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