libringbuffer/frontend_api.h

   1 #ifndef _LINUX_RING_BUFFER_FRONTEND_API_H
   2 #define _LINUX_RING_BUFFER_FRONTEND_API_H
   3
   4 /*
   5  * linux/ringbuffer/frontend_api.h
   6  *
   7  * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   8  *
   9  * Ring Buffer Library Synchronization Header (buffer write API).
  10  *
  11  * Author:
  12  *      Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  13  *
  14  * See ring_buffer_frontend.c for more information on wait-free algorithms.
  15  * See linux/ringbuffer/frontend.h for channel allocation and read-side API.
  16  *
  17  * Dual LGPL v2.1/GPL v2 license.
  18  */
  19
  20 #include "frontend.h"
  21 #include <urcu-bp.h>
  22 #include <urcu/compiler.h>
  23
  24 /**
  25  * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit.
  26  *
  27  * Grabs RCU read-side lock and keeps a ring buffer nesting count as
  28  * supplementary safety net to ensure tracer client code will never
  29  * trigger an endless recursion. Returns the processor ID on success,
  30  * -EPERM on failure (nesting count too high).
  31  *
  32  * asm volatile and "memory" clobber prevent the compiler from moving
  33  * instructions out of the ring buffer nesting count. This is required to ensure
  34  * that probe side-effects which can cause recursion (e.g. unforeseen traps,
  35  * divisions by 0, ...) are triggered within the incremented nesting count
  36  * section.
  37  */
  38 static inline
  39 int lib_ring_buffer_get_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
  40 {
  41         int cpu, nesting;
  42
  43         rcu_read_lock();
  44         cpu = lttng_ust_get_cpu();
  45         nesting = ++lib_ring_buffer_nesting;    /* TLS */
  46         cmm_barrier();
  47
  48         if (caa_unlikely(nesting > 4)) {
  49                 WARN_ON_ONCE(1);
  50                 lib_ring_buffer_nesting--;      /* TLS */
  51                 rcu_read_unlock();
  52                 return -EPERM;
  53         } else
  54                 return cpu;
  55 }
  56
  57 /**
  58  * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
  59  */
  60 static inline
  61 void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
  62 {
  63         cmm_barrier();
  64         lib_ring_buffer_nesting--;              /* TLS */
  65         rcu_read_unlock();
  66 }
  67
  68 /*
  69  * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
  70  * part of the API per se.
  71  *
  72  * returns 0 if reserve ok, or 1 if the slow path must be taken.
  73  */
  74 static inline
  75 int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
  76                                 struct lttng_ust_lib_ring_buffer_ctx *ctx,
  77                                 unsigned long *o_begin, unsigned long *o_end,
  78                                 unsigned long *o_old, size_t *before_hdr_pad)
  79 {
  80         struct channel *chan = ctx->chan;
  81         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
  82         *o_begin = v_read(config, &buf->offset);
  83         *o_old = *o_begin;
  84
  85         ctx->tsc = lib_ring_buffer_clock_read(chan);
  86         if ((int64_t) ctx->tsc == -EIO)
  87                 return 1;
  88
  89         /*
  90          * Prefetch cacheline for read because we have to read the previous
  91          * commit counter to increment it and commit seq value to compare it to
  92          * the commit counter.
  93          */
  94         //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
  95
  96         /*
  97          * Because we don't use any timer in the application, we
  98          * currently cannot guarantee that we have frequent
  99          * events that let us detect 27-bit overflows.
 100          * Therefore, for now, we force event headers
 101          * to contain 64-bit timestamps.
 102          */
 103         ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
 104 #if 0
 105         if (last_tsc_overflow(config, buf, ctx->tsc))
 106                 ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
 107 #endif //0
 108
 109         if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0))
 110                 return 1;
 111
 112         ctx->slot_size = record_header_size(config, chan, *o_begin,
 113                                             before_hdr_pad, ctx);
 114         ctx->slot_size +=
 115                 lib_ring_buffer_align(*o_begin + ctx->slot_size,
 116                                       ctx->largest_align) + ctx->data_size;
 117         if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size)
 118                      > chan->backend.subbuf_size))
 119                 return 1;
 120
 121         /*
 122          * Record fits in the current buffer and we are not on a switch
 123          * boundary. It's safe to write.
 124          */
 125         *o_end = *o_begin + ctx->slot_size;
 126
 127         if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0))
 128                 /*
 129                  * The offset_end will fall at the very beginning of the next
 130                  * subbuffer.
 131                  */
 132                 return 1;
 133
 134         return 0;
 135 }
 136
 137 /**
 138  * lib_ring_buffer_reserve - Reserve space in a ring buffer.
 139  * @config: ring buffer instance configuration.
 140  * @ctx: ring buffer context. (input and output) Must be already initialized.
 141  *
 142  * Atomic wait-free slot reservation. The reserved space starts at the context
 143  * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
 144  *
 145  * Return :
 146  *  0 on success.
 147  * -EAGAIN if channel is disabled.
 148  * -ENOSPC if event size is too large for packet.
 149  * -ENOBUFS if there is currently not enough space in buffer for the event.
 150  * -EIO if data cannot be written into the buffer for any other reason.
 151  */
 152
 153 static inline
 154 int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 155                             struct lttng_ust_lib_ring_buffer_ctx *ctx)
 156 {
 157         struct channel *chan = ctx->chan;
 158         struct lttng_ust_shm_handle *handle = ctx->handle;
 159         struct lttng_ust_lib_ring_buffer *buf;
 160         unsigned long o_begin, o_end, o_old;
 161         size_t before_hdr_pad = 0;
 162
 163         if (uatomic_read(&chan->record_disabled))
 164                 return -EAGAIN;
 165
 166         if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
 167                 buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp);
 168         else
 169                 buf = shmp(handle, chan->backend.buf[0].shmp);
 170         if (uatomic_read(&buf->record_disabled))
 171                 return -EAGAIN;
 172         ctx->buf = buf;
 173
 174         /*
 175          * Perform retryable operations.
 176          */
 177         if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, &o_begin,
 178                                                  &o_end, &o_old, &before_hdr_pad)))
 179                 goto slow_path;
 180
 181         if (caa_unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end)
 182                      != o_old))
 183                 goto slow_path;
 184
 185         /*
 186          * Atomically update last_tsc. This update races against concurrent
 187          * atomic updates, but the race will always cause supplementary full TSC
 188          * record headers, never the opposite (missing a full TSC record header
 189          * when it would be needed).
 190          */
 191         save_last_tsc(config, ctx->buf, ctx->tsc);
 192
 193         /*
 194          * Push the reader if necessary
 195          */
 196         lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1);
 197
 198         /*
 199          * Clear noref flag for this subbuffer.
 200          */
 201         lib_ring_buffer_clear_noref(config, &ctx->buf->backend,
 202                                 subbuf_index(o_end - 1, chan), handle);
 203
 204         ctx->pre_offset = o_begin;
 205         ctx->buf_offset = o_begin + before_hdr_pad;
 206         return 0;
 207 slow_path:
 208         return lib_ring_buffer_reserve_slow(ctx);
 209 }
 210
 211 /**
 212  * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
 213  * @config: ring buffer instance configuration.
 214  * @buf: buffer
 215  * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
 216  *
 217  * This operation is completely reentrant : can be called while tracing is
 218  * active with absolutely no lock held.
 219  *
 220  * Note, however, that as a v_cmpxchg is used for some atomic operations and
 221  * requires to be executed locally for per-CPU buffers, this function must be
 222  * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
 223  * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
 224  */
 225 static inline
 226 void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config,
 227                             struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode,
 228                             struct lttng_ust_shm_handle *handle)
 229 {
 230         lib_ring_buffer_switch_slow(buf, mode, handle);
 231 }
 232
 233 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 234
 235 /**
 236  * lib_ring_buffer_commit - Commit an record.
 237  * @config: ring buffer instance configuration.
 238  * @ctx: ring buffer context. (input arguments only)
 239  *
 240  * Atomic unordered slot commit. Increments the commit count in the
 241  * specified sub-buffer, and delivers it if necessary.
 242  */
 243 static inline
 244 void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config,
 245                             const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 246 {
 247         struct channel *chan = ctx->chan;
 248         struct lttng_ust_shm_handle *handle = ctx->handle;
 249         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
 250         unsigned long offset_end = ctx->buf_offset;
 251         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 252         unsigned long commit_count;
 253
 254         /*
 255          * Must count record before incrementing the commit count.
 256          */
 257         subbuffer_count_record(config, &buf->backend, endidx, handle);
 258
 259         /*
 260          * Order all writes to buffer before the commit count update that will
 261          * determine that the subbuffer is full.
 262          */
 263         cmm_smp_wmb();
 264
 265         v_add(config, ctx->slot_size, &shmp_index(handle, buf->commit_hot, endidx)->cc);
 266
 267         /*
 268          * commit count read can race with concurrent OOO commit count updates.
 269          * This is only needed for lib_ring_buffer_check_deliver (for
 270          * non-polling delivery only) and for
 271          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 272          * counter to be read with the same value more than once, which could
 273          * cause :
 274          * - Multiple delivery for the same sub-buffer (which is handled
 275          *   gracefully by the reader code) if the value is for a full
 276          *   sub-buffer. It's important that we can never miss a sub-buffer
 277          *   delivery. Re-reading the value after the v_add ensures this.
 278          * - Reading a commit_count with a higher value that what was actually
 279          *   added to it for the lib_ring_buffer_write_commit_counter call
 280          *   (again caused by a concurrent committer). It does not matter,
 281          *   because this function is interested in the fact that the commit
 282          *   count reaches back the reserve offset for a specific sub-buffer,
 283          *   which is completely independent of the order.
 284          */
 285         commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc);
 286
 287         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 288                                       commit_count, endidx, handle);
 289         /*
 290          * Update used size at each commit. It's needed only for extracting
 291          * ring_buffer buffers from vmcore, after crash.
 292          */
 293         lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
 294                                              ctx->buf_offset, commit_count,
 295                                              ctx->slot_size, handle);
 296 }
 297
 298 /**
 299  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 300  * @config: ring buffer instance configuration.
 301  * @ctx: ring buffer context. (input arguments only)
 302  *
 303  * Only succeeds if no other record has been written after the record to
 304  * discard. If discard fails, the record must be committed to the buffer.
 305  *
 306  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 307  */
 308 static inline
 309 int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 310                                         const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 311 {
 312         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
 313         unsigned long end_offset = ctx->pre_offset + ctx->slot_size;
 314
 315         /*
 316          * We need to ensure that if the cmpxchg succeeds and discards the
 317          * record, the next record will record a full TSC, because it cannot
 318          * rely on the last_tsc associated with the discarded record to detect
 319          * overflows. The only way to ensure this is to set the last_tsc to 0
 320          * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
 321          * timestamp in the next record.
 322          *
 323          * Note: if discard fails, we must leave the TSC in the record header.
 324          * It is needed to keep track of TSC overflows for the following
 325          * records.
 326          */
 327         save_last_tsc(config, buf, 0ULL);
 328
 329         if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset)
 330                    != end_offset))
 331                 return -EPERM;
 332         else
 333                 return 0;
 334 }
 335
 336 static inline
 337 void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 338                             struct channel *chan)
 339 {
 340         uatomic_inc(&chan->record_disabled);
 341 }
 342
 343 static inline
 344 void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 345                            struct channel *chan)
 346 {
 347         uatomic_dec(&chan->record_disabled);
 348 }
 349
 350 static inline
 351 void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 352                                     struct lttng_ust_lib_ring_buffer *buf)
 353 {
 354         uatomic_inc(&buf->record_disabled);
 355 }
 356
 357 static inline
 358 void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 359                                    struct lttng_ust_lib_ring_buffer *buf)
 360 {
 361         uatomic_dec(&buf->record_disabled);
 362 }
 363
 364 #endif /* _LINUX_RING_BUFFER_FRONTEND_API_H */