yuezonghe | 824eb0c | 2024-06-27 02:32:26 -0700 | [diff] [blame] | 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/spinlock.h> |
| 9 | #include <linux/debugfs.h> |
| 10 | #include <linux/uaccess.h> |
| 11 | #include <linux/hardirq.h> |
| 12 | #include <linux/kmemcheck.h> |
| 13 | #include <linux/module.h> |
| 14 | #include <linux/percpu.h> |
| 15 | #include <linux/mutex.h> |
| 16 | #include <linux/slab.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 <asm/local.h> |
| 24 | #include "trace.h" |
| 25 | |
| 26 | /* |
| 27 | * The ring buffer header is special. We must manually up keep it. |
| 28 | */ |
| 29 | int ring_buffer_print_entry_header(struct trace_seq *s) |
| 30 | { |
| 31 | int ret; |
| 32 | |
| 33 | ret = trace_seq_printf(s, "# compressed entry header\n"); |
| 34 | ret = trace_seq_printf(s, "\ttype_len : 5 bits\n"); |
| 35 | ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n"); |
| 36 | ret = trace_seq_printf(s, "\tarray : 32 bits\n"); |
| 37 | ret = trace_seq_printf(s, "\n"); |
| 38 | ret = trace_seq_printf(s, "\tpadding : type == %d\n", |
| 39 | RINGBUF_TYPE_PADDING); |
| 40 | ret = trace_seq_printf(s, "\ttime_extend : type == %d\n", |
| 41 | RINGBUF_TYPE_TIME_EXTEND); |
| 42 | ret = trace_seq_printf(s, "\tdata max type_len == %d\n", |
| 43 | RINGBUF_TYPE_DATA_TYPE_LEN_MAX); |
| 44 | |
| 45 | return ret; |
| 46 | } |
| 47 | |
| 48 | /* |
| 49 | * The ring buffer is made up of a list of pages. A separate list of pages is |
| 50 | * allocated for each CPU. A writer may only write to a buffer that is |
| 51 | * associated with the CPU it is currently executing on. A reader may read |
| 52 | * from any per cpu buffer. |
| 53 | * |
| 54 | * The reader is special. For each per cpu buffer, the reader has its own |
| 55 | * reader page. When a reader has read the entire reader page, this reader |
| 56 | * page is swapped with another page in the ring buffer. |
| 57 | * |
| 58 | * Now, as long as the writer is off the reader page, the reader can do what |
| 59 | * ever it wants with that page. The writer will never write to that page |
| 60 | * again (as long as it is out of the ring buffer). |
| 61 | * |
| 62 | * Here's some silly ASCII art. |
| 63 | * |
| 64 | * +------+ |
| 65 | * |reader| RING BUFFER |
| 66 | * |page | |
| 67 | * +------+ +---+ +---+ +---+ |
| 68 | * | |-->| |-->| | |
| 69 | * +---+ +---+ +---+ |
| 70 | * ^ | |
| 71 | * | | |
| 72 | * +---------------+ |
| 73 | * |
| 74 | * |
| 75 | * +------+ |
| 76 | * |reader| RING BUFFER |
| 77 | * |page |------------------v |
| 78 | * +------+ +---+ +---+ +---+ |
| 79 | * | |-->| |-->| | |
| 80 | * +---+ +---+ +---+ |
| 81 | * ^ | |
| 82 | * | | |
| 83 | * +---------------+ |
| 84 | * |
| 85 | * |
| 86 | * +------+ |
| 87 | * |reader| RING BUFFER |
| 88 | * |page |------------------v |
| 89 | * +------+ +---+ +---+ +---+ |
| 90 | * ^ | |-->| |-->| | |
| 91 | * | +---+ +---+ +---+ |
| 92 | * | | |
| 93 | * | | |
| 94 | * +------------------------------+ |
| 95 | * |
| 96 | * |
| 97 | * +------+ |
| 98 | * |buffer| RING BUFFER |
| 99 | * |page |------------------v |
| 100 | * +------+ +---+ +---+ +---+ |
| 101 | * ^ | | | |-->| | |
| 102 | * | New +---+ +---+ +---+ |
| 103 | * | Reader------^ | |
| 104 | * | page | |
| 105 | * +------------------------------+ |
| 106 | * |
| 107 | * |
| 108 | * After we make this swap, the reader can hand this page off to the splice |
| 109 | * code and be done with it. It can even allocate a new page if it needs to |
| 110 | * and swap that into the ring buffer. |
| 111 | * |
| 112 | * We will be using cmpxchg soon to make all this lockless. |
| 113 | * |
| 114 | */ |
| 115 | |
| 116 | /* |
| 117 | * A fast way to enable or disable all ring buffers is to |
| 118 | * call tracing_on or tracing_off. Turning off the ring buffers |
| 119 | * prevents all ring buffers from being recorded to. |
| 120 | * Turning this switch on, makes it OK to write to the |
| 121 | * ring buffer, if the ring buffer is enabled itself. |
| 122 | * |
| 123 | * There's three layers that must be on in order to write |
| 124 | * to the ring buffer. |
| 125 | * |
| 126 | * 1) This global flag must be set. |
| 127 | * 2) The ring buffer must be enabled for recording. |
| 128 | * 3) The per cpu buffer must be enabled for recording. |
| 129 | * |
| 130 | * In case of an anomaly, this global flag has a bit set that |
| 131 | * will permantly disable all ring buffers. |
| 132 | */ |
| 133 | |
| 134 | /* |
| 135 | * Global flag to disable all recording to ring buffers |
| 136 | * This has two bits: ON, DISABLED |
| 137 | * |
| 138 | * ON DISABLED |
| 139 | * ---- ---------- |
| 140 | * 0 0 : ring buffers are off |
| 141 | * 1 0 : ring buffers are on |
| 142 | * X 1 : ring buffers are permanently disabled |
| 143 | */ |
| 144 | |
| 145 | enum { |
| 146 | RB_BUFFERS_ON_BIT = 0, |
| 147 | RB_BUFFERS_DISABLED_BIT = 1, |
| 148 | }; |
| 149 | |
| 150 | enum { |
| 151 | RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT, |
| 152 | RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT, |
| 153 | }; |
| 154 | |
| 155 | static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON; |
| 156 | |
| 157 | /* Used for individual buffers (after the counter) */ |
| 158 | #define RB_BUFFER_OFF (1 << 20) |
| 159 | |
| 160 | #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data) |
| 161 | |
| 162 | /** |
| 163 | * tracing_off_permanent - permanently disable ring buffers |
| 164 | * |
| 165 | * This function, once called, will disable all ring buffers |
| 166 | * permanently. |
| 167 | */ |
| 168 | void tracing_off_permanent(void) |
| 169 | { |
| 170 | set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags); |
| 171 | } |
| 172 | |
| 173 | #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array)) |
| 174 | #define RB_ALIGNMENT 4U |
| 175 | #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX) |
| 176 | #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */ |
| 177 | |
| 178 | #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) |
| 179 | # define RB_FORCE_8BYTE_ALIGNMENT 0 |
| 180 | # define RB_ARCH_ALIGNMENT RB_ALIGNMENT |
| 181 | #else |
| 182 | # define RB_FORCE_8BYTE_ALIGNMENT 1 |
| 183 | # define RB_ARCH_ALIGNMENT 8U |
| 184 | #endif |
| 185 | |
| 186 | /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */ |
| 187 | #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX |
| 188 | |
| 189 | enum { |
| 190 | RB_LEN_TIME_EXTEND = 8, |
| 191 | RB_LEN_TIME_STAMP = 16, |
| 192 | }; |
| 193 | |
| 194 | #define skip_time_extend(event) \ |
| 195 | ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND)) |
| 196 | |
| 197 | static inline int rb_null_event(struct ring_buffer_event *event) |
| 198 | { |
| 199 | return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta; |
| 200 | } |
| 201 | |
| 202 | static void rb_event_set_padding(struct ring_buffer_event *event) |
| 203 | { |
| 204 | /* padding has a NULL time_delta */ |
| 205 | event->type_len = RINGBUF_TYPE_PADDING; |
| 206 | event->time_delta = 0; |
| 207 | } |
| 208 | |
| 209 | static unsigned |
| 210 | rb_event_data_length(struct ring_buffer_event *event) |
| 211 | { |
| 212 | unsigned length; |
| 213 | |
| 214 | if (event->type_len) |
| 215 | length = event->type_len * RB_ALIGNMENT; |
| 216 | else |
| 217 | length = event->array[0]; |
| 218 | return length + RB_EVNT_HDR_SIZE; |
| 219 | } |
| 220 | |
| 221 | /* |
| 222 | * Return the length of the given event. Will return |
| 223 | * the length of the time extend if the event is a |
| 224 | * time extend. |
| 225 | */ |
| 226 | static inline unsigned |
| 227 | rb_event_length(struct ring_buffer_event *event) |
| 228 | { |
| 229 | switch (event->type_len) { |
| 230 | case RINGBUF_TYPE_PADDING: |
| 231 | if (rb_null_event(event)) |
| 232 | /* undefined */ |
| 233 | return -1; |
| 234 | return event->array[0] + RB_EVNT_HDR_SIZE; |
| 235 | |
| 236 | case RINGBUF_TYPE_TIME_EXTEND: |
| 237 | return RB_LEN_TIME_EXTEND; |
| 238 | |
| 239 | case RINGBUF_TYPE_TIME_STAMP: |
| 240 | return RB_LEN_TIME_STAMP; |
| 241 | |
| 242 | case RINGBUF_TYPE_DATA: |
| 243 | return rb_event_data_length(event); |
| 244 | default: |
| 245 | BUG(); |
| 246 | } |
| 247 | /* not hit */ |
| 248 | return 0; |
| 249 | } |
| 250 | |
| 251 | /* |
| 252 | * Return total length of time extend and data, |
| 253 | * or just the event length for all other events. |
| 254 | */ |
| 255 | static inline unsigned |
| 256 | rb_event_ts_length(struct ring_buffer_event *event) |
| 257 | { |
| 258 | unsigned len = 0; |
| 259 | |
| 260 | if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) { |
| 261 | /* time extends include the data event after it */ |
| 262 | len = RB_LEN_TIME_EXTEND; |
| 263 | event = skip_time_extend(event); |
| 264 | } |
| 265 | return len + rb_event_length(event); |
| 266 | } |
| 267 | |
| 268 | /** |
| 269 | * ring_buffer_event_length - return the length of the event |
| 270 | * @event: the event to get the length of |
| 271 | * |
| 272 | * Returns the size of the data load of a data event. |
| 273 | * If the event is something other than a data event, it |
| 274 | * returns the size of the event itself. With the exception |
| 275 | * of a TIME EXTEND, where it still returns the size of the |
| 276 | * data load of the data event after it. |
| 277 | */ |
| 278 | unsigned ring_buffer_event_length(struct ring_buffer_event *event) |
| 279 | { |
| 280 | unsigned length; |
| 281 | |
| 282 | if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) |
| 283 | event = skip_time_extend(event); |
| 284 | |
| 285 | length = rb_event_length(event); |
| 286 | if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX) |
| 287 | return length; |
| 288 | length -= RB_EVNT_HDR_SIZE; |
| 289 | if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0])) |
| 290 | length -= sizeof(event->array[0]); |
| 291 | return length; |
| 292 | } |
| 293 | EXPORT_SYMBOL_GPL(ring_buffer_event_length); |
| 294 | |
| 295 | /* inline for ring buffer fast paths */ |
| 296 | static void * |
| 297 | rb_event_data(struct ring_buffer_event *event) |
| 298 | { |
| 299 | if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) |
| 300 | event = skip_time_extend(event); |
| 301 | BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX); |
| 302 | /* If length is in len field, then array[0] has the data */ |
| 303 | if (event->type_len) |
| 304 | return (void *)&event->array[0]; |
| 305 | /* Otherwise length is in array[0] and array[1] has the data */ |
| 306 | return (void *)&event->array[1]; |
| 307 | } |
| 308 | |
| 309 | /** |
| 310 | * ring_buffer_event_data - return the data of the event |
| 311 | * @event: the event to get the data from |
| 312 | */ |
| 313 | void *ring_buffer_event_data(struct ring_buffer_event *event) |
| 314 | { |
| 315 | return rb_event_data(event); |
| 316 | } |
| 317 | EXPORT_SYMBOL_GPL(ring_buffer_event_data); |
| 318 | |
| 319 | #define for_each_buffer_cpu(buffer, cpu) \ |
| 320 | for_each_cpu(cpu, buffer->cpumask) |
| 321 | |
| 322 | #define TS_SHIFT 27 |
| 323 | #define TS_MASK ((1ULL << TS_SHIFT) - 1) |
| 324 | #define TS_DELTA_TEST (~TS_MASK) |
| 325 | |
| 326 | /* Flag when events were overwritten */ |
| 327 | #define RB_MISSED_EVENTS (1 << 31) |
| 328 | /* Missed count stored at end */ |
| 329 | #define RB_MISSED_STORED (1 << 30) |
| 330 | |
| 331 | struct buffer_data_page { |
| 332 | u64 time_stamp; /* page time stamp */ |
| 333 | local_t commit; /* write committed index */ |
| 334 | unsigned char data[]; /* data of buffer page */ |
| 335 | }; |
| 336 | |
| 337 | /* |
| 338 | * Note, the buffer_page list must be first. The buffer pages |
| 339 | * are allocated in cache lines, which means that each buffer |
| 340 | * page will be at the beginning of a cache line, and thus |
| 341 | * the least significant bits will be zero. We use this to |
| 342 | * add flags in the list struct pointers, to make the ring buffer |
| 343 | * lockless. |
| 344 | */ |
| 345 | struct buffer_page { |
| 346 | struct list_head list; /* list of buffer pages */ |
| 347 | local_t write; /* index for next write */ |
| 348 | unsigned read; /* index for next read */ |
| 349 | local_t entries; /* entries on this page */ |
| 350 | unsigned long real_end; /* real end of data */ |
| 351 | struct buffer_data_page *page; /* Actual data page */ |
| 352 | }; |
| 353 | |
| 354 | /* |
| 355 | * The buffer page counters, write and entries, must be reset |
| 356 | * atomically when crossing page boundaries. To synchronize this |
| 357 | * update, two counters are inserted into the number. One is |
| 358 | * the actual counter for the write position or count on the page. |
| 359 | * |
| 360 | * The other is a counter of updaters. Before an update happens |
| 361 | * the update partition of the counter is incremented. This will |
| 362 | * allow the updater to update the counter atomically. |
| 363 | * |
| 364 | * The counter is 20 bits, and the state data is 12. |
| 365 | */ |
| 366 | #define RB_WRITE_MASK 0xfffff |
| 367 | #define RB_WRITE_INTCNT (1 << 20) |
| 368 | |
| 369 | static void rb_init_page(struct buffer_data_page *bpage) |
| 370 | { |
| 371 | local_set(&bpage->commit, 0); |
| 372 | } |
| 373 | |
| 374 | /** |
| 375 | * ring_buffer_page_len - the size of data on the page. |
| 376 | * @page: The page to read |
| 377 | * |
| 378 | * Returns the amount of data on the page, including buffer page header. |
| 379 | */ |
| 380 | size_t ring_buffer_page_len(void *page) |
| 381 | { |
| 382 | return local_read(&((struct buffer_data_page *)page)->commit) |
| 383 | + BUF_PAGE_HDR_SIZE; |
| 384 | } |
| 385 | |
| 386 | /* |
| 387 | * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing |
| 388 | * this issue out. |
| 389 | */ |
| 390 | static void free_buffer_page(struct buffer_page *bpage) |
| 391 | { |
| 392 | free_page((unsigned long)bpage->page); |
| 393 | kfree(bpage); |
| 394 | } |
| 395 | |
| 396 | /* |
| 397 | * We need to fit the time_stamp delta into 27 bits. |
| 398 | */ |
| 399 | static inline int test_time_stamp(u64 delta) |
| 400 | { |
| 401 | if (delta & TS_DELTA_TEST) |
| 402 | return 1; |
| 403 | return 0; |
| 404 | } |
| 405 | |
| 406 | #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE) |
| 407 | |
| 408 | /* Max payload is BUF_PAGE_SIZE - header (8bytes) */ |
| 409 | #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2)) |
| 410 | |
| 411 | int ring_buffer_print_page_header(struct trace_seq *s) |
| 412 | { |
| 413 | struct buffer_data_page field; |
| 414 | int ret; |
| 415 | |
| 416 | ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t" |
| 417 | "offset:0;\tsize:%u;\tsigned:%u;\n", |
| 418 | (unsigned int)sizeof(field.time_stamp), |
| 419 | (unsigned int)is_signed_type(u64)); |
| 420 | |
| 421 | ret = trace_seq_printf(s, "\tfield: local_t commit;\t" |
| 422 | "offset:%u;\tsize:%u;\tsigned:%u;\n", |
| 423 | (unsigned int)offsetof(typeof(field), commit), |
| 424 | (unsigned int)sizeof(field.commit), |
| 425 | (unsigned int)is_signed_type(long)); |
| 426 | |
| 427 | ret = trace_seq_printf(s, "\tfield: int overwrite;\t" |
| 428 | "offset:%u;\tsize:%u;\tsigned:%u;\n", |
| 429 | (unsigned int)offsetof(typeof(field), commit), |
| 430 | 1, |
| 431 | (unsigned int)is_signed_type(long)); |
| 432 | |
| 433 | ret = trace_seq_printf(s, "\tfield: char data;\t" |
| 434 | "offset:%u;\tsize:%u;\tsigned:%u;\n", |
| 435 | (unsigned int)offsetof(typeof(field), data), |
| 436 | (unsigned int)BUF_PAGE_SIZE, |
| 437 | (unsigned int)is_signed_type(char)); |
| 438 | |
| 439 | return ret; |
| 440 | } |
| 441 | |
| 442 | /* |
| 443 | * head_page == tail_page && head == tail then buffer is empty. |
| 444 | */ |
| 445 | struct ring_buffer_per_cpu { |
| 446 | int cpu; |
| 447 | atomic_t record_disabled; |
| 448 | struct ring_buffer *buffer; |
| 449 | spinlock_t reader_lock; /* serialize readers */ |
| 450 | arch_spinlock_t lock; |
| 451 | struct lock_class_key lock_key; |
| 452 | struct list_head *pages; |
| 453 | struct buffer_page *head_page; /* read from head */ |
| 454 | struct buffer_page *tail_page; /* write to tail */ |
| 455 | struct buffer_page *commit_page; /* committed pages */ |
| 456 | struct buffer_page *reader_page; |
| 457 | unsigned long lost_events; |
| 458 | unsigned long last_overrun; |
| 459 | local_t entries_bytes; |
| 460 | local_t commit_overrun; |
| 461 | local_t overrun; |
| 462 | local_t entries; |
| 463 | local_t committing; |
| 464 | local_t commits; |
| 465 | unsigned long read; |
| 466 | unsigned long read_bytes; |
| 467 | u64 write_stamp; |
| 468 | u64 read_stamp; |
| 469 | }; |
| 470 | |
| 471 | struct ring_buffer { |
| 472 | unsigned pages; |
| 473 | unsigned flags; |
| 474 | int cpus; |
| 475 | atomic_t record_disabled; |
| 476 | cpumask_var_t cpumask; |
| 477 | |
| 478 | struct lock_class_key *reader_lock_key; |
| 479 | |
| 480 | struct mutex mutex; |
| 481 | |
| 482 | struct ring_buffer_per_cpu **buffers; |
| 483 | |
| 484 | #ifdef CONFIG_HOTPLUG_CPU |
| 485 | struct notifier_block cpu_notify; |
| 486 | #endif |
| 487 | u64 (*clock)(void); |
| 488 | }; |
| 489 | |
| 490 | struct ring_buffer_iter { |
| 491 | struct ring_buffer_per_cpu *cpu_buffer; |
| 492 | unsigned long head; |
| 493 | struct buffer_page *head_page; |
| 494 | struct buffer_page *cache_reader_page; |
| 495 | unsigned long cache_read; |
| 496 | u64 read_stamp; |
| 497 | }; |
| 498 | |
| 499 | /* buffer may be either ring_buffer or ring_buffer_per_cpu */ |
| 500 | #define RB_WARN_ON(b, cond) \ |
| 501 | ({ \ |
| 502 | int _____ret = unlikely(cond); \ |
| 503 | if (_____ret) { \ |
| 504 | if (__same_type(*(b), struct ring_buffer_per_cpu)) { \ |
| 505 | struct ring_buffer_per_cpu *__b = \ |
| 506 | (void *)b; \ |
| 507 | atomic_inc(&__b->buffer->record_disabled); \ |
| 508 | } else \ |
| 509 | atomic_inc(&b->record_disabled); \ |
| 510 | WARN_ON(1); \ |
| 511 | } \ |
| 512 | _____ret; \ |
| 513 | }) |
| 514 | |
| 515 | /* Up this if you want to test the TIME_EXTENTS and normalization */ |
| 516 | #define DEBUG_SHIFT 0 |
| 517 | |
| 518 | static inline u64 rb_time_stamp(struct ring_buffer *buffer) |
| 519 | { |
| 520 | /* shift to debug/test normalization and TIME_EXTENTS */ |
| 521 | return buffer->clock() << DEBUG_SHIFT; |
| 522 | } |
| 523 | |
| 524 | u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu) |
| 525 | { |
| 526 | u64 time; |
| 527 | |
| 528 | preempt_disable_notrace(); |
| 529 | time = rb_time_stamp(buffer); |
| 530 | preempt_enable_no_resched_notrace(); |
| 531 | |
| 532 | return time; |
| 533 | } |
| 534 | EXPORT_SYMBOL_GPL(ring_buffer_time_stamp); |
| 535 | |
| 536 | void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer, |
| 537 | int cpu, u64 *ts) |
| 538 | { |
| 539 | /* Just stupid testing the normalize function and deltas */ |
| 540 | *ts >>= DEBUG_SHIFT; |
| 541 | } |
| 542 | EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp); |
| 543 | |
| 544 | /* |
| 545 | * Making the ring buffer lockless makes things tricky. |
| 546 | * Although writes only happen on the CPU that they are on, |
| 547 | * and they only need to worry about interrupts. Reads can |
| 548 | * happen on any CPU. |
| 549 | * |
| 550 | * The reader page is always off the ring buffer, but when the |
| 551 | * reader finishes with a page, it needs to swap its page with |
| 552 | * a new one from the buffer. The reader needs to take from |
| 553 | * the head (writes go to the tail). But if a writer is in overwrite |
| 554 | * mode and wraps, it must push the head page forward. |
| 555 | * |
| 556 | * Here lies the problem. |
| 557 | * |
| 558 | * The reader must be careful to replace only the head page, and |
| 559 | * not another one. As described at the top of the file in the |
| 560 | * ASCII art, the reader sets its old page to point to the next |
| 561 | * page after head. It then sets the page after head to point to |
| 562 | * the old reader page. But if the writer moves the head page |
| 563 | * during this operation, the reader could end up with the tail. |
| 564 | * |
| 565 | * We use cmpxchg to help prevent this race. We also do something |
| 566 | * special with the page before head. We set the LSB to 1. |
| 567 | * |
| 568 | * When the writer must push the page forward, it will clear the |
| 569 | * bit that points to the head page, move the head, and then set |
| 570 | * the bit that points to the new head page. |
| 571 | * |
| 572 | * We also don't want an interrupt coming in and moving the head |
| 573 | * page on another writer. Thus we use the second LSB to catch |
| 574 | * that too. Thus: |
| 575 | * |
| 576 | * head->list->prev->next bit 1 bit 0 |
| 577 | * ------- ------- |
| 578 | * Normal page 0 0 |
| 579 | * Points to head page 0 1 |
| 580 | * New head page 1 0 |
| 581 | * |
| 582 | * Note we can not trust the prev pointer of the head page, because: |
| 583 | * |
| 584 | * +----+ +-----+ +-----+ |
| 585 | * | |------>| T |---X--->| N | |
| 586 | * | |<------| | | | |
| 587 | * +----+ +-----+ +-----+ |
| 588 | * ^ ^ | |
| 589 | * | +-----+ | | |
| 590 | * +----------| R |----------+ | |
| 591 | * | |<-----------+ |
| 592 | * +-----+ |
| 593 | * |
| 594 | * Key: ---X--> HEAD flag set in pointer |
| 595 | * T Tail page |
| 596 | * R Reader page |
| 597 | * N Next page |
| 598 | * |
| 599 | * (see __rb_reserve_next() to see where this happens) |
| 600 | * |
| 601 | * What the above shows is that the reader just swapped out |
| 602 | * the reader page with a page in the buffer, but before it |
| 603 | * could make the new header point back to the new page added |
| 604 | * it was preempted by a writer. The writer moved forward onto |
| 605 | * the new page added by the reader and is about to move forward |
| 606 | * again. |
| 607 | * |
| 608 | * You can see, it is legitimate for the previous pointer of |
| 609 | * the head (or any page) not to point back to itself. But only |
| 610 | * temporarially. |
| 611 | */ |
| 612 | |
| 613 | #define RB_PAGE_NORMAL 0UL |
| 614 | #define RB_PAGE_HEAD 1UL |
| 615 | #define RB_PAGE_UPDATE 2UL |
| 616 | |
| 617 | |
| 618 | #define RB_FLAG_MASK 3UL |
| 619 | |
| 620 | /* PAGE_MOVED is not part of the mask */ |
| 621 | #define RB_PAGE_MOVED 4UL |
| 622 | |
| 623 | /* |
| 624 | * rb_list_head - remove any bit |
| 625 | */ |
| 626 | static struct list_head *rb_list_head(struct list_head *list) |
| 627 | { |
| 628 | unsigned long val = (unsigned long)list; |
| 629 | |
| 630 | return (struct list_head *)(val & ~RB_FLAG_MASK); |
| 631 | } |
| 632 | |
| 633 | /* |
| 634 | * rb_is_head_page - test if the given page is the head page |
| 635 | * |
| 636 | * Because the reader may move the head_page pointer, we can |
| 637 | * not trust what the head page is (it may be pointing to |
| 638 | * the reader page). But if the next page is a header page, |
| 639 | * its flags will be non zero. |
| 640 | */ |
| 641 | static inline int |
| 642 | rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer, |
| 643 | struct buffer_page *page, struct list_head *list) |
| 644 | { |
| 645 | unsigned long val; |
| 646 | |
| 647 | val = (unsigned long)list->next; |
| 648 | |
| 649 | if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list) |
| 650 | return RB_PAGE_MOVED; |
| 651 | |
| 652 | return val & RB_FLAG_MASK; |
| 653 | } |
| 654 | |
| 655 | /* |
| 656 | * rb_is_reader_page |
| 657 | * |
| 658 | * The unique thing about the reader page, is that, if the |
| 659 | * writer is ever on it, the previous pointer never points |
| 660 | * back to the reader page. |
| 661 | */ |
| 662 | static int rb_is_reader_page(struct buffer_page *page) |
| 663 | { |
| 664 | struct list_head *list = page->list.prev; |
| 665 | |
| 666 | return rb_list_head(list->next) != &page->list; |
| 667 | } |
| 668 | |
| 669 | /* |
| 670 | * rb_set_list_to_head - set a list_head to be pointing to head. |
| 671 | */ |
| 672 | static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer, |
| 673 | struct list_head *list) |
| 674 | { |
| 675 | unsigned long *ptr; |
| 676 | |
| 677 | ptr = (unsigned long *)&list->next; |
| 678 | *ptr |= RB_PAGE_HEAD; |
| 679 | *ptr &= ~RB_PAGE_UPDATE; |
| 680 | } |
| 681 | |
| 682 | /* |
| 683 | * rb_head_page_activate - sets up head page |
| 684 | */ |
| 685 | static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer) |
| 686 | { |
| 687 | struct buffer_page *head; |
| 688 | |
| 689 | head = cpu_buffer->head_page; |
| 690 | if (!head) |
| 691 | return; |
| 692 | |
| 693 | /* |
| 694 | * Set the previous list pointer to have the HEAD flag. |
| 695 | */ |
| 696 | rb_set_list_to_head(cpu_buffer, head->list.prev); |
| 697 | } |
| 698 | |
| 699 | static void rb_list_head_clear(struct list_head *list) |
| 700 | { |
| 701 | unsigned long *ptr = (unsigned long *)&list->next; |
| 702 | |
| 703 | *ptr &= ~RB_FLAG_MASK; |
| 704 | } |
| 705 | |
| 706 | /* |
| 707 | * rb_head_page_dactivate - clears head page ptr (for free list) |
| 708 | */ |
| 709 | static void |
| 710 | rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer) |
| 711 | { |
| 712 | struct list_head *hd; |
| 713 | |
| 714 | /* Go through the whole list and clear any pointers found. */ |
| 715 | rb_list_head_clear(cpu_buffer->pages); |
| 716 | |
| 717 | list_for_each(hd, cpu_buffer->pages) |
| 718 | rb_list_head_clear(hd); |
| 719 | } |
| 720 | |
| 721 | static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer, |
| 722 | struct buffer_page *head, |
| 723 | struct buffer_page *prev, |
| 724 | int old_flag, int new_flag) |
| 725 | { |
| 726 | struct list_head *list; |
| 727 | unsigned long val = (unsigned long)&head->list; |
| 728 | unsigned long ret; |
| 729 | |
| 730 | list = &prev->list; |
| 731 | |
| 732 | val &= ~RB_FLAG_MASK; |
| 733 | |
| 734 | ret = cmpxchg((unsigned long *)&list->next, |
| 735 | val | old_flag, val | new_flag); |
| 736 | |
| 737 | /* check if the reader took the page */ |
| 738 | if ((ret & ~RB_FLAG_MASK) != val) |
| 739 | return RB_PAGE_MOVED; |
| 740 | |
| 741 | return ret & RB_FLAG_MASK; |
| 742 | } |
| 743 | |
| 744 | static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer, |
| 745 | struct buffer_page *head, |
| 746 | struct buffer_page *prev, |
| 747 | int old_flag) |
| 748 | { |
| 749 | return rb_head_page_set(cpu_buffer, head, prev, |
| 750 | old_flag, RB_PAGE_UPDATE); |
| 751 | } |
| 752 | |
| 753 | static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer, |
| 754 | struct buffer_page *head, |
| 755 | struct buffer_page *prev, |
| 756 | int old_flag) |
| 757 | { |
| 758 | return rb_head_page_set(cpu_buffer, head, prev, |
| 759 | old_flag, RB_PAGE_HEAD); |
| 760 | } |
| 761 | |
| 762 | static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer, |
| 763 | struct buffer_page *head, |
| 764 | struct buffer_page *prev, |
| 765 | int old_flag) |
| 766 | { |
| 767 | return rb_head_page_set(cpu_buffer, head, prev, |
| 768 | old_flag, RB_PAGE_NORMAL); |
| 769 | } |
| 770 | |
| 771 | static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer, |
| 772 | struct buffer_page **bpage) |
| 773 | { |
| 774 | struct list_head *p = rb_list_head((*bpage)->list.next); |
| 775 | |
| 776 | *bpage = list_entry(p, struct buffer_page, list); |
| 777 | } |
| 778 | |
| 779 | static struct buffer_page * |
| 780 | rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer) |
| 781 | { |
| 782 | struct buffer_page *head; |
| 783 | struct buffer_page *page; |
| 784 | struct list_head *list; |
| 785 | int i; |
| 786 | |
| 787 | if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page)) |
| 788 | return NULL; |
| 789 | |
| 790 | /* sanity check */ |
| 791 | list = cpu_buffer->pages; |
| 792 | if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list)) |
| 793 | return NULL; |
| 794 | |
| 795 | page = head = cpu_buffer->head_page; |
| 796 | /* |
| 797 | * It is possible that the writer moves the header behind |
| 798 | * where we started, and we miss in one loop. |
| 799 | * A second loop should grab the header, but we'll do |
| 800 | * three loops just because I'm paranoid. |
| 801 | */ |
| 802 | for (i = 0; i < 3; i++) { |
| 803 | do { |
| 804 | if (rb_is_head_page(cpu_buffer, page, page->list.prev)) { |
| 805 | cpu_buffer->head_page = page; |
| 806 | return page; |
| 807 | } |
| 808 | rb_inc_page(cpu_buffer, &page); |
| 809 | } while (page != head); |
| 810 | } |
| 811 | |
| 812 | RB_WARN_ON(cpu_buffer, 1); |
| 813 | |
| 814 | return NULL; |
| 815 | } |
| 816 | |
| 817 | static int rb_head_page_replace(struct buffer_page *old, |
| 818 | struct buffer_page *new) |
| 819 | { |
| 820 | unsigned long *ptr = (unsigned long *)&old->list.prev->next; |
| 821 | unsigned long val; |
| 822 | unsigned long ret; |
| 823 | |
| 824 | val = *ptr & ~RB_FLAG_MASK; |
| 825 | val |= RB_PAGE_HEAD; |
| 826 | |
| 827 | ret = cmpxchg(ptr, val, (unsigned long)&new->list); |
| 828 | |
| 829 | return ret == val; |
| 830 | } |
| 831 | |
| 832 | /* |
| 833 | * rb_tail_page_update - move the tail page forward |
| 834 | * |
| 835 | * Returns 1 if moved tail page, 0 if someone else did. |
| 836 | */ |
| 837 | static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer, |
| 838 | struct buffer_page *tail_page, |
| 839 | struct buffer_page *next_page) |
| 840 | { |
| 841 | struct buffer_page *old_tail; |
| 842 | unsigned long old_entries; |
| 843 | unsigned long old_write; |
| 844 | int ret = 0; |
| 845 | |
| 846 | /* |
| 847 | * The tail page now needs to be moved forward. |
| 848 | * |
| 849 | * We need to reset the tail page, but without messing |
| 850 | * with possible erasing of data brought in by interrupts |
| 851 | * that have moved the tail page and are currently on it. |
| 852 | * |
| 853 | * We add a counter to the write field to denote this. |
| 854 | */ |
| 855 | old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write); |
| 856 | old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries); |
| 857 | |
| 858 | /* |
| 859 | * Just make sure we have seen our old_write and synchronize |
| 860 | * with any interrupts that come in. |
| 861 | */ |
| 862 | barrier(); |
| 863 | |
| 864 | /* |
| 865 | * If the tail page is still the same as what we think |
| 866 | * it is, then it is up to us to update the tail |
| 867 | * pointer. |
| 868 | */ |
| 869 | if (tail_page == cpu_buffer->tail_page) { |
| 870 | /* Zero the write counter */ |
| 871 | unsigned long val = old_write & ~RB_WRITE_MASK; |
| 872 | unsigned long eval = old_entries & ~RB_WRITE_MASK; |
| 873 | |
| 874 | /* |
| 875 | * This will only succeed if an interrupt did |
| 876 | * not come in and change it. In which case, we |
| 877 | * do not want to modify it. |
| 878 | * |
| 879 | * We add (void) to let the compiler know that we do not care |
| 880 | * about the return value of these functions. We use the |
| 881 | * cmpxchg to only update if an interrupt did not already |
| 882 | * do it for us. If the cmpxchg fails, we don't care. |
| 883 | */ |
| 884 | (void)local_cmpxchg(&next_page->write, old_write, val); |
| 885 | (void)local_cmpxchg(&next_page->entries, old_entries, eval); |
| 886 | |
| 887 | /* |
| 888 | * No need to worry about races with clearing out the commit. |
| 889 | * it only can increment when a commit takes place. But that |
| 890 | * only happens in the outer most nested commit. |
| 891 | */ |
| 892 | local_set(&next_page->page->commit, 0); |
| 893 | |
| 894 | old_tail = cmpxchg(&cpu_buffer->tail_page, |
| 895 | tail_page, next_page); |
| 896 | |
| 897 | if (old_tail == tail_page) |
| 898 | ret = 1; |
| 899 | } |
| 900 | |
| 901 | return ret; |
| 902 | } |
| 903 | |
| 904 | static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer, |
| 905 | struct buffer_page *bpage) |
| 906 | { |
| 907 | unsigned long val = (unsigned long)bpage; |
| 908 | |
| 909 | if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK)) |
| 910 | return 1; |
| 911 | |
| 912 | return 0; |
| 913 | } |
| 914 | |
| 915 | /** |
| 916 | * rb_check_list - make sure a pointer to a list has the last bits zero |
| 917 | */ |
| 918 | static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer, |
| 919 | struct list_head *list) |
| 920 | { |
| 921 | if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev)) |
| 922 | return 1; |
| 923 | if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next)) |
| 924 | return 1; |
| 925 | return 0; |
| 926 | } |
| 927 | |
| 928 | /** |
| 929 | * check_pages - integrity check of buffer pages |
| 930 | * @cpu_buffer: CPU buffer with pages to test |
| 931 | * |
| 932 | * As a safety measure we check to make sure the data pages have not |
| 933 | * been corrupted. |
| 934 | */ |
| 935 | static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer) |
| 936 | { |
| 937 | struct list_head *head = cpu_buffer->pages; |
| 938 | struct buffer_page *bpage, *tmp; |
| 939 | |
| 940 | rb_head_page_deactivate(cpu_buffer); |
| 941 | |
| 942 | if (RB_WARN_ON(cpu_buffer, head->next->prev != head)) |
| 943 | return -1; |
| 944 | if (RB_WARN_ON(cpu_buffer, head->prev->next != head)) |
| 945 | return -1; |
| 946 | |
| 947 | if (rb_check_list(cpu_buffer, head)) |
| 948 | return -1; |
| 949 | |
| 950 | list_for_each_entry_safe(bpage, tmp, head, list) { |
| 951 | if (RB_WARN_ON(cpu_buffer, |
| 952 | bpage->list.next->prev != &bpage->list)) |
| 953 | return -1; |
| 954 | if (RB_WARN_ON(cpu_buffer, |
| 955 | bpage->list.prev->next != &bpage->list)) |
| 956 | return -1; |
| 957 | if (rb_check_list(cpu_buffer, &bpage->list)) |
| 958 | return -1; |
| 959 | } |
| 960 | |
| 961 | rb_head_page_activate(cpu_buffer); |
| 962 | |
| 963 | return 0; |
| 964 | } |
| 965 | |
| 966 | static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer, |
| 967 | unsigned nr_pages) |
| 968 | { |
| 969 | struct buffer_page *bpage, *tmp; |
| 970 | LIST_HEAD(pages); |
| 971 | unsigned i; |
| 972 | |
| 973 | WARN_ON(!nr_pages); |
| 974 | |
| 975 | for (i = 0; i < nr_pages; i++) { |
| 976 | struct page *page; |
| 977 | /* |
| 978 | * __GFP_NORETRY flag makes sure that the allocation fails |
| 979 | * gracefully without invoking oom-killer and the system is |
| 980 | * not destabilized. |
| 981 | */ |
| 982 | bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()), |
| 983 | GFP_KERNEL | __GFP_NORETRY, |
| 984 | cpu_to_node(cpu_buffer->cpu)); |
| 985 | if (!bpage) |
| 986 | goto free_pages; |
| 987 | |
| 988 | rb_check_bpage(cpu_buffer, bpage); |
| 989 | |
| 990 | list_add(&bpage->list, &pages); |
| 991 | |
| 992 | page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), |
| 993 | GFP_KERNEL | __GFP_NORETRY, 0); |
| 994 | if (!page) |
| 995 | goto free_pages; |
| 996 | bpage->page = page_address(page); |
| 997 | rb_init_page(bpage->page); |
| 998 | } |
| 999 | |
| 1000 | /* |
| 1001 | * The ring buffer page list is a circular list that does not |
| 1002 | * start and end with a list head. All page list items point to |
| 1003 | * other pages. |
| 1004 | */ |
| 1005 | cpu_buffer->pages = pages.next; |
| 1006 | list_del(&pages); |
| 1007 | |
| 1008 | rb_check_pages(cpu_buffer); |
| 1009 | |
| 1010 | return 0; |
| 1011 | |
| 1012 | free_pages: |
| 1013 | list_for_each_entry_safe(bpage, tmp, &pages, list) { |
| 1014 | list_del_init(&bpage->list); |
| 1015 | free_buffer_page(bpage); |
| 1016 | } |
| 1017 | return -ENOMEM; |
| 1018 | } |
| 1019 | |
| 1020 | static inline int ok_to_lock(void) |
| 1021 | { |
| 1022 | if (in_nmi()) |
| 1023 | return 0; |
| 1024 | #ifdef CONFIG_PREEMPT_RT_FULL |
| 1025 | if (in_atomic() || irqs_disabled()) |
| 1026 | return 0; |
| 1027 | #endif |
| 1028 | return 1; |
| 1029 | } |
| 1030 | |
| 1031 | static int |
| 1032 | read_buffer_lock(struct ring_buffer_per_cpu *cpu_buffer, |
| 1033 | unsigned long *flags) |
| 1034 | { |
| 1035 | /* |
| 1036 | * If an NMI die dumps out the content of the ring buffer |
| 1037 | * do not grab locks. We also permanently disable the ring |
| 1038 | * buffer too. A one time deal is all you get from reading |
| 1039 | * the ring buffer from an NMI. |
| 1040 | */ |
| 1041 | if (!ok_to_lock()) { |
| 1042 | if (spin_trylock_irqsave(&cpu_buffer->reader_lock, *flags)) |
| 1043 | return 1; |
| 1044 | tracing_off_permanent(); |
| 1045 | return 0; |
| 1046 | } |
| 1047 | spin_lock_irqsave(&cpu_buffer->reader_lock, *flags); |
| 1048 | return 1; |
| 1049 | } |
| 1050 | |
| 1051 | static void |
| 1052 | read_buffer_unlock(struct ring_buffer_per_cpu *cpu_buffer, |
| 1053 | unsigned long flags, int locked) |
| 1054 | { |
| 1055 | if (locked) |
| 1056 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); |
| 1057 | } |
| 1058 | static struct ring_buffer_per_cpu * |
| 1059 | rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu) |
| 1060 | { |
| 1061 | struct ring_buffer_per_cpu *cpu_buffer; |
| 1062 | struct buffer_page *bpage; |
| 1063 | struct page *page; |
| 1064 | int ret; |
| 1065 | |
| 1066 | cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()), |
| 1067 | GFP_KERNEL, cpu_to_node(cpu)); |
| 1068 | if (!cpu_buffer) |
| 1069 | return NULL; |
| 1070 | |
| 1071 | cpu_buffer->cpu = cpu; |
| 1072 | cpu_buffer->buffer = buffer; |
| 1073 | spin_lock_init(&cpu_buffer->reader_lock); |
| 1074 | lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key); |
| 1075 | cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED; |
| 1076 | |
| 1077 | bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()), |
| 1078 | GFP_KERNEL, cpu_to_node(cpu)); |
| 1079 | if (!bpage) |
| 1080 | goto fail_free_buffer; |
| 1081 | |
| 1082 | rb_check_bpage(cpu_buffer, bpage); |
| 1083 | |
| 1084 | cpu_buffer->reader_page = bpage; |
| 1085 | page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0); |
| 1086 | if (!page) |
| 1087 | goto fail_free_reader; |
| 1088 | bpage->page = page_address(page); |
| 1089 | rb_init_page(bpage->page); |
| 1090 | |
| 1091 | INIT_LIST_HEAD(&cpu_buffer->reader_page->list); |
| 1092 | |
| 1093 | ret = rb_allocate_pages(cpu_buffer, buffer->pages); |
| 1094 | if (ret < 0) |
| 1095 | goto fail_free_reader; |
| 1096 | |
| 1097 | cpu_buffer->head_page |
| 1098 | = list_entry(cpu_buffer->pages, struct buffer_page, list); |
| 1099 | cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page; |
| 1100 | |
| 1101 | rb_head_page_activate(cpu_buffer); |
| 1102 | |
| 1103 | return cpu_buffer; |
| 1104 | |
| 1105 | fail_free_reader: |
| 1106 | free_buffer_page(cpu_buffer->reader_page); |
| 1107 | |
| 1108 | fail_free_buffer: |
| 1109 | kfree(cpu_buffer); |
| 1110 | return NULL; |
| 1111 | } |
| 1112 | |
| 1113 | static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer) |
| 1114 | { |
| 1115 | struct list_head *head = cpu_buffer->pages; |
| 1116 | struct buffer_page *bpage, *tmp; |
| 1117 | |
| 1118 | free_buffer_page(cpu_buffer->reader_page); |
| 1119 | |
| 1120 | rb_head_page_deactivate(cpu_buffer); |
| 1121 | |
| 1122 | if (head) { |
| 1123 | list_for_each_entry_safe(bpage, tmp, head, list) { |
| 1124 | list_del_init(&bpage->list); |
| 1125 | free_buffer_page(bpage); |
| 1126 | } |
| 1127 | bpage = list_entry(head, struct buffer_page, list); |
| 1128 | free_buffer_page(bpage); |
| 1129 | } |
| 1130 | |
| 1131 | kfree(cpu_buffer); |
| 1132 | } |
| 1133 | |
| 1134 | #ifdef CONFIG_HOTPLUG_CPU |
| 1135 | static int rb_cpu_notify(struct notifier_block *self, |
| 1136 | unsigned long action, void *hcpu); |
| 1137 | #endif |
| 1138 | |
| 1139 | /** |
| 1140 | * ring_buffer_alloc - allocate a new ring_buffer |
| 1141 | * @size: the size in bytes per cpu that is needed. |
| 1142 | * @flags: attributes to set for the ring buffer. |
| 1143 | * |
| 1144 | * Currently the only flag that is available is the RB_FL_OVERWRITE |
| 1145 | * flag. This flag means that the buffer will overwrite old data |
| 1146 | * when the buffer wraps. If this flag is not set, the buffer will |
| 1147 | * drop data when the tail hits the head. |
| 1148 | */ |
| 1149 | struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags, |
| 1150 | struct lock_class_key *key) |
| 1151 | { |
| 1152 | struct ring_buffer *buffer; |
| 1153 | int bsize; |
| 1154 | int cpu; |
| 1155 | |
| 1156 | /* keep it in its own cache line */ |
| 1157 | buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()), |
| 1158 | GFP_KERNEL); |
| 1159 | if (!buffer) |
| 1160 | return NULL; |
| 1161 | |
| 1162 | if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL)) |
| 1163 | goto fail_free_buffer; |
| 1164 | |
| 1165 | buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE); |
| 1166 | buffer->flags = flags; |
| 1167 | buffer->clock = trace_clock_local; |
| 1168 | buffer->reader_lock_key = key; |
| 1169 | |
| 1170 | /* need at least two pages */ |
| 1171 | if (buffer->pages < 2) |
| 1172 | buffer->pages = 2; |
| 1173 | |
| 1174 | /* |
| 1175 | * In case of non-hotplug cpu, if the ring-buffer is allocated |
| 1176 | * in early initcall, it will not be notified of secondary cpus. |
| 1177 | * In that off case, we need to allocate for all possible cpus. |
| 1178 | */ |
| 1179 | #ifdef CONFIG_HOTPLUG_CPU |
| 1180 | get_online_cpus(); |
| 1181 | cpumask_copy(buffer->cpumask, cpu_online_mask); |
| 1182 | #else |
| 1183 | cpumask_copy(buffer->cpumask, cpu_possible_mask); |
| 1184 | #endif |
| 1185 | buffer->cpus = nr_cpu_ids; |
| 1186 | |
| 1187 | bsize = sizeof(void *) * nr_cpu_ids; |
| 1188 | buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()), |
| 1189 | GFP_KERNEL); |
| 1190 | if (!buffer->buffers) |
| 1191 | goto fail_free_cpumask; |
| 1192 | |
| 1193 | for_each_buffer_cpu(buffer, cpu) { |
| 1194 | buffer->buffers[cpu] = |
| 1195 | rb_allocate_cpu_buffer(buffer, cpu); |
| 1196 | if (!buffer->buffers[cpu]) |
| 1197 | goto fail_free_buffers; |
| 1198 | } |
| 1199 | |
| 1200 | #ifdef CONFIG_HOTPLUG_CPU |
| 1201 | buffer->cpu_notify.notifier_call = rb_cpu_notify; |
| 1202 | buffer->cpu_notify.priority = 0; |
| 1203 | register_cpu_notifier(&buffer->cpu_notify); |
| 1204 | #endif |
| 1205 | |
| 1206 | put_online_cpus(); |
| 1207 | mutex_init(&buffer->mutex); |
| 1208 | |
| 1209 | return buffer; |
| 1210 | |
| 1211 | fail_free_buffers: |
| 1212 | for_each_buffer_cpu(buffer, cpu) { |
| 1213 | if (buffer->buffers[cpu]) |
| 1214 | rb_free_cpu_buffer(buffer->buffers[cpu]); |
| 1215 | } |
| 1216 | kfree(buffer->buffers); |
| 1217 | |
| 1218 | fail_free_cpumask: |
| 1219 | free_cpumask_var(buffer->cpumask); |
| 1220 | put_online_cpus(); |
| 1221 | |
| 1222 | fail_free_buffer: |
| 1223 | kfree(buffer); |
| 1224 | return NULL; |
| 1225 | } |
| 1226 | EXPORT_SYMBOL_GPL(__ring_buffer_alloc); |
| 1227 | |
| 1228 | /** |
| 1229 | * ring_buffer_free - free a ring buffer. |
| 1230 | * @buffer: the buffer to free. |
| 1231 | */ |
| 1232 | void |
| 1233 | ring_buffer_free(struct ring_buffer *buffer) |
| 1234 | { |
| 1235 | int cpu; |
| 1236 | |
| 1237 | get_online_cpus(); |
| 1238 | |
| 1239 | #ifdef CONFIG_HOTPLUG_CPU |
| 1240 | unregister_cpu_notifier(&buffer->cpu_notify); |
| 1241 | #endif |
| 1242 | |
| 1243 | for_each_buffer_cpu(buffer, cpu) |
| 1244 | rb_free_cpu_buffer(buffer->buffers[cpu]); |
| 1245 | |
| 1246 | put_online_cpus(); |
| 1247 | |
| 1248 | kfree(buffer->buffers); |
| 1249 | free_cpumask_var(buffer->cpumask); |
| 1250 | |
| 1251 | kfree(buffer); |
| 1252 | } |
| 1253 | EXPORT_SYMBOL_GPL(ring_buffer_free); |
| 1254 | |
| 1255 | void ring_buffer_set_clock(struct ring_buffer *buffer, |
| 1256 | u64 (*clock)(void)) |
| 1257 | { |
| 1258 | buffer->clock = clock; |
| 1259 | } |
| 1260 | |
| 1261 | static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer); |
| 1262 | |
| 1263 | static void |
| 1264 | rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages) |
| 1265 | { |
| 1266 | struct buffer_page *bpage; |
| 1267 | struct list_head *p; |
| 1268 | unsigned long flags; |
| 1269 | unsigned i; |
| 1270 | int locked; |
| 1271 | |
| 1272 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 1273 | rb_head_page_deactivate(cpu_buffer); |
| 1274 | |
| 1275 | for (i = 0; i < nr_pages; i++) { |
| 1276 | if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages))) |
| 1277 | goto out; |
| 1278 | p = cpu_buffer->pages->next; |
| 1279 | bpage = list_entry(p, struct buffer_page, list); |
| 1280 | list_del_init(&bpage->list); |
| 1281 | free_buffer_page(bpage); |
| 1282 | } |
| 1283 | if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages))) |
| 1284 | goto out; |
| 1285 | |
| 1286 | rb_reset_cpu(cpu_buffer); |
| 1287 | rb_check_pages(cpu_buffer); |
| 1288 | |
| 1289 | out: |
| 1290 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 1291 | } |
| 1292 | |
| 1293 | static void |
| 1294 | rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer, |
| 1295 | struct list_head *pages, unsigned nr_pages) |
| 1296 | { |
| 1297 | struct buffer_page *bpage; |
| 1298 | struct list_head *p; |
| 1299 | unsigned long flags; |
| 1300 | unsigned i; |
| 1301 | int locked; |
| 1302 | |
| 1303 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 1304 | rb_head_page_deactivate(cpu_buffer); |
| 1305 | |
| 1306 | for (i = 0; i < nr_pages; i++) { |
| 1307 | if (RB_WARN_ON(cpu_buffer, list_empty(pages))) |
| 1308 | goto out; |
| 1309 | p = pages->next; |
| 1310 | bpage = list_entry(p, struct buffer_page, list); |
| 1311 | list_del_init(&bpage->list); |
| 1312 | list_add_tail(&bpage->list, cpu_buffer->pages); |
| 1313 | } |
| 1314 | rb_reset_cpu(cpu_buffer); |
| 1315 | rb_check_pages(cpu_buffer); |
| 1316 | |
| 1317 | out: |
| 1318 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 1319 | } |
| 1320 | |
| 1321 | /** |
| 1322 | * ring_buffer_resize - resize the ring buffer |
| 1323 | * @buffer: the buffer to resize. |
| 1324 | * @size: the new size. |
| 1325 | * |
| 1326 | * Minimum size is 2 * BUF_PAGE_SIZE. |
| 1327 | * |
| 1328 | * Returns -1 on failure. |
| 1329 | */ |
| 1330 | int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size) |
| 1331 | { |
| 1332 | struct ring_buffer_per_cpu *cpu_buffer; |
| 1333 | unsigned nr_pages, rm_pages, new_pages; |
| 1334 | struct buffer_page *bpage, *tmp; |
| 1335 | unsigned long buffer_size; |
| 1336 | LIST_HEAD(pages); |
| 1337 | int i, cpu; |
| 1338 | |
| 1339 | /* |
| 1340 | * Always succeed at resizing a non-existent buffer: |
| 1341 | */ |
| 1342 | if (!buffer) |
| 1343 | return size; |
| 1344 | |
| 1345 | size = DIV_ROUND_UP(size, BUF_PAGE_SIZE); |
| 1346 | size *= BUF_PAGE_SIZE; |
| 1347 | buffer_size = buffer->pages * BUF_PAGE_SIZE; |
| 1348 | |
| 1349 | /* we need a minimum of two pages */ |
| 1350 | if (size < BUF_PAGE_SIZE * 2) |
| 1351 | size = BUF_PAGE_SIZE * 2; |
| 1352 | |
| 1353 | if (size == buffer_size) |
| 1354 | return size; |
| 1355 | |
| 1356 | atomic_inc(&buffer->record_disabled); |
| 1357 | |
| 1358 | /* Make sure all writers are done with this buffer. */ |
| 1359 | synchronize_sched(); |
| 1360 | |
| 1361 | mutex_lock(&buffer->mutex); |
| 1362 | get_online_cpus(); |
| 1363 | |
| 1364 | nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE); |
| 1365 | |
| 1366 | if (size < buffer_size) { |
| 1367 | |
| 1368 | /* easy case, just free pages */ |
| 1369 | if (RB_WARN_ON(buffer, nr_pages >= buffer->pages)) |
| 1370 | goto out_fail; |
| 1371 | |
| 1372 | rm_pages = buffer->pages - nr_pages; |
| 1373 | |
| 1374 | for_each_buffer_cpu(buffer, cpu) { |
| 1375 | cpu_buffer = buffer->buffers[cpu]; |
| 1376 | rb_remove_pages(cpu_buffer, rm_pages); |
| 1377 | } |
| 1378 | goto out; |
| 1379 | } |
| 1380 | |
| 1381 | /* |
| 1382 | * This is a bit more difficult. We only want to add pages |
| 1383 | * when we can allocate enough for all CPUs. We do this |
| 1384 | * by allocating all the pages and storing them on a local |
| 1385 | * link list. If we succeed in our allocation, then we |
| 1386 | * add these pages to the cpu_buffers. Otherwise we just free |
| 1387 | * them all and return -ENOMEM; |
| 1388 | */ |
| 1389 | if (RB_WARN_ON(buffer, nr_pages <= buffer->pages)) |
| 1390 | goto out_fail; |
| 1391 | |
| 1392 | new_pages = nr_pages - buffer->pages; |
| 1393 | |
| 1394 | for_each_buffer_cpu(buffer, cpu) { |
| 1395 | for (i = 0; i < new_pages; i++) { |
| 1396 | struct page *page; |
| 1397 | /* |
| 1398 | * __GFP_NORETRY flag makes sure that the allocation |
| 1399 | * fails gracefully without invoking oom-killer and |
| 1400 | * the system is not destabilized. |
| 1401 | */ |
| 1402 | bpage = kzalloc_node(ALIGN(sizeof(*bpage), |
| 1403 | cache_line_size()), |
| 1404 | GFP_KERNEL | __GFP_NORETRY, |
| 1405 | cpu_to_node(cpu)); |
| 1406 | if (!bpage) |
| 1407 | goto free_pages; |
| 1408 | list_add(&bpage->list, &pages); |
| 1409 | page = alloc_pages_node(cpu_to_node(cpu), |
| 1410 | GFP_KERNEL | __GFP_NORETRY, 0); |
| 1411 | if (!page) |
| 1412 | goto free_pages; |
| 1413 | bpage->page = page_address(page); |
| 1414 | rb_init_page(bpage->page); |
| 1415 | } |
| 1416 | } |
| 1417 | |
| 1418 | for_each_buffer_cpu(buffer, cpu) { |
| 1419 | cpu_buffer = buffer->buffers[cpu]; |
| 1420 | rb_insert_pages(cpu_buffer, &pages, new_pages); |
| 1421 | } |
| 1422 | |
| 1423 | if (RB_WARN_ON(buffer, !list_empty(&pages))) |
| 1424 | goto out_fail; |
| 1425 | |
| 1426 | out: |
| 1427 | buffer->pages = nr_pages; |
| 1428 | put_online_cpus(); |
| 1429 | mutex_unlock(&buffer->mutex); |
| 1430 | |
| 1431 | atomic_dec(&buffer->record_disabled); |
| 1432 | |
| 1433 | return size; |
| 1434 | |
| 1435 | free_pages: |
| 1436 | list_for_each_entry_safe(bpage, tmp, &pages, list) { |
| 1437 | list_del_init(&bpage->list); |
| 1438 | free_buffer_page(bpage); |
| 1439 | } |
| 1440 | put_online_cpus(); |
| 1441 | mutex_unlock(&buffer->mutex); |
| 1442 | atomic_dec(&buffer->record_disabled); |
| 1443 | return -ENOMEM; |
| 1444 | |
| 1445 | /* |
| 1446 | * Something went totally wrong, and we are too paranoid |
| 1447 | * to even clean up the mess. |
| 1448 | */ |
| 1449 | out_fail: |
| 1450 | put_online_cpus(); |
| 1451 | mutex_unlock(&buffer->mutex); |
| 1452 | atomic_dec(&buffer->record_disabled); |
| 1453 | return -1; |
| 1454 | } |
| 1455 | EXPORT_SYMBOL_GPL(ring_buffer_resize); |
| 1456 | |
| 1457 | void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val) |
| 1458 | { |
| 1459 | mutex_lock(&buffer->mutex); |
| 1460 | if (val) |
| 1461 | buffer->flags |= RB_FL_OVERWRITE; |
| 1462 | else |
| 1463 | buffer->flags &= ~RB_FL_OVERWRITE; |
| 1464 | mutex_unlock(&buffer->mutex); |
| 1465 | } |
| 1466 | EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite); |
| 1467 | |
| 1468 | static inline void * |
| 1469 | __rb_data_page_index(struct buffer_data_page *bpage, unsigned index) |
| 1470 | { |
| 1471 | return bpage->data + index; |
| 1472 | } |
| 1473 | |
| 1474 | static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index) |
| 1475 | { |
| 1476 | return bpage->page->data + index; |
| 1477 | } |
| 1478 | |
| 1479 | static inline struct ring_buffer_event * |
| 1480 | rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer) |
| 1481 | { |
| 1482 | return __rb_page_index(cpu_buffer->reader_page, |
| 1483 | cpu_buffer->reader_page->read); |
| 1484 | } |
| 1485 | |
| 1486 | static inline struct ring_buffer_event * |
| 1487 | rb_iter_head_event(struct ring_buffer_iter *iter) |
| 1488 | { |
| 1489 | return __rb_page_index(iter->head_page, iter->head); |
| 1490 | } |
| 1491 | |
| 1492 | static inline unsigned long rb_page_write(struct buffer_page *bpage) |
| 1493 | { |
| 1494 | return local_read(&bpage->write) & RB_WRITE_MASK; |
| 1495 | } |
| 1496 | |
| 1497 | static inline unsigned rb_page_commit(struct buffer_page *bpage) |
| 1498 | { |
| 1499 | return local_read(&bpage->page->commit); |
| 1500 | } |
| 1501 | |
| 1502 | static inline unsigned long rb_page_entries(struct buffer_page *bpage) |
| 1503 | { |
| 1504 | return local_read(&bpage->entries) & RB_WRITE_MASK; |
| 1505 | } |
| 1506 | |
| 1507 | /* Size is determined by what has been committed */ |
| 1508 | static inline unsigned rb_page_size(struct buffer_page *bpage) |
| 1509 | { |
| 1510 | return rb_page_commit(bpage); |
| 1511 | } |
| 1512 | |
| 1513 | static inline unsigned |
| 1514 | rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer) |
| 1515 | { |
| 1516 | return rb_page_commit(cpu_buffer->commit_page); |
| 1517 | } |
| 1518 | |
| 1519 | static inline unsigned |
| 1520 | rb_event_index(struct ring_buffer_event *event) |
| 1521 | { |
| 1522 | unsigned long addr = (unsigned long)event; |
| 1523 | |
| 1524 | return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE; |
| 1525 | } |
| 1526 | |
| 1527 | static inline int |
| 1528 | rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer, |
| 1529 | struct ring_buffer_event *event) |
| 1530 | { |
| 1531 | unsigned long addr = (unsigned long)event; |
| 1532 | unsigned long index; |
| 1533 | |
| 1534 | index = rb_event_index(event); |
| 1535 | addr &= PAGE_MASK; |
| 1536 | |
| 1537 | return cpu_buffer->commit_page->page == (void *)addr && |
| 1538 | rb_commit_index(cpu_buffer) == index; |
| 1539 | } |
| 1540 | |
| 1541 | static void |
| 1542 | rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer) |
| 1543 | { |
| 1544 | unsigned long max_count; |
| 1545 | |
| 1546 | /* |
| 1547 | * We only race with interrupts and NMIs on this CPU. |
| 1548 | * If we own the commit event, then we can commit |
| 1549 | * all others that interrupted us, since the interruptions |
| 1550 | * are in stack format (they finish before they come |
| 1551 | * back to us). This allows us to do a simple loop to |
| 1552 | * assign the commit to the tail. |
| 1553 | */ |
| 1554 | again: |
| 1555 | max_count = cpu_buffer->buffer->pages * 100; |
| 1556 | |
| 1557 | while (cpu_buffer->commit_page != cpu_buffer->tail_page) { |
| 1558 | if (RB_WARN_ON(cpu_buffer, !(--max_count))) |
| 1559 | return; |
| 1560 | if (RB_WARN_ON(cpu_buffer, |
| 1561 | rb_is_reader_page(cpu_buffer->tail_page))) |
| 1562 | return; |
| 1563 | local_set(&cpu_buffer->commit_page->page->commit, |
| 1564 | rb_page_write(cpu_buffer->commit_page)); |
| 1565 | rb_inc_page(cpu_buffer, &cpu_buffer->commit_page); |
| 1566 | cpu_buffer->write_stamp = |
| 1567 | cpu_buffer->commit_page->page->time_stamp; |
| 1568 | /* add barrier to keep gcc from optimizing too much */ |
| 1569 | barrier(); |
| 1570 | } |
| 1571 | while (rb_commit_index(cpu_buffer) != |
| 1572 | rb_page_write(cpu_buffer->commit_page)) { |
| 1573 | |
| 1574 | local_set(&cpu_buffer->commit_page->page->commit, |
| 1575 | rb_page_write(cpu_buffer->commit_page)); |
| 1576 | RB_WARN_ON(cpu_buffer, |
| 1577 | local_read(&cpu_buffer->commit_page->page->commit) & |
| 1578 | ~RB_WRITE_MASK); |
| 1579 | barrier(); |
| 1580 | } |
| 1581 | |
| 1582 | /* again, keep gcc from optimizing */ |
| 1583 | barrier(); |
| 1584 | |
| 1585 | /* |
| 1586 | * If an interrupt came in just after the first while loop |
| 1587 | * and pushed the tail page forward, we will be left with |
| 1588 | * a dangling commit that will never go forward. |
| 1589 | */ |
| 1590 | if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page)) |
| 1591 | goto again; |
| 1592 | } |
| 1593 | |
| 1594 | static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer) |
| 1595 | { |
| 1596 | cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp; |
| 1597 | cpu_buffer->reader_page->read = 0; |
| 1598 | } |
| 1599 | |
| 1600 | static void rb_inc_iter(struct ring_buffer_iter *iter) |
| 1601 | { |
| 1602 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; |
| 1603 | |
| 1604 | /* |
| 1605 | * The iterator could be on the reader page (it starts there). |
| 1606 | * But the head could have moved, since the reader was |
| 1607 | * found. Check for this case and assign the iterator |
| 1608 | * to the head page instead of next. |
| 1609 | */ |
| 1610 | if (iter->head_page == cpu_buffer->reader_page) |
| 1611 | iter->head_page = rb_set_head_page(cpu_buffer); |
| 1612 | else |
| 1613 | rb_inc_page(cpu_buffer, &iter->head_page); |
| 1614 | |
| 1615 | iter->read_stamp = iter->head_page->page->time_stamp; |
| 1616 | iter->head = 0; |
| 1617 | } |
| 1618 | |
| 1619 | /* Slow path, do not inline */ |
| 1620 | static noinline struct ring_buffer_event * |
| 1621 | rb_add_time_stamp(struct ring_buffer_event *event, u64 delta) |
| 1622 | { |
| 1623 | event->type_len = RINGBUF_TYPE_TIME_EXTEND; |
| 1624 | |
| 1625 | /* Not the first event on the page? */ |
| 1626 | if (rb_event_index(event)) { |
| 1627 | event->time_delta = delta & TS_MASK; |
| 1628 | event->array[0] = delta >> TS_SHIFT; |
| 1629 | } else { |
| 1630 | /* nope, just zero it */ |
| 1631 | event->time_delta = 0; |
| 1632 | event->array[0] = 0; |
| 1633 | } |
| 1634 | |
| 1635 | return skip_time_extend(event); |
| 1636 | } |
| 1637 | |
| 1638 | /** |
| 1639 | * ring_buffer_update_event - update event type and data |
| 1640 | * @event: the even to update |
| 1641 | * @type: the type of event |
| 1642 | * @length: the size of the event field in the ring buffer |
| 1643 | * |
| 1644 | * Update the type and data fields of the event. The length |
| 1645 | * is the actual size that is written to the ring buffer, |
| 1646 | * and with this, we can determine what to place into the |
| 1647 | * data field. |
| 1648 | */ |
| 1649 | static void |
| 1650 | rb_update_event(struct ring_buffer_per_cpu *cpu_buffer, |
| 1651 | struct ring_buffer_event *event, unsigned length, |
| 1652 | int add_timestamp, u64 delta) |
| 1653 | { |
| 1654 | /* Only a commit updates the timestamp */ |
| 1655 | if (unlikely(!rb_event_is_commit(cpu_buffer, event))) |
| 1656 | delta = 0; |
| 1657 | |
| 1658 | /* |
| 1659 | * If we need to add a timestamp, then we |
| 1660 | * add it to the start of the resevered space. |
| 1661 | */ |
| 1662 | if (unlikely(add_timestamp)) { |
| 1663 | event = rb_add_time_stamp(event, delta); |
| 1664 | length -= RB_LEN_TIME_EXTEND; |
| 1665 | delta = 0; |
| 1666 | } |
| 1667 | |
| 1668 | event->time_delta = delta; |
| 1669 | length -= RB_EVNT_HDR_SIZE; |
| 1670 | if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) { |
| 1671 | event->type_len = 0; |
| 1672 | event->array[0] = length; |
| 1673 | } else |
| 1674 | event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT); |
| 1675 | } |
| 1676 | |
| 1677 | /* |
| 1678 | * rb_handle_head_page - writer hit the head page |
| 1679 | * |
| 1680 | * Returns: +1 to retry page |
| 1681 | * 0 to continue |
| 1682 | * -1 on error |
| 1683 | */ |
| 1684 | static int |
| 1685 | rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer, |
| 1686 | struct buffer_page *tail_page, |
| 1687 | struct buffer_page *next_page) |
| 1688 | { |
| 1689 | struct buffer_page *new_head; |
| 1690 | int entries; |
| 1691 | int type; |
| 1692 | int ret; |
| 1693 | |
| 1694 | entries = rb_page_entries(next_page); |
| 1695 | |
| 1696 | /* |
| 1697 | * The hard part is here. We need to move the head |
| 1698 | * forward, and protect against both readers on |
| 1699 | * other CPUs and writers coming in via interrupts. |
| 1700 | */ |
| 1701 | type = rb_head_page_set_update(cpu_buffer, next_page, tail_page, |
| 1702 | RB_PAGE_HEAD); |
| 1703 | |
| 1704 | /* |
| 1705 | * type can be one of four: |
| 1706 | * NORMAL - an interrupt already moved it for us |
| 1707 | * HEAD - we are the first to get here. |
| 1708 | * UPDATE - we are the interrupt interrupting |
| 1709 | * a current move. |
| 1710 | * MOVED - a reader on another CPU moved the next |
| 1711 | * pointer to its reader page. Give up |
| 1712 | * and try again. |
| 1713 | */ |
| 1714 | |
| 1715 | switch (type) { |
| 1716 | case RB_PAGE_HEAD: |
| 1717 | /* |
| 1718 | * We changed the head to UPDATE, thus |
| 1719 | * it is our responsibility to update |
| 1720 | * the counters. |
| 1721 | */ |
| 1722 | local_add(entries, &cpu_buffer->overrun); |
| 1723 | local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes); |
| 1724 | |
| 1725 | /* |
| 1726 | * The entries will be zeroed out when we move the |
| 1727 | * tail page. |
| 1728 | */ |
| 1729 | |
| 1730 | /* still more to do */ |
| 1731 | break; |
| 1732 | |
| 1733 | case RB_PAGE_UPDATE: |
| 1734 | /* |
| 1735 | * This is an interrupt that interrupt the |
| 1736 | * previous update. Still more to do. |
| 1737 | */ |
| 1738 | break; |
| 1739 | case RB_PAGE_NORMAL: |
| 1740 | /* |
| 1741 | * An interrupt came in before the update |
| 1742 | * and processed this for us. |
| 1743 | * Nothing left to do. |
| 1744 | */ |
| 1745 | return 1; |
| 1746 | case RB_PAGE_MOVED: |
| 1747 | /* |
| 1748 | * The reader is on another CPU and just did |
| 1749 | * a swap with our next_page. |
| 1750 | * Try again. |
| 1751 | */ |
| 1752 | return 1; |
| 1753 | default: |
| 1754 | RB_WARN_ON(cpu_buffer, 1); /* WTF??? */ |
| 1755 | return -1; |
| 1756 | } |
| 1757 | |
| 1758 | /* |
| 1759 | * Now that we are here, the old head pointer is |
| 1760 | * set to UPDATE. This will keep the reader from |
| 1761 | * swapping the head page with the reader page. |
| 1762 | * The reader (on another CPU) will spin till |
| 1763 | * we are finished. |
| 1764 | * |
| 1765 | * We just need to protect against interrupts |
| 1766 | * doing the job. We will set the next pointer |
| 1767 | * to HEAD. After that, we set the old pointer |
| 1768 | * to NORMAL, but only if it was HEAD before. |
| 1769 | * otherwise we are an interrupt, and only |
| 1770 | * want the outer most commit to reset it. |
| 1771 | */ |
| 1772 | new_head = next_page; |
| 1773 | rb_inc_page(cpu_buffer, &new_head); |
| 1774 | |
| 1775 | ret = rb_head_page_set_head(cpu_buffer, new_head, next_page, |
| 1776 | RB_PAGE_NORMAL); |
| 1777 | |
| 1778 | /* |
| 1779 | * Valid returns are: |
| 1780 | * HEAD - an interrupt came in and already set it. |
| 1781 | * NORMAL - One of two things: |
| 1782 | * 1) We really set it. |
| 1783 | * 2) A bunch of interrupts came in and moved |
| 1784 | * the page forward again. |
| 1785 | */ |
| 1786 | switch (ret) { |
| 1787 | case RB_PAGE_HEAD: |
| 1788 | case RB_PAGE_NORMAL: |
| 1789 | /* OK */ |
| 1790 | break; |
| 1791 | default: |
| 1792 | RB_WARN_ON(cpu_buffer, 1); |
| 1793 | return -1; |
| 1794 | } |
| 1795 | |
| 1796 | /* |
| 1797 | * It is possible that an interrupt came in, |
| 1798 | * set the head up, then more interrupts came in |
| 1799 | * and moved it again. When we get back here, |
| 1800 | * the page would have been set to NORMAL but we |
| 1801 | * just set it back to HEAD. |
| 1802 | * |
| 1803 | * How do you detect this? Well, if that happened |
| 1804 | * the tail page would have moved. |
| 1805 | */ |
| 1806 | if (ret == RB_PAGE_NORMAL) { |
| 1807 | /* |
| 1808 | * If the tail had moved passed next, then we need |
| 1809 | * to reset the pointer. |
| 1810 | */ |
| 1811 | if (cpu_buffer->tail_page != tail_page && |
| 1812 | cpu_buffer->tail_page != next_page) |
| 1813 | rb_head_page_set_normal(cpu_buffer, new_head, |
| 1814 | next_page, |
| 1815 | RB_PAGE_HEAD); |
| 1816 | } |
| 1817 | |
| 1818 | /* |
| 1819 | * If this was the outer most commit (the one that |
| 1820 | * changed the original pointer from HEAD to UPDATE), |
| 1821 | * then it is up to us to reset it to NORMAL. |
| 1822 | */ |
| 1823 | if (type == RB_PAGE_HEAD) { |
| 1824 | ret = rb_head_page_set_normal(cpu_buffer, next_page, |
| 1825 | tail_page, |
| 1826 | RB_PAGE_UPDATE); |
| 1827 | if (RB_WARN_ON(cpu_buffer, |
| 1828 | ret != RB_PAGE_UPDATE)) |
| 1829 | return -1; |
| 1830 | } |
| 1831 | |
| 1832 | return 0; |
| 1833 | } |
| 1834 | |
| 1835 | static unsigned rb_calculate_event_length(unsigned length) |
| 1836 | { |
| 1837 | struct ring_buffer_event event; /* Used only for sizeof array */ |
| 1838 | |
| 1839 | /* zero length can cause confusions */ |
| 1840 | if (!length) |
| 1841 | length = 1; |
| 1842 | |
| 1843 | if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) |
| 1844 | length += sizeof(event.array[0]); |
| 1845 | |
| 1846 | length += RB_EVNT_HDR_SIZE; |
| 1847 | length = ALIGN(length, RB_ARCH_ALIGNMENT); |
| 1848 | |
| 1849 | return length; |
| 1850 | } |
| 1851 | |
| 1852 | static inline void |
| 1853 | rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer, |
| 1854 | struct buffer_page *tail_page, |
| 1855 | unsigned long tail, unsigned long length) |
| 1856 | { |
| 1857 | struct ring_buffer_event *event; |
| 1858 | |
| 1859 | /* |
| 1860 | * Only the event that crossed the page boundary |
| 1861 | * must fill the old tail_page with padding. |
| 1862 | */ |
| 1863 | if (tail >= BUF_PAGE_SIZE) { |
| 1864 | /* |
| 1865 | * If the page was filled, then we still need |
| 1866 | * to update the real_end. Reset it to zero |
| 1867 | * and the reader will ignore it. |
| 1868 | */ |
| 1869 | if (tail == BUF_PAGE_SIZE) |
| 1870 | tail_page->real_end = 0; |
| 1871 | |
| 1872 | local_sub(length, &tail_page->write); |
| 1873 | return; |
| 1874 | } |
| 1875 | |
| 1876 | event = __rb_page_index(tail_page, tail); |
| 1877 | kmemcheck_annotate_bitfield(event, bitfield); |
| 1878 | |
| 1879 | /* account for padding bytes */ |
| 1880 | local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes); |
| 1881 | |
| 1882 | /* |
| 1883 | * Save the original length to the meta data. |
| 1884 | * This will be used by the reader to add lost event |
| 1885 | * counter. |
| 1886 | */ |
| 1887 | tail_page->real_end = tail; |
| 1888 | |
| 1889 | /* |
| 1890 | * If this event is bigger than the minimum size, then |
| 1891 | * we need to be careful that we don't subtract the |
| 1892 | * write counter enough to allow another writer to slip |
| 1893 | * in on this page. |
| 1894 | * We put in a discarded commit instead, to make sure |
| 1895 | * that this space is not used again. |
| 1896 | * |
| 1897 | * If we are less than the minimum size, we don't need to |
| 1898 | * worry about it. |
| 1899 | */ |
| 1900 | if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) { |
| 1901 | /* No room for any events */ |
| 1902 | |
| 1903 | /* Mark the rest of the page with padding */ |
| 1904 | rb_event_set_padding(event); |
| 1905 | |
| 1906 | /* Set the write back to the previous setting */ |
| 1907 | local_sub(length, &tail_page->write); |
| 1908 | return; |
| 1909 | } |
| 1910 | |
| 1911 | /* Put in a discarded event */ |
| 1912 | event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE; |
| 1913 | event->type_len = RINGBUF_TYPE_PADDING; |
| 1914 | /* time delta must be non zero */ |
| 1915 | event->time_delta = 1; |
| 1916 | |
| 1917 | /* Set write to end of buffer */ |
| 1918 | length = (tail + length) - BUF_PAGE_SIZE; |
| 1919 | local_sub(length, &tail_page->write); |
| 1920 | } |
| 1921 | |
| 1922 | /* |
| 1923 | * This is the slow path, force gcc not to inline it. |
| 1924 | */ |
| 1925 | static noinline struct ring_buffer_event * |
| 1926 | rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer, |
| 1927 | unsigned long length, unsigned long tail, |
| 1928 | struct buffer_page *tail_page, u64 ts) |
| 1929 | { |
| 1930 | struct buffer_page *commit_page = cpu_buffer->commit_page; |
| 1931 | struct ring_buffer *buffer = cpu_buffer->buffer; |
| 1932 | struct buffer_page *next_page; |
| 1933 | int ret; |
| 1934 | |
| 1935 | next_page = tail_page; |
| 1936 | |
| 1937 | rb_inc_page(cpu_buffer, &next_page); |
| 1938 | |
| 1939 | /* |
| 1940 | * If for some reason, we had an interrupt storm that made |
| 1941 | * it all the way around the buffer, bail, and warn |
| 1942 | * about it. |
| 1943 | */ |
| 1944 | if (unlikely(next_page == commit_page)) { |
| 1945 | local_inc(&cpu_buffer->commit_overrun); |
| 1946 | goto out_reset; |
| 1947 | } |
| 1948 | |
| 1949 | /* |
| 1950 | * This is where the fun begins! |
| 1951 | * |
| 1952 | * We are fighting against races between a reader that |
| 1953 | * could be on another CPU trying to swap its reader |
| 1954 | * page with the buffer head. |
| 1955 | * |
| 1956 | * We are also fighting against interrupts coming in and |
| 1957 | * moving the head or tail on us as well. |
| 1958 | * |
| 1959 | * If the next page is the head page then we have filled |
| 1960 | * the buffer, unless the commit page is still on the |
| 1961 | * reader page. |
| 1962 | */ |
| 1963 | if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) { |
| 1964 | |
| 1965 | /* |
| 1966 | * If the commit is not on the reader page, then |
| 1967 | * move the header page. |
| 1968 | */ |
| 1969 | if (!rb_is_reader_page(cpu_buffer->commit_page)) { |
| 1970 | /* |
| 1971 | * If we are not in overwrite mode, |
| 1972 | * this is easy, just stop here. |
| 1973 | */ |
| 1974 | if (!(buffer->flags & RB_FL_OVERWRITE)) |
| 1975 | goto out_reset; |
| 1976 | |
| 1977 | ret = rb_handle_head_page(cpu_buffer, |
| 1978 | tail_page, |
| 1979 | next_page); |
| 1980 | if (ret < 0) |
| 1981 | goto out_reset; |
| 1982 | if (ret) |
| 1983 | goto out_again; |
| 1984 | } else { |
| 1985 | /* |
| 1986 | * We need to be careful here too. The |
| 1987 | * commit page could still be on the reader |
| 1988 | * page. We could have a small buffer, and |
| 1989 | * have filled up the buffer with events |
| 1990 | * from interrupts and such, and wrapped. |
| 1991 | * |
| 1992 | * Note, if the tail page is also the on the |
| 1993 | * reader_page, we let it move out. |
| 1994 | */ |
| 1995 | if (unlikely((cpu_buffer->commit_page != |
| 1996 | cpu_buffer->tail_page) && |
| 1997 | (cpu_buffer->commit_page == |
| 1998 | cpu_buffer->reader_page))) { |
| 1999 | local_inc(&cpu_buffer->commit_overrun); |
| 2000 | goto out_reset; |
| 2001 | } |
| 2002 | } |
| 2003 | } |
| 2004 | |
| 2005 | ret = rb_tail_page_update(cpu_buffer, tail_page, next_page); |
| 2006 | if (ret) { |
| 2007 | /* |
| 2008 | * Nested commits always have zero deltas, so |
| 2009 | * just reread the time stamp |
| 2010 | */ |
| 2011 | ts = rb_time_stamp(buffer); |
| 2012 | next_page->page->time_stamp = ts; |
| 2013 | } |
| 2014 | |
| 2015 | out_again: |
| 2016 | |
| 2017 | rb_reset_tail(cpu_buffer, tail_page, tail, length); |
| 2018 | |
| 2019 | /* fail and let the caller try again */ |
| 2020 | return ERR_PTR(-EAGAIN); |
| 2021 | |
| 2022 | out_reset: |
| 2023 | /* reset write */ |
| 2024 | rb_reset_tail(cpu_buffer, tail_page, tail, length); |
| 2025 | |
| 2026 | return NULL; |
| 2027 | } |
| 2028 | |
| 2029 | static struct ring_buffer_event * |
| 2030 | __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer, |
| 2031 | unsigned long length, u64 ts, |
| 2032 | u64 delta, int add_timestamp) |
| 2033 | { |
| 2034 | struct buffer_page *tail_page; |
| 2035 | struct ring_buffer_event *event; |
| 2036 | unsigned long tail, write; |
| 2037 | |
| 2038 | /* |
| 2039 | * If the time delta since the last event is too big to |
| 2040 | * hold in the time field of the event, then we append a |
| 2041 | * TIME EXTEND event ahead of the data event. |
| 2042 | */ |
| 2043 | if (unlikely(add_timestamp)) |
| 2044 | length += RB_LEN_TIME_EXTEND; |
| 2045 | |
| 2046 | tail_page = cpu_buffer->tail_page; |
| 2047 | write = local_add_return(length, &tail_page->write); |
| 2048 | |
| 2049 | /* set write to only the index of the write */ |
| 2050 | write &= RB_WRITE_MASK; |
| 2051 | tail = write - length; |
| 2052 | |
| 2053 | /* |
| 2054 | * If this is the first commit on the page, then it has the same |
| 2055 | * timestamp as the page itself. |
| 2056 | */ |
| 2057 | if (!tail) |
| 2058 | delta = 0; |
| 2059 | |
| 2060 | /* See if we shot pass the end of this buffer page */ |
| 2061 | if (unlikely(write > BUF_PAGE_SIZE)) |
| 2062 | return rb_move_tail(cpu_buffer, length, tail, |
| 2063 | tail_page, ts); |
| 2064 | |
| 2065 | /* We reserved something on the buffer */ |
| 2066 | |
| 2067 | event = __rb_page_index(tail_page, tail); |
| 2068 | kmemcheck_annotate_bitfield(event, bitfield); |
| 2069 | rb_update_event(cpu_buffer, event, length, add_timestamp, delta); |
| 2070 | |
| 2071 | local_inc(&tail_page->entries); |
| 2072 | |
| 2073 | /* |
| 2074 | * If this is the first commit on the page, then update |
| 2075 | * its timestamp. |
| 2076 | */ |
| 2077 | if (!tail) |
| 2078 | tail_page->page->time_stamp = ts; |
| 2079 | |
| 2080 | /* account for these added bytes */ |
| 2081 | local_add(length, &cpu_buffer->entries_bytes); |
| 2082 | |
| 2083 | return event; |
| 2084 | } |
| 2085 | |
| 2086 | static inline int |
| 2087 | rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer, |
| 2088 | struct ring_buffer_event *event) |
| 2089 | { |
| 2090 | unsigned long new_index, old_index; |
| 2091 | struct buffer_page *bpage; |
| 2092 | unsigned long index; |
| 2093 | unsigned long addr; |
| 2094 | |
| 2095 | new_index = rb_event_index(event); |
| 2096 | old_index = new_index + rb_event_ts_length(event); |
| 2097 | addr = (unsigned long)event; |
| 2098 | addr &= PAGE_MASK; |
| 2099 | |
| 2100 | bpage = cpu_buffer->tail_page; |
| 2101 | |
| 2102 | if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) { |
| 2103 | unsigned long write_mask = |
| 2104 | local_read(&bpage->write) & ~RB_WRITE_MASK; |
| 2105 | unsigned long event_length = rb_event_length(event); |
| 2106 | /* |
| 2107 | * This is on the tail page. It is possible that |
| 2108 | * a write could come in and move the tail page |
| 2109 | * and write to the next page. That is fine |
| 2110 | * because we just shorten what is on this page. |
| 2111 | */ |
| 2112 | old_index += write_mask; |
| 2113 | new_index += write_mask; |
| 2114 | index = local_cmpxchg(&bpage->write, old_index, new_index); |
| 2115 | if (index == old_index) { |
| 2116 | /* update counters */ |
| 2117 | local_sub(event_length, &cpu_buffer->entries_bytes); |
| 2118 | return 1; |
| 2119 | } |
| 2120 | } |
| 2121 | |
| 2122 | /* could not discard */ |
| 2123 | return 0; |
| 2124 | } |
| 2125 | |
| 2126 | static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer) |
| 2127 | { |
| 2128 | local_inc(&cpu_buffer->committing); |
| 2129 | local_inc(&cpu_buffer->commits); |
| 2130 | } |
| 2131 | |
| 2132 | static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer) |
| 2133 | { |
| 2134 | unsigned long commits; |
| 2135 | |
| 2136 | if (RB_WARN_ON(cpu_buffer, |
| 2137 | !local_read(&cpu_buffer->committing))) |
| 2138 | return; |
| 2139 | |
| 2140 | again: |
| 2141 | commits = local_read(&cpu_buffer->commits); |
| 2142 | /* synchronize with interrupts */ |
| 2143 | barrier(); |
| 2144 | if (local_read(&cpu_buffer->committing) == 1) |
| 2145 | rb_set_commit_to_write(cpu_buffer); |
| 2146 | |
| 2147 | local_dec(&cpu_buffer->committing); |
| 2148 | |
| 2149 | /* synchronize with interrupts */ |
| 2150 | barrier(); |
| 2151 | |
| 2152 | /* |
| 2153 | * Need to account for interrupts coming in between the |
| 2154 | * updating of the commit page and the clearing of the |
| 2155 | * committing counter. |
| 2156 | */ |
| 2157 | if (unlikely(local_read(&cpu_buffer->commits) != commits) && |
| 2158 | !local_read(&cpu_buffer->committing)) { |
| 2159 | local_inc(&cpu_buffer->committing); |
| 2160 | goto again; |
| 2161 | } |
| 2162 | } |
| 2163 | |
| 2164 | static struct ring_buffer_event * |
| 2165 | rb_reserve_next_event(struct ring_buffer *buffer, |
| 2166 | struct ring_buffer_per_cpu *cpu_buffer, |
| 2167 | unsigned long length) |
| 2168 | { |
| 2169 | struct ring_buffer_event *event; |
| 2170 | u64 ts, delta; |
| 2171 | int nr_loops = 0; |
| 2172 | int add_timestamp; |
| 2173 | u64 diff; |
| 2174 | |
| 2175 | rb_start_commit(cpu_buffer); |
| 2176 | |
| 2177 | #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP |
| 2178 | /* |
| 2179 | * Due to the ability to swap a cpu buffer from a buffer |
| 2180 | * it is possible it was swapped before we committed. |
| 2181 | * (committing stops a swap). We check for it here and |
| 2182 | * if it happened, we have to fail the write. |
| 2183 | */ |
| 2184 | barrier(); |
| 2185 | if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) { |
| 2186 | local_dec(&cpu_buffer->committing); |
| 2187 | local_dec(&cpu_buffer->commits); |
| 2188 | return NULL; |
| 2189 | } |
| 2190 | #endif |
| 2191 | |
| 2192 | length = rb_calculate_event_length(length); |
| 2193 | again: |
| 2194 | add_timestamp = 0; |
| 2195 | delta = 0; |
| 2196 | |
| 2197 | /* |
| 2198 | * We allow for interrupts to reenter here and do a trace. |
| 2199 | * If one does, it will cause this original code to loop |
| 2200 | * back here. Even with heavy interrupts happening, this |
| 2201 | * should only happen a few times in a row. If this happens |
| 2202 | * 1000 times in a row, there must be either an interrupt |
| 2203 | * storm or we have something buggy. |
| 2204 | * Bail! |
| 2205 | */ |
| 2206 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000)) |
| 2207 | goto out_fail; |
| 2208 | |
| 2209 | ts = rb_time_stamp(cpu_buffer->buffer); |
| 2210 | diff = ts - cpu_buffer->write_stamp; |
| 2211 | |
| 2212 | /* make sure this diff is calculated here */ |
| 2213 | barrier(); |
| 2214 | |
| 2215 | /* Did the write stamp get updated already? */ |
| 2216 | if (likely(ts >= cpu_buffer->write_stamp)) { |
| 2217 | delta = diff; |
| 2218 | if (unlikely(test_time_stamp(delta))) { |
| 2219 | int local_clock_stable = 1; |
| 2220 | #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK |
| 2221 | local_clock_stable = sched_clock_stable; |
| 2222 | #endif |
| 2223 | WARN_ONCE(delta > (1ULL << 59), |
| 2224 | KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s", |
| 2225 | (unsigned long long)delta, |
| 2226 | (unsigned long long)ts, |
| 2227 | (unsigned long long)cpu_buffer->write_stamp, |
| 2228 | local_clock_stable ? "" : |
| 2229 | "If you just came from a suspend/resume,\n" |
| 2230 | "please switch to the trace global clock:\n" |
| 2231 | " echo global > /sys/kernel/debug/tracing/trace_clock\n"); |
| 2232 | add_timestamp = 1; |
| 2233 | } |
| 2234 | } |
| 2235 | |
| 2236 | event = __rb_reserve_next(cpu_buffer, length, ts, |
| 2237 | delta, add_timestamp); |
| 2238 | if (unlikely(PTR_ERR(event) == -EAGAIN)) |
| 2239 | goto again; |
| 2240 | |
| 2241 | if (!event) |
| 2242 | goto out_fail; |
| 2243 | |
| 2244 | return event; |
| 2245 | |
| 2246 | out_fail: |
| 2247 | rb_end_commit(cpu_buffer); |
| 2248 | return NULL; |
| 2249 | } |
| 2250 | |
| 2251 | #ifdef CONFIG_TRACING |
| 2252 | |
| 2253 | #define TRACE_RECURSIVE_DEPTH 16 |
| 2254 | |
| 2255 | /* Keep this code out of the fast path cache */ |
| 2256 | static noinline void trace_recursive_fail(void) |
| 2257 | { |
| 2258 | /* Disable all tracing before we do anything else */ |
| 2259 | tracing_off_permanent(); |
| 2260 | |
| 2261 | printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:" |
| 2262 | "HC[%lu]:SC[%lu]:NMI[%lu]\n", |
| 2263 | trace_recursion_buffer(), |
| 2264 | hardirq_count() >> HARDIRQ_SHIFT, |
| 2265 | softirq_count() >> SOFTIRQ_SHIFT, |
| 2266 | in_nmi()); |
| 2267 | |
| 2268 | WARN_ON_ONCE(1); |
| 2269 | } |
| 2270 | |
| 2271 | static inline int trace_recursive_lock(void) |
| 2272 | { |
| 2273 | trace_recursion_inc(); |
| 2274 | |
| 2275 | if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH)) |
| 2276 | return 0; |
| 2277 | |
| 2278 | trace_recursive_fail(); |
| 2279 | |
| 2280 | return -1; |
| 2281 | } |
| 2282 | |
| 2283 | static inline void trace_recursive_unlock(void) |
| 2284 | { |
| 2285 | WARN_ON_ONCE(!trace_recursion_buffer()); |
| 2286 | |
| 2287 | trace_recursion_dec(); |
| 2288 | } |
| 2289 | |
| 2290 | #else |
| 2291 | |
| 2292 | #define trace_recursive_lock() (0) |
| 2293 | #define trace_recursive_unlock() do { } while (0) |
| 2294 | |
| 2295 | #endif |
| 2296 | |
| 2297 | /** |
| 2298 | * ring_buffer_lock_reserve - reserve a part of the buffer |
| 2299 | * @buffer: the ring buffer to reserve from |
| 2300 | * @length: the length of the data to reserve (excluding event header) |
| 2301 | * |
| 2302 | * Returns a reseverd event on the ring buffer to copy directly to. |
| 2303 | * The user of this interface will need to get the body to write into |
| 2304 | * and can use the ring_buffer_event_data() interface. |
| 2305 | * |
| 2306 | * The length is the length of the data needed, not the event length |
| 2307 | * which also includes the event header. |
| 2308 | * |
| 2309 | * Must be paired with ring_buffer_unlock_commit, unless NULL is returned. |
| 2310 | * If NULL is returned, then nothing has been allocated or locked. |
| 2311 | */ |
| 2312 | struct ring_buffer_event * |
| 2313 | ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length) |
| 2314 | { |
| 2315 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2316 | struct ring_buffer_event *event; |
| 2317 | int cpu; |
| 2318 | |
| 2319 | if (ring_buffer_flags != RB_BUFFERS_ON) |
| 2320 | return NULL; |
| 2321 | |
| 2322 | /* If we are tracing schedule, we don't want to recurse */ |
| 2323 | preempt_disable_notrace(); |
| 2324 | |
| 2325 | if (atomic_read(&buffer->record_disabled)) |
| 2326 | goto out_nocheck; |
| 2327 | |
| 2328 | if (trace_recursive_lock()) |
| 2329 | goto out_nocheck; |
| 2330 | |
| 2331 | cpu = raw_smp_processor_id(); |
| 2332 | |
| 2333 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 2334 | goto out; |
| 2335 | |
| 2336 | cpu_buffer = buffer->buffers[cpu]; |
| 2337 | |
| 2338 | if (atomic_read(&cpu_buffer->record_disabled)) |
| 2339 | goto out; |
| 2340 | |
| 2341 | if (length > BUF_MAX_DATA_SIZE) |
| 2342 | goto out; |
| 2343 | |
| 2344 | event = rb_reserve_next_event(buffer, cpu_buffer, length); |
| 2345 | if (!event) |
| 2346 | goto out; |
| 2347 | |
| 2348 | return event; |
| 2349 | |
| 2350 | out: |
| 2351 | trace_recursive_unlock(); |
| 2352 | |
| 2353 | out_nocheck: |
| 2354 | preempt_enable_notrace(); |
| 2355 | return NULL; |
| 2356 | } |
| 2357 | EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve); |
| 2358 | |
| 2359 | static void |
| 2360 | rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer, |
| 2361 | struct ring_buffer_event *event) |
| 2362 | { |
| 2363 | u64 delta; |
| 2364 | |
| 2365 | /* |
| 2366 | * The event first in the commit queue updates the |
| 2367 | * time stamp. |
| 2368 | */ |
| 2369 | if (rb_event_is_commit(cpu_buffer, event)) { |
| 2370 | /* |
| 2371 | * A commit event that is first on a page |
| 2372 | * updates the write timestamp with the page stamp |
| 2373 | */ |
| 2374 | if (!rb_event_index(event)) |
| 2375 | cpu_buffer->write_stamp = |
| 2376 | cpu_buffer->commit_page->page->time_stamp; |
| 2377 | else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) { |
| 2378 | delta = event->array[0]; |
| 2379 | delta <<= TS_SHIFT; |
| 2380 | delta += event->time_delta; |
| 2381 | cpu_buffer->write_stamp += delta; |
| 2382 | } else |
| 2383 | cpu_buffer->write_stamp += event->time_delta; |
| 2384 | } |
| 2385 | } |
| 2386 | |
| 2387 | static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer, |
| 2388 | struct ring_buffer_event *event) |
| 2389 | { |
| 2390 | local_inc(&cpu_buffer->entries); |
| 2391 | rb_update_write_stamp(cpu_buffer, event); |
| 2392 | rb_end_commit(cpu_buffer); |
| 2393 | } |
| 2394 | |
| 2395 | /** |
| 2396 | * ring_buffer_unlock_commit - commit a reserved |
| 2397 | * @buffer: The buffer to commit to |
| 2398 | * @event: The event pointer to commit. |
| 2399 | * |
| 2400 | * This commits the data to the ring buffer, and releases any locks held. |
| 2401 | * |
| 2402 | * Must be paired with ring_buffer_lock_reserve. |
| 2403 | */ |
| 2404 | int ring_buffer_unlock_commit(struct ring_buffer *buffer, |
| 2405 | struct ring_buffer_event *event) |
| 2406 | { |
| 2407 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2408 | int cpu = raw_smp_processor_id(); |
| 2409 | |
| 2410 | cpu_buffer = buffer->buffers[cpu]; |
| 2411 | |
| 2412 | rb_commit(cpu_buffer, event); |
| 2413 | |
| 2414 | trace_recursive_unlock(); |
| 2415 | |
| 2416 | preempt_enable_notrace(); |
| 2417 | |
| 2418 | return 0; |
| 2419 | } |
| 2420 | EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit); |
| 2421 | |
| 2422 | static inline void rb_event_discard(struct ring_buffer_event *event) |
| 2423 | { |
| 2424 | if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) |
| 2425 | event = skip_time_extend(event); |
| 2426 | |
| 2427 | /* array[0] holds the actual length for the discarded event */ |
| 2428 | event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE; |
| 2429 | event->type_len = RINGBUF_TYPE_PADDING; |
| 2430 | /* time delta must be non zero */ |
| 2431 | if (!event->time_delta) |
| 2432 | event->time_delta = 1; |
| 2433 | } |
| 2434 | |
| 2435 | /* |
| 2436 | * Decrement the entries to the page that an event is on. |
| 2437 | * The event does not even need to exist, only the pointer |
| 2438 | * to the page it is on. This may only be called before the commit |
| 2439 | * takes place. |
| 2440 | */ |
| 2441 | static inline void |
| 2442 | rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer, |
| 2443 | struct ring_buffer_event *event) |
| 2444 | { |
| 2445 | unsigned long addr = (unsigned long)event; |
| 2446 | struct buffer_page *bpage = cpu_buffer->commit_page; |
| 2447 | struct buffer_page *start; |
| 2448 | |
| 2449 | addr &= PAGE_MASK; |
| 2450 | |
| 2451 | /* Do the likely case first */ |
| 2452 | if (likely(bpage->page == (void *)addr)) { |
| 2453 | local_dec(&bpage->entries); |
| 2454 | return; |
| 2455 | } |
| 2456 | |
| 2457 | /* |
| 2458 | * Because the commit page may be on the reader page we |
| 2459 | * start with the next page and check the end loop there. |
| 2460 | */ |
| 2461 | rb_inc_page(cpu_buffer, &bpage); |
| 2462 | start = bpage; |
| 2463 | do { |
| 2464 | if (bpage->page == (void *)addr) { |
| 2465 | local_dec(&bpage->entries); |
| 2466 | return; |
| 2467 | } |
| 2468 | rb_inc_page(cpu_buffer, &bpage); |
| 2469 | } while (bpage != start); |
| 2470 | |
| 2471 | /* commit not part of this buffer?? */ |
| 2472 | RB_WARN_ON(cpu_buffer, 1); |
| 2473 | } |
| 2474 | |
| 2475 | /** |
| 2476 | * ring_buffer_commit_discard - discard an event that has not been committed |
| 2477 | * @buffer: the ring buffer |
| 2478 | * @event: non committed event to discard |
| 2479 | * |
| 2480 | * Sometimes an event that is in the ring buffer needs to be ignored. |
| 2481 | * This function lets the user discard an event in the ring buffer |
| 2482 | * and then that event will not be read later. |
| 2483 | * |
| 2484 | * This function only works if it is called before the the item has been |
| 2485 | * committed. It will try to free the event from the ring buffer |
| 2486 | * if another event has not been added behind it. |
| 2487 | * |
| 2488 | * If another event has been added behind it, it will set the event |
| 2489 | * up as discarded, and perform the commit. |
| 2490 | * |
| 2491 | * If this function is called, do not call ring_buffer_unlock_commit on |
| 2492 | * the event. |
| 2493 | */ |
| 2494 | void ring_buffer_discard_commit(struct ring_buffer *buffer, |
| 2495 | struct ring_buffer_event *event) |
| 2496 | { |
| 2497 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2498 | int cpu; |
| 2499 | |
| 2500 | /* The event is discarded regardless */ |
| 2501 | rb_event_discard(event); |
| 2502 | |
| 2503 | cpu = smp_processor_id(); |
| 2504 | cpu_buffer = buffer->buffers[cpu]; |
| 2505 | |
| 2506 | /* |
| 2507 | * This must only be called if the event has not been |
| 2508 | * committed yet. Thus we can assume that preemption |
| 2509 | * is still disabled. |
| 2510 | */ |
| 2511 | RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing)); |
| 2512 | |
| 2513 | rb_decrement_entry(cpu_buffer, event); |
| 2514 | if (rb_try_to_discard(cpu_buffer, event)) |
| 2515 | goto out; |
| 2516 | |
| 2517 | /* |
| 2518 | * The commit is still visible by the reader, so we |
| 2519 | * must still update the timestamp. |
| 2520 | */ |
| 2521 | rb_update_write_stamp(cpu_buffer, event); |
| 2522 | out: |
| 2523 | rb_end_commit(cpu_buffer); |
| 2524 | |
| 2525 | trace_recursive_unlock(); |
| 2526 | |
| 2527 | preempt_enable_notrace(); |
| 2528 | |
| 2529 | } |
| 2530 | EXPORT_SYMBOL_GPL(ring_buffer_discard_commit); |
| 2531 | |
| 2532 | /** |
| 2533 | * ring_buffer_write - write data to the buffer without reserving |
| 2534 | * @buffer: The ring buffer to write to. |
| 2535 | * @length: The length of the data being written (excluding the event header) |
| 2536 | * @data: The data to write to the buffer. |
| 2537 | * |
| 2538 | * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as |
| 2539 | * one function. If you already have the data to write to the buffer, it |
| 2540 | * may be easier to simply call this function. |
| 2541 | * |
| 2542 | * Note, like ring_buffer_lock_reserve, the length is the length of the data |
| 2543 | * and not the length of the event which would hold the header. |
| 2544 | */ |
| 2545 | int ring_buffer_write(struct ring_buffer *buffer, |
| 2546 | unsigned long length, |
| 2547 | void *data) |
| 2548 | { |
| 2549 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2550 | struct ring_buffer_event *event; |
| 2551 | void *body; |
| 2552 | int ret = -EBUSY; |
| 2553 | int cpu; |
| 2554 | |
| 2555 | if (ring_buffer_flags != RB_BUFFERS_ON) |
| 2556 | return -EBUSY; |
| 2557 | |
| 2558 | preempt_disable_notrace(); |
| 2559 | |
| 2560 | if (atomic_read(&buffer->record_disabled)) |
| 2561 | goto out; |
| 2562 | |
| 2563 | cpu = raw_smp_processor_id(); |
| 2564 | |
| 2565 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 2566 | goto out; |
| 2567 | |
| 2568 | cpu_buffer = buffer->buffers[cpu]; |
| 2569 | |
| 2570 | if (atomic_read(&cpu_buffer->record_disabled)) |
| 2571 | goto out; |
| 2572 | |
| 2573 | if (length > BUF_MAX_DATA_SIZE) |
| 2574 | goto out; |
| 2575 | |
| 2576 | event = rb_reserve_next_event(buffer, cpu_buffer, length); |
| 2577 | if (!event) |
| 2578 | goto out; |
| 2579 | |
| 2580 | body = rb_event_data(event); |
| 2581 | |
| 2582 | memcpy(body, data, length); |
| 2583 | |
| 2584 | rb_commit(cpu_buffer, event); |
| 2585 | |
| 2586 | ret = 0; |
| 2587 | out: |
| 2588 | preempt_enable_notrace(); |
| 2589 | |
| 2590 | return ret; |
| 2591 | } |
| 2592 | EXPORT_SYMBOL_GPL(ring_buffer_write); |
| 2593 | |
| 2594 | static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer) |
| 2595 | { |
| 2596 | struct buffer_page *reader = cpu_buffer->reader_page; |
| 2597 | struct buffer_page *head = rb_set_head_page(cpu_buffer); |
| 2598 | struct buffer_page *commit = cpu_buffer->commit_page; |
| 2599 | |
| 2600 | /* In case of error, head will be NULL */ |
| 2601 | if (unlikely(!head)) |
| 2602 | return 1; |
| 2603 | |
| 2604 | return reader->read == rb_page_commit(reader) && |
| 2605 | (commit == reader || |
| 2606 | (commit == head && |
| 2607 | head->read == rb_page_commit(commit))); |
| 2608 | } |
| 2609 | |
| 2610 | /** |
| 2611 | * ring_buffer_record_disable - stop all writes into the buffer |
| 2612 | * @buffer: The ring buffer to stop writes to. |
| 2613 | * |
| 2614 | * This prevents all writes to the buffer. Any attempt to write |
| 2615 | * to the buffer after this will fail and return NULL. |
| 2616 | * |
| 2617 | * The caller should call synchronize_sched() after this. |
| 2618 | */ |
| 2619 | void ring_buffer_record_disable(struct ring_buffer *buffer) |
| 2620 | { |
| 2621 | atomic_inc(&buffer->record_disabled); |
| 2622 | } |
| 2623 | EXPORT_SYMBOL_GPL(ring_buffer_record_disable); |
| 2624 | |
| 2625 | /** |
| 2626 | * ring_buffer_record_enable - enable writes to the buffer |
| 2627 | * @buffer: The ring buffer to enable writes |
| 2628 | * |
| 2629 | * Note, multiple disables will need the same number of enables |
| 2630 | * to truly enable the writing (much like preempt_disable). |
| 2631 | */ |
| 2632 | void ring_buffer_record_enable(struct ring_buffer *buffer) |
| 2633 | { |
| 2634 | atomic_dec(&buffer->record_disabled); |
| 2635 | } |
| 2636 | EXPORT_SYMBOL_GPL(ring_buffer_record_enable); |
| 2637 | |
| 2638 | /** |
| 2639 | * ring_buffer_record_off - stop all writes into the buffer |
| 2640 | * @buffer: The ring buffer to stop writes to. |
| 2641 | * |
| 2642 | * This prevents all writes to the buffer. Any attempt to write |
| 2643 | * to the buffer after this will fail and return NULL. |
| 2644 | * |
| 2645 | * This is different than ring_buffer_record_disable() as |
| 2646 | * it works like an on/off switch, where as the disable() verison |
| 2647 | * must be paired with a enable(). |
| 2648 | */ |
| 2649 | void ring_buffer_record_off(struct ring_buffer *buffer) |
| 2650 | { |
| 2651 | unsigned int rd; |
| 2652 | unsigned int new_rd; |
| 2653 | |
| 2654 | do { |
| 2655 | rd = atomic_read(&buffer->record_disabled); |
| 2656 | new_rd = rd | RB_BUFFER_OFF; |
| 2657 | } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd); |
| 2658 | } |
| 2659 | EXPORT_SYMBOL_GPL(ring_buffer_record_off); |
| 2660 | |
| 2661 | /** |
| 2662 | * ring_buffer_record_on - restart writes into the buffer |
| 2663 | * @buffer: The ring buffer to start writes to. |
| 2664 | * |
| 2665 | * This enables all writes to the buffer that was disabled by |
| 2666 | * ring_buffer_record_off(). |
| 2667 | * |
| 2668 | * This is different than ring_buffer_record_enable() as |
| 2669 | * it works like an on/off switch, where as the enable() verison |
| 2670 | * must be paired with a disable(). |
| 2671 | */ |
| 2672 | void ring_buffer_record_on(struct ring_buffer *buffer) |
| 2673 | { |
| 2674 | unsigned int rd; |
| 2675 | unsigned int new_rd; |
| 2676 | |
| 2677 | do { |
| 2678 | rd = atomic_read(&buffer->record_disabled); |
| 2679 | new_rd = rd & ~RB_BUFFER_OFF; |
| 2680 | } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd); |
| 2681 | } |
| 2682 | EXPORT_SYMBOL_GPL(ring_buffer_record_on); |
| 2683 | |
| 2684 | /** |
| 2685 | * ring_buffer_record_is_on - return true if the ring buffer can write |
| 2686 | * @buffer: The ring buffer to see if write is enabled |
| 2687 | * |
| 2688 | * Returns true if the ring buffer is in a state that it accepts writes. |
| 2689 | */ |
| 2690 | int ring_buffer_record_is_on(struct ring_buffer *buffer) |
| 2691 | { |
| 2692 | return !atomic_read(&buffer->record_disabled); |
| 2693 | } |
| 2694 | |
| 2695 | /** |
| 2696 | * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer |
| 2697 | * @buffer: The ring buffer to stop writes to. |
| 2698 | * @cpu: The CPU buffer to stop |
| 2699 | * |
| 2700 | * This prevents all writes to the buffer. Any attempt to write |
| 2701 | * to the buffer after this will fail and return NULL. |
| 2702 | * |
| 2703 | * The caller should call synchronize_sched() after this. |
| 2704 | */ |
| 2705 | void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu) |
| 2706 | { |
| 2707 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2708 | |
| 2709 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 2710 | return; |
| 2711 | |
| 2712 | cpu_buffer = buffer->buffers[cpu]; |
| 2713 | atomic_inc(&cpu_buffer->record_disabled); |
| 2714 | } |
| 2715 | EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu); |
| 2716 | |
| 2717 | /** |
| 2718 | * ring_buffer_record_enable_cpu - enable writes to the buffer |
| 2719 | * @buffer: The ring buffer to enable writes |
| 2720 | * @cpu: The CPU to enable. |
| 2721 | * |
| 2722 | * Note, multiple disables will need the same number of enables |
| 2723 | * to truly enable the writing (much like preempt_disable). |
| 2724 | */ |
| 2725 | void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu) |
| 2726 | { |
| 2727 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2728 | |
| 2729 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 2730 | return; |
| 2731 | |
| 2732 | cpu_buffer = buffer->buffers[cpu]; |
| 2733 | atomic_dec(&cpu_buffer->record_disabled); |
| 2734 | } |
| 2735 | EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu); |
| 2736 | |
| 2737 | /* |
| 2738 | * The total entries in the ring buffer is the running counter |
| 2739 | * of entries entered into the ring buffer, minus the sum of |
| 2740 | * the entries read from the ring buffer and the number of |
| 2741 | * entries that were overwritten. |
| 2742 | */ |
| 2743 | static inline unsigned long |
| 2744 | rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer) |
| 2745 | { |
| 2746 | return local_read(&cpu_buffer->entries) - |
| 2747 | (local_read(&cpu_buffer->overrun) + cpu_buffer->read); |
| 2748 | } |
| 2749 | |
| 2750 | /** |
| 2751 | * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer |
| 2752 | * @buffer: The ring buffer |
| 2753 | * @cpu: The per CPU buffer to read from. |
| 2754 | */ |
| 2755 | unsigned long ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu) |
| 2756 | { |
| 2757 | unsigned long flags; |
| 2758 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2759 | struct buffer_page *bpage; |
| 2760 | unsigned long ret = 0; |
| 2761 | |
| 2762 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 2763 | return 0; |
| 2764 | |
| 2765 | cpu_buffer = buffer->buffers[cpu]; |
| 2766 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags); |
| 2767 | /* |
| 2768 | * if the tail is on reader_page, oldest time stamp is on the reader |
| 2769 | * page |
| 2770 | */ |
| 2771 | if (cpu_buffer->tail_page == cpu_buffer->reader_page) |
| 2772 | bpage = cpu_buffer->reader_page; |
| 2773 | else |
| 2774 | bpage = rb_set_head_page(cpu_buffer); |
| 2775 | if (bpage) |
| 2776 | ret = bpage->page->time_stamp; |
| 2777 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); |
| 2778 | |
| 2779 | return ret; |
| 2780 | } |
| 2781 | EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts); |
| 2782 | |
| 2783 | /** |
| 2784 | * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer |
| 2785 | * @buffer: The ring buffer |
| 2786 | * @cpu: The per CPU buffer to read from. |
| 2787 | */ |
| 2788 | unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu) |
| 2789 | { |
| 2790 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2791 | unsigned long ret; |
| 2792 | |
| 2793 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 2794 | return 0; |
| 2795 | |
| 2796 | cpu_buffer = buffer->buffers[cpu]; |
| 2797 | ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes; |
| 2798 | |
| 2799 | return ret; |
| 2800 | } |
| 2801 | EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu); |
| 2802 | |
| 2803 | /** |
| 2804 | * ring_buffer_entries_cpu - get the number of entries in a cpu buffer |
| 2805 | * @buffer: The ring buffer |
| 2806 | * @cpu: The per CPU buffer to get the entries from. |
| 2807 | */ |
| 2808 | unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu) |
| 2809 | { |
| 2810 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2811 | |
| 2812 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 2813 | return 0; |
| 2814 | |
| 2815 | cpu_buffer = buffer->buffers[cpu]; |
| 2816 | |
| 2817 | return rb_num_of_entries(cpu_buffer); |
| 2818 | } |
| 2819 | EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu); |
| 2820 | |
| 2821 | /** |
| 2822 | * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer |
| 2823 | * @buffer: The ring buffer |
| 2824 | * @cpu: The per CPU buffer to get the number of overruns from |
| 2825 | */ |
| 2826 | unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu) |
| 2827 | { |
| 2828 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2829 | unsigned long ret; |
| 2830 | |
| 2831 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 2832 | return 0; |
| 2833 | |
| 2834 | cpu_buffer = buffer->buffers[cpu]; |
| 2835 | ret = local_read(&cpu_buffer->overrun); |
| 2836 | |
| 2837 | return ret; |
| 2838 | } |
| 2839 | EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu); |
| 2840 | |
| 2841 | /** |
| 2842 | * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits |
| 2843 | * @buffer: The ring buffer |
| 2844 | * @cpu: The per CPU buffer to get the number of overruns from |
| 2845 | */ |
| 2846 | unsigned long |
| 2847 | ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu) |
| 2848 | { |
| 2849 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2850 | unsigned long ret; |
| 2851 | |
| 2852 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 2853 | return 0; |
| 2854 | |
| 2855 | cpu_buffer = buffer->buffers[cpu]; |
| 2856 | ret = local_read(&cpu_buffer->commit_overrun); |
| 2857 | |
| 2858 | return ret; |
| 2859 | } |
| 2860 | EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu); |
| 2861 | |
| 2862 | /** |
| 2863 | * ring_buffer_entries - get the number of entries in a buffer |
| 2864 | * @buffer: The ring buffer |
| 2865 | * |
| 2866 | * Returns the total number of entries in the ring buffer |
| 2867 | * (all CPU entries) |
| 2868 | */ |
| 2869 | unsigned long ring_buffer_entries(struct ring_buffer *buffer) |
| 2870 | { |
| 2871 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2872 | unsigned long entries = 0; |
| 2873 | int cpu; |
| 2874 | |
| 2875 | /* if you care about this being correct, lock the buffer */ |
| 2876 | for_each_buffer_cpu(buffer, cpu) { |
| 2877 | cpu_buffer = buffer->buffers[cpu]; |
| 2878 | entries += rb_num_of_entries(cpu_buffer); |
| 2879 | } |
| 2880 | |
| 2881 | return entries; |
| 2882 | } |
| 2883 | EXPORT_SYMBOL_GPL(ring_buffer_entries); |
| 2884 | |
| 2885 | /** |
| 2886 | * ring_buffer_overruns - get the number of overruns in buffer |
| 2887 | * @buffer: The ring buffer |
| 2888 | * |
| 2889 | * Returns the total number of overruns in the ring buffer |
| 2890 | * (all CPU entries) |
| 2891 | */ |
| 2892 | unsigned long ring_buffer_overruns(struct ring_buffer *buffer) |
| 2893 | { |
| 2894 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2895 | unsigned long overruns = 0; |
| 2896 | int cpu; |
| 2897 | |
| 2898 | /* if you care about this being correct, lock the buffer */ |
| 2899 | for_each_buffer_cpu(buffer, cpu) { |
| 2900 | cpu_buffer = buffer->buffers[cpu]; |
| 2901 | overruns += local_read(&cpu_buffer->overrun); |
| 2902 | } |
| 2903 | |
| 2904 | return overruns; |
| 2905 | } |
| 2906 | EXPORT_SYMBOL_GPL(ring_buffer_overruns); |
| 2907 | |
| 2908 | static void rb_iter_reset(struct ring_buffer_iter *iter) |
| 2909 | { |
| 2910 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; |
| 2911 | |
| 2912 | /* Iterator usage is expected to have record disabled */ |
| 2913 | if (list_empty(&cpu_buffer->reader_page->list)) { |
| 2914 | iter->head_page = rb_set_head_page(cpu_buffer); |
| 2915 | if (unlikely(!iter->head_page)) |
| 2916 | return; |
| 2917 | iter->head = iter->head_page->read; |
| 2918 | } else { |
| 2919 | iter->head_page = cpu_buffer->reader_page; |
| 2920 | iter->head = cpu_buffer->reader_page->read; |
| 2921 | } |
| 2922 | if (iter->head) |
| 2923 | iter->read_stamp = cpu_buffer->read_stamp; |
| 2924 | else |
| 2925 | iter->read_stamp = iter->head_page->page->time_stamp; |
| 2926 | iter->cache_reader_page = cpu_buffer->reader_page; |
| 2927 | iter->cache_read = cpu_buffer->read; |
| 2928 | } |
| 2929 | |
| 2930 | /** |
| 2931 | * ring_buffer_iter_reset - reset an iterator |
| 2932 | * @iter: The iterator to reset |
| 2933 | * |
| 2934 | * Resets the iterator, so that it will start from the beginning |
| 2935 | * again. |
| 2936 | */ |
| 2937 | void ring_buffer_iter_reset(struct ring_buffer_iter *iter) |
| 2938 | { |
| 2939 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2940 | unsigned long flags; |
| 2941 | int locked; |
| 2942 | |
| 2943 | if (!iter) |
| 2944 | return; |
| 2945 | |
| 2946 | cpu_buffer = iter->cpu_buffer; |
| 2947 | |
| 2948 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 2949 | rb_iter_reset(iter); |
| 2950 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 2951 | } |
| 2952 | EXPORT_SYMBOL_GPL(ring_buffer_iter_reset); |
| 2953 | |
| 2954 | /** |
| 2955 | * ring_buffer_iter_empty - check if an iterator has no more to read |
| 2956 | * @iter: The iterator to check |
| 2957 | */ |
| 2958 | int ring_buffer_iter_empty(struct ring_buffer_iter *iter) |
| 2959 | { |
| 2960 | struct ring_buffer_per_cpu *cpu_buffer; |
| 2961 | |
| 2962 | cpu_buffer = iter->cpu_buffer; |
| 2963 | |
| 2964 | return iter->head_page == cpu_buffer->commit_page && |
| 2965 | iter->head == rb_commit_index(cpu_buffer); |
| 2966 | } |
| 2967 | EXPORT_SYMBOL_GPL(ring_buffer_iter_empty); |
| 2968 | |
| 2969 | static void |
| 2970 | rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer, |
| 2971 | struct ring_buffer_event *event) |
| 2972 | { |
| 2973 | u64 delta; |
| 2974 | |
| 2975 | switch (event->type_len) { |
| 2976 | case RINGBUF_TYPE_PADDING: |
| 2977 | return; |
| 2978 | |
| 2979 | case RINGBUF_TYPE_TIME_EXTEND: |
| 2980 | delta = event->array[0]; |
| 2981 | delta <<= TS_SHIFT; |
| 2982 | delta += event->time_delta; |
| 2983 | cpu_buffer->read_stamp += delta; |
| 2984 | return; |
| 2985 | |
| 2986 | case RINGBUF_TYPE_TIME_STAMP: |
| 2987 | /* FIXME: not implemented */ |
| 2988 | return; |
| 2989 | |
| 2990 | case RINGBUF_TYPE_DATA: |
| 2991 | cpu_buffer->read_stamp += event->time_delta; |
| 2992 | return; |
| 2993 | |
| 2994 | default: |
| 2995 | BUG(); |
| 2996 | } |
| 2997 | return; |
| 2998 | } |
| 2999 | |
| 3000 | static void |
| 3001 | rb_update_iter_read_stamp(struct ring_buffer_iter *iter, |
| 3002 | struct ring_buffer_event *event) |
| 3003 | { |
| 3004 | u64 delta; |
| 3005 | |
| 3006 | switch (event->type_len) { |
| 3007 | case RINGBUF_TYPE_PADDING: |
| 3008 | return; |
| 3009 | |
| 3010 | case RINGBUF_TYPE_TIME_EXTEND: |
| 3011 | delta = event->array[0]; |
| 3012 | delta <<= TS_SHIFT; |
| 3013 | delta += event->time_delta; |
| 3014 | iter->read_stamp += delta; |
| 3015 | return; |
| 3016 | |
| 3017 | case RINGBUF_TYPE_TIME_STAMP: |
| 3018 | /* FIXME: not implemented */ |
| 3019 | return; |
| 3020 | |
| 3021 | case RINGBUF_TYPE_DATA: |
| 3022 | iter->read_stamp += event->time_delta; |
| 3023 | return; |
| 3024 | |
| 3025 | default: |
| 3026 | BUG(); |
| 3027 | } |
| 3028 | return; |
| 3029 | } |
| 3030 | |
| 3031 | static struct buffer_page * |
| 3032 | rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer) |
| 3033 | { |
| 3034 | struct buffer_page *reader = NULL; |
| 3035 | unsigned long overwrite; |
| 3036 | unsigned long flags; |
| 3037 | int nr_loops = 0; |
| 3038 | int ret; |
| 3039 | |
| 3040 | local_irq_save(flags); |
| 3041 | arch_spin_lock(&cpu_buffer->lock); |
| 3042 | |
| 3043 | again: |
| 3044 | /* |
| 3045 | * This should normally only loop twice. But because the |
| 3046 | * start of the reader inserts an empty page, it causes |
| 3047 | * a case where we will loop three times. There should be no |
| 3048 | * reason to loop four times (that I know of). |
| 3049 | */ |
| 3050 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) { |
| 3051 | reader = NULL; |
| 3052 | goto out; |
| 3053 | } |
| 3054 | |
| 3055 | reader = cpu_buffer->reader_page; |
| 3056 | |
| 3057 | /* If there's more to read, return this page */ |
| 3058 | if (cpu_buffer->reader_page->read < rb_page_size(reader)) |
| 3059 | goto out; |
| 3060 | |
| 3061 | /* Never should we have an index greater than the size */ |
| 3062 | if (RB_WARN_ON(cpu_buffer, |
| 3063 | cpu_buffer->reader_page->read > rb_page_size(reader))) |
| 3064 | goto out; |
| 3065 | |
| 3066 | /* check if we caught up to the tail */ |
| 3067 | reader = NULL; |
| 3068 | if (cpu_buffer->commit_page == cpu_buffer->reader_page) |
| 3069 | goto out; |
| 3070 | |
| 3071 | /* |
| 3072 | * Reset the reader page to size zero. |
| 3073 | */ |
| 3074 | local_set(&cpu_buffer->reader_page->write, 0); |
| 3075 | local_set(&cpu_buffer->reader_page->entries, 0); |
| 3076 | local_set(&cpu_buffer->reader_page->page->commit, 0); |
| 3077 | cpu_buffer->reader_page->real_end = 0; |
| 3078 | |
| 3079 | spin: |
| 3080 | /* |
| 3081 | * Splice the empty reader page into the list around the head. |
| 3082 | */ |
| 3083 | reader = rb_set_head_page(cpu_buffer); |
| 3084 | if (!reader) |
| 3085 | goto out; |
| 3086 | cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next); |
| 3087 | cpu_buffer->reader_page->list.prev = reader->list.prev; |
| 3088 | |
| 3089 | /* |
| 3090 | * cpu_buffer->pages just needs to point to the buffer, it |
| 3091 | * has no specific buffer page to point to. Lets move it out |
| 3092 | * of our way so we don't accidentally swap it. |
| 3093 | */ |
| 3094 | cpu_buffer->pages = reader->list.prev; |
| 3095 | |
| 3096 | /* The reader page will be pointing to the new head */ |
| 3097 | rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list); |
| 3098 | |
| 3099 | /* |
| 3100 | * We want to make sure we read the overruns after we set up our |
| 3101 | * pointers to the next object. The writer side does a |
| 3102 | * cmpxchg to cross pages which acts as the mb on the writer |
| 3103 | * side. Note, the reader will constantly fail the swap |
| 3104 | * while the writer is updating the pointers, so this |
| 3105 | * guarantees that the overwrite recorded here is the one we |
| 3106 | * want to compare with the last_overrun. |
| 3107 | */ |
| 3108 | smp_mb(); |
| 3109 | overwrite = local_read(&(cpu_buffer->overrun)); |
| 3110 | |
| 3111 | /* |
| 3112 | * Here's the tricky part. |
| 3113 | * |
| 3114 | * We need to move the pointer past the header page. |
| 3115 | * But we can only do that if a writer is not currently |
| 3116 | * moving it. The page before the header page has the |
| 3117 | * flag bit '1' set if it is pointing to the page we want. |
| 3118 | * but if the writer is in the process of moving it |
| 3119 | * than it will be '2' or already moved '0'. |
| 3120 | */ |
| 3121 | |
| 3122 | ret = rb_head_page_replace(reader, cpu_buffer->reader_page); |
| 3123 | |
| 3124 | /* |
| 3125 | * If we did not convert it, then we must try again. |
| 3126 | */ |
| 3127 | if (!ret) |
| 3128 | goto spin; |
| 3129 | |
| 3130 | /* |
| 3131 | * Yeah! We succeeded in replacing the page. |
| 3132 | * |
| 3133 | * Now make the new head point back to the reader page. |
| 3134 | */ |
| 3135 | rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list; |
| 3136 | rb_inc_page(cpu_buffer, &cpu_buffer->head_page); |
| 3137 | |
| 3138 | /* Finally update the reader page to the new head */ |
| 3139 | cpu_buffer->reader_page = reader; |
| 3140 | rb_reset_reader_page(cpu_buffer); |
| 3141 | |
| 3142 | if (overwrite != cpu_buffer->last_overrun) { |
| 3143 | cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun; |
| 3144 | cpu_buffer->last_overrun = overwrite; |
| 3145 | } |
| 3146 | |
| 3147 | goto again; |
| 3148 | |
| 3149 | out: |
| 3150 | arch_spin_unlock(&cpu_buffer->lock); |
| 3151 | local_irq_restore(flags); |
| 3152 | |
| 3153 | return reader; |
| 3154 | } |
| 3155 | |
| 3156 | static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer) |
| 3157 | { |
| 3158 | struct ring_buffer_event *event; |
| 3159 | struct buffer_page *reader; |
| 3160 | unsigned length; |
| 3161 | |
| 3162 | reader = rb_get_reader_page(cpu_buffer); |
| 3163 | |
| 3164 | /* This function should not be called when buffer is empty */ |
| 3165 | if (RB_WARN_ON(cpu_buffer, !reader)) |
| 3166 | return; |
| 3167 | |
| 3168 | event = rb_reader_event(cpu_buffer); |
| 3169 | |
| 3170 | if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX) |
| 3171 | cpu_buffer->read++; |
| 3172 | |
| 3173 | rb_update_read_stamp(cpu_buffer, event); |
| 3174 | |
| 3175 | length = rb_event_length(event); |
| 3176 | cpu_buffer->reader_page->read += length; |
| 3177 | } |
| 3178 | |
| 3179 | static void rb_advance_iter(struct ring_buffer_iter *iter) |
| 3180 | { |
| 3181 | struct ring_buffer_per_cpu *cpu_buffer; |
| 3182 | struct ring_buffer_event *event; |
| 3183 | unsigned length; |
| 3184 | |
| 3185 | cpu_buffer = iter->cpu_buffer; |
| 3186 | |
| 3187 | /* |
| 3188 | * Check if we are at the end of the buffer. |
| 3189 | */ |
| 3190 | if (iter->head >= rb_page_size(iter->head_page)) { |
| 3191 | /* discarded commits can make the page empty */ |
| 3192 | if (iter->head_page == cpu_buffer->commit_page) |
| 3193 | return; |
| 3194 | rb_inc_iter(iter); |
| 3195 | return; |
| 3196 | } |
| 3197 | |
| 3198 | event = rb_iter_head_event(iter); |
| 3199 | |
| 3200 | length = rb_event_length(event); |
| 3201 | |
| 3202 | /* |
| 3203 | * This should not be called to advance the header if we are |
| 3204 | * at the tail of the buffer. |
| 3205 | */ |
| 3206 | if (RB_WARN_ON(cpu_buffer, |
| 3207 | (iter->head_page == cpu_buffer->commit_page) && |
| 3208 | (iter->head + length > rb_commit_index(cpu_buffer)))) |
| 3209 | return; |
| 3210 | |
| 3211 | rb_update_iter_read_stamp(iter, event); |
| 3212 | |
| 3213 | iter->head += length; |
| 3214 | |
| 3215 | /* check for end of page padding */ |
| 3216 | if ((iter->head >= rb_page_size(iter->head_page)) && |
| 3217 | (iter->head_page != cpu_buffer->commit_page)) |
| 3218 | rb_advance_iter(iter); |
| 3219 | } |
| 3220 | |
| 3221 | static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer) |
| 3222 | { |
| 3223 | return cpu_buffer->lost_events; |
| 3224 | } |
| 3225 | |
| 3226 | static struct ring_buffer_event * |
| 3227 | rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts, |
| 3228 | unsigned long *lost_events) |
| 3229 | { |
| 3230 | struct ring_buffer_event *event; |
| 3231 | struct buffer_page *reader; |
| 3232 | int nr_loops = 0; |
| 3233 | |
| 3234 | again: |
| 3235 | /* |
| 3236 | * We repeat when a time extend is encountered. |
| 3237 | * Since the time extend is always attached to a data event, |
| 3238 | * we should never loop more than once. |
| 3239 | * (We never hit the following condition more than twice). |
| 3240 | */ |
| 3241 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2)) |
| 3242 | return NULL; |
| 3243 | |
| 3244 | reader = rb_get_reader_page(cpu_buffer); |
| 3245 | if (!reader) |
| 3246 | return NULL; |
| 3247 | |
| 3248 | event = rb_reader_event(cpu_buffer); |
| 3249 | |
| 3250 | switch (event->type_len) { |
| 3251 | case RINGBUF_TYPE_PADDING: |
| 3252 | if (rb_null_event(event)) |
| 3253 | RB_WARN_ON(cpu_buffer, 1); |
| 3254 | /* |
| 3255 | * Because the writer could be discarding every |
| 3256 | * event it creates (which would probably be bad) |
| 3257 | * if we were to go back to "again" then we may never |
| 3258 | * catch up, and will trigger the warn on, or lock |
| 3259 | * the box. Return the padding, and we will release |
| 3260 | * the current locks, and try again. |
| 3261 | */ |
| 3262 | return event; |
| 3263 | |
| 3264 | case RINGBUF_TYPE_TIME_EXTEND: |
| 3265 | /* Internal data, OK to advance */ |
| 3266 | rb_advance_reader(cpu_buffer); |
| 3267 | goto again; |
| 3268 | |
| 3269 | case RINGBUF_TYPE_TIME_STAMP: |
| 3270 | /* FIXME: not implemented */ |
| 3271 | rb_advance_reader(cpu_buffer); |
| 3272 | goto again; |
| 3273 | |
| 3274 | case RINGBUF_TYPE_DATA: |
| 3275 | if (ts) { |
| 3276 | *ts = cpu_buffer->read_stamp + event->time_delta; |
| 3277 | ring_buffer_normalize_time_stamp(cpu_buffer->buffer, |
| 3278 | cpu_buffer->cpu, ts); |
| 3279 | } |
| 3280 | if (lost_events) |
| 3281 | *lost_events = rb_lost_events(cpu_buffer); |
| 3282 | return event; |
| 3283 | |
| 3284 | default: |
| 3285 | BUG(); |
| 3286 | } |
| 3287 | |
| 3288 | return NULL; |
| 3289 | } |
| 3290 | EXPORT_SYMBOL_GPL(ring_buffer_peek); |
| 3291 | |
| 3292 | static struct ring_buffer_event * |
| 3293 | rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts) |
| 3294 | { |
| 3295 | struct ring_buffer *buffer; |
| 3296 | struct ring_buffer_per_cpu *cpu_buffer; |
| 3297 | struct ring_buffer_event *event; |
| 3298 | int nr_loops = 0; |
| 3299 | |
| 3300 | cpu_buffer = iter->cpu_buffer; |
| 3301 | buffer = cpu_buffer->buffer; |
| 3302 | |
| 3303 | /* |
| 3304 | * Check if someone performed a consuming read to |
| 3305 | * the buffer. A consuming read invalidates the iterator |
| 3306 | * and we need to reset the iterator in this case. |
| 3307 | */ |
| 3308 | if (unlikely(iter->cache_read != cpu_buffer->read || |
| 3309 | iter->cache_reader_page != cpu_buffer->reader_page)) |
| 3310 | rb_iter_reset(iter); |
| 3311 | |
| 3312 | again: |
| 3313 | if (ring_buffer_iter_empty(iter)) |
| 3314 | return NULL; |
| 3315 | |
| 3316 | /* |
| 3317 | * We repeat when a time extend is encountered. |
| 3318 | * Since the time extend is always attached to a data event, |
| 3319 | * we should never loop more than once. |
| 3320 | * (We never hit the following condition more than twice). |
| 3321 | */ |
| 3322 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2)) |
| 3323 | return NULL; |
| 3324 | |
| 3325 | if (rb_per_cpu_empty(cpu_buffer)) |
| 3326 | return NULL; |
| 3327 | |
| 3328 | if (iter->head >= local_read(&iter->head_page->page->commit)) { |
| 3329 | rb_inc_iter(iter); |
| 3330 | goto again; |
| 3331 | } |
| 3332 | |
| 3333 | event = rb_iter_head_event(iter); |
| 3334 | |
| 3335 | switch (event->type_len) { |
| 3336 | case RINGBUF_TYPE_PADDING: |
| 3337 | if (rb_null_event(event)) { |
| 3338 | rb_inc_iter(iter); |
| 3339 | goto again; |
| 3340 | } |
| 3341 | rb_advance_iter(iter); |
| 3342 | return event; |
| 3343 | |
| 3344 | case RINGBUF_TYPE_TIME_EXTEND: |
| 3345 | /* Internal data, OK to advance */ |
| 3346 | rb_advance_iter(iter); |
| 3347 | goto again; |
| 3348 | |
| 3349 | case RINGBUF_TYPE_TIME_STAMP: |
| 3350 | /* FIXME: not implemented */ |
| 3351 | rb_advance_iter(iter); |
| 3352 | goto again; |
| 3353 | |
| 3354 | case RINGBUF_TYPE_DATA: |
| 3355 | if (ts) { |
| 3356 | *ts = iter->read_stamp + event->time_delta; |
| 3357 | ring_buffer_normalize_time_stamp(buffer, |
| 3358 | cpu_buffer->cpu, ts); |
| 3359 | } |
| 3360 | return event; |
| 3361 | |
| 3362 | default: |
| 3363 | BUG(); |
| 3364 | } |
| 3365 | |
| 3366 | return NULL; |
| 3367 | } |
| 3368 | EXPORT_SYMBOL_GPL(ring_buffer_iter_peek); |
| 3369 | |
| 3370 | /** |
| 3371 | * ring_buffer_peek - peek at the next event to be read |
| 3372 | * @buffer: The ring buffer to read |
| 3373 | * @cpu: The cpu to peak at |
| 3374 | * @ts: The timestamp counter of this event. |
| 3375 | * @lost_events: a variable to store if events were lost (may be NULL) |
| 3376 | * |
| 3377 | * This will return the event that will be read next, but does |
| 3378 | * not consume the data. |
| 3379 | */ |
| 3380 | struct ring_buffer_event * |
| 3381 | ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts, |
| 3382 | unsigned long *lost_events) |
| 3383 | { |
| 3384 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu]; |
| 3385 | struct ring_buffer_event *event; |
| 3386 | unsigned long flags; |
| 3387 | int locked; |
| 3388 | |
| 3389 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 3390 | return NULL; |
| 3391 | |
| 3392 | again: |
| 3393 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 3394 | event = rb_buffer_peek(cpu_buffer, ts, lost_events); |
| 3395 | if (event && event->type_len == RINGBUF_TYPE_PADDING) |
| 3396 | rb_advance_reader(cpu_buffer); |
| 3397 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 3398 | |
| 3399 | if (event && event->type_len == RINGBUF_TYPE_PADDING) |
| 3400 | goto again; |
| 3401 | |
| 3402 | return event; |
| 3403 | } |
| 3404 | |
| 3405 | /** |
| 3406 | * ring_buffer_iter_peek - peek at the next event to be read |
| 3407 | * @iter: The ring buffer iterator |
| 3408 | * @ts: The timestamp counter of this event. |
| 3409 | * |
| 3410 | * This will return the event that will be read next, but does |
| 3411 | * not increment the iterator. |
| 3412 | */ |
| 3413 | struct ring_buffer_event * |
| 3414 | ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts) |
| 3415 | { |
| 3416 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; |
| 3417 | struct ring_buffer_event *event; |
| 3418 | unsigned long flags; |
| 3419 | int locked; |
| 3420 | |
| 3421 | again: |
| 3422 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 3423 | event = rb_iter_peek(iter, ts); |
| 3424 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 3425 | |
| 3426 | if (event && event->type_len == RINGBUF_TYPE_PADDING) |
| 3427 | goto again; |
| 3428 | |
| 3429 | return event; |
| 3430 | } |
| 3431 | |
| 3432 | /** |
| 3433 | * ring_buffer_consume - return an event and consume it |
| 3434 | * @buffer: The ring buffer to get the next event from |
| 3435 | * @cpu: the cpu to read the buffer from |
| 3436 | * @ts: a variable to store the timestamp (may be NULL) |
| 3437 | * @lost_events: a variable to store if events were lost (may be NULL) |
| 3438 | * |
| 3439 | * Returns the next event in the ring buffer, and that event is consumed. |
| 3440 | * Meaning, that sequential reads will keep returning a different event, |
| 3441 | * and eventually empty the ring buffer if the producer is slower. |
| 3442 | */ |
| 3443 | struct ring_buffer_event * |
| 3444 | ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts, |
| 3445 | unsigned long *lost_events) |
| 3446 | { |
| 3447 | struct ring_buffer_per_cpu *cpu_buffer; |
| 3448 | struct ring_buffer_event *event = NULL; |
| 3449 | unsigned long flags; |
| 3450 | int locked; |
| 3451 | |
| 3452 | again: |
| 3453 | /* might be called in atomic */ |
| 3454 | preempt_disable(); |
| 3455 | |
| 3456 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 3457 | goto out; |
| 3458 | |
| 3459 | cpu_buffer = buffer->buffers[cpu]; |
| 3460 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 3461 | |
| 3462 | event = rb_buffer_peek(cpu_buffer, ts, lost_events); |
| 3463 | if (event) { |
| 3464 | cpu_buffer->lost_events = 0; |
| 3465 | rb_advance_reader(cpu_buffer); |
| 3466 | } |
| 3467 | |
| 3468 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 3469 | |
| 3470 | |
| 3471 | out: |
| 3472 | preempt_enable(); |
| 3473 | |
| 3474 | if (event && event->type_len == RINGBUF_TYPE_PADDING) |
| 3475 | goto again; |
| 3476 | |
| 3477 | return event; |
| 3478 | } |
| 3479 | EXPORT_SYMBOL_GPL(ring_buffer_consume); |
| 3480 | |
| 3481 | /** |
| 3482 | * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer |
| 3483 | * @buffer: The ring buffer to read from |
| 3484 | * @cpu: The cpu buffer to iterate over |
| 3485 | * |
| 3486 | * This performs the initial preparations necessary to iterate |
| 3487 | * through the buffer. Memory is allocated, buffer recording |
| 3488 | * is disabled, and the iterator pointer is returned to the caller. |
| 3489 | * |
| 3490 | * Disabling buffer recordng prevents the reading from being |
| 3491 | * corrupted. This is not a consuming read, so a producer is not |
| 3492 | * expected. |
| 3493 | * |
| 3494 | * After a sequence of ring_buffer_read_prepare calls, the user is |
| 3495 | * expected to make at least one call to ring_buffer_prepare_sync. |
| 3496 | * Afterwards, ring_buffer_read_start is invoked to get things going |
| 3497 | * for real. |
| 3498 | * |
| 3499 | * This overall must be paired with ring_buffer_finish. |
| 3500 | */ |
| 3501 | struct ring_buffer_iter * |
| 3502 | ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu) |
| 3503 | { |
| 3504 | struct ring_buffer_per_cpu *cpu_buffer; |
| 3505 | struct ring_buffer_iter *iter; |
| 3506 | |
| 3507 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 3508 | return NULL; |
| 3509 | |
| 3510 | iter = kmalloc(sizeof(*iter), GFP_KERNEL); |
| 3511 | if (!iter) |
| 3512 | return NULL; |
| 3513 | |
| 3514 | cpu_buffer = buffer->buffers[cpu]; |
| 3515 | |
| 3516 | iter->cpu_buffer = cpu_buffer; |
| 3517 | |
| 3518 | atomic_inc(&cpu_buffer->record_disabled); |
| 3519 | |
| 3520 | return iter; |
| 3521 | } |
| 3522 | EXPORT_SYMBOL_GPL(ring_buffer_read_prepare); |
| 3523 | |
| 3524 | /** |
| 3525 | * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls |
| 3526 | * |
| 3527 | * All previously invoked ring_buffer_read_prepare calls to prepare |
| 3528 | * iterators will be synchronized. Afterwards, read_buffer_read_start |
| 3529 | * calls on those iterators are allowed. |
| 3530 | */ |
| 3531 | void |
| 3532 | ring_buffer_read_prepare_sync(void) |
| 3533 | { |
| 3534 | synchronize_sched(); |
| 3535 | } |
| 3536 | EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync); |
| 3537 | |
| 3538 | /** |
| 3539 | * ring_buffer_read_start - start a non consuming read of the buffer |
| 3540 | * @iter: The iterator returned by ring_buffer_read_prepare |
| 3541 | * |
| 3542 | * This finalizes the startup of an iteration through the buffer. |
| 3543 | * The iterator comes from a call to ring_buffer_read_prepare and |
| 3544 | * an intervening ring_buffer_read_prepare_sync must have been |
| 3545 | * performed. |
| 3546 | * |
| 3547 | * Must be paired with ring_buffer_finish. |
| 3548 | */ |
| 3549 | void |
| 3550 | ring_buffer_read_start(struct ring_buffer_iter *iter) |
| 3551 | { |
| 3552 | struct ring_buffer_per_cpu *cpu_buffer; |
| 3553 | unsigned long flags; |
| 3554 | int locked; |
| 3555 | |
| 3556 | if (!iter) |
| 3557 | return; |
| 3558 | |
| 3559 | cpu_buffer = iter->cpu_buffer; |
| 3560 | |
| 3561 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 3562 | arch_spin_lock(&cpu_buffer->lock); |
| 3563 | rb_iter_reset(iter); |
| 3564 | arch_spin_unlock(&cpu_buffer->lock); |
| 3565 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 3566 | } |
| 3567 | EXPORT_SYMBOL_GPL(ring_buffer_read_start); |
| 3568 | |
| 3569 | /** |
| 3570 | * ring_buffer_finish - finish reading the iterator of the buffer |
| 3571 | * @iter: The iterator retrieved by ring_buffer_start |
| 3572 | * |
| 3573 | * This re-enables the recording to the buffer, and frees the |
| 3574 | * iterator. |
| 3575 | */ |
| 3576 | void |
| 3577 | ring_buffer_read_finish(struct ring_buffer_iter *iter) |
| 3578 | { |
| 3579 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; |
| 3580 | |
| 3581 | atomic_dec(&cpu_buffer->record_disabled); |
| 3582 | kfree(iter); |
| 3583 | } |
| 3584 | EXPORT_SYMBOL_GPL(ring_buffer_read_finish); |
| 3585 | |
| 3586 | /** |
| 3587 | * ring_buffer_read - read the next item in the ring buffer by the iterator |
| 3588 | * @iter: The ring buffer iterator |
| 3589 | * @ts: The time stamp of the event read. |
| 3590 | * |
| 3591 | * This reads the next event in the ring buffer and increments the iterator. |
| 3592 | */ |
| 3593 | struct ring_buffer_event * |
| 3594 | ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts) |
| 3595 | { |
| 3596 | struct ring_buffer_event *event; |
| 3597 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; |
| 3598 | unsigned long flags; |
| 3599 | int locked; |
| 3600 | |
| 3601 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 3602 | again: |
| 3603 | event = rb_iter_peek(iter, ts); |
| 3604 | if (!event) |
| 3605 | goto out; |
| 3606 | |
| 3607 | if (event->type_len == RINGBUF_TYPE_PADDING) |
| 3608 | goto again; |
| 3609 | |
| 3610 | rb_advance_iter(iter); |
| 3611 | out: |
| 3612 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 3613 | |
| 3614 | return event; |
| 3615 | } |
| 3616 | EXPORT_SYMBOL_GPL(ring_buffer_read); |
| 3617 | |
| 3618 | /** |
| 3619 | * ring_buffer_size - return the size of the ring buffer (in bytes) |
| 3620 | * @buffer: The ring buffer. |
| 3621 | */ |
| 3622 | unsigned long ring_buffer_size(struct ring_buffer *buffer) |
| 3623 | { |
| 3624 | return BUF_PAGE_SIZE * buffer->pages; |
| 3625 | } |
| 3626 | EXPORT_SYMBOL_GPL(ring_buffer_size); |
| 3627 | |
| 3628 | static void |
| 3629 | rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer) |
| 3630 | { |
| 3631 | rb_head_page_deactivate(cpu_buffer); |
| 3632 | |
| 3633 | cpu_buffer->head_page |
| 3634 | = list_entry(cpu_buffer->pages, struct buffer_page, list); |
| 3635 | local_set(&cpu_buffer->head_page->write, 0); |
| 3636 | local_set(&cpu_buffer->head_page->entries, 0); |
| 3637 | local_set(&cpu_buffer->head_page->page->commit, 0); |
| 3638 | |
| 3639 | cpu_buffer->head_page->read = 0; |
| 3640 | |
| 3641 | cpu_buffer->tail_page = cpu_buffer->head_page; |
| 3642 | cpu_buffer->commit_page = cpu_buffer->head_page; |
| 3643 | |
| 3644 | INIT_LIST_HEAD(&cpu_buffer->reader_page->list); |
| 3645 | local_set(&cpu_buffer->reader_page->write, 0); |
| 3646 | local_set(&cpu_buffer->reader_page->entries, 0); |
| 3647 | local_set(&cpu_buffer->reader_page->page->commit, 0); |
| 3648 | cpu_buffer->reader_page->read = 0; |
| 3649 | |
| 3650 | local_set(&cpu_buffer->commit_overrun, 0); |
| 3651 | local_set(&cpu_buffer->entries_bytes, 0); |
| 3652 | local_set(&cpu_buffer->overrun, 0); |
| 3653 | local_set(&cpu_buffer->entries, 0); |
| 3654 | local_set(&cpu_buffer->committing, 0); |
| 3655 | local_set(&cpu_buffer->commits, 0); |
| 3656 | cpu_buffer->read = 0; |
| 3657 | cpu_buffer->read_bytes = 0; |
| 3658 | |
| 3659 | cpu_buffer->write_stamp = 0; |
| 3660 | cpu_buffer->read_stamp = 0; |
| 3661 | |
| 3662 | cpu_buffer->lost_events = 0; |
| 3663 | cpu_buffer->last_overrun = 0; |
| 3664 | |
| 3665 | rb_head_page_activate(cpu_buffer); |
| 3666 | } |
| 3667 | |
| 3668 | /** |
| 3669 | * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer |
| 3670 | * @buffer: The ring buffer to reset a per cpu buffer of |
| 3671 | * @cpu: The CPU buffer to be reset |
| 3672 | */ |
| 3673 | void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu) |
| 3674 | { |
| 3675 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu]; |
| 3676 | unsigned long flags; |
| 3677 | int locked; |
| 3678 | |
| 3679 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 3680 | return; |
| 3681 | |
| 3682 | atomic_inc(&cpu_buffer->record_disabled); |
| 3683 | |
| 3684 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 3685 | |
| 3686 | if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing))) |
| 3687 | goto out; |
| 3688 | |
| 3689 | arch_spin_lock(&cpu_buffer->lock); |
| 3690 | |
| 3691 | rb_reset_cpu(cpu_buffer); |
| 3692 | |
| 3693 | arch_spin_unlock(&cpu_buffer->lock); |
| 3694 | |
| 3695 | out: |
| 3696 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 3697 | |
| 3698 | atomic_dec(&cpu_buffer->record_disabled); |
| 3699 | } |
| 3700 | EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu); |
| 3701 | |
| 3702 | /** |
| 3703 | * ring_buffer_reset - reset a ring buffer |
| 3704 | * @buffer: The ring buffer to reset all cpu buffers |
| 3705 | */ |
| 3706 | void ring_buffer_reset(struct ring_buffer *buffer) |
| 3707 | { |
| 3708 | int cpu; |
| 3709 | |
| 3710 | for_each_buffer_cpu(buffer, cpu) |
| 3711 | ring_buffer_reset_cpu(buffer, cpu); |
| 3712 | } |
| 3713 | EXPORT_SYMBOL_GPL(ring_buffer_reset); |
| 3714 | |
| 3715 | /** |
| 3716 | * rind_buffer_empty - is the ring buffer empty? |
| 3717 | * @buffer: The ring buffer to test |
| 3718 | */ |
| 3719 | int ring_buffer_empty(struct ring_buffer *buffer) |
| 3720 | { |
| 3721 | struct ring_buffer_per_cpu *cpu_buffer; |
| 3722 | unsigned long flags; |
| 3723 | int locked; |
| 3724 | int cpu; |
| 3725 | int ret; |
| 3726 | |
| 3727 | /* yes this is racy, but if you don't like the race, lock the buffer */ |
| 3728 | for_each_buffer_cpu(buffer, cpu) { |
| 3729 | cpu_buffer = buffer->buffers[cpu]; |
| 3730 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 3731 | ret = rb_per_cpu_empty(cpu_buffer); |
| 3732 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 3733 | |
| 3734 | if (!ret) |
| 3735 | return 0; |
| 3736 | } |
| 3737 | |
| 3738 | return 1; |
| 3739 | } |
| 3740 | EXPORT_SYMBOL_GPL(ring_buffer_empty); |
| 3741 | |
| 3742 | /** |
| 3743 | * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty? |
| 3744 | * @buffer: The ring buffer |
| 3745 | * @cpu: The CPU buffer to test |
| 3746 | */ |
| 3747 | int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu) |
| 3748 | { |
| 3749 | struct ring_buffer_per_cpu *cpu_buffer; |
| 3750 | unsigned long flags; |
| 3751 | int locked; |
| 3752 | int ret; |
| 3753 | |
| 3754 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 3755 | return 1; |
| 3756 | |
| 3757 | cpu_buffer = buffer->buffers[cpu]; |
| 3758 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 3759 | ret = rb_per_cpu_empty(cpu_buffer); |
| 3760 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 3761 | |
| 3762 | return ret; |
| 3763 | } |
| 3764 | EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu); |
| 3765 | |
| 3766 | #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP |
| 3767 | /** |
| 3768 | * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers |
| 3769 | * @buffer_a: One buffer to swap with |
| 3770 | * @buffer_b: The other buffer to swap with |
| 3771 | * |
| 3772 | * This function is useful for tracers that want to take a "snapshot" |
| 3773 | * of a CPU buffer and has another back up buffer lying around. |
| 3774 | * it is expected that the tracer handles the cpu buffer not being |
| 3775 | * used at the moment. |
| 3776 | */ |
| 3777 | int ring_buffer_swap_cpu(struct ring_buffer *buffer_a, |
| 3778 | struct ring_buffer *buffer_b, int cpu) |
| 3779 | { |
| 3780 | struct ring_buffer_per_cpu *cpu_buffer_a; |
| 3781 | struct ring_buffer_per_cpu *cpu_buffer_b; |
| 3782 | int ret = -EINVAL; |
| 3783 | |
| 3784 | if (!cpumask_test_cpu(cpu, buffer_a->cpumask) || |
| 3785 | !cpumask_test_cpu(cpu, buffer_b->cpumask)) |
| 3786 | goto out; |
| 3787 | |
| 3788 | /* At least make sure the two buffers are somewhat the same */ |
| 3789 | if (buffer_a->pages != buffer_b->pages) |
| 3790 | goto out; |
| 3791 | |
| 3792 | ret = -EAGAIN; |
| 3793 | |
| 3794 | if (ring_buffer_flags != RB_BUFFERS_ON) |
| 3795 | goto out; |
| 3796 | |
| 3797 | if (atomic_read(&buffer_a->record_disabled)) |
| 3798 | goto out; |
| 3799 | |
| 3800 | if (atomic_read(&buffer_b->record_disabled)) |
| 3801 | goto out; |
| 3802 | |
| 3803 | cpu_buffer_a = buffer_a->buffers[cpu]; |
| 3804 | cpu_buffer_b = buffer_b->buffers[cpu]; |
| 3805 | |
| 3806 | if (atomic_read(&cpu_buffer_a->record_disabled)) |
| 3807 | goto out; |
| 3808 | |
| 3809 | if (atomic_read(&cpu_buffer_b->record_disabled)) |
| 3810 | goto out; |
| 3811 | |
| 3812 | /* |
| 3813 | * We can't do a synchronize_sched here because this |
| 3814 | * function can be called in atomic context. |
| 3815 | * Normally this will be called from the same CPU as cpu. |
| 3816 | * If not it's up to the caller to protect this. |
| 3817 | */ |
| 3818 | atomic_inc(&cpu_buffer_a->record_disabled); |
| 3819 | atomic_inc(&cpu_buffer_b->record_disabled); |
| 3820 | |
| 3821 | ret = -EBUSY; |
| 3822 | if (local_read(&cpu_buffer_a->committing)) |
| 3823 | goto out_dec; |
| 3824 | if (local_read(&cpu_buffer_b->committing)) |
| 3825 | goto out_dec; |
| 3826 | |
| 3827 | buffer_a->buffers[cpu] = cpu_buffer_b; |
| 3828 | buffer_b->buffers[cpu] = cpu_buffer_a; |
| 3829 | |
| 3830 | cpu_buffer_b->buffer = buffer_a; |
| 3831 | cpu_buffer_a->buffer = buffer_b; |
| 3832 | |
| 3833 | ret = 0; |
| 3834 | |
| 3835 | out_dec: |
| 3836 | atomic_dec(&cpu_buffer_a->record_disabled); |
| 3837 | atomic_dec(&cpu_buffer_b->record_disabled); |
| 3838 | out: |
| 3839 | return ret; |
| 3840 | } |
| 3841 | EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu); |
| 3842 | #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */ |
| 3843 | |
| 3844 | /** |
| 3845 | * ring_buffer_alloc_read_page - allocate a page to read from buffer |
| 3846 | * @buffer: the buffer to allocate for. |
| 3847 | * |
| 3848 | * This function is used in conjunction with ring_buffer_read_page. |
| 3849 | * When reading a full page from the ring buffer, these functions |
| 3850 | * can be used to speed up the process. The calling function should |
| 3851 | * allocate a few pages first with this function. Then when it |
| 3852 | * needs to get pages from the ring buffer, it passes the result |
| 3853 | * of this function into ring_buffer_read_page, which will swap |
| 3854 | * the page that was allocated, with the read page of the buffer. |
| 3855 | * |
| 3856 | * Returns: |
| 3857 | * The page allocated, or NULL on error. |
| 3858 | */ |
| 3859 | void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu) |
| 3860 | { |
| 3861 | struct buffer_data_page *bpage; |
| 3862 | struct page *page; |
| 3863 | |
| 3864 | page = alloc_pages_node(cpu_to_node(cpu), |
| 3865 | GFP_KERNEL | __GFP_NORETRY, 0); |
| 3866 | if (!page) |
| 3867 | return NULL; |
| 3868 | |
| 3869 | bpage = page_address(page); |
| 3870 | |
| 3871 | rb_init_page(bpage); |
| 3872 | |
| 3873 | return bpage; |
| 3874 | } |
| 3875 | EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page); |
| 3876 | |
| 3877 | /** |
| 3878 | * ring_buffer_free_read_page - free an allocated read page |
| 3879 | * @buffer: the buffer the page was allocate for |
| 3880 | * @data: the page to free |
| 3881 | * |
| 3882 | * Free a page allocated from ring_buffer_alloc_read_page. |
| 3883 | */ |
| 3884 | void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data) |
| 3885 | { |
| 3886 | free_page((unsigned long)data); |
| 3887 | } |
| 3888 | EXPORT_SYMBOL_GPL(ring_buffer_free_read_page); |
| 3889 | |
| 3890 | /** |
| 3891 | * ring_buffer_read_page - extract a page from the ring buffer |
| 3892 | * @buffer: buffer to extract from |
| 3893 | * @data_page: the page to use allocated from ring_buffer_alloc_read_page |
| 3894 | * @len: amount to extract |
| 3895 | * @cpu: the cpu of the buffer to extract |
| 3896 | * @full: should the extraction only happen when the page is full. |
| 3897 | * |
| 3898 | * This function will pull out a page from the ring buffer and consume it. |
| 3899 | * @data_page must be the address of the variable that was returned |
| 3900 | * from ring_buffer_alloc_read_page. This is because the page might be used |
| 3901 | * to swap with a page in the ring buffer. |
| 3902 | * |
| 3903 | * for example: |
| 3904 | * rpage = ring_buffer_alloc_read_page(buffer); |
| 3905 | * if (!rpage) |
| 3906 | * return error; |
| 3907 | * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0); |
| 3908 | * if (ret >= 0) |
| 3909 | * process_page(rpage, ret); |
| 3910 | * |
| 3911 | * When @full is set, the function will not return true unless |
| 3912 | * the writer is off the reader page. |
| 3913 | * |
| 3914 | * Note: it is up to the calling functions to handle sleeps and wakeups. |
| 3915 | * The ring buffer can be used anywhere in the kernel and can not |
| 3916 | * blindly call wake_up. The layer that uses the ring buffer must be |
| 3917 | * responsible for that. |
| 3918 | * |
| 3919 | * Returns: |
| 3920 | * >=0 if data has been transferred, returns the offset of consumed data. |
| 3921 | * <0 if no data has been transferred. |
| 3922 | */ |
| 3923 | int ring_buffer_read_page(struct ring_buffer *buffer, |
| 3924 | void **data_page, size_t len, int cpu, int full) |
| 3925 | { |
| 3926 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu]; |
| 3927 | struct ring_buffer_event *event; |
| 3928 | struct buffer_data_page *bpage; |
| 3929 | struct buffer_page *reader; |
| 3930 | unsigned long missed_events; |
| 3931 | unsigned long flags; |
| 3932 | unsigned int commit; |
| 3933 | unsigned int read; |
| 3934 | u64 save_timestamp; |
| 3935 | int locked; |
| 3936 | int ret = -1; |
| 3937 | |
| 3938 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) |
| 3939 | goto out; |
| 3940 | |
| 3941 | /* |
| 3942 | * If len is not big enough to hold the page header, then |
| 3943 | * we can not copy anything. |
| 3944 | */ |
| 3945 | if (len <= BUF_PAGE_HDR_SIZE) |
| 3946 | goto out; |
| 3947 | |
| 3948 | len -= BUF_PAGE_HDR_SIZE; |
| 3949 | |
| 3950 | if (!data_page) |
| 3951 | goto out; |
| 3952 | |
| 3953 | bpage = *data_page; |
| 3954 | if (!bpage) |
| 3955 | goto out; |
| 3956 | |
| 3957 | locked = read_buffer_lock(cpu_buffer, &flags); |
| 3958 | |
| 3959 | reader = rb_get_reader_page(cpu_buffer); |
| 3960 | if (!reader) |
| 3961 | goto out_unlock; |
| 3962 | |
| 3963 | event = rb_reader_event(cpu_buffer); |
| 3964 | |
| 3965 | read = reader->read; |
| 3966 | commit = rb_page_commit(reader); |
| 3967 | |
| 3968 | /* Check if any events were dropped */ |
| 3969 | missed_events = cpu_buffer->lost_events; |
| 3970 | |
| 3971 | /* |
| 3972 | * If this page has been partially read or |
| 3973 | * if len is not big enough to read the rest of the page or |
| 3974 | * a writer is still on the page, then |
| 3975 | * we must copy the data from the page to the buffer. |
| 3976 | * Otherwise, we can simply swap the page with the one passed in. |
| 3977 | */ |
| 3978 | if (read || (len < (commit - read)) || |
| 3979 | cpu_buffer->reader_page == cpu_buffer->commit_page) { |
| 3980 | struct buffer_data_page *rpage = cpu_buffer->reader_page->page; |
| 3981 | unsigned int rpos = read; |
| 3982 | unsigned int pos = 0; |
| 3983 | unsigned int size; |
| 3984 | |
| 3985 | if (full) |
| 3986 | goto out_unlock; |
| 3987 | |
| 3988 | if (len > (commit - read)) |
| 3989 | len = (commit - read); |
| 3990 | |
| 3991 | /* Always keep the time extend and data together */ |
| 3992 | size = rb_event_ts_length(event); |
| 3993 | |
| 3994 | if (len < size) |
| 3995 | goto out_unlock; |
| 3996 | |
| 3997 | /* save the current timestamp, since the user will need it */ |
| 3998 | save_timestamp = cpu_buffer->read_stamp; |
| 3999 | |
| 4000 | /* Need to copy one event at a time */ |
| 4001 | do { |
| 4002 | /* We need the size of one event, because |
| 4003 | * rb_advance_reader only advances by one event, |
| 4004 | * whereas rb_event_ts_length may include the size of |
| 4005 | * one or two events. |
| 4006 | * We have already ensured there's enough space if this |
| 4007 | * is a time extend. */ |
| 4008 | size = rb_event_length(event); |
| 4009 | memcpy(bpage->data + pos, rpage->data + rpos, size); |
| 4010 | |
| 4011 | len -= size; |
| 4012 | |
| 4013 | rb_advance_reader(cpu_buffer); |
| 4014 | rpos = reader->read; |
| 4015 | pos += size; |
| 4016 | |
| 4017 | if (rpos >= commit) |
| 4018 | break; |
| 4019 | |
| 4020 | event = rb_reader_event(cpu_buffer); |
| 4021 | /* Always keep the time extend and data together */ |
| 4022 | size = rb_event_ts_length(event); |
| 4023 | } while (len >= size); |
| 4024 | |
| 4025 | /* update bpage */ |
| 4026 | local_set(&bpage->commit, pos); |
| 4027 | bpage->time_stamp = save_timestamp; |
| 4028 | |
| 4029 | /* we copied everything to the beginning */ |
| 4030 | read = 0; |
| 4031 | } else { |
| 4032 | /* update the entry counter */ |
| 4033 | cpu_buffer->read += rb_page_entries(reader); |
| 4034 | cpu_buffer->read_bytes += BUF_PAGE_SIZE; |
| 4035 | |
| 4036 | /* swap the pages */ |
| 4037 | rb_init_page(bpage); |
| 4038 | bpage = reader->page; |
| 4039 | reader->page = *data_page; |
| 4040 | local_set(&reader->write, 0); |
| 4041 | local_set(&reader->entries, 0); |
| 4042 | reader->read = 0; |
| 4043 | *data_page = bpage; |
| 4044 | |
| 4045 | /* |
| 4046 | * Use the real_end for the data size, |
| 4047 | * This gives us a chance to store the lost events |
| 4048 | * on the page. |
| 4049 | */ |
| 4050 | if (reader->real_end) |
| 4051 | local_set(&bpage->commit, reader->real_end); |
| 4052 | } |
| 4053 | ret = read; |
| 4054 | |
| 4055 | cpu_buffer->lost_events = 0; |
| 4056 | |
| 4057 | commit = local_read(&bpage->commit); |
| 4058 | /* |
| 4059 | * Set a flag in the commit field if we lost events |
| 4060 | */ |
| 4061 | if (missed_events) { |
| 4062 | /* If there is room at the end of the page to save the |
| 4063 | * missed events, then record it there. |
| 4064 | */ |
| 4065 | if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) { |
| 4066 | memcpy(&bpage->data[commit], &missed_events, |
| 4067 | sizeof(missed_events)); |
| 4068 | local_add(RB_MISSED_STORED, &bpage->commit); |
| 4069 | commit += sizeof(missed_events); |
| 4070 | } |
| 4071 | local_add(RB_MISSED_EVENTS, &bpage->commit); |
| 4072 | } |
| 4073 | |
| 4074 | /* |
| 4075 | * This page may be off to user land. Zero it out here. |
| 4076 | */ |
| 4077 | if (commit < BUF_PAGE_SIZE) |
| 4078 | memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit); |
| 4079 | |
| 4080 | out_unlock: |
| 4081 | read_buffer_unlock(cpu_buffer, flags, locked); |
| 4082 | |
| 4083 | out: |
| 4084 | return ret; |
| 4085 | } |
| 4086 | EXPORT_SYMBOL_GPL(ring_buffer_read_page); |
| 4087 | |
| 4088 | #ifdef CONFIG_HOTPLUG_CPU |
| 4089 | static int rb_cpu_notify(struct notifier_block *self, |
| 4090 | unsigned long action, void *hcpu) |
| 4091 | { |
| 4092 | struct ring_buffer *buffer = |
| 4093 | container_of(self, struct ring_buffer, cpu_notify); |
| 4094 | long cpu = (long)hcpu; |
| 4095 | |
| 4096 | switch (action) { |
| 4097 | case CPU_UP_PREPARE: |
| 4098 | case CPU_UP_PREPARE_FROZEN: |
| 4099 | if (cpumask_test_cpu(cpu, buffer->cpumask)) |
| 4100 | return NOTIFY_OK; |
| 4101 | |
| 4102 | buffer->buffers[cpu] = |
| 4103 | rb_allocate_cpu_buffer(buffer, cpu); |
| 4104 | if (!buffer->buffers[cpu]) { |
| 4105 | WARN(1, "failed to allocate ring buffer on CPU %ld\n", |
| 4106 | cpu); |
| 4107 | return NOTIFY_OK; |
| 4108 | } |
| 4109 | smp_wmb(); |
| 4110 | cpumask_set_cpu(cpu, buffer->cpumask); |
| 4111 | break; |
| 4112 | case CPU_DOWN_PREPARE: |
| 4113 | case CPU_DOWN_PREPARE_FROZEN: |
| 4114 | /* |
| 4115 | * Do nothing. |
| 4116 | * If we were to free the buffer, then the user would |
| 4117 | * lose any trace that was in the buffer. |
| 4118 | */ |
| 4119 | break; |
| 4120 | default: |
| 4121 | break; |
| 4122 | } |
| 4123 | return NOTIFY_OK; |
| 4124 | } |
| 4125 | #endif |