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 <urcu-bp.h>
  22 #include <urcu/compiler.h>
  23
  24 /**
  25  * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit.
  26  *
  27  * Grabs RCU read-side lock and keeps a ring buffer nesting count as
  28  * supplementary safety net to ensure tracer client code will never
  29  * trigger an endless recursion. Returns the processor ID on success,
  30  * -EPERM on failure (nesting count too high).
  31  *
  32  * asm volatile and "memory" clobber prevent the compiler from moving
  33  * instructions out of the ring buffer nesting count. This is required to ensure
  34  * that probe side-effects which can cause recursion (e.g. unforeseen traps,
  35  * divisions by 0, ...) are triggered within the incremented nesting count
  36  * section.
  37  */
  38 static inline
  39 int lib_ring_buffer_get_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
  40 {
  41         int cpu, nesting;
  42
  43         rcu_read_lock();
  44         cpu = lttng_ust_get_cpu();
  45         nesting = ++lib_ring_buffer_nesting;    /* TLS */
  46         cmm_barrier();
  47
  48         if (caa_unlikely(nesting > 4)) {
  49                 WARN_ON_ONCE(1);
  50                 lib_ring_buffer_nesting--;      /* TLS */
  51                 rcu_read_unlock();
  52                 return -EPERM;
  53         } else
  54                 return cpu;
  55 }
  56
  57 /**
  58  * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
  59  */
  60 static inline
  61 void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
  62 {
  63         cmm_barrier();
  64         lib_ring_buffer_nesting--;              /* TLS */
  65         rcu_read_unlock();
  66 }
  67
  68 /*
  69  * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
  70  * part of the API per se.
  71  *
  72  * returns 0 if reserve ok, or 1 if the slow path must be taken.
  73  */
  74 static inline
  75 int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
  76                                 struct lttng_ust_lib_ring_buffer_ctx *ctx,
  77                                 unsigned long *o_begin, unsigned long *o_end,
  78                                 unsigned long *o_old, size_t *before_hdr_pad)
  79 {
  80         struct channel *chan = ctx->chan;
  81         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
  82         *o_begin = v_read(config, &buf->offset);
  83         *o_old = *o_begin;
  84
  85         ctx->tsc = lib_ring_buffer_clock_read(chan);
  86         if ((int64_t) ctx->tsc == -EIO)
  87                 return 1;
  88
  89         /*
  90          * Prefetch cacheline for read because we have to read the previous
  91          * commit counter to increment it and commit seq value to compare it to
  92          * the commit counter.
  93          */
  94         //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
  95
  96         if (last_tsc_overflow(config, buf, ctx->tsc))
  97                 ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
  98
  99         if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0))
 100                 return 1;
 101
 102         ctx->slot_size = record_header_size(config, chan, *o_begin,
 103                                             before_hdr_pad, ctx);
 104         ctx->slot_size +=
 105                 lib_ring_buffer_align(*o_begin + ctx->slot_size,
 106                                       ctx->largest_align) + ctx->data_size;
 107         if (caa_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 (caa_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 lttng_ust_lib_ring_buffer_config *config,
 145                             struct lttng_ust_lib_ring_buffer_ctx *ctx)
 146 {
 147         struct channel *chan = ctx->chan;
 148         struct lttng_ust_shm_handle *handle = ctx->handle;
 149         struct lttng_ust_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(handle, chan->backend.buf[ctx->cpu].shmp);
 158         else
 159                 buf = shmp(handle, chan->backend.buf[0].shmp);
 160         if (uatomic_read(&buf->record_disabled))
 161                 return -EAGAIN;
 162         ctx->buf = buf;
 163
 164         /*
 165          * Perform retryable operations.
 166          */
 167         if (caa_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 (caa_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), handle);
 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 lttng_ust_lib_ring_buffer_config *config,
 217                             struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode,
 218                             struct lttng_ust_shm_handle *handle)
 219 {
 220         lib_ring_buffer_switch_slow(buf, mode, handle);
 221 }
 222
 223 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 224
 225 /**
 226  * lib_ring_buffer_commit - Commit an record.
 227  * @config: ring buffer instance configuration.
 228  * @ctx: ring buffer context. (input arguments only)
 229  *
 230  * Atomic unordered slot commit. Increments the commit count in the
 231  * specified sub-buffer, and delivers it if necessary.
 232  */
 233 static inline
 234 void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config,
 235                             const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 236 {
 237         struct channel *chan = ctx->chan;
 238         struct lttng_ust_shm_handle *handle = ctx->handle;
 239         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
 240         unsigned long offset_end = ctx->buf_offset;
 241         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 242         unsigned long commit_count;
 243
 244         /*
 245          * Must count record before incrementing the commit count.
 246          */
 247         subbuffer_count_record(config, &buf->backend, endidx, handle);
 248
 249         /*
 250          * Order all writes to buffer before the commit count update that will
 251          * determine that the subbuffer is full.
 252          */
 253         cmm_smp_wmb();
 254
 255         v_add(config, ctx->slot_size, &shmp_index(handle, buf->commit_hot, endidx)->cc);
 256
 257         /*
 258          * commit count read can race with concurrent OOO commit count updates.
 259          * This is only needed for lib_ring_buffer_check_deliver (for
 260          * non-polling delivery only) and for
 261          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 262          * counter to be read with the same value more than once, which could
 263          * cause :
 264          * - Multiple delivery for the same sub-buffer (which is handled
 265          *   gracefully by the reader code) if the value is for a full
 266          *   sub-buffer. It's important that we can never miss a sub-buffer
 267          *   delivery. Re-reading the value after the v_add ensures this.
 268          * - Reading a commit_count with a higher value that what was actually
 269          *   added to it for the lib_ring_buffer_write_commit_counter call
 270          *   (again caused by a concurrent committer). It does not matter,
 271          *   because this function is interested in the fact that the commit
 272          *   count reaches back the reserve offset for a specific sub-buffer,
 273          *   which is completely independent of the order.
 274          */
 275         commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc);
 276
 277         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 278                                       commit_count, endidx, handle);
 279         /*
 280          * Update used size at each commit. It's needed only for extracting
 281          * ring_buffer buffers from vmcore, after crash.
 282          */
 283         lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
 284                                              ctx->buf_offset, commit_count,
 285                                              ctx->slot_size, handle);
 286 }
 287
 288 /**
 289  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 290  * @config: ring buffer instance configuration.
 291  * @ctx: ring buffer context. (input arguments only)
 292  *
 293  * Only succeeds if no other record has been written after the record to
 294  * discard. If discard fails, the record must be committed to the buffer.
 295  *
 296  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 297  */
 298 static inline
 299 int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 300                                         const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 301 {
 302         struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
 303         unsigned long end_offset = ctx->pre_offset + ctx->slot_size;
 304
 305         /*
 306          * We need to ensure that if the cmpxchg succeeds and discards the
 307          * record, the next record will record a full TSC, because it cannot
 308          * rely on the last_tsc associated with the discarded record to detect
 309          * overflows. The only way to ensure this is to set the last_tsc to 0
 310          * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
 311          * timestamp in the next record.
 312          *
 313          * Note: if discard fails, we must leave the TSC in the record header.
 314          * It is needed to keep track of TSC overflows for the following
 315          * records.
 316          */
 317         save_last_tsc(config, buf, 0ULL);
 318
 319         if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset)
 320                    != end_offset))
 321                 return -EPERM;
 322         else
 323                 return 0;
 324 }
 325
 326 static inline
 327 void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 328                             struct channel *chan)
 329 {
 330         uatomic_inc(&chan->record_disabled);
 331 }
 332
 333 static inline
 334 void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 335                            struct channel *chan)
 336 {
 337         uatomic_dec(&chan->record_disabled);
 338 }
 339
 340 static inline
 341 void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 342                                     struct lttng_ust_lib_ring_buffer *buf)
 343 {
 344         uatomic_inc(&buf->record_disabled);
 345 }
 346
 347 static inline
 348 void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 349                                    struct lttng_ust_lib_ring_buffer *buf)
 350 {
 351         uatomic_dec(&buf->record_disabled);
 352 }
 353
 354 #endif /* _LINUX_RING_BUFFER_FRONTEND_API_H */