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 <stddef.h>
  34
  35 #include <urcu-bp.h>
  36 #include <urcu/compiler.h>
  37
  38 #include "frontend.h"
  39
  40 /**
  41  * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit.
  42  *
  43  * Keeps a ring buffer nesting count as supplementary safety net to
  44  * ensure tracer client code will never trigger an endless recursion.
  45  * Returns the processor ID on success, -EPERM on failure (nesting count
  46  * too high).
  47  *
  48  * asm volatile and "memory" clobber prevent the compiler from moving
  49  * instructions out of the ring buffer nesting count. This is required to ensure
  50  * that probe side-effects which can cause recursion (e.g. unforeseen traps,
  51  * divisions by 0, ...) are triggered within the incremented nesting count
  52  * section.
  53  */
  54 static inline
  55 int lib_ring_buffer_get_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
  56 {
  57         int cpu, nesting;
  58
  59         cpu = lttng_ust_get_cpu();
  60         nesting = ++URCU_TLS(lib_ring_buffer_nesting);
  61         cmm_barrier();
  62
  63         if (caa_unlikely(nesting > 4)) {
  64                 WARN_ON_ONCE(1);
  65                 URCU_TLS(lib_ring_buffer_nesting)--;
  66                 return -EPERM;
  67         } else
  68                 return cpu;
  69 }
  70
  71 /**
  72  * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
  73  */
  74 static inline
  75 void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
  76 {
  77         cmm_barrier();
  78         URCU_TLS(lib_ring_buffer_nesting)--;            /* TLS */
  79 }
  80
  81 /*
  82  * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
  83  * part of the API per se.
  84  *
  85  * returns 0 if reserve ok, or 1 if the slow path must be taken.
  86  */
  87 static inline
  88 int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
  89                                 struct lttng_ust_lib_ring_buffer_ctx *ctx,
  90                                 void *client_ctx,
  91                                 unsigned long *o_begin, unsigned long *o_end,
  92                                 unsigned long *o_old, size_t *before_hdr_pad)
  93 {
  94         struct channel *chan = ctx->chan;
  95         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
  96         *o_begin = v_read(config, &buf->offset);
  97         *o_old = *o_begin;
  98
  99         ctx->tsc = lib_ring_buffer_clock_read(chan);
 100         if ((int64_t) ctx->tsc == -EIO)
 101                 return 1;
 102
 103         /*
 104          * Prefetch cacheline for read because we have to read the previous
 105          * commit counter to increment it and commit seq value to compare it to
 106          * the commit counter.
 107          */
 108         //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
 109
 110         if (last_tsc_overflow(config, buf, ctx->tsc))
 111                 ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
 112
 113         if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0))
 114                 return 1;
 115
 116         ctx->slot_size = record_header_size(config, chan, *o_begin,
 117                                             before_hdr_pad, ctx, client_ctx);
 118         ctx->slot_size +=
 119                 lib_ring_buffer_align(*o_begin + ctx->slot_size,
 120                                       ctx->largest_align) + ctx->data_size;
 121         if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size)
 122                      > chan->backend.subbuf_size))
 123                 return 1;
 124
 125         /*
 126          * Record fits in the current buffer and we are not on a switch
 127          * boundary. It's safe to write.
 128          */
 129         *o_end = *o_begin + ctx->slot_size;
 130
 131         if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0))
 132                 /*
 133                  * The offset_end will fall at the very beginning of the next
 134                  * subbuffer.
 135                  */
 136                 return 1;
 137
 138         return 0;
 139 }
 140
 141 /**
 142  * lib_ring_buffer_reserve - Reserve space in a ring buffer.
 143  * @config: ring buffer instance configuration.
 144  * @ctx: ring buffer context. (input and output) Must be already initialized.
 145  *
 146  * Atomic wait-free slot reservation. The reserved space starts at the context
 147  * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
 148  *
 149  * Return :
 150  *  0 on success.
 151  * -EAGAIN if channel is disabled.
 152  * -ENOSPC if event size is too large for packet.
 153  * -ENOBUFS if there is currently not enough space in buffer for the event.
 154  * -EIO if data cannot be written into the buffer for any other reason.
 155  */
 156
 157 static inline
 158 int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 159                             struct lttng_ust_lib_ring_buffer_ctx *ctx,
 160                             void *client_ctx)
 161 {
 162         struct channel *chan = ctx->chan;
 163         struct lttng_ust_shm_handle *handle = ctx->handle;
 164         struct lttng_ust_lib_ring_buffer *buf;
 165         unsigned long o_begin, o_end, o_old;
 166         size_t before_hdr_pad = 0;
 167
 168         if (caa_unlikely(uatomic_read(&chan->record_disabled)))
 169                 return -EAGAIN;
 170
 171         if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
 172                 buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp);
 173         else
 174                 buf = shmp(handle, chan->backend.buf[0].shmp);
 175         if (caa_unlikely(!buf))
 176                 return -EIO;
 177         if (caa_unlikely(uatomic_read(&buf->record_disabled)))
 178                 return -EAGAIN;
 179         ctx->buf = buf;
 180
 181         /*
 182          * Perform retryable operations.
 183          */
 184         if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, client_ctx, &o_begin,
 185                                                  &o_end, &o_old, &before_hdr_pad)))
 186                 goto slow_path;
 187
 188         if (caa_unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end)
 189                      != o_old))
 190                 goto slow_path;
 191
 192         /*
 193          * Atomically update last_tsc. This update races against concurrent
 194          * atomic updates, but the race will always cause supplementary full TSC
 195          * record headers, never the opposite (missing a full TSC record header
 196          * when it would be needed).
 197          */
 198         save_last_tsc(config, ctx->buf, ctx->tsc);
 199
 200         /*
 201          * Push the reader if necessary
 202          */
 203         lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1);
 204
 205         /*
 206          * Clear noref flag for this subbuffer.
 207          */
 208         lib_ring_buffer_clear_noref(config, &ctx->buf->backend,
 209                                 subbuf_index(o_end - 1, chan), handle);
 210
 211         ctx->pre_offset = o_begin;
 212         ctx->buf_offset = o_begin + before_hdr_pad;
 213         return 0;
 214 slow_path:
 215         return lib_ring_buffer_reserve_slow(ctx, client_ctx);
 216 }
 217
 218 /**
 219  * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
 220  * @config: ring buffer instance configuration.
 221  * @buf: buffer
 222  * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
 223  *
 224  * This operation is completely reentrant : can be called while tracing is
 225  * active with absolutely no lock held.
 226  *
 227  * Note, however, that as a v_cmpxchg is used for some atomic operations and
 228  * requires to be executed locally for per-CPU buffers, this function must be
 229  * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
 230  * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
 231  */
 232 static inline
 233 void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config,
 234                             struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode,
 235                             struct lttng_ust_shm_handle *handle)
 236 {
 237         lib_ring_buffer_switch_slow(buf, mode, handle);
 238 }
 239
 240 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 241
 242 /**
 243  * lib_ring_buffer_commit - Commit an record.
 244  * @config: ring buffer instance configuration.
 245  * @ctx: ring buffer context. (input arguments only)
 246  *
 247  * Atomic unordered slot commit. Increments the commit count in the
 248  * specified sub-buffer, and delivers it if necessary.
 249  */
 250 static inline
 251 void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config,
 252                             const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 253 {
 254         struct channel *chan = ctx->chan;
 255         struct lttng_ust_shm_handle *handle = ctx->handle;
 256         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
 257         unsigned long offset_end = ctx->buf_offset;
 258         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 259         unsigned long commit_count;
 260         struct commit_counters_hot *cc_hot = shmp_index(handle,
 261                                                 buf->commit_hot, endidx);
 262
 263         if (caa_unlikely(!cc_hot))
 264                 return;
 265
 266         /*
 267          * Must count record before incrementing the commit count.
 268          */
 269         subbuffer_count_record(config, ctx, &buf->backend, endidx, handle);
 270
 271         /*
 272          * Order all writes to buffer before the commit count update that will
 273          * determine that the subbuffer is full.
 274          */
 275         cmm_smp_wmb();
 276
 277         v_add(config, ctx->slot_size, &cc_hot->cc);
 278
 279         /*
 280          * commit count read can race with concurrent OOO commit count updates.
 281          * This is only needed for lib_ring_buffer_check_deliver (for
 282          * non-polling delivery only) and for
 283          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 284          * counter to be read with the same value more than once, which could
 285          * cause :
 286          * - Multiple delivery for the same sub-buffer (which is handled
 287          *   gracefully by the reader code) if the value is for a full
 288          *   sub-buffer. It's important that we can never miss a sub-buffer
 289          *   delivery. Re-reading the value after the v_add ensures this.
 290          * - Reading a commit_count with a higher value that what was actually
 291          *   added to it for the lib_ring_buffer_write_commit_counter call
 292          *   (again caused by a concurrent committer). It does not matter,
 293          *   because this function is interested in the fact that the commit
 294          *   count reaches back the reserve offset for a specific sub-buffer,
 295          *   which is completely independent of the order.
 296          */
 297         commit_count = v_read(config, &cc_hot->cc);
 298
 299         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 300                                       commit_count, endidx, handle, ctx->tsc);
 301         /*
 302          * Update used size at each commit. It's needed only for extracting
 303          * ring_buffer buffers from vmcore, after crash.
 304          */
 305         lib_ring_buffer_write_commit_counter(config, buf, chan,
 306                         offset_end, commit_count, handle, cc_hot);
 307 }
 308
 309 /**
 310  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 311  * @config: ring buffer instance configuration.
 312  * @ctx: ring buffer context. (input arguments only)
 313  *
 314  * Only succeeds if no other record has been written after the record to
 315  * discard. If discard fails, the record must be committed to the buffer.
 316  *
 317  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 318  */
 319 static inline
 320 int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 321                                         const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 322 {
 323         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
 324         unsigned long end_offset = ctx->pre_offset + ctx->slot_size;
 325
 326         /*
 327          * We need to ensure that if the cmpxchg succeeds and discards the
 328          * record, the next record will record a full TSC, because it cannot
 329          * rely on the last_tsc associated with the discarded record to detect
 330          * overflows. The only way to ensure this is to set the last_tsc to 0
 331          * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
 332          * timestamp in the next record.
 333          *
 334          * Note: if discard fails, we must leave the TSC in the record header.
 335          * It is needed to keep track of TSC overflows for the following
 336          * records.
 337          */
 338         save_last_tsc(config, buf, 0ULL);
 339
 340         if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset)
 341                    != end_offset))
 342                 return -EPERM;
 343         else
 344                 return 0;
 345 }
 346
 347 static inline
 348 void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 349                             struct channel *chan)
 350 {
 351         uatomic_inc(&chan->record_disabled);
 352 }
 353
 354 static inline
 355 void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 356                            struct channel *chan)
 357 {
 358         uatomic_dec(&chan->record_disabled);
 359 }
 360
 361 static inline
 362 void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 363                                     struct lttng_ust_lib_ring_buffer *buf)
 364 {
 365         uatomic_inc(&buf->record_disabled);
 366 }
 367
 368 static inline
 369 void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 370                                    struct lttng_ust_lib_ring_buffer *buf)
 371 {
 372         uatomic_dec(&buf->record_disabled);
 373 }
 374
 375 #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */