libringbuffer/frontend_api.h

   1 #ifndef _LTTNG_RING_BUFFER_FRONTEND_API_H
   2 #define _LTTNG_RING_BUFFER_FRONTEND_API_H
   3
   4 /*
   5  * libringbuffer/frontend_api.h
   6  *
   7  * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   8  *
   9  * This library is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; only
  12  * version 2.1 of the License.
  13  *
  14  * This library is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with this library; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  *
  23  * Ring Buffer Library Synchronization Header (buffer write API).
  24  *
  25  * Author:
  26  *      Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  27  *
  28  * See ring_buffer_frontend.c for more information on wait-free
  29  * algorithms.
  30  * See frontend.h for channel allocation and read-side API.
  31  */
  32
  33 #include "frontend.h"
  34 #include <urcu-bp.h>
  35 #include <urcu/compiler.h>
  36
  37 /**
  38  * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit.
  39  *
  40  * Grabs RCU read-side lock and keeps a ring buffer nesting count as
  41  * supplementary safety net to ensure tracer client code will never
  42  * trigger an endless recursion. Returns the processor ID on success,
  43  * -EPERM on failure (nesting count too high).
  44  *
  45  * asm volatile and "memory" clobber prevent the compiler from moving
  46  * instructions out of the ring buffer nesting count. This is required to ensure
  47  * that probe side-effects which can cause recursion (e.g. unforeseen traps,
  48  * divisions by 0, ...) are triggered within the incremented nesting count
  49  * section.
  50  */
  51 static inline
  52 int lib_ring_buffer_get_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
  53 {
  54         int cpu, nesting;
  55
  56         rcu_read_lock();
  57         cpu = lttng_ust_get_cpu();
  58         nesting = ++URCU_TLS(lib_ring_buffer_nesting);
  59         cmm_barrier();
  60
  61         if (caa_unlikely(nesting > 4)) {
  62                 WARN_ON_ONCE(1);
  63                 URCU_TLS(lib_ring_buffer_nesting)--;
  64                 return -EPERM;
  65         } else
  66                 return cpu;
  67 }
  68
  69 /**
  70  * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
  71  */
  72 static inline
  73 void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
  74 {
  75         cmm_barrier();
  76         URCU_TLS(lib_ring_buffer_nesting)--;            /* TLS */
  77         rcu_read_unlock();
  78 }
  79
  80 /*
  81  * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
  82  * part of the API per se.
  83  *
  84  * returns 0 if reserve ok, or 1 if the slow path must be taken.
  85  */
  86 static inline
  87 int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
  88                                 struct lttng_ust_lib_ring_buffer_ctx *ctx,
  89                                 unsigned long *o_begin, unsigned long *o_end,
  90                                 unsigned long *o_old, size_t *before_hdr_pad)
  91 {
  92         struct channel *chan = ctx->chan;
  93         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
  94         *o_begin = v_read(config, &buf->offset);
  95         *o_old = *o_begin;
  96
  97         ctx->tsc = lib_ring_buffer_clock_read(chan);
  98         if ((int64_t) ctx->tsc == -EIO)
  99                 return 1;
 100
 101         /*
 102          * Prefetch cacheline for read because we have to read the previous
 103          * commit counter to increment it and commit seq value to compare it to
 104          * the commit counter.
 105          */
 106         //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
 107
 108         if (last_tsc_overflow(config, buf, ctx->tsc))
 109                 ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
 110
 111         if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0))
 112                 return 1;
 113
 114         ctx->slot_size = record_header_size(config, chan, *o_begin,
 115                                             before_hdr_pad, ctx);
 116         ctx->slot_size +=
 117                 lib_ring_buffer_align(*o_begin + ctx->slot_size,
 118                                       ctx->largest_align) + ctx->data_size;
 119         if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size)
 120                      > chan->backend.subbuf_size))
 121                 return 1;
 122
 123         /*
 124          * Record fits in the current buffer and we are not on a switch
 125          * boundary. It's safe to write.
 126          */
 127         *o_end = *o_begin + ctx->slot_size;
 128
 129         if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0))
 130                 /*
 131                  * The offset_end will fall at the very beginning of the next
 132                  * subbuffer.
 133                  */
 134                 return 1;
 135
 136         return 0;
 137 }
 138
 139 /**
 140  * lib_ring_buffer_reserve - Reserve space in a ring buffer.
 141  * @config: ring buffer instance configuration.
 142  * @ctx: ring buffer context. (input and output) Must be already initialized.
 143  *
 144  * Atomic wait-free slot reservation. The reserved space starts at the context
 145  * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
 146  *
 147  * Return :
 148  *  0 on success.
 149  * -EAGAIN if channel is disabled.
 150  * -ENOSPC if event size is too large for packet.
 151  * -ENOBUFS if there is currently not enough space in buffer for the event.
 152  * -EIO if data cannot be written into the buffer for any other reason.
 153  */
 154
 155 static inline
 156 int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 157                             struct lttng_ust_lib_ring_buffer_ctx *ctx)
 158 {
 159         struct channel *chan = ctx->chan;
 160         struct lttng_ust_shm_handle *handle = ctx->handle;
 161         struct lttng_ust_lib_ring_buffer *buf;
 162         unsigned long o_begin, o_end, o_old;
 163         size_t before_hdr_pad = 0;
 164
 165         if (uatomic_read(&chan->record_disabled))
 166                 return -EAGAIN;
 167
 168         if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
 169                 buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp);
 170         else
 171                 buf = shmp(handle, chan->backend.buf[0].shmp);
 172         if (uatomic_read(&buf->record_disabled))
 173                 return -EAGAIN;
 174         ctx->buf = buf;
 175
 176         /*
 177          * Perform retryable operations.
 178          */
 179         if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, &o_begin,
 180                                                  &o_end, &o_old, &before_hdr_pad)))
 181                 goto slow_path;
 182
 183         if (caa_unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end)
 184                      != o_old))
 185                 goto slow_path;
 186
 187         /*
 188          * Atomically update last_tsc. This update races against concurrent
 189          * atomic updates, but the race will always cause supplementary full TSC
 190          * record headers, never the opposite (missing a full TSC record header
 191          * when it would be needed).
 192          */
 193         save_last_tsc(config, ctx->buf, ctx->tsc);
 194
 195         /*
 196          * Push the reader if necessary
 197          */
 198         lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1);
 199
 200         /*
 201          * Clear noref flag for this subbuffer.
 202          */
 203         lib_ring_buffer_clear_noref(config, &ctx->buf->backend,
 204                                 subbuf_index(o_end - 1, chan), handle);
 205
 206         ctx->pre_offset = o_begin;
 207         ctx->buf_offset = o_begin + before_hdr_pad;
 208         return 0;
 209 slow_path:
 210         return lib_ring_buffer_reserve_slow(ctx);
 211 }
 212
 213 /**
 214  * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
 215  * @config: ring buffer instance configuration.
 216  * @buf: buffer
 217  * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
 218  *
 219  * This operation is completely reentrant : can be called while tracing is
 220  * active with absolutely no lock held.
 221  *
 222  * Note, however, that as a v_cmpxchg is used for some atomic operations and
 223  * requires to be executed locally for per-CPU buffers, this function must be
 224  * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
 225  * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
 226  */
 227 static inline
 228 void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config,
 229                             struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode,
 230                             struct lttng_ust_shm_handle *handle)
 231 {
 232         lib_ring_buffer_switch_slow(buf, mode, handle);
 233 }
 234
 235 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 236
 237 /**
 238  * lib_ring_buffer_commit - Commit an record.
 239  * @config: ring buffer instance configuration.
 240  * @ctx: ring buffer context. (input arguments only)
 241  *
 242  * Atomic unordered slot commit. Increments the commit count in the
 243  * specified sub-buffer, and delivers it if necessary.
 244  */
 245 static inline
 246 void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config,
 247                             const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 248 {
 249         struct channel *chan = ctx->chan;
 250         struct lttng_ust_shm_handle *handle = ctx->handle;
 251         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
 252         unsigned long offset_end = ctx->buf_offset;
 253         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 254         unsigned long commit_count;
 255
 256         /*
 257          * Must count record before incrementing the commit count.
 258          */
 259         subbuffer_count_record(config, &buf->backend, endidx, handle);
 260
 261         /*
 262          * Order all writes to buffer before the commit count update that will
 263          * determine that the subbuffer is full.
 264          */
 265         cmm_smp_wmb();
 266
 267         v_add(config, ctx->slot_size, &shmp_index(handle, buf->commit_hot, endidx)->cc);
 268
 269         /*
 270          * commit count read can race with concurrent OOO commit count updates.
 271          * This is only needed for lib_ring_buffer_check_deliver (for
 272          * non-polling delivery only) and for
 273          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 274          * counter to be read with the same value more than once, which could
 275          * cause :
 276          * - Multiple delivery for the same sub-buffer (which is handled
 277          *   gracefully by the reader code) if the value is for a full
 278          *   sub-buffer. It's important that we can never miss a sub-buffer
 279          *   delivery. Re-reading the value after the v_add ensures this.
 280          * - Reading a commit_count with a higher value that what was actually
 281          *   added to it for the lib_ring_buffer_write_commit_counter call
 282          *   (again caused by a concurrent committer). It does not matter,
 283          *   because this function is interested in the fact that the commit
 284          *   count reaches back the reserve offset for a specific sub-buffer,
 285          *   which is completely independent of the order.
 286          */
 287         commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc);
 288
 289         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 290                                       commit_count, endidx, handle, ctx->tsc);
 291         /*
 292          * Update used size at each commit. It's needed only for extracting
 293          * ring_buffer buffers from vmcore, after crash.
 294          */
 295         lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
 296                         offset_end, commit_count, handle);
 297 }
 298
 299 /**
 300  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 301  * @config: ring buffer instance configuration.
 302  * @ctx: ring buffer context. (input arguments only)
 303  *
 304  * Only succeeds if no other record has been written after the record to
 305  * discard. If discard fails, the record must be committed to the buffer.
 306  *
 307  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 308  */
 309 static inline
 310 int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 311                                         const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 312 {
 313         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
 314         unsigned long end_offset = ctx->pre_offset + ctx->slot_size;
 315
 316         /*
 317          * We need to ensure that if the cmpxchg succeeds and discards the
 318          * record, the next record will record a full TSC, because it cannot
 319          * rely on the last_tsc associated with the discarded record to detect
 320          * overflows. The only way to ensure this is to set the last_tsc to 0
 321          * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
 322          * timestamp in the next record.
 323          *
 324          * Note: if discard fails, we must leave the TSC in the record header.
 325          * It is needed to keep track of TSC overflows for the following
 326          * records.
 327          */
 328         save_last_tsc(config, buf, 0ULL);
 329
 330         if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset)
 331                    != end_offset))
 332                 return -EPERM;
 333         else
 334                 return 0;
 335 }
 336
 337 static inline
 338 void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 339                             struct channel *chan)
 340 {
 341         uatomic_inc(&chan->record_disabled);
 342 }
 343
 344 static inline
 345 void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 346                            struct channel *chan)
 347 {
 348         uatomic_dec(&chan->record_disabled);
 349 }
 350
 351 static inline
 352 void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 353                                     struct lttng_ust_lib_ring_buffer *buf)
 354 {
 355         uatomic_inc(&buf->record_disabled);
 356 }
 357
 358 static inline
 359 void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 360                                    struct lttng_ust_lib_ring_buffer *buf)
 361 {
 362         uatomic_dec(&buf->record_disabled);
 363 }
 364
 365 #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */