src/common/ringbuffer/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 "common/getcpu.h"
  19 #include "frontend.h"
  20
  21 /**
  22  * lib_ring_buffer_nesting_inc - Ring buffer recursive use protection.
  23  *
  24  * The rint buffer buffer nesting count is a safety net to ensure tracer
  25  * client code will never trigger an endless recursion.
  26  * Returns a nesting level >= 0 on success, -EPERM on failure (nesting
  27  * count too high).
  28  *
  29  * __asm__ __volatile__ and "memory" clobber prevent the compiler from moving
  30  * instructions out of the ring buffer nesting count. This is required to ensure
  31  * that probe side-effects which can cause recursion (e.g. unforeseen traps,
  32  * divisions by 0, ...) are triggered within the incremented nesting count
  33  * section.
  34  */
  35 static inline
  36 int lib_ring_buffer_nesting_inc(
  37                 const struct lttng_ust_ring_buffer_config *config __attribute__((unused)))
  38 {
  39         int nesting;
  40
  41         nesting = ++URCU_TLS(lib_ring_buffer_nesting);
  42         cmm_barrier();
  43         if (caa_unlikely(nesting >= LIB_RING_BUFFER_MAX_NESTING)) {
  44                 WARN_ON_ONCE(1);
  45                 URCU_TLS(lib_ring_buffer_nesting)--;
  46                 return -EPERM;
  47         }
  48         return nesting - 1;
  49 }
  50
  51 static inline
  52 int lib_ring_buffer_nesting_count(
  53                 const struct lttng_ust_ring_buffer_config *config __attribute__((unused)))
  54 {
  55         return URCU_TLS(lib_ring_buffer_nesting);
  56 }
  57
  58 static inline
  59 void lib_ring_buffer_nesting_dec(
  60                 const struct lttng_ust_ring_buffer_config *config __attribute__((unused)))
  61 {
  62         cmm_barrier();
  63         URCU_TLS(lib_ring_buffer_nesting)--;            /* TLS */
  64 }
  65
  66 /*
  67  * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
  68  * part of the API per se.
  69  *
  70  * returns 0 if reserve ok, or 1 if the slow path must be taken.
  71  */
  72 static inline
  73 int lib_ring_buffer_try_reserve(const struct lttng_ust_ring_buffer_config *config,
  74                                 struct lttng_ust_ring_buffer_ctx *ctx,
  75                                 void *client_ctx,
  76                                 unsigned long *o_begin, unsigned long *o_end,
  77                                 unsigned long *o_old, size_t *before_hdr_pad)
  78 {
  79         struct lttng_ust_ring_buffer_ctx_private *ctx_private = ctx->priv;
  80         struct lttng_ust_ring_buffer_channel *chan = ctx_private->chan;
  81         struct lttng_ust_ring_buffer *buf = ctx_private->buf;
  82         *o_begin = v_read(config, &buf->offset);
  83         *o_old = *o_begin;
  84
  85         ctx_private->timestamp = lib_ring_buffer_clock_read(chan);
  86         if ((int64_t) ctx_private->timestamp == -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_timestamp_overflow(config, buf, ctx_private->timestamp))
  97                 ctx_private->rflags |= RING_BUFFER_RFLAG_FULL_TIMESTAMP;
  98
  99         if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0))
 100                 return 1;
 101
 102         ctx_private->slot_size = record_header_size(config, chan, *o_begin,
 103                                             before_hdr_pad, ctx, client_ctx);
 104         ctx_private->slot_size +=
 105                 lttng_ust_ring_buffer_align(*o_begin + ctx_private->slot_size,
 106                                       ctx->largest_align) + ctx->data_size;
 107         if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx_private->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_private->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
 134  * "timestamp".
 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 lttng_ust_ring_buffer_config *config,
 146                             struct lttng_ust_ring_buffer_ctx *ctx,
 147                             void *client_ctx)
 148 {
 149         struct lttng_ust_ring_buffer_ctx_private *ctx_private = ctx->priv;
 150         struct lttng_ust_ring_buffer_channel *chan = ctx_private->chan;
 151         struct lttng_ust_shm_handle *handle = chan->handle;
 152         struct lttng_ust_ring_buffer *buf;
 153         unsigned long o_begin, o_end, o_old;
 154         size_t before_hdr_pad = 0;
 155
 156         if (caa_unlikely(uatomic_read(&chan->record_disabled)))
 157                 return -EAGAIN;
 158
 159         if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) {
 160                 ctx_private->reserve_cpu = lttng_ust_get_cpu();
 161                 buf = shmp(handle, chan->backend.buf[ctx_private->reserve_cpu].shmp);
 162         } else {
 163                 buf = shmp(handle, chan->backend.buf[0].shmp);
 164         }
 165         if (caa_unlikely(!buf))
 166                 return -EIO;
 167         if (caa_unlikely(uatomic_read(&buf->record_disabled)))
 168                 return -EAGAIN;
 169         ctx_private->buf = buf;
 170
 171         /*
 172          * Perform retryable operations.
 173          */
 174         if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, client_ctx, &o_begin,
 175                                                  &o_end, &o_old, &before_hdr_pad)))
 176                 goto slow_path;
 177
 178         if (caa_unlikely(v_cmpxchg(config, &buf->offset, o_old, o_end)
 179                      != o_old))
 180                 goto slow_path;
 181
 182         /*
 183          * Atomically update last_timestamp. This update races against concurrent
 184          * atomic updates, but the race will always cause supplementary full
 185          * timestamp record headers, never the opposite (missing a full
 186          * timestamp record header when it would be needed).
 187          */
 188         save_last_timestamp(config, buf, ctx_private->timestamp);
 189
 190         /*
 191          * Push the reader if necessary
 192          */
 193         lib_ring_buffer_reserve_push_reader(buf, chan, o_end - 1);
 194
 195         /*
 196          * Clear noref flag for this subbuffer.
 197          */
 198         lib_ring_buffer_clear_noref(config, &buf->backend,
 199                                 subbuf_index(o_end - 1, chan), handle);
 200
 201         ctx_private->pre_offset = o_begin;
 202         ctx_private->buf_offset = o_begin + before_hdr_pad;
 203         return 0;
 204 slow_path:
 205         return lib_ring_buffer_reserve_slow(ctx, client_ctx);
 206 }
 207
 208 /**
 209  * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
 210  * @config: ring buffer instance configuration.
 211  * @buf: buffer
 212  * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
 213  *
 214  * This operation is completely reentrant : can be called while tracing is
 215  * active with absolutely no lock held.
 216  *
 217  * Note, however, that as a v_cmpxchg is used for some atomic operations and
 218  * requires to be executed locally for per-CPU buffers, this function must be
 219  * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
 220  * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
 221  */
 222 static inline
 223 void lib_ring_buffer_switch(
 224                 const struct lttng_ust_ring_buffer_config *config __attribute__((unused)),
 225                 struct lttng_ust_ring_buffer *buf, enum switch_mode mode,
 226                 struct lttng_ust_shm_handle *handle)
 227 {
 228         lib_ring_buffer_switch_slow(buf, mode, handle);
 229 }
 230
 231 /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
 232
 233 /**
 234  * lib_ring_buffer_commit - Commit an record.
 235  * @config: ring buffer instance configuration.
 236  * @ctx: ring buffer context. (input arguments only)
 237  *
 238  * Atomic unordered slot commit. Increments the commit count in the
 239  * specified sub-buffer, and delivers it if necessary.
 240  */
 241 static inline
 242 void lib_ring_buffer_commit(const struct lttng_ust_ring_buffer_config *config,
 243                             const struct lttng_ust_ring_buffer_ctx *ctx)
 244 {
 245         struct lttng_ust_ring_buffer_ctx_private *ctx_private = ctx->priv;
 246         struct lttng_ust_ring_buffer_channel *chan = ctx_private->chan;
 247         struct lttng_ust_shm_handle *handle = chan->handle;
 248         struct lttng_ust_ring_buffer *buf = ctx_private->buf;
 249         unsigned long offset_end = ctx_private->buf_offset;
 250         unsigned long endidx = subbuf_index(offset_end - 1, chan);
 251         unsigned long commit_count;
 252         struct commit_counters_hot *cc_hot = shmp_index(handle,
 253                                                 buf->commit_hot, endidx);
 254
 255         if (caa_unlikely(!cc_hot))
 256                 return;
 257
 258         /*
 259          * Must count record before incrementing the commit count.
 260          */
 261         subbuffer_count_record(config, ctx);
 262
 263         /*
 264          * Order all writes to buffer before the commit count update that will
 265          * determine that the subbuffer is full.
 266          */
 267         cmm_smp_wmb();
 268
 269         v_add(config, ctx_private->slot_size, &cc_hot->cc);
 270
 271         /*
 272          * commit count read can race with concurrent OOO commit count updates.
 273          * This is only needed for lib_ring_buffer_check_deliver (for
 274          * non-polling delivery only) and for
 275          * lib_ring_buffer_write_commit_counter.  The race can only cause the
 276          * counter to be read with the same value more than once, which could
 277          * cause :
 278          * - Multiple delivery for the same sub-buffer (which is handled
 279          *   gracefully by the reader code) if the value is for a full
 280          *   sub-buffer. It's important that we can never miss a sub-buffer
 281          *   delivery. Re-reading the value after the v_add ensures this.
 282          * - Reading a commit_count with a higher value that what was actually
 283          *   added to it for the lib_ring_buffer_write_commit_counter call
 284          *   (again caused by a concurrent committer). It does not matter,
 285          *   because this function is interested in the fact that the commit
 286          *   count reaches back the reserve offset for a specific sub-buffer,
 287          *   which is completely independent of the order.
 288          */
 289         commit_count = v_read(config, &cc_hot->cc);
 290
 291         lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
 292                                       commit_count, endidx, handle, ctx);
 293         /*
 294          * Update used size at each commit. It's needed only for extracting
 295          * ring_buffer buffers from vmcore, after crash.
 296          */
 297         lib_ring_buffer_write_commit_counter(config, buf, chan,
 298                         offset_end, commit_count, handle, cc_hot);
 299 }
 300
 301 /**
 302  * lib_ring_buffer_try_discard_reserve - Try discarding a record.
 303  * @config: ring buffer instance configuration.
 304  * @ctx: ring buffer context. (input arguments only)
 305  *
 306  * Only succeeds if no other record has been written after the record to
 307  * discard. If discard fails, the record must be committed to the buffer.
 308  *
 309  * Returns 0 upon success, -EPERM if the record cannot be discarded.
 310  */
 311 static inline
 312 int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_ring_buffer_config *config,
 313                                         const struct lttng_ust_ring_buffer_ctx *ctx)
 314 {
 315         struct lttng_ust_ring_buffer_ctx_private *ctx_private = ctx->priv;
 316         struct lttng_ust_ring_buffer *buf = ctx_private->buf;
 317         unsigned long end_offset = ctx_private->pre_offset + ctx_private->slot_size;
 318
 319         /*
 320          * We need to ensure that if the cmpxchg succeeds and discards the
 321          * record, the next record will record a full timestamp, because it cannot
 322          * rely on the last_timestamp associated with the discarded record to detect
 323          * overflows. The only way to ensure this is to set the last_timestamp to 0
 324          * (assuming no 64-bit timestamp overflow), which forces to write a 64-bit
 325          * timestamp in the next record.
 326          *
 327          * Note: if discard fails, we must leave the timestamp in the record header.
 328          * It is needed to keep track of timestamp overflows for the following
 329          * records.
 330          */
 331         save_last_timestamp(config, buf, 0ULL);
 332
 333         if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx_private->pre_offset)
 334                    != end_offset))
 335                 return -EPERM;
 336         else
 337                 return 0;
 338 }
 339
 340 static inline
 341 void channel_record_disable(
 342                 const struct lttng_ust_ring_buffer_config *config __attribute__((unused)),
 343                 struct lttng_ust_ring_buffer_channel *chan)
 344 {
 345         uatomic_inc(&chan->record_disabled);
 346 }
 347
 348 static inline
 349 void channel_record_enable(
 350                 const struct lttng_ust_ring_buffer_config *config __attribute__((unused)),
 351                 struct lttng_ust_ring_buffer_channel *chan)
 352 {
 353         uatomic_dec(&chan->record_disabled);
 354 }
 355
 356 static inline
 357 void lib_ring_buffer_record_disable(
 358                 const struct lttng_ust_ring_buffer_config *config __attribute__((unused)),
 359                 struct lttng_ust_ring_buffer *buf)
 360 {
 361         uatomic_inc(&buf->record_disabled);
 362 }
 363
 364 static inline
 365 void lib_ring_buffer_record_enable(
 366                 const struct lttng_ust_ring_buffer_config *config __attribute__((unused)),
 367                 struct lttng_ust_ring_buffer *buf)
 368 {
 369         uatomic_dec(&buf->record_disabled);
 370 }
 371
 372 #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */