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 | ||
c075712b MD |
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> | |
63 | #include <wrapper/atomic.h> | |
64 | #include <wrapper/kref.h> | |
65 | #include <wrapper/percpu-defs.h> | |
f3bc08c5 MD |
66 | |
67 | /* | |
68 | * Internal structure representing offsets to use at a sub-buffer switch. | |
69 | */ | |
70 | struct switch_offsets { | |
71 | unsigned long begin, end, old; | |
72 | size_t pre_header_padding, size; | |
f5ea5800 MD |
73 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, |
74 | switch_old_end:1; | |
f3bc08c5 MD |
75 | }; |
76 | ||
77 | #ifdef CONFIG_NO_HZ | |
78 | enum tick_nohz_val { | |
79 | TICK_NOHZ_STOP, | |
80 | TICK_NOHZ_FLUSH, | |
81 | TICK_NOHZ_RESTART, | |
82 | }; | |
83 | ||
84 | static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier); | |
85 | #endif /* CONFIG_NO_HZ */ | |
86 | ||
87 | static DEFINE_PER_CPU(spinlock_t, ring_buffer_nohz_lock); | |
88 | ||
89 | DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting); | |
90 | EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting); | |
91 | ||
92 | static | |
93 | void lib_ring_buffer_print_errors(struct channel *chan, | |
94 | struct lib_ring_buffer *buf, int cpu); | |
64af2437 MD |
95 | static |
96 | void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf, | |
97 | enum switch_mode mode); | |
f3bc08c5 MD |
98 | |
99 | /* | |
100 | * Must be called under cpu hotplug protection. | |
101 | */ | |
102 | void lib_ring_buffer_free(struct lib_ring_buffer *buf) | |
103 | { | |
104 | struct channel *chan = buf->backend.chan; | |
105 | ||
106 | lib_ring_buffer_print_errors(chan, buf, buf->backend.cpu); | |
107 | kfree(buf->commit_hot); | |
108 | kfree(buf->commit_cold); | |
109 | ||
110 | lib_ring_buffer_backend_free(&buf->backend); | |
111 | } | |
112 | ||
113 | /** | |
114 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
115 | * @buf: Ring buffer. | |
116 | * | |
117 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
118 | * used for writing. The ring buffer can be opened for reading, but the reader | |
119 | * should not be using the iterator concurrently with reset. The previous | |
120 | * current iterator record is reset. | |
121 | */ | |
122 | void lib_ring_buffer_reset(struct lib_ring_buffer *buf) | |
123 | { | |
124 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 125 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
126 | unsigned int i; |
127 | ||
128 | /* | |
129 | * Reset iterator first. It will put the subbuffer if it currently holds | |
130 | * it. | |
131 | */ | |
132 | lib_ring_buffer_iterator_reset(buf); | |
133 | v_set(config, &buf->offset, 0); | |
134 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
135 | v_set(config, &buf->commit_hot[i].cc, 0); | |
136 | v_set(config, &buf->commit_hot[i].seq, 0); | |
137 | v_set(config, &buf->commit_cold[i].cc_sb, 0); | |
138 | } | |
139 | atomic_long_set(&buf->consumed, 0); | |
140 | atomic_set(&buf->record_disabled, 0); | |
141 | v_set(config, &buf->last_tsc, 0); | |
142 | lib_ring_buffer_backend_reset(&buf->backend); | |
143 | /* Don't reset number of active readers */ | |
144 | v_set(config, &buf->records_lost_full, 0); | |
145 | v_set(config, &buf->records_lost_wrap, 0); | |
146 | v_set(config, &buf->records_lost_big, 0); | |
147 | v_set(config, &buf->records_count, 0); | |
148 | v_set(config, &buf->records_overrun, 0); | |
149 | buf->finalized = 0; | |
150 | } | |
151 | EXPORT_SYMBOL_GPL(lib_ring_buffer_reset); | |
152 | ||
153 | /** | |
154 | * channel_reset - Reset channel to initial values. | |
155 | * @chan: Channel. | |
156 | * | |
157 | * Effectively empty the channel. Should be called when the channel is not used | |
158 | * for writing. The channel can be opened for reading, but the reader should not | |
159 | * be using the iterator concurrently with reset. The previous current iterator | |
160 | * record is reset. | |
161 | */ | |
162 | void channel_reset(struct channel *chan) | |
163 | { | |
164 | /* | |
165 | * Reset iterators first. Will put the subbuffer if held for reading. | |
166 | */ | |
167 | channel_iterator_reset(chan); | |
168 | atomic_set(&chan->record_disabled, 0); | |
169 | /* Don't reset commit_count_mask, still valid */ | |
170 | channel_backend_reset(&chan->backend); | |
171 | /* Don't reset switch/read timer interval */ | |
172 | /* Don't reset notifiers and notifier enable bits */ | |
173 | /* Don't reset reader reference count */ | |
174 | } | |
175 | EXPORT_SYMBOL_GPL(channel_reset); | |
176 | ||
177 | /* | |
178 | * Must be called under cpu hotplug protection. | |
179 | */ | |
180 | int lib_ring_buffer_create(struct lib_ring_buffer *buf, | |
181 | struct channel_backend *chanb, int cpu) | |
182 | { | |
5a8fd222 | 183 | const struct lib_ring_buffer_config *config = &chanb->config; |
f3bc08c5 MD |
184 | struct channel *chan = container_of(chanb, struct channel, backend); |
185 | void *priv = chanb->priv; | |
f3bc08c5 MD |
186 | size_t subbuf_header_size; |
187 | u64 tsc; | |
188 | int ret; | |
189 | ||
190 | /* Test for cpu hotplug */ | |
191 | if (buf->backend.allocated) | |
192 | return 0; | |
193 | ||
194 | /* | |
195 | * Paranoia: per cpu dynamic allocation is not officially documented as | |
196 | * zeroing the memory, so let's do it here too, just in case. | |
197 | */ | |
198 | memset(buf, 0, sizeof(*buf)); | |
199 | ||
200 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, cpu); | |
201 | if (ret) | |
202 | return ret; | |
203 | ||
204 | buf->commit_hot = | |
205 | kzalloc_node(ALIGN(sizeof(*buf->commit_hot) | |
206 | * chan->backend.num_subbuf, | |
207 | 1 << INTERNODE_CACHE_SHIFT), | |
208 | GFP_KERNEL, cpu_to_node(max(cpu, 0))); | |
209 | if (!buf->commit_hot) { | |
210 | ret = -ENOMEM; | |
211 | goto free_chanbuf; | |
212 | } | |
213 | ||
214 | buf->commit_cold = | |
215 | kzalloc_node(ALIGN(sizeof(*buf->commit_cold) | |
216 | * chan->backend.num_subbuf, | |
217 | 1 << INTERNODE_CACHE_SHIFT), | |
218 | GFP_KERNEL, cpu_to_node(max(cpu, 0))); | |
219 | if (!buf->commit_cold) { | |
220 | ret = -ENOMEM; | |
221 | goto free_commit; | |
222 | } | |
223 | ||
f3bc08c5 | 224 | init_waitqueue_head(&buf->read_wait); |
71c1d843 | 225 | init_waitqueue_head(&buf->write_wait); |
f3bc08c5 MD |
226 | raw_spin_lock_init(&buf->raw_tick_nohz_spinlock); |
227 | ||
228 | /* | |
229 | * Write the subbuffer header for first subbuffer so we know the total | |
230 | * duration of data gathering. | |
231 | */ | |
232 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
233 | v_set(config, &buf->offset, subbuf_header_size); | |
234 | subbuffer_id_clear_noref(config, &buf->backend.buf_wsb[0].id); | |
235 | tsc = config->cb.ring_buffer_clock_read(buf->backend.chan); | |
236 | config->cb.buffer_begin(buf, tsc, 0); | |
237 | v_add(config, subbuf_header_size, &buf->commit_hot[0].cc); | |
238 | ||
239 | if (config->cb.buffer_create) { | |
240 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name); | |
241 | if (ret) | |
242 | goto free_init; | |
243 | } | |
244 | ||
245 | /* | |
246 | * Ensure the buffer is ready before setting it to allocated and setting | |
247 | * the cpumask. | |
248 | * Used for cpu hotplug vs cpumask iteration. | |
249 | */ | |
250 | smp_wmb(); | |
251 | buf->backend.allocated = 1; | |
252 | ||
253 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
254 | CHAN_WARN_ON(chan, cpumask_test_cpu(cpu, | |
255 | chan->backend.cpumask)); | |
256 | cpumask_set_cpu(cpu, chan->backend.cpumask); | |
257 | } | |
258 | ||
259 | return 0; | |
260 | ||
261 | /* Error handling */ | |
262 | free_init: | |
263 | kfree(buf->commit_cold); | |
264 | free_commit: | |
265 | kfree(buf->commit_hot); | |
266 | free_chanbuf: | |
267 | lib_ring_buffer_backend_free(&buf->backend); | |
268 | return ret; | |
269 | } | |
270 | ||
271 | static void switch_buffer_timer(unsigned long data) | |
272 | { | |
273 | struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data; | |
274 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 275 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
276 | |
277 | /* | |
278 | * Only flush buffers periodically if readers are active. | |
279 | */ | |
280 | if (atomic_long_read(&buf->active_readers)) | |
281 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
282 | ||
283 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
284 | mod_timer_pinned(&buf->switch_timer, | |
285 | jiffies + chan->switch_timer_interval); | |
286 | else | |
287 | mod_timer(&buf->switch_timer, | |
288 | jiffies + chan->switch_timer_interval); | |
289 | } | |
290 | ||
291 | /* | |
292 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
293 | */ | |
294 | static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer *buf) | |
295 | { | |
296 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 297 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
298 | |
299 | if (!chan->switch_timer_interval || buf->switch_timer_enabled) | |
300 | return; | |
301 | init_timer(&buf->switch_timer); | |
302 | buf->switch_timer.function = switch_buffer_timer; | |
303 | buf->switch_timer.expires = jiffies + chan->switch_timer_interval; | |
304 | buf->switch_timer.data = (unsigned long)buf; | |
305 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
306 | add_timer_on(&buf->switch_timer, buf->backend.cpu); | |
307 | else | |
308 | add_timer(&buf->switch_timer); | |
309 | buf->switch_timer_enabled = 1; | |
310 | } | |
311 | ||
312 | /* | |
313 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
314 | */ | |
315 | static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer *buf) | |
316 | { | |
317 | struct channel *chan = buf->backend.chan; | |
318 | ||
319 | if (!chan->switch_timer_interval || !buf->switch_timer_enabled) | |
320 | return; | |
321 | ||
322 | del_timer_sync(&buf->switch_timer); | |
323 | buf->switch_timer_enabled = 0; | |
324 | } | |
325 | ||
326 | /* | |
327 | * Polling timer to check the channels for data. | |
328 | */ | |
329 | static void read_buffer_timer(unsigned long data) | |
330 | { | |
331 | struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data; | |
332 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 333 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
334 | |
335 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
336 | ||
337 | if (atomic_long_read(&buf->active_readers) | |
338 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
339 | wake_up_interruptible(&buf->read_wait); | |
340 | wake_up_interruptible(&chan->read_wait); | |
341 | } | |
342 | ||
343 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
344 | mod_timer_pinned(&buf->read_timer, | |
345 | jiffies + chan->read_timer_interval); | |
346 | else | |
347 | mod_timer(&buf->read_timer, | |
348 | jiffies + chan->read_timer_interval); | |
349 | } | |
350 | ||
351 | /* | |
352 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
353 | */ | |
354 | static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf) | |
355 | { | |
356 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 357 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
358 | |
359 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
360 | || !chan->read_timer_interval | |
361 | || buf->read_timer_enabled) | |
362 | return; | |
363 | ||
364 | init_timer(&buf->read_timer); | |
365 | buf->read_timer.function = read_buffer_timer; | |
366 | buf->read_timer.expires = jiffies + chan->read_timer_interval; | |
367 | buf->read_timer.data = (unsigned long)buf; | |
368 | ||
369 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
370 | add_timer_on(&buf->read_timer, buf->backend.cpu); | |
371 | else | |
372 | add_timer(&buf->read_timer); | |
373 | buf->read_timer_enabled = 1; | |
374 | } | |
375 | ||
376 | /* | |
377 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
378 | */ | |
379 | static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer *buf) | |
380 | { | |
381 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 382 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
383 | |
384 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
385 | || !chan->read_timer_interval | |
386 | || !buf->read_timer_enabled) | |
387 | return; | |
388 | ||
389 | del_timer_sync(&buf->read_timer); | |
390 | /* | |
391 | * do one more check to catch data that has been written in the last | |
392 | * timer period. | |
393 | */ | |
394 | if (lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
395 | wake_up_interruptible(&buf->read_wait); | |
396 | wake_up_interruptible(&chan->read_wait); | |
397 | } | |
398 | buf->read_timer_enabled = 0; | |
399 | } | |
400 | ||
401 | #ifdef CONFIG_HOTPLUG_CPU | |
402 | /** | |
403 | * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback | |
404 | * @nb: notifier block | |
405 | * @action: hotplug action to take | |
406 | * @hcpu: CPU number | |
407 | * | |
408 | * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD) | |
409 | */ | |
410 | static | |
e8f071d5 | 411 | int lib_ring_buffer_cpu_hp_callback(struct notifier_block *nb, |
f3bc08c5 MD |
412 | unsigned long action, |
413 | void *hcpu) | |
414 | { | |
415 | unsigned int cpu = (unsigned long)hcpu; | |
416 | struct channel *chan = container_of(nb, struct channel, | |
417 | cpu_hp_notifier); | |
418 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
5a8fd222 | 419 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
420 | |
421 | if (!chan->cpu_hp_enable) | |
422 | return NOTIFY_DONE; | |
423 | ||
424 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
425 | ||
426 | switch (action) { | |
427 | case CPU_DOWN_FAILED: | |
428 | case CPU_DOWN_FAILED_FROZEN: | |
429 | case CPU_ONLINE: | |
430 | case CPU_ONLINE_FROZEN: | |
24cedcfe | 431 | wake_up_interruptible(&chan->hp_wait); |
f3bc08c5 MD |
432 | lib_ring_buffer_start_switch_timer(buf); |
433 | lib_ring_buffer_start_read_timer(buf); | |
434 | return NOTIFY_OK; | |
435 | ||
436 | case CPU_DOWN_PREPARE: | |
437 | case CPU_DOWN_PREPARE_FROZEN: | |
438 | lib_ring_buffer_stop_switch_timer(buf); | |
439 | lib_ring_buffer_stop_read_timer(buf); | |
440 | return NOTIFY_OK; | |
441 | ||
442 | case CPU_DEAD: | |
443 | case CPU_DEAD_FROZEN: | |
444 | /* | |
445 | * Performing a buffer switch on a remote CPU. Performed by | |
446 | * the CPU responsible for doing the hotunplug after the target | |
447 | * CPU stopped running completely. Ensures that all data | |
448 | * from that remote CPU is flushed. | |
449 | */ | |
450 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
451 | return NOTIFY_OK; | |
452 | ||
453 | default: | |
454 | return NOTIFY_DONE; | |
455 | } | |
456 | } | |
457 | #endif | |
458 | ||
23b908b0 | 459 | #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) |
f3bc08c5 MD |
460 | /* |
461 | * For per-cpu buffers, call the reader wakeups before switching the buffer, so | |
462 | * that wake-up-tracing generated events are flushed before going idle (in | |
463 | * tick_nohz). We test if the spinlock is locked to deal with the race where | |
464 | * readers try to sample the ring buffer before we perform the switch. We let | |
465 | * the readers retry in that case. If there is data in the buffer, the wake up | |
466 | * is going to forbid the CPU running the reader thread from going idle. | |
467 | */ | |
468 | static int notrace ring_buffer_tick_nohz_callback(struct notifier_block *nb, | |
469 | unsigned long val, | |
470 | void *data) | |
471 | { | |
472 | struct channel *chan = container_of(nb, struct channel, | |
473 | tick_nohz_notifier); | |
5a8fd222 | 474 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
475 | struct lib_ring_buffer *buf; |
476 | int cpu = smp_processor_id(); | |
477 | ||
478 | if (config->alloc != RING_BUFFER_ALLOC_PER_CPU) { | |
479 | /* | |
480 | * We don't support keeping the system idle with global buffers | |
481 | * and streaming active. In order to do so, we would need to | |
482 | * sample a non-nohz-cpumask racelessly with the nohz updates | |
483 | * without adding synchronization overhead to nohz. Leave this | |
484 | * use-case out for now. | |
485 | */ | |
486 | return 0; | |
487 | } | |
488 | ||
489 | buf = channel_get_ring_buffer(config, chan, cpu); | |
490 | switch (val) { | |
491 | case TICK_NOHZ_FLUSH: | |
492 | raw_spin_lock(&buf->raw_tick_nohz_spinlock); | |
493 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_TIMER | |
494 | && chan->read_timer_interval | |
495 | && atomic_long_read(&buf->active_readers) | |
496 | && (lib_ring_buffer_poll_deliver(config, buf, chan) | |
497 | || lib_ring_buffer_pending_data(config, buf, chan))) { | |
498 | wake_up_interruptible(&buf->read_wait); | |
499 | wake_up_interruptible(&chan->read_wait); | |
500 | } | |
501 | if (chan->switch_timer_interval) | |
502 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
503 | raw_spin_unlock(&buf->raw_tick_nohz_spinlock); | |
504 | break; | |
505 | case TICK_NOHZ_STOP: | |
e6b06d7d | 506 | spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
507 | lib_ring_buffer_stop_switch_timer(buf); |
508 | lib_ring_buffer_stop_read_timer(buf); | |
e6b06d7d | 509 | spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
510 | break; |
511 | case TICK_NOHZ_RESTART: | |
e6b06d7d | 512 | spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
513 | lib_ring_buffer_start_read_timer(buf); |
514 | lib_ring_buffer_start_switch_timer(buf); | |
e6b06d7d | 515 | spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
516 | break; |
517 | } | |
518 | ||
519 | return 0; | |
520 | } | |
521 | ||
522 | void notrace lib_ring_buffer_tick_nohz_flush(void) | |
523 | { | |
524 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_FLUSH, | |
525 | NULL); | |
526 | } | |
527 | ||
528 | void notrace lib_ring_buffer_tick_nohz_stop(void) | |
529 | { | |
530 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_STOP, | |
531 | NULL); | |
532 | } | |
533 | ||
534 | void notrace lib_ring_buffer_tick_nohz_restart(void) | |
535 | { | |
536 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_RESTART, | |
537 | NULL); | |
538 | } | |
23b908b0 | 539 | #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */ |
f3bc08c5 MD |
540 | |
541 | /* | |
542 | * Holds CPU hotplug. | |
543 | */ | |
544 | static void channel_unregister_notifiers(struct channel *chan) | |
545 | { | |
5a8fd222 | 546 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
547 | int cpu; |
548 | ||
549 | channel_iterator_unregister_notifiers(chan); | |
550 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
551 | #ifdef CONFIG_NO_HZ | |
552 | /* | |
553 | * Remove the nohz notifier first, so we are certain we stop | |
554 | * the timers. | |
555 | */ | |
556 | atomic_notifier_chain_unregister(&tick_nohz_notifier, | |
557 | &chan->tick_nohz_notifier); | |
558 | /* | |
559 | * ring_buffer_nohz_lock will not be needed below, because | |
560 | * we just removed the notifiers, which were the only source of | |
561 | * concurrency. | |
562 | */ | |
563 | #endif /* CONFIG_NO_HZ */ | |
564 | #ifdef CONFIG_HOTPLUG_CPU | |
565 | get_online_cpus(); | |
566 | chan->cpu_hp_enable = 0; | |
567 | for_each_online_cpu(cpu) { | |
568 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
569 | cpu); | |
570 | lib_ring_buffer_stop_switch_timer(buf); | |
571 | lib_ring_buffer_stop_read_timer(buf); | |
572 | } | |
573 | put_online_cpus(); | |
574 | unregister_cpu_notifier(&chan->cpu_hp_notifier); | |
575 | #else | |
576 | for_each_possible_cpu(cpu) { | |
577 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
578 | cpu); | |
579 | lib_ring_buffer_stop_switch_timer(buf); | |
580 | lib_ring_buffer_stop_read_timer(buf); | |
581 | } | |
582 | #endif | |
583 | } else { | |
584 | struct lib_ring_buffer *buf = chan->backend.buf; | |
585 | ||
586 | lib_ring_buffer_stop_switch_timer(buf); | |
587 | lib_ring_buffer_stop_read_timer(buf); | |
588 | } | |
589 | channel_backend_unregister_notifiers(&chan->backend); | |
590 | } | |
591 | ||
64af2437 MD |
592 | static void lib_ring_buffer_set_quiescent(struct lib_ring_buffer *buf) |
593 | { | |
594 | if (!buf->quiescent) { | |
595 | buf->quiescent = true; | |
596 | _lib_ring_buffer_switch_remote(buf, SWITCH_FLUSH); | |
597 | } | |
598 | } | |
599 | ||
600 | static void lib_ring_buffer_clear_quiescent(struct lib_ring_buffer *buf) | |
601 | { | |
602 | buf->quiescent = false; | |
603 | } | |
604 | ||
605 | void lib_ring_buffer_set_quiescent_channel(struct channel *chan) | |
606 | { | |
607 | int cpu; | |
608 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
609 | ||
610 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
611 | get_online_cpus(); | |
612 | for_each_channel_cpu(cpu, chan) { | |
613 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
614 | cpu); | |
615 | ||
616 | lib_ring_buffer_set_quiescent(buf); | |
617 | } | |
618 | put_online_cpus(); | |
619 | } else { | |
620 | struct lib_ring_buffer *buf = chan->backend.buf; | |
621 | ||
622 | lib_ring_buffer_set_quiescent(buf); | |
623 | } | |
624 | } | |
625 | EXPORT_SYMBOL_GPL(lib_ring_buffer_set_quiescent_channel); | |
626 | ||
627 | void lib_ring_buffer_clear_quiescent_channel(struct channel *chan) | |
628 | { | |
629 | int cpu; | |
630 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
631 | ||
632 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
633 | get_online_cpus(); | |
634 | for_each_channel_cpu(cpu, chan) { | |
635 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
636 | cpu); | |
637 | ||
638 | lib_ring_buffer_clear_quiescent(buf); | |
639 | } | |
640 | put_online_cpus(); | |
641 | } else { | |
642 | struct lib_ring_buffer *buf = chan->backend.buf; | |
643 | ||
644 | lib_ring_buffer_clear_quiescent(buf); | |
645 | } | |
646 | } | |
647 | EXPORT_SYMBOL_GPL(lib_ring_buffer_clear_quiescent_channel); | |
648 | ||
f3bc08c5 MD |
649 | static void channel_free(struct channel *chan) |
650 | { | |
dd5a0db3 MD |
651 | if (chan->backend.release_priv_ops) { |
652 | chan->backend.release_priv_ops(chan->backend.priv_ops); | |
653 | } | |
f3bc08c5 MD |
654 | channel_iterator_free(chan); |
655 | channel_backend_free(&chan->backend); | |
656 | kfree(chan); | |
657 | } | |
658 | ||
659 | /** | |
660 | * channel_create - Create channel. | |
661 | * @config: ring buffer instance configuration | |
662 | * @name: name of the channel | |
663 | * @priv: ring buffer client private data | |
664 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous | |
665 | * address mapping. It is used only by RING_BUFFER_STATIC | |
666 | * configuration. It can be set to NULL for other backends. | |
667 | * @subbuf_size: subbuffer size | |
668 | * @num_subbuf: number of subbuffers | |
669 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
670 | * padding to let readers get those sub-buffers. | |
671 | * Used for live streaming. | |
672 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
673 | * | |
674 | * Holds cpu hotplug. | |
675 | * Returns NULL on failure. | |
676 | */ | |
677 | struct channel *channel_create(const struct lib_ring_buffer_config *config, | |
678 | const char *name, void *priv, void *buf_addr, | |
679 | size_t subbuf_size, | |
680 | size_t num_subbuf, unsigned int switch_timer_interval, | |
681 | unsigned int read_timer_interval) | |
682 | { | |
683 | int ret, cpu; | |
684 | struct channel *chan; | |
685 | ||
686 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
687 | read_timer_interval)) | |
688 | return NULL; | |
689 | ||
690 | chan = kzalloc(sizeof(struct channel), GFP_KERNEL); | |
691 | if (!chan) | |
692 | return NULL; | |
693 | ||
694 | ret = channel_backend_init(&chan->backend, name, config, priv, | |
695 | subbuf_size, num_subbuf); | |
696 | if (ret) | |
697 | goto error; | |
698 | ||
699 | ret = channel_iterator_init(chan); | |
700 | if (ret) | |
701 | goto error_free_backend; | |
702 | ||
703 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
704 | chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
705 | chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
f40270ad | 706 | kref_init(&chan->ref); |
f3bc08c5 | 707 | init_waitqueue_head(&chan->read_wait); |
24cedcfe | 708 | init_waitqueue_head(&chan->hp_wait); |
f3bc08c5 MD |
709 | |
710 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
23b908b0 | 711 | #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) |
f3bc08c5 MD |
712 | /* Only benefit from NO_HZ idle with per-cpu buffers for now. */ |
713 | chan->tick_nohz_notifier.notifier_call = | |
714 | ring_buffer_tick_nohz_callback; | |
715 | chan->tick_nohz_notifier.priority = ~0U; | |
716 | atomic_notifier_chain_register(&tick_nohz_notifier, | |
717 | &chan->tick_nohz_notifier); | |
23b908b0 | 718 | #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */ |
f3bc08c5 MD |
719 | |
720 | /* | |
721 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
722 | * in early initcall, it will not be notified of secondary cpus. | |
723 | * In that off case, we need to allocate for all possible cpus. | |
724 | */ | |
725 | #ifdef CONFIG_HOTPLUG_CPU | |
726 | chan->cpu_hp_notifier.notifier_call = | |
727 | lib_ring_buffer_cpu_hp_callback; | |
728 | chan->cpu_hp_notifier.priority = 6; | |
729 | register_cpu_notifier(&chan->cpu_hp_notifier); | |
730 | ||
731 | get_online_cpus(); | |
732 | for_each_online_cpu(cpu) { | |
733 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
734 | cpu); | |
735 | spin_lock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
736 | lib_ring_buffer_start_switch_timer(buf); | |
737 | lib_ring_buffer_start_read_timer(buf); | |
738 | spin_unlock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
739 | } | |
740 | chan->cpu_hp_enable = 1; | |
741 | put_online_cpus(); | |
742 | #else | |
743 | for_each_possible_cpu(cpu) { | |
744 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
745 | cpu); | |
746 | spin_lock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
747 | lib_ring_buffer_start_switch_timer(buf); | |
748 | lib_ring_buffer_start_read_timer(buf); | |
749 | spin_unlock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
750 | } | |
751 | #endif | |
752 | } else { | |
753 | struct lib_ring_buffer *buf = chan->backend.buf; | |
754 | ||
755 | lib_ring_buffer_start_switch_timer(buf); | |
756 | lib_ring_buffer_start_read_timer(buf); | |
757 | } | |
758 | ||
759 | return chan; | |
760 | ||
761 | error_free_backend: | |
762 | channel_backend_free(&chan->backend); | |
763 | error: | |
764 | kfree(chan); | |
765 | return NULL; | |
766 | } | |
767 | EXPORT_SYMBOL_GPL(channel_create); | |
768 | ||
f40270ad MD |
769 | static |
770 | void channel_release(struct kref *kref) | |
771 | { | |
772 | struct channel *chan = container_of(kref, struct channel, ref); | |
773 | channel_free(chan); | |
774 | } | |
775 | ||
f3bc08c5 MD |
776 | /** |
777 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
778 | * @chan: channel to destroy | |
779 | * | |
780 | * Holds cpu hotplug. | |
9a0df743 MD |
781 | * Call "destroy" callback, finalize channels, and then decrement the |
782 | * channel reference count. Note that when readers have completed data | |
783 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
784 | * They should release their handle at that point. Returns the private | |
785 | * data pointer. | |
f3bc08c5 MD |
786 | */ |
787 | void *channel_destroy(struct channel *chan) | |
788 | { | |
789 | int cpu; | |
5a8fd222 | 790 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
791 | void *priv; |
792 | ||
793 | channel_unregister_notifiers(chan); | |
794 | ||
795 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
796 | /* | |
797 | * No need to hold cpu hotplug, because all notifiers have been | |
798 | * unregistered. | |
799 | */ | |
800 | for_each_channel_cpu(cpu, chan) { | |
801 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
802 | cpu); | |
803 | ||
804 | if (config->cb.buffer_finalize) | |
805 | config->cb.buffer_finalize(buf, | |
806 | chan->backend.priv, | |
807 | cpu); | |
808 | if (buf->backend.allocated) | |
64af2437 | 809 | lib_ring_buffer_set_quiescent(buf); |
f3bc08c5 MD |
810 | /* |
811 | * Perform flush before writing to finalized. | |
812 | */ | |
813 | smp_wmb(); | |
814 | ACCESS_ONCE(buf->finalized) = 1; | |
815 | wake_up_interruptible(&buf->read_wait); | |
816 | } | |
817 | } else { | |
818 | struct lib_ring_buffer *buf = chan->backend.buf; | |
819 | ||
820 | if (config->cb.buffer_finalize) | |
821 | config->cb.buffer_finalize(buf, chan->backend.priv, -1); | |
822 | if (buf->backend.allocated) | |
64af2437 | 823 | lib_ring_buffer_set_quiescent(buf); |
f3bc08c5 MD |
824 | /* |
825 | * Perform flush before writing to finalized. | |
826 | */ | |
827 | smp_wmb(); | |
828 | ACCESS_ONCE(buf->finalized) = 1; | |
829 | wake_up_interruptible(&buf->read_wait); | |
830 | } | |
24cedcfe MD |
831 | ACCESS_ONCE(chan->finalized) = 1; |
832 | wake_up_interruptible(&chan->hp_wait); | |
f3bc08c5 | 833 | wake_up_interruptible(&chan->read_wait); |
f3bc08c5 | 834 | priv = chan->backend.priv; |
ba1d61bc | 835 | kref_put(&chan->ref, channel_release); |
f3bc08c5 MD |
836 | return priv; |
837 | } | |
838 | EXPORT_SYMBOL_GPL(channel_destroy); | |
839 | ||
840 | struct lib_ring_buffer *channel_get_ring_buffer( | |
841 | const struct lib_ring_buffer_config *config, | |
842 | struct channel *chan, int cpu) | |
843 | { | |
844 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) | |
845 | return chan->backend.buf; | |
846 | else | |
847 | return per_cpu_ptr(chan->backend.buf, cpu); | |
848 | } | |
849 | EXPORT_SYMBOL_GPL(channel_get_ring_buffer); | |
850 | ||
851 | int lib_ring_buffer_open_read(struct lib_ring_buffer *buf) | |
852 | { | |
853 | struct channel *chan = buf->backend.chan; | |
854 | ||
855 | if (!atomic_long_add_unless(&buf->active_readers, 1, 1)) | |
856 | return -EBUSY; | |
9c1f4643 MD |
857 | if (!lttng_kref_get(&chan->ref)) { |
858 | atomic_long_dec(&buf->active_readers); | |
859 | return -EOVERFLOW; | |
860 | } | |
505fb410 | 861 | lttng_smp_mb__after_atomic(); |
f3bc08c5 MD |
862 | return 0; |
863 | } | |
864 | EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read); | |
865 | ||
866 | void lib_ring_buffer_release_read(struct lib_ring_buffer *buf) | |
867 | { | |
868 | struct channel *chan = buf->backend.chan; | |
869 | ||
870 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
505fb410 | 871 | lttng_smp_mb__before_atomic(); |
f3bc08c5 | 872 | atomic_long_dec(&buf->active_readers); |
f40270ad | 873 | kref_put(&chan->ref, channel_release); |
f3bc08c5 MD |
874 | } |
875 | EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read); | |
876 | ||
877 | /* | |
878 | * Promote compiler barrier to a smp_mb(). | |
879 | * For the specific ring buffer case, this IPI call should be removed if the | |
880 | * architecture does not reorder writes. This should eventually be provided by | |
881 | * a separate architecture-specific infrastructure. | |
882 | */ | |
883 | static void remote_mb(void *info) | |
884 | { | |
885 | smp_mb(); | |
886 | } | |
887 | ||
888 | /** | |
889 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
890 | * @buf: ring buffer | |
891 | * @consumed: consumed count indicating the position where to read | |
892 | * @produced: produced count, indicates position when to stop reading | |
893 | * | |
894 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
895 | * data to read at consumed position, or 0 if the get operation succeeds. | |
896 | * Busy-loop trying to get data if the tick_nohz sequence lock is held. | |
897 | */ | |
898 | ||
899 | int lib_ring_buffer_snapshot(struct lib_ring_buffer *buf, | |
900 | unsigned long *consumed, unsigned long *produced) | |
901 | { | |
902 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 903 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
904 | unsigned long consumed_cur, write_offset; |
905 | int finalized; | |
906 | ||
dc5cd570 MD |
907 | /* |
908 | * First, ensure we perform a "final" flush onto the stream. This will | |
909 | * ensure we create a packet of padding if we encounter an empty | |
910 | * packet. This ensures the time-stamps right before the snapshot is | |
911 | * used as end of packet timestamp. | |
912 | */ | |
913 | if (!buf->quiescent) | |
914 | _lib_ring_buffer_switch_remote(buf, SWITCH_FLUSH); | |
915 | ||
f3bc08c5 MD |
916 | retry: |
917 | finalized = ACCESS_ONCE(buf->finalized); | |
918 | /* | |
919 | * Read finalized before counters. | |
920 | */ | |
921 | smp_rmb(); | |
922 | consumed_cur = atomic_long_read(&buf->consumed); | |
923 | /* | |
924 | * No need to issue a memory barrier between consumed count read and | |
925 | * write offset read, because consumed count can only change | |
926 | * concurrently in overwrite mode, and we keep a sequence counter | |
927 | * identifier derived from the write offset to check we are getting | |
928 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
929 | * "tagged" upon writes, tags are checked upon read). | |
930 | */ | |
931 | write_offset = v_read(config, &buf->offset); | |
932 | ||
933 | /* | |
934 | * Check that we are not about to read the same subbuffer in | |
935 | * which the writer head is. | |
936 | */ | |
937 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
938 | == 0) | |
939 | goto nodata; | |
940 | ||
941 | *consumed = consumed_cur; | |
942 | *produced = subbuf_trunc(write_offset, chan); | |
943 | ||
944 | return 0; | |
945 | ||
946 | nodata: | |
947 | /* | |
948 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
949 | * of "raw_spin_is_locked" memory ordering. | |
950 | */ | |
951 | if (finalized) | |
952 | return -ENODATA; | |
953 | else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock)) | |
954 | goto retry; | |
955 | else | |
956 | return -EAGAIN; | |
957 | } | |
958 | EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot); | |
959 | ||
960 | /** | |
961 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
71c1d843 MD |
962 | * |
963 | * Should only be called from consumer context. | |
f3bc08c5 MD |
964 | * @buf: ring buffer |
965 | * @consumed_new: new consumed count value | |
966 | */ | |
967 | void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf, | |
968 | unsigned long consumed_new) | |
969 | { | |
970 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
971 | struct channel *chan = bufb->chan; | |
972 | unsigned long consumed; | |
973 | ||
974 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
975 | ||
976 | /* | |
977 | * Only push the consumed value forward. | |
978 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
979 | * the writer in flight recorder mode. | |
980 | */ | |
981 | consumed = atomic_long_read(&buf->consumed); | |
982 | while ((long) consumed - (long) consumed_new < 0) | |
983 | consumed = atomic_long_cmpxchg(&buf->consumed, consumed, | |
984 | consumed_new); | |
71c1d843 MD |
985 | /* Wake-up the metadata producer */ |
986 | wake_up_interruptible(&buf->write_wait); | |
f3bc08c5 MD |
987 | } |
988 | EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer); | |
989 | ||
990 | /** | |
991 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
992 | * @buf: ring buffer | |
993 | * @consumed: consumed count indicating the position where to read | |
994 | * | |
995 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
996 | * data to read at consumed position, or 0 if the get operation succeeds. | |
997 | * Busy-loop trying to get data if the tick_nohz sequence lock is held. | |
998 | */ | |
999 | int lib_ring_buffer_get_subbuf(struct lib_ring_buffer *buf, | |
1000 | unsigned long consumed) | |
1001 | { | |
1002 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1003 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1004 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
1005 | int ret; | |
1006 | int finalized; | |
1007 | ||
8202b8a0 MD |
1008 | if (buf->get_subbuf) { |
1009 | /* | |
1010 | * Reader is trying to get a subbuffer twice. | |
1011 | */ | |
1012 | CHAN_WARN_ON(chan, 1); | |
1013 | return -EBUSY; | |
1014 | } | |
f3bc08c5 MD |
1015 | retry: |
1016 | finalized = ACCESS_ONCE(buf->finalized); | |
1017 | /* | |
1018 | * Read finalized before counters. | |
1019 | */ | |
1020 | smp_rmb(); | |
1021 | consumed_cur = atomic_long_read(&buf->consumed); | |
1022 | consumed_idx = subbuf_index(consumed, chan); | |
1023 | commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); | |
1024 | /* | |
1025 | * Make sure we read the commit count before reading the buffer | |
1026 | * data and the write offset. Correct consumed offset ordering | |
1027 | * wrt commit count is insured by the use of cmpxchg to update | |
1028 | * the consumed offset. | |
1029 | * smp_call_function_single can fail if the remote CPU is offline, | |
1030 | * this is OK because then there is no wmb to execute there. | |
1031 | * If our thread is executing on the same CPU as the on the buffers | |
1032 | * belongs to, we don't have to synchronize it at all. If we are | |
1033 | * migrated, the scheduler will take care of the memory barriers. | |
1034 | * Normally, smp_call_function_single() should ensure program order when | |
1035 | * executing the remote function, which implies that it surrounds the | |
1036 | * function execution with : | |
1037 | * smp_mb() | |
1038 | * send IPI | |
1039 | * csd_lock_wait | |
1040 | * recv IPI | |
1041 | * smp_mb() | |
1042 | * exec. function | |
1043 | * smp_mb() | |
1044 | * csd unlock | |
1045 | * smp_mb() | |
1046 | * | |
1047 | * However, smp_call_function_single() does not seem to clearly execute | |
1048 | * such barriers. It depends on spinlock semantic to provide the barrier | |
1049 | * before executing the IPI and, when busy-looping, csd_lock_wait only | |
1050 | * executes smp_mb() when it has to wait for the other CPU. | |
1051 | * | |
1052 | * I don't trust this code. Therefore, let's add the smp_mb() sequence | |
1053 | * required ourself, even if duplicated. It has no performance impact | |
1054 | * anyway. | |
1055 | * | |
1056 | * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs | |
1057 | * read and write vs write. They do not ensure core synchronization. We | |
1058 | * really have to ensure total order between the 3 barriers running on | |
1059 | * the 2 CPUs. | |
1060 | */ | |
1061 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1062 | if (config->sync == RING_BUFFER_SYNC_PER_CPU | |
1063 | && config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
1064 | if (raw_smp_processor_id() != buf->backend.cpu) { | |
1065 | /* Total order with IPI handler smp_mb() */ | |
1066 | smp_mb(); | |
1067 | smp_call_function_single(buf->backend.cpu, | |
1068 | remote_mb, NULL, 1); | |
1069 | /* Total order with IPI handler smp_mb() */ | |
1070 | smp_mb(); | |
1071 | } | |
1072 | } else { | |
1073 | /* Total order with IPI handler smp_mb() */ | |
1074 | smp_mb(); | |
1075 | smp_call_function(remote_mb, NULL, 1); | |
1076 | /* Total order with IPI handler smp_mb() */ | |
1077 | smp_mb(); | |
1078 | } | |
1079 | } else { | |
1080 | /* | |
1081 | * Local rmb to match the remote wmb to read the commit count | |
1082 | * before the buffer data and the write offset. | |
1083 | */ | |
1084 | smp_rmb(); | |
1085 | } | |
1086 | ||
1087 | write_offset = v_read(config, &buf->offset); | |
1088 | ||
1089 | /* | |
1090 | * Check that the buffer we are getting is after or at consumed_cur | |
1091 | * position. | |
1092 | */ | |
1093 | if ((long) subbuf_trunc(consumed, chan) | |
1094 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
1095 | goto nodata; | |
1096 | ||
1097 | /* | |
1098 | * Check that the subbuffer we are trying to consume has been | |
1099 | * already fully committed. | |
1100 | */ | |
1101 | if (((commit_count - chan->backend.subbuf_size) | |
1102 | & chan->commit_count_mask) | |
c9b3b5e2 | 1103 | - (buf_trunc(consumed, chan) |
f3bc08c5 MD |
1104 | >> chan->backend.num_subbuf_order) |
1105 | != 0) | |
1106 | goto nodata; | |
1107 | ||
1108 | /* | |
1109 | * Check that we are not about to read the same subbuffer in | |
1110 | * which the writer head is. | |
1111 | */ | |
c9b3b5e2 | 1112 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed, chan) |
f3bc08c5 MD |
1113 | == 0) |
1114 | goto nodata; | |
1115 | ||
1116 | /* | |
1117 | * Failure to get the subbuffer causes a busy-loop retry without going | |
1118 | * to a wait queue. These are caused by short-lived race windows where | |
1119 | * the writer is getting access to a subbuffer we were trying to get | |
1120 | * access to. Also checks that the "consumed" buffer count we are | |
1121 | * looking for matches the one contained in the subbuffer id. | |
1122 | */ | |
1123 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1124 | consumed_idx, buf_trunc_val(consumed, chan)); | |
1125 | if (ret) | |
1126 | goto retry; | |
1127 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
1128 | ||
1129 | buf->get_subbuf_consumed = consumed; | |
1130 | buf->get_subbuf = 1; | |
1131 | ||
1132 | return 0; | |
1133 | ||
1134 | nodata: | |
1135 | /* | |
1136 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1137 | * of "raw_spin_is_locked" memory ordering. | |
1138 | */ | |
1139 | if (finalized) | |
1140 | return -ENODATA; | |
1141 | else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock)) | |
1142 | goto retry; | |
1143 | else | |
1144 | return -EAGAIN; | |
1145 | } | |
1146 | EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf); | |
1147 | ||
1148 | /** | |
1149 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
1150 | * @buf: ring buffer | |
1151 | */ | |
1152 | void lib_ring_buffer_put_subbuf(struct lib_ring_buffer *buf) | |
1153 | { | |
1154 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1155 | struct channel *chan = bufb->chan; | |
5a8fd222 | 1156 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1157 | unsigned long read_sb_bindex, consumed_idx, consumed; |
1158 | ||
1159 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
1160 | ||
1161 | if (!buf->get_subbuf) { | |
1162 | /* | |
1163 | * Reader puts a subbuffer it did not get. | |
1164 | */ | |
1165 | CHAN_WARN_ON(chan, 1); | |
1166 | return; | |
1167 | } | |
1168 | consumed = buf->get_subbuf_consumed; | |
1169 | buf->get_subbuf = 0; | |
1170 | ||
1171 | /* | |
1172 | * Clear the records_unread counter. (overruns counter) | |
1173 | * Can still be non-zero if a file reader simply grabbed the data | |
1174 | * without using iterators. | |
1175 | * Can be below zero if an iterator is used on a snapshot more than | |
1176 | * once. | |
1177 | */ | |
1178 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
1179 | v_add(config, v_read(config, | |
1180 | &bufb->array[read_sb_bindex]->records_unread), | |
1181 | &bufb->records_read); | |
1182 | v_set(config, &bufb->array[read_sb_bindex]->records_unread, 0); | |
1183 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE | |
1184 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
1185 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
1186 | ||
1187 | /* | |
1188 | * Exchange the reader subbuffer with the one we put in its place in the | |
1189 | * writer subbuffer table. Expect the original consumed count. If | |
1190 | * update_read_sb_index fails, this is because the writer updated the | |
1191 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
1192 | * currently have: it has become invalid to try reading this sub-buffer | |
1193 | * consumed count value anyway. | |
1194 | */ | |
1195 | consumed_idx = subbuf_index(consumed, chan); | |
1196 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1197 | consumed_idx, buf_trunc_val(consumed, chan)); | |
1198 | /* | |
1199 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
1200 | * if the writer concurrently updated it. | |
1201 | */ | |
1202 | } | |
1203 | EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf); | |
1204 | ||
1205 | /* | |
1206 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
1207 | * position and the writer position. (inclusive) | |
1208 | */ | |
1209 | static | |
1210 | void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer *buf, | |
1211 | struct channel *chan, | |
1212 | unsigned long cons_offset, | |
1213 | int cpu) | |
1214 | { | |
5a8fd222 | 1215 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1216 | unsigned long cons_idx, commit_count, commit_count_sb; |
1217 | ||
1218 | cons_idx = subbuf_index(cons_offset, chan); | |
1219 | commit_count = v_read(config, &buf->commit_hot[cons_idx].cc); | |
1220 | commit_count_sb = v_read(config, &buf->commit_cold[cons_idx].cc_sb); | |
1221 | ||
1222 | if (subbuf_offset(commit_count, chan) != 0) | |
1223 | printk(KERN_WARNING | |
1224 | "ring buffer %s, cpu %d: " | |
1225 | "commit count in subbuffer %lu,\n" | |
1226 | "expecting multiples of %lu bytes\n" | |
1227 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
1228 | chan->backend.name, cpu, cons_idx, | |
1229 | chan->backend.subbuf_size, | |
1230 | commit_count, commit_count_sb); | |
1231 | ||
1232 | printk(KERN_DEBUG "ring buffer: %s, cpu %d: %lu bytes committed\n", | |
1233 | chan->backend.name, cpu, commit_count); | |
1234 | } | |
1235 | ||
1236 | static | |
1237 | void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf, | |
1238 | struct channel *chan, | |
1239 | void *priv, int cpu) | |
1240 | { | |
5a8fd222 | 1241 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1242 | unsigned long write_offset, cons_offset; |
1243 | ||
f3bc08c5 MD |
1244 | /* |
1245 | * No need to order commit_count, write_offset and cons_offset reads | |
1246 | * because we execute at teardown when no more writer nor reader | |
1247 | * references are left. | |
1248 | */ | |
1249 | write_offset = v_read(config, &buf->offset); | |
1250 | cons_offset = atomic_long_read(&buf->consumed); | |
1251 | if (write_offset != cons_offset) | |
05aad775 | 1252 | printk(KERN_DEBUG |
f3bc08c5 MD |
1253 | "ring buffer %s, cpu %d: " |
1254 | "non-consumed data\n" | |
1255 | " [ %lu bytes written, %lu bytes read ]\n", | |
1256 | chan->backend.name, cpu, write_offset, cons_offset); | |
1257 | ||
1258 | for (cons_offset = atomic_long_read(&buf->consumed); | |
1259 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), | |
1260 | chan) | |
1261 | - cons_offset) > 0; | |
1262 | cons_offset = subbuf_align(cons_offset, chan)) | |
1263 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1264 | cpu); | |
1265 | } | |
1266 | ||
1267 | static | |
1268 | void lib_ring_buffer_print_errors(struct channel *chan, | |
1269 | struct lib_ring_buffer *buf, int cpu) | |
1270 | { | |
5a8fd222 | 1271 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1272 | void *priv = chan->backend.priv; |
1273 | ||
ec01ec93 MD |
1274 | if (!strcmp(chan->backend.name, "relay-metadata")) { |
1275 | printk(KERN_DEBUG "ring buffer %s: %lu records written, " | |
1276 | "%lu records overrun\n", | |
1277 | chan->backend.name, | |
1278 | v_read(config, &buf->records_count), | |
1279 | v_read(config, &buf->records_overrun)); | |
1280 | } else { | |
1281 | printk(KERN_DEBUG "ring buffer %s, cpu %d: %lu records written, " | |
1282 | "%lu records overrun\n", | |
1283 | chan->backend.name, cpu, | |
1284 | v_read(config, &buf->records_count), | |
1285 | v_read(config, &buf->records_overrun)); | |
1286 | ||
1287 | if (v_read(config, &buf->records_lost_full) | |
1288 | || v_read(config, &buf->records_lost_wrap) | |
1289 | || v_read(config, &buf->records_lost_big)) | |
1290 | printk(KERN_WARNING | |
1291 | "ring buffer %s, cpu %d: records were lost. Caused by:\n" | |
1292 | " [ %lu buffer full, %lu nest buffer wrap-around, " | |
1293 | "%lu event too big ]\n", | |
1294 | chan->backend.name, cpu, | |
1295 | v_read(config, &buf->records_lost_full), | |
1296 | v_read(config, &buf->records_lost_wrap), | |
1297 | v_read(config, &buf->records_lost_big)); | |
1298 | } | |
f3bc08c5 MD |
1299 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu); |
1300 | } | |
1301 | ||
1302 | /* | |
1303 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1304 | * | |
dc5cd570 MD |
1305 | * Only executed by SWITCH_FLUSH, which can be issued while tracing is active |
1306 | * or at buffer finalization (destroy). | |
f3bc08c5 MD |
1307 | */ |
1308 | static | |
1309 | void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf, | |
1310 | struct channel *chan, | |
1311 | struct switch_offsets *offsets, | |
1312 | u64 tsc) | |
1313 | { | |
5a8fd222 | 1314 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1315 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1316 | unsigned long commit_count; | |
1317 | ||
1318 | config->cb.buffer_begin(buf, tsc, oldidx); | |
1319 | ||
1320 | /* | |
1321 | * Order all writes to buffer before the commit count update that will | |
1322 | * determine that the subbuffer is full. | |
1323 | */ | |
1324 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1325 | /* | |
1326 | * Must write slot data before incrementing commit count. This | |
1327 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1328 | * by get_subbuf(). | |
1329 | */ | |
1330 | barrier(); | |
1331 | } else | |
1332 | smp_wmb(); | |
1333 | v_add(config, config->cb.subbuffer_header_size(), | |
1334 | &buf->commit_hot[oldidx].cc); | |
1335 | commit_count = v_read(config, &buf->commit_hot[oldidx].cc); | |
1336 | /* Check if the written buffer has to be delivered */ | |
1337 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
635e457c | 1338 | commit_count, oldidx, tsc); |
f3bc08c5 | 1339 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
7915e163 MD |
1340 | offsets->old + config->cb.subbuffer_header_size(), |
1341 | commit_count); | |
f3bc08c5 MD |
1342 | } |
1343 | ||
1344 | /* | |
1345 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1346 | * | |
1347 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1348 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1349 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1350 | * subbuffer. | |
1351 | */ | |
1352 | static | |
1353 | void lib_ring_buffer_switch_old_end(struct lib_ring_buffer *buf, | |
1354 | struct channel *chan, | |
1355 | struct switch_offsets *offsets, | |
1356 | u64 tsc) | |
1357 | { | |
5a8fd222 | 1358 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1359 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1360 | unsigned long commit_count, padding_size, data_size; | |
1361 | ||
1362 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1363 | padding_size = chan->backend.subbuf_size - data_size; | |
1364 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size); | |
1365 | ||
1366 | /* | |
1367 | * Order all writes to buffer before the commit count update that will | |
1368 | * determine that the subbuffer is full. | |
1369 | */ | |
1370 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1371 | /* | |
1372 | * Must write slot data before incrementing commit count. This | |
1373 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1374 | * by get_subbuf(). | |
1375 | */ | |
1376 | barrier(); | |
1377 | } else | |
1378 | smp_wmb(); | |
1379 | v_add(config, padding_size, &buf->commit_hot[oldidx].cc); | |
1380 | commit_count = v_read(config, &buf->commit_hot[oldidx].cc); | |
1381 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, | |
635e457c | 1382 | commit_count, oldidx, tsc); |
f3bc08c5 | 1383 | lib_ring_buffer_write_commit_counter(config, buf, chan, oldidx, |
7915e163 | 1384 | offsets->old + padding_size, commit_count); |
f3bc08c5 MD |
1385 | } |
1386 | ||
1387 | /* | |
1388 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1389 | * | |
1390 | * This code can be executed unordered : writers may already have written to the | |
1391 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1392 | * that this code is executed before the deliver of this sub-buffer. | |
1393 | */ | |
1394 | static | |
1395 | void lib_ring_buffer_switch_new_start(struct lib_ring_buffer *buf, | |
1396 | struct channel *chan, | |
1397 | struct switch_offsets *offsets, | |
1398 | u64 tsc) | |
1399 | { | |
5a8fd222 | 1400 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1401 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1402 | unsigned long commit_count; | |
1403 | ||
1404 | config->cb.buffer_begin(buf, tsc, beginidx); | |
1405 | ||
1406 | /* | |
1407 | * Order all writes to buffer before the commit count update that will | |
1408 | * determine that the subbuffer is full. | |
1409 | */ | |
1410 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1411 | /* | |
1412 | * Must write slot data before incrementing commit count. This | |
1413 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1414 | * by get_subbuf(). | |
1415 | */ | |
1416 | barrier(); | |
1417 | } else | |
1418 | smp_wmb(); | |
1419 | v_add(config, config->cb.subbuffer_header_size(), | |
1420 | &buf->commit_hot[beginidx].cc); | |
1421 | commit_count = v_read(config, &buf->commit_hot[beginidx].cc); | |
1422 | /* Check if the written buffer has to be delivered */ | |
1423 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
635e457c | 1424 | commit_count, beginidx, tsc); |
f3bc08c5 | 1425 | lib_ring_buffer_write_commit_counter(config, buf, chan, beginidx, |
7915e163 MD |
1426 | offsets->begin + config->cb.subbuffer_header_size(), |
1427 | commit_count); | |
f3bc08c5 MD |
1428 | } |
1429 | ||
f5ea5800 MD |
1430 | /* |
1431 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1432 | * | |
768b05c9 MD |
1433 | * Calls subbuffer_set_data_size() to set the data size of the current |
1434 | * sub-buffer. We do not need to perform check_deliver nor commit here, | |
1435 | * since this task will be done by the "commit" of the event for which | |
1436 | * we are currently doing the space reservation. | |
f5ea5800 MD |
1437 | */ |
1438 | static | |
1439 | void lib_ring_buffer_switch_new_end(struct lib_ring_buffer *buf, | |
1440 | struct channel *chan, | |
1441 | struct switch_offsets *offsets, | |
1442 | u64 tsc) | |
1443 | { | |
1444 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
768b05c9 | 1445 | unsigned long endidx, data_size; |
f5ea5800 | 1446 | |
768b05c9 | 1447 | endidx = subbuf_index(offsets->end - 1, chan); |
f5ea5800 | 1448 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; |
f5ea5800 | 1449 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size); |
f5ea5800 MD |
1450 | } |
1451 | ||
f3bc08c5 MD |
1452 | /* |
1453 | * Returns : | |
1454 | * 0 if ok | |
1455 | * !0 if execution must be aborted. | |
1456 | */ | |
1457 | static | |
1458 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
1459 | struct lib_ring_buffer *buf, | |
1460 | struct channel *chan, | |
1461 | struct switch_offsets *offsets, | |
1462 | u64 *tsc) | |
1463 | { | |
5a8fd222 | 1464 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
5334a2c5 | 1465 | unsigned long off, reserve_commit_diff; |
f3bc08c5 MD |
1466 | |
1467 | offsets->begin = v_read(config, &buf->offset); | |
1468 | offsets->old = offsets->begin; | |
1469 | offsets->switch_old_start = 0; | |
1470 | off = subbuf_offset(offsets->begin, chan); | |
1471 | ||
1472 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1473 | ||
1474 | /* | |
1475 | * Ensure we flush the header of an empty subbuffer when doing the | |
1476 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1477 | * total data gathering duration even if there were no records saved | |
1478 | * after the last buffer switch. | |
1479 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1480 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1481 | * subbuffer header as appropriate. | |
1482 | * The next record that reserves space will be responsible for | |
1483 | * populating the following subbuffer header. We choose not to populate | |
1484 | * the next subbuffer header here because we want to be able to use | |
1485 | * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop | |
1486 | * buffer flush, which must guarantee that all the buffer content | |
1487 | * (records and header timestamps) are visible to the reader. This is | |
1488 | * required for quiescence guarantees for the fusion merge. | |
1489 | */ | |
5334a2c5 MD |
1490 | if (mode != SWITCH_FLUSH && !off) |
1491 | return -1; /* we do not have to switch : buffer is empty */ | |
1492 | ||
1493 | if (unlikely(off == 0)) { | |
1494 | unsigned long sb_index, commit_count; | |
1495 | ||
1496 | /* | |
dc5cd570 MD |
1497 | * We are performing a SWITCH_FLUSH. There may be concurrent |
1498 | * writes into the buffer if e.g. invoked while performing a | |
1499 | * snapshot on an active trace. | |
5334a2c5 | 1500 | * |
dc5cd570 MD |
1501 | * If the client does not save any header information (sub-buffer |
1502 | * header size == 0), don't switch empty subbuffer on finalize, | |
1503 | * because it is invalid to deliver a completely empty | |
1504 | * subbuffer. | |
5334a2c5 MD |
1505 | */ |
1506 | if (!config->cb.subbuffer_header_size()) | |
1507 | return -1; | |
1508 | ||
1509 | /* Test new buffer integrity */ | |
1510 | sb_index = subbuf_index(offsets->begin, chan); | |
1511 | commit_count = v_read(config, | |
1512 | &buf->commit_cold[sb_index].cc_sb); | |
1513 | reserve_commit_diff = | |
1514 | (buf_trunc(offsets->begin, chan) | |
1515 | >> chan->backend.num_subbuf_order) | |
1516 | - (commit_count & chan->commit_count_mask); | |
1517 | if (likely(reserve_commit_diff == 0)) { | |
1518 | /* Next subbuffer not being written to. */ | |
1519 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1520 | subbuf_trunc(offsets->begin, chan) | |
1521 | - subbuf_trunc((unsigned long) | |
1522 | atomic_long_read(&buf->consumed), chan) | |
1523 | >= chan->backend.buf_size)) { | |
1524 | /* | |
1525 | * We do not overwrite non consumed buffers | |
1526 | * and we are full : don't switch. | |
1527 | */ | |
f3bc08c5 | 1528 | return -1; |
5334a2c5 MD |
1529 | } else { |
1530 | /* | |
1531 | * Next subbuffer not being written to, and we | |
1532 | * are either in overwrite mode or the buffer is | |
1533 | * not full. It's safe to write in this new | |
1534 | * subbuffer. | |
1535 | */ | |
1536 | } | |
1537 | } else { | |
f3bc08c5 | 1538 | /* |
5334a2c5 MD |
1539 | * Next subbuffer reserve offset does not match the |
1540 | * commit offset. Don't perform switch in | |
1541 | * producer-consumer and overwrite mode. Caused by | |
1542 | * either a writer OOPS or too many nested writes over a | |
1543 | * reserve/commit pair. | |
f3bc08c5 | 1544 | */ |
5334a2c5 | 1545 | return -1; |
f3bc08c5 | 1546 | } |
5334a2c5 MD |
1547 | |
1548 | /* | |
1549 | * Need to write the subbuffer start header on finalize. | |
1550 | */ | |
1551 | offsets->switch_old_start = 1; | |
1552 | } | |
1553 | offsets->begin = subbuf_align(offsets->begin, chan); | |
f3bc08c5 MD |
1554 | /* Note: old points to the next subbuf at offset 0 */ |
1555 | offsets->end = offsets->begin; | |
1556 | return 0; | |
1557 | } | |
1558 | ||
1559 | /* | |
1560 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1561 | * called while tracing is active with absolutely no lock held. | |
1562 | * | |
1563 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1564 | * operations, this function must be called from the CPU which owns the buffer | |
1565 | * for a ACTIVE flush. | |
1566 | */ | |
1567 | void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode) | |
1568 | { | |
1569 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1570 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1571 | struct switch_offsets offsets; |
1572 | unsigned long oldidx; | |
1573 | u64 tsc; | |
1574 | ||
1575 | offsets.size = 0; | |
1576 | ||
1577 | /* | |
1578 | * Perform retryable operations. | |
1579 | */ | |
1580 | do { | |
1581 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1582 | &tsc)) | |
1583 | return; /* Switch not needed */ | |
1584 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1585 | != offsets.old); | |
1586 | ||
1587 | /* | |
1588 | * Atomically update last_tsc. This update races against concurrent | |
1589 | * atomic updates, but the race will always cause supplementary full TSC | |
1590 | * records, never the opposite (missing a full TSC record when it would | |
1591 | * be needed). | |
1592 | */ | |
1593 | save_last_tsc(config, buf, tsc); | |
1594 | ||
1595 | /* | |
1596 | * Push the reader if necessary | |
1597 | */ | |
1598 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1599 | ||
1600 | oldidx = subbuf_index(offsets.old, chan); | |
1601 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx); | |
1602 | ||
1603 | /* | |
1604 | * May need to populate header start on SWITCH_FLUSH. | |
1605 | */ | |
1606 | if (offsets.switch_old_start) { | |
1607 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc); | |
1608 | offsets.old += config->cb.subbuffer_header_size(); | |
1609 | } | |
1610 | ||
1611 | /* | |
1612 | * Switch old subbuffer. | |
1613 | */ | |
1614 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc); | |
1615 | } | |
1616 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow); | |
1617 | ||
5e391252 MD |
1618 | static void remote_switch(void *info) |
1619 | { | |
1620 | struct lib_ring_buffer *buf = info; | |
1621 | ||
1622 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
1623 | } | |
1624 | ||
64af2437 MD |
1625 | static void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf, |
1626 | enum switch_mode mode) | |
5e391252 MD |
1627 | { |
1628 | struct channel *chan = buf->backend.chan; | |
1629 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1630 | int ret; | |
1631 | ||
1632 | /* | |
1633 | * With global synchronization we don't need to use the IPI scheme. | |
1634 | */ | |
1635 | if (config->sync == RING_BUFFER_SYNC_GLOBAL) { | |
64af2437 | 1636 | lib_ring_buffer_switch_slow(buf, mode); |
5e391252 MD |
1637 | return; |
1638 | } | |
1639 | ||
1640 | /* | |
1641 | * Taking lock on CPU hotplug to ensure two things: first, that the | |
1642 | * target cpu is not taken concurrently offline while we are within | |
1643 | * smp_call_function_single() (I don't trust that get_cpu() on the | |
1644 | * _local_ CPU actually inhibit CPU hotplug for the _remote_ CPU (to be | |
1645 | * confirmed)). Secondly, if it happens that the CPU is not online, our | |
1646 | * own call to lib_ring_buffer_switch_slow() needs to be protected from | |
1647 | * CPU hotplug handlers, which can also perform a remote subbuffer | |
1648 | * switch. | |
1649 | */ | |
1650 | get_online_cpus(); | |
1651 | ret = smp_call_function_single(buf->backend.cpu, | |
1652 | remote_switch, buf, 1); | |
1653 | if (ret) { | |
1654 | /* Remote CPU is offline, do it ourself. */ | |
64af2437 | 1655 | lib_ring_buffer_switch_slow(buf, mode); |
5e391252 MD |
1656 | } |
1657 | put_online_cpus(); | |
1658 | } | |
64af2437 MD |
1659 | |
1660 | void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf) | |
1661 | { | |
1662 | _lib_ring_buffer_switch_remote(buf, SWITCH_ACTIVE); | |
1663 | } | |
5e391252 MD |
1664 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote); |
1665 | ||
f3bc08c5 MD |
1666 | /* |
1667 | * Returns : | |
1668 | * 0 if ok | |
97ca2c54 MD |
1669 | * -ENOSPC if event size is too large for packet. |
1670 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1671 | * -EIO if data cannot be written into the buffer for any other reason. | |
f3bc08c5 MD |
1672 | */ |
1673 | static | |
1674 | int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, | |
1675 | struct channel *chan, | |
1676 | struct switch_offsets *offsets, | |
1677 | struct lib_ring_buffer_ctx *ctx) | |
1678 | { | |
5a8fd222 | 1679 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
0fdec686 | 1680 | unsigned long reserve_commit_diff, offset_cmp; |
f3bc08c5 | 1681 | |
0fdec686 MD |
1682 | retry: |
1683 | offsets->begin = offset_cmp = v_read(config, &buf->offset); | |
f3bc08c5 MD |
1684 | offsets->old = offsets->begin; |
1685 | offsets->switch_new_start = 0; | |
f5ea5800 | 1686 | offsets->switch_new_end = 0; |
f3bc08c5 MD |
1687 | offsets->switch_old_end = 0; |
1688 | offsets->pre_header_padding = 0; | |
1689 | ||
1690 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
97ca2c54 MD |
1691 | if ((int64_t) ctx->tsc == -EIO) |
1692 | return -EIO; | |
f3bc08c5 MD |
1693 | |
1694 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
64c796d8 | 1695 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; |
f3bc08c5 MD |
1696 | |
1697 | if (unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { | |
1698 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1699 | } else { | |
1700 | offsets->size = config->cb.record_header_size(config, chan, | |
1701 | offsets->begin, | |
f3bc08c5 | 1702 | &offsets->pre_header_padding, |
64c796d8 | 1703 | ctx); |
f3bc08c5 MD |
1704 | offsets->size += |
1705 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1706 | ctx->largest_align) | |
1707 | + ctx->data_size; | |
1708 | if (unlikely(subbuf_offset(offsets->begin, chan) + | |
1709 | offsets->size > chan->backend.subbuf_size)) { | |
1710 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1711 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1712 | } | |
1713 | } | |
1714 | if (unlikely(offsets->switch_new_start)) { | |
0fdec686 | 1715 | unsigned long sb_index, commit_count; |
f3bc08c5 MD |
1716 | |
1717 | /* | |
1718 | * We are typically not filling the previous buffer completely. | |
1719 | */ | |
1720 | if (likely(offsets->switch_old_end)) | |
1721 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1722 | offsets->begin = offsets->begin | |
1723 | + config->cb.subbuffer_header_size(); | |
1724 | /* Test new buffer integrity */ | |
1725 | sb_index = subbuf_index(offsets->begin, chan); | |
0fdec686 MD |
1726 | /* |
1727 | * Read buf->offset before buf->commit_cold[sb_index].cc_sb. | |
1728 | * lib_ring_buffer_check_deliver() has the matching | |
1729 | * memory barriers required around commit_cold cc_sb | |
1730 | * updates to ensure reserve and commit counter updates | |
1731 | * are not seen reordered when updated by another CPU. | |
1732 | */ | |
1733 | smp_rmb(); | |
1734 | commit_count = v_read(config, | |
1735 | &buf->commit_cold[sb_index].cc_sb); | |
1736 | /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */ | |
1737 | smp_rmb(); | |
1738 | if (unlikely(offset_cmp != v_read(config, &buf->offset))) { | |
1739 | /* | |
1740 | * The reserve counter have been concurrently updated | |
1741 | * while we read the commit counter. This means the | |
1742 | * commit counter we read might not match buf->offset | |
1743 | * due to concurrent update. We therefore need to retry. | |
1744 | */ | |
1745 | goto retry; | |
1746 | } | |
f3bc08c5 MD |
1747 | reserve_commit_diff = |
1748 | (buf_trunc(offsets->begin, chan) | |
1749 | >> chan->backend.num_subbuf_order) | |
0fdec686 | 1750 | - (commit_count & chan->commit_count_mask); |
f3bc08c5 MD |
1751 | if (likely(reserve_commit_diff == 0)) { |
1752 | /* Next subbuffer not being written to. */ | |
1753 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1754 | subbuf_trunc(offsets->begin, chan) | |
1755 | - subbuf_trunc((unsigned long) | |
1756 | atomic_long_read(&buf->consumed), chan) | |
1757 | >= chan->backend.buf_size)) { | |
1758 | /* | |
1759 | * We do not overwrite non consumed buffers | |
1760 | * and we are full : record is lost. | |
1761 | */ | |
1762 | v_inc(config, &buf->records_lost_full); | |
97ca2c54 | 1763 | return -ENOBUFS; |
f3bc08c5 MD |
1764 | } else { |
1765 | /* | |
1766 | * Next subbuffer not being written to, and we | |
1767 | * are either in overwrite mode or the buffer is | |
1768 | * not full. It's safe to write in this new | |
1769 | * subbuffer. | |
1770 | */ | |
1771 | } | |
1772 | } else { | |
1773 | /* | |
1774 | * Next subbuffer reserve offset does not match the | |
0fdec686 MD |
1775 | * commit offset, and this did not involve update to the |
1776 | * reserve counter. Drop record in producer-consumer and | |
1777 | * overwrite mode. Caused by either a writer OOPS or | |
1778 | * too many nested writes over a reserve/commit pair. | |
f3bc08c5 MD |
1779 | */ |
1780 | v_inc(config, &buf->records_lost_wrap); | |
97ca2c54 | 1781 | return -EIO; |
f3bc08c5 MD |
1782 | } |
1783 | offsets->size = | |
1784 | config->cb.record_header_size(config, chan, | |
1785 | offsets->begin, | |
f3bc08c5 | 1786 | &offsets->pre_header_padding, |
64c796d8 | 1787 | ctx); |
f3bc08c5 MD |
1788 | offsets->size += |
1789 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1790 | ctx->largest_align) | |
1791 | + ctx->data_size; | |
1792 | if (unlikely(subbuf_offset(offsets->begin, chan) | |
1793 | + offsets->size > chan->backend.subbuf_size)) { | |
1794 | /* | |
1795 | * Record too big for subbuffers, report error, don't | |
1796 | * complete the sub-buffer switch. | |
1797 | */ | |
1798 | v_inc(config, &buf->records_lost_big); | |
97ca2c54 | 1799 | return -ENOSPC; |
f3bc08c5 MD |
1800 | } else { |
1801 | /* | |
1802 | * We just made a successful buffer switch and the | |
1803 | * record fits in the new subbuffer. Let's write. | |
1804 | */ | |
1805 | } | |
1806 | } else { | |
1807 | /* | |
1808 | * Record fits in the current buffer and we are not on a switch | |
1809 | * boundary. It's safe to write. | |
1810 | */ | |
1811 | } | |
1812 | offsets->end = offsets->begin + offsets->size; | |
f5ea5800 MD |
1813 | |
1814 | if (unlikely(subbuf_offset(offsets->end, chan) == 0)) { | |
1815 | /* | |
1816 | * The offset_end will fall at the very beginning of the next | |
1817 | * subbuffer. | |
1818 | */ | |
1819 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1820 | } | |
f3bc08c5 MD |
1821 | return 0; |
1822 | } | |
1823 | ||
d7e74017 MD |
1824 | static struct lib_ring_buffer *get_current_buf(struct channel *chan, int cpu) |
1825 | { | |
1826 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1827 | ||
1828 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1829 | return per_cpu_ptr(chan->backend.buf, cpu); | |
1830 | else | |
1831 | return chan->backend.buf; | |
1832 | } | |
1833 | ||
1834 | void lib_ring_buffer_lost_event_too_big(struct channel *chan) | |
1835 | { | |
1836 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1837 | struct lib_ring_buffer *buf = get_current_buf(chan, smp_processor_id()); | |
1838 | ||
1839 | v_inc(config, &buf->records_lost_big); | |
1840 | } | |
1841 | EXPORT_SYMBOL_GPL(lib_ring_buffer_lost_event_too_big); | |
1842 | ||
f3bc08c5 MD |
1843 | /** |
1844 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
1845 | * @ctx: ring buffer context. | |
1846 | * | |
97ca2c54 MD |
1847 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, |
1848 | * -EIO for other errors, else returns 0. | |
f3bc08c5 MD |
1849 | * It will take care of sub-buffer switching. |
1850 | */ | |
1851 | int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx) | |
1852 | { | |
1853 | struct channel *chan = ctx->chan; | |
5a8fd222 | 1854 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1855 | struct lib_ring_buffer *buf; |
1856 | struct switch_offsets offsets; | |
c099397a | 1857 | int ret; |
f3bc08c5 | 1858 | |
d7e74017 | 1859 | ctx->buf = buf = get_current_buf(chan, ctx->cpu); |
f3bc08c5 MD |
1860 | offsets.size = 0; |
1861 | ||
1862 | do { | |
97ca2c54 MD |
1863 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, |
1864 | ctx); | |
1865 | if (unlikely(ret)) | |
1866 | return ret; | |
f3bc08c5 MD |
1867 | } while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
1868 | offsets.end) | |
1869 | != offsets.old)); | |
1870 | ||
1871 | /* | |
1872 | * Atomically update last_tsc. This update races against concurrent | |
1873 | * atomic updates, but the race will always cause supplementary full TSC | |
1874 | * records, never the opposite (missing a full TSC record when it would | |
1875 | * be needed). | |
1876 | */ | |
1877 | save_last_tsc(config, buf, ctx->tsc); | |
1878 | ||
1879 | /* | |
1880 | * Push the reader if necessary | |
1881 | */ | |
1882 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
1883 | ||
1884 | /* | |
1885 | * Clear noref flag for this subbuffer. | |
1886 | */ | |
1887 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1888 | subbuf_index(offsets.end - 1, chan)); | |
1889 | ||
1890 | /* | |
1891 | * Switch old subbuffer if needed. | |
1892 | */ | |
1893 | if (unlikely(offsets.switch_old_end)) { | |
1894 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
1895 | subbuf_index(offsets.old - 1, chan)); | |
1896 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc); | |
1897 | } | |
1898 | ||
1899 | /* | |
1900 | * Populate new subbuffer. | |
1901 | */ | |
1902 | if (unlikely(offsets.switch_new_start)) | |
1903 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc); | |
1904 | ||
f5ea5800 MD |
1905 | if (unlikely(offsets.switch_new_end)) |
1906 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc); | |
1907 | ||
f3bc08c5 MD |
1908 | ctx->slot_size = offsets.size; |
1909 | ctx->pre_offset = offsets.begin; | |
1910 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
1911 | return 0; | |
1912 | } | |
1913 | EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow); | |
6fb8de4b | 1914 | |
02a766bb | 1915 | int __init init_lib_ring_buffer_frontend(void) |
6fb8de4b MD |
1916 | { |
1917 | int cpu; | |
1918 | ||
1919 | for_each_possible_cpu(cpu) | |
1920 | spin_lock_init(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
02a766bb | 1921 | return 0; |
6fb8de4b | 1922 | } |
02a766bb MD |
1923 | |
1924 | module_init(init_lib_ring_buffer_frontend); | |
1a5db82d MD |
1925 | |
1926 | void __exit exit_lib_ring_buffer_frontend(void) | |
1927 | { | |
1928 | } | |
1929 | ||
1930 | module_exit(exit_lib_ring_buffer_frontend); |