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
  86         /*
  87          * Prefetch cacheline for read because we have to read the previous
  88          * commit counter to increment it and commit seq value to compare it to
  89          * the commit counter.
  90          */
  91         prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
  92
  93         if (last_tsc_overflow(config, buf, ctx->tsc))
  94                 ctx->rflags = RING_BUFFER_RFLAG_FULL_TSC;
  95
  96         if (unlikely(subbuf_offset(*o_begin, chan) == 0))
  97                 return 1;
  98
  99         ctx->slot_size = record_header_size(config, chan, *o_begin,
 100                                             ctx->data_size, before_hdr_pad,
 101                                             ctx->rflags, ctx);
 102         ctx->slot_size +=
 103                 lib_ring_buffer_align(*o_begin + ctx->slot_size,
 104                                       ctx->largest_align) + ctx->data_size;
 105         if (unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size)
 106                      > chan->backend.subbuf_size))
 107                 return 1;
 108
 109         /*
 110          * Record fits in the current buffer and we are not on a switch
 111          * boundary. It's safe to write.
 112          */
 113         *o_end = *o_begin + ctx->slot_size;
 114
 115         if (unlikely((subbuf_offset(*o_end, chan)) == 0))
 116                 /*
 117                  * The offset_end will fall at the very beginning of the next
 118                  * subbuffer.
 119                  */
 120                 return 1;
 121
 122         return 0;
 123 }
 124
 125 /**
 126  * lib_ring_buffer_reserve - Reserve space in a ring buffer.
 127  * @config: ring buffer instance configuration.
 128  * @ctx: ring buffer context. (input and output) Must be already initialized.
 129  *
 130  * Atomic wait-free slot reservation. The reserved space starts at the context
 131  * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
 132  *
 133  * Return : -ENOSPC if not enough space, -EAGAIN if channel is disabled.
 134  *          Returns 0 on success.
 135  */
 136
 137 static inline
 138 int lib_ring_buffer_reserve(const struct lib_ring_buffer_config *config,
 139                             struct lib_ring_buffer_ctx *ctx)
 140 {
 141         struct channel *chan = ctx->chan;
 142         struct lib_ring_buffer *buf;
 143         unsigned long o_begin, o_end, o_old;
 144         size_t before_hdr_pad = 0;
 145
 146         if (atomic_read(&chan->record_disabled))
 147                 return -EAGAIN;
 148
 149         if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
 150                 buf = per_cpu_ptr(chan->backend.buf, ctx->cpu);
 151         else
 152                 buf = chan->backend.buf;
 153         if (atomic_read(&buf->record_disabled))
 154                 return -EAGAIN;
 155         ctx->buf = buf;
 156
 157         /*
 158          * Perform retryable operations.
 159          */
 160         if (unlikely(lib_ring_buffer_try_reserve(config, ctx, &o_begin,
 161                                                  &o_end, &o_old, &before_hdr_pad)))
 162                 goto slow_path;
 163
 164         if (unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end)
 165                      != o_old))
 166                 goto slow_path;
 167
 168         /*
 169          * Atomically update last_tsc. This update races against concurrent
 170          * atomic updates, but the race will always cause supplementary full TSC
 171          * record headers, never the opposite (missing a full TSC record header
 172          * when it would be needed).
 173          */
 174         save_last_tsc(config, ctx->buf, ctx->tsc);
 175
 176         /*
 177          * Push the reader if necessary
 178          */
 179         lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1);
 180
 181         /*
 182          * Clear noref flag for this subbuffer.
 183          */
 184         lib_ring_buffer_clear_noref(config, &ctx->buf->backend,
 185                                 subbuf_index(o_end - 1, chan));
 186
 187         ctx->pre_offset = o_begin;
 188         ctx->buf_offset = o_begin + before_hdr_pad;
 189         return 0;
 190 slow_path:
 191         return lib_ring_buffer_reserve_slow(ctx);
 192 }
 193
 194 /**
 195  * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
 196  * @config: ring buffer instance configuration.
 197  * @buf: buffer
 198  * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
 199  *
 200  * This operation is completely reentrant : can be called while tracing is
 201  * active with absolutely no lock held.
 202  *
 203  * Note, however, that as a v_cmpxchg is used for some atomic operations and
 204  * requires to be executed locally for per-CPU buffers, this function must be
 205  * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
 206  * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
 207  */
 208 static inline
 209 void lib_ring_buffer_switch(const struct lib_ring_buffer_config *config,
 210                             struct lib_ring_buffer *buf, enum switch_mode mode)
 211 {
 212         lib_ring_buffer_switch_slow(buf, mode);
 213 }
 214
 215 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 216
 217 /**
 218  * lib_ring_buffer_commit - Commit an record.
 219  * @config: ring buffer instance configuration.
 220  * @ctx: ring buffer context. (input arguments only)
 221  *
 222  * Atomic unordered slot commit. Increments the commit count in the
 223  * specified sub-buffer, and delivers it if necessary.
 224  */
 225 static inline
 226 void lib_ring_buffer_commit(const struct lib_ring_buffer_config *config,
 227                             const struct lib_ring_buffer_ctx *ctx)
 228 {
 229         struct channel *chan = ctx->chan;
 230         struct lib_ring_buffer *buf = ctx->buf;
 231         unsigned long offset_end = ctx->buf_offset;
 232         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 233         unsigned long commit_count;
 234
 235         /*
 236          * Must count record before incrementing the commit count.
 237          */
 238         subbuffer_count_record(config, &buf->backend, endidx);
 239
 240         /*
 241          * Order all writes to buffer before the commit count update that will
 242          * determine that the subbuffer is full.
 243          */
 244         if (config->ipi == RING_BUFFER_IPI_BARRIER) {
 245                 /*
 246                  * Must write slot data before incrementing commit count.  This
 247                  * compiler barrier is upgraded into a smp_mb() by the IPI sent
 248                  * by get_subbuf().
 249                  */
 250                 barrier();
 251         } else
 252                 smp_wmb();
 253
 254         v_add(config, ctx->slot_size, &buf->commit_hot[endidx].cc);
 255
 256         /*
 257          * commit count read can race with concurrent OOO commit count updates.
 258          * This is only needed for lib_ring_buffer_check_deliver (for
 259          * non-polling delivery only) and for
 260          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 261          * counter to be read with the same value more than once, which could
 262          * cause :
 263          * - Multiple delivery for the same sub-buffer (which is handled
 264          *   gracefully by the reader code) if the value is for a full
 265          *   sub-buffer. It's important that we can never miss a sub-buffer
 266          *   delivery. Re-reading the value after the v_add ensures this.
 267          * - Reading a commit_count with a higher value that what was actually
 268          *   added to it for the lib_ring_buffer_write_commit_counter call
 269          *   (again caused by a concurrent committer). It does not matter,
 270          *   because this function is interested in the fact that the commit
 271          *   count reaches back the reserve offset for a specific sub-buffer,
 272          *   which is completely independent of the order.
 273          */
 274         commit_count = v_read(config, &buf->commit_hot[endidx].cc);
 275
 276         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 277                                       commit_count, endidx);
 278         /*
 279          * Update used size at each commit. It's needed only for extracting
 280          * ring_buffer buffers from vmcore, after crash.
 281          */
 282         lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
 283                                              ctx->buf_offset, commit_count,
 284                                          ctx->slot_size);
 285 }
 286
 287 /**
 288  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 289  * @config: ring buffer instance configuration.
 290  * @ctx: ring buffer context. (input arguments only)
 291  *
 292  * Only succeeds if no other record has been written after the record to
 293  * discard. If discard fails, the record must be committed to the buffer.
 294  *
 295  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 296  */
 297 static inline
 298 int lib_ring_buffer_try_discard_reserve(const struct lib_ring_buffer_config *config,
 299                                         const struct lib_ring_buffer_ctx *ctx)
 300 {
 301         struct lib_ring_buffer *buf = ctx->buf;
 302         unsigned long end_offset = ctx->pre_offset + ctx->slot_size;
 303
 304         /*
 305          * We need to ensure that if the cmpxchg succeeds and discards the
 306          * record, the next record will record a full TSC, because it cannot
 307          * rely on the last_tsc associated with the discarded record to detect
 308          * overflows. The only way to ensure this is to set the last_tsc to 0
 309          * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
 310          * timestamp in the next record.
 311          *
 312          * Note: if discard fails, we must leave the TSC in the record header.
 313          * It is needed to keep track of TSC overflows for the following
 314          * records.
 315          */
 316         save_last_tsc(config, buf, 0ULL);
 317
 318         if (likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset)
 319                    != end_offset))
 320                 return -EPERM;
 321         else
 322                 return 0;
 323 }
 324
 325 static inline
 326 void channel_record_disable(const struct lib_ring_buffer_config *config,
 327                             struct channel *chan)
 328 {
 329         atomic_inc(&chan->record_disabled);
 330 }
 331
 332 static inline
 333 void channel_record_enable(const struct lib_ring_buffer_config *config,
 334                            struct channel *chan)
 335 {
 336         atomic_dec(&chan->record_disabled);
 337 }
 338
 339 static inline
 340 void lib_ring_buffer_record_disable(const struct lib_ring_buffer_config *config,
 341                                     struct lib_ring_buffer *buf)
 342 {
 343         atomic_inc(&buf->record_disabled);
 344 }
 345
 346 static inline
 347 void lib_ring_buffer_record_enable(const struct lib_ring_buffer_config *config,
 348                                    struct lib_ring_buffer *buf)
 349 {
 350         atomic_dec(&buf->record_disabled);
 351 }
 352
 353 #endif /* _LINUX_RING_BUFFER_FRONTEND_API_H */