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 | * | |
9f3fdbc6 MD |
40 | * Grabs RCU read-side lock and keeps a ring buffer nesting count as |
41 | * supplementary safety net to ensure tracer client code will never | |
42 | * trigger an endless recursion. Returns the processor ID on success, | |
43 | * -EPERM on failure (nesting count 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 | ||
9f3fdbc6 | 56 | rcu_read_lock(); |
2b2d6ff7 | 57 | cpu = lttng_ust_get_cpu(); |
8c90a710 | 58 | nesting = ++URCU_TLS(lib_ring_buffer_nesting); |
9f3fdbc6 | 59 | cmm_barrier(); |
852c2936 | 60 | |
b5a3dfa5 | 61 | if (caa_unlikely(nesting > 4)) { |
852c2936 | 62 | WARN_ON_ONCE(1); |
8c90a710 | 63 | URCU_TLS(lib_ring_buffer_nesting)--; |
9f3fdbc6 | 64 | rcu_read_unlock(); |
852c2936 MD |
65 | return -EPERM; |
66 | } else | |
67 | return cpu; | |
68 | } | |
69 | ||
70 | /** | |
71 | * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit. | |
72 | */ | |
73 | static inline | |
4cfec15c | 74 | void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config *config) |
852c2936 | 75 | { |
9f3fdbc6 | 76 | cmm_barrier(); |
8c90a710 | 77 | URCU_TLS(lib_ring_buffer_nesting)--; /* TLS */ |
9f3fdbc6 | 78 | rcu_read_unlock(); |
852c2936 MD |
79 | } |
80 | ||
81 | /* | |
82 | * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not | |
83 | * part of the API per se. | |
84 | * | |
85 | * returns 0 if reserve ok, or 1 if the slow path must be taken. | |
86 | */ | |
87 | static inline | |
4cfec15c MD |
88 | int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
89 | struct lttng_ust_lib_ring_buffer_ctx *ctx, | |
852c2936 MD |
90 | unsigned long *o_begin, unsigned long *o_end, |
91 | unsigned long *o_old, size_t *before_hdr_pad) | |
92 | { | |
93 | struct channel *chan = ctx->chan; | |
4cfec15c | 94 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
95 | *o_begin = v_read(config, &buf->offset); |
96 | *o_old = *o_begin; | |
97 | ||
98 | ctx->tsc = lib_ring_buffer_clock_read(chan); | |
99 | if ((int64_t) ctx->tsc == -EIO) | |
100 | return 1; | |
101 | ||
102 | /* | |
103 | * Prefetch cacheline for read because we have to read the previous | |
104 | * commit counter to increment it and commit seq value to compare it to | |
105 | * the commit counter. | |
106 | */ | |
9f3fdbc6 | 107 | //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]); |
852c2936 MD |
108 | |
109 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
110 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
111 | ||
b5a3dfa5 | 112 | if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0)) |
852c2936 MD |
113 | return 1; |
114 | ||
115 | ctx->slot_size = record_header_size(config, chan, *o_begin, | |
116 | before_hdr_pad, ctx); | |
117 | ctx->slot_size += | |
118 | lib_ring_buffer_align(*o_begin + ctx->slot_size, | |
119 | ctx->largest_align) + ctx->data_size; | |
b5a3dfa5 | 120 | if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size) |
852c2936 MD |
121 | > chan->backend.subbuf_size)) |
122 | return 1; | |
123 | ||
124 | /* | |
125 | * Record fits in the current buffer and we are not on a switch | |
126 | * boundary. It's safe to write. | |
127 | */ | |
128 | *o_end = *o_begin + ctx->slot_size; | |
1ad21f70 MD |
129 | |
130 | if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0)) | |
131 | /* | |
132 | * The offset_end will fall at the very beginning of the next | |
133 | * subbuffer. | |
134 | */ | |
135 | return 1; | |
136 | ||
852c2936 MD |
137 | return 0; |
138 | } | |
139 | ||
140 | /** | |
141 | * lib_ring_buffer_reserve - Reserve space in a ring buffer. | |
142 | * @config: ring buffer instance configuration. | |
143 | * @ctx: ring buffer context. (input and output) Must be already initialized. | |
144 | * | |
145 | * Atomic wait-free slot reservation. The reserved space starts at the context | |
146 | * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc". | |
147 | * | |
148 | * Return : | |
149 | * 0 on success. | |
150 | * -EAGAIN if channel is disabled. | |
151 | * -ENOSPC if event size is too large for packet. | |
152 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
153 | * -EIO if data cannot be written into the buffer for any other reason. | |
154 | */ | |
155 | ||
156 | static inline | |
4cfec15c MD |
157 | int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
158 | struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 MD |
159 | { |
160 | struct channel *chan = ctx->chan; | |
38fae1d3 | 161 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c | 162 | struct lttng_ust_lib_ring_buffer *buf; |
852c2936 MD |
163 | unsigned long o_begin, o_end, o_old; |
164 | size_t before_hdr_pad = 0; | |
165 | ||
9f3fdbc6 | 166 | if (uatomic_read(&chan->record_disabled)) |
852c2936 MD |
167 | return -EAGAIN; |
168 | ||
169 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 170 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 171 | else |
1d498196 | 172 | buf = shmp(handle, chan->backend.buf[0].shmp); |
9f3fdbc6 | 173 | if (uatomic_read(&buf->record_disabled)) |
852c2936 MD |
174 | return -EAGAIN; |
175 | ctx->buf = buf; | |
176 | ||
177 | /* | |
178 | * Perform retryable operations. | |
179 | */ | |
b5a3dfa5 | 180 | if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, &o_begin, |
852c2936 MD |
181 | &o_end, &o_old, &before_hdr_pad))) |
182 | goto slow_path; | |
183 | ||
b5a3dfa5 | 184 | if (caa_unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end) |
852c2936 MD |
185 | != o_old)) |
186 | goto slow_path; | |
187 | ||
188 | /* | |
189 | * Atomically update last_tsc. This update races against concurrent | |
190 | * atomic updates, but the race will always cause supplementary full TSC | |
191 | * record headers, never the opposite (missing a full TSC record header | |
192 | * when it would be needed). | |
193 | */ | |
194 | save_last_tsc(config, ctx->buf, ctx->tsc); | |
195 | ||
196 | /* | |
197 | * Push the reader if necessary | |
198 | */ | |
199 | lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1); | |
200 | ||
201 | /* | |
202 | * Clear noref flag for this subbuffer. | |
203 | */ | |
204 | lib_ring_buffer_clear_noref(config, &ctx->buf->backend, | |
1d498196 | 205 | subbuf_index(o_end - 1, chan), handle); |
852c2936 MD |
206 | |
207 | ctx->pre_offset = o_begin; | |
208 | ctx->buf_offset = o_begin + before_hdr_pad; | |
209 | return 0; | |
210 | slow_path: | |
211 | return lib_ring_buffer_reserve_slow(ctx); | |
212 | } | |
213 | ||
214 | /** | |
215 | * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer. | |
216 | * @config: ring buffer instance configuration. | |
217 | * @buf: buffer | |
218 | * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH) | |
219 | * | |
220 | * This operation is completely reentrant : can be called while tracing is | |
221 | * active with absolutely no lock held. | |
222 | * | |
223 | * Note, however, that as a v_cmpxchg is used for some atomic operations and | |
224 | * requires to be executed locally for per-CPU buffers, this function must be | |
225 | * called from the CPU which owns the buffer for a ACTIVE flush, with preemption | |
226 | * disabled, for RING_BUFFER_SYNC_PER_CPU configuration. | |
227 | */ | |
228 | static inline | |
4cfec15c MD |
229 | void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config, |
230 | struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, | |
38fae1d3 | 231 | struct lttng_ust_shm_handle *handle) |
852c2936 | 232 | { |
1d498196 | 233 | lib_ring_buffer_switch_slow(buf, mode, handle); |
852c2936 MD |
234 | } |
235 | ||
236 | /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */ | |
237 | ||
238 | /** | |
239 | * lib_ring_buffer_commit - Commit an record. | |
240 | * @config: ring buffer instance configuration. | |
241 | * @ctx: ring buffer context. (input arguments only) | |
242 | * | |
243 | * Atomic unordered slot commit. Increments the commit count in the | |
244 | * specified sub-buffer, and delivers it if necessary. | |
245 | */ | |
246 | static inline | |
4cfec15c MD |
247 | void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config, |
248 | const struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 MD |
249 | { |
250 | struct channel *chan = ctx->chan; | |
38fae1d3 | 251 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c | 252 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
253 | unsigned long offset_end = ctx->buf_offset; |
254 | unsigned long endidx = subbuf_index(offset_end - 1, chan); | |
255 | unsigned long commit_count; | |
256 | ||
257 | /* | |
258 | * Must count record before incrementing the commit count. | |
259 | */ | |
1d498196 | 260 | subbuffer_count_record(config, &buf->backend, endidx, handle); |
852c2936 MD |
261 | |
262 | /* | |
263 | * Order all writes to buffer before the commit count update that will | |
264 | * determine that the subbuffer is full. | |
265 | */ | |
9f3fdbc6 | 266 | cmm_smp_wmb(); |
852c2936 | 267 | |
4746ae29 | 268 | v_add(config, ctx->slot_size, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
852c2936 MD |
269 | |
270 | /* | |
271 | * commit count read can race with concurrent OOO commit count updates. | |
272 | * This is only needed for lib_ring_buffer_check_deliver (for | |
273 | * non-polling delivery only) and for | |
274 | * lib_ring_buffer_write_commit_counter. The race can only cause the | |
275 | * counter to be read with the same value more than once, which could | |
276 | * cause : | |
277 | * - Multiple delivery for the same sub-buffer (which is handled | |
278 | * gracefully by the reader code) if the value is for a full | |
279 | * sub-buffer. It's important that we can never miss a sub-buffer | |
280 | * delivery. Re-reading the value after the v_add ensures this. | |
281 | * - Reading a commit_count with a higher value that what was actually | |
282 | * added to it for the lib_ring_buffer_write_commit_counter call | |
283 | * (again caused by a concurrent committer). It does not matter, | |
284 | * because this function is interested in the fact that the commit | |
285 | * count reaches back the reserve offset for a specific sub-buffer, | |
286 | * which is completely independent of the order. | |
287 | */ | |
4746ae29 | 288 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
852c2936 MD |
289 | |
290 | lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1, | |
1b7b0501 | 291 | commit_count, endidx, handle, ctx->tsc); |
852c2936 MD |
292 | /* |
293 | * Update used size at each commit. It's needed only for extracting | |
294 | * ring_buffer buffers from vmcore, after crash. | |
295 | */ | |
296 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, | |
80249235 | 297 | offset_end, commit_count, handle); |
852c2936 MD |
298 | } |
299 | ||
300 | /** | |
301 | * lib_ring_buffer_try_discard_reserve - Try discarding a record. | |
302 | * @config: ring buffer instance configuration. | |
303 | * @ctx: ring buffer context. (input arguments only) | |
304 | * | |
305 | * Only succeeds if no other record has been written after the record to | |
306 | * discard. If discard fails, the record must be committed to the buffer. | |
307 | * | |
308 | * Returns 0 upon success, -EPERM if the record cannot be discarded. | |
309 | */ | |
310 | static inline | |
4cfec15c MD |
311 | int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
312 | const struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 | 313 | { |
4cfec15c | 314 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
315 | unsigned long end_offset = ctx->pre_offset + ctx->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 | ||
b5a3dfa5 | 331 | if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset) |
852c2936 MD |
332 | != end_offset)) |
333 | return -EPERM; | |
334 | else | |
335 | return 0; | |
336 | } | |
337 | ||
338 | static inline | |
4cfec15c | 339 | void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config, |
852c2936 MD |
340 | struct channel *chan) |
341 | { | |
9f3fdbc6 | 342 | uatomic_inc(&chan->record_disabled); |
852c2936 MD |
343 | } |
344 | ||
345 | static inline | |
4cfec15c | 346 | void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config, |
852c2936 MD |
347 | struct channel *chan) |
348 | { | |
9f3fdbc6 | 349 | uatomic_dec(&chan->record_disabled); |
852c2936 MD |
350 | } |
351 | ||
352 | static inline | |
4cfec15c MD |
353 | void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config, |
354 | struct lttng_ust_lib_ring_buffer *buf) | |
852c2936 | 355 | { |
9f3fdbc6 | 356 | uatomic_inc(&buf->record_disabled); |
852c2936 MD |
357 | } |
358 | ||
359 | static inline | |
4cfec15c MD |
360 | void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config, |
361 | struct lttng_ust_lib_ring_buffer *buf) | |
852c2936 | 362 | { |
9f3fdbc6 | 363 | uatomic_dec(&buf->record_disabled); |
852c2936 MD |
364 | } |
365 | ||
e92f3e28 | 366 | #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */ |