lib/ringbuffer/frontend_api.h

   1 #ifndef _LIB_RING_BUFFER_FRONTEND_API_H
   2 #define _LIB_RING_BUFFER_FRONTEND_API_H
   3
   4 /*
   5  * lib/ringbuffer/frontend_api.h
   6  *
   7  * Ring Buffer Library Synchronization Header (buffer write API).
   8  *
   9  * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  10  *
  11  * This library is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public
  13  * License as published by the Free Software Foundation; only
  14  * version 2.1 of the License.
  15  *
  16  * This library is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with this library; if not, write to the Free Software
  23  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  24  *
  25  * Author:
  26  *      Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  27  *
  28  * See ring_buffer_frontend.c for more information on wait-free algorithms.
  29  * See linux/ringbuffer/frontend.h for channel allocation and read-side API.
  30  */
  31
  32 #include "../../wrapper/ringbuffer/frontend.h"
  33 #include "../../wrapper/percpu-defs.h"
  34 #include <linux/errno.h>
  35 #include <linux/prefetch.h>
  36
  37 /**
  38  * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit.
  39  *
  40  * Disables preemption (acts as a RCU read-side critical section) and keeps a
  41  * ring buffer nesting count as supplementary safety net to ensure tracer client
  42  * code will never trigger an endless recursion. Returns the processor ID on
  43  * success, -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 lib_ring_buffer_config *config)
  53 {
  54         int cpu, nesting;
  55
  56         rcu_read_lock_sched_notrace();
  57         cpu = smp_processor_id();
  58         nesting = ++per_cpu(lib_ring_buffer_nesting, cpu);
  59         barrier();
  60
  61         if (unlikely(nesting > 4)) {
  62                 WARN_ON_ONCE(1);
  63                 per_cpu(lib_ring_buffer_nesting, cpu)--;
  64                 rcu_read_unlock_sched_notrace();
  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 lib_ring_buffer_config *config)
  75 {
  76         barrier();
  77         (*lttng_this_cpu_ptr(&lib_ring_buffer_nesting))--;
  78         rcu_read_unlock_sched_notrace();
  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 lib_ring_buffer_config *config,
  89                                 struct 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 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 (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 (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
 130         if (unlikely((subbuf_offset(*o_end, chan)) == 0))
 131                 /*
 132                  * The offset_end will fall at the very beginning of the next
 133                  * subbuffer.
 134                  */
 135                 return 1;
 136
 137         return 0;
 138 }
 139
 140 /**
 141  * lib_ring_buffer_reserve - Reserve space in a ring buffer.
 142  * @config: ring buffer instance configuration.
 143  * @ctx: ring buffer context. (input and output) Must be already initialized.
 144  *
 145  * Atomic wait-free slot reservation. The reserved space starts at the context
 146  * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
 147  *
 148  * Return :
 149  *  0 on success.
 150  * -EAGAIN if channel is disabled.
 151  * -ENOSPC if event size is too large for packet.
 152  * -ENOBUFS if there is currently not enough space in buffer for the event.
 153  * -EIO if data cannot be written into the buffer for any other reason.
 154  */
 155
 156 static inline
 157 int lib_ring_buffer_reserve(const struct lib_ring_buffer_config *config,
 158                             struct lib_ring_buffer_ctx *ctx)
 159 {
 160         struct channel *chan = ctx->chan;
 161         struct lib_ring_buffer *buf;
 162         unsigned long o_begin, o_end, o_old;
 163         size_t before_hdr_pad = 0;
 164
 165         if (atomic_read(&chan->record_disabled))
 166                 return -EAGAIN;
 167
 168         if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
 169                 buf = per_cpu_ptr(chan->backend.buf, ctx->cpu);
 170         else
 171                 buf = chan->backend.buf;
 172         if (atomic_read(&buf->record_disabled))
 173                 return -EAGAIN;
 174         ctx->buf = buf;
 175
 176         /*
 177          * Perform retryable operations.
 178          */
 179         if (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 (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));
 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 lib_ring_buffer_config *config,
 229                             struct lib_ring_buffer *buf, enum switch_mode mode)
 230 {
 231         lib_ring_buffer_switch_slow(buf, mode);
 232 }
 233
 234 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 235
 236 /**
 237  * lib_ring_buffer_commit - Commit an record.
 238  * @config: ring buffer instance configuration.
 239  * @ctx: ring buffer context. (input arguments only)
 240  *
 241  * Atomic unordered slot commit. Increments the commit count in the
 242  * specified sub-buffer, and delivers it if necessary.
 243  */
 244 static inline
 245 void lib_ring_buffer_commit(const struct lib_ring_buffer_config *config,
 246                             const struct lib_ring_buffer_ctx *ctx)
 247 {
 248         struct channel *chan = ctx->chan;
 249         struct 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);
 258
 259         /*
 260          * Order all writes to buffer before the commit count update that will
 261          * determine that the subbuffer is full.
 262          */
 263         if (config->ipi == RING_BUFFER_IPI_BARRIER) {
 264                 /*
 265                  * Must write slot data before incrementing commit count.  This
 266                  * compiler barrier is upgraded into a smp_mb() by the IPI sent
 267                  * by get_subbuf().
 268                  */
 269                 barrier();
 270         } else
 271                 smp_wmb();
 272
 273         v_add(config, ctx->slot_size, &buf->commit_hot[endidx].cc);
 274
 275         /*
 276          * commit count read can race with concurrent OOO commit count updates.
 277          * This is only needed for lib_ring_buffer_check_deliver (for
 278          * non-polling delivery only) and for
 279          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 280          * counter to be read with the same value more than once, which could
 281          * cause :
 282          * - Multiple delivery for the same sub-buffer (which is handled
 283          *   gracefully by the reader code) if the value is for a full
 284          *   sub-buffer. It's important that we can never miss a sub-buffer
 285          *   delivery. Re-reading the value after the v_add ensures this.
 286          * - Reading a commit_count with a higher value that what was actually
 287          *   added to it for the lib_ring_buffer_write_commit_counter call
 288          *   (again caused by a concurrent committer). It does not matter,
 289          *   because this function is interested in the fact that the commit
 290          *   count reaches back the reserve offset for a specific sub-buffer,
 291          *   which is completely independent of the order.
 292          */
 293         commit_count = v_read(config, &buf->commit_hot[endidx].cc);
 294
 295         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 296                                       commit_count, endidx, ctx->tsc);
 297         /*
 298          * Update used size at each commit. It's needed only for extracting
 299          * ring_buffer buffers from vmcore, after crash.
 300          */
 301         lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
 302                         offset_end, commit_count);
 303 }
 304
 305 /**
 306  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 307  * @config: ring buffer instance configuration.
 308  * @ctx: ring buffer context. (input arguments only)
 309  *
 310  * Only succeeds if no other record has been written after the record to
 311  * discard. If discard fails, the record must be committed to the buffer.
 312  *
 313  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 314  */
 315 static inline
 316 int lib_ring_buffer_try_discard_reserve(const struct lib_ring_buffer_config *config,
 317                                         const struct lib_ring_buffer_ctx *ctx)
 318 {
 319         struct lib_ring_buffer *buf = ctx->buf;
 320         unsigned long end_offset = ctx->pre_offset + ctx->slot_size;
 321
 322         /*
 323          * We need to ensure that if the cmpxchg succeeds and discards the
 324          * record, the next record will record a full TSC, because it cannot
 325          * rely on the last_tsc associated with the discarded record to detect
 326          * overflows. The only way to ensure this is to set the last_tsc to 0
 327          * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
 328          * timestamp in the next record.
 329          *
 330          * Note: if discard fails, we must leave the TSC in the record header.
 331          * It is needed to keep track of TSC overflows for the following
 332          * records.
 333          */
 334         save_last_tsc(config, buf, 0ULL);
 335
 336         if (likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset)
 337                    != end_offset))
 338                 return -EPERM;
 339         else
 340                 return 0;
 341 }
 342
 343 static inline
 344 void channel_record_disable(const struct lib_ring_buffer_config *config,
 345                             struct channel *chan)
 346 {
 347         atomic_inc(&chan->record_disabled);
 348 }
 349
 350 static inline
 351 void channel_record_enable(const struct lib_ring_buffer_config *config,
 352                            struct channel *chan)
 353 {
 354         atomic_dec(&chan->record_disabled);
 355 }
 356
 357 static inline
 358 void lib_ring_buffer_record_disable(const struct lib_ring_buffer_config *config,
 359                                     struct lib_ring_buffer *buf)
 360 {
 361         atomic_inc(&buf->record_disabled);
 362 }
 363
 364 static inline
 365 void lib_ring_buffer_record_enable(const struct lib_ring_buffer_config *config,
 366                                    struct lib_ring_buffer *buf)
 367 {
 368         atomic_dec(&buf->record_disabled);
 369 }
 370
 371 #endif /* _LIB_RING_BUFFER_FRONTEND_API_H */