1; 1959 } 1960 1961 /* could not discard */ 1962 return 0; 1963 } 1964 1965 static int 1966 rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer, 1967 u64 *ts, u64 *delta) 1968 { 1969 struct ring_buffer_event *event; 1970 static int once; 1971 int ret; 1972 1973 if (unlikely(*delta > (1ULL << 59) && !once++)) { 1974 printk(KERN_WARNING "Delta way too big! %llu" 1975 " ts=%llu write stamp = %llu\n", 1976 (unsigned long long)*delta, 1977 (unsigned long long)*ts, 1978 (unsigned long long)cpu_buffer->write_stamp); 1979 WARN_ON(1); 1980 } 1981 1982 /* 1983 * The delta is too big, we to add a 1984 * new timestamp. 1985 */ 1986 event = __rb_reserve_next(cpu_buffer, 1987 RINGBUF_TYPE_TIME_EXTEND, 1988 RB_LEN_TIME_EXTEND, 1989 ts); 1990 if (!event) 1991 return -EBUSY; 1992 1993 if (PTR_ERR(event) == -EAGAIN) 1994 return -EAGAIN; 1995 1996 /* Only a commited time event can update the write stamp */ 1997 if (rb_event_is_commit(cpu_buffer, event)) { 1998 /* 1999 * If this is the first on the page, then it was 2000 * updated with the page itself. Try to discard it 2001 * and if we can't just make it zero. 2002 */ 2003 if (rb_event_index(event)) { 2004 event->time_delta = *delta & TS_MASK; 2005 event->array[0] = *delta >> TS_SHIFT; 2006 } else { 2007 /* try to discard, since we do not need this */ 2008 if (!rb_try_to_discard(cpu_buffer, event)) { 2009 /* nope, just zero it */ 2010 event->time_delta = 0; 2011 event->array[0] = 0; 2012 } 2013 } 2014 cpu_buffer->write_stamp = *ts; 2015 /* let the caller know this was the commit */ 2016 ret = 1; 2017 } else { 2018 /* Try to discard the event */ 2019 if (!rb_try_to_discard(cpu_buffer, event)) { 2020 /* Darn, this is just wasted space */ 2021 event->time_delta = 0; 2022 event->array[0] = 0; 2023 } 2024 ret = 0; 2025 } 2026 2027 *delta = 0; 2028 2029 return ret; 2030 } 2031 2032 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer) 2033 { 2034 local_inc(&cpu_buffer->committing); 2035 local_inc(&cpu_buffer->commits); 2036 } 2037 2038 static void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer) 2039 { 2040 unsigned long commits; 2041 2042 if (RB_WARN_ON(cpu_buffer, 2043 !local_read(&cpu_buffer->committing))) 2044 return; 2045 2046 again: 2047 commits = local_read(&cpu_buffer->commits); 2048 /* synchronize with interrupts */ 2049 barrier(); 2050 if (local_read(&cpu_buffer->committing) == 1) 2051 rb_set_commit_to_write(cpu_buffer); 2052 2053 local_dec(&cpu_buffer->committing); 2054 2055 /* synchronize with interrupts */ 2056 barrier(); 2057 2058 /* 2059 * Need to account for interrupts coming in between the 2060 * updating of the commit page and the clearing of the 2061 * committing counter. 2062 */ 2063 if (unlikely(local_read(&cpu_buffer->commits) != commits) && 2064 !local_read(&cpu_buffer->committing)) { 2065 local_inc(&cpu_buffer->committing); 2066 goto again; 2067 } 2068 } 2069 2070 static struct ring_buffer_event * 2071 rb_reserve_next_event(struct ring_buffer *buffer, 2072 struct ring_buffer_per_cpu *cpu_buffer, 2073 unsigned long length) 2074 { 2075 struct ring_buffer_event *event; 2076 u64 ts, delta = 0; 2077 int commit = 0; 2078 int nr_loops = 0; 2079 2080 rb_start_commit(cpu_buffer); 2081 2082 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP 2083 /* 2084 * Due to the ability to swap a cpu buffer from a buffer 2085 * it is possible it was swapped before we committed. 2086 * (committing stops a swap). We check for it here and 2087 * if it happened, we have to fail the write. 2088 */ 2089 barrier(); 2090 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) { 2091 local_dec(&cpu_buffer->committing); 2092 local_dec(&cpu_buffer->commits); 2093 return NULL; 2094 } 2095 #endif 2096 2097 length = rb_calculate_event_length(length); 2098 again: 2099 /* 2100 * We allow for interrupts to reenter here and do a trace. 2101 * If one does, it will cause this original code to loop 2102 * back here. Even with heavy interrupts happening, this 2103 * should only happen a few times in a row. If this happens 2104 * 1000 times in a row, there must be either an interrupt 2105 * storm or we have something buggy. 2106 * Bail! 2107 */ 2108 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000)) 2109 goto out_fail; 2110 2111 ts = rb_time_stamp(cpu_buffer->buffer); 2112 2113 /* 2114 * Only the first commit can update the timestamp. 2115 * Yes there is a race here. If an interrupt comes in 2116 * just after the conditional and it traces too, then it 2117 * will also check the deltas. More than one timestamp may 2118 * also be made. But only the entry that did the actual 2119 * commit will be something other than zero. 2120 */ 2121 if (likely(cpu_buffer->tail_page == cpu_buffer->commit_page && 2122 rb_page_write(cpu_buffer->tail_page) == 2123 rb_commit_index(cpu_buffer))) { 2124 u64 diff; 2125 2126 diff = ts - cpu_buffer->write_stamp; 2127 2128 /* make sure this diff is calculated here */ 2129 barrier(); 2130 2131 /* Did the write stamp get updated already? */ 2132 if (unlikely(ts < cpu_buffer->write_stamp)) 2133 goto get_event; 2134 2135 delta = diff; 2136 if (unlikely(test_time_stamp(delta))) { 2137 2138 commit = rb_add_time_stamp(cpu_buffer, &ts, &delta); 2139 if (commit == -EBUSY) 2140 goto out_fail; 2141 2142 if (commit == -EAGAIN) 2143 goto again; 2144 2145 RB_WARN_ON(cpu_buffer, commit < 0); 2146 } 2147 } 2148 2149 get_event: 2150 event = __rb_reserve_next(cpu_buffer, 0, length, &ts); 2151 if (unlikely(PTR_ERR(event) == -EAGAIN)) 2152 goto again; 2153 2154 if (!event) 2155 goto out_fail; 2156 2157 if (!rb_event_is_commit(cpu_buffer, event)) 2158 delta = 0; 2159 2160 event->time_delta = delta; 2161 2162 return event; 2163 2164 out_fail: 2165 rb_end_commit(cpu_buffer); 2166 return NULL; 2167 } 2168 2169 #ifdef CONFIG_TRACING 2170 2171 #define TRACE_RECURSIVE_DEPTH 16 2172 2173 static int trace_recursive_lock(void) 2174 { 2175 current->trace_recursion++; 2176 2177 if (likely(current->trace_recursion < TRACE_RECURSIVE_DEPTH)) 2178 return 0; 2179 2180 /* Disable all tracing before we do anything else */ 2181 tracing_off_permanent(); 2182 2183 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:" 2184 "HC[%lu]:SC[%lu]:NMI[%lu]\n", 2185 current->trace_recursion, 2186 hardirq_count() >> HARDIRQ_SHIFT, 2187 softirq_count() >> SOFTIRQ_SHIFT, 2188 in_nmi()); 2189 2190 WARN_ON_ONCE(1); 2191 return -1; 2192 } 2193 2194 static void trace_recursive_unlock(void) 2195 { 2196 WARN_ON_ONCE(!current->trace_recursion); 2197 2198 current->trace_recursion--; 2199 } 2200 2201 #else 2202 2203 #define trace_recursive_lock() (0) 2204 #define trace_recursive_unlock() do { } while (0) 2205 2206 #endif 2207 2208 static DEFINE_PER_CPU(int, rb_need_resched); 2209 2210 /** 2211 * ring_buffer_lock_reserve - reserve a part of the buffer 2212 * @buffer: the ring buffer to reserve from 2213 * @length: the length of the data to reserve (excluding event header) 2214 * 2215 * Returns a reseverd event on the ring buffer to copy directly to. 2216 * The user of this interface will need to get the body to write into 2217 * and can use the ring_buffer_event_data() interface. 2218 * 2219 * The length is the length of the data needed, not the event length 2220 * which also includes the event header. 2221 * 2222 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned. 2223 * If NULL is returned, then nothing has been allocated or locked. 2224 */ 2225 struct ring_buffer_event * 2226 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length) 2227 { 2228 struct ring_buffer_per_cpu *cpu_buffer; 2229 struct ring_buffer_event *event; 2230 int cpu, resched; 2231 2232 if (ring_buffer_flags != RB_BUFFERS_ON) 2233 return NULL; 2234 2235 if (atomic_read(&buffer->record_disabled)) 2236 return NULL; 2237 2238 /* If we are tracing schedule, we don't want to recurse */ 2239 resched = ftrace_preempt_disable(); 2240 2241 if (trace_recursive_lock()) 2242 goto out_nocheck; 2243 2244 cpu = raw_smp_processor_id(); 2245 2246 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 2247 goto out; 2248 2249 cpu_buffer = buffer->buffers[cpu]; 2250 2251 if (atomic_read(&cpu_buffer->record_disabled)) 2252 goto out; 2253 2254 if (length > BUF_MAX_DATA_SIZE) 2255 goto out; 2256 2257 event = rb_reserve_next_event(buffer, cpu_buffer, length); 2258 if (!event) 2259 goto out; 2260 2261 /* 2262 * Need to store resched state on this cpu. 2263 * Only the first needs to. 2264 */ 2265 2266 if (preempt_count() == 1) 2267 per_cpu(rb_need_resched, cpu) = resched; 2268 2269 return event; 2270 2271 out: 2272 trace_recursive_unlock(); 2273 2274 out_nocheck: 2275 ftrace_preempt_enable(resched); 2276 return NULL; 2277 } 2278 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve); 2279 2280 static void 2281 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer, 2282 struct ring_buffer_event *event) 2283 { 2284 /* 2285 * The event first in the commit queue updates the 2286 * time stamp. 2287 */ 2288 if (rb_event_is_commit(cpu_buffer, event)) 2289 cpu_buffer->write_stamp += event->time_delta; 2290 } 2291 2292 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer, 2293 struct ring_buffer_event *event) 2294 { 2295 local_inc(&cpu_buffer->entries); 2296 rb_update_write_stamp(cpu_buffer, event); 2297 rb_end_commit(cpu_buffer); 2298 } 2299 2300 /** 2301 * ring_buffer_unlock_commit - commit a reserved 2302 * @buffer: The buffer to commit to 2303 * @event: The event pointer to commit. 2304 * 2305 * This commits the data to the ring buffer, and releases any locks held. 2306 * 2307 * Must be paired with ring_buffer_lock_reserve. 2308 */ 2309 int ring_buffer_unlock_commit(struct ring_buffer *buffer, 2310 struct ring_buffer_event *event) 2311 { 2312 struct ring_buffer_per_cpu *cpu_buffer; 2313 int cpu = raw_smp_processor_id(); 2314 2315 cpu_buffer = buffer->buffers[cpu]; 2316 2317 rb_commit(cpu_buffer, event); 2318 2319 trace_recursive_unlock(); 2320 2321 /* 2322 * Only the last preempt count needs to restore preemption. 2323 */ 2324 if (preempt_count() == 1) 2325 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu)); 2326 else 2327 preempt_enable_no_resched_notrace(); 2328 2329 return 0; 2330 } 2331 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit); 2332 2333 static inline void rb_event_discard(struct ring_buffer_event *event) 2334 { 2335 /* array[0] holds the actual length for the discarded event */ 2336 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE; 2337 event->type_len = RINGBUF_TYPE_PADDING; 2338 /* time delta must be non zero */ 2339 if (!event->time_delta) 2340 event->time_delta = 1; 2341 } 2342 2343 /* 2344 * Decrement the entries to the page that an event is on. 2345 * The event does not even need to exist, only the pointer 2346 * to the page it is on. This may only be called before the commit 2347 * takes place. 2348 */ 2349 static inline void 2350 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer, 2351 struct ring_buffer_event *event) 2352 { 2353 unsigned long addr = (unsigned long)event; 2354 struct buffer_page *bpage = cpu_buffer->commit_page; 2355 struct buffer_page *start; 2356 2357 addr &= PAGE_MASK; 2358 2359 /* Do the likely case first */ 2360 if (likely(bpage->page == (void *)addr)) { 2361 local_dec(&bpage->entries); 2362 return; 2363 } 2364 2365 /* 2366 * Because the commit page may be on the reader page we 2367 * start with the next page and check the end loop there. 2368 */ 2369 rb_inc_page(cpu_buffer, &bpage); 2370 start = bpage; 2371 do { 2372 if (bpage->page == (void *)addr) { 2373 local_dec(&bpage->entries); 2374 return; 2375 } 2376 rb_inc_page(cpu_buffer, &bpage); 2377 } while (bpage != start); 2378 2379 /* commit not part of this buffer?? */ 2380 RB_WARN_ON(cpu_buffer, 1); 2381 } 2382 2383 /** 2384 * ring_buffer_commit_discard - discard an event that has not been committed 2385 * @buffer: the ring buffer 2386 * @event: non committed event to discard 2387 * 2388 * Sometimes an event that is in the ring buffer needs to be ignored. 2389 * This function lets the user discard an event in the ring buffer 2390 * and then that event will not be read later. 2391 * 2392 * This function only works if it is called before the the item has been 2393 * committed. It will try to free the event from the ring buffer 2394 * if another event has not been added behind it. 2395 * 2396 * If another event has been added behind it, it will set the event 2397 * up as discarded, and perform the commit. 2398 * 2399 * If this function is called, do not call ring_buffer_unlock_commit on 2400 * the event. 2401 */ 2402 void ring_buffer_discard_commit(struct ring_buffer *buffer, 2403 struct ring_buffer_event *event) 2404 { 2405 struct ring_buffer_per_cpu *cpu_buffer; 2406 int cpu; 2407 2408 /* The event is discarded regardless */ 2409 rb_event_discard(event); 2410 2411 cpu = smp_processor_id(); 2412 cpu_buffer = buffer->buffers[cpu]; 2413 2414 /* 2415 * This must only be called if the event has not been 2416 * committed yet. Thus we can assume that preemption 2417 * is still disabled. 2418 */ 2419 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing)); 2420 2421 rb_decrement_entry(cpu_buffer, event); 2422 if (rb_try_to_discard(cpu_buffer, event)) 2423 goto out; 2424 2425 /* 2426 * The commit is still visible by the reader, so we 2427 * must still update the timestamp. 2428 */ 2429 rb_update_write_stamp(cpu_buffer, event); 2430 out: 2431 rb_end_commit(cpu_buffer); 2432 2433 trace_recursive_unlock(); 2434 2435 /* 2436 * Only the last preempt count needs to restore preemption. 2437 */ 2438 if (preempt_count() == 1) 2439 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu)); 2440 else 2441 preempt_enable_no_resched_notrace(); 2442 2443 } 2444 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit); 2445 2446 /** 2447 * ring_buffer_write - write data to the buffer without reserving 2448 * @buffer: The ring buffer to write to. 2449 * @length: The length of the data being written (excluding the event header) 2450 * @data: The data to write to the buffer. 2451 * 2452 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as 2453 * one function. If you already have the data to write to the buffer, it 2454 * may be easier to simply call this function. 2455 * 2456 * Note, like ring_buffer_lock_reserve, the length is the length of the data 2457 * and not the length of the event which would hold the header. 2458 */ 2459 int ring_buffer_write(struct ring_buffer *buffer, 2460 unsigned long length, 2461 void *data) 2462 { 2463 struct ring_buffer_per_cpu *cpu_buffer; 2464 struct ring_buffer_event *event; 2465 void *body; 2466 int ret = -EBUSY; 2467 int cpu, resched; 2468 2469 if (ring_buffer_flags != RB_BUFFERS_ON) 2470 return -EBUSY; 2471 2472 if (atomic_read(&buffer->record_disabled)) 2473 return -EBUSY; 2474 2475 resched = ftrace_preempt_disable(); 2476 2477 cpu = raw_smp_processor_id(); 2478 2479 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 2480 goto out; 2481 2482 cpu_buffer = buffer->buffers[cpu]; 2483 2484 if (atomic_read(&cpu_buffer->record_disabled)) 2485 goto out; 2486 2487 if (length > BUF_MAX_DATA_SIZE) 2488 goto out; 2489 2490 event = rb_reserve_next_event(buffer, cpu_buffer, length); 2491 if (!event) 2492 goto out; 2493 2494 body = rb_event_data(event); 2495 2496 memcpy(body, data, length); 2497 2498 rb_commit(cpu_buffer, event); 2499 2500 ret = 0; 2501 out: 2502 ftrace_preempt_enable(resched); 2503 2504 return ret; 2505 } 2506 EXPORT_SYMBOL_GPL(ring_buffer_write); 2507 2508 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer) 2509 { 2510 struct buffer_page *reader = cpu_buffer->reader_page; 2511 struct buffer_page *head = rb_set_head_page(cpu_buffer); 2512 struct buffer_page *commit = cpu_buffer->commit_page; 2513 2514 /* In case of error, head will be NULL */ 2515 if (unlikely(!head)) 2516 return 1; 2517 2518 return reader->read == rb_page_commit(reader) && 2519 (commit == reader || 2520 (commit == head && 2521 head->read == rb_page_commit(commit))); 2522 } 2523 2524 /** 2525 * ring_buffer_record_disable - stop all writes into the buffer 2526 * @buffer: The ring buffer to stop writes to. 2527 * 2528 * This prevents all writes to the buffer. Any attempt to write 2529 * to the buffer after this will fail and return NULL. 2530 * 2531 * The caller should call synchronize_sched() after this. 2532 */ 2533 void ring_buffer_record_disable(struct ring_buffer *buffer) 2534 { 2535 atomic_inc(&buffer->record_disabled); 2536 } 2537 EXPORT_SYMBOL_GPL(ring_buffer_record_disable); 2538 2539 /** 2540 * ring_buffer_record_enable - enable writes to the buffer 2541 * @buffer: The ring buffer to enable writes 2542 * 2543 * Note, multiple disables will need the same number of enables 2544 * to truely enable the writing (much like preempt_disable). 2545 */ 2546 void ring_buffer_record_enable(struct ring_buffer *buffer) 2547 { 2548 atomic_dec(&buffer->record_disabled); 2549 } 2550 EXPORT_SYMBOL_GPL(ring_buffer_record_enable); 2551 2552 /** 2553 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer 2554 * @buffer: The ring buffer to stop writes to. 2555 * @cpu: The CPU buffer to stop 2556 * 2557 * This prevents all writes to the buffer. Any attempt to write 2558 * to the buffer after this will fail and return NULL. 2559 * 2560 * The caller should call synchronize_sched() after this. 2561 */ 2562 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu) 2563 { 2564 struct ring_buffer_per_cpu *cpu_buffer; 2565 2566 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 2567 return; 2568 2569 cpu_buffer = buffer->buffers[cpu]; 2570 atomic_inc(&cpu_buffer->record_disabled); 2571 } 2572 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu); 2573 2574 /** 2575 * ring_buffer_record_enable_cpu - enable writes to the buffer 2576 * @buffer: The ring buffer to enable writes 2577 * @cpu: The CPU to enable. 2578 * 2579 * Note, multiple disables will need the same number of enables 2580 * to truely enable the writing (much like preempt_disable). 2581 */ 2582 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu) 2583 { 2584 struct ring_buffer_per_cpu *cpu_buffer; 2585 2586 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 2587 return; 2588 2589 cpu_buffer = buffer->buffers[cpu]; 2590 atomic_dec(&cpu_buffer->record_disabled); 2591 } 2592 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu); 2593 2594 /** 2595 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer 2596 * @buffer: The ring buffer 2597 * @cpu: The per CPU buffer to get the entries from. 2598 */ 2599 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu) 2600 { 2601 struct ring_buffer_per_cpu *cpu_buffer; 2602 unsigned long ret; 2603 2604 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 2605 return 0; 2606 2607 cpu_buffer = buffer->buffers[cpu]; 2608 ret = (local_read(&cpu_buffer->entries) - local_read(&cpu_buffer->overrun)) 2609 - cpu_buffer->read; 2610 2611 return ret; 2612 } 2613 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu); 2614 2615 /** 2616 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer 2617 * @buffer: The ring buffer 2618 * @cpu: The per CPU buffer to get the number of overruns from 2619 */ 2620 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu) 2621 { 2622 struct ring_buffer_per_cpu *cpu_buffer; 2623 unsigned long ret; 2624 2625 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 2626 return 0; 2627 2628 cpu_buffer = buffer->buffers[cpu]; 2629 ret = local_read(&cpu_buffer->overrun); 2630 2631 return ret; 2632 } 2633 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu); 2634 2635 /** 2636 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits 2637 * @buffer: The ring buffer 2638 * @cpu: The per CPU buffer to get the number of overruns from 2639 */ 2640 unsigned long 2641 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu) 2642 { 2643 struct ring_buffer_per_cpu *cpu_buffer; 2644 unsigned long ret; 2645 2646 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 2647 return 0; 2648 2649 cpu_buffer = buffer->buffers[cpu]; 2650 ret = local_read(&cpu_buffer->commit_overrun); 2651 2652 return ret; 2653 } 2654 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu); 2655 2656 /** 2657 * ring_buffer_entries - get the number of entries in a buffer 2658 * @buffer: The ring buffer 2659 * 2660 * Returns the total number of entries in the ring buffer 2661 * (all CPU entries) 2662 */ 2663 unsigned long ring_buffer_entries(struct ring_buffer *buffer) 2664 { 2665 struct ring_buffer_per_cpu *cpu_buffer; 2666 unsigned long entries = 0; 2667 int cpu; 2668 2669 /* if you care about this being correct, lock the buffer */ 2670 for_each_buffer_cpu(buffer, cpu) { 2671 cpu_buffer = buffer->buffers[cpu]; 2672 entries += (local_read(&cpu_buffer->entries) - 2673 local_read(&cpu_buffer->overrun)) - cpu_buffer->read; 2674 } 2675 2676 return entries; 2677 } 2678 EXPORT_SYMBOL_GPL(ring_buffer_entries); 2679 2680 /** 2681 * ring_buffer_overruns - get the number of overruns in buffer 2682 * @buffer: The ring buffer 2683 * 2684 * Returns the total number of overruns in the ring buffer 2685 * (all CPU entries) 2686 */ 2687 unsigned long ring_buffer_overruns(struct ring_buffer *buffer) 2688 { 2689 struct ring_buffer_per_cpu *cpu_buffer; 2690 unsigned long overruns = 0; 2691 int cpu; 2692 2693 /* if you care about this being correct, lock the buffer */ 2694 for_each_buffer_cpu(buffer, cpu) { 2695 cpu_buffer = buffer->buffers[cpu]; 2696 overruns += local_read(&cpu_buffer->overrun); 2697 } 2698 2699 return overruns; 2700 } 2701 EXPORT_SYMBOL_GPL(ring_buffer_overruns); 2702 2703 static void rb_iter_reset(struct ring_buffer_iter *iter) 2704 { 2705 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; 2706 2707 /* Iterator usage is expected to have record disabled */ 2708 if (list_empty(&cpu_buffer->reader_page->list)) { 2709 iter->head_page = rb_set_head_page(cpu_buffer); 2710 if (unlikely(!iter->head_page)) 2711 return; 2712 iter->head = iter->head_page->read; 2713 } else { 2714 iter->head_page = cpu_buffer->reader_page; 2715 iter->head = cpu_buffer->reader_page->read; 2716 } 2717 if (iter->head) 2718 iter->read_stamp = cpu_buffer->read_stamp; 2719 else 2720 iter->read_stamp = iter->head_page->page->time_stamp; 2721 iter->cache_reader_page = cpu_buffer->reader_page; 2722 iter->cache_read = cpu_buffer->read; 2723 } 2724 2725 /** 2726 * ring_buffer_iter_reset - reset an iterator 2727 * @iter: The iterator to reset 2728 * 2729 * Resets the iterator, so that it will start from the beginning 2730 * again. 2731 */ 2732 void ring_buffer_iter_reset(struct ring_buffer_iter *iter) 2733 { 2734 struct ring_buffer_per_cpu *cpu_buffer; 2735 unsigned long flags; 2736 2737 if (!iter) 2738 return; 2739 2740 cpu_buffer = iter->cpu_buffer; 2741 2742 spin_lock_irqsave(&cpu_buffer->reader_lock, flags); 2743 rb_iter_reset(iter); 2744 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); 2745 } 2746 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset); 2747 2748 /** 2749 * ring_buffer_iter_empty - check if an iterator has no more to read 2750 * @iter: The iterator to check 2751 */ 2752 int ring_buffer_iter_empty(struct ring_buffer_iter *iter) 2753 { 2754 struct ring_buffer_per_cpu *cpu_buffer; 2755 2756 cpu_buffer = iter->cpu_buffer; 2757 2758 return iter->head_page == cpu_buffer->commit_page && 2759 iter->head == rb_commit_index(cpu_buffer); 2760 } 2761 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty); 2762 2763 static void 2764 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer, 2765 struct ring_buffer_event *event) 2766 { 2767 u64 delta; 2768 2769 switch (event->type_len) { 2770 case RINGBUF_TYPE_PADDING: 2771 return; 2772 2773 case RINGBUF_TYPE_TIME_EXTEND: 2774 delta = event->array[0]; 2775 delta <<= TS_SHIFT; 2776 delta += event->time_delta; 2777 cpu_buffer->read_stamp += delta; 2778 return; 2779 2780 case RINGBUF_TYPE_TIME_STAMP: 2781 /* FIXME: not implemented */ 2782 return; 2783 2784 case RINGBUF_TYPE_DATA: 2785 cpu_buffer->read_stamp += event->time_delta; 2786 return; 2787 2788 default: 2789 BUG(); 2790 } 2791 return; 2792 } 2793 2794 static void 2795 rb_update_iter_read_stamp(struct ring_buffer_iter *iter, 2796 struct ring_buffer_event *event) 2797 { 2798 u64 delta; 2799 2800 switch (event->type_len) { 2801 case RINGBUF_TYPE_PADDING: 2802 return; 2803 2804 case RINGBUF_TYPE_TIME_EXTEND: 2805 delta = event->array[0]; 2806 delta <<= TS_SHIFT; 2807 delta += event->time_delta; 2808 iter->read_stamp += delta; 2809 return; 2810 2811 case RINGBUF_TYPE_TIME_STAMP: 2812 /* FIXME: not implemented */ 2813 return; 2814 2815 case RINGBUF_TYPE_DATA: 2816 iter->read_stamp += event->time_delta; 2817 return; 2818 2819 default: 2820 BUG(); 2821 } 2822 return; 2823 } 2824 2825 static struct buffer_page * 2826 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer) 2827 { 2828 struct buffer_page *reader = NULL; 2829 unsigned long flags; 2830 int nr_loops = 0; 2831 int ret; 2832 2833 local_irq_save(flags); 2834 arch_spin_lock(&cpu_buffer->lock); 2835 2836 again: 2837 /* 2838 * This should normally only loop twice. But because the 2839 * start of the reader inserts an empty page, it causes 2840 * a case where we will loop three times. There should be no 2841 * reason to loop four times (that I know of). 2842 */ 2843 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) { 2844 reader = NULL; 2845 goto out; 2846 } 2847 2848 reader = cpu_buffer->reader_page; 2849 2850 /* If there's more to read, return this page */ 2851 if (cpu_buffer->reader_page->read < rb_page_size(reader)) 2852 goto out; 2853 2854 /* Never should we have an index greater than the size */ 2855 if (RB_WARN_ON(cpu_buffer, 2856 cpu_buffer->reader_page->read > rb_page_size(reader))) 2857 goto out; 2858 2859 /* check if we caught up to the tail */ 2860 reader = NULL; 2861 if (cpu_buffer->commit_page == cpu_buffer->reader_page) 2862 goto out; 2863 2864 /* 2865 * Reset the reader page to size zero. 2866 */ 2867 local_set(&cpu_buffer->reader_page->write, 0); 2868 local_set(&cpu_buffer->reader_page->entries, 0); 2869 local_set(&cpu_buffer->reader_page->page->commit, 0); 2870 2871 spin: 2872 /* 2873 * Splice the empty reader page into the list around the head. 2874 */ 2875 reader = rb_set_head_page(cpu_buffer); 2876 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next); 2877 cpu_buffer->reader_page->list.prev = reader->list.prev; 2878 2879 /* 2880 * cpu_buffer->pages just needs to point to the buffer, it 2881 * has no specific buffer page to point to. Lets move it out 2882 * of our way so we don't accidently swap it. 2883 */ 2884 cpu_buffer->pages = reader->list.prev; 2885 2886 /* The reader page will be pointing to the new head */ 2887 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list); 2888 2889 /* 2890 * Here's the tricky part. 2891 * 2892 * We need to move the pointer past the header page. 2893 * But we can only do that if a writer is not currently 2894 * moving it. The page before the header page has the 2895 * flag bit '1' set if it is pointing to the page we want. 2896 * but if the writer is in the process of moving it 2897 * than it will be '2' or already moved '0'. 2898 */ 2899 2900 ret = rb_head_page_replace(reader, cpu_buffer->reader_page); 2901 2902 /* 2903 * If we did not convert it, then we must try again. 2904 */ 2905 if (!ret) 2906 goto spin; 2907 2908 /* 2909 * Yeah! We succeeded in replacing the page. 2910 * 2911 * Now make the new head point back to the reader page. 2912 */ 2913 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list; 2914 rb_inc_page(cpu_buffer, &cpu_buffer->head_page); 2915 2916 /* Finally update the reader page to the new head */ 2917 cpu_buffer->reader_page = reader; 2918 rb_reset_reader_page(cpu_buffer); 2919 2920 goto again; 2921 2922 out: 2923 arch_spin_unlock(&cpu_buffer->lock); 2924 local_irq_restore(flags); 2925 2926 return reader; 2927 } 2928 2929 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer) 2930 { 2931 struct ring_buffer_event *event; 2932 struct buffer_page *reader; 2933 unsigned length; 2934 2935 reader = rb_get_reader_page(cpu_buffer); 2936 2937 /* This function should not be called when buffer is empty */ 2938 if (RB_WARN_ON(cpu_buffer, !reader)) 2939 return; 2940 2941 event = rb_reader_event(cpu_buffer); 2942 2943 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX) 2944 cpu_buffer->read++; 2945 2946 rb_update_read_stamp(cpu_buffer, event); 2947 2948 length = rb_event_length(event); 2949 cpu_buffer->reader_page->read += length; 2950 } 2951 2952 static void rb_advance_iter(struct ring_buffer_iter *iter) 2953 { 2954 struct ring_buffer *buffer; 2955 struct ring_buffer_per_cpu *cpu_buffer; 2956 struct ring_buffer_event *event; 2957 unsigned length; 2958 2959 cpu_buffer = iter->cpu_buffer; 2960 buffer = cpu_buffer->buffer; 2961 2962 /* 2963 * Check if we are at the end of the buffer. 2964 */ 2965 if (iter->head >= rb_page_size(iter->head_page)) { 2966 /* discarded commits can make the page empty */ 2967 if (iter->head_page == cpu_buffer->commit_page) 2968 return; 2969 rb_inc_iter(iter); 2970 return; 2971 } 2972 2973 event = rb_iter_head_event(iter); 2974 2975 length = rb_event_length(event); 2976 2977 /* 2978 * This should not be called to advance the header if we are 2979 * at the tail of the buffer. 2980 */ 2981 if (RB_WARN_ON(cpu_buffer, 2982 (iter->head_page == cpu_buffer->commit_page) && 2983 (iter->head + length > rb_commit_index(cpu_buffer)))) 2984 return; 2985 2986 rb_update_iter_read_stamp(iter, event); 2987 2988 iter->head += length; 2989 2990 /* check for end of page padding */ 2991 if ((iter->head >= rb_page_size(iter->head_page)) && 2992 (iter->head_page != cpu_buffer->commit_page)) 2993 rb_advance_iter(iter); 2994 } 2995 2996 static struct ring_buffer_event * 2997 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts) 2998 { 2999 struct ring_buffer_event *event; 3000 struct buffer_page *reader; 3001 int nr_loops = 0; 3002 3003 again: 3004 /* 3005 * We repeat when a timestamp is encountered. It is possible 3006 * to get multiple timestamps from an interrupt entering just 3007 * as one timestamp is about to be written, or from discarded 3008 * commits. The most that we can have is the number on a single page. 3009 */ 3010 if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE)) 3011 return NULL; 3012 3013 reader = rb_get_reader_page(cpu_buffer); 3014 if (!reader) 3015 return NULL; 3016 3017 event = rb_reader_event(cpu_buffer); 3018 3019 switch (event->type_len) { 3020 case RINGBUF_TYPE_PADDING: 3021 if (rb_null_event(event)) 3022 RB_WARN_ON(cpu_buffer, 1); 3023 /* 3024 * Because the writer could be discarding every 3025 * event it creates (which would probably be bad) 3026 * if we were to go back to "again" then we may never 3027 * catch up, and will trigger the warn on, or lock 3028 * the box. Return the padding, and we will release 3029 * the current locks, and try again. 3030 */ 3031 return event; 3032 3033 case RINGBUF_TYPE_TIME_EXTEND: 3034 /* Internal data, OK to advance */ 3035 rb_advance_reader(cpu_buffer); 3036 goto again; 3037 3038 case RINGBUF_TYPE_TIME_STAMP: 3039 /* FIXME: not implemented */ 3040 rb_advance_reader(cpu_buffer); 3041 goto again; 3042 3043 case RINGBUF_TYPE_DATA: 3044 if (ts) { 3045 *ts = cpu_buffer->read_stamp + event->time_delta; 3046 ring_buffer_normalize_time_stamp(cpu_buffer->buffer, 3047 cpu_buffer->cpu, ts); 3048 } 3049 return event; 3050 3051 default: 3052 BUG(); 3053 } 3054 3055 return NULL; 3056 } 3057 EXPORT_SYMBOL_GPL(ring_buffer_peek); 3058 3059 static struct ring_buffer_event * 3060 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts) 3061 { 3062 struct ring_buffer *buffer; 3063 struct ring_buffer_per_cpu *cpu_buffer; 3064 struct ring_buffer_event *event; 3065 int nr_loops = 0; 3066 3067 cpu_buffer = iter->cpu_buffer; 3068 buffer = cpu_buffer->buffer; 3069 3070 /* 3071 * Check if someone performed a consuming read to 3072 * the buffer. A consuming read invalidates the iterator 3073 * and we need to reset the iterator in this case. 3074 */ 3075 if (unlikely(iter->cache_read != cpu_buffer->read || 3076 iter->cache_reader_page != cpu_buffer->reader_page)) 3077 rb_iter_reset(iter); 3078 3079 again: 3080 if (ring_buffer_iter_empty(iter)) 3081 return NULL; 3082 3083 /* 3084 * We repeat when a timestamp is encountered. 3085 * We can get multiple timestamps by nested interrupts or also 3086 * if filtering is on (discarding commits). Since discarding 3087 * commits can be frequent we can get a lot of timestamps. 3088 * But we limit them by not adding timestamps if they begin 3089 * at the start of a page. 3090 */ 3091 if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE)) 3092 return NULL; 3093 3094 if (rb_per_cpu_empty(cpu_buffer)) 3095 return NULL; 3096 3097 if (iter->head >= local_read(&iter->head_page->page->commit)) { 3098 rb_inc_iter(iter); 3099 goto again; 3100 } 3101 3102 event = rb_iter_head_event(iter); 3103 3104 switch (event->type_len) { 3105 case RINGBUF_TYPE_PADDING: 3106 if (rb_null_event(event)) { 3107 rb_inc_iter(iter); 3108 goto again; 3109 } 3110 rb_advance_iter(iter); 3111 return event; 3112 3113 case RINGBUF_TYPE_TIME_EXTEND: 3114 /* Internal data, OK to advance */ 3115 rb_advance_iter(iter); 3116 goto again; 3117 3118 case RINGBUF_TYPE_TIME_STAMP: 3119 /* FIXME: not implemented */ 3120 rb_advance_iter(iter); 3121 goto again; 3122 3123 case RINGBUF_TYPE_DATA: 3124 if (ts) { 3125 *ts = iter->read_stamp + event->time_delta; 3126 ring_buffer_normalize_time_stamp(buffer, 3127 cpu_buffer->cpu, ts); 3128 } 3129 return event; 3130 3131 default: 3132 BUG(); 3133 } 3134 3135 return NULL; 3136 } 3137 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek); 3138 3139 static inline int rb_ok_to_lock(void) 3140 { 3141 /* 3142 * If an NMI die dumps out the content of the ring buffer 3143 * do not grab locks. We also permanently disable the ring 3144 * buffer too. A one time deal is all you get from reading 3145 * the ring buffer from an NMI. 3146 */ 3147 if (likely(!in_nmi())) 3148 return 1; 3149 3150 tracing_off_permanent(); 3151 return 0; 3152 } 3153 3154 /** 3155 * ring_buffer_peek - peek at the next event to be read 3156 * @buffer: The ring buffer to read 3157 * @cpu: The cpu to peak at 3158 * @ts: The timestamp counter of this event. 3159 * 3160 * This will return the event that will be read next, but does 3161 * not consume the data. 3162 */ 3163 struct ring_buffer_event * 3164 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts) 3165 { 3166 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu]; 3167 struct ring_buffer_event *event; 3168 unsigned long flags; 3169 int dolock; 3170 3171 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 3172 return NULL; 3173 3174 dolock = rb_ok_to_lock(); 3175 again: 3176 local_irq_save(flags); 3177 if (dolock) 3178 spin_lock(&cpu_buffer->reader_lock); 3179 event = rb_buffer_peek(cpu_buffer, ts); 3180 if (event && event->type_len == RINGBUF_TYPE_PADDING) 3181 rb_advance_reader(cpu_buffer); 3182 if (dolock) 3183 spin_unlock(&cpu_buffer->reader_lock); 3184 local_irq_restore(flags); 3185 3186 if (event && event->type_len == RINGBUF_TYPE_PADDING) 3187 goto again; 3188 3189 return event; 3190 } 3191 3192 /** 3193 * ring_buffer_iter_peek - peek at the next event to be read 3194 * @iter: The ring buffer iterator 3195 * @ts: The timestamp counter of this event. 3196 * 3197 * This will return the event that will be read next, but does 3198 * not increment the iterator. 3199 */ 3200 struct ring_buffer_event * 3201 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts) 3202 { 3203 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; 3204 struct ring_buffer_event *event; 3205 unsigned long flags; 3206 3207 again: 3208 spin_lock_irqsave(&cpu_buffer->reader_lock, flags); 3209 event = rb_iter_peek(iter, ts); 3210 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); 3211 3212 if (event && event->type_len == RINGBUF_TYPE_PADDING) 3213 goto again; 3214 3215 return event; 3216 } 3217 3218 /** 3219 * ring_buffer_consume - return an event and consume it 3220 * @buffer: The ring buffer to get the next event from 3221 * 3222 * Returns the next event in the ring buffer, and that event is consumed. 3223 * Meaning, that sequential reads will keep returning a different event, 3224 * and eventually empty the ring buffer if the producer is slower. 3225 */ 3226 struct ring_buffer_event * 3227 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts) 3228 { 3229 struct ring_buffer_per_cpu *cpu_buffer; 3230 struct ring_buffer_event *event = NULL; 3231 unsigned long flags; 3232 int dolock; 3233 3234 dolock = rb_ok_to_lock(); 3235 3236 again: 3237 /* might be called in atomic */ 3238 preempt_disable(); 3239 3240 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 3241 goto out; 3242 3243 cpu_buffer = buffer->buffers[cpu]; 3244 local_irq_save(flags); 3245 if (dolock) 3246 spin_lock(&cpu_buffer->reader_lock); 3247 3248 event = rb_buffer_peek(cpu_buffer, ts); 3249 if (event) 3250 rb_advance_reader(cpu_buffer); 3251 3252 if (dolock) 3253 spin_unlock(&cpu_buffer->reader_lock); 3254 local_irq_restore(flags); 3255 3256 out: 3257 preempt_enable(); 3258 3259 if (event && event->type_len == RINGBUF_TYPE_PADDING) 3260 goto again; 3261 3262 return event; 3263 } 3264 EXPORT_SYMBOL_GPL(ring_buffer_consume); 3265 3266 /** 3267 * ring_buffer_read_start - start a non consuming read of the buffer 3268 * @buffer: The ring buffer to read from 3269 * @cpu: The cpu buffer to iterate over 3270 * 3271 * This starts up an iteration through the buffer. It also disables 3272 * the recording to the buffer until the reading is finished. 3273 * This prevents the reading from being corrupted. This is not 3274 * a consuming read, so a producer is not expected. 3275 * 3276 * Must be paired with ring_buffer_finish. 3277 */ 3278 struct ring_buffer_iter * 3279 ring_buffer_read_start(struct ring_buffer *buffer, int cpu) 3280 { 3281 struct ring_buffer_per_cpu *cpu_buffer; 3282 struct ring_buffer_iter *iter; 3283 unsigned long flags; 3284 3285 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 3286 return NULL; 3287 3288 iter = kmalloc(sizeof(*iter), GFP_KERNEL); 3289 if (!iter) 3290 return NULL; 3291 3292 cpu_buffer = buffer->buffers[cpu]; 3293 3294 iter->cpu_buffer = cpu_buffer; 3295 3296 atomic_inc(&cpu_buffer->record_disabled); 3297 synchronize_sched(); 3298 3299 spin_lock_irqsave(&cpu_buffer->reader_lock, flags); 3300 arch_spin_lock(&cpu_buffer->lock); 3301 rb_iter_reset(iter); 3302 arch_spin_unlock(&cpu_buffer->lock); 3303 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); 3304 3305 return iter; 3306 } 3307 EXPORT_SYMBOL_GPL(ring_buffer_read_start); 3308 3309 /** 3310 * ring_buffer_finish - finish reading the iterator of the buffer 3311 * @iter: The iterator retrieved by ring_buffer_start 3312 * 3313 * This re-enables the recording to the buffer, and frees the 3314 * iterator. 3315 */ 3316 void 3317 ring_buffer_read_finish(struct ring_buffer_iter *iter) 3318 { 3319 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; 3320 3321 atomic_dec(&cpu_buffer->record_disabled); 3322 kfree(iter); 3323 } 3324 EXPORT_SYMBOL_GPL(ring_buffer_read_finish); 3325 3326 /** 3327 * ring_buffer_read - read the next item in the ring buffer by the iterator 3328 * @iter: The ring buffer iterator 3329 * @ts: The time stamp of the event read. 3330 * 3331 * This reads the next event in the ring buffer and increments the iterator. 3332 */ 3333 struct ring_buffer_event * 3334 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts) 3335 { 3336 struct ring_buffer_event *event; 3337 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; 3338 unsigned long flags; 3339 3340 spin_lock_irqsave(&cpu_buffer->reader_lock, flags); 3341 again: 3342 event = rb_iter_peek(iter, ts); 3343 if (!event) 3344 goto out; 3345 3346 if (event->type_len == RINGBUF_TYPE_PADDING) 3347 goto again; 3348 3349 rb_advance_iter(iter); 3350 out: 3351 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); 3352 3353 return event; 3354 } 3355 EXPORT_SYMBOL_GPL(ring_buffer_read); 3356 3357 /** 3358 * ring_buffer_size - return the size of the ring buffer (in bytes) 3359 * @buffer: The ring buffer. 3360 */ 3361 unsigned long ring_buffer_size(struct ring_buffer *buffer) 3362 { 3363 return BUF_PAGE_SIZE * buffer->pages; 3364 } 3365 EXPORT_SYMBOL_GPL(ring_buffer_size); 3366 3367 static void 3368 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer) 3369 { 3370 rb_head_page_deactivate(cpu_buffer); 3371 3372 cpu_buffer->head_page 3373 = list_entry(cpu_buffer->pages, struct buffer_page, list); 3374 local_set(&cpu_buffer->head_page->write, 0); 3375 local_set(&cpu_buffer->head_page->entries, 0); 3376 local_set(&cpu_buffer->head_page->page->commit, 0); 3377 3378 cpu_buffer->head_page->read = 0; 3379 3380 cpu_buffer->tail_page = cpu_buffer->head_page; 3381 cpu_buffer->commit_page = cpu_buffer->head_page; 3382 3383 INIT_LIST_HEAD(&cpu_buffer->reader_page->list); 3384 local_set(&cpu_buffer->reader_page->write, 0); 3385 local_set(&cpu_buffer->reader_page->entries, 0); 3386 local_set(&cpu_buffer->reader_page->page->commit, 0); 3387 cpu_buffer->reader_page->read = 0; 3388 3389 local_set(&cpu_buffer->commit_overrun, 0); 3390 local_set(&cpu_buffer->overrun, 0); 3391 local_set(&cpu_buffer->entries, 0); 3392 local_set(&cpu_buffer->committing, 0); 3393 local_set(&cpu_buffer->commits, 0); 3394 cpu_buffer->read = 0; 3395 3396 cpu_buffer->write_stamp = 0; 3397 cpu_buffer->read_stamp = 0; 3398 3399 rb_head_page_activate(cpu_buffer); 3400 } 3401 3402 /** 3403 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer 3404 * @buffer: The ring buffer to reset a per cpu buffer of 3405 * @cpu: The CPU buffer to be reset 3406 */ 3407 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu) 3408 { 3409 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu]; 3410 unsigned long flags; 3411 3412 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 3413 return; 3414 3415 atomic_inc(&cpu_buffer->record_disabled); 3416 3417 spin_lock_irqsave(&cpu_buffer->reader_lock, flags); 3418 3419 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing))) 3420 goto out; 3421 3422 arch_spin_lock(&cpu_buffer->lock); 3423 3424 rb_reset_cpu(cpu_buffer); 3425 3426 arch_spin_unlock(&cpu_buffer->lock); 3427 3428 out: 3429 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); 3430 3431 atomic_dec(&cpu_buffer->record_disabled); 3432 } 3433 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu); 3434 3435 /** 3436 * ring_buffer_reset - reset a ring buffer 3437 * @buffer: The ring buffer to reset all cpu buffers 3438 */ 3439 void ring_buffer_reset(struct ring_buffer *buffer) 3440 { 3441 int cpu; 3442 3443 for_each_buffer_cpu(buffer, cpu) 3444 ring_buffer_reset_cpu(buffer, cpu); 3445 } 3446 EXPORT_SYMBOL_GPL(ring_buffer_reset); 3447 3448 /** 3449 * rind_buffer_empty - is the ring buffer empty? 3450 * @buffer: The ring buffer to test 3451 */ 3452 int ring_buffer_empty(struct ring_buffer *buffer) 3453 { 3454 struct ring_buffer_per_cpu *cpu_buffer; 3455 unsigned long flags; 3456 int dolock; 3457 int cpu; 3458 int ret; 3459 3460 dolock = rb_ok_to_lock(); 3461 3462 /* yes this is racy, but if you don't like the race, lock the buffer */ 3463 for_each_buffer_cpu(buffer, cpu) { 3464 cpu_buffer = buffer->buffers[cpu]; 3465 local_irq_save(flags); 3466 if (dolock) 3467 spin_lock(&cpu_buffer->reader_lock); 3468 ret = rb_per_cpu_empty(cpu_buffer); 3469 if (dolock) 3470 spin_unlock(&cpu_buffer->reader_lock); 3471 local_irq_restore(flags); 3472 3473 if (!ret) 3474 return 0; 3475 } 3476 3477 return 1; 3478 } 3479 EXPORT_SYMBOL_GPL(ring_buffer_empty); 3480 3481 /** 3482 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty? 3483 * @buffer: The ring buffer 3484 * @cpu: The CPU buffer to test 3485 */ 3486 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu) 3487 { 3488 struct ring_buffer_per_cpu *cpu_buffer; 3489 unsigned long flags; 3490 int dolock; 3491 int ret; 3492 3493 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 3494 return 1; 3495 3496 dolock = rb_ok_to_lock(); 3497 3498 cpu_buffer = buffer->buffers[cpu]; 3499 local_irq_save(flags); 3500 if (dolock) 3501 spin_lock(&cpu_buffer->reader_lock); 3502 ret = rb_per_cpu_empty(cpu_buffer); 3503 if (dolock) 3504 spin_unlock(&cpu_buffer->reader_lock); 3505 local_irq_restore(flags); 3506 3507 return ret; 3508 } 3509 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu); 3510 3511 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP 3512 /** 3513 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers 3514 * @buffer_a: One buffer to swap with 3515 * @buffer_b: The other buffer to swap with 3516 * 3517 * This function is useful for tracers that want to take a "snapshot" 3518 * of a CPU buffer and has another back up buffer lying around. 3519 * it is expected that the tracer handles the cpu buffer not being 3520 * used at the moment. 3521 */ 3522 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a, 3523 struct ring_buffer *buffer_b, int cpu) 3524 { 3525 struct ring_buffer_per_cpu *cpu_buffer_a; 3526 struct ring_buffer_per_cpu *cpu_buffer_b; 3527 int ret = -EINVAL; 3528 3529 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) || 3530 !cpumask_test_cpu(cpu, buffer_b->cpumask)) 3531 goto out; 3532 3533 /* At least make sure the two buffers are somewhat the same */ 3534 if (buffer_a->pages != buffer_b->pages) 3535 goto out; 3536 3537 ret = -EAGAIN; 3538 3539 if (ring_buffer_flags != RB_BUFFERS_ON) 3540 goto out; 3541 3542 if (atomic_read(&buffer_a->record_disabled)) 3543 goto out; 3544 3545 if (atomic_read(&buffer_b->record_disabled)) 3546 goto out; 3547 3548 cpu_buffer_a = buffer_a->buffers[cpu]; 3549 cpu_buffer_b = buffer_b->buffers[cpu]; 3550 3551 if (atomic_read(&cpu_buffer_a->record_disabled)) 3552 goto out; 3553 3554 if (atomic_read(&cpu_buffer_b->record_disabled)) 3555 goto out; 3556 3557 /* 3558 * We can't do a synchronize_sched here because this 3559 * function can be called in atomic context. 3560 * Normally this will be called from the same CPU as cpu. 3561 * If not it's up to the caller to protect this. 3562 */ 3563 atomic_inc(&cpu_buffer_a->record_disabled); 3564 atomic_inc(&cpu_buffer_b->record_disabled); 3565 3566 ret = -EBUSY; 3567 if (local_read(&cpu_buffer_a->committing)) 3568 goto out_dec; 3569 if (local_read(&cpu_buffer_b->committing)) 3570 goto out_dec; 3571 3572 buffer_a->buffers[cpu] = cpu_buffer_b; 3573 buffer_b->buffers[cpu] = cpu_buffer_a; 3574 3575 cpu_buffer_b->buffer = buffer_a; 3576 cpu_buffer_a->buffer = buffer_b; 3577 3578 ret = 0; 3579 3580 out_dec: 3581 atomic_dec(&cpu_buffer_a->record_disabled); 3582 atomic_dec(&cpu_buffer_b->record_disabled); 3583 out: 3584 return ret; 3585 } 3586 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu); 3587 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */ 3588 3589 /** 3590 * ring_buffer_alloc_read_page - allocate a page to read from buffer 3591 * @buffer: the buffer to allocate for. 3592 * 3593 * This function is used in conjunction with ring_buffer_read_page. 3594 * When reading a full page from the ring buffer, these functions 3595 * can be used to speed up the process. The calling function should 3596 * allocate a few pages first with this function. Then when it 3597 * needs to get pages from the ring buffer, it passes the result 3598 * of this function into ring_buffer_read_page, which will swap 3599 * the page that was allocated, with the read page of the buffer. 3600 * 3601 * Returns: 3602 * The page allocated, or NULL on error. 3603 */ 3604 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer) 3605 { 3606 struct buffer_data_page *bpage; 3607 unsigned long addr; 3608 3609 addr = __get_free_page(GFP_KERNEL); 3610 if (!addr) 3611 return NULL; 3612 3613 bpage = (void *)addr; 3614 3615 rb_init_page(bpage); 3616 3617 return bpage; 3618 } 3619 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page); 3620 3621 /** 3622 * ring_buffer_free_read_page - free an allocated read page 3623 * @buffer: the buffer the page was allocate for 3624 * @data: the page to free 3625 * 3626 * Free a page allocated from ring_buffer_alloc_read_page. 3627 */ 3628 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data) 3629 { 3630 free_page((unsigned long)data); 3631 } 3632 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page); 3633 3634 /** 3635 * ring_buffer_read_page - extract a page from the ring buffer 3636 * @buffer: buffer to extract from 3637 * @data_page: the page to use allocated from ring_buffer_alloc_read_page 3638 * @len: amount to extract 3639 * @cpu: the cpu of the buffer to extract 3640 * @full: should the extraction only happen when the page is full. 3641 * 3642 * This function will pull out a page from the ring buffer and consume it. 3643 * @data_page must be the address of the variable that was returned 3644 * from ring_buffer_alloc_read_page. This is because the page might be used 3645 * to swap with a page in the ring buffer. 3646 * 3647 * for example: 3648 * rpage = ring_buffer_alloc_read_page(buffer); 3649 * if (!rpage) 3650 * return error; 3651 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0); 3652 * if (ret >= 0) 3653 * process_page(rpage, ret); 3654 * 3655 * When @full is set, the function will not return true unless 3656 * the writer is off the reader page. 3657 * 3658 * Note: it is up to the calling functions to handle sleeps and wakeups. 3659 * The ring buffer can be used anywhere in the kernel and can not 3660 * blindly call wake_up. The layer that uses the ring buffer must be 3661 * responsible for that. 3662 * 3663 * Returns: 3664 * >=0 if data has been transferred, returns the offset of consumed data. 3665 * <0 if no data has been transferred. 3666 */ 3667 int ring_buffer_read_page(struct ring_buffer *buffer, 3668 void **data_page, size_t len, int cpu, int full) 3669 { 3670 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu]; 3671 struct ring_buffer_event *event; 3672 struct buffer_data_page *bpage; 3673 struct buffer_page *reader; 3674 unsigned long flags; 3675 unsigned int commit; 3676 unsigned int read; 3677 u64 save_timestamp; 3678 int ret = -1; 3679 3680 if (!cpumask_test_cpu(cpu, buffer->cpumask)) 3681 goto out; 3682 3683 /* 3684 * If len is not big enough to hold the page header, then 3685 * we can not copy anything. 3686 */ 3687 if (len <= BUF_PAGE_HDR_SIZE) 3688 goto out; 3689 3690 len -= BUF_PAGE_HDR_SIZE; 3691 3692 if (!data_page) 3693 goto out; 3694 3695 bpage = *data_page; 3696 if (!bpage) 3697 goto out; 3698 3699 spin_lock_irqsave(&cpu_buffer->reader_lock, flags); 3700 3701 reader = rb_get_reader_page(cpu_buffer); 3702 if (!reader) 3703 goto out_unlock; 3704 3705 event = rb_reader_event(cpu_buffer); 3706 3707 read = reader->read; 3708 commit = rb_page_commit(reader); 3709 3710 /* 3711 * If this page has been partially read or 3712 * if len is not big enough to read the rest of the page or 3713 * a writer is still on the page, then 3714 * we must copy the data from the page to the buffer. 3715 * Otherwise, we can simply swap the page with the one passed in. 3716 */ 3717 if (read || (len < (commit - read)) || 3718 cpu_buffer->reader_page == cpu_buffer->commit_page) { 3719 struct buffer_data_page *rpage = cpu_buffer->reader_page->page; 3720 unsigned int rpos = read; 3721 unsigned int pos = 0; 3722 unsigned int size; 3723 3724 if (full) 3725 goto out_unlock; 3726 3727 if (len > (commit - read)) 3728 len = (commit - read); 3729 3730 size = rb_event_length(event); 3731 3732 if (len < size) 3733 goto out_unlock; 3734 3735 /* save the current timestamp, since the user will need it */ 3736 save_timestamp = cpu_buffer->read_stamp; 3737 3738 /* Need to copy one event at a time */ 3739 do { 3740 memcpy(bpage->data + pos, rpage->data + rpos, size); 3741 3742 len -= size; 3743 3744 rb_advance_reader(cpu_buffer); 3745 rpos = reader->read; 3746 pos += size; 3747 3748 event = rb_reader_event(cpu_buffer); 3749 size = rb_event_length(event); 3750 } while (len > size); 3751 3752 /* update bpage */ 3753 local_set(&bpage->commit, pos); 3754 bpage->time_stamp = save_timestamp; 3755 3756 /* we copied everything to the beginning */ 3757 read = 0; 3758 } else { 3759 /* update the entry counter */ 3760 cpu_buffer->read += rb_page_entries(reader); 3761 3762 /* swap the pages */ 3763 rb_init_page(bpage); 3764 bpage = reader->page; 3765 reader->page = *data_page; 3766 local_set(&reader->write, 0); 3767 local_set(&reader->entries, 0); 3768 reader->read = 0; 3769 *data_page = bpage; 3770 } 3771 ret = read; 3772 3773 out_unlock: 3774 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); 3775 3776 out: 3777 return ret; 3778 } 3779 EXPORT_SYMBOL_GPL(ring_buffer_read_page); 3780 3781 #ifdef CONFIG_TRACING 3782 static ssize_t 3783 rb_simple_read(struct file *filp, char __user *ubuf, 3784 size_t cnt, loff_t *ppos) 3785 { 3786 unsigned long *p = filp->private_data; 3787 char buf[64]; 3788 int r; 3789 3790 if (test_bit(RB_BUFFERS_DISABLED_BIT, p)) 3791 r = sprintf(buf, "permanently disabled\n"); 3792 else 3793 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p)); 3794 3795 return simple_read_from_buffer(ubuf, cnt, ppos, buf, r); 3796 } 3797 3798 static ssize_t 3799 rb_simple_write(struct file *filp, const char __user *ubuf, 3800 size_t cnt, loff_t *ppos) 3801 { 3802 unsigned long *p = filp->private_data; 3803 char buf[64]; 3804 unsigned long val; 3805 int ret; 3806 3807 if (cnt >= sizeof(buf)) 3808 return -EINVAL; 3809 3810 if (copy_from_user(&buf, ubuf, cnt)) 3811 return -EFAULT; 3812 3813 buf[cnt] = 0; 3814 3815 ret = strict_strtoul(buf, 10, &val); 3816 if (ret < 0) 3817 return ret; 3818 3819 if (val) 3820 set_bit(RB_BUFFERS_ON_BIT, p); 3821 else 3822 clear_bit(RB_BUFFERS_ON_BIT, p); 3823 3824 (*ppos)++; 3825 3826 return cnt; 3827 } 3828 3829 static const struct file_operations rb_simple_fops = { 3830 .open = tracing_open_generic, 3831 .read = rb_simple_read, 3832 .write = rb_simple_write, 3833 }; 3834 3835 3836 static __init int rb_init_debugfs(void) 3837 { 3838 struct dentry *d_tracer; 3839 3840 d_tracer = tracing_init_dentry(); 3841 3842 trace_create_file("tracing_on", 0644, d_tracer, 3843 &ring_buffer_flags, &rb_simple_fops); 3844 3845 return 0; 3846 } 3847 3848 fs_initcall(rb_init_debugfs); 3849 #endif 3850 3851 #ifdef CONFIG_HOTPLUG_CPU 3852 static int rb_cpu_notify(struct notifier_block *self, 3853 unsigned long action, void *hcpu) 3854 { 3855 struct ring_buffer *buffer = 3856 container_of(self, struct ring_buffer, cpu_notify); 3857 long cpu = (long)hcpu; 3858 3859 switch (action) { 3860 case CPU_UP_PREPARE: 3861 case CPU_UP_PREPARE_FROZEN: 3862 if (cpumask_test_cpu(cpu, buffer->cpumask)) 3863 return NOTIFY_OK; 3864 3865 buffer->buffers[cpu] = 3866 rb_allocate_cpu_buffer(buffer, cpu); 3867 if (!buffer->buffers[cpu]) { 3868 WARN(1, "failed to allocate ring buffer on CPU %ld\n", 3869 cpu); 3870 return NOTIFY_OK; 3871 } 3872 smp_wmb(); 3873 cpumask_set_cpu(cpu, buffer->cpumask); 3874 break; 3875 case CPU_DOWN_PREPARE: 3876 case CPU_DOWN_PREPARE_FROZEN: 3877 /* 3878 * Do nothing. 3879 * If we were to free the buffer, then the user would 3880 * lose any trace that was in the buffer. 3881 */ 3882 break; 3883 default: 3884 break; 3885 } 3886 return NOTIFY_OK; 3887 } 3888 #endif 3889