libringbuffer/frontend_api.h

   1 /*
   2  * SPDX-License-Identifier: LGPL-2.1-only
   3  *
   4  * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   5  *
   6  * See ring_buffer_frontend.c for more information on wait-free
   7  * algorithms.
   8  * See frontend.h for channel allocation and read-side API.
   9  */
  10
  11 #ifndef _LTTNG_RING_BUFFER_FRONTEND_API_H
  12 #define _LTTNG_RING_BUFFER_FRONTEND_API_H
  13
  14 #include <stddef.h>
  15
  16 #include <urcu/compiler.h>
  17
  18 #include "frontend.h"
  19
  20 /**
  21  * lib_ring_buffer_nesting_inc - Ring buffer recursive use protection.
  22  *
  23  * The rint buffer buffer nesting count is a safety net to ensure tracer
  24  * client code will never trigger an endless recursion.
  25  * Returns a nesting level >= 0 on success, -EPERM on failure (nesting
  26  * count too high).
  27  *
  28  * asm volatile and "memory" clobber prevent the compiler from moving
  29  * instructions out of the ring buffer nesting count. This is required to ensure
  30  * that probe side-effects which can cause recursion (e.g. unforeseen traps,
  31  * divisions by 0, ...) are triggered within the incremented nesting count
  32  * section.
  33  */
  34 static inline
  35 int lib_ring_buffer_nesting_inc(
  36                 const struct lttng_ust_lib_ring_buffer_config *config __attribute__((unused)))
  37 {
  38         int nesting;
  39
  40         nesting = ++URCU_TLS(lib_ring_buffer_nesting);
  41         cmm_barrier();
  42         if (caa_unlikely(nesting >= LIB_RING_BUFFER_MAX_NESTING)) {
  43                 WARN_ON_ONCE(1);
  44                 URCU_TLS(lib_ring_buffer_nesting)--;
  45                 return -EPERM;
  46         }
  47         return nesting - 1;
  48 }
  49
  50 static inline
  51 int lib_ring_buffer_nesting_count(
  52                 const struct lttng_ust_lib_ring_buffer_config *config __attribute__((unused)))
  53 {
  54         return URCU_TLS(lib_ring_buffer_nesting);
  55 }
  56
  57 static inline
  58 void lib_ring_buffer_nesting_dec(
  59                 const struct lttng_ust_lib_ring_buffer_config *config __attribute__((unused)))
  60 {
  61         cmm_barrier();
  62         URCU_TLS(lib_ring_buffer_nesting)--;            /* TLS */
  63 }
  64
  65 /*
  66  * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
  67  * part of the API per se.
  68  *
  69  * returns 0 if reserve ok, or 1 if the slow path must be taken.
  70  */
  71 static inline
  72 int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
  73                                 struct lttng_ust_lib_ring_buffer_ctx *ctx,
  74                                 void *client_ctx,
  75                                 unsigned long *o_begin, unsigned long *o_end,
  76                                 unsigned long *o_old, size_t *before_hdr_pad)
  77 {
  78         struct lttng_ust_lib_ring_buffer_ctx_private *ctx_private = ctx->priv;
  79         struct lttng_ust_lib_ring_buffer_channel *chan = ctx_private->chan;
  80         struct lttng_ust_lib_ring_buffer *buf = ctx_private->buf;
  81         *o_begin = v_read(config, &buf->offset);
  82         *o_old = *o_begin;
  83
  84         ctx_private->tsc = lib_ring_buffer_clock_read(chan);
  85         if ((int64_t) ctx_private->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_private->tsc))
  96                 ctx_private->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
  97
  98         if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0))
  99                 return 1;
 100
 101         ctx_private->slot_size = record_header_size(config, chan, *o_begin,
 102                                             before_hdr_pad, ctx, client_ctx);
 103         ctx_private->slot_size +=
 104                 lttng_ust_lib_ring_buffer_align(*o_begin + ctx_private->slot_size,
 105                                       ctx->largest_align) + ctx->data_size;
 106         if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx_private->slot_size)
 107                      > chan->backend.subbuf_size))
 108                 return 1;
 109
 110         /*
 111          * Record fits in the current buffer and we are not on a switch
 112          * boundary. It's safe to write.
 113          */
 114         *o_end = *o_begin + ctx_private->slot_size;
 115
 116         if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0))
 117                 /*
 118                  * The offset_end will fall at the very beginning of the next
 119                  * subbuffer.
 120                  */
 121                 return 1;
 122
 123         return 0;
 124 }
 125
 126 /**
 127  * lib_ring_buffer_reserve - Reserve space in a ring buffer.
 128  * @config: ring buffer instance configuration.
 129  * @ctx: ring buffer context. (input and output) Must be already initialized.
 130  *
 131  * Atomic wait-free slot reservation. The reserved space starts at the context
 132  * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
 133  *
 134  * Return :
 135  *  0 on success.
 136  * -EAGAIN if channel is disabled.
 137  * -ENOSPC if event size is too large for packet.
 138  * -ENOBUFS if there is currently not enough space in buffer for the event.
 139  * -EIO if data cannot be written into the buffer for any other reason.
 140  */
 141
 142 static inline
 143 int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 144                             struct lttng_ust_lib_ring_buffer_ctx *ctx,
 145                             void *client_ctx)
 146 {
 147         struct lttng_ust_lib_ring_buffer_ctx_private *ctx_private = ctx->priv;
 148         struct lttng_ust_lib_ring_buffer_channel *chan = ctx_private->chan;
 149         struct lttng_ust_shm_handle *handle = chan->handle;
 150         struct lttng_ust_lib_ring_buffer *buf;
 151         unsigned long o_begin, o_end, o_old;
 152         size_t before_hdr_pad = 0;
 153
 154         if (caa_unlikely(uatomic_read(&chan->record_disabled)))
 155                 return -EAGAIN;
 156
 157         if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
 158                 ctx_private->reserve_cpu = lttng_ust_get_cpu();
 159                 buf = shmp(handle, chan->backend.buf[ctx_private->reserve_cpu].shmp);
 160         } else {
 161                 buf = shmp(handle, chan->backend.buf[0].shmp);
 162         }
 163         if (caa_unlikely(!buf))
 164                 return -EIO;
 165         if (caa_unlikely(uatomic_read(&buf->record_disabled)))
 166                 return -EAGAIN;
 167         ctx_private->buf = buf;
 168
 169         /*
 170          * Perform retryable operations.
 171          */
 172         if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, client_ctx, &o_begin,
 173                                                  &o_end, &o_old, &before_hdr_pad)))
 174                 goto slow_path;
 175
 176         if (caa_unlikely(v_cmpxchg(config, &buf->offset, o_old, o_end)
 177                      != o_old))
 178                 goto slow_path;
 179
 180         /*
 181          * Atomically update last_tsc. This update races against concurrent
 182          * atomic updates, but the race will always cause supplementary full TSC
 183          * record headers, never the opposite (missing a full TSC record header
 184          * when it would be needed).
 185          */
 186         save_last_tsc(config, buf, ctx_private->tsc);
 187
 188         /*
 189          * Push the reader if necessary
 190          */
 191         lib_ring_buffer_reserve_push_reader(buf, chan, o_end - 1);
 192
 193         /*
 194          * Clear noref flag for this subbuffer.
 195          */
 196         lib_ring_buffer_clear_noref(config, &buf->backend,
 197                                 subbuf_index(o_end - 1, chan), handle);
 198
 199         ctx_private->pre_offset = o_begin;
 200         ctx_private->buf_offset = o_begin + before_hdr_pad;
 201         return 0;
 202 slow_path:
 203         return lib_ring_buffer_reserve_slow(ctx, client_ctx);
 204 }
 205
 206 /**
 207  * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
 208  * @config: ring buffer instance configuration.
 209  * @buf: buffer
 210  * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
 211  *
 212  * This operation is completely reentrant : can be called while tracing is
 213  * active with absolutely no lock held.
 214  *
 215  * Note, however, that as a v_cmpxchg is used for some atomic operations and
 216  * requires to be executed locally for per-CPU buffers, this function must be
 217  * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
 218  * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
 219  */
 220 static inline
 221 void lib_ring_buffer_switch(
 222                 const struct lttng_ust_lib_ring_buffer_config *config __attribute__((unused)),
 223                 struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode,
 224                 struct lttng_ust_shm_handle *handle)
 225 {
 226         lib_ring_buffer_switch_slow(buf, mode, handle);
 227 }
 228
 229 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 230
 231 /**
 232  * lib_ring_buffer_commit - Commit an record.
 233  * @config: ring buffer instance configuration.
 234  * @ctx: ring buffer context. (input arguments only)
 235  *
 236  * Atomic unordered slot commit. Increments the commit count in the
 237  * specified sub-buffer, and delivers it if necessary.
 238  */
 239 static inline
 240 void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config,
 241                             const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 242 {
 243         struct lttng_ust_lib_ring_buffer_ctx_private *ctx_private = ctx->priv;
 244         struct lttng_ust_lib_ring_buffer_channel *chan = ctx_private->chan;
 245         struct lttng_ust_shm_handle *handle = chan->handle;
 246         struct lttng_ust_lib_ring_buffer *buf = ctx_private->buf;
 247         unsigned long offset_end = ctx_private->buf_offset;
 248         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 249         unsigned long commit_count;
 250         struct commit_counters_hot *cc_hot = shmp_index(handle,
 251                                                 buf->commit_hot, endidx);
 252
 253         if (caa_unlikely(!cc_hot))
 254                 return;
 255
 256         /*
 257          * Must count record before incrementing the commit count.
 258          */
 259         subbuffer_count_record(config, ctx, &buf->backend, endidx, handle);
 260
 261         /*
 262          * Order all writes to buffer before the commit count update that will
 263          * determine that the subbuffer is full.
 264          */
 265         cmm_smp_wmb();
 266
 267         v_add(config, ctx_private->slot_size, &cc_hot->cc);
 268
 269         /*
 270          * commit count read can race with concurrent OOO commit count updates.
 271          * This is only needed for lib_ring_buffer_check_deliver (for
 272          * non-polling delivery only) and for
 273          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 274          * counter to be read with the same value more than once, which could
 275          * cause :
 276          * - Multiple delivery for the same sub-buffer (which is handled
 277          *   gracefully by the reader code) if the value is for a full
 278          *   sub-buffer. It's important that we can never miss a sub-buffer
 279          *   delivery. Re-reading the value after the v_add ensures this.
 280          * - Reading a commit_count with a higher value that what was actually
 281          *   added to it for the lib_ring_buffer_write_commit_counter call
 282          *   (again caused by a concurrent committer). It does not matter,
 283          *   because this function is interested in the fact that the commit
 284          *   count reaches back the reserve offset for a specific sub-buffer,
 285          *   which is completely independent of the order.
 286          */
 287         commit_count = v_read(config, &cc_hot->cc);
 288
 289         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 290                                       commit_count, endidx, handle, ctx_private->tsc);
 291         /*
 292          * Update used size at each commit. It's needed only for extracting
 293          * ring_buffer buffers from vmcore, after crash.
 294          */
 295         lib_ring_buffer_write_commit_counter(config, buf, chan,
 296                         offset_end, commit_count, handle, cc_hot);
 297 }
 298
 299 /**
 300  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 301  * @config: ring buffer instance configuration.
 302  * @ctx: ring buffer context. (input arguments only)
 303  *
 304  * Only succeeds if no other record has been written after the record to
 305  * discard. If discard fails, the record must be committed to the buffer.
 306  *
 307  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 308  */
 309 static inline
 310 int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 311                                         const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 312 {
 313         struct lttng_ust_lib_ring_buffer_ctx_private *ctx_private = ctx->priv;
 314         struct lttng_ust_lib_ring_buffer *buf = ctx_private->buf;
 315         unsigned long end_offset = ctx_private->pre_offset + ctx_private->slot_size;
 316
 317         /*
 318          * We need to ensure that if the cmpxchg succeeds and discards the
 319          * record, the next record will record a full TSC, because it cannot
 320          * rely on the last_tsc associated with the discarded record to detect
 321          * overflows. The only way to ensure this is to set the last_tsc to 0
 322          * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
 323          * timestamp in the next record.
 324          *
 325          * Note: if discard fails, we must leave the TSC in the record header.
 326          * It is needed to keep track of TSC overflows for the following
 327          * records.
 328          */
 329         save_last_tsc(config, buf, 0ULL);
 330
 331         if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx_private->pre_offset)
 332                    != end_offset))
 333                 return -EPERM;
 334         else
 335                 return 0;
 336 }
 337
 338 static inline
 339 void channel_record_disable(
 340                 const struct lttng_ust_lib_ring_buffer_config *config __attribute__((unused)),
 341                 struct lttng_ust_lib_ring_buffer_channel *chan)
 342 {
 343         uatomic_inc(&chan->record_disabled);
 344 }
 345
 346 static inline
 347 void channel_record_enable(
 348                 const struct lttng_ust_lib_ring_buffer_config *config __attribute__((unused)),
 349                 struct lttng_ust_lib_ring_buffer_channel *chan)
 350 {
 351         uatomic_dec(&chan->record_disabled);
 352 }
 353
 354 static inline
 355 void lib_ring_buffer_record_disable(
 356                 const struct lttng_ust_lib_ring_buffer_config *config __attribute__((unused)),
 357                 struct lttng_ust_lib_ring_buffer *buf)
 358 {
 359         uatomic_inc(&buf->record_disabled);
 360 }
 361
 362 static inline
 363 void lib_ring_buffer_record_enable(
 364                 const struct lttng_ust_lib_ring_buffer_config *config __attribute__((unused)),
 365                 struct lttng_ust_lib_ring_buffer *buf)
 366 {
 367         uatomic_dec(&buf->record_disabled);
 368 }
 369
 370 #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */