| 1 | |
| 2 | |
| 3 | * Generic ring buffer
|
| 4 | *
|
| 5 | * Copyright (C) 2008 Steven Rostedt <[email protected]>
|
| 6 |
|
| 7 | #include <linux/trace_recursion.h>
|
| 8 | #include <linux/trace_events.h>
|
| 9 | #include <linux/ring_buffer.h>
|
| 10 | #include <linux/trace_clock.h>
|
| 11 | #include <linux/sched/clock.h>
|
| 12 | #include <linux/trace_seq.h>
|
| 13 | #include <linux/spinlock.h>
|
| 14 | #include <linux/irq_work.h>
|
| 15 | #include <linux/security.h>
|
| 16 | #include <linux/uaccess.h>
|
| 17 | #include <linux/hardirq.h>
|
| 18 | #include <linux/kthread.h>
|
| 19 | #include <linux/module.h>
|
| 20 | #include <linux/percpu.h>
|
| 21 | #include <linux/mutex.h>
|
| 22 | #include <linux/delay.h>
|
| 23 | #include <linux/slab.h>
|
| 24 | #include <linux/init.h>
|
| 25 | #include <linux/hash.h>
|
| 26 | #include <linux/list.h>
|
| 27 | #include <linux/cpu.h>
|
| 28 | #include <linux/oom.h>
|
| 29 |
|
| 30 | #include <asm/local.h>
|
| 31 |
|
| 32 | |
| 33 | * The "absolute" timestamp in the buffer is only 59 bits.
|
| 34 | * If a clock has the 5 MSBs set, it needs to be saved and
|
| 35 | * reinserted.
|
| 36 |
|
| 37 | #define TS_MSB (0xf8ULL << 56)
|
| 38 | #define ABS_TS_MASK (~TS_MSB)
|
| 39 |
|
| 40 | static void update_pages_handler(struct work_struct *work);
|
| 41 |
|
| 42 | |
| 43 | * The ring buffer header is special. We must manually up keep it.
|
| 44 |
|
| 45 | int ring_buffer_print_entry_header(struct trace_seq *s)
|
| 46 | {
|
| 47 | trace_seq_puts(s, "# compressed entry header\n");
|
| 48 | trace_seq_puts(s, "\ttype_len : 5 bits\n");
|
| 49 | trace_seq_puts(s, "\ttime_delta : 27 bits\n");
|
| 50 | trace_seq_puts(s, "\tarray : 32 bits\n");
|
| 51 | trace_seq_putc(s, '\n');
|
| 52 | trace_seq_printf(s, "\tpadding : type == %d\n",
|
| 53 | RINGBUF_TYPE_PADDING);
|
| 54 | trace_seq_printf(s, "\ttime_extend : type == %d\n",
|
| 55 | RINGBUF_TYPE_TIME_EXTEND);
|
| 56 | trace_seq_printf(s, "\ttime_stamp : type == %d\n",
|
| 57 | RINGBUF_TYPE_TIME_STAMP);
|
| 58 | trace_seq_printf(s, "\tdata max type_len == %d\n",
|
| 59 | RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
|
| 60 |
|
| 61 | return !trace_seq_has_overflowed(s);
|
| 62 | }
|
| 63 |
|
| 64 | |
| 65 | * The ring buffer is made up of a list of pages. A separate list of pages is
|
| 66 | * allocated for each CPU. A writer may only write to a buffer that is
|
| 67 | * associated with the CPU it is currently executing on. A reader may read
|
| 68 | * from any per cpu buffer.
|
| 69 | *
|
| 70 | * The reader is special. For each per cpu buffer, the reader has its own
|
| 71 | * reader page. When a reader has read the entire reader page, this reader
|
| 72 | * page is swapped with another page in the ring buffer.
|
| 73 | *
|
| 74 | * Now, as long as the writer is off the reader page, the reader can do what
|
| 75 | * ever it wants with that page. The writer will never write to that page
|
| 76 | * again (as long as it is out of the ring buffer).
|
| 77 | *
|
| 78 | * Here's some silly ASCII art.
|
| 79 | *
|
| 80 | * +------+
|
| 81 | * |reader| RING BUFFER
|
| 82 | * |page |
|
| 83 | * +------+ +---+ +---+ +---+
|
| 84 | * | |-->| |-->| |
|
| 85 | * +---+ +---+ +---+
|
| 86 | * ^ |
|
| 87 | * | |
|
| 88 | * +---------------+
|
| 89 | *
|
| 90 | *
|
| 91 | * +------+
|
| 92 | * |reader| RING BUFFER
|
| 93 | * |page |------------------v
|
| 94 | * +------+ +---+ +---+ +---+
|
| 95 | * | |-->| |-->| |
|
| 96 | * +---+ +---+ +---+
|
| 97 | * ^ |
|
| 98 | * | |
|
| 99 | * +---------------+
|
| 100 | *
|
| 101 | *
|
| 102 | * +------+
|
| 103 | * |reader| RING BUFFER
|
| 104 | * |page |------------------v
|
| 105 | * +------+ +---+ +---+ +---+
|
| 106 | * ^ | |-->| |-->| |
|
| 107 | * | +---+ +---+ +---+
|
| 108 | * | |
|
| 109 | * | |
|
| 110 | * +------------------------------+
|
| 111 | *
|
| 112 | *
|
| 113 | * +------+
|
| 114 | * |buffer| RING BUFFER
|
| 115 | * |page |------------------v
|
| 116 | * +------+ +---+ +---+ +---+
|
| 117 | * ^ | | | |-->| |
|
| 118 | * | New +---+ +---+ +---+
|
| 119 | * | Reader------^ |
|
| 120 | * | page |
|
| 121 | * +------------------------------+
|
| 122 | *
|
| 123 | *
|
| 124 | * After we make this swap, the reader can hand this page off to the splice
|
| 125 | * code and be done with it. It can even allocate a new page if it needs to
|
| 126 | * and swap that into the ring buffer.
|
| 127 | *
|
| 128 | * We will be using cmpxchg soon to make all this lockless.
|
| 129 | *
|
| 130 |
|
| 131 |
|
| 132 |
|
| 133 | #define RB_BUFFER_OFF (1 << 20)
|
| 134 |
|
| 135 | #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
|
| 136 |
|
| 137 | #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
|
| 138 | #define RB_ALIGNMENT 4U
|
| 139 | #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
|
| 140 | #define RB_EVNT_MIN_SIZE 8U
|
| 141 |
|
| 142 | #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
|
| 143 | # define RB_FORCE_8BYTE_ALIGNMENT 0
|
| 144 | # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
|
| 145 | #else
|
| 146 | # define RB_FORCE_8BYTE_ALIGNMENT 1
|
| 147 | # define RB_ARCH_ALIGNMENT 8U
|
| 148 | #endif
|
| 149 |
|
| 150 | #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
|
| 151 |
|
| 152 |
|
| 153 | #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
|
| 154 |
|
| 155 | enum {
|
| 156 | RB_LEN_TIME_EXTEND = 8,
|
| 157 | RB_LEN_TIME_STAMP = 8,
|
| 158 | };
|
| 159 |
|
| 160 | #define skip_time_extend(event) \
|
| 161 | ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
|
| 162 |
|
| 163 | #define extended_time(event) \
|
| 164 | (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
|
| 165 |
|
| 166 | static inline bool rb_null_event(struct ring_buffer_event *event)
|
| 167 | {
|
| 168 | return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
|
| 169 | }
|
| 170 |
|
| 171 | static void rb_event_set_padding(struct ring_buffer_event *event)
|
| 172 | {
|
| 173 |
|
| 174 | event->type_len = RINGBUF_TYPE_PADDING;
|
| 175 | event->time_delta = 0;
|
| 176 | }
|
| 177 |
|
| 178 | static unsigned
|
| 179 | rb_event_data_length(struct ring_buffer_event *event)
|
| 180 | {
|
| 181 | unsigned length;
|
| 182 |
|
| 183 | if (event->type_len)
|
| 184 | length = event->type_len * RB_ALIGNMENT;
|
| 185 | else
|
| 186 | length = event->array[0];
|
| 187 | return length + RB_EVNT_HDR_SIZE;
|
| 188 | }
|
| 189 |
|
| 190 | |
| 191 | * Return the length of the given event. Will return
|
| 192 | * the length of the time extend if the event is a
|
| 193 | * time extend.
|
| 194 |
|
| 195 | static inline unsigned
|
| 196 | rb_event_length(struct ring_buffer_event *event)
|
| 197 | {
|
| 198 | switch (event->type_len) {
|
| 199 | case RINGBUF_TYPE_PADDING:
|
| 200 | if (rb_null_event(event))
|
| 201 |
|
| 202 | return -1;
|
| 203 | return event->array[0] + RB_EVNT_HDR_SIZE;
|
| 204 |
|
| 205 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 206 | return RB_LEN_TIME_EXTEND;
|
| 207 |
|
| 208 | case RINGBUF_TYPE_TIME_STAMP:
|
| 209 | return RB_LEN_TIME_STAMP;
|
| 210 |
|
| 211 | case RINGBUF_TYPE_DATA:
|
| 212 | return rb_event_data_length(event);
|
| 213 | default:
|
| 214 | WARN_ON_ONCE(1);
|
| 215 | }
|
| 216 |
|
| 217 | return 0;
|
| 218 | }
|
| 219 |
|
| 220 | |
| 221 | * Return total length of time extend and data,
|
| 222 | * or just the event length for all other events.
|
| 223 |
|
| 224 | static inline unsigned
|
| 225 | rb_event_ts_length(struct ring_buffer_event *event)
|
| 226 | {
|
| 227 | unsigned len = 0;
|
| 228 |
|
| 229 | if (extended_time(event)) {
|
| 230 |
|
| 231 | len = RB_LEN_TIME_EXTEND;
|
| 232 | event = skip_time_extend(event);
|
| 233 | }
|
| 234 | return len + rb_event_length(event);
|
| 235 | }
|
| 236 |
|
| 237 | |
| 238 | * ring_buffer_event_length - return the length of the event
|
| 239 | * @event: the event to get the length of
|
| 240 | *
|
| 241 | * Returns the size of the data load of a data event.
|
| 242 | * If the event is something other than a data event, it
|
| 243 | * returns the size of the event itself. With the exception
|
| 244 | * of a TIME EXTEND, where it still returns the size of the
|
| 245 | * data load of the data event after it.
|
| 246 |
|
| 247 | unsigned ring_buffer_event_length(struct ring_buffer_event *event)
|
| 248 | {
|
| 249 | unsigned length;
|
| 250 |
|
| 251 | if (extended_time(event))
|
| 252 | event = skip_time_extend(event);
|
| 253 |
|
| 254 | length = rb_event_length(event);
|
| 255 | if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
|
| 256 | return length;
|
| 257 | length -= RB_EVNT_HDR_SIZE;
|
| 258 | if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
|
| 259 | length -= sizeof(event->array[0]);
|
| 260 | return length;
|
| 261 | }
|
| 262 | EXPORT_SYMBOL_GPL(ring_buffer_event_length);
|
| 263 |
|
| 264 |
|
| 265 | static __always_inline void *
|
| 266 | rb_event_data(struct ring_buffer_event *event)
|
| 267 | {
|
| 268 | if (extended_time(event))
|
| 269 | event = skip_time_extend(event);
|
| 270 | WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
|
| 271 |
|
| 272 | if (event->type_len)
|
| 273 | return (void *)&event->array[0];
|
| 274 |
|
| 275 | return (void *)&event->array[1];
|
| 276 | }
|
| 277 |
|
| 278 | |
| 279 | * ring_buffer_event_data - return the data of the event
|
| 280 | * @event: the event to get the data from
|
| 281 |
|
| 282 | void *ring_buffer_event_data(struct ring_buffer_event *event)
|
| 283 | {
|
| 284 | return rb_event_data(event);
|
| 285 | }
|
| 286 | EXPORT_SYMBOL_GPL(ring_buffer_event_data);
|
| 287 |
|
| 288 | #define for_each_buffer_cpu(buffer, cpu) \
|
| 289 | for_each_cpu(cpu, buffer->cpumask)
|
| 290 |
|
| 291 | #define for_each_online_buffer_cpu(buffer, cpu) \
|
| 292 | for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
|
| 293 |
|
| 294 | #define TS_SHIFT 27
|
| 295 | #define TS_MASK ((1ULL << TS_SHIFT) - 1)
|
| 296 | #define TS_DELTA_TEST (~TS_MASK)
|
| 297 |
|
| 298 | static u64 rb_event_time_stamp(struct ring_buffer_event *event)
|
| 299 | {
|
| 300 | u64 ts;
|
| 301 |
|
| 302 | ts = event->array[0];
|
| 303 | ts <<= TS_SHIFT;
|
| 304 | ts += event->time_delta;
|
| 305 |
|
| 306 | return ts;
|
| 307 | }
|
| 308 |
|
| 309 |
|
| 310 | #define RB_MISSED_EVENTS (1 << 31)
|
| 311 |
|
| 312 | #define RB_MISSED_STORED (1 << 30)
|
| 313 |
|
| 314 | struct buffer_data_page {
|
| 315 | u64 time_stamp;
|
| 316 | local_t commit;
|
| 317 | unsigned char data[] RB_ALIGN_DATA;
|
| 318 | };
|
| 319 |
|
| 320 | |
| 321 | * Note, the buffer_page list must be first. The buffer pages
|
| 322 | * are allocated in cache lines, which means that each buffer
|
| 323 | * page will be at the beginning of a cache line, and thus
|
| 324 | * the least significant bits will be zero. We use this to
|
| 325 | * add flags in the list struct pointers, to make the ring buffer
|
| 326 | * lockless.
|
| 327 |
|
| 328 | struct buffer_page {
|
| 329 | struct list_head list;
|
| 330 | local_t write;
|
| 331 | unsigned read;
|
| 332 | local_t entries;
|
| 333 | unsigned long real_end;
|
| 334 | struct buffer_data_page *page;
|
| 335 | };
|
| 336 |
|
| 337 | |
| 338 | * The buffer page counters, write and entries, must be reset
|
| 339 | * atomically when crossing page boundaries. To synchronize this
|
| 340 | * update, two counters are inserted into the number. One is
|
| 341 | * the actual counter for the write position or count on the page.
|
| 342 | *
|
| 343 | * The other is a counter of updaters. Before an update happens
|
| 344 | * the update partition of the counter is incremented. This will
|
| 345 | * allow the updater to update the counter atomically.
|
| 346 | *
|
| 347 | * The counter is 20 bits, and the state data is 12.
|
| 348 |
|
| 349 | #define RB_WRITE_MASK 0xfffff
|
| 350 | #define RB_WRITE_INTCNT (1 << 20)
|
| 351 |
|
| 352 | static void rb_init_page(struct buffer_data_page *bpage)
|
| 353 | {
|
| 354 | local_set(&bpage->commit, 0);
|
| 355 | }
|
| 356 |
|
| 357 | static void free_buffer_page(struct buffer_page *bpage)
|
| 358 | {
|
| 359 | free_page((unsigned long)bpage->page);
|
| 360 | kfree(bpage);
|
| 361 | }
|
| 362 |
|
| 363 | |
| 364 | * We need to fit the time_stamp delta into 27 bits.
|
| 365 |
|
| 366 | static inline bool test_time_stamp(u64 delta)
|
| 367 | {
|
| 368 | return !!(delta & TS_DELTA_TEST);
|
| 369 | }
|
| 370 |
|
| 371 | #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
|
| 372 |
|
| 373 |
|
| 374 | #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
|
| 375 |
|
| 376 | int ring_buffer_print_page_header(struct trace_seq *s)
|
| 377 | {
|
| 378 | struct buffer_data_page field;
|
| 379 |
|
| 380 | trace_seq_printf(s, "\tfield: u64 timestamp;\t"
|
| 381 | "offset:0;\tsize:%u;\tsigned:%u;\n",
|
| 382 | (unsigned int)sizeof(field.time_stamp),
|
| 383 | (unsigned int)is_signed_type(u64));
|
| 384 |
|
| 385 | trace_seq_printf(s, "\tfield: local_t commit;\t"
|
| 386 | "offset:%u;\tsize:%u;\tsigned:%u;\n",
|
| 387 | (unsigned int)offsetof(typeof(field), commit),
|
| 388 | (unsigned int)sizeof(field.commit),
|
| 389 | (unsigned int)is_signed_type(long));
|
| 390 |
|
| 391 | trace_seq_printf(s, "\tfield: int overwrite;\t"
|
| 392 | "offset:%u;\tsize:%u;\tsigned:%u;\n",
|
| 393 | (unsigned int)offsetof(typeof(field), commit),
|
| 394 | 1,
|
| 395 | (unsigned int)is_signed_type(long));
|
| 396 |
|
| 397 | trace_seq_printf(s, "\tfield: char data;\t"
|
| 398 | "offset:%u;\tsize:%u;\tsigned:%u;\n",
|
| 399 | (unsigned int)offsetof(typeof(field), data),
|
| 400 | (unsigned int)BUF_PAGE_SIZE,
|
| 401 | (unsigned int)is_signed_type(char));
|
| 402 |
|
| 403 | return !trace_seq_has_overflowed(s);
|
| 404 | }
|
| 405 |
|
| 406 | struct rb_irq_work {
|
| 407 | struct irq_work work;
|
| 408 | wait_queue_head_t waiters;
|
| 409 | wait_queue_head_t full_waiters;
|
| 410 | long wait_index;
|
| 411 | bool waiters_pending;
|
| 412 | bool full_waiters_pending;
|
| 413 | bool wakeup_full;
|
| 414 | };
|
| 415 |
|
| 416 | |
| 417 | * Structure to hold event state and handle nested events.
|
| 418 |
|
| 419 | struct rb_event_info {
|
| 420 | u64 ts;
|
| 421 | u64 delta;
|
| 422 | u64 before;
|
| 423 | u64 after;
|
| 424 | unsigned long length;
|
| 425 | struct buffer_page *tail_page;
|
| 426 | int add_timestamp;
|
| 427 | };
|
| 428 |
|
| 429 | |
| 430 | * Used for the add_timestamp
|
| 431 | * NONE
|
| 432 | * EXTEND - wants a time extend
|
| 433 | * ABSOLUTE - the buffer requests all events to have absolute time stamps
|
| 434 | * FORCE - force a full time stamp.
|
| 435 |
|
| 436 | enum {
|
| 437 | RB_ADD_STAMP_NONE = 0,
|
| 438 | RB_ADD_STAMP_EXTEND = BIT(1),
|
| 439 | RB_ADD_STAMP_ABSOLUTE = BIT(2),
|
| 440 | RB_ADD_STAMP_FORCE = BIT(3)
|
| 441 | };
|
| 442 | |
| 443 | * Used for which event context the event is in.
|
| 444 | * TRANSITION = 0
|
| 445 | * NMI = 1
|
| 446 | * IRQ = 2
|
| 447 | * SOFTIRQ = 3
|
| 448 | * NORMAL = 4
|
| 449 | *
|
| 450 | * See trace_recursive_lock() comment below for more details.
|
| 451 |
|
| 452 | enum {
|
| 453 | RB_CTX_TRANSITION,
|
| 454 | RB_CTX_NMI,
|
| 455 | RB_CTX_IRQ,
|
| 456 | RB_CTX_SOFTIRQ,
|
| 457 | RB_CTX_NORMAL,
|
| 458 | RB_CTX_MAX
|
| 459 | };
|
| 460 |
|
| 461 | #if BITS_PER_LONG == 32
|
| 462 | #define RB_TIME_32
|
| 463 | #endif
|
| 464 |
|
| 465 |
|
| 466 | |
| 467 |
|
| 468 | #ifdef RB_TIME_32
|
| 469 |
|
| 470 | struct rb_time_struct {
|
| 471 | local_t cnt;
|
| 472 | local_t top;
|
| 473 | local_t bottom;
|
| 474 | local_t msb;
|
| 475 | };
|
| 476 | #else
|
| 477 | #include <asm/local64.h>
|
| 478 | struct rb_time_struct {
|
| 479 | local64_t time;
|
| 480 | };
|
| 481 | #endif
|
| 482 | typedef struct rb_time_struct rb_time_t;
|
| 483 |
|
| 484 | #define MAX_NEST 5
|
| 485 |
|
| 486 | |
| 487 | * head_page == tail_page && head == tail then buffer is empty.
|
| 488 |
|
| 489 | struct ring_buffer_per_cpu {
|
| 490 | int cpu;
|
| 491 | atomic_t record_disabled;
|
| 492 | atomic_t resize_disabled;
|
| 493 | struct trace_buffer *buffer;
|
| 494 | raw_spinlock_t reader_lock;
|
| 495 | arch_spinlock_t lock;
|
| 496 | struct lock_class_key lock_key;
|
| 497 | struct buffer_data_page *free_page;
|
| 498 | unsigned long nr_pages;
|
| 499 | unsigned int current_context;
|
| 500 | struct list_head *pages;
|
| 501 | struct buffer_page *head_page;
|
| 502 | struct buffer_page *tail_page;
|
| 503 | struct buffer_page *commit_page;
|
| 504 | struct buffer_page *reader_page;
|
| 505 | unsigned long lost_events;
|
| 506 | unsigned long last_overrun;
|
| 507 | unsigned long nest;
|
| 508 | local_t entries_bytes;
|
| 509 | local_t entries;
|
| 510 | local_t overrun;
|
| 511 | local_t commit_overrun;
|
| 512 | local_t dropped_events;
|
| 513 | local_t committing;
|
| 514 | local_t commits;
|
| 515 | local_t pages_touched;
|
| 516 | local_t pages_lost;
|
| 517 | local_t pages_read;
|
| 518 | long last_pages_touch;
|
| 519 | size_t shortest_full;
|
| 520 | unsigned long read;
|
| 521 | unsigned long read_bytes;
|
| 522 | rb_time_t write_stamp;
|
| 523 | rb_time_t before_stamp;
|
| 524 | u64 event_stamp[MAX_NEST];
|
| 525 | u64 read_stamp;
|
| 526 |
|
| 527 | long nr_pages_to_update;
|
| 528 | struct list_head new_pages;
|
| 529 | struct work_struct update_pages_work;
|
| 530 | struct completion update_done;
|
| 531 |
|
| 532 | struct rb_irq_work irq_work;
|
| 533 | };
|
| 534 |
|
| 535 | struct trace_buffer {
|
| 536 | unsigned flags;
|
| 537 | int cpus;
|
| 538 | atomic_t record_disabled;
|
| 539 | atomic_t resizing;
|
| 540 | cpumask_var_t cpumask;
|
| 541 |
|
| 542 | struct lock_class_key *reader_lock_key;
|
| 543 |
|
| 544 | struct mutex mutex;
|
| 545 |
|
| 546 | struct ring_buffer_per_cpu **buffers;
|
| 547 |
|
| 548 | struct hlist_node node;
|
| 549 | u64 (*clock)(void);
|
| 550 |
|
| 551 | struct rb_irq_work irq_work;
|
| 552 | bool time_stamp_abs;
|
| 553 | };
|
| 554 |
|
| 555 | struct ring_buffer_iter {
|
| 556 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 557 | unsigned long head;
|
| 558 | unsigned long next_event;
|
| 559 | struct buffer_page *head_page;
|
| 560 | struct buffer_page *cache_reader_page;
|
| 561 | unsigned long cache_read;
|
| 562 | u64 read_stamp;
|
| 563 | u64 page_stamp;
|
| 564 | struct ring_buffer_event *event;
|
| 565 | int missed_events;
|
| 566 | };
|
| 567 |
|
| 568 | #ifdef RB_TIME_32
|
| 569 |
|
| 570 | |
| 571 | * On 32 bit machines, local64_t is very expensive. As the ring
|
| 572 | * buffer doesn't need all the features of a true 64 bit atomic,
|
| 573 | * on 32 bit, it uses these functions (64 still uses local64_t).
|
| 574 | *
|
| 575 | * For the ring buffer, 64 bit required operations for the time is
|
| 576 | * the following:
|
| 577 | *
|
| 578 | * - Reads may fail if it interrupted a modification of the time stamp.
|
| 579 | * It will succeed if it did not interrupt another write even if
|
| 580 | * the read itself is interrupted by a write.
|
| 581 | * It returns whether it was successful or not.
|
| 582 | *
|
| 583 | * - Writes always succeed and will overwrite other writes and writes
|
| 584 | * that were done by events interrupting the current write.
|
| 585 | *
|
| 586 | * - A write followed by a read of the same time stamp will always succeed,
|
| 587 | * but may not contain the same value.
|
| 588 | *
|
| 589 | * - A cmpxchg will fail if it interrupted another write or cmpxchg.
|
| 590 | * Other than that, it acts like a normal cmpxchg.
|
| 591 | *
|
| 592 | * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
|
| 593 | * (bottom being the least significant 30 bits of the 60 bit time stamp).
|
| 594 | *
|
| 595 | * The two most significant bits of each half holds a 2 bit counter (0-3).
|
| 596 | * Each update will increment this counter by one.
|
| 597 | * When reading the top and bottom, if the two counter bits match then the
|
| 598 | * top and bottom together make a valid 60 bit number.
|
| 599 |
|
| 600 | #define RB_TIME_SHIFT 30
|
| 601 | #define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
|
| 602 | #define RB_TIME_MSB_SHIFT 60
|
| 603 |
|
| 604 | static inline int rb_time_cnt(unsigned long val)
|
| 605 | {
|
| 606 | return (val >> RB_TIME_SHIFT) & 3;
|
| 607 | }
|
| 608 |
|
| 609 | static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
|
| 610 | {
|
| 611 | u64 val;
|
| 612 |
|
| 613 | val = top & RB_TIME_VAL_MASK;
|
| 614 | val <<= RB_TIME_SHIFT;
|
| 615 | val |= bottom & RB_TIME_VAL_MASK;
|
| 616 |
|
| 617 | return val;
|
| 618 | }
|
| 619 |
|
| 620 | static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
|
| 621 | {
|
| 622 | unsigned long top, bottom, msb;
|
| 623 | unsigned long c;
|
| 624 |
|
| 625 | |
| 626 | * If the read is interrupted by a write, then the cnt will
|
| 627 | * be different. Loop until both top and bottom have been read
|
| 628 | * without interruption.
|
| 629 |
|
| 630 | do {
|
| 631 | c = local_read(&t->cnt);
|
| 632 | top = local_read(&t->top);
|
| 633 | bottom = local_read(&t->bottom);
|
| 634 | msb = local_read(&t->msb);
|
| 635 | } while (c != local_read(&t->cnt));
|
| 636 |
|
| 637 | *cnt = rb_time_cnt(top);
|
| 638 |
|
| 639 |
|
| 640 | if (*cnt != rb_time_cnt(bottom))
|
| 641 | return false;
|
| 642 |
|
| 643 |
|
| 644 | *ret = rb_time_val(top, bottom) | ((u64)msb << RB_TIME_MSB_SHIFT);
|
| 645 | return true;
|
| 646 | }
|
| 647 |
|
| 648 | static bool rb_time_read(rb_time_t *t, u64 *ret)
|
| 649 | {
|
| 650 | unsigned long cnt;
|
| 651 |
|
| 652 | return __rb_time_read(t, ret, &cnt);
|
| 653 | }
|
| 654 |
|
| 655 | static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
|
| 656 | {
|
| 657 | return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
|
| 658 | }
|
| 659 |
|
| 660 | static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom,
|
| 661 | unsigned long *msb)
|
| 662 | {
|
| 663 | *top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
|
| 664 | *bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
|
| 665 | *msb = (unsigned long)(val >> RB_TIME_MSB_SHIFT);
|
| 666 | }
|
| 667 |
|
| 668 | static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
|
| 669 | {
|
| 670 | val = rb_time_val_cnt(val, cnt);
|
| 671 | local_set(t, val);
|
| 672 | }
|
| 673 |
|
| 674 | static void rb_time_set(rb_time_t *t, u64 val)
|
| 675 | {
|
| 676 | unsigned long cnt, top, bottom, msb;
|
| 677 |
|
| 678 | rb_time_split(val, &top, &bottom, &msb);
|
| 679 |
|
| 680 |
|
| 681 | do {
|
| 682 | cnt = local_inc_return(&t->cnt);
|
| 683 | rb_time_val_set(&t->top, top, cnt);
|
| 684 | rb_time_val_set(&t->bottom, bottom, cnt);
|
| 685 | rb_time_val_set(&t->msb, val >> RB_TIME_MSB_SHIFT, cnt);
|
| 686 | } while (cnt != local_read(&t->cnt));
|
| 687 | }
|
| 688 |
|
| 689 | static inline bool
|
| 690 | rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
|
| 691 | {
|
| 692 | unsigned long ret;
|
| 693 |
|
| 694 | ret = local_cmpxchg(l, expect, set);
|
| 695 | return ret == expect;
|
| 696 | }
|
| 697 |
|
| 698 | static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
|
| 699 | {
|
| 700 | unsigned long cnt, top, bottom, msb;
|
| 701 | unsigned long cnt2, top2, bottom2, msb2;
|
| 702 | u64 val;
|
| 703 |
|
| 704 |
|
| 705 | if (!__rb_time_read(t, &val, &cnt2))
|
| 706 | return false;
|
| 707 |
|
| 708 | if (val != expect)
|
| 709 | return false;
|
| 710 |
|
| 711 | cnt = local_read(&t->cnt);
|
| 712 | if ((cnt & 3) != cnt2)
|
| 713 | return false;
|
| 714 |
|
| 715 | cnt2 = cnt + 1;
|
| 716 |
|
| 717 | rb_time_split(val, &top, &bottom, &msb);
|
| 718 | top = rb_time_val_cnt(top, cnt);
|
| 719 | bottom = rb_time_val_cnt(bottom, cnt);
|
| 720 |
|
| 721 | rb_time_split(set, &top2, &bottom2, &msb2);
|
| 722 | top2 = rb_time_val_cnt(top2, cnt2);
|
| 723 | bottom2 = rb_time_val_cnt(bottom2, cnt2);
|
| 724 |
|
| 725 | if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
|
| 726 | return false;
|
| 727 | if (!rb_time_read_cmpxchg(&t->msb, msb, msb2))
|
| 728 | return false;
|
| 729 | if (!rb_time_read_cmpxchg(&t->top, top, top2))
|
| 730 | return false;
|
| 731 | if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
|
| 732 | return false;
|
| 733 | return true;
|
| 734 | }
|
| 735 |
|
| 736 | #else
|
| 737 |
|
| 738 |
|
| 739 |
|
| 740 | static inline bool rb_time_read(rb_time_t *t, u64 *ret)
|
| 741 | {
|
| 742 | *ret = local64_read(&t->time);
|
| 743 | return true;
|
| 744 | }
|
| 745 | static void rb_time_set(rb_time_t *t, u64 val)
|
| 746 | {
|
| 747 | local64_set(&t->time, val);
|
| 748 | }
|
| 749 |
|
| 750 | static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
|
| 751 | {
|
| 752 | u64 val;
|
| 753 | val = local64_cmpxchg(&t->time, expect, set);
|
| 754 | return val == expect;
|
| 755 | }
|
| 756 | #endif
|
| 757 |
|
| 758 | |
| 759 | * Enable this to make sure that the event passed to
|
| 760 | * ring_buffer_event_time_stamp() is not committed and also
|
| 761 | * is on the buffer that it passed in.
|
| 762 |
|
| 763 | |
| 764 | #ifdef RB_VERIFY_EVENT
|
| 765 | static struct list_head *rb_list_head(struct list_head *list);
|
| 766 | static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
|
| 767 | void *event)
|
| 768 | {
|
| 769 | struct buffer_page *page = cpu_buffer->commit_page;
|
| 770 | struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
|
| 771 | struct list_head *next;
|
| 772 | long commit, write;
|
| 773 | unsigned long addr = (unsigned long)event;
|
| 774 | bool done = false;
|
| 775 | int stop = 0;
|
| 776 |
|
| 777 |
|
| 778 | do {
|
| 779 | if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
|
| 780 | done = true;
|
| 781 | commit = local_read(&page->page->commit);
|
| 782 | write = local_read(&page->write);
|
| 783 | if (addr >= (unsigned long)&page->page->data[commit] &&
|
| 784 | addr < (unsigned long)&page->page->data[write])
|
| 785 | return;
|
| 786 |
|
| 787 | next = rb_list_head(page->list.next);
|
| 788 | page = list_entry(next, struct buffer_page, list);
|
| 789 | } while (!done);
|
| 790 | WARN_ON_ONCE(1);
|
| 791 | }
|
| 792 | #else
|
| 793 | static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
|
| 794 | void *event)
|
| 795 | {
|
| 796 | }
|
| 797 | #endif
|
| 798 |
|
| 799 | |
| 800 | * The absolute time stamp drops the 5 MSBs and some clocks may
|
| 801 | * require them. The rb_fix_abs_ts() will take a previous full
|
| 802 | * time stamp, and add the 5 MSB of that time stamp on to the
|
| 803 | * saved absolute time stamp. Then they are compared in case of
|
| 804 | * the unlikely event that the latest time stamp incremented
|
| 805 | * the 5 MSB.
|
| 806 |
|
| 807 | static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
|
| 808 | {
|
| 809 | if (save_ts & TS_MSB) {
|
| 810 | abs |= save_ts & TS_MSB;
|
| 811 |
|
| 812 | if (unlikely(abs < save_ts))
|
| 813 | abs += 1ULL << 59;
|
| 814 | }
|
| 815 | return abs;
|
| 816 | }
|
| 817 |
|
| 818 | static inline u64 rb_time_stamp(struct trace_buffer *buffer);
|
| 819 |
|
| 820 | |
| 821 | * ring_buffer_event_time_stamp - return the event's current time stamp
|
| 822 | * @buffer: The buffer that the event is on
|
| 823 | * @event: the event to get the time stamp of
|
| 824 | *
|
| 825 | * Note, this must be called after @event is reserved, and before it is
|
| 826 | * committed to the ring buffer. And must be called from the same
|
| 827 | * context where the event was reserved (normal, softirq, irq, etc).
|
| 828 | *
|
| 829 | * Returns the time stamp associated with the current event.
|
| 830 | * If the event has an extended time stamp, then that is used as
|
| 831 | * the time stamp to return.
|
| 832 | * In the highly unlikely case that the event was nested more than
|
| 833 | * the max nesting, then the write_stamp of the buffer is returned,
|
| 834 | * otherwise current time is returned, but that really neither of
|
| 835 | * the last two cases should ever happen.
|
| 836 |
|
| 837 | u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
|
| 838 | struct ring_buffer_event *event)
|
| 839 | {
|
| 840 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
|
| 841 | unsigned int nest;
|
| 842 | u64 ts;
|
| 843 |
|
| 844 |
|
| 845 | if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
|
| 846 | ts = rb_event_time_stamp(event);
|
| 847 | return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
|
| 848 | }
|
| 849 |
|
| 850 | nest = local_read(&cpu_buffer->committing);
|
| 851 | verify_event(cpu_buffer, event);
|
| 852 | if (WARN_ON_ONCE(!nest))
|
| 853 | goto fail;
|
| 854 |
|
| 855 |
|
| 856 | if (likely(--nest < MAX_NEST))
|
| 857 | return cpu_buffer->event_stamp[nest];
|
| 858 |
|
| 859 |
|
| 860 | WARN_ONCE(1, "nest (%d) greater than max", nest);
|
| 861 |
|
| 862 | fail:
|
| 863 |
|
| 864 | if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
|
| 865 |
|
| 866 | ts = rb_time_stamp(cpu_buffer->buffer);
|
| 867 |
|
| 868 | return ts;
|
| 869 | }
|
| 870 |
|
| 871 | |
| 872 | * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
|
| 873 | * @buffer: The ring_buffer to get the number of pages from
|
| 874 | * @cpu: The cpu of the ring_buffer to get the number of pages from
|
| 875 | *
|
| 876 | * Returns the number of pages used by a per_cpu buffer of the ring buffer.
|
| 877 |
|
| 878 | size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
|
| 879 | {
|
| 880 | return buffer->buffers[cpu]->nr_pages;
|
| 881 | }
|
| 882 |
|
| 883 | |
| 884 | * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
|
| 885 | * @buffer: The ring_buffer to get the number of pages from
|
| 886 | * @cpu: The cpu of the ring_buffer to get the number of pages from
|
| 887 | *
|
| 888 | * Returns the number of pages that have content in the ring buffer.
|
| 889 |
|
| 890 | size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
|
| 891 | {
|
| 892 | size_t read;
|
| 893 | size_t lost;
|
| 894 | size_t cnt;
|
| 895 |
|
| 896 | read = local_read(&buffer->buffers[cpu]->pages_read);
|
| 897 | lost = local_read(&buffer->buffers[cpu]->pages_lost);
|
| 898 | cnt = local_read(&buffer->buffers[cpu]->pages_touched);
|
| 899 |
|
| 900 | if (WARN_ON_ONCE(cnt < lost))
|
| 901 | return 0;
|
| 902 |
|
| 903 | cnt -= lost;
|
| 904 |
|
| 905 |
|
| 906 | if (cnt < read) {
|
| 907 | WARN_ON_ONCE(read > cnt + 1);
|
| 908 | return 0;
|
| 909 | }
|
| 910 |
|
| 911 | return cnt - read;
|
| 912 | }
|
| 913 |
|
| 914 | static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
|
| 915 | {
|
| 916 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
|
| 917 | size_t nr_pages;
|
| 918 | size_t dirty;
|
| 919 |
|
| 920 | nr_pages = cpu_buffer->nr_pages;
|
| 921 | if (!nr_pages || !full)
|
| 922 | return true;
|
| 923 |
|
| 924 | dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
|
| 925 |
|
| 926 | return (dirty * 100) > (full * nr_pages);
|
| 927 | }
|
| 928 |
|
| 929 | |
| 930 | * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
|
| 931 | *
|
| 932 | * Schedules a delayed work to wake up any task that is blocked on the
|
| 933 | * ring buffer waiters queue.
|
| 934 |
|
| 935 | static void rb_wake_up_waiters(struct irq_work *work)
|
| 936 | {
|
| 937 | struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
|
| 938 |
|
| 939 | wake_up_all(&rbwork->waiters);
|
| 940 | if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
|
| 941 | rbwork->wakeup_full = false;
|
| 942 | rbwork->full_waiters_pending = false;
|
| 943 | wake_up_all(&rbwork->full_waiters);
|
| 944 | }
|
| 945 | }
|
| 946 |
|
| 947 | |
| 948 | * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
|
| 949 | * @buffer: The ring buffer to wake waiters on
|
| 950 | *
|
| 951 | * In the case of a file that represents a ring buffer is closing,
|
| 952 | * it is prudent to wake up any waiters that are on this.
|
| 953 |
|
| 954 | void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
|
| 955 | {
|
| 956 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 957 | struct rb_irq_work *rbwork;
|
| 958 |
|
| 959 | if (!buffer)
|
| 960 | return;
|
| 961 |
|
| 962 | if (cpu == RING_BUFFER_ALL_CPUS) {
|
| 963 |
|
| 964 |
|
| 965 | for_each_buffer_cpu(buffer, cpu)
|
| 966 | ring_buffer_wake_waiters(buffer, cpu);
|
| 967 |
|
| 968 | rbwork = &buffer->irq_work;
|
| 969 | } else {
|
| 970 | if (WARN_ON_ONCE(!buffer->buffers))
|
| 971 | return;
|
| 972 | if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
|
| 973 | return;
|
| 974 |
|
| 975 | cpu_buffer = buffer->buffers[cpu];
|
| 976 |
|
| 977 | if (!cpu_buffer)
|
| 978 | return;
|
| 979 | rbwork = &cpu_buffer->irq_work;
|
| 980 | }
|
| 981 |
|
| 982 | rbwork->wait_index++;
|
| 983 |
|
| 984 | smp_wmb();
|
| 985 |
|
| 986 | rb_wake_up_waiters(&rbwork->work);
|
| 987 | }
|
| 988 |
|
| 989 | |
| 990 | * ring_buffer_wait - wait for input to the ring buffer
|
| 991 | * @buffer: buffer to wait on
|
| 992 | * @cpu: the cpu buffer to wait on
|
| 993 | * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
|
| 994 | *
|
| 995 | * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
|
| 996 | * as data is added to any of the @buffer's cpu buffers. Otherwise
|
| 997 | * it will wait for data to be added to a specific cpu buffer.
|
| 998 |
|
| 999 | int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
|
| 1000 | {
|
| 1001 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 1002 | DEFINE_WAIT(wait);
|
| 1003 | struct rb_irq_work *work;
|
| 1004 | long wait_index;
|
| 1005 | int ret = 0;
|
| 1006 |
|
| 1007 | |
| 1008 | * Depending on what the caller is waiting for, either any
|
| 1009 | * data in any cpu buffer, or a specific buffer, put the
|
| 1010 | * caller on the appropriate wait queue.
|
| 1011 |
|
| 1012 | if (cpu == RING_BUFFER_ALL_CPUS) {
|
| 1013 | work = &buffer->irq_work;
|
| 1014 |
|
| 1015 | full = 0;
|
| 1016 | } else {
|
| 1017 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 1018 | return -ENODEV;
|
| 1019 | cpu_buffer = buffer->buffers[cpu];
|
| 1020 | work = &cpu_buffer->irq_work;
|
| 1021 | }
|
| 1022 |
|
| 1023 | wait_index = READ_ONCE(work->wait_index);
|
| 1024 |
|
| 1025 | while (true) {
|
| 1026 | if (full)
|
| 1027 | prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
|
| 1028 | else
|
| 1029 | prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
|
| 1030 |
|
| 1031 | |
| 1032 | * The events can happen in critical sections where
|
| 1033 | * checking a work queue can cause deadlocks.
|
| 1034 | * After adding a task to the queue, this flag is set
|
| 1035 | * only to notify events to try to wake up the queue
|
| 1036 | * using irq_work.
|
| 1037 | *
|
| 1038 | * We don't clear it even if the buffer is no longer
|
| 1039 | * empty. The flag only causes the next event to run
|
| 1040 | * irq_work to do the work queue wake up. The worse
|
| 1041 | * that can happen if we race with !trace_empty() is that
|
| 1042 | * an event will cause an irq_work to try to wake up
|
| 1043 | * an empty queue.
|
| 1044 | *
|
| 1045 | * There's no reason to protect this flag either, as
|
| 1046 | * the work queue and irq_work logic will do the necessary
|
| 1047 | * synchronization for the wake ups. The only thing
|
| 1048 | * that is necessary is that the wake up happens after
|
| 1049 | * a task has been queued. It's OK for spurious wake ups.
|
| 1050 |
|
| 1051 | if (full)
|
| 1052 | work->full_waiters_pending = true;
|
| 1053 | else
|
| 1054 | work->waiters_pending = true;
|
| 1055 |
|
| 1056 | if (signal_pending(current)) {
|
| 1057 | ret = -EINTR;
|
| 1058 | break;
|
| 1059 | }
|
| 1060 |
|
| 1061 | if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
|
| 1062 | break;
|
| 1063 |
|
| 1064 | if (cpu != RING_BUFFER_ALL_CPUS &&
|
| 1065 | !ring_buffer_empty_cpu(buffer, cpu)) {
|
| 1066 | unsigned long flags;
|
| 1067 | bool pagebusy;
|
| 1068 | bool done;
|
| 1069 |
|
| 1070 | if (!full)
|
| 1071 | break;
|
| 1072 |
|
| 1073 | raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 1074 | pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
|
| 1075 | done = !pagebusy && full_hit(buffer, cpu, full);
|
| 1076 |
|
| 1077 | if (!cpu_buffer->shortest_full ||
|
| 1078 | cpu_buffer->shortest_full > full)
|
| 1079 | cpu_buffer->shortest_full = full;
|
| 1080 | raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 1081 | if (done)
|
| 1082 | break;
|
| 1083 | }
|
| 1084 |
|
| 1085 | schedule();
|
| 1086 |
|
| 1087 |
|
| 1088 | smp_rmb();
|
| 1089 | if (wait_index != work->wait_index)
|
| 1090 | break;
|
| 1091 | }
|
| 1092 |
|
| 1093 | if (full)
|
| 1094 | finish_wait(&work->full_waiters, &wait);
|
| 1095 | else
|
| 1096 | finish_wait(&work->waiters, &wait);
|
| 1097 |
|
| 1098 | return ret;
|
| 1099 | }
|
| 1100 |
|
| 1101 | |
| 1102 | * ring_buffer_poll_wait - poll on buffer input
|
| 1103 | * @buffer: buffer to wait on
|
| 1104 | * @cpu: the cpu buffer to wait on
|
| 1105 | * @filp: the file descriptor
|
| 1106 | * @poll_table: The poll descriptor
|
| 1107 | * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
|
| 1108 | *
|
| 1109 | * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
|
| 1110 | * as data is added to any of the @buffer's cpu buffers. Otherwise
|
| 1111 | * it will wait for data to be added to a specific cpu buffer.
|
| 1112 | *
|
| 1113 | * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
|
| 1114 | * zero otherwise.
|
| 1115 |
|
| 1116 | __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
|
| 1117 | struct file *filp, poll_table *poll_table, int full)
|
| 1118 | {
|
| 1119 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 1120 | struct rb_irq_work *work;
|
| 1121 |
|
| 1122 | if (cpu == RING_BUFFER_ALL_CPUS) {
|
| 1123 | work = &buffer->irq_work;
|
| 1124 | full = 0;
|
| 1125 | } else {
|
| 1126 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 1127 | return -EINVAL;
|
| 1128 |
|
| 1129 | cpu_buffer = buffer->buffers[cpu];
|
| 1130 | work = &cpu_buffer->irq_work;
|
| 1131 | }
|
| 1132 |
|
| 1133 | if (full) {
|
| 1134 | poll_wait(filp, &work->full_waiters, poll_table);
|
| 1135 | work->full_waiters_pending = true;
|
| 1136 | } else {
|
| 1137 | poll_wait(filp, &work->waiters, poll_table);
|
| 1138 | work->waiters_pending = true;
|
| 1139 | }
|
| 1140 |
|
| 1141 | |
| 1142 | * There's a tight race between setting the waiters_pending and
|
| 1143 | * checking if the ring buffer is empty. Once the waiters_pending bit
|
| 1144 | * is set, the next event will wake the task up, but we can get stuck
|
| 1145 | * if there's only a single event in.
|
| 1146 | *
|
| 1147 | * FIXME: Ideally, we need a memory barrier on the writer side as well,
|
| 1148 | * but adding a memory barrier to all events will cause too much of a
|
| 1149 | * performance hit in the fast path. We only need a memory barrier when
|
| 1150 | * the buffer goes from empty to having content. But as this race is
|
| 1151 | * extremely small, and it's not a problem if another event comes in, we
|
| 1152 | * will fix it later.
|
| 1153 |
|
| 1154 | smp_mb();
|
| 1155 |
|
| 1156 | if (full)
|
| 1157 | return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
|
| 1158 |
|
| 1159 | if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
|
| 1160 | (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
|
| 1161 | return EPOLLIN | EPOLLRDNORM;
|
| 1162 | return 0;
|
| 1163 | }
|
| 1164 |
|
| 1165 |
|
| 1166 | #define RB_WARN_ON(b, cond) \
|
| 1167 | ({ \
|
| 1168 | int _____ret = unlikely(cond); \
|
| 1169 | if (_____ret) { \
|
| 1170 | if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
|
| 1171 | struct ring_buffer_per_cpu *__b = \
|
| 1172 | (void *)b; \
|
| 1173 | atomic_inc(&__b->buffer->record_disabled); \
|
| 1174 | } else \
|
| 1175 | atomic_inc(&b->record_disabled); \
|
| 1176 | WARN_ON(1); \
|
| 1177 | } \
|
| 1178 | _____ret; \
|
| 1179 | })
|
| 1180 |
|
| 1181 |
|
| 1182 | #define DEBUG_SHIFT 0
|
| 1183 |
|
| 1184 | static inline u64 rb_time_stamp(struct trace_buffer *buffer)
|
| 1185 | {
|
| 1186 | u64 ts;
|
| 1187 |
|
| 1188 |
|
| 1189 | if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
|
| 1190 | ts = trace_clock_local();
|
| 1191 | else
|
| 1192 | ts = buffer->clock();
|
| 1193 |
|
| 1194 |
|
| 1195 | return ts << DEBUG_SHIFT;
|
| 1196 | }
|
| 1197 |
|
| 1198 | u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
|
| 1199 | {
|
| 1200 | u64 time;
|
| 1201 |
|
| 1202 | preempt_disable_notrace();
|
| 1203 | time = rb_time_stamp(buffer);
|
| 1204 | preempt_enable_notrace();
|
| 1205 |
|
| 1206 | return time;
|
| 1207 | }
|
| 1208 | EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
|
| 1209 |
|
| 1210 | void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
|
| 1211 | int cpu, u64 *ts)
|
| 1212 | {
|
| 1213 |
|
| 1214 | *ts >>= DEBUG_SHIFT;
|
| 1215 | }
|
| 1216 | EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
|
| 1217 |
|
| 1218 | |
| 1219 | * Making the ring buffer lockless makes things tricky.
|
| 1220 | * Although writes only happen on the CPU that they are on,
|
| 1221 | * and they only need to worry about interrupts. Reads can
|
| 1222 | * happen on any CPU.
|
| 1223 | *
|
| 1224 | * The reader page is always off the ring buffer, but when the
|
| 1225 | * reader finishes with a page, it needs to swap its page with
|
| 1226 | * a new one from the buffer. The reader needs to take from
|
| 1227 | * the head (writes go to the tail). But if a writer is in overwrite
|
| 1228 | * mode and wraps, it must push the head page forward.
|
| 1229 | *
|
| 1230 | * Here lies the problem.
|
| 1231 | *
|
| 1232 | * The reader must be careful to replace only the head page, and
|
| 1233 | * not another one. As described at the top of the file in the
|
| 1234 | * ASCII art, the reader sets its old page to point to the next
|
| 1235 | * page after head. It then sets the page after head to point to
|
| 1236 | * the old reader page. But if the writer moves the head page
|
| 1237 | * during this operation, the reader could end up with the tail.
|
| 1238 | *
|
| 1239 | * We use cmpxchg to help prevent this race. We also do something
|
| 1240 | * special with the page before head. We set the LSB to 1.
|
| 1241 | *
|
| 1242 | * When the writer must push the page forward, it will clear the
|
| 1243 | * bit that points to the head page, move the head, and then set
|
| 1244 | * the bit that points to the new head page.
|
| 1245 | *
|
| 1246 | * We also don't want an interrupt coming in and moving the head
|
| 1247 | * page on another writer. Thus we use the second LSB to catch
|
| 1248 | * that too. Thus:
|
| 1249 | *
|
| 1250 | * head->list->prev->next bit 1 bit 0
|
| 1251 | * ------- -------
|
| 1252 | * Normal page 0 0
|
| 1253 | * Points to head page 0 1
|
| 1254 | * New head page 1 0
|
| 1255 | *
|
| 1256 | * Note we can not trust the prev pointer of the head page, because:
|
| 1257 | *
|
| 1258 | * +----+ +-----+ +-----+
|
| 1259 | * | |------>| T |---X--->| N |
|
| 1260 | * | |<------| | | |
|
| 1261 | * +----+ +-----+ +-----+
|
| 1262 | * ^ ^ |
|
| 1263 | * | +-----+ | |
|
| 1264 | * +----------| R |----------+ |
|
| 1265 | * | |<-----------+
|
| 1266 | * +-----+
|
| 1267 | *
|
| 1268 | * Key: ---X--> HEAD flag set in pointer
|
| 1269 | * T Tail page
|
| 1270 | * R Reader page
|
| 1271 | * N Next page
|
| 1272 | *
|
| 1273 | * (see __rb_reserve_next() to see where this happens)
|
| 1274 | *
|
| 1275 | * What the above shows is that the reader just swapped out
|
| 1276 | * the reader page with a page in the buffer, but before it
|
| 1277 | * could make the new header point back to the new page added
|
| 1278 | * it was preempted by a writer. The writer moved forward onto
|
| 1279 | * the new page added by the reader and is about to move forward
|
| 1280 | * again.
|
| 1281 | *
|
| 1282 | * You can see, it is legitimate for the previous pointer of
|
| 1283 | * the head (or any page) not to point back to itself. But only
|
| 1284 | * temporarily.
|
| 1285 |
|
| 1286 |
|
| 1287 | #define RB_PAGE_NORMAL 0UL
|
| 1288 | #define RB_PAGE_HEAD 1UL
|
| 1289 | #define RB_PAGE_UPDATE 2UL
|
| 1290 |
|
| 1291 |
|
| 1292 | #define RB_FLAG_MASK 3UL
|
| 1293 |
|
| 1294 |
|
| 1295 | #define RB_PAGE_MOVED 4UL
|
| 1296 |
|
| 1297 | |
| 1298 | * rb_list_head - remove any bit
|
| 1299 |
|
| 1300 | static struct list_head *rb_list_head(struct list_head *list)
|
| 1301 | {
|
| 1302 | unsigned long val = (unsigned long)list;
|
| 1303 |
|
| 1304 | return (struct list_head *)(val & ~RB_FLAG_MASK);
|
| 1305 | }
|
| 1306 |
|
| 1307 | |
| 1308 | * rb_is_head_page - test if the given page is the head page
|
| 1309 | *
|
| 1310 | * Because the reader may move the head_page pointer, we can
|
| 1311 | * not trust what the head page is (it may be pointing to
|
| 1312 | * the reader page). But if the next page is a header page,
|
| 1313 | * its flags will be non zero.
|
| 1314 |
|
| 1315 | static inline int
|
| 1316 | rb_is_head_page(struct buffer_page *page, struct list_head *list)
|
| 1317 | {
|
| 1318 | unsigned long val;
|
| 1319 |
|
| 1320 | val = (unsigned long)list->next;
|
| 1321 |
|
| 1322 | if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
|
| 1323 | return RB_PAGE_MOVED;
|
| 1324 |
|
| 1325 | return val & RB_FLAG_MASK;
|
| 1326 | }
|
| 1327 |
|
| 1328 | |
| 1329 | * rb_is_reader_page
|
| 1330 | *
|
| 1331 | * The unique thing about the reader page, is that, if the
|
| 1332 | * writer is ever on it, the previous pointer never points
|
| 1333 | * back to the reader page.
|
| 1334 |
|
| 1335 | static bool rb_is_reader_page(struct buffer_page *page)
|
| 1336 | {
|
| 1337 | struct list_head *list = page->list.prev;
|
| 1338 |
|
| 1339 | return rb_list_head(list->next) != &page->list;
|
| 1340 | }
|
| 1341 |
|
| 1342 | |
| 1343 | * rb_set_list_to_head - set a list_head to be pointing to head.
|
| 1344 |
|
| 1345 | static void rb_set_list_to_head(struct list_head *list)
|
| 1346 | {
|
| 1347 | unsigned long *ptr;
|
| 1348 |
|
| 1349 | ptr = (unsigned long *)&list->next;
|
| 1350 | *ptr |= RB_PAGE_HEAD;
|
| 1351 | *ptr &= ~RB_PAGE_UPDATE;
|
| 1352 | }
|
| 1353 |
|
| 1354 | |
| 1355 | * rb_head_page_activate - sets up head page
|
| 1356 |
|
| 1357 | static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
|
| 1358 | {
|
| 1359 | struct buffer_page *head;
|
| 1360 |
|
| 1361 | head = cpu_buffer->head_page;
|
| 1362 | if (!head)
|
| 1363 | return;
|
| 1364 |
|
| 1365 | |
| 1366 | * Set the previous list pointer to have the HEAD flag.
|
| 1367 |
|
| 1368 | rb_set_list_to_head(head->list.prev);
|
| 1369 | }
|
| 1370 |
|
| 1371 | static void rb_list_head_clear(struct list_head *list)
|
| 1372 | {
|
| 1373 | unsigned long *ptr = (unsigned long *)&list->next;
|
| 1374 |
|
| 1375 | *ptr &= ~RB_FLAG_MASK;
|
| 1376 | }
|
| 1377 |
|
| 1378 | |
| 1379 | * rb_head_page_deactivate - clears head page ptr (for free list)
|
| 1380 |
|
| 1381 | static void
|
| 1382 | rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
|
| 1383 | {
|
| 1384 | struct list_head *hd;
|
| 1385 |
|
| 1386 |
|
| 1387 | rb_list_head_clear(cpu_buffer->pages);
|
| 1388 |
|
| 1389 | list_for_each(hd, cpu_buffer->pages)
|
| 1390 | rb_list_head_clear(hd);
|
| 1391 | }
|
| 1392 |
|
| 1393 | static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1394 | struct buffer_page *head,
|
| 1395 | struct buffer_page *prev,
|
| 1396 | int old_flag, int new_flag)
|
| 1397 | {
|
| 1398 | struct list_head *list;
|
| 1399 | unsigned long val = (unsigned long)&head->list;
|
| 1400 | unsigned long ret;
|
| 1401 |
|
| 1402 | list = &prev->list;
|
| 1403 |
|
| 1404 | val &= ~RB_FLAG_MASK;
|
| 1405 |
|
| 1406 | ret = cmpxchg((unsigned long *)&list->next,
|
| 1407 | val | old_flag, val | new_flag);
|
| 1408 |
|
| 1409 |
|
| 1410 | if ((ret & ~RB_FLAG_MASK) != val)
|
| 1411 | return RB_PAGE_MOVED;
|
| 1412 |
|
| 1413 | return ret & RB_FLAG_MASK;
|
| 1414 | }
|
| 1415 |
|
| 1416 | static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1417 | struct buffer_page *head,
|
| 1418 | struct buffer_page *prev,
|
| 1419 | int old_flag)
|
| 1420 | {
|
| 1421 | return rb_head_page_set(cpu_buffer, head, prev,
|
| 1422 | old_flag, RB_PAGE_UPDATE);
|
| 1423 | }
|
| 1424 |
|
| 1425 | static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1426 | struct buffer_page *head,
|
| 1427 | struct buffer_page *prev,
|
| 1428 | int old_flag)
|
| 1429 | {
|
| 1430 | return rb_head_page_set(cpu_buffer, head, prev,
|
| 1431 | old_flag, RB_PAGE_HEAD);
|
| 1432 | }
|
| 1433 |
|
| 1434 | static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1435 | struct buffer_page *head,
|
| 1436 | struct buffer_page *prev,
|
| 1437 | int old_flag)
|
| 1438 | {
|
| 1439 | return rb_head_page_set(cpu_buffer, head, prev,
|
| 1440 | old_flag, RB_PAGE_NORMAL);
|
| 1441 | }
|
| 1442 |
|
| 1443 | static inline void rb_inc_page(struct buffer_page **bpage)
|
| 1444 | {
|
| 1445 | struct list_head *p = rb_list_head((*bpage)->list.next);
|
| 1446 |
|
| 1447 | *bpage = list_entry(p, struct buffer_page, list);
|
| 1448 | }
|
| 1449 |
|
| 1450 | static struct buffer_page *
|
| 1451 | rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
|
| 1452 | {
|
| 1453 | struct buffer_page *head;
|
| 1454 | struct buffer_page *page;
|
| 1455 | struct list_head *list;
|
| 1456 | int i;
|
| 1457 |
|
| 1458 | if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
|
| 1459 | return NULL;
|
| 1460 |
|
| 1461 |
|
| 1462 | list = cpu_buffer->pages;
|
| 1463 | if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
|
| 1464 | return NULL;
|
| 1465 |
|
| 1466 | page = head = cpu_buffer->head_page;
|
| 1467 | |
| 1468 | * It is possible that the writer moves the header behind
|
| 1469 | * where we started, and we miss in one loop.
|
| 1470 | * A second loop should grab the header, but we'll do
|
| 1471 | * three loops just because I'm paranoid.
|
| 1472 |
|
| 1473 | for (i = 0; i < 3; i++) {
|
| 1474 | do {
|
| 1475 | if (rb_is_head_page(page, page->list.prev)) {
|
| 1476 | cpu_buffer->head_page = page;
|
| 1477 | return page;
|
| 1478 | }
|
| 1479 | rb_inc_page(&page);
|
| 1480 | } while (page != head);
|
| 1481 | }
|
| 1482 |
|
| 1483 | RB_WARN_ON(cpu_buffer, 1);
|
| 1484 |
|
| 1485 | return NULL;
|
| 1486 | }
|
| 1487 |
|
| 1488 | static bool rb_head_page_replace(struct buffer_page *old,
|
| 1489 | struct buffer_page *new)
|
| 1490 | {
|
| 1491 | unsigned long *ptr = (unsigned long *)&old->list.prev->next;
|
| 1492 | unsigned long val;
|
| 1493 | unsigned long ret;
|
| 1494 |
|
| 1495 | val = *ptr & ~RB_FLAG_MASK;
|
| 1496 | val |= RB_PAGE_HEAD;
|
| 1497 |
|
| 1498 | ret = cmpxchg(ptr, val, (unsigned long)&new->list);
|
| 1499 |
|
| 1500 | return ret == val;
|
| 1501 | }
|
| 1502 |
|
| 1503 | |
| 1504 | * rb_tail_page_update - move the tail page forward
|
| 1505 |
|
| 1506 | static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1507 | struct buffer_page *tail_page,
|
| 1508 | struct buffer_page *next_page)
|
| 1509 | {
|
| 1510 | unsigned long old_entries;
|
| 1511 | unsigned long old_write;
|
| 1512 |
|
| 1513 | |
| 1514 | * The tail page now needs to be moved forward.
|
| 1515 | *
|
| 1516 | * We need to reset the tail page, but without messing
|
| 1517 | * with possible erasing of data brought in by interrupts
|
| 1518 | * that have moved the tail page and are currently on it.
|
| 1519 | *
|
| 1520 | * We add a counter to the write field to denote this.
|
| 1521 |
|
| 1522 | old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
|
| 1523 | old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
|
| 1524 |
|
| 1525 | local_inc(&cpu_buffer->pages_touched);
|
| 1526 | |
| 1527 | * Just make sure we have seen our old_write and synchronize
|
| 1528 | * with any interrupts that come in.
|
| 1529 |
|
| 1530 | barrier();
|
| 1531 |
|
| 1532 | |
| 1533 | * If the tail page is still the same as what we think
|
| 1534 | * it is, then it is up to us to update the tail
|
| 1535 | * pointer.
|
| 1536 |
|
| 1537 | if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
|
| 1538 |
|
| 1539 | unsigned long val = old_write & ~RB_WRITE_MASK;
|
| 1540 | unsigned long eval = old_entries & ~RB_WRITE_MASK;
|
| 1541 |
|
| 1542 | |
| 1543 | * This will only succeed if an interrupt did
|
| 1544 | * not come in and change it. In which case, we
|
| 1545 | * do not want to modify it.
|
| 1546 | *
|
| 1547 | * We add (void) to let the compiler know that we do not care
|
| 1548 | * about the return value of these functions. We use the
|
| 1549 | * cmpxchg to only update if an interrupt did not already
|
| 1550 | * do it for us. If the cmpxchg fails, we don't care.
|
| 1551 |
|
| 1552 | (void)local_cmpxchg(&next_page->write, old_write, val);
|
| 1553 | (void)local_cmpxchg(&next_page->entries, old_entries, eval);
|
| 1554 |
|
| 1555 | |
| 1556 | * No need to worry about races with clearing out the commit.
|
| 1557 | * it only can increment when a commit takes place. But that
|
| 1558 | * only happens in the outer most nested commit.
|
| 1559 |
|
| 1560 | local_set(&next_page->page->commit, 0);
|
| 1561 |
|
| 1562 |
|
| 1563 | (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
|
| 1564 | }
|
| 1565 | }
|
| 1566 |
|
| 1567 | static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1568 | struct buffer_page *bpage)
|
| 1569 | {
|
| 1570 | unsigned long val = (unsigned long)bpage;
|
| 1571 |
|
| 1572 | RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
|
| 1573 | }
|
| 1574 |
|
| 1575 | |
| 1576 | * rb_check_pages - integrity check of buffer pages
|
| 1577 | * @cpu_buffer: CPU buffer with pages to test
|
| 1578 | *
|
| 1579 | * As a safety measure we check to make sure the data pages have not
|
| 1580 | * been corrupted.
|
| 1581 |
|
| 1582 | static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
|
| 1583 | {
|
| 1584 | struct list_head *head = rb_list_head(cpu_buffer->pages);
|
| 1585 | struct list_head *tmp;
|
| 1586 |
|
| 1587 | if (RB_WARN_ON(cpu_buffer,
|
| 1588 | rb_list_head(rb_list_head(head->next)->prev) != head))
|
| 1589 | return;
|
| 1590 |
|
| 1591 | if (RB_WARN_ON(cpu_buffer,
|
| 1592 | rb_list_head(rb_list_head(head->prev)->next) != head))
|
| 1593 | return;
|
| 1594 |
|
| 1595 | for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
|
| 1596 | if (RB_WARN_ON(cpu_buffer,
|
| 1597 | rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
|
| 1598 | return;
|
| 1599 |
|
| 1600 | if (RB_WARN_ON(cpu_buffer,
|
| 1601 | rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
|
| 1602 | return;
|
| 1603 | }
|
| 1604 | }
|
| 1605 |
|
| 1606 | static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1607 | long nr_pages, struct list_head *pages)
|
| 1608 | {
|
| 1609 | struct buffer_page *bpage, *tmp;
|
| 1610 | bool user_thread = current->mm != NULL;
|
| 1611 | gfp_t mflags;
|
| 1612 | long i;
|
| 1613 |
|
| 1614 | |
| 1615 | * Check if the available memory is there first.
|
| 1616 | * Note, si_mem_available() only gives us a rough estimate of available
|
| 1617 | * memory. It may not be accurate. But we don't care, we just want
|
| 1618 | * to prevent doing any allocation when it is obvious that it is
|
| 1619 | * not going to succeed.
|
| 1620 |
|
| 1621 | i = si_mem_available();
|
| 1622 | if (i < nr_pages)
|
| 1623 | return -ENOMEM;
|
| 1624 |
|
| 1625 | |
| 1626 | * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
|
| 1627 | * gracefully without invoking oom-killer and the system is not
|
| 1628 | * destabilized.
|
| 1629 |
|
| 1630 | mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
|
| 1631 |
|
| 1632 | |
| 1633 | * If a user thread allocates too much, and si_mem_available()
|
| 1634 | * reports there's enough memory, even though there is not.
|
| 1635 | * Make sure the OOM killer kills this thread. This can happen
|
| 1636 | * even with RETRY_MAYFAIL because another task may be doing
|
| 1637 | * an allocation after this task has taken all memory.
|
| 1638 | * This is the task the OOM killer needs to take out during this
|
| 1639 | * loop, even if it was triggered by an allocation somewhere else.
|
| 1640 |
|
| 1641 | if (user_thread)
|
| 1642 | set_current_oom_origin();
|
| 1643 | for (i = 0; i < nr_pages; i++) {
|
| 1644 | struct page *page;
|
| 1645 |
|
| 1646 | bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
|
| 1647 | mflags, cpu_to_node(cpu_buffer->cpu));
|
| 1648 | if (!bpage)
|
| 1649 | goto free_pages;
|
| 1650 |
|
| 1651 | rb_check_bpage(cpu_buffer, bpage);
|
| 1652 |
|
| 1653 | list_add(&bpage->list, pages);
|
| 1654 |
|
| 1655 | page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
|
| 1656 | if (!page)
|
| 1657 | goto free_pages;
|
| 1658 | bpage->page = page_address(page);
|
| 1659 | rb_init_page(bpage->page);
|
| 1660 |
|
| 1661 | if (user_thread && fatal_signal_pending(current))
|
| 1662 | goto free_pages;
|
| 1663 | }
|
| 1664 | if (user_thread)
|
| 1665 | clear_current_oom_origin();
|
| 1666 |
|
| 1667 | return 0;
|
| 1668 |
|
| 1669 | free_pages:
|
| 1670 | list_for_each_entry_safe(bpage, tmp, pages, list) {
|
| 1671 | list_del_init(&bpage->list);
|
| 1672 | free_buffer_page(bpage);
|
| 1673 | }
|
| 1674 | if (user_thread)
|
| 1675 | clear_current_oom_origin();
|
| 1676 |
|
| 1677 | return -ENOMEM;
|
| 1678 | }
|
| 1679 |
|
| 1680 | static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
|
| 1681 | unsigned long nr_pages)
|
| 1682 | {
|
| 1683 | LIST_HEAD(pages);
|
| 1684 |
|
| 1685 | WARN_ON(!nr_pages);
|
| 1686 |
|
| 1687 | if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
|
| 1688 | return -ENOMEM;
|
| 1689 |
|
| 1690 | |
| 1691 | * The ring buffer page list is a circular list that does not
|
| 1692 | * start and end with a list head. All page list items point to
|
| 1693 | * other pages.
|
| 1694 |
|
| 1695 | cpu_buffer->pages = pages.next;
|
| 1696 | list_del(&pages);
|
| 1697 |
|
| 1698 | cpu_buffer->nr_pages = nr_pages;
|
| 1699 |
|
| 1700 | rb_check_pages(cpu_buffer);
|
| 1701 |
|
| 1702 | return 0;
|
| 1703 | }
|
| 1704 |
|
| 1705 | static struct ring_buffer_per_cpu *
|
| 1706 | rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
|
| 1707 | {
|
| 1708 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 1709 | struct buffer_page *bpage;
|
| 1710 | struct page *page;
|
| 1711 | int ret;
|
| 1712 |
|
| 1713 | cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
|
| 1714 | GFP_KERNEL, cpu_to_node(cpu));
|
| 1715 | if (!cpu_buffer)
|
| 1716 | return NULL;
|
| 1717 |
|
| 1718 | cpu_buffer->cpu = cpu;
|
| 1719 | cpu_buffer->buffer = buffer;
|
| 1720 | raw_spin_lock_init(&cpu_buffer->reader_lock);
|
| 1721 | lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
|
| 1722 | cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
|
| 1723 | INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
|
| 1724 | init_completion(&cpu_buffer->update_done);
|
| 1725 | init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
|
| 1726 | init_waitqueue_head(&cpu_buffer->irq_work.waiters);
|
| 1727 | init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
|
| 1728 |
|
| 1729 | bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
|
| 1730 | GFP_KERNEL, cpu_to_node(cpu));
|
| 1731 | if (!bpage)
|
| 1732 | goto fail_free_buffer;
|
| 1733 |
|
| 1734 | rb_check_bpage(cpu_buffer, bpage);
|
| 1735 |
|
| 1736 | cpu_buffer->reader_page = bpage;
|
| 1737 | page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
|
| 1738 | if (!page)
|
| 1739 | goto fail_free_reader;
|
| 1740 | bpage->page = page_address(page);
|
| 1741 | rb_init_page(bpage->page);
|
| 1742 |
|
| 1743 | INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
|
| 1744 | INIT_LIST_HEAD(&cpu_buffer->new_pages);
|
| 1745 |
|
| 1746 | ret = rb_allocate_pages(cpu_buffer, nr_pages);
|
| 1747 | if (ret < 0)
|
| 1748 | goto fail_free_reader;
|
| 1749 |
|
| 1750 | cpu_buffer->head_page
|
| 1751 | = list_entry(cpu_buffer->pages, struct buffer_page, list);
|
| 1752 | cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
|
| 1753 |
|
| 1754 | rb_head_page_activate(cpu_buffer);
|
| 1755 |
|
| 1756 | return cpu_buffer;
|
| 1757 |
|
| 1758 | fail_free_reader:
|
| 1759 | free_buffer_page(cpu_buffer->reader_page);
|
| 1760 |
|
| 1761 | fail_free_buffer:
|
| 1762 | kfree(cpu_buffer);
|
| 1763 | return NULL;
|
| 1764 | }
|
| 1765 |
|
| 1766 | static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
|
| 1767 | {
|
| 1768 | struct list_head *head = cpu_buffer->pages;
|
| 1769 | struct buffer_page *bpage, *tmp;
|
| 1770 |
|
| 1771 | irq_work_sync(&cpu_buffer->irq_work.work);
|
| 1772 |
|
| 1773 | free_buffer_page(cpu_buffer->reader_page);
|
| 1774 |
|
| 1775 | if (head) {
|
| 1776 | rb_head_page_deactivate(cpu_buffer);
|
| 1777 |
|
| 1778 | list_for_each_entry_safe(bpage, tmp, head, list) {
|
| 1779 | list_del_init(&bpage->list);
|
| 1780 | free_buffer_page(bpage);
|
| 1781 | }
|
| 1782 | bpage = list_entry(head, struct buffer_page, list);
|
| 1783 | free_buffer_page(bpage);
|
| 1784 | }
|
| 1785 |
|
| 1786 | kfree(cpu_buffer);
|
| 1787 | }
|
| 1788 |
|
| 1789 | |
| 1790 | * __ring_buffer_alloc - allocate a new ring_buffer
|
| 1791 | * @size: the size in bytes per cpu that is needed.
|
| 1792 | * @flags: attributes to set for the ring buffer.
|
| 1793 | * @key: ring buffer reader_lock_key.
|
| 1794 | *
|
| 1795 | * Currently the only flag that is available is the RB_FL_OVERWRITE
|
| 1796 | * flag. This flag means that the buffer will overwrite old data
|
| 1797 | * when the buffer wraps. If this flag is not set, the buffer will
|
| 1798 | * drop data when the tail hits the head.
|
| 1799 |
|
| 1800 | struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
|
| 1801 | struct lock_class_key *key)
|
| 1802 | {
|
| 1803 | struct trace_buffer *buffer;
|
| 1804 | long nr_pages;
|
| 1805 | int bsize;
|
| 1806 | int cpu;
|
| 1807 | int ret;
|
| 1808 |
|
| 1809 |
|
| 1810 | buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
|
| 1811 | GFP_KERNEL);
|
| 1812 | if (!buffer)
|
| 1813 | return NULL;
|
| 1814 |
|
| 1815 | if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
|
| 1816 | goto fail_free_buffer;
|
| 1817 |
|
| 1818 | nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
|
| 1819 | buffer->flags = flags;
|
| 1820 | buffer->clock = trace_clock_local;
|
| 1821 | buffer->reader_lock_key = key;
|
| 1822 |
|
| 1823 | init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
|
| 1824 | init_waitqueue_head(&buffer->irq_work.waiters);
|
| 1825 |
|
| 1826 |
|
| 1827 | if (nr_pages < 2)
|
| 1828 | nr_pages = 2;
|
| 1829 |
|
| 1830 | buffer->cpus = nr_cpu_ids;
|
| 1831 |
|
| 1832 | bsize = sizeof(void *) * nr_cpu_ids;
|
| 1833 | buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
|
| 1834 | GFP_KERNEL);
|
| 1835 | if (!buffer->buffers)
|
| 1836 | goto fail_free_cpumask;
|
| 1837 |
|
| 1838 | cpu = raw_smp_processor_id();
|
| 1839 | cpumask_set_cpu(cpu, buffer->cpumask);
|
| 1840 | buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
|
| 1841 | if (!buffer->buffers[cpu])
|
| 1842 | goto fail_free_buffers;
|
| 1843 |
|
| 1844 | ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
|
| 1845 | if (ret < 0)
|
| 1846 | goto fail_free_buffers;
|
| 1847 |
|
| 1848 | mutex_init(&buffer->mutex);
|
| 1849 |
|
| 1850 | return buffer;
|
| 1851 |
|
| 1852 | fail_free_buffers:
|
| 1853 | for_each_buffer_cpu(buffer, cpu) {
|
| 1854 | if (buffer->buffers[cpu])
|
| 1855 | rb_free_cpu_buffer(buffer->buffers[cpu]);
|
| 1856 | }
|
| 1857 | kfree(buffer->buffers);
|
| 1858 |
|
| 1859 | fail_free_cpumask:
|
| 1860 | free_cpumask_var(buffer->cpumask);
|
| 1861 |
|
| 1862 | fail_free_buffer:
|
| 1863 | kfree(buffer);
|
| 1864 | return NULL;
|
| 1865 | }
|
| 1866 | EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
|
| 1867 |
|
| 1868 | |
| 1869 | * ring_buffer_free - free a ring buffer.
|
| 1870 | * @buffer: the buffer to free.
|
| 1871 |
|
| 1872 | void
|
| 1873 | ring_buffer_free(struct trace_buffer *buffer)
|
| 1874 | {
|
| 1875 | int cpu;
|
| 1876 |
|
| 1877 | cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
|
| 1878 |
|
| 1879 | irq_work_sync(&buffer->irq_work.work);
|
| 1880 |
|
| 1881 | for_each_buffer_cpu(buffer, cpu)
|
| 1882 | rb_free_cpu_buffer(buffer->buffers[cpu]);
|
| 1883 |
|
| 1884 | kfree(buffer->buffers);
|
| 1885 | free_cpumask_var(buffer->cpumask);
|
| 1886 |
|
| 1887 | kfree(buffer);
|
| 1888 | }
|
| 1889 | EXPORT_SYMBOL_GPL(ring_buffer_free);
|
| 1890 |
|
| 1891 | void ring_buffer_set_clock(struct trace_buffer *buffer,
|
| 1892 | u64 (*clock)(void))
|
| 1893 | {
|
| 1894 | buffer->clock = clock;
|
| 1895 | }
|
| 1896 |
|
| 1897 | void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
|
| 1898 | {
|
| 1899 | buffer->time_stamp_abs = abs;
|
| 1900 | }
|
| 1901 |
|
| 1902 | bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
|
| 1903 | {
|
| 1904 | return buffer->time_stamp_abs;
|
| 1905 | }
|
| 1906 |
|
| 1907 | static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
|
| 1908 |
|
| 1909 | static inline unsigned long rb_page_entries(struct buffer_page *bpage)
|
| 1910 | {
|
| 1911 | return local_read(&bpage->entries) & RB_WRITE_MASK;
|
| 1912 | }
|
| 1913 |
|
| 1914 | static inline unsigned long rb_page_write(struct buffer_page *bpage)
|
| 1915 | {
|
| 1916 | return local_read(&bpage->write) & RB_WRITE_MASK;
|
| 1917 | }
|
| 1918 |
|
| 1919 | static bool
|
| 1920 | rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
|
| 1921 | {
|
| 1922 | struct list_head *tail_page, *to_remove, *next_page;
|
| 1923 | struct buffer_page *to_remove_page, *tmp_iter_page;
|
| 1924 | struct buffer_page *last_page, *first_page;
|
| 1925 | unsigned long nr_removed;
|
| 1926 | unsigned long head_bit;
|
| 1927 | int page_entries;
|
| 1928 |
|
| 1929 | head_bit = 0;
|
| 1930 |
|
| 1931 | raw_spin_lock_irq(&cpu_buffer->reader_lock);
|
| 1932 | atomic_inc(&cpu_buffer->record_disabled);
|
| 1933 | |
| 1934 | * We don't race with the readers since we have acquired the reader
|
| 1935 | * lock. We also don't race with writers after disabling recording.
|
| 1936 | * This makes it easy to figure out the first and the last page to be
|
| 1937 | * removed from the list. We unlink all the pages in between including
|
| 1938 | * the first and last pages. This is done in a busy loop so that we
|
| 1939 | * lose the least number of traces.
|
| 1940 | * The pages are freed after we restart recording and unlock readers.
|
| 1941 |
|
| 1942 | tail_page = &cpu_buffer->tail_page->list;
|
| 1943 |
|
| 1944 | |
| 1945 | * tail page might be on reader page, we remove the next page
|
| 1946 | * from the ring buffer
|
| 1947 |
|
| 1948 | if (cpu_buffer->tail_page == cpu_buffer->reader_page)
|
| 1949 | tail_page = rb_list_head(tail_page->next);
|
| 1950 | to_remove = tail_page;
|
| 1951 |
|
| 1952 |
|
| 1953 | first_page = list_entry(rb_list_head(to_remove->next),
|
| 1954 | struct buffer_page, list);
|
| 1955 |
|
| 1956 | for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
|
| 1957 | to_remove = rb_list_head(to_remove)->next;
|
| 1958 | head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
|
| 1959 | }
|
| 1960 |
|
| 1961 | next_page = rb_list_head(to_remove)->next;
|
| 1962 |
|
| 1963 | |
| 1964 | * Now we remove all pages between tail_page and next_page.
|
| 1965 | * Make sure that we have head_bit value preserved for the
|
| 1966 | * next page
|
| 1967 |
|
| 1968 | tail_page->next = (struct list_head *)((unsigned long)next_page |
|
| 1969 | head_bit);
|
| 1970 | next_page = rb_list_head(next_page);
|
| 1971 | next_page->prev = tail_page;
|
| 1972 |
|
| 1973 |
|
| 1974 | cpu_buffer->pages = next_page;
|
| 1975 |
|
| 1976 |
|
| 1977 | if (head_bit)
|
| 1978 | cpu_buffer->head_page = list_entry(next_page,
|
| 1979 | struct buffer_page, list);
|
| 1980 |
|
| 1981 | |
| 1982 | * change read pointer to make sure any read iterators reset
|
| 1983 | * themselves
|
| 1984 |
|
| 1985 | cpu_buffer->read = 0;
|
| 1986 |
|
| 1987 |
|
| 1988 | atomic_dec(&cpu_buffer->record_disabled);
|
| 1989 | raw_spin_unlock_irq(&cpu_buffer->reader_lock);
|
| 1990 |
|
| 1991 | RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
|
| 1992 |
|
| 1993 |
|
| 1994 | last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
|
| 1995 | list);
|
| 1996 | tmp_iter_page = first_page;
|
| 1997 |
|
| 1998 | do {
|
| 1999 | cond_resched();
|
| 2000 |
|
| 2001 | to_remove_page = tmp_iter_page;
|
| 2002 | rb_inc_page(&tmp_iter_page);
|
| 2003 |
|
| 2004 |
|
| 2005 | page_entries = rb_page_entries(to_remove_page);
|
| 2006 | if (page_entries) {
|
| 2007 | |
| 2008 | * If something was added to this page, it was full
|
| 2009 | * since it is not the tail page. So we deduct the
|
| 2010 | * bytes consumed in ring buffer from here.
|
| 2011 | * Increment overrun to account for the lost events.
|
| 2012 |
|
| 2013 | local_add(page_entries, &cpu_buffer->overrun);
|
| 2014 | local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
|
| 2015 | local_inc(&cpu_buffer->pages_lost);
|
| 2016 | }
|
| 2017 |
|
| 2018 | |
| 2019 | * We have already removed references to this list item, just
|
| 2020 | * free up the buffer_page and its page
|
| 2021 |
|
| 2022 | free_buffer_page(to_remove_page);
|
| 2023 | nr_removed--;
|
| 2024 |
|
| 2025 | } while (to_remove_page != last_page);
|
| 2026 |
|
| 2027 | RB_WARN_ON(cpu_buffer, nr_removed);
|
| 2028 |
|
| 2029 | return nr_removed == 0;
|
| 2030 | }
|
| 2031 |
|
| 2032 | static bool
|
| 2033 | rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
|
| 2034 | {
|
| 2035 | struct list_head *pages = &cpu_buffer->new_pages;
|
| 2036 | unsigned long flags;
|
| 2037 | bool success;
|
| 2038 | int retries;
|
| 2039 |
|
| 2040 |
|
| 2041 | raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 2042 | |
| 2043 | * We are holding the reader lock, so the reader page won't be swapped
|
| 2044 | * in the ring buffer. Now we are racing with the writer trying to
|
| 2045 | * move head page and the tail page.
|
| 2046 | * We are going to adapt the reader page update process where:
|
| 2047 | * 1. We first splice the start and end of list of new pages between
|
| 2048 | * the head page and its previous page.
|
| 2049 | * 2. We cmpxchg the prev_page->next to point from head page to the
|
| 2050 | * start of new pages list.
|
| 2051 | * 3. Finally, we update the head->prev to the end of new list.
|
| 2052 | *
|
| 2053 | * We will try this process 10 times, to make sure that we don't keep
|
| 2054 | * spinning.
|
| 2055 |
|
| 2056 | retries = 10;
|
| 2057 | success = false;
|
| 2058 | while (retries--) {
|
| 2059 | struct list_head *head_page, *prev_page, *r;
|
| 2060 | struct list_head *last_page, *first_page;
|
| 2061 | struct list_head *head_page_with_bit;
|
| 2062 | struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
|
| 2063 |
|
| 2064 | if (!hpage)
|
| 2065 | break;
|
| 2066 | head_page = &hpage->list;
|
| 2067 | prev_page = head_page->prev;
|
| 2068 |
|
| 2069 | first_page = pages->next;
|
| 2070 | last_page = pages->prev;
|
| 2071 |
|
| 2072 | head_page_with_bit = (struct list_head *)
|
| 2073 | ((unsigned long)head_page | RB_PAGE_HEAD);
|
| 2074 |
|
| 2075 | last_page->next = head_page_with_bit;
|
| 2076 | first_page->prev = prev_page;
|
| 2077 |
|
| 2078 | r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
|
| 2079 |
|
| 2080 | if (r == head_page_with_bit) {
|
| 2081 | |
| 2082 | * yay, we replaced the page pointer to our new list,
|
| 2083 | * now, we just have to update to head page's prev
|
| 2084 | * pointer to point to end of list
|
| 2085 |
|
| 2086 | head_page->prev = last_page;
|
| 2087 | success = true;
|
| 2088 | break;
|
| 2089 | }
|
| 2090 | }
|
| 2091 |
|
| 2092 | if (success)
|
| 2093 | INIT_LIST_HEAD(pages);
|
| 2094 | |
| 2095 | * If we weren't successful in adding in new pages, warn and stop
|
| 2096 | * tracing
|
| 2097 |
|
| 2098 | RB_WARN_ON(cpu_buffer, !success);
|
| 2099 | raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 2100 |
|
| 2101 |
|
| 2102 | if (!success) {
|
| 2103 | struct buffer_page *bpage, *tmp;
|
| 2104 | list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
|
| 2105 | list) {
|
| 2106 | list_del_init(&bpage->list);
|
| 2107 | free_buffer_page(bpage);
|
| 2108 | }
|
| 2109 | }
|
| 2110 | return success;
|
| 2111 | }
|
| 2112 |
|
| 2113 | static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
|
| 2114 | {
|
| 2115 | bool success;
|
| 2116 |
|
| 2117 | if (cpu_buffer->nr_pages_to_update > 0)
|
| 2118 | success = rb_insert_pages(cpu_buffer);
|
| 2119 | else
|
| 2120 | success = rb_remove_pages(cpu_buffer,
|
| 2121 | -cpu_buffer->nr_pages_to_update);
|
| 2122 |
|
| 2123 | if (success)
|
| 2124 | cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
|
| 2125 | }
|
| 2126 |
|
| 2127 | static void update_pages_handler(struct work_struct *work)
|
| 2128 | {
|
| 2129 | struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
|
| 2130 | struct ring_buffer_per_cpu, update_pages_work);
|
| 2131 | rb_update_pages(cpu_buffer);
|
| 2132 | complete(&cpu_buffer->update_done);
|
| 2133 | }
|
| 2134 |
|
| 2135 | |
| 2136 | * ring_buffer_resize - resize the ring buffer
|
| 2137 | * @buffer: the buffer to resize.
|
| 2138 | * @size: the new size.
|
| 2139 | * @cpu_id: the cpu buffer to resize
|
| 2140 | *
|
| 2141 | * Minimum size is 2 * BUF_PAGE_SIZE.
|
| 2142 | *
|
| 2143 | * Returns 0 on success and < 0 on failure.
|
| 2144 |
|
| 2145 | int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
|
| 2146 | int cpu_id)
|
| 2147 | {
|
| 2148 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 2149 | unsigned long nr_pages;
|
| 2150 | int cpu, err;
|
| 2151 |
|
| 2152 | |
| 2153 | * Always succeed at resizing a non-existent buffer:
|
| 2154 |
|
| 2155 | if (!buffer)
|
| 2156 | return 0;
|
| 2157 |
|
| 2158 |
|
| 2159 | if (cpu_id != RING_BUFFER_ALL_CPUS &&
|
| 2160 | !cpumask_test_cpu(cpu_id, buffer->cpumask))
|
| 2161 | return 0;
|
| 2162 |
|
| 2163 | nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
|
| 2164 |
|
| 2165 |
|
| 2166 | if (nr_pages < 2)
|
| 2167 | nr_pages = 2;
|
| 2168 |
|
| 2169 |
|
| 2170 | mutex_lock(&buffer->mutex);
|
| 2171 | atomic_inc(&buffer->resizing);
|
| 2172 |
|
| 2173 | if (cpu_id == RING_BUFFER_ALL_CPUS) {
|
| 2174 | |
| 2175 | * Don't succeed if resizing is disabled, as a reader might be
|
| 2176 | * manipulating the ring buffer and is expecting a sane state while
|
| 2177 | * this is true.
|
| 2178 |
|
| 2179 | for_each_buffer_cpu(buffer, cpu) {
|
| 2180 | cpu_buffer = buffer->buffers[cpu];
|
| 2181 | if (atomic_read(&cpu_buffer->resize_disabled)) {
|
| 2182 | err = -EBUSY;
|
| 2183 | goto out_err_unlock;
|
| 2184 | }
|
| 2185 | }
|
| 2186 |
|
| 2187 |
|
| 2188 | for_each_buffer_cpu(buffer, cpu) {
|
| 2189 | cpu_buffer = buffer->buffers[cpu];
|
| 2190 |
|
| 2191 | cpu_buffer->nr_pages_to_update = nr_pages -
|
| 2192 | cpu_buffer->nr_pages;
|
| 2193 | |
| 2194 | * nothing more to do for removing pages or no update
|
| 2195 |
|
| 2196 | if (cpu_buffer->nr_pages_to_update <= 0)
|
| 2197 | continue;
|
| 2198 | |
| 2199 | * to add pages, make sure all new pages can be
|
| 2200 | * allocated without receiving ENOMEM
|
| 2201 |
|
| 2202 | INIT_LIST_HEAD(&cpu_buffer->new_pages);
|
| 2203 | if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
|
| 2204 | &cpu_buffer->new_pages)) {
|
| 2205 |
|
| 2206 | err = -ENOMEM;
|
| 2207 | goto out_err;
|
| 2208 | }
|
| 2209 | }
|
| 2210 |
|
| 2211 | cpus_read_lock();
|
| 2212 | |
| 2213 | * Fire off all the required work handlers
|
| 2214 | * We can't schedule on offline CPUs, but it's not necessary
|
| 2215 | * since we can change their buffer sizes without any race.
|
| 2216 |
|
| 2217 | for_each_buffer_cpu(buffer, cpu) {
|
| 2218 | cpu_buffer = buffer->buffers[cpu];
|
| 2219 | if (!cpu_buffer->nr_pages_to_update)
|
| 2220 | continue;
|
| 2221 |
|
| 2222 |
|
| 2223 | if (!cpu_online(cpu)) {
|
| 2224 | rb_update_pages(cpu_buffer);
|
| 2225 | cpu_buffer->nr_pages_to_update = 0;
|
| 2226 | } else {
|
| 2227 |
|
| 2228 | migrate_disable();
|
| 2229 | if (cpu != smp_processor_id()) {
|
| 2230 | migrate_enable();
|
| 2231 | schedule_work_on(cpu,
|
| 2232 | &cpu_buffer->update_pages_work);
|
| 2233 | } else {
|
| 2234 | update_pages_handler(&cpu_buffer->update_pages_work);
|
| 2235 | migrate_enable();
|
| 2236 | }
|
| 2237 | }
|
| 2238 | }
|
| 2239 |
|
| 2240 |
|
| 2241 | for_each_buffer_cpu(buffer, cpu) {
|
| 2242 | cpu_buffer = buffer->buffers[cpu];
|
| 2243 | if (!cpu_buffer->nr_pages_to_update)
|
| 2244 | continue;
|
| 2245 |
|
| 2246 | if (cpu_online(cpu))
|
| 2247 | wait_for_completion(&cpu_buffer->update_done);
|
| 2248 | cpu_buffer->nr_pages_to_update = 0;
|
| 2249 | }
|
| 2250 |
|
| 2251 | cpus_read_unlock();
|
| 2252 | } else {
|
| 2253 | cpu_buffer = buffer->buffers[cpu_id];
|
| 2254 |
|
| 2255 | if (nr_pages == cpu_buffer->nr_pages)
|
| 2256 | goto out;
|
| 2257 |
|
| 2258 | |
| 2259 | * Don't succeed if resizing is disabled, as a reader might be
|
| 2260 | * manipulating the ring buffer and is expecting a sane state while
|
| 2261 | * this is true.
|
| 2262 |
|
| 2263 | if (atomic_read(&cpu_buffer->resize_disabled)) {
|
| 2264 | err = -EBUSY;
|
| 2265 | goto out_err_unlock;
|
| 2266 | }
|
| 2267 |
|
| 2268 | cpu_buffer->nr_pages_to_update = nr_pages -
|
| 2269 | cpu_buffer->nr_pages;
|
| 2270 |
|
| 2271 | INIT_LIST_HEAD(&cpu_buffer->new_pages);
|
| 2272 | if (cpu_buffer->nr_pages_to_update > 0 &&
|
| 2273 | __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
|
| 2274 | &cpu_buffer->new_pages)) {
|
| 2275 | err = -ENOMEM;
|
| 2276 | goto out_err;
|
| 2277 | }
|
| 2278 |
|
| 2279 | cpus_read_lock();
|
| 2280 |
|
| 2281 |
|
| 2282 | if (!cpu_online(cpu_id))
|
| 2283 | rb_update_pages(cpu_buffer);
|
| 2284 | else {
|
| 2285 |
|
| 2286 | migrate_disable();
|
| 2287 | if (cpu_id == smp_processor_id()) {
|
| 2288 | rb_update_pages(cpu_buffer);
|
| 2289 | migrate_enable();
|
| 2290 | } else {
|
| 2291 | migrate_enable();
|
| 2292 | schedule_work_on(cpu_id,
|
| 2293 | &cpu_buffer->update_pages_work);
|
| 2294 | wait_for_completion(&cpu_buffer->update_done);
|
| 2295 | }
|
| 2296 | }
|
| 2297 |
|
| 2298 | cpu_buffer->nr_pages_to_update = 0;
|
| 2299 | cpus_read_unlock();
|
| 2300 | }
|
| 2301 |
|
| 2302 | out:
|
| 2303 | |
| 2304 | * The ring buffer resize can happen with the ring buffer
|
| 2305 | * enabled, so that the update disturbs the tracing as little
|
| 2306 | * as possible. But if the buffer is disabled, we do not need
|
| 2307 | * to worry about that, and we can take the time to verify
|
| 2308 | * that the buffer is not corrupt.
|
| 2309 |
|
| 2310 | if (atomic_read(&buffer->record_disabled)) {
|
| 2311 | atomic_inc(&buffer->record_disabled);
|
| 2312 | |
| 2313 | * Even though the buffer was disabled, we must make sure
|
| 2314 | * that it is truly disabled before calling rb_check_pages.
|
| 2315 | * There could have been a race between checking
|
| 2316 | * record_disable and incrementing it.
|
| 2317 |
|
| 2318 | synchronize_rcu();
|
| 2319 | for_each_buffer_cpu(buffer, cpu) {
|
| 2320 | cpu_buffer = buffer->buffers[cpu];
|
| 2321 | rb_check_pages(cpu_buffer);
|
| 2322 | }
|
| 2323 | atomic_dec(&buffer->record_disabled);
|
| 2324 | }
|
| 2325 |
|
| 2326 | atomic_dec(&buffer->resizing);
|
| 2327 | mutex_unlock(&buffer->mutex);
|
| 2328 | return 0;
|
| 2329 |
|
| 2330 | out_err:
|
| 2331 | for_each_buffer_cpu(buffer, cpu) {
|
| 2332 | struct buffer_page *bpage, *tmp;
|
| 2333 |
|
| 2334 | cpu_buffer = buffer->buffers[cpu];
|
| 2335 | cpu_buffer->nr_pages_to_update = 0;
|
| 2336 |
|
| 2337 | if (list_empty(&cpu_buffer->new_pages))
|
| 2338 | continue;
|
| 2339 |
|
| 2340 | list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
|
| 2341 | list) {
|
| 2342 | list_del_init(&bpage->list);
|
| 2343 | free_buffer_page(bpage);
|
| 2344 | }
|
| 2345 | }
|
| 2346 | out_err_unlock:
|
| 2347 | atomic_dec(&buffer->resizing);
|
| 2348 | mutex_unlock(&buffer->mutex);
|
| 2349 | return err;
|
| 2350 | }
|
| 2351 | EXPORT_SYMBOL_GPL(ring_buffer_resize);
|
| 2352 |
|
| 2353 | void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
|
| 2354 | {
|
| 2355 | mutex_lock(&buffer->mutex);
|
| 2356 | if (val)
|
| 2357 | buffer->flags |= RB_FL_OVERWRITE;
|
| 2358 | else
|
| 2359 | buffer->flags &= ~RB_FL_OVERWRITE;
|
| 2360 | mutex_unlock(&buffer->mutex);
|
| 2361 | }
|
| 2362 | EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
|
| 2363 |
|
| 2364 | static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
|
| 2365 | {
|
| 2366 | return bpage->page->data + index;
|
| 2367 | }
|
| 2368 |
|
| 2369 | static __always_inline struct ring_buffer_event *
|
| 2370 | rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
|
| 2371 | {
|
| 2372 | return __rb_page_index(cpu_buffer->reader_page,
|
| 2373 | cpu_buffer->reader_page->read);
|
| 2374 | }
|
| 2375 |
|
| 2376 | static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
|
| 2377 | {
|
| 2378 | return local_read(&bpage->page->commit);
|
| 2379 | }
|
| 2380 |
|
| 2381 | static struct ring_buffer_event *
|
| 2382 | rb_iter_head_event(struct ring_buffer_iter *iter)
|
| 2383 | {
|
| 2384 | struct ring_buffer_event *event;
|
| 2385 | struct buffer_page *iter_head_page = iter->head_page;
|
| 2386 | unsigned long commit;
|
| 2387 | unsigned length;
|
| 2388 |
|
| 2389 | if (iter->head != iter->next_event)
|
| 2390 | return iter->event;
|
| 2391 |
|
| 2392 | |
| 2393 | * When the writer goes across pages, it issues a cmpxchg which
|
| 2394 | * is a mb(), which will synchronize with the rmb here.
|
| 2395 | * (see rb_tail_page_update() and __rb_reserve_next())
|
| 2396 |
|
| 2397 | commit = rb_page_commit(iter_head_page);
|
| 2398 | smp_rmb();
|
| 2399 | event = __rb_page_index(iter_head_page, iter->head);
|
| 2400 | length = rb_event_length(event);
|
| 2401 |
|
| 2402 | |
| 2403 | * READ_ONCE() doesn't work on functions and we don't want the
|
| 2404 | * compiler doing any crazy optimizations with length.
|
| 2405 |
|
| 2406 | barrier();
|
| 2407 |
|
| 2408 | if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
|
| 2409 |
|
| 2410 | goto reset;
|
| 2411 |
|
| 2412 | memcpy(iter->event, event, length);
|
| 2413 | |
| 2414 | * If the page stamp is still the same after this rmb() then the
|
| 2415 | * event was safely copied without the writer entering the page.
|
| 2416 |
|
| 2417 | smp_rmb();
|
| 2418 |
|
| 2419 |
|
| 2420 | if (iter->page_stamp != iter_head_page->page->time_stamp ||
|
| 2421 | commit > rb_page_commit(iter_head_page))
|
| 2422 | goto reset;
|
| 2423 |
|
| 2424 | iter->next_event = iter->head + length;
|
| 2425 | return iter->event;
|
| 2426 | reset:
|
| 2427 |
|
| 2428 | iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
|
| 2429 | iter->head = 0;
|
| 2430 | iter->next_event = 0;
|
| 2431 | iter->missed_events = 1;
|
| 2432 | return NULL;
|
| 2433 | }
|
| 2434 |
|
| 2435 |
|
| 2436 | static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
|
| 2437 | {
|
| 2438 | return rb_page_commit(bpage);
|
| 2439 | }
|
| 2440 |
|
| 2441 | static __always_inline unsigned
|
| 2442 | rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
|
| 2443 | {
|
| 2444 | return rb_page_commit(cpu_buffer->commit_page);
|
| 2445 | }
|
| 2446 |
|
| 2447 | static __always_inline unsigned
|
| 2448 | rb_event_index(struct ring_buffer_event *event)
|
| 2449 | {
|
| 2450 | unsigned long addr = (unsigned long)event;
|
| 2451 |
|
| 2452 | return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
|
| 2453 | }
|
| 2454 |
|
| 2455 | static void rb_inc_iter(struct ring_buffer_iter *iter)
|
| 2456 | {
|
| 2457 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
|
| 2458 |
|
| 2459 | |
| 2460 | * The iterator could be on the reader page (it starts there).
|
| 2461 | * But the head could have moved, since the reader was
|
| 2462 | * found. Check for this case and assign the iterator
|
| 2463 | * to the head page instead of next.
|
| 2464 |
|
| 2465 | if (iter->head_page == cpu_buffer->reader_page)
|
| 2466 | iter->head_page = rb_set_head_page(cpu_buffer);
|
| 2467 | else
|
| 2468 | rb_inc_page(&iter->head_page);
|
| 2469 |
|
| 2470 | iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
|
| 2471 | iter->head = 0;
|
| 2472 | iter->next_event = 0;
|
| 2473 | }
|
| 2474 |
|
| 2475 | |
| 2476 | * rb_handle_head_page - writer hit the head page
|
| 2477 | *
|
| 2478 | * Returns: +1 to retry page
|
| 2479 | * 0 to continue
|
| 2480 | * -1 on error
|
| 2481 |
|
| 2482 | static int
|
| 2483 | rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2484 | struct buffer_page *tail_page,
|
| 2485 | struct buffer_page *next_page)
|
| 2486 | {
|
| 2487 | struct buffer_page *new_head;
|
| 2488 | int entries;
|
| 2489 | int type;
|
| 2490 | int ret;
|
| 2491 |
|
| 2492 | entries = rb_page_entries(next_page);
|
| 2493 |
|
| 2494 | |
| 2495 | * The hard part is here. We need to move the head
|
| 2496 | * forward, and protect against both readers on
|
| 2497 | * other CPUs and writers coming in via interrupts.
|
| 2498 |
|
| 2499 | type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
|
| 2500 | RB_PAGE_HEAD);
|
| 2501 |
|
| 2502 | |
| 2503 | * type can be one of four:
|
| 2504 | * NORMAL - an interrupt already moved it for us
|
| 2505 | * HEAD - we are the first to get here.
|
| 2506 | * UPDATE - we are the interrupt interrupting
|
| 2507 | * a current move.
|
| 2508 | * MOVED - a reader on another CPU moved the next
|
| 2509 | * pointer to its reader page. Give up
|
| 2510 | * and try again.
|
| 2511 |
|
| 2512 |
|
| 2513 | switch (type) {
|
| 2514 | case RB_PAGE_HEAD:
|
| 2515 | |
| 2516 | * We changed the head to UPDATE, thus
|
| 2517 | * it is our responsibility to update
|
| 2518 | * the counters.
|
| 2519 |
|
| 2520 | local_add(entries, &cpu_buffer->overrun);
|
| 2521 | local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
|
| 2522 | local_inc(&cpu_buffer->pages_lost);
|
| 2523 |
|
| 2524 | |
| 2525 | * The entries will be zeroed out when we move the
|
| 2526 | * tail page.
|
| 2527 |
|
| 2528 |
|
| 2529 |
|
| 2530 | break;
|
| 2531 |
|
| 2532 | case RB_PAGE_UPDATE:
|
| 2533 | |
| 2534 | * This is an interrupt that interrupt the
|
| 2535 | * previous update. Still more to do.
|
| 2536 |
|
| 2537 | break;
|
| 2538 | case RB_PAGE_NORMAL:
|
| 2539 | |
| 2540 | * An interrupt came in before the update
|
| 2541 | * and processed this for us.
|
| 2542 | * Nothing left to do.
|
| 2543 |
|
| 2544 | return 1;
|
| 2545 | case RB_PAGE_MOVED:
|
| 2546 | |
| 2547 | * The reader is on another CPU and just did
|
| 2548 | * a swap with our next_page.
|
| 2549 | * Try again.
|
| 2550 |
|
| 2551 | return 1;
|
| 2552 | default:
|
| 2553 | RB_WARN_ON(cpu_buffer, 1);
|
| 2554 | return -1;
|
| 2555 | }
|
| 2556 |
|
| 2557 | |
| 2558 | * Now that we are here, the old head pointer is
|
| 2559 | * set to UPDATE. This will keep the reader from
|
| 2560 | * swapping the head page with the reader page.
|
| 2561 | * The reader (on another CPU) will spin till
|
| 2562 | * we are finished.
|
| 2563 | *
|
| 2564 | * We just need to protect against interrupts
|
| 2565 | * doing the job. We will set the next pointer
|
| 2566 | * to HEAD. After that, we set the old pointer
|
| 2567 | * to NORMAL, but only if it was HEAD before.
|
| 2568 | * otherwise we are an interrupt, and only
|
| 2569 | * want the outer most commit to reset it.
|
| 2570 |
|
| 2571 | new_head = next_page;
|
| 2572 | rb_inc_page(&new_head);
|
| 2573 |
|
| 2574 | ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
|
| 2575 | RB_PAGE_NORMAL);
|
| 2576 |
|
| 2577 | |
| 2578 | * Valid returns are:
|
| 2579 | * HEAD - an interrupt came in and already set it.
|
| 2580 | * NORMAL - One of two things:
|
| 2581 | * 1) We really set it.
|
| 2582 | * 2) A bunch of interrupts came in and moved
|
| 2583 | * the page forward again.
|
| 2584 |
|
| 2585 | switch (ret) {
|
| 2586 | case RB_PAGE_HEAD:
|
| 2587 | case RB_PAGE_NORMAL:
|
| 2588 |
|
| 2589 | break;
|
| 2590 | default:
|
| 2591 | RB_WARN_ON(cpu_buffer, 1);
|
| 2592 | return -1;
|
| 2593 | }
|
| 2594 |
|
| 2595 | |
| 2596 | * It is possible that an interrupt came in,
|
| 2597 | * set the head up, then more interrupts came in
|
| 2598 | * and moved it again. When we get back here,
|
| 2599 | * the page would have been set to NORMAL but we
|
| 2600 | * just set it back to HEAD.
|
| 2601 | *
|
| 2602 | * How do you detect this? Well, if that happened
|
| 2603 | * the tail page would have moved.
|
| 2604 |
|
| 2605 | if (ret == RB_PAGE_NORMAL) {
|
| 2606 | struct buffer_page *buffer_tail_page;
|
| 2607 |
|
| 2608 | buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
|
| 2609 | |
| 2610 | * If the tail had moved passed next, then we need
|
| 2611 | * to reset the pointer.
|
| 2612 |
|
| 2613 | if (buffer_tail_page != tail_page &&
|
| 2614 | buffer_tail_page != next_page)
|
| 2615 | rb_head_page_set_normal(cpu_buffer, new_head,
|
| 2616 | next_page,
|
| 2617 | RB_PAGE_HEAD);
|
| 2618 | }
|
| 2619 |
|
| 2620 | |
| 2621 | * If this was the outer most commit (the one that
|
| 2622 | * changed the original pointer from HEAD to UPDATE),
|
| 2623 | * then it is up to us to reset it to NORMAL.
|
| 2624 |
|
| 2625 | if (type == RB_PAGE_HEAD) {
|
| 2626 | ret = rb_head_page_set_normal(cpu_buffer, next_page,
|
| 2627 | tail_page,
|
| 2628 | RB_PAGE_UPDATE);
|
| 2629 | if (RB_WARN_ON(cpu_buffer,
|
| 2630 | ret != RB_PAGE_UPDATE))
|
| 2631 | return -1;
|
| 2632 | }
|
| 2633 |
|
| 2634 | return 0;
|
| 2635 | }
|
| 2636 |
|
| 2637 | static inline void
|
| 2638 | rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2639 | unsigned long tail, struct rb_event_info *info)
|
| 2640 | {
|
| 2641 | struct buffer_page *tail_page = info->tail_page;
|
| 2642 | struct ring_buffer_event *event;
|
| 2643 | unsigned long length = info->length;
|
| 2644 |
|
| 2645 | |
| 2646 | * Only the event that crossed the page boundary
|
| 2647 | * must fill the old tail_page with padding.
|
| 2648 |
|
| 2649 | if (tail >= BUF_PAGE_SIZE) {
|
| 2650 | |
| 2651 | * If the page was filled, then we still need
|
| 2652 | * to update the real_end. Reset it to zero
|
| 2653 | * and the reader will ignore it.
|
| 2654 |
|
| 2655 | if (tail == BUF_PAGE_SIZE)
|
| 2656 | tail_page->real_end = 0;
|
| 2657 |
|
| 2658 | local_sub(length, &tail_page->write);
|
| 2659 | return;
|
| 2660 | }
|
| 2661 |
|
| 2662 | event = __rb_page_index(tail_page, tail);
|
| 2663 |
|
| 2664 |
|
| 2665 | local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
|
| 2666 |
|
| 2667 | |
| 2668 | * Save the original length to the meta data.
|
| 2669 | * This will be used by the reader to add lost event
|
| 2670 | * counter.
|
| 2671 |
|
| 2672 | tail_page->real_end = tail;
|
| 2673 |
|
| 2674 | |
| 2675 | * If this event is bigger than the minimum size, then
|
| 2676 | * we need to be careful that we don't subtract the
|
| 2677 | * write counter enough to allow another writer to slip
|
| 2678 | * in on this page.
|
| 2679 | * We put in a discarded commit instead, to make sure
|
| 2680 | * that this space is not used again.
|
| 2681 | *
|
| 2682 | * If we are less than the minimum size, we don't need to
|
| 2683 | * worry about it.
|
| 2684 |
|
| 2685 | if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
|
| 2686 |
|
| 2687 |
|
| 2688 |
|
| 2689 | rb_event_set_padding(event);
|
| 2690 |
|
| 2691 |
|
| 2692 | smp_wmb();
|
| 2693 |
|
| 2694 |
|
| 2695 | local_sub(length, &tail_page->write);
|
| 2696 | return;
|
| 2697 | }
|
| 2698 |
|
| 2699 |
|
| 2700 | event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
|
| 2701 | event->type_len = RINGBUF_TYPE_PADDING;
|
| 2702 |
|
| 2703 | event->time_delta = 1;
|
| 2704 |
|
| 2705 |
|
| 2706 | smp_wmb();
|
| 2707 |
|
| 2708 |
|
| 2709 | length = (tail + length) - BUF_PAGE_SIZE;
|
| 2710 | local_sub(length, &tail_page->write);
|
| 2711 | }
|
| 2712 |
|
| 2713 | static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
|
| 2714 |
|
| 2715 | |
| 2716 | * This is the slow path, force gcc not to inline it.
|
| 2717 |
|
| 2718 | static noinline struct ring_buffer_event *
|
| 2719 | rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2720 | unsigned long tail, struct rb_event_info *info)
|
| 2721 | {
|
| 2722 | struct buffer_page *tail_page = info->tail_page;
|
| 2723 | struct buffer_page *commit_page = cpu_buffer->commit_page;
|
| 2724 | struct trace_buffer *buffer = cpu_buffer->buffer;
|
| 2725 | struct buffer_page *next_page;
|
| 2726 | int ret;
|
| 2727 |
|
| 2728 | next_page = tail_page;
|
| 2729 |
|
| 2730 | rb_inc_page(&next_page);
|
| 2731 |
|
| 2732 | |
| 2733 | * If for some reason, we had an interrupt storm that made
|
| 2734 | * it all the way around the buffer, bail, and warn
|
| 2735 | * about it.
|
| 2736 |
|
| 2737 | if (unlikely(next_page == commit_page)) {
|
| 2738 | local_inc(&cpu_buffer->commit_overrun);
|
| 2739 | goto out_reset;
|
| 2740 | }
|
| 2741 |
|
| 2742 | |
| 2743 | * This is where the fun begins!
|
| 2744 | *
|
| 2745 | * We are fighting against races between a reader that
|
| 2746 | * could be on another CPU trying to swap its reader
|
| 2747 | * page with the buffer head.
|
| 2748 | *
|
| 2749 | * We are also fighting against interrupts coming in and
|
| 2750 | * moving the head or tail on us as well.
|
| 2751 | *
|
| 2752 | * If the next page is the head page then we have filled
|
| 2753 | * the buffer, unless the commit page is still on the
|
| 2754 | * reader page.
|
| 2755 |
|
| 2756 | if (rb_is_head_page(next_page, &tail_page->list)) {
|
| 2757 |
|
| 2758 | |
| 2759 | * If the commit is not on the reader page, then
|
| 2760 | * move the header page.
|
| 2761 |
|
| 2762 | if (!rb_is_reader_page(cpu_buffer->commit_page)) {
|
| 2763 | |
| 2764 | * If we are not in overwrite mode,
|
| 2765 | * this is easy, just stop here.
|
| 2766 |
|
| 2767 | if (!(buffer->flags & RB_FL_OVERWRITE)) {
|
| 2768 | local_inc(&cpu_buffer->dropped_events);
|
| 2769 | goto out_reset;
|
| 2770 | }
|
| 2771 |
|
| 2772 | ret = rb_handle_head_page(cpu_buffer,
|
| 2773 | tail_page,
|
| 2774 | next_page);
|
| 2775 | if (ret < 0)
|
| 2776 | goto out_reset;
|
| 2777 | if (ret)
|
| 2778 | goto out_again;
|
| 2779 | } else {
|
| 2780 | |
| 2781 | * We need to be careful here too. The
|
| 2782 | * commit page could still be on the reader
|
| 2783 | * page. We could have a small buffer, and
|
| 2784 | * have filled up the buffer with events
|
| 2785 | * from interrupts and such, and wrapped.
|
| 2786 | *
|
| 2787 | * Note, if the tail page is also on the
|
| 2788 | * reader_page, we let it move out.
|
| 2789 |
|
| 2790 | if (unlikely((cpu_buffer->commit_page !=
|
| 2791 | cpu_buffer->tail_page) &&
|
| 2792 | (cpu_buffer->commit_page ==
|
| 2793 | cpu_buffer->reader_page))) {
|
| 2794 | local_inc(&cpu_buffer->commit_overrun);
|
| 2795 | goto out_reset;
|
| 2796 | }
|
| 2797 | }
|
| 2798 | }
|
| 2799 |
|
| 2800 | rb_tail_page_update(cpu_buffer, tail_page, next_page);
|
| 2801 |
|
| 2802 | out_again:
|
| 2803 |
|
| 2804 | rb_reset_tail(cpu_buffer, tail, info);
|
| 2805 |
|
| 2806 |
|
| 2807 | rb_end_commit(cpu_buffer);
|
| 2808 |
|
| 2809 | local_inc(&cpu_buffer->committing);
|
| 2810 |
|
| 2811 |
|
| 2812 | return ERR_PTR(-EAGAIN);
|
| 2813 |
|
| 2814 | out_reset:
|
| 2815 |
|
| 2816 | rb_reset_tail(cpu_buffer, tail, info);
|
| 2817 |
|
| 2818 | return NULL;
|
| 2819 | }
|
| 2820 |
|
| 2821 |
|
| 2822 | static struct ring_buffer_event *
|
| 2823 | rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
|
| 2824 | {
|
| 2825 | if (abs)
|
| 2826 | event->type_len = RINGBUF_TYPE_TIME_STAMP;
|
| 2827 | else
|
| 2828 | event->type_len = RINGBUF_TYPE_TIME_EXTEND;
|
| 2829 |
|
| 2830 |
|
| 2831 | if (abs || rb_event_index(event)) {
|
| 2832 | event->time_delta = delta & TS_MASK;
|
| 2833 | event->array[0] = delta >> TS_SHIFT;
|
| 2834 | } else {
|
| 2835 |
|
| 2836 | event->time_delta = 0;
|
| 2837 | event->array[0] = 0;
|
| 2838 | }
|
| 2839 |
|
| 2840 | return skip_time_extend(event);
|
| 2841 | }
|
| 2842 |
|
| 2843 | #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
|
| 2844 | static inline bool sched_clock_stable(void)
|
| 2845 | {
|
| 2846 | return true;
|
| 2847 | }
|
| 2848 | #endif
|
| 2849 |
|
| 2850 | static void
|
| 2851 | rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2852 | struct rb_event_info *info)
|
| 2853 | {
|
| 2854 | u64 write_stamp;
|
| 2855 |
|
| 2856 | WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
|
| 2857 | (unsigned long long)info->delta,
|
| 2858 | (unsigned long long)info->ts,
|
| 2859 | (unsigned long long)info->before,
|
| 2860 | (unsigned long long)info->after,
|
| 2861 | (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
|
| 2862 | sched_clock_stable() ? "" :
|
| 2863 | "If you just came from a suspend/resume,\n"
|
| 2864 | "please switch to the trace global clock:\n"
|
| 2865 | " echo global > /sys/kernel/tracing/trace_clock\n"
|
| 2866 | "or add trace_clock=global to the kernel command line\n");
|
| 2867 | }
|
| 2868 |
|
| 2869 | static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2870 | struct ring_buffer_event **event,
|
| 2871 | struct rb_event_info *info,
|
| 2872 | u64 *delta,
|
| 2873 | unsigned int *length)
|
| 2874 | {
|
| 2875 | bool abs = info->add_timestamp &
|
| 2876 | (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
|
| 2877 |
|
| 2878 | if (unlikely(info->delta > (1ULL << 59))) {
|
| 2879 | |
| 2880 | * Some timers can use more than 59 bits, and when a timestamp
|
| 2881 | * is added to the buffer, it will lose those bits.
|
| 2882 |
|
| 2883 | if (abs && (info->ts & TS_MSB)) {
|
| 2884 | info->delta &= ABS_TS_MASK;
|
| 2885 |
|
| 2886 |
|
| 2887 | } else if (info->before == info->after && info->before > info->ts) {
|
| 2888 |
|
| 2889 | static int once;
|
| 2890 |
|
| 2891 | |
| 2892 | * This is possible with a recalibrating of the TSC.
|
| 2893 | * Do not produce a call stack, but just report it.
|
| 2894 |
|
| 2895 | if (!once) {
|
| 2896 | once++;
|
| 2897 | pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
|
| 2898 | info->before, info->ts);
|
| 2899 | }
|
| 2900 | } else
|
| 2901 | rb_check_timestamp(cpu_buffer, info);
|
| 2902 | if (!abs)
|
| 2903 | info->delta = 0;
|
| 2904 | }
|
| 2905 | *event = rb_add_time_stamp(*event, info->delta, abs);
|
| 2906 | *length -= RB_LEN_TIME_EXTEND;
|
| 2907 | *delta = 0;
|
| 2908 | }
|
| 2909 |
|
| 2910 | |
| 2911 | * rb_update_event - update event type and data
|
| 2912 | * @cpu_buffer: The per cpu buffer of the @event
|
| 2913 | * @event: the event to update
|
| 2914 | * @info: The info to update the @event with (contains length and delta)
|
| 2915 | *
|
| 2916 | * Update the type and data fields of the @event. The length
|
| 2917 | * is the actual size that is written to the ring buffer,
|
| 2918 | * and with this, we can determine what to place into the
|
| 2919 | * data field.
|
| 2920 |
|
| 2921 | static void
|
| 2922 | rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
|
| 2923 | struct ring_buffer_event *event,
|
| 2924 | struct rb_event_info *info)
|
| 2925 | {
|
| 2926 | unsigned length = info->length;
|
| 2927 | u64 delta = info->delta;
|
| 2928 | unsigned int nest = local_read(&cpu_buffer->committing) - 1;
|
| 2929 |
|
| 2930 | if (!WARN_ON_ONCE(nest >= MAX_NEST))
|
| 2931 | cpu_buffer->event_stamp[nest] = info->ts;
|
| 2932 |
|
| 2933 | |
| 2934 | * If we need to add a timestamp, then we
|
| 2935 | * add it to the start of the reserved space.
|
| 2936 |
|
| 2937 | if (unlikely(info->add_timestamp))
|
| 2938 | rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
|
| 2939 |
|
| 2940 | event->time_delta = delta;
|
| 2941 | length -= RB_EVNT_HDR_SIZE;
|
| 2942 | if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
|
| 2943 | event->type_len = 0;
|
| 2944 | event->array[0] = length;
|
| 2945 | } else
|
| 2946 | event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
|
| 2947 | }
|
| 2948 |
|
| 2949 | static unsigned rb_calculate_event_length(unsigned length)
|
| 2950 | {
|
| 2951 | struct ring_buffer_event event;
|
| 2952 |
|
| 2953 |
|
| 2954 | if (!length)
|
| 2955 | length++;
|
| 2956 |
|
| 2957 | if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
|
| 2958 | length += sizeof(event.array[0]);
|
| 2959 |
|
| 2960 | length += RB_EVNT_HDR_SIZE;
|
| 2961 | length = ALIGN(length, RB_ARCH_ALIGNMENT);
|
| 2962 |
|
| 2963 | |
| 2964 | * In case the time delta is larger than the 27 bits for it
|
| 2965 | * in the header, we need to add a timestamp. If another
|
| 2966 | * event comes in when trying to discard this one to increase
|
| 2967 | * the length, then the timestamp will be added in the allocated
|
| 2968 | * space of this event. If length is bigger than the size needed
|
| 2969 | * for the TIME_EXTEND, then padding has to be used. The events
|
| 2970 | * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
|
| 2971 | * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
|
| 2972 | * As length is a multiple of 4, we only need to worry if it
|
| 2973 | * is 12 (RB_LEN_TIME_EXTEND + 4).
|
| 2974 |
|
| 2975 | if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
|
| 2976 | length += RB_ALIGNMENT;
|
| 2977 |
|
| 2978 | return length;
|
| 2979 | }
|
| 2980 |
|
| 2981 | static u64 rb_time_delta(struct ring_buffer_event *event)
|
| 2982 | {
|
| 2983 | switch (event->type_len) {
|
| 2984 | case RINGBUF_TYPE_PADDING:
|
| 2985 | return 0;
|
| 2986 |
|
| 2987 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 2988 | return rb_event_time_stamp(event);
|
| 2989 |
|
| 2990 | case RINGBUF_TYPE_TIME_STAMP:
|
| 2991 | return 0;
|
| 2992 |
|
| 2993 | case RINGBUF_TYPE_DATA:
|
| 2994 | return event->time_delta;
|
| 2995 | default:
|
| 2996 | return 0;
|
| 2997 | }
|
| 2998 | }
|
| 2999 |
|
| 3000 | static inline bool
|
| 3001 | rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
|
| 3002 | struct ring_buffer_event *event)
|
| 3003 | {
|
| 3004 | unsigned long new_index, old_index;
|
| 3005 | struct buffer_page *bpage;
|
| 3006 | unsigned long index;
|
| 3007 | unsigned long addr;
|
| 3008 | u64 write_stamp;
|
| 3009 | u64 delta;
|
| 3010 |
|
| 3011 | new_index = rb_event_index(event);
|
| 3012 | old_index = new_index + rb_event_ts_length(event);
|
| 3013 | addr = (unsigned long)event;
|
| 3014 | addr &= PAGE_MASK;
|
| 3015 |
|
| 3016 | bpage = READ_ONCE(cpu_buffer->tail_page);
|
| 3017 |
|
| 3018 | delta = rb_time_delta(event);
|
| 3019 |
|
| 3020 | if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
|
| 3021 | return false;
|
| 3022 |
|
| 3023 |
|
| 3024 | barrier();
|
| 3025 |
|
| 3026 | if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
|
| 3027 | unsigned long write_mask =
|
| 3028 | local_read(&bpage->write) & ~RB_WRITE_MASK;
|
| 3029 | unsigned long event_length = rb_event_length(event);
|
| 3030 |
|
| 3031 |
|
| 3032 | if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
|
| 3033 | write_stamp, write_stamp - delta))
|
| 3034 | return false;
|
| 3035 |
|
| 3036 | |
| 3037 | * It's possible that the event time delta is zero
|
| 3038 | * (has the same time stamp as the previous event)
|
| 3039 | * in which case write_stamp and before_stamp could
|
| 3040 | * be the same. In such a case, force before_stamp
|
| 3041 | * to be different than write_stamp. It doesn't
|
| 3042 | * matter what it is, as long as its different.
|
| 3043 |
|
| 3044 | if (!delta)
|
| 3045 | rb_time_set(&cpu_buffer->before_stamp, 0);
|
| 3046 |
|
| 3047 | |
| 3048 | * If an event were to come in now, it would see that the
|
| 3049 | * write_stamp and the before_stamp are different, and assume
|
| 3050 | * that this event just added itself before updating
|
| 3051 | * the write stamp. The interrupting event will fix the
|
| 3052 | * write stamp for us, and use the before stamp as its delta.
|
| 3053 |
|
| 3054 |
|
| 3055 | |
| 3056 | * This is on the tail page. It is possible that
|
| 3057 | * a write could come in and move the tail page
|
| 3058 | * and write to the next page. That is fine
|
| 3059 | * because we just shorten what is on this page.
|
| 3060 |
|
| 3061 | old_index += write_mask;
|
| 3062 | new_index += write_mask;
|
| 3063 | index = local_cmpxchg(&bpage->write, old_index, new_index);
|
| 3064 | if (index == old_index) {
|
| 3065 |
|
| 3066 | local_sub(event_length, &cpu_buffer->entries_bytes);
|
| 3067 | return true;
|
| 3068 | }
|
| 3069 | }
|
| 3070 |
|
| 3071 |
|
| 3072 | return false;
|
| 3073 | }
|
| 3074 |
|
| 3075 | static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
|
| 3076 | {
|
| 3077 | local_inc(&cpu_buffer->committing);
|
| 3078 | local_inc(&cpu_buffer->commits);
|
| 3079 | }
|
| 3080 |
|
| 3081 | static __always_inline void
|
| 3082 | rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
|
| 3083 | {
|
| 3084 | unsigned long max_count;
|
| 3085 |
|
| 3086 | |
| 3087 | * We only race with interrupts and NMIs on this CPU.
|
| 3088 | * If we own the commit event, then we can commit
|
| 3089 | * all others that interrupted us, since the interruptions
|
| 3090 | * are in stack format (they finish before they come
|
| 3091 | * back to us). This allows us to do a simple loop to
|
| 3092 | * assign the commit to the tail.
|
| 3093 |
|
| 3094 | again:
|
| 3095 | max_count = cpu_buffer->nr_pages * 100;
|
| 3096 |
|
| 3097 | while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
|
| 3098 | if (RB_WARN_ON(cpu_buffer, !(--max_count)))
|
| 3099 | return;
|
| 3100 | if (RB_WARN_ON(cpu_buffer,
|
| 3101 | rb_is_reader_page(cpu_buffer->tail_page)))
|
| 3102 | return;
|
| 3103 | |
| 3104 | * No need for a memory barrier here, as the update
|
| 3105 | * of the tail_page did it for this page.
|
| 3106 |
|
| 3107 | local_set(&cpu_buffer->commit_page->page->commit,
|
| 3108 | rb_page_write(cpu_buffer->commit_page));
|
| 3109 | rb_inc_page(&cpu_buffer->commit_page);
|
| 3110 |
|
| 3111 | barrier();
|
| 3112 | }
|
| 3113 | while (rb_commit_index(cpu_buffer) !=
|
| 3114 | rb_page_write(cpu_buffer->commit_page)) {
|
| 3115 |
|
| 3116 |
|
| 3117 | smp_wmb();
|
| 3118 | local_set(&cpu_buffer->commit_page->page->commit,
|
| 3119 | rb_page_write(cpu_buffer->commit_page));
|
| 3120 | RB_WARN_ON(cpu_buffer,
|
| 3121 | local_read(&cpu_buffer->commit_page->page->commit) &
|
| 3122 | ~RB_WRITE_MASK);
|
| 3123 | barrier();
|
| 3124 | }
|
| 3125 |
|
| 3126 |
|
| 3127 | barrier();
|
| 3128 |
|
| 3129 | |
| 3130 | * If an interrupt came in just after the first while loop
|
| 3131 | * and pushed the tail page forward, we will be left with
|
| 3132 | * a dangling commit that will never go forward.
|
| 3133 |
|
| 3134 | if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
|
| 3135 | goto again;
|
| 3136 | }
|
| 3137 |
|
| 3138 | static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
|
| 3139 | {
|
| 3140 | unsigned long commits;
|
| 3141 |
|
| 3142 | if (RB_WARN_ON(cpu_buffer,
|
| 3143 | !local_read(&cpu_buffer->committing)))
|
| 3144 | return;
|
| 3145 |
|
| 3146 | again:
|
| 3147 | commits = local_read(&cpu_buffer->commits);
|
| 3148 |
|
| 3149 | barrier();
|
| 3150 | if (local_read(&cpu_buffer->committing) == 1)
|
| 3151 | rb_set_commit_to_write(cpu_buffer);
|
| 3152 |
|
| 3153 | local_dec(&cpu_buffer->committing);
|
| 3154 |
|
| 3155 |
|
| 3156 | barrier();
|
| 3157 |
|
| 3158 | |
| 3159 | * Need to account for interrupts coming in between the
|
| 3160 | * updating of the commit page and the clearing of the
|
| 3161 | * committing counter.
|
| 3162 |
|
| 3163 | if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
|
| 3164 | !local_read(&cpu_buffer->committing)) {
|
| 3165 | local_inc(&cpu_buffer->committing);
|
| 3166 | goto again;
|
| 3167 | }
|
| 3168 | }
|
| 3169 |
|
| 3170 | static inline void rb_event_discard(struct ring_buffer_event *event)
|
| 3171 | {
|
| 3172 | if (extended_time(event))
|
| 3173 | event = skip_time_extend(event);
|
| 3174 |
|
| 3175 |
|
| 3176 | event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
|
| 3177 | event->type_len = RINGBUF_TYPE_PADDING;
|
| 3178 |
|
| 3179 | if (!event->time_delta)
|
| 3180 | event->time_delta = 1;
|
| 3181 | }
|
| 3182 |
|
| 3183 | static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
|
| 3184 | {
|
| 3185 | local_inc(&cpu_buffer->entries);
|
| 3186 | rb_end_commit(cpu_buffer);
|
| 3187 | }
|
| 3188 |
|
| 3189 | static __always_inline void
|
| 3190 | rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
|
| 3191 | {
|
| 3192 | if (buffer->irq_work.waiters_pending) {
|
| 3193 | buffer->irq_work.waiters_pending = false;
|
| 3194 |
|
| 3195 | irq_work_queue(&buffer->irq_work.work);
|
| 3196 | }
|
| 3197 |
|
| 3198 | if (cpu_buffer->irq_work.waiters_pending) {
|
| 3199 | cpu_buffer->irq_work.waiters_pending = false;
|
| 3200 |
|
| 3201 | irq_work_queue(&cpu_buffer->irq_work.work);
|
| 3202 | }
|
| 3203 |
|
| 3204 | if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
|
| 3205 | return;
|
| 3206 |
|
| 3207 | if (cpu_buffer->reader_page == cpu_buffer->commit_page)
|
| 3208 | return;
|
| 3209 |
|
| 3210 | if (!cpu_buffer->irq_work.full_waiters_pending)
|
| 3211 | return;
|
| 3212 |
|
| 3213 | cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
|
| 3214 |
|
| 3215 | if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
|
| 3216 | return;
|
| 3217 |
|
| 3218 | cpu_buffer->irq_work.wakeup_full = true;
|
| 3219 | cpu_buffer->irq_work.full_waiters_pending = false;
|
| 3220 |
|
| 3221 | irq_work_queue(&cpu_buffer->irq_work.work);
|
| 3222 | }
|
| 3223 |
|
| 3224 | #ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
|
| 3225 | # define do_ring_buffer_record_recursion() \
|
| 3226 | do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
|
| 3227 | #else
|
| 3228 | # define do_ring_buffer_record_recursion() do { } while (0)
|
| 3229 | #endif
|
| 3230 |
|
| 3231 | |
| 3232 | * The lock and unlock are done within a preempt disable section.
|
| 3233 | * The current_context per_cpu variable can only be modified
|
| 3234 | * by the current task between lock and unlock. But it can
|
| 3235 | * be modified more than once via an interrupt. To pass this
|
| 3236 | * information from the lock to the unlock without having to
|
| 3237 | * access the 'in_interrupt()' functions again (which do show
|
| 3238 | * a bit of overhead in something as critical as function tracing,
|
| 3239 | * we use a bitmask trick.
|
| 3240 | *
|
| 3241 | * bit 1 = NMI context
|
| 3242 | * bit 2 = IRQ context
|
| 3243 | * bit 3 = SoftIRQ context
|
| 3244 | * bit 4 = normal context.
|
| 3245 | *
|
| 3246 | * This works because this is the order of contexts that can
|
| 3247 | * preempt other contexts. A SoftIRQ never preempts an IRQ
|
| 3248 | * context.
|
| 3249 | *
|
| 3250 | * When the context is determined, the corresponding bit is
|
| 3251 | * checked and set (if it was set, then a recursion of that context
|
| 3252 | * happened).
|
| 3253 | *
|
| 3254 | * On unlock, we need to clear this bit. To do so, just subtract
|
| 3255 | * 1 from the current_context and AND it to itself.
|
| 3256 | *
|
| 3257 | * (binary)
|
| 3258 | * 101 - 1 = 100
|
| 3259 | * 101 & 100 = 100 (clearing bit zero)
|
| 3260 | *
|
| 3261 | * 1010 - 1 = 1001
|
| 3262 | * 1010 & 1001 = 1000 (clearing bit 1)
|
| 3263 | *
|
| 3264 | * The least significant bit can be cleared this way, and it
|
| 3265 | * just so happens that it is the same bit corresponding to
|
| 3266 | * the current context.
|
| 3267 | *
|
| 3268 | * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
|
| 3269 | * is set when a recursion is detected at the current context, and if
|
| 3270 | * the TRANSITION bit is already set, it will fail the recursion.
|
| 3271 | * This is needed because there's a lag between the changing of
|
| 3272 | * interrupt context and updating the preempt count. In this case,
|
| 3273 | * a false positive will be found. To handle this, one extra recursion
|
| 3274 | * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
|
| 3275 | * bit is already set, then it is considered a recursion and the function
|
| 3276 | * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
|
| 3277 | *
|
| 3278 | * On the trace_recursive_unlock(), the TRANSITION bit will be the first
|
| 3279 | * to be cleared. Even if it wasn't the context that set it. That is,
|
| 3280 | * if an interrupt comes in while NORMAL bit is set and the ring buffer
|
| 3281 | * is called before preempt_count() is updated, since the check will
|
| 3282 | * be on the NORMAL bit, the TRANSITION bit will then be set. If an
|
| 3283 | * NMI then comes in, it will set the NMI bit, but when the NMI code
|
| 3284 | * does the trace_recursive_unlock() it will clear the TRANSITION bit
|
| 3285 | * and leave the NMI bit set. But this is fine, because the interrupt
|
| 3286 | * code that set the TRANSITION bit will then clear the NMI bit when it
|
| 3287 | * calls trace_recursive_unlock(). If another NMI comes in, it will
|
| 3288 | * set the TRANSITION bit and continue.
|
| 3289 | *
|
| 3290 | * Note: The TRANSITION bit only handles a single transition between context.
|
| 3291 |
|
| 3292 |
|
| 3293 | static __always_inline bool
|
| 3294 | trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
|
| 3295 | {
|
| 3296 | unsigned int val = cpu_buffer->current_context;
|
| 3297 | int bit = interrupt_context_level();
|
| 3298 |
|
| 3299 | bit = RB_CTX_NORMAL - bit;
|
| 3300 |
|
| 3301 | if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
|
| 3302 | |
| 3303 | * It is possible that this was called by transitioning
|
| 3304 | * between interrupt context, and preempt_count() has not
|
| 3305 | * been updated yet. In this case, use the TRANSITION bit.
|
| 3306 |
|
| 3307 | bit = RB_CTX_TRANSITION;
|
| 3308 | if (val & (1 << (bit + cpu_buffer->nest))) {
|
| 3309 | do_ring_buffer_record_recursion();
|
| 3310 | return true;
|
| 3311 | }
|
| 3312 | }
|
| 3313 |
|
| 3314 | val |= (1 << (bit + cpu_buffer->nest));
|
| 3315 | cpu_buffer->current_context = val;
|
| 3316 |
|
| 3317 | return false;
|
| 3318 | }
|
| 3319 |
|
| 3320 | static __always_inline void
|
| 3321 | trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
|
| 3322 | {
|
| 3323 | cpu_buffer->current_context &=
|
| 3324 | cpu_buffer->current_context - (1 << cpu_buffer->nest);
|
| 3325 | }
|
| 3326 |
|
| 3327 |
|
| 3328 | #define NESTED_BITS 5
|
| 3329 |
|
| 3330 | |
| 3331 | * ring_buffer_nest_start - Allow to trace while nested
|
| 3332 | * @buffer: The ring buffer to modify
|
| 3333 | *
|
| 3334 | * The ring buffer has a safety mechanism to prevent recursion.
|
| 3335 | * But there may be a case where a trace needs to be done while
|
| 3336 | * tracing something else. In this case, calling this function
|
| 3337 | * will allow this function to nest within a currently active
|
| 3338 | * ring_buffer_lock_reserve().
|
| 3339 | *
|
| 3340 | * Call this function before calling another ring_buffer_lock_reserve() and
|
| 3341 | * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
|
| 3342 |
|
| 3343 | void ring_buffer_nest_start(struct trace_buffer *buffer)
|
| 3344 | {
|
| 3345 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3346 | int cpu;
|
| 3347 |
|
| 3348 |
|
| 3349 | preempt_disable_notrace();
|
| 3350 | cpu = raw_smp_processor_id();
|
| 3351 | cpu_buffer = buffer->buffers[cpu];
|
| 3352 |
|
| 3353 | cpu_buffer->nest += NESTED_BITS;
|
| 3354 | }
|
| 3355 |
|
| 3356 | |
| 3357 | * ring_buffer_nest_end - Allow to trace while nested
|
| 3358 | * @buffer: The ring buffer to modify
|
| 3359 | *
|
| 3360 | * Must be called after ring_buffer_nest_start() and after the
|
| 3361 | * ring_buffer_unlock_commit().
|
| 3362 |
|
| 3363 | void ring_buffer_nest_end(struct trace_buffer *buffer)
|
| 3364 | {
|
| 3365 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3366 | int cpu;
|
| 3367 |
|
| 3368 |
|
| 3369 | cpu = raw_smp_processor_id();
|
| 3370 | cpu_buffer = buffer->buffers[cpu];
|
| 3371 |
|
| 3372 | cpu_buffer->nest -= NESTED_BITS;
|
| 3373 | preempt_enable_notrace();
|
| 3374 | }
|
| 3375 |
|
| 3376 | |
| 3377 | * ring_buffer_unlock_commit - commit a reserved
|
| 3378 | * @buffer: The buffer to commit to
|
| 3379 | * @event: The event pointer to commit.
|
| 3380 | *
|
| 3381 | * This commits the data to the ring buffer, and releases any locks held.
|
| 3382 | *
|
| 3383 | * Must be paired with ring_buffer_lock_reserve.
|
| 3384 |
|
| 3385 | int ring_buffer_unlock_commit(struct trace_buffer *buffer)
|
| 3386 | {
|
| 3387 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3388 | int cpu = raw_smp_processor_id();
|
| 3389 |
|
| 3390 | cpu_buffer = buffer->buffers[cpu];
|
| 3391 |
|
| 3392 | rb_commit(cpu_buffer);
|
| 3393 |
|
| 3394 | rb_wakeups(buffer, cpu_buffer);
|
| 3395 |
|
| 3396 | trace_recursive_unlock(cpu_buffer);
|
| 3397 |
|
| 3398 | preempt_enable_notrace();
|
| 3399 |
|
| 3400 | return 0;
|
| 3401 | }
|
| 3402 | EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
|
| 3403 |
|
| 3404 |
|
| 3405 | #define CHECK_FULL_PAGE 1L
|
| 3406 |
|
| 3407 | #ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
|
| 3408 | static void dump_buffer_page(struct buffer_data_page *bpage,
|
| 3409 | struct rb_event_info *info,
|
| 3410 | unsigned long tail)
|
| 3411 | {
|
| 3412 | struct ring_buffer_event *event;
|
| 3413 | u64 ts, delta;
|
| 3414 | int e;
|
| 3415 |
|
| 3416 | ts = bpage->time_stamp;
|
| 3417 | pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
|
| 3418 |
|
| 3419 | for (e = 0; e < tail; e += rb_event_length(event)) {
|
| 3420 |
|
| 3421 | event = (struct ring_buffer_event *)(bpage->data + e);
|
| 3422 |
|
| 3423 | switch (event->type_len) {
|
| 3424 |
|
| 3425 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 3426 | delta = rb_event_time_stamp(event);
|
| 3427 | ts += delta;
|
| 3428 | pr_warn(" [%lld] delta:%lld TIME EXTEND\n", ts, delta);
|
| 3429 | break;
|
| 3430 |
|
| 3431 | case RINGBUF_TYPE_TIME_STAMP:
|
| 3432 | delta = rb_event_time_stamp(event);
|
| 3433 | ts = rb_fix_abs_ts(delta, ts);
|
| 3434 | pr_warn(" [%lld] absolute:%lld TIME STAMP\n", ts, delta);
|
| 3435 | break;
|
| 3436 |
|
| 3437 | case RINGBUF_TYPE_PADDING:
|
| 3438 | ts += event->time_delta;
|
| 3439 | pr_warn(" [%lld] delta:%d PADDING\n", ts, event->time_delta);
|
| 3440 | break;
|
| 3441 |
|
| 3442 | case RINGBUF_TYPE_DATA:
|
| 3443 | ts += event->time_delta;
|
| 3444 | pr_warn(" [%lld] delta:%d\n", ts, event->time_delta);
|
| 3445 | break;
|
| 3446 |
|
| 3447 | default:
|
| 3448 | break;
|
| 3449 | }
|
| 3450 | }
|
| 3451 | }
|
| 3452 |
|
| 3453 | static DEFINE_PER_CPU(atomic_t, checking);
|
| 3454 | static atomic_t ts_dump;
|
| 3455 |
|
| 3456 | |
| 3457 | * Check if the current event time stamp matches the deltas on
|
| 3458 | * the buffer page.
|
| 3459 |
|
| 3460 | static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
|
| 3461 | struct rb_event_info *info,
|
| 3462 | unsigned long tail)
|
| 3463 | {
|
| 3464 | struct ring_buffer_event *event;
|
| 3465 | struct buffer_data_page *bpage;
|
| 3466 | u64 ts, delta;
|
| 3467 | bool full = false;
|
| 3468 | int e;
|
| 3469 |
|
| 3470 | bpage = info->tail_page->page;
|
| 3471 |
|
| 3472 | if (tail == CHECK_FULL_PAGE) {
|
| 3473 | full = true;
|
| 3474 | tail = local_read(&bpage->commit);
|
| 3475 | } else if (info->add_timestamp &
|
| 3476 | (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
|
| 3477 |
|
| 3478 | return;
|
| 3479 | }
|
| 3480 |
|
| 3481 | |
| 3482 | * Do not check the first event (skip possible extends too).
|
| 3483 | * Also do not check if previous events have not been committed.
|
| 3484 |
|
| 3485 | if (tail <= 8 || tail > local_read(&bpage->commit))
|
| 3486 | return;
|
| 3487 |
|
| 3488 | |
| 3489 | * If this interrupted another event,
|
| 3490 |
|
| 3491 | if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
|
| 3492 | goto out;
|
| 3493 |
|
| 3494 | ts = bpage->time_stamp;
|
| 3495 |
|
| 3496 | for (e = 0; e < tail; e += rb_event_length(event)) {
|
| 3497 |
|
| 3498 | event = (struct ring_buffer_event *)(bpage->data + e);
|
| 3499 |
|
| 3500 | switch (event->type_len) {
|
| 3501 |
|
| 3502 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 3503 | delta = rb_event_time_stamp(event);
|
| 3504 | ts += delta;
|
| 3505 | break;
|
| 3506 |
|
| 3507 | case RINGBUF_TYPE_TIME_STAMP:
|
| 3508 | delta = rb_event_time_stamp(event);
|
| 3509 | ts = rb_fix_abs_ts(delta, ts);
|
| 3510 | break;
|
| 3511 |
|
| 3512 | case RINGBUF_TYPE_PADDING:
|
| 3513 | if (event->time_delta == 1)
|
| 3514 | break;
|
| 3515 | fallthrough;
|
| 3516 | case RINGBUF_TYPE_DATA:
|
| 3517 | ts += event->time_delta;
|
| 3518 | break;
|
| 3519 |
|
| 3520 | default:
|
| 3521 | RB_WARN_ON(cpu_buffer, 1);
|
| 3522 | }
|
| 3523 | }
|
| 3524 | if ((full && ts > info->ts) ||
|
| 3525 | (!full && ts + info->delta != info->ts)) {
|
| 3526 |
|
| 3527 | if (atomic_inc_return(&ts_dump) != 1) {
|
| 3528 | atomic_dec(&ts_dump);
|
| 3529 | goto out;
|
| 3530 | }
|
| 3531 | atomic_inc(&cpu_buffer->record_disabled);
|
| 3532 |
|
| 3533 | WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
|
| 3534 | pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
|
| 3535 | cpu_buffer->cpu,
|
| 3536 | ts + info->delta, info->ts, info->delta,
|
| 3537 | info->before, info->after,
|
| 3538 | full ? " (full)" : "");
|
| 3539 | dump_buffer_page(bpage, info, tail);
|
| 3540 | atomic_dec(&ts_dump);
|
| 3541 |
|
| 3542 | return;
|
| 3543 | }
|
| 3544 | out:
|
| 3545 | atomic_dec(this_cpu_ptr(&checking));
|
| 3546 | }
|
| 3547 | #else
|
| 3548 | static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
|
| 3549 | struct rb_event_info *info,
|
| 3550 | unsigned long tail)
|
| 3551 | {
|
| 3552 | }
|
| 3553 | #endif
|
| 3554 |
|
| 3555 | static struct ring_buffer_event *
|
| 3556 | __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
|
| 3557 | struct rb_event_info *info)
|
| 3558 | {
|
| 3559 | struct ring_buffer_event *event;
|
| 3560 | struct buffer_page *tail_page;
|
| 3561 | unsigned long tail, write, w;
|
| 3562 | bool a_ok;
|
| 3563 | bool b_ok;
|
| 3564 |
|
| 3565 |
|
| 3566 | tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
|
| 3567 |
|
| 3568 | w = local_read(&tail_page->write) & RB_WRITE_MASK;
|
| 3569 | barrier();
|
| 3570 | b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
|
| 3571 | a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
|
| 3572 | barrier();
|
| 3573 | info->ts = rb_time_stamp(cpu_buffer->buffer);
|
| 3574 |
|
| 3575 | if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
|
| 3576 | info->delta = info->ts;
|
| 3577 | } else {
|
| 3578 | |
| 3579 | * If interrupting an event time update, we may need an
|
| 3580 | * absolute timestamp.
|
| 3581 | * Don't bother if this is the start of a new page (w == 0).
|
| 3582 |
|
| 3583 | if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
|
| 3584 | info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
|
| 3585 | info->length += RB_LEN_TIME_EXTEND;
|
| 3586 | } else {
|
| 3587 | info->delta = info->ts - info->after;
|
| 3588 | if (unlikely(test_time_stamp(info->delta))) {
|
| 3589 | info->add_timestamp |= RB_ADD_STAMP_EXTEND;
|
| 3590 | info->length += RB_LEN_TIME_EXTEND;
|
| 3591 | }
|
| 3592 | }
|
| 3593 | }
|
| 3594 |
|
| 3595 | rb_time_set(&cpu_buffer->before_stamp, info->ts);
|
| 3596 |
|
| 3597 | write = local_add_return(info->length, &tail_page->write);
|
| 3598 |
|
| 3599 |
|
| 3600 | write &= RB_WRITE_MASK;
|
| 3601 |
|
| 3602 | tail = write - info->length;
|
| 3603 |
|
| 3604 |
|
| 3605 | if (unlikely(write > BUF_PAGE_SIZE)) {
|
| 3606 |
|
| 3607 | b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
|
| 3608 | a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
|
| 3609 | if (a_ok && b_ok && info->before != info->after)
|
| 3610 | (void)rb_time_cmpxchg(&cpu_buffer->before_stamp,
|
| 3611 | info->before, info->after);
|
| 3612 | if (a_ok && b_ok)
|
| 3613 | check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
|
| 3614 | return rb_move_tail(cpu_buffer, tail, info);
|
| 3615 | }
|
| 3616 |
|
| 3617 | if (likely(tail == w)) {
|
| 3618 | u64 save_before;
|
| 3619 | bool s_ok;
|
| 3620 |
|
| 3621 |
|
| 3622 | rb_time_set(&cpu_buffer->write_stamp, info->ts);
|
| 3623 | barrier();
|
| 3624 | s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
|
| 3625 | RB_WARN_ON(cpu_buffer, !s_ok);
|
| 3626 | if (likely(!(info->add_timestamp &
|
| 3627 | (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
|
| 3628 |
|
| 3629 | info->delta = info->ts - info->after;
|
| 3630 | else
|
| 3631 |
|
| 3632 | info->delta = info->ts;
|
| 3633 | barrier();
|
| 3634 | check_buffer(cpu_buffer, info, tail);
|
| 3635 | if (unlikely(info->ts != save_before)) {
|
| 3636 |
|
| 3637 |
|
| 3638 | a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
|
| 3639 | RB_WARN_ON(cpu_buffer, !a_ok);
|
| 3640 |
|
| 3641 |
|
| 3642 | if (save_before > info->after) {
|
| 3643 | |
| 3644 | * We do not care about the result, only that
|
| 3645 | * it gets updated atomically.
|
| 3646 |
|
| 3647 | (void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
|
| 3648 | info->after, save_before);
|
| 3649 | }
|
| 3650 | }
|
| 3651 | } else {
|
| 3652 | u64 ts;
|
| 3653 |
|
| 3654 | a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
|
| 3655 |
|
| 3656 | RB_WARN_ON(cpu_buffer, !a_ok);
|
| 3657 | ts = rb_time_stamp(cpu_buffer->buffer);
|
| 3658 | barrier();
|
| 3659 | if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
|
| 3660 | info->after < ts &&
|
| 3661 | rb_time_cmpxchg(&cpu_buffer->write_stamp,
|
| 3662 | info->after, ts)) {
|
| 3663 |
|
| 3664 | info->delta = ts - info->after;
|
| 3665 | } else {
|
| 3666 | |
| 3667 | * Interrupted between C and E:
|
| 3668 | * Lost the previous events time stamp. Just set the
|
| 3669 | * delta to zero, and this will be the same time as
|
| 3670 | * the event this event interrupted. And the events that
|
| 3671 | * came after this will still be correct (as they would
|
| 3672 | * have built their delta on the previous event.
|
| 3673 |
|
| 3674 | info->delta = 0;
|
| 3675 | }
|
| 3676 | info->ts = ts;
|
| 3677 | info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
|
| 3678 | }
|
| 3679 |
|
| 3680 | |
| 3681 | * If this is the first commit on the page, then it has the same
|
| 3682 | * timestamp as the page itself.
|
| 3683 |
|
| 3684 | if (unlikely(!tail && !(info->add_timestamp &
|
| 3685 | (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
|
| 3686 | info->delta = 0;
|
| 3687 |
|
| 3688 |
|
| 3689 |
|
| 3690 | event = __rb_page_index(tail_page, tail);
|
| 3691 | rb_update_event(cpu_buffer, event, info);
|
| 3692 |
|
| 3693 | local_inc(&tail_page->entries);
|
| 3694 |
|
| 3695 | |
| 3696 | * If this is the first commit on the page, then update
|
| 3697 | * its timestamp.
|
| 3698 |
|
| 3699 | if (unlikely(!tail))
|
| 3700 | tail_page->page->time_stamp = info->ts;
|
| 3701 |
|
| 3702 |
|
| 3703 | local_add(info->length, &cpu_buffer->entries_bytes);
|
| 3704 |
|
| 3705 | return event;
|
| 3706 | }
|
| 3707 |
|
| 3708 | static __always_inline struct ring_buffer_event *
|
| 3709 | rb_reserve_next_event(struct trace_buffer *buffer,
|
| 3710 | struct ring_buffer_per_cpu *cpu_buffer,
|
| 3711 | unsigned long length)
|
| 3712 | {
|
| 3713 | struct ring_buffer_event *event;
|
| 3714 | struct rb_event_info info;
|
| 3715 | int nr_loops = 0;
|
| 3716 | int add_ts_default;
|
| 3717 |
|
| 3718 | rb_start_commit(cpu_buffer);
|
| 3719 |
|
| 3720 |
|
| 3721 | #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
|
| 3722 | |
| 3723 | * Due to the ability to swap a cpu buffer from a buffer
|
| 3724 | * it is possible it was swapped before we committed.
|
| 3725 | * (committing stops a swap). We check for it here and
|
| 3726 | * if it happened, we have to fail the write.
|
| 3727 |
|
| 3728 | barrier();
|
| 3729 | if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
|
| 3730 | local_dec(&cpu_buffer->committing);
|
| 3731 | local_dec(&cpu_buffer->commits);
|
| 3732 | return NULL;
|
| 3733 | }
|
| 3734 | #endif
|
| 3735 |
|
| 3736 | info.length = rb_calculate_event_length(length);
|
| 3737 |
|
| 3738 | if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
|
| 3739 | add_ts_default = RB_ADD_STAMP_ABSOLUTE;
|
| 3740 | info.length += RB_LEN_TIME_EXTEND;
|
| 3741 | } else {
|
| 3742 | add_ts_default = RB_ADD_STAMP_NONE;
|
| 3743 | }
|
| 3744 |
|
| 3745 | again:
|
| 3746 | info.add_timestamp = add_ts_default;
|
| 3747 | info.delta = 0;
|
| 3748 |
|
| 3749 | |
| 3750 | * We allow for interrupts to reenter here and do a trace.
|
| 3751 | * If one does, it will cause this original code to loop
|
| 3752 | * back here. Even with heavy interrupts happening, this
|
| 3753 | * should only happen a few times in a row. If this happens
|
| 3754 | * 1000 times in a row, there must be either an interrupt
|
| 3755 | * storm or we have something buggy.
|
| 3756 | * Bail!
|
| 3757 |
|
| 3758 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
|
| 3759 | goto out_fail;
|
| 3760 |
|
| 3761 | event = __rb_reserve_next(cpu_buffer, &info);
|
| 3762 |
|
| 3763 | if (unlikely(PTR_ERR(event) == -EAGAIN)) {
|
| 3764 | if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
|
| 3765 | info.length -= RB_LEN_TIME_EXTEND;
|
| 3766 | goto again;
|
| 3767 | }
|
| 3768 |
|
| 3769 | if (likely(event))
|
| 3770 | return event;
|
| 3771 | out_fail:
|
| 3772 | rb_end_commit(cpu_buffer);
|
| 3773 | return NULL;
|
| 3774 | }
|
| 3775 |
|
| 3776 | |
| 3777 | * ring_buffer_lock_reserve - reserve a part of the buffer
|
| 3778 | * @buffer: the ring buffer to reserve from
|
| 3779 | * @length: the length of the data to reserve (excluding event header)
|
| 3780 | *
|
| 3781 | * Returns a reserved event on the ring buffer to copy directly to.
|
| 3782 | * The user of this interface will need to get the body to write into
|
| 3783 | * and can use the ring_buffer_event_data() interface.
|
| 3784 | *
|
| 3785 | * The length is the length of the data needed, not the event length
|
| 3786 | * which also includes the event header.
|
| 3787 | *
|
| 3788 | * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
|
| 3789 | * If NULL is returned, then nothing has been allocated or locked.
|
| 3790 |
|
| 3791 | struct ring_buffer_event *
|
| 3792 | ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
|
| 3793 | {
|
| 3794 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3795 | struct ring_buffer_event *event;
|
| 3796 | int cpu;
|
| 3797 |
|
| 3798 |
|
| 3799 | preempt_disable_notrace();
|
| 3800 |
|
| 3801 | if (unlikely(atomic_read(&buffer->record_disabled)))
|
| 3802 | goto out;
|
| 3803 |
|
| 3804 | cpu = raw_smp_processor_id();
|
| 3805 |
|
| 3806 | if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
|
| 3807 | goto out;
|
| 3808 |
|
| 3809 | cpu_buffer = buffer->buffers[cpu];
|
| 3810 |
|
| 3811 | if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
|
| 3812 | goto out;
|
| 3813 |
|
| 3814 | if (unlikely(length > BUF_MAX_DATA_SIZE))
|
| 3815 | goto out;
|
| 3816 |
|
| 3817 | if (unlikely(trace_recursive_lock(cpu_buffer)))
|
| 3818 | goto out;
|
| 3819 |
|
| 3820 | event = rb_reserve_next_event(buffer, cpu_buffer, length);
|
| 3821 | if (!event)
|
| 3822 | goto out_unlock;
|
| 3823 |
|
| 3824 | return event;
|
| 3825 |
|
| 3826 | out_unlock:
|
| 3827 | trace_recursive_unlock(cpu_buffer);
|
| 3828 | out:
|
| 3829 | preempt_enable_notrace();
|
| 3830 | return NULL;
|
| 3831 | }
|
| 3832 | EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
|
| 3833 |
|
| 3834 | |
| 3835 | * Decrement the entries to the page that an event is on.
|
| 3836 | * The event does not even need to exist, only the pointer
|
| 3837 | * to the page it is on. This may only be called before the commit
|
| 3838 | * takes place.
|
| 3839 |
|
| 3840 | static inline void
|
| 3841 | rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
|
| 3842 | struct ring_buffer_event *event)
|
| 3843 | {
|
| 3844 | unsigned long addr = (unsigned long)event;
|
| 3845 | struct buffer_page *bpage = cpu_buffer->commit_page;
|
| 3846 | struct buffer_page *start;
|
| 3847 |
|
| 3848 | addr &= PAGE_MASK;
|
| 3849 |
|
| 3850 |
|
| 3851 | if (likely(bpage->page == (void *)addr)) {
|
| 3852 | local_dec(&bpage->entries);
|
| 3853 | return;
|
| 3854 | }
|
| 3855 |
|
| 3856 | |
| 3857 | * Because the commit page may be on the reader page we
|
| 3858 | * start with the next page and check the end loop there.
|
| 3859 |
|
| 3860 | rb_inc_page(&bpage);
|
| 3861 | start = bpage;
|
| 3862 | do {
|
| 3863 | if (bpage->page == (void *)addr) {
|
| 3864 | local_dec(&bpage->entries);
|
| 3865 | return;
|
| 3866 | }
|
| 3867 | rb_inc_page(&bpage);
|
| 3868 | } while (bpage != start);
|
| 3869 |
|
| 3870 |
|
| 3871 | RB_WARN_ON(cpu_buffer, 1);
|
| 3872 | }
|
| 3873 |
|
| 3874 | |
| 3875 | * ring_buffer_discard_commit - discard an event that has not been committed
|
| 3876 | * @buffer: the ring buffer
|
| 3877 | * @event: non committed event to discard
|
| 3878 | *
|
| 3879 | * Sometimes an event that is in the ring buffer needs to be ignored.
|
| 3880 | * This function lets the user discard an event in the ring buffer
|
| 3881 | * and then that event will not be read later.
|
| 3882 | *
|
| 3883 | * This function only works if it is called before the item has been
|
| 3884 | * committed. It will try to free the event from the ring buffer
|
| 3885 | * if another event has not been added behind it.
|
| 3886 | *
|
| 3887 | * If another event has been added behind it, it will set the event
|
| 3888 | * up as discarded, and perform the commit.
|
| 3889 | *
|
| 3890 | * If this function is called, do not call ring_buffer_unlock_commit on
|
| 3891 | * the event.
|
| 3892 |
|
| 3893 | void ring_buffer_discard_commit(struct trace_buffer *buffer,
|
| 3894 | struct ring_buffer_event *event)
|
| 3895 | {
|
| 3896 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3897 | int cpu;
|
| 3898 |
|
| 3899 |
|
| 3900 | rb_event_discard(event);
|
| 3901 |
|
| 3902 | cpu = smp_processor_id();
|
| 3903 | cpu_buffer = buffer->buffers[cpu];
|
| 3904 |
|
| 3905 | |
| 3906 | * This must only be called if the event has not been
|
| 3907 | * committed yet. Thus we can assume that preemption
|
| 3908 | * is still disabled.
|
| 3909 |
|
| 3910 | RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
|
| 3911 |
|
| 3912 | rb_decrement_entry(cpu_buffer, event);
|
| 3913 | if (rb_try_to_discard(cpu_buffer, event))
|
| 3914 | goto out;
|
| 3915 |
|
| 3916 | out:
|
| 3917 | rb_end_commit(cpu_buffer);
|
| 3918 |
|
| 3919 | trace_recursive_unlock(cpu_buffer);
|
| 3920 |
|
| 3921 | preempt_enable_notrace();
|
| 3922 |
|
| 3923 | }
|
| 3924 | EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
|
| 3925 |
|
| 3926 | |
| 3927 | * ring_buffer_write - write data to the buffer without reserving
|
| 3928 | * @buffer: The ring buffer to write to.
|
| 3929 | * @length: The length of the data being written (excluding the event header)
|
| 3930 | * @data: The data to write to the buffer.
|
| 3931 | *
|
| 3932 | * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
|
| 3933 | * one function. If you already have the data to write to the buffer, it
|
| 3934 | * may be easier to simply call this function.
|
| 3935 | *
|
| 3936 | * Note, like ring_buffer_lock_reserve, the length is the length of the data
|
| 3937 | * and not the length of the event which would hold the header.
|
| 3938 |
|
| 3939 | int ring_buffer_write(struct trace_buffer *buffer,
|
| 3940 | unsigned long length,
|
| 3941 | void *data)
|
| 3942 | {
|
| 3943 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 3944 | struct ring_buffer_event *event;
|
| 3945 | void *body;
|
| 3946 | int ret = -EBUSY;
|
| 3947 | int cpu;
|
| 3948 |
|
| 3949 | preempt_disable_notrace();
|
| 3950 |
|
| 3951 | if (atomic_read(&buffer->record_disabled))
|
| 3952 | goto out;
|
| 3953 |
|
| 3954 | cpu = raw_smp_processor_id();
|
| 3955 |
|
| 3956 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 3957 | goto out;
|
| 3958 |
|
| 3959 | cpu_buffer = buffer->buffers[cpu];
|
| 3960 |
|
| 3961 | if (atomic_read(&cpu_buffer->record_disabled))
|
| 3962 | goto out;
|
| 3963 |
|
| 3964 | if (length > BUF_MAX_DATA_SIZE)
|
| 3965 | goto out;
|
| 3966 |
|
| 3967 | if (unlikely(trace_recursive_lock(cpu_buffer)))
|
| 3968 | goto out;
|
| 3969 |
|
| 3970 | event = rb_reserve_next_event(buffer, cpu_buffer, length);
|
| 3971 | if (!event)
|
| 3972 | goto out_unlock;
|
| 3973 |
|
| 3974 | body = rb_event_data(event);
|
| 3975 |
|
| 3976 | memcpy(body, data, length);
|
| 3977 |
|
| 3978 | rb_commit(cpu_buffer);
|
| 3979 |
|
| 3980 | rb_wakeups(buffer, cpu_buffer);
|
| 3981 |
|
| 3982 | ret = 0;
|
| 3983 |
|
| 3984 | out_unlock:
|
| 3985 | trace_recursive_unlock(cpu_buffer);
|
| 3986 |
|
| 3987 | out:
|
| 3988 | preempt_enable_notrace();
|
| 3989 |
|
| 3990 | return ret;
|
| 3991 | }
|
| 3992 | EXPORT_SYMBOL_GPL(ring_buffer_write);
|
| 3993 |
|
| 3994 | static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
|
| 3995 | {
|
| 3996 | struct buffer_page *reader = cpu_buffer->reader_page;
|
| 3997 | struct buffer_page *head = rb_set_head_page(cpu_buffer);
|
| 3998 | struct buffer_page *commit = cpu_buffer->commit_page;
|
| 3999 |
|
| 4000 |
|
| 4001 | if (unlikely(!head))
|
| 4002 | return true;
|
| 4003 |
|
| 4004 |
|
| 4005 | if (reader->read != rb_page_commit(reader))
|
| 4006 | return false;
|
| 4007 |
|
| 4008 | |
| 4009 | * If writers are committing on the reader page, knowing all
|
| 4010 | * committed content has been read, the ring buffer is empty.
|
| 4011 |
|
| 4012 | if (commit == reader)
|
| 4013 | return true;
|
| 4014 |
|
| 4015 | |
| 4016 | * If writers are committing on a page other than reader page
|
| 4017 | * and head page, there should always be content to read.
|
| 4018 |
|
| 4019 | if (commit != head)
|
| 4020 | return false;
|
| 4021 |
|
| 4022 | |
| 4023 | * Writers are committing on the head page, we just need
|
| 4024 | * to care about there're committed data, and the reader will
|
| 4025 | * swap reader page with head page when it is to read data.
|
| 4026 |
|
| 4027 | return rb_page_commit(commit) == 0;
|
| 4028 | }
|
| 4029 |
|
| 4030 | |
| 4031 | * ring_buffer_record_disable - stop all writes into the buffer
|
| 4032 | * @buffer: The ring buffer to stop writes to.
|
| 4033 | *
|
| 4034 | * This prevents all writes to the buffer. Any attempt to write
|
| 4035 | * to the buffer after this will fail and return NULL.
|
| 4036 | *
|
| 4037 | * The caller should call synchronize_rcu() after this.
|
| 4038 |
|
| 4039 | void ring_buffer_record_disable(struct trace_buffer *buffer)
|
| 4040 | {
|
| 4041 | atomic_inc(&buffer->record_disabled);
|
| 4042 | }
|
| 4043 | EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
|
| 4044 |
|
| 4045 | |
| 4046 | * ring_buffer_record_enable - enable writes to the buffer
|
| 4047 | * @buffer: The ring buffer to enable writes
|
| 4048 | *
|
| 4049 | * Note, multiple disables will need the same number of enables
|
| 4050 | * to truly enable the writing (much like preempt_disable).
|
| 4051 |
|
| 4052 | void ring_buffer_record_enable(struct trace_buffer *buffer)
|
| 4053 | {
|
| 4054 | atomic_dec(&buffer->record_disabled);
|
| 4055 | }
|
| 4056 | EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
|
| 4057 |
|
| 4058 | |
| 4059 | * ring_buffer_record_off - stop all writes into the buffer
|
| 4060 | * @buffer: The ring buffer to stop writes to.
|
| 4061 | *
|
| 4062 | * This prevents all writes to the buffer. Any attempt to write
|
| 4063 | * to the buffer after this will fail and return NULL.
|
| 4064 | *
|
| 4065 | * This is different than ring_buffer_record_disable() as
|
| 4066 | * it works like an on/off switch, where as the disable() version
|
| 4067 | * must be paired with a enable().
|
| 4068 |
|
| 4069 | void ring_buffer_record_off(struct trace_buffer *buffer)
|
| 4070 | {
|
| 4071 | unsigned int rd;
|
| 4072 | unsigned int new_rd;
|
| 4073 |
|
| 4074 | rd = atomic_read(&buffer->record_disabled);
|
| 4075 | do {
|
| 4076 | new_rd = rd | RB_BUFFER_OFF;
|
| 4077 | } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
|
| 4078 | }
|
| 4079 | EXPORT_SYMBOL_GPL(ring_buffer_record_off);
|
| 4080 |
|
| 4081 | |
| 4082 | * ring_buffer_record_on - restart writes into the buffer
|
| 4083 | * @buffer: The ring buffer to start writes to.
|
| 4084 | *
|
| 4085 | * This enables all writes to the buffer that was disabled by
|
| 4086 | * ring_buffer_record_off().
|
| 4087 | *
|
| 4088 | * This is different than ring_buffer_record_enable() as
|
| 4089 | * it works like an on/off switch, where as the enable() version
|
| 4090 | * must be paired with a disable().
|
| 4091 |
|
| 4092 | void ring_buffer_record_on(struct trace_buffer *buffer)
|
| 4093 | {
|
| 4094 | unsigned int rd;
|
| 4095 | unsigned int new_rd;
|
| 4096 |
|
| 4097 | rd = atomic_read(&buffer->record_disabled);
|
| 4098 | do {
|
| 4099 | new_rd = rd & ~RB_BUFFER_OFF;
|
| 4100 | } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
|
| 4101 | }
|
| 4102 | EXPORT_SYMBOL_GPL(ring_buffer_record_on);
|
| 4103 |
|
| 4104 | |
| 4105 | * ring_buffer_record_is_on - return true if the ring buffer can write
|
| 4106 | * @buffer: The ring buffer to see if write is enabled
|
| 4107 | *
|
| 4108 | * Returns true if the ring buffer is in a state that it accepts writes.
|
| 4109 |
|
| 4110 | bool ring_buffer_record_is_on(struct trace_buffer *buffer)
|
| 4111 | {
|
| 4112 | return !atomic_read(&buffer->record_disabled);
|
| 4113 | }
|
| 4114 |
|
| 4115 | |
| 4116 | * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
|
| 4117 | * @buffer: The ring buffer to see if write is set enabled
|
| 4118 | *
|
| 4119 | * Returns true if the ring buffer is set writable by ring_buffer_record_on().
|
| 4120 | * Note that this does NOT mean it is in a writable state.
|
| 4121 | *
|
| 4122 | * It may return true when the ring buffer has been disabled by
|
| 4123 | * ring_buffer_record_disable(), as that is a temporary disabling of
|
| 4124 | * the ring buffer.
|
| 4125 |
|
| 4126 | bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
|
| 4127 | {
|
| 4128 | return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
|
| 4129 | }
|
| 4130 |
|
| 4131 | |
| 4132 | * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
|
| 4133 | * @buffer: The ring buffer to stop writes to.
|
| 4134 | * @cpu: The CPU buffer to stop
|
| 4135 | *
|
| 4136 | * This prevents all writes to the buffer. Any attempt to write
|
| 4137 | * to the buffer after this will fail and return NULL.
|
| 4138 | *
|
| 4139 | * The caller should call synchronize_rcu() after this.
|
| 4140 |
|
| 4141 | void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
|
| 4142 | {
|
| 4143 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4144 |
|
| 4145 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4146 | return;
|
| 4147 |
|
| 4148 | cpu_buffer = buffer->buffers[cpu];
|
| 4149 | atomic_inc(&cpu_buffer->record_disabled);
|
| 4150 | }
|
| 4151 | EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
|
| 4152 |
|
| 4153 | |
| 4154 | * ring_buffer_record_enable_cpu - enable writes to the buffer
|
| 4155 | * @buffer: The ring buffer to enable writes
|
| 4156 | * @cpu: The CPU to enable.
|
| 4157 | *
|
| 4158 | * Note, multiple disables will need the same number of enables
|
| 4159 | * to truly enable the writing (much like preempt_disable).
|
| 4160 |
|
| 4161 | void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
|
| 4162 | {
|
| 4163 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4164 |
|
| 4165 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4166 | return;
|
| 4167 |
|
| 4168 | cpu_buffer = buffer->buffers[cpu];
|
| 4169 | atomic_dec(&cpu_buffer->record_disabled);
|
| 4170 | }
|
| 4171 | EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
|
| 4172 |
|
| 4173 | |
| 4174 | * The total entries in the ring buffer is the running counter
|
| 4175 | * of entries entered into the ring buffer, minus the sum of
|
| 4176 | * the entries read from the ring buffer and the number of
|
| 4177 | * entries that were overwritten.
|
| 4178 |
|
| 4179 | static inline unsigned long
|
| 4180 | rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
|
| 4181 | {
|
| 4182 | return local_read(&cpu_buffer->entries) -
|
| 4183 | (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
|
| 4184 | }
|
| 4185 |
|
| 4186 | |
| 4187 | * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
|
| 4188 | * @buffer: The ring buffer
|
| 4189 | * @cpu: The per CPU buffer to read from.
|
| 4190 |
|
| 4191 | u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
|
| 4192 | {
|
| 4193 | unsigned long flags;
|
| 4194 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4195 | struct buffer_page *bpage;
|
| 4196 | u64 ret = 0;
|
| 4197 |
|
| 4198 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4199 | return 0;
|
| 4200 |
|
| 4201 | cpu_buffer = buffer->buffers[cpu];
|
| 4202 | raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 4203 | |
| 4204 | * if the tail is on reader_page, oldest time stamp is on the reader
|
| 4205 | * page
|
| 4206 |
|
| 4207 | if (cpu_buffer->tail_page == cpu_buffer->reader_page)
|
| 4208 | bpage = cpu_buffer->reader_page;
|
| 4209 | else
|
| 4210 | bpage = rb_set_head_page(cpu_buffer);
|
| 4211 | if (bpage)
|
| 4212 | ret = bpage->page->time_stamp;
|
| 4213 | raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 4214 |
|
| 4215 | return ret;
|
| 4216 | }
|
| 4217 | EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
|
| 4218 |
|
| 4219 | |
| 4220 | * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
|
| 4221 | * @buffer: The ring buffer
|
| 4222 | * @cpu: The per CPU buffer to read from.
|
| 4223 |
|
| 4224 | unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
|
| 4225 | {
|
| 4226 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4227 | unsigned long ret;
|
| 4228 |
|
| 4229 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4230 | return 0;
|
| 4231 |
|
| 4232 | cpu_buffer = buffer->buffers[cpu];
|
| 4233 | ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
|
| 4234 |
|
| 4235 | return ret;
|
| 4236 | }
|
| 4237 | EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
|
| 4238 |
|
| 4239 | |
| 4240 | * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
|
| 4241 | * @buffer: The ring buffer
|
| 4242 | * @cpu: The per CPU buffer to get the entries from.
|
| 4243 |
|
| 4244 | unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
|
| 4245 | {
|
| 4246 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4247 |
|
| 4248 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4249 | return 0;
|
| 4250 |
|
| 4251 | cpu_buffer = buffer->buffers[cpu];
|
| 4252 |
|
| 4253 | return rb_num_of_entries(cpu_buffer);
|
| 4254 | }
|
| 4255 | EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
|
| 4256 |
|
| 4257 | |
| 4258 | * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
|
| 4259 | * buffer wrapping around (only if RB_FL_OVERWRITE is on).
|
| 4260 | * @buffer: The ring buffer
|
| 4261 | * @cpu: The per CPU buffer to get the number of overruns from
|
| 4262 |
|
| 4263 | unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
|
| 4264 | {
|
| 4265 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4266 | unsigned long ret;
|
| 4267 |
|
| 4268 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4269 | return 0;
|
| 4270 |
|
| 4271 | cpu_buffer = buffer->buffers[cpu];
|
| 4272 | ret = local_read(&cpu_buffer->overrun);
|
| 4273 |
|
| 4274 | return ret;
|
| 4275 | }
|
| 4276 | EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
|
| 4277 |
|
| 4278 | |
| 4279 | * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
|
| 4280 | * commits failing due to the buffer wrapping around while there are uncommitted
|
| 4281 | * events, such as during an interrupt storm.
|
| 4282 | * @buffer: The ring buffer
|
| 4283 | * @cpu: The per CPU buffer to get the number of overruns from
|
| 4284 |
|
| 4285 | unsigned long
|
| 4286 | ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
|
| 4287 | {
|
| 4288 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4289 | unsigned long ret;
|
| 4290 |
|
| 4291 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4292 | return 0;
|
| 4293 |
|
| 4294 | cpu_buffer = buffer->buffers[cpu];
|
| 4295 | ret = local_read(&cpu_buffer->commit_overrun);
|
| 4296 |
|
| 4297 | return ret;
|
| 4298 | }
|
| 4299 | EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
|
| 4300 |
|
| 4301 | |
| 4302 | * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
|
| 4303 | * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
|
| 4304 | * @buffer: The ring buffer
|
| 4305 | * @cpu: The per CPU buffer to get the number of overruns from
|
| 4306 |
|
| 4307 | unsigned long
|
| 4308 | ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
|
| 4309 | {
|
| 4310 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4311 | unsigned long ret;
|
| 4312 |
|
| 4313 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4314 | return 0;
|
| 4315 |
|
| 4316 | cpu_buffer = buffer->buffers[cpu];
|
| 4317 | ret = local_read(&cpu_buffer->dropped_events);
|
| 4318 |
|
| 4319 | return ret;
|
| 4320 | }
|
| 4321 | EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
|
| 4322 |
|
| 4323 | |
| 4324 | * ring_buffer_read_events_cpu - get the number of events successfully read
|
| 4325 | * @buffer: The ring buffer
|
| 4326 | * @cpu: The per CPU buffer to get the number of events read
|
| 4327 |
|
| 4328 | unsigned long
|
| 4329 | ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
|
| 4330 | {
|
| 4331 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4332 |
|
| 4333 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4334 | return 0;
|
| 4335 |
|
| 4336 | cpu_buffer = buffer->buffers[cpu];
|
| 4337 | return cpu_buffer->read;
|
| 4338 | }
|
| 4339 | EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
|
| 4340 |
|
| 4341 | |
| 4342 | * ring_buffer_entries - get the number of entries in a buffer
|
| 4343 | * @buffer: The ring buffer
|
| 4344 | *
|
| 4345 | * Returns the total number of entries in the ring buffer
|
| 4346 | * (all CPU entries)
|
| 4347 |
|
| 4348 | unsigned long ring_buffer_entries(struct trace_buffer *buffer)
|
| 4349 | {
|
| 4350 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4351 | unsigned long entries = 0;
|
| 4352 | int cpu;
|
| 4353 |
|
| 4354 |
|
| 4355 | for_each_buffer_cpu(buffer, cpu) {
|
| 4356 | cpu_buffer = buffer->buffers[cpu];
|
| 4357 | entries += rb_num_of_entries(cpu_buffer);
|
| 4358 | }
|
| 4359 |
|
| 4360 | return entries;
|
| 4361 | }
|
| 4362 | EXPORT_SYMBOL_GPL(ring_buffer_entries);
|
| 4363 |
|
| 4364 | |
| 4365 | * ring_buffer_overruns - get the number of overruns in buffer
|
| 4366 | * @buffer: The ring buffer
|
| 4367 | *
|
| 4368 | * Returns the total number of overruns in the ring buffer
|
| 4369 | * (all CPU entries)
|
| 4370 |
|
| 4371 | unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
|
| 4372 | {
|
| 4373 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4374 | unsigned long overruns = 0;
|
| 4375 | int cpu;
|
| 4376 |
|
| 4377 |
|
| 4378 | for_each_buffer_cpu(buffer, cpu) {
|
| 4379 | cpu_buffer = buffer->buffers[cpu];
|
| 4380 | overruns += local_read(&cpu_buffer->overrun);
|
| 4381 | }
|
| 4382 |
|
| 4383 | return overruns;
|
| 4384 | }
|
| 4385 | EXPORT_SYMBOL_GPL(ring_buffer_overruns);
|
| 4386 |
|
| 4387 | static void rb_iter_reset(struct ring_buffer_iter *iter)
|
| 4388 | {
|
| 4389 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
|
| 4390 |
|
| 4391 |
|
| 4392 | iter->head_page = cpu_buffer->reader_page;
|
| 4393 | iter->head = cpu_buffer->reader_page->read;
|
| 4394 | iter->next_event = iter->head;
|
| 4395 |
|
| 4396 | iter->cache_reader_page = iter->head_page;
|
| 4397 | iter->cache_read = cpu_buffer->read;
|
| 4398 |
|
| 4399 | if (iter->head) {
|
| 4400 | iter->read_stamp = cpu_buffer->read_stamp;
|
| 4401 | iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
|
| 4402 | } else {
|
| 4403 | iter->read_stamp = iter->head_page->page->time_stamp;
|
| 4404 | iter->page_stamp = iter->read_stamp;
|
| 4405 | }
|
| 4406 | }
|
| 4407 |
|
| 4408 | |
| 4409 | * ring_buffer_iter_reset - reset an iterator
|
| 4410 | * @iter: The iterator to reset
|
| 4411 | *
|
| 4412 | * Resets the iterator, so that it will start from the beginning
|
| 4413 | * again.
|
| 4414 |
|
| 4415 | void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
|
| 4416 | {
|
| 4417 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4418 | unsigned long flags;
|
| 4419 |
|
| 4420 | if (!iter)
|
| 4421 | return;
|
| 4422 |
|
| 4423 | cpu_buffer = iter->cpu_buffer;
|
| 4424 |
|
| 4425 | raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 4426 | rb_iter_reset(iter);
|
| 4427 | raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 4428 | }
|
| 4429 | EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
|
| 4430 |
|
| 4431 | |
| 4432 | * ring_buffer_iter_empty - check if an iterator has no more to read
|
| 4433 | * @iter: The iterator to check
|
| 4434 |
|
| 4435 | int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
|
| 4436 | {
|
| 4437 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4438 | struct buffer_page *reader;
|
| 4439 | struct buffer_page *head_page;
|
| 4440 | struct buffer_page *commit_page;
|
| 4441 | struct buffer_page *curr_commit_page;
|
| 4442 | unsigned commit;
|
| 4443 | u64 curr_commit_ts;
|
| 4444 | u64 commit_ts;
|
| 4445 |
|
| 4446 | cpu_buffer = iter->cpu_buffer;
|
| 4447 | reader = cpu_buffer->reader_page;
|
| 4448 | head_page = cpu_buffer->head_page;
|
| 4449 | commit_page = cpu_buffer->commit_page;
|
| 4450 | commit_ts = commit_page->page->time_stamp;
|
| 4451 |
|
| 4452 | |
| 4453 | * When the writer goes across pages, it issues a cmpxchg which
|
| 4454 | * is a mb(), which will synchronize with the rmb here.
|
| 4455 | * (see rb_tail_page_update())
|
| 4456 |
|
| 4457 | smp_rmb();
|
| 4458 | commit = rb_page_commit(commit_page);
|
| 4459 |
|
| 4460 | smp_rmb();
|
| 4461 |
|
| 4462 |
|
| 4463 | curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
|
| 4464 | curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
|
| 4465 |
|
| 4466 |
|
| 4467 | if (curr_commit_page != commit_page ||
|
| 4468 | curr_commit_ts != commit_ts)
|
| 4469 | return 0;
|
| 4470 |
|
| 4471 |
|
| 4472 | return ((iter->head_page == commit_page && iter->head >= commit) ||
|
| 4473 | (iter->head_page == reader && commit_page == head_page &&
|
| 4474 | head_page->read == commit &&
|
| 4475 | iter->head == rb_page_commit(cpu_buffer->reader_page)));
|
| 4476 | }
|
| 4477 | EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
|
| 4478 |
|
| 4479 | static void
|
| 4480 | rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
|
| 4481 | struct ring_buffer_event *event)
|
| 4482 | {
|
| 4483 | u64 delta;
|
| 4484 |
|
| 4485 | switch (event->type_len) {
|
| 4486 | case RINGBUF_TYPE_PADDING:
|
| 4487 | return;
|
| 4488 |
|
| 4489 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 4490 | delta = rb_event_time_stamp(event);
|
| 4491 | cpu_buffer->read_stamp += delta;
|
| 4492 | return;
|
| 4493 |
|
| 4494 | case RINGBUF_TYPE_TIME_STAMP:
|
| 4495 | delta = rb_event_time_stamp(event);
|
| 4496 | delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
|
| 4497 | cpu_buffer->read_stamp = delta;
|
| 4498 | return;
|
| 4499 |
|
| 4500 | case RINGBUF_TYPE_DATA:
|
| 4501 | cpu_buffer->read_stamp += event->time_delta;
|
| 4502 | return;
|
| 4503 |
|
| 4504 | default:
|
| 4505 | RB_WARN_ON(cpu_buffer, 1);
|
| 4506 | }
|
| 4507 | }
|
| 4508 |
|
| 4509 | static void
|
| 4510 | rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
|
| 4511 | struct ring_buffer_event *event)
|
| 4512 | {
|
| 4513 | u64 delta;
|
| 4514 |
|
| 4515 | switch (event->type_len) {
|
| 4516 | case RINGBUF_TYPE_PADDING:
|
| 4517 | return;
|
| 4518 |
|
| 4519 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 4520 | delta = rb_event_time_stamp(event);
|
| 4521 | iter->read_stamp += delta;
|
| 4522 | return;
|
| 4523 |
|
| 4524 | case RINGBUF_TYPE_TIME_STAMP:
|
| 4525 | delta = rb_event_time_stamp(event);
|
| 4526 | delta = rb_fix_abs_ts(delta, iter->read_stamp);
|
| 4527 | iter->read_stamp = delta;
|
| 4528 | return;
|
| 4529 |
|
| 4530 | case RINGBUF_TYPE_DATA:
|
| 4531 | iter->read_stamp += event->time_delta;
|
| 4532 | return;
|
| 4533 |
|
| 4534 | default:
|
| 4535 | RB_WARN_ON(iter->cpu_buffer, 1);
|
| 4536 | }
|
| 4537 | }
|
| 4538 |
|
| 4539 | static struct buffer_page *
|
| 4540 | rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
|
| 4541 | {
|
| 4542 | struct buffer_page *reader = NULL;
|
| 4543 | unsigned long overwrite;
|
| 4544 | unsigned long flags;
|
| 4545 | int nr_loops = 0;
|
| 4546 | bool ret;
|
| 4547 |
|
| 4548 | local_irq_save(flags);
|
| 4549 | arch_spin_lock(&cpu_buffer->lock);
|
| 4550 |
|
| 4551 | again:
|
| 4552 | |
| 4553 | * This should normally only loop twice. But because the
|
| 4554 | * start of the reader inserts an empty page, it causes
|
| 4555 | * a case where we will loop three times. There should be no
|
| 4556 | * reason to loop four times (that I know of).
|
| 4557 |
|
| 4558 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
|
| 4559 | reader = NULL;
|
| 4560 | goto out;
|
| 4561 | }
|
| 4562 |
|
| 4563 | reader = cpu_buffer->reader_page;
|
| 4564 |
|
| 4565 |
|
| 4566 | if (cpu_buffer->reader_page->read < rb_page_size(reader))
|
| 4567 | goto out;
|
| 4568 |
|
| 4569 |
|
| 4570 | if (RB_WARN_ON(cpu_buffer,
|
| 4571 | cpu_buffer->reader_page->read > rb_page_size(reader)))
|
| 4572 | goto out;
|
| 4573 |
|
| 4574 |
|
| 4575 | reader = NULL;
|
| 4576 | if (cpu_buffer->commit_page == cpu_buffer->reader_page)
|
| 4577 | goto out;
|
| 4578 |
|
| 4579 |
|
| 4580 | if (rb_num_of_entries(cpu_buffer) == 0)
|
| 4581 | goto out;
|
| 4582 |
|
| 4583 | |
| 4584 | * Reset the reader page to size zero.
|
| 4585 |
|
| 4586 | local_set(&cpu_buffer->reader_page->write, 0);
|
| 4587 | local_set(&cpu_buffer->reader_page->entries, 0);
|
| 4588 | local_set(&cpu_buffer->reader_page->page->commit, 0);
|
| 4589 | cpu_buffer->reader_page->real_end = 0;
|
| 4590 |
|
| 4591 | spin:
|
| 4592 | |
| 4593 | * Splice the empty reader page into the list around the head.
|
| 4594 |
|
| 4595 | reader = rb_set_head_page(cpu_buffer);
|
| 4596 | if (!reader)
|
| 4597 | goto out;
|
| 4598 | cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
|
| 4599 | cpu_buffer->reader_page->list.prev = reader->list.prev;
|
| 4600 |
|
| 4601 | |
| 4602 | * cpu_buffer->pages just needs to point to the buffer, it
|
| 4603 | * has no specific buffer page to point to. Lets move it out
|
| 4604 | * of our way so we don't accidentally swap it.
|
| 4605 |
|
| 4606 | cpu_buffer->pages = reader->list.prev;
|
| 4607 |
|
| 4608 |
|
| 4609 | rb_set_list_to_head(&cpu_buffer->reader_page->list);
|
| 4610 |
|
| 4611 | |
| 4612 | * We want to make sure we read the overruns after we set up our
|
| 4613 | * pointers to the next object. The writer side does a
|
| 4614 | * cmpxchg to cross pages which acts as the mb on the writer
|
| 4615 | * side. Note, the reader will constantly fail the swap
|
| 4616 | * while the writer is updating the pointers, so this
|
| 4617 | * guarantees that the overwrite recorded here is the one we
|
| 4618 | * want to compare with the last_overrun.
|
| 4619 |
|
| 4620 | smp_mb();
|
| 4621 | overwrite = local_read(&(cpu_buffer->overrun));
|
| 4622 |
|
| 4623 | |
| 4624 | * Here's the tricky part.
|
| 4625 | *
|
| 4626 | * We need to move the pointer past the header page.
|
| 4627 | * But we can only do that if a writer is not currently
|
| 4628 | * moving it. The page before the header page has the
|
| 4629 | * flag bit '1' set if it is pointing to the page we want.
|
| 4630 | * but if the writer is in the process of moving it
|
| 4631 | * than it will be '2' or already moved '0'.
|
| 4632 |
|
| 4633 |
|
| 4634 | ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
|
| 4635 |
|
| 4636 | |
| 4637 | * If we did not convert it, then we must try again.
|
| 4638 |
|
| 4639 | if (!ret)
|
| 4640 | goto spin;
|
| 4641 |
|
| 4642 | |
| 4643 | * Yay! We succeeded in replacing the page.
|
| 4644 | *
|
| 4645 | * Now make the new head point back to the reader page.
|
| 4646 |
|
| 4647 | rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
|
| 4648 | rb_inc_page(&cpu_buffer->head_page);
|
| 4649 |
|
| 4650 | local_inc(&cpu_buffer->pages_read);
|
| 4651 |
|
| 4652 |
|
| 4653 | cpu_buffer->reader_page = reader;
|
| 4654 | cpu_buffer->reader_page->read = 0;
|
| 4655 |
|
| 4656 | if (overwrite != cpu_buffer->last_overrun) {
|
| 4657 | cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
|
| 4658 | cpu_buffer->last_overrun = overwrite;
|
| 4659 | }
|
| 4660 |
|
| 4661 | goto again;
|
| 4662 |
|
| 4663 | out:
|
| 4664 |
|
| 4665 | if (reader && reader->read == 0)
|
| 4666 | cpu_buffer->read_stamp = reader->page->time_stamp;
|
| 4667 |
|
| 4668 | arch_spin_unlock(&cpu_buffer->lock);
|
| 4669 | local_irq_restore(flags);
|
| 4670 |
|
| 4671 | |
| 4672 | * The writer has preempt disable, wait for it. But not forever
|
| 4673 | * Although, 1 second is pretty much "forever"
|
| 4674 |
|
| 4675 | #define USECS_WAIT 1000000
|
| 4676 | for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
|
| 4677 |
|
| 4678 | if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
|
| 4679 | break;
|
| 4680 |
|
| 4681 | udelay(1);
|
| 4682 |
|
| 4683 |
|
| 4684 | smp_rmb();
|
| 4685 | }
|
| 4686 |
|
| 4687 |
|
| 4688 | if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
|
| 4689 | reader = NULL;
|
| 4690 |
|
| 4691 | |
| 4692 | * Make sure we see any padding after the write update
|
| 4693 | * (see rb_reset_tail()).
|
| 4694 | *
|
| 4695 | * In addition, a writer may be writing on the reader page
|
| 4696 | * if the page has not been fully filled, so the read barrier
|
| 4697 | * is also needed to make sure we see the content of what is
|
| 4698 | * committed by the writer (see rb_set_commit_to_write()).
|
| 4699 |
|
| 4700 | smp_rmb();
|
| 4701 |
|
| 4702 |
|
| 4703 | return reader;
|
| 4704 | }
|
| 4705 |
|
| 4706 | static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
|
| 4707 | {
|
| 4708 | struct ring_buffer_event *event;
|
| 4709 | struct buffer_page *reader;
|
| 4710 | unsigned length;
|
| 4711 |
|
| 4712 | reader = rb_get_reader_page(cpu_buffer);
|
| 4713 |
|
| 4714 |
|
| 4715 | if (RB_WARN_ON(cpu_buffer, !reader))
|
| 4716 | return;
|
| 4717 |
|
| 4718 | event = rb_reader_event(cpu_buffer);
|
| 4719 |
|
| 4720 | if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
|
| 4721 | cpu_buffer->read++;
|
| 4722 |
|
| 4723 | rb_update_read_stamp(cpu_buffer, event);
|
| 4724 |
|
| 4725 | length = rb_event_length(event);
|
| 4726 | cpu_buffer->reader_page->read += length;
|
| 4727 | }
|
| 4728 |
|
| 4729 | static void rb_advance_iter(struct ring_buffer_iter *iter)
|
| 4730 | {
|
| 4731 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4732 |
|
| 4733 | cpu_buffer = iter->cpu_buffer;
|
| 4734 |
|
| 4735 |
|
| 4736 | if (iter->head == iter->next_event) {
|
| 4737 |
|
| 4738 | if (rb_iter_head_event(iter) == NULL)
|
| 4739 | return;
|
| 4740 | }
|
| 4741 |
|
| 4742 | iter->head = iter->next_event;
|
| 4743 |
|
| 4744 | |
| 4745 | * Check if we are at the end of the buffer.
|
| 4746 |
|
| 4747 | if (iter->next_event >= rb_page_size(iter->head_page)) {
|
| 4748 |
|
| 4749 | if (iter->head_page == cpu_buffer->commit_page)
|
| 4750 | return;
|
| 4751 | rb_inc_iter(iter);
|
| 4752 | return;
|
| 4753 | }
|
| 4754 |
|
| 4755 | rb_update_iter_read_stamp(iter, iter->event);
|
| 4756 | }
|
| 4757 |
|
| 4758 | static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
|
| 4759 | {
|
| 4760 | return cpu_buffer->lost_events;
|
| 4761 | }
|
| 4762 |
|
| 4763 | static struct ring_buffer_event *
|
| 4764 | rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
|
| 4765 | unsigned long *lost_events)
|
| 4766 | {
|
| 4767 | struct ring_buffer_event *event;
|
| 4768 | struct buffer_page *reader;
|
| 4769 | int nr_loops = 0;
|
| 4770 |
|
| 4771 | if (ts)
|
| 4772 | *ts = 0;
|
| 4773 | again:
|
| 4774 | |
| 4775 | * We repeat when a time extend is encountered.
|
| 4776 | * Since the time extend is always attached to a data event,
|
| 4777 | * we should never loop more than once.
|
| 4778 | * (We never hit the following condition more than twice).
|
| 4779 |
|
| 4780 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
|
| 4781 | return NULL;
|
| 4782 |
|
| 4783 | reader = rb_get_reader_page(cpu_buffer);
|
| 4784 | if (!reader)
|
| 4785 | return NULL;
|
| 4786 |
|
| 4787 | event = rb_reader_event(cpu_buffer);
|
| 4788 |
|
| 4789 | switch (event->type_len) {
|
| 4790 | case RINGBUF_TYPE_PADDING:
|
| 4791 | if (rb_null_event(event))
|
| 4792 | RB_WARN_ON(cpu_buffer, 1);
|
| 4793 | |
| 4794 | * Because the writer could be discarding every
|
| 4795 | * event it creates (which would probably be bad)
|
| 4796 | * if we were to go back to "again" then we may never
|
| 4797 | * catch up, and will trigger the warn on, or lock
|
| 4798 | * the box. Return the padding, and we will release
|
| 4799 | * the current locks, and try again.
|
| 4800 |
|
| 4801 | return event;
|
| 4802 |
|
| 4803 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 4804 |
|
| 4805 | rb_advance_reader(cpu_buffer);
|
| 4806 | goto again;
|
| 4807 |
|
| 4808 | case RINGBUF_TYPE_TIME_STAMP:
|
| 4809 | if (ts) {
|
| 4810 | *ts = rb_event_time_stamp(event);
|
| 4811 | *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
|
| 4812 | ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
|
| 4813 | cpu_buffer->cpu, ts);
|
| 4814 | }
|
| 4815 |
|
| 4816 | rb_advance_reader(cpu_buffer);
|
| 4817 | goto again;
|
| 4818 |
|
| 4819 | case RINGBUF_TYPE_DATA:
|
| 4820 | if (ts && !(*ts)) {
|
| 4821 | *ts = cpu_buffer->read_stamp + event->time_delta;
|
| 4822 | ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
|
| 4823 | cpu_buffer->cpu, ts);
|
| 4824 | }
|
| 4825 | if (lost_events)
|
| 4826 | *lost_events = rb_lost_events(cpu_buffer);
|
| 4827 | return event;
|
| 4828 |
|
| 4829 | default:
|
| 4830 | RB_WARN_ON(cpu_buffer, 1);
|
| 4831 | }
|
| 4832 |
|
| 4833 | return NULL;
|
| 4834 | }
|
| 4835 | EXPORT_SYMBOL_GPL(ring_buffer_peek);
|
| 4836 |
|
| 4837 | static struct ring_buffer_event *
|
| 4838 | rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
|
| 4839 | {
|
| 4840 | struct trace_buffer *buffer;
|
| 4841 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 4842 | struct ring_buffer_event *event;
|
| 4843 | int nr_loops = 0;
|
| 4844 |
|
| 4845 | if (ts)
|
| 4846 | *ts = 0;
|
| 4847 |
|
| 4848 | cpu_buffer = iter->cpu_buffer;
|
| 4849 | buffer = cpu_buffer->buffer;
|
| 4850 |
|
| 4851 | |
| 4852 | * Check if someone performed a consuming read to
|
| 4853 | * the buffer. A consuming read invalidates the iterator
|
| 4854 | * and we need to reset the iterator in this case.
|
| 4855 |
|
| 4856 | if (unlikely(iter->cache_read != cpu_buffer->read ||
|
| 4857 | iter->cache_reader_page != cpu_buffer->reader_page))
|
| 4858 | rb_iter_reset(iter);
|
| 4859 |
|
| 4860 | again:
|
| 4861 | if (ring_buffer_iter_empty(iter))
|
| 4862 | return NULL;
|
| 4863 |
|
| 4864 | |
| 4865 | * As the writer can mess with what the iterator is trying
|
| 4866 | * to read, just give up if we fail to get an event after
|
| 4867 | * three tries. The iterator is not as reliable when reading
|
| 4868 | * the ring buffer with an active write as the consumer is.
|
| 4869 | * Do not warn if the three failures is reached.
|
| 4870 |
|
| 4871 | if (++nr_loops > 3)
|
| 4872 | return NULL;
|
| 4873 |
|
| 4874 | if (rb_per_cpu_empty(cpu_buffer))
|
| 4875 | return NULL;
|
| 4876 |
|
| 4877 | if (iter->head >= rb_page_size(iter->head_page)) {
|
| 4878 | rb_inc_iter(iter);
|
| 4879 | goto again;
|
| 4880 | }
|
| 4881 |
|
| 4882 | event = rb_iter_head_event(iter);
|
| 4883 | if (!event)
|
| 4884 | goto again;
|
| 4885 |
|
| 4886 | switch (event->type_len) {
|
| 4887 | case RINGBUF_TYPE_PADDING:
|
| 4888 | if (rb_null_event(event)) {
|
| 4889 | rb_inc_iter(iter);
|
| 4890 | goto again;
|
| 4891 | }
|
| 4892 | rb_advance_iter(iter);
|
| 4893 | return event;
|
| 4894 |
|
| 4895 | case RINGBUF_TYPE_TIME_EXTEND:
|
| 4896 |
|
| 4897 | rb_advance_iter(iter);
|
| 4898 | goto again;
|
| 4899 |
|
| 4900 | case RINGBUF_TYPE_TIME_STAMP:
|
| 4901 | if (ts) {
|
| 4902 | *ts = rb_event_time_stamp(event);
|
| 4903 | *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
|
| 4904 | ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
|
| 4905 | cpu_buffer->cpu, ts);
|
| 4906 | }
|
| 4907 |
|
| 4908 | rb_advance_iter(iter);
|
| 4909 | goto again;
|
| 4910 |
|
| 4911 | case RINGBUF_TYPE_DATA:
|
| 4912 | if (ts && !(*ts)) {
|
| 4913 | *ts = iter->read_stamp + event->time_delta;
|
| 4914 | ring_buffer_normalize_time_stamp(buffer,
|
| 4915 | cpu_buffer->cpu, ts);
|
| 4916 | }
|
| 4917 | return event;
|
| 4918 |
|
| 4919 | default:
|
| 4920 | RB_WARN_ON(cpu_buffer, 1);
|
| 4921 | }
|
| 4922 |
|
| 4923 | return NULL;
|
| 4924 | }
|
| 4925 | EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
|
| 4926 |
|
| 4927 | static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
|
| 4928 | {
|
| 4929 | if (likely(!in_nmi())) {
|
| 4930 | raw_spin_lock(&cpu_buffer->reader_lock);
|
| 4931 | return true;
|
| 4932 | }
|
| 4933 |
|
| 4934 | |
| 4935 | * If an NMI die dumps out the content of the ring buffer
|
| 4936 | * trylock must be used to prevent a deadlock if the NMI
|
| 4937 | * preempted a task that holds the ring buffer locks. If
|
| 4938 | * we get the lock then all is fine, if not, then continue
|
| 4939 | * to do the read, but this can corrupt the ring buffer,
|
| 4940 | * so it must be permanently disabled from future writes.
|
| 4941 | * Reading from NMI is a oneshot deal.
|
| 4942 |
|
| 4943 | if (raw_spin_trylock(&cpu_buffer->reader_lock))
|
| 4944 | return true;
|
| 4945 |
|
| 4946 |
|
| 4947 | atomic_inc(&cpu_buffer->record_disabled);
|
| 4948 | return false;
|
| 4949 | }
|
| 4950 |
|
| 4951 | static inline void
|
| 4952 | rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
|
| 4953 | {
|
| 4954 | if (likely(locked))
|
| 4955 | raw_spin_unlock(&cpu_buffer->reader_lock);
|
| 4956 | }
|
| 4957 |
|
| 4958 | |
| 4959 | * ring_buffer_peek - peek at the next event to be read
|
| 4960 | * @buffer: The ring buffer to read
|
| 4961 | * @cpu: The cpu to peak at
|
| 4962 | * @ts: The timestamp counter of this event.
|
| 4963 | * @lost_events: a variable to store if events were lost (may be NULL)
|
| 4964 | *
|
| 4965 | * This will return the event that will be read next, but does
|
| 4966 | * not consume the data.
|
| 4967 |
|
| 4968 | struct ring_buffer_event *
|
| 4969 | ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
|
| 4970 | unsigned long *lost_events)
|
| 4971 | {
|
| 4972 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
|
| 4973 | struct ring_buffer_event *event;
|
| 4974 | unsigned long flags;
|
| 4975 | bool dolock;
|
| 4976 |
|
| 4977 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 4978 | return NULL;
|
| 4979 |
|
| 4980 | again:
|
| 4981 | local_irq_save(flags);
|
| 4982 | dolock = rb_reader_lock(cpu_buffer);
|
| 4983 | event = rb_buffer_peek(cpu_buffer, ts, lost_events);
|
| 4984 | if (event && event->type_len == RINGBUF_TYPE_PADDING)
|
| 4985 | rb_advance_reader(cpu_buffer);
|
| 4986 | rb_reader_unlock(cpu_buffer, dolock);
|
| 4987 | local_irq_restore(flags);
|
| 4988 |
|
| 4989 | if (event && event->type_len == RINGBUF_TYPE_PADDING)
|
| 4990 | goto again;
|
| 4991 |
|
| 4992 | return event;
|
| 4993 | }
|
| 4994 |
|
| 4995 | |
| 4996 | * @iter: The ring buffer iterator
|
| 4997 | *
|
| 4998 | * Returns true if there was dropped events since the last peek.
|
| 4999 |
|
| 5000 | bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
|
| 5001 | {
|
| 5002 | bool ret = iter->missed_events != 0;
|
| 5003 |
|
| 5004 | iter->missed_events = 0;
|
| 5005 | return ret;
|
| 5006 | }
|
| 5007 | EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
|
| 5008 |
|
| 5009 | |
| 5010 | * ring_buffer_iter_peek - peek at the next event to be read
|
| 5011 | * @iter: The ring buffer iterator
|
| 5012 | * @ts: The timestamp counter of this event.
|
| 5013 | *
|
| 5014 | * This will return the event that will be read next, but does
|
| 5015 | * not increment the iterator.
|
| 5016 |
|
| 5017 | struct ring_buffer_event *
|
| 5018 | ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
|
| 5019 | {
|
| 5020 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
|
| 5021 | struct ring_buffer_event *event;
|
| 5022 | unsigned long flags;
|
| 5023 |
|
| 5024 | again:
|
| 5025 | raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 5026 | event = rb_iter_peek(iter, ts);
|
| 5027 | raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 5028 |
|
| 5029 | if (event && event->type_len == RINGBUF_TYPE_PADDING)
|
| 5030 | goto again;
|
| 5031 |
|
| 5032 | return event;
|
| 5033 | }
|
| 5034 |
|
| 5035 | |
| 5036 | * ring_buffer_consume - return an event and consume it
|
| 5037 | * @buffer: The ring buffer to get the next event from
|
| 5038 | * @cpu: the cpu to read the buffer from
|
| 5039 | * @ts: a variable to store the timestamp (may be NULL)
|
| 5040 | * @lost_events: a variable to store if events were lost (may be NULL)
|
| 5041 | *
|
| 5042 | * Returns the next event in the ring buffer, and that event is consumed.
|
| 5043 | * Meaning, that sequential reads will keep returning a different event,
|
| 5044 | * and eventually empty the ring buffer if the producer is slower.
|
| 5045 |
|
| 5046 | struct ring_buffer_event *
|
| 5047 | ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
|
| 5048 | unsigned long *lost_events)
|
| 5049 | {
|
| 5050 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 5051 | struct ring_buffer_event *event = NULL;
|
| 5052 | unsigned long flags;
|
| 5053 | bool dolock;
|
| 5054 |
|
| 5055 | again:
|
| 5056 |
|
| 5057 | preempt_disable();
|
| 5058 |
|
| 5059 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 5060 | goto out;
|
| 5061 |
|
| 5062 | cpu_buffer = buffer->buffers[cpu];
|
| 5063 | local_irq_save(flags);
|
| 5064 | dolock = rb_reader_lock(cpu_buffer);
|
| 5065 |
|
| 5066 | event = rb_buffer_peek(cpu_buffer, ts, lost_events);
|
| 5067 | if (event) {
|
| 5068 | cpu_buffer->lost_events = 0;
|
| 5069 | rb_advance_reader(cpu_buffer);
|
| 5070 | }
|
| 5071 |
|
| 5072 | rb_reader_unlock(cpu_buffer, dolock);
|
| 5073 | local_irq_restore(flags);
|
| 5074 |
|
| 5075 | out:
|
| 5076 | preempt_enable();
|
| 5077 |
|
| 5078 | if (event && event->type_len == RINGBUF_TYPE_PADDING)
|
| 5079 | goto again;
|
| 5080 |
|
| 5081 | return event;
|
| 5082 | }
|
| 5083 | EXPORT_SYMBOL_GPL(ring_buffer_consume);
|
| 5084 |
|
| 5085 | |
| 5086 | * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
|
| 5087 | * @buffer: The ring buffer to read from
|
| 5088 | * @cpu: The cpu buffer to iterate over
|
| 5089 | * @flags: gfp flags to use for memory allocation
|
| 5090 | *
|
| 5091 | * This performs the initial preparations necessary to iterate
|
| 5092 | * through the buffer. Memory is allocated, buffer recording
|
| 5093 | * is disabled, and the iterator pointer is returned to the caller.
|
| 5094 | *
|
| 5095 | * Disabling buffer recording prevents the reading from being
|
| 5096 | * corrupted. This is not a consuming read, so a producer is not
|
| 5097 | * expected.
|
| 5098 | *
|
| 5099 | * After a sequence of ring_buffer_read_prepare calls, the user is
|
| 5100 | * expected to make at least one call to ring_buffer_read_prepare_sync.
|
| 5101 | * Afterwards, ring_buffer_read_start is invoked to get things going
|
| 5102 | * for real.
|
| 5103 | *
|
| 5104 | * This overall must be paired with ring_buffer_read_finish.
|
| 5105 |
|
| 5106 | struct ring_buffer_iter *
|
| 5107 | ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
|
| 5108 | {
|
| 5109 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 5110 | struct ring_buffer_iter *iter;
|
| 5111 |
|
| 5112 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 5113 | return NULL;
|
| 5114 |
|
| 5115 | iter = kzalloc(sizeof(*iter), flags);
|
| 5116 | if (!iter)
|
| 5117 | return NULL;
|
| 5118 |
|
| 5119 | iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
|
| 5120 | if (!iter->event) {
|
| 5121 | kfree(iter);
|
| 5122 | return NULL;
|
| 5123 | }
|
| 5124 |
|
| 5125 | cpu_buffer = buffer->buffers[cpu];
|
| 5126 |
|
| 5127 | iter->cpu_buffer = cpu_buffer;
|
| 5128 |
|
| 5129 | atomic_inc(&cpu_buffer->resize_disabled);
|
| 5130 |
|
| 5131 | return iter;
|
| 5132 | }
|
| 5133 | EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
|
| 5134 |
|
| 5135 | |
| 5136 | * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
|
| 5137 | *
|
| 5138 | * All previously invoked ring_buffer_read_prepare calls to prepare
|
| 5139 | * iterators will be synchronized. Afterwards, read_buffer_read_start
|
| 5140 | * calls on those iterators are allowed.
|
| 5141 |
|
| 5142 | void
|
| 5143 | ring_buffer_read_prepare_sync(void)
|
| 5144 | {
|
| 5145 | synchronize_rcu();
|
| 5146 | }
|
| 5147 | EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
|
| 5148 |
|
| 5149 | |
| 5150 | * ring_buffer_read_start - start a non consuming read of the buffer
|
| 5151 | * @iter: The iterator returned by ring_buffer_read_prepare
|
| 5152 | *
|
| 5153 | * This finalizes the startup of an iteration through the buffer.
|
| 5154 | * The iterator comes from a call to ring_buffer_read_prepare and
|
| 5155 | * an intervening ring_buffer_read_prepare_sync must have been
|
| 5156 | * performed.
|
| 5157 | *
|
| 5158 | * Must be paired with ring_buffer_read_finish.
|
| 5159 |
|
| 5160 | void
|
| 5161 | ring_buffer_read_start(struct ring_buffer_iter *iter)
|
| 5162 | {
|
| 5163 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 5164 | unsigned long flags;
|
| 5165 |
|
| 5166 | if (!iter)
|
| 5167 | return;
|
| 5168 |
|
| 5169 | cpu_buffer = iter->cpu_buffer;
|
| 5170 |
|
| 5171 | raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 5172 | arch_spin_lock(&cpu_buffer->lock);
|
| 5173 | rb_iter_reset(iter);
|
| 5174 | arch_spin_unlock(&cpu_buffer->lock);
|
| 5175 | raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 5176 | }
|
| 5177 | EXPORT_SYMBOL_GPL(ring_buffer_read_start);
|
| 5178 |
|
| 5179 | |
| 5180 | * ring_buffer_read_finish - finish reading the iterator of the buffer
|
| 5181 | * @iter: The iterator retrieved by ring_buffer_start
|
| 5182 | *
|
| 5183 | * This re-enables the recording to the buffer, and frees the
|
| 5184 | * iterator.
|
| 5185 |
|
| 5186 | void
|
| 5187 | ring_buffer_read_finish(struct ring_buffer_iter *iter)
|
| 5188 | {
|
| 5189 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
|
| 5190 | unsigned long flags;
|
| 5191 |
|
| 5192 | |
| 5193 | * Ring buffer is disabled from recording, here's a good place
|
| 5194 | * to check the integrity of the ring buffer.
|
| 5195 | * Must prevent readers from trying to read, as the check
|
| 5196 | * clears the HEAD page and readers require it.
|
| 5197 |
|
| 5198 | raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 5199 | rb_check_pages(cpu_buffer);
|
| 5200 | raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 5201 |
|
| 5202 | atomic_dec(&cpu_buffer->resize_disabled);
|
| 5203 | kfree(iter->event);
|
| 5204 | kfree(iter);
|
| 5205 | }
|
| 5206 | EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
|
| 5207 |
|
| 5208 | |
| 5209 | * ring_buffer_iter_advance - advance the iterator to the next location
|
| 5210 | * @iter: The ring buffer iterator
|
| 5211 | *
|
| 5212 | * Move the location of the iterator such that the next read will
|
| 5213 | * be the next location of the iterator.
|
| 5214 |
|
| 5215 | void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
|
| 5216 | {
|
| 5217 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
|
| 5218 | unsigned long flags;
|
| 5219 |
|
| 5220 | raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 5221 |
|
| 5222 | rb_advance_iter(iter);
|
| 5223 |
|
| 5224 | raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 5225 | }
|
| 5226 | EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
|
| 5227 |
|
| 5228 | |
| 5229 | * ring_buffer_size - return the size of the ring buffer (in bytes)
|
| 5230 | * @buffer: The ring buffer.
|
| 5231 | * @cpu: The CPU to get ring buffer size from.
|
| 5232 |
|
| 5233 | unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
|
| 5234 | {
|
| 5235 | |
| 5236 | * Earlier, this method returned
|
| 5237 | * BUF_PAGE_SIZE * buffer->nr_pages
|
| 5238 | * Since the nr_pages field is now removed, we have converted this to
|
| 5239 | * return the per cpu buffer value.
|
| 5240 |
|
| 5241 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 5242 | return 0;
|
| 5243 |
|
| 5244 | return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
|
| 5245 | }
|
| 5246 | EXPORT_SYMBOL_GPL(ring_buffer_size);
|
| 5247 |
|
| 5248 | static void rb_clear_buffer_page(struct buffer_page *page)
|
| 5249 | {
|
| 5250 | local_set(&page->write, 0);
|
| 5251 | local_set(&page->entries, 0);
|
| 5252 | rb_init_page(page->page);
|
| 5253 | page->read = 0;
|
| 5254 | }
|
| 5255 |
|
| 5256 | static void
|
| 5257 | rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
|
| 5258 | {
|
| 5259 | struct buffer_page *page;
|
| 5260 |
|
| 5261 | rb_head_page_deactivate(cpu_buffer);
|
| 5262 |
|
| 5263 | cpu_buffer->head_page
|
| 5264 | = list_entry(cpu_buffer->pages, struct buffer_page, list);
|
| 5265 | rb_clear_buffer_page(cpu_buffer->head_page);
|
| 5266 | list_for_each_entry(page, cpu_buffer->pages, list) {
|
| 5267 | rb_clear_buffer_page(page);
|
| 5268 | }
|
| 5269 |
|
| 5270 | cpu_buffer->tail_page = cpu_buffer->head_page;
|
| 5271 | cpu_buffer->commit_page = cpu_buffer->head_page;
|
| 5272 |
|
| 5273 | INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
|
| 5274 | INIT_LIST_HEAD(&cpu_buffer->new_pages);
|
| 5275 | rb_clear_buffer_page(cpu_buffer->reader_page);
|
| 5276 |
|
| 5277 | local_set(&cpu_buffer->entries_bytes, 0);
|
| 5278 | local_set(&cpu_buffer->overrun, 0);
|
| 5279 | local_set(&cpu_buffer->commit_overrun, 0);
|
| 5280 | local_set(&cpu_buffer->dropped_events, 0);
|
| 5281 | local_set(&cpu_buffer->entries, 0);
|
| 5282 | local_set(&cpu_buffer->committing, 0);
|
| 5283 | local_set(&cpu_buffer->commits, 0);
|
| 5284 | local_set(&cpu_buffer->pages_touched, 0);
|
| 5285 | local_set(&cpu_buffer->pages_lost, 0);
|
| 5286 | local_set(&cpu_buffer->pages_read, 0);
|
| 5287 | cpu_buffer->last_pages_touch = 0;
|
| 5288 | cpu_buffer->shortest_full = 0;
|
| 5289 | cpu_buffer->read = 0;
|
| 5290 | cpu_buffer->read_bytes = 0;
|
| 5291 |
|
| 5292 | rb_time_set(&cpu_buffer->write_stamp, 0);
|
| 5293 | rb_time_set(&cpu_buffer->before_stamp, 0);
|
| 5294 |
|
| 5295 | memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
|
| 5296 |
|
| 5297 | cpu_buffer->lost_events = 0;
|
| 5298 | cpu_buffer->last_overrun = 0;
|
| 5299 |
|
| 5300 | rb_head_page_activate(cpu_buffer);
|
| 5301 | }
|
| 5302 |
|
| 5303 |
|
| 5304 | static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
|
| 5305 | {
|
| 5306 | unsigned long flags;
|
| 5307 |
|
| 5308 | raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 5309 |
|
| 5310 | if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
|
| 5311 | goto out;
|
| 5312 |
|
| 5313 | arch_spin_lock(&cpu_buffer->lock);
|
| 5314 |
|
| 5315 | rb_reset_cpu(cpu_buffer);
|
| 5316 |
|
| 5317 | arch_spin_unlock(&cpu_buffer->lock);
|
| 5318 |
|
| 5319 | out:
|
| 5320 | raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 5321 | }
|
| 5322 |
|
| 5323 | |
| 5324 | * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
|
| 5325 | * @buffer: The ring buffer to reset a per cpu buffer of
|
| 5326 | * @cpu: The CPU buffer to be reset
|
| 5327 |
|
| 5328 | void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
|
| 5329 | {
|
| 5330 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
|
| 5331 |
|
| 5332 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 5333 | return;
|
| 5334 |
|
| 5335 |
|
| 5336 | mutex_lock(&buffer->mutex);
|
| 5337 |
|
| 5338 | atomic_inc(&cpu_buffer->resize_disabled);
|
| 5339 | atomic_inc(&cpu_buffer->record_disabled);
|
| 5340 |
|
| 5341 |
|
| 5342 | synchronize_rcu();
|
| 5343 |
|
| 5344 | reset_disabled_cpu_buffer(cpu_buffer);
|
| 5345 |
|
| 5346 | atomic_dec(&cpu_buffer->record_disabled);
|
| 5347 | atomic_dec(&cpu_buffer->resize_disabled);
|
| 5348 |
|
| 5349 | mutex_unlock(&buffer->mutex);
|
| 5350 | }
|
| 5351 | EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
|
| 5352 |
|
| 5353 |
|
| 5354 | #define RESET_BIT (1 << 30)
|
| 5355 |
|
| 5356 | |
| 5357 | * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
|
| 5358 | * @buffer: The ring buffer to reset a per cpu buffer of
|
| 5359 | * @cpu: The CPU buffer to be reset
|
| 5360 |
|
| 5361 | void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
|
| 5362 | {
|
| 5363 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 5364 | int cpu;
|
| 5365 |
|
| 5366 |
|
| 5367 | mutex_lock(&buffer->mutex);
|
| 5368 |
|
| 5369 | for_each_online_buffer_cpu(buffer, cpu) {
|
| 5370 | cpu_buffer = buffer->buffers[cpu];
|
| 5371 |
|
| 5372 | atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
|
| 5373 | atomic_inc(&cpu_buffer->record_disabled);
|
| 5374 | }
|
| 5375 |
|
| 5376 |
|
| 5377 | synchronize_rcu();
|
| 5378 |
|
| 5379 | for_each_buffer_cpu(buffer, cpu) {
|
| 5380 | cpu_buffer = buffer->buffers[cpu];
|
| 5381 |
|
| 5382 | |
| 5383 | * If a CPU came online during the synchronize_rcu(), then
|
| 5384 | * ignore it.
|
| 5385 |
|
| 5386 | if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
|
| 5387 | continue;
|
| 5388 |
|
| 5389 | reset_disabled_cpu_buffer(cpu_buffer);
|
| 5390 |
|
| 5391 | atomic_dec(&cpu_buffer->record_disabled);
|
| 5392 | atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
|
| 5393 | }
|
| 5394 |
|
| 5395 | mutex_unlock(&buffer->mutex);
|
| 5396 | }
|
| 5397 |
|
| 5398 | |
| 5399 | * ring_buffer_reset - reset a ring buffer
|
| 5400 | * @buffer: The ring buffer to reset all cpu buffers
|
| 5401 |
|
| 5402 | void ring_buffer_reset(struct trace_buffer *buffer)
|
| 5403 | {
|
| 5404 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 5405 | int cpu;
|
| 5406 |
|
| 5407 |
|
| 5408 | mutex_lock(&buffer->mutex);
|
| 5409 |
|
| 5410 | for_each_buffer_cpu(buffer, cpu) {
|
| 5411 | cpu_buffer = buffer->buffers[cpu];
|
| 5412 |
|
| 5413 | atomic_inc(&cpu_buffer->resize_disabled);
|
| 5414 | atomic_inc(&cpu_buffer->record_disabled);
|
| 5415 | }
|
| 5416 |
|
| 5417 |
|
| 5418 | synchronize_rcu();
|
| 5419 |
|
| 5420 | for_each_buffer_cpu(buffer, cpu) {
|
| 5421 | cpu_buffer = buffer->buffers[cpu];
|
| 5422 |
|
| 5423 | reset_disabled_cpu_buffer(cpu_buffer);
|
| 5424 |
|
| 5425 | atomic_dec(&cpu_buffer->record_disabled);
|
| 5426 | atomic_dec(&cpu_buffer->resize_disabled);
|
| 5427 | }
|
| 5428 |
|
| 5429 | mutex_unlock(&buffer->mutex);
|
| 5430 | }
|
| 5431 | EXPORT_SYMBOL_GPL(ring_buffer_reset);
|
| 5432 |
|
| 5433 | |
| 5434 | * ring_buffer_empty - is the ring buffer empty?
|
| 5435 | * @buffer: The ring buffer to test
|
| 5436 |
|
| 5437 | bool ring_buffer_empty(struct trace_buffer *buffer)
|
| 5438 | {
|
| 5439 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 5440 | unsigned long flags;
|
| 5441 | bool dolock;
|
| 5442 | bool ret;
|
| 5443 | int cpu;
|
| 5444 |
|
| 5445 |
|
| 5446 | for_each_buffer_cpu(buffer, cpu) {
|
| 5447 | cpu_buffer = buffer->buffers[cpu];
|
| 5448 | local_irq_save(flags);
|
| 5449 | dolock = rb_reader_lock(cpu_buffer);
|
| 5450 | ret = rb_per_cpu_empty(cpu_buffer);
|
| 5451 | rb_reader_unlock(cpu_buffer, dolock);
|
| 5452 | local_irq_restore(flags);
|
| 5453 |
|
| 5454 | if (!ret)
|
| 5455 | return false;
|
| 5456 | }
|
| 5457 |
|
| 5458 | return true;
|
| 5459 | }
|
| 5460 | EXPORT_SYMBOL_GPL(ring_buffer_empty);
|
| 5461 |
|
| 5462 | |
| 5463 | * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
|
| 5464 | * @buffer: The ring buffer
|
| 5465 | * @cpu: The CPU buffer to test
|
| 5466 |
|
| 5467 | bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
|
| 5468 | {
|
| 5469 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 5470 | unsigned long flags;
|
| 5471 | bool dolock;
|
| 5472 | bool ret;
|
| 5473 |
|
| 5474 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 5475 | return true;
|
| 5476 |
|
| 5477 | cpu_buffer = buffer->buffers[cpu];
|
| 5478 | local_irq_save(flags);
|
| 5479 | dolock = rb_reader_lock(cpu_buffer);
|
| 5480 | ret = rb_per_cpu_empty(cpu_buffer);
|
| 5481 | rb_reader_unlock(cpu_buffer, dolock);
|
| 5482 | local_irq_restore(flags);
|
| 5483 |
|
| 5484 | return ret;
|
| 5485 | }
|
| 5486 | EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
|
| 5487 |
|
| 5488 | #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
|
| 5489 | |
| 5490 | * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
|
| 5491 | * @buffer_a: One buffer to swap with
|
| 5492 | * @buffer_b: The other buffer to swap with
|
| 5493 | * @cpu: the CPU of the buffers to swap
|
| 5494 | *
|
| 5495 | * This function is useful for tracers that want to take a "snapshot"
|
| 5496 | * of a CPU buffer and has another back up buffer lying around.
|
| 5497 | * it is expected that the tracer handles the cpu buffer not being
|
| 5498 | * used at the moment.
|
| 5499 |
|
| 5500 | int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
|
| 5501 | struct trace_buffer *buffer_b, int cpu)
|
| 5502 | {
|
| 5503 | struct ring_buffer_per_cpu *cpu_buffer_a;
|
| 5504 | struct ring_buffer_per_cpu *cpu_buffer_b;
|
| 5505 | int ret = -EINVAL;
|
| 5506 |
|
| 5507 | if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
|
| 5508 | !cpumask_test_cpu(cpu, buffer_b->cpumask))
|
| 5509 | goto out;
|
| 5510 |
|
| 5511 | cpu_buffer_a = buffer_a->buffers[cpu];
|
| 5512 | cpu_buffer_b = buffer_b->buffers[cpu];
|
| 5513 |
|
| 5514 |
|
| 5515 | if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
|
| 5516 | goto out;
|
| 5517 |
|
| 5518 | ret = -EAGAIN;
|
| 5519 |
|
| 5520 | if (atomic_read(&buffer_a->record_disabled))
|
| 5521 | goto out;
|
| 5522 |
|
| 5523 | if (atomic_read(&buffer_b->record_disabled))
|
| 5524 | goto out;
|
| 5525 |
|
| 5526 | if (atomic_read(&cpu_buffer_a->record_disabled))
|
| 5527 | goto out;
|
| 5528 |
|
| 5529 | if (atomic_read(&cpu_buffer_b->record_disabled))
|
| 5530 | goto out;
|
| 5531 |
|
| 5532 | |
| 5533 | * We can't do a synchronize_rcu here because this
|
| 5534 | * function can be called in atomic context.
|
| 5535 | * Normally this will be called from the same CPU as cpu.
|
| 5536 | * If not it's up to the caller to protect this.
|
| 5537 |
|
| 5538 | atomic_inc(&cpu_buffer_a->record_disabled);
|
| 5539 | atomic_inc(&cpu_buffer_b->record_disabled);
|
| 5540 |
|
| 5541 | ret = -EBUSY;
|
| 5542 | if (local_read(&cpu_buffer_a->committing))
|
| 5543 | goto out_dec;
|
| 5544 | if (local_read(&cpu_buffer_b->committing))
|
| 5545 | goto out_dec;
|
| 5546 |
|
| 5547 | |
| 5548 | * When resize is in progress, we cannot swap it because
|
| 5549 | * it will mess the state of the cpu buffer.
|
| 5550 |
|
| 5551 | if (atomic_read(&buffer_a->resizing))
|
| 5552 | goto out_dec;
|
| 5553 | if (atomic_read(&buffer_b->resizing))
|
| 5554 | goto out_dec;
|
| 5555 |
|
| 5556 | buffer_a->buffers[cpu] = cpu_buffer_b;
|
| 5557 | buffer_b->buffers[cpu] = cpu_buffer_a;
|
| 5558 |
|
| 5559 | cpu_buffer_b->buffer = buffer_a;
|
| 5560 | cpu_buffer_a->buffer = buffer_b;
|
| 5561 |
|
| 5562 | ret = 0;
|
| 5563 |
|
| 5564 | out_dec:
|
| 5565 | atomic_dec(&cpu_buffer_a->record_disabled);
|
| 5566 | atomic_dec(&cpu_buffer_b->record_disabled);
|
| 5567 | out:
|
| 5568 | return ret;
|
| 5569 | }
|
| 5570 | EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
|
| 5571 | #endif
|
| 5572 |
|
| 5573 | |
| 5574 | * ring_buffer_alloc_read_page - allocate a page to read from buffer
|
| 5575 | * @buffer: the buffer to allocate for.
|
| 5576 | * @cpu: the cpu buffer to allocate.
|
| 5577 | *
|
| 5578 | * This function is used in conjunction with ring_buffer_read_page.
|
| 5579 | * When reading a full page from the ring buffer, these functions
|
| 5580 | * can be used to speed up the process. The calling function should
|
| 5581 | * allocate a few pages first with this function. Then when it
|
| 5582 | * needs to get pages from the ring buffer, it passes the result
|
| 5583 | * of this function into ring_buffer_read_page, which will swap
|
| 5584 | * the page that was allocated, with the read page of the buffer.
|
| 5585 | *
|
| 5586 | * Returns:
|
| 5587 | * The page allocated, or ERR_PTR
|
| 5588 |
|
| 5589 | void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
|
| 5590 | {
|
| 5591 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 5592 | struct buffer_data_page *bpage = NULL;
|
| 5593 | unsigned long flags;
|
| 5594 | struct page *page;
|
| 5595 |
|
| 5596 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 5597 | return ERR_PTR(-ENODEV);
|
| 5598 |
|
| 5599 | cpu_buffer = buffer->buffers[cpu];
|
| 5600 | local_irq_save(flags);
|
| 5601 | arch_spin_lock(&cpu_buffer->lock);
|
| 5602 |
|
| 5603 | if (cpu_buffer->free_page) {
|
| 5604 | bpage = cpu_buffer->free_page;
|
| 5605 | cpu_buffer->free_page = NULL;
|
| 5606 | }
|
| 5607 |
|
| 5608 | arch_spin_unlock(&cpu_buffer->lock);
|
| 5609 | local_irq_restore(flags);
|
| 5610 |
|
| 5611 | if (bpage)
|
| 5612 | goto out;
|
| 5613 |
|
| 5614 | page = alloc_pages_node(cpu_to_node(cpu),
|
| 5615 | GFP_KERNEL | __GFP_NORETRY, 0);
|
| 5616 | if (!page)
|
| 5617 | return ERR_PTR(-ENOMEM);
|
| 5618 |
|
| 5619 | bpage = page_address(page);
|
| 5620 |
|
| 5621 | out:
|
| 5622 | rb_init_page(bpage);
|
| 5623 |
|
| 5624 | return bpage;
|
| 5625 | }
|
| 5626 | EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
|
| 5627 |
|
| 5628 | |
| 5629 | * ring_buffer_free_read_page - free an allocated read page
|
| 5630 | * @buffer: the buffer the page was allocate for
|
| 5631 | * @cpu: the cpu buffer the page came from
|
| 5632 | * @data: the page to free
|
| 5633 | *
|
| 5634 | * Free a page allocated from ring_buffer_alloc_read_page.
|
| 5635 |
|
| 5636 | void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
|
| 5637 | {
|
| 5638 | struct ring_buffer_per_cpu *cpu_buffer;
|
| 5639 | struct buffer_data_page *bpage = data;
|
| 5640 | struct page *page = virt_to_page(bpage);
|
| 5641 | unsigned long flags;
|
| 5642 |
|
| 5643 | if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
|
| 5644 | return;
|
| 5645 |
|
| 5646 | cpu_buffer = buffer->buffers[cpu];
|
| 5647 |
|
| 5648 |
|
| 5649 | if (page_ref_count(page) > 1)
|
| 5650 | goto out;
|
| 5651 |
|
| 5652 | local_irq_save(flags);
|
| 5653 | arch_spin_lock(&cpu_buffer->lock);
|
| 5654 |
|
| 5655 | if (!cpu_buffer->free_page) {
|
| 5656 | cpu_buffer->free_page = bpage;
|
| 5657 | bpage = NULL;
|
| 5658 | }
|
| 5659 |
|
| 5660 | arch_spin_unlock(&cpu_buffer->lock);
|
| 5661 | local_irq_restore(flags);
|
| 5662 |
|
| 5663 | out:
|
| 5664 | free_page((unsigned long)bpage);
|
| 5665 | }
|
| 5666 | EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
|
| 5667 |
|
| 5668 | |
| 5669 | * ring_buffer_read_page - extract a page from the ring buffer
|
| 5670 | * @buffer: buffer to extract from
|
| 5671 | * @data_page: the page to use allocated from ring_buffer_alloc_read_page
|
| 5672 | * @len: amount to extract
|
| 5673 | * @cpu: the cpu of the buffer to extract
|
| 5674 | * @full: should the extraction only happen when the page is full.
|
| 5675 | *
|
| 5676 | * This function will pull out a page from the ring buffer and consume it.
|
| 5677 | * @data_page must be the address of the variable that was returned
|
| 5678 | * from ring_buffer_alloc_read_page. This is because the page might be used
|
| 5679 | * to swap with a page in the ring buffer.
|
| 5680 | *
|
| 5681 | * for example:
|
| 5682 | * rpage = ring_buffer_alloc_read_page(buffer, cpu);
|
| 5683 | * if (IS_ERR(rpage))
|
| 5684 | * return PTR_ERR(rpage);
|
| 5685 | * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
|
| 5686 | * if (ret >= 0)
|
| 5687 | * process_page(rpage, ret);
|
| 5688 | *
|
| 5689 | * When @full is set, the function will not return true unless
|
| 5690 | * the writer is off the reader page.
|
| 5691 | *
|
| 5692 | * Note: it is up to the calling functions to handle sleeps and wakeups.
|
| 5693 | * The ring buffer can be used anywhere in the kernel and can not
|
| 5694 | * blindly call wake_up. The layer that uses the ring buffer must be
|
| 5695 | * responsible for that.
|
| 5696 | *
|
| 5697 | * Returns:
|
| 5698 | * >=0 if data has been transferred, returns the offset of consumed data.
|
| 5699 | * <0 if no data has been transferred.
|
| 5700 |
|
| 5701 | int ring_buffer_read_page(struct trace_buffer *buffer,
|
| 5702 | void **data_page, size_t len, int cpu, int full)
|
| 5703 | {
|
| 5704 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
|
| 5705 | struct ring_buffer_event *event;
|
| 5706 | struct buffer_data_page *bpage;
|
| 5707 | struct buffer_page *reader;
|
| 5708 | unsigned long missed_events;
|
| 5709 | unsigned long flags;
|
| 5710 | unsigned int commit;
|
| 5711 | unsigned int read;
|
| 5712 | u64 save_timestamp;
|
| 5713 | int ret = -1;
|
| 5714 |
|
| 5715 | if (!cpumask_test_cpu(cpu, buffer->cpumask))
|
| 5716 | goto out;
|
| 5717 |
|
| 5718 | |
| 5719 | * If len is not big enough to hold the page header, then
|
| 5720 | * we can not copy anything.
|
| 5721 |
|
| 5722 | if (len <= BUF_PAGE_HDR_SIZE)
|
| 5723 | goto out;
|
| 5724 |
|
| 5725 | len -= BUF_PAGE_HDR_SIZE;
|
| 5726 |
|
| 5727 | if (!data_page)
|
| 5728 | goto out;
|
| 5729 |
|
| 5730 | bpage = *data_page;
|
| 5731 | if (!bpage)
|
| 5732 | goto out;
|
| 5733 |
|
| 5734 | raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
|
| 5735 |
|
| 5736 | reader = rb_get_reader_page(cpu_buffer);
|
| 5737 | if (!reader)
|
| 5738 | goto out_unlock;
|
| 5739 |
|
| 5740 | event = rb_reader_event(cpu_buffer);
|
| 5741 |
|
| 5742 | read = reader->read;
|
| 5743 | commit = rb_page_commit(reader);
|
| 5744 |
|
| 5745 |
|
| 5746 | missed_events = cpu_buffer->lost_events;
|
| 5747 |
|
| 5748 | |
| 5749 | * If this page has been partially read or
|
| 5750 | * if len is not big enough to read the rest of the page or
|
| 5751 | * a writer is still on the page, then
|
| 5752 | * we must copy the data from the page to the buffer.
|
| 5753 | * Otherwise, we can simply swap the page with the one passed in.
|
| 5754 |
|
| 5755 | if (read || (len < (commit - read)) ||
|
| 5756 | cpu_buffer->reader_page == cpu_buffer->commit_page) {
|
| 5757 | struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
|
| 5758 | unsigned int rpos = read;
|
| 5759 | unsigned int pos = 0;
|
| 5760 | unsigned int size;
|
| 5761 |
|
| 5762 | |
| 5763 | * If a full page is expected, this can still be returned
|
| 5764 | * if there's been a previous partial read and the
|
| 5765 | * rest of the page can be read and the commit page is off
|
| 5766 | * the reader page.
|
| 5767 |
|
| 5768 | if (full &&
|
| 5769 | (!read || (len < (commit - read)) ||
|
| 5770 | cpu_buffer->reader_page == cpu_buffer->commit_page))
|
| 5771 | goto out_unlock;
|
| 5772 |
|
| 5773 | if (len > (commit - read))
|
| 5774 | len = (commit - read);
|
| 5775 |
|
| 5776 |
|
| 5777 | size = rb_event_ts_length(event);
|
| 5778 |
|
| 5779 | if (len < size)
|
| 5780 | goto out_unlock;
|
| 5781 |
|
| 5782 |
|
| 5783 | save_timestamp = cpu_buffer->read_stamp;
|
| 5784 |
|
| 5785 |
|
| 5786 | do {
|
| 5787 | |
| 5788 | * rb_advance_reader only advances by one event,
|
| 5789 | * whereas rb_event_ts_length may include the size of
|
| 5790 | * one or two events.
|
| 5791 | * We have already ensured there's enough space if this
|
| 5792 | * is a time extend.
|
| 5793 | size = rb_event_length(event);
|
| 5794 | memcpy(bpage->data + pos, rpage->data + rpos, size);
|
| 5795 |
|
| 5796 | len -= size;
|
| 5797 |
|
| 5798 | rb_advance_reader(cpu_buffer);
|
| 5799 | rpos = reader->read;
|
| 5800 | pos += size;
|
| 5801 |
|
| 5802 | if (rpos >= commit)
|
| 5803 | break;
|
| 5804 |
|
| 5805 | event = rb_reader_event(cpu_buffer);
|
| 5806 |
|
| 5807 | size = rb_event_ts_length(event);
|
| 5808 | } while (len >= size);
|
| 5809 |
|
| 5810 |
|
| 5811 | local_set(&bpage->commit, pos);
|
| 5812 | bpage->time_stamp = save_timestamp;
|
| 5813 |
|
| 5814 |
|
| 5815 | read = 0;
|
| 5816 | } else {
|
| 5817 |
|
| 5818 | cpu_buffer->read += rb_page_entries(reader);
|
| 5819 | cpu_buffer->read_bytes += BUF_PAGE_SIZE;
|
| 5820 |
|
| 5821 |
|
| 5822 | rb_init_page(bpage);
|
| 5823 | bpage = reader->page;
|
| 5824 | reader->page = *data_page;
|
| 5825 | local_set(&reader->write, 0);
|
| 5826 | local_set(&reader->entries, 0);
|
| 5827 | reader->read = 0;
|
| 5828 | *data_page = bpage;
|
| 5829 |
|
| 5830 | |
| 5831 | * Use the real_end for the data size,
|
| 5832 | * This gives us a chance to store the lost events
|
| 5833 | * on the page.
|
| 5834 |
|
| 5835 | if (reader->real_end)
|
| 5836 | local_set(&bpage->commit, reader->real_end);
|
| 5837 | }
|
| 5838 | ret = read;
|
| 5839 |
|
| 5840 | cpu_buffer->lost_events = 0;
|
| 5841 |
|
| 5842 | commit = local_read(&bpage->commit);
|
| 5843 | |
| 5844 | * Set a flag in the commit field if we lost events
|
| 5845 |
|
| 5846 | if (missed_events) {
|
| 5847 | |
| 5848 | * missed events, then record it there.
|
| 5849 |
|
| 5850 | if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
|
| 5851 | memcpy(&bpage->data[commit], &missed_events,
|
| 5852 | sizeof(missed_events));
|
| 5853 | local_add(RB_MISSED_STORED, &bpage->commit);
|
| 5854 | commit += sizeof(missed_events);
|
| 5855 | }
|
| 5856 | local_add(RB_MISSED_EVENTS, &bpage->commit);
|
| 5857 | }
|
| 5858 |
|
| 5859 | |
| 5860 | * This page may be off to user land. Zero it out here.
|
| 5861 |
|
| 5862 | if (commit < BUF_PAGE_SIZE)
|
| 5863 | memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
|
| 5864 |
|
| 5865 | out_unlock:
|
| 5866 | raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
|
| 5867 |
|
| 5868 | out:
|
| 5869 | return ret;
|
| 5870 | }
|
| 5871 | EXPORT_SYMBOL_GPL(ring_buffer_read_page);
|
| 5872 |
|
| 5873 | |
| 5874 | * We only allocate new buffers, never free them if the CPU goes down.
|
| 5875 | * If we were to free the buffer, then the user would lose any trace that was in
|
| 5876 | * the buffer.
|
| 5877 |
|
| 5878 | int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
|
| 5879 | {
|
| 5880 | struct trace_buffer *buffer;
|
| 5881 | long nr_pages_same;
|
| 5882 | int cpu_i;
|
| 5883 | unsigned long nr_pages;
|
| 5884 |
|
| 5885 | buffer = container_of(node, struct trace_buffer, node);
|
| 5886 | if (cpumask_test_cpu(cpu, buffer->cpumask))
|
| 5887 | return 0;
|
| 5888 |
|
| 5889 | nr_pages = 0;
|
| 5890 | nr_pages_same = 1;
|
| 5891 |
|
| 5892 | for_each_buffer_cpu(buffer, cpu_i) {
|
| 5893 |
|
| 5894 | if (nr_pages == 0)
|
| 5895 | nr_pages = buffer->buffers[cpu_i]->nr_pages;
|
| 5896 | if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
|
| 5897 | nr_pages_same = 0;
|
| 5898 | break;
|
| 5899 | }
|
| 5900 | }
|
| 5901 |
|
| 5902 | if (!nr_pages_same)
|
| 5903 | nr_pages = 2;
|
| 5904 | buffer->buffers[cpu] =
|
| 5905 | rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
|
| 5906 | if (!buffer->buffers[cpu]) {
|
| 5907 | WARN(1, "failed to allocate ring buffer on CPU %u\n",
|
| 5908 | cpu);
|
| 5909 | return -ENOMEM;
|
| 5910 | }
|
| 5911 | smp_wmb();
|
| 5912 | cpumask_set_cpu(cpu, buffer->cpumask);
|
| 5913 | return 0;
|
| 5914 | }
|
| 5915 |
|
| 5916 | #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
|
| 5917 | |
| 5918 | * This is a basic integrity check of the ring buffer.
|
| 5919 | * Late in the boot cycle this test will run when configured in.
|
| 5920 | * It will kick off a thread per CPU that will go into a loop
|
| 5921 | * writing to the per cpu ring buffer various sizes of data.
|
| 5922 | * Some of the data will be large items, some small.
|
| 5923 | *
|
| 5924 | * Another thread is created that goes into a spin, sending out
|
| 5925 | * IPIs to the other CPUs to also write into the ring buffer.
|
| 5926 | * this is to test the nesting ability of the buffer.
|
| 5927 | *
|
| 5928 | * Basic stats are recorded and reported. If something in the
|
| 5929 | * ring buffer should happen that's not expected, a big warning
|
| 5930 | * is displayed and all ring buffers are disabled.
|
| 5931 |
|
| 5932 | static struct task_struct *rb_threads[NR_CPUS] __initdata;
|
| 5933 |
|
| 5934 | struct rb_test_data {
|
| 5935 | struct trace_buffer *buffer;
|
| 5936 | unsigned long events;
|
| 5937 | unsigned long bytes_written;
|
| 5938 | unsigned long bytes_alloc;
|
| 5939 | unsigned long bytes_dropped;
|
| 5940 | unsigned long events_nested;
|
| 5941 | unsigned long bytes_written_nested;
|
| 5942 | unsigned long bytes_alloc_nested;
|
| 5943 | unsigned long bytes_dropped_nested;
|
| 5944 | int min_size_nested;
|
| 5945 | int max_size_nested;
|
| 5946 | int max_size;
|
| 5947 | int min_size;
|
| 5948 | int cpu;
|
| 5949 | int cnt;
|
| 5950 | };
|
| 5951 |
|
| 5952 | static struct rb_test_data rb_data[NR_CPUS] __initdata;
|
| 5953 |
|
| 5954 |
|
| 5955 | #define RB_TEST_BUFFER_SIZE 1048576
|
| 5956 |
|
| 5957 | static char rb_string[] __initdata =
|
| 5958 | "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
|
| 5959 | "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
|
| 5960 | "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
|
| 5961 |
|
| 5962 | static bool rb_test_started __initdata;
|
| 5963 |
|
| 5964 | struct rb_item {
|
| 5965 | int size;
|
| 5966 | char str[];
|
| 5967 | };
|
| 5968 |
|
| 5969 | static __init int rb_write_something(struct rb_test_data *data, bool nested)
|
| 5970 | {
|
| 5971 | struct ring_buffer_event *event;
|
| 5972 | struct rb_item *item;
|
| 5973 | bool started;
|
| 5974 | int event_len;
|
| 5975 | int size;
|
| 5976 | int len;
|
| 5977 | int cnt;
|
| 5978 |
|
| 5979 |
|
| 5980 | cnt = data->cnt + (nested ? 27 : 0);
|
| 5981 |
|
| 5982 |
|
| 5983 | size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
|
| 5984 |
|
| 5985 | len = size + sizeof(struct rb_item);
|
| 5986 |
|
| 5987 | started = rb_test_started;
|
| 5988 |
|
| 5989 | smp_rmb();
|
| 5990 |
|
| 5991 | event = ring_buffer_lock_reserve(data->buffer, len);
|
| 5992 | if (!event) {
|
| 5993 |
|
| 5994 | if (started) {
|
| 5995 | if (nested)
|
| 5996 | data->bytes_dropped += len;
|
| 5997 | else
|
| 5998 | data->bytes_dropped_nested += len;
|
| 5999 | }
|
| 6000 | return len;
|
| 6001 | }
|
| 6002 |
|
| 6003 | event_len = ring_buffer_event_length(event);
|
| 6004 |
|
| 6005 | if (RB_WARN_ON(data->buffer, event_len < len))
|
| 6006 | goto out;
|
| 6007 |
|
| 6008 | item = ring_buffer_event_data(event);
|
| 6009 | item->size = size;
|
| 6010 | memcpy(item->str, rb_string, size);
|
| 6011 |
|
| 6012 | if (nested) {
|
| 6013 | data->bytes_alloc_nested += event_len;
|
| 6014 | data->bytes_written_nested += len;
|
| 6015 | data->events_nested++;
|
| 6016 | if (!data->min_size_nested || len < data->min_size_nested)
|
| 6017 | data->min_size_nested = len;
|
| 6018 | if (len > data->max_size_nested)
|
| 6019 | data->max_size_nested = len;
|
| 6020 | } else {
|
| 6021 | data->bytes_alloc += event_len;
|
| 6022 | data->bytes_written += len;
|
| 6023 | data->events++;
|
| 6024 | if (!data->min_size || len < data->min_size)
|
| 6025 | data->max_size = len;
|
| 6026 | if (len > data->max_size)
|
| 6027 | data->max_size = len;
|
| 6028 | }
|
| 6029 |
|
| 6030 | out:
|
| 6031 | ring_buffer_unlock_commit(data->buffer);
|
| 6032 |
|
| 6033 | return 0;
|
| 6034 | }
|
| 6035 |
|
| 6036 | static __init int rb_test(void *arg)
|
| 6037 | {
|
| 6038 | struct rb_test_data *data = arg;
|
| 6039 |
|
| 6040 | while (!kthread_should_stop()) {
|
| 6041 | rb_write_something(data, false);
|
| 6042 | data->cnt++;
|
| 6043 |
|
| 6044 | set_current_state(TASK_INTERRUPTIBLE);
|
| 6045 |
|
| 6046 | usleep_range(((data->cnt % 3) + 1) * 100, 1000);
|
| 6047 | }
|
| 6048 |
|
| 6049 | return 0;
|
| 6050 | }
|
| 6051 |
|
| 6052 | static __init void rb_ipi(void *ignore)
|
| 6053 | {
|
| 6054 | struct rb_test_data *data;
|
| 6055 | int cpu = smp_processor_id();
|
| 6056 |
|
| 6057 | data = &rb_data[cpu];
|
| 6058 | rb_write_something(data, true);
|
| 6059 | }
|
| 6060 |
|
| 6061 | static __init int rb_hammer_test(void *arg)
|
| 6062 | {
|
| 6063 | while (!kthread_should_stop()) {
|
| 6064 |
|
| 6065 |
|
| 6066 | smp_call_function(rb_ipi, NULL, 1);
|
| 6067 |
|
| 6068 | schedule();
|
| 6069 | }
|
| 6070 |
|
| 6071 | return 0;
|
| 6072 | }
|
| 6073 |
|
| 6074 | static __init int test_ringbuffer(void)
|
| 6075 | {
|
| 6076 | struct task_struct *rb_hammer;
|
| 6077 | struct trace_buffer *buffer;
|
| 6078 | int cpu;
|
| 6079 | int ret = 0;
|
| 6080 |
|
| 6081 | if (security_locked_down(LOCKDOWN_TRACEFS)) {
|
| 6082 | pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
|
| 6083 | return 0;
|
| 6084 | }
|
| 6085 |
|
| 6086 | pr_info("Running ring buffer tests...\n");
|
| 6087 |
|
| 6088 | buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
|
| 6089 | if (WARN_ON(!buffer))
|
| 6090 | return 0;
|
| 6091 |
|
| 6092 |
|
| 6093 | ring_buffer_record_off(buffer);
|
| 6094 |
|
| 6095 | for_each_online_cpu(cpu) {
|
| 6096 | rb_data[cpu].buffer = buffer;
|
| 6097 | rb_data[cpu].cpu = cpu;
|
| 6098 | rb_data[cpu].cnt = cpu;
|
| 6099 | rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
|
| 6100 | cpu, "rbtester/%u");
|
| 6101 | if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
|
| 6102 | pr_cont("FAILED\n");
|
| 6103 | ret = PTR_ERR(rb_threads[cpu]);
|
| 6104 | goto out_free;
|
| 6105 | }
|
| 6106 | }
|
| 6107 |
|
| 6108 |
|
| 6109 | rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
|
| 6110 | if (WARN_ON(IS_ERR(rb_hammer))) {
|
| 6111 | pr_cont("FAILED\n");
|
| 6112 | ret = PTR_ERR(rb_hammer);
|
| 6113 | goto out_free;
|
| 6114 | }
|
| 6115 |
|
| 6116 | ring_buffer_record_on(buffer);
|
| 6117 | |
| 6118 | * Show buffer is enabled before setting rb_test_started.
|
| 6119 | * Yes there's a small race window where events could be
|
| 6120 | * dropped and the thread wont catch it. But when a ring
|
| 6121 | * buffer gets enabled, there will always be some kind of
|
| 6122 | * delay before other CPUs see it. Thus, we don't care about
|
| 6123 | * those dropped events. We care about events dropped after
|
| 6124 | * the threads see that the buffer is active.
|
| 6125 |
|
| 6126 | smp_wmb();
|
| 6127 | rb_test_started = true;
|
| 6128 |
|
| 6129 | set_current_state(TASK_INTERRUPTIBLE);
|
| 6130 | ;
|
| 6131 | schedule_timeout(10 * HZ);
|
| 6132 |
|
| 6133 | kthread_stop(rb_hammer);
|
| 6134 |
|
| 6135 | out_free:
|
| 6136 | for_each_online_cpu(cpu) {
|
| 6137 | if (!rb_threads[cpu])
|
| 6138 | break;
|
| 6139 | kthread_stop(rb_threads[cpu]);
|
| 6140 | }
|
| 6141 | if (ret) {
|
| 6142 | ring_buffer_free(buffer);
|
| 6143 | return ret;
|
| 6144 | }
|
| 6145 |
|
| 6146 |
|
| 6147 | pr_info("finished\n");
|
| 6148 | for_each_online_cpu(cpu) {
|
| 6149 | struct ring_buffer_event *event;
|
| 6150 | struct rb_test_data *data = &rb_data[cpu];
|
| 6151 | struct rb_item *item;
|
| 6152 | unsigned long total_events;
|
| 6153 | unsigned long total_dropped;
|
| 6154 | unsigned long total_written;
|
| 6155 | unsigned long total_alloc;
|
| 6156 | unsigned long total_read = 0;
|
| 6157 | unsigned long total_size = 0;
|
| 6158 | unsigned long total_len = 0;
|
| 6159 | unsigned long total_lost = 0;
|
| 6160 | unsigned long lost;
|
| 6161 | int big_event_size;
|
| 6162 | int small_event_size;
|
| 6163 |
|
| 6164 | ret = -1;
|
| 6165 |
|
| 6166 | total_events = data->events + data->events_nested;
|
| 6167 | total_written = data->bytes_written + data->bytes_written_nested;
|
| 6168 | total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
|
| 6169 | total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
|
| 6170 |
|
| 6171 | big_event_size = data->max_size + data->max_size_nested;
|
| 6172 | small_event_size = data->min_size + data->min_size_nested;
|
| 6173 |
|
| 6174 | pr_info("CPU %d:\n", cpu);
|
| 6175 | pr_info(" events: %ld\n", total_events);
|
| 6176 | pr_info(" dropped bytes: %ld\n", total_dropped);
|
| 6177 | pr_info(" alloced bytes: %ld\n", total_alloc);
|
| 6178 | pr_info(" written bytes: %ld\n", total_written);
|
| 6179 | pr_info(" biggest event: %d\n", big_event_size);
|
| 6180 | pr_info(" smallest event: %d\n", small_event_size);
|
| 6181 |
|
| 6182 | if (RB_WARN_ON(buffer, total_dropped))
|
| 6183 | break;
|
| 6184 |
|
| 6185 | ret = 0;
|
| 6186 |
|
| 6187 | while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
|
| 6188 | total_lost += lost;
|
| 6189 | item = ring_buffer_event_data(event);
|
| 6190 | total_len += ring_buffer_event_length(event);
|
| 6191 | total_size += item->size + sizeof(struct rb_item);
|
| 6192 | if (memcmp(&item->str[0], rb_string, item->size) != 0) {
|
| 6193 | pr_info("FAILED!\n");
|
| 6194 | pr_info("buffer had: %.*s\n", item->size, item->str);
|
| 6195 | pr_info("expected: %.*s\n", item->size, rb_string);
|
| 6196 | RB_WARN_ON(buffer, 1);
|
| 6197 | ret = -1;
|
| 6198 | break;
|
| 6199 | }
|
| 6200 | total_read++;
|
| 6201 | }
|
| 6202 | if (ret)
|
| 6203 | break;
|
| 6204 |
|
| 6205 | ret = -1;
|
| 6206 |
|
| 6207 | pr_info(" read events: %ld\n", total_read);
|
| 6208 | pr_info(" lost events: %ld\n", total_lost);
|
| 6209 | pr_info(" total events: %ld\n", total_lost + total_read);
|
| 6210 | pr_info(" recorded len bytes: %ld\n", total_len);
|
| 6211 | pr_info(" recorded size bytes: %ld\n", total_size);
|
| 6212 | if (total_lost) {
|
| 6213 | pr_info(" With dropped events, record len and size may not match\n"
|
| 6214 | " alloced and written from above\n");
|
| 6215 | } else {
|
| 6216 | if (RB_WARN_ON(buffer, total_len != total_alloc ||
|
| 6217 | total_size != total_written))
|
| 6218 | break;
|
| 6219 | }
|
| 6220 | if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
|
| 6221 | break;
|
| 6222 |
|
| 6223 | ret = 0;
|
| 6224 | }
|
| 6225 | if (!ret)
|
| 6226 | pr_info("Ring buffer PASSED!\n");
|
| 6227 |
|
| 6228 | ring_buffer_free(buffer);
|
| 6229 | return 0;
|
| 6230 | }
|
| 6231 |
|
| 6232 | late_initcall(test_ringbuffer);
|
| 6233 | #endif
|