libringbuffer/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 "frontend.h"
  21 #include "ust/core.h"
  22 #include <urcu-bp.h>
  23 #include <urcu/compiler.h>
  24
  25 /**
  26  * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit.
  27  *
  28  * Grabs RCU read-side lock and keeps a ring buffer nesting count as
  29  * supplementary safety net to ensure tracer client code will never
  30  * trigger an endless recursion. Returns the processor ID on success,
  31  * -EPERM on failure (nesting count too high).
  32  *
  33  * asm volatile and "memory" clobber prevent the compiler from moving
  34  * instructions out of the ring buffer nesting count. This is required to ensure
  35  * that probe side-effects which can cause recursion (e.g. unforeseen traps,
  36  * divisions by 0, ...) are triggered within the incremented nesting count
  37  * section.
  38  */
  39 static inline
  40 int lib_ring_buffer_get_cpu(const struct lib_ring_buffer_config *config)
  41 {
  42         int cpu, nesting;
  43
  44         rcu_read_lock();
  45         cpu = ust_get_cpu();
  46         nesting = ++lib_ring_buffer_nesting;    /* TLS */
  47         cmm_barrier();
  48
  49         if (unlikely(nesting > 4)) {
  50                 WARN_ON_ONCE(1);
  51                 lib_ring_buffer_nesting--;      /* TLS */
  52                 rcu_read_unlock();
  53                 return -EPERM;
  54         } else
  55                 return cpu;
  56 }
  57
  58 /**
  59  * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
  60  */
  61 static inline
  62 void lib_ring_buffer_put_cpu(const struct lib_ring_buffer_config *config)
  63 {
  64         cmm_barrier();
  65         lib_ring_buffer_nesting--;              /* TLS */
  66         rcu_read_unlock();
  67 }
  68
  69 /*
  70  * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
  71  * part of the API per se.
  72  *
  73  * returns 0 if reserve ok, or 1 if the slow path must be taken.
  74  */
  75 static inline
  76 int lib_ring_buffer_try_reserve(const struct lib_ring_buffer_config *config,
  77                                 struct lib_ring_buffer_ctx *ctx,
  78                                 unsigned long *o_begin, unsigned long *o_end,
  79                                 unsigned long *o_old, size_t *before_hdr_pad)
  80 {
  81         struct channel *chan = ctx->chan;
  82         struct lib_ring_buffer *buf = ctx->buf;
  83         *o_begin = v_read(config, &buf->offset);
  84         *o_old = *o_begin;
  85
  86         ctx->tsc = lib_ring_buffer_clock_read(chan);
  87         if ((int64_t) ctx->tsc == -EIO)
  88                 return 1;
  89
  90         /*
  91          * Prefetch cacheline for read because we have to read the previous
  92          * commit counter to increment it and commit seq value to compare it to
  93          * the commit counter.
  94          */
  95         //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
  96
  97         if (last_tsc_overflow(config, buf, ctx->tsc))
  98                 ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
  99
 100         if (unlikely(subbuf_offset(*o_begin, chan) == 0))
 101                 return 1;
 102
 103         ctx->slot_size = record_header_size(config, chan, *o_begin,
 104                                             before_hdr_pad, ctx);
 105         ctx->slot_size +=
 106                 lib_ring_buffer_align(*o_begin + ctx->slot_size,
 107                                       ctx->largest_align) + ctx->data_size;
 108         if (unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size)
 109                      > chan->backend.subbuf_size))
 110                 return 1;
 111
 112         /*
 113          * Record fits in the current buffer and we are not on a switch
 114          * boundary. It's safe to write.
 115          */
 116         *o_end = *o_begin + ctx->slot_size;
 117
 118         if (unlikely((subbuf_offset(*o_end, chan)) == 0))
 119                 /*
 120                  * The offset_end will fall at the very beginning of the next
 121                  * subbuffer.
 122                  */
 123                 return 1;
 124
 125         return 0;
 126 }
 127
 128 /**
 129  * lib_ring_buffer_reserve - Reserve space in a ring buffer.
 130  * @config: ring buffer instance configuration.
 131  * @ctx: ring buffer context. (input and output) Must be already initialized.
 132  *
 133  * Atomic wait-free slot reservation. The reserved space starts at the context
 134  * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
 135  *
 136  * Return :
 137  *  0 on success.
 138  * -EAGAIN if channel is disabled.
 139  * -ENOSPC if event size is too large for packet.
 140  * -ENOBUFS if there is currently not enough space in buffer for the event.
 141  * -EIO if data cannot be written into the buffer for any other reason.
 142  */
 143
 144 static inline
 145 int lib_ring_buffer_reserve(const struct lib_ring_buffer_config *config,
 146                             struct lib_ring_buffer_ctx *ctx)
 147 {
 148         struct channel *chan = ctx->chan;
 149         struct lib_ring_buffer *buf;
 150         unsigned long o_begin, o_end, o_old;
 151         size_t before_hdr_pad = 0;
 152
 153         if (uatomic_read(&chan->record_disabled))
 154                 return -EAGAIN;
 155
 156         if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
 157                 buf = &shmp(chan->backend.buf)[ctx->cpu];
 158         else
 159                 buf = shmp(chan->backend.buf);
 160         if (uatomic_read(&buf->record_disabled))
 161                 return -EAGAIN;
 162         ctx->buf = buf;
 163
 164         /*
 165          * Perform retryable operations.
 166          */
 167         if (unlikely(lib_ring_buffer_try_reserve(config, ctx, &o_begin,
 168                                                  &o_end, &o_old, &before_hdr_pad)))
 169                 goto slow_path;
 170
 171         if (unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end)
 172                      != o_old))
 173                 goto slow_path;
 174
 175         /*
 176          * Atomically update last_tsc. This update races against concurrent
 177          * atomic updates, but the race will always cause supplementary full TSC
 178          * record headers, never the opposite (missing a full TSC record header
 179          * when it would be needed).
 180          */
 181         save_last_tsc(config, ctx->buf, ctx->tsc);
 182
 183         /*
 184          * Push the reader if necessary
 185          */
 186         lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1);
 187
 188         /*
 189          * Clear noref flag for this subbuffer.
 190          */
 191         lib_ring_buffer_clear_noref(config, &ctx->buf->backend,
 192                                 subbuf_index(o_end - 1, chan));
 193
 194         ctx->pre_offset = o_begin;
 195         ctx->buf_offset = o_begin + before_hdr_pad;
 196         return 0;
 197 slow_path:
 198         return lib_ring_buffer_reserve_slow(ctx);
 199 }
 200
 201 /**
 202  * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
 203  * @config: ring buffer instance configuration.
 204  * @buf: buffer
 205  * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
 206  *
 207  * This operation is completely reentrant : can be called while tracing is
 208  * active with absolutely no lock held.
 209  *
 210  * Note, however, that as a v_cmpxchg is used for some atomic operations and
 211  * requires to be executed locally for per-CPU buffers, this function must be
 212  * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
 213  * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
 214  */
 215 static inline
 216 void lib_ring_buffer_switch(const struct lib_ring_buffer_config *config,
 217                             struct lib_ring_buffer *buf, enum switch_mode mode)
 218 {
 219         lib_ring_buffer_switch_slow(buf, mode);
 220 }
 221
 222 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 223
 224 /**
 225  * lib_ring_buffer_commit - Commit an record.
 226  * @config: ring buffer instance configuration.
 227  * @ctx: ring buffer context. (input arguments only)
 228  *
 229  * Atomic unordered slot commit. Increments the commit count in the
 230  * specified sub-buffer, and delivers it if necessary.
 231  */
 232 static inline
 233 void lib_ring_buffer_commit(const struct lib_ring_buffer_config *config,
 234                             const struct lib_ring_buffer_ctx *ctx)
 235 {
 236         struct channel *chan = ctx->chan;
 237         struct lib_ring_buffer *buf = ctx->buf;
 238         unsigned long offset_end = ctx->buf_offset;
 239         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 240         unsigned long commit_count;
 241
 242         /*
 243          * Must count record before incrementing the commit count.
 244          */
 245         subbuffer_count_record(config, &buf->backend, endidx);
 246
 247         /*
 248          * Order all writes to buffer before the commit count update that will
 249          * determine that the subbuffer is full.
 250          */
 251         cmm_smp_wmb();
 252
 253         v_add(config, ctx->slot_size, &shmp(buf->commit_hot)[endidx].cc);
 254
 255         /*
 256          * commit count read can race with concurrent OOO commit count updates.
 257          * This is only needed for lib_ring_buffer_check_deliver (for
 258          * non-polling delivery only) and for
 259          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 260          * counter to be read with the same value more than once, which could
 261          * cause :
 262          * - Multiple delivery for the same sub-buffer (which is handled
 263          *   gracefully by the reader code) if the value is for a full
 264          *   sub-buffer. It's important that we can never miss a sub-buffer
 265          *   delivery. Re-reading the value after the v_add ensures this.
 266          * - Reading a commit_count with a higher value that what was actually
 267          *   added to it for the lib_ring_buffer_write_commit_counter call
 268          *   (again caused by a concurrent committer). It does not matter,
 269          *   because this function is interested in the fact that the commit
 270          *   count reaches back the reserve offset for a specific sub-buffer,
 271          *   which is completely independent of the order.
 272          */
 273         commit_count = v_read(config, &shmp(buf->commit_hot)[endidx].cc);
 274
 275         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 276                                       commit_count, endidx);
 277         /*
 278          * Update used size at each commit. It's needed only for extracting
 279          * ring_buffer buffers from vmcore, after crash.
 280          */
 281         lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
 282                                              ctx->buf_offset, commit_count,
 283                                          ctx->slot_size);
 284 }
 285
 286 /**
 287  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 288  * @config: ring buffer instance configuration.
 289  * @ctx: ring buffer context. (input arguments only)
 290  *
 291  * Only succeeds if no other record has been written after the record to
 292  * discard. If discard fails, the record must be committed to the buffer.
 293  *
 294  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 295  */
 296 static inline
 297 int lib_ring_buffer_try_discard_reserve(const struct lib_ring_buffer_config *config,
 298                                         const struct lib_ring_buffer_ctx *ctx)
 299 {
 300         struct lib_ring_buffer *buf = ctx->buf;
 301         unsigned long end_offset = ctx->pre_offset + ctx->slot_size;
 302
 303         /*
 304          * We need to ensure that if the cmpxchg succeeds and discards the
 305          * record, the next record will record a full TSC, because it cannot
 306          * rely on the last_tsc associated with the discarded record to detect
 307          * overflows. The only way to ensure this is to set the last_tsc to 0
 308          * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
 309          * timestamp in the next record.
 310          *
 311          * Note: if discard fails, we must leave the TSC in the record header.
 312          * It is needed to keep track of TSC overflows for the following
 313          * records.
 314          */
 315         save_last_tsc(config, buf, 0ULL);
 316
 317         if (likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset)
 318                    != end_offset))
 319                 return -EPERM;
 320         else
 321                 return 0;
 322 }
 323
 324 static inline
 325 void channel_record_disable(const struct lib_ring_buffer_config *config,
 326                             struct channel *chan)
 327 {
 328         uatomic_inc(&chan->record_disabled);
 329 }
 330
 331 static inline
 332 void channel_record_enable(const struct lib_ring_buffer_config *config,
 333                            struct channel *chan)
 334 {
 335         uatomic_dec(&chan->record_disabled);
 336 }
 337
 338 static inline
 339 void lib_ring_buffer_record_disable(const struct lib_ring_buffer_config *config,
 340                                     struct lib_ring_buffer *buf)
 341 {
 342         uatomic_inc(&buf->record_disabled);
 343 }
 344
 345 static inline
 346 void lib_ring_buffer_record_enable(const struct lib_ring_buffer_config *config,
 347                                    struct lib_ring_buffer *buf)
 348 {
 349         uatomic_dec(&buf->record_disabled);
 350 }
 351
 352 #endif /* _LINUX_RING_BUFFER_FRONTEND_API_H */