Commit | Line | Data |
---|---|---|
852c2936 | 1 | /* |
c0c0989a | 2 | * SPDX-License-Identifier: LGPL-2.1-only |
852c2936 | 3 | * |
e92f3e28 MD |
4 | * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
5 | * | |
e92f3e28 MD |
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. | |
852c2936 MD |
9 | */ |
10 | ||
c0c0989a MJ |
11 | #ifndef _LTTNG_RING_BUFFER_FRONTEND_API_H |
12 | #define _LTTNG_RING_BUFFER_FRONTEND_API_H | |
13 | ||
b4051ad8 FD |
14 | #include <stddef.h> |
15 | ||
9f3fdbc6 | 16 | #include <urcu/compiler.h> |
852c2936 | 17 | |
b4051ad8 FD |
18 | #include "frontend.h" |
19 | ||
852c2936 | 20 | /** |
7489fcb4 | 21 | * lib_ring_buffer_nesting_inc - Ring buffer recursive use protection. |
852c2936 | 22 | * |
7489fcb4 MD |
23 | * The rint buffer buffer nesting count is a safety net to ensure tracer |
24 | * client code will never trigger an endless recursion. | |
25 | * Returns 0 on success, -EPERM on failure (nesting count too high). | |
852c2936 MD |
26 | * |
27 | * asm volatile and "memory" clobber prevent the compiler from moving | |
28 | * instructions out of the ring buffer nesting count. This is required to ensure | |
29 | * that probe side-effects which can cause recursion (e.g. unforeseen traps, | |
30 | * divisions by 0, ...) are triggered within the incremented nesting count | |
31 | * section. | |
32 | */ | |
33 | static inline | |
7489fcb4 | 34 | int lib_ring_buffer_nesting_inc(const struct lttng_ust_lib_ring_buffer_config *config) |
852c2936 | 35 | { |
7489fcb4 | 36 | int nesting; |
852c2936 | 37 | |
8c90a710 | 38 | nesting = ++URCU_TLS(lib_ring_buffer_nesting); |
9f3fdbc6 | 39 | cmm_barrier(); |
b5a3dfa5 | 40 | if (caa_unlikely(nesting > 4)) { |
852c2936 | 41 | WARN_ON_ONCE(1); |
8c90a710 | 42 | URCU_TLS(lib_ring_buffer_nesting)--; |
852c2936 | 43 | return -EPERM; |
7489fcb4 MD |
44 | } |
45 | return 0; | |
852c2936 MD |
46 | } |
47 | ||
852c2936 | 48 | static inline |
7489fcb4 | 49 | void lib_ring_buffer_nesting_dec(const struct lttng_ust_lib_ring_buffer_config *config) |
852c2936 | 50 | { |
9f3fdbc6 | 51 | cmm_barrier(); |
8c90a710 | 52 | URCU_TLS(lib_ring_buffer_nesting)--; /* TLS */ |
852c2936 MD |
53 | } |
54 | ||
55 | /* | |
56 | * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not | |
57 | * part of the API per se. | |
58 | * | |
59 | * returns 0 if reserve ok, or 1 if the slow path must be taken. | |
60 | */ | |
61 | static inline | |
4cfec15c MD |
62 | int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
63 | struct lttng_ust_lib_ring_buffer_ctx *ctx, | |
e56bb47c | 64 | void *client_ctx, |
852c2936 MD |
65 | unsigned long *o_begin, unsigned long *o_end, |
66 | unsigned long *o_old, size_t *before_hdr_pad) | |
67 | { | |
5198080d | 68 | struct lttng_ust_lib_ring_buffer_channel *chan = ctx->chan; |
4cfec15c | 69 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
70 | *o_begin = v_read(config, &buf->offset); |
71 | *o_old = *o_begin; | |
72 | ||
73 | ctx->tsc = lib_ring_buffer_clock_read(chan); | |
74 | if ((int64_t) ctx->tsc == -EIO) | |
75 | return 1; | |
76 | ||
77 | /* | |
78 | * Prefetch cacheline for read because we have to read the previous | |
79 | * commit counter to increment it and commit seq value to compare it to | |
80 | * the commit counter. | |
81 | */ | |
9f3fdbc6 | 82 | //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]); |
852c2936 MD |
83 | |
84 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
85 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
86 | ||
b5a3dfa5 | 87 | if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0)) |
852c2936 MD |
88 | return 1; |
89 | ||
90 | ctx->slot_size = record_header_size(config, chan, *o_begin, | |
e56bb47c | 91 | before_hdr_pad, ctx, client_ctx); |
852c2936 MD |
92 | ctx->slot_size += |
93 | lib_ring_buffer_align(*o_begin + ctx->slot_size, | |
94 | ctx->largest_align) + ctx->data_size; | |
b5a3dfa5 | 95 | if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size) |
852c2936 MD |
96 | > chan->backend.subbuf_size)) |
97 | return 1; | |
98 | ||
99 | /* | |
100 | * Record fits in the current buffer and we are not on a switch | |
101 | * boundary. It's safe to write. | |
102 | */ | |
103 | *o_end = *o_begin + ctx->slot_size; | |
1ad21f70 MD |
104 | |
105 | if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0)) | |
106 | /* | |
107 | * The offset_end will fall at the very beginning of the next | |
108 | * subbuffer. | |
109 | */ | |
110 | return 1; | |
111 | ||
852c2936 MD |
112 | return 0; |
113 | } | |
114 | ||
115 | /** | |
116 | * lib_ring_buffer_reserve - Reserve space in a ring buffer. | |
117 | * @config: ring buffer instance configuration. | |
118 | * @ctx: ring buffer context. (input and output) Must be already initialized. | |
119 | * | |
120 | * Atomic wait-free slot reservation. The reserved space starts at the context | |
121 | * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc". | |
122 | * | |
123 | * Return : | |
124 | * 0 on success. | |
125 | * -EAGAIN if channel is disabled. | |
126 | * -ENOSPC if event size is too large for packet. | |
127 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
128 | * -EIO if data cannot be written into the buffer for any other reason. | |
129 | */ | |
130 | ||
131 | static inline | |
4cfec15c | 132 | int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
e56bb47c MD |
133 | struct lttng_ust_lib_ring_buffer_ctx *ctx, |
134 | void *client_ctx) | |
852c2936 | 135 | { |
5198080d | 136 | struct lttng_ust_lib_ring_buffer_channel *chan = ctx->chan; |
aab8b172 | 137 | struct lttng_ust_shm_handle *handle = ctx->chan->handle; |
4cfec15c | 138 | struct lttng_ust_lib_ring_buffer *buf; |
852c2936 MD |
139 | unsigned long o_begin, o_end, o_old; |
140 | size_t before_hdr_pad = 0; | |
141 | ||
f52a5702 | 142 | if (caa_unlikely(uatomic_read(&chan->record_disabled))) |
852c2936 MD |
143 | return -EAGAIN; |
144 | ||
7489fcb4 MD |
145 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { |
146 | ctx->reserve_cpu = lttng_ust_get_cpu(); | |
147 | buf = shmp(handle, chan->backend.buf[ctx->reserve_cpu].shmp); | |
148 | } else { | |
1d498196 | 149 | buf = shmp(handle, chan->backend.buf[0].shmp); |
7489fcb4 | 150 | } |
15500a1b MD |
151 | if (caa_unlikely(!buf)) |
152 | return -EIO; | |
f52a5702 | 153 | if (caa_unlikely(uatomic_read(&buf->record_disabled))) |
852c2936 MD |
154 | return -EAGAIN; |
155 | ctx->buf = buf; | |
156 | ||
157 | /* | |
158 | * Perform retryable operations. | |
159 | */ | |
e56bb47c | 160 | if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, client_ctx, &o_begin, |
852c2936 MD |
161 | &o_end, &o_old, &before_hdr_pad))) |
162 | goto slow_path; | |
163 | ||
b5a3dfa5 | 164 | if (caa_unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end) |
852c2936 MD |
165 | != o_old)) |
166 | goto slow_path; | |
167 | ||
168 | /* | |
169 | * Atomically update last_tsc. This update races against concurrent | |
170 | * atomic updates, but the race will always cause supplementary full TSC | |
171 | * record headers, never the opposite (missing a full TSC record header | |
172 | * when it would be needed). | |
173 | */ | |
174 | save_last_tsc(config, ctx->buf, ctx->tsc); | |
175 | ||
176 | /* | |
177 | * Push the reader if necessary | |
178 | */ | |
179 | lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1); | |
180 | ||
181 | /* | |
182 | * Clear noref flag for this subbuffer. | |
183 | */ | |
184 | lib_ring_buffer_clear_noref(config, &ctx->buf->backend, | |
1d498196 | 185 | subbuf_index(o_end - 1, chan), handle); |
852c2936 MD |
186 | |
187 | ctx->pre_offset = o_begin; | |
188 | ctx->buf_offset = o_begin + before_hdr_pad; | |
189 | return 0; | |
190 | slow_path: | |
e56bb47c | 191 | return lib_ring_buffer_reserve_slow(ctx, client_ctx); |
852c2936 MD |
192 | } |
193 | ||
194 | /** | |
195 | * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer. | |
196 | * @config: ring buffer instance configuration. | |
197 | * @buf: buffer | |
198 | * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH) | |
199 | * | |
200 | * This operation is completely reentrant : can be called while tracing is | |
201 | * active with absolutely no lock held. | |
202 | * | |
203 | * Note, however, that as a v_cmpxchg is used for some atomic operations and | |
204 | * requires to be executed locally for per-CPU buffers, this function must be | |
205 | * called from the CPU which owns the buffer for a ACTIVE flush, with preemption | |
206 | * disabled, for RING_BUFFER_SYNC_PER_CPU configuration. | |
207 | */ | |
208 | static inline | |
4cfec15c MD |
209 | void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config, |
210 | struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, | |
38fae1d3 | 211 | struct lttng_ust_shm_handle *handle) |
852c2936 | 212 | { |
1d498196 | 213 | lib_ring_buffer_switch_slow(buf, mode, handle); |
852c2936 MD |
214 | } |
215 | ||
216 | /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */ | |
217 | ||
218 | /** | |
219 | * lib_ring_buffer_commit - Commit an record. | |
220 | * @config: ring buffer instance configuration. | |
221 | * @ctx: ring buffer context. (input arguments only) | |
222 | * | |
223 | * Atomic unordered slot commit. Increments the commit count in the | |
224 | * specified sub-buffer, and delivers it if necessary. | |
225 | */ | |
226 | static inline | |
4cfec15c MD |
227 | void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config, |
228 | const struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 | 229 | { |
5198080d | 230 | struct lttng_ust_lib_ring_buffer_channel *chan = ctx->chan; |
aab8b172 | 231 | struct lttng_ust_shm_handle *handle = ctx->chan->handle; |
4cfec15c | 232 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
233 | unsigned long offset_end = ctx->buf_offset; |
234 | unsigned long endidx = subbuf_index(offset_end - 1, chan); | |
235 | unsigned long commit_count; | |
d2fe4771 MD |
236 | struct commit_counters_hot *cc_hot = shmp_index(handle, |
237 | buf->commit_hot, endidx); | |
852c2936 | 238 | |
15500a1b MD |
239 | if (caa_unlikely(!cc_hot)) |
240 | return; | |
241 | ||
852c2936 MD |
242 | /* |
243 | * Must count record before incrementing the commit count. | |
244 | */ | |
15500a1b | 245 | subbuffer_count_record(config, ctx, &buf->backend, endidx, handle); |
852c2936 MD |
246 | |
247 | /* | |
248 | * Order all writes to buffer before the commit count update that will | |
249 | * determine that the subbuffer is full. | |
250 | */ | |
9f3fdbc6 | 251 | cmm_smp_wmb(); |
852c2936 | 252 | |
d2fe4771 | 253 | v_add(config, ctx->slot_size, &cc_hot->cc); |
852c2936 MD |
254 | |
255 | /* | |
256 | * commit count read can race with concurrent OOO commit count updates. | |
257 | * This is only needed for lib_ring_buffer_check_deliver (for | |
258 | * non-polling delivery only) and for | |
259 | * lib_ring_buffer_write_commit_counter. The race can only cause the | |
260 | * counter to be read with the same value more than once, which could | |
261 | * cause : | |
262 | * - Multiple delivery for the same sub-buffer (which is handled | |
263 | * gracefully by the reader code) if the value is for a full | |
264 | * sub-buffer. It's important that we can never miss a sub-buffer | |
265 | * delivery. Re-reading the value after the v_add ensures this. | |
266 | * - Reading a commit_count with a higher value that what was actually | |
267 | * added to it for the lib_ring_buffer_write_commit_counter call | |
268 | * (again caused by a concurrent committer). It does not matter, | |
269 | * because this function is interested in the fact that the commit | |
270 | * count reaches back the reserve offset for a specific sub-buffer, | |
271 | * which is completely independent of the order. | |
272 | */ | |
d2fe4771 | 273 | commit_count = v_read(config, &cc_hot->cc); |
852c2936 MD |
274 | |
275 | lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1, | |
1b7b0501 | 276 | commit_count, endidx, handle, ctx->tsc); |
852c2936 MD |
277 | /* |
278 | * Update used size at each commit. It's needed only for extracting | |
279 | * ring_buffer buffers from vmcore, after crash. | |
280 | */ | |
d2fe4771 MD |
281 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
282 | offset_end, commit_count, handle, cc_hot); | |
852c2936 MD |
283 | } |
284 | ||
285 | /** | |
286 | * lib_ring_buffer_try_discard_reserve - Try discarding a record. | |
287 | * @config: ring buffer instance configuration. | |
288 | * @ctx: ring buffer context. (input arguments only) | |
289 | * | |
290 | * Only succeeds if no other record has been written after the record to | |
291 | * discard. If discard fails, the record must be committed to the buffer. | |
292 | * | |
293 | * Returns 0 upon success, -EPERM if the record cannot be discarded. | |
294 | */ | |
295 | static inline | |
4cfec15c MD |
296 | int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
297 | const struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 | 298 | { |
4cfec15c | 299 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
300 | unsigned long end_offset = ctx->pre_offset + ctx->slot_size; |
301 | ||
302 | /* | |
303 | * We need to ensure that if the cmpxchg succeeds and discards the | |
304 | * record, the next record will record a full TSC, because it cannot | |
305 | * rely on the last_tsc associated with the discarded record to detect | |
306 | * overflows. The only way to ensure this is to set the last_tsc to 0 | |
307 | * (assuming no 64-bit TSC overflow), which forces to write a 64-bit | |
308 | * timestamp in the next record. | |
309 | * | |
310 | * Note: if discard fails, we must leave the TSC in the record header. | |
311 | * It is needed to keep track of TSC overflows for the following | |
312 | * records. | |
313 | */ | |
314 | save_last_tsc(config, buf, 0ULL); | |
315 | ||
b5a3dfa5 | 316 | if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset) |
852c2936 MD |
317 | != end_offset)) |
318 | return -EPERM; | |
319 | else | |
320 | return 0; | |
321 | } | |
322 | ||
323 | static inline | |
4cfec15c | 324 | void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config, |
5198080d | 325 | struct lttng_ust_lib_ring_buffer_channel *chan) |
852c2936 | 326 | { |
9f3fdbc6 | 327 | uatomic_inc(&chan->record_disabled); |
852c2936 MD |
328 | } |
329 | ||
330 | static inline | |
4cfec15c | 331 | void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config, |
5198080d | 332 | struct lttng_ust_lib_ring_buffer_channel *chan) |
852c2936 | 333 | { |
9f3fdbc6 | 334 | uatomic_dec(&chan->record_disabled); |
852c2936 MD |
335 | } |
336 | ||
337 | static inline | |
4cfec15c MD |
338 | void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config, |
339 | struct lttng_ust_lib_ring_buffer *buf) | |
852c2936 | 340 | { |
9f3fdbc6 | 341 | uatomic_inc(&buf->record_disabled); |
852c2936 MD |
342 | } |
343 | ||
344 | static inline | |
4cfec15c MD |
345 | void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config, |
346 | struct lttng_ust_lib_ring_buffer *buf) | |
852c2936 | 347 | { |
9f3fdbc6 | 348 | uatomic_dec(&buf->record_disabled); |
852c2936 MD |
349 | } |
350 | ||
e92f3e28 | 351 | #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */ |