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 #include "ltt-events.h"
  33
  34 /* Number of bytes to log with a read/write event */
  35 #define LTT_LOG_RW_SIZE                 32L
  36
  37 /* Maximum number of callbacks per marker */
  38 #define LTT_NR_CALLBACKS        10
  39
  40 struct ltt_serialize_closure;
  41
  42 /* Serialization callback */
  43 typedef size_t (*ltt_serialize_cb)(struct lib_ring_buffer *buf,
  44                                    size_t buf_offset,
  45                                    struct ltt_serialize_closure *closure,
  46                                    void *serialize_private,
  47                                    unsigned int stack_pos_ctx,
  48                                    int *largest_align,
  49                                    const char *fmt, va_list *args);
  50
  51 struct ltt_serialize_closure {
  52         ltt_serialize_cb *callbacks;
  53         long cb_args[LTT_NR_CALLBACKS];
  54         unsigned int cb_idx;
  55 };
  56
  57 size_t ltt_serialize_data(struct lib_ring_buffer *buf, size_t buf_offset,
  58                           struct ltt_serialize_closure *closure,
  59                           void *serialize_private, unsigned int stack_pos_ctx,
  60                           int *largest_align, const char *fmt, va_list *args);
  61
  62 enum ltt_channels {
  63         LTT_CHANNEL_METADATA,
  64         LTT_CHANNEL_FD_STATE,
  65         LTT_CHANNEL_GLOBAL_STATE,
  66         LTT_CHANNEL_IRQ_STATE,
  67         LTT_CHANNEL_MODULE_STATE,
  68         LTT_CHANNEL_NETIF_STATE,
  69         LTT_CHANNEL_SOFTIRQ_STATE,
  70         LTT_CHANNEL_SWAP_STATE,
  71         LTT_CHANNEL_SYSCALL_STATE,
  72         LTT_CHANNEL_TASK_STATE,
  73         LTT_CHANNEL_VM_STATE,
  74         LTT_CHANNEL_FS,
  75         LTT_CHANNEL_INPUT,
  76         LTT_CHANNEL_IPC,
  77         LTT_CHANNEL_KERNEL,
  78         LTT_CHANNEL_MM,
  79         LTT_CHANNEL_RCU,
  80         LTT_CHANNEL_DEFAULT,
  81 };
  82
  83 #if 0
  84 size_t ltt_serialize_printf(struct lib_ring_buffer *buf, unsigned long buf_offset,
  85                             size_t *msg_size, char *output, size_t outlen,
  86                             const char *fmt);
  87
  88 /*
  89  * Unique ID assigned to each registered probe.
  90  */
  91 enum marker_id {
  92         MARKER_ID_SET_MARKER_ID = 0,    /* Static IDs available (range 0-7) */
  93         MARKER_ID_SET_MARKER_FORMAT,
  94         MARKER_ID_COMPACT,              /* Compact IDs (range: 8-127)       */
  95         MARKER_ID_DYNAMIC,              /* Dynamic IDs (range: 128-65535)   */
  96 };
  97
  98 /* static ids 0-1 reserved for internal use. */
  99 #define MARKER_CORE_IDS         2
 100 static __inline__ enum marker_id marker_id_type(uint16_t id)
 101 {
 102         if (id < MARKER_CORE_IDS)
 103                 return (enum marker_id)id;
 104         else
 105                 return MARKER_ID_DYNAMIC;
 106 }
 107
 108 struct user_dbg_data {
 109         unsigned long avail_size;
 110         unsigned long write;
 111         unsigned long read;
 112 };
 113
 114 enum trace_mode { LTT_TRACE_NORMAL, LTT_TRACE_FLIGHT, LTT_TRACE_HYBRID };
 115
 116 #define CHANNEL_FLAG_ENABLE     (1U<<0)
 117 #define CHANNEL_FLAG_OVERWRITE  (1U<<1)
 118 #endif //0
 119
 120 #if 0
 121 /* Per-trace information - each trace/flight recorder represented by one */
 122 struct ltt_trace {
 123         /* First 32 bytes cache-hot cacheline */
 124         struct list_head list;
 125         struct ltt_chan **channels;
 126         unsigned int nr_channels;
 127         int active;
 128         /* Second 32 bytes cache-hot cacheline */
 129         struct ltt_trace_ops *ops;
 130         u32 freq_scale;
 131         u64 start_freq;
 132         u64 start_tsc;
 133         unsigned long long start_monotonic;
 134         struct timeval          start_time;
 135         struct ltt_channel_setting *settings;
 136         struct {
 137                 struct dentry                   *trace_root;
 138                 struct dentry                   *ascii_root;
 139         } dentry;
 140         struct kref kref; /* Each channel has a kref of the trace struct */
 141         struct ltt_transport *transport;
 142         struct kref ltt_transport_kref;
 143         wait_queue_head_t kref_wq; /* Place for ltt_trace_destroy to sleep */
 144         char trace_name[NAME_MAX];
 145 } ____cacheline_aligned;
 146 #endif //0
 147
 148 /*
 149  * Hardcoded event headers
 150  *
 151  * event header for a trace with active heartbeat : 27 bits timestamps
 152  *
 153  * headers are 32-bits aligned. In order to insure such alignment, a dynamic per
 154  * trace alignment value must be done.
 155  *
 156  * Remember that the C compiler does align each member on the boundary
 157  * equivalent to their own size.
 158  *
 159  * As relay subbuffers are aligned on pages, we are sure that they are 4 and 8
 160  * bytes aligned, so the buffer header and trace header are aligned.
 161  *
 162  * Event headers are aligned depending on the trace alignment option.
 163  *
 164  * Note using C structure bitfields for cross-endianness and portability
 165  * concerns.
 166  */
 167
 168 #define LTT_RESERVED_EVENTS     3
 169 #define LTT_EVENT_BITS          5
 170 #define LTT_FREE_EVENTS         ((1 << LTT_EVENT_BITS) - LTT_RESERVED_EVENTS)
 171 #define LTT_TSC_BITS            27
 172 #define LTT_TSC_MASK            ((1 << LTT_TSC_BITS) - 1)
 173
 174 struct event_header {
 175         u32 id_time;            /* 5 bits event id (MSB); 27 bits time (LSB) */
 176 };
 177
 178 /* Reservation flags */
 179 #define LTT_RFLAG_ID                    (1 << 0)
 180 #define LTT_RFLAG_ID_SIZE               (1 << 1)
 181 #define LTT_RFLAG_ID_SIZE_TSC           (1 << 2)
 182
 183 #define LTT_MAX_SMALL_SIZE              0xFFFFU
 184
 185 /*
 186  * We use asm/timex.h : cpu_khz/HZ variable in here : we might have to deal
 187  * specifically with CPU frequency scaling someday, so using an interpolation
 188  * between the start and end of buffer values is not flexible enough. Using an
 189  * immediate frequency value permits to calculate directly the times for parts
 190  * of a buffer that would be before a frequency change.
 191  *
 192  * Keep the natural field alignment for _each field_ within this structure if
 193  * you ever add/remove a field from this header. Packed attribute is not used
 194  * because gcc generates poor code on at least powerpc and mips. Don't ever
 195  * let gcc add padding between the structure elements.
 196  */
 197 struct packet_header {
 198         uint32_t magic;                 /*
 199                                          * Trace magic number.
 200                                          * contains endianness information.
 201                                          */
 202         uint8_t trace_uuid[16];
 203         uint32_t stream_id;
 204         uint64_t timestamp_begin;       /* Cycle count at subbuffer start */
 205         uint64_t timestamp_end; /* Cycle count at subbuffer end */
 206         uint32_t content_size;          /* Size of data in subbuffer */
 207         uint32_t packet_size;           /* Subbuffer size (include padding) */
 208         uint32_t events_lost;           /*
 209                                          * Events lost in this subbuffer since
 210                                          * the beginning of the trace.
 211                                          * (may overflow)
 212                                          */
 213         /* TODO: move to metadata */
 214 #if 0
 215         uint8_t major_version;
 216         uint8_t minor_version;
 217         uint8_t arch_size;              /* Architecture pointer size */
 218         uint8_t alignment;              /* LTT data alignment */
 219         uint64_t start_time_sec;        /* NTP-corrected start time */
 220         uint64_t start_time_usec;
 221         uint64_t start_freq;            /*
 222                                          * Frequency at trace start,
 223                                          * used all along the trace.
 224                                          */
 225         uint32_t freq_scale;            /* Frequency scaling (divisor) */
 226 #endif //0
 227         uint8_t header_end[0];          /* End of header */
 228 };
 229
 230 static inline notrace u64 lib_ring_buffer_clock_read(struct channel *chan)
 231 {
 232         return trace_clock_read64();
 233 }
 234
 235 /*
 236  * record_header_size - Calculate the header size and padding necessary.
 237  * @config: ring buffer instance configuration
 238  * @chan: channel
 239  * @offset: offset in the write buffer
 240  * @data_size: size of the payload
 241  * @pre_header_padding: padding to add before the header (output)
 242  * @rflags: reservation flags
 243  * @ctx: reservation context
 244  *
 245  * Returns the event header size (including padding).
 246  *
 247  * Important note :
 248  * The event header must be 32-bits. The total offset calculated here :
 249  *
 250  * Alignment of header struct on 32 bits (min arch size, header size)
 251  * + sizeof(header struct)  (32-bits)
 252  * + (opt) u16 (ext. event id)
 253  * + (opt) u16 (event_size)
 254  *             (if event_size == LTT_MAX_SMALL_SIZE, has ext. event size)
 255  * + (opt) u32 (ext. event size)
 256  * + (opt) u64 full TSC (aligned on min(64-bits, arch size))
 257  *
 258  * The payload must itself determine its own alignment from the biggest type it
 259  * contains.
 260  */
 261 static __inline__
 262 unsigned char record_header_size(const struct lib_ring_buffer_config *config,
 263                                  struct channel *chan, size_t offset,
 264                                  size_t data_size, size_t *pre_header_padding,
 265                                  unsigned int rflags,
 266                                  struct lib_ring_buffer_ctx *ctx)
 267 {
 268         size_t orig_offset = offset;
 269         size_t padding;
 270
 271         BUILD_BUG_ON(sizeof(struct event_header) != sizeof(u32));
 272
 273         padding = lib_ring_buffer_align(config, offset,
 274                                         sizeof(struct event_header));
 275         offset += padding;
 276         offset += sizeof(struct event_header);
 277
 278         if (unlikely(rflags)) {
 279                 switch (rflags) {
 280                 case LTT_RFLAG_ID_SIZE_TSC:
 281                         offset += sizeof(u16) + sizeof(u16);
 282                         if (data_size >= LTT_MAX_SMALL_SIZE)
 283                                 offset += sizeof(u32);
 284                         offset += ltt_align(offset, sizeof(u64));
 285                         offset += sizeof(u64);
 286                         break;
 287                 case LTT_RFLAG_ID_SIZE:
 288                         offset += sizeof(u16) + sizeof(u16);
 289                         if (data_size >= LTT_MAX_SMALL_SIZE)
 290                                 offset += sizeof(u32);
 291                         break;
 292                 case LTT_RFLAG_ID:
 293                         offset += sizeof(u16);
 294                         break;
 295                 }
 296         }
 297
 298         *pre_header_padding = padding;
 299         return offset - orig_offset;
 300 }
 301
 302 #include <linux/ringbuffer/api.h>
 303
 304 extern
 305 void ltt_write_event_header_slow(const struct lib_ring_buffer_config *config,
 306                                  struct lib_ring_buffer_ctx *ctx,
 307                                  u16 eID, u32 event_size);
 308
 309 /*
 310  * ltt_write_event_header
 311  *
 312  * Writes the event header to the offset (already aligned on 32-bits).
 313  *
 314  * @config: ring buffer instance configuration
 315  * @ctx: reservation context
 316  * @eID : event ID
 317  * @event_size : size of the event, excluding the event header.
 318  */
 319 static __inline__
 320 void ltt_write_event_header(const struct lib_ring_buffer_config *config,
 321                             struct lib_ring_buffer_ctx *ctx,
 322                             u16 eID, u32 event_size)
 323 {
 324         struct event_header header;
 325
 326         if (unlikely(ctx->rflags))
 327                 goto slow_path;
 328
 329         header.id_time = eID << LTT_TSC_BITS;
 330         header.id_time |= (u32)ctx->tsc & LTT_TSC_MASK;
 331         lib_ring_buffer_write(config, ctx, &header, sizeof(header));
 332
 333 slow_path:
 334         ltt_write_event_header_slow(config, ctx, eID, event_size);
 335 }
 336
 337 #if 0
 338 /*
 339  * ltt_read_event_header
 340  * buf_offset must aligned on 32 bits
 341  */
 342 static __inline__
 343 size_t ltt_read_event_header(struct ltt_chanbuf_alloc *bufa, long buf_offset,
 344                              u64 *tsc, u32 *event_size, u16 *eID,
 345                              unsigned int *rflags)
 346 {
 347         struct ltt_event_header header;
 348         u16 small_size;
 349
 350         ltt_relay_read(bufa, buf_offset, &header, sizeof(header));
 351         buf_offset += sizeof(header);
 352
 353         *event_size = INT_MAX;
 354         *eID = header.id_time >> LTT_TSC_BITS;
 355         *tsc = header.id_time & LTT_TSC_MASK;
 356
 357         switch (*eID) {
 358         case 29:
 359                 *rflags = LTT_RFLAG_ID_SIZE_TSC;
 360                 ltt_relay_read(bufa, buf_offset, eID, sizeof(u16));
 361                 buf_offset += sizeof(u16);
 362                 ltt_relay_read(bufa, buf_offset, &small_size, sizeof(u16));
 363                 buf_offset += sizeof(u16);
 364                 if (small_size == LTT_MAX_SMALL_SIZE) {
 365                         ltt_relay_read(bufa, buf_offset, event_size,
 366                                         sizeof(u32));
 367                         buf_offset += sizeof(u32);
 368                 } else
 369                         *event_size = small_size;
 370                 buf_offset += ltt_align(buf_offset, sizeof(u64));
 371                 ltt_relay_read(bufa, buf_offset, tsc, sizeof(u64));
 372                 buf_offset += sizeof(u64);
 373                 break;
 374         case 30:
 375                 *rflags = LTT_RFLAG_ID_SIZE;
 376                 ltt_relay_read(bufa, buf_offset, eID, sizeof(u16));
 377                 buf_offset += sizeof(u16);
 378                 ltt_relay_read(bufa, buf_offset, &small_size, sizeof(u16));
 379                 buf_offset += sizeof(u16);
 380                 if (small_size == LTT_MAX_SMALL_SIZE) {
 381                         ltt_relay_read(bufa, buf_offset, event_size,
 382                                         sizeof(u32));
 383                         buf_offset += sizeof(u32);
 384                 } else
 385                         *event_size = small_size;
 386                 break;
 387         case 31:
 388                 *rflags = LTT_RFLAG_ID;
 389                 ltt_relay_read(bufa, buf_offset, eID, sizeof(u16));
 390                 buf_offset += sizeof(u16);
 391                 break;
 392         default:
 393                 *rflags = 0;
 394                 break;
 395         }
 396
 397         return buf_offset;
 398 }
 399 #endif //0
 400
 401 /* Tracer properties */
 402 #define CTF_MAGIC_NUMBER                0xC1FC1FC1
 403 #define LTT_TRACER_VERSION_MAJOR        3
 404 #define LTT_TRACER_VERSION_MINOR        0
 405
 406 /**
 407  * ltt_write_trace_header - Write trace header
 408  * @priv: Private data (struct trace)
 409  * @header: Memory address where the information must be written to
 410  */
 411 static __inline__
 412 void write_trace_header(const struct lib_ring_buffer_config *config,
 413                         struct packet_header *header)
 414 {
 415         header->magic = CTF_MAGIC_NUMBER;
 416 #if 0
 417         /* TODO: move start time to metadata */
 418         header->major_version = LTT_TRACER_VERSION_MAJOR;
 419         header->minor_version = LTT_TRACER_VERSION_MINOR;
 420         header->arch_size = sizeof(void *);
 421         header->alignment = lib_ring_buffer_get_alignment(config);
 422         header->start_time_sec = ltt_chan->session->start_time.tv_sec;
 423         header->start_time_usec = ltt_chan->session->start_time.tv_usec;
 424         header->start_freq = ltt_chan->session->start_freq;
 425         header->freq_scale = ltt_chan->session->freq_scale;
 426 #endif //0
 427 }
 428
 429 /*
 430  * Size reserved for high priority events (interrupts, NMI, BH) at the end of a
 431  * nearly full buffer. User space won't use this last amount of space when in
 432  * blocking mode. This space also includes the event header that would be
 433  * written by this user space event.
 434  */
 435 #define LTT_RESERVE_CRITICAL            4096
 436
 437 /* Register and unregister function pointers */
 438
 439 enum ltt_module_function {
 440         LTT_FUNCTION_RUN_FILTER,
 441         LTT_FUNCTION_FILTER_CONTROL,
 442         LTT_FUNCTION_STATEDUMP
 443 };
 444
 445 extern int ltt_module_register(enum ltt_module_function name, void *function,
 446                                struct module *owner);
 447 extern void ltt_module_unregister(enum ltt_module_function name);
 448
 449 /* Exported control function */
 450
 451 void ltt_core_register(int (*function)(u8, void *));
 452
 453 void ltt_core_unregister(void);
 454
 455 extern
 456 void ltt_statedump_register_kprobes_dump(void (*callback)(void *call_data));
 457 extern
 458 void ltt_statedump_unregister_kprobes_dump(void (*callback)(void *call_data));
 459
 460 extern void ltt_dump_softirq_vec(void *call_data);
 461
 462 #ifdef CONFIG_HAVE_LTT_DUMP_TABLES
 463 extern void ltt_dump_sys_call_table(void *call_data);
 464 extern void ltt_dump_idt_table(void *call_data);
 465 #else
 466 static inline void ltt_dump_sys_call_table(void *call_data)
 467 {
 468 }
 469
 470 static inline void ltt_dump_idt_table(void *call_data)
 471 {
 472 }
 473 #endif
 474
 475 #endif /* _LTT_TRACER_H */