lib/ringbuffer/frontend_api.h

   1 #ifndef _LINUX_RING_BUFFER_FRONTEND_API_H
   2 #define _LINUX_RING_BUFFER_FRONTEND_API_H
   3
   4 /*
   5  * linux/ringbuffer/frontend_api.h
   6  *
   7  * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   8  *
   9  * Ring Buffer Library Synchronization Header (buffer write API).
  10  *
  11  * Author:
  12  *      Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  13  *
  14  * See ring_buffer_frontend.c for more information on wait-free algorithms.
  15  * See linux/ringbuffer/frontend.h for channel allocation and read-side API.
  16  *
  17  * Dual LGPL v2.1/GPL v2 license.
  18  */
  19
  20 #include "../../wrapper/ringbuffer/frontend.h"
  21 #include <linux/errno.h>
  22
  23 /**
  24  * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit.
  25  *
  26  * Disables preemption (acts as a RCU read-side critical section) and keeps a
  27  * ring buffer nesting count as supplementary safety net to ensure tracer client
  28  * code will never trigger an endless recursion. Returns the processor ID on
  29  * success, -EPERM on failure (nesting count too high).
  30  *
  31  * asm volatile and "memory" clobber prevent the compiler from moving
  32  * instructions out of the ring buffer nesting count. This is required to ensure
  33  * that probe side-effects which can cause recursion (e.g. unforeseen traps,
  34  * divisions by 0, ...) are triggered within the incremented nesting count
  35  * section.
  36  */
  37 static inline
  38 int lib_ring_buffer_get_cpu(const struct lib_ring_buffer_config *config)
  39 {
  40         int cpu, nesting;
  41
  42         rcu_read_lock_sched_notrace();
  43         cpu = smp_processor_id();
  44         nesting = ++per_cpu(lib_ring_buffer_nesting, cpu);
  45         barrier();
  46
  47         if (unlikely(nesting > 4)) {
  48                 WARN_ON_ONCE(1);
  49                 per_cpu(lib_ring_buffer_nesting, cpu)--;
  50                 rcu_read_unlock_sched_notrace();
  51                 return -EPERM;
  52         } else
  53                 return cpu;
  54 }
  55
  56 /**
  57  * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
  58  */
  59 static inline
  60 void lib_ring_buffer_put_cpu(const struct lib_ring_buffer_config *config)
  61 {
  62         barrier();
  63         __get_cpu_var(lib_ring_buffer_nesting)--;
  64         rcu_read_unlock_sched_notrace();
  65 }
  66
  67 /*
  68  * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
  69  * part of the API per se.
  70  *
  71  * returns 0 if reserve ok, or 1 if the slow path must be taken.
  72  */
  73 static inline
  74 int lib_ring_buffer_try_reserve(const struct lib_ring_buffer_config *config,
  75                                 struct lib_ring_buffer_ctx *ctx,
  76                                 unsigned long *o_begin, unsigned long *o_end,
  77                                 unsigned long *o_old, size_t *before_hdr_pad)
  78 {
  79         struct channel *chan = ctx->chan;
  80         struct lib_ring_buffer *buf = ctx->buf;
  81         *o_begin = v_read(config, &buf->offset);
  82         *o_old = *o_begin;
  83
  84         ctx->tsc = lib_ring_buffer_clock_read(chan);
  85         if ((int64_t) ctx->tsc == -EIO)
  86                 return 1;
  87
  88         /*
  89          * Prefetch cacheline for read because we have to read the previous
  90          * commit counter to increment it and commit seq value to compare it to
  91          * the commit counter.
  92          */
  93         prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
  94
  95         if (last_tsc_overflow(config, buf, ctx->tsc))
  96                 ctx->rflags = RING_BUFFER_RFLAG_FULL_TSC;
  97
  98         if (unlikely(subbuf_offset(*o_begin, chan) == 0))
  99                 return 1;
 100
 101         ctx->slot_size = record_header_size(config, chan, *o_begin,
 102                                             ctx->data_size, before_hdr_pad,
 103                                             ctx->rflags, ctx);
 104         ctx->slot_size +=
 105                 lib_ring_buffer_align(*o_begin + ctx->slot_size,
 106                                       ctx->largest_align) + ctx->data_size;
 107         if (unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size)
 108                      > chan->backend.subbuf_size))
 109                 return 1;
 110
 111         /*
 112          * Record fits in the current buffer and we are not on a switch
 113          * boundary. It's safe to write.
 114          */
 115         *o_end = *o_begin + ctx->slot_size;
 116
 117         if (unlikely((subbuf_offset(*o_end, chan)) == 0))
 118                 /*
 119                  * The offset_end will fall at the very beginning of the next
 120                  * subbuffer.
 121                  */
 122                 return 1;
 123
 124         return 0;
 125 }
 126
 127 /**
 128  * lib_ring_buffer_reserve - Reserve space in a ring buffer.
 129  * @config: ring buffer instance configuration.
 130  * @ctx: ring buffer context. (input and output) Must be already initialized.
 131  *
 132  * Atomic wait-free slot reservation. The reserved space starts at the context
 133  * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
 134  *
 135  * Return :
 136  *  0 on success.
 137  * -EAGAIN if channel is disabled.
 138  * -ENOSPC if event size is too large for packet.
 139  * -ENOBUFS if there is currently not enough space in buffer for the event.
 140  * -EIO if data cannot be written into the buffer for any other reason.
 141  */
 142
 143 static inline
 144 int lib_ring_buffer_reserve(const struct lib_ring_buffer_config *config,
 145                             struct lib_ring_buffer_ctx *ctx)
 146 {
 147         struct channel *chan = ctx->chan;
 148         struct lib_ring_buffer *buf;
 149         unsigned long o_begin, o_end, o_old;
 150         size_t before_hdr_pad = 0;
 151
 152         if (atomic_read(&chan->record_disabled))
 153                 return -EAGAIN;
 154
 155         if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
 156                 buf = per_cpu_ptr(chan->backend.buf, ctx->cpu);
 157         else
 158                 buf = chan->backend.buf;
 159         if (atomic_read(&buf->record_disabled))
 160                 return -EAGAIN;
 161         ctx->buf = buf;
 162
 163         /*
 164          * Perform retryable operations.
 165          */
 166         if (unlikely(lib_ring_buffer_try_reserve(config, ctx, &o_begin,
 167                                                  &o_end, &o_old, &before_hdr_pad)))
 168                 goto slow_path;
 169
 170         if (unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end)
 171                      != o_old))
 172                 goto slow_path;
 173
 174         /*
 175          * Atomically update last_tsc. This update races against concurrent
 176          * atomic updates, but the race will always cause supplementary full TSC
 177          * record headers, never the opposite (missing a full TSC record header
 178          * when it would be needed).
 179          */
 180         save_last_tsc(config, ctx->buf, ctx->tsc);
 181
 182         /*
 183          * Push the reader if necessary
 184          */
 185         lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1);
 186
 187         /*
 188          * Clear noref flag for this subbuffer.
 189          */
 190         lib_ring_buffer_clear_noref(config, &ctx->buf->backend,
 191                                 subbuf_index(o_end - 1, chan));
 192
 193         ctx->pre_offset = o_begin;
 194         ctx->buf_offset = o_begin + before_hdr_pad;
 195         return 0;
 196 slow_path:
 197         return lib_ring_buffer_reserve_slow(ctx);
 198 }
 199
 200 /**
 201  * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
 202  * @config: ring buffer instance configuration.
 203  * @buf: buffer
 204  * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
 205  *
 206  * This operation is completely reentrant : can be called while tracing is
 207  * active with absolutely no lock held.
 208  *
 209  * Note, however, that as a v_cmpxchg is used for some atomic operations and
 210  * requires to be executed locally for per-CPU buffers, this function must be
 211  * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
 212  * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
 213  */
 214 static inline
 215 void lib_ring_buffer_switch(const struct lib_ring_buffer_config *config,
 216                             struct lib_ring_buffer *buf, enum switch_mode mode)
 217 {
 218         lib_ring_buffer_switch_slow(buf, mode);
 219 }
 220
 221 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 222
 223 /**
 224  * lib_ring_buffer_commit - Commit an record.
 225  * @config: ring buffer instance configuration.
 226  * @ctx: ring buffer context. (input arguments only)
 227  *
 228  * Atomic unordered slot commit. Increments the commit count in the
 229  * specified sub-buffer, and delivers it if necessary.
 230  */
 231 static inline
 232 void lib_ring_buffer_commit(const struct lib_ring_buffer_config *config,
 233                             const struct lib_ring_buffer_ctx *ctx)
 234 {
 235         struct channel *chan = ctx->chan;
 236         struct lib_ring_buffer *buf = ctx->buf;
 237         unsigned long offset_end = ctx->buf_offset;
 238         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 239         unsigned long commit_count;
 240
 241         /*
 242          * Must count record before incrementing the commit count.
 243          */
 244         subbuffer_count_record(config, &buf->backend, endidx);
 245
 246         /*
 247          * Order all writes to buffer before the commit count update that will
 248          * determine that the subbuffer is full.
 249          */
 250         if (config->ipi == RING_BUFFER_IPI_BARRIER) {
 251                 /*
 252                  * Must write slot data before incrementing commit count.  This
 253                  * compiler barrier is upgraded into a smp_mb() by the IPI sent
 254                  * by get_subbuf().
 255                  */
 256                 barrier();
 257         } else
 258                 smp_wmb();
 259
 260         v_add(config, ctx->slot_size, &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, &buf->commit_hot[endidx].cc);
 281
 282         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 283                                       commit_count, endidx);
 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);
 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 lib_ring_buffer_config *config,
 305                                         const struct lib_ring_buffer_ctx *ctx)
 306 {
 307         struct 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 (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 lib_ring_buffer_config *config,
 333                             struct channel *chan)
 334 {
 335         atomic_inc(&chan->record_disabled);
 336 }
 337
 338 static inline
 339 void channel_record_enable(const struct lib_ring_buffer_config *config,
 340                            struct channel *chan)
 341 {
 342         atomic_dec(&chan->record_disabled);
 343 }
 344
 345 static inline
 346 void lib_ring_buffer_record_disable(const struct lib_ring_buffer_config *config,
 347                                     struct lib_ring_buffer *buf)
 348 {
 349         atomic_inc(&buf->record_disabled);
 350 }
 351
 352 static inline
 353 void lib_ring_buffer_record_enable(const struct lib_ring_buffer_config *config,
 354                                    struct lib_ring_buffer *buf)
 355 {
 356         atomic_dec(&buf->record_disabled);
 357 }
 358
 359 #endif /* _LINUX_RING_BUFFER_FRONTEND_API_H */