ltt-tracer.h

   1 /*
   2  * Copyright (C) 2005,2006,2008 Mathieu Desnoyers (mathieu.desnoyers@polymtl.ca)
   3  *
   4  * This contains the definitions for the Linux Trace Toolkit tracer.
   5  *
   6  * Dual LGPL v2.1/GPL v2 license.
   7  */
   8
   9 #ifndef _LTT_TRACER_H
  10 #define _LTT_TRACER_H
  11
  12 #ifndef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
  13 /* Align data on its natural alignment */
  14 #define RING_BUFFER_ALIGN
  15 #endif
  16
  17 #include <linux/ringbuffer/config.h>
  18
  19 #include <stdarg.h>
  20 #include <linux/types.h>
  21 #include <linux/limits.h>
  22 #include <linux/list.h>
  23 #include <linux/cache.h>
  24 #include <linux/kernel.h>
  25 #include <linux/timex.h>
  26 #include <linux/wait.h>
  27 #include <linux/trace-clock.h>
  28 #include <asm/atomic.h>
  29 #include <asm/local.h>
  30
  31 #include "ltt-tracer-core.h"
  32
  33 /* Number of bytes to log with a read/write event */
  34 #define LTT_LOG_RW_SIZE                 32L
  35
  36 /* Maximum number of callbacks per marker */
  37 #define LTT_NR_CALLBACKS        10
  38
  39 struct ltt_serialize_closure;
  40
  41 /* Serialization callback */
  42 typedef size_t (*ltt_serialize_cb)(struct lib_ring_buffer *buf,
  43                                    size_t buf_offset,
  44                                    struct ltt_serialize_closure *closure,
  45                                    void *serialize_private,
  46                                    unsigned int stack_pos_ctx,
  47                                    int *largest_align,
  48                                    const char *fmt, va_list *args);
  49
  50 struct ltt_serialize_closure {
  51         ltt_serialize_cb *callbacks;
  52         long cb_args[LTT_NR_CALLBACKS];
  53         unsigned int cb_idx;
  54 };
  55
  56 size_t ltt_serialize_data(struct lib_ring_buffer *buf, size_t buf_offset,
  57                           struct ltt_serialize_closure *closure,
  58                           void *serialize_private, unsigned int stack_pos_ctx,
  59                           int *largest_align, const char *fmt, va_list *args);
  60
  61 enum ltt_channels {
  62         LTT_CHANNEL_METADATA,
  63         LTT_CHANNEL_FD_STATE,
  64         LTT_CHANNEL_GLOBAL_STATE,
  65         LTT_CHANNEL_IRQ_STATE,
  66         LTT_CHANNEL_MODULE_STATE,
  67         LTT_CHANNEL_NETIF_STATE,
  68         LTT_CHANNEL_SOFTIRQ_STATE,
  69         LTT_CHANNEL_SWAP_STATE,
  70         LTT_CHANNEL_SYSCALL_STATE,
  71         LTT_CHANNEL_TASK_STATE,
  72         LTT_CHANNEL_VM_STATE,
  73         LTT_CHANNEL_FS,
  74         LTT_CHANNEL_INPUT,
  75         LTT_CHANNEL_IPC,
  76         LTT_CHANNEL_KERNEL,
  77         LTT_CHANNEL_MM,
  78         LTT_CHANNEL_RCU,
  79         LTT_CHANNEL_DEFAULT,
  80 };
  81
  82 #if 0
  83 size_t ltt_serialize_printf(struct lib_ring_buffer *buf, unsigned long buf_offset,
  84                             size_t *msg_size, char *output, size_t outlen,
  85                             const char *fmt);
  86
  87 /*
  88  * Unique ID assigned to each registered probe.
  89  */
  90 enum marker_id {
  91         MARKER_ID_SET_MARKER_ID = 0,    /* Static IDs available (range 0-7) */
  92         MARKER_ID_SET_MARKER_FORMAT,
  93         MARKER_ID_COMPACT,              /* Compact IDs (range: 8-127)       */
  94         MARKER_ID_DYNAMIC,              /* Dynamic IDs (range: 128-65535)   */
  95 };
  96
  97 /* static ids 0-1 reserved for internal use. */
  98 #define MARKER_CORE_IDS         2
  99 static __inline__ enum marker_id marker_id_type(uint16_t id)
 100 {
 101         if (id < MARKER_CORE_IDS)
 102                 return (enum marker_id)id;
 103         else
 104                 return MARKER_ID_DYNAMIC;
 105 }
 106
 107 struct user_dbg_data {
 108         unsigned long avail_size;
 109         unsigned long write;
 110         unsigned long read;
 111 };
 112
 113 enum trace_mode { LTT_TRACE_NORMAL, LTT_TRACE_FLIGHT, LTT_TRACE_HYBRID };
 114
 115 #define CHANNEL_FLAG_ENABLE     (1U<<0)
 116 #define CHANNEL_FLAG_OVERWRITE  (1U<<1)
 117 #endif //0
 118
 119 #if 0
 120 /* Per-trace information - each trace/flight recorder represented by one */
 121 struct ltt_trace {
 122         /* First 32 bytes cache-hot cacheline */
 123         struct list_head list;
 124         struct ltt_chan **channels;
 125         unsigned int nr_channels;
 126         int active;
 127         /* Second 32 bytes cache-hot cacheline */
 128         struct ltt_trace_ops *ops;
 129         u32 freq_scale;
 130         u64 start_freq;
 131         u64 start_tsc;
 132         unsigned long long start_monotonic;
 133         struct timeval          start_time;
 134         struct ltt_channel_setting *settings;
 135         struct {
 136                 struct dentry                   *trace_root;
 137                 struct dentry                   *ascii_root;
 138         } dentry;
 139         struct kref kref; /* Each channel has a kref of the trace struct */
 140         struct ltt_transport *transport;
 141         struct kref ltt_transport_kref;
 142         wait_queue_head_t kref_wq; /* Place for ltt_trace_destroy to sleep */
 143         char trace_name[NAME_MAX];
 144 } ____cacheline_aligned;
 145 #endif //0
 146
 147 /*
 148  * Hardcoded event headers
 149  *
 150  * event header for a trace with active heartbeat : 27 bits timestamps
 151  *
 152  * headers are 32-bits aligned. In order to insure such alignment, a dynamic per
 153  * trace alignment value must be done.
 154  *
 155  * Remember that the C compiler does align each member on the boundary
 156  * equivalent to their own size.
 157  *
 158  * As relay subbuffers are aligned on pages, we are sure that they are 4 and 8
 159  * bytes aligned, so the buffer header and trace header are aligned.
 160  *
 161  * Event headers are aligned depending on the trace alignment option.
 162  *
 163  * Note using C structure bitfields for cross-endianness and portability
 164  * concerns.
 165  */
 166
 167 #define LTT_RESERVED_EVENTS     3
 168 #define LTT_EVENT_BITS          5
 169 #define LTT_FREE_EVENTS         ((1 << LTT_EVENT_BITS) - LTT_RESERVED_EVENTS)
 170 #define LTT_TSC_BITS            27
 171 #define LTT_TSC_MASK            ((1 << LTT_TSC_BITS) - 1)
 172
 173 struct event_header {
 174         u32 id_time;            /* 5 bits event id (MSB); 27 bits time (LSB) */
 175 };
 176
 177 /* Reservation flags */
 178 #define LTT_RFLAG_ID                    (1 << 0)
 179 #define LTT_RFLAG_ID_SIZE               (1 << 1)
 180 #define LTT_RFLAG_ID_SIZE_TSC           (1 << 2)
 181
 182 #define LTT_MAX_SMALL_SIZE              0xFFFFU
 183
 184 /*
 185  * We use asm/timex.h : cpu_khz/HZ variable in here : we might have to deal
 186  * specifically with CPU frequency scaling someday, so using an interpolation
 187  * between the start and end of buffer values is not flexible enough. Using an
 188  * immediate frequency value permits to calculate directly the times for parts
 189  * of a buffer that would be before a frequency change.
 190  *
 191  * Keep the natural field alignment for _each field_ within this structure if
 192  * you ever add/remove a field from this header. Packed attribute is not used
 193  * because gcc generates poor code on at least powerpc and mips. Don't ever
 194  * let gcc add padding between the structure elements.
 195  */
 196 struct subbuffer_header {
 197         uint64_t cycle_count_begin;     /* Cycle count at subbuffer start */
 198         uint64_t cycle_count_end;       /* Cycle count at subbuffer end */
 199         uint32_t magic_number;          /*
 200                                          * Trace magic number.
 201                                          * contains endianness information.
 202                                          */
 203         uint8_t major_version;
 204         uint8_t minor_version;
 205         uint8_t arch_size;              /* Architecture pointer size */
 206         uint8_t alignment;              /* LTT data alignment */
 207         uint64_t start_time_sec;        /* NTP-corrected start time */
 208         uint64_t start_time_usec;
 209         uint64_t start_freq;            /*
 210                                          * Frequency at trace start,
 211                                          * used all along the trace.
 212                                          */
 213         uint32_t freq_scale;            /* Frequency scaling (divisor) */
 214         uint32_t data_size;             /* Size of data in subbuffer */
 215         uint32_t sb_size;               /* Subbuffer size (include padding) */
 216         uint32_t events_lost;           /*
 217                                          * Events lost in this subbuffer since
 218                                          * the beginning of the trace.
 219                                          * (may overflow)
 220                                          */
 221         uint32_t subbuf_corrupt;        /*
 222                                          * Corrupted (lost) subbuffers since
 223                                          * the begginig of the trace.
 224                                          * (may overflow)
 225                                          */
 226         uint8_t header_end[0];          /* End of header */
 227 };
 228
 229 static inline notrace u64 lib_ring_buffer_clock_read(struct channel *chan)
 230 {
 231         return trace_clock_read64();
 232 }
 233
 234 /*
 235  * record_header_size - Calculate the header size and padding necessary.
 236  * @config: ring buffer instance configuration
 237  * @chan: channel
 238  * @offset: offset in the write buffer
 239  * @data_size: size of the payload
 240  * @pre_header_padding: padding to add before the header (output)
 241  * @rflags: reservation flags
 242  * @ctx: reservation context
 243  *
 244  * Returns the event header size (including padding).
 245  *
 246  * Important note :
 247  * The event header must be 32-bits. The total offset calculated here :
 248  *
 249  * Alignment of header struct on 32 bits (min arch size, header size)
 250  * + sizeof(header struct)  (32-bits)
 251  * + (opt) u16 (ext. event id)
 252  * + (opt) u16 (event_size)
 253  *             (if event_size == LTT_MAX_SMALL_SIZE, has ext. event size)
 254  * + (opt) u32 (ext. event size)
 255  * + (opt) u64 full TSC (aligned on min(64-bits, arch size))
 256  *
 257  * The payload must itself determine its own alignment from the biggest type it
 258  * contains.
 259  */
 260 static __inline__
 261 unsigned char record_header_size(const struct lib_ring_buffer_config *config,
 262                                  struct channel *chan, size_t offset,
 263                                  size_t data_size, size_t *pre_header_padding,
 264                                  unsigned int rflags,
 265                                  struct lib_ring_buffer_ctx *ctx)
 266 {
 267         size_t orig_offset = offset;
 268         size_t padding;
 269
 270         BUILD_BUG_ON(sizeof(struct event_header) != sizeof(u32));
 271
 272         padding = lib_ring_buffer_align(config, offset,
 273                                         sizeof(struct event_header));
 274         offset += padding;
 275         offset += sizeof(struct event_header);
 276
 277         if (unlikely(rflags)) {
 278                 switch (rflags) {
 279                 case LTT_RFLAG_ID_SIZE_TSC:
 280                         offset += sizeof(u16) + sizeof(u16);
 281                         if (data_size >= LTT_MAX_SMALL_SIZE)
 282                                 offset += sizeof(u32);
 283                         offset += ltt_align(offset, sizeof(u64));
 284                         offset += sizeof(u64);
 285                         break;
 286                 case LTT_RFLAG_ID_SIZE:
 287                         offset += sizeof(u16) + sizeof(u16);
 288                         if (data_size >= LTT_MAX_SMALL_SIZE)
 289                                 offset += sizeof(u32);
 290                         break;
 291                 case LTT_RFLAG_ID:
 292                         offset += sizeof(u16);
 293                         break;
 294                 }
 295         }
 296
 297         *pre_header_padding = padding;
 298         return offset - orig_offset;
 299 }
 300
 301 #include <linux/ringbuffer/api.h>
 302
 303 extern
 304 size_t ltt_write_event_header_slow(const struct lib_ring_buffer_config *config,
 305                                    struct lib_ring_buffer_ctx *ctx,
 306                                    u16 eID, u32 event_size);
 307
 308 /*
 309  * ltt_write_event_header
 310  *
 311  * Writes the event header to the offset (already aligned on 32-bits).
 312  *
 313  * @config: ring buffer instance configuration
 314  * @ctx: reservation context
 315  * @eID : event ID
 316  * @event_size : size of the event, excluding the event header.
 317  */
 318 static __inline__
 319 void ltt_write_event_header(const struct lib_ring_buffer_config *config,
 320                             struct lib_ring_buffer_ctx *ctx,
 321                             u16 eID, u32 event_size)
 322 {
 323         struct event_header header;
 324
 325         if (unlikely(ctx->rflags))
 326                 goto slow_path;
 327
 328         header.id_time = eID << LTT_TSC_BITS;
 329         header.id_time |= (u32)ctx->tsc & LTT_TSC_MASK;
 330         lib_ring_buffer_write(config, ctx, &header, sizeof(header));
 331
 332 slow_path:
 333         ltt_write_event_header_slow(config, ctx, eID, event_size);
 334 }
 335
 336 #if 0
 337 /*
 338  * ltt_read_event_header
 339  * buf_offset must aligned on 32 bits
 340  */
 341 static __inline__
 342 size_t ltt_read_event_header(struct ltt_chanbuf_alloc *bufa, long buf_offset,
 343                              u64 *tsc, u32 *event_size, u16 *eID,
 344                              unsigned int *rflags)
 345 {
 346         struct ltt_event_header header;
 347         u16 small_size;
 348
 349         ltt_relay_read(bufa, buf_offset, &header, sizeof(header));
 350         buf_offset += sizeof(header);
 351
 352         *event_size = INT_MAX;
 353         *eID = header.id_time >> LTT_TSC_BITS;
 354         *tsc = header.id_time & LTT_TSC_MASK;
 355
 356         switch (*eID) {
 357         case 29:
 358                 *rflags = LTT_RFLAG_ID_SIZE_TSC;
 359                 ltt_relay_read(bufa, buf_offset, eID, sizeof(u16));
 360                 buf_offset += sizeof(u16);
 361                 ltt_relay_read(bufa, buf_offset, &small_size, sizeof(u16));
 362                 buf_offset += sizeof(u16);
 363                 if (small_size == LTT_MAX_SMALL_SIZE) {
 364                         ltt_relay_read(bufa, buf_offset, event_size,
 365                                         sizeof(u32));
 366                         buf_offset += sizeof(u32);
 367                 } else
 368                         *event_size = small_size;
 369                 buf_offset += ltt_align(buf_offset, sizeof(u64));
 370                 ltt_relay_read(bufa, buf_offset, tsc, sizeof(u64));
 371                 buf_offset += sizeof(u64);
 372                 break;
 373         case 30:
 374                 *rflags = LTT_RFLAG_ID_SIZE;
 375                 ltt_relay_read(bufa, buf_offset, eID, sizeof(u16));
 376                 buf_offset += sizeof(u16);
 377                 ltt_relay_read(bufa, buf_offset, &small_size, sizeof(u16));
 378                 buf_offset += sizeof(u16);
 379                 if (small_size == LTT_MAX_SMALL_SIZE) {
 380                         ltt_relay_read(bufa, buf_offset, event_size,
 381                                         sizeof(u32));
 382                         buf_offset += sizeof(u32);
 383                 } else
 384                         *event_size = small_size;
 385                 break;
 386         case 31:
 387                 *rflags = LTT_RFLAG_ID;
 388                 ltt_relay_read(bufa, buf_offset, eID, sizeof(u16));
 389                 buf_offset += sizeof(u16);
 390                 break;
 391         default:
 392                 *rflags = 0;
 393                 break;
 394         }
 395
 396         return buf_offset;
 397 }
 398 #endif //0
 399
 400 /*
 401  * Control channels :
 402  * control/metadata
 403  * control/interrupts
 404  * control/...
 405  *
 406  * cpu channel :
 407  * cpu
 408  */
 409 #define LTT_RELAY_ROOT                  "ltt"
 410
 411 #define LTT_METADATA_CHANNEL            "metadata_state"
 412 #define LTT_FD_STATE_CHANNEL            "fd_state"
 413 #define LTT_GLOBAL_STATE_CHANNEL        "global_state"
 414 #define LTT_IRQ_STATE_CHANNEL           "irq_state"
 415 #define LTT_MODULE_STATE_CHANNEL        "module_state"
 416 #define LTT_NETIF_STATE_CHANNEL         "netif_state"
 417 #define LTT_SOFTIRQ_STATE_CHANNEL       "softirq_state"
 418 #define LTT_SWAP_STATE_CHANNEL          "swap_state"
 419 #define LTT_SYSCALL_STATE_CHANNEL       "syscall_state"
 420 #define LTT_TASK_STATE_CHANNEL          "task_state"
 421 #define LTT_VM_STATE_CHANNEL            "vm_state"
 422 #define LTT_FS_CHANNEL                  "fs"
 423 #define LTT_INPUT_CHANNEL               "input"
 424 #define LTT_IPC_CHANNEL                 "ipc"
 425 #define LTT_KERNEL_CHANNEL              "kernel"
 426 #define LTT_MM_CHANNEL                  "mm"
 427 #define LTT_RCU_CHANNEL                 "rcu"
 428
 429 #define LTT_FLIGHT_PREFIX               "flight-"
 430
 431 #define LTT_ASCII                       "ascii"
 432
 433 /* Tracer properties */
 434 #define LTT_DEFAULT_SUBBUF_SIZE_LOW     65536
 435 #define LTT_DEFAULT_N_SUBBUFS_LOW       2
 436 #define LTT_DEFAULT_SUBBUF_SIZE_MED     262144
 437 #define LTT_DEFAULT_N_SUBBUFS_MED       2
 438 #define LTT_DEFAULT_SUBBUF_SIZE_HIGH    1048576
 439 #define LTT_DEFAULT_N_SUBBUFS_HIGH      2
 440 #define LTT_TRACER_MAGIC_NUMBER         0x00D6B7ED
 441 #define LTT_TRACER_VERSION_MAJOR        2
 442 #define LTT_TRACER_VERSION_MINOR        6
 443
 444 /**
 445  * ltt_write_trace_header - Write trace header
 446  * @priv: Private data (struct trace)
 447  * @header: Memory address where the information must be written to
 448  */
 449 static __inline__
 450 void ltt_write_trace_header(void *priv,
 451                             struct subbuffer_header *header)
 452 {
 453         struct ltt_session *session = priv;
 454
 455         header->magic_number = LTT_TRACER_MAGIC_NUMBER;
 456         header->major_version = LTT_TRACER_VERSION_MAJOR;
 457         header->minor_version = LTT_TRACER_VERSION_MINOR;
 458         header->arch_size = sizeof(void *);
 459         header->alignment = lib_ring_buffer_get_alignment();
 460         header->start_time_sec = session->start_time.tv_sec;
 461         header->start_time_usec = session->start_time.tv_usec;
 462         header->start_freq = session->start_freq;
 463         header->freq_scale = session->freq_scale;
 464 }
 465
 466 /*
 467  * Size reserved for high priority events (interrupts, NMI, BH) at the end of a
 468  * nearly full buffer. User space won't use this last amount of space when in
 469  * blocking mode. This space also includes the event header that would be
 470  * written by this user space event.
 471  */
 472 #define LTT_RESERVE_CRITICAL            4096
 473
 474 /* Register and unregister function pointers */
 475
 476 enum ltt_module_function {
 477         LTT_FUNCTION_RUN_FILTER,
 478         LTT_FUNCTION_FILTER_CONTROL,
 479         LTT_FUNCTION_STATEDUMP
 480 };
 481
 482 extern int ltt_module_register(enum ltt_module_function name, void *function,
 483                                struct module *owner);
 484 extern void ltt_module_unregister(enum ltt_module_function name);
 485
 486 /* Exported control function */
 487
 488 enum ltt_control_msg {
 489         LTT_CONTROL_START,
 490         LTT_CONTROL_STOP,
 491         LTT_CONTROL_CREATE_TRACE,
 492         LTT_CONTROL_DESTROY_TRACE
 493 };
 494
 495 union ltt_control_args {
 496         struct {
 497                 enum trace_mode mode;
 498                 unsigned int subbuf_size_low;
 499                 unsigned int n_subbufs_low;
 500                 unsigned int subbuf_size_med;
 501                 unsigned int n_subbufs_med;
 502                 unsigned int subbuf_size_high;
 503                 unsigned int n_subbufs_high;
 504         } new_trace;
 505 };
 506
 507 void ltt_core_register(int (*function)(u8, void *));
 508
 509 void ltt_core_unregister(void);
 510
 511 extern int ltt_marker_connect(const char *channel, const char *mname,
 512                               const char *pname);
 513 extern int ltt_marker_disconnect(const char *channel, const char *mname,
 514                                  const char *pname);
 515 extern void ltt_dump_marker_state(struct ltt_trace *trace);
 516
 517 extern
 518 void ltt_statedump_register_kprobes_dump(void (*callback)(void *call_data));
 519 extern
 520 void ltt_statedump_unregister_kprobes_dump(void (*callback)(void *call_data));
 521
 522 extern void ltt_dump_softirq_vec(void *call_data);
 523
 524 #ifdef CONFIG_HAVE_LTT_DUMP_TABLES
 525 extern void ltt_dump_sys_call_table(void *call_data);
 526 extern void ltt_dump_idt_table(void *call_data);
 527 #else
 528 static inline void ltt_dump_sys_call_table(void *call_data)
 529 {
 530 }
 531
 532 static inline void ltt_dump_idt_table(void *call_data)
 533 {
 534 }
 535 #endif
 536
 537 #endif /* _LTT_TRACER_H */