| 1 | |
| 2 | * Generic ring buffer
|
| 3 | *
|
| 4 | * Copyright (C) 2008 Steven Rostedt <[email protected]>
|
| 5 |
|
| 6 | #include <linux/ring_buffer.h>
|
| 7 | #include <linux/trace_clock.h>
|
| 8 | #include <linux/ftrace_irq.h>
|
| 9 | #include <linux/spinlock.h>
|
| 10 | #include <linux/debugfs.h>
|
| 11 | #include <linux/uaccess.h>
|
| 12 | #include <linux/hardirq.h>
|
| 13 | #include <linux/kmemcheck.h>
|
| 14 | #include <linux/module.h>
|
| 15 | #include <linux/percpu.h>
|
| 16 | #include <linux/mutex.h>
|
| 17 | #include <linux/slab.h>
|
| 18 | #include <linux/init.h>
|
| 19 | #include <linux/hash.h>
|
| 20 | #include <linux/list.h>
|
| 21 | #include <linux/cpu.h>
|
| 22 | #include <linux/fs.h>
|
| 23 |
|
| 24 | #include <asm/local.h>
|
| 25 | #include "trace.h"
|
| 26 |
|
| 27 | |
| 28 | * The ring buffer header is special. We must manually up keep it.
|
| 29 |
|
| 30 | int ring_buffer_print_entry_header(struct trace_seq *s)
|
| 31 | {
|
| 32 | int ret;
|
| 33 |
|
| 34 | ret = trace_seq_printf(s, "# compressed entry header\n");
|
| 35 | ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
|
| 36 | ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
|
| 37 | ret = trace_seq_printf(s, "\tarray : 32 bits\n");
|
| 38 | ret = trace_seq_printf(s, "\n");
|
| 39 | ret = trace_seq_printf(s, "\tpadding : type == %d\n",
|
| 40 | RINGBUF_TYPE_PADDING);
|
| 41 | ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
|
| 42 | RINGBUF_TYPE_TIME_EXTEND);
|
| 43 | ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
|
| 44 | RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
|
| 45 |
|
| 46 | return ret;
|
| 47 | }
|
| 48 |
|
| 49 | |
| 50 | * The ring buffer is made up of a list of pages. A separate list of pages is
|
| 51 | * allocated for each CPU. A writer may only write to a buffer that is
|
| 52 | * associated with the CPU it is currently executing on. A reader may read
|
| 53 | * from any per cpu buffer.
|
| 54 | *
|
| 55 | * The reader is special. For each per cpu buffer, the reader has its own
|
| 56 | * reader page. When a reader has read the entire reader page, this reader
|
| 57 | * page is swapped with another page in the ring buffer.
|
| 58 | *
|
| 59 | * Now, as long as the writer is off the reader page, the reader can do what
|
| 60 | * ever it wants with that page. The writer will never write to that page
|
| 61 | * again (as long as it is out of the ring buffer).
|
| 62 | *
|
| 63 | * Here's some silly ASCII art.
|
| 64 | *
|
| 65 | * +------+
|
| 66 | * |reader| RING BUFFER
|
| 67 | * |page |
|
| 68 | * +------+ +---+ +---+ +---+
|
| 69 | * | |-->| |-->| |
|
| 70 | * +---+ +---+ +---+
|
| 71 | * ^ |
|
| 72 | * | |
|
| 73 | * +---------------+
|
| 74 | *
|
| 75 | *
|
| 76 | * +------+
|
| 77 | * |reader| RING BUFFER
|
| 78 | * |page |------------------v
|
| 79 | * +------+ +---+ +---+ +---+
|
| 80 | * | |-->| |-->| |
|
| 81 | * +---+ +---+ +---+
|
| 82 | * ^ |
|
| 83 | * | |
|
| 84 | * +---------------+
|
| 85 | *
|
| 86 | *
|
| 87 | * +------+
|
| 88 | * |reader| RING BUFFER
|
| 89 | * |page |------------------v
|
| 90 | * +------+ +---+ +---+ +---+
|
| 91 | * ^ | |-->| |-->| |
|
| 92 | * | +---+ +---+ +---+
|
| 93 | * | |
|
| 94 | * | |
|
| 95 | * +------------------------------+
|
| 96 | *
|
| 97 | *
|
| 98 | * +------+
|
| 99 | * |buffer| RING BUFFER
|
| 100 | * |page |------------------v
|
| 101 | * +------+ +---+ +---+ +---+
|
| 102 | * ^ | | | |-->| |
|
| 103 | * | New +---+ +---+ +---+
|
| 104 | * | Reader------^ |
|
| 105 | * | page |
|
| 106 | * +------------------------------+
|
| 107 | *
|
| 108 | *
|
| 109 | * After we make this swap, the reader can hand this page off to the splice
|
| 110 | * code and be done with it. It can even allocate a new page if it needs to
|
| 111 | * and swap that into the ring buffer.
|
| 112 | *
|
| 113 | * We will be using cmpxchg soon to make all this lockless.
|
| 114 | *
|
| 115 |
|
| 116 |
|
| 117 | |
| 118 | * A fast way to enable or disable all ring buffers is to
|
| 119 | * call tracing_on or tracing_off. Turning off the ring buffers
|
| 120 | * prevents all ring buffers from being recorded to.
|
| 121 | * Turning this switch on, makes it OK to write to the
|
| 122 | * ring buffer, if the ring buffer is enabled itself.
|
| 123 | *
|
| 124 | * There's three layers that must be on in order to write
|
| 125 | * to the ring buffer.
|
| 126 | *
|
| 127 | * 1) This global flag must be set.
|
| 128 | * 2) The ring buffer must be enabled for recording.
|
| 129 | * 3) The per cpu buffer must be enabled for recording.
|
| 130 | *
|
| 131 | * In case of an anomaly, this global flag has a bit set that
|
| 132 | * will permantly disable all ring buffers.
|
| 133 |
|
| 134 |
|
| 135 | |
| 136 | * Global flag to disable all recording to ring buffers
|
| 137 | * This has two bits: ON, DISABLED
|
| 138 | *
|
| 139 | * ON DISABLED
|
| 140 | * ---- ----------
|
| 141 | * 0 0 : ring buffers are off
|
| 142 | * 1 0 : ring buffers are on
|
| 143 | * X 1 : ring buffers are permanently disabled
|
| 144 |
|
| 145 |
|
| 146 | enum {
|
| 147 | RB_BUFFERS_ON_BIT = 0,
|
| 148 | RB_BUFFERS_DISABLED_BIT = 1,
|
| 149 | };
|
| 150 |
|
| 151 | enum {
|
| 152 | RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
|
| 153 | RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
|
| 154 | };
|
| 155 |
|
| 156 | static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
|
| 157 |
|
| 158 | #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
|
| 159 |
|
| 160 | |
| 161 | * tracing_on - enable all tracing buffers
|
| 162 | *
|
| 163 | * This function enables all tracing buffers that may have been
|
| 164 | * disabled with tracing_off.
|
| 165 |
|
| 166 | void tracing_on(void)
|
| 167 | {
|
| 168 | set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
|
| 169 | }
|
| 170 | EXPORT_SYMBOL_GPL(tracing_on);
|
| 171 |
|
| 172 | |
| 173 | * tracing_off - turn off all tracing buffers
|
| 174 | *
|
| 175 | * This function stops all tracing buffers from recording data.
|
| 176 | * It does not disable any overhead the tracers themselves may
|
| 177 | * be causing. This function simply causes all recording to
|
| 178 | * the ring buffers to fail.
|
| 179 |
|
| 180 | void tracing_off(void)
|
| 181 | {
|
| 182 | clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
|
| 183 | }
|
| 184 | EXPORT_SYMBOL_GPL(tracing_off);
|
| 185 |
|
| 186 | |
| 187 | * tracing_off_permanent - permanently disable ring buffers
|
| 188 | *
|
| 189 | * This function, once called, will disable all ring buffers
|
| 190 | * permanently.
|
| 191 |
|
| 192 | void tracing_off_permanent(void)
|
| 193 | {
|
| 194 | set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
|
| 195 | }
|
| 196 |
|
| 197 | |
| 198 | * tracing_is_on - show state of ring buffers enabled
|
| 199 |
|
| 200 | int tracing_is_on(void)
|
| 201 | {
|
| 202 | return ring_buffer_flags == RB_BUFFERS_ON;
|
| 203 | }
|
| 204 | EXPORT_SYMBOL_GPL(tracing_is_on);
|
| 205 |
|
| 206 | #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
|
| 207 | #define RB_ALIGNMENT 4U
|
| 208 | #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
|
| 209 | #define RB_EVNT_MIN_SIZE 8U
|
| 210 |
|
| 211 | #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
|
| 212 | # define RB_FORCE_8BYTE_ALIGNMENT 0
|
| 213 | # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
|
| 214 | #else
|
| 215 | # define RB_FORCE_8BYTE_ALIGNMENT 1
|
| 216 | # define RB_ARCH_ALIGNMENT 8U
|
| 217 | #endif
|
| 218 |
|
| 219 |
|
| 220 | #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
|
| 221 |
|
| 222 | enum {
|
| 223 | RB_LEN_TIME_EXTEND = 8,
|
| 224 | RB_LEN_TIME_STAMP = 16,
|
| 225 | };
|
| 226 |
|
| 227 | static inline int rb_null_event(struct ring_buffer_event *event)
|
| 228 | {
|
| 229 | return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
|
| 230 | }
|
| 231 |
|
| 232 | static void rb_event_set_padding(struct ring_buffer_event *event)
|
| 233 | {
|
| 234 |
|
| 235 | event->type_len = RINGBUF_TYPE_PADDING;
|
| 236 | event->time_delta = 0;
|
| 237 | }
|
| 238 |
|
| 239 | static unsigned
|
| 240 | rb_event_data_length(struct ring_buffer_event *event)
|
| 241 | {
|
| 242 | unsigned length;
|
| 243 |
|
| 244 | if (event->type_len)
|
| 245 | length = event->type_len * RB_ALIGNMENT;
|
| 246 | else
|
| 247 | length = event->array[0];
|
| 248 | return length + RB_EVNT_HDR_SIZE;
|
| 249 | }
|
| 250 |
|
| 251 |
|
| 252 | static unsigned
|
| 253 | rb_event_length(struct ring_buffer_event *event)
|
| 254 | {
|
| 255 | switch (event->type_len) {
|
| 256 | case RINGBUF_TYPE_PADDING:
|
| 257 | if (rb_null_event(event))
|
| 258 |
|
| 259 | return -1;
|
| 260 | return event->array[0] + RB_EVNT_HDR_SIZE;
|
| 261 |
|
| 262 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 263 | return RB_LEN_TIME_EXTEND;
|
| 264 |
|
| 265 | case RINGBUF_TYPE_TIME_STAMP:
|
| 266 | return RB_LEN_TIME_STAMP;
|
| 267 |
|
| 268 | case RINGBUF_TYPE_DATA:
|
| 269 | return rb_event_data_length(event);
|
| 270 | default:
|
| 271 | BUG();
|
| 272 | }
|
| 273 |
|
| 274 | return 0;
|
| 275 | }
|
| 276 |
|
| 277 | |
| 278 | * ring_buffer_event_length - return the length of the event
|
| 279 | * @event: the event to get the length of
|
| 280 |
|
| 281 | unsigned ring_buffer_event_length(struct ring_buffer_event *event)
|
| 282 | {
|
| 283 | unsigned length = rb_event_length(event);
|
| 284 | if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
|
| 285 | return length;
|
| 286 | length -= RB_EVNT_HDR_SIZE;
|
| 287 | if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
|
| 288 | length -= sizeof(event->array[0]);
|
| 289 | return length;
|
| 290 | }
|
| 291 | EXPORT_SYMBOL_GPL(ring_buffer_event_length);
|
| 292 |
|
| 293 |
|
| 294 | static void *
|
| 295 | rb_event_data(struct ring_buffer_event *event)
|
| 296 | {
|
| 297 | BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
|
| 298 |
|
| 299 | if (event->type_len)
|
| 300 | return (void *)&event->array[0];
|
| 301 |
|
| 302 | return (void *)&event->array[1];
|
| 303 | }
|
| 304 |
|
| 305 | |
| 306 | * ring_buffer_event_data - return the data of the event
|
| 307 | * @event: the event to get the data from
|
| 308 |
|
| 309 | void *ring_buffer_event_data(struct ring_buffer_event *event)
|
| 310 | {
|
| 311 | return rb_event_data(event);
|
| 312 | }
|
| 313 | EXPORT_SYMBOL_GPL(ring_buffer_event_data);
|
| 314 |
|
| 315 | #define for_each_buffer_cpu(buffer, cpu) \
|
| 316 | for_each_cpu(cpu, buffer->cpumask)
|
| 317 |
|
| 318 | #define TS_SHIFT 27
|
| 319 | #define TS_MASK ((1ULL << TS_SHIFT) - 1)
|
| 320 | #define TS_DELTA_TEST (~TS_MASK)
|
| 321 |
|
| 322 |
|
| 323 | #define RB_MISSED_EVENTS (1 << 31)
|
| 324 |
|
| 325 | #define RB_MISSED_STORED (1 << 30)
|
| 326 |
|
| 327 | struct buffer_data_page {
|
| 328 | u64 time_stamp;
|
| 329 | local_t commit;
|
| 330 | unsigned char data[];
|
| 331 | };
|
| 332 |
|
| 333 | |
| 334 | * Note, the buffer_page list must be first. The buffer pages
|
| 335 | * are allocated in cache lines, which means that each buffer
|
| 336 | * page will be at the beginning of a cache line, and thus
|
| 337 | * the least significant bits will be zero. We use this to
|
| 338 | * add flags in the list struct pointers, to make the ring buffer
|
| 339 | * lockless.
|
| 340 |
|
| 341 | struct buffer_page {
|
| 342 | struct list_head list;
|
| 343 | local_t write;
|
| 344 | unsigned read;
|
| 345 | local_t entries;
|
| 346 | unsigned long real_end;
|
| 347 | struct buffer_data_page *page;
|
| 348 | };
|
| 349 |
|
| 350 | |
| 351 | * The buffer page counters, write and entries, must be reset
|
| 352 | * atomically when crossing page boundaries. To synchronize this
|
| 353 | * update, two counters are inserted into the number. One is
|
| 354 | * the actual counter for the write position or count on the page.
|
| 355 | *
|
| 356 | * The other is a counter of updaters. Before an update happens
|
| 357 | * the update partition of the counter is incremented. This will
|
| 358 | * allow the updater to update the counter atomically.
|
| 359 | *
|
| 360 | * The counter is 20 bits, and the state data is 12.
|
| 361 |
|
| 362 | #define RB_WRITE_MASK 0xfffff
|
| 363 | #define RB_WRITE_INTCNT (1 << 20)
|
| 364 |
|
| 365 | static void rb_init_page(struct buffer_data_page *bpage)
|
| 366 | {
|
| 367 | local_set(&bpage->commit, 0);
|
| 368 | }
|
| 369 |
|
| 370 | |
| 371 | * ring_buffer_page_len - the size of data on the page.
|
| 372 | * @page: The page to read
|
| 373 | *
|
| 374 | * Returns the amount of data on the page, including buffer page header.
|
| 375 |
|
| 376 | size_t ring_buffer_page_len(void *page)
|
| 377 | {
|
| 378 | return local_read(&((struct buffer_data_page *)page)->commit)
|
| 379 | + BUF_PAGE_HDR_SIZE;
|
| 380 | }
|
| 381 |
|
| 382 | |
| 383 | * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
|
| 384 | * this issue out.
|
| 385 |
|
| 386 | static void free_buffer_page(struct buffer_page *bpage)
|
| 387 | {
|
| 388 | free_page((unsigned long)bpage->page);
|
| 389 | kfree(bpage);
|
| 390 | }
|
| 391 |
|
| 392 | |
| 393 | * We need to fit the time_stamp delta into 27 bits.
|
| 394 |
|
| 395 | static inline int test_time_stamp(u64 delta)
|
| 396 | {
|
| 397 | if (delta & TS_DELTA_TEST)
|
| 398 | return 1;
|
| 399 | return 0;
|
| 400 | }
|
| 401 |
|
| 402 | #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
|
| 403 |
|
| 404 |
|
| 405 | #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
|
| 406 |
|
| 407 |
|
| 408 | #define RB_TIMESTAMPS_PER_PAGE (BUF_PAGE_SIZE / RB_LEN_TIME_STAMP)
|
| 409 |
|
| 410 | int ring_buffer_print_page_header(struct trace_seq *s)
|
| 411 | {
|
| 412 | struct buffer_data_page field;
|
| 413 | int ret;
|
| 414 |
|
| 415 | ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
|
| 416 | "offset:0;\tsize:%u;\tsigned:%u;\n",
|
| 417 | (unsigned int)sizeof(field.time_stamp),
|
| 418 | (unsigned int)is_signed_type(u64));
|
| 419 |
|
| 420 | ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
|
| 421 | "offset:%u;\tsize:%u;\tsigned:%u;\n",
|
| 422 | (unsigned int)offsetof(typeof(field), commit),
|
| 423 | (unsigned int)sizeof(field.commit),
|
| 424 | (unsigned int)is_signed_type(long));
|
| 425 |
|
| 426 | ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
|
| 427 | "offset:%u;\tsize:%u;\tsigned:%u;\n",
|
| 428 | (unsigned int)offsetof(typeof(field), commit),
|
| 429 | 1,
|
| 430 | (unsigned int)is_signed_type(long));
|
| 431 |
|
| 432 | ret = trace_seq_printf(s, "\tfield: char data;\t"
|
| 433 | "offset:%u;\tsize:%u;\tsigned:%u;\n",
|
| 434 | (unsigned int)offsetof(typeof(field), data),
|
| 435 | (unsigned int)BUF_PAGE_SIZE,
|
| 436 | (unsigned int)is_signed_type(char));
|
| 437 |
|
| 438 | return ret;
|
| 439 | }
|
| 440 |
|
| 441 | |
| 442 | * head_page == tail_page && head == tail then buffer is empty.
|
| 443 |
|
| 444 | struct ring_buffer_per_cpu {
|
| 445 | int cpu;
|
| 446 | struct ring_buffer *buffer;
|
| 447 | spinlock_t reader_lock;
|
| 448 | arch_spinlock_t lock;
|
| 449 | struct lock_class_key lock_key;
|
| 450 | struct list_head *pages;
|
| 451 | struct buffer_page *head_page;
|
| 452 | struct buffer_page *tail_page;
|
| 453 | struct buffer_page *commit_page;
|
| 454 | struct buffer_page *reader_page;
|
| 455 | unsigned long lost_events;
|
| 456 | unsigned long last_overrun;
|
| 457 | local_t commit_overrun;
|
| 458 | local_t overrun;
|
| 459 | local_t entries;
|
| 460 | local_t committing;
|
| 461 | local_t commits;
|
| 462 | unsigned long read;
|
| 463 | u64 write_stamp;
|
| 464 | u64 read_stamp;
|
| 465 | atomic_t record_disabled;
|
| 466 | };
|
| 467 |
|
| 468 | struct ring_buffer {
|
| 469 | unsigned pages;
|
| 470 | unsigned flags;
|
| 471 | int cpus;
|
| 472 | atomic_t record_disabled;
|
| 473 | cpumask_var_t cpumask;
|
| 474 |
|
| 475 | struct lock_class_key *reader_lock_key;
|
| 476 |
|
| 477 | struct mutex mutex;
|
| 478 |
|
| 479 | struct ring_buffer_per_cpu **buffers;
|
| 480 |
|
| 481 | #ifdef CONFIG_HOTPLUG_CPU
|
| 482 | struct notifier_block cpu_notify;
|
| 483 | #endif
|
| 484 | u64 (*clock)(void);
|
| 485 | };
|
| 486 |
|
| 487 | struct ring_buffer_iter {
|
| 488 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 489 | unsigned long head;
|
| 490 | struct buffer_page *head_page;
|
| 491 | struct buffer_page *cache_reader_page;
|
| 492 | unsigned long cache_read;
|
| 493 | u64 read_stamp;
|
| 494 | };
|
| 495 |
|
| 496 |
|
| 497 | #define RB_WARN_ON(b, cond) \
|
| 498 | ({ \
|
| 499 | int _____ret = unlikely(cond); \
|
| 500 | if (_____ret) { \
|
| 501 | if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
|
| 502 | struct ring_buffer_per_cpu *__b = \
|
| 503 | (void *)b; \
|
| 504 | atomic_inc(&__b->buffer->record_disabled); \
|
| 505 | } else \
|
| 506 | atomic_inc(&b->record_disabled); \
|
| 507 | WARN_ON(1); \
|
| 508 | } \
|
| 509 | _____ret; \
|
| 510 | })
|
| 511 |
|
| 512 |
|
| 513 | #define DEBUG_SHIFT 0
|
| 514 |
|
| 515 | static inline u64 rb_time_stamp(struct ring_buffer *buffer)
|
| 516 | {
|
| 517 |
|
| 518 | return buffer->clock() << DEBUG_SHIFT;
|
| 519 | }
|
| 520 |
|
| 521 | u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
|
| 522 | {
|
| 523 | u64 time;
|
| 524 |
|
| 525 | preempt_disable_notrace();
|
| 526 | time = rb_time_stamp(buffer);
|
| 527 | preempt_enable_no_resched_notrace();
|
| 528 |
|
| 529 | return time;
|
| 530 | }
|
| 531 | EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
|
| 532 |
|
| 533 | void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
|
| 534 | int cpu, u64 *ts)
|
| 535 | {
|
| 536 |
|
| 537 | *ts >>= DEBUG_SHIFT;
|
| 538 | }
|
| 539 | EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
|
| 540 |
|
| 541 | |
| 542 | * Making the ring buffer lockless makes things tricky.
|
| 543 | * Although writes only happen on the CPU that they are on,
|
| 544 | * and they only need to worry about interrupts. Reads can
|
| 545 | * happen on any CPU.
|
| 546 | *
|
| 547 | * The reader page is always off the ring buffer, but when the
|
| 548 | * reader finishes with a page, it needs to swap its page with
|
| 549 | * a new one from the buffer. The reader needs to take from
|
| 550 | * the head (writes go to the tail). But if a writer is in overwrite
|
| 551 | * mode and wraps, it must push the head page forward.
|
| 552 | *
|
| 553 | * Here lies the problem.
|
| 554 | *
|
| 555 | * The reader must be careful to replace only the head page, and
|
| 556 | * not another one. As described at the top of the file in the
|
| 557 | * ASCII art, the reader sets its old page to point to the next
|
| 558 | * page after head. It then sets the page after head to point to
|
| 559 | * the old reader page. But if the writer moves the head page
|
| 560 | * during this operation, the reader could end up with the tail.
|
| 561 | *
|
| 562 | * We use cmpxchg to help prevent this race. We also do something
|
| 563 | * special with the page before head. We set the LSB to 1.
|
| 564 | *
|
| 565 | * When the writer must push the page forward, it will clear the
|
| 566 | * bit that points to the head page, move the head, and then set
|
| 567 | * the bit that points to the new head page.
|
| 568 | *
|
| 569 | * We also don't want an interrupt coming in and moving the head
|
| 570 | * page on another writer. Thus we use the second LSB to catch
|
| 571 | * that too. Thus:
|
| 572 | *
|
| 573 | * head->list->prev->next bit 1 bit 0
|
| 574 | * ------- -------
|
| 575 | * Normal page 0 0
|
| 576 | * Points to head page 0 1
|
| 577 | * New head page 1 0
|
| 578 | *
|
| 579 | * Note we can not trust the prev pointer of the head page, because:
|
| 580 | *
|
| 581 | * +----+ +-----+ +-----+
|
| 582 | * | |------>| T |---X--->| N |
|
| 583 | * | |<------| | | |
|
| 584 | * +----+ +-----+ +-----+
|
| 585 | * ^ ^ |
|
| 586 | * | +-----+ | |
|
| 587 | * +----------| R |----------+ |
|
| 588 | * | |<-----------+
|
| 589 | * +-----+
|
| 590 | *
|
| 591 | * Key: ---X--> HEAD flag set in pointer
|
| 592 | * T Tail page
|
| 593 | * R Reader page
|
| 594 | * N Next page
|
| 595 | *
|
| 596 | * (see __rb_reserve_next() to see where this happens)
|
| 597 | *
|
| 598 | * What the above shows is that the reader just swapped out
|
| 599 | * the reader page with a page in the buffer, but before it
|
| 600 | * could make the new header point back to the new page added
|
| 601 | * it was preempted by a writer. The writer moved forward onto
|
| 602 | * the new page added by the reader and is about to move forward
|
| 603 | * again.
|
| 604 | *
|
| 605 | * You can see, it is legitimate for the previous pointer of
|
| 606 | * the head (or any page) not to point back to itself. But only
|
| 607 | * temporarially.
|
| 608 |
|
| 609 |
|
| 610 | #define RB_PAGE_NORMAL 0UL
|
| 611 | #define RB_PAGE_HEAD 1UL
|
| 612 | #define RB_PAGE_UPDATE 2UL
|
| 613 |
|
| 614 |
|
| 615 | #define RB_FLAG_MASK 3UL
|
| 616 |
|
| 617 |
|
| 618 | #define RB_PAGE_MOVED 4UL
|
| 619 |
|
| 620 | |
| 621 | * rb_list_head - remove any bit
|
| 622 |
|
| 623 | static struct list_head *rb_list_head(struct list_head *list)
|
| 624 | {
|
| 625 | unsigned long val = (unsigned long)list;
|
| 626 |
|
| 627 | return (struct list_head *)(val & ~RB_FLAG_MASK);
|
| 628 | }
|
| 629 |
|
| 630 | |
| 631 | * rb_is_head_page - test if the given page is the head page
|
| 632 | *
|
| 633 | * Because the reader may move the head_page pointer, we can
|
| 634 | * not trust what the head page is (it may be pointing to
|
| 635 | * the reader page). But if the next page is a header page,
|
| 636 | * its flags will be non zero.
|
| 637 |
|
| 638 | static int inline
|
| 639 | rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
|
| 640 | struct buffer_page *page, struct list_head *list)
|
| 641 | {
|
| 642 | unsigned long val;
|
| 643 |
|
| 644 | val = (unsigned long)list->next;
|
| 645 |
|
| 646 | if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
|
| 647 | return RB_PAGE_MOVED;
|
| 648 |
|
| 649 | return val & RB_FLAG_MASK;
|
| 650 | }
|
| 651 |
|
| 652 | |
| 653 | * rb_is_reader_page
|
| 654 | *
|
| 655 | * The unique thing about the reader page, is that, if the
|
| 656 | * writer is ever on it, the previous pointer never points
|
| 657 | * back to the reader page.
|
| 658 |
|
| 659 | static int rb_is_reader_page(struct buffer_page *page)
|
| 660 | {
|
| 661 | struct list_head *list = page->list.prev;
|
| 662 |
|
| 663 | return rb_list_head(list->next) != &page->list;
|
| 664 | }
|
| 665 |
|
| 666 | |
| 667 | * rb_set_list_to_head - set a list_head to be pointing to head.
|
| 668 |
|
| 669 | static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
|
| 670 | struct list_head *list)
|
| 671 | {
|
| 672 | unsigned long *ptr;
|
| 673 |
|
| 674 | ptr = (unsigned long *)&list->next;
|
| 675 | *ptr |= RB_PAGE_HEAD;
|
| 676 | *ptr &= ~RB_PAGE_UPDATE;
|
| 677 | }
|
| 678 |
|
| 679 | |
| 680 | * rb_head_page_activate - sets up head page
|
| 681 |
|
| 682 | static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
|
| 683 | {
|
| 684 | struct buffer_page *head;
|
| 685 |
|
| 686 | head = cpu_buffer->head_page;
|
| 687 | if (!head)
|
| 688 | return;
|
| 689 |
|
| 690 | |
| 691 | * Set the previous list pointer to have the HEAD flag.
|
| 692 |
|
| 693 | rb_set_list_to_head(cpu_buffer, head->list.prev);
|
| 694 | }
|
| 695 |
|
| 696 | static void rb_list_head_clear(struct list_head *list)
|
| 697 | {
|
| 698 | unsigned long *ptr = (unsigned long *)&list->next;
|
| 699 |
|
| 700 | *ptr &= ~RB_FLAG_MASK;
|
| 701 | }
|
| 702 |
|
| 703 | |
| 704 | * rb_head_page_dactivate - clears head page ptr (for free list)
|
| 705 |
|
| 706 | static void
|
| 707 | rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
|
| 708 | {
|
| 709 | struct list_head *hd;
|
| 710 |
|
| 711 |
|
| 712 | rb_list_head_clear(cpu_buffer->pages);
|
| 713 |
|
| 714 | list_for_each(hd, cpu_buffer->pages)
|
| 715 | rb_list_head_clear(hd);
|
| 716 | }
|
| 717 |
|
| 718 | static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
|
| 719 | struct buffer_page *head,
|
| 720 | struct buffer_page *prev,
|
| 721 | int old_flag, int new_flag)
|
| 722 | {
|
| 723 | struct list_head *list;
|
| 724 | unsigned long val = (unsigned long)&head->list;
|
| 725 | unsigned long ret;
|
| 726 |
|
| 727 | list = &prev->list;
|
| 728 |
|
| 729 | val &= ~RB_FLAG_MASK;
|
| 730 |
|
| 731 | ret = cmpxchg((unsigned long *)&list->next,
|
| 732 | val | old_flag, val | new_flag);
|
| 733 |
|
| 734 |
|
| 735 | if ((ret & ~RB_FLAG_MASK) != val)
|
| 736 | return RB_PAGE_MOVED;
|
| 737 |
|
| 738 | return ret & RB_FLAG_MASK;
|
| 739 | }
|
| 740 |
|
| 741 | static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
|
| 742 | struct buffer_page *head,
|
| 743 | struct buffer_page *prev,
|
| 744 | int old_flag)
|
| 745 | {
|
| 746 | return rb_head_page_set(cpu_buffer, head, prev,
|
| 747 | old_flag, RB_PAGE_UPDATE);
|
| 748 | }
|
| 749 |
|
| 750 | static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
|
| 751 | struct buffer_page *head,
|
| 752 | struct buffer_page *prev,
|
| 753 | int old_flag)
|
| 754 | {
|
| 755 | return rb_head_page_set(cpu_buffer, head, prev,
|
| 756 | old_flag, RB_PAGE_HEAD);
|
| 757 | }
|
| 758 |
|
| 759 | static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
|
| 760 | struct buffer_page *head,
|
| 761 | struct buffer_page *prev,
|
| 762 | int old_flag)
|
| 763 | {
|
| 764 | return rb_head_page_set(cpu_buffer, head, prev,
|
| 765 | old_flag, RB_PAGE_NORMAL);
|
| 766 | }
|
| 767 |
|
| 768 | static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
|
| 769 | struct buffer_page **bpage)
|
| 770 | {
|
| 771 | struct list_head *p = rb_list_head((*bpage)->list.next);
|
| 772 |
|
| 773 | *bpage = list_entry(p, struct buffer_page, list);
|
| 774 | }
|
| 775 |
|
| 776 | static struct buffer_page *
|
| 777 | rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
|
| 778 | {
|
| 779 | struct buffer_page *head;
|
| 780 | struct buffer_page *page;
|
| 781 | struct list_head *list;
|
| 782 | int i;
|
| 783 |
|
| 784 | if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
|
| 785 | return NULL;
|
| 786 |
|
| 787 |
|
| 788 | list = cpu_buffer->pages;
|
| 789 | if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
|
| 790 | return NULL;
|
| 791 |
|
| 792 | page = head = cpu_buffer->head_page;
|
| 793 | |
| 794 | * It is possible that the writer moves the header behind
|
| 795 | * where we started, and we miss in one loop.
|
| 796 | * A second loop should grab the header, but we'll do
|
| 797 | * three loops just because I'm paranoid.
|
| 798 |
|
| 799 | for (i = 0; i < 3; i++) {
|
| 800 | do {
|
| 801 | if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
|
| 802 | cpu_buffer->head_page = page;
|
| 803 | return page;
|
| 804 | }
|
| 805 | rb_inc_page(cpu_buffer, &page);
|
| 806 | } while (page != head);
|
| 807 | }
|
| 808 |
|
| 809 | RB_WARN_ON(cpu_buffer, 1);
|
| 810 |
|
| 811 | return NULL;
|
| 812 | }
|
| 813 |
|
| 814 | static int rb_head_page_replace(struct buffer_page *old,
|
| 815 | struct buffer_page *new)
|
| 816 | {
|
| 817 | unsigned long *ptr = (unsigned long *)&old->list.prev->next;
|
| 818 | unsigned long val;
|
| 819 | unsigned long ret;
|
| 820 |
|
| 821 | val = *ptr & ~RB_FLAG_MASK;
|
| 822 | val |= RB_PAGE_HEAD;
|
| 823 |
|
| 824 | ret = cmpxchg(ptr, val, (unsigned long)&new->list);
|
| 825 |
|
| 826 | return ret == val;
|
| 827 | }
|
| 828 |
|
| 829 | |
| 830 | * rb_tail_page_update - move the tail page forward
|
| 831 | *
|
| 832 | * Returns 1 if moved tail page, 0 if someone else did.
|
| 833 |
|
| 834 | static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
|
| 835 | struct buffer_page *tail_page,
|
| 836 | struct buffer_page *next_page)
|
| 837 | {
|
| 838 | struct buffer_page *old_tail;
|
| 839 | unsigned long old_entries;
|
| 840 | unsigned long old_write;
|
| 841 | int ret = 0;
|
| 842 |
|
| 843 | |
| 844 | * The tail page now needs to be moved forward.
|
| 845 | *
|
| 846 | * We need to reset the tail page, but without messing
|
| 847 | * with possible erasing of data brought in by interrupts
|
| 848 | * that have moved the tail page and are currently on it.
|
| 849 | *
|
| 850 | * We add a counter to the write field to denote this.
|
| 851 |
|
| 852 | old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
|
| 853 | old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
|
| 854 |
|
| 855 | |
| 856 | * Just make sure we have seen our old_write and synchronize
|
| 857 | * with any interrupts that come in.
|
| 858 |
|
| 859 | barrier();
|
| 860 |
|
| 861 | |
| 862 | * If the tail page is still the same as what we think
|
| 863 | * it is, then it is up to us to update the tail
|
| 864 | * pointer.
|
| 865 |
|
| 866 | if (tail_page == cpu_buffer->tail_page) {
|
| 867 |
|
| 868 | unsigned long val = old_write & ~RB_WRITE_MASK;
|
| 869 | unsigned long eval = old_entries & ~RB_WRITE_MASK;
|
| 870 |
|
| 871 | |
| 872 | * This will only succeed if an interrupt did
|
| 873 | * not come in and change it. In which case, we
|
| 874 | * do not want to modify it.
|
| 875 | *
|
| 876 | * We add (void) to let the compiler know that we do not care
|
| 877 | * about the return value of these functions. We use the
|
| 878 | * cmpxchg to only update if an interrupt did not already
|
| 879 | * do it for us. If the cmpxchg fails, we don't care.
|
| 880 |
|
| 881 | (void)local_cmpxchg(&next_page->write, old_write, val);
|
| 882 | (void)local_cmpxchg(&next_page->entries, old_entries, eval);
|
| 883 |
|
| 884 | |
| 885 | * No need to worry about races with clearing out the commit.
|
| 886 | * it only can increment when a commit takes place. But that
|
| 887 | * only happens in the outer most nested commit.
|
| 888 |
|
| 889 | local_set(&next_page->page->commit, 0);
|
| 890 |
|
| 891 | old_tail = cmpxchg(&cpu_buffer->tail_page,
|
| 892 | tail_page, next_page);
|
| 893 |
|
| 894 | if (old_tail == tail_page)
|
| 895 | ret = 1;
|
| 896 | }
|
| 897 |
|
| 898 | return ret;
|
| 899 | }
|
| 900 |
|
| 901 | static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
|
| 902 | struct buffer_page *bpage)
|
| 903 | {
|
| 904 | unsigned long val = (unsigned long)bpage;
|
| 905 |
|
| 906 | if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
|
| 907 | return 1;
|
| 908 |
|
| 909 | return 0;
|
| 910 | }
|
| 911 |
|
| 912 | |
| 913 | * rb_check_list - make sure a pointer to a list has the last bits zero
|
| 914 |
|
| 915 | static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
|
| 916 | struct list_head *list)
|
| 917 | {
|
| 918 | if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
|
| 919 | return 1;
|
| 920 | if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
|
| 921 | return 1;
|
| 922 | return 0;
|
| 923 | }
|
| 924 |
|
| 925 | |
| 926 | * check_pages - integrity check of buffer pages
|
| 927 | * @cpu_buffer: CPU buffer with pages to test
|
| 928 | *
|
| 929 | * As a safety measure we check to make sure the data pages have not
|
| 930 | * been corrupted.
|
| 931 |
|
| 932 | static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
|
| 933 | {
|
| 934 | struct list_head *head = cpu_buffer->pages;
|
| 935 | struct buffer_page *bpage, *tmp;
|
| 936 |
|
| 937 | rb_head_page_deactivate(cpu_buffer);
|
| 938 |
|
| 939 | if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
|
| 940 | return -1;
|
| 941 | if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
|
| 942 | return -1;
|
| 943 |
|
| 944 | if (rb_check_list(cpu_buffer, head))
|
| 945 | return -1;
|
| 946 |
|
| 947 | list_for_each_entry_safe(bpage, tmp, head, list) {
|
| 948 | if (RB_WARN_ON(cpu_buffer,
|
| 949 | bpage->list.next->prev != &bpage->list))
|
| 950 | return -1;
|
| 951 | if (RB_WARN_ON(cpu_buffer,
|
| 952 | bpage->list.prev->next != &bpage->list))
|
| 953 | return -1;
|
| 954 | if (rb_check_list(cpu_buffer, &bpage->list))
|
| 955 | return -1;
|
| 956 | }
|
| 957 |
|
| 958 | rb_head_page_activate(cpu_buffer);
|
| 959 |
|
| 960 | return 0;
|
| 961 | }
|
| 962 |
|
| 963 | static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
|
| 964 | unsigned nr_pages)
|
| 965 | {
|
| 966 | struct buffer_page *bpage, *tmp;
|
| 967 | unsigned long addr;
|
| 968 | LIST_HEAD(pages);
|
| 969 | unsigned i;
|
| 970 |
|
| 971 | WARN_ON(!nr_pages);
|
| 972 |
|
| 973 | for (i = 0; i < nr_pages; i++) {
|
| 974 | bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
|
| 975 | GFP_KERNEL, cpu_to_node(cpu_buffer->cpu));
|
| 976 | if (!bpage)
|
| 977 | goto free_pages;
|
| 978 |
|
| 979 | rb_check_bpage(cpu_buffer, bpage);
|
| 980 |
|
| 981 | list_add(&bpage->list, &pages);
|
| 982 |
|
| 983 | addr = __get_free_page(GFP_KERNEL);
|
| 984 | if (!addr)
|
| 985 | goto free_pages;
|
| 986 | bpage->page = (void *)addr;
|
| 987 | rb_init_page(bpage->page);
|
| 988 | }
|
| 989 |
|
| 990 | |
| 991 | * The ring buffer page list is a circular list that does not
|
| 992 | * start and end with a list head. All page list items point to
|
| 993 | * other pages.
|
| 994 |
|
| 995 | cpu_buffer->pages = pages.next;
|
| 996 | list_del(&pages);
|
| 997 |
|
| 998 | rb_check_pages(cpu_buffer);
|
| 999 |
|
| 1000 | return 0;
|
| 1001 |
|
| 1002 | free_pages:
|
| 1003 | list_for_each_entry_safe(bpage, tmp, &pages, list) {
|
| 1004 | list_del_init(&bpage->list);
|
| 1005 | free_buffer_page(bpage);
|
| 1006 | }
|
| 1007 | return -ENOMEM;
|
| 1008 | }
|
| 1009 |
|
| 1010 | static struct ring_buffer_per_cpu *
|
| 1011 | rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
|
| 1012 | {
|
| 1013 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 1014 | struct buffer_page *bpage;
|
| 1015 | unsigned long addr;
|
| 1016 | int ret;
|
| 1017 |
|
| 1018 | cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
|
| 1019 | GFP_KERNEL, cpu_to_node(cpu));
|
| 1020 | if (!cpu_buffer)
|
| 1021 | return NULL;
|
| 1022 |
|
| 1023 | cpu_buffer->cpu = cpu;
|
| 1024 | cpu_buffer->buffer = buffer;
|
| 1025 | spin_lock_init(&cpu_buffer->reader_lock);
|
| 1026 | lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
|
| 1027 | cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
|
| 1028 |
|
| 1029 | bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
|
| 1030 | GFP_KERNEL, cpu_to_node(cpu));
|
| 1031 | if (!bpage)
|
| 1032 | goto fail_free_buffer;
|
| 1033 |
|
| 1034 | rb_check_bpage(cpu_buffer, bpage);
|
| 1035 |
|
| 1036 | cpu_buffer->reader_page = bpage;
|
| 1037 | addr = __get_free_page(GFP_KERNEL);
|
| 1038 | if (!addr)
|
| 1039 | goto fail_free_reader;
|
| 1040 | bpage->page = (void *)addr;
|
| 1041 | rb_init_page(bpage->page);
|
| 1042 |
|
| 1043 | INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
|
| 1044 |
|
| 1045 | ret = rb_allocate_pages(cpu_buffer, buffer->pages);
|
| 1046 | if (ret < 0)
|
| 1047 | goto fail_free_reader;
|
| 1048 |
|
| 1049 | cpu_buffer->head_page
|
| 1050 | = list_entry(cpu_buffer->pages, struct buffer_page, list);
|
| 1051 | cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
|
| 1052 |
|
| 1053 | rb_head_page_activate(cpu_buffer);
|
| 1054 |
|
| 1055 | return cpu_buffer;
|
| 1056 |
|
| 1057 | fail_free_reader:
|
| 1058 | free_buffer_page(cpu_buffer->reader_page);
|
| 1059 |
|
| 1060 | fail_free_buffer:
|
| 1061 | kfree(cpu_buffer);
|
| 1062 | return NULL;
|
| 1063 | }
|
| 1064 |
|
| 1065 | static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
|
| 1066 | {
|
| 1067 | struct list_head *head = cpu_buffer->pages;
|
| 1068 | struct buffer_page *bpage, *tmp;
|
| 1069 |
|
| 1070 | free_buffer_page(cpu_buffer->reader_page);
|
| 1071 |
|
| 1072 | rb_head_page_deactivate(cpu_buffer);
|
| 1073 |
|
| 1074 | if (head) {
|
| 1075 | list_for_each_entry_safe(bpage, tmp, head, list) {
|
| 1076 | list_del_init(&bpage->list);
|
| 1077 | free_buffer_page(bpage);
|
| 1078 | }
|
| 1079 | bpage = list_entry(head, struct buffer_page, list);
|
| 1080 | free_buffer_page(bpage);
|
| 1081 | }
|
| 1082 |
|
| 1083 | kfree(cpu_buffer);
|
| 1084 | }
|
| 1085 |
|
| 1086 | #ifdef CONFIG_HOTPLUG_CPU
|
| 1087 | static int rb_cpu_notify(struct notifier_block *self,
|
| 1088 | unsigned long action, void *hcpu);
|
| 1089 | #endif
|
| 1090 |
|
| 1091 | |
| 1092 | * ring_buffer_alloc - allocate a new ring_buffer
|
| 1093 | * @size: the size in bytes per cpu that is needed.
|
| 1094 | * @flags: attributes to set for the ring buffer.
|
| 1095 | *
|
| 1096 | * Currently the only flag that is available is the RB_FL_OVERWRITE
|
| 1097 | * flag. This flag means that the buffer will overwrite old data
|
| 1098 | * when the buffer wraps. If this flag is not set, the buffer will
|
| 1099 | * drop data when the tail hits the head.
|
| 1100 |
|
| 1101 | struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
|
| 1102 | struct lock_class_key *key)
|
| 1103 | {
|
| 1104 | struct ring_buffer *buffer;
|
| 1105 | int bsize;
|
| 1106 | int cpu;
|
| 1107 |
|
| 1108 |
|
| 1109 | buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
|
| 1110 | GFP_KERNEL);
|
| 1111 | if (!buffer)
|
| 1112 | return NULL;
|
| 1113 |
|
| 1114 | if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
|
| 1115 | goto fail_free_buffer;
|
| 1116 |
|
| 1117 | buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
|
| 1118 | buffer->flags = flags;
|
| 1119 | buffer->clock = trace_clock_local;
|
| 1120 | buffer->reader_lock_key = key;
|
| 1121 |
|
| 1122 |
|
| 1123 | if (buffer->pages < 2)
|
| 1124 | buffer->pages = 2;
|
| 1125 |
|
| 1126 | |
| 1127 | * In case of non-hotplug cpu, if the ring-buffer is allocated
|
| 1128 | * in early initcall, it will not be notified of secondary cpus.
|
| 1129 | * In that off case, we need to allocate for all possible cpus.
|
| 1130 |
|
| 1131 | #ifdef CONFIG_HOTPLUG_CPU
|
| 1132 | get_online_cpus();
|
| 1133 | cpumask_copy(buffer->cpumask, cpu_online_mask);
|
| 1134 | #else
|
| 1135 | cpumask_copy(buffer->cpumask, cpu_possible_mask);
|
| 1136 | #endif
|
| 1137 | buffer->cpus = nr_cpu_ids;
|
| 1138 |
|
| 1139 | bsize = sizeof(void *) * nr_cpu_ids;
|
| 1140 | buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
|
| 1141 | GFP_KERNEL);
|
| 1142 | if (!buffer->buffers)
|
| 1143 | goto fail_free_cpumask;
|
| 1144 |
|
| 1145 | for_each_buffer_cpu(buffer, cpu) {
|
| 1146 | buffer->buffers[cpu] =
|
| 1147 | rb_allocate_cpu_buffer(buffer, cpu);
|
| 1148 | if (!buffer->buffers[cpu])
|
| 1149 | goto fail_free_buffers;
|
| 1150 | }
|
| 1151 |
|
| 1152 | #ifdef CONFIG_HOTPLUG_CPU
|
| 1153 | buffer->cpu_notify.notifier_call = rb_cpu_notify;
|
| 1154 | buffer->cpu_notify.priority = 0;
|
| 1155 | register_cpu_notifier(&buffer->cpu_notify);
|
| 1156 | #endif
|
| 1157 |
|
| 1158 | put_online_cpus();
|
| 1159 | mutex_init(&buffer->mutex);
|
| 1160 |
|
| 1161 | return buffer;
|
| 1162 |
|
| 1163 | fail_free_buffers:
|
| 1164 | for_each_buffer_cpu(buffer, cpu) {
|
| 1165 | if (buffer->buffers[cpu])
|
| 1166 | rb_free_cpu_buffer(buffer->buffers[cpu]);
|
| 1167 | }
|
| 1168 | kfree(buffer->buffers);
|
| 1169 |
|
| 1170 | fail_free_cpumask:
|
| 1171 | free_cpumask_var(buffer->cpumask);
|
| 1172 | put_online_cpus();
|
| 1173 |
|
| 1174 | fail_free_buffer:
|
| 1175 | kfree(buffer);
|
| 1176 | return NULL;
|
| 1177 | }
|
| 1178 | EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
|
| 1179 |
|
| 1180 | |
| 1181 | * ring_buffer_free - free a ring buffer.
|
| 1182 | * @buffer: the buffer to free.
|
| 1183 |
|
| 1184 | void
|
| 1185 | ring_buffer_free(struct ring_buffer *buffer)
|
| 1186 | {
|
| 1187 | int cpu;
|
| 1188 |
|
| 1189 | get_online_cpus();
|
| 1190 |
|
| 1191 | #ifdef CONFIG_HOTPLUG_CPU
|
| 1192 | unregister_cpu_notifier(&buffer->cpu_notify);
|
| 1193 | #endif
|
| 1194 |
|
| 1195 | for_each_buffer_cpu(buffer, cpu)
|
| 1196 | rb_free_cpu_buffer(buffer->buffers[cpu]);
|
| 1197 |
|
| 1198 | put_online_cpus();
|
| 1199 |
|
| 1200 | kfree(buffer->buffers);
|
| 1201 | free_cpumask_var(buffer->cpumask);
|
| 1202 |
|
| 1203 | kfree(buffer);
|
| 1204 | }
|
| 1205 | EXPORT_SYMBOL_GPL(ring_buffer_free);
|
| 1206 |
|
| 1207 | void ring_buffer_set_clock(struct ring_buffer *buffer,
|
| 1208 | u64 (*clock)(void))
|
| 1209 | {
|
| 1210 | buffer->clock = clock;
|
| 1211 | }
|
| 1212 |
|
| 1213 | static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
|
| 1214 |
|
| 1215 | static void
|
| 1216 | rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
|
| 1217 | {
|
| 1218 | struct buffer_page *bpage;
|
| 1219 | struct list_head *p;
|
| 1220 | unsigned i;
|
| 1221 |
|
| 1222 | spin_lock_irq(&cpu_buffer->reader_lock);
|
| 1223 | rb_head_page_deactivate(cpu_buffer);
|
| 1224 |
|
| 1225 | for (i = 0; i < nr_pages; i++) {
|
| 1226 | if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
|
| 1227 | goto out;
|
| 1228 | p = cpu_buffer->pages->next;
|
| 1229 | bpage = list_entry(p, struct buffer_page, list);
|
| 1230 | list_del_init(&bpage->list);
|
| 1231 | free_buffer_page(bpage);
|
| 1232 | }
|
| 1233 | if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
|
| 1234 | goto out;
|
| 1235 |
|
| 1236 | rb_reset_cpu(cpu_buffer);
|
| 1237 | rb_check_pages(cpu_buffer);
|
| 1238 |
|
| 1239 | out:
|
| 1240 | spin_unlock_irq(&cpu_buffer->reader_lock);
|
| 1241 | }
|
| 1242 |
|
| 1243 | static void
|
| 1244 | rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1245 | struct list_head *pages, unsigned nr_pages)
|
| 1246 | {
|
| 1247 | struct buffer_page *bpage;
|
| 1248 | struct list_head *p;
|
| 1249 | unsigned i;
|
| 1250 |
|
| 1251 | spin_lock_irq(&cpu_buffer->reader_lock);
|
| 1252 | rb_head_page_deactivate(cpu_buffer);
|
| 1253 |
|
| 1254 | for (i = 0; i < nr_pages; i++) {
|
| 1255 | if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
|
| 1256 | goto out;
|
| 1257 | p = pages->next;
|
| 1258 | bpage = list_entry(p, struct buffer_page, list);
|
| 1259 | list_del_init(&bpage->list);
|
| 1260 | list_add_tail(&bpage->list, cpu_buffer->pages);
|
| 1261 | }
|
| 1262 | rb_reset_cpu(cpu_buffer);
|
| 1263 | rb_check_pages(cpu_buffer);
|
| 1264 |
|
| 1265 | out:
|
| 1266 | spin_unlock_irq(&cpu_buffer->reader_lock);
|
| 1267 | }
|
| 1268 |
|
| 1269 | |
| 1270 | * ring_buffer_resize - resize the ring buffer
|
| 1271 | * @buffer: the buffer to resize.
|
| 1272 | * @size: the new size.
|
| 1273 | *
|
| 1274 | * Minimum size is 2 * BUF_PAGE_SIZE.
|
| 1275 | *
|
| 1276 | * Returns -1 on failure.
|
| 1277 |
|
| 1278 | int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
|
| 1279 | {
|
| 1280 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 1281 | unsigned nr_pages, rm_pages, new_pages;
|
| 1282 | struct buffer_page *bpage, *tmp;
|
| 1283 | unsigned long buffer_size;
|
| 1284 | unsigned long addr;
|
| 1285 | LIST_HEAD(pages);
|
| 1286 | int i, cpu;
|
| 1287 |
|
| 1288 | |
| 1289 | * Always succeed at resizing a non-existent buffer:
|
| 1290 |
|
| 1291 | if (!buffer)
|
| 1292 | return size;
|
| 1293 |
|
| 1294 | size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
|
| 1295 | size *= BUF_PAGE_SIZE;
|
| 1296 | buffer_size = buffer->pages * BUF_PAGE_SIZE;
|
| 1297 |
|
| 1298 |
|
| 1299 | if (size < BUF_PAGE_SIZE * 2)
|
| 1300 | size = BUF_PAGE_SIZE * 2;
|
| 1301 |
|
| 1302 | if (size == buffer_size)
|
| 1303 | return size;
|
| 1304 |
|
| 1305 | atomic_inc(&buffer->record_disabled);
|
| 1306 |
|
| 1307 |
|
| 1308 | synchronize_sched();
|
| 1309 |
|
| 1310 | mutex_lock(&buffer->mutex);
|
| 1311 | get_online_cpus();
|
| 1312 |
|
| 1313 | nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
|
| 1314 |
|
| 1315 | if (size < buffer_size) {
|
| 1316 |
|
| 1317 |
|
| 1318 | if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
|
| 1319 | goto out_fail;
|
| 1320 |
|
| 1321 | rm_pages = buffer->pages - nr_pages;
|
| 1322 |
|
| 1323 | for_each_buffer_cpu(buffer, cpu) {
|
| 1324 | cpu_buffer = buffer->buffers[cpu];
|
| 1325 | rb_remove_pages(cpu_buffer, rm_pages);
|
| 1326 | }
|
| 1327 | goto out;
|
| 1328 | }
|
| 1329 |
|
| 1330 | |
| 1331 | * This is a bit more difficult. We only want to add pages
|
| 1332 | * when we can allocate enough for all CPUs. We do this
|
| 1333 | * by allocating all the pages and storing them on a local
|
| 1334 | * link list. If we succeed in our allocation, then we
|
| 1335 | * add these pages to the cpu_buffers. Otherwise we just free
|
| 1336 | * them all and return -ENOMEM;
|
| 1337 |
|
| 1338 | if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
|
| 1339 | goto out_fail;
|
| 1340 |
|
| 1341 | new_pages = nr_pages - buffer->pages;
|
| 1342 |
|
| 1343 | for_each_buffer_cpu(buffer, cpu) {
|
| 1344 | for (i = 0; i < new_pages; i++) {
|
| 1345 | bpage = kzalloc_node(ALIGN(sizeof(*bpage),
|
| 1346 | cache_line_size()),
|
| 1347 | GFP_KERNEL, cpu_to_node(cpu));
|
| 1348 | if (!bpage)
|
| 1349 | goto free_pages;
|
| 1350 | list_add(&bpage->list, &pages);
|
| 1351 | addr = __get_free_page(GFP_KERNEL);
|
| 1352 | if (!addr)
|
| 1353 | goto free_pages;
|
| 1354 | bpage->page = (void *)addr;
|
| 1355 | rb_init_page(bpage->page);
|
| 1356 | }
|
| 1357 | }
|
| 1358 |
|
| 1359 | for_each_buffer_cpu(buffer, cpu) {
|
| 1360 | cpu_buffer = buffer->buffers[cpu];
|
| 1361 | rb_insert_pages(cpu_buffer, &pages, new_pages);
|
| 1362 | }
|
| 1363 |
|
| 1364 | if (RB_WARN_ON(buffer, !list_empty(&pages)))
|
| 1365 | goto out_fail;
|
| 1366 |
|
| 1367 | out:
|
| 1368 | buffer->pages = nr_pages;
|
| 1369 | put_online_cpus();
|
| 1370 | mutex_unlock(&buffer->mutex);
|
| 1371 |
|
| 1372 | atomic_dec(&buffer->record_disabled);
|
| 1373 |
|
| 1374 | return size;
|
| 1375 |
|
| 1376 | free_pages:
|
| 1377 | list_for_each_entry_safe(bpage, tmp, &pages, list) {
|
| 1378 | list_del_init(&bpage->list);
|
| 1379 | free_buffer_page(bpage);
|
| 1380 | }
|
| 1381 | put_online_cpus();
|
| 1382 | mutex_unlock(&buffer->mutex);
|
| 1383 | atomic_dec(&buffer->record_disabled);
|
| 1384 | return -ENOMEM;
|
| 1385 |
|
| 1386 | |
| 1387 | * Something went totally wrong, and we are too paranoid
|
| 1388 | * to even clean up the mess.
|
| 1389 |
|
| 1390 | out_fail:
|
| 1391 | put_online_cpus();
|
| 1392 | mutex_unlock(&buffer->mutex);
|
| 1393 | atomic_dec(&buffer->record_disabled);
|
| 1394 | return -1;
|
| 1395 | }
|
| 1396 | EXPORT_SYMBOL_GPL(ring_buffer_resize);
|
| 1397 |
|
| 1398 | static inline void *
|
| 1399 | __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
|
| 1400 | {
|
| 1401 | return bpage->data + index;
|
| 1402 | }
|
| 1403 |
|
| 1404 | static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
|
| 1405 | {
|
| 1406 | return bpage->page->data + index;
|
| 1407 | }
|
| 1408 |
|
| 1409 | static inline struct ring_buffer_event *
|
| 1410 | rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
|
| 1411 | {
|
| 1412 | return __rb_page_index(cpu_buffer->reader_page,
|
| 1413 | cpu_buffer->reader_page->read);
|
| 1414 | }
|
| 1415 |
|
| 1416 | static inline struct ring_buffer_event *
|
| 1417 | rb_iter_head_event(struct ring_buffer_iter *iter)
|
| 1418 | {
|
| 1419 | return __rb_page_index(iter->head_page, iter->head);
|
| 1420 | }
|
| 1421 |
|
| 1422 | static inline unsigned long rb_page_write(struct buffer_page *bpage)
|
| 1423 | {
|
| 1424 | return local_read(&bpage->write) & RB_WRITE_MASK;
|
| 1425 | }
|
| 1426 |
|
| 1427 | static inline unsigned rb_page_commit(struct buffer_page *bpage)
|
| 1428 | {
|
| 1429 | return local_read(&bpage->page->commit);
|
| 1430 | }
|
| 1431 |
|
| 1432 | static inline unsigned long rb_page_entries(struct buffer_page *bpage)
|
| 1433 | {
|
| 1434 | return local_read(&bpage->entries) & RB_WRITE_MASK;
|
| 1435 | }
|
| 1436 |
|
| 1437 |
|
| 1438 | static inline unsigned rb_page_size(struct buffer_page *bpage)
|
| 1439 | {
|
| 1440 | return rb_page_commit(bpage);
|
| 1441 | }
|
| 1442 |
|
| 1443 | static inline unsigned
|
| 1444 | rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
|
| 1445 | {
|
| 1446 | return rb_page_commit(cpu_buffer->commit_page);
|
| 1447 | }
|
| 1448 |
|
| 1449 | static inline unsigned
|
| 1450 | rb_event_index(struct ring_buffer_event *event)
|
| 1451 | {
|
| 1452 | unsigned long addr = (unsigned long)event;
|
| 1453 |
|
| 1454 | return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
|
| 1455 | }
|
| 1456 |
|
| 1457 | static inline int
|
| 1458 | rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1459 | struct ring_buffer_event *event)
|
| 1460 | {
|
| 1461 | unsigned long addr = (unsigned long)event;
|
| 1462 | unsigned long index;
|
| 1463 |
|
| 1464 | index = rb_event_index(event);
|
| 1465 | addr &= PAGE_MASK;
|
| 1466 |
|
| 1467 | return cpu_buffer->commit_page->page == (void *)addr &&
|
| 1468 | rb_commit_index(cpu_buffer) == index;
|
| 1469 | }
|
| 1470 |
|
| 1471 | static void
|
| 1472 | rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
|
| 1473 | {
|
| 1474 | unsigned long max_count;
|
| 1475 |
|
| 1476 | |
| 1477 | * We only race with interrupts and NMIs on this CPU.
|
| 1478 | * If we own the commit event, then we can commit
|
| 1479 | * all others that interrupted us, since the interruptions
|
| 1480 | * are in stack format (they finish before they come
|
| 1481 | * back to us). This allows us to do a simple loop to
|
| 1482 | * assign the commit to the tail.
|
| 1483 |
|
| 1484 | again:
|
| 1485 | max_count = cpu_buffer->buffer->pages * 100;
|
| 1486 |
|
| 1487 | while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
|
| 1488 | if (RB_WARN_ON(cpu_buffer, !(--max_count)))
|
| 1489 | return;
|
| 1490 | if (RB_WARN_ON(cpu_buffer,
|
| 1491 | rb_is_reader_page(cpu_buffer->tail_page)))
|
| 1492 | return;
|
| 1493 | local_set(&cpu_buffer->commit_page->page->commit,
|
| 1494 | rb_page_write(cpu_buffer->commit_page));
|
| 1495 | rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
|
| 1496 | cpu_buffer->write_stamp =
|
| 1497 | cpu_buffer->commit_page->page->time_stamp;
|
| 1498 |
|
| 1499 | barrier();
|
| 1500 | }
|
| 1501 | while (rb_commit_index(cpu_buffer) !=
|
| 1502 | rb_page_write(cpu_buffer->commit_page)) {
|
| 1503 |
|
| 1504 | local_set(&cpu_buffer->commit_page->page->commit,
|
| 1505 | rb_page_write(cpu_buffer->commit_page));
|
| 1506 | RB_WARN_ON(cpu_buffer,
|
| 1507 | local_read(&cpu_buffer->commit_page->page->commit) &
|
| 1508 | ~RB_WRITE_MASK);
|
| 1509 | barrier();
|
| 1510 | }
|
| 1511 |
|
| 1512 |
|
| 1513 | barrier();
|
| 1514 |
|
| 1515 | |
| 1516 | * If an interrupt came in just after the first while loop
|
| 1517 | * and pushed the tail page forward, we will be left with
|
| 1518 | * a dangling commit that will never go forward.
|
| 1519 |
|
| 1520 | if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
|
| 1521 | goto again;
|
| 1522 | }
|
| 1523 |
|
| 1524 | static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
|
| 1525 | {
|
| 1526 | cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
|
| 1527 | cpu_buffer->reader_page->read = 0;
|
| 1528 | }
|
| 1529 |
|
| 1530 | static void rb_inc_iter(struct ring_buffer_iter *iter)
|
| 1531 | {
|
| 1532 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
|
| 1533 |
|
| 1534 | |
| 1535 | * The iterator could be on the reader page (it starts there).
|
| 1536 | * But the head could have moved, since the reader was
|
| 1537 | * found. Check for this case and assign the iterator
|
| 1538 | * to the head page instead of next.
|
| 1539 |
|
| 1540 | if (iter->head_page == cpu_buffer->reader_page)
|
| 1541 | iter->head_page = rb_set_head_page(cpu_buffer);
|
| 1542 | else
|
| 1543 | rb_inc_page(cpu_buffer, &iter->head_page);
|
| 1544 |
|
| 1545 | iter->read_stamp = iter->head_page->page->time_stamp;
|
| 1546 | iter->head = 0;
|
| 1547 | }
|
| 1548 |
|
| 1549 | |
| 1550 | * ring_buffer_update_event - update event type and data
|
| 1551 | * @event: the even to update
|
| 1552 | * @type: the type of event
|
| 1553 | * @length: the size of the event field in the ring buffer
|
| 1554 | *
|
| 1555 | * Update the type and data fields of the event. The length
|
| 1556 | * is the actual size that is written to the ring buffer,
|
| 1557 | * and with this, we can determine what to place into the
|
| 1558 | * data field.
|
| 1559 |
|
| 1560 | static void
|
| 1561 | rb_update_event(struct ring_buffer_event *event,
|
| 1562 | unsigned type, unsigned length)
|
| 1563 | {
|
| 1564 | event->type_len = type;
|
| 1565 |
|
| 1566 | switch (type) {
|
| 1567 |
|
| 1568 | case RINGBUF_TYPE_PADDING:
|
| 1569 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 1570 | case RINGBUF_TYPE_TIME_STAMP:
|
| 1571 | break;
|
| 1572 |
|
| 1573 | case 0:
|
| 1574 | length -= RB_EVNT_HDR_SIZE;
|
| 1575 | if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
|
| 1576 | event->array[0] = length;
|
| 1577 | else
|
| 1578 | event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
|
| 1579 | break;
|
| 1580 | default:
|
| 1581 | BUG();
|
| 1582 | }
|
| 1583 | }
|
| 1584 |
|
| 1585 | |
| 1586 | * rb_handle_head_page - writer hit the head page
|
| 1587 | *
|
| 1588 | * Returns: +1 to retry page
|
| 1589 | * 0 to continue
|
| 1590 | * -1 on error
|
| 1591 |
|
| 1592 | static int
|
| 1593 | rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1594 | struct buffer_page *tail_page,
|
| 1595 | struct buffer_page *next_page)
|
| 1596 | {
|
| 1597 | struct buffer_page *new_head;
|
| 1598 | int entries;
|
| 1599 | int type;
|
| 1600 | int ret;
|
| 1601 |
|
| 1602 | entries = rb_page_entries(next_page);
|
| 1603 |
|
| 1604 | |
| 1605 | * The hard part is here. We need to move the head
|
| 1606 | * forward, and protect against both readers on
|
| 1607 | * other CPUs and writers coming in via interrupts.
|
| 1608 |
|
| 1609 | type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
|
| 1610 | RB_PAGE_HEAD);
|
| 1611 |
|
| 1612 | |
| 1613 | * type can be one of four:
|
| 1614 | * NORMAL - an interrupt already moved it for us
|
| 1615 | * HEAD - we are the first to get here.
|
| 1616 | * UPDATE - we are the interrupt interrupting
|
| 1617 | * a current move.
|
| 1618 | * MOVED - a reader on another CPU moved the next
|
| 1619 | * pointer to its reader page. Give up
|
| 1620 | * and try again.
|
| 1621 |
|
| 1622 |
|
| 1623 | switch (type) {
|
| 1624 | case RB_PAGE_HEAD:
|
| 1625 | |
| 1626 | * We changed the head to UPDATE, thus
|
| 1627 | * it is our responsibility to update
|
| 1628 | * the counters.
|
| 1629 |
|
| 1630 | local_add(entries, &cpu_buffer->overrun);
|
| 1631 |
|
| 1632 | |
| 1633 | * The entries will be zeroed out when we move the
|
| 1634 | * tail page.
|
| 1635 |
|
| 1636 |
|
| 1637 |
|
| 1638 | break;
|
| 1639 |
|
| 1640 | case RB_PAGE_UPDATE:
|
| 1641 | |
| 1642 | * This is an interrupt that interrupt the
|
| 1643 | * previous update. Still more to do.
|
| 1644 |
|
| 1645 | break;
|
| 1646 | case RB_PAGE_NORMAL:
|
| 1647 | |
| 1648 | * An interrupt came in before the update
|
| 1649 | * and processed this for us.
|
| 1650 | * Nothing left to do.
|
| 1651 |
|
| 1652 | return 1;
|
| 1653 | case RB_PAGE_MOVED:
|
| 1654 | |
| 1655 | * The reader is on another CPU and just did
|
| 1656 | * a swap with our next_page.
|
| 1657 | * Try again.
|
| 1658 |
|
| 1659 | return 1;
|
| 1660 | default:
|
| 1661 | RB_WARN_ON(cpu_buffer, 1);
|
| 1662 | return -1;
|
| 1663 | }
|
| 1664 |
|
| 1665 | |
| 1666 | * Now that we are here, the old head pointer is
|
| 1667 | * set to UPDATE. This will keep the reader from
|
| 1668 | * swapping the head page with the reader page.
|
| 1669 | * The reader (on another CPU) will spin till
|
| 1670 | * we are finished.
|
| 1671 | *
|
| 1672 | * We just need to protect against interrupts
|
| 1673 | * doing the job. We will set the next pointer
|
| 1674 | * to HEAD. After that, we set the old pointer
|
| 1675 | * to NORMAL, but only if it was HEAD before.
|
| 1676 | * otherwise we are an interrupt, and only
|
| 1677 | * want the outer most commit to reset it.
|
| 1678 |
|
| 1679 | new_head = next_page;
|
| 1680 | rb_inc_page(cpu_buffer, &new_head);
|
| 1681 |
|
| 1682 | ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
|
| 1683 | RB_PAGE_NORMAL);
|
| 1684 |
|
| 1685 | |
| 1686 | * Valid returns are:
|
| 1687 | * HEAD - an interrupt came in and already set it.
|
| 1688 | * NORMAL - One of two things:
|
| 1689 | * 1) We really set it.
|
| 1690 | * 2) A bunch of interrupts came in and moved
|
| 1691 | * the page forward again.
|
| 1692 |
|
| 1693 | switch (ret) {
|
| 1694 | case RB_PAGE_HEAD:
|
| 1695 | case RB_PAGE_NORMAL:
|
| 1696 |
|
| 1697 | break;
|
| 1698 | default:
|
| 1699 | RB_WARN_ON(cpu_buffer, 1);
|
| 1700 | return -1;
|
| 1701 | }
|
| 1702 |
|
| 1703 | |
| 1704 | * It is possible that an interrupt came in,
|
| 1705 | * set the head up, then more interrupts came in
|
| 1706 | * and moved it again. When we get back here,
|
| 1707 | * the page would have been set to NORMAL but we
|
| 1708 | * just set it back to HEAD.
|
| 1709 | *
|
| 1710 | * How do you detect this? Well, if that happened
|
| 1711 | * the tail page would have moved.
|
| 1712 |
|
| 1713 | if (ret == RB_PAGE_NORMAL) {
|
| 1714 | |
| 1715 | * If the tail had moved passed next, then we need
|
| 1716 | * to reset the pointer.
|
| 1717 |
|
| 1718 | if (cpu_buffer->tail_page != tail_page &&
|
| 1719 | cpu_buffer->tail_page != next_page)
|
| 1720 | rb_head_page_set_normal(cpu_buffer, new_head,
|
| 1721 | next_page,
|
| 1722 | RB_PAGE_HEAD);
|
| 1723 | }
|
| 1724 |
|
| 1725 | |
| 1726 | * If this was the outer most commit (the one that
|
| 1727 | * changed the original pointer from HEAD to UPDATE),
|
| 1728 | * then it is up to us to reset it to NORMAL.
|
| 1729 |
|
| 1730 | if (type == RB_PAGE_HEAD) {
|
| 1731 | ret = rb_head_page_set_normal(cpu_buffer, next_page,
|
| 1732 | tail_page,
|
| 1733 | RB_PAGE_UPDATE);
|
| 1734 | if (RB_WARN_ON(cpu_buffer,
|
| 1735 | ret != RB_PAGE_UPDATE))
|
| 1736 | return -1;
|
| 1737 | }
|
| 1738 |
|
| 1739 | return 0;
|
| 1740 | }
|
| 1741 |
|
| 1742 | static unsigned rb_calculate_event_length(unsigned length)
|
| 1743 | {
|
| 1744 | struct ring_buffer_event event;
|
| 1745 |
|
| 1746 |
|
| 1747 | if (!length)
|
| 1748 | length = 1;
|
| 1749 |
|
| 1750 | if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
|
| 1751 | length += sizeof(event.array[0]);
|
| 1752 |
|
| 1753 | length += RB_EVNT_HDR_SIZE;
|
| 1754 | length = ALIGN(length, RB_ARCH_ALIGNMENT);
|
| 1755 |
|
| 1756 | return length;
|
| 1757 | }
|
| 1758 |
|
| 1759 | static inline void
|
| 1760 | rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1761 | struct buffer_page *tail_page,
|
| 1762 | unsigned long tail, unsigned long length)
|
| 1763 | {
|
| 1764 | struct ring_buffer_event *event;
|
| 1765 |
|
| 1766 | |
| 1767 | * Only the event that crossed the page boundary
|
| 1768 | * must fill the old tail_page with padding.
|
| 1769 |
|
| 1770 | if (tail >= BUF_PAGE_SIZE) {
|
| 1771 | local_sub(length, &tail_page->write);
|
| 1772 | return;
|
| 1773 | }
|
| 1774 |
|
| 1775 | event = __rb_page_index(tail_page, tail);
|
| 1776 | kmemcheck_annotate_bitfield(event, bitfield);
|
| 1777 |
|
| 1778 | |
| 1779 | * Save the original length to the meta data.
|
| 1780 | * This will be used by the reader to add lost event
|
| 1781 | * counter.
|
| 1782 |
|
| 1783 | tail_page->real_end = tail;
|
| 1784 |
|
| 1785 | |
| 1786 | * If this event is bigger than the minimum size, then
|
| 1787 | * we need to be careful that we don't subtract the
|
| 1788 | * write counter enough to allow another writer to slip
|
| 1789 | * in on this page.
|
| 1790 | * We put in a discarded commit instead, to make sure
|
| 1791 | * that this space is not used again.
|
| 1792 | *
|
| 1793 | * If we are less than the minimum size, we don't need to
|
| 1794 | * worry about it.
|
| 1795 |
|
| 1796 | if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
|
| 1797 |
|
| 1798 |
|
| 1799 |
|
| 1800 | rb_event_set_padding(event);
|
| 1801 |
|
| 1802 |
|
| 1803 | local_sub(length, &tail_page->write);
|
| 1804 | return;
|
| 1805 | }
|
| 1806 |
|
| 1807 |
|
| 1808 | event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
|
| 1809 | event->type_len = RINGBUF_TYPE_PADDING;
|
| 1810 |
|
| 1811 | event->time_delta = 1;
|
| 1812 |
|
| 1813 |
|
| 1814 | length = (tail + length) - BUF_PAGE_SIZE;
|
| 1815 | local_sub(length, &tail_page->write);
|
| 1816 | }
|
| 1817 |
|
| 1818 | static struct ring_buffer_event *
|
| 1819 | rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1820 | unsigned long length, unsigned long tail,
|
| 1821 | struct buffer_page *tail_page, u64 *ts)
|
| 1822 | {
|
| 1823 | struct buffer_page *commit_page = cpu_buffer->commit_page;
|
| 1824 | struct ring_buffer *buffer = cpu_buffer->buffer;
|
| 1825 | struct buffer_page *next_page;
|
| 1826 | int ret;
|
| 1827 |
|
| 1828 | next_page = tail_page;
|
| 1829 |
|
| 1830 | rb_inc_page(cpu_buffer, &next_page);
|
| 1831 |
|
| 1832 | |
| 1833 | * If for some reason, we had an interrupt storm that made
|
| 1834 | * it all the way around the buffer, bail, and warn
|
| 1835 | * about it.
|
| 1836 |
|
| 1837 | if (unlikely(next_page == commit_page)) {
|
| 1838 | local_inc(&cpu_buffer->commit_overrun);
|
| 1839 | goto out_reset;
|
| 1840 | }
|
| 1841 |
|
| 1842 | |
| 1843 | * This is where the fun begins!
|
| 1844 | *
|
| 1845 | * We are fighting against races between a reader that
|
| 1846 | * could be on another CPU trying to swap its reader
|
| 1847 | * page with the buffer head.
|
| 1848 | *
|
| 1849 | * We are also fighting against interrupts coming in and
|
| 1850 | * moving the head or tail on us as well.
|
| 1851 | *
|
| 1852 | * If the next page is the head page then we have filled
|
| 1853 | * the buffer, unless the commit page is still on the
|
| 1854 | * reader page.
|
| 1855 |
|
| 1856 | if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
|
| 1857 |
|
| 1858 | |
| 1859 | * If the commit is not on the reader page, then
|
| 1860 | * move the header page.
|
| 1861 |
|
| 1862 | if (!rb_is_reader_page(cpu_buffer->commit_page)) {
|
| 1863 | |
| 1864 | * If we are not in overwrite mode,
|
| 1865 | * this is easy, just stop here.
|
| 1866 |
|
| 1867 | if (!(buffer->flags & RB_FL_OVERWRITE))
|
| 1868 | goto out_reset;
|
| 1869 |
|
| 1870 | ret = rb_handle_head_page(cpu_buffer,
|
| 1871 | tail_page,
|
| 1872 | next_page);
|
| 1873 | if (ret < 0)
|
| 1874 | goto out_reset;
|
| 1875 | if (ret)
|
| 1876 | goto out_again;
|
| 1877 | } else {
|
| 1878 | |
| 1879 | * We need to be careful here too. The
|
| 1880 | * commit page could still be on the reader
|
| 1881 | * page. We could have a small buffer, and
|
| 1882 | * have filled up the buffer with events
|
| 1883 | * from interrupts and such, and wrapped.
|
| 1884 | *
|
| 1885 | * Note, if the tail page is also the on the
|
| 1886 | * reader_page, we let it move out.
|
| 1887 |
|
| 1888 | if (unlikely((cpu_buffer->commit_page !=
|
| 1889 | cpu_buffer->tail_page) &&
|
| 1890 | (cpu_buffer->commit_page ==
|
| 1891 | cpu_buffer->reader_page))) {
|
| 1892 | local_inc(&cpu_buffer->commit_overrun);
|
| 1893 | goto out_reset;
|
| 1894 | }
|
| 1895 | }
|
| 1896 | }
|
| 1897 |
|
| 1898 | ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
|
| 1899 | if (ret) {
|
| 1900 | |
| 1901 | * Nested commits always have zero deltas, so
|
| 1902 | * just reread the time stamp
|
| 1903 |
|
| 1904 | *ts = rb_time_stamp(buffer);
|
| 1905 | next_page->page->time_stamp = *ts;
|
| 1906 | }
|
| 1907 |
|
| 1908 | out_again:
|
| 1909 |
|
| 1910 | rb_reset_tail(cpu_buffer, tail_page, tail, length);
|
| 1911 |
|
| 1912 |
|
| 1913 | return ERR_PTR(-EAGAIN);
|
| 1914 |
|
| 1915 | out_reset:
|
| 1916 |
|
| 1917 | rb_reset_tail(cpu_buffer, tail_page, tail, length);
|
| 1918 |
|
| 1919 | return NULL;
|
| 1920 | }
|
| 1921 |
|
| 1922 | static struct ring_buffer_event *
|
| 1923 | __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1924 | unsigned type, unsigned long length, u64 *ts)
|
| 1925 | {
|
| 1926 | struct buffer_page *tail_page;
|
| 1927 | struct ring_buffer_event *event;
|
| 1928 | unsigned long tail, write;
|
| 1929 |
|
| 1930 | tail_page = cpu_buffer->tail_page;
|
| 1931 | write = local_add_return(length, &tail_page->write);
|
| 1932 |
|
| 1933 |
|
| 1934 | write &= RB_WRITE_MASK;
|
| 1935 | tail = write - length;
|
| 1936 |
|
| 1937 |
|
| 1938 | if (write > BUF_PAGE_SIZE)
|
| 1939 | return rb_move_tail(cpu_buffer, length, tail,
|
| 1940 | tail_page, ts);
|
| 1941 |
|
| 1942 |
|
| 1943 |
|
| 1944 | event = __rb_page_index(tail_page, tail);
|
| 1945 | kmemcheck_annotate_bitfield(event, bitfield);
|
| 1946 | rb_update_event(event, type, length);
|
| 1947 |
|
| 1948 |
|
| 1949 | if (likely(!type))
|
| 1950 | local_inc(&tail_page->entries);
|
| 1951 |
|
| 1952 | |
| 1953 | * If this is the first commit on the page, then update
|
| 1954 | * its timestamp.
|
| 1955 |
|
| 1956 | if (!tail)
|
| 1957 | tail_page->page->time_stamp = *ts;
|
| 1958 |
|
| 1959 | return event;
|
| 1960 | }
|
| 1961 |
|
| 1962 | static inline int
|
| 1963 | rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1964 | struct ring_buffer_event *event)
|
| 1965 | {
|
| 1966 | unsigned long new_index, old_index;
|
| 1967 | struct buffer_page *bpage;
|
| 1968 | unsigned long index;
|
| 1969 | unsigned long addr;
|
| 1970 |
|
| 1971 | new_index = rb_event_index(event);
|
| 1972 | old_index = new_index + rb_event_length(event);
|
| 1973 | addr = (unsigned long)event;
|
| 1974 | addr &= PAGE_MASK;
|
| 1975 |
|
| 1976 | bpage = cpu_buffer->tail_page;
|
| 1977 |
|
| 1978 | if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
|
| 1979 | unsigned long write_mask =
|
| 1980 | local_read(&bpage->write) & ~RB_WRITE_MASK;
|
| 1981 | |
| 1982 | * This is on the tail page. It is possible that
|
| 1983 | * a write could come in and move the tail page
|
| 1984 | * and write to the next page. That is fine
|
| 1985 | * because we just shorten what is on this page.
|
| 1986 |
|
| 1987 | old_index += write_mask;
|
| 1988 | new_index += write_mask;
|
| 1989 | index = local_cmpxchg(&bpage->write, old_index, new_index);
|
| 1990 | if (index == old_index)
|
| 1991 | return 1;
|
| 1992 | }
|
| 1993 |
|
| 1994 |
|
| 1995 | return 0;
|
| 1996 | }
|
| 1997 |
|
| 1998 | static int
|
| 1999 | rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2000 | u64 *ts, u64 *delta)
|
| 2001 | {
|
| 2002 | struct ring_buffer_event *event;
|
| 2003 | int ret;
|
| 2004 |
|
| 2005 | WARN_ONCE(*delta > (1ULL << 59),
|
| 2006 | KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n",
|
| 2007 | (unsigned long long)*delta,
|
| 2008 | (unsigned long long)*ts,
|
| 2009 | (unsigned long long)cpu_buffer->write_stamp);
|
| 2010 |
|
| 2011 | |
| 2012 | * The delta is too big, we to add a
|
| 2013 | * new timestamp.
|
| 2014 |
|
| 2015 | event = __rb_reserve_next(cpu_buffer,
|
| 2016 | RINGBUF_TYPE_TIME_EXTEND,
|
| 2017 | RB_LEN_TIME_EXTEND,
|
| 2018 | ts);
|
| 2019 | if (!event)
|
| 2020 | return -EBUSY;
|
| 2021 |
|
| 2022 | if (PTR_ERR(event) == -EAGAIN)
|
| 2023 | return -EAGAIN;
|
| 2024 |
|
| 2025 |
|
| 2026 | if (rb_event_is_commit(cpu_buffer, event)) {
|
| 2027 | |
| 2028 | * If this is the first on the page, then it was
|
| 2029 | * updated with the page itself. Try to discard it
|
| 2030 | * and if we can't just make it zero.
|
| 2031 |
|
| 2032 | if (rb_event_index(event)) {
|
| 2033 | event->time_delta = *delta & TS_MASK;
|
| 2034 | event->array[0] = *delta >> TS_SHIFT;
|
| 2035 | } else {
|
| 2036 |
|
| 2037 | if (!rb_try_to_discard(cpu_buffer, event)) {
|
| 2038 |
|
| 2039 | event->time_delta = 0;
|
| 2040 | event->array[0] = 0;
|
| 2041 | }
|
| 2042 | }
|
| 2043 | cpu_buffer->write_stamp = *ts;
|
| 2044 |
|
| 2045 | ret = 1;
|
| 2046 | } else {
|
| 2047 |
|
| 2048 | if (!rb_try_to_discard(cpu_buffer, event)) {
|
| 2049 |
|
| 2050 | event->time_delta = 0;
|
| 2051 | event->array[0] = 0;
|
| 2052 | }
|
| 2053 | ret = 0;
|
| 2054 | }
|
| 2055 |
|
| 2056 | *delta = 0;
|
| 2057 |
|
| 2058 | return ret;
|
| 2059 | }
|
| 2060 |
|
| 2061 | static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
|
| 2062 | {
|
| 2063 | local_inc(&cpu_buffer->committing);
|
| 2064 | local_inc(&cpu_buffer->commits);
|
| 2065 | }
|
| 2066 |
|
| 2067 | static void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
|
| 2068 | {
|
| 2069 | unsigned long commits;
|
| 2070 |
|
| 2071 | if (RB_WARN_ON(cpu_buffer,
|
| 2072 | !local_read(&cpu_buffer->committing)))
|
| 2073 | return;
|
| 2074 |
|
| 2075 | again:
|
| 2076 | commits = local_read(&cpu_buffer->commits);
|
| 2077 |
|
| 2078 | barrier();
|
| 2079 | if (local_read(&cpu_buffer->committing) == 1)
|
| 2080 | rb_set_commit_to_write(cpu_buffer);
|
| 2081 |
|
| 2082 | local_dec(&cpu_buffer->committing);
|
| 2083 |
|
| 2084 |
|
| 2085 | barrier();
|
| 2086 |
|
| 2087 | |
| 2088 | * Need to account for interrupts coming in between the
|
| 2089 | * updating of the commit page and the clearing of the
|
| 2090 | * committing counter.
|
| 2091 |
|
| 2092 | if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
|
| 2093 | !local_read(&cpu_buffer->committing)) {
|
| 2094 | local_inc(&cpu_buffer->committing);
|
| 2095 | goto again;
|
| 2096 | }
|
| 2097 | }
|
| 2098 |
|
| 2099 | static struct ring_buffer_event *
|
| 2100 | rb_reserve_next_event(struct ring_buffer *buffer,
|
| 2101 | struct ring_buffer_per_cpu *cpu_buffer,
|
| 2102 | unsigned long length)
|
| 2103 | {
|
| 2104 | struct ring_buffer_event *event;
|
| 2105 | u64 ts, delta = 0;
|
| 2106 | int commit = 0;
|
| 2107 | int nr_loops = 0;
|
| 2108 |
|
| 2109 | rb_start_commit(cpu_buffer);
|
| 2110 |
|
| 2111 | #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
|
| 2112 | |
| 2113 | * Due to the ability to swap a cpu buffer from a buffer
|
| 2114 | * it is possible it was swapped before we committed.
|
| 2115 | * (committing stops a swap). We check for it here and
|
| 2116 | * if it happened, we have to fail the write.
|
| 2117 |
|
| 2118 | barrier();
|
| 2119 | if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
|
| 2120 | local_dec(&cpu_buffer->committing);
|
| 2121 | local_dec(&cpu_buffer->commits);
|
| 2122 | return NULL;
|
| 2123 | }
|
| 2124 | #endif
|
| 2125 |
|
| 2126 | length = rb_calculate_event_length(length);
|
| 2127 | again:
|
| 2128 | |
| 2129 | * We allow for interrupts to reenter here and do a trace.
|
| 2130 | * If one does, it will cause this original code to loop
|
| 2131 | * back here. Even with heavy interrupts happening, this
|
| 2132 | * should only happen a few times in a row. If this happens
|
| 2133 | * 1000 times in a row, there must be either an interrupt
|
| 2134 | * storm or we have something buggy.
|
| 2135 | * Bail!
|
| 2136 |
|
| 2137 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
|
| 2138 | goto out_fail;
|
| 2139 |
|
| 2140 | ts = rb_time_stamp(cpu_buffer->buffer);
|
| 2141 |
|
| 2142 | |
| 2143 | * Only the first commit can update the timestamp.
|
| 2144 | * Yes there is a race here. If an interrupt comes in
|
| 2145 | * just after the conditional and it traces too, then it
|
| 2146 | * will also check the deltas. More than one timestamp may
|
| 2147 | * also be made. But only the entry that did the actual
|
| 2148 | * commit will be something other than zero.
|
| 2149 |
|
| 2150 | if (likely(cpu_buffer->tail_page == cpu_buffer->commit_page &&
|
| 2151 | rb_page_write(cpu_buffer->tail_page) ==
|
| 2152 | rb_commit_index(cpu_buffer))) {
|
| 2153 | u64 diff;
|
| 2154 |
|
| 2155 | diff = ts - cpu_buffer->write_stamp;
|
| 2156 |
|
| 2157 |
|
| 2158 | barrier();
|
| 2159 |
|
| 2160 |
|
| 2161 | if (unlikely(ts < cpu_buffer->write_stamp))
|
| 2162 | goto get_event;
|
| 2163 |
|
| 2164 | delta = diff;
|
| 2165 | if (unlikely(test_time_stamp(delta))) {
|
| 2166 |
|
| 2167 | commit = rb_add_time_stamp(cpu_buffer, &ts, &delta);
|
| 2168 | if (commit == -EBUSY)
|
| 2169 | goto out_fail;
|
| 2170 |
|
| 2171 | if (commit == -EAGAIN)
|
| 2172 | goto again;
|
| 2173 |
|
| 2174 | RB_WARN_ON(cpu_buffer, commit < 0);
|
| 2175 | }
|
| 2176 | }
|
| 2177 |
|
| 2178 | get_event:
|
| 2179 | event = __rb_reserve_next(cpu_buffer, 0, length, &ts);
|
| 2180 | if (unlikely(PTR_ERR(event) == -EAGAIN))
|
| 2181 | goto again;
|
| 2182 |
|
| 2183 | if (!event)
|
| 2184 | goto out_fail;
|
| 2185 |
|
| 2186 | if (!rb_event_is_commit(cpu_buffer, event))
|
| 2187 | delta = 0;
|
| 2188 |
|
| 2189 | event->time_delta = delta;
|
| 2190 |
|
| 2191 | return event;
|
| 2192 |
|
| 2193 | out_fail:
|
| 2194 | rb_end_commit(cpu_buffer);
|
| 2195 | return NULL;
|
| 2196 | }
|
| 2197 |
|
| 2198 | #ifdef CONFIG_TRACING
|
| 2199 |
|
| 2200 | #define TRACE_RECURSIVE_DEPTH 16
|
| 2201 |
|
| 2202 | static int trace_recursive_lock(void)
|
| 2203 | {
|
| 2204 | current->trace_recursion++;
|
| 2205 |
|
| 2206 | if (likely(current->trace_recursion < TRACE_RECURSIVE_DEPTH))
|
| 2207 | return 0;
|
| 2208 |
|
| 2209 |
|
| 2210 | tracing_off_permanent();
|
| 2211 |
|
| 2212 | printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
|
| 2213 | "HC[%lu]:SC[%lu]:NMI[%lu]\n",
|
| 2214 | current->trace_recursion,
|
| 2215 | hardirq_count() >> HARDIRQ_SHIFT,
|
| 2216 | softirq_count() >> SOFTIRQ_SHIFT,
|
| 2217 | in_nmi());
|
| 2218 |
|
| 2219 | WARN_ON_ONCE(1);
|
| 2220 | return -1;
|
| 2221 | }
|
| 2222 |
|
| 2223 | static void trace_recursive_unlock(void)
|
| 2224 | {
|
| 2225 | WARN_ON_ONCE(!current->trace_recursion);
|
| 2226 |
|
| 2227 | current->trace_recursion--;
|
| 2228 | }
|
| 2229 |
|
| 2230 | #else
|
| 2231 |
|
| 2232 | #define trace_recursive_lock() (0)
|
| 2233 | #define trace_recursive_unlock() do { } while (0)
|
| 2234 |
|
| 2235 | #endif
|
| 2236 |
|
| 2237 | static DEFINE_PER_CPU(int, rb_need_resched);
|
| 2238 |
|
| 2239 | |
| 2240 | * ring_buffer_lock_reserve - reserve a part of the buffer
|
| 2241 | * @buffer: the ring buffer to reserve from
|
| 2242 | * @length: the length of the data to reserve (excluding event header)
|
| 2243 | *
|
| 2244 | * Returns a reseverd event on the ring buffer to copy directly to.
|
| 2245 | * The user of this interface will need to get the body to write into
|
| 2246 | * and can use the ring_buffer_event_data() interface.
|
| 2247 | *
|
| 2248 | * The length is the length of the data needed, not the event length
|
| 2249 | * which also includes the event header.
|
| 2250 | *
|
| 2251 | * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
|
| 2252 | * If NULL is returned, then nothing has been allocated or locked.
|
| 2253 |
|
| 2254 | struct ring_buffer_event *
|
| 2255 | ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
|
| 2256 | {
|
| 2257 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2258 | struct ring_buffer_event *event;
|
| 2259 | int cpu, resched;
|
| 2260 |
|
| 2261 | if (ring_buffer_flags != RB_BUFFERS_ON)
|
| 2262 | return NULL;
|
| 2263 |
|
| 2264 |
|
| 2265 | resched = ftrace_preempt_disable();
|
| 2266 |
|
| 2267 | if (atomic_read(&buffer->record_disabled))
|
| 2268 | goto out_nocheck;
|
| 2269 |
|
| 2270 | if (trace_recursive_lock())
|
| 2271 | goto out_nocheck;
|
| 2272 |
|
| 2273 | cpu = raw_smp_processor_id();
|
| 2274 |
|
| 2275 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 2276 | goto out;
|
| 2277 |
|
| 2278 | cpu_buffer = buffer->buffers[cpu];
|
| 2279 |
|
| 2280 | if (atomic_read(&cpu_buffer->record_disabled))
|
| 2281 | goto out;
|
| 2282 |
|
| 2283 | if (length > BUF_MAX_DATA_SIZE)
|
| 2284 | goto out;
|
| 2285 |
|
| 2286 | event = rb_reserve_next_event(buffer, cpu_buffer, length);
|
| 2287 | if (!event)
|
| 2288 | goto out;
|
| 2289 |
|
| 2290 | |
| 2291 | * Need to store resched state on this cpu.
|
| 2292 | * Only the first needs to.
|
| 2293 |
|
| 2294 |
|
| 2295 | if (preempt_count() == 1)
|
| 2296 | per_cpu(rb_need_resched, cpu) = resched;
|
| 2297 |
|
| 2298 | return event;
|
| 2299 |
|
| 2300 | out:
|
| 2301 | trace_recursive_unlock();
|
| 2302 |
|
| 2303 | out_nocheck:
|
| 2304 | ftrace_preempt_enable(resched);
|
| 2305 | return NULL;
|
| 2306 | }
|
| 2307 | EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
|
| 2308 |
|
| 2309 | static void
|
| 2310 | rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2311 | struct ring_buffer_event *event)
|
| 2312 | {
|
| 2313 | |
| 2314 | * The event first in the commit queue updates the
|
| 2315 | * time stamp.
|
| 2316 |
|
| 2317 | if (rb_event_is_commit(cpu_buffer, event))
|
| 2318 | cpu_buffer->write_stamp += event->time_delta;
|
| 2319 | }
|
| 2320 |
|
| 2321 | static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2322 | struct ring_buffer_event *event)
|
| 2323 | {
|
| 2324 | local_inc(&cpu_buffer->entries);
|
| 2325 | rb_update_write_stamp(cpu_buffer, event);
|
| 2326 | rb_end_commit(cpu_buffer);
|
| 2327 | }
|
| 2328 |
|
| 2329 | |
| 2330 | * ring_buffer_unlock_commit - commit a reserved
|
| 2331 | * @buffer: The buffer to commit to
|
| 2332 | * @event: The event pointer to commit.
|
| 2333 | *
|
| 2334 | * This commits the data to the ring buffer, and releases any locks held.
|
| 2335 | *
|
| 2336 | * Must be paired with ring_buffer_lock_reserve.
|
| 2337 |
|
| 2338 | int ring_buffer_unlock_commit(struct ring_buffer *buffer,
|
| 2339 | struct ring_buffer_event *event)
|
| 2340 | {
|
| 2341 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2342 | int cpu = raw_smp_processor_id();
|
| 2343 |
|
| 2344 | cpu_buffer = buffer->buffers[cpu];
|
| 2345 |
|
| 2346 | rb_commit(cpu_buffer, event);
|
| 2347 |
|
| 2348 | trace_recursive_unlock();
|
| 2349 |
|
| 2350 | |
| 2351 | * Only the last preempt count needs to restore preemption.
|
| 2352 |
|
| 2353 | if (preempt_count() == 1)
|
| 2354 | ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
|
| 2355 | else
|
| 2356 | preempt_enable_no_resched_notrace();
|
| 2357 |
|
| 2358 | return 0;
|
| 2359 | }
|
| 2360 | EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
|
| 2361 |
|
| 2362 | static inline void rb_event_discard(struct ring_buffer_event *event)
|
| 2363 | {
|
| 2364 |
|
| 2365 | event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
|
| 2366 | event->type_len = RINGBUF_TYPE_PADDING;
|
| 2367 |
|
| 2368 | if (!event->time_delta)
|
| 2369 | event->time_delta = 1;
|
| 2370 | }
|
| 2371 |
|
| 2372 | |
| 2373 | * Decrement the entries to the page that an event is on.
|
| 2374 | * The event does not even need to exist, only the pointer
|
| 2375 | * to the page it is on. This may only be called before the commit
|
| 2376 | * takes place.
|
| 2377 |
|
| 2378 | static inline void
|
| 2379 | rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2380 | struct ring_buffer_event *event)
|
| 2381 | {
|
| 2382 | unsigned long addr = (unsigned long)event;
|
| 2383 | struct buffer_page *bpage = cpu_buffer->commit_page;
|
| 2384 | struct buffer_page *start;
|
| 2385 |
|
| 2386 | addr &= PAGE_MASK;
|
| 2387 |
|
| 2388 |
|
| 2389 | if (likely(bpage->page == (void *)addr)) {
|
| 2390 | local_dec(&bpage->entries);
|
| 2391 | return;
|
| 2392 | }
|
| 2393 |
|
| 2394 | |
| 2395 | * Because the commit page may be on the reader page we
|
| 2396 | * start with the next page and check the end loop there.
|
| 2397 |
|
| 2398 | rb_inc_page(cpu_buffer, &bpage);
|
| 2399 | start = bpage;
|
| 2400 | do {
|
| 2401 | if (bpage->page == (void *)addr) {
|
| 2402 | local_dec(&bpage->entries);
|
| 2403 | return;
|
| 2404 | }
|
| 2405 | rb_inc_page(cpu_buffer, &bpage);
|
| 2406 | } while (bpage != start);
|
| 2407 |
|
| 2408 |
|
| 2409 | RB_WARN_ON(cpu_buffer, 1);
|
| 2410 | }
|
| 2411 |
|
| 2412 | |
| 2413 | * ring_buffer_commit_discard - discard an event that has not been committed
|
| 2414 | * @buffer: the ring buffer
|
| 2415 | * @event: non committed event to discard
|
| 2416 | *
|
| 2417 | * Sometimes an event that is in the ring buffer needs to be ignored.
|
| 2418 | * This function lets the user discard an event in the ring buffer
|
| 2419 | * and then that event will not be read later.
|
| 2420 | *
|
| 2421 | * This function only works if it is called before the the item has been
|
| 2422 | * committed. It will try to free the event from the ring buffer
|
| 2423 | * if another event has not been added behind it.
|
| 2424 | *
|
| 2425 | * If another event has been added behind it, it will set the event
|
| 2426 | * up as discarded, and perform the commit.
|
| 2427 | *
|
| 2428 | * If this function is called, do not call ring_buffer_unlock_commit on
|
| 2429 | * the event.
|
| 2430 |
|
| 2431 | void ring_buffer_discard_commit(struct ring_buffer *buffer,
|
| 2432 | struct ring_buffer_event *event)
|
| 2433 | {
|
| 2434 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2435 | int cpu;
|
| 2436 |
|
| 2437 |
|
| 2438 | rb_event_discard(event);
|
| 2439 |
|
| 2440 | cpu = smp_processor_id();
|
| 2441 | cpu_buffer = buffer->buffers[cpu];
|
| 2442 |
|
| 2443 | |
| 2444 | * This must only be called if the event has not been
|
| 2445 | * committed yet. Thus we can assume that preemption
|
| 2446 | * is still disabled.
|
| 2447 |
|
| 2448 | RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
|
| 2449 |
|
| 2450 | rb_decrement_entry(cpu_buffer, event);
|
| 2451 | if (rb_try_to_discard(cpu_buffer, event))
|
| 2452 | goto out;
|
| 2453 |
|
| 2454 | |
| 2455 | * The commit is still visible by the reader, so we
|
| 2456 | * must still update the timestamp.
|
| 2457 |
|
| 2458 | rb_update_write_stamp(cpu_buffer, event);
|
| 2459 | out:
|
| 2460 | rb_end_commit(cpu_buffer);
|
| 2461 |
|
| 2462 | trace_recursive_unlock();
|
| 2463 |
|
| 2464 | |
| 2465 | * Only the last preempt count needs to restore preemption.
|
| 2466 |
|
| 2467 | if (preempt_count() == 1)
|
| 2468 | ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
|
| 2469 | else
|
| 2470 | preempt_enable_no_resched_notrace();
|
| 2471 |
|
| 2472 | }
|
| 2473 | EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
|
| 2474 |
|
| 2475 | |
| 2476 | * ring_buffer_write - write data to the buffer without reserving
|
| 2477 | * @buffer: The ring buffer to write to.
|
| 2478 | * @length: The length of the data being written (excluding the event header)
|
| 2479 | * @data: The data to write to the buffer.
|
| 2480 | *
|
| 2481 | * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
|
| 2482 | * one function. If you already have the data to write to the buffer, it
|
| 2483 | * may be easier to simply call this function.
|
| 2484 | *
|
| 2485 | * Note, like ring_buffer_lock_reserve, the length is the length of the data
|
| 2486 | * and not the length of the event which would hold the header.
|
| 2487 |
|
| 2488 | int ring_buffer_write(struct ring_buffer *buffer,
|
| 2489 | unsigned long length,
|
| 2490 | void *data)
|
| 2491 | {
|
| 2492 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2493 | struct ring_buffer_event *event;
|
| 2494 | void *body;
|
| 2495 | int ret = -EBUSY;
|
| 2496 | int cpu, resched;
|
| 2497 |
|
| 2498 | if (ring_buffer_flags != RB_BUFFERS_ON)
|
| 2499 | return -EBUSY;
|
| 2500 |
|
| 2501 | resched = ftrace_preempt_disable();
|
| 2502 |
|
| 2503 | if (atomic_read(&buffer->record_disabled))
|
| 2504 | goto out;
|
| 2505 |
|
| 2506 | cpu = raw_smp_processor_id();
|
| 2507 |
|
| 2508 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 2509 | goto out;
|
| 2510 |
|
| 2511 | cpu_buffer = buffer->buffers[cpu];
|
| 2512 |
|
| 2513 | if (atomic_read(&cpu_buffer->record_disabled))
|
| 2514 | goto out;
|
| 2515 |
|
| 2516 | if (length > BUF_MAX_DATA_SIZE)
|
| 2517 | goto out;
|
| 2518 |
|
| 2519 | event = rb_reserve_next_event(buffer, cpu_buffer, length);
|
| 2520 | if (!event)
|
| 2521 | goto out;
|
| 2522 |
|
| 2523 | body = rb_event_data(event);
|
| 2524 |
|
| 2525 | memcpy(body, data, length);
|
| 2526 |
|
| 2527 | rb_commit(cpu_buffer, event);
|
| 2528 |
|
| 2529 | ret = 0;
|
| 2530 | out:
|
| 2531 | ftrace_preempt_enable(resched);
|
| 2532 |
|
| 2533 | return ret;
|
| 2534 | }
|
| 2535 | EXPORT_SYMBOL_GPL(ring_buffer_write);
|
| 2536 |
|
| 2537 | static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
|
| 2538 | {
|
| 2539 | struct buffer_page *reader = cpu_buffer->reader_page;
|
| 2540 | struct buffer_page *head = rb_set_head_page(cpu_buffer);
|
| 2541 | struct buffer_page *commit = cpu_buffer->commit_page;
|
| 2542 |
|
| 2543 |
|
| 2544 | if (unlikely(!head))
|
| 2545 | return 1;
|
| 2546 |
|
| 2547 | return reader->read == rb_page_commit(reader) &&
|
| 2548 | (commit == reader ||
|
| 2549 | (commit == head &&
|
| 2550 | head->read == rb_page_commit(commit)));
|
| 2551 | }
|
| 2552 |
|
| 2553 | |
| 2554 | * ring_buffer_record_disable - stop all writes into the buffer
|
| 2555 | * @buffer: The ring buffer to stop writes to.
|
| 2556 | *
|
| 2557 | * This prevents all writes to the buffer. Any attempt to write
|
| 2558 | * to the buffer after this will fail and return NULL.
|
| 2559 | *
|
| 2560 | * The caller should call synchronize_sched() after this.
|
| 2561 |
|
| 2562 | void ring_buffer_record_disable(struct ring_buffer *buffer)
|
| 2563 | {
|
| 2564 | atomic_inc(&buffer->record_disabled);
|
| 2565 | }
|
| 2566 | EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
|
| 2567 |
|
| 2568 | |
| 2569 | * ring_buffer_record_enable - enable writes to the buffer
|
| 2570 | * @buffer: The ring buffer to enable writes
|
| 2571 | *
|
| 2572 | * Note, multiple disables will need the same number of enables
|
| 2573 | * to truly enable the writing (much like preempt_disable).
|
| 2574 |
|
| 2575 | void ring_buffer_record_enable(struct ring_buffer *buffer)
|
| 2576 | {
|
| 2577 | atomic_dec(&buffer->record_disabled);
|
| 2578 | }
|
| 2579 | EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
|
| 2580 |
|
| 2581 | |
| 2582 | * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
|
| 2583 | * @buffer: The ring buffer to stop writes to.
|
| 2584 | * @cpu: The CPU buffer to stop
|
| 2585 | *
|
| 2586 | * This prevents all writes to the buffer. Any attempt to write
|
| 2587 | * to the buffer after this will fail and return NULL.
|
| 2588 | *
|
| 2589 | * The caller should call synchronize_sched() after this.
|
| 2590 |
|
| 2591 | void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
|
| 2592 | {
|
| 2593 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2594 |
|
| 2595 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 2596 | return;
|
| 2597 |
|
| 2598 | cpu_buffer = buffer->buffers[cpu];
|
| 2599 | atomic_inc(&cpu_buffer->record_disabled);
|
| 2600 | }
|
| 2601 | EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
|
| 2602 |
|
| 2603 | |
| 2604 | * ring_buffer_record_enable_cpu - enable writes to the buffer
|
| 2605 | * @buffer: The ring buffer to enable writes
|
| 2606 | * @cpu: The CPU to enable.
|
| 2607 | *
|
| 2608 | * Note, multiple disables will need the same number of enables
|
| 2609 | * to truly enable the writing (much like preempt_disable).
|
| 2610 |
|
| 2611 | void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
|
| 2612 | {
|
| 2613 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2614 |
|
| 2615 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 2616 | return;
|
| 2617 |
|
| 2618 | cpu_buffer = buffer->buffers[cpu];
|
| 2619 | atomic_dec(&cpu_buffer->record_disabled);
|
| 2620 | }
|
| 2621 | EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
|
| 2622 |
|
| 2623 | |
| 2624 | * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
|
| 2625 | * @buffer: The ring buffer
|
| 2626 | * @cpu: The per CPU buffer to get the entries from.
|
| 2627 |
|
| 2628 | unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
|
| 2629 | {
|
| 2630 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2631 | unsigned long ret;
|
| 2632 |
|
| 2633 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 2634 | return 0;
|
| 2635 |
|
| 2636 | cpu_buffer = buffer->buffers[cpu];
|
| 2637 | ret = (local_read(&cpu_buffer->entries) - local_read(&cpu_buffer->overrun))
|
| 2638 | - cpu_buffer->read;
|
| 2639 |
|
| 2640 | return ret;
|
| 2641 | }
|
| 2642 | EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
|
| 2643 |
|
| 2644 | |
| 2645 | * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
|
| 2646 | * @buffer: The ring buffer
|
| 2647 | * @cpu: The per CPU buffer to get the number of overruns from
|
| 2648 |
|
| 2649 | unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
|
| 2650 | {
|
| 2651 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2652 | unsigned long ret;
|
| 2653 |
|
| 2654 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 2655 | return 0;
|
| 2656 |
|
| 2657 | cpu_buffer = buffer->buffers[cpu];
|
| 2658 | ret = local_read(&cpu_buffer->overrun);
|
| 2659 |
|
| 2660 | return ret;
|
| 2661 | }
|
| 2662 | EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
|
| 2663 |
|
| 2664 | |
| 2665 | * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
|
| 2666 | * @buffer: The ring buffer
|
| 2667 | * @cpu: The per CPU buffer to get the number of overruns from
|
| 2668 |
|
| 2669 | unsigned long
|
| 2670 | ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
|
| 2671 | {
|
| 2672 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2673 | unsigned long ret;
|
| 2674 |
|
| 2675 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 2676 | return 0;
|
| 2677 |
|
| 2678 | cpu_buffer = buffer->buffers[cpu];
|
| 2679 | ret = local_read(&cpu_buffer->commit_overrun);
|
| 2680 |
|
| 2681 | return ret;
|
| 2682 | }
|
| 2683 | EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
|
| 2684 |
|
| 2685 | |
| 2686 | * ring_buffer_entries - get the number of entries in a buffer
|
| 2687 | * @buffer: The ring buffer
|
| 2688 | *
|
| 2689 | * Returns the total number of entries in the ring buffer
|
| 2690 | * (all CPU entries)
|
| 2691 |
|
| 2692 | unsigned long ring_buffer_entries(struct ring_buffer *buffer)
|
| 2693 | {
|
| 2694 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2695 | unsigned long entries = 0;
|
| 2696 | int cpu;
|
| 2697 |
|
| 2698 |
|
| 2699 | for_each_buffer_cpu(buffer, cpu) {
|
| 2700 | cpu_buffer = buffer->buffers[cpu];
|
| 2701 | entries += (local_read(&cpu_buffer->entries) -
|
| 2702 | local_read(&cpu_buffer->overrun)) - cpu_buffer->read;
|
| 2703 | }
|
| 2704 |
|
| 2705 | return entries;
|
| 2706 | }
|
| 2707 | EXPORT_SYMBOL_GPL(ring_buffer_entries);
|
| 2708 |
|
| 2709 | |
| 2710 | * ring_buffer_overruns - get the number of overruns in buffer
|
| 2711 | * @buffer: The ring buffer
|
| 2712 | *
|
| 2713 | * Returns the total number of overruns in the ring buffer
|
| 2714 | * (all CPU entries)
|
| 2715 |
|
| 2716 | unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
|
| 2717 | {
|
| 2718 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2719 | unsigned long overruns = 0;
|
| 2720 | int cpu;
|
| 2721 |
|
| 2722 |
|
| 2723 | for_each_buffer_cpu(buffer, cpu) {
|
| 2724 | cpu_buffer = buffer->buffers[cpu];
|
| 2725 | overruns += local_read(&cpu_buffer->overrun);
|
| 2726 | }
|
| 2727 |
|
| 2728 | return overruns;
|
| 2729 | }
|
| 2730 | EXPORT_SYMBOL_GPL(ring_buffer_overruns);
|
| 2731 |
|
| 2732 | static void rb_iter_reset(struct ring_buffer_iter *iter)
|
| 2733 | {
|
| 2734 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
|
| 2735 |
|
| 2736 |
|
| 2737 | if (list_empty(&cpu_buffer->reader_page->list)) {
|
| 2738 | iter->head_page = rb_set_head_page(cpu_buffer);
|
| 2739 | if (unlikely(!iter->head_page))
|
| 2740 | return;
|
| 2741 | iter->head = iter->head_page->read;
|
| 2742 | } else {
|
| 2743 | iter->head_page = cpu_buffer->reader_page;
|
| 2744 | iter->head = cpu_buffer->reader_page->read;
|
| 2745 | }
|
| 2746 | if (iter->head)
|
| 2747 | iter->read_stamp = cpu_buffer->read_stamp;
|
| 2748 | else
|
| 2749 | iter->read_stamp = iter->head_page->page->time_stamp;
|
| 2750 | iter->cache_reader_page = cpu_buffer->reader_page;
|
| 2751 | iter->cache_read = cpu_buffer->read;
|
| 2752 | }
|
| 2753 |
|
| 2754 | |
| 2755 | * ring_buffer_iter_reset - reset an iterator
|
| 2756 | * @iter: The iterator to reset
|
| 2757 | *
|
| 2758 | * Resets the iterator, so that it will start from the beginning
|
| 2759 | * again.
|
| 2760 |
|
| 2761 | void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
|
| 2762 | {
|
| 2763 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2764 | unsigned long flags;
|
| 2765 |
|
| 2766 | if (!iter)
|
| 2767 | return;
|
| 2768 |
|
| 2769 | cpu_buffer = iter->cpu_buffer;
|
| 2770 |
|
| 2771 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 2772 | rb_iter_reset(iter);
|
| 2773 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 2774 | }
|
| 2775 | EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
|
| 2776 |
|
| 2777 | |
| 2778 | * ring_buffer_iter_empty - check if an iterator has no more to read
|
| 2779 | * @iter: The iterator to check
|
| 2780 |
|
| 2781 | int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
|
| 2782 | {
|
| 2783 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2784 |
|
| 2785 | cpu_buffer = iter->cpu_buffer;
|
| 2786 |
|
| 2787 | return iter->head_page == cpu_buffer->commit_page &&
|
| 2788 | iter->head == rb_commit_index(cpu_buffer);
|
| 2789 | }
|
| 2790 | EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
|
| 2791 |
|
| 2792 | static void
|
| 2793 | rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2794 | struct ring_buffer_event *event)
|
| 2795 | {
|
| 2796 | u64 delta;
|
| 2797 |
|
| 2798 | switch (event->type_len) {
|
| 2799 | case RINGBUF_TYPE_PADDING:
|
| 2800 | return;
|
| 2801 |
|
| 2802 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 2803 | delta = event->array[0];
|
| 2804 | delta <<= TS_SHIFT;
|
| 2805 | delta += event->time_delta;
|
| 2806 | cpu_buffer->read_stamp += delta;
|
| 2807 | return;
|
| 2808 |
|
| 2809 | case RINGBUF_TYPE_TIME_STAMP:
|
| 2810 |
|
| 2811 | return;
|
| 2812 |
|
| 2813 | case RINGBUF_TYPE_DATA:
|
| 2814 | cpu_buffer->read_stamp += event->time_delta;
|
| 2815 | return;
|
| 2816 |
|
| 2817 | default:
|
| 2818 | BUG();
|
| 2819 | }
|
| 2820 | return;
|
| 2821 | }
|
| 2822 |
|
| 2823 | static void
|
| 2824 | rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
|
| 2825 | struct ring_buffer_event *event)
|
| 2826 | {
|
| 2827 | u64 delta;
|
| 2828 |
|
| 2829 | switch (event->type_len) {
|
| 2830 | case RINGBUF_TYPE_PADDING:
|
| 2831 | return;
|
| 2832 |
|
| 2833 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 2834 | delta = event->array[0];
|
| 2835 | delta <<= TS_SHIFT;
|
| 2836 | delta += event->time_delta;
|
| 2837 | iter->read_stamp += delta;
|
| 2838 | return;
|
| 2839 |
|
| 2840 | case RINGBUF_TYPE_TIME_STAMP:
|
| 2841 |
|
| 2842 | return;
|
| 2843 |
|
| 2844 | case RINGBUF_TYPE_DATA:
|
| 2845 | iter->read_stamp += event->time_delta;
|
| 2846 | return;
|
| 2847 |
|
| 2848 | default:
|
| 2849 | BUG();
|
| 2850 | }
|
| 2851 | return;
|
| 2852 | }
|
| 2853 |
|
| 2854 | static struct buffer_page *
|
| 2855 | rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
|
| 2856 | {
|
| 2857 | struct buffer_page *reader = NULL;
|
| 2858 | unsigned long overwrite;
|
| 2859 | unsigned long flags;
|
| 2860 | int nr_loops = 0;
|
| 2861 | int ret;
|
| 2862 |
|
| 2863 | local_irq_save(flags);
|
| 2864 | arch_spin_lock(&cpu_buffer->lock);
|
| 2865 |
|
| 2866 | again:
|
| 2867 | |
| 2868 | * This should normally only loop twice. But because the
|
| 2869 | * start of the reader inserts an empty page, it causes
|
| 2870 | * a case where we will loop three times. There should be no
|
| 2871 | * reason to loop four times (that I know of).
|
| 2872 |
|
| 2873 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
|
| 2874 | reader = NULL;
|
| 2875 | goto out;
|
| 2876 | }
|
| 2877 |
|
| 2878 | reader = cpu_buffer->reader_page;
|
| 2879 |
|
| 2880 |
|
| 2881 | if (cpu_buffer->reader_page->read < rb_page_size(reader))
|
| 2882 | goto out;
|
| 2883 |
|
| 2884 |
|
| 2885 | if (RB_WARN_ON(cpu_buffer,
|
| 2886 | cpu_buffer->reader_page->read > rb_page_size(reader)))
|
| 2887 | goto out;
|
| 2888 |
|
| 2889 |
|
| 2890 | reader = NULL;
|
| 2891 | if (cpu_buffer->commit_page == cpu_buffer->reader_page)
|
| 2892 | goto out;
|
| 2893 |
|
| 2894 | |
| 2895 | * Reset the reader page to size zero.
|
| 2896 |
|
| 2897 | local_set(&cpu_buffer->reader_page->write, 0);
|
| 2898 | local_set(&cpu_buffer->reader_page->entries, 0);
|
| 2899 | local_set(&cpu_buffer->reader_page->page->commit, 0);
|
| 2900 | cpu_buffer->reader_page->real_end = 0;
|
| 2901 |
|
| 2902 | spin:
|
| 2903 | |
| 2904 | * Splice the empty reader page into the list around the head.
|
| 2905 |
|
| 2906 | reader = rb_set_head_page(cpu_buffer);
|
| 2907 | cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
|
| 2908 | cpu_buffer->reader_page->list.prev = reader->list.prev;
|
| 2909 |
|
| 2910 | |
| 2911 | * cpu_buffer->pages just needs to point to the buffer, it
|
| 2912 | * has no specific buffer page to point to. Lets move it out
|
| 2913 | * of our way so we don't accidently swap it.
|
| 2914 |
|
| 2915 | cpu_buffer->pages = reader->list.prev;
|
| 2916 |
|
| 2917 |
|
| 2918 | rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
|
| 2919 |
|
| 2920 | |
| 2921 | * We want to make sure we read the overruns after we set up our
|
| 2922 | * pointers to the next object. The writer side does a
|
| 2923 | * cmpxchg to cross pages which acts as the mb on the writer
|
| 2924 | * side. Note, the reader will constantly fail the swap
|
| 2925 | * while the writer is updating the pointers, so this
|
| 2926 | * guarantees that the overwrite recorded here is the one we
|
| 2927 | * want to compare with the last_overrun.
|
| 2928 |
|
| 2929 | smp_mb();
|
| 2930 | overwrite = local_read(&(cpu_buffer->overrun));
|
| 2931 |
|
| 2932 | |
| 2933 | * Here's the tricky part.
|
| 2934 | *
|
| 2935 | * We need to move the pointer past the header page.
|
| 2936 | * But we can only do that if a writer is not currently
|
| 2937 | * moving it. The page before the header page has the
|
| 2938 | * flag bit '1' set if it is pointing to the page we want.
|
| 2939 | * but if the writer is in the process of moving it
|
| 2940 | * than it will be '2' or already moved '0'.
|
| 2941 |
|
| 2942 |
|
| 2943 | ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
|
| 2944 |
|
| 2945 | |
| 2946 | * If we did not convert it, then we must try again.
|
| 2947 |
|
| 2948 | if (!ret)
|
| 2949 | goto spin;
|
| 2950 |
|
| 2951 | |
| 2952 | * Yeah! We succeeded in replacing the page.
|
| 2953 | *
|
| 2954 | * Now make the new head point back to the reader page.
|
| 2955 |
|
| 2956 | rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
|
| 2957 | rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
|
| 2958 |
|
| 2959 |
|
| 2960 | cpu_buffer->reader_page = reader;
|
| 2961 | rb_reset_reader_page(cpu_buffer);
|
| 2962 |
|
| 2963 | if (overwrite != cpu_buffer->last_overrun) {
|
| 2964 | cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
|
| 2965 | cpu_buffer->last_overrun = overwrite;
|
| 2966 | }
|
| 2967 |
|
| 2968 | goto again;
|
| 2969 |
|
| 2970 | out:
|
| 2971 | arch_spin_unlock(&cpu_buffer->lock);
|
| 2972 | local_irq_restore(flags);
|
| 2973 |
|
| 2974 | return reader;
|
| 2975 | }
|
| 2976 |
|
| 2977 | static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
|
| 2978 | {
|
| 2979 | struct ring_buffer_event *event;
|
| 2980 | struct buffer_page *reader;
|
| 2981 | unsigned length;
|
| 2982 |
|
| 2983 | reader = rb_get_reader_page(cpu_buffer);
|
| 2984 |
|
| 2985 |
|
| 2986 | if (RB_WARN_ON(cpu_buffer, !reader))
|
| 2987 | return;
|
| 2988 |
|
| 2989 | event = rb_reader_event(cpu_buffer);
|
| 2990 |
|
| 2991 | if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
|
| 2992 | cpu_buffer->read++;
|
| 2993 |
|
| 2994 | rb_update_read_stamp(cpu_buffer, event);
|
| 2995 |
|
| 2996 | length = rb_event_length(event);
|
| 2997 | cpu_buffer->reader_page->read += length;
|
| 2998 | }
|
| 2999 |
|
| 3000 | static void rb_advance_iter(struct ring_buffer_iter *iter)
|
| 3001 | {
|
| 3002 | struct ring_buffer *buffer;
|
| 3003 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3004 | struct ring_buffer_event *event;
|
| 3005 | unsigned length;
|
| 3006 |
|
| 3007 | cpu_buffer = iter->cpu_buffer;
|
| 3008 | buffer = cpu_buffer->buffer;
|
| 3009 |
|
| 3010 | |
| 3011 | * Check if we are at the end of the buffer.
|
| 3012 |
|
| 3013 | if (iter->head >= rb_page_size(iter->head_page)) {
|
| 3014 |
|
| 3015 | if (iter->head_page == cpu_buffer->commit_page)
|
| 3016 | return;
|
| 3017 | rb_inc_iter(iter);
|
| 3018 | return;
|
| 3019 | }
|
| 3020 |
|
| 3021 | event = rb_iter_head_event(iter);
|
| 3022 |
|
| 3023 | length = rb_event_length(event);
|
| 3024 |
|
| 3025 | |
| 3026 | * This should not be called to advance the header if we are
|
| 3027 | * at the tail of the buffer.
|
| 3028 |
|
| 3029 | if (RB_WARN_ON(cpu_buffer,
|
| 3030 | (iter->head_page == cpu_buffer->commit_page) &&
|
| 3031 | (iter->head + length > rb_commit_index(cpu_buffer))))
|
| 3032 | return;
|
| 3033 |
|
| 3034 | rb_update_iter_read_stamp(iter, event);
|
| 3035 |
|
| 3036 | iter->head += length;
|
| 3037 |
|
| 3038 |
|
| 3039 | if ((iter->head >= rb_page_size(iter->head_page)) &&
|
| 3040 | (iter->head_page != cpu_buffer->commit_page))
|
| 3041 | rb_advance_iter(iter);
|
| 3042 | }
|
| 3043 |
|
| 3044 | static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
|
| 3045 | {
|
| 3046 | return cpu_buffer->lost_events;
|
| 3047 | }
|
| 3048 |
|
| 3049 | static struct ring_buffer_event *
|
| 3050 | rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
|
| 3051 | unsigned long *lost_events)
|
| 3052 | {
|
| 3053 | struct ring_buffer_event *event;
|
| 3054 | struct buffer_page *reader;
|
| 3055 | int nr_loops = 0;
|
| 3056 |
|
| 3057 | again:
|
| 3058 | |
| 3059 | * We repeat when a timestamp is encountered. It is possible
|
| 3060 | * to get multiple timestamps from an interrupt entering just
|
| 3061 | * as one timestamp is about to be written, or from discarded
|
| 3062 | * commits. The most that we can have is the number on a single page.
|
| 3063 |
|
| 3064 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
|
| 3065 | return NULL;
|
| 3066 |
|
| 3067 | reader = rb_get_reader_page(cpu_buffer);
|
| 3068 | if (!reader)
|
| 3069 | return NULL;
|
| 3070 |
|
| 3071 | event = rb_reader_event(cpu_buffer);
|
| 3072 |
|
| 3073 | switch (event->type_len) {
|
| 3074 | case RINGBUF_TYPE_PADDING:
|
| 3075 | if (rb_null_event(event))
|
| 3076 | RB_WARN_ON(cpu_buffer, 1);
|
| 3077 | |
| 3078 | * Because the writer could be discarding every
|
| 3079 | * event it creates (which would probably be bad)
|
| 3080 | * if we were to go back to "again" then we may never
|
| 3081 | * catch up, and will trigger the warn on, or lock
|
| 3082 | * the box. Return the padding, and we will release
|
| 3083 | * the current locks, and try again.
|
| 3084 |
|
| 3085 | return event;
|
| 3086 |
|
| 3087 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 3088 |
|
| 3089 | rb_advance_reader(cpu_buffer);
|
| 3090 | goto again;
|
| 3091 |
|
| 3092 | case RINGBUF_TYPE_TIME_STAMP:
|
| 3093 |
|
| 3094 | rb_advance_reader(cpu_buffer);
|
| 3095 | goto again;
|
| 3096 |
|
| 3097 | case RINGBUF_TYPE_DATA:
|
| 3098 | if (ts) {
|
| 3099 | *ts = cpu_buffer->read_stamp + event->time_delta;
|
| 3100 | ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
|
| 3101 | cpu_buffer->cpu, ts);
|
| 3102 | }
|
| 3103 | if (lost_events)
|
| 3104 | *lost_events = rb_lost_events(cpu_buffer);
|
| 3105 | return event;
|
| 3106 |
|
| 3107 | default:
|
| 3108 | BUG();
|
| 3109 | }
|
| 3110 |
|
| 3111 | return NULL;
|
| 3112 | }
|
| 3113 | EXPORT_SYMBOL_GPL(ring_buffer_peek);
|
| 3114 |
|
| 3115 | static struct ring_buffer_event *
|
| 3116 | rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
|
| 3117 | {
|
| 3118 | struct ring_buffer *buffer;
|
| 3119 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3120 | struct ring_buffer_event *event;
|
| 3121 | int nr_loops = 0;
|
| 3122 |
|
| 3123 | cpu_buffer = iter->cpu_buffer;
|
| 3124 | buffer = cpu_buffer->buffer;
|
| 3125 |
|
| 3126 | |
| 3127 | * Check if someone performed a consuming read to
|
| 3128 | * the buffer. A consuming read invalidates the iterator
|
| 3129 | * and we need to reset the iterator in this case.
|
| 3130 |
|
| 3131 | if (unlikely(iter->cache_read != cpu_buffer->read ||
|
| 3132 | iter->cache_reader_page != cpu_buffer->reader_page))
|
| 3133 | rb_iter_reset(iter);
|
| 3134 |
|
| 3135 | again:
|
| 3136 | if (ring_buffer_iter_empty(iter))
|
| 3137 | return NULL;
|
| 3138 |
|
| 3139 | |
| 3140 | * We repeat when a timestamp is encountered.
|
| 3141 | * We can get multiple timestamps by nested interrupts or also
|
| 3142 | * if filtering is on (discarding commits). Since discarding
|
| 3143 | * commits can be frequent we can get a lot of timestamps.
|
| 3144 | * But we limit them by not adding timestamps if they begin
|
| 3145 | * at the start of a page.
|
| 3146 |
|
| 3147 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
|
| 3148 | return NULL;
|
| 3149 |
|
| 3150 | if (rb_per_cpu_empty(cpu_buffer))
|
| 3151 | return NULL;
|
| 3152 |
|
| 3153 | if (iter->head >= local_read(&iter->head_page->page->commit)) {
|
| 3154 | rb_inc_iter(iter);
|
| 3155 | goto again;
|
| 3156 | }
|
| 3157 |
|
| 3158 | event = rb_iter_head_event(iter);
|
| 3159 |
|
| 3160 | switch (event->type_len) {
|
| 3161 | case RINGBUF_TYPE_PADDING:
|
| 3162 | if (rb_null_event(event)) {
|
| 3163 | rb_inc_iter(iter);
|
| 3164 | goto again;
|
| 3165 | }
|
| 3166 | rb_advance_iter(iter);
|
| 3167 | return event;
|
| 3168 |
|
| 3169 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 3170 |
|
| 3171 | rb_advance_iter(iter);
|
| 3172 | goto again;
|
| 3173 |
|
| 3174 | case RINGBUF_TYPE_TIME_STAMP:
|
| 3175 |
|
| 3176 | rb_advance_iter(iter);
|
| 3177 | goto again;
|
| 3178 |
|
| 3179 | case RINGBUF_TYPE_DATA:
|
| 3180 | if (ts) {
|
| 3181 | *ts = iter->read_stamp + event->time_delta;
|
| 3182 | ring_buffer_normalize_time_stamp(buffer,
|
| 3183 | cpu_buffer->cpu, ts);
|
| 3184 | }
|
| 3185 | return event;
|
| 3186 |
|
| 3187 | default:
|
| 3188 | BUG();
|
| 3189 | }
|
| 3190 |
|
| 3191 | return NULL;
|
| 3192 | }
|
| 3193 | EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
|
| 3194 |
|
| 3195 | static inline int rb_ok_to_lock(void)
|
| 3196 | {
|
| 3197 | |
| 3198 | * If an NMI die dumps out the content of the ring buffer
|
| 3199 | * do not grab locks. We also permanently disable the ring
|
| 3200 | * buffer too. A one time deal is all you get from reading
|
| 3201 | * the ring buffer from an NMI.
|
| 3202 |
|
| 3203 | if (likely(!in_nmi()))
|
| 3204 | return 1;
|
| 3205 |
|
| 3206 | tracing_off_permanent();
|
| 3207 | return 0;
|
| 3208 | }
|
| 3209 |
|
| 3210 | |
| 3211 | * ring_buffer_peek - peek at the next event to be read
|
| 3212 | * @buffer: The ring buffer to read
|
| 3213 | * @cpu: The cpu to peak at
|
| 3214 | * @ts: The timestamp counter of this event.
|
| 3215 | * @lost_events: a variable to store if events were lost (may be NULL)
|
| 3216 | *
|
| 3217 | * This will return the event that will be read next, but does
|
| 3218 | * not consume the data.
|
| 3219 |
|
| 3220 | struct ring_buffer_event *
|
| 3221 | ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
|
| 3222 | unsigned long *lost_events)
|
| 3223 | {
|
| 3224 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
|
| 3225 | struct ring_buffer_event *event;
|
| 3226 | unsigned long flags;
|
| 3227 | int dolock;
|
| 3228 |
|
| 3229 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 3230 | return NULL;
|
| 3231 |
|
| 3232 | dolock = rb_ok_to_lock();
|
| 3233 | again:
|
| 3234 | local_irq_save(flags);
|
| 3235 | if (dolock)
|
| 3236 | spin_lock(&cpu_buffer->reader_lock);
|
| 3237 | event = rb_buffer_peek(cpu_buffer, ts, lost_events);
|
| 3238 | if (event && event->type_len == RINGBUF_TYPE_PADDING)
|
| 3239 | rb_advance_reader(cpu_buffer);
|
| 3240 | if (dolock)
|
| 3241 | spin_unlock(&cpu_buffer->reader_lock);
|
| 3242 | local_irq_restore(flags);
|
| 3243 |
|
| 3244 | if (event && event->type_len == RINGBUF_TYPE_PADDING)
|
| 3245 | goto again;
|
| 3246 |
|
| 3247 | return event;
|
| 3248 | }
|
| 3249 |
|
| 3250 | |
| 3251 | * ring_buffer_iter_peek - peek at the next event to be read
|
| 3252 | * @iter: The ring buffer iterator
|
| 3253 | * @ts: The timestamp counter of this event.
|
| 3254 | *
|
| 3255 | * This will return the event that will be read next, but does
|
| 3256 | * not increment the iterator.
|
| 3257 |
|
| 3258 | struct ring_buffer_event *
|
| 3259 | ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
|
| 3260 | {
|
| 3261 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
|
| 3262 | struct ring_buffer_event *event;
|
| 3263 | unsigned long flags;
|
| 3264 |
|
| 3265 | again:
|
| 3266 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 3267 | event = rb_iter_peek(iter, ts);
|
| 3268 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 3269 |
|
| 3270 | if (event && event->type_len == RINGBUF_TYPE_PADDING)
|
| 3271 | goto again;
|
| 3272 |
|
| 3273 | return event;
|
| 3274 | }
|
| 3275 |
|
| 3276 | |
| 3277 | * ring_buffer_consume - return an event and consume it
|
| 3278 | * @buffer: The ring buffer to get the next event from
|
| 3279 | * @cpu: the cpu to read the buffer from
|
| 3280 | * @ts: a variable to store the timestamp (may be NULL)
|
| 3281 | * @lost_events: a variable to store if events were lost (may be NULL)
|
| 3282 | *
|
| 3283 | * Returns the next event in the ring buffer, and that event is consumed.
|
| 3284 | * Meaning, that sequential reads will keep returning a different event,
|
| 3285 | * and eventually empty the ring buffer if the producer is slower.
|
| 3286 |
|
| 3287 | struct ring_buffer_event *
|
| 3288 | ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
|
| 3289 | unsigned long *lost_events)
|
| 3290 | {
|
| 3291 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3292 | struct ring_buffer_event *event = NULL;
|
| 3293 | unsigned long flags;
|
| 3294 | int dolock;
|
| 3295 |
|
| 3296 | dolock = rb_ok_to_lock();
|
| 3297 |
|
| 3298 | again:
|
| 3299 |
|
| 3300 | preempt_disable();
|
| 3301 |
|
| 3302 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 3303 | goto out;
|
| 3304 |
|
| 3305 | cpu_buffer = buffer->buffers[cpu];
|
| 3306 | local_irq_save(flags);
|
| 3307 | if (dolock)
|
| 3308 | spin_lock(&cpu_buffer->reader_lock);
|
| 3309 |
|
| 3310 | event = rb_buffer_peek(cpu_buffer, ts, lost_events);
|
| 3311 | if (event) {
|
| 3312 | cpu_buffer->lost_events = 0;
|
| 3313 | rb_advance_reader(cpu_buffer);
|
| 3314 | }
|
| 3315 |
|
| 3316 | if (dolock)
|
| 3317 | spin_unlock(&cpu_buffer->reader_lock);
|
| 3318 | local_irq_restore(flags);
|
| 3319 |
|
| 3320 | out:
|
| 3321 | preempt_enable();
|
| 3322 |
|
| 3323 | if (event && event->type_len == RINGBUF_TYPE_PADDING)
|
| 3324 | goto again;
|
| 3325 |
|
| 3326 | return event;
|
| 3327 | }
|
| 3328 | EXPORT_SYMBOL_GPL(ring_buffer_consume);
|
| 3329 |
|
| 3330 | |
| 3331 | * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
|
| 3332 | * @buffer: The ring buffer to read from
|
| 3333 | * @cpu: The cpu buffer to iterate over
|
| 3334 | *
|
| 3335 | * This performs the initial preparations necessary to iterate
|
| 3336 | * through the buffer. Memory is allocated, buffer recording
|
| 3337 | * is disabled, and the iterator pointer is returned to the caller.
|
| 3338 | *
|
| 3339 | * Disabling buffer recordng prevents the reading from being
|
| 3340 | * corrupted. This is not a consuming read, so a producer is not
|
| 3341 | * expected.
|
| 3342 | *
|
| 3343 | * After a sequence of ring_buffer_read_prepare calls, the user is
|
| 3344 | * expected to make at least one call to ring_buffer_prepare_sync.
|
| 3345 | * Afterwards, ring_buffer_read_start is invoked to get things going
|
| 3346 | * for real.
|
| 3347 | *
|
| 3348 | * This overall must be paired with ring_buffer_finish.
|
| 3349 |
|
| 3350 | struct ring_buffer_iter *
|
| 3351 | ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
|
| 3352 | {
|
| 3353 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3354 | struct ring_buffer_iter *iter;
|
| 3355 |
|
| 3356 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 3357 | return NULL;
|
| 3358 |
|
| 3359 | iter = kmalloc(sizeof(*iter), GFP_KERNEL);
|
| 3360 | if (!iter)
|
| 3361 | return NULL;
|
| 3362 |
|
| 3363 | cpu_buffer = buffer->buffers[cpu];
|
| 3364 |
|
| 3365 | iter->cpu_buffer = cpu_buffer;
|
| 3366 |
|
| 3367 | atomic_inc(&cpu_buffer->record_disabled);
|
| 3368 |
|
| 3369 | return iter;
|
| 3370 | }
|
| 3371 | EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
|
| 3372 |
|
| 3373 | |
| 3374 | * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
|
| 3375 | *
|
| 3376 | * All previously invoked ring_buffer_read_prepare calls to prepare
|
| 3377 | * iterators will be synchronized. Afterwards, read_buffer_read_start
|
| 3378 | * calls on those iterators are allowed.
|
| 3379 |
|
| 3380 | void
|
| 3381 | ring_buffer_read_prepare_sync(void)
|
| 3382 | {
|
| 3383 | synchronize_sched();
|
| 3384 | }
|
| 3385 | EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
|
| 3386 |
|
| 3387 | |
| 3388 | * ring_buffer_read_start - start a non consuming read of the buffer
|
| 3389 | * @iter: The iterator returned by ring_buffer_read_prepare
|
| 3390 | *
|
| 3391 | * This finalizes the startup of an iteration through the buffer.
|
| 3392 | * The iterator comes from a call to ring_buffer_read_prepare and
|
| 3393 | * an intervening ring_buffer_read_prepare_sync must have been
|
| 3394 | * performed.
|
| 3395 | *
|
| 3396 | * Must be paired with ring_buffer_finish.
|
| 3397 |
|
| 3398 | void
|
| 3399 | ring_buffer_read_start(struct ring_buffer_iter *iter)
|
| 3400 | {
|
| 3401 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3402 | unsigned long flags;
|
| 3403 |
|
| 3404 | if (!iter)
|
| 3405 | return;
|
| 3406 |
|
| 3407 | cpu_buffer = iter->cpu_buffer;
|
| 3408 |
|
| 3409 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 3410 | arch_spin_lock(&cpu_buffer->lock);
|
| 3411 | rb_iter_reset(iter);
|
| 3412 | arch_spin_unlock(&cpu_buffer->lock);
|
| 3413 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 3414 | }
|
| 3415 | EXPORT_SYMBOL_GPL(ring_buffer_read_start);
|
| 3416 |
|
| 3417 | |
| 3418 | * ring_buffer_finish - finish reading the iterator of the buffer
|
| 3419 | * @iter: The iterator retrieved by ring_buffer_start
|
| 3420 | *
|
| 3421 | * This re-enables the recording to the buffer, and frees the
|
| 3422 | * iterator.
|
| 3423 |
|
| 3424 | void
|
| 3425 | ring_buffer_read_finish(struct ring_buffer_iter *iter)
|
| 3426 | {
|
| 3427 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
|
| 3428 |
|
| 3429 | atomic_dec(&cpu_buffer->record_disabled);
|
| 3430 | kfree(iter);
|
| 3431 | }
|
| 3432 | EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
|
| 3433 |
|
| 3434 | |
| 3435 | * ring_buffer_read - read the next item in the ring buffer by the iterator
|
| 3436 | * @iter: The ring buffer iterator
|
| 3437 | * @ts: The time stamp of the event read.
|
| 3438 | *
|
| 3439 | * This reads the next event in the ring buffer and increments the iterator.
|
| 3440 |
|
| 3441 | struct ring_buffer_event *
|
| 3442 | ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
|
| 3443 | {
|
| 3444 | struct ring_buffer_event *event;
|
| 3445 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
|
| 3446 | unsigned long flags;
|
| 3447 |
|
| 3448 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 3449 | again:
|
| 3450 | event = rb_iter_peek(iter, ts);
|
| 3451 | if (!event)
|
| 3452 | goto out;
|
| 3453 |
|
| 3454 | if (event->type_len == RINGBUF_TYPE_PADDING)
|
| 3455 | goto again;
|
| 3456 |
|
| 3457 | rb_advance_iter(iter);
|
| 3458 | out:
|
| 3459 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 3460 |
|
| 3461 | return event;
|
| 3462 | }
|
| 3463 | EXPORT_SYMBOL_GPL(ring_buffer_read);
|
| 3464 |
|
| 3465 | |
| 3466 | * ring_buffer_size - return the size of the ring buffer (in bytes)
|
| 3467 | * @buffer: The ring buffer.
|
| 3468 |
|
| 3469 | unsigned long ring_buffer_size(struct ring_buffer *buffer)
|
| 3470 | {
|
| 3471 | return BUF_PAGE_SIZE * buffer->pages;
|
| 3472 | }
|
| 3473 | EXPORT_SYMBOL_GPL(ring_buffer_size);
|
| 3474 |
|
| 3475 | static void
|
| 3476 | rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
|
| 3477 | {
|
| 3478 | rb_head_page_deactivate(cpu_buffer);
|
| 3479 |
|
| 3480 | cpu_buffer->head_page
|
| 3481 | = list_entry(cpu_buffer->pages, struct buffer_page, list);
|
| 3482 | local_set(&cpu_buffer->head_page->write, 0);
|
| 3483 | local_set(&cpu_buffer->head_page->entries, 0);
|
| 3484 | local_set(&cpu_buffer->head_page->page->commit, 0);
|
| 3485 |
|
| 3486 | cpu_buffer->head_page->read = 0;
|
| 3487 |
|
| 3488 | cpu_buffer->tail_page = cpu_buffer->head_page;
|
| 3489 | cpu_buffer->commit_page = cpu_buffer->head_page;
|
| 3490 |
|
| 3491 | INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
|
| 3492 | local_set(&cpu_buffer->reader_page->write, 0);
|
| 3493 | local_set(&cpu_buffer->reader_page->entries, 0);
|
| 3494 | local_set(&cpu_buffer->reader_page->page->commit, 0);
|
| 3495 | cpu_buffer->reader_page->read = 0;
|
| 3496 |
|
| 3497 | local_set(&cpu_buffer->commit_overrun, 0);
|
| 3498 | local_set(&cpu_buffer->overrun, 0);
|
| 3499 | local_set(&cpu_buffer->entries, 0);
|
| 3500 | local_set(&cpu_buffer->committing, 0);
|
| 3501 | local_set(&cpu_buffer->commits, 0);
|
| 3502 | cpu_buffer->read = 0;
|
| 3503 |
|
| 3504 | cpu_buffer->write_stamp = 0;
|
| 3505 | cpu_buffer->read_stamp = 0;
|
| 3506 |
|
| 3507 | cpu_buffer->lost_events = 0;
|
| 3508 | cpu_buffer->last_overrun = 0;
|
| 3509 |
|
| 3510 | rb_head_page_activate(cpu_buffer);
|
| 3511 | }
|
| 3512 |
|
| 3513 | |
| 3514 | * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
|
| 3515 | * @buffer: The ring buffer to reset a per cpu buffer of
|
| 3516 | * @cpu: The CPU buffer to be reset
|
| 3517 |
|
| 3518 | void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
|
| 3519 | {
|
| 3520 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
|
| 3521 | unsigned long flags;
|
| 3522 |
|
| 3523 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 3524 | return;
|
| 3525 |
|
| 3526 | atomic_inc(&cpu_buffer->record_disabled);
|
| 3527 |
|
| 3528 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 3529 |
|
| 3530 | if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
|
| 3531 | goto out;
|
| 3532 |
|
| 3533 | arch_spin_lock(&cpu_buffer->lock);
|
| 3534 |
|
| 3535 | rb_reset_cpu(cpu_buffer);
|
| 3536 |
|
| 3537 | arch_spin_unlock(&cpu_buffer->lock);
|
| 3538 |
|
| 3539 | out:
|
| 3540 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 3541 |
|
| 3542 | atomic_dec(&cpu_buffer->record_disabled);
|
| 3543 | }
|
| 3544 | EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
|
| 3545 |
|
| 3546 | |
| 3547 | * ring_buffer_reset - reset a ring buffer
|
| 3548 | * @buffer: The ring buffer to reset all cpu buffers
|
| 3549 |
|
| 3550 | void ring_buffer_reset(struct ring_buffer *buffer)
|
| 3551 | {
|
| 3552 | int cpu;
|
| 3553 |
|
| 3554 | for_each_buffer_cpu(buffer, cpu)
|
| 3555 | ring_buffer_reset_cpu(buffer, cpu);
|
| 3556 | }
|
| 3557 | EXPORT_SYMBOL_GPL(ring_buffer_reset);
|
| 3558 |
|
| 3559 | |
| 3560 | * rind_buffer_empty - is the ring buffer empty?
|
| 3561 | * @buffer: The ring buffer to test
|
| 3562 |
|
| 3563 | int ring_buffer_empty(struct ring_buffer *buffer)
|
| 3564 | {
|
| 3565 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3566 | unsigned long flags;
|
| 3567 | int dolock;
|
| 3568 | int cpu;
|
| 3569 | int ret;
|
| 3570 |
|
| 3571 | dolock = rb_ok_to_lock();
|
| 3572 |
|
| 3573 |
|
| 3574 | for_each_buffer_cpu(buffer, cpu) {
|
| 3575 | cpu_buffer = buffer->buffers[cpu];
|
| 3576 | local_irq_save(flags);
|
| 3577 | if (dolock)
|
| 3578 | spin_lock(&cpu_buffer->reader_lock);
|
| 3579 | ret = rb_per_cpu_empty(cpu_buffer);
|
| 3580 | if (dolock)
|
| 3581 | spin_unlock(&cpu_buffer->reader_lock);
|
| 3582 | local_irq_restore(flags);
|
| 3583 |
|
| 3584 | if (!ret)
|
| 3585 | return 0;
|
| 3586 | }
|
| 3587 |
|
| 3588 | return 1;
|
| 3589 | }
|
| 3590 | EXPORT_SYMBOL_GPL(ring_buffer_empty);
|
| 3591 |
|
| 3592 | |
| 3593 | * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
|
| 3594 | * @buffer: The ring buffer
|
| 3595 | * @cpu: The CPU buffer to test
|
| 3596 |
|
| 3597 | int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
|
| 3598 | {
|
| 3599 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3600 | unsigned long flags;
|
| 3601 | int dolock;
|
| 3602 | int ret;
|
| 3603 |
|
| 3604 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 3605 | return 1;
|
| 3606 |
|
| 3607 | dolock = rb_ok_to_lock();
|
| 3608 |
|
| 3609 | cpu_buffer = buffer->buffers[cpu];
|
| 3610 | local_irq_save(flags);
|
| 3611 | if (dolock)
|
| 3612 | spin_lock(&cpu_buffer->reader_lock);
|
| 3613 | ret = rb_per_cpu_empty(cpu_buffer);
|
| 3614 | if (dolock)
|
| 3615 | spin_unlock(&cpu_buffer->reader_lock);
|
| 3616 | local_irq_restore(flags);
|
| 3617 |
|
| 3618 | return ret;
|
| 3619 | }
|
| 3620 | EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
|
| 3621 |
|
| 3622 | #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
|
| 3623 | |
| 3624 | * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
|
| 3625 | * @buffer_a: One buffer to swap with
|
| 3626 | * @buffer_b: The other buffer to swap with
|
| 3627 | *
|
| 3628 | * This function is useful for tracers that want to take a "snapshot"
|
| 3629 | * of a CPU buffer and has another back up buffer lying around.
|
| 3630 | * it is expected that the tracer handles the cpu buffer not being
|
| 3631 | * used at the moment.
|
| 3632 |
|
| 3633 | int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
|
| 3634 | struct ring_buffer *buffer_b, int cpu)
|
| 3635 | {
|
| 3636 | struct ring_buffer_per_cpu *cpu_buffer_a;
|
| 3637 | struct ring_buffer_per_cpu *cpu_buffer_b;
|
| 3638 | int ret = -EINVAL;
|
| 3639 |
|
| 3640 | if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
|
| 3641 | !cpumask_test_cpu(cpu, buffer_b->cpumask))
|
| 3642 | goto out;
|
| 3643 |
|
| 3644 |
|
| 3645 | if (buffer_a->pages != buffer_b->pages)
|
| 3646 | goto out;
|
| 3647 |
|
| 3648 | ret = -EAGAIN;
|
| 3649 |
|
| 3650 | if (ring_buffer_flags != RB_BUFFERS_ON)
|
| 3651 | goto out;
|
| 3652 |
|
| 3653 | if (atomic_read(&buffer_a->record_disabled))
|
| 3654 | goto out;
|
| 3655 |
|
| 3656 | if (atomic_read(&buffer_b->record_disabled))
|
| 3657 | goto out;
|
| 3658 |
|
| 3659 | cpu_buffer_a = buffer_a->buffers[cpu];
|
| 3660 | cpu_buffer_b = buffer_b->buffers[cpu];
|
| 3661 |
|
| 3662 | if (atomic_read(&cpu_buffer_a->record_disabled))
|
| 3663 | goto out;
|
| 3664 |
|
| 3665 | if (atomic_read(&cpu_buffer_b->record_disabled))
|
| 3666 | goto out;
|
| 3667 |
|
| 3668 | |
| 3669 | * We can't do a synchronize_sched here because this
|
| 3670 | * function can be called in atomic context.
|
| 3671 | * Normally this will be called from the same CPU as cpu.
|
| 3672 | * If not it's up to the caller to protect this.
|
| 3673 |
|
| 3674 | atomic_inc(&cpu_buffer_a->record_disabled);
|
| 3675 | atomic_inc(&cpu_buffer_b->record_disabled);
|
| 3676 |
|
| 3677 | ret = -EBUSY;
|
| 3678 | if (local_read(&cpu_buffer_a->committing))
|
| 3679 | goto out_dec;
|
| 3680 | if (local_read(&cpu_buffer_b->committing))
|
| 3681 | goto out_dec;
|
| 3682 |
|
| 3683 | buffer_a->buffers[cpu] = cpu_buffer_b;
|
| 3684 | buffer_b->buffers[cpu] = cpu_buffer_a;
|
| 3685 |
|
| 3686 | cpu_buffer_b->buffer = buffer_a;
|
| 3687 | cpu_buffer_a->buffer = buffer_b;
|
| 3688 |
|
| 3689 | ret = 0;
|
| 3690 |
|
| 3691 | out_dec:
|
| 3692 | atomic_dec(&cpu_buffer_a->record_disabled);
|
| 3693 | atomic_dec(&cpu_buffer_b->record_disabled);
|
| 3694 | out:
|
| 3695 | return ret;
|
| 3696 | }
|
| 3697 | EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
|
| 3698 | #endif
|
| 3699 |
|
| 3700 | |
| 3701 | * ring_buffer_alloc_read_page - allocate a page to read from buffer
|
| 3702 | * @buffer: the buffer to allocate for.
|
| 3703 | *
|
| 3704 | * This function is used in conjunction with ring_buffer_read_page.
|
| 3705 | * When reading a full page from the ring buffer, these functions
|
| 3706 | * can be used to speed up the process. The calling function should
|
| 3707 | * allocate a few pages first with this function. Then when it
|
| 3708 | * needs to get pages from the ring buffer, it passes the result
|
| 3709 | * of this function into ring_buffer_read_page, which will swap
|
| 3710 | * the page that was allocated, with the read page of the buffer.
|
| 3711 | *
|
| 3712 | * Returns:
|
| 3713 | * The page allocated, or NULL on error.
|
| 3714 |
|
| 3715 | void *ring_buffer_alloc_read_page(struct ring_buffer *buffer)
|
| 3716 | {
|
| 3717 | struct buffer_data_page *bpage;
|
| 3718 | unsigned long addr;
|
| 3719 |
|
| 3720 | addr = __get_free_page(GFP_KERNEL);
|
| 3721 | if (!addr)
|
| 3722 | return NULL;
|
| 3723 |
|
| 3724 | bpage = (void *)addr;
|
| 3725 |
|
| 3726 | rb_init_page(bpage);
|
| 3727 |
|
| 3728 | return bpage;
|
| 3729 | }
|
| 3730 | EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
|
| 3731 |
|
| 3732 | |
| 3733 | * ring_buffer_free_read_page - free an allocated read page
|
| 3734 | * @buffer: the buffer the page was allocate for
|
| 3735 | * @data: the page to free
|
| 3736 | *
|
| 3737 | * Free a page allocated from ring_buffer_alloc_read_page.
|
| 3738 |
|
| 3739 | void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
|
| 3740 | {
|
| 3741 | free_page((unsigned long)data);
|
| 3742 | }
|
| 3743 | EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
|
| 3744 |
|
| 3745 | |
| 3746 | * ring_buffer_read_page - extract a page from the ring buffer
|
| 3747 | * @buffer: buffer to extract from
|
| 3748 | * @data_page: the page to use allocated from ring_buffer_alloc_read_page
|
| 3749 | * @len: amount to extract
|
| 3750 | * @cpu: the cpu of the buffer to extract
|
| 3751 | * @full: should the extraction only happen when the page is full.
|
| 3752 | *
|
| 3753 | * This function will pull out a page from the ring buffer and consume it.
|
| 3754 | * @data_page must be the address of the variable that was returned
|
| 3755 | * from ring_buffer_alloc_read_page. This is because the page might be used
|
| 3756 | * to swap with a page in the ring buffer.
|
| 3757 | *
|
| 3758 | * for example:
|
| 3759 | * rpage = ring_buffer_alloc_read_page(buffer);
|
| 3760 | * if (!rpage)
|
| 3761 | * return error;
|
| 3762 | * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
|
| 3763 | * if (ret >= 0)
|
| 3764 | * process_page(rpage, ret);
|
| 3765 | *
|
| 3766 | * When @full is set, the function will not return true unless
|
| 3767 | * the writer is off the reader page.
|
| 3768 | *
|
| 3769 | * Note: it is up to the calling functions to handle sleeps and wakeups.
|
| 3770 | * The ring buffer can be used anywhere in the kernel and can not
|
| 3771 | * blindly call wake_up. The layer that uses the ring buffer must be
|
| 3772 | * responsible for that.
|
| 3773 | *
|
| 3774 | * Returns:
|
| 3775 | * >=0 if data has been transferred, returns the offset of consumed data.
|
| 3776 | * <0 if no data has been transferred.
|
| 3777 |
|
| 3778 | int ring_buffer_read_page(struct ring_buffer *buffer,
|
| 3779 | void **data_page, size_t len, int cpu, int full)
|
| 3780 | {
|
| 3781 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
|
| 3782 | struct ring_buffer_event *event;
|
| 3783 | struct buffer_data_page *bpage;
|
| 3784 | struct buffer_page *reader;
|
| 3785 | unsigned long missed_events;
|
| 3786 | unsigned long flags;
|
| 3787 | unsigned int commit;
|
| 3788 | unsigned int read;
|
| 3789 | u64 save_timestamp;
|
| 3790 | int ret = -1;
|
| 3791 |
|
| 3792 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 3793 | goto out;
|
| 3794 |
|
| 3795 | |
| 3796 | * If len is not big enough to hold the page header, then
|
| 3797 | * we can not copy anything.
|
| 3798 |
|
| 3799 | if (len <= BUF_PAGE_HDR_SIZE)
|
| 3800 | goto out;
|
| 3801 |
|
| 3802 | len -= BUF_PAGE_HDR_SIZE;
|
| 3803 |
|
| 3804 | if (!data_page)
|
| 3805 | goto out;
|
| 3806 |
|
| 3807 | bpage = *data_page;
|
| 3808 | if (!bpage)
|
| 3809 | goto out;
|
| 3810 |
|
| 3811 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 3812 |
|
| 3813 | reader = rb_get_reader_page(cpu_buffer);
|
| 3814 | if (!reader)
|
| 3815 | goto out_unlock;
|
| 3816 |
|
| 3817 | event = rb_reader_event(cpu_buffer);
|
| 3818 |
|
| 3819 | read = reader->read;
|
| 3820 | commit = rb_page_commit(reader);
|
| 3821 |
|
| 3822 |
|
| 3823 | missed_events = cpu_buffer->lost_events;
|
| 3824 |
|
| 3825 | |
| 3826 | * If this page has been partially read or
|
| 3827 | * if len is not big enough to read the rest of the page or
|
| 3828 | * a writer is still on the page, then
|
| 3829 | * we must copy the data from the page to the buffer.
|
| 3830 | * Otherwise, we can simply swap the page with the one passed in.
|
| 3831 |
|
| 3832 | if (read || (len < (commit - read)) ||
|
| 3833 | cpu_buffer->reader_page == cpu_buffer->commit_page) {
|
| 3834 | struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
|
| 3835 | unsigned int rpos = read;
|
| 3836 | unsigned int pos = 0;
|
| 3837 | unsigned int size;
|
| 3838 |
|
| 3839 | if (full)
|
| 3840 | goto out_unlock;
|
| 3841 |
|
| 3842 | if (len > (commit - read))
|
| 3843 | len = (commit - read);
|
| 3844 |
|
| 3845 | size = rb_event_length(event);
|
| 3846 |
|
| 3847 | if (len < size)
|
| 3848 | goto out_unlock;
|
| 3849 |
|
| 3850 |
|
| 3851 | save_timestamp = cpu_buffer->read_stamp;
|
| 3852 |
|
| 3853 |
|
| 3854 | do {
|
| 3855 | memcpy(bpage->data + pos, rpage->data + rpos, size);
|
| 3856 |
|
| 3857 | len -= size;
|
| 3858 |
|
| 3859 | rb_advance_reader(cpu_buffer);
|
| 3860 | rpos = reader->read;
|
| 3861 | pos += size;
|
| 3862 |
|
| 3863 | event = rb_reader_event(cpu_buffer);
|
| 3864 | size = rb_event_length(event);
|
| 3865 | } while (len > size);
|
| 3866 |
|
| 3867 |
|
| 3868 | local_set(&bpage->commit, pos);
|
| 3869 | bpage->time_stamp = save_timestamp;
|
| 3870 |
|
| 3871 |
|
| 3872 | read = 0;
|
| 3873 | } else {
|
| 3874 |
|
| 3875 | cpu_buffer->read += rb_page_entries(reader);
|
| 3876 |
|
| 3877 |
|
| 3878 | rb_init_page(bpage);
|
| 3879 | bpage = reader->page;
|
| 3880 | reader->page = *data_page;
|
| 3881 | local_set(&reader->write, 0);
|
| 3882 | local_set(&reader->entries, 0);
|
| 3883 | reader->read = 0;
|
| 3884 | *data_page = bpage;
|
| 3885 |
|
| 3886 | |
| 3887 | * Use the real_end for the data size,
|
| 3888 | * This gives us a chance to store the lost events
|
| 3889 | * on the page.
|
| 3890 |
|
| 3891 | if (reader->real_end)
|
| 3892 | local_set(&bpage->commit, reader->real_end);
|
| 3893 | }
|
| 3894 | ret = read;
|
| 3895 |
|
| 3896 | cpu_buffer->lost_events = 0;
|
| 3897 | |
| 3898 | * Set a flag in the commit field if we lost events
|
| 3899 |
|
| 3900 | if (missed_events) {
|
| 3901 | commit = local_read(&bpage->commit);
|
| 3902 |
|
| 3903 | |
| 3904 | * missed events, then record it there.
|
| 3905 |
|
| 3906 | if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
|
| 3907 | memcpy(&bpage->data[commit], &missed_events,
|
| 3908 | sizeof(missed_events));
|
| 3909 | local_add(RB_MISSED_STORED, &bpage->commit);
|
| 3910 | }
|
| 3911 | local_add(RB_MISSED_EVENTS, &bpage->commit);
|
| 3912 | }
|
| 3913 |
|
| 3914 | out_unlock:
|
| 3915 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 3916 |
|
| 3917 | out:
|
| 3918 | return ret;
|
| 3919 | }
|
| 3920 | EXPORT_SYMBOL_GPL(ring_buffer_read_page);
|
| 3921 |
|
| 3922 | #ifdef CONFIG_TRACING
|
| 3923 | static ssize_t
|
| 3924 | rb_simple_read(struct file *filp, char __user *ubuf,
|
| 3925 | size_t cnt, loff_t *ppos)
|
| 3926 | {
|
| 3927 | unsigned long *p = filp->private_data;
|
| 3928 | char buf[64];
|
| 3929 | int r;
|
| 3930 |
|
| 3931 | if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
|
| 3932 | r = sprintf(buf, "permanently disabled\n");
|
| 3933 | else
|
| 3934 | r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
|
| 3935 |
|
| 3936 | return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
|
| 3937 | }
|
| 3938 |
|
| 3939 | static ssize_t
|
| 3940 | rb_simple_write(struct file *filp, const char __user *ubuf,
|
| 3941 | size_t cnt, loff_t *ppos)
|
| 3942 | {
|
| 3943 | unsigned long *p = filp->private_data;
|
| 3944 | char buf[64];
|
| 3945 | unsigned long val;
|
| 3946 | int ret;
|
| 3947 |
|
| 3948 | if (cnt >= sizeof(buf))
|
| 3949 | return -EINVAL;
|
| 3950 |
|
| 3951 | if (copy_from_user(&buf, ubuf, cnt))
|
| 3952 | return -EFAULT;
|
| 3953 |
|
| 3954 | buf[cnt] = 0;
|
| 3955 |
|
| 3956 | ret = strict_strtoul(buf, 10, &val);
|
| 3957 | if (ret < 0)
|
| 3958 | return ret;
|
| 3959 |
|
| 3960 | if (val)
|
| 3961 | set_bit(RB_BUFFERS_ON_BIT, p);
|
| 3962 | else
|
| 3963 | clear_bit(RB_BUFFERS_ON_BIT, p);
|
| 3964 |
|
| 3965 | (*ppos)++;
|
| 3966 |
|
| 3967 | return cnt;
|
| 3968 | }
|
| 3969 |
|
| 3970 | static const struct file_operations rb_simple_fops = {
|
| 3971 | .open = tracing_open_generic,
|
| 3972 | .read = rb_simple_read,
|
| 3973 | .write = rb_simple_write,
|
| 3974 | };
|
| 3975 |
|
| 3976 |
|
| 3977 | static __init int rb_init_debugfs(void)
|
| 3978 | {
|
| 3979 | struct dentry *d_tracer;
|
| 3980 |
|
| 3981 | d_tracer = tracing_init_dentry();
|
| 3982 |
|
| 3983 | trace_create_file("tracing_on", 0644, d_tracer,
|
| 3984 | &ring_buffer_flags, &rb_simple_fops);
|
| 3985 |
|
| 3986 | return 0;
|
| 3987 | }
|
| 3988 |
|
| 3989 | fs_initcall(rb_init_debugfs);
|
| 3990 | #endif
|
| 3991 |
|
| 3992 | #ifdef CONFIG_HOTPLUG_CPU
|
| 3993 | static int rb_cpu_notify(struct notifier_block *self,
|
| 3994 | unsigned long action, void *hcpu)
|
| 3995 | {
|
| 3996 | struct ring_buffer *buffer =
|
| 3997 | container_of(self, struct ring_buffer, cpu_notify);
|
| 3998 | long cpu = (long)hcpu;
|
| 3999 |
|
| 4000 | switch (action) {
|
| 4001 | case CPU_UP_PREPARE:
|
| 4002 | case CPU_UP_PREPARE_FROZEN:
|
| 4003 | if (cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4004 | return NOTIFY_OK;
|
| 4005 |
|
| 4006 | buffer->buffers[cpu] =
|
| 4007 | rb_allocate_cpu_buffer(buffer, cpu);
|
| 4008 | if (!buffer->buffers[cpu]) {
|
| 4009 | WARN(1, "failed to allocate ring buffer on CPU %ld\n",
|
| 4010 | cpu);
|
| 4011 | return NOTIFY_OK;
|
| 4012 | }
|
| 4013 | smp_wmb();
|
| 4014 | cpumask_set_cpu(cpu, buffer->cpumask);
|
| 4015 | break;
|
| 4016 | case CPU_DOWN_PREPARE:
|
| 4017 | case CPU_DOWN_PREPARE_FROZEN:
|
| 4018 | |
| 4019 | * Do nothing.
|
| 4020 | * If we were to free the buffer, then the user would
|
| 4021 | * lose any trace that was in the buffer.
|
| 4022 |
|
| 4023 | break;
|
| 4024 | default:
|
| 4025 | break;
|
| 4026 | }
|
| 4027 | return NOTIFY_OK;
|
| 4028 | }
|
| 4029 | #endif
|