Commit | Line | Data |
---|---|---|
f3bc08c5 MD |
1 | /* |
2 | * ring_buffer_frontend.c | |
3 | * | |
886d51a3 MD |
4 | * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
5 | * | |
6 | * This library is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU Lesser General Public | |
8 | * License as published by the Free Software Foundation; only | |
9 | * version 2.1 of the License. | |
10 | * | |
11 | * This library is distributed in the hope that it will be useful, | |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
14 | * Lesser General Public License for more details. | |
15 | * | |
16 | * You should have received a copy of the GNU Lesser General Public | |
17 | * License along with this library; if not, write to the Free Software | |
18 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
19 | * | |
f3bc08c5 MD |
20 | * |
21 | * Ring buffer wait-free buffer synchronization. Producer-consumer and flight | |
22 | * recorder (overwrite) modes. See thesis: | |
23 | * | |
24 | * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D. | |
25 | * dissertation, Ecole Polytechnique de Montreal. | |
26 | * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf | |
27 | * | |
28 | * - Algorithm presentation in Chapter 5: | |
29 | * "Lockless Multi-Core High-Throughput Buffering". | |
30 | * - Algorithm formal verification in Section 8.6: | |
31 | * "Formal verification of LTTng" | |
32 | * | |
33 | * Author: | |
34 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
35 | * | |
36 | * Inspired from LTT and RelayFS: | |
37 | * Karim Yaghmour <karim@opersys.com> | |
38 | * Tom Zanussi <zanussi@us.ibm.com> | |
39 | * Bob Wisniewski <bob@watson.ibm.com> | |
40 | * And from K42 : | |
41 | * Bob Wisniewski <bob@watson.ibm.com> | |
42 | * | |
43 | * Buffer reader semantic : | |
44 | * | |
45 | * - get_subbuf_size | |
46 | * while buffer is not finalized and empty | |
47 | * - get_subbuf | |
48 | * - if return value != 0, continue | |
49 | * - splice one subbuffer worth of data to a pipe | |
50 | * - splice the data from pipe to disk/network | |
51 | * - put_subbuf | |
f3bc08c5 MD |
52 | */ |
53 | ||
54 | #include <linux/delay.h> | |
55 | #include <linux/module.h> | |
56 | #include <linux/percpu.h> | |
57 | ||
58 | #include "../../wrapper/ringbuffer/config.h" | |
59 | #include "../../wrapper/ringbuffer/backend.h" | |
60 | #include "../../wrapper/ringbuffer/frontend.h" | |
61 | #include "../../wrapper/ringbuffer/iterator.h" | |
62 | #include "../../wrapper/ringbuffer/nohz.h" | |
505fb410 | 63 | #include "../../wrapper/atomic.h" |
f3bc08c5 MD |
64 | |
65 | /* | |
66 | * Internal structure representing offsets to use at a sub-buffer switch. | |
67 | */ | |
68 | struct switch_offsets { | |
69 | unsigned long begin, end, old; | |
70 | size_t pre_header_padding, size; | |
f5ea5800 MD |
71 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, |
72 | switch_old_end:1; | |
f3bc08c5 MD |
73 | }; |
74 | ||
75 | #ifdef CONFIG_NO_HZ | |
76 | enum tick_nohz_val { | |
77 | TICK_NOHZ_STOP, | |
78 | TICK_NOHZ_FLUSH, | |
79 | TICK_NOHZ_RESTART, | |
80 | }; | |
81 | ||
82 | static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier); | |
83 | #endif /* CONFIG_NO_HZ */ | |
84 | ||
85 | static DEFINE_PER_CPU(spinlock_t, ring_buffer_nohz_lock); | |
86 | ||
87 | DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting); | |
88 | EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting); | |
89 | ||
90 | static | |
91 | void lib_ring_buffer_print_errors(struct channel *chan, | |
92 | struct lib_ring_buffer *buf, int cpu); | |
93 | ||
94 | /* | |
95 | * Must be called under cpu hotplug protection. | |
96 | */ | |
97 | void lib_ring_buffer_free(struct lib_ring_buffer *buf) | |
98 | { | |
99 | struct channel *chan = buf->backend.chan; | |
100 | ||
101 | lib_ring_buffer_print_errors(chan, buf, buf->backend.cpu); | |
102 | kfree(buf->commit_hot); | |
103 | kfree(buf->commit_cold); | |
104 | ||
105 | lib_ring_buffer_backend_free(&buf->backend); | |
106 | } | |
107 | ||
108 | /** | |
109 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
110 | * @buf: Ring buffer. | |
111 | * | |
112 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
113 | * used for writing. The ring buffer can be opened for reading, but the reader | |
114 | * should not be using the iterator concurrently with reset. The previous | |
115 | * current iterator record is reset. | |
116 | */ | |
117 | void lib_ring_buffer_reset(struct lib_ring_buffer *buf) | |
118 | { | |
119 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 120 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
121 | unsigned int i; |
122 | ||
123 | /* | |
124 | * Reset iterator first. It will put the subbuffer if it currently holds | |
125 | * it. | |
126 | */ | |
127 | lib_ring_buffer_iterator_reset(buf); | |
128 | v_set(config, &buf->offset, 0); | |
129 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
130 | v_set(config, &buf->commit_hot[i].cc, 0); | |
131 | v_set(config, &buf->commit_hot[i].seq, 0); | |
132 | v_set(config, &buf->commit_cold[i].cc_sb, 0); | |
133 | } | |
134 | atomic_long_set(&buf->consumed, 0); | |
135 | atomic_set(&buf->record_disabled, 0); | |
136 | v_set(config, &buf->last_tsc, 0); | |
137 | lib_ring_buffer_backend_reset(&buf->backend); | |
138 | /* Don't reset number of active readers */ | |
139 | v_set(config, &buf->records_lost_full, 0); | |
140 | v_set(config, &buf->records_lost_wrap, 0); | |
141 | v_set(config, &buf->records_lost_big, 0); | |
142 | v_set(config, &buf->records_count, 0); | |
143 | v_set(config, &buf->records_overrun, 0); | |
144 | buf->finalized = 0; | |
145 | } | |
146 | EXPORT_SYMBOL_GPL(lib_ring_buffer_reset); | |
147 | ||
148 | /** | |
149 | * channel_reset - Reset channel to initial values. | |
150 | * @chan: Channel. | |
151 | * | |
152 | * Effectively empty the channel. Should be called when the channel is not used | |
153 | * for writing. The channel can be opened for reading, but the reader should not | |
154 | * be using the iterator concurrently with reset. The previous current iterator | |
155 | * record is reset. | |
156 | */ | |
157 | void channel_reset(struct channel *chan) | |
158 | { | |
159 | /* | |
160 | * Reset iterators first. Will put the subbuffer if held for reading. | |
161 | */ | |
162 | channel_iterator_reset(chan); | |
163 | atomic_set(&chan->record_disabled, 0); | |
164 | /* Don't reset commit_count_mask, still valid */ | |
165 | channel_backend_reset(&chan->backend); | |
166 | /* Don't reset switch/read timer interval */ | |
167 | /* Don't reset notifiers and notifier enable bits */ | |
168 | /* Don't reset reader reference count */ | |
169 | } | |
170 | EXPORT_SYMBOL_GPL(channel_reset); | |
171 | ||
172 | /* | |
173 | * Must be called under cpu hotplug protection. | |
174 | */ | |
175 | int lib_ring_buffer_create(struct lib_ring_buffer *buf, | |
176 | struct channel_backend *chanb, int cpu) | |
177 | { | |
5a8fd222 | 178 | const struct lib_ring_buffer_config *config = &chanb->config; |
f3bc08c5 MD |
179 | struct channel *chan = container_of(chanb, struct channel, backend); |
180 | void *priv = chanb->priv; | |
f3bc08c5 MD |
181 | size_t subbuf_header_size; |
182 | u64 tsc; | |
183 | int ret; | |
184 | ||
185 | /* Test for cpu hotplug */ | |
186 | if (buf->backend.allocated) | |
187 | return 0; | |
188 | ||
189 | /* | |
190 | * Paranoia: per cpu dynamic allocation is not officially documented as | |
191 | * zeroing the memory, so let's do it here too, just in case. | |
192 | */ | |
193 | memset(buf, 0, sizeof(*buf)); | |
194 | ||
195 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, cpu); | |
196 | if (ret) | |
197 | return ret; | |
198 | ||
199 | buf->commit_hot = | |
200 | kzalloc_node(ALIGN(sizeof(*buf->commit_hot) | |
201 | * chan->backend.num_subbuf, | |
202 | 1 << INTERNODE_CACHE_SHIFT), | |
203 | GFP_KERNEL, cpu_to_node(max(cpu, 0))); | |
204 | if (!buf->commit_hot) { | |
205 | ret = -ENOMEM; | |
206 | goto free_chanbuf; | |
207 | } | |
208 | ||
209 | buf->commit_cold = | |
210 | kzalloc_node(ALIGN(sizeof(*buf->commit_cold) | |
211 | * chan->backend.num_subbuf, | |
212 | 1 << INTERNODE_CACHE_SHIFT), | |
213 | GFP_KERNEL, cpu_to_node(max(cpu, 0))); | |
214 | if (!buf->commit_cold) { | |
215 | ret = -ENOMEM; | |
216 | goto free_commit; | |
217 | } | |
218 | ||
f3bc08c5 | 219 | init_waitqueue_head(&buf->read_wait); |
71c1d843 | 220 | init_waitqueue_head(&buf->write_wait); |
f3bc08c5 MD |
221 | raw_spin_lock_init(&buf->raw_tick_nohz_spinlock); |
222 | ||
223 | /* | |
224 | * Write the subbuffer header for first subbuffer so we know the total | |
225 | * duration of data gathering. | |
226 | */ | |
227 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
228 | v_set(config, &buf->offset, subbuf_header_size); | |
229 | subbuffer_id_clear_noref(config, &buf->backend.buf_wsb[0].id); | |
230 | tsc = config->cb.ring_buffer_clock_read(buf->backend.chan); | |
231 | config->cb.buffer_begin(buf, tsc, 0); | |
232 | v_add(config, subbuf_header_size, &buf->commit_hot[0].cc); | |
233 | ||
234 | if (config->cb.buffer_create) { | |
235 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name); | |
236 | if (ret) | |
237 | goto free_init; | |
238 | } | |
239 | ||
240 | /* | |
241 | * Ensure the buffer is ready before setting it to allocated and setting | |
242 | * the cpumask. | |
243 | * Used for cpu hotplug vs cpumask iteration. | |
244 | */ | |
245 | smp_wmb(); | |
246 | buf->backend.allocated = 1; | |
247 | ||
248 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
249 | CHAN_WARN_ON(chan, cpumask_test_cpu(cpu, | |
250 | chan->backend.cpumask)); | |
251 | cpumask_set_cpu(cpu, chan->backend.cpumask); | |
252 | } | |
253 | ||
254 | return 0; | |
255 | ||
256 | /* Error handling */ | |
257 | free_init: | |
258 | kfree(buf->commit_cold); | |
259 | free_commit: | |
260 | kfree(buf->commit_hot); | |
261 | free_chanbuf: | |
262 | lib_ring_buffer_backend_free(&buf->backend); | |
263 | return ret; | |
264 | } | |
265 | ||
266 | static void switch_buffer_timer(unsigned long data) | |
267 | { | |
268 | struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data; | |
269 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 270 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
271 | |
272 | /* | |
273 | * Only flush buffers periodically if readers are active. | |
274 | */ | |
275 | if (atomic_long_read(&buf->active_readers)) | |
276 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
277 | ||
278 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
279 | mod_timer_pinned(&buf->switch_timer, | |
280 | jiffies + chan->switch_timer_interval); | |
281 | else | |
282 | mod_timer(&buf->switch_timer, | |
283 | jiffies + chan->switch_timer_interval); | |
284 | } | |
285 | ||
286 | /* | |
287 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
288 | */ | |
289 | static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer *buf) | |
290 | { | |
291 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 292 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
293 | |
294 | if (!chan->switch_timer_interval || buf->switch_timer_enabled) | |
295 | return; | |
296 | init_timer(&buf->switch_timer); | |
297 | buf->switch_timer.function = switch_buffer_timer; | |
298 | buf->switch_timer.expires = jiffies + chan->switch_timer_interval; | |
299 | buf->switch_timer.data = (unsigned long)buf; | |
300 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
301 | add_timer_on(&buf->switch_timer, buf->backend.cpu); | |
302 | else | |
303 | add_timer(&buf->switch_timer); | |
304 | buf->switch_timer_enabled = 1; | |
305 | } | |
306 | ||
307 | /* | |
308 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
309 | */ | |
310 | static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer *buf) | |
311 | { | |
312 | struct channel *chan = buf->backend.chan; | |
313 | ||
314 | if (!chan->switch_timer_interval || !buf->switch_timer_enabled) | |
315 | return; | |
316 | ||
317 | del_timer_sync(&buf->switch_timer); | |
318 | buf->switch_timer_enabled = 0; | |
319 | } | |
320 | ||
321 | /* | |
322 | * Polling timer to check the channels for data. | |
323 | */ | |
324 | static void read_buffer_timer(unsigned long data) | |
325 | { | |
326 | struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data; | |
327 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 328 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
329 | |
330 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
331 | ||
332 | if (atomic_long_read(&buf->active_readers) | |
333 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
334 | wake_up_interruptible(&buf->read_wait); | |
335 | wake_up_interruptible(&chan->read_wait); | |
336 | } | |
337 | ||
338 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
339 | mod_timer_pinned(&buf->read_timer, | |
340 | jiffies + chan->read_timer_interval); | |
341 | else | |
342 | mod_timer(&buf->read_timer, | |
343 | jiffies + chan->read_timer_interval); | |
344 | } | |
345 | ||
346 | /* | |
347 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
348 | */ | |
349 | static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf) | |
350 | { | |
351 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 352 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
353 | |
354 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
355 | || !chan->read_timer_interval | |
356 | || buf->read_timer_enabled) | |
357 | return; | |
358 | ||
359 | init_timer(&buf->read_timer); | |
360 | buf->read_timer.function = read_buffer_timer; | |
361 | buf->read_timer.expires = jiffies + chan->read_timer_interval; | |
362 | buf->read_timer.data = (unsigned long)buf; | |
363 | ||
364 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
365 | add_timer_on(&buf->read_timer, buf->backend.cpu); | |
366 | else | |
367 | add_timer(&buf->read_timer); | |
368 | buf->read_timer_enabled = 1; | |
369 | } | |
370 | ||
371 | /* | |
372 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
373 | */ | |
374 | static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer *buf) | |
375 | { | |
376 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 377 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
378 | |
379 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
380 | || !chan->read_timer_interval | |
381 | || !buf->read_timer_enabled) | |
382 | return; | |
383 | ||
384 | del_timer_sync(&buf->read_timer); | |
385 | /* | |
386 | * do one more check to catch data that has been written in the last | |
387 | * timer period. | |
388 | */ | |
389 | if (lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
390 | wake_up_interruptible(&buf->read_wait); | |
391 | wake_up_interruptible(&chan->read_wait); | |
392 | } | |
393 | buf->read_timer_enabled = 0; | |
394 | } | |
395 | ||
396 | #ifdef CONFIG_HOTPLUG_CPU | |
397 | /** | |
398 | * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback | |
399 | * @nb: notifier block | |
400 | * @action: hotplug action to take | |
401 | * @hcpu: CPU number | |
402 | * | |
403 | * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD) | |
404 | */ | |
405 | static | |
e8f071d5 | 406 | int lib_ring_buffer_cpu_hp_callback(struct notifier_block *nb, |
f3bc08c5 MD |
407 | unsigned long action, |
408 | void *hcpu) | |
409 | { | |
410 | unsigned int cpu = (unsigned long)hcpu; | |
411 | struct channel *chan = container_of(nb, struct channel, | |
412 | cpu_hp_notifier); | |
413 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
5a8fd222 | 414 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
415 | |
416 | if (!chan->cpu_hp_enable) | |
417 | return NOTIFY_DONE; | |
418 | ||
419 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
420 | ||
421 | switch (action) { | |
422 | case CPU_DOWN_FAILED: | |
423 | case CPU_DOWN_FAILED_FROZEN: | |
424 | case CPU_ONLINE: | |
425 | case CPU_ONLINE_FROZEN: | |
24cedcfe | 426 | wake_up_interruptible(&chan->hp_wait); |
f3bc08c5 MD |
427 | lib_ring_buffer_start_switch_timer(buf); |
428 | lib_ring_buffer_start_read_timer(buf); | |
429 | return NOTIFY_OK; | |
430 | ||
431 | case CPU_DOWN_PREPARE: | |
432 | case CPU_DOWN_PREPARE_FROZEN: | |
433 | lib_ring_buffer_stop_switch_timer(buf); | |
434 | lib_ring_buffer_stop_read_timer(buf); | |
435 | return NOTIFY_OK; | |
436 | ||
437 | case CPU_DEAD: | |
438 | case CPU_DEAD_FROZEN: | |
439 | /* | |
440 | * Performing a buffer switch on a remote CPU. Performed by | |
441 | * the CPU responsible for doing the hotunplug after the target | |
442 | * CPU stopped running completely. Ensures that all data | |
443 | * from that remote CPU is flushed. | |
444 | */ | |
445 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
446 | return NOTIFY_OK; | |
447 | ||
448 | default: | |
449 | return NOTIFY_DONE; | |
450 | } | |
451 | } | |
452 | #endif | |
453 | ||
23b908b0 | 454 | #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) |
f3bc08c5 MD |
455 | /* |
456 | * For per-cpu buffers, call the reader wakeups before switching the buffer, so | |
457 | * that wake-up-tracing generated events are flushed before going idle (in | |
458 | * tick_nohz). We test if the spinlock is locked to deal with the race where | |
459 | * readers try to sample the ring buffer before we perform the switch. We let | |
460 | * the readers retry in that case. If there is data in the buffer, the wake up | |
461 | * is going to forbid the CPU running the reader thread from going idle. | |
462 | */ | |
463 | static int notrace ring_buffer_tick_nohz_callback(struct notifier_block *nb, | |
464 | unsigned long val, | |
465 | void *data) | |
466 | { | |
467 | struct channel *chan = container_of(nb, struct channel, | |
468 | tick_nohz_notifier); | |
5a8fd222 | 469 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
470 | struct lib_ring_buffer *buf; |
471 | int cpu = smp_processor_id(); | |
472 | ||
473 | if (config->alloc != RING_BUFFER_ALLOC_PER_CPU) { | |
474 | /* | |
475 | * We don't support keeping the system idle with global buffers | |
476 | * and streaming active. In order to do so, we would need to | |
477 | * sample a non-nohz-cpumask racelessly with the nohz updates | |
478 | * without adding synchronization overhead to nohz. Leave this | |
479 | * use-case out for now. | |
480 | */ | |
481 | return 0; | |
482 | } | |
483 | ||
484 | buf = channel_get_ring_buffer(config, chan, cpu); | |
485 | switch (val) { | |
486 | case TICK_NOHZ_FLUSH: | |
487 | raw_spin_lock(&buf->raw_tick_nohz_spinlock); | |
488 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_TIMER | |
489 | && chan->read_timer_interval | |
490 | && atomic_long_read(&buf->active_readers) | |
491 | && (lib_ring_buffer_poll_deliver(config, buf, chan) | |
492 | || lib_ring_buffer_pending_data(config, buf, chan))) { | |
493 | wake_up_interruptible(&buf->read_wait); | |
494 | wake_up_interruptible(&chan->read_wait); | |
495 | } | |
496 | if (chan->switch_timer_interval) | |
497 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
498 | raw_spin_unlock(&buf->raw_tick_nohz_spinlock); | |
499 | break; | |
500 | case TICK_NOHZ_STOP: | |
501 | spin_lock(&__get_cpu_var(ring_buffer_nohz_lock)); | |
502 | lib_ring_buffer_stop_switch_timer(buf); | |
503 | lib_ring_buffer_stop_read_timer(buf); | |
504 | spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock)); | |
505 | break; | |
506 | case TICK_NOHZ_RESTART: | |
507 | spin_lock(&__get_cpu_var(ring_buffer_nohz_lock)); | |
508 | lib_ring_buffer_start_read_timer(buf); | |
509 | lib_ring_buffer_start_switch_timer(buf); | |
510 | spin_unlock(&__get_cpu_var(ring_buffer_nohz_lock)); | |
511 | break; | |
512 | } | |
513 | ||
514 | return 0; | |
515 | } | |
516 | ||
517 | void notrace lib_ring_buffer_tick_nohz_flush(void) | |
518 | { | |
519 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_FLUSH, | |
520 | NULL); | |
521 | } | |
522 | ||
523 | void notrace lib_ring_buffer_tick_nohz_stop(void) | |
524 | { | |
525 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_STOP, | |
526 | NULL); | |
527 | } | |
528 | ||
529 | void notrace lib_ring_buffer_tick_nohz_restart(void) | |
530 | { | |
531 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_RESTART, | |
532 | NULL); | |
533 | } | |
23b908b0 | 534 | #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */ |
f3bc08c5 MD |
535 | |
536 | /* | |
537 | * Holds CPU hotplug. | |
538 | */ | |
539 | static void channel_unregister_notifiers(struct channel *chan) | |
540 | { | |
5a8fd222 | 541 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
542 | int cpu; |
543 | ||
544 | channel_iterator_unregister_notifiers(chan); | |
545 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
546 | #ifdef CONFIG_NO_HZ | |
547 | /* | |
548 | * Remove the nohz notifier first, so we are certain we stop | |
549 | * the timers. | |
550 | */ | |
551 | atomic_notifier_chain_unregister(&tick_nohz_notifier, | |
552 | &chan->tick_nohz_notifier); | |
553 | /* | |
554 | * ring_buffer_nohz_lock will not be needed below, because | |
555 | * we just removed the notifiers, which were the only source of | |
556 | * concurrency. | |
557 | */ | |
558 | #endif /* CONFIG_NO_HZ */ | |
559 | #ifdef CONFIG_HOTPLUG_CPU | |
560 | get_online_cpus(); | |
561 | chan->cpu_hp_enable = 0; | |
562 | for_each_online_cpu(cpu) { | |
563 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
564 | cpu); | |
565 | lib_ring_buffer_stop_switch_timer(buf); | |
566 | lib_ring_buffer_stop_read_timer(buf); | |
567 | } | |
568 | put_online_cpus(); | |
569 | unregister_cpu_notifier(&chan->cpu_hp_notifier); | |
570 | #else | |
571 | for_each_possible_cpu(cpu) { | |
572 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
573 | cpu); | |
574 | lib_ring_buffer_stop_switch_timer(buf); | |
575 | lib_ring_buffer_stop_read_timer(buf); | |
576 | } | |
577 | #endif | |
578 | } else { | |
579 | struct lib_ring_buffer *buf = chan->backend.buf; | |
580 | ||
581 | lib_ring_buffer_stop_switch_timer(buf); | |
582 | lib_ring_buffer_stop_read_timer(buf); | |
583 | } | |
584 | channel_backend_unregister_notifiers(&chan->backend); | |
585 | } | |
586 | ||
587 | static void channel_free(struct channel *chan) | |
588 | { | |
dd5a0db3 MD |
589 | if (chan->backend.release_priv_ops) { |
590 | chan->backend.release_priv_ops(chan->backend.priv_ops); | |
591 | } | |
f3bc08c5 MD |
592 | channel_iterator_free(chan); |
593 | channel_backend_free(&chan->backend); | |
594 | kfree(chan); | |
595 | } | |
596 | ||
597 | /** | |
598 | * channel_create - Create channel. | |
599 | * @config: ring buffer instance configuration | |
600 | * @name: name of the channel | |
601 | * @priv: ring buffer client private data | |
602 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous | |
603 | * address mapping. It is used only by RING_BUFFER_STATIC | |
604 | * configuration. It can be set to NULL for other backends. | |
605 | * @subbuf_size: subbuffer size | |
606 | * @num_subbuf: number of subbuffers | |
607 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
608 | * padding to let readers get those sub-buffers. | |
609 | * Used for live streaming. | |
610 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
611 | * | |
612 | * Holds cpu hotplug. | |
613 | * Returns NULL on failure. | |
614 | */ | |
615 | struct channel *channel_create(const struct lib_ring_buffer_config *config, | |
616 | const char *name, void *priv, void *buf_addr, | |
617 | size_t subbuf_size, | |
618 | size_t num_subbuf, unsigned int switch_timer_interval, | |
619 | unsigned int read_timer_interval) | |
620 | { | |
621 | int ret, cpu; | |
622 | struct channel *chan; | |
623 | ||
624 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
625 | read_timer_interval)) | |
626 | return NULL; | |
627 | ||
628 | chan = kzalloc(sizeof(struct channel), GFP_KERNEL); | |
629 | if (!chan) | |
630 | return NULL; | |
631 | ||
632 | ret = channel_backend_init(&chan->backend, name, config, priv, | |
633 | subbuf_size, num_subbuf); | |
634 | if (ret) | |
635 | goto error; | |
636 | ||
637 | ret = channel_iterator_init(chan); | |
638 | if (ret) | |
639 | goto error_free_backend; | |
640 | ||
641 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
642 | chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
643 | chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
f40270ad | 644 | kref_init(&chan->ref); |
f3bc08c5 | 645 | init_waitqueue_head(&chan->read_wait); |
24cedcfe | 646 | init_waitqueue_head(&chan->hp_wait); |
f3bc08c5 MD |
647 | |
648 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
23b908b0 | 649 | #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) |
f3bc08c5 MD |
650 | /* Only benefit from NO_HZ idle with per-cpu buffers for now. */ |
651 | chan->tick_nohz_notifier.notifier_call = | |
652 | ring_buffer_tick_nohz_callback; | |
653 | chan->tick_nohz_notifier.priority = ~0U; | |
654 | atomic_notifier_chain_register(&tick_nohz_notifier, | |
655 | &chan->tick_nohz_notifier); | |
23b908b0 | 656 | #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */ |
f3bc08c5 MD |
657 | |
658 | /* | |
659 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
660 | * in early initcall, it will not be notified of secondary cpus. | |
661 | * In that off case, we need to allocate for all possible cpus. | |
662 | */ | |
663 | #ifdef CONFIG_HOTPLUG_CPU | |
664 | chan->cpu_hp_notifier.notifier_call = | |
665 | lib_ring_buffer_cpu_hp_callback; | |
666 | chan->cpu_hp_notifier.priority = 6; | |
667 | register_cpu_notifier(&chan->cpu_hp_notifier); | |
668 | ||
669 | get_online_cpus(); | |
670 | for_each_online_cpu(cpu) { | |
671 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
672 | cpu); | |
673 | spin_lock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
674 | lib_ring_buffer_start_switch_timer(buf); | |
675 | lib_ring_buffer_start_read_timer(buf); | |
676 | spin_unlock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
677 | } | |
678 | chan->cpu_hp_enable = 1; | |
679 | put_online_cpus(); | |
680 | #else | |
681 | for_each_possible_cpu(cpu) { | |
682 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
683 | cpu); | |
684 | spin_lock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
685 | lib_ring_buffer_start_switch_timer(buf); | |
686 | lib_ring_buffer_start_read_timer(buf); | |
687 | spin_unlock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
688 | } | |
689 | #endif | |
690 | } else { | |
691 | struct lib_ring_buffer *buf = chan->backend.buf; | |
692 | ||
693 | lib_ring_buffer_start_switch_timer(buf); | |
694 | lib_ring_buffer_start_read_timer(buf); | |
695 | } | |
696 | ||
697 | return chan; | |
698 | ||
699 | error_free_backend: | |
700 | channel_backend_free(&chan->backend); | |
701 | error: | |
702 | kfree(chan); | |
703 | return NULL; | |
704 | } | |
705 | EXPORT_SYMBOL_GPL(channel_create); | |
706 | ||
f40270ad MD |
707 | static |
708 | void channel_release(struct kref *kref) | |
709 | { | |
710 | struct channel *chan = container_of(kref, struct channel, ref); | |
711 | channel_free(chan); | |
712 | } | |
713 | ||
f3bc08c5 MD |
714 | /** |
715 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
716 | * @chan: channel to destroy | |
717 | * | |
718 | * Holds cpu hotplug. | |
9a0df743 MD |
719 | * Call "destroy" callback, finalize channels, and then decrement the |
720 | * channel reference count. Note that when readers have completed data | |
721 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
722 | * They should release their handle at that point. Returns the private | |
723 | * data pointer. | |
f3bc08c5 MD |
724 | */ |
725 | void *channel_destroy(struct channel *chan) | |
726 | { | |
727 | int cpu; | |
5a8fd222 | 728 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
729 | void *priv; |
730 | ||
731 | channel_unregister_notifiers(chan); | |
732 | ||
733 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
734 | /* | |
735 | * No need to hold cpu hotplug, because all notifiers have been | |
736 | * unregistered. | |
737 | */ | |
738 | for_each_channel_cpu(cpu, chan) { | |
739 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
740 | cpu); | |
741 | ||
742 | if (config->cb.buffer_finalize) | |
743 | config->cb.buffer_finalize(buf, | |
744 | chan->backend.priv, | |
745 | cpu); | |
746 | if (buf->backend.allocated) | |
747 | lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH); | |
748 | /* | |
749 | * Perform flush before writing to finalized. | |
750 | */ | |
751 | smp_wmb(); | |
752 | ACCESS_ONCE(buf->finalized) = 1; | |
753 | wake_up_interruptible(&buf->read_wait); | |
754 | } | |
755 | } else { | |
756 | struct lib_ring_buffer *buf = chan->backend.buf; | |
757 | ||
758 | if (config->cb.buffer_finalize) | |
759 | config->cb.buffer_finalize(buf, chan->backend.priv, -1); | |
760 | if (buf->backend.allocated) | |
761 | lib_ring_buffer_switch_slow(buf, SWITCH_FLUSH); | |
762 | /* | |
763 | * Perform flush before writing to finalized. | |
764 | */ | |
765 | smp_wmb(); | |
766 | ACCESS_ONCE(buf->finalized) = 1; | |
767 | wake_up_interruptible(&buf->read_wait); | |
768 | } | |
24cedcfe MD |
769 | ACCESS_ONCE(chan->finalized) = 1; |
770 | wake_up_interruptible(&chan->hp_wait); | |
f3bc08c5 | 771 | wake_up_interruptible(&chan->read_wait); |
f3bc08c5 | 772 | priv = chan->backend.priv; |
ba1d61bc | 773 | kref_put(&chan->ref, channel_release); |
f3bc08c5 MD |
774 | return priv; |
775 | } | |
776 | EXPORT_SYMBOL_GPL(channel_destroy); | |
777 | ||
778 | struct lib_ring_buffer *channel_get_ring_buffer( | |
779 | const struct lib_ring_buffer_config *config, | |
780 | struct channel *chan, int cpu) | |
781 | { | |
782 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) | |
783 | return chan->backend.buf; | |
784 | else | |
785 | return per_cpu_ptr(chan->backend.buf, cpu); | |
786 | } | |
787 | EXPORT_SYMBOL_GPL(channel_get_ring_buffer); | |
788 | ||
789 | int lib_ring_buffer_open_read(struct lib_ring_buffer *buf) | |
790 | { | |
791 | struct channel *chan = buf->backend.chan; | |
792 | ||
793 | if (!atomic_long_add_unless(&buf->active_readers, 1, 1)) | |
794 | return -EBUSY; | |
f40270ad | 795 | kref_get(&chan->ref); |
505fb410 | 796 | lttng_smp_mb__after_atomic(); |
f3bc08c5 MD |
797 | return 0; |
798 | } | |
799 | EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read); | |
800 | ||
801 | void lib_ring_buffer_release_read(struct lib_ring_buffer *buf) | |
802 | { | |
803 | struct channel *chan = buf->backend.chan; | |
804 | ||
805 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
505fb410 | 806 | lttng_smp_mb__before_atomic(); |
f3bc08c5 | 807 | atomic_long_dec(&buf->active_readers); |
f40270ad | 808 | kref_put(&chan->ref, channel_release); |
f3bc08c5 MD |
809 | } |
810 | EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read); | |
811 | ||
812 | /* | |
813 | * Promote compiler barrier to a smp_mb(). | |
814 | * For the specific ring buffer case, this IPI call should be removed if the | |
815 | * architecture does not reorder writes. This should eventually be provided by | |
816 | * a separate architecture-specific infrastructure. | |
817 | */ | |
818 | static void remote_mb(void *info) | |
819 | { | |
820 | smp_mb(); | |
821 | } | |
822 | ||
823 | /** | |
824 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
825 | * @buf: ring buffer | |
826 | * @consumed: consumed count indicating the position where to read | |
827 | * @produced: produced count, indicates position when to stop reading | |
828 | * | |
829 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
830 | * data to read at consumed position, or 0 if the get operation succeeds. | |
831 | * Busy-loop trying to get data if the tick_nohz sequence lock is held. | |
832 | */ | |
833 | ||
834 | int lib_ring_buffer_snapshot(struct lib_ring_buffer *buf, | |
835 | unsigned long *consumed, unsigned long *produced) | |
836 | { | |
837 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 838 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
839 | unsigned long consumed_cur, write_offset; |
840 | int finalized; | |
841 | ||
842 | retry: | |
843 | finalized = ACCESS_ONCE(buf->finalized); | |
844 | /* | |
845 | * Read finalized before counters. | |
846 | */ | |
847 | smp_rmb(); | |
848 | consumed_cur = atomic_long_read(&buf->consumed); | |
849 | /* | |
850 | * No need to issue a memory barrier between consumed count read and | |
851 | * write offset read, because consumed count can only change | |
852 | * concurrently in overwrite mode, and we keep a sequence counter | |
853 | * identifier derived from the write offset to check we are getting | |
854 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
855 | * "tagged" upon writes, tags are checked upon read). | |
856 | */ | |
857 | write_offset = v_read(config, &buf->offset); | |
858 | ||
859 | /* | |
860 | * Check that we are not about to read the same subbuffer in | |
861 | * which the writer head is. | |
862 | */ | |
863 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
864 | == 0) | |
865 | goto nodata; | |
866 | ||
867 | *consumed = consumed_cur; | |
868 | *produced = subbuf_trunc(write_offset, chan); | |
869 | ||
870 | return 0; | |
871 | ||
872 | nodata: | |
873 | /* | |
874 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
875 | * of "raw_spin_is_locked" memory ordering. | |
876 | */ | |
877 | if (finalized) | |
878 | return -ENODATA; | |
879 | else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock)) | |
880 | goto retry; | |
881 | else | |
882 | return -EAGAIN; | |
883 | } | |
884 | EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot); | |
885 | ||
886 | /** | |
887 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
71c1d843 MD |
888 | * |
889 | * Should only be called from consumer context. | |
f3bc08c5 MD |
890 | * @buf: ring buffer |
891 | * @consumed_new: new consumed count value | |
892 | */ | |
893 | void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf, | |
894 | unsigned long consumed_new) | |
895 | { | |
896 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
897 | struct channel *chan = bufb->chan; | |
898 | unsigned long consumed; | |
899 | ||
900 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
901 | ||
902 | /* | |
903 | * Only push the consumed value forward. | |
904 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
905 | * the writer in flight recorder mode. | |
906 | */ | |
907 | consumed = atomic_long_read(&buf->consumed); | |
908 | while ((long) consumed - (long) consumed_new < 0) | |
909 | consumed = atomic_long_cmpxchg(&buf->consumed, consumed, | |
910 | consumed_new); | |
71c1d843 MD |
911 | /* Wake-up the metadata producer */ |
912 | wake_up_interruptible(&buf->write_wait); | |
f3bc08c5 MD |
913 | } |
914 | EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer); | |
915 | ||
916 | /** | |
917 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
918 | * @buf: ring buffer | |
919 | * @consumed: consumed count indicating the position where to read | |
920 | * | |
921 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
922 | * data to read at consumed position, or 0 if the get operation succeeds. | |
923 | * Busy-loop trying to get data if the tick_nohz sequence lock is held. | |
924 | */ | |
925 | int lib_ring_buffer_get_subbuf(struct lib_ring_buffer *buf, | |
926 | unsigned long consumed) | |
927 | { | |
928 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 929 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
930 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
931 | int ret; | |
932 | int finalized; | |
933 | ||
8202b8a0 MD |
934 | if (buf->get_subbuf) { |
935 | /* | |
936 | * Reader is trying to get a subbuffer twice. | |
937 | */ | |
938 | CHAN_WARN_ON(chan, 1); | |
939 | return -EBUSY; | |
940 | } | |
f3bc08c5 MD |
941 | retry: |
942 | finalized = ACCESS_ONCE(buf->finalized); | |
943 | /* | |
944 | * Read finalized before counters. | |
945 | */ | |
946 | smp_rmb(); | |
947 | consumed_cur = atomic_long_read(&buf->consumed); | |
948 | consumed_idx = subbuf_index(consumed, chan); | |
949 | commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); | |
950 | /* | |
951 | * Make sure we read the commit count before reading the buffer | |
952 | * data and the write offset. Correct consumed offset ordering | |
953 | * wrt commit count is insured by the use of cmpxchg to update | |
954 | * the consumed offset. | |
955 | * smp_call_function_single can fail if the remote CPU is offline, | |
956 | * this is OK because then there is no wmb to execute there. | |
957 | * If our thread is executing on the same CPU as the on the buffers | |
958 | * belongs to, we don't have to synchronize it at all. If we are | |
959 | * migrated, the scheduler will take care of the memory barriers. | |
960 | * Normally, smp_call_function_single() should ensure program order when | |
961 | * executing the remote function, which implies that it surrounds the | |
962 | * function execution with : | |
963 | * smp_mb() | |
964 | * send IPI | |
965 | * csd_lock_wait | |
966 | * recv IPI | |
967 | * smp_mb() | |
968 | * exec. function | |
969 | * smp_mb() | |
970 | * csd unlock | |
971 | * smp_mb() | |
972 | * | |
973 | * However, smp_call_function_single() does not seem to clearly execute | |
974 | * such barriers. It depends on spinlock semantic to provide the barrier | |
975 | * before executing the IPI and, when busy-looping, csd_lock_wait only | |
976 | * executes smp_mb() when it has to wait for the other CPU. | |
977 | * | |
978 | * I don't trust this code. Therefore, let's add the smp_mb() sequence | |
979 | * required ourself, even if duplicated. It has no performance impact | |
980 | * anyway. | |
981 | * | |
982 | * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs | |
983 | * read and write vs write. They do not ensure core synchronization. We | |
984 | * really have to ensure total order between the 3 barriers running on | |
985 | * the 2 CPUs. | |
986 | */ | |
987 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
988 | if (config->sync == RING_BUFFER_SYNC_PER_CPU | |
989 | && config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
990 | if (raw_smp_processor_id() != buf->backend.cpu) { | |
991 | /* Total order with IPI handler smp_mb() */ | |
992 | smp_mb(); | |
993 | smp_call_function_single(buf->backend.cpu, | |
994 | remote_mb, NULL, 1); | |
995 | /* Total order with IPI handler smp_mb() */ | |
996 | smp_mb(); | |
997 | } | |
998 | } else { | |
999 | /* Total order with IPI handler smp_mb() */ | |
1000 | smp_mb(); | |
1001 | smp_call_function(remote_mb, NULL, 1); | |
1002 | /* Total order with IPI handler smp_mb() */ | |
1003 | smp_mb(); | |
1004 | } | |
1005 | } else { | |
1006 | /* | |
1007 | * Local rmb to match the remote wmb to read the commit count | |
1008 | * before the buffer data and the write offset. | |
1009 | */ | |
1010 | smp_rmb(); | |
1011 | } | |
1012 | ||
1013 | write_offset = v_read(config, &buf->offset); | |
1014 | ||
1015 | /* | |
1016 | * Check that the buffer we are getting is after or at consumed_cur | |
1017 | * position. | |
1018 | */ | |
1019 | if ((long) subbuf_trunc(consumed, chan) | |
1020 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
1021 | goto nodata; | |
1022 | ||
1023 | /* | |
1024 | * Check that the subbuffer we are trying to consume has been | |
1025 | * already fully committed. | |
1026 | */ | |
1027 | if (((commit_count - chan->backend.subbuf_size) | |
1028 | & chan->commit_count_mask) | |
c9b3b5e2 | 1029 | - (buf_trunc(consumed, chan) |
f3bc08c5 MD |
1030 | >> chan->backend.num_subbuf_order) |
1031 | != 0) | |
1032 | goto nodata; | |
1033 | ||
1034 | /* | |
1035 | * Check that we are not about to read the same subbuffer in | |
1036 | * which the writer head is. | |
1037 | */ | |
c9b3b5e2 | 1038 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed, chan) |
f3bc08c5 MD |
1039 | == 0) |
1040 | goto nodata; | |
1041 | ||
1042 | /* | |
1043 | * Failure to get the subbuffer causes a busy-loop retry without going | |
1044 | * to a wait queue. These are caused by short-lived race windows where | |
1045 | * the writer is getting access to a subbuffer we were trying to get | |
1046 | * access to. Also checks that the "consumed" buffer count we are | |
1047 | * looking for matches the one contained in the subbuffer id. | |
1048 | */ | |
1049 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1050 | consumed_idx, buf_trunc_val(consumed, chan)); | |
1051 | if (ret) | |
1052 | goto retry; | |
1053 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
1054 | ||
1055 | buf->get_subbuf_consumed = consumed; | |
1056 | buf->get_subbuf = 1; | |
1057 | ||
1058 | return 0; | |
1059 | ||
1060 | nodata: | |
1061 | /* | |
1062 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1063 | * of "raw_spin_is_locked" memory ordering. | |
1064 | */ | |
1065 | if (finalized) | |
1066 | return -ENODATA; | |
1067 | else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock)) | |
1068 | goto retry; | |
1069 | else | |
1070 | return -EAGAIN; | |
1071 | } | |
1072 | EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf); | |
1073 | ||
1074 | /** | |
1075 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
1076 | * @buf: ring buffer | |
1077 | */ | |
1078 | void lib_ring_buffer_put_subbuf(struct lib_ring_buffer *buf) | |
1079 | { | |
1080 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1081 | struct channel *chan = bufb->chan; | |
5a8fd222 | 1082 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1083 | unsigned long read_sb_bindex, consumed_idx, consumed; |
1084 | ||
1085 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
1086 | ||
1087 | if (!buf->get_subbuf) { | |
1088 | /* | |
1089 | * Reader puts a subbuffer it did not get. | |
1090 | */ | |
1091 | CHAN_WARN_ON(chan, 1); | |
1092 | return; | |
1093 | } | |
1094 | consumed = buf->get_subbuf_consumed; | |
1095 | buf->get_subbuf = 0; | |
1096 | ||
1097 | /* | |
1098 | * Clear the records_unread counter. (overruns counter) | |
1099 | * Can still be non-zero if a file reader simply grabbed the data | |
1100 | * without using iterators. | |
1101 | * Can be below zero if an iterator is used on a snapshot more than | |
1102 | * once. | |
1103 | */ | |
1104 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
1105 | v_add(config, v_read(config, | |
1106 | &bufb->array[read_sb_bindex]->records_unread), | |
1107 | &bufb->records_read); | |
1108 | v_set(config, &bufb->array[read_sb_bindex]->records_unread, 0); | |
1109 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE | |
1110 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
1111 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
1112 | ||
1113 | /* | |
1114 | * Exchange the reader subbuffer with the one we put in its place in the | |
1115 | * writer subbuffer table. Expect the original consumed count. If | |
1116 | * update_read_sb_index fails, this is because the writer updated the | |
1117 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
1118 | * currently have: it has become invalid to try reading this sub-buffer | |
1119 | * consumed count value anyway. | |
1120 | */ | |
1121 | consumed_idx = subbuf_index(consumed, chan); | |
1122 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1123 | consumed_idx, buf_trunc_val(consumed, chan)); | |
1124 | /* | |
1125 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
1126 | * if the writer concurrently updated it. | |
1127 | */ | |
1128 | } | |
1129 | EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf); | |
1130 | ||
1131 | /* | |
1132 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
1133 | * position and the writer position. (inclusive) | |
1134 | */ | |
1135 | static | |
1136 | void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer *buf, | |
1137 | struct channel *chan, | |
1138 | unsigned long cons_offset, | |
1139 | int cpu) | |
1140 | { | |
5a8fd222 | 1141 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1142 | unsigned long cons_idx, commit_count, commit_count_sb; |
1143 | ||
1144 | cons_idx = subbuf_index(cons_offset, chan); | |
1145 | commit_count = v_read(config, &buf->commit_hot[cons_idx].cc); | |
1146 | commit_count_sb = v_read(config, &buf->commit_cold[cons_idx].cc_sb); | |
1147 | ||
1148 | if (subbuf_offset(commit_count, chan) != 0) | |
1149 | printk(KERN_WARNING | |
1150 | "ring buffer %s, cpu %d: " | |
1151 | "commit count in subbuffer %lu,\n" | |
1152 | "expecting multiples of %lu bytes\n" | |
1153 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
1154 | chan->backend.name, cpu, cons_idx, | |
1155 | chan->backend.subbuf_size, | |
1156 | commit_count, commit_count_sb); | |
1157 | ||
1158 | printk(KERN_DEBUG "ring buffer: %s, cpu %d: %lu bytes committed\n", | |
1159 | chan->backend.name, cpu, commit_count); | |
1160 | } | |
1161 | ||
1162 | static | |
1163 | void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf, | |
1164 | struct channel *chan, | |
1165 | void *priv, int cpu) | |
1166 | { | |
5a8fd222 | 1167 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1168 | unsigned long write_offset, cons_offset; |
1169 | ||
f3bc08c5 MD |
1170 | /* |
1171 | * No need to order commit_count, write_offset and cons_offset reads | |
1172 | * because we execute at teardown when no more writer nor reader | |
1173 | * references are left. | |
1174 | */ | |
1175 | write_offset = v_read(config, &buf->offset); | |
1176 | cons_offset = atomic_long_read(&buf->consumed); | |
1177 | if (write_offset != cons_offset) | |
05aad775 | 1178 | printk(KERN_DEBUG |
f3bc08c5 MD |
1179 | "ring buffer %s, cpu %d: " |
1180 | "non-consumed data\n" | |
1181 | " [ %lu bytes written, %lu bytes read ]\n", | |
1182 | chan->backend.name, cpu, write_offset, cons_offset); | |
1183 | ||
1184 | for (cons_offset = atomic_long_read(&buf->consumed); | |
1185 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), | |
1186 | chan) | |
1187 | - cons_offset) > 0; | |
1188 | cons_offset = subbuf_align(cons_offset, chan)) | |
1189 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1190 | cpu); | |
1191 | } | |
1192 | ||
1193 | static | |
1194 | void lib_ring_buffer_print_errors(struct channel *chan, | |
1195 | struct lib_ring_buffer *buf, int cpu) | |
1196 | { | |
5a8fd222 | 1197 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1198 | void *priv = chan->backend.priv; |
1199 | ||
ec01ec93 MD |
1200 | if (!strcmp(chan->backend.name, "relay-metadata")) { |
1201 | printk(KERN_DEBUG "ring buffer %s: %lu records written, " | |
1202 | "%lu records overrun\n", | |
1203 | chan->backend.name, | |
1204 | v_read(config, &buf->records_count), | |
1205 | v_read(config, &buf->records_overrun)); | |
1206 | } else { | |
1207 | printk(KERN_DEBUG "ring buffer %s, cpu %d: %lu records written, " | |
1208 | "%lu records overrun\n", | |
1209 | chan->backend.name, cpu, | |
1210 | v_read(config, &buf->records_count), | |
1211 | v_read(config, &buf->records_overrun)); | |
1212 | ||
1213 | if (v_read(config, &buf->records_lost_full) | |
1214 | || v_read(config, &buf->records_lost_wrap) | |
1215 | || v_read(config, &buf->records_lost_big)) | |
1216 | printk(KERN_WARNING | |
1217 | "ring buffer %s, cpu %d: records were lost. Caused by:\n" | |
1218 | " [ %lu buffer full, %lu nest buffer wrap-around, " | |
1219 | "%lu event too big ]\n", | |
1220 | chan->backend.name, cpu, | |
1221 | v_read(config, &buf->records_lost_full), | |
1222 | v_read(config, &buf->records_lost_wrap), | |
1223 | v_read(config, &buf->records_lost_big)); | |
1224 | } | |
f3bc08c5 MD |
1225 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu); |
1226 | } | |
1227 | ||
1228 | /* | |
1229 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1230 | * | |
1231 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. | |
1232 | */ | |
1233 | static | |
1234 | void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf, | |
1235 | struct channel *chan, | |
1236 | struct switch_offsets *offsets, | |
1237 | u64 tsc) | |
1238 | { | |
5a8fd222 | 1239 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1240 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1241 | unsigned long commit_count; | |
1242 | ||
1243 | config->cb.buffer_begin(buf, tsc, oldidx); | |
1244 | ||
1245 | /* | |
1246 | * Order all writes to buffer before the commit count update that will | |
1247 | * determine that the subbuffer is full. | |
1248 | */ | |
1249 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1250 | /* | |
1251 | * Must write slot data before incrementing commit count. This | |
1252 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1253 | * by get_subbuf(). | |
1254 | */ | |
1255 | barrier(); | |
1256 | } else | |
1257 | smp_wmb(); | |
1258 | v_add(config, config->cb.subbuffer_header_size(), | |
1259 | &buf->commit_hot[oldidx].cc); | |
1260 | commit_count = v_read(config, &buf->commit_hot[oldidx].cc); | |
1261 | /* Check if the written buffer has to be delivered */ | |
1262 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
635e457c | 1263 | commit_count, oldidx, tsc); |
f3bc08c5 | 1264 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
7915e163 MD |
1265 | offsets->old + config->cb.subbuffer_header_size(), |
1266 | commit_count); | |
f3bc08c5 MD |
1267 | } |
1268 | ||
1269 | /* | |
1270 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1271 | * | |
1272 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1273 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1274 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1275 | * subbuffer. | |
1276 | */ | |
1277 | static | |
1278 | void lib_ring_buffer_switch_old_end(struct lib_ring_buffer *buf, | |
1279 | struct channel *chan, | |
1280 | struct switch_offsets *offsets, | |
1281 | u64 tsc) | |
1282 | { | |
5a8fd222 | 1283 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1284 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1285 | unsigned long commit_count, padding_size, data_size; | |
1286 | ||
1287 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1288 | padding_size = chan->backend.subbuf_size - data_size; | |
1289 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size); | |
1290 | ||
1291 | /* | |
1292 | * Order all writes to buffer before the commit count update that will | |
1293 | * determine that the subbuffer is full. | |
1294 | */ | |
1295 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1296 | /* | |
1297 | * Must write slot data before incrementing commit count. This | |
1298 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1299 | * by get_subbuf(). | |
1300 | */ | |
1301 | barrier(); | |
1302 | } else | |
1303 | smp_wmb(); | |
1304 | v_add(config, padding_size, &buf->commit_hot[oldidx].cc); | |
1305 | commit_count = v_read(config, &buf->commit_hot[oldidx].cc); | |
1306 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, | |
635e457c | 1307 | commit_count, oldidx, tsc); |
f3bc08c5 | 1308 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
7915e163 | 1309 | offsets->old + padding_size, commit_count); |
f3bc08c5 MD |
1310 | } |
1311 | ||
1312 | /* | |
1313 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1314 | * | |
1315 | * This code can be executed unordered : writers may already have written to the | |
1316 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1317 | * that this code is executed before the deliver of this sub-buffer. | |
1318 | */ | |
1319 | static | |
1320 | void lib_ring_buffer_switch_new_start(struct lib_ring_buffer *buf, | |
1321 | struct channel *chan, | |
1322 | struct switch_offsets *offsets, | |
1323 | u64 tsc) | |
1324 | { | |
5a8fd222 | 1325 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1326 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1327 | unsigned long commit_count; | |
1328 | ||
1329 | config->cb.buffer_begin(buf, tsc, beginidx); | |
1330 | ||
1331 | /* | |
1332 | * Order all writes to buffer before the commit count update that will | |
1333 | * determine that the subbuffer is full. | |
1334 | */ | |
1335 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1336 | /* | |
1337 | * Must write slot data before incrementing commit count. This | |
1338 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1339 | * by get_subbuf(). | |
1340 | */ | |
1341 | barrier(); | |
1342 | } else | |
1343 | smp_wmb(); | |
1344 | v_add(config, config->cb.subbuffer_header_size(), | |
1345 | &buf->commit_hot[beginidx].cc); | |
1346 | commit_count = v_read(config, &buf->commit_hot[beginidx].cc); | |
1347 | /* Check if the written buffer has to be delivered */ | |
1348 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
635e457c | 1349 | commit_count, beginidx, tsc); |
f3bc08c5 | 1350 | lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx, |
7915e163 MD |
1351 | offsets->begin + config->cb.subbuffer_header_size(), |
1352 | commit_count); | |
f3bc08c5 MD |
1353 | } |
1354 | ||
f5ea5800 MD |
1355 | /* |
1356 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1357 | * | |
768b05c9 MD |
1358 | * Calls subbuffer_set_data_size() to set the data size of the current |
1359 | * sub-buffer. We do not need to perform check_deliver nor commit here, | |
1360 | * since this task will be done by the "commit" of the event for which | |
1361 | * we are currently doing the space reservation. | |
f5ea5800 MD |
1362 | */ |
1363 | static | |
1364 | void lib_ring_buffer_switch_new_end(struct lib_ring_buffer *buf, | |
1365 | struct channel *chan, | |
1366 | struct switch_offsets *offsets, | |
1367 | u64 tsc) | |
1368 | { | |
1369 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
768b05c9 | 1370 | unsigned long endidx, data_size; |
f5ea5800 | 1371 | |
768b05c9 | 1372 | endidx = subbuf_index(offsets->end - 1, chan); |
f5ea5800 | 1373 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; |
f5ea5800 | 1374 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size); |
f5ea5800 MD |
1375 | } |
1376 | ||
f3bc08c5 MD |
1377 | /* |
1378 | * Returns : | |
1379 | * 0 if ok | |
1380 | * !0 if execution must be aborted. | |
1381 | */ | |
1382 | static | |
1383 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
1384 | struct lib_ring_buffer *buf, | |
1385 | struct channel *chan, | |
1386 | struct switch_offsets *offsets, | |
1387 | u64 *tsc) | |
1388 | { | |
5a8fd222 | 1389 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
5334a2c5 | 1390 | unsigned long off, reserve_commit_diff; |
f3bc08c5 MD |
1391 | |
1392 | offsets->begin = v_read(config, &buf->offset); | |
1393 | offsets->old = offsets->begin; | |
1394 | offsets->switch_old_start = 0; | |
1395 | off = subbuf_offset(offsets->begin, chan); | |
1396 | ||
1397 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1398 | ||
1399 | /* | |
1400 | * Ensure we flush the header of an empty subbuffer when doing the | |
1401 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1402 | * total data gathering duration even if there were no records saved | |
1403 | * after the last buffer switch. | |
1404 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1405 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1406 | * subbuffer header as appropriate. | |
1407 | * The next record that reserves space will be responsible for | |
1408 | * populating the following subbuffer header. We choose not to populate | |
1409 | * the next subbuffer header here because we want to be able to use | |
1410 | * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop | |
1411 | * buffer flush, which must guarantee that all the buffer content | |
1412 | * (records and header timestamps) are visible to the reader. This is | |
1413 | * required for quiescence guarantees for the fusion merge. | |
1414 | */ | |
5334a2c5 MD |
1415 | if (mode != SWITCH_FLUSH && !off) |
1416 | return -1; /* we do not have to switch : buffer is empty */ | |
1417 | ||
1418 | if (unlikely(off == 0)) { | |
1419 | unsigned long sb_index, commit_count; | |
1420 | ||
1421 | /* | |
1422 | * We are performing a SWITCH_FLUSH. At this stage, there are no | |
1423 | * concurrent writes into the buffer. | |
1424 | * | |
1425 | * The client does not save any header information. Don't | |
1426 | * switch empty subbuffer on finalize, because it is invalid to | |
1427 | * deliver a completely empty subbuffer. | |
1428 | */ | |
1429 | if (!config->cb.subbuffer_header_size()) | |
1430 | return -1; | |
1431 | ||
1432 | /* Test new buffer integrity */ | |
1433 | sb_index = subbuf_index(offsets->begin, chan); | |
1434 | commit_count = v_read(config, | |
1435 | &buf->commit_cold[sb_index].cc_sb); | |
1436 | reserve_commit_diff = | |
1437 | (buf_trunc(offsets->begin, chan) | |
1438 | >> chan->backend.num_subbuf_order) | |
1439 | - (commit_count & chan->commit_count_mask); | |
1440 | if (likely(reserve_commit_diff == 0)) { | |
1441 | /* Next subbuffer not being written to. */ | |
1442 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1443 | subbuf_trunc(offsets->begin, chan) | |
1444 | - subbuf_trunc((unsigned long) | |
1445 | atomic_long_read(&buf->consumed), chan) | |
1446 | >= chan->backend.buf_size)) { | |
1447 | /* | |
1448 | * We do not overwrite non consumed buffers | |
1449 | * and we are full : don't switch. | |
1450 | */ | |
f3bc08c5 | 1451 | return -1; |
5334a2c5 MD |
1452 | } else { |
1453 | /* | |
1454 | * Next subbuffer not being written to, and we | |
1455 | * are either in overwrite mode or the buffer is | |
1456 | * not full. It's safe to write in this new | |
1457 | * subbuffer. | |
1458 | */ | |
1459 | } | |
1460 | } else { | |
f3bc08c5 | 1461 | /* |
5334a2c5 MD |
1462 | * Next subbuffer reserve offset does not match the |
1463 | * commit offset. Don't perform switch in | |
1464 | * producer-consumer and overwrite mode. Caused by | |
1465 | * either a writer OOPS or too many nested writes over a | |
1466 | * reserve/commit pair. | |
f3bc08c5 | 1467 | */ |
5334a2c5 | 1468 | return -1; |
f3bc08c5 | 1469 | } |
5334a2c5 MD |
1470 | |
1471 | /* | |
1472 | * Need to write the subbuffer start header on finalize. | |
1473 | */ | |
1474 | offsets->switch_old_start = 1; | |
1475 | } | |
1476 | offsets->begin = subbuf_align(offsets->begin, chan); | |
f3bc08c5 MD |
1477 | /* Note: old points to the next subbuf at offset 0 */ |
1478 | offsets->end = offsets->begin; | |
1479 | return 0; | |
1480 | } | |
1481 | ||
1482 | /* | |
1483 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1484 | * called while tracing is active with absolutely no lock held. | |
1485 | * | |
1486 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1487 | * operations, this function must be called from the CPU which owns the buffer | |
1488 | * for a ACTIVE flush. | |
1489 | */ | |
1490 | void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode) | |
1491 | { | |
1492 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1493 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1494 | struct switch_offsets offsets; |
1495 | unsigned long oldidx; | |
1496 | u64 tsc; | |
1497 | ||
1498 | offsets.size = 0; | |
1499 | ||
1500 | /* | |
1501 | * Perform retryable operations. | |
1502 | */ | |
1503 | do { | |
1504 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1505 | &tsc)) | |
1506 | return; /* Switch not needed */ | |
1507 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1508 | != offsets.old); | |
1509 | ||
1510 | /* | |
1511 | * Atomically update last_tsc. This update races against concurrent | |
1512 | * atomic updates, but the race will always cause supplementary full TSC | |
1513 | * records, never the opposite (missing a full TSC record when it would | |
1514 | * be needed). | |
1515 | */ | |
1516 | save_last_tsc(config, buf, tsc); | |
1517 | ||
1518 | /* | |
1519 | * Push the reader if necessary | |
1520 | */ | |
1521 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1522 | ||
1523 | oldidx = subbuf_index(offsets.old, chan); | |
1524 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx); | |
1525 | ||
1526 | /* | |
1527 | * May need to populate header start on SWITCH_FLUSH. | |
1528 | */ | |
1529 | if (offsets.switch_old_start) { | |
1530 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc); | |
1531 | offsets.old += config->cb.subbuffer_header_size(); | |
1532 | } | |
1533 | ||
1534 | /* | |
1535 | * Switch old subbuffer. | |
1536 | */ | |
1537 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc); | |
1538 | } | |
1539 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow); | |
1540 | ||
5e391252 MD |
1541 | static void remote_switch(void *info) |
1542 | { | |
1543 | struct lib_ring_buffer *buf = info; | |
1544 | ||
1545 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
1546 | } | |
1547 | ||
1548 | void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf) | |
1549 | { | |
1550 | struct channel *chan = buf->backend.chan; | |
1551 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1552 | int ret; | |
1553 | ||
1554 | /* | |
1555 | * With global synchronization we don't need to use the IPI scheme. | |
1556 | */ | |
1557 | if (config->sync == RING_BUFFER_SYNC_GLOBAL) { | |
1558 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
1559 | return; | |
1560 | } | |
1561 | ||
1562 | /* | |
1563 | * Taking lock on CPU hotplug to ensure two things: first, that the | |
1564 | * target cpu is not taken concurrently offline while we are within | |
1565 | * smp_call_function_single() (I don't trust that get_cpu() on the | |
1566 | * _local_ CPU actually inhibit CPU hotplug for the _remote_ CPU (to be | |
1567 | * confirmed)). Secondly, if it happens that the CPU is not online, our | |
1568 | * own call to lib_ring_buffer_switch_slow() needs to be protected from | |
1569 | * CPU hotplug handlers, which can also perform a remote subbuffer | |
1570 | * switch. | |
1571 | */ | |
1572 | get_online_cpus(); | |
1573 | ret = smp_call_function_single(buf->backend.cpu, | |
1574 | remote_switch, buf, 1); | |
1575 | if (ret) { | |
1576 | /* Remote CPU is offline, do it ourself. */ | |
1577 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
1578 | } | |
1579 | put_online_cpus(); | |
1580 | } | |
1581 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote); | |
1582 | ||
f3bc08c5 MD |
1583 | /* |
1584 | * Returns : | |
1585 | * 0 if ok | |
97ca2c54 MD |
1586 | * -ENOSPC if event size is too large for packet. |
1587 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1588 | * -EIO if data cannot be written into the buffer for any other reason. | |
f3bc08c5 MD |
1589 | */ |
1590 | static | |
1591 | int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, | |
1592 | struct channel *chan, | |
1593 | struct switch_offsets *offsets, | |
1594 | struct lib_ring_buffer_ctx *ctx) | |
1595 | { | |
5a8fd222 | 1596 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
0fdec686 | 1597 | unsigned long reserve_commit_diff, offset_cmp; |
f3bc08c5 | 1598 | |
0fdec686 MD |
1599 | retry: |
1600 | offsets->begin = offset_cmp = v_read(config, &buf->offset); | |
f3bc08c5 MD |
1601 | offsets->old = offsets->begin; |
1602 | offsets->switch_new_start = 0; | |
f5ea5800 | 1603 | offsets->switch_new_end = 0; |
f3bc08c5 MD |
1604 | offsets->switch_old_end = 0; |
1605 | offsets->pre_header_padding = 0; | |
1606 | ||
1607 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
97ca2c54 MD |
1608 | if ((int64_t) ctx->tsc == -EIO) |
1609 | return -EIO; | |
f3bc08c5 MD |
1610 | |
1611 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
64c796d8 | 1612 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; |
f3bc08c5 MD |
1613 | |
1614 | if (unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { | |
1615 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1616 | } else { | |
1617 | offsets->size = config->cb.record_header_size(config, chan, | |
1618 | offsets->begin, | |
f3bc08c5 | 1619 | &offsets->pre_header_padding, |
64c796d8 | 1620 | ctx); |
f3bc08c5 MD |
1621 | offsets->size += |
1622 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1623 | ctx->largest_align) | |
1624 | + ctx->data_size; | |
1625 | if (unlikely(subbuf_offset(offsets->begin, chan) + | |
1626 | offsets->size > chan->backend.subbuf_size)) { | |
1627 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1628 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1629 | } | |
1630 | } | |
1631 | if (unlikely(offsets->switch_new_start)) { | |
0fdec686 | 1632 | unsigned long sb_index, commit_count; |
f3bc08c5 MD |
1633 | |
1634 | /* | |
1635 | * We are typically not filling the previous buffer completely. | |
1636 | */ | |
1637 | if (likely(offsets->switch_old_end)) | |
1638 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1639 | offsets->begin = offsets->begin | |
1640 | + config->cb.subbuffer_header_size(); | |
1641 | /* Test new buffer integrity */ | |
1642 | sb_index = subbuf_index(offsets->begin, chan); | |
0fdec686 MD |
1643 | /* |
1644 | * Read buf->offset before buf->commit_cold[sb_index].cc_sb. | |
1645 | * lib_ring_buffer_check_deliver() has the matching | |
1646 | * memory barriers required around commit_cold cc_sb | |
1647 | * updates to ensure reserve and commit counter updates | |
1648 | * are not seen reordered when updated by another CPU. | |
1649 | */ | |
1650 | smp_rmb(); | |
1651 | commit_count = v_read(config, | |
1652 | &buf->commit_cold[sb_index].cc_sb); | |
1653 | /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */ | |
1654 | smp_rmb(); | |
1655 | if (unlikely(offset_cmp != v_read(config, &buf->offset))) { | |
1656 | /* | |
1657 | * The reserve counter have been concurrently updated | |
1658 | * while we read the commit counter. This means the | |
1659 | * commit counter we read might not match buf->offset | |
1660 | * due to concurrent update. We therefore need to retry. | |
1661 | */ | |
1662 | goto retry; | |
1663 | } | |
f3bc08c5 MD |
1664 | reserve_commit_diff = |
1665 | (buf_trunc(offsets->begin, chan) | |
1666 | >> chan->backend.num_subbuf_order) | |
0fdec686 | 1667 | - (commit_count & chan->commit_count_mask); |
f3bc08c5 MD |
1668 | if (likely(reserve_commit_diff == 0)) { |
1669 | /* Next subbuffer not being written to. */ | |
1670 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1671 | subbuf_trunc(offsets->begin, chan) | |
1672 | - subbuf_trunc((unsigned long) | |
1673 | atomic_long_read(&buf->consumed), chan) | |
1674 | >= chan->backend.buf_size)) { | |
1675 | /* | |
1676 | * We do not overwrite non consumed buffers | |
1677 | * and we are full : record is lost. | |
1678 | */ | |
1679 | v_inc(config, &buf->records_lost_full); | |
97ca2c54 | 1680 | return -ENOBUFS; |
f3bc08c5 MD |
1681 | } else { |
1682 | /* | |
1683 | * Next subbuffer not being written to, and we | |
1684 | * are either in overwrite mode or the buffer is | |
1685 | * not full. It's safe to write in this new | |
1686 | * subbuffer. | |
1687 | */ | |
1688 | } | |
1689 | } else { | |
1690 | /* | |
1691 | * Next subbuffer reserve offset does not match the | |
0fdec686 MD |
1692 | * commit offset, and this did not involve update to the |
1693 | * reserve counter. Drop record in producer-consumer and | |
1694 | * overwrite mode. Caused by either a writer OOPS or | |
1695 | * too many nested writes over a reserve/commit pair. | |
f3bc08c5 MD |
1696 | */ |
1697 | v_inc(config, &buf->records_lost_wrap); | |
97ca2c54 | 1698 | return -EIO; |
f3bc08c5 MD |
1699 | } |
1700 | offsets->size = | |
1701 | config->cb.record_header_size(config, chan, | |
1702 | offsets->begin, | |
f3bc08c5 | 1703 | &offsets->pre_header_padding, |
64c796d8 | 1704 | ctx); |
f3bc08c5 MD |
1705 | offsets->size += |
1706 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1707 | ctx->largest_align) | |
1708 | + ctx->data_size; | |
1709 | if (unlikely(subbuf_offset(offsets->begin, chan) | |
1710 | + offsets->size > chan->backend.subbuf_size)) { | |
1711 | /* | |
1712 | * Record too big for subbuffers, report error, don't | |
1713 | * complete the sub-buffer switch. | |
1714 | */ | |
1715 | v_inc(config, &buf->records_lost_big); | |
97ca2c54 | 1716 | return -ENOSPC; |
f3bc08c5 MD |
1717 | } else { |
1718 | /* | |
1719 | * We just made a successful buffer switch and the | |
1720 | * record fits in the new subbuffer. Let's write. | |
1721 | */ | |
1722 | } | |
1723 | } else { | |
1724 | /* | |
1725 | * Record fits in the current buffer and we are not on a switch | |
1726 | * boundary. It's safe to write. | |
1727 | */ | |
1728 | } | |
1729 | offsets->end = offsets->begin + offsets->size; | |
f5ea5800 MD |
1730 | |
1731 | if (unlikely(subbuf_offset(offsets->end, chan) == 0)) { | |
1732 | /* | |
1733 | * The offset_end will fall at the very beginning of the next | |
1734 | * subbuffer. | |
1735 | */ | |
1736 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1737 | } | |
f3bc08c5 MD |
1738 | return 0; |
1739 | } | |
1740 | ||
1741 | /** | |
1742 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1743 | * @ctx: ring buffer context. | |
1744 | * | |
97ca2c54 MD |
1745 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, |
1746 | * -EIO for other errors, else returns 0. | |
f3bc08c5 MD |
1747 | * It will take care of sub-buffer switching. |
1748 | */ | |
1749 | int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx) | |
1750 | { | |
1751 | struct channel *chan = ctx->chan; | |
5a8fd222 | 1752 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1753 | struct lib_ring_buffer *buf; |
1754 | struct switch_offsets offsets; | |
c099397a | 1755 | int ret; |
f3bc08c5 MD |
1756 | |
1757 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1758 | buf = per_cpu_ptr(chan->backend.buf, ctx->cpu); | |
1759 | else | |
1760 | buf = chan->backend.buf; | |
1761 | ctx->buf = buf; | |
1762 | ||
1763 | offsets.size = 0; | |
1764 | ||
1765 | do { | |
97ca2c54 MD |
1766 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, |
1767 | ctx); | |
1768 | if (unlikely(ret)) | |
1769 | return ret; | |
f3bc08c5 MD |
1770 | } while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
1771 | offsets.end) | |
1772 | != offsets.old)); | |
1773 | ||
1774 | /* | |
1775 | * Atomically update last_tsc. This update races against concurrent | |
1776 | * atomic updates, but the race will always cause supplementary full TSC | |
1777 | * records, never the opposite (missing a full TSC record when it would | |
1778 | * be needed). | |
1779 | */ | |
1780 | save_last_tsc(config, buf, ctx->tsc); | |
1781 | ||
1782 | /* | |
1783 | * Push the reader if necessary | |
1784 | */ | |
1785 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
1786 | ||
1787 | /* | |
1788 | * Clear noref flag for this subbuffer. | |
1789 | */ | |
1790 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1791 | subbuf_index(offsets.end - 1, chan)); | |
1792 | ||
1793 | /* | |
1794 | * Switch old subbuffer if needed. | |
1795 | */ | |
1796 | if (unlikely(offsets.switch_old_end)) { | |
1797 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1798 | subbuf_index(offsets.old - 1, chan)); | |
1799 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc); | |
1800 | } | |
1801 | ||
1802 | /* | |
1803 | * Populate new subbuffer. | |
1804 | */ | |
1805 | if (unlikely(offsets.switch_new_start)) | |
1806 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc); | |
1807 | ||
f5ea5800 MD |
1808 | if (unlikely(offsets.switch_new_end)) |
1809 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc); | |
1810 | ||
f3bc08c5 MD |
1811 | ctx->slot_size = offsets.size; |
1812 | ctx->pre_offset = offsets.begin; | |
1813 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
1814 | return 0; | |
1815 | } | |
1816 | EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow); | |
6fb8de4b | 1817 | |
02a766bb | 1818 | int __init init_lib_ring_buffer_frontend(void) |
6fb8de4b MD |
1819 | { |
1820 | int cpu; | |
1821 | ||
1822 | for_each_possible_cpu(cpu) | |
1823 | spin_lock_init(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
02a766bb | 1824 | return 0; |
6fb8de4b | 1825 | } |
02a766bb MD |
1826 | |
1827 | module_init(init_lib_ring_buffer_frontend); | |
1a5db82d MD |
1828 | |
1829 | void __exit exit_lib_ring_buffer_frontend(void) | |
1830 | { | |
1831 | } | |
1832 | ||
1833 | module_exit(exit_lib_ring_buffer_frontend); |