Commit | Line | Data |
---|---|---|
886d51a3 MD |
1 | #ifndef _LIB_RING_BUFFER_FRONTEND_API_H |
2 | #define _LIB_RING_BUFFER_FRONTEND_API_H | |
f3bc08c5 MD |
3 | |
4 | /* | |
886d51a3 | 5 | * lib/ringbuffer/frontend_api.h |
f3bc08c5 MD |
6 | * |
7 | * Ring Buffer Library Synchronization Header (buffer write API). | |
8 | * | |
886d51a3 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 | * | |
f3bc08c5 MD |
25 | * Author: |
26 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
27 | * | |
28 | * See ring_buffer_frontend.c for more information on wait-free algorithms. | |
29 | * See linux/ringbuffer/frontend.h for channel allocation and read-side API. | |
f3bc08c5 MD |
30 | */ |
31 | ||
32 | #include "../../wrapper/ringbuffer/frontend.h" | |
33 | #include <linux/errno.h> | |
dfbc2ec7 | 34 | #include <linux/prefetch.h> |
f3bc08c5 MD |
35 | |
36 | /** | |
37 | * lib_ring_buffer_get_cpu - Precedes ring buffer reserve/commit. | |
38 | * | |
39 | * Disables preemption (acts as a RCU read-side critical section) and keeps a | |
40 | * ring buffer nesting count as supplementary safety net to ensure tracer client | |
41 | * code will never trigger an endless recursion. Returns the processor ID on | |
42 | * success, -EPERM on failure (nesting count too high). | |
43 | * | |
44 | * asm volatile and "memory" clobber prevent the compiler from moving | |
45 | * instructions out of the ring buffer nesting count. This is required to ensure | |
46 | * that probe side-effects which can cause recursion (e.g. unforeseen traps, | |
47 | * divisions by 0, ...) are triggered within the incremented nesting count | |
48 | * section. | |
49 | */ | |
50 | static inline | |
51 | int lib_ring_buffer_get_cpu(const struct lib_ring_buffer_config *config) | |
52 | { | |
53 | int cpu, nesting; | |
54 | ||
55 | rcu_read_lock_sched_notrace(); | |
56 | cpu = smp_processor_id(); | |
57 | nesting = ++per_cpu(lib_ring_buffer_nesting, cpu); | |
58 | barrier(); | |
59 | ||
60 | if (unlikely(nesting > 4)) { | |
61 | WARN_ON_ONCE(1); | |
62 | per_cpu(lib_ring_buffer_nesting, cpu)--; | |
63 | rcu_read_unlock_sched_notrace(); | |
64 | return -EPERM; | |
65 | } else | |
66 | return cpu; | |
67 | } | |
68 | ||
69 | /** | |
70 | * lib_ring_buffer_put_cpu - Follows ring buffer reserve/commit. | |
71 | */ | |
72 | static inline | |
73 | void lib_ring_buffer_put_cpu(const struct lib_ring_buffer_config *config) | |
74 | { | |
75 | barrier(); | |
76 | __get_cpu_var(lib_ring_buffer_nesting)--; | |
77 | rcu_read_unlock_sched_notrace(); | |
78 | } | |
79 | ||
80 | /* | |
81 | * lib_ring_buffer_try_reserve is called by lib_ring_buffer_reserve(). It is not | |
82 | * part of the API per se. | |
83 | * | |
84 | * returns 0 if reserve ok, or 1 if the slow path must be taken. | |
85 | */ | |
86 | static inline | |
87 | int lib_ring_buffer_try_reserve(const struct lib_ring_buffer_config *config, | |
88 | struct lib_ring_buffer_ctx *ctx, | |
89 | unsigned long *o_begin, unsigned long *o_end, | |
90 | unsigned long *o_old, size_t *before_hdr_pad) | |
91 | { | |
92 | struct channel *chan = ctx->chan; | |
93 | struct lib_ring_buffer *buf = ctx->buf; | |
94 | *o_begin = v_read(config, &buf->offset); | |
95 | *o_old = *o_begin; | |
96 | ||
97 | ctx->tsc = lib_ring_buffer_clock_read(chan); | |
97ca2c54 MD |
98 | if ((int64_t) ctx->tsc == -EIO) |
99 | return 1; | |
f3bc08c5 MD |
100 | |
101 | /* | |
102 | * Prefetch cacheline for read because we have to read the previous | |
103 | * commit counter to increment it and commit seq value to compare it to | |
104 | * the commit counter. | |
105 | */ | |
106 | prefetch(&buf->commit_hot[subbuf_index(*o_begin, chan)]); | |
107 | ||
108 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
64c796d8 | 109 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; |
f3bc08c5 MD |
110 | |
111 | if (unlikely(subbuf_offset(*o_begin, chan) == 0)) | |
112 | return 1; | |
113 | ||
114 | ctx->slot_size = record_header_size(config, chan, *o_begin, | |
64c796d8 | 115 | before_hdr_pad, ctx); |
f3bc08c5 MD |
116 | ctx->slot_size += |
117 | lib_ring_buffer_align(*o_begin + ctx->slot_size, | |
118 | ctx->largest_align) + ctx->data_size; | |
119 | if (unlikely((subbuf_offset(*o_begin, chan) + ctx->slot_size) | |
120 | > chan->backend.subbuf_size)) | |
121 | return 1; | |
122 | ||
123 | /* | |
124 | * Record fits in the current buffer and we are not on a switch | |
125 | * boundary. It's safe to write. | |
126 | */ | |
127 | *o_end = *o_begin + ctx->slot_size; | |
128 | ||
129 | if (unlikely((subbuf_offset(*o_end, chan)) == 0)) | |
130 | /* | |
131 | * The offset_end will fall at the very beginning of the next | |
132 | * subbuffer. | |
133 | */ | |
134 | return 1; | |
135 | ||
136 | return 0; | |
137 | } | |
138 | ||
139 | /** | |
140 | * lib_ring_buffer_reserve - Reserve space in a ring buffer. | |
141 | * @config: ring buffer instance configuration. | |
142 | * @ctx: ring buffer context. (input and output) Must be already initialized. | |
143 | * | |
144 | * Atomic wait-free slot reservation. The reserved space starts at the context | |
145 | * "pre_offset". Its length is "slot_size". The associated time-stamp is "tsc". | |
146 | * | |
97ca2c54 MD |
147 | * Return : |
148 | * 0 on success. | |
149 | * -EAGAIN if channel is disabled. | |
150 | * -ENOSPC if event size is too large for packet. | |
151 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
152 | * -EIO if data cannot be written into the buffer for any other reason. | |
f3bc08c5 MD |
153 | */ |
154 | ||
155 | static inline | |
156 | int lib_ring_buffer_reserve(const struct lib_ring_buffer_config *config, | |
157 | struct lib_ring_buffer_ctx *ctx) | |
158 | { | |
159 | struct channel *chan = ctx->chan; | |
160 | struct lib_ring_buffer *buf; | |
161 | unsigned long o_begin, o_end, o_old; | |
162 | size_t before_hdr_pad = 0; | |
163 | ||
164 | if (atomic_read(&chan->record_disabled)) | |
165 | return -EAGAIN; | |
166 | ||
167 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
168 | buf = per_cpu_ptr(chan->backend.buf, ctx->cpu); | |
169 | else | |
170 | buf = chan->backend.buf; | |
171 | if (atomic_read(&buf->record_disabled)) | |
172 | return -EAGAIN; | |
173 | ctx->buf = buf; | |
174 | ||
175 | /* | |
176 | * Perform retryable operations. | |
177 | */ | |
178 | if (unlikely(lib_ring_buffer_try_reserve(config, ctx, &o_begin, | |
179 | &o_end, &o_old, &before_hdr_pad))) | |
180 | goto slow_path; | |
181 | ||
182 | if (unlikely(v_cmpxchg(config, &ctx->buf->offset, o_old, o_end) | |
183 | != o_old)) | |
184 | goto slow_path; | |
185 | ||
186 | /* | |
187 | * Atomically update last_tsc. This update races against concurrent | |
188 | * atomic updates, but the race will always cause supplementary full TSC | |
189 | * record headers, never the opposite (missing a full TSC record header | |
190 | * when it would be needed). | |
191 | */ | |
192 | save_last_tsc(config, ctx->buf, ctx->tsc); | |
193 | ||
194 | /* | |
195 | * Push the reader if necessary | |
196 | */ | |
197 | lib_ring_buffer_reserve_push_reader(ctx->buf, chan, o_end - 1); | |
198 | ||
199 | /* | |
200 | * Clear noref flag for this subbuffer. | |
201 | */ | |
202 | lib_ring_buffer_clear_noref(config, &ctx->buf->backend, | |
203 | subbuf_index(o_end - 1, chan)); | |
204 | ||
205 | ctx->pre_offset = o_begin; | |
206 | ctx->buf_offset = o_begin + before_hdr_pad; | |
207 | return 0; | |
208 | slow_path: | |
209 | return lib_ring_buffer_reserve_slow(ctx); | |
210 | } | |
211 | ||
212 | /** | |
213 | * lib_ring_buffer_switch - Perform a sub-buffer switch for a per-cpu buffer. | |
214 | * @config: ring buffer instance configuration. | |
215 | * @buf: buffer | |
216 | * @mode: buffer switch mode (SWITCH_ACTIVE or SWITCH_FLUSH) | |
217 | * | |
218 | * This operation is completely reentrant : can be called while tracing is | |
219 | * active with absolutely no lock held. | |
220 | * | |
221 | * Note, however, that as a v_cmpxchg is used for some atomic operations and | |
222 | * requires to be executed locally for per-CPU buffers, this function must be | |
223 | * called from the CPU which owns the buffer for a ACTIVE flush, with preemption | |
224 | * disabled, for RING_BUFFER_SYNC_PER_CPU configuration. | |
225 | */ | |
226 | static inline | |
227 | void lib_ring_buffer_switch(const struct lib_ring_buffer_config *config, | |
228 | struct lib_ring_buffer *buf, enum switch_mode mode) | |
229 | { | |
230 | lib_ring_buffer_switch_slow(buf, mode); | |
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 | |
244 | void lib_ring_buffer_commit(const struct lib_ring_buffer_config *config, | |
245 | const struct lib_ring_buffer_ctx *ctx) | |
246 | { | |
247 | struct channel *chan = ctx->chan; | |
248 | struct lib_ring_buffer *buf = ctx->buf; | |
249 | unsigned long offset_end = ctx->buf_offset; | |
250 | unsigned long endidx = subbuf_index(offset_end - 1, chan); | |
251 | unsigned long commit_count; | |
252 | ||
253 | /* | |
254 | * Must count record before incrementing the commit count. | |
255 | */ | |
256 | subbuffer_count_record(config, &buf->backend, endidx); | |
257 | ||
258 | /* | |
259 | * Order all writes to buffer before the commit count update that will | |
260 | * determine that the subbuffer is full. | |
261 | */ | |
262 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
263 | /* | |
264 | * Must write slot data before incrementing commit count. This | |
265 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
266 | * by get_subbuf(). | |
267 | */ | |
268 | barrier(); | |
269 | } else | |
270 | smp_wmb(); | |
271 | ||
272 | v_add(config, ctx->slot_size, &buf->commit_hot[endidx].cc); | |
273 | ||
274 | /* | |
275 | * commit count read can race with concurrent OOO commit count updates. | |
276 | * This is only needed for lib_ring_buffer_check_deliver (for | |
277 | * non-polling delivery only) and for | |
278 | * lib_ring_buffer_write_commit_counter. The race can only cause the | |
279 | * counter to be read with the same value more than once, which could | |
280 | * cause : | |
281 | * - Multiple delivery for the same sub-buffer (which is handled | |
282 | * gracefully by the reader code) if the value is for a full | |
283 | * sub-buffer. It's important that we can never miss a sub-buffer | |
284 | * delivery. Re-reading the value after the v_add ensures this. | |
285 | * - Reading a commit_count with a higher value that what was actually | |
286 | * added to it for the lib_ring_buffer_write_commit_counter call | |
287 | * (again caused by a concurrent committer). It does not matter, | |
288 | * because this function is interested in the fact that the commit | |
289 | * count reaches back the reserve offset for a specific sub-buffer, | |
290 | * which is completely independent of the order. | |
291 | */ | |
292 | commit_count = v_read(config, &buf->commit_hot[endidx].cc); | |
293 | ||
294 | lib_ring_buffer_check_deliver(config, buf, chan, offset_end - 1, | |
295 | commit_count, endidx); | |
296 | /* | |
297 | * Update used size at each commit. It's needed only for extracting | |
298 | * ring_buffer buffers from vmcore, after crash. | |
299 | */ | |
300 | lib_ring_buffer_write_commit_counter(config, buf, chan, endidx, | |
301 | ctx->buf_offset, commit_count, | |
302 | ctx->slot_size); | |
303 | } | |
304 | ||
305 | /** | |
306 | * lib_ring_buffer_try_discard_reserve - Try discarding a record. | |
307 | * @config: ring buffer instance configuration. | |
308 | * @ctx: ring buffer context. (input arguments only) | |
309 | * | |
310 | * Only succeeds if no other record has been written after the record to | |
311 | * discard. If discard fails, the record must be committed to the buffer. | |
312 | * | |
313 | * Returns 0 upon success, -EPERM if the record cannot be discarded. | |
314 | */ | |
315 | static inline | |
316 | int lib_ring_buffer_try_discard_reserve(const struct lib_ring_buffer_config *config, | |
317 | const struct lib_ring_buffer_ctx *ctx) | |
318 | { | |
319 | struct lib_ring_buffer *buf = ctx->buf; | |
320 | unsigned long end_offset = ctx->pre_offset + ctx->slot_size; | |
321 | ||
322 | /* | |
323 | * We need to ensure that if the cmpxchg succeeds and discards the | |
324 | * record, the next record will record a full TSC, because it cannot | |
325 | * rely on the last_tsc associated with the discarded record to detect | |
326 | * overflows. The only way to ensure this is to set the last_tsc to 0 | |
327 | * (assuming no 64-bit TSC overflow), which forces to write a 64-bit | |
328 | * timestamp in the next record. | |
329 | * | |
330 | * Note: if discard fails, we must leave the TSC in the record header. | |
331 | * It is needed to keep track of TSC overflows for the following | |
332 | * records. | |
333 | */ | |
334 | save_last_tsc(config, buf, 0ULL); | |
335 | ||
336 | if (likely(v_cmpxchg(config, &buf->offset, end_offset, ctx->pre_offset) | |
337 | != end_offset)) | |
338 | return -EPERM; | |
339 | else | |
340 | return 0; | |
341 | } | |
342 | ||
343 | static inline | |
344 | void channel_record_disable(const struct lib_ring_buffer_config *config, | |
345 | struct channel *chan) | |
346 | { | |
347 | atomic_inc(&chan->record_disabled); | |
348 | } | |
349 | ||
350 | static inline | |
351 | void channel_record_enable(const struct lib_ring_buffer_config *config, | |
352 | struct channel *chan) | |
353 | { | |
354 | atomic_dec(&chan->record_disabled); | |
355 | } | |
356 | ||
357 | static inline | |
358 | void lib_ring_buffer_record_disable(const struct lib_ring_buffer_config *config, | |
359 | struct lib_ring_buffer *buf) | |
360 | { | |
361 | atomic_inc(&buf->record_disabled); | |
362 | } | |
363 | ||
364 | static inline | |
365 | void lib_ring_buffer_record_enable(const struct lib_ring_buffer_config *config, | |
366 | struct lib_ring_buffer *buf) | |
367 | { | |
368 | atomic_dec(&buf->record_disabled); | |
369 | } | |
370 | ||
886d51a3 | 371 | #endif /* _LIB_RING_BUFFER_FRONTEND_API_H */ |