Commit | Line | Data |
---|---|---|
852c2936 MD |
1 | #ifndef _LINUX_RING_BUFFER_FRONTEND_API_H |
2 | #define _LINUX_RING_BUFFER_FRONTEND_API_H | |
3 | ||
4 | /* | |
5 | * linux/ringbuffer/frontend_api.h | |
6 | * | |
7 | * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
8 | * | |
9 | * Ring Buffer Library Synchronization Header (buffer write API). | |
10 | * | |
11 | * Author: | |
12 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
13 | * | |
14 | * See ring_buffer_frontend.c for more information on wait-free algorithms. | |
15 | * See linux/ringbuffer/frontend.h for channel allocation and read-side API. | |
16 | * | |
17 | * Dual LGPL v2.1/GPL v2 license. | |
18 | */ | |
19 | ||
4931a13e | 20 | #include "frontend.h" |
9f3fdbc6 MD |
21 | #include <urcu-bp.h> |
22 | #include <urcu/compiler.h> | |
852c2936 MD |
23 | |
24 | /** | |
25 | * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit. | |
26 | * | |
9f3fdbc6 MD |
27 | * Grabs RCU read-side lock and keeps a ring buffer nesting count as |
28 | * supplementary safety net to ensure tracer client code will never | |
29 | * trigger an endless recursion. Returns the processor ID on success, | |
30 | * -EPERM on failure (nesting count too high). | |
852c2936 MD |
31 | * |
32 | * asm volatile and "memory" clobber prevent the compiler from moving | |
33 | * instructions out of the ring buffer nesting count. This is required to ensure | |
34 | * that probe side-effects which can cause recursion (e.g. unforeseen traps, | |
35 | * divisions by 0, ...) are triggered within the incremented nesting count | |
36 | * section. | |
37 | */ | |
38 | static inline | |
4cfec15c | 39 | int lib_ring_buffer_get_cpu(const struct lttng_ust_lib_ring_buffer_config *config) |
852c2936 MD |
40 | { |
41 | int cpu, nesting; | |
42 | ||
9f3fdbc6 | 43 | rcu_read_lock(); |
2b2d6ff7 | 44 | cpu = lttng_ust_get_cpu(); |
b9b858b8 | 45 | nesting = ++lib_ring_buffer_nesting; /* TLS */ |
9f3fdbc6 | 46 | cmm_barrier(); |
852c2936 | 47 | |
b5a3dfa5 | 48 | if (caa_unlikely(nesting > 4)) { |
852c2936 | 49 | WARN_ON_ONCE(1); |
b9b858b8 | 50 | lib_ring_buffer_nesting--; /* TLS */ |
9f3fdbc6 | 51 | rcu_read_unlock(); |
852c2936 MD |
52 | return -EPERM; |
53 | } else | |
54 | return cpu; | |
55 | } | |
56 | ||
57 | /** | |
58 | * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit. | |
59 | */ | |
60 | static inline | |
4cfec15c | 61 | void lib_ring_buffer_put_cpu(const struct lttng_ust_lib_ring_buffer_config *config) |
852c2936 | 62 | { |
9f3fdbc6 | 63 | cmm_barrier(); |
b9b858b8 | 64 | lib_ring_buffer_nesting--; /* TLS */ |
9f3fdbc6 | 65 | rcu_read_unlock(); |
852c2936 MD |
66 | } |
67 | ||
68 | /* | |
69 | * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not | |
70 | * part of the API per se. | |
71 | * | |
72 | * returns 0 if reserve ok, or 1 if the slow path must be taken. | |
73 | */ | |
74 | static inline | |
4cfec15c MD |
75 | int lib_ring_buffer_try_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
76 | struct lttng_ust_lib_ring_buffer_ctx *ctx, | |
852c2936 MD |
77 | unsigned long *o_begin, unsigned long *o_end, |
78 | unsigned long *o_old, size_t *before_hdr_pad) | |
79 | { | |
80 | struct channel *chan = ctx->chan; | |
4cfec15c | 81 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
82 | *o_begin = v_read(config, &buf->offset); |
83 | *o_old = *o_begin; | |
84 | ||
85 | ctx->tsc = lib_ring_buffer_clock_read(chan); | |
86 | if ((int64_t) ctx->tsc == -EIO) | |
87 | return 1; | |
88 | ||
89 | /* | |
90 | * Prefetch cacheline for read because we have to read the previous | |
91 | * commit counter to increment it and commit seq value to compare it to | |
92 | * the commit counter. | |
93 | */ | |
9f3fdbc6 | 94 | //prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]); |
852c2936 | 95 | |
aae88c70 MD |
96 | /* |
97 | * Because we don't use any timer in the application, we | |
98 | * currently cannot guarantee that we have frequent | |
99 | * events that let us detect 27-bit overflows. | |
100 | * Therefore, for now, we force event headers | |
101 | * to contain 64-bit timestamps. | |
102 | */ | |
103 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
104 | #if 0 | |
852c2936 MD |
105 | if (last_tsc_overflow(config, buf, ctx->tsc)) |
106 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; | |
aae88c70 | 107 | #endif //0 |
852c2936 | 108 | |
b5a3dfa5 | 109 | if (caa_unlikely(subbuf_offset(*o_begin, chan) == 0)) |
852c2936 MD |
110 | return 1; |
111 | ||
112 | ctx->slot_size = record_header_size(config, chan, *o_begin, | |
113 | before_hdr_pad, ctx); | |
114 | ctx->slot_size += | |
115 | lib_ring_buffer_align(*o_begin + ctx->slot_size, | |
116 | ctx->largest_align) + ctx->data_size; | |
b5a3dfa5 | 117 | if (caa_unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size) |
852c2936 MD |
118 | > chan->backend.subbuf_size)) |
119 | return 1; | |
120 | ||
121 | /* | |
122 | * Record fits in the current buffer and we are not on a switch | |
123 | * boundary. It's safe to write. | |
124 | */ | |
125 | *o_end = *o_begin + ctx->slot_size; | |
126 | ||
b5a3dfa5 | 127 | if (caa_unlikely((subbuf_offset(*o_end, chan)) == 0)) |
852c2936 MD |
128 | /* |
129 | * The offset_end will fall at the very beginning of the next | |
130 | * subbuffer. | |
131 | */ | |
132 | return 1; | |
133 | ||
134 | return 0; | |
135 | } | |
136 | ||
137 | /** | |
138 | * lib_ring_buffer_reserve - Reserve space in a ring buffer. | |
139 | * @config: ring buffer instance configuration. | |
140 | * @ctx: ring buffer context. (input and output) Must be already initialized. | |
141 | * | |
142 | * Atomic wait-free slot reservation. The reserved space starts at the context | |
143 | * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc". | |
144 | * | |
145 | * Return : | |
146 | * 0 on success. | |
147 | * -EAGAIN if channel is disabled. | |
148 | * -ENOSPC if event size is too large for packet. | |
149 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
150 | * -EIO if data cannot be written into the buffer for any other reason. | |
151 | */ | |
152 | ||
153 | static inline | |
4cfec15c MD |
154 | int lib_ring_buffer_reserve(const struct lttng_ust_lib_ring_buffer_config *config, |
155 | struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 MD |
156 | { |
157 | struct channel *chan = ctx->chan; | |
38fae1d3 | 158 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c | 159 | struct lttng_ust_lib_ring_buffer *buf; |
852c2936 MD |
160 | unsigned long o_begin, o_end, o_old; |
161 | size_t before_hdr_pad = 0; | |
162 | ||
9f3fdbc6 | 163 | if (uatomic_read(&chan->record_disabled)) |
852c2936 MD |
164 | return -EAGAIN; |
165 | ||
166 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1d498196 | 167 | buf = shmp(handle, chan->backend.buf[ctx->cpu].shmp); |
852c2936 | 168 | else |
1d498196 | 169 | buf = shmp(handle, chan->backend.buf[0].shmp); |
9f3fdbc6 | 170 | if (uatomic_read(&buf->record_disabled)) |
852c2936 MD |
171 | return -EAGAIN; |
172 | ctx->buf = buf; | |
173 | ||
174 | /* | |
175 | * Perform retryable operations. | |
176 | */ | |
b5a3dfa5 | 177 | if (caa_unlikely(lib_ring_buffer_try_reserve(config, ctx, &o_begin, |
852c2936 MD |
178 | &o_end, &o_old, &before_hdr_pad))) |
179 | goto slow_path; | |
180 | ||
b5a3dfa5 | 181 | if (caa_unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end) |
852c2936 MD |
182 | != o_old)) |
183 | goto slow_path; | |
184 | ||
185 | /* | |
186 | * Atomically update last_tsc. This update races against concurrent | |
187 | * atomic updates, but the race will always cause supplementary full TSC | |
188 | * record headers, never the opposite (missing a full TSC record header | |
189 | * when it would be needed). | |
190 | */ | |
191 | save_last_tsc(config, ctx->buf, ctx->tsc); | |
192 | ||
193 | /* | |
194 | * Push the reader if necessary | |
195 | */ | |
196 | lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1); | |
197 | ||
198 | /* | |
199 | * Clear noref flag for this subbuffer. | |
200 | */ | |
201 | lib_ring_buffer_clear_noref(config, &ctx->buf->backend, | |
1d498196 | 202 | subbuf_index(o_end - 1, chan), handle); |
852c2936 MD |
203 | |
204 | ctx->pre_offset = o_begin; | |
205 | ctx->buf_offset = o_begin + before_hdr_pad; | |
206 | return 0; | |
207 | slow_path: | |
208 | return lib_ring_buffer_reserve_slow(ctx); | |
209 | } | |
210 | ||
211 | /** | |
212 | * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer. | |
213 | * @config: ring buffer instance configuration. | |
214 | * @buf: buffer | |
215 | * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH) | |
216 | * | |
217 | * This operation is completely reentrant : can be called while tracing is | |
218 | * active with absolutely no lock held. | |
219 | * | |
220 | * Note, however, that as a v_cmpxchg is used for some atomic operations and | |
221 | * requires to be executed locally for per-CPU buffers, this function must be | |
222 | * called from the CPU which owns the buffer for a ACTIVE flush, with preemption | |
223 | * disabled, for RING_BUFFER_SYNC_PER_CPU configuration. | |
224 | */ | |
225 | static inline | |
4cfec15c MD |
226 | void lib_ring_buffer_switch(const struct lttng_ust_lib_ring_buffer_config *config, |
227 | struct lttng_ust_lib_ring_buffer *buf, enum switch_mode mode, | |
38fae1d3 | 228 | struct lttng_ust_shm_handle *handle) |
852c2936 | 229 | { |
1d498196 | 230 | lib_ring_buffer_switch_slow(buf, mode, handle); |
852c2936 MD |
231 | } |
232 | ||
233 | /* See ring_buffer_frontend_api.h for lib_ring_buffer_reserve(). */ | |
234 | ||
235 | /** | |
236 | * lib_ring_buffer_commit - Commit an record. | |
237 | * @config: ring buffer instance configuration. | |
238 | * @ctx: ring buffer context. (input arguments only) | |
239 | * | |
240 | * Atomic unordered slot commit. Increments the commit count in the | |
241 | * specified sub-buffer, and delivers it if necessary. | |
242 | */ | |
243 | static inline | |
4cfec15c MD |
244 | void lib_ring_buffer_commit(const struct lttng_ust_lib_ring_buffer_config *config, |
245 | const struct lttng_ust_lib_ring_buffer_ctx *ctx) | |
852c2936 MD |
246 | { |
247 | struct channel *chan = ctx->chan; | |
38fae1d3 | 248 | struct lttng_ust_shm_handle *handle = ctx->handle; |
4cfec15c | 249 | struct lttng_ust_lib_ring_buffer *buf = ctx->buf; |
852c2936 MD |
250 | unsigned long offset_end = ctx->buf_offset; |
251 | unsigned long endidx = subbuf_index(offset_end - 1, chan); | |
252 | unsigned long commit_count; | |
253 | ||
254 | /* | |
255 | * Must count record before incrementing the commit count. | |
256 | */ | |
1d498196 | 257 | subbuffer_count_record(config, &buf->backend, endidx, handle); |
852c2936 MD |
258 | |
259 | /* | |
260 | * Order all writes to buffer before the commit count update that will | |
261 | * determine that the subbuffer is full. | |
262 | */ | |
9f3fdbc6 | 263 | cmm_smp_wmb(); |
852c2936 | 264 | |
4746ae29 | 265 | v_add(config, ctx->slot_size, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
852c2936 MD |
266 | |
267 | /* | |
268 | * commit count read can race with concurrent OOO commit count updates. | |
269 | * This is only needed for lib_ring_buffer_check_deliver (for | |
270 | * non-polling delivery only) and for | |
271 | * lib_ring_buffer_write_commit_counter. The race can only cause the | |
272 | * counter to be read with the same value more than once, which could | |
273 | * cause : | |
274 | * - Multiple delivery for the same sub-buffer (which is handled | |
275 | * gracefully by the reader code) if the value is for a full | |
276 | * sub-buffer. It's important that we can never miss a sub-buffer | |
277 | * delivery. Re-reading the value after the v_add ensures this. | |
278 | * - Reading a commit_count with a higher value that what was actually | |
279 | * added to it for the lib_ring_buffer_write_commit_counter call | |
280 | * (again caused by a concurrent committer). It does not matter, | |
281 | * because this function is interested in the fact that the commit | |
282 | * count reaches back the reserve offset for a specific sub-buffer, | |
283 | * which is completely independent of the order. | |
284 | */ | |
4746ae29 | 285 | commit_count = v_read(config, &shmp_index(handle, buf->commit_hot, endidx)->cc); |
852c2936 MD |
286 | |
287 | lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1, | |
1d498196 | 288 | commit_count, endidx, handle); |
852c2936 MD |
289 | /* |
290 | * Update used size at each commit. It's needed only for extracting | |
291 | * ring_buffer buffers from vmcore, after crash. | |
292 | */ | |
293 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, | |
294 | ctx->buf_offset, commit_count, | |
1d498196 | 295 | ctx->slot_size, 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 | ||
364 | #endif /* _LINUX_RING_BUFFER_FRONTEND_API_H */ |