Commit | Line | Data |
---|---|---|
e92f3e28 MD |
1 | #ifndef _LTTNG_RING_BUFFER_FRONTEND_INTERNAL_H |
2 | #define _LTTNG_RING_BUFFER_FRONTEND_INTERNAL_H | |
852c2936 MD |
3 | |
4 | /* | |
e92f3e28 | 5 | * libringbuffer/frontend_internal.h |
852c2936 MD |
6 | * |
7 | * Ring Buffer Library Synchronization Header (internal helpers). | |
8 | * | |
e92f3e28 MD |
9 | * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
10 | * | |
11 | * This library is free software; you can redistribute it and/or | |
12 | * modify it under the terms of the GNU Lesser General Public | |
13 | * License as published by the Free Software Foundation; only | |
14 | * version 2.1 of the License. | |
15 | * | |
16 | * This library is distributed in the hope that it will be useful, | |
17 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
18 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
19 | * Lesser General Public License for more details. | |
20 | * | |
21 | * You should have received a copy of the GNU Lesser General Public | |
22 | * License along with this library; if not, write to the Free Software | |
23 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
24 | * | |
25 | * | |
852c2936 MD |
26 | * Author: |
27 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
28 | * | |
29 | * See ring_buffer_frontend.c for more information on wait-free algorithms. | |
30 | * | |
31 | * Dual LGPL v2.1/GPL v2 license. | |
32 | */ | |
33 | ||
14641deb | 34 | #include <urcu/compiler.h> |
8c90a710 | 35 | #include <urcu/tls-compat.h> |
2c44f5b9 | 36 | #include <signal.h> |
fb31eb73 | 37 | #include <stdint.h> |
2c44f5b9 | 38 | #include <pthread.h> |
14641deb | 39 | |
4318ae1b | 40 | #include <lttng/ringbuffer-config.h> |
4931a13e MD |
41 | #include "backend_types.h" |
42 | #include "frontend_types.h" | |
a6352fd4 | 43 | #include "shm.h" |
852c2936 MD |
44 | |
45 | /* Buffer offset macros */ | |
46 | ||
47 | /* buf_trunc mask selects only the buffer number. */ | |
48 | static inline | |
49 | unsigned long buf_trunc(unsigned long offset, struct channel *chan) | |
50 | { | |
51 | return offset & ~(chan->backend.buf_size - 1); | |
52 | ||
53 | } | |
54 | ||
55 | /* Select the buffer number value (counter). */ | |
56 | static inline | |
57 | unsigned long buf_trunc_val(unsigned long offset, struct channel *chan) | |
58 | { | |
59 | return buf_trunc(offset, chan) >> chan->backend.buf_size_order; | |
60 | } | |
61 | ||
62 | /* buf_offset mask selects only the offset within the current buffer. */ | |
63 | static inline | |
64 | unsigned long buf_offset(unsigned long offset, struct channel *chan) | |
65 | { | |
66 | return offset & (chan->backend.buf_size - 1); | |
67 | } | |
68 | ||
69 | /* subbuf_offset mask selects the offset within the current subbuffer. */ | |
70 | static inline | |
71 | unsigned long subbuf_offset(unsigned long offset, struct channel *chan) | |
72 | { | |
73 | return offset & (chan->backend.subbuf_size - 1); | |
74 | } | |
75 | ||
76 | /* subbuf_trunc mask selects the subbuffer number. */ | |
77 | static inline | |
78 | unsigned long subbuf_trunc(unsigned long offset, struct channel *chan) | |
79 | { | |
80 | return offset & ~(chan->backend.subbuf_size - 1); | |
81 | } | |
82 | ||
83 | /* subbuf_align aligns the offset to the next subbuffer. */ | |
84 | static inline | |
85 | unsigned long subbuf_align(unsigned long offset, struct channel *chan) | |
86 | { | |
87 | return (offset + chan->backend.subbuf_size) | |
88 | & ~(chan->backend.subbuf_size - 1); | |
89 | } | |
90 | ||
91 | /* subbuf_index returns the index of the current subbuffer within the buffer. */ | |
92 | static inline | |
93 | unsigned long subbuf_index(unsigned long offset, struct channel *chan) | |
94 | { | |
95 | return buf_offset(offset, chan) >> chan->backend.subbuf_size_order; | |
96 | } | |
97 | ||
98 | /* | |
99 | * Last TSC comparison functions. Check if the current TSC overflows tsc_bits | |
100 | * bits from the last TSC read. When overflows are detected, the full 64-bit | |
101 | * timestamp counter should be written in the record header. Reads and writes | |
102 | * last_tsc atomically. | |
103 | */ | |
104 | ||
14641deb | 105 | #if (CAA_BITS_PER_LONG == 32) |
852c2936 | 106 | static inline |
4cfec15c | 107 | void save_last_tsc(const struct lttng_ust_lib_ring_buffer_config *config, |
2fed87ae | 108 | struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc) |
852c2936 MD |
109 | { |
110 | if (config->tsc_bits == 0 || config->tsc_bits == 64) | |
111 | return; | |
112 | ||
113 | /* | |
114 | * Ensure the compiler performs this update in a single instruction. | |
115 | */ | |
116 | v_set(config, &buf->last_tsc, (unsigned long)(tsc >> config->tsc_bits)); | |
117 | } | |
118 | ||
119 | static inline | |
4cfec15c | 120 | int last_tsc_overflow(const struct lttng_ust_lib_ring_buffer_config *config, |
2fed87ae | 121 | struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc) |
852c2936 MD |
122 | { |
123 | unsigned long tsc_shifted; | |
124 | ||
125 | if (config->tsc_bits == 0 || config->tsc_bits == 64) | |
126 | return 0; | |
127 | ||
128 | tsc_shifted = (unsigned long)(tsc >> config->tsc_bits); | |
b5a3dfa5 | 129 | if (caa_unlikely(tsc_shifted |
852c2936 MD |
130 | - (unsigned long)v_read(config, &buf->last_tsc))) |
131 | return 1; | |
132 | else | |
133 | return 0; | |
134 | } | |
135 | #else | |
136 | static inline | |
4cfec15c | 137 | void save_last_tsc(const struct lttng_ust_lib_ring_buffer_config *config, |
2fed87ae | 138 | struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc) |
852c2936 MD |
139 | { |
140 | if (config->tsc_bits == 0 || config->tsc_bits == 64) | |
141 | return; | |
142 | ||
143 | v_set(config, &buf->last_tsc, (unsigned long)tsc); | |
144 | } | |
145 | ||
146 | static inline | |
4cfec15c | 147 | int last_tsc_overflow(const struct lttng_ust_lib_ring_buffer_config *config, |
2fed87ae | 148 | struct lttng_ust_lib_ring_buffer *buf, uint64_t tsc) |
852c2936 MD |
149 | { |
150 | if (config->tsc_bits == 0 || config->tsc_bits == 64) | |
151 | return 0; | |
152 | ||
b5a3dfa5 | 153 | if (caa_unlikely((tsc - v_read(config, &buf->last_tsc)) |
852c2936 MD |
154 | >> config->tsc_bits)) |
155 | return 1; | |
156 | else | |
157 | return 0; | |
158 | } | |
159 | #endif | |
160 | ||
161 | extern | |
e56bb47c MD |
162 | int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx, |
163 | void *client_ctx); | |
852c2936 MD |
164 | |
165 | extern | |
4cfec15c | 166 | void lib_ring_buffer_switch_slow(struct lttng_ust_lib_ring_buffer *buf, |
1d498196 | 167 | enum switch_mode mode, |
38fae1d3 | 168 | struct lttng_ust_shm_handle *handle); |
852c2936 | 169 | |
b07cd987 MD |
170 | void lib_ring_buffer_check_deliver_slow(const struct lttng_ust_lib_ring_buffer_config *config, |
171 | struct lttng_ust_lib_ring_buffer *buf, | |
172 | struct channel *chan, | |
173 | unsigned long offset, | |
174 | unsigned long commit_count, | |
175 | unsigned long idx, | |
176 | struct lttng_ust_shm_handle *handle, | |
177 | uint64_t tsc); | |
178 | ||
852c2936 MD |
179 | /* Buffer write helpers */ |
180 | ||
181 | static inline | |
4cfec15c | 182 | void lib_ring_buffer_reserve_push_reader(struct lttng_ust_lib_ring_buffer *buf, |
852c2936 MD |
183 | struct channel *chan, |
184 | unsigned long offset) | |
185 | { | |
186 | unsigned long consumed_old, consumed_new; | |
187 | ||
188 | do { | |
14641deb | 189 | consumed_old = uatomic_read(&buf->consumed); |
852c2936 MD |
190 | /* |
191 | * If buffer is in overwrite mode, push the reader consumed | |
192 | * count if the write position has reached it and we are not | |
193 | * at the first iteration (don't push the reader farther than | |
194 | * the writer). This operation can be done concurrently by many | |
195 | * writers in the same buffer, the writer being at the farthest | |
196 | * write position sub-buffer index in the buffer being the one | |
197 | * which will win this loop. | |
198 | */ | |
b5a3dfa5 | 199 | if (caa_unlikely(subbuf_trunc(offset, chan) |
852c2936 MD |
200 | - subbuf_trunc(consumed_old, chan) |
201 | >= chan->backend.buf_size)) | |
202 | consumed_new = subbuf_align(consumed_old, chan); | |
203 | else | |
204 | return; | |
b5a3dfa5 | 205 | } while (caa_unlikely(uatomic_cmpxchg(&buf->consumed, consumed_old, |
852c2936 MD |
206 | consumed_new) != consumed_old)); |
207 | } | |
208 | ||
4c742ffd MD |
209 | /* |
210 | * Move consumed position to the beginning of subbuffer in which the | |
211 | * write offset is. Should only be used on ring buffers that are not | |
212 | * actively being written into, because clear_reader does not take into | |
213 | * account the commit counters when moving the consumed position, which | |
214 | * can make concurrent trace producers or consumers observe consumed | |
215 | * position further than the write offset, which breaks ring buffer | |
216 | * algorithm guarantees. | |
217 | */ | |
beca55a1 MD |
218 | static inline |
219 | void lib_ring_buffer_clear_reader(struct lttng_ust_lib_ring_buffer *buf, | |
220 | struct lttng_ust_shm_handle *handle) | |
221 | { | |
222 | struct channel *chan; | |
223 | const struct lttng_ust_lib_ring_buffer_config *config; | |
224 | unsigned long offset, consumed_old, consumed_new; | |
225 | ||
226 | chan = shmp(handle, buf->backend.chan); | |
227 | if (!chan) | |
228 | return; | |
229 | config = &chan->backend.config; | |
230 | ||
231 | do { | |
232 | offset = v_read(config, &buf->offset); | |
233 | consumed_old = uatomic_read(&buf->consumed); | |
4c742ffd MD |
234 | CHAN_WARN_ON(chan, (long) (subbuf_trunc(offset, chan) |
235 | - subbuf_trunc(consumed_old, chan)) | |
236 | < 0); | |
237 | consumed_new = subbuf_trunc(offset, chan); | |
beca55a1 MD |
238 | } while (caa_unlikely(uatomic_cmpxchg(&buf->consumed, consumed_old, |
239 | consumed_new) != consumed_old)); | |
240 | } | |
241 | ||
852c2936 | 242 | static inline |
4cfec15c MD |
243 | int lib_ring_buffer_pending_data(const struct lttng_ust_lib_ring_buffer_config *config, |
244 | struct lttng_ust_lib_ring_buffer *buf, | |
852c2936 MD |
245 | struct channel *chan) |
246 | { | |
247 | return !!subbuf_offset(v_read(config, &buf->offset), chan); | |
248 | } | |
249 | ||
250 | static inline | |
4cfec15c MD |
251 | unsigned long lib_ring_buffer_get_data_size(const struct lttng_ust_lib_ring_buffer_config *config, |
252 | struct lttng_ust_lib_ring_buffer *buf, | |
1d498196 | 253 | unsigned long idx, |
38fae1d3 | 254 | struct lttng_ust_shm_handle *handle) |
852c2936 | 255 | { |
1d498196 | 256 | return subbuffer_get_data_size(config, &buf->backend, idx, handle); |
852c2936 MD |
257 | } |
258 | ||
259 | /* | |
260 | * Check if all space reservation in a buffer have been committed. This helps | |
261 | * knowing if an execution context is nested (for per-cpu buffers only). | |
262 | * This is a very specific ftrace use-case, so we keep this as "internal" API. | |
263 | */ | |
264 | static inline | |
4cfec15c MD |
265 | int lib_ring_buffer_reserve_committed(const struct lttng_ust_lib_ring_buffer_config *config, |
266 | struct lttng_ust_lib_ring_buffer *buf, | |
1d498196 | 267 | struct channel *chan, |
38fae1d3 | 268 | struct lttng_ust_shm_handle *handle) |
852c2936 MD |
269 | { |
270 | unsigned long offset, idx, commit_count; | |
730be651 | 271 | struct commit_counters_hot *cc_hot; |
852c2936 MD |
272 | |
273 | CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU); | |
274 | CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU); | |
275 | ||
276 | /* | |
277 | * Read offset and commit count in a loop so they are both read | |
278 | * atomically wrt interrupts. By deal with interrupt concurrency by | |
279 | * restarting both reads if the offset has been pushed. Note that given | |
280 | * we only have to deal with interrupt concurrency here, an interrupt | |
281 | * modifying the commit count will also modify "offset", so it is safe | |
282 | * to only check for offset modifications. | |
283 | */ | |
284 | do { | |
285 | offset = v_read(config, &buf->offset); | |
286 | idx = subbuf_index(offset, chan); | |
730be651 MD |
287 | cc_hot = shmp_index(handle, buf->commit_hot, idx); |
288 | if (caa_unlikely(!cc_hot)) | |
289 | return 0; | |
15500a1b | 290 | commit_count = v_read(config, &cc_hot->cc); |
852c2936 MD |
291 | } while (offset != v_read(config, &buf->offset)); |
292 | ||
293 | return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order) | |
294 | - (commit_count & chan->commit_count_mask) == 0); | |
295 | } | |
296 | ||
1b7b0501 MD |
297 | /* |
298 | * Receive end of subbuffer TSC as parameter. It has been read in the | |
299 | * space reservation loop of either reserve or switch, which ensures it | |
300 | * progresses monotonically with event records in the buffer. Therefore, | |
301 | * it ensures that the end timestamp of a subbuffer is <= begin | |
302 | * timestamp of the following subbuffers. | |
303 | */ | |
852c2936 | 304 | static inline |
4cfec15c MD |
305 | void lib_ring_buffer_check_deliver(const struct lttng_ust_lib_ring_buffer_config *config, |
306 | struct lttng_ust_lib_ring_buffer *buf, | |
852c2936 MD |
307 | struct channel *chan, |
308 | unsigned long offset, | |
309 | unsigned long commit_count, | |
1d498196 | 310 | unsigned long idx, |
1b7b0501 MD |
311 | struct lttng_ust_shm_handle *handle, |
312 | uint64_t tsc) | |
852c2936 MD |
313 | { |
314 | unsigned long old_commit_count = commit_count | |
315 | - chan->backend.subbuf_size; | |
852c2936 MD |
316 | |
317 | /* Check if all commits have been done */ | |
b5a3dfa5 | 318 | if (caa_unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order) |
b07cd987 MD |
319 | - (old_commit_count & chan->commit_count_mask) == 0)) |
320 | lib_ring_buffer_check_deliver_slow(config, buf, chan, offset, | |
321 | commit_count, idx, handle, tsc); | |
852c2936 MD |
322 | } |
323 | ||
324 | /* | |
325 | * lib_ring_buffer_write_commit_counter | |
326 | * | |
327 | * For flight recording. must be called after commit. | |
328 | * This function increments the subbuffer's commit_seq counter each time the | |
329 | * commit count reaches back the reserve offset (modulo subbuffer size). It is | |
330 | * useful for crash dump. | |
331 | */ | |
332 | static inline | |
4cfec15c MD |
333 | void lib_ring_buffer_write_commit_counter(const struct lttng_ust_lib_ring_buffer_config *config, |
334 | struct lttng_ust_lib_ring_buffer *buf, | |
852c2936 | 335 | struct channel *chan, |
852c2936 MD |
336 | unsigned long buf_offset, |
337 | unsigned long commit_count, | |
d2fe4771 MD |
338 | struct lttng_ust_shm_handle *handle, |
339 | struct commit_counters_hot *cc_hot) | |
852c2936 | 340 | { |
80249235 | 341 | unsigned long commit_seq_old; |
852c2936 MD |
342 | |
343 | if (config->oops != RING_BUFFER_OOPS_CONSISTENCY) | |
344 | return; | |
345 | ||
852c2936 MD |
346 | /* |
347 | * subbuf_offset includes commit_count_mask. We can simply | |
348 | * compare the offsets within the subbuffer without caring about | |
349 | * buffer full/empty mismatch because offset is never zero here | |
350 | * (subbuffer header and record headers have non-zero length). | |
351 | */ | |
80249235 | 352 | if (caa_unlikely(subbuf_offset(buf_offset - commit_count, chan))) |
852c2936 MD |
353 | return; |
354 | ||
d2fe4771 | 355 | commit_seq_old = v_read(config, &cc_hot->seq); |
16ec84c8 | 356 | if (caa_likely((long) (commit_seq_old - commit_count) < 0)) |
d2fe4771 | 357 | v_set(config, &cc_hot->seq, commit_count); |
852c2936 MD |
358 | } |
359 | ||
4cfec15c | 360 | extern int lib_ring_buffer_create(struct lttng_ust_lib_ring_buffer *buf, |
a6352fd4 | 361 | struct channel_backend *chanb, int cpu, |
38fae1d3 | 362 | struct lttng_ust_shm_handle *handle, |
1d498196 | 363 | struct shm_object *shmobj); |
4cfec15c | 364 | extern void lib_ring_buffer_free(struct lttng_ust_lib_ring_buffer *buf, |
38fae1d3 | 365 | struct lttng_ust_shm_handle *handle); |
852c2936 MD |
366 | |
367 | /* Keep track of trap nesting inside ring buffer code */ | |
8c90a710 | 368 | extern DECLARE_URCU_TLS(unsigned int, lib_ring_buffer_nesting); |
852c2936 | 369 | |
e92f3e28 | 370 | #endif /* _LTTNG_RING_BUFFER_FRONTEND_INTERNAL_H */ |