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  * Keeps a ring buffer nesting count as supplementary safety net to
  41  * ensure tracer client code will never trigger an endless recursion.
  42  * Returns the processor ID on success, -EPERM on failure (nesting count
  43  * 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         cpu = lttng_ust_get_cpu();
  57         nesting = ++URCU_TLS(lib_ring_buffer_nesting);
  58         cmm_barrier();
  59
  60         if (caa_unlikely(nesting > 4)) {
  61                 WARN_ON_ONCE(1);
  62                 URCU_TLS(lib_ring_buffer_nesting)--;
  63                 return -EPERM;
  64         } else
  65                 return cpu;
  66 }
  67
  68 /**
  69  * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
  70  */
  71 static inline
  72 void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
  73 {
  74         cmm_barrier();
  75         URCU_TLS(lib_ring_buffer_nesting)--;            /* TLS */
  76 }
  77
  78 /*
  79  * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
  80  * part of the API per se.
  81  *
  82  * returns 0 if reserve ok, or 1 if the slow path must be taken.
  83  */
  84 static inline
  85 int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
  86                                 struct lttng_ust_lib_ring_buffer_ctx *ctx,
  87                                 void *client_ctx,
  88                                 unsigned long *o_begin, unsigned long *o_end,
  89                                 unsigned long *o_old, size_t *before_hdr_pad)
  90 {
  91         struct channel *chan = ctx->chan;
  92         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
  93         *o_begin = v_read(config, &buf->offset);
  94         *o_old = *o_begin;
  95
  96         ctx->tsc = lib_ring_buffer_clock_read(chan);
  97         if ((int64_t) ctx->tsc == -EIO)
  98                 return 1;
  99
 100         /*
 101          * Prefetch cacheline for read because we have to read the previous
 102          * commit counter to increment it and commit seq value to compare it to
 103          * the commit counter.
 104          */
 105         //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
 106
 107         if (last_tsc_overflow(config, buf, ctx->tsc))
 108                 ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
 109
 110         if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0))
 111                 return 1;
 112
 113         ctx->slot_size = record_header_size(config, chan, *o_begin,
 114                                             before_hdr_pad, ctx, client_ctx);
 115         ctx->slot_size +=
 116                 lib_ring_buffer_align(*o_begin + ctx->slot_size,
 117                                       ctx->largest_align) + ctx->data_size;
 118         if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size)
 119                      > chan->backend.subbuf_size))
 120                 return 1;
 121
 122         /*
 123          * Record fits in the current buffer and we are not on a switch
 124          * boundary. It's safe to write.
 125          */
 126         *o_end = *o_begin + ctx->slot_size;
 127
 128         if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0))
 129                 /*
 130                  * The offset_end will fall at the very beginning of the next
 131                  * subbuffer.
 132                  */
 133                 return 1;
 134
 135         return 0;
 136 }
 137
 138 /**
 139  * lib_ring_buffer_reserve - Reserve space in a ring buffer.
 140  * @config: ring buffer instance configuration.
 141  * @ctx: ring buffer context. (input and output) Must be already initialized.
 142  *
 143  * Atomic wait-free slot reservation. The reserved space starts at the context
 144  * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
 145  *
 146  * Return :
 147  *  0 on success.
 148  * -EAGAIN if channel is disabled.
 149  * -ENOSPC if event size is too large for packet.
 150  * -ENOBUFS if there is currently not enough space in buffer for the event.
 151  * -EIO if data cannot be written into the buffer for any other reason.
 152  */
 153
 154 static inline
 155 int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 156                             struct lttng_ust_lib_ring_buffer_ctx *ctx,
 157                             void *client_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 (caa_unlikely(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 (caa_unlikely(!buf))
 173                 return -EIO;
 174         if (caa_unlikely(uatomic_read(&buf->record_disabled)))
 175                 return -EAGAIN;
 176         ctx->buf = buf;
 177
 178         /*
 179          * Perform retryable operations.
 180          */
 181         if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, client_ctx, &o_begin,
 182                                                  &o_end, &o_old, &before_hdr_pad)))
 183                 goto slow_path;
 184
 185         if (caa_unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end)
 186                      != o_old))
 187                 goto slow_path;
 188
 189         /*
 190          * Atomically update last_tsc. This update races against concurrent
 191          * atomic updates, but the race will always cause supplementary full TSC
 192          * record headers, never the opposite (missing a full TSC record header
 193          * when it would be needed).
 194          */
 195         save_last_tsc(config, ctx->buf, ctx->tsc);
 196
 197         /*
 198          * Push the reader if necessary
 199          */
 200         lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1);
 201
 202         /*
 203          * Clear noref flag for this subbuffer.
 204          */
 205         lib_ring_buffer_clear_noref(config, &ctx->buf->backend,
 206                                 subbuf_index(o_end - 1, chan), handle);
 207
 208         ctx->pre_offset = o_begin;
 209         ctx->buf_offset = o_begin + before_hdr_pad;
 210         return 0;
 211 slow_path:
 212         return lib_ring_buffer_reserve_slow(ctx, client_ctx);
 213 }
 214
 215 /**
 216  * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
 217  * @config: ring buffer instance configuration.
 218  * @buf: buffer
 219  * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
 220  *
 221  * This operation is completely reentrant : can be called while tracing is
 222  * active with absolutely no lock held.
 223  *
 224  * Note, however, that as a v_cmpxchg is used for some atomic operations and
 225  * requires to be executed locally for per-CPU buffers, this function must be
 226  * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
 227  * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
 228  */
 229 static inline
 230 void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config,
 231                             struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode,
 232                             struct lttng_ust_shm_handle *handle)
 233 {
 234         lib_ring_buffer_switch_slow(buf, mode, handle);
 235 }
 236
 237 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 238
 239 /**
 240  * lib_ring_buffer_commit - Commit an record.
 241  * @config: ring buffer instance configuration.
 242  * @ctx: ring buffer context. (input arguments only)
 243  *
 244  * Atomic unordered slot commit. Increments the commit count in the
 245  * specified sub-buffer, and delivers it if necessary.
 246  */
 247 static inline
 248 void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config,
 249                             const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 250 {
 251         struct channel *chan = ctx->chan;
 252         struct lttng_ust_shm_handle *handle = ctx->handle;
 253         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
 254         unsigned long offset_end = ctx->buf_offset;
 255         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 256         unsigned long commit_count;
 257         struct commit_counters_hot *cc_hot = shmp_index(handle,
 258                                                 buf->commit_hot, endidx);
 259
 260         if (caa_unlikely(!cc_hot))
 261                 return;
 262
 263         /*
 264          * Must count record before incrementing the commit count.
 265          */
 266         subbuffer_count_record(config, ctx, &buf->backend, endidx, handle);
 267
 268         /*
 269          * Order all writes to buffer before the commit count update that will
 270          * determine that the subbuffer is full.
 271          */
 272         cmm_smp_wmb();
 273
 274         v_add(config, ctx->slot_size, &cc_hot->cc);
 275
 276         /*
 277          * commit count read can race with concurrent OOO commit count updates.
 278          * This is only needed for lib_ring_buffer_check_deliver (for
 279          * non-polling delivery only) and for
 280          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 281          * counter to be read with the same value more than once, which could
 282          * cause :
 283          * - Multiple delivery for the same sub-buffer (which is handled
 284          *   gracefully by the reader code) if the value is for a full
 285          *   sub-buffer. It's important that we can never miss a sub-buffer
 286          *   delivery. Re-reading the value after the v_add ensures this.
 287          * - Reading a commit_count with a higher value that what was actually
 288          *   added to it for the lib_ring_buffer_write_commit_counter call
 289          *   (again caused by a concurrent committer). It does not matter,
 290          *   because this function is interested in the fact that the commit
 291          *   count reaches back the reserve offset for a specific sub-buffer,
 292          *   which is completely independent of the order.
 293          */
 294         commit_count = v_read(config, &cc_hot->cc);
 295
 296         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 297                                       commit_count, endidx, handle, ctx->tsc);
 298         /*
 299          * Update used size at each commit. It's needed only for extracting
 300          * ring_buffer buffers from vmcore, after crash.
 301          */
 302         lib_ring_buffer_write_commit_counter(config, buf, chan,
 303                         offset_end, commit_count, handle, cc_hot);
 304 }
 305
 306 /**
 307  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 308  * @config: ring buffer instance configuration.
 309  * @ctx: ring buffer context. (input arguments only)
 310  *
 311  * Only succeeds if no other record has been written after the record to
 312  * discard. If discard fails, the record must be committed to the buffer.
 313  *
 314  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 315  */
 316 static inline
 317 int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 318                                         const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 319 {
 320         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
 321         unsigned long end_offset = ctx->pre_offset + ctx->slot_size;
 322
 323         /*
 324          * We need to ensure that if the cmpxchg succeeds and discards the
 325          * record, the next record will record a full TSC, because it cannot
 326          * rely on the last_tsc associated with the discarded record to detect
 327          * overflows. The only way to ensure this is to set the last_tsc to 0
 328          * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
 329          * timestamp in the next record.
 330          *
 331          * Note: if discard fails, we must leave the TSC in the record header.
 332          * It is needed to keep track of TSC overflows for the following
 333          * records.
 334          */
 335         save_last_tsc(config, buf, 0ULL);
 336
 337         if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset)
 338                    != end_offset))
 339                 return -EPERM;
 340         else
 341                 return 0;
 342 }
 343
 344 static inline
 345 void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 346                             struct channel *chan)
 347 {
 348         uatomic_inc(&chan->record_disabled);
 349 }
 350
 351 static inline
 352 void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 353                            struct channel *chan)
 354 {
 355         uatomic_dec(&chan->record_disabled);
 356 }
 357
 358 static inline
 359 void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 360                                     struct lttng_ust_lib_ring_buffer *buf)
 361 {
 362         uatomic_inc(&buf->record_disabled);
 363 }
 364
 365 static inline
 366 void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 367                                    struct lttng_ust_lib_ring_buffer *buf)
 368 {
 369         uatomic_dec(&buf->record_disabled);
 370 }
 371
 372 #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */