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 MD |
20 | /** |
21 | * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit. | |
22 | * | |
36cb884c MD |
23 | * Keeps a ring buffer nesting count as supplementary safety net to |
24 | * ensure tracer client code will never trigger an endless recursion. | |
25 | * Returns the processor ID on success, -EPERM on failure (nesting count | |
26 | * too high). | |
852c2936 MD |
27 | * |
28 | * asm volatile and "memory" clobber prevent the compiler from moving | |
29 | * instructions out of the ring buffer nesting count. This is required to ensure | |
30 | * that probe side-effects which can cause recursion (e.g. unforeseen traps, | |
31 | * divisions by 0, ...) are triggered within the incremented nesting count | |
32 | * section. | |
33 | */ | |
34 | static inline | |
4cfec15c | 35 | int lib_ring_buffer_get_cpu(const struct lttng_ust_lib_ring_buffer_config *config) |
852c2936 MD |
36 | { |
37 | int cpu, nesting; | |
38 | ||
2b2d6ff7 | 39 | cpu = lttng_ust_get_cpu(); |
8c90a710 | 40 | nesting = ++URCU_TLS(lib_ring_buffer_nesting); |
9f3fdbc6 | 41 | cmm_barrier(); |
852c2936 | 42 | |
b5a3dfa5 | 43 | if (caa_unlikely(nesting > 4)) { |
852c2936 | 44 | WARN_ON_ONCE(1); |
8c90a710 | 45 | URCU_TLS(lib_ring_buffer_nesting)--; |
852c2936 MD |
46 | return -EPERM; |
47 | } else | |
48 | return cpu; | |
49 | } | |
50 | ||
51 | /** | |
52 | * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit. | |
53 | */ | |
54 | static inline | |
4cfec15c | 55 | void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config *config) |
852c2936 | 56 | { |
9f3fdbc6 | 57 | cmm_barrier(); |
8c90a710 | 58 | URCU_TLS(lib_ring_buffer_nesting)--; /* TLS */ |
852c2936 MD |
59 | } |
60 | ||
61 | /* | |
62 | * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not | |
63 | * part of the API per se. | |
64 | * | |
65 | * returns 0 if reserve ok, or 1 if the slow path must be taken. | |
66 | */ | |
67 | static inline | |
4cfec15c MD |
68 | int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
69 | struct lttng_ust_lib_ring_buffer_ctx *ctx, | |
e56bb47c | 70 | void *client_ctx, |
852c2936 MD |
71 | unsigned long *o_begin, unsigned long *o_end, |
72 | unsigned long *o_old, size_t *before_hdr_pad) | |
73 | { | |
5198080d | 74 | struct lttng_ust_lib_ring_buffer_channel *chan = ctx->chan; |
4cfec15c | 75 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
76 | *o_begin = v_read(config, &buf->offset); |
77 | *o_old = *o_begin; | |
78 | ||
79 | ctx->tsc = lib_ring_buffer_clock_read(chan); | |
80 | if ((int64_t) ctx->tsc == -EIO) | |
81 | return 1; | |
82 | ||
83 | /* | |
84 | * Prefetch cacheline for read because we have to read the previous | |
85 | * commit counter to increment it and commit seq value to compare it to | |
86 | * the commit counter. | |
87 | */ | |
9f3fdbc6 | 88 | //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]); |
852c2936 MD |
89 | |
90 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
91 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
92 | ||
b5a3dfa5 | 93 | if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0)) |
852c2936 MD |
94 | return 1; |
95 | ||
96 | ctx->slot_size = record_header_size(config, chan, *o_begin, | |
e56bb47c | 97 | before_hdr_pad, ctx, client_ctx); |
852c2936 MD |
98 | ctx->slot_size += |
99 | lib_ring_buffer_align(*o_begin + ctx->slot_size, | |
100 | ctx->largest_align) + ctx->data_size; | |
b5a3dfa5 | 101 | if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size) |
852c2936 MD |
102 | > chan->backend.subbuf_size)) |
103 | return 1; | |
104 | ||
105 | /* | |
106 | * Record fits in the current buffer and we are not on a switch | |
107 | * boundary. It's safe to write. | |
108 | */ | |
109 | *o_end = *o_begin + ctx->slot_size; | |
1ad21f70 MD |
110 | |
111 | if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0)) | |
112 | /* | |
113 | * The offset_end will fall at the very beginning of the next | |
114 | * subbuffer. | |
115 | */ | |
116 | return 1; | |
117 | ||
852c2936 MD |
118 | return 0; |
119 | } | |
120 | ||
121 | /** | |
122 | * lib_ring_buffer_reserve - Reserve space in a ring buffer. | |
123 | * @config: ring buffer instance configuration. | |
124 | * @ctx: ring buffer context. (input and output) Must be already initialized. | |
125 | * | |
126 | * Atomic wait-free slot reservation. The reserved space starts at the context | |
127 | * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc". | |
128 | * | |
129 | * Return : | |
130 | * 0 on success. | |
131 | * -EAGAIN if channel is disabled. | |
132 | * -ENOSPC if event size is too large for packet. | |
133 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
134 | * -EIO if data cannot be written into the buffer for any other reason. | |
135 | */ | |
136 | ||
137 | static inline | |
4cfec15c | 138 | int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
e56bb47c MD |
139 | struct lttng_ust_lib_ring_buffer_ctx *ctx, |
140 | void *client_ctx) | |
852c2936 | 141 | { |
5198080d | 142 | struct lttng_ust_lib_ring_buffer_channel *chan = ctx->chan; |
38fae1d3 | 143 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c | 144 | struct lttng_ust_lib_ring_buffer *buf; |
852c2936 MD |
145 | unsigned long o_begin, o_end, o_old; |
146 | size_t before_hdr_pad = 0; | |
147 | ||
f52a5702 | 148 | if (caa_unlikely(uatomic_read(&chan->record_disabled))) |
852c2936 MD |
149 | return -EAGAIN; |
150 | ||
151 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 152 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 153 | else |
1d498196 | 154 | buf = shmp(handle, chan->backend.buf[0].shmp); |
15500a1b MD |
155 | if (caa_unlikely(!buf)) |
156 | return -EIO; | |
f52a5702 | 157 | if (caa_unlikely(uatomic_read(&buf->record_disabled))) |
852c2936 MD |
158 | return -EAGAIN; |
159 | ctx->buf = buf; | |
160 | ||
161 | /* | |
162 | * Perform retryable operations. | |
163 | */ | |
e56bb47c | 164 | if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, client_ctx, &o_begin, |
852c2936 MD |
165 | &o_end, &o_old, &before_hdr_pad))) |
166 | goto slow_path; | |
167 | ||
b5a3dfa5 | 168 | if (caa_unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end) |
852c2936 MD |
169 | != o_old)) |
170 | goto slow_path; | |
171 | ||
172 | /* | |
173 | * Atomically update last_tsc. This update races against concurrent | |
174 | * atomic updates, but the race will always cause supplementary full TSC | |
175 | * record headers, never the opposite (missing a full TSC record header | |
176 | * when it would be needed). | |
177 | */ | |
178 | save_last_tsc(config, ctx->buf, ctx->tsc); | |
179 | ||
180 | /* | |
181 | * Push the reader if necessary | |
182 | */ | |
183 | lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1); | |
184 | ||
185 | /* | |
186 | * Clear noref flag for this subbuffer. | |
187 | */ | |
188 | lib_ring_buffer_clear_noref(config, &ctx->buf->backend, | |
1d498196 | 189 | subbuf_index(o_end - 1, chan), handle); |
852c2936 MD |
190 | |
191 | ctx->pre_offset = o_begin; | |
192 | ctx->buf_offset = o_begin + before_hdr_pad; | |
193 | return 0; | |
194 | slow_path: | |
e56bb47c | 195 | return lib_ring_buffer_reserve_slow(ctx, client_ctx); |
852c2936 MD |
196 | } |
197 | ||
198 | /** | |
199 | * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer. | |
200 | * @config: ring buffer instance configuration. | |
201 | * @buf: buffer | |
202 | * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH) | |
203 | * | |
204 | * This operation is completely reentrant : can be called while tracing is | |
205 | * active with absolutely no lock held. | |
206 | * | |
207 | * Note, however, that as a v_cmpxchg is used for some atomic operations and | |
208 | * requires to be executed locally for per-CPU buffers, this function must be | |
209 | * called from the CPU which owns the buffer for a ACTIVE flush, with preemption | |
210 | * disabled, for RING_BUFFER_SYNC_PER_CPU configuration. | |
211 | */ | |
212 | static inline | |
4cfec15c MD |
213 | void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config, |
214 | struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, | |
38fae1d3 | 215 | struct lttng_ust_shm_handle *handle) |
852c2936 | 216 | { |
1d498196 | 217 | lib_ring_buffer_switch_slow(buf, mode, handle); |
852c2936 MD |
218 | } |
219 | ||
220 | /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */ | |
221 | ||
222 | /** | |
223 | * lib_ring_buffer_commit - Commit an record. | |
224 | * @config: ring buffer instance configuration. | |
225 | * @ctx: ring buffer context. (input arguments only) | |
226 | * | |
227 | * Atomic unordered slot commit. Increments the commit count in the | |
228 | * specified sub-buffer, and delivers it if necessary. | |
229 | */ | |
230 | static inline | |
4cfec15c MD |
231 | void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config, |
232 | const struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 | 233 | { |
5198080d | 234 | struct lttng_ust_lib_ring_buffer_channel *chan = ctx->chan; |
38fae1d3 | 235 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c | 236 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
237 | unsigned long offset_end = ctx->buf_offset; |
238 | unsigned long endidx = subbuf_index(offset_end - 1, chan); | |
239 | unsigned long commit_count; | |
d2fe4771 MD |
240 | struct commit_counters_hot *cc_hot = shmp_index(handle, |
241 | buf->commit_hot, endidx); | |
852c2936 | 242 | |
15500a1b MD |
243 | if (caa_unlikely(!cc_hot)) |
244 | return; | |
245 | ||
852c2936 MD |
246 | /* |
247 | * Must count record before incrementing the commit count. | |
248 | */ | |
15500a1b | 249 | subbuffer_count_record(config, ctx, &buf->backend, endidx, handle); |
852c2936 MD |
250 | |
251 | /* | |
252 | * Order all writes to buffer before the commit count update that will | |
253 | * determine that the subbuffer is full. | |
254 | */ | |
9f3fdbc6 | 255 | cmm_smp_wmb(); |
852c2936 | 256 | |
d2fe4771 | 257 | v_add(config, ctx->slot_size, &cc_hot->cc); |
852c2936 MD |
258 | |
259 | /* | |
260 | * commit count read can race with concurrent OOO commit count updates. | |
261 | * This is only needed for lib_ring_buffer_check_deliver (for | |
262 | * non-polling delivery only) and for | |
263 | * lib_ring_buffer_write_commit_counter. The race can only cause the | |
264 | * counter to be read with the same value more than once, which could | |
265 | * cause : | |
266 | * - Multiple delivery for the same sub-buffer (which is handled | |
267 | * gracefully by the reader code) if the value is for a full | |
268 | * sub-buffer. It's important that we can never miss a sub-buffer | |
269 | * delivery. Re-reading the value after the v_add ensures this. | |
270 | * - Reading a commit_count with a higher value that what was actually | |
271 | * added to it for the lib_ring_buffer_write_commit_counter call | |
272 | * (again caused by a concurrent committer). It does not matter, | |
273 | * because this function is interested in the fact that the commit | |
274 | * count reaches back the reserve offset for a specific sub-buffer, | |
275 | * which is completely independent of the order. | |
276 | */ | |
d2fe4771 | 277 | commit_count = v_read(config, &cc_hot->cc); |
852c2936 MD |
278 | |
279 | lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1, | |
1b7b0501 | 280 | commit_count, endidx, handle, ctx->tsc); |
852c2936 MD |
281 | /* |
282 | * Update used size at each commit. It's needed only for extracting | |
283 | * ring_buffer buffers from vmcore, after crash. | |
284 | */ | |
d2fe4771 MD |
285 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
286 | offset_end, commit_count, handle, cc_hot); | |
852c2936 MD |
287 | } |
288 | ||
289 | /** | |
290 | * lib_ring_buffer_try_discard_reserve - Try discarding a record. | |
291 | * @config: ring buffer instance configuration. | |
292 | * @ctx: ring buffer context. (input arguments only) | |
293 | * | |
294 | * Only succeeds if no other record has been written after the record to | |
295 | * discard. If discard fails, the record must be committed to the buffer. | |
296 | * | |
297 | * Returns 0 upon success, -EPERM if the record cannot be discarded. | |
298 | */ | |
299 | static inline | |
4cfec15c MD |
300 | int lib_ring_buffer_try_discard_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
301 | const struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 | 302 | { |
4cfec15c | 303 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
304 | unsigned long end_offset = ctx->pre_offset + ctx->slot_size; |
305 | ||
306 | /* | |
307 | * We need to ensure that if the cmpxchg succeeds and discards the | |
308 | * record, the next record will record a full TSC, because it cannot | |
309 | * rely on the last_tsc associated with the discarded record to detect | |
310 | * overflows. The only way to ensure this is to set the last_tsc to 0 | |
311 | * (assuming no 64-bit TSC overflow), which forces to write a 64-bit | |
312 | * timestamp in the next record. | |
313 | * | |
314 | * Note: if discard fails, we must leave the TSC in the record header. | |
315 | * It is needed to keep track of TSC overflows for the following | |
316 | * records. | |
317 | */ | |
318 | save_last_tsc(config, buf, 0ULL); | |
319 | ||
b5a3dfa5 | 320 | if (caa_likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset) |
852c2936 MD |
321 | != end_offset)) |
322 | return -EPERM; | |
323 | else | |
324 | return 0; | |
325 | } | |
326 | ||
327 | static inline | |
4cfec15c | 328 | void channel_record_disable(const struct lttng_ust_lib_ring_buffer_config *config, |
5198080d | 329 | struct lttng_ust_lib_ring_buffer_channel *chan) |
852c2936 | 330 | { |
9f3fdbc6 | 331 | uatomic_inc(&chan->record_disabled); |
852c2936 MD |
332 | } |
333 | ||
334 | static inline | |
4cfec15c | 335 | void channel_record_enable(const struct lttng_ust_lib_ring_buffer_config *config, |
5198080d | 336 | struct lttng_ust_lib_ring_buffer_channel *chan) |
852c2936 | 337 | { |
9f3fdbc6 | 338 | uatomic_dec(&chan->record_disabled); |
852c2936 MD |
339 | } |
340 | ||
341 | static inline | |
4cfec15c MD |
342 | void lib_ring_buffer_record_disable(const struct lttng_ust_lib_ring_buffer_config *config, |
343 | struct lttng_ust_lib_ring_buffer *buf) | |
852c2936 | 344 | { |
9f3fdbc6 | 345 | uatomic_inc(&buf->record_disabled); |
852c2936 MD |
346 | } |
347 | ||
348 | static inline | |
4cfec15c MD |
349 | void lib_ring_buffer_record_enable(const struct lttng_ust_lib_ring_buffer_config *config, |
350 | struct lttng_ust_lib_ring_buffer *buf) | |
852c2936 | 351 | { |
9f3fdbc6 | 352 | uatomic_dec(&buf->record_disabled); |
852c2936 MD |
353 | } |
354 | ||
e92f3e28 | 355 | #endif /* _LTTNG_RING_BUFFER_FRONTEND_API_H */ |