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> | |
152fe7fc | 66 | #include <wrapper/timer.h> |
f3bc08c5 MD |
67 | |
68 | /* | |
69 | * Internal structure representing offsets to use at a sub-buffer switch. | |
70 | */ | |
71 | struct switch_offsets { | |
72 | unsigned long begin, end, old; | |
73 | size_t pre_header_padding, size; | |
f5ea5800 MD |
74 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, |
75 | switch_old_end:1; | |
f3bc08c5 MD |
76 | }; |
77 | ||
78 | #ifdef CONFIG_NO_HZ | |
79 | enum tick_nohz_val { | |
80 | TICK_NOHZ_STOP, | |
81 | TICK_NOHZ_FLUSH, | |
82 | TICK_NOHZ_RESTART, | |
83 | }; | |
84 | ||
85 | static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier); | |
86 | #endif /* CONFIG_NO_HZ */ | |
87 | ||
88 | static DEFINE_PER_CPU(spinlock_t, ring_buffer_nohz_lock); | |
89 | ||
90 | DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting); | |
91 | EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting); | |
92 | ||
93 | static | |
94 | void lib_ring_buffer_print_errors(struct channel *chan, | |
95 | struct lib_ring_buffer *buf, int cpu); | |
64af2437 MD |
96 | static |
97 | void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf, | |
98 | enum switch_mode mode); | |
f3bc08c5 | 99 | |
aece661f MD |
100 | static |
101 | int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config, | |
102 | struct lib_ring_buffer *buf, | |
103 | struct channel *chan) | |
104 | { | |
105 | unsigned long consumed_old, consumed_idx, commit_count, write_offset; | |
106 | ||
107 | consumed_old = atomic_long_read(&buf->consumed); | |
108 | consumed_idx = subbuf_index(consumed_old, chan); | |
109 | commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); | |
110 | /* | |
111 | * No memory barrier here, since we are only interested | |
112 | * in a statistically correct polling result. The next poll will | |
113 | * get the data is we are racing. The mb() that ensures correct | |
114 | * memory order is in get_subbuf. | |
115 | */ | |
116 | write_offset = v_read(config, &buf->offset); | |
117 | ||
118 | /* | |
119 | * Check that the subbuffer we are trying to consume has been | |
120 | * already fully committed. | |
121 | */ | |
122 | ||
123 | if (((commit_count - chan->backend.subbuf_size) | |
124 | & chan->commit_count_mask) | |
125 | - (buf_trunc(consumed_old, chan) | |
126 | >> chan->backend.num_subbuf_order) | |
127 | != 0) | |
128 | return 0; | |
129 | ||
130 | /* | |
131 | * Check that we are not about to read the same subbuffer in | |
132 | * which the writer head is. | |
133 | */ | |
134 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan) | |
135 | == 0) | |
136 | return 0; | |
137 | ||
138 | return 1; | |
139 | } | |
140 | ||
f3bc08c5 MD |
141 | /* |
142 | * Must be called under cpu hotplug protection. | |
143 | */ | |
144 | void lib_ring_buffer_free(struct lib_ring_buffer *buf) | |
145 | { | |
146 | struct channel *chan = buf->backend.chan; | |
147 | ||
148 | lib_ring_buffer_print_errors(chan, buf, buf->backend.cpu); | |
149 | kfree(buf->commit_hot); | |
150 | kfree(buf->commit_cold); | |
151 | ||
152 | lib_ring_buffer_backend_free(&buf->backend); | |
153 | } | |
154 | ||
155 | /** | |
156 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
157 | * @buf: Ring buffer. | |
158 | * | |
159 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
160 | * used for writing. The ring buffer can be opened for reading, but the reader | |
161 | * should not be using the iterator concurrently with reset. The previous | |
162 | * current iterator record is reset. | |
163 | */ | |
164 | void lib_ring_buffer_reset(struct lib_ring_buffer *buf) | |
165 | { | |
166 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 167 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
168 | unsigned int i; |
169 | ||
170 | /* | |
171 | * Reset iterator first. It will put the subbuffer if it currently holds | |
172 | * it. | |
173 | */ | |
174 | lib_ring_buffer_iterator_reset(buf); | |
175 | v_set(config, &buf->offset, 0); | |
176 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
177 | v_set(config, &buf->commit_hot[i].cc, 0); | |
178 | v_set(config, &buf->commit_hot[i].seq, 0); | |
179 | v_set(config, &buf->commit_cold[i].cc_sb, 0); | |
180 | } | |
181 | atomic_long_set(&buf->consumed, 0); | |
182 | atomic_set(&buf->record_disabled, 0); | |
183 | v_set(config, &buf->last_tsc, 0); | |
184 | lib_ring_buffer_backend_reset(&buf->backend); | |
185 | /* Don't reset number of active readers */ | |
186 | v_set(config, &buf->records_lost_full, 0); | |
187 | v_set(config, &buf->records_lost_wrap, 0); | |
188 | v_set(config, &buf->records_lost_big, 0); | |
189 | v_set(config, &buf->records_count, 0); | |
190 | v_set(config, &buf->records_overrun, 0); | |
191 | buf->finalized = 0; | |
192 | } | |
193 | EXPORT_SYMBOL_GPL(lib_ring_buffer_reset); | |
194 | ||
195 | /** | |
196 | * channel_reset - Reset channel to initial values. | |
197 | * @chan: Channel. | |
198 | * | |
199 | * Effectively empty the channel. Should be called when the channel is not used | |
200 | * for writing. The channel can be opened for reading, but the reader should not | |
201 | * be using the iterator concurrently with reset. The previous current iterator | |
202 | * record is reset. | |
203 | */ | |
204 | void channel_reset(struct channel *chan) | |
205 | { | |
206 | /* | |
207 | * Reset iterators first. Will put the subbuffer if held for reading. | |
208 | */ | |
209 | channel_iterator_reset(chan); | |
210 | atomic_set(&chan->record_disabled, 0); | |
211 | /* Don't reset commit_count_mask, still valid */ | |
212 | channel_backend_reset(&chan->backend); | |
213 | /* Don't reset switch/read timer interval */ | |
214 | /* Don't reset notifiers and notifier enable bits */ | |
215 | /* Don't reset reader reference count */ | |
216 | } | |
217 | EXPORT_SYMBOL_GPL(channel_reset); | |
218 | ||
219 | /* | |
220 | * Must be called under cpu hotplug protection. | |
221 | */ | |
222 | int lib_ring_buffer_create(struct lib_ring_buffer *buf, | |
223 | struct channel_backend *chanb, int cpu) | |
224 | { | |
5a8fd222 | 225 | const struct lib_ring_buffer_config *config = &chanb->config; |
f3bc08c5 MD |
226 | struct channel *chan = container_of(chanb, struct channel, backend); |
227 | void *priv = chanb->priv; | |
f3bc08c5 MD |
228 | size_t subbuf_header_size; |
229 | u64 tsc; | |
230 | int ret; | |
231 | ||
232 | /* Test for cpu hotplug */ | |
233 | if (buf->backend.allocated) | |
234 | return 0; | |
235 | ||
236 | /* | |
237 | * Paranoia: per cpu dynamic allocation is not officially documented as | |
238 | * zeroing the memory, so let's do it here too, just in case. | |
239 | */ | |
240 | memset(buf, 0, sizeof(*buf)); | |
241 | ||
242 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, cpu); | |
243 | if (ret) | |
244 | return ret; | |
245 | ||
246 | buf->commit_hot = | |
247 | kzalloc_node(ALIGN(sizeof(*buf->commit_hot) | |
248 | * chan->backend.num_subbuf, | |
249 | 1 << INTERNODE_CACHE_SHIFT), | |
df388b78 MD |
250 | GFP_KERNEL | __GFP_NOWARN, |
251 | cpu_to_node(max(cpu, 0))); | |
f3bc08c5 MD |
252 | if (!buf->commit_hot) { |
253 | ret = -ENOMEM; | |
254 | goto free_chanbuf; | |
255 | } | |
256 | ||
257 | buf->commit_cold = | |
258 | kzalloc_node(ALIGN(sizeof(*buf->commit_cold) | |
259 | * chan->backend.num_subbuf, | |
260 | 1 << INTERNODE_CACHE_SHIFT), | |
df388b78 MD |
261 | GFP_KERNEL | __GFP_NOWARN, |
262 | cpu_to_node(max(cpu, 0))); | |
f3bc08c5 MD |
263 | if (!buf->commit_cold) { |
264 | ret = -ENOMEM; | |
265 | goto free_commit; | |
266 | } | |
267 | ||
f3bc08c5 | 268 | init_waitqueue_head(&buf->read_wait); |
71c1d843 | 269 | init_waitqueue_head(&buf->write_wait); |
f3bc08c5 MD |
270 | raw_spin_lock_init(&buf->raw_tick_nohz_spinlock); |
271 | ||
272 | /* | |
273 | * Write the subbuffer header for first subbuffer so we know the total | |
274 | * duration of data gathering. | |
275 | */ | |
276 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
277 | v_set(config, &buf->offset, subbuf_header_size); | |
278 | subbuffer_id_clear_noref(config, &buf->backend.buf_wsb[0].id); | |
279 | tsc = config->cb.ring_buffer_clock_read(buf->backend.chan); | |
280 | config->cb.buffer_begin(buf, tsc, 0); | |
281 | v_add(config, subbuf_header_size, &buf->commit_hot[0].cc); | |
282 | ||
283 | if (config->cb.buffer_create) { | |
284 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name); | |
285 | if (ret) | |
286 | goto free_init; | |
287 | } | |
288 | ||
289 | /* | |
290 | * Ensure the buffer is ready before setting it to allocated and setting | |
291 | * the cpumask. | |
292 | * Used for cpu hotplug vs cpumask iteration. | |
293 | */ | |
294 | smp_wmb(); | |
295 | buf->backend.allocated = 1; | |
296 | ||
297 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
298 | CHAN_WARN_ON(chan, cpumask_test_cpu(cpu, | |
299 | chan->backend.cpumask)); | |
300 | cpumask_set_cpu(cpu, chan->backend.cpumask); | |
301 | } | |
302 | ||
303 | return 0; | |
304 | ||
305 | /* Error handling */ | |
306 | free_init: | |
307 | kfree(buf->commit_cold); | |
308 | free_commit: | |
309 | kfree(buf->commit_hot); | |
310 | free_chanbuf: | |
311 | lib_ring_buffer_backend_free(&buf->backend); | |
312 | return ret; | |
313 | } | |
314 | ||
315 | static void switch_buffer_timer(unsigned long data) | |
316 | { | |
317 | struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data; | |
318 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 319 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
320 | |
321 | /* | |
322 | * Only flush buffers periodically if readers are active. | |
323 | */ | |
324 | if (atomic_long_read(&buf->active_readers)) | |
325 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
326 | ||
327 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
152fe7fc | 328 | lttng_mod_timer_pinned(&buf->switch_timer, |
f3bc08c5 MD |
329 | jiffies + chan->switch_timer_interval); |
330 | else | |
331 | mod_timer(&buf->switch_timer, | |
332 | jiffies + chan->switch_timer_interval); | |
333 | } | |
334 | ||
335 | /* | |
336 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
337 | */ | |
338 | static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer *buf) | |
339 | { | |
340 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 341 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
342 | |
343 | if (!chan->switch_timer_interval || buf->switch_timer_enabled) | |
344 | return; | |
152fe7fc MJ |
345 | |
346 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
347 | lttng_init_timer_pinned(&buf->switch_timer); | |
348 | else | |
349 | init_timer(&buf->switch_timer); | |
350 | ||
f3bc08c5 MD |
351 | buf->switch_timer.function = switch_buffer_timer; |
352 | buf->switch_timer.expires = jiffies + chan->switch_timer_interval; | |
353 | buf->switch_timer.data = (unsigned long)buf; | |
354 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
355 | add_timer_on(&buf->switch_timer, buf->backend.cpu); | |
356 | else | |
357 | add_timer(&buf->switch_timer); | |
358 | buf->switch_timer_enabled = 1; | |
359 | } | |
360 | ||
361 | /* | |
362 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
363 | */ | |
364 | static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer *buf) | |
365 | { | |
366 | struct channel *chan = buf->backend.chan; | |
367 | ||
368 | if (!chan->switch_timer_interval || !buf->switch_timer_enabled) | |
369 | return; | |
370 | ||
371 | del_timer_sync(&buf->switch_timer); | |
372 | buf->switch_timer_enabled = 0; | |
373 | } | |
374 | ||
375 | /* | |
376 | * Polling timer to check the channels for data. | |
377 | */ | |
378 | static void read_buffer_timer(unsigned long data) | |
379 | { | |
380 | struct lib_ring_buffer *buf = (struct lib_ring_buffer *)data; | |
381 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 382 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
383 | |
384 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
385 | ||
386 | if (atomic_long_read(&buf->active_readers) | |
387 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
388 | wake_up_interruptible(&buf->read_wait); | |
389 | wake_up_interruptible(&chan->read_wait); | |
390 | } | |
391 | ||
392 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
152fe7fc | 393 | lttng_mod_timer_pinned(&buf->read_timer, |
f3bc08c5 MD |
394 | jiffies + chan->read_timer_interval); |
395 | else | |
396 | mod_timer(&buf->read_timer, | |
397 | jiffies + chan->read_timer_interval); | |
398 | } | |
399 | ||
400 | /* | |
401 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
402 | */ | |
403 | static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf) | |
404 | { | |
405 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 406 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
407 | |
408 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
409 | || !chan->read_timer_interval | |
410 | || buf->read_timer_enabled) | |
411 | return; | |
412 | ||
152fe7fc MJ |
413 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
414 | lttng_init_timer_pinned(&buf->read_timer); | |
415 | else | |
416 | init_timer(&buf->read_timer); | |
417 | ||
f3bc08c5 MD |
418 | buf->read_timer.function = read_buffer_timer; |
419 | buf->read_timer.expires = jiffies + chan->read_timer_interval; | |
420 | buf->read_timer.data = (unsigned long)buf; | |
421 | ||
422 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
423 | add_timer_on(&buf->read_timer, buf->backend.cpu); | |
424 | else | |
425 | add_timer(&buf->read_timer); | |
426 | buf->read_timer_enabled = 1; | |
427 | } | |
428 | ||
429 | /* | |
430 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
431 | */ | |
432 | static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer *buf) | |
433 | { | |
434 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 435 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
436 | |
437 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
438 | || !chan->read_timer_interval | |
439 | || !buf->read_timer_enabled) | |
440 | return; | |
441 | ||
442 | del_timer_sync(&buf->read_timer); | |
443 | /* | |
444 | * do one more check to catch data that has been written in the last | |
445 | * timer period. | |
446 | */ | |
447 | if (lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
448 | wake_up_interruptible(&buf->read_wait); | |
449 | wake_up_interruptible(&chan->read_wait); | |
450 | } | |
451 | buf->read_timer_enabled = 0; | |
452 | } | |
453 | ||
1e367326 MD |
454 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
455 | ||
456 | enum cpuhp_state lttng_rb_hp_prepare; | |
457 | enum cpuhp_state lttng_rb_hp_online; | |
458 | ||
459 | void lttng_rb_set_hp_prepare(enum cpuhp_state val) | |
460 | { | |
461 | lttng_rb_hp_prepare = val; | |
462 | } | |
463 | EXPORT_SYMBOL_GPL(lttng_rb_set_hp_prepare); | |
464 | ||
465 | void lttng_rb_set_hp_online(enum cpuhp_state val) | |
466 | { | |
467 | lttng_rb_hp_online = val; | |
468 | } | |
469 | EXPORT_SYMBOL_GPL(lttng_rb_set_hp_online); | |
470 | ||
471 | int lttng_cpuhp_rb_frontend_dead(unsigned int cpu, | |
472 | struct lttng_cpuhp_node *node) | |
473 | { | |
474 | struct channel *chan = container_of(node, struct channel, | |
475 | cpuhp_prepare); | |
476 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
477 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
478 | ||
479 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
480 | ||
481 | /* | |
482 | * Performing a buffer switch on a remote CPU. Performed by | |
483 | * the CPU responsible for doing the hotunplug after the target | |
484 | * CPU stopped running completely. Ensures that all data | |
485 | * from that remote CPU is flushed. | |
486 | */ | |
487 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
488 | return 0; | |
489 | } | |
490 | EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_dead); | |
491 | ||
492 | int lttng_cpuhp_rb_frontend_online(unsigned int cpu, | |
493 | struct lttng_cpuhp_node *node) | |
494 | { | |
495 | struct channel *chan = container_of(node, struct channel, | |
496 | cpuhp_online); | |
497 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
498 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
499 | ||
500 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
501 | ||
502 | wake_up_interruptible(&chan->hp_wait); | |
503 | lib_ring_buffer_start_switch_timer(buf); | |
504 | lib_ring_buffer_start_read_timer(buf); | |
505 | return 0; | |
506 | } | |
507 | EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_online); | |
508 | ||
509 | int lttng_cpuhp_rb_frontend_offline(unsigned int cpu, | |
510 | struct lttng_cpuhp_node *node) | |
511 | { | |
512 | struct channel *chan = container_of(node, struct channel, | |
513 | cpuhp_online); | |
514 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
515 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
516 | ||
517 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
518 | ||
519 | lib_ring_buffer_stop_switch_timer(buf); | |
520 | lib_ring_buffer_stop_read_timer(buf); | |
521 | return 0; | |
522 | } | |
523 | EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_offline); | |
524 | ||
525 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ | |
526 | ||
f3bc08c5 | 527 | #ifdef CONFIG_HOTPLUG_CPU |
1e367326 | 528 | |
f3bc08c5 MD |
529 | /** |
530 | * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback | |
531 | * @nb: notifier block | |
532 | * @action: hotplug action to take | |
533 | * @hcpu: CPU number | |
534 | * | |
535 | * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD) | |
536 | */ | |
537 | static | |
e8f071d5 | 538 | int lib_ring_buffer_cpu_hp_callback(struct notifier_block *nb, |
f3bc08c5 MD |
539 | unsigned long action, |
540 | void *hcpu) | |
541 | { | |
542 | unsigned int cpu = (unsigned long)hcpu; | |
543 | struct channel *chan = container_of(nb, struct channel, | |
544 | cpu_hp_notifier); | |
545 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
5a8fd222 | 546 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
547 | |
548 | if (!chan->cpu_hp_enable) | |
549 | return NOTIFY_DONE; | |
550 | ||
551 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
552 | ||
553 | switch (action) { | |
554 | case CPU_DOWN_FAILED: | |
555 | case CPU_DOWN_FAILED_FROZEN: | |
556 | case CPU_ONLINE: | |
557 | case CPU_ONLINE_FROZEN: | |
24cedcfe | 558 | wake_up_interruptible(&chan->hp_wait); |
f3bc08c5 MD |
559 | lib_ring_buffer_start_switch_timer(buf); |
560 | lib_ring_buffer_start_read_timer(buf); | |
561 | return NOTIFY_OK; | |
562 | ||
563 | case CPU_DOWN_PREPARE: | |
564 | case CPU_DOWN_PREPARE_FROZEN: | |
565 | lib_ring_buffer_stop_switch_timer(buf); | |
566 | lib_ring_buffer_stop_read_timer(buf); | |
567 | return NOTIFY_OK; | |
568 | ||
569 | case CPU_DEAD: | |
570 | case CPU_DEAD_FROZEN: | |
571 | /* | |
572 | * Performing a buffer switch on a remote CPU. Performed by | |
573 | * the CPU responsible for doing the hotunplug after the target | |
574 | * CPU stopped running completely. Ensures that all data | |
575 | * from that remote CPU is flushed. | |
576 | */ | |
577 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
578 | return NOTIFY_OK; | |
579 | ||
580 | default: | |
581 | return NOTIFY_DONE; | |
582 | } | |
583 | } | |
1e367326 | 584 | |
f3bc08c5 MD |
585 | #endif |
586 | ||
1e367326 MD |
587 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
588 | ||
23b908b0 | 589 | #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) |
f3bc08c5 MD |
590 | /* |
591 | * For per-cpu buffers, call the reader wakeups before switching the buffer, so | |
592 | * that wake-up-tracing generated events are flushed before going idle (in | |
593 | * tick_nohz). We test if the spinlock is locked to deal with the race where | |
594 | * readers try to sample the ring buffer before we perform the switch. We let | |
595 | * the readers retry in that case. If there is data in the buffer, the wake up | |
596 | * is going to forbid the CPU running the reader thread from going idle. | |
597 | */ | |
598 | static int notrace ring_buffer_tick_nohz_callback(struct notifier_block *nb, | |
599 | unsigned long val, | |
600 | void *data) | |
601 | { | |
602 | struct channel *chan = container_of(nb, struct channel, | |
603 | tick_nohz_notifier); | |
5a8fd222 | 604 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
605 | struct lib_ring_buffer *buf; |
606 | int cpu = smp_processor_id(); | |
607 | ||
608 | if (config->alloc != RING_BUFFER_ALLOC_PER_CPU) { | |
609 | /* | |
610 | * We don't support keeping the system idle with global buffers | |
611 | * and streaming active. In order to do so, we would need to | |
612 | * sample a non-nohz-cpumask racelessly with the nohz updates | |
613 | * without adding synchronization overhead to nohz. Leave this | |
614 | * use-case out for now. | |
615 | */ | |
616 | return 0; | |
617 | } | |
618 | ||
619 | buf = channel_get_ring_buffer(config, chan, cpu); | |
620 | switch (val) { | |
621 | case TICK_NOHZ_FLUSH: | |
622 | raw_spin_lock(&buf->raw_tick_nohz_spinlock); | |
623 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_TIMER | |
624 | && chan->read_timer_interval | |
625 | && atomic_long_read(&buf->active_readers) | |
626 | && (lib_ring_buffer_poll_deliver(config, buf, chan) | |
627 | || lib_ring_buffer_pending_data(config, buf, chan))) { | |
628 | wake_up_interruptible(&buf->read_wait); | |
629 | wake_up_interruptible(&chan->read_wait); | |
630 | } | |
631 | if (chan->switch_timer_interval) | |
632 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
633 | raw_spin_unlock(&buf->raw_tick_nohz_spinlock); | |
634 | break; | |
635 | case TICK_NOHZ_STOP: | |
e6b06d7d | 636 | spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
637 | lib_ring_buffer_stop_switch_timer(buf); |
638 | lib_ring_buffer_stop_read_timer(buf); | |
e6b06d7d | 639 | spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
640 | break; |
641 | case TICK_NOHZ_RESTART: | |
e6b06d7d | 642 | spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
643 | lib_ring_buffer_start_read_timer(buf); |
644 | lib_ring_buffer_start_switch_timer(buf); | |
e6b06d7d | 645 | spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
646 | break; |
647 | } | |
648 | ||
649 | return 0; | |
650 | } | |
651 | ||
652 | void notrace lib_ring_buffer_tick_nohz_flush(void) | |
653 | { | |
654 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_FLUSH, | |
655 | NULL); | |
656 | } | |
657 | ||
658 | void notrace lib_ring_buffer_tick_nohz_stop(void) | |
659 | { | |
660 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_STOP, | |
661 | NULL); | |
662 | } | |
663 | ||
664 | void notrace lib_ring_buffer_tick_nohz_restart(void) | |
665 | { | |
666 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_RESTART, | |
667 | NULL); | |
668 | } | |
23b908b0 | 669 | #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */ |
f3bc08c5 MD |
670 | |
671 | /* | |
672 | * Holds CPU hotplug. | |
673 | */ | |
674 | static void channel_unregister_notifiers(struct channel *chan) | |
675 | { | |
5a8fd222 | 676 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
677 | |
678 | channel_iterator_unregister_notifiers(chan); | |
679 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
680 | #ifdef CONFIG_NO_HZ | |
681 | /* | |
682 | * Remove the nohz notifier first, so we are certain we stop | |
683 | * the timers. | |
684 | */ | |
685 | atomic_notifier_chain_unregister(&tick_nohz_notifier, | |
686 | &chan->tick_nohz_notifier); | |
687 | /* | |
688 | * ring_buffer_nohz_lock will not be needed below, because | |
689 | * we just removed the notifiers, which were the only source of | |
690 | * concurrency. | |
691 | */ | |
692 | #endif /* CONFIG_NO_HZ */ | |
1e367326 MD |
693 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
694 | { | |
695 | int ret; | |
696 | ||
697 | ret = cpuhp_state_remove_instance(lttng_rb_hp_online, | |
698 | &chan->cpuhp_online.node); | |
699 | WARN_ON(ret); | |
700 | ret = cpuhp_state_remove_instance_nocalls(lttng_rb_hp_prepare, | |
701 | &chan->cpuhp_prepare.node); | |
702 | WARN_ON(ret); | |
f3bc08c5 | 703 | } |
1e367326 MD |
704 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ |
705 | { | |
706 | int cpu; | |
707 | ||
708 | #ifdef CONFIG_HOTPLUG_CPU | |
709 | get_online_cpus(); | |
710 | chan->cpu_hp_enable = 0; | |
711 | for_each_online_cpu(cpu) { | |
712 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
713 | cpu); | |
714 | lib_ring_buffer_stop_switch_timer(buf); | |
715 | lib_ring_buffer_stop_read_timer(buf); | |
716 | } | |
717 | put_online_cpus(); | |
718 | unregister_cpu_notifier(&chan->cpu_hp_notifier); | |
f3bc08c5 | 719 | #else |
1e367326 MD |
720 | for_each_possible_cpu(cpu) { |
721 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
722 | cpu); | |
723 | lib_ring_buffer_stop_switch_timer(buf); | |
724 | lib_ring_buffer_stop_read_timer(buf); | |
725 | } | |
f3bc08c5 | 726 | #endif |
1e367326 MD |
727 | } |
728 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ | |
f3bc08c5 MD |
729 | } else { |
730 | struct lib_ring_buffer *buf = chan->backend.buf; | |
731 | ||
732 | lib_ring_buffer_stop_switch_timer(buf); | |
733 | lib_ring_buffer_stop_read_timer(buf); | |
734 | } | |
735 | channel_backend_unregister_notifiers(&chan->backend); | |
736 | } | |
737 | ||
64af2437 MD |
738 | static void lib_ring_buffer_set_quiescent(struct lib_ring_buffer *buf) |
739 | { | |
740 | if (!buf->quiescent) { | |
741 | buf->quiescent = true; | |
742 | _lib_ring_buffer_switch_remote(buf, SWITCH_FLUSH); | |
743 | } | |
744 | } | |
745 | ||
746 | static void lib_ring_buffer_clear_quiescent(struct lib_ring_buffer *buf) | |
747 | { | |
748 | buf->quiescent = false; | |
749 | } | |
750 | ||
751 | void lib_ring_buffer_set_quiescent_channel(struct channel *chan) | |
752 | { | |
753 | int cpu; | |
754 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
755 | ||
756 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
757 | get_online_cpus(); | |
758 | for_each_channel_cpu(cpu, chan) { | |
759 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
760 | cpu); | |
761 | ||
762 | lib_ring_buffer_set_quiescent(buf); | |
763 | } | |
764 | put_online_cpus(); | |
765 | } else { | |
766 | struct lib_ring_buffer *buf = chan->backend.buf; | |
767 | ||
768 | lib_ring_buffer_set_quiescent(buf); | |
769 | } | |
770 | } | |
771 | EXPORT_SYMBOL_GPL(lib_ring_buffer_set_quiescent_channel); | |
772 | ||
773 | void lib_ring_buffer_clear_quiescent_channel(struct channel *chan) | |
774 | { | |
775 | int cpu; | |
776 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
777 | ||
778 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
779 | get_online_cpus(); | |
780 | for_each_channel_cpu(cpu, chan) { | |
781 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
782 | cpu); | |
783 | ||
784 | lib_ring_buffer_clear_quiescent(buf); | |
785 | } | |
786 | put_online_cpus(); | |
787 | } else { | |
788 | struct lib_ring_buffer *buf = chan->backend.buf; | |
789 | ||
790 | lib_ring_buffer_clear_quiescent(buf); | |
791 | } | |
792 | } | |
793 | EXPORT_SYMBOL_GPL(lib_ring_buffer_clear_quiescent_channel); | |
794 | ||
f3bc08c5 MD |
795 | static void channel_free(struct channel *chan) |
796 | { | |
dd5a0db3 MD |
797 | if (chan->backend.release_priv_ops) { |
798 | chan->backend.release_priv_ops(chan->backend.priv_ops); | |
799 | } | |
f3bc08c5 MD |
800 | channel_iterator_free(chan); |
801 | channel_backend_free(&chan->backend); | |
802 | kfree(chan); | |
803 | } | |
804 | ||
805 | /** | |
806 | * channel_create - Create channel. | |
807 | * @config: ring buffer instance configuration | |
808 | * @name: name of the channel | |
809 | * @priv: ring buffer client private data | |
810 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous | |
811 | * address mapping. It is used only by RING_BUFFER_STATIC | |
812 | * configuration. It can be set to NULL for other backends. | |
813 | * @subbuf_size: subbuffer size | |
814 | * @num_subbuf: number of subbuffers | |
815 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
816 | * padding to let readers get those sub-buffers. | |
817 | * Used for live streaming. | |
818 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
819 | * | |
820 | * Holds cpu hotplug. | |
821 | * Returns NULL on failure. | |
822 | */ | |
823 | struct channel *channel_create(const struct lib_ring_buffer_config *config, | |
824 | const char *name, void *priv, void *buf_addr, | |
825 | size_t subbuf_size, | |
826 | size_t num_subbuf, unsigned int switch_timer_interval, | |
827 | unsigned int read_timer_interval) | |
828 | { | |
1e367326 | 829 | int ret; |
f3bc08c5 MD |
830 | struct channel *chan; |
831 | ||
832 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
833 | read_timer_interval)) | |
834 | return NULL; | |
835 | ||
836 | chan = kzalloc(sizeof(struct channel), GFP_KERNEL); | |
837 | if (!chan) | |
838 | return NULL; | |
839 | ||
840 | ret = channel_backend_init(&chan->backend, name, config, priv, | |
841 | subbuf_size, num_subbuf); | |
842 | if (ret) | |
843 | goto error; | |
844 | ||
845 | ret = channel_iterator_init(chan); | |
846 | if (ret) | |
847 | goto error_free_backend; | |
848 | ||
849 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
850 | chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
851 | chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
f40270ad | 852 | kref_init(&chan->ref); |
f3bc08c5 | 853 | init_waitqueue_head(&chan->read_wait); |
24cedcfe | 854 | init_waitqueue_head(&chan->hp_wait); |
f3bc08c5 MD |
855 | |
856 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
1e367326 MD |
857 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
858 | chan->cpuhp_prepare.component = LTTNG_RING_BUFFER_FRONTEND; | |
859 | ret = cpuhp_state_add_instance_nocalls(lttng_rb_hp_prepare, | |
860 | &chan->cpuhp_prepare.node); | |
861 | if (ret) | |
862 | goto cpuhp_prepare_error; | |
863 | ||
864 | chan->cpuhp_online.component = LTTNG_RING_BUFFER_FRONTEND; | |
865 | ret = cpuhp_state_add_instance(lttng_rb_hp_online, | |
866 | &chan->cpuhp_online.node); | |
867 | if (ret) | |
868 | goto cpuhp_online_error; | |
869 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ | |
870 | { | |
871 | int cpu; | |
872 | /* | |
873 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
874 | * in early initcall, it will not be notified of secondary cpus. | |
875 | * In that off case, we need to allocate for all possible cpus. | |
876 | */ | |
877 | #ifdef CONFIG_HOTPLUG_CPU | |
878 | chan->cpu_hp_notifier.notifier_call = | |
879 | lib_ring_buffer_cpu_hp_callback; | |
880 | chan->cpu_hp_notifier.priority = 6; | |
881 | register_cpu_notifier(&chan->cpu_hp_notifier); | |
882 | ||
883 | get_online_cpus(); | |
884 | for_each_online_cpu(cpu) { | |
885 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
886 | cpu); | |
887 | spin_lock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
888 | lib_ring_buffer_start_switch_timer(buf); | |
889 | lib_ring_buffer_start_read_timer(buf); | |
890 | spin_unlock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
891 | } | |
892 | chan->cpu_hp_enable = 1; | |
893 | put_online_cpus(); | |
894 | #else | |
895 | for_each_possible_cpu(cpu) { | |
896 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
897 | cpu); | |
898 | spin_lock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
899 | lib_ring_buffer_start_switch_timer(buf); | |
900 | lib_ring_buffer_start_read_timer(buf); | |
901 | spin_unlock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
902 | } | |
903 | #endif | |
904 | } | |
905 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ | |
906 | ||
23b908b0 | 907 | #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) |
f3bc08c5 MD |
908 | /* Only benefit from NO_HZ idle with per-cpu buffers for now. */ |
909 | chan->tick_nohz_notifier.notifier_call = | |
910 | ring_buffer_tick_nohz_callback; | |
911 | chan->tick_nohz_notifier.priority = ~0U; | |
912 | atomic_notifier_chain_register(&tick_nohz_notifier, | |
913 | &chan->tick_nohz_notifier); | |
23b908b0 | 914 | #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */ |
f3bc08c5 | 915 | |
f3bc08c5 MD |
916 | } else { |
917 | struct lib_ring_buffer *buf = chan->backend.buf; | |
918 | ||
919 | lib_ring_buffer_start_switch_timer(buf); | |
920 | lib_ring_buffer_start_read_timer(buf); | |
921 | } | |
922 | ||
923 | return chan; | |
924 | ||
1e367326 MD |
925 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) |
926 | cpuhp_online_error: | |
927 | ret = cpuhp_state_remove_instance_nocalls(lttng_rb_hp_prepare, | |
928 | &chan->cpuhp_prepare.node); | |
929 | WARN_ON(ret); | |
930 | cpuhp_prepare_error: | |
931 | #endif /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,10,0)) */ | |
f3bc08c5 MD |
932 | error_free_backend: |
933 | channel_backend_free(&chan->backend); | |
934 | error: | |
935 | kfree(chan); | |
936 | return NULL; | |
937 | } | |
938 | EXPORT_SYMBOL_GPL(channel_create); | |
939 | ||
f40270ad MD |
940 | static |
941 | void channel_release(struct kref *kref) | |
942 | { | |
943 | struct channel *chan = container_of(kref, struct channel, ref); | |
944 | channel_free(chan); | |
945 | } | |
946 | ||
f3bc08c5 MD |
947 | /** |
948 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
949 | * @chan: channel to destroy | |
950 | * | |
951 | * Holds cpu hotplug. | |
9a0df743 MD |
952 | * Call "destroy" callback, finalize channels, and then decrement the |
953 | * channel reference count. Note that when readers have completed data | |
954 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
955 | * They should release their handle at that point. Returns the private | |
956 | * data pointer. | |
f3bc08c5 MD |
957 | */ |
958 | void *channel_destroy(struct channel *chan) | |
959 | { | |
960 | int cpu; | |
5a8fd222 | 961 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
962 | void *priv; |
963 | ||
964 | channel_unregister_notifiers(chan); | |
965 | ||
966 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
967 | /* | |
968 | * No need to hold cpu hotplug, because all notifiers have been | |
969 | * unregistered. | |
970 | */ | |
971 | for_each_channel_cpu(cpu, chan) { | |
972 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
973 | cpu); | |
974 | ||
975 | if (config->cb.buffer_finalize) | |
976 | config->cb.buffer_finalize(buf, | |
977 | chan->backend.priv, | |
978 | cpu); | |
979 | if (buf->backend.allocated) | |
64af2437 | 980 | lib_ring_buffer_set_quiescent(buf); |
f3bc08c5 MD |
981 | /* |
982 | * Perform flush before writing to finalized. | |
983 | */ | |
984 | smp_wmb(); | |
985 | ACCESS_ONCE(buf->finalized) = 1; | |
986 | wake_up_interruptible(&buf->read_wait); | |
987 | } | |
988 | } else { | |
989 | struct lib_ring_buffer *buf = chan->backend.buf; | |
990 | ||
991 | if (config->cb.buffer_finalize) | |
992 | config->cb.buffer_finalize(buf, chan->backend.priv, -1); | |
993 | if (buf->backend.allocated) | |
64af2437 | 994 | lib_ring_buffer_set_quiescent(buf); |
f3bc08c5 MD |
995 | /* |
996 | * Perform flush before writing to finalized. | |
997 | */ | |
998 | smp_wmb(); | |
999 | ACCESS_ONCE(buf->finalized) = 1; | |
1000 | wake_up_interruptible(&buf->read_wait); | |
1001 | } | |
24cedcfe MD |
1002 | ACCESS_ONCE(chan->finalized) = 1; |
1003 | wake_up_interruptible(&chan->hp_wait); | |
f3bc08c5 | 1004 | wake_up_interruptible(&chan->read_wait); |
f3bc08c5 | 1005 | priv = chan->backend.priv; |
ba1d61bc | 1006 | kref_put(&chan->ref, channel_release); |
f3bc08c5 MD |
1007 | return priv; |
1008 | } | |
1009 | EXPORT_SYMBOL_GPL(channel_destroy); | |
1010 | ||
1011 | struct lib_ring_buffer *channel_get_ring_buffer( | |
1012 | const struct lib_ring_buffer_config *config, | |
1013 | struct channel *chan, int cpu) | |
1014 | { | |
1015 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) | |
1016 | return chan->backend.buf; | |
1017 | else | |
1018 | return per_cpu_ptr(chan->backend.buf, cpu); | |
1019 | } | |
1020 | EXPORT_SYMBOL_GPL(channel_get_ring_buffer); | |
1021 | ||
1022 | int lib_ring_buffer_open_read(struct lib_ring_buffer *buf) | |
1023 | { | |
1024 | struct channel *chan = buf->backend.chan; | |
1025 | ||
1026 | if (!atomic_long_add_unless(&buf->active_readers, 1, 1)) | |
1027 | return -EBUSY; | |
9c1f4643 MD |
1028 | if (!lttng_kref_get(&chan->ref)) { |
1029 | atomic_long_dec(&buf->active_readers); | |
1030 | return -EOVERFLOW; | |
1031 | } | |
505fb410 | 1032 | lttng_smp_mb__after_atomic(); |
f3bc08c5 MD |
1033 | return 0; |
1034 | } | |
1035 | EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read); | |
1036 | ||
1037 | void lib_ring_buffer_release_read(struct lib_ring_buffer *buf) | |
1038 | { | |
1039 | struct channel *chan = buf->backend.chan; | |
1040 | ||
1041 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
505fb410 | 1042 | lttng_smp_mb__before_atomic(); |
f3bc08c5 | 1043 | atomic_long_dec(&buf->active_readers); |
f40270ad | 1044 | kref_put(&chan->ref, channel_release); |
f3bc08c5 MD |
1045 | } |
1046 | EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read); | |
1047 | ||
1048 | /* | |
1049 | * Promote compiler barrier to a smp_mb(). | |
1050 | * For the specific ring buffer case, this IPI call should be removed if the | |
1051 | * architecture does not reorder writes. This should eventually be provided by | |
1052 | * a separate architecture-specific infrastructure. | |
1053 | */ | |
1054 | static void remote_mb(void *info) | |
1055 | { | |
1056 | smp_mb(); | |
1057 | } | |
1058 | ||
1059 | /** | |
1060 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
1061 | * @buf: ring buffer | |
1062 | * @consumed: consumed count indicating the position where to read | |
1063 | * @produced: produced count, indicates position when to stop reading | |
1064 | * | |
1065 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1066 | * data to read at consumed position, or 0 if the get operation succeeds. | |
1067 | * Busy-loop trying to get data if the tick_nohz sequence lock is held. | |
1068 | */ | |
1069 | ||
1070 | int lib_ring_buffer_snapshot(struct lib_ring_buffer *buf, | |
1071 | unsigned long *consumed, unsigned long *produced) | |
1072 | { | |
1073 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1074 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1075 | unsigned long consumed_cur, write_offset; |
1076 | int finalized; | |
1077 | ||
1078 | retry: | |
1079 | finalized = ACCESS_ONCE(buf->finalized); | |
1080 | /* | |
1081 | * Read finalized before counters. | |
1082 | */ | |
1083 | smp_rmb(); | |
1084 | consumed_cur = atomic_long_read(&buf->consumed); | |
1085 | /* | |
1086 | * No need to issue a memory barrier between consumed count read and | |
1087 | * write offset read, because consumed count can only change | |
1088 | * concurrently in overwrite mode, and we keep a sequence counter | |
1089 | * identifier derived from the write offset to check we are getting | |
1090 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
1091 | * "tagged" upon writes, tags are checked upon read). | |
1092 | */ | |
1093 | write_offset = v_read(config, &buf->offset); | |
1094 | ||
1095 | /* | |
1096 | * Check that we are not about to read the same subbuffer in | |
1097 | * which the writer head is. | |
1098 | */ | |
1099 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
1100 | == 0) | |
1101 | goto nodata; | |
1102 | ||
1103 | *consumed = consumed_cur; | |
1104 | *produced = subbuf_trunc(write_offset, chan); | |
1105 | ||
1106 | return 0; | |
1107 | ||
1108 | nodata: | |
1109 | /* | |
1110 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1111 | * of "raw_spin_is_locked" memory ordering. | |
1112 | */ | |
1113 | if (finalized) | |
1114 | return -ENODATA; | |
1115 | else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock)) | |
1116 | goto retry; | |
1117 | else | |
1118 | return -EAGAIN; | |
1119 | } | |
1120 | EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot); | |
1121 | ||
1122 | /** | |
1123 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
71c1d843 MD |
1124 | * |
1125 | * Should only be called from consumer context. | |
f3bc08c5 MD |
1126 | * @buf: ring buffer |
1127 | * @consumed_new: new consumed count value | |
1128 | */ | |
1129 | void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf, | |
1130 | unsigned long consumed_new) | |
1131 | { | |
1132 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1133 | struct channel *chan = bufb->chan; | |
1134 | unsigned long consumed; | |
1135 | ||
1136 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
1137 | ||
1138 | /* | |
1139 | * Only push the consumed value forward. | |
1140 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
1141 | * the writer in flight recorder mode. | |
1142 | */ | |
1143 | consumed = atomic_long_read(&buf->consumed); | |
1144 | while ((long) consumed - (long) consumed_new < 0) | |
1145 | consumed = atomic_long_cmpxchg(&buf->consumed, consumed, | |
1146 | consumed_new); | |
71c1d843 MD |
1147 | /* Wake-up the metadata producer */ |
1148 | wake_up_interruptible(&buf->write_wait); | |
f3bc08c5 MD |
1149 | } |
1150 | EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer); | |
1151 | ||
1152 | /** | |
1153 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
1154 | * @buf: ring buffer | |
1155 | * @consumed: consumed count indicating the position where to read | |
1156 | * | |
1157 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1158 | * data to read at consumed position, or 0 if the get operation succeeds. | |
1159 | * Busy-loop trying to get data if the tick_nohz sequence lock is held. | |
1160 | */ | |
1161 | int lib_ring_buffer_get_subbuf(struct lib_ring_buffer *buf, | |
1162 | unsigned long consumed) | |
1163 | { | |
1164 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1165 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1166 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
1167 | int ret; | |
1168 | int finalized; | |
1169 | ||
8202b8a0 MD |
1170 | if (buf->get_subbuf) { |
1171 | /* | |
1172 | * Reader is trying to get a subbuffer twice. | |
1173 | */ | |
1174 | CHAN_WARN_ON(chan, 1); | |
1175 | return -EBUSY; | |
1176 | } | |
f3bc08c5 MD |
1177 | retry: |
1178 | finalized = ACCESS_ONCE(buf->finalized); | |
1179 | /* | |
1180 | * Read finalized before counters. | |
1181 | */ | |
1182 | smp_rmb(); | |
1183 | consumed_cur = atomic_long_read(&buf->consumed); | |
1184 | consumed_idx = subbuf_index(consumed, chan); | |
1185 | commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); | |
1186 | /* | |
1187 | * Make sure we read the commit count before reading the buffer | |
1188 | * data and the write offset. Correct consumed offset ordering | |
1189 | * wrt commit count is insured by the use of cmpxchg to update | |
1190 | * the consumed offset. | |
1191 | * smp_call_function_single can fail if the remote CPU is offline, | |
1192 | * this is OK because then there is no wmb to execute there. | |
1193 | * If our thread is executing on the same CPU as the on the buffers | |
1194 | * belongs to, we don't have to synchronize it at all. If we are | |
1195 | * migrated, the scheduler will take care of the memory barriers. | |
1196 | * Normally, smp_call_function_single() should ensure program order when | |
1197 | * executing the remote function, which implies that it surrounds the | |
1198 | * function execution with : | |
1199 | * smp_mb() | |
1200 | * send IPI | |
1201 | * csd_lock_wait | |
1202 | * recv IPI | |
1203 | * smp_mb() | |
1204 | * exec. function | |
1205 | * smp_mb() | |
1206 | * csd unlock | |
1207 | * smp_mb() | |
1208 | * | |
1209 | * However, smp_call_function_single() does not seem to clearly execute | |
1210 | * such barriers. It depends on spinlock semantic to provide the barrier | |
1211 | * before executing the IPI and, when busy-looping, csd_lock_wait only | |
1212 | * executes smp_mb() when it has to wait for the other CPU. | |
1213 | * | |
1214 | * I don't trust this code. Therefore, let's add the smp_mb() sequence | |
1215 | * required ourself, even if duplicated. It has no performance impact | |
1216 | * anyway. | |
1217 | * | |
1218 | * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs | |
1219 | * read and write vs write. They do not ensure core synchronization. We | |
1220 | * really have to ensure total order between the 3 barriers running on | |
1221 | * the 2 CPUs. | |
1222 | */ | |
1223 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1224 | if (config->sync == RING_BUFFER_SYNC_PER_CPU | |
1225 | && config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
1226 | if (raw_smp_processor_id() != buf->backend.cpu) { | |
1227 | /* Total order with IPI handler smp_mb() */ | |
1228 | smp_mb(); | |
1229 | smp_call_function_single(buf->backend.cpu, | |
1230 | remote_mb, NULL, 1); | |
1231 | /* Total order with IPI handler smp_mb() */ | |
1232 | smp_mb(); | |
1233 | } | |
1234 | } else { | |
1235 | /* Total order with IPI handler smp_mb() */ | |
1236 | smp_mb(); | |
1237 | smp_call_function(remote_mb, NULL, 1); | |
1238 | /* Total order with IPI handler smp_mb() */ | |
1239 | smp_mb(); | |
1240 | } | |
1241 | } else { | |
1242 | /* | |
1243 | * Local rmb to match the remote wmb to read the commit count | |
1244 | * before the buffer data and the write offset. | |
1245 | */ | |
1246 | smp_rmb(); | |
1247 | } | |
1248 | ||
1249 | write_offset = v_read(config, &buf->offset); | |
1250 | ||
1251 | /* | |
1252 | * Check that the buffer we are getting is after or at consumed_cur | |
1253 | * position. | |
1254 | */ | |
1255 | if ((long) subbuf_trunc(consumed, chan) | |
1256 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
1257 | goto nodata; | |
1258 | ||
1259 | /* | |
1260 | * Check that the subbuffer we are trying to consume has been | |
1261 | * already fully committed. | |
1262 | */ | |
1263 | if (((commit_count - chan->backend.subbuf_size) | |
1264 | & chan->commit_count_mask) | |
c9b3b5e2 | 1265 | - (buf_trunc(consumed, chan) |
f3bc08c5 MD |
1266 | >> chan->backend.num_subbuf_order) |
1267 | != 0) | |
1268 | goto nodata; | |
1269 | ||
1270 | /* | |
1271 | * Check that we are not about to read the same subbuffer in | |
1272 | * which the writer head is. | |
1273 | */ | |
c9b3b5e2 | 1274 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed, chan) |
f3bc08c5 MD |
1275 | == 0) |
1276 | goto nodata; | |
1277 | ||
1278 | /* | |
1279 | * Failure to get the subbuffer causes a busy-loop retry without going | |
1280 | * to a wait queue. These are caused by short-lived race windows where | |
1281 | * the writer is getting access to a subbuffer we were trying to get | |
1282 | * access to. Also checks that the "consumed" buffer count we are | |
1283 | * looking for matches the one contained in the subbuffer id. | |
1284 | */ | |
1285 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1286 | consumed_idx, buf_trunc_val(consumed, chan)); | |
1287 | if (ret) | |
1288 | goto retry; | |
1289 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
1290 | ||
1291 | buf->get_subbuf_consumed = consumed; | |
1292 | buf->get_subbuf = 1; | |
1293 | ||
1294 | return 0; | |
1295 | ||
1296 | nodata: | |
1297 | /* | |
1298 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1299 | * of "raw_spin_is_locked" memory ordering. | |
1300 | */ | |
1301 | if (finalized) | |
1302 | return -ENODATA; | |
1303 | else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock)) | |
1304 | goto retry; | |
1305 | else | |
1306 | return -EAGAIN; | |
1307 | } | |
1308 | EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf); | |
1309 | ||
1310 | /** | |
1311 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
1312 | * @buf: ring buffer | |
1313 | */ | |
1314 | void lib_ring_buffer_put_subbuf(struct lib_ring_buffer *buf) | |
1315 | { | |
1316 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1317 | struct channel *chan = bufb->chan; | |
5a8fd222 | 1318 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1319 | unsigned long read_sb_bindex, consumed_idx, consumed; |
1320 | ||
1321 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
1322 | ||
1323 | if (!buf->get_subbuf) { | |
1324 | /* | |
1325 | * Reader puts a subbuffer it did not get. | |
1326 | */ | |
1327 | CHAN_WARN_ON(chan, 1); | |
1328 | return; | |
1329 | } | |
1330 | consumed = buf->get_subbuf_consumed; | |
1331 | buf->get_subbuf = 0; | |
1332 | ||
1333 | /* | |
1334 | * Clear the records_unread counter. (overruns counter) | |
1335 | * Can still be non-zero if a file reader simply grabbed the data | |
1336 | * without using iterators. | |
1337 | * Can be below zero if an iterator is used on a snapshot more than | |
1338 | * once. | |
1339 | */ | |
1340 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
1341 | v_add(config, v_read(config, | |
1342 | &bufb->array[read_sb_bindex]->records_unread), | |
1343 | &bufb->records_read); | |
1344 | v_set(config, &bufb->array[read_sb_bindex]->records_unread, 0); | |
1345 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE | |
1346 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
1347 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
1348 | ||
1349 | /* | |
1350 | * Exchange the reader subbuffer with the one we put in its place in the | |
1351 | * writer subbuffer table. Expect the original consumed count. If | |
1352 | * update_read_sb_index fails, this is because the writer updated the | |
1353 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
1354 | * currently have: it has become invalid to try reading this sub-buffer | |
1355 | * consumed count value anyway. | |
1356 | */ | |
1357 | consumed_idx = subbuf_index(consumed, chan); | |
1358 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1359 | consumed_idx, buf_trunc_val(consumed, chan)); | |
1360 | /* | |
1361 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
1362 | * if the writer concurrently updated it. | |
1363 | */ | |
1364 | } | |
1365 | EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf); | |
1366 | ||
1367 | /* | |
1368 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
1369 | * position and the writer position. (inclusive) | |
1370 | */ | |
1371 | static | |
1372 | void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer *buf, | |
1373 | struct channel *chan, | |
1374 | unsigned long cons_offset, | |
1375 | int cpu) | |
1376 | { | |
5a8fd222 | 1377 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1378 | unsigned long cons_idx, commit_count, commit_count_sb; |
1379 | ||
1380 | cons_idx = subbuf_index(cons_offset, chan); | |
1381 | commit_count = v_read(config, &buf->commit_hot[cons_idx].cc); | |
1382 | commit_count_sb = v_read(config, &buf->commit_cold[cons_idx].cc_sb); | |
1383 | ||
1384 | if (subbuf_offset(commit_count, chan) != 0) | |
1385 | printk(KERN_WARNING | |
1386 | "ring buffer %s, cpu %d: " | |
1387 | "commit count in subbuffer %lu,\n" | |
1388 | "expecting multiples of %lu bytes\n" | |
1389 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
1390 | chan->backend.name, cpu, cons_idx, | |
1391 | chan->backend.subbuf_size, | |
1392 | commit_count, commit_count_sb); | |
1393 | ||
1394 | printk(KERN_DEBUG "ring buffer: %s, cpu %d: %lu bytes committed\n", | |
1395 | chan->backend.name, cpu, commit_count); | |
1396 | } | |
1397 | ||
1398 | static | |
1399 | void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf, | |
1400 | struct channel *chan, | |
1401 | void *priv, int cpu) | |
1402 | { | |
5a8fd222 | 1403 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1404 | unsigned long write_offset, cons_offset; |
1405 | ||
f3bc08c5 MD |
1406 | /* |
1407 | * No need to order commit_count, write_offset and cons_offset reads | |
1408 | * because we execute at teardown when no more writer nor reader | |
1409 | * references are left. | |
1410 | */ | |
1411 | write_offset = v_read(config, &buf->offset); | |
1412 | cons_offset = atomic_long_read(&buf->consumed); | |
1413 | if (write_offset != cons_offset) | |
05aad775 | 1414 | printk(KERN_DEBUG |
f3bc08c5 MD |
1415 | "ring buffer %s, cpu %d: " |
1416 | "non-consumed data\n" | |
1417 | " [ %lu bytes written, %lu bytes read ]\n", | |
1418 | chan->backend.name, cpu, write_offset, cons_offset); | |
1419 | ||
1420 | for (cons_offset = atomic_long_read(&buf->consumed); | |
1421 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), | |
1422 | chan) | |
1423 | - cons_offset) > 0; | |
1424 | cons_offset = subbuf_align(cons_offset, chan)) | |
1425 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1426 | cpu); | |
1427 | } | |
1428 | ||
1429 | static | |
1430 | void lib_ring_buffer_print_errors(struct channel *chan, | |
1431 | struct lib_ring_buffer *buf, int cpu) | |
1432 | { | |
5a8fd222 | 1433 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1434 | void *priv = chan->backend.priv; |
1435 | ||
ec01ec93 MD |
1436 | if (!strcmp(chan->backend.name, "relay-metadata")) { |
1437 | printk(KERN_DEBUG "ring buffer %s: %lu records written, " | |
1438 | "%lu records overrun\n", | |
1439 | chan->backend.name, | |
1440 | v_read(config, &buf->records_count), | |
1441 | v_read(config, &buf->records_overrun)); | |
1442 | } else { | |
1443 | printk(KERN_DEBUG "ring buffer %s, cpu %d: %lu records written, " | |
1444 | "%lu records overrun\n", | |
1445 | chan->backend.name, cpu, | |
1446 | v_read(config, &buf->records_count), | |
1447 | v_read(config, &buf->records_overrun)); | |
1448 | ||
1449 | if (v_read(config, &buf->records_lost_full) | |
1450 | || v_read(config, &buf->records_lost_wrap) | |
1451 | || v_read(config, &buf->records_lost_big)) | |
1452 | printk(KERN_WARNING | |
1453 | "ring buffer %s, cpu %d: records were lost. Caused by:\n" | |
1454 | " [ %lu buffer full, %lu nest buffer wrap-around, " | |
1455 | "%lu event too big ]\n", | |
1456 | chan->backend.name, cpu, | |
1457 | v_read(config, &buf->records_lost_full), | |
1458 | v_read(config, &buf->records_lost_wrap), | |
1459 | v_read(config, &buf->records_lost_big)); | |
1460 | } | |
f3bc08c5 MD |
1461 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu); |
1462 | } | |
1463 | ||
1464 | /* | |
1465 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1466 | * | |
dc5cd570 MD |
1467 | * Only executed by SWITCH_FLUSH, which can be issued while tracing is active |
1468 | * or at buffer finalization (destroy). | |
f3bc08c5 MD |
1469 | */ |
1470 | static | |
1471 | void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf, | |
1472 | struct channel *chan, | |
1473 | struct switch_offsets *offsets, | |
1474 | u64 tsc) | |
1475 | { | |
5a8fd222 | 1476 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1477 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1478 | unsigned long commit_count; | |
fbe84fd7 | 1479 | struct commit_counters_hot *cc_hot; |
f3bc08c5 MD |
1480 | |
1481 | config->cb.buffer_begin(buf, tsc, oldidx); | |
1482 | ||
1483 | /* | |
1484 | * Order all writes to buffer before the commit count update that will | |
1485 | * determine that the subbuffer is full. | |
1486 | */ | |
1487 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1488 | /* | |
1489 | * Must write slot data before incrementing commit count. This | |
1490 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1491 | * by get_subbuf(). | |
1492 | */ | |
1493 | barrier(); | |
1494 | } else | |
1495 | smp_wmb(); | |
fbe84fd7 MD |
1496 | cc_hot = &buf->commit_hot[oldidx]; |
1497 | v_add(config, config->cb.subbuffer_header_size(), &cc_hot->cc); | |
1498 | commit_count = v_read(config, &cc_hot->cc); | |
f3bc08c5 MD |
1499 | /* Check if the written buffer has to be delivered */ |
1500 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
635e457c | 1501 | commit_count, oldidx, tsc); |
8ec496cf | 1502 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
7915e163 | 1503 | offsets->old + config->cb.subbuffer_header_size(), |
fbe84fd7 | 1504 | commit_count, cc_hot); |
f3bc08c5 MD |
1505 | } |
1506 | ||
1507 | /* | |
1508 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1509 | * | |
1510 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1511 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1512 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1513 | * subbuffer. | |
1514 | */ | |
1515 | static | |
1516 | void lib_ring_buffer_switch_old_end(struct lib_ring_buffer *buf, | |
1517 | struct channel *chan, | |
1518 | struct switch_offsets *offsets, | |
1519 | u64 tsc) | |
1520 | { | |
5a8fd222 | 1521 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1522 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1523 | unsigned long commit_count, padding_size, data_size; | |
fbe84fd7 | 1524 | struct commit_counters_hot *cc_hot; |
f3bc08c5 MD |
1525 | |
1526 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1527 | padding_size = chan->backend.subbuf_size - data_size; | |
1528 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size); | |
1529 | ||
1530 | /* | |
1531 | * Order all writes to buffer before the commit count update that will | |
1532 | * determine that the subbuffer is full. | |
1533 | */ | |
1534 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1535 | /* | |
1536 | * Must write slot data before incrementing commit count. This | |
1537 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1538 | * by get_subbuf(). | |
1539 | */ | |
1540 | barrier(); | |
1541 | } else | |
1542 | smp_wmb(); | |
fbe84fd7 MD |
1543 | cc_hot = &buf->commit_hot[oldidx]; |
1544 | v_add(config, padding_size, &cc_hot->cc); | |
1545 | commit_count = v_read(config, &cc_hot->cc); | |
f3bc08c5 | 1546 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
635e457c | 1547 | commit_count, oldidx, tsc); |
8ec496cf MD |
1548 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
1549 | offsets->old + padding_size, commit_count, | |
fbe84fd7 | 1550 | cc_hot); |
f3bc08c5 MD |
1551 | } |
1552 | ||
1553 | /* | |
1554 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1555 | * | |
1556 | * This code can be executed unordered : writers may already have written to the | |
1557 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1558 | * that this code is executed before the deliver of this sub-buffer. | |
1559 | */ | |
1560 | static | |
1561 | void lib_ring_buffer_switch_new_start(struct lib_ring_buffer *buf, | |
1562 | struct channel *chan, | |
1563 | struct switch_offsets *offsets, | |
1564 | u64 tsc) | |
1565 | { | |
5a8fd222 | 1566 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1567 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1568 | unsigned long commit_count; | |
fbe84fd7 | 1569 | struct commit_counters_hot *cc_hot; |
f3bc08c5 MD |
1570 | |
1571 | config->cb.buffer_begin(buf, tsc, beginidx); | |
1572 | ||
1573 | /* | |
1574 | * Order all writes to buffer before the commit count update that will | |
1575 | * determine that the subbuffer is full. | |
1576 | */ | |
1577 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1578 | /* | |
1579 | * Must write slot data before incrementing commit count. This | |
1580 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1581 | * by get_subbuf(). | |
1582 | */ | |
1583 | barrier(); | |
1584 | } else | |
1585 | smp_wmb(); | |
fbe84fd7 MD |
1586 | cc_hot = &buf->commit_hot[beginidx]; |
1587 | v_add(config, config->cb.subbuffer_header_size(), &cc_hot->cc); | |
1588 | commit_count = v_read(config, &cc_hot->cc); | |
f3bc08c5 MD |
1589 | /* Check if the written buffer has to be delivered */ |
1590 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
635e457c | 1591 | commit_count, beginidx, tsc); |
8ec496cf | 1592 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
7915e163 | 1593 | offsets->begin + config->cb.subbuffer_header_size(), |
fbe84fd7 | 1594 | commit_count, cc_hot); |
f3bc08c5 MD |
1595 | } |
1596 | ||
f5ea5800 MD |
1597 | /* |
1598 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1599 | * | |
768b05c9 MD |
1600 | * Calls subbuffer_set_data_size() to set the data size of the current |
1601 | * sub-buffer. We do not need to perform check_deliver nor commit here, | |
1602 | * since this task will be done by the "commit" of the event for which | |
1603 | * we are currently doing the space reservation. | |
f5ea5800 MD |
1604 | */ |
1605 | static | |
1606 | void lib_ring_buffer_switch_new_end(struct lib_ring_buffer *buf, | |
1607 | struct channel *chan, | |
1608 | struct switch_offsets *offsets, | |
1609 | u64 tsc) | |
1610 | { | |
1611 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
768b05c9 | 1612 | unsigned long endidx, data_size; |
f5ea5800 | 1613 | |
768b05c9 | 1614 | endidx = subbuf_index(offsets->end - 1, chan); |
f5ea5800 | 1615 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; |
f5ea5800 | 1616 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size); |
f5ea5800 MD |
1617 | } |
1618 | ||
f3bc08c5 MD |
1619 | /* |
1620 | * Returns : | |
1621 | * 0 if ok | |
1622 | * !0 if execution must be aborted. | |
1623 | */ | |
1624 | static | |
1625 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
1626 | struct lib_ring_buffer *buf, | |
1627 | struct channel *chan, | |
1628 | struct switch_offsets *offsets, | |
1629 | u64 *tsc) | |
1630 | { | |
5a8fd222 | 1631 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
5334a2c5 | 1632 | unsigned long off, reserve_commit_diff; |
f3bc08c5 MD |
1633 | |
1634 | offsets->begin = v_read(config, &buf->offset); | |
1635 | offsets->old = offsets->begin; | |
1636 | offsets->switch_old_start = 0; | |
1637 | off = subbuf_offset(offsets->begin, chan); | |
1638 | ||
1639 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1640 | ||
1641 | /* | |
1642 | * Ensure we flush the header of an empty subbuffer when doing the | |
1643 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1644 | * total data gathering duration even if there were no records saved | |
1645 | * after the last buffer switch. | |
1646 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1647 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1648 | * subbuffer header as appropriate. | |
1649 | * The next record that reserves space will be responsible for | |
1650 | * populating the following subbuffer header. We choose not to populate | |
1651 | * the next subbuffer header here because we want to be able to use | |
1652 | * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop | |
1653 | * buffer flush, which must guarantee that all the buffer content | |
1654 | * (records and header timestamps) are visible to the reader. This is | |
1655 | * required for quiescence guarantees for the fusion merge. | |
1656 | */ | |
5334a2c5 MD |
1657 | if (mode != SWITCH_FLUSH && !off) |
1658 | return -1; /* we do not have to switch : buffer is empty */ | |
1659 | ||
1660 | if (unlikely(off == 0)) { | |
1661 | unsigned long sb_index, commit_count; | |
1662 | ||
1663 | /* | |
dc5cd570 MD |
1664 | * We are performing a SWITCH_FLUSH. There may be concurrent |
1665 | * writes into the buffer if e.g. invoked while performing a | |
1666 | * snapshot on an active trace. | |
5334a2c5 | 1667 | * |
dc5cd570 MD |
1668 | * If the client does not save any header information (sub-buffer |
1669 | * header size == 0), don't switch empty subbuffer on finalize, | |
1670 | * because it is invalid to deliver a completely empty | |
1671 | * subbuffer. | |
5334a2c5 MD |
1672 | */ |
1673 | if (!config->cb.subbuffer_header_size()) | |
1674 | return -1; | |
1675 | ||
1676 | /* Test new buffer integrity */ | |
1677 | sb_index = subbuf_index(offsets->begin, chan); | |
1678 | commit_count = v_read(config, | |
1679 | &buf->commit_cold[sb_index].cc_sb); | |
1680 | reserve_commit_diff = | |
1681 | (buf_trunc(offsets->begin, chan) | |
1682 | >> chan->backend.num_subbuf_order) | |
1683 | - (commit_count & chan->commit_count_mask); | |
1684 | if (likely(reserve_commit_diff == 0)) { | |
1685 | /* Next subbuffer not being written to. */ | |
1686 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1687 | subbuf_trunc(offsets->begin, chan) | |
1688 | - subbuf_trunc((unsigned long) | |
1689 | atomic_long_read(&buf->consumed), chan) | |
1690 | >= chan->backend.buf_size)) { | |
1691 | /* | |
1692 | * We do not overwrite non consumed buffers | |
1693 | * and we are full : don't switch. | |
1694 | */ | |
f3bc08c5 | 1695 | return -1; |
5334a2c5 MD |
1696 | } else { |
1697 | /* | |
1698 | * Next subbuffer not being written to, and we | |
1699 | * are either in overwrite mode or the buffer is | |
1700 | * not full. It's safe to write in this new | |
1701 | * subbuffer. | |
1702 | */ | |
1703 | } | |
1704 | } else { | |
f3bc08c5 | 1705 | /* |
5334a2c5 MD |
1706 | * Next subbuffer reserve offset does not match the |
1707 | * commit offset. Don't perform switch in | |
1708 | * producer-consumer and overwrite mode. Caused by | |
1709 | * either a writer OOPS or too many nested writes over a | |
1710 | * reserve/commit pair. | |
f3bc08c5 | 1711 | */ |
5334a2c5 | 1712 | return -1; |
f3bc08c5 | 1713 | } |
5334a2c5 MD |
1714 | |
1715 | /* | |
1716 | * Need to write the subbuffer start header on finalize. | |
1717 | */ | |
1718 | offsets->switch_old_start = 1; | |
1719 | } | |
1720 | offsets->begin = subbuf_align(offsets->begin, chan); | |
f3bc08c5 MD |
1721 | /* Note: old points to the next subbuf at offset 0 */ |
1722 | offsets->end = offsets->begin; | |
1723 | return 0; | |
1724 | } | |
1725 | ||
1726 | /* | |
1727 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1728 | * called while tracing is active with absolutely no lock held. | |
1729 | * | |
1730 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1731 | * operations, this function must be called from the CPU which owns the buffer | |
1732 | * for a ACTIVE flush. | |
1733 | */ | |
1734 | void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode) | |
1735 | { | |
1736 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1737 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1738 | struct switch_offsets offsets; |
1739 | unsigned long oldidx; | |
1740 | u64 tsc; | |
1741 | ||
1742 | offsets.size = 0; | |
1743 | ||
1744 | /* | |
1745 | * Perform retryable operations. | |
1746 | */ | |
1747 | do { | |
1748 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1749 | &tsc)) | |
1750 | return; /* Switch not needed */ | |
1751 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1752 | != offsets.old); | |
1753 | ||
1754 | /* | |
1755 | * Atomically update last_tsc. This update races against concurrent | |
1756 | * atomic updates, but the race will always cause supplementary full TSC | |
1757 | * records, never the opposite (missing a full TSC record when it would | |
1758 | * be needed). | |
1759 | */ | |
1760 | save_last_tsc(config, buf, tsc); | |
1761 | ||
1762 | /* | |
1763 | * Push the reader if necessary | |
1764 | */ | |
1765 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1766 | ||
1767 | oldidx = subbuf_index(offsets.old, chan); | |
1768 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx); | |
1769 | ||
1770 | /* | |
1771 | * May need to populate header start on SWITCH_FLUSH. | |
1772 | */ | |
1773 | if (offsets.switch_old_start) { | |
1774 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc); | |
1775 | offsets.old += config->cb.subbuffer_header_size(); | |
1776 | } | |
1777 | ||
1778 | /* | |
1779 | * Switch old subbuffer. | |
1780 | */ | |
1781 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc); | |
1782 | } | |
1783 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow); | |
1784 | ||
9dded431 MD |
1785 | struct switch_param { |
1786 | struct lib_ring_buffer *buf; | |
1787 | enum switch_mode mode; | |
1788 | }; | |
1789 | ||
5e391252 MD |
1790 | static void remote_switch(void *info) |
1791 | { | |
9dded431 MD |
1792 | struct switch_param *param = info; |
1793 | struct lib_ring_buffer *buf = param->buf; | |
5e391252 | 1794 | |
9dded431 | 1795 | lib_ring_buffer_switch_slow(buf, param->mode); |
5e391252 MD |
1796 | } |
1797 | ||
64af2437 MD |
1798 | static void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf, |
1799 | enum switch_mode mode) | |
5e391252 MD |
1800 | { |
1801 | struct channel *chan = buf->backend.chan; | |
1802 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1803 | int ret; | |
9dded431 | 1804 | struct switch_param param; |
5e391252 MD |
1805 | |
1806 | /* | |
1807 | * With global synchronization we don't need to use the IPI scheme. | |
1808 | */ | |
1809 | if (config->sync == RING_BUFFER_SYNC_GLOBAL) { | |
64af2437 | 1810 | lib_ring_buffer_switch_slow(buf, mode); |
5e391252 MD |
1811 | return; |
1812 | } | |
1813 | ||
1814 | /* | |
1815 | * Taking lock on CPU hotplug to ensure two things: first, that the | |
1816 | * target cpu is not taken concurrently offline while we are within | |
1817 | * smp_call_function_single() (I don't trust that get_cpu() on the | |
1818 | * _local_ CPU actually inhibit CPU hotplug for the _remote_ CPU (to be | |
1819 | * confirmed)). Secondly, if it happens that the CPU is not online, our | |
1820 | * own call to lib_ring_buffer_switch_slow() needs to be protected from | |
1821 | * CPU hotplug handlers, which can also perform a remote subbuffer | |
1822 | * switch. | |
1823 | */ | |
1824 | get_online_cpus(); | |
9dded431 MD |
1825 | param.buf = buf; |
1826 | param.mode = mode; | |
5e391252 | 1827 | ret = smp_call_function_single(buf->backend.cpu, |
9dded431 | 1828 | remote_switch, ¶m, 1); |
5e391252 MD |
1829 | if (ret) { |
1830 | /* Remote CPU is offline, do it ourself. */ | |
64af2437 | 1831 | lib_ring_buffer_switch_slow(buf, mode); |
5e391252 MD |
1832 | } |
1833 | put_online_cpus(); | |
1834 | } | |
64af2437 MD |
1835 | |
1836 | void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf) | |
1837 | { | |
1838 | _lib_ring_buffer_switch_remote(buf, SWITCH_ACTIVE); | |
1839 | } | |
5e391252 MD |
1840 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote); |
1841 | ||
7cf44d03 MD |
1842 | /* Switch sub-buffer even if current sub-buffer is empty. */ |
1843 | void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer *buf) | |
1844 | { | |
1845 | _lib_ring_buffer_switch_remote(buf, SWITCH_FLUSH); | |
1846 | } | |
1847 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote_empty); | |
1848 | ||
f3bc08c5 MD |
1849 | /* |
1850 | * Returns : | |
1851 | * 0 if ok | |
97ca2c54 MD |
1852 | * -ENOSPC if event size is too large for packet. |
1853 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1854 | * -EIO if data cannot be written into the buffer for any other reason. | |
f3bc08c5 MD |
1855 | */ |
1856 | static | |
1857 | int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, | |
1858 | struct channel *chan, | |
1859 | struct switch_offsets *offsets, | |
1860 | struct lib_ring_buffer_ctx *ctx) | |
1861 | { | |
5a8fd222 | 1862 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
0fdec686 | 1863 | unsigned long reserve_commit_diff, offset_cmp; |
f3bc08c5 | 1864 | |
0fdec686 MD |
1865 | retry: |
1866 | offsets->begin = offset_cmp = v_read(config, &buf->offset); | |
f3bc08c5 MD |
1867 | offsets->old = offsets->begin; |
1868 | offsets->switch_new_start = 0; | |
f5ea5800 | 1869 | offsets->switch_new_end = 0; |
f3bc08c5 MD |
1870 | offsets->switch_old_end = 0; |
1871 | offsets->pre_header_padding = 0; | |
1872 | ||
1873 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
97ca2c54 MD |
1874 | if ((int64_t) ctx->tsc == -EIO) |
1875 | return -EIO; | |
f3bc08c5 MD |
1876 | |
1877 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
64c796d8 | 1878 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; |
f3bc08c5 MD |
1879 | |
1880 | if (unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { | |
1881 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1882 | } else { | |
1883 | offsets->size = config->cb.record_header_size(config, chan, | |
1884 | offsets->begin, | |
f3bc08c5 | 1885 | &offsets->pre_header_padding, |
64c796d8 | 1886 | ctx); |
f3bc08c5 MD |
1887 | offsets->size += |
1888 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1889 | ctx->largest_align) | |
1890 | + ctx->data_size; | |
1891 | if (unlikely(subbuf_offset(offsets->begin, chan) + | |
1892 | offsets->size > chan->backend.subbuf_size)) { | |
1893 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1894 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1895 | } | |
1896 | } | |
1897 | if (unlikely(offsets->switch_new_start)) { | |
0fdec686 | 1898 | unsigned long sb_index, commit_count; |
f3bc08c5 MD |
1899 | |
1900 | /* | |
1901 | * We are typically not filling the previous buffer completely. | |
1902 | */ | |
1903 | if (likely(offsets->switch_old_end)) | |
1904 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1905 | offsets->begin = offsets->begin | |
1906 | + config->cb.subbuffer_header_size(); | |
1907 | /* Test new buffer integrity */ | |
1908 | sb_index = subbuf_index(offsets->begin, chan); | |
0fdec686 MD |
1909 | /* |
1910 | * Read buf->offset before buf->commit_cold[sb_index].cc_sb. | |
1911 | * lib_ring_buffer_check_deliver() has the matching | |
1912 | * memory barriers required around commit_cold cc_sb | |
1913 | * updates to ensure reserve and commit counter updates | |
1914 | * are not seen reordered when updated by another CPU. | |
1915 | */ | |
1916 | smp_rmb(); | |
1917 | commit_count = v_read(config, | |
1918 | &buf->commit_cold[sb_index].cc_sb); | |
1919 | /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */ | |
1920 | smp_rmb(); | |
1921 | if (unlikely(offset_cmp != v_read(config, &buf->offset))) { | |
1922 | /* | |
1923 | * The reserve counter have been concurrently updated | |
1924 | * while we read the commit counter. This means the | |
1925 | * commit counter we read might not match buf->offset | |
1926 | * due to concurrent update. We therefore need to retry. | |
1927 | */ | |
1928 | goto retry; | |
1929 | } | |
f3bc08c5 MD |
1930 | reserve_commit_diff = |
1931 | (buf_trunc(offsets->begin, chan) | |
1932 | >> chan->backend.num_subbuf_order) | |
0fdec686 | 1933 | - (commit_count & chan->commit_count_mask); |
f3bc08c5 MD |
1934 | if (likely(reserve_commit_diff == 0)) { |
1935 | /* Next subbuffer not being written to. */ | |
1936 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1937 | subbuf_trunc(offsets->begin, chan) | |
1938 | - subbuf_trunc((unsigned long) | |
1939 | atomic_long_read(&buf->consumed), chan) | |
1940 | >= chan->backend.buf_size)) { | |
1941 | /* | |
1942 | * We do not overwrite non consumed buffers | |
1943 | * and we are full : record is lost. | |
1944 | */ | |
1945 | v_inc(config, &buf->records_lost_full); | |
97ca2c54 | 1946 | return -ENOBUFS; |
f3bc08c5 MD |
1947 | } else { |
1948 | /* | |
1949 | * Next subbuffer not being written to, and we | |
1950 | * are either in overwrite mode or the buffer is | |
1951 | * not full. It's safe to write in this new | |
1952 | * subbuffer. | |
1953 | */ | |
1954 | } | |
1955 | } else { | |
1956 | /* | |
1957 | * Next subbuffer reserve offset does not match the | |
0fdec686 MD |
1958 | * commit offset, and this did not involve update to the |
1959 | * reserve counter. Drop record in producer-consumer and | |
1960 | * overwrite mode. Caused by either a writer OOPS or | |
1961 | * too many nested writes over a reserve/commit pair. | |
f3bc08c5 MD |
1962 | */ |
1963 | v_inc(config, &buf->records_lost_wrap); | |
97ca2c54 | 1964 | return -EIO; |
f3bc08c5 MD |
1965 | } |
1966 | offsets->size = | |
1967 | config->cb.record_header_size(config, chan, | |
1968 | offsets->begin, | |
f3bc08c5 | 1969 | &offsets->pre_header_padding, |
64c796d8 | 1970 | ctx); |
f3bc08c5 MD |
1971 | offsets->size += |
1972 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1973 | ctx->largest_align) | |
1974 | + ctx->data_size; | |
1975 | if (unlikely(subbuf_offset(offsets->begin, chan) | |
1976 | + offsets->size > chan->backend.subbuf_size)) { | |
1977 | /* | |
1978 | * Record too big for subbuffers, report error, don't | |
1979 | * complete the sub-buffer switch. | |
1980 | */ | |
1981 | v_inc(config, &buf->records_lost_big); | |
97ca2c54 | 1982 | return -ENOSPC; |
f3bc08c5 MD |
1983 | } else { |
1984 | /* | |
1985 | * We just made a successful buffer switch and the | |
1986 | * record fits in the new subbuffer. Let's write. | |
1987 | */ | |
1988 | } | |
1989 | } else { | |
1990 | /* | |
1991 | * Record fits in the current buffer and we are not on a switch | |
1992 | * boundary. It's safe to write. | |
1993 | */ | |
1994 | } | |
1995 | offsets->end = offsets->begin + offsets->size; | |
f5ea5800 MD |
1996 | |
1997 | if (unlikely(subbuf_offset(offsets->end, chan) == 0)) { | |
1998 | /* | |
1999 | * The offset_end will fall at the very beginning of the next | |
2000 | * subbuffer. | |
2001 | */ | |
2002 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
2003 | } | |
f3bc08c5 MD |
2004 | return 0; |
2005 | } | |
2006 | ||
d7e74017 MD |
2007 | static struct lib_ring_buffer *get_current_buf(struct channel *chan, int cpu) |
2008 | { | |
2009 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
2010 | ||
2011 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
2012 | return per_cpu_ptr(chan->backend.buf, cpu); | |
2013 | else | |
2014 | return chan->backend.buf; | |
2015 | } | |
2016 | ||
2017 | void lib_ring_buffer_lost_event_too_big(struct channel *chan) | |
2018 | { | |
2019 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
2020 | struct lib_ring_buffer *buf = get_current_buf(chan, smp_processor_id()); | |
2021 | ||
2022 | v_inc(config, &buf->records_lost_big); | |
2023 | } | |
2024 | EXPORT_SYMBOL_GPL(lib_ring_buffer_lost_event_too_big); | |
2025 | ||
f3bc08c5 MD |
2026 | /** |
2027 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
2028 | * @ctx: ring buffer context. | |
2029 | * | |
97ca2c54 MD |
2030 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, |
2031 | * -EIO for other errors, else returns 0. | |
f3bc08c5 MD |
2032 | * It will take care of sub-buffer switching. |
2033 | */ | |
2034 | int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx) | |
2035 | { | |
2036 | struct channel *chan = ctx->chan; | |
5a8fd222 | 2037 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
2038 | struct lib_ring_buffer *buf; |
2039 | struct switch_offsets offsets; | |
c099397a | 2040 | int ret; |
f3bc08c5 | 2041 | |
d7e74017 | 2042 | ctx->buf = buf = get_current_buf(chan, ctx->cpu); |
f3bc08c5 MD |
2043 | offsets.size = 0; |
2044 | ||
2045 | do { | |
97ca2c54 MD |
2046 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, |
2047 | ctx); | |
2048 | if (unlikely(ret)) | |
2049 | return ret; | |
f3bc08c5 MD |
2050 | } while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
2051 | offsets.end) | |
2052 | != offsets.old)); | |
2053 | ||
2054 | /* | |
2055 | * Atomically update last_tsc. This update races against concurrent | |
2056 | * atomic updates, but the race will always cause supplementary full TSC | |
2057 | * records, never the opposite (missing a full TSC record when it would | |
2058 | * be needed). | |
2059 | */ | |
2060 | save_last_tsc(config, buf, ctx->tsc); | |
2061 | ||
2062 | /* | |
2063 | * Push the reader if necessary | |
2064 | */ | |
2065 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
2066 | ||
2067 | /* | |
2068 | * Clear noref flag for this subbuffer. | |
2069 | */ | |
2070 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
2071 | subbuf_index(offsets.end - 1, chan)); | |
2072 | ||
2073 | /* | |
2074 | * Switch old subbuffer if needed. | |
2075 | */ | |
2076 | if (unlikely(offsets.switch_old_end)) { | |
2077 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
2078 | subbuf_index(offsets.old - 1, chan)); | |
2079 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc); | |
2080 | } | |
2081 | ||
2082 | /* | |
2083 | * Populate new subbuffer. | |
2084 | */ | |
2085 | if (unlikely(offsets.switch_new_start)) | |
2086 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc); | |
2087 | ||
f5ea5800 MD |
2088 | if (unlikely(offsets.switch_new_end)) |
2089 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc); | |
2090 | ||
f3bc08c5 MD |
2091 | ctx->slot_size = offsets.size; |
2092 | ctx->pre_offset = offsets.begin; | |
2093 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
2094 | return 0; | |
2095 | } | |
2096 | EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow); | |
6fb8de4b | 2097 | |
aece661f MD |
2098 | static |
2099 | void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config, | |
2100 | struct lib_ring_buffer *buf, | |
2101 | unsigned long commit_count, | |
2102 | unsigned long idx) | |
2103 | { | |
2104 | if (config->oops == RING_BUFFER_OOPS_CONSISTENCY) | |
2105 | v_set(config, &buf->commit_hot[idx].seq, commit_count); | |
2106 | } | |
2107 | ||
25337cb5 MD |
2108 | /* |
2109 | * The ring buffer can count events recorded and overwritten per buffer, | |
2110 | * but it is disabled by default due to its performance overhead. | |
2111 | */ | |
2112 | #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS | |
2113 | static | |
2114 | void deliver_count_events(const struct lib_ring_buffer_config *config, | |
2115 | struct lib_ring_buffer *buf, | |
2116 | unsigned long idx) | |
2117 | { | |
2118 | v_add(config, subbuffer_get_records_count(config, | |
2119 | &buf->backend, idx), | |
2120 | &buf->records_count); | |
2121 | v_add(config, subbuffer_count_records_overrun(config, | |
2122 | &buf->backend, idx), | |
2123 | &buf->records_overrun); | |
2124 | } | |
2125 | #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */ | |
2126 | static | |
2127 | void deliver_count_events(const struct lib_ring_buffer_config *config, | |
2128 | struct lib_ring_buffer *buf, | |
2129 | unsigned long idx) | |
2130 | { | |
2131 | } | |
2132 | #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */ | |
2133 | ||
2134 | ||
aece661f MD |
2135 | void lib_ring_buffer_check_deliver_slow(const struct lib_ring_buffer_config *config, |
2136 | struct lib_ring_buffer *buf, | |
2137 | struct channel *chan, | |
2138 | unsigned long offset, | |
2139 | unsigned long commit_count, | |
2140 | unsigned long idx, | |
2141 | u64 tsc) | |
2142 | { | |
2143 | unsigned long old_commit_count = commit_count | |
2144 | - chan->backend.subbuf_size; | |
2145 | ||
2146 | /* | |
2147 | * If we succeeded at updating cc_sb below, we are the subbuffer | |
2148 | * writer delivering the subbuffer. Deals with concurrent | |
2149 | * updates of the "cc" value without adding a add_return atomic | |
2150 | * operation to the fast path. | |
2151 | * | |
2152 | * We are doing the delivery in two steps: | |
2153 | * - First, we cmpxchg() cc_sb to the new value | |
2154 | * old_commit_count + 1. This ensures that we are the only | |
2155 | * subbuffer user successfully filling the subbuffer, but we | |
2156 | * do _not_ set the cc_sb value to "commit_count" yet. | |
2157 | * Therefore, other writers that would wrap around the ring | |
2158 | * buffer and try to start writing to our subbuffer would | |
2159 | * have to drop records, because it would appear as | |
2160 | * non-filled. | |
2161 | * We therefore have exclusive access to the subbuffer control | |
2162 | * structures. This mutual exclusion with other writers is | |
2163 | * crucially important to perform record overruns count in | |
2164 | * flight recorder mode locklessly. | |
2165 | * - When we are ready to release the subbuffer (either for | |
2166 | * reading or for overrun by other writers), we simply set the | |
2167 | * cc_sb value to "commit_count" and perform delivery. | |
2168 | * | |
2169 | * The subbuffer size is least 2 bytes (minimum size: 1 page). | |
2170 | * This guarantees that old_commit_count + 1 != commit_count. | |
2171 | */ | |
2172 | ||
2173 | /* | |
2174 | * Order prior updates to reserve count prior to the | |
2175 | * commit_cold cc_sb update. | |
2176 | */ | |
2177 | smp_wmb(); | |
2178 | if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb, | |
2179 | old_commit_count, old_commit_count + 1) | |
2180 | == old_commit_count)) { | |
2181 | /* | |
2182 | * Start of exclusive subbuffer access. We are | |
2183 | * guaranteed to be the last writer in this subbuffer | |
2184 | * and any other writer trying to access this subbuffer | |
2185 | * in this state is required to drop records. | |
2186 | */ | |
25337cb5 | 2187 | deliver_count_events(config, buf, idx); |
aece661f MD |
2188 | config->cb.buffer_end(buf, tsc, idx, |
2189 | lib_ring_buffer_get_data_size(config, | |
2190 | buf, | |
2191 | idx)); | |
2192 | ||
2193 | /* | |
2194 | * Increment the packet counter while we have exclusive | |
2195 | * access. | |
2196 | */ | |
2197 | subbuffer_inc_packet_count(config, &buf->backend, idx); | |
2198 | ||
2199 | /* | |
2200 | * Set noref flag and offset for this subbuffer id. | |
2201 | * Contains a memory barrier that ensures counter stores | |
2202 | * are ordered before set noref and offset. | |
2203 | */ | |
2204 | lib_ring_buffer_set_noref_offset(config, &buf->backend, idx, | |
2205 | buf_trunc_val(offset, chan)); | |
2206 | ||
2207 | /* | |
2208 | * Order set_noref and record counter updates before the | |
2209 | * end of subbuffer exclusive access. Orders with | |
2210 | * respect to writers coming into the subbuffer after | |
2211 | * wrap around, and also order wrt concurrent readers. | |
2212 | */ | |
2213 | smp_mb(); | |
2214 | /* End of exclusive subbuffer access */ | |
2215 | v_set(config, &buf->commit_cold[idx].cc_sb, | |
2216 | commit_count); | |
2217 | /* | |
2218 | * Order later updates to reserve count after | |
2219 | * the commit_cold cc_sb update. | |
2220 | */ | |
2221 | smp_wmb(); | |
2222 | lib_ring_buffer_vmcore_check_deliver(config, buf, | |
2223 | commit_count, idx); | |
2224 | ||
2225 | /* | |
2226 | * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free. | |
2227 | */ | |
2228 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER | |
2229 | && atomic_long_read(&buf->active_readers) | |
2230 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
2231 | wake_up_interruptible(&buf->read_wait); | |
2232 | wake_up_interruptible(&chan->read_wait); | |
2233 | } | |
2234 | ||
2235 | } | |
2236 | } | |
2237 | EXPORT_SYMBOL_GPL(lib_ring_buffer_check_deliver_slow); | |
2238 | ||
02a766bb | 2239 | int __init init_lib_ring_buffer_frontend(void) |
6fb8de4b MD |
2240 | { |
2241 | int cpu; | |
2242 | ||
2243 | for_each_possible_cpu(cpu) | |
2244 | spin_lock_init(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
02a766bb | 2245 | return 0; |
6fb8de4b | 2246 | } |
02a766bb MD |
2247 | |
2248 | module_init(init_lib_ring_buffer_frontend); | |
1a5db82d MD |
2249 | |
2250 | void __exit exit_lib_ring_buffer_frontend(void) | |
2251 | { | |
2252 | } | |
2253 | ||
2254 | module_exit(exit_lib_ring_buffer_frontend); |