Commit | Line | Data |
---|---|---|
f3bc08c5 MD |
1 | #ifndef _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H |
2 | #define _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H | |
3 | ||
4 | /* | |
5 | * linux/ringbuffer/frontend_internal.h | |
6 | * | |
7 | * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
8 | * | |
9 | * Ring Buffer Library Synchronization Header (internal helpers). | |
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 | * | |
16 | * Dual LGPL v2.1/GPL v2 license. | |
17 | */ | |
18 | ||
19 | #include "../../wrapper/ringbuffer/config.h" | |
20 | #include "../../wrapper/ringbuffer/backend_types.h" | |
21 | #include "../../wrapper/ringbuffer/frontend_types.h" | |
a88db018 | 22 | #include "../../lib/prio_heap/lttng_prio_heap.h" /* For per-CPU read-side iterator */ |
f3bc08c5 MD |
23 | |
24 | /* Buffer offset macros */ | |
25 | ||
26 | /* buf_trunc mask selects only the buffer number. */ | |
27 | static inline | |
28 | unsigned long buf_trunc(unsigned long offset, struct channel *chan) | |
29 | { | |
30 | return offset & ~(chan->backend.buf_size - 1); | |
31 | ||
32 | } | |
33 | ||
34 | /* Select the buffer number value (counter). */ | |
35 | static inline | |
36 | unsigned long buf_trunc_val(unsigned long offset, struct channel *chan) | |
37 | { | |
38 | return buf_trunc(offset, chan) >> chan->backend.buf_size_order; | |
39 | } | |
40 | ||
41 | /* buf_offset mask selects only the offset within the current buffer. */ | |
42 | static inline | |
43 | unsigned long buf_offset(unsigned long offset, struct channel *chan) | |
44 | { | |
45 | return offset & (chan->backend.buf_size - 1); | |
46 | } | |
47 | ||
48 | /* subbuf_offset mask selects the offset within the current subbuffer. */ | |
49 | static inline | |
50 | unsigned long subbuf_offset(unsigned long offset, struct channel *chan) | |
51 | { | |
52 | return offset & (chan->backend.subbuf_size - 1); | |
53 | } | |
54 | ||
55 | /* subbuf_trunc mask selects the subbuffer number. */ | |
56 | static inline | |
57 | unsigned long subbuf_trunc(unsigned long offset, struct channel *chan) | |
58 | { | |
59 | return offset & ~(chan->backend.subbuf_size - 1); | |
60 | } | |
61 | ||
62 | /* subbuf_align aligns the offset to the next subbuffer. */ | |
63 | static inline | |
64 | unsigned long subbuf_align(unsigned long offset, struct channel *chan) | |
65 | { | |
66 | return (offset + chan->backend.subbuf_size) | |
67 | & ~(chan->backend.subbuf_size - 1); | |
68 | } | |
69 | ||
70 | /* subbuf_index returns the index of the current subbuffer within the buffer. */ | |
71 | static inline | |
72 | unsigned long subbuf_index(unsigned long offset, struct channel *chan) | |
73 | { | |
74 | return buf_offset(offset, chan) >> chan->backend.subbuf_size_order; | |
75 | } | |
76 | ||
77 | /* | |
78 | * Last TSC comparison functions. Check if the current TSC overflows tsc_bits | |
79 | * bits from the last TSC read. When overflows are detected, the full 64-bit | |
80 | * timestamp counter should be written in the record header. Reads and writes | |
81 | * last_tsc atomically. | |
82 | */ | |
83 | ||
84 | #if (BITS_PER_LONG == 32) | |
85 | static inline | |
86 | void save_last_tsc(const struct lib_ring_buffer_config *config, | |
87 | struct lib_ring_buffer *buf, u64 tsc) | |
88 | { | |
89 | if (config->tsc_bits == 0 || config->tsc_bits == 64) | |
90 | return; | |
91 | ||
92 | /* | |
93 | * Ensure the compiler performs this update in a single instruction. | |
94 | */ | |
95 | v_set(config, &buf->last_tsc, (unsigned long)(tsc >> config->tsc_bits)); | |
96 | } | |
97 | ||
98 | static inline | |
99 | int last_tsc_overflow(const struct lib_ring_buffer_config *config, | |
100 | struct lib_ring_buffer *buf, u64 tsc) | |
101 | { | |
102 | unsigned long tsc_shifted; | |
103 | ||
104 | if (config->tsc_bits == 0 || config->tsc_bits == 64) | |
105 | return 0; | |
106 | ||
107 | tsc_shifted = (unsigned long)(tsc >> config->tsc_bits); | |
108 | if (unlikely(tsc_shifted | |
109 | - (unsigned long)v_read(config, &buf->last_tsc))) | |
110 | return 1; | |
111 | else | |
112 | return 0; | |
113 | } | |
114 | #else | |
115 | static inline | |
116 | void save_last_tsc(const struct lib_ring_buffer_config *config, | |
117 | struct lib_ring_buffer *buf, u64 tsc) | |
118 | { | |
119 | if (config->tsc_bits == 0 || config->tsc_bits == 64) | |
120 | return; | |
121 | ||
122 | v_set(config, &buf->last_tsc, (unsigned long)tsc); | |
123 | } | |
124 | ||
125 | static inline | |
126 | int last_tsc_overflow(const struct lib_ring_buffer_config *config, | |
127 | struct lib_ring_buffer *buf, u64 tsc) | |
128 | { | |
129 | if (config->tsc_bits == 0 || config->tsc_bits == 64) | |
130 | return 0; | |
131 | ||
132 | if (unlikely((tsc - v_read(config, &buf->last_tsc)) | |
133 | >> config->tsc_bits)) | |
134 | return 1; | |
135 | else | |
136 | return 0; | |
137 | } | |
138 | #endif | |
139 | ||
140 | extern | |
141 | int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx); | |
142 | ||
143 | extern | |
144 | void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, | |
145 | enum switch_mode mode); | |
146 | ||
147 | /* Buffer write helpers */ | |
148 | ||
149 | static inline | |
150 | void lib_ring_buffer_reserve_push_reader(struct lib_ring_buffer *buf, | |
151 | struct channel *chan, | |
152 | unsigned long offset) | |
153 | { | |
154 | unsigned long consumed_old, consumed_new; | |
155 | ||
156 | do { | |
157 | consumed_old = atomic_long_read(&buf->consumed); | |
158 | /* | |
159 | * If buffer is in overwrite mode, push the reader consumed | |
160 | * count if the write position has reached it and we are not | |
161 | * at the first iteration (don't push the reader farther than | |
162 | * the writer). This operation can be done concurrently by many | |
163 | * writers in the same buffer, the writer being at the farthest | |
164 | * write position sub-buffer index in the buffer being the one | |
165 | * which will win this loop. | |
166 | */ | |
167 | if (unlikely(subbuf_trunc(offset, chan) | |
168 | - subbuf_trunc(consumed_old, chan) | |
169 | >= chan->backend.buf_size)) | |
170 | consumed_new = subbuf_align(consumed_old, chan); | |
171 | else | |
172 | return; | |
173 | } while (unlikely(atomic_long_cmpxchg(&buf->consumed, consumed_old, | |
174 | consumed_new) != consumed_old)); | |
175 | } | |
176 | ||
177 | static inline | |
178 | void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config, | |
179 | struct lib_ring_buffer *buf, | |
180 | unsigned long commit_count, | |
181 | unsigned long idx) | |
182 | { | |
183 | if (config->oops == RING_BUFFER_OOPS_CONSISTENCY) | |
184 | v_set(config, &buf->commit_hot[idx].seq, commit_count); | |
185 | } | |
186 | ||
187 | static inline | |
188 | int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config, | |
189 | struct lib_ring_buffer *buf, | |
190 | struct channel *chan) | |
191 | { | |
192 | unsigned long consumed_old, consumed_idx, commit_count, write_offset; | |
193 | ||
194 | consumed_old = atomic_long_read(&buf->consumed); | |
195 | consumed_idx = subbuf_index(consumed_old, chan); | |
196 | commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); | |
197 | /* | |
198 | * No memory barrier here, since we are only interested | |
199 | * in a statistically correct polling result. The next poll will | |
200 | * get the data is we are racing. The mb() that ensures correct | |
201 | * memory order is in get_subbuf. | |
202 | */ | |
203 | write_offset = v_read(config, &buf->offset); | |
204 | ||
205 | /* | |
206 | * Check that the subbuffer we are trying to consume has been | |
207 | * already fully committed. | |
208 | */ | |
209 | ||
210 | if (((commit_count - chan->backend.subbuf_size) | |
211 | & chan->commit_count_mask) | |
212 | - (buf_trunc(consumed_old, chan) | |
213 | >> chan->backend.num_subbuf_order) | |
214 | != 0) | |
215 | return 0; | |
216 | ||
217 | /* | |
218 | * Check that we are not about to read the same subbuffer in | |
219 | * which the writer head is. | |
220 | */ | |
221 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan) | |
222 | == 0) | |
223 | return 0; | |
224 | ||
225 | return 1; | |
226 | ||
227 | } | |
228 | ||
229 | static inline | |
230 | int lib_ring_buffer_pending_data(const struct lib_ring_buffer_config *config, | |
231 | struct lib_ring_buffer *buf, | |
232 | struct channel *chan) | |
233 | { | |
234 | return !!subbuf_offset(v_read(config, &buf->offset), chan); | |
235 | } | |
236 | ||
237 | static inline | |
238 | unsigned long lib_ring_buffer_get_data_size(const struct lib_ring_buffer_config *config, | |
239 | struct lib_ring_buffer *buf, | |
240 | unsigned long idx) | |
241 | { | |
242 | return subbuffer_get_data_size(config, &buf->backend, idx); | |
243 | } | |
244 | ||
245 | /* | |
246 | * Check if all space reservation in a buffer have been committed. This helps | |
247 | * knowing if an execution context is nested (for per-cpu buffers only). | |
248 | * This is a very specific ftrace use-case, so we keep this as "internal" API. | |
249 | */ | |
250 | static inline | |
251 | int lib_ring_buffer_reserve_committed(const struct lib_ring_buffer_config *config, | |
252 | struct lib_ring_buffer *buf, | |
253 | struct channel *chan) | |
254 | { | |
255 | unsigned long offset, idx, commit_count; | |
256 | ||
257 | CHAN_WARN_ON(chan, config->alloc != RING_BUFFER_ALLOC_PER_CPU); | |
258 | CHAN_WARN_ON(chan, config->sync != RING_BUFFER_SYNC_PER_CPU); | |
259 | ||
260 | /* | |
261 | * Read offset and commit count in a loop so they are both read | |
262 | * atomically wrt interrupts. By deal with interrupt concurrency by | |
263 | * restarting both reads if the offset has been pushed. Note that given | |
264 | * we only have to deal with interrupt concurrency here, an interrupt | |
265 | * modifying the commit count will also modify "offset", so it is safe | |
266 | * to only check for offset modifications. | |
267 | */ | |
268 | do { | |
269 | offset = v_read(config, &buf->offset); | |
270 | idx = subbuf_index(offset, chan); | |
271 | commit_count = v_read(config, &buf->commit_hot[idx].cc); | |
272 | } while (offset != v_read(config, &buf->offset)); | |
273 | ||
274 | return ((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order) | |
275 | - (commit_count & chan->commit_count_mask) == 0); | |
276 | } | |
277 | ||
278 | static inline | |
279 | void lib_ring_buffer_check_deliver(const struct lib_ring_buffer_config *config, | |
280 | struct lib_ring_buffer *buf, | |
281 | struct channel *chan, | |
282 | unsigned long offset, | |
283 | unsigned long commit_count, | |
284 | unsigned long idx) | |
285 | { | |
286 | unsigned long old_commit_count = commit_count | |
287 | - chan->backend.subbuf_size; | |
288 | u64 tsc; | |
289 | ||
290 | /* Check if all commits have been done */ | |
291 | if (unlikely((buf_trunc(offset, chan) >> chan->backend.num_subbuf_order) | |
292 | - (old_commit_count & chan->commit_count_mask) == 0)) { | |
293 | /* | |
294 | * If we succeeded at updating cc_sb below, we are the subbuffer | |
295 | * writer delivering the subbuffer. Deals with concurrent | |
296 | * updates of the "cc" value without adding a add_return atomic | |
297 | * operation to the fast path. | |
298 | * | |
299 | * We are doing the delivery in two steps: | |
300 | * - First, we cmpxchg() cc_sb to the new value | |
301 | * old_commit_count + 1. This ensures that we are the only | |
302 | * subbuffer user successfully filling the subbuffer, but we | |
303 | * do _not_ set the cc_sb value to "commit_count" yet. | |
304 | * Therefore, other writers that would wrap around the ring | |
305 | * buffer and try to start writing to our subbuffer would | |
306 | * have to drop records, because it would appear as | |
307 | * non-filled. | |
308 | * We therefore have exclusive access to the subbuffer control | |
309 | * structures. This mutual exclusion with other writers is | |
310 | * crucially important to perform record overruns count in | |
311 | * flight recorder mode locklessly. | |
312 | * - When we are ready to release the subbuffer (either for | |
313 | * reading or for overrun by other writers), we simply set the | |
314 | * cc_sb value to "commit_count" and perform delivery. | |
315 | * | |
316 | * The subbuffer size is least 2 bytes (minimum size: 1 page). | |
317 | * This guarantees that old_commit_count + 1 != commit_count. | |
318 | */ | |
319 | if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb, | |
320 | old_commit_count, old_commit_count + 1) | |
321 | == old_commit_count)) { | |
322 | /* | |
323 | * Start of exclusive subbuffer access. We are | |
324 | * guaranteed to be the last writer in this subbuffer | |
325 | * and any other writer trying to access this subbuffer | |
326 | * in this state is required to drop records. | |
327 | */ | |
328 | tsc = config->cb.ring_buffer_clock_read(chan); | |
329 | v_add(config, | |
330 | subbuffer_get_records_count(config, | |
331 | &buf->backend, idx), | |
332 | &buf->records_count); | |
333 | v_add(config, | |
334 | subbuffer_count_records_overrun(config, | |
335 | &buf->backend, | |
336 | idx), | |
337 | &buf->records_overrun); | |
338 | config->cb.buffer_end(buf, tsc, idx, | |
339 | lib_ring_buffer_get_data_size(config, | |
340 | buf, | |
341 | idx)); | |
342 | ||
343 | /* | |
344 | * Set noref flag and offset for this subbuffer id. | |
345 | * Contains a memory barrier that ensures counter stores | |
346 | * are ordered before set noref and offset. | |
347 | */ | |
348 | lib_ring_buffer_set_noref_offset(config, &buf->backend, idx, | |
349 | buf_trunc_val(offset, chan)); | |
350 | ||
351 | /* | |
352 | * Order set_noref and record counter updates before the | |
353 | * end of subbuffer exclusive access. Orders with | |
354 | * respect to writers coming into the subbuffer after | |
355 | * wrap around, and also order wrt concurrent readers. | |
356 | */ | |
357 | smp_mb(); | |
358 | /* End of exclusive subbuffer access */ | |
359 | v_set(config, &buf->commit_cold[idx].cc_sb, | |
360 | commit_count); | |
361 | lib_ring_buffer_vmcore_check_deliver(config, buf, | |
362 | commit_count, idx); | |
363 | ||
364 | /* | |
365 | * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free. | |
366 | */ | |
367 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER | |
368 | && atomic_long_read(&buf->active_readers) | |
369 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
370 | wake_up_interruptible(&buf->read_wait); | |
371 | wake_up_interruptible(&chan->read_wait); | |
372 | } | |
373 | ||
374 | } | |
375 | } | |
376 | } | |
377 | ||
378 | /* | |
379 | * lib_ring_buffer_write_commit_counter | |
380 | * | |
381 | * For flight recording. must be called after commit. | |
382 | * This function increments the subbuffer's commit_seq counter each time the | |
383 | * commit count reaches back the reserve offset (modulo subbuffer size). It is | |
384 | * useful for crash dump. | |
385 | */ | |
386 | static inline | |
387 | void lib_ring_buffer_write_commit_counter(const struct lib_ring_buffer_config *config, | |
388 | struct lib_ring_buffer *buf, | |
389 | struct channel *chan, | |
390 | unsigned long idx, | |
391 | unsigned long buf_offset, | |
392 | unsigned long commit_count, | |
393 | size_t slot_size) | |
394 | { | |
395 | unsigned long offset, commit_seq_old; | |
396 | ||
397 | if (config->oops != RING_BUFFER_OOPS_CONSISTENCY) | |
398 | return; | |
399 | ||
400 | offset = buf_offset + slot_size; | |
401 | ||
402 | /* | |
403 | * subbuf_offset includes commit_count_mask. We can simply | |
404 | * compare the offsets within the subbuffer without caring about | |
405 | * buffer full/empty mismatch because offset is never zero here | |
406 | * (subbuffer header and record headers have non-zero length). | |
407 | */ | |
408 | if (unlikely(subbuf_offset(offset - commit_count, chan))) | |
409 | return; | |
410 | ||
411 | commit_seq_old = v_read(config, &buf->commit_hot[idx].seq); | |
412 | while ((long) (commit_seq_old - commit_count) < 0) | |
413 | commit_seq_old = v_cmpxchg(config, &buf->commit_hot[idx].seq, | |
414 | commit_seq_old, commit_count); | |
415 | } | |
416 | ||
417 | extern int lib_ring_buffer_create(struct lib_ring_buffer *buf, | |
418 | struct channel_backend *chanb, int cpu); | |
419 | extern void lib_ring_buffer_free(struct lib_ring_buffer *buf); | |
420 | ||
421 | /* Keep track of trap nesting inside ring buffer code */ | |
422 | DECLARE_PER_CPU(unsigned int, lib_ring_buffer_nesting); | |
423 | ||
424 | #endif /* _LINUX_RING_BUFFER_FRONTEND_INTERNAL_H */ |