Commit | Line | Data |
---|---|---|
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 | */ | |
51 | static inline | |
4cfec15c | 52 | int 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)--; |
9f3fdbc6 | 63 | rcu_read_unlock(); |
852c2936 MD |
64 | return -EPERM; |
65 | } else | |
66 | return cpu; | |
67 | } | |
68 | ||
69 | /** | |
70 | * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit. | |
71 | */ | |
72 | static inline | |
4cfec15c | 73 | void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config *config) |
852c2936 | 74 | { |
9f3fdbc6 | 75 | cmm_barrier(); |
8c90a710 | 76 | URCU_TLS(lib_ring_buffer_nesting)--; /* TLS */ |
852c2936 MD |
77 | } |
78 | ||
79 | /* | |
80 | * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not | |
81 | * part of the API per se. | |
82 | * | |
83 | * returns 0 if reserve ok, or 1 if the slow path must be taken. | |
84 | */ | |
85 | static inline | |
4cfec15c MD |
86 | int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
87 | struct lttng_ust_lib_ring_buffer_ctx *ctx, | |
852c2936 MD |
88 | unsigned long *o_begin, unsigned long *o_end, |
89 | unsigned long *o_old, size_t *before_hdr_pad) | |
90 | { | |
91 | struct channel *chan = ctx->chan; | |
4cfec15c | 92 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
93 | *o_begin = v_read(config, &buf->offset); |
94 | *o_old = *o_begin; | |
95 | ||
96 | ctx->tsc = lib_ring_buffer_clock_read(chan); | |
97 | if ((int64_t) ctx->tsc == -EIO) | |
98 | return 1; | |
99 | ||
100 | /* | |
101 | * Prefetch cacheline for read because we have to read the previous | |
102 | * commit counter to increment it and commit seq value to compare it to | |
103 | * the commit counter. | |
104 | */ | |
9f3fdbc6 | 105 | //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]); |
852c2936 MD |
106 | |
107 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
108 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
109 | ||
b5a3dfa5 | 110 | if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0)) |
852c2936 MD |
111 | return 1; |
112 | ||
113 | ctx->slot_size = record_header_size(config, chan, *o_begin, | |
114 | before_hdr_pad, ctx); | |
115 | ctx->slot_size += | |
116 | lib_ring_buffer_align(*o_begin + ctx->slot_size, | |
117 | ctx->largest_align) + ctx->data_size; | |
b5a3dfa5 | 118 | if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size) |
852c2936 MD |
119 | > chan->backend.subbuf_size)) |
120 | return 1; | |
121 | ||
122 | /* | |
123 | * Record fits in the current buffer and we are not on a switch | |
124 | * boundary. It's safe to write. | |
125 | */ | |
126 | *o_end = *o_begin + ctx->slot_size; | |
1ad21f70 MD |
127 | |
128 | if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0)) | |
129 | /* | |
130 | * The offset_end will fall at the very beginning of the next | |
131 | * subbuffer. | |
132 | */ | |
133 | return 1; | |
134 | ||
852c2936 MD |
135 | return 0; |
136 | } | |
137 | ||
138 | /** | |
139 | * lib_ring_buffer_reserve - Reserve space in a ring buffer. | |
140 | * @config: ring buffer instance configuration. | |
141 | * @ctx: ring buffer context. (input and output) Must be already initialized. | |
142 | * | |
143 | * Atomic wait-free slot reservation. The reserved space starts at the context | |
144 | * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc". | |
145 | * | |
146 | * Return : | |
147 | * 0 on success. | |
148 | * -EAGAIN if channel is disabled. | |
149 | * -ENOSPC if event size is too large for packet. | |
150 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
151 | * -EIO if data cannot be written into the buffer for any other reason. | |
152 | */ | |
153 | ||
154 | static inline | |
4cfec15c MD |
155 | int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
156 | struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 MD |
157 | { |
158 | struct channel *chan = ctx->chan; | |
38fae1d3 | 159 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c | 160 | struct lttng_ust_lib_ring_buffer *buf; |
852c2936 MD |
161 | unsigned long o_begin, o_end, o_old; |
162 | size_t before_hdr_pad = 0; | |
163 | ||
9f3fdbc6 | 164 | if (uatomic_read(&chan->record_disabled)) |
852c2936 MD |
165 | return -EAGAIN; |
166 | ||
167 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 168 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 169 | else |
1d498196 | 170 | buf = shmp(handle, chan->backend.buf[0].shmp); |
9f3fdbc6 | 171 | if (uatomic_read(&buf->record_disabled)) |
852c2936 MD |
172 | return -EAGAIN; |
173 | ctx->buf = buf; | |
174 | ||
175 | /* | |
176 | * Perform retryable operations. | |
177 | */ | |
b5a3dfa5 | 178 | if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, &o_begin, |
852c2936 MD |
179 | &o_end, &o_old, &before_hdr_pad))) |
180 | goto slow_path; | |
181 | ||
b5a3dfa5 | 182 | if (caa_unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end) |
852c2936 MD |
183 | != o_old)) |
184 | goto slow_path; | |
185 | ||
186 | /* | |
187 | * Atomically update last_tsc. This update races against concurrent | |
188 | * atomic updates, but the race will always cause supplementary full TSC | |
189 | * record headers, never the opposite (missing a full TSC record header | |
190 | * when it would be needed). | |
191 | */ | |
192 | save_last_tsc(config, ctx->buf, ctx->tsc); | |
193 | ||
194 | /* | |
195 | * Push the reader if necessary | |
196 | */ | |
197 | lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1); | |
198 | ||
199 | /* | |
200 | * Clear noref flag for this subbuffer. | |
201 | */ | |
202 | lib_ring_buffer_clear_noref(config, &ctx->buf->backend, | |
1d498196 | 203 | subbuf_index(o_end - 1, chan), handle); |
852c2936 MD |
204 | |
205 | ctx->pre_offset = o_begin; | |
206 | ctx->buf_offset = o_begin + before_hdr_pad; | |
207 | return 0; | |
208 | slow_path: | |
209 | return lib_ring_buffer_reserve_slow(ctx); | |
210 | } | |
211 | ||
212 | /** | |
213 | * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer. | |
214 | * @config: ring buffer instance configuration. | |
215 | * @buf: buffer | |
216 | * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH) | |
217 | * | |
218 | * This operation is completely reentrant : can be called while tracing is | |
219 | * active with absolutely no lock held. | |
220 | * | |
221 | * Note, however, that as a v_cmpxchg is used for some atomic operations and | |
222 | * requires to be executed locally for per-CPU buffers, this function must be | |
223 | * called from the CPU which owns the buffer for a ACTIVE flush, with preemption | |
224 | * disabled, for RING_BUFFER_SYNC_PER_CPU configuration. | |
225 | */ | |
226 | static inline | |
4cfec15c MD |
227 | void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config, |
228 | struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, | |
38fae1d3 | 229 | struct lttng_ust_shm_handle *handle) |
852c2936 | 230 | { |
1d498196 | 231 | lib_ring_buffer_switch_slow(buf, mode, handle); |
852c2936 MD |
232 | } |
233 | ||
234 | /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */ | |
235 | ||
236 | /** | |
237 | * lib_ring_buffer_commit - Commit an record. | |
238 | * @config: ring buffer instance configuration. | |
239 | * @ctx: ring buffer context. (input arguments only) | |
240 | * | |
241 | * Atomic unordered slot commit. Increments the commit count in the | |
242 | * specified sub-buffer, and delivers it if necessary. | |
243 | */ | |
244 | static inline | |
4cfec15c MD |
245 | void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config, |
246 | const struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 MD |
247 | { |
248 | struct channel *chan = ctx->chan; | |
38fae1d3 | 249 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c | 250 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
251 | unsigned long offset_end = ctx->buf_offset; |
252 | unsigned long endidx = subbuf_index(offset_end - 1, chan); | |
253 | unsigned long commit_count; | |
254 | ||
255 | /* | |
256 | * Must count record before incrementing the commit count. | |
257 | */ | |
1d498196 | 258 | subbuffer_count_record(config, &buf->backend, endidx, handle); |
852c2936 MD |
259 | |
260 | /* | |
261 | * Order all writes to buffer before the commit count update that will | |
262 | * determine that the subbuffer is full. | |
263 | */ | |
9f3fdbc6 | 264 | cmm_smp_wmb(); |
852c2936 | 265 | |
4746ae29 | 266 | v_add(config, ctx->slot_size, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
852c2936 MD |
267 | |
268 | /* | |
269 | * commit count read can race with concurrent OOO commit count updates. | |
270 | * This is only needed for lib_ring_buffer_check_deliver (for | |
271 | * non-polling delivery only) and for | |
272 | * lib_ring_buffer_write_commit_counter. The race can only cause the | |
273 | * counter to be read with the same value more than once, which could | |
274 | * cause : | |
275 | * - Multiple delivery for the same sub-buffer (which is handled | |
276 | * gracefully by the reader code) if the value is for a full | |
277 | * sub-buffer. It's important that we can never miss a sub-buffer | |
278 | * delivery. Re-reading the value after the v_add ensures this. | |
279 | * - Reading a commit_count with a higher value that what was actually | |
280 | * added to it for the lib_ring_buffer_write_commit_counter call | |
281 | * (again caused by a concurrent committer). It does not matter, | |
282 | * because this function is interested in the fact that the commit | |
283 | * count reaches back the reserve offset for a specific sub-buffer, | |
284 | * which is completely independent of the order. | |
285 | */ | |
4746ae29 | 286 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
852c2936 MD |
287 | |
288 | lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1, | |
1b7b0501 | 289 | commit_count, endidx, handle, ctx->tsc); |
852c2936 MD |
290 | /* |
291 | * Update used size at each commit. It's needed only for extracting | |
292 | * ring_buffer buffers from vmcore, after crash. | |
293 | */ | |
294 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, | |
80249235 | 295 | offset_end, commit_count, handle); |
852c2936 MD |
296 | } |
297 | ||
298 | /** | |
299 | * lib_ring_buffer_try_discard_reserve - Try discarding a record. | |
300 | * @config: ring buffer instance configuration. | |
301 | * @ctx: ring buffer context. (input arguments only) | |
302 | * | |
303 | * Only succeeds if no other record has been written after the record to | |
304 | * discard. If discard fails, the record must be committed to the buffer. | |
305 | * | |
306 | * Returns 0 upon success, -EPERM if the record cannot be discarded. | |
307 | */ | |
308 | static inline | |
4cfec15c MD |
309 | int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
310 | const struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 | 311 | { |
4cfec15c | 312 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
313 | unsigned long end_offset = ctx->pre_offset + ctx->slot_size; |
314 | ||
315 | /* | |
316 | * We need to ensure that if the cmpxchg succeeds and discards the | |
317 | * record, the next record will record a full TSC, because it cannot | |
318 | * rely on the last_tsc associated with the discarded record to detect | |
319 | * overflows. The only way to ensure this is to set the last_tsc to 0 | |
320 | * (assuming no 64-bit TSC overflow), which forces to write a 64-bit | |
321 | * timestamp in the next record. | |
322 | * | |
323 | * Note: if discard fails, we must leave the TSC in the record header. | |
324 | * It is needed to keep track of TSC overflows for the following | |
325 | * records. | |
326 | */ | |
327 | save_last_tsc(config, buf, 0ULL); | |
328 | ||
b5a3dfa5 | 329 | if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset) |
852c2936 MD |
330 | != end_offset)) |
331 | return -EPERM; | |
332 | else | |
333 | return 0; | |
334 | } | |
335 | ||
336 | static inline | |
4cfec15c | 337 | void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config, |
852c2936 MD |
338 | struct channel *chan) |
339 | { | |
9f3fdbc6 | 340 | uatomic_inc(&chan->record_disabled); |
852c2936 MD |
341 | } |
342 | ||
343 | static inline | |
4cfec15c | 344 | void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config, |
852c2936 MD |
345 | struct channel *chan) |
346 | { | |
9f3fdbc6 | 347 | uatomic_dec(&chan->record_disabled); |
852c2936 MD |
348 | } |
349 | ||
350 | static inline | |
4cfec15c MD |
351 | void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config, |
352 | struct lttng_ust_lib_ring_buffer *buf) | |
852c2936 | 353 | { |
9f3fdbc6 | 354 | uatomic_inc(&buf->record_disabled); |
852c2936 MD |
355 | } |
356 | ||
357 | static inline | |
4cfec15c MD |
358 | void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config, |
359 | struct lttng_ust_lib_ring_buffer *buf) | |
852c2936 | 360 | { |
9f3fdbc6 | 361 | uatomic_dec(&buf->record_disabled); |
852c2936 MD |
362 | } |
363 | ||
e92f3e28 | 364 | #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */ |