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(const struct lttng_ust_lib_ring_buffer_config *config)
  36 {
  37         int nesting;
  38
  39         nesting = ++URCU_TLS(lib_ring_buffer_nesting);
  40         cmm_barrier();
  41         if (caa_unlikely(nesting >= LIB_RING_BUFFER_MAX_NESTING)) {
  42                 WARN_ON_ONCE(1);
  43                 URCU_TLS(lib_ring_buffer_nesting)--;
  44                 return -EPERM;
  45         }
  46         return nesting - 1;
  47 }
  48
  49 static inline
  50 int lib_ring_buffer_nesting_count(const struct lttng_ust_lib_ring_buffer_config *config)
  51 {
  52         return URCU_TLS(lib_ring_buffer_nesting);
  53 }
  54
  55 static inline
  56 void lib_ring_buffer_nesting_dec(const struct lttng_ust_lib_ring_buffer_config *config)
  57 {
  58         cmm_barrier();
  59         URCU_TLS(lib_ring_buffer_nesting)--;            /* TLS */
  60 }
  61
  62 /*
  63  * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
  64  * part of the API per se.
  65  *
  66  * returns 0 if reserve ok, or 1 if the slow path must be taken.
  67  */
  68 static inline
  69 int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
  70                                 struct lttng_ust_lib_ring_buffer_ctx *ctx,
  71                                 void *client_ctx,
  72                                 unsigned long *o_begin, unsigned long *o_end,
  73                                 unsigned long *o_old, size_t *before_hdr_pad)
  74 {
  75         struct lttng_ust_lib_ring_buffer_ctx_private *ctx_private = ctx->priv;
  76         struct lttng_ust_lib_ring_buffer_channel *chan = ctx_private->chan;
  77         struct lttng_ust_lib_ring_buffer *buf = ctx_private->buf;
  78         *o_begin = v_read(config, &buf->offset);
  79         *o_old = *o_begin;
  80
  81         ctx_private->tsc = lib_ring_buffer_clock_read(chan);
  82         if ((int64_t) ctx_private->tsc == -EIO)
  83                 return 1;
  84
  85         /*
  86          * Prefetch cacheline for read because we have to read the previous
  87          * commit counter to increment it and commit seq value to compare it to
  88          * the commit counter.
  89          */
  90         //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
  91
  92         if (last_tsc_overflow(config, buf, ctx_private->tsc))
  93                 ctx_private->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
  94
  95         if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0))
  96                 return 1;
  97
  98         ctx_private->slot_size = record_header_size(config, chan, *o_begin,
  99                                             before_hdr_pad, ctx, client_ctx);
 100         ctx_private->slot_size +=
 101                 lttng_ust_lib_ring_buffer_align(*o_begin + ctx_private->slot_size,
 102                                       ctx->largest_align) + ctx->data_size;
 103         if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx_private->slot_size)
 104                      > chan->backend.subbuf_size))
 105                 return 1;
 106
 107         /*
 108          * Record fits in the current buffer and we are not on a switch
 109          * boundary. It's safe to write.
 110          */
 111         *o_end = *o_begin + ctx_private->slot_size;
 112
 113         if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0))
 114                 /*
 115                  * The offset_end will fall at the very beginning of the next
 116                  * subbuffer.
 117                  */
 118                 return 1;
 119
 120         return 0;
 121 }
 122
 123 /**
 124  * lib_ring_buffer_reserve - Reserve space in a ring buffer.
 125  * @config: ring buffer instance configuration.
 126  * @ctx: ring buffer context. (input and output) Must be already initialized.
 127  *
 128  * Atomic wait-free slot reservation. The reserved space starts at the context
 129  * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
 130  *
 131  * Return :
 132  *  0 on success.
 133  * -EAGAIN if channel is disabled.
 134  * -ENOSPC if event size is too large for packet.
 135  * -ENOBUFS if there is currently not enough space in buffer for the event.
 136  * -EIO if data cannot be written into the buffer for any other reason.
 137  */
 138
 139 static inline
 140 int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 141                             struct lttng_ust_lib_ring_buffer_ctx *ctx,
 142                             void *client_ctx)
 143 {
 144         struct lttng_ust_lib_ring_buffer_ctx_private *ctx_private = ctx->priv;
 145         struct lttng_ust_lib_ring_buffer_channel *chan = ctx_private->chan;
 146         struct lttng_ust_shm_handle *handle = chan->handle;
 147         struct lttng_ust_lib_ring_buffer *buf;
 148         unsigned long o_begin, o_end, o_old;
 149         size_t before_hdr_pad = 0;
 150
 151         if (caa_unlikely(uatomic_read(&chan->record_disabled)))
 152                 return -EAGAIN;
 153
 154         if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
 155                 ctx_private->reserve_cpu = lttng_ust_get_cpu();
 156                 buf = shmp(handle, chan->backend.buf[ctx_private->reserve_cpu].shmp);
 157         } else {
 158                 buf = shmp(handle, chan->backend.buf[0].shmp);
 159         }
 160         if (caa_unlikely(!buf))
 161                 return -EIO;
 162         if (caa_unlikely(uatomic_read(&buf->record_disabled)))
 163                 return -EAGAIN;
 164         ctx_private->buf = buf;
 165
 166         /*
 167          * Perform retryable operations.
 168          */
 169         if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, client_ctx, &o_begin,
 170                                                  &o_end, &o_old, &before_hdr_pad)))
 171                 goto slow_path;
 172
 173         if (caa_unlikely(v_cmpxchg(config, &buf->offset, o_old, o_end)
 174                      != o_old))
 175                 goto slow_path;
 176
 177         /*
 178          * Atomically update last_tsc. This update races against concurrent
 179          * atomic updates, but the race will always cause supplementary full TSC
 180          * record headers, never the opposite (missing a full TSC record header
 181          * when it would be needed).
 182          */
 183         save_last_tsc(config, buf, ctx_private->tsc);
 184
 185         /*
 186          * Push the reader if necessary
 187          */
 188         lib_ring_buffer_reserve_push_reader(buf, chan, o_end - 1);
 189
 190         /*
 191          * Clear noref flag for this subbuffer.
 192          */
 193         lib_ring_buffer_clear_noref(config, &buf->backend,
 194                                 subbuf_index(o_end - 1, chan), handle);
 195
 196         ctx_private->pre_offset = o_begin;
 197         ctx_private->buf_offset = o_begin + before_hdr_pad;
 198         return 0;
 199 slow_path:
 200         return lib_ring_buffer_reserve_slow(ctx, client_ctx);
 201 }
 202
 203 /**
 204  * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
 205  * @config: ring buffer instance configuration.
 206  * @buf: buffer
 207  * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
 208  *
 209  * This operation is completely reentrant : can be called while tracing is
 210  * active with absolutely no lock held.
 211  *
 212  * Note, however, that as a v_cmpxchg is used for some atomic operations and
 213  * requires to be executed locally for per-CPU buffers, this function must be
 214  * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
 215  * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
 216  */
 217 static inline
 218 void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config,
 219                             struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode,
 220                             struct lttng_ust_shm_handle *handle)
 221 {
 222         lib_ring_buffer_switch_slow(buf, mode, handle);
 223 }
 224
 225 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 226
 227 /**
 228  * lib_ring_buffer_commit - Commit an record.
 229  * @config: ring buffer instance configuration.
 230  * @ctx: ring buffer context. (input arguments only)
 231  *
 232  * Atomic unordered slot commit. Increments the commit count in the
 233  * specified sub-buffer, and delivers it if necessary.
 234  */
 235 static inline
 236 void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config,
 237                             const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 238 {
 239         struct lttng_ust_lib_ring_buffer_ctx_private *ctx_private = ctx->priv;
 240         struct lttng_ust_lib_ring_buffer_channel *chan = ctx_private->chan;
 241         struct lttng_ust_shm_handle *handle = chan->handle;
 242         struct lttng_ust_lib_ring_buffer *buf = ctx_private->buf;
 243         unsigned long offset_end = ctx_private->buf_offset;
 244         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 245         unsigned long commit_count;
 246         struct commit_counters_hot *cc_hot = shmp_index(handle,
 247                                                 buf->commit_hot, endidx);
 248
 249         if (caa_unlikely(!cc_hot))
 250                 return;
 251
 252         /*
 253          * Must count record before incrementing the commit count.
 254          */
 255         subbuffer_count_record(config, ctx, &buf->backend, endidx, handle);
 256
 257         /*
 258          * Order all writes to buffer before the commit count update that will
 259          * determine that the subbuffer is full.
 260          */
 261         cmm_smp_wmb();
 262
 263         v_add(config, ctx_private->slot_size, &cc_hot->cc);
 264
 265         /*
 266          * commit count read can race with concurrent OOO commit count updates.
 267          * This is only needed for lib_ring_buffer_check_deliver (for
 268          * non-polling delivery only) and for
 269          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 270          * counter to be read with the same value more than once, which could
 271          * cause :
 272          * - Multiple delivery for the same sub-buffer (which is handled
 273          *   gracefully by the reader code) if the value is for a full
 274          *   sub-buffer. It's important that we can never miss a sub-buffer
 275          *   delivery. Re-reading the value after the v_add ensures this.
 276          * - Reading a commit_count with a higher value that what was actually
 277          *   added to it for the lib_ring_buffer_write_commit_counter call
 278          *   (again caused by a concurrent committer). It does not matter,
 279          *   because this function is interested in the fact that the commit
 280          *   count reaches back the reserve offset for a specific sub-buffer,
 281          *   which is completely independent of the order.
 282          */
 283         commit_count = v_read(config, &cc_hot->cc);
 284
 285         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 286                                       commit_count, endidx, handle, ctx_private->tsc);
 287         /*
 288          * Update used size at each commit. It's needed only for extracting
 289          * ring_buffer buffers from vmcore, after crash.
 290          */
 291         lib_ring_buffer_write_commit_counter(config, buf, chan,
 292                         offset_end, commit_count, handle, cc_hot);
 293 }
 294
 295 /**
 296  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 297  * @config: ring buffer instance configuration.
 298  * @ctx: ring buffer context. (input arguments only)
 299  *
 300  * Only succeeds if no other record has been written after the record to
 301  * discard. If discard fails, the record must be committed to the buffer.
 302  *
 303  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 304  */
 305 static inline
 306 int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
 307                                         const struct lttng_ust_lib_ring_buffer_ctx *ctx)
 308 {
 309         struct lttng_ust_lib_ring_buffer_ctx_private *ctx_private = ctx->priv;
 310         struct lttng_ust_lib_ring_buffer *buf = ctx_private->buf;
 311         unsigned long end_offset = ctx_private->pre_offset + ctx_private->slot_size;
 312
 313         /*
 314          * We need to ensure that if the cmpxchg succeeds and discards the
 315          * record, the next record will record a full TSC, because it cannot
 316          * rely on the last_tsc associated with the discarded record to detect
 317          * overflows. The only way to ensure this is to set the last_tsc to 0
 318          * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
 319          * timestamp in the next record.
 320          *
 321          * Note: if discard fails, we must leave the TSC in the record header.
 322          * It is needed to keep track of TSC overflows for the following
 323          * records.
 324          */
 325         save_last_tsc(config, buf, 0ULL);
 326
 327         if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx_private->pre_offset)
 328                    != end_offset))
 329                 return -EPERM;
 330         else
 331                 return 0;
 332 }
 333
 334 static inline
 335 void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 336                             struct lttng_ust_lib_ring_buffer_channel *chan)
 337 {
 338         uatomic_inc(&chan->record_disabled);
 339 }
 340
 341 static inline
 342 void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 343                            struct lttng_ust_lib_ring_buffer_channel *chan)
 344 {
 345         uatomic_dec(&chan->record_disabled);
 346 }
 347
 348 static inline
 349 void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
 350                                     struct lttng_ust_lib_ring_buffer *buf)
 351 {
 352         uatomic_inc(&buf->record_disabled);
 353 }
 354
 355 static inline
 356 void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
 357                                    struct lttng_ust_lib_ring_buffer *buf)
 358 {
 359         uatomic_dec(&buf->record_disabled);
 360 }
 361
 362 #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */