Performance: Relax atomicity constraints for crash handling
[lttng-ust.git] / libringbuffer / frontend_api.h
CommitLineData
e92f3e28
MD
1#ifndef _LTTNG_RING_BUFFER_FRONTEND_API_H
2#define _LTTNG_RING_BUFFER_FRONTEND_API_H
852c2936
MD
3
4/*
e92f3e28 5 * libringbuffer/frontend_api.h
852c2936 6 *
e92f3e28
MD
7 * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
8 *
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; only
12 * version 2.1 of the License.
13 *
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
852c2936
MD
22 *
23 * Ring Buffer Library Synchronization Header (buffer write API).
24 *
25 * Author:
e92f3e28 26 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
852c2936 27 *
e92f3e28
MD
28 * See ring_buffer_frontend.c for more information on wait-free
29 * algorithms.
30 * See frontend.h for channel allocation and read-side API.
852c2936
MD
31 */
32
4931a13e 33#include "frontend.h"
9f3fdbc6
MD
34#include <urcu-bp.h>
35#include <urcu/compiler.h>
852c2936
MD
36
37/**
38 * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit.
39 *
36cb884c
MD
40 * Keeps a ring buffer nesting count as supplementary safety net to
41 * ensure tracer client code will never trigger an endless recursion.
42 * Returns the processor ID on success, -EPERM on failure (nesting count
43 * too high).
852c2936
MD
44 *
45 * asm volatile and "memory" clobber prevent the compiler from moving
46 * instructions out of the ring buffer nesting count. This is required to ensure
47 * that probe side-effects which can cause recursion (e.g. unforeseen traps,
48 * divisions by 0, ...) are triggered within the incremented nesting count
49 * section.
50 */
51static inline
4cfec15c 52int lib_ring_buffer_get_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
852c2936
MD
53{
54 int cpu, nesting;
55
2b2d6ff7 56 cpu = lttng_ust_get_cpu();
8c90a710 57 nesting = ++URCU_TLS(lib_ring_buffer_nesting);
9f3fdbc6 58 cmm_barrier();
852c2936 59
b5a3dfa5 60 if (caa_unlikely(nesting > 4)) {
852c2936 61 WARN_ON_ONCE(1);
8c90a710 62 URCU_TLS(lib_ring_buffer_nesting)--;
852c2936
MD
63 return -EPERM;
64 } else
65 return cpu;
66}
67
68/**
69 * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit.
70 */
71static inline
4cfec15c 72void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config *config)
852c2936 73{
9f3fdbc6 74 cmm_barrier();
8c90a710 75 URCU_TLS(lib_ring_buffer_nesting)--; /* TLS */
852c2936
MD
76}
77
78/*
79 * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not
80 * part of the API per se.
81 *
82 * returns 0 if reserve ok, or 1 if the slow path must be taken.
83 */
84static inline
4cfec15c
MD
85int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
86 struct lttng_ust_lib_ring_buffer_ctx *ctx,
852c2936
MD
87 unsigned long *o_begin, unsigned long *o_end,
88 unsigned long *o_old, size_t *before_hdr_pad)
89{
90 struct channel *chan = ctx->chan;
4cfec15c 91 struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
852c2936
MD
92 *o_begin = v_read(config, &buf->offset);
93 *o_old = *o_begin;
94
95 ctx->tsc = lib_ring_buffer_clock_read(chan);
96 if ((int64_t) ctx->tsc == -EIO)
97 return 1;
98
99 /*
100 * Prefetch cacheline for read because we have to read the previous
101 * commit counter to increment it and commit seq value to compare it to
102 * the commit counter.
103 */
9f3fdbc6 104 //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]);
852c2936
MD
105
106 if (last_tsc_overflow(config, buf, ctx->tsc))
107 ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC;
108
b5a3dfa5 109 if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0))
852c2936
MD
110 return 1;
111
112 ctx->slot_size = record_header_size(config, chan, *o_begin,
113 before_hdr_pad, ctx);
114 ctx->slot_size +=
115 lib_ring_buffer_align(*o_begin + ctx->slot_size,
116 ctx->largest_align) + ctx->data_size;
b5a3dfa5 117 if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size)
852c2936
MD
118 > chan->backend.subbuf_size))
119 return 1;
120
121 /*
122 * Record fits in the current buffer and we are not on a switch
123 * boundary. It's safe to write.
124 */
125 *o_end = *o_begin + ctx->slot_size;
1ad21f70
MD
126
127 if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0))
128 /*
129 * The offset_end will fall at the very beginning of the next
130 * subbuffer.
131 */
132 return 1;
133
852c2936
MD
134 return 0;
135}
136
137/**
138 * lib_ring_buffer_reserve - Reserve space in a ring buffer.
139 * @config: ring buffer instance configuration.
140 * @ctx: ring buffer context. (input and output) Must be already initialized.
141 *
142 * Atomic wait-free slot reservation. The reserved space starts at the context
143 * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc".
144 *
145 * Return :
146 * 0 on success.
147 * -EAGAIN if channel is disabled.
148 * -ENOSPC if event size is too large for packet.
149 * -ENOBUFS if there is currently not enough space in buffer for the event.
150 * -EIO if data cannot be written into the buffer for any other reason.
151 */
152
153static inline
4cfec15c
MD
154int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
155 struct lttng_ust_lib_ring_buffer_ctx *ctx)
852c2936
MD
156{
157 struct channel *chan = ctx->chan;
38fae1d3 158 struct lttng_ust_shm_handle *handle = ctx->handle;
4cfec15c 159 struct lttng_ust_lib_ring_buffer *buf;
852c2936
MD
160 unsigned long o_begin, o_end, o_old;
161 size_t before_hdr_pad = 0;
162
9f3fdbc6 163 if (uatomic_read(&chan->record_disabled))
852c2936
MD
164 return -EAGAIN;
165
166 if (config->alloc == RING_BUFFER_ALLOC_PER_CPU)
1d498196 167 buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp);
852c2936 168 else
1d498196 169 buf = shmp(handle, chan->backend.buf[0].shmp);
9f3fdbc6 170 if (uatomic_read(&buf->record_disabled))
852c2936
MD
171 return -EAGAIN;
172 ctx->buf = buf;
173
174 /*
175 * Perform retryable operations.
176 */
b5a3dfa5 177 if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, &o_begin,
852c2936
MD
178 &o_end, &o_old, &before_hdr_pad)))
179 goto slow_path;
180
b5a3dfa5 181 if (caa_unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end)
852c2936
MD
182 != o_old))
183 goto slow_path;
184
185 /*
186 * Atomically update last_tsc. This update races against concurrent
187 * atomic updates, but the race will always cause supplementary full TSC
188 * record headers, never the opposite (missing a full TSC record header
189 * when it would be needed).
190 */
191 save_last_tsc(config, ctx->buf, ctx->tsc);
192
193 /*
194 * Push the reader if necessary
195 */
196 lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1);
197
198 /*
199 * Clear noref flag for this subbuffer.
200 */
201 lib_ring_buffer_clear_noref(config, &ctx->buf->backend,
1d498196 202 subbuf_index(o_end - 1, chan), handle);
852c2936
MD
203
204 ctx->pre_offset = o_begin;
205 ctx->buf_offset = o_begin + before_hdr_pad;
206 return 0;
207slow_path:
208 return lib_ring_buffer_reserve_slow(ctx);
209}
210
211/**
212 * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer.
213 * @config: ring buffer instance configuration.
214 * @buf: buffer
215 * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH)
216 *
217 * This operation is completely reentrant : can be called while tracing is
218 * active with absolutely no lock held.
219 *
220 * Note, however, that as a v_cmpxchg is used for some atomic operations and
221 * requires to be executed locally for per-CPU buffers, this function must be
222 * called from the CPU which owns the buffer for a ACTIVE flush, with preemption
223 * disabled, for RING_BUFFER_SYNC_PER_CPU configuration.
224 */
225static inline
4cfec15c
MD
226void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config,
227 struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode,
38fae1d3 228 struct lttng_ust_shm_handle *handle)
852c2936 229{
1d498196 230 lib_ring_buffer_switch_slow(buf, mode, handle);
852c2936
MD
231}
232
233/* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */
234
235/**
236 * lib_ring_buffer_commit - Commit an record.
237 * @config: ring buffer instance configuration.
238 * @ctx: ring buffer context. (input arguments only)
239 *
240 * Atomic unordered slot commit. Increments the commit count in the
241 * specified sub-buffer, and delivers it if necessary.
242 */
243static inline
4cfec15c
MD
244void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config,
245 const struct lttng_ust_lib_ring_buffer_ctx *ctx)
852c2936
MD
246{
247 struct channel *chan = ctx->chan;
38fae1d3 248 struct lttng_ust_shm_handle *handle = ctx->handle;
4cfec15c 249 struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
852c2936
MD
250 unsigned long offset_end = ctx->buf_offset;
251 unsigned long endidx = subbuf_index(offset_end - 1, chan);
252 unsigned long commit_count;
253
254 /*
255 * Must count record before incrementing the commit count.
256 */
1d498196 257 subbuffer_count_record(config, &buf->backend, endidx, handle);
852c2936
MD
258
259 /*
260 * Order all writes to buffer before the commit count update that will
261 * determine that the subbuffer is full.
262 */
9f3fdbc6 263 cmm_smp_wmb();
852c2936 264
4746ae29 265 v_add(config, ctx->slot_size, &shmp_index(handle, buf->commit_hot, endidx)->cc);
852c2936
MD
266
267 /*
268 * commit count read can race with concurrent OOO commit count updates.
269 * This is only needed for lib_ring_buffer_check_deliver (for
270 * non-polling delivery only) and for
271 * lib_ring_buffer_write_commit_counter. The race can only cause the
272 * counter to be read with the same value more than once, which could
273 * cause :
274 * - Multiple delivery for the same sub-buffer (which is handled
275 * gracefully by the reader code) if the value is for a full
276 * sub-buffer. It's important that we can never miss a sub-buffer
277 * delivery. Re-reading the value after the v_add ensures this.
278 * - Reading a commit_count with a higher value that what was actually
279 * added to it for the lib_ring_buffer_write_commit_counter call
280 * (again caused by a concurrent committer). It does not matter,
281 * because this function is interested in the fact that the commit
282 * count reaches back the reserve offset for a specific sub-buffer,
283 * which is completely independent of the order.
284 */
4746ae29 285 commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc);
852c2936
MD
286
287 lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1,
1b7b0501 288 commit_count, endidx, handle, ctx->tsc);
852c2936
MD
289 /*
290 * Update used size at each commit. It's needed only for extracting
291 * ring_buffer buffers from vmcore, after crash.
292 */
293 lib_ring_buffer_write_commit_counter(config, buf, chan, endidx,
80249235 294 offset_end, commit_count, handle);
852c2936
MD
295}
296
297/**
298 * lib_ring_buffer_try_discard_reserve - Try discarding a record.
299 * @config: ring buffer instance configuration.
300 * @ctx: ring buffer context. (input arguments only)
301 *
302 * Only succeeds if no other record has been written after the record to
303 * discard. If discard fails, the record must be committed to the buffer.
304 *
305 * Returns 0 upon success, -EPERM if the record cannot be discarded.
306 */
307static inline
4cfec15c
MD
308int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config,
309 const struct lttng_ust_lib_ring_buffer_ctx *ctx)
852c2936 310{
4cfec15c 311 struct lttng_ust_lib_ring_buffer *buf = ctx->buf;
852c2936
MD
312 unsigned long end_offset = ctx->pre_offset + ctx->slot_size;
313
314 /*
315 * We need to ensure that if the cmpxchg succeeds and discards the
316 * record, the next record will record a full TSC, because it cannot
317 * rely on the last_tsc associated with the discarded record to detect
318 * overflows. The only way to ensure this is to set the last_tsc to 0
319 * (assuming no 64-bit TSC overflow), which forces to write a 64-bit
320 * timestamp in the next record.
321 *
322 * Note: if discard fails, we must leave the TSC in the record header.
323 * It is needed to keep track of TSC overflows for the following
324 * records.
325 */
326 save_last_tsc(config, buf, 0ULL);
327
b5a3dfa5 328 if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset)
852c2936
MD
329 != end_offset))
330 return -EPERM;
331 else
332 return 0;
333}
334
335static inline
4cfec15c 336void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
852c2936
MD
337 struct channel *chan)
338{
9f3fdbc6 339 uatomic_inc(&chan->record_disabled);
852c2936
MD
340}
341
342static inline
4cfec15c 343void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
852c2936
MD
344 struct channel *chan)
345{
9f3fdbc6 346 uatomic_dec(&chan->record_disabled);
852c2936
MD
347}
348
349static inline
4cfec15c
MD
350void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config,
351 struct lttng_ust_lib_ring_buffer *buf)
852c2936 352{
9f3fdbc6 353 uatomic_inc(&buf->record_disabled);
852c2936
MD
354}
355
356static inline
4cfec15c
MD
357void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config,
358 struct lttng_ust_lib_ring_buffer *buf)
852c2936 359{
9f3fdbc6 360 uatomic_dec(&buf->record_disabled);
852c2936
MD
361}
362
e92f3e28 363#endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */
This page took 0.062261 seconds and 4 git commands to generate.