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> | |
ec364dc9 | 57 | #include <asm/cacheflush.h> |
f3bc08c5 | 58 | |
c075712b MD |
59 | #include <wrapper/ringbuffer/config.h> |
60 | #include <wrapper/ringbuffer/backend.h> | |
61 | #include <wrapper/ringbuffer/frontend.h> | |
62 | #include <wrapper/ringbuffer/iterator.h> | |
63 | #include <wrapper/ringbuffer/nohz.h> | |
64 | #include <wrapper/atomic.h> | |
65 | #include <wrapper/kref.h> | |
66 | #include <wrapper/percpu-defs.h> | |
152fe7fc | 67 | #include <wrapper/timer.h> |
fd0cebd8 | 68 | #include <wrapper/vmalloc.h> |
f3bc08c5 MD |
69 | |
70 | /* | |
71 | * Internal structure representing offsets to use at a sub-buffer switch. | |
72 | */ | |
73 | struct switch_offsets { | |
74 | unsigned long begin, end, old; | |
75 | size_t pre_header_padding, size; | |
f5ea5800 MD |
76 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, |
77 | switch_old_end:1; | |
f3bc08c5 MD |
78 | }; |
79 | ||
80 | #ifdef CONFIG_NO_HZ | |
81 | enum tick_nohz_val { | |
82 | TICK_NOHZ_STOP, | |
83 | TICK_NOHZ_FLUSH, | |
84 | TICK_NOHZ_RESTART, | |
85 | }; | |
86 | ||
87 | static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier); | |
88 | #endif /* CONFIG_NO_HZ */ | |
89 | ||
90 | static DEFINE_PER_CPU(spinlock_t, ring_buffer_nohz_lock); | |
91 | ||
92 | DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting); | |
93 | EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting); | |
94 | ||
95 | static | |
96 | void lib_ring_buffer_print_errors(struct channel *chan, | |
97 | struct lib_ring_buffer *buf, int cpu); | |
64af2437 MD |
98 | static |
99 | void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf, | |
100 | enum switch_mode mode); | |
f3bc08c5 | 101 | |
aece661f MD |
102 | static |
103 | int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config, | |
104 | struct lib_ring_buffer *buf, | |
105 | struct channel *chan) | |
106 | { | |
107 | unsigned long consumed_old, consumed_idx, commit_count, write_offset; | |
108 | ||
109 | consumed_old = atomic_long_read(&buf->consumed); | |
110 | consumed_idx = subbuf_index(consumed_old, chan); | |
111 | commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); | |
112 | /* | |
113 | * No memory barrier here, since we are only interested | |
114 | * in a statistically correct polling result. The next poll will | |
115 | * get the data is we are racing. The mb() that ensures correct | |
116 | * memory order is in get_subbuf. | |
117 | */ | |
118 | write_offset = v_read(config, &buf->offset); | |
119 | ||
120 | /* | |
121 | * Check that the subbuffer we are trying to consume has been | |
122 | * already fully committed. | |
123 | */ | |
124 | ||
125 | if (((commit_count - chan->backend.subbuf_size) | |
126 | & chan->commit_count_mask) | |
127 | - (buf_trunc(consumed_old, chan) | |
128 | >> chan->backend.num_subbuf_order) | |
129 | != 0) | |
130 | return 0; | |
131 | ||
132 | /* | |
133 | * Check that we are not about to read the same subbuffer in | |
134 | * which the writer head is. | |
135 | */ | |
136 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan) | |
137 | == 0) | |
138 | return 0; | |
139 | ||
140 | return 1; | |
141 | } | |
142 | ||
f3bc08c5 MD |
143 | /* |
144 | * Must be called under cpu hotplug protection. | |
145 | */ | |
146 | void lib_ring_buffer_free(struct lib_ring_buffer *buf) | |
147 | { | |
148 | struct channel *chan = buf->backend.chan; | |
149 | ||
150 | lib_ring_buffer_print_errors(chan, buf, buf->backend.cpu); | |
fd0cebd8 MJ |
151 | lttng_kvfree(buf->commit_hot); |
152 | lttng_kvfree(buf->commit_cold); | |
a1817dcb | 153 | lttng_kvfree(buf->ts_end); |
f3bc08c5 MD |
154 | |
155 | lib_ring_buffer_backend_free(&buf->backend); | |
156 | } | |
157 | ||
158 | /** | |
159 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
160 | * @buf: Ring buffer. | |
161 | * | |
162 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
163 | * used for writing. The ring buffer can be opened for reading, but the reader | |
164 | * should not be using the iterator concurrently with reset. The previous | |
165 | * current iterator record is reset. | |
166 | */ | |
167 | void lib_ring_buffer_reset(struct lib_ring_buffer *buf) | |
168 | { | |
169 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 170 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
171 | unsigned int i; |
172 | ||
173 | /* | |
174 | * Reset iterator first. It will put the subbuffer if it currently holds | |
175 | * it. | |
176 | */ | |
177 | lib_ring_buffer_iterator_reset(buf); | |
178 | v_set(config, &buf->offset, 0); | |
179 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
180 | v_set(config, &buf->commit_hot[i].cc, 0); | |
181 | v_set(config, &buf->commit_hot[i].seq, 0); | |
182 | v_set(config, &buf->commit_cold[i].cc_sb, 0); | |
a1817dcb | 183 | buf->ts_end[i] = 0; |
f3bc08c5 MD |
184 | } |
185 | atomic_long_set(&buf->consumed, 0); | |
186 | atomic_set(&buf->record_disabled, 0); | |
187 | v_set(config, &buf->last_tsc, 0); | |
188 | lib_ring_buffer_backend_reset(&buf->backend); | |
189 | /* Don't reset number of active readers */ | |
190 | v_set(config, &buf->records_lost_full, 0); | |
191 | v_set(config, &buf->records_lost_wrap, 0); | |
192 | v_set(config, &buf->records_lost_big, 0); | |
193 | v_set(config, &buf->records_count, 0); | |
194 | v_set(config, &buf->records_overrun, 0); | |
195 | buf->finalized = 0; | |
196 | } | |
197 | EXPORT_SYMBOL_GPL(lib_ring_buffer_reset); | |
198 | ||
199 | /** | |
200 | * channel_reset - Reset channel to initial values. | |
201 | * @chan: Channel. | |
202 | * | |
203 | * Effectively empty the channel. Should be called when the channel is not used | |
204 | * for writing. The channel can be opened for reading, but the reader should not | |
205 | * be using the iterator concurrently with reset. The previous current iterator | |
206 | * record is reset. | |
207 | */ | |
208 | void channel_reset(struct channel *chan) | |
209 | { | |
210 | /* | |
211 | * Reset iterators first. Will put the subbuffer if held for reading. | |
212 | */ | |
213 | channel_iterator_reset(chan); | |
214 | atomic_set(&chan->record_disabled, 0); | |
215 | /* Don't reset commit_count_mask, still valid */ | |
216 | channel_backend_reset(&chan->backend); | |
217 | /* Don't reset switch/read timer interval */ | |
218 | /* Don't reset notifiers and notifier enable bits */ | |
219 | /* Don't reset reader reference count */ | |
220 | } | |
221 | EXPORT_SYMBOL_GPL(channel_reset); | |
222 | ||
223 | /* | |
224 | * Must be called under cpu hotplug protection. | |
225 | */ | |
226 | int lib_ring_buffer_create(struct lib_ring_buffer *buf, | |
227 | struct channel_backend *chanb, int cpu) | |
228 | { | |
5a8fd222 | 229 | const struct lib_ring_buffer_config *config = &chanb->config; |
f3bc08c5 MD |
230 | struct channel *chan = container_of(chanb, struct channel, backend); |
231 | void *priv = chanb->priv; | |
f3bc08c5 MD |
232 | size_t subbuf_header_size; |
233 | u64 tsc; | |
234 | int ret; | |
235 | ||
236 | /* Test for cpu hotplug */ | |
237 | if (buf->backend.allocated) | |
238 | return 0; | |
239 | ||
240 | /* | |
241 | * Paranoia: per cpu dynamic allocation is not officially documented as | |
242 | * zeroing the memory, so let's do it here too, just in case. | |
243 | */ | |
244 | memset(buf, 0, sizeof(*buf)); | |
245 | ||
246 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, cpu); | |
247 | if (ret) | |
248 | return ret; | |
249 | ||
250 | buf->commit_hot = | |
fd0cebd8 | 251 | lttng_kvzalloc_node(ALIGN(sizeof(*buf->commit_hot) |
f3bc08c5 MD |
252 | * chan->backend.num_subbuf, |
253 | 1 << INTERNODE_CACHE_SHIFT), | |
df388b78 MD |
254 | GFP_KERNEL | __GFP_NOWARN, |
255 | cpu_to_node(max(cpu, 0))); | |
f3bc08c5 MD |
256 | if (!buf->commit_hot) { |
257 | ret = -ENOMEM; | |
258 | goto free_chanbuf; | |
259 | } | |
260 | ||
261 | buf->commit_cold = | |
fd0cebd8 | 262 | lttng_kvzalloc_node(ALIGN(sizeof(*buf->commit_cold) |
f3bc08c5 MD |
263 | * chan->backend.num_subbuf, |
264 | 1 << INTERNODE_CACHE_SHIFT), | |
df388b78 MD |
265 | GFP_KERNEL | __GFP_NOWARN, |
266 | cpu_to_node(max(cpu, 0))); | |
f3bc08c5 MD |
267 | if (!buf->commit_cold) { |
268 | ret = -ENOMEM; | |
269 | goto free_commit; | |
270 | } | |
271 | ||
a1817dcb MD |
272 | buf->ts_end = |
273 | lttng_kvzalloc_node(ALIGN(sizeof(*buf->ts_end) | |
274 | * chan->backend.num_subbuf, | |
275 | 1 << INTERNODE_CACHE_SHIFT), | |
276 | GFP_KERNEL | __GFP_NOWARN, | |
277 | cpu_to_node(max(cpu, 0))); | |
278 | if (!buf->ts_end) { | |
279 | ret = -ENOMEM; | |
280 | goto free_commit_cold; | |
281 | } | |
282 | ||
f3bc08c5 | 283 | init_waitqueue_head(&buf->read_wait); |
71c1d843 | 284 | init_waitqueue_head(&buf->write_wait); |
f3bc08c5 MD |
285 | raw_spin_lock_init(&buf->raw_tick_nohz_spinlock); |
286 | ||
287 | /* | |
288 | * Write the subbuffer header for first subbuffer so we know the total | |
289 | * duration of data gathering. | |
290 | */ | |
291 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
292 | v_set(config, &buf->offset, subbuf_header_size); | |
293 | subbuffer_id_clear_noref(config, &buf->backend.buf_wsb[0].id); | |
294 | tsc = config->cb.ring_buffer_clock_read(buf->backend.chan); | |
295 | config->cb.buffer_begin(buf, tsc, 0); | |
296 | v_add(config, subbuf_header_size, &buf->commit_hot[0].cc); | |
297 | ||
298 | if (config->cb.buffer_create) { | |
299 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name); | |
300 | if (ret) | |
301 | goto free_init; | |
302 | } | |
303 | ||
304 | /* | |
305 | * Ensure the buffer is ready before setting it to allocated and setting | |
306 | * the cpumask. | |
307 | * Used for cpu hotplug vs cpumask iteration. | |
308 | */ | |
309 | smp_wmb(); | |
310 | buf->backend.allocated = 1; | |
311 | ||
312 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
313 | CHAN_WARN_ON(chan, cpumask_test_cpu(cpu, | |
314 | chan->backend.cpumask)); | |
315 | cpumask_set_cpu(cpu, chan->backend.cpumask); | |
316 | } | |
317 | ||
318 | return 0; | |
319 | ||
320 | /* Error handling */ | |
321 | free_init: | |
a1817dcb MD |
322 | lttng_kvfree(buf->ts_end); |
323 | free_commit_cold: | |
fd0cebd8 | 324 | lttng_kvfree(buf->commit_cold); |
f3bc08c5 | 325 | free_commit: |
fd0cebd8 | 326 | lttng_kvfree(buf->commit_hot); |
f3bc08c5 MD |
327 | free_chanbuf: |
328 | lib_ring_buffer_backend_free(&buf->backend); | |
329 | return ret; | |
330 | } | |
331 | ||
df93abe8 | 332 | static void switch_buffer_timer(LTTNG_TIMER_FUNC_ARG_TYPE t) |
f3bc08c5 | 333 | { |
df93abe8 | 334 | struct lib_ring_buffer *buf = lttng_from_timer(buf, t, switch_timer); |
f3bc08c5 | 335 | struct channel *chan = buf->backend.chan; |
5a8fd222 | 336 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
337 | |
338 | /* | |
339 | * Only flush buffers periodically if readers are active. | |
340 | */ | |
341 | if (atomic_long_read(&buf->active_readers)) | |
342 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
343 | ||
344 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
152fe7fc | 345 | lttng_mod_timer_pinned(&buf->switch_timer, |
f3bc08c5 MD |
346 | jiffies + chan->switch_timer_interval); |
347 | else | |
348 | mod_timer(&buf->switch_timer, | |
349 | jiffies + chan->switch_timer_interval); | |
350 | } | |
351 | ||
352 | /* | |
353 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
354 | */ | |
355 | static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer *buf) | |
356 | { | |
357 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 358 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
df93abe8 | 359 | unsigned int flags = 0; |
f3bc08c5 MD |
360 | |
361 | if (!chan->switch_timer_interval || buf->switch_timer_enabled) | |
362 | return; | |
152fe7fc MJ |
363 | |
364 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
df93abe8 | 365 | flags = LTTNG_TIMER_PINNED; |
152fe7fc | 366 | |
df93abe8 | 367 | lttng_timer_setup(&buf->switch_timer, switch_buffer_timer, flags, buf); |
f3bc08c5 | 368 | buf->switch_timer.expires = jiffies + chan->switch_timer_interval; |
df93abe8 | 369 | |
f3bc08c5 MD |
370 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
371 | add_timer_on(&buf->switch_timer, buf->backend.cpu); | |
372 | else | |
373 | add_timer(&buf->switch_timer); | |
df93abe8 | 374 | |
f3bc08c5 MD |
375 | buf->switch_timer_enabled = 1; |
376 | } | |
377 | ||
378 | /* | |
379 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
380 | */ | |
381 | static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer *buf) | |
382 | { | |
383 | struct channel *chan = buf->backend.chan; | |
384 | ||
385 | if (!chan->switch_timer_interval || !buf->switch_timer_enabled) | |
386 | return; | |
387 | ||
388 | del_timer_sync(&buf->switch_timer); | |
389 | buf->switch_timer_enabled = 0; | |
390 | } | |
391 | ||
392 | /* | |
393 | * Polling timer to check the channels for data. | |
394 | */ | |
df93abe8 | 395 | static void read_buffer_timer(LTTNG_TIMER_FUNC_ARG_TYPE t) |
f3bc08c5 | 396 | { |
df93abe8 | 397 | struct lib_ring_buffer *buf = lttng_from_timer(buf, t, read_timer); |
f3bc08c5 | 398 | struct channel *chan = buf->backend.chan; |
5a8fd222 | 399 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
400 | |
401 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
402 | ||
403 | if (atomic_long_read(&buf->active_readers) | |
404 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
405 | wake_up_interruptible(&buf->read_wait); | |
406 | wake_up_interruptible(&chan->read_wait); | |
407 | } | |
408 | ||
409 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
152fe7fc | 410 | lttng_mod_timer_pinned(&buf->read_timer, |
f3bc08c5 MD |
411 | jiffies + chan->read_timer_interval); |
412 | else | |
413 | mod_timer(&buf->read_timer, | |
414 | jiffies + chan->read_timer_interval); | |
415 | } | |
416 | ||
417 | /* | |
418 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
419 | */ | |
420 | static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf) | |
421 | { | |
422 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 423 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
ddef58a9 | 424 | unsigned int flags = 0; |
f3bc08c5 MD |
425 | |
426 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
427 | || !chan->read_timer_interval | |
428 | || buf->read_timer_enabled) | |
429 | return; | |
430 | ||
152fe7fc | 431 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
df93abe8 | 432 | flags = LTTNG_TIMER_PINNED; |
152fe7fc | 433 | |
df93abe8 | 434 | lttng_timer_setup(&buf->read_timer, read_buffer_timer, flags, buf); |
f3bc08c5 | 435 | buf->read_timer.expires = jiffies + chan->read_timer_interval; |
f3bc08c5 MD |
436 | |
437 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
438 | add_timer_on(&buf->read_timer, buf->backend.cpu); | |
439 | else | |
440 | add_timer(&buf->read_timer); | |
df93abe8 | 441 | |
f3bc08c5 MD |
442 | buf->read_timer_enabled = 1; |
443 | } | |
444 | ||
445 | /* | |
446 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
447 | */ | |
448 | static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer *buf) | |
449 | { | |
450 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 451 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
452 | |
453 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
454 | || !chan->read_timer_interval | |
455 | || !buf->read_timer_enabled) | |
456 | return; | |
457 | ||
458 | del_timer_sync(&buf->read_timer); | |
459 | /* | |
460 | * do one more check to catch data that has been written in the last | |
461 | * timer period. | |
462 | */ | |
463 | if (lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
464 | wake_up_interruptible(&buf->read_wait); | |
465 | wake_up_interruptible(&chan->read_wait); | |
466 | } | |
467 | buf->read_timer_enabled = 0; | |
468 | } | |
469 | ||
360d3efe | 470 | #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) |
1e367326 MD |
471 | |
472 | enum cpuhp_state lttng_rb_hp_prepare; | |
473 | enum cpuhp_state lttng_rb_hp_online; | |
474 | ||
475 | void lttng_rb_set_hp_prepare(enum cpuhp_state val) | |
476 | { | |
477 | lttng_rb_hp_prepare = val; | |
478 | } | |
479 | EXPORT_SYMBOL_GPL(lttng_rb_set_hp_prepare); | |
480 | ||
481 | void lttng_rb_set_hp_online(enum cpuhp_state val) | |
482 | { | |
483 | lttng_rb_hp_online = val; | |
484 | } | |
485 | EXPORT_SYMBOL_GPL(lttng_rb_set_hp_online); | |
486 | ||
487 | int lttng_cpuhp_rb_frontend_dead(unsigned int cpu, | |
488 | struct lttng_cpuhp_node *node) | |
489 | { | |
490 | struct channel *chan = container_of(node, struct channel, | |
491 | cpuhp_prepare); | |
492 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
493 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
494 | ||
495 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
496 | ||
497 | /* | |
498 | * Performing a buffer switch on a remote CPU. Performed by | |
499 | * the CPU responsible for doing the hotunplug after the target | |
500 | * CPU stopped running completely. Ensures that all data | |
501 | * from that remote CPU is flushed. | |
502 | */ | |
503 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
504 | return 0; | |
505 | } | |
506 | EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_dead); | |
507 | ||
508 | int lttng_cpuhp_rb_frontend_online(unsigned int cpu, | |
509 | struct lttng_cpuhp_node *node) | |
510 | { | |
511 | struct channel *chan = container_of(node, struct channel, | |
512 | cpuhp_online); | |
513 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
514 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
515 | ||
516 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
517 | ||
518 | wake_up_interruptible(&chan->hp_wait); | |
519 | lib_ring_buffer_start_switch_timer(buf); | |
520 | lib_ring_buffer_start_read_timer(buf); | |
521 | return 0; | |
522 | } | |
523 | EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_online); | |
524 | ||
525 | int lttng_cpuhp_rb_frontend_offline(unsigned int cpu, | |
526 | struct lttng_cpuhp_node *node) | |
527 | { | |
528 | struct channel *chan = container_of(node, struct channel, | |
529 | cpuhp_online); | |
530 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
531 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
532 | ||
533 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
534 | ||
535 | lib_ring_buffer_stop_switch_timer(buf); | |
536 | lib_ring_buffer_stop_read_timer(buf); | |
537 | return 0; | |
538 | } | |
539 | EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_offline); | |
540 | ||
360d3efe | 541 | #else /* #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
1e367326 | 542 | |
f3bc08c5 | 543 | #ifdef CONFIG_HOTPLUG_CPU |
1e367326 | 544 | |
f3bc08c5 MD |
545 | /** |
546 | * lib_ring_buffer_cpu_hp_callback - CPU hotplug callback | |
547 | * @nb: notifier block | |
548 | * @action: hotplug action to take | |
549 | * @hcpu: CPU number | |
550 | * | |
551 | * Returns the success/failure of the operation. (%NOTIFY_OK, %NOTIFY_BAD) | |
552 | */ | |
553 | static | |
e8f071d5 | 554 | int lib_ring_buffer_cpu_hp_callback(struct notifier_block *nb, |
f3bc08c5 MD |
555 | unsigned long action, |
556 | void *hcpu) | |
557 | { | |
558 | unsigned int cpu = (unsigned long)hcpu; | |
559 | struct channel *chan = container_of(nb, struct channel, | |
560 | cpu_hp_notifier); | |
561 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
5a8fd222 | 562 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
563 | |
564 | if (!chan->cpu_hp_enable) | |
565 | return NOTIFY_DONE; | |
566 | ||
567 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
568 | ||
569 | switch (action) { | |
570 | case CPU_DOWN_FAILED: | |
571 | case CPU_DOWN_FAILED_FROZEN: | |
572 | case CPU_ONLINE: | |
573 | case CPU_ONLINE_FROZEN: | |
24cedcfe | 574 | wake_up_interruptible(&chan->hp_wait); |
f3bc08c5 MD |
575 | lib_ring_buffer_start_switch_timer(buf); |
576 | lib_ring_buffer_start_read_timer(buf); | |
577 | return NOTIFY_OK; | |
578 | ||
579 | case CPU_DOWN_PREPARE: | |
580 | case CPU_DOWN_PREPARE_FROZEN: | |
581 | lib_ring_buffer_stop_switch_timer(buf); | |
582 | lib_ring_buffer_stop_read_timer(buf); | |
583 | return NOTIFY_OK; | |
584 | ||
585 | case CPU_DEAD: | |
586 | case CPU_DEAD_FROZEN: | |
587 | /* | |
588 | * Performing a buffer switch on a remote CPU. Performed by | |
589 | * the CPU responsible for doing the hotunplug after the target | |
590 | * CPU stopped running completely. Ensures that all data | |
591 | * from that remote CPU is flushed. | |
592 | */ | |
593 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
594 | return NOTIFY_OK; | |
595 | ||
596 | default: | |
597 | return NOTIFY_DONE; | |
598 | } | |
599 | } | |
1e367326 | 600 | |
f3bc08c5 MD |
601 | #endif |
602 | ||
360d3efe | 603 | #endif /* #else #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
1e367326 | 604 | |
23b908b0 | 605 | #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) |
f3bc08c5 MD |
606 | /* |
607 | * For per-cpu buffers, call the reader wakeups before switching the buffer, so | |
608 | * that wake-up-tracing generated events are flushed before going idle (in | |
609 | * tick_nohz). We test if the spinlock is locked to deal with the race where | |
610 | * readers try to sample the ring buffer before we perform the switch. We let | |
611 | * the readers retry in that case. If there is data in the buffer, the wake up | |
612 | * is going to forbid the CPU running the reader thread from going idle. | |
613 | */ | |
614 | static int notrace ring_buffer_tick_nohz_callback(struct notifier_block *nb, | |
615 | unsigned long val, | |
616 | void *data) | |
617 | { | |
618 | struct channel *chan = container_of(nb, struct channel, | |
619 | tick_nohz_notifier); | |
5a8fd222 | 620 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
621 | struct lib_ring_buffer *buf; |
622 | int cpu = smp_processor_id(); | |
623 | ||
624 | if (config->alloc != RING_BUFFER_ALLOC_PER_CPU) { | |
625 | /* | |
626 | * We don't support keeping the system idle with global buffers | |
627 | * and streaming active. In order to do so, we would need to | |
628 | * sample a non-nohz-cpumask racelessly with the nohz updates | |
629 | * without adding synchronization overhead to nohz. Leave this | |
630 | * use-case out for now. | |
631 | */ | |
632 | return 0; | |
633 | } | |
634 | ||
635 | buf = channel_get_ring_buffer(config, chan, cpu); | |
636 | switch (val) { | |
637 | case TICK_NOHZ_FLUSH: | |
638 | raw_spin_lock(&buf->raw_tick_nohz_spinlock); | |
639 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_TIMER | |
640 | && chan->read_timer_interval | |
641 | && atomic_long_read(&buf->active_readers) | |
642 | && (lib_ring_buffer_poll_deliver(config, buf, chan) | |
643 | || lib_ring_buffer_pending_data(config, buf, chan))) { | |
644 | wake_up_interruptible(&buf->read_wait); | |
645 | wake_up_interruptible(&chan->read_wait); | |
646 | } | |
647 | if (chan->switch_timer_interval) | |
648 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
649 | raw_spin_unlock(&buf->raw_tick_nohz_spinlock); | |
650 | break; | |
651 | case TICK_NOHZ_STOP: | |
e6b06d7d | 652 | spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
653 | lib_ring_buffer_stop_switch_timer(buf); |
654 | lib_ring_buffer_stop_read_timer(buf); | |
e6b06d7d | 655 | spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
656 | break; |
657 | case TICK_NOHZ_RESTART: | |
e6b06d7d | 658 | spin_lock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
659 | lib_ring_buffer_start_read_timer(buf); |
660 | lib_ring_buffer_start_switch_timer(buf); | |
e6b06d7d | 661 | spin_unlock(lttng_this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
662 | break; |
663 | } | |
664 | ||
665 | return 0; | |
666 | } | |
667 | ||
668 | void notrace lib_ring_buffer_tick_nohz_flush(void) | |
669 | { | |
670 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_FLUSH, | |
671 | NULL); | |
672 | } | |
673 | ||
674 | void notrace lib_ring_buffer_tick_nohz_stop(void) | |
675 | { | |
676 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_STOP, | |
677 | NULL); | |
678 | } | |
679 | ||
680 | void notrace lib_ring_buffer_tick_nohz_restart(void) | |
681 | { | |
682 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_RESTART, | |
683 | NULL); | |
684 | } | |
23b908b0 | 685 | #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */ |
f3bc08c5 MD |
686 | |
687 | /* | |
688 | * Holds CPU hotplug. | |
689 | */ | |
690 | static void channel_unregister_notifiers(struct channel *chan) | |
691 | { | |
5a8fd222 | 692 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
693 | |
694 | channel_iterator_unregister_notifiers(chan); | |
695 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
696 | #ifdef CONFIG_NO_HZ | |
697 | /* | |
698 | * Remove the nohz notifier first, so we are certain we stop | |
699 | * the timers. | |
700 | */ | |
701 | atomic_notifier_chain_unregister(&tick_nohz_notifier, | |
702 | &chan->tick_nohz_notifier); | |
703 | /* | |
704 | * ring_buffer_nohz_lock will not be needed below, because | |
705 | * we just removed the notifiers, which were the only source of | |
706 | * concurrency. | |
707 | */ | |
708 | #endif /* CONFIG_NO_HZ */ | |
360d3efe | 709 | #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) |
1e367326 MD |
710 | { |
711 | int ret; | |
712 | ||
713 | ret = cpuhp_state_remove_instance(lttng_rb_hp_online, | |
714 | &chan->cpuhp_online.node); | |
715 | WARN_ON(ret); | |
716 | ret = cpuhp_state_remove_instance_nocalls(lttng_rb_hp_prepare, | |
717 | &chan->cpuhp_prepare.node); | |
718 | WARN_ON(ret); | |
f3bc08c5 | 719 | } |
360d3efe | 720 | #else /* #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
1e367326 MD |
721 | { |
722 | int cpu; | |
723 | ||
724 | #ifdef CONFIG_HOTPLUG_CPU | |
725 | get_online_cpus(); | |
726 | chan->cpu_hp_enable = 0; | |
727 | for_each_online_cpu(cpu) { | |
728 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
729 | cpu); | |
730 | lib_ring_buffer_stop_switch_timer(buf); | |
731 | lib_ring_buffer_stop_read_timer(buf); | |
732 | } | |
733 | put_online_cpus(); | |
734 | unregister_cpu_notifier(&chan->cpu_hp_notifier); | |
f3bc08c5 | 735 | #else |
1e367326 MD |
736 | for_each_possible_cpu(cpu) { |
737 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
738 | cpu); | |
739 | lib_ring_buffer_stop_switch_timer(buf); | |
740 | lib_ring_buffer_stop_read_timer(buf); | |
741 | } | |
f3bc08c5 | 742 | #endif |
1e367326 | 743 | } |
360d3efe | 744 | #endif /* #else #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
f3bc08c5 MD |
745 | } else { |
746 | struct lib_ring_buffer *buf = chan->backend.buf; | |
747 | ||
748 | lib_ring_buffer_stop_switch_timer(buf); | |
749 | lib_ring_buffer_stop_read_timer(buf); | |
750 | } | |
751 | channel_backend_unregister_notifiers(&chan->backend); | |
752 | } | |
753 | ||
64af2437 MD |
754 | static void lib_ring_buffer_set_quiescent(struct lib_ring_buffer *buf) |
755 | { | |
756 | if (!buf->quiescent) { | |
757 | buf->quiescent = true; | |
758 | _lib_ring_buffer_switch_remote(buf, SWITCH_FLUSH); | |
759 | } | |
760 | } | |
761 | ||
762 | static void lib_ring_buffer_clear_quiescent(struct lib_ring_buffer *buf) | |
763 | { | |
764 | buf->quiescent = false; | |
765 | } | |
766 | ||
767 | void lib_ring_buffer_set_quiescent_channel(struct channel *chan) | |
768 | { | |
769 | int cpu; | |
770 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
771 | ||
772 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
773 | get_online_cpus(); | |
774 | for_each_channel_cpu(cpu, chan) { | |
775 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
776 | cpu); | |
777 | ||
778 | lib_ring_buffer_set_quiescent(buf); | |
779 | } | |
780 | put_online_cpus(); | |
781 | } else { | |
782 | struct lib_ring_buffer *buf = chan->backend.buf; | |
783 | ||
784 | lib_ring_buffer_set_quiescent(buf); | |
785 | } | |
786 | } | |
787 | EXPORT_SYMBOL_GPL(lib_ring_buffer_set_quiescent_channel); | |
788 | ||
789 | void lib_ring_buffer_clear_quiescent_channel(struct channel *chan) | |
790 | { | |
791 | int cpu; | |
792 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
793 | ||
794 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
795 | get_online_cpus(); | |
796 | for_each_channel_cpu(cpu, chan) { | |
797 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
798 | cpu); | |
799 | ||
800 | lib_ring_buffer_clear_quiescent(buf); | |
801 | } | |
802 | put_online_cpus(); | |
803 | } else { | |
804 | struct lib_ring_buffer *buf = chan->backend.buf; | |
805 | ||
806 | lib_ring_buffer_clear_quiescent(buf); | |
807 | } | |
808 | } | |
809 | EXPORT_SYMBOL_GPL(lib_ring_buffer_clear_quiescent_channel); | |
810 | ||
f3bc08c5 MD |
811 | static void channel_free(struct channel *chan) |
812 | { | |
dd5a0db3 MD |
813 | if (chan->backend.release_priv_ops) { |
814 | chan->backend.release_priv_ops(chan->backend.priv_ops); | |
815 | } | |
f3bc08c5 MD |
816 | channel_iterator_free(chan); |
817 | channel_backend_free(&chan->backend); | |
818 | kfree(chan); | |
819 | } | |
820 | ||
821 | /** | |
822 | * channel_create - Create channel. | |
823 | * @config: ring buffer instance configuration | |
824 | * @name: name of the channel | |
825 | * @priv: ring buffer client private data | |
826 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous | |
827 | * address mapping. It is used only by RING_BUFFER_STATIC | |
828 | * configuration. It can be set to NULL for other backends. | |
829 | * @subbuf_size: subbuffer size | |
830 | * @num_subbuf: number of subbuffers | |
831 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
832 | * padding to let readers get those sub-buffers. | |
833 | * Used for live streaming. | |
834 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
835 | * | |
836 | * Holds cpu hotplug. | |
837 | * Returns NULL on failure. | |
838 | */ | |
839 | struct channel *channel_create(const struct lib_ring_buffer_config *config, | |
840 | const char *name, void *priv, void *buf_addr, | |
841 | size_t subbuf_size, | |
842 | size_t num_subbuf, unsigned int switch_timer_interval, | |
843 | unsigned int read_timer_interval) | |
844 | { | |
1e367326 | 845 | int ret; |
f3bc08c5 MD |
846 | struct channel *chan; |
847 | ||
848 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
849 | read_timer_interval)) | |
850 | return NULL; | |
851 | ||
852 | chan = kzalloc(sizeof(struct channel), GFP_KERNEL); | |
853 | if (!chan) | |
854 | return NULL; | |
855 | ||
856 | ret = channel_backend_init(&chan->backend, name, config, priv, | |
857 | subbuf_size, num_subbuf); | |
858 | if (ret) | |
859 | goto error; | |
860 | ||
861 | ret = channel_iterator_init(chan); | |
862 | if (ret) | |
863 | goto error_free_backend; | |
864 | ||
865 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
866 | chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
867 | chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
f40270ad | 868 | kref_init(&chan->ref); |
f3bc08c5 | 869 | init_waitqueue_head(&chan->read_wait); |
24cedcfe | 870 | init_waitqueue_head(&chan->hp_wait); |
f3bc08c5 MD |
871 | |
872 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
360d3efe | 873 | #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) |
1e367326 MD |
874 | chan->cpuhp_prepare.component = LTTNG_RING_BUFFER_FRONTEND; |
875 | ret = cpuhp_state_add_instance_nocalls(lttng_rb_hp_prepare, | |
876 | &chan->cpuhp_prepare.node); | |
877 | if (ret) | |
878 | goto cpuhp_prepare_error; | |
879 | ||
880 | chan->cpuhp_online.component = LTTNG_RING_BUFFER_FRONTEND; | |
881 | ret = cpuhp_state_add_instance(lttng_rb_hp_online, | |
882 | &chan->cpuhp_online.node); | |
883 | if (ret) | |
884 | goto cpuhp_online_error; | |
360d3efe | 885 | #else /* #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
1e367326 MD |
886 | { |
887 | int cpu; | |
888 | /* | |
889 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
890 | * in early initcall, it will not be notified of secondary cpus. | |
891 | * In that off case, we need to allocate for all possible cpus. | |
892 | */ | |
893 | #ifdef CONFIG_HOTPLUG_CPU | |
894 | chan->cpu_hp_notifier.notifier_call = | |
895 | lib_ring_buffer_cpu_hp_callback; | |
896 | chan->cpu_hp_notifier.priority = 6; | |
897 | register_cpu_notifier(&chan->cpu_hp_notifier); | |
898 | ||
899 | get_online_cpus(); | |
900 | for_each_online_cpu(cpu) { | |
901 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
902 | cpu); | |
903 | spin_lock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
904 | lib_ring_buffer_start_switch_timer(buf); | |
905 | lib_ring_buffer_start_read_timer(buf); | |
906 | spin_unlock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
907 | } | |
908 | chan->cpu_hp_enable = 1; | |
909 | put_online_cpus(); | |
910 | #else | |
911 | for_each_possible_cpu(cpu) { | |
912 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
913 | cpu); | |
914 | spin_lock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
915 | lib_ring_buffer_start_switch_timer(buf); | |
916 | lib_ring_buffer_start_read_timer(buf); | |
917 | spin_unlock(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
918 | } | |
919 | #endif | |
920 | } | |
360d3efe | 921 | #endif /* #else #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
1e367326 | 922 | |
23b908b0 | 923 | #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) |
f3bc08c5 MD |
924 | /* Only benefit from NO_HZ idle with per-cpu buffers for now. */ |
925 | chan->tick_nohz_notifier.notifier_call = | |
926 | ring_buffer_tick_nohz_callback; | |
927 | chan->tick_nohz_notifier.priority = ~0U; | |
928 | atomic_notifier_chain_register(&tick_nohz_notifier, | |
929 | &chan->tick_nohz_notifier); | |
23b908b0 | 930 | #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */ |
f3bc08c5 | 931 | |
f3bc08c5 MD |
932 | } else { |
933 | struct lib_ring_buffer *buf = chan->backend.buf; | |
934 | ||
935 | lib_ring_buffer_start_switch_timer(buf); | |
936 | lib_ring_buffer_start_read_timer(buf); | |
937 | } | |
938 | ||
939 | return chan; | |
940 | ||
360d3efe | 941 | #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) |
1e367326 MD |
942 | cpuhp_online_error: |
943 | ret = cpuhp_state_remove_instance_nocalls(lttng_rb_hp_prepare, | |
944 | &chan->cpuhp_prepare.node); | |
945 | WARN_ON(ret); | |
946 | cpuhp_prepare_error: | |
360d3efe | 947 | #endif /* #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(4,10,0)) */ |
f3bc08c5 MD |
948 | error_free_backend: |
949 | channel_backend_free(&chan->backend); | |
950 | error: | |
951 | kfree(chan); | |
952 | return NULL; | |
953 | } | |
954 | EXPORT_SYMBOL_GPL(channel_create); | |
955 | ||
f40270ad MD |
956 | static |
957 | void channel_release(struct kref *kref) | |
958 | { | |
959 | struct channel *chan = container_of(kref, struct channel, ref); | |
960 | channel_free(chan); | |
961 | } | |
962 | ||
f3bc08c5 MD |
963 | /** |
964 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
965 | * @chan: channel to destroy | |
966 | * | |
967 | * Holds cpu hotplug. | |
9a0df743 MD |
968 | * Call "destroy" callback, finalize channels, and then decrement the |
969 | * channel reference count. Note that when readers have completed data | |
970 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
971 | * They should release their handle at that point. Returns the private | |
972 | * data pointer. | |
f3bc08c5 MD |
973 | */ |
974 | void *channel_destroy(struct channel *chan) | |
975 | { | |
976 | int cpu; | |
5a8fd222 | 977 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
978 | void *priv; |
979 | ||
980 | channel_unregister_notifiers(chan); | |
981 | ||
982 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
983 | /* | |
984 | * No need to hold cpu hotplug, because all notifiers have been | |
985 | * unregistered. | |
986 | */ | |
987 | for_each_channel_cpu(cpu, chan) { | |
988 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
989 | cpu); | |
990 | ||
991 | if (config->cb.buffer_finalize) | |
992 | config->cb.buffer_finalize(buf, | |
993 | chan->backend.priv, | |
994 | cpu); | |
f3bc08c5 MD |
995 | /* |
996 | * Perform flush before writing to finalized. | |
997 | */ | |
998 | smp_wmb(); | |
acb98454 | 999 | WRITE_ONCE(buf->finalized, 1); |
f3bc08c5 MD |
1000 | wake_up_interruptible(&buf->read_wait); |
1001 | } | |
1002 | } else { | |
1003 | struct lib_ring_buffer *buf = chan->backend.buf; | |
1004 | ||
1005 | if (config->cb.buffer_finalize) | |
1006 | config->cb.buffer_finalize(buf, chan->backend.priv, -1); | |
f3bc08c5 MD |
1007 | /* |
1008 | * Perform flush before writing to finalized. | |
1009 | */ | |
1010 | smp_wmb(); | |
acb98454 | 1011 | WRITE_ONCE(buf->finalized, 1); |
f3bc08c5 MD |
1012 | wake_up_interruptible(&buf->read_wait); |
1013 | } | |
acb98454 | 1014 | WRITE_ONCE(chan->finalized, 1); |
24cedcfe | 1015 | wake_up_interruptible(&chan->hp_wait); |
f3bc08c5 | 1016 | wake_up_interruptible(&chan->read_wait); |
f3bc08c5 | 1017 | priv = chan->backend.priv; |
ba1d61bc | 1018 | kref_put(&chan->ref, channel_release); |
f3bc08c5 MD |
1019 | return priv; |
1020 | } | |
1021 | EXPORT_SYMBOL_GPL(channel_destroy); | |
1022 | ||
1023 | struct lib_ring_buffer *channel_get_ring_buffer( | |
1024 | const struct lib_ring_buffer_config *config, | |
1025 | struct channel *chan, int cpu) | |
1026 | { | |
1027 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) | |
1028 | return chan->backend.buf; | |
1029 | else | |
1030 | return per_cpu_ptr(chan->backend.buf, cpu); | |
1031 | } | |
1032 | EXPORT_SYMBOL_GPL(channel_get_ring_buffer); | |
1033 | ||
1034 | int lib_ring_buffer_open_read(struct lib_ring_buffer *buf) | |
1035 | { | |
1036 | struct channel *chan = buf->backend.chan; | |
1037 | ||
1038 | if (!atomic_long_add_unless(&buf->active_readers, 1, 1)) | |
1039 | return -EBUSY; | |
9c1f4643 MD |
1040 | if (!lttng_kref_get(&chan->ref)) { |
1041 | atomic_long_dec(&buf->active_readers); | |
1042 | return -EOVERFLOW; | |
1043 | } | |
505fb410 | 1044 | lttng_smp_mb__after_atomic(); |
f3bc08c5 MD |
1045 | return 0; |
1046 | } | |
1047 | EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read); | |
1048 | ||
1049 | void lib_ring_buffer_release_read(struct lib_ring_buffer *buf) | |
1050 | { | |
1051 | struct channel *chan = buf->backend.chan; | |
1052 | ||
1053 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
505fb410 | 1054 | lttng_smp_mb__before_atomic(); |
f3bc08c5 | 1055 | atomic_long_dec(&buf->active_readers); |
f40270ad | 1056 | kref_put(&chan->ref, channel_release); |
f3bc08c5 MD |
1057 | } |
1058 | EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read); | |
1059 | ||
1060 | /* | |
1061 | * Promote compiler barrier to a smp_mb(). | |
1062 | * For the specific ring buffer case, this IPI call should be removed if the | |
1063 | * architecture does not reorder writes. This should eventually be provided by | |
1064 | * a separate architecture-specific infrastructure. | |
1065 | */ | |
1066 | static void remote_mb(void *info) | |
1067 | { | |
1068 | smp_mb(); | |
1069 | } | |
1070 | ||
1071 | /** | |
1072 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
1073 | * @buf: ring buffer | |
1074 | * @consumed: consumed count indicating the position where to read | |
1075 | * @produced: produced count, indicates position when to stop reading | |
1076 | * | |
1077 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1078 | * data to read at consumed position, or 0 if the get operation succeeds. | |
1079 | * Busy-loop trying to get data if the tick_nohz sequence lock is held. | |
1080 | */ | |
1081 | ||
1082 | int lib_ring_buffer_snapshot(struct lib_ring_buffer *buf, | |
1083 | unsigned long *consumed, unsigned long *produced) | |
1084 | { | |
1085 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1086 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1087 | unsigned long consumed_cur, write_offset; |
1088 | int finalized; | |
1089 | ||
1090 | retry: | |
acb98454 | 1091 | finalized = READ_ONCE(buf->finalized); |
f3bc08c5 MD |
1092 | /* |
1093 | * Read finalized before counters. | |
1094 | */ | |
1095 | smp_rmb(); | |
1096 | consumed_cur = atomic_long_read(&buf->consumed); | |
1097 | /* | |
1098 | * No need to issue a memory barrier between consumed count read and | |
1099 | * write offset read, because consumed count can only change | |
1100 | * concurrently in overwrite mode, and we keep a sequence counter | |
1101 | * identifier derived from the write offset to check we are getting | |
1102 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
1103 | * "tagged" upon writes, tags are checked upon read). | |
1104 | */ | |
1105 | write_offset = v_read(config, &buf->offset); | |
1106 | ||
1107 | /* | |
1108 | * Check that we are not about to read the same subbuffer in | |
1109 | * which the writer head is. | |
1110 | */ | |
1111 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
1112 | == 0) | |
1113 | goto nodata; | |
1114 | ||
1115 | *consumed = consumed_cur; | |
1116 | *produced = subbuf_trunc(write_offset, chan); | |
1117 | ||
1118 | return 0; | |
1119 | ||
1120 | nodata: | |
1121 | /* | |
1122 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1123 | * of "raw_spin_is_locked" memory ordering. | |
1124 | */ | |
1125 | if (finalized) | |
1126 | return -ENODATA; | |
1127 | else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock)) | |
1128 | goto retry; | |
1129 | else | |
1130 | return -EAGAIN; | |
1131 | } | |
1132 | EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot); | |
1133 | ||
4dce5a48 JG |
1134 | /** |
1135 | * Performs the same function as lib_ring_buffer_snapshot(), but the positions | |
1136 | * are saved regardless of whether the consumed and produced positions are | |
1137 | * in the same subbuffer. | |
1138 | * @buf: ring buffer | |
1139 | * @consumed: consumed byte count indicating the last position read | |
1140 | * @produced: produced byte count indicating the last position written | |
1141 | * | |
1142 | * This function is meant to provide information on the exact producer and | |
1143 | * consumer positions without regard for the "snapshot" feature. | |
1144 | */ | |
1145 | int lib_ring_buffer_snapshot_sample_positions(struct lib_ring_buffer *buf, | |
1146 | unsigned long *consumed, unsigned long *produced) | |
1147 | { | |
1148 | struct channel *chan = buf->backend.chan; | |
1149 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1150 | ||
1151 | smp_rmb(); | |
1152 | *consumed = atomic_long_read(&buf->consumed); | |
1153 | /* | |
1154 | * No need to issue a memory barrier between consumed count read and | |
1155 | * write offset read, because consumed count can only change | |
1156 | * concurrently in overwrite mode, and we keep a sequence counter | |
1157 | * identifier derived from the write offset to check we are getting | |
1158 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
1159 | * "tagged" upon writes, tags are checked upon read). | |
1160 | */ | |
1161 | *produced = v_read(config, &buf->offset); | |
1162 | return 0; | |
1163 | } | |
1164 | ||
f3bc08c5 MD |
1165 | /** |
1166 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
71c1d843 MD |
1167 | * |
1168 | * Should only be called from consumer context. | |
f3bc08c5 MD |
1169 | * @buf: ring buffer |
1170 | * @consumed_new: new consumed count value | |
1171 | */ | |
1172 | void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf, | |
1173 | unsigned long consumed_new) | |
1174 | { | |
1175 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1176 | struct channel *chan = bufb->chan; | |
1177 | unsigned long consumed; | |
1178 | ||
1179 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
1180 | ||
1181 | /* | |
1182 | * Only push the consumed value forward. | |
1183 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
1184 | * the writer in flight recorder mode. | |
1185 | */ | |
1186 | consumed = atomic_long_read(&buf->consumed); | |
1187 | while ((long) consumed - (long) consumed_new < 0) | |
1188 | consumed = atomic_long_cmpxchg(&buf->consumed, consumed, | |
1189 | consumed_new); | |
71c1d843 MD |
1190 | /* Wake-up the metadata producer */ |
1191 | wake_up_interruptible(&buf->write_wait); | |
f3bc08c5 MD |
1192 | } |
1193 | EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer); | |
1194 | ||
ec364dc9 MD |
1195 | #if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE |
1196 | static void lib_ring_buffer_flush_read_subbuf_dcache( | |
1197 | const struct lib_ring_buffer_config *config, | |
1198 | struct channel *chan, | |
1199 | struct lib_ring_buffer *buf) | |
1200 | { | |
1201 | struct lib_ring_buffer_backend_pages *pages; | |
1202 | unsigned long sb_bindex, id, i, nr_pages; | |
1203 | ||
1204 | if (config->output != RING_BUFFER_MMAP) | |
1205 | return; | |
1206 | ||
1207 | /* | |
1208 | * Architectures with caches aliased on virtual addresses may | |
1209 | * use different cache lines for the linear mapping vs | |
1210 | * user-space memory mapping. Given that the ring buffer is | |
1211 | * based on the kernel linear mapping, aligning it with the | |
1212 | * user-space mapping is not straightforward, and would require | |
1213 | * extra TLB entries. Therefore, simply flush the dcache for the | |
1214 | * entire sub-buffer before reading it. | |
1215 | */ | |
1216 | id = buf->backend.buf_rsb.id; | |
1217 | sb_bindex = subbuffer_id_get_index(config, id); | |
1218 | pages = buf->backend.array[sb_bindex]; | |
1219 | nr_pages = buf->backend.num_pages_per_subbuf; | |
1220 | for (i = 0; i < nr_pages; i++) { | |
1221 | struct lib_ring_buffer_backend_page *backend_page; | |
1222 | ||
1223 | backend_page = &pages->p[i]; | |
1224 | flush_dcache_page(pfn_to_page(backend_page->pfn)); | |
1225 | } | |
1226 | } | |
1227 | #else | |
1228 | static void lib_ring_buffer_flush_read_subbuf_dcache( | |
1229 | const struct lib_ring_buffer_config *config, | |
1230 | struct channel *chan, | |
1231 | struct lib_ring_buffer *buf) | |
1232 | { | |
1233 | } | |
1234 | #endif | |
1235 | ||
f3bc08c5 MD |
1236 | /** |
1237 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
1238 | * @buf: ring buffer | |
1239 | * @consumed: consumed count indicating the position where to read | |
1240 | * | |
1241 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1242 | * data to read at consumed position, or 0 if the get operation succeeds. | |
1243 | * Busy-loop trying to get data if the tick_nohz sequence lock is held. | |
1244 | */ | |
1245 | int lib_ring_buffer_get_subbuf(struct lib_ring_buffer *buf, | |
1246 | unsigned long consumed) | |
1247 | { | |
1248 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1249 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1250 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
1251 | int ret; | |
1252 | int finalized; | |
1253 | ||
8202b8a0 MD |
1254 | if (buf->get_subbuf) { |
1255 | /* | |
1256 | * Reader is trying to get a subbuffer twice. | |
1257 | */ | |
1258 | CHAN_WARN_ON(chan, 1); | |
1259 | return -EBUSY; | |
1260 | } | |
f3bc08c5 | 1261 | retry: |
acb98454 | 1262 | finalized = READ_ONCE(buf->finalized); |
f3bc08c5 MD |
1263 | /* |
1264 | * Read finalized before counters. | |
1265 | */ | |
1266 | smp_rmb(); | |
1267 | consumed_cur = atomic_long_read(&buf->consumed); | |
1268 | consumed_idx = subbuf_index(consumed, chan); | |
1269 | commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); | |
1270 | /* | |
1271 | * Make sure we read the commit count before reading the buffer | |
1272 | * data and the write offset. Correct consumed offset ordering | |
1273 | * wrt commit count is insured by the use of cmpxchg to update | |
1274 | * the consumed offset. | |
1275 | * smp_call_function_single can fail if the remote CPU is offline, | |
1276 | * this is OK because then there is no wmb to execute there. | |
1277 | * If our thread is executing on the same CPU as the on the buffers | |
1278 | * belongs to, we don't have to synchronize it at all. If we are | |
1279 | * migrated, the scheduler will take care of the memory barriers. | |
1280 | * Normally, smp_call_function_single() should ensure program order when | |
1281 | * executing the remote function, which implies that it surrounds the | |
1282 | * function execution with : | |
1283 | * smp_mb() | |
1284 | * send IPI | |
1285 | * csd_lock_wait | |
1286 | * recv IPI | |
1287 | * smp_mb() | |
1288 | * exec. function | |
1289 | * smp_mb() | |
1290 | * csd unlock | |
1291 | * smp_mb() | |
1292 | * | |
1293 | * However, smp_call_function_single() does not seem to clearly execute | |
1294 | * such barriers. It depends on spinlock semantic to provide the barrier | |
1295 | * before executing the IPI and, when busy-looping, csd_lock_wait only | |
1296 | * executes smp_mb() when it has to wait for the other CPU. | |
1297 | * | |
1298 | * I don't trust this code. Therefore, let's add the smp_mb() sequence | |
1299 | * required ourself, even if duplicated. It has no performance impact | |
1300 | * anyway. | |
1301 | * | |
1302 | * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs | |
1303 | * read and write vs write. They do not ensure core synchronization. We | |
1304 | * really have to ensure total order between the 3 barriers running on | |
1305 | * the 2 CPUs. | |
1306 | */ | |
1307 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1308 | if (config->sync == RING_BUFFER_SYNC_PER_CPU | |
1309 | && config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
1310 | if (raw_smp_processor_id() != buf->backend.cpu) { | |
1311 | /* Total order with IPI handler smp_mb() */ | |
1312 | smp_mb(); | |
1313 | smp_call_function_single(buf->backend.cpu, | |
1314 | remote_mb, NULL, 1); | |
1315 | /* Total order with IPI handler smp_mb() */ | |
1316 | smp_mb(); | |
1317 | } | |
1318 | } else { | |
1319 | /* Total order with IPI handler smp_mb() */ | |
1320 | smp_mb(); | |
1321 | smp_call_function(remote_mb, NULL, 1); | |
1322 | /* Total order with IPI handler smp_mb() */ | |
1323 | smp_mb(); | |
1324 | } | |
1325 | } else { | |
1326 | /* | |
1327 | * Local rmb to match the remote wmb to read the commit count | |
1328 | * before the buffer data and the write offset. | |
1329 | */ | |
1330 | smp_rmb(); | |
1331 | } | |
1332 | ||
1333 | write_offset = v_read(config, &buf->offset); | |
1334 | ||
1335 | /* | |
1336 | * Check that the buffer we are getting is after or at consumed_cur | |
1337 | * position. | |
1338 | */ | |
1339 | if ((long) subbuf_trunc(consumed, chan) | |
1340 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
1341 | goto nodata; | |
1342 | ||
1343 | /* | |
1344 | * Check that the subbuffer we are trying to consume has been | |
1345 | * already fully committed. | |
1346 | */ | |
1347 | if (((commit_count - chan->backend.subbuf_size) | |
1348 | & chan->commit_count_mask) | |
c9b3b5e2 | 1349 | - (buf_trunc(consumed, chan) |
f3bc08c5 MD |
1350 | >> chan->backend.num_subbuf_order) |
1351 | != 0) | |
1352 | goto nodata; | |
1353 | ||
1354 | /* | |
1355 | * Check that we are not about to read the same subbuffer in | |
1356 | * which the writer head is. | |
1357 | */ | |
c9b3b5e2 | 1358 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed, chan) |
f3bc08c5 MD |
1359 | == 0) |
1360 | goto nodata; | |
1361 | ||
1362 | /* | |
1363 | * Failure to get the subbuffer causes a busy-loop retry without going | |
1364 | * to a wait queue. These are caused by short-lived race windows where | |
1365 | * the writer is getting access to a subbuffer we were trying to get | |
1366 | * access to. Also checks that the "consumed" buffer count we are | |
1367 | * looking for matches the one contained in the subbuffer id. | |
1368 | */ | |
1369 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1370 | consumed_idx, buf_trunc_val(consumed, chan)); | |
1371 | if (ret) | |
1372 | goto retry; | |
1373 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
1374 | ||
1375 | buf->get_subbuf_consumed = consumed; | |
1376 | buf->get_subbuf = 1; | |
1377 | ||
ec364dc9 MD |
1378 | lib_ring_buffer_flush_read_subbuf_dcache(config, chan, buf); |
1379 | ||
f3bc08c5 MD |
1380 | return 0; |
1381 | ||
1382 | nodata: | |
1383 | /* | |
1384 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1385 | * of "raw_spin_is_locked" memory ordering. | |
1386 | */ | |
1387 | if (finalized) | |
1388 | return -ENODATA; | |
1389 | else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock)) | |
1390 | goto retry; | |
1391 | else | |
1392 | return -EAGAIN; | |
1393 | } | |
1394 | EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf); | |
1395 | ||
1396 | /** | |
1397 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
1398 | * @buf: ring buffer | |
1399 | */ | |
1400 | void lib_ring_buffer_put_subbuf(struct lib_ring_buffer *buf) | |
1401 | { | |
1402 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1403 | struct channel *chan = bufb->chan; | |
5a8fd222 | 1404 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1405 | unsigned long read_sb_bindex, consumed_idx, consumed; |
1406 | ||
1407 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
1408 | ||
1409 | if (!buf->get_subbuf) { | |
1410 | /* | |
1411 | * Reader puts a subbuffer it did not get. | |
1412 | */ | |
1413 | CHAN_WARN_ON(chan, 1); | |
1414 | return; | |
1415 | } | |
1416 | consumed = buf->get_subbuf_consumed; | |
1417 | buf->get_subbuf = 0; | |
1418 | ||
1419 | /* | |
1420 | * Clear the records_unread counter. (overruns counter) | |
1421 | * Can still be non-zero if a file reader simply grabbed the data | |
1422 | * without using iterators. | |
1423 | * Can be below zero if an iterator is used on a snapshot more than | |
1424 | * once. | |
1425 | */ | |
1426 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
1427 | v_add(config, v_read(config, | |
1428 | &bufb->array[read_sb_bindex]->records_unread), | |
1429 | &bufb->records_read); | |
1430 | v_set(config, &bufb->array[read_sb_bindex]->records_unread, 0); | |
1431 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE | |
1432 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
1433 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
1434 | ||
1435 | /* | |
1436 | * Exchange the reader subbuffer with the one we put in its place in the | |
1437 | * writer subbuffer table. Expect the original consumed count. If | |
1438 | * update_read_sb_index fails, this is because the writer updated the | |
1439 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
1440 | * currently have: it has become invalid to try reading this sub-buffer | |
1441 | * consumed count value anyway. | |
1442 | */ | |
1443 | consumed_idx = subbuf_index(consumed, chan); | |
1444 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1445 | consumed_idx, buf_trunc_val(consumed, chan)); | |
1446 | /* | |
1447 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
1448 | * if the writer concurrently updated it. | |
1449 | */ | |
1450 | } | |
1451 | EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf); | |
1452 | ||
1453 | /* | |
1454 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
1455 | * position and the writer position. (inclusive) | |
1456 | */ | |
1457 | static | |
1458 | void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer *buf, | |
1459 | struct channel *chan, | |
1460 | unsigned long cons_offset, | |
1461 | int cpu) | |
1462 | { | |
5a8fd222 | 1463 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1464 | unsigned long cons_idx, commit_count, commit_count_sb; |
1465 | ||
1466 | cons_idx = subbuf_index(cons_offset, chan); | |
1467 | commit_count = v_read(config, &buf->commit_hot[cons_idx].cc); | |
1468 | commit_count_sb = v_read(config, &buf->commit_cold[cons_idx].cc_sb); | |
1469 | ||
1470 | if (subbuf_offset(commit_count, chan) != 0) | |
1471 | printk(KERN_WARNING | |
1472 | "ring buffer %s, cpu %d: " | |
1473 | "commit count in subbuffer %lu,\n" | |
1474 | "expecting multiples of %lu bytes\n" | |
1475 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
1476 | chan->backend.name, cpu, cons_idx, | |
1477 | chan->backend.subbuf_size, | |
1478 | commit_count, commit_count_sb); | |
1479 | ||
1480 | printk(KERN_DEBUG "ring buffer: %s, cpu %d: %lu bytes committed\n", | |
1481 | chan->backend.name, cpu, commit_count); | |
1482 | } | |
1483 | ||
1484 | static | |
1485 | void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf, | |
1486 | struct channel *chan, | |
1487 | void *priv, int cpu) | |
1488 | { | |
5a8fd222 | 1489 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1490 | unsigned long write_offset, cons_offset; |
1491 | ||
f3bc08c5 MD |
1492 | /* |
1493 | * No need to order commit_count, write_offset and cons_offset reads | |
1494 | * because we execute at teardown when no more writer nor reader | |
1495 | * references are left. | |
1496 | */ | |
1497 | write_offset = v_read(config, &buf->offset); | |
1498 | cons_offset = atomic_long_read(&buf->consumed); | |
1499 | if (write_offset != cons_offset) | |
05aad775 | 1500 | printk(KERN_DEBUG |
f3bc08c5 MD |
1501 | "ring buffer %s, cpu %d: " |
1502 | "non-consumed data\n" | |
1503 | " [ %lu bytes written, %lu bytes read ]\n", | |
1504 | chan->backend.name, cpu, write_offset, cons_offset); | |
1505 | ||
1506 | for (cons_offset = atomic_long_read(&buf->consumed); | |
1507 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), | |
1508 | chan) | |
1509 | - cons_offset) > 0; | |
1510 | cons_offset = subbuf_align(cons_offset, chan)) | |
1511 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1512 | cpu); | |
1513 | } | |
1514 | ||
12e1cd7b | 1515 | #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS |
f3bc08c5 | 1516 | static |
12e1cd7b JG |
1517 | void lib_ring_buffer_print_records_count(struct channel *chan, |
1518 | struct lib_ring_buffer *buf, | |
1519 | int cpu) | |
f3bc08c5 | 1520 | { |
5a8fd222 | 1521 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 | 1522 | |
ec01ec93 MD |
1523 | if (!strcmp(chan->backend.name, "relay-metadata")) { |
1524 | printk(KERN_DEBUG "ring buffer %s: %lu records written, " | |
1525 | "%lu records overrun\n", | |
1526 | chan->backend.name, | |
1527 | v_read(config, &buf->records_count), | |
1528 | v_read(config, &buf->records_overrun)); | |
1529 | } else { | |
1530 | printk(KERN_DEBUG "ring buffer %s, cpu %d: %lu records written, " | |
1531 | "%lu records overrun\n", | |
1532 | chan->backend.name, cpu, | |
1533 | v_read(config, &buf->records_count), | |
1534 | v_read(config, &buf->records_overrun)); | |
12e1cd7b JG |
1535 | } |
1536 | } | |
1537 | #else | |
1538 | static | |
1539 | void lib_ring_buffer_print_records_count(struct channel *chan, | |
1540 | struct lib_ring_buffer *buf, | |
1541 | int cpu) | |
1542 | { | |
1543 | } | |
1544 | #endif | |
1545 | ||
1546 | static | |
1547 | void lib_ring_buffer_print_errors(struct channel *chan, | |
1548 | struct lib_ring_buffer *buf, int cpu) | |
1549 | { | |
1550 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1551 | void *priv = chan->backend.priv; | |
ec01ec93 | 1552 | |
12e1cd7b JG |
1553 | lib_ring_buffer_print_records_count(chan, buf, cpu); |
1554 | if (strcmp(chan->backend.name, "relay-metadata")) { | |
ec01ec93 MD |
1555 | if (v_read(config, &buf->records_lost_full) |
1556 | || v_read(config, &buf->records_lost_wrap) | |
1557 | || v_read(config, &buf->records_lost_big)) | |
1558 | printk(KERN_WARNING | |
1559 | "ring buffer %s, cpu %d: records were lost. Caused by:\n" | |
1560 | " [ %lu buffer full, %lu nest buffer wrap-around, " | |
1561 | "%lu event too big ]\n", | |
1562 | chan->backend.name, cpu, | |
1563 | v_read(config, &buf->records_lost_full), | |
1564 | v_read(config, &buf->records_lost_wrap), | |
1565 | v_read(config, &buf->records_lost_big)); | |
1566 | } | |
f3bc08c5 MD |
1567 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu); |
1568 | } | |
1569 | ||
1570 | /* | |
1571 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1572 | * | |
c11ab967 | 1573 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. |
f3bc08c5 MD |
1574 | */ |
1575 | static | |
1576 | void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf, | |
1577 | struct channel *chan, | |
1578 | struct switch_offsets *offsets, | |
1579 | u64 tsc) | |
1580 | { | |
5a8fd222 | 1581 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1582 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1583 | unsigned long commit_count; | |
fbe84fd7 | 1584 | struct commit_counters_hot *cc_hot; |
f3bc08c5 MD |
1585 | |
1586 | config->cb.buffer_begin(buf, tsc, oldidx); | |
1587 | ||
1588 | /* | |
1589 | * Order all writes to buffer before the commit count update that will | |
1590 | * determine that the subbuffer is full. | |
1591 | */ | |
1592 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1593 | /* | |
1594 | * Must write slot data before incrementing commit count. This | |
1595 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1596 | * by get_subbuf(). | |
1597 | */ | |
1598 | barrier(); | |
1599 | } else | |
1600 | smp_wmb(); | |
fbe84fd7 MD |
1601 | cc_hot = &buf->commit_hot[oldidx]; |
1602 | v_add(config, config->cb.subbuffer_header_size(), &cc_hot->cc); | |
1603 | commit_count = v_read(config, &cc_hot->cc); | |
f3bc08c5 MD |
1604 | /* Check if the written buffer has to be delivered */ |
1605 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
635e457c | 1606 | commit_count, oldidx, tsc); |
8ec496cf | 1607 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
7915e163 | 1608 | offsets->old + config->cb.subbuffer_header_size(), |
fbe84fd7 | 1609 | commit_count, cc_hot); |
f3bc08c5 MD |
1610 | } |
1611 | ||
1612 | /* | |
1613 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1614 | * | |
1615 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1616 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1617 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1618 | * subbuffer. | |
1619 | */ | |
1620 | static | |
1621 | void lib_ring_buffer_switch_old_end(struct lib_ring_buffer *buf, | |
1622 | struct channel *chan, | |
1623 | struct switch_offsets *offsets, | |
1624 | u64 tsc) | |
1625 | { | |
5a8fd222 | 1626 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1627 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1628 | unsigned long commit_count, padding_size, data_size; | |
fbe84fd7 | 1629 | struct commit_counters_hot *cc_hot; |
a1817dcb | 1630 | u64 *ts_end; |
f3bc08c5 MD |
1631 | |
1632 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1633 | padding_size = chan->backend.subbuf_size - data_size; | |
1634 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size); | |
1635 | ||
a1817dcb | 1636 | ts_end = &buf->ts_end[oldidx]; |
f3bc08c5 | 1637 | /* |
a1817dcb MD |
1638 | * This is the last space reservation in that sub-buffer before |
1639 | * it gets delivered. This provides exclusive access to write to | |
1640 | * this sub-buffer's ts_end. There are also no concurrent | |
1641 | * readers of that ts_end because delivery of that sub-buffer is | |
1642 | * postponed until the commit counter is incremented for the | |
1643 | * current space reservation. | |
1644 | */ | |
1645 | *ts_end = tsc; | |
1646 | ||
1647 | /* | |
1648 | * Order all writes to buffer and store to ts_end before the commit | |
1649 | * count update that will determine that the subbuffer is full. | |
f3bc08c5 MD |
1650 | */ |
1651 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1652 | /* | |
1653 | * Must write slot data before incrementing commit count. This | |
1654 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1655 | * by get_subbuf(). | |
1656 | */ | |
1657 | barrier(); | |
1658 | } else | |
1659 | smp_wmb(); | |
fbe84fd7 MD |
1660 | cc_hot = &buf->commit_hot[oldidx]; |
1661 | v_add(config, padding_size, &cc_hot->cc); | |
1662 | commit_count = v_read(config, &cc_hot->cc); | |
f3bc08c5 | 1663 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
635e457c | 1664 | commit_count, oldidx, tsc); |
8ec496cf MD |
1665 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
1666 | offsets->old + padding_size, commit_count, | |
fbe84fd7 | 1667 | cc_hot); |
f3bc08c5 MD |
1668 | } |
1669 | ||
1670 | /* | |
1671 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1672 | * | |
1673 | * This code can be executed unordered : writers may already have written to the | |
1674 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1675 | * that this code is executed before the deliver of this sub-buffer. | |
1676 | */ | |
1677 | static | |
1678 | void lib_ring_buffer_switch_new_start(struct lib_ring_buffer *buf, | |
1679 | struct channel *chan, | |
1680 | struct switch_offsets *offsets, | |
1681 | u64 tsc) | |
1682 | { | |
5a8fd222 | 1683 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1684 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1685 | unsigned long commit_count; | |
fbe84fd7 | 1686 | struct commit_counters_hot *cc_hot; |
f3bc08c5 MD |
1687 | |
1688 | config->cb.buffer_begin(buf, tsc, beginidx); | |
1689 | ||
1690 | /* | |
1691 | * Order all writes to buffer before the commit count update that will | |
1692 | * determine that the subbuffer is full. | |
1693 | */ | |
1694 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1695 | /* | |
1696 | * Must write slot data before incrementing commit count. This | |
1697 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1698 | * by get_subbuf(). | |
1699 | */ | |
1700 | barrier(); | |
1701 | } else | |
1702 | smp_wmb(); | |
fbe84fd7 MD |
1703 | cc_hot = &buf->commit_hot[beginidx]; |
1704 | v_add(config, config->cb.subbuffer_header_size(), &cc_hot->cc); | |
1705 | commit_count = v_read(config, &cc_hot->cc); | |
f3bc08c5 MD |
1706 | /* Check if the written buffer has to be delivered */ |
1707 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
635e457c | 1708 | commit_count, beginidx, tsc); |
8ec496cf | 1709 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
7915e163 | 1710 | offsets->begin + config->cb.subbuffer_header_size(), |
fbe84fd7 | 1711 | commit_count, cc_hot); |
f3bc08c5 MD |
1712 | } |
1713 | ||
f5ea5800 MD |
1714 | /* |
1715 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1716 | * | |
768b05c9 MD |
1717 | * Calls subbuffer_set_data_size() to set the data size of the current |
1718 | * sub-buffer. We do not need to perform check_deliver nor commit here, | |
1719 | * since this task will be done by the "commit" of the event for which | |
1720 | * we are currently doing the space reservation. | |
f5ea5800 MD |
1721 | */ |
1722 | static | |
1723 | void lib_ring_buffer_switch_new_end(struct lib_ring_buffer *buf, | |
1724 | struct channel *chan, | |
1725 | struct switch_offsets *offsets, | |
1726 | u64 tsc) | |
1727 | { | |
1728 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
768b05c9 | 1729 | unsigned long endidx, data_size; |
a1817dcb | 1730 | u64 *ts_end; |
f5ea5800 | 1731 | |
768b05c9 | 1732 | endidx = subbuf_index(offsets->end - 1, chan); |
f5ea5800 | 1733 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; |
f5ea5800 | 1734 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size); |
a1817dcb MD |
1735 | ts_end = &buf->ts_end[endidx]; |
1736 | /* | |
1737 | * This is the last space reservation in that sub-buffer before | |
1738 | * it gets delivered. This provides exclusive access to write to | |
1739 | * this sub-buffer's ts_end. There are also no concurrent | |
1740 | * readers of that ts_end because delivery of that sub-buffer is | |
1741 | * postponed until the commit counter is incremented for the | |
1742 | * current space reservation. | |
1743 | */ | |
1744 | *ts_end = tsc; | |
f5ea5800 MD |
1745 | } |
1746 | ||
f3bc08c5 MD |
1747 | /* |
1748 | * Returns : | |
1749 | * 0 if ok | |
1750 | * !0 if execution must be aborted. | |
1751 | */ | |
1752 | static | |
1753 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
1754 | struct lib_ring_buffer *buf, | |
1755 | struct channel *chan, | |
1756 | struct switch_offsets *offsets, | |
1757 | u64 *tsc) | |
1758 | { | |
5a8fd222 | 1759 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
5334a2c5 | 1760 | unsigned long off, reserve_commit_diff; |
f3bc08c5 MD |
1761 | |
1762 | offsets->begin = v_read(config, &buf->offset); | |
1763 | offsets->old = offsets->begin; | |
1764 | offsets->switch_old_start = 0; | |
1765 | off = subbuf_offset(offsets->begin, chan); | |
1766 | ||
1767 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1768 | ||
1769 | /* | |
1770 | * Ensure we flush the header of an empty subbuffer when doing the | |
1771 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1772 | * total data gathering duration even if there were no records saved | |
1773 | * after the last buffer switch. | |
1774 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1775 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1776 | * subbuffer header as appropriate. | |
1777 | * The next record that reserves space will be responsible for | |
1778 | * populating the following subbuffer header. We choose not to populate | |
1779 | * the next subbuffer header here because we want to be able to use | |
1780 | * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop | |
1781 | * buffer flush, which must guarantee that all the buffer content | |
1782 | * (records and header timestamps) are visible to the reader. This is | |
1783 | * required for quiescence guarantees for the fusion merge. | |
1784 | */ | |
5334a2c5 MD |
1785 | if (mode != SWITCH_FLUSH && !off) |
1786 | return -1; /* we do not have to switch : buffer is empty */ | |
1787 | ||
1788 | if (unlikely(off == 0)) { | |
1789 | unsigned long sb_index, commit_count; | |
1790 | ||
1791 | /* | |
c11ab967 MD |
1792 | * We are performing a SWITCH_FLUSH. At this stage, there are no |
1793 | * concurrent writes into the buffer. | |
5334a2c5 | 1794 | * |
c11ab967 MD |
1795 | * The client does not save any header information. Don't |
1796 | * switch empty subbuffer on finalize, because it is invalid to | |
1797 | * deliver a completely empty subbuffer. | |
5334a2c5 MD |
1798 | */ |
1799 | if (!config->cb.subbuffer_header_size()) | |
1800 | return -1; | |
1801 | ||
1802 | /* Test new buffer integrity */ | |
1803 | sb_index = subbuf_index(offsets->begin, chan); | |
1804 | commit_count = v_read(config, | |
1805 | &buf->commit_cold[sb_index].cc_sb); | |
1806 | reserve_commit_diff = | |
1807 | (buf_trunc(offsets->begin, chan) | |
1808 | >> chan->backend.num_subbuf_order) | |
1809 | - (commit_count & chan->commit_count_mask); | |
1810 | if (likely(reserve_commit_diff == 0)) { | |
1811 | /* Next subbuffer not being written to. */ | |
1812 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1813 | subbuf_trunc(offsets->begin, chan) | |
1814 | - subbuf_trunc((unsigned long) | |
1815 | atomic_long_read(&buf->consumed), chan) | |
1816 | >= chan->backend.buf_size)) { | |
1817 | /* | |
1818 | * We do not overwrite non consumed buffers | |
1819 | * and we are full : don't switch. | |
1820 | */ | |
f3bc08c5 | 1821 | return -1; |
5334a2c5 MD |
1822 | } else { |
1823 | /* | |
1824 | * Next subbuffer not being written to, and we | |
1825 | * are either in overwrite mode or the buffer is | |
1826 | * not full. It's safe to write in this new | |
1827 | * subbuffer. | |
1828 | */ | |
1829 | } | |
1830 | } else { | |
f3bc08c5 | 1831 | /* |
5334a2c5 MD |
1832 | * Next subbuffer reserve offset does not match the |
1833 | * commit offset. Don't perform switch in | |
1834 | * producer-consumer and overwrite mode. Caused by | |
1835 | * either a writer OOPS or too many nested writes over a | |
1836 | * reserve/commit pair. | |
f3bc08c5 | 1837 | */ |
5334a2c5 | 1838 | return -1; |
f3bc08c5 | 1839 | } |
5334a2c5 MD |
1840 | |
1841 | /* | |
1842 | * Need to write the subbuffer start header on finalize. | |
1843 | */ | |
1844 | offsets->switch_old_start = 1; | |
1845 | } | |
1846 | offsets->begin = subbuf_align(offsets->begin, chan); | |
f3bc08c5 MD |
1847 | /* Note: old points to the next subbuf at offset 0 */ |
1848 | offsets->end = offsets->begin; | |
1849 | return 0; | |
1850 | } | |
1851 | ||
1852 | /* | |
1853 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1854 | * called while tracing is active with absolutely no lock held. | |
1855 | * | |
1856 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1857 | * operations, this function must be called from the CPU which owns the buffer | |
1858 | * for a ACTIVE flush. | |
1859 | */ | |
1860 | void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode) | |
1861 | { | |
1862 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1863 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1864 | struct switch_offsets offsets; |
1865 | unsigned long oldidx; | |
1866 | u64 tsc; | |
1867 | ||
1868 | offsets.size = 0; | |
1869 | ||
1870 | /* | |
1871 | * Perform retryable operations. | |
1872 | */ | |
1873 | do { | |
1874 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1875 | &tsc)) | |
1876 | return; /* Switch not needed */ | |
1877 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1878 | != offsets.old); | |
1879 | ||
1880 | /* | |
1881 | * Atomically update last_tsc. This update races against concurrent | |
1882 | * atomic updates, but the race will always cause supplementary full TSC | |
1883 | * records, never the opposite (missing a full TSC record when it would | |
1884 | * be needed). | |
1885 | */ | |
1886 | save_last_tsc(config, buf, tsc); | |
1887 | ||
1888 | /* | |
1889 | * Push the reader if necessary | |
1890 | */ | |
1891 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1892 | ||
1893 | oldidx = subbuf_index(offsets.old, chan); | |
1894 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx); | |
1895 | ||
1896 | /* | |
1897 | * May need to populate header start on SWITCH_FLUSH. | |
1898 | */ | |
1899 | if (offsets.switch_old_start) { | |
1900 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc); | |
1901 | offsets.old += config->cb.subbuffer_header_size(); | |
1902 | } | |
1903 | ||
1904 | /* | |
1905 | * Switch old subbuffer. | |
1906 | */ | |
1907 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc); | |
1908 | } | |
1909 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow); | |
1910 | ||
9dded431 MD |
1911 | struct switch_param { |
1912 | struct lib_ring_buffer *buf; | |
1913 | enum switch_mode mode; | |
1914 | }; | |
1915 | ||
5e391252 MD |
1916 | static void remote_switch(void *info) |
1917 | { | |
9dded431 MD |
1918 | struct switch_param *param = info; |
1919 | struct lib_ring_buffer *buf = param->buf; | |
5e391252 | 1920 | |
9dded431 | 1921 | lib_ring_buffer_switch_slow(buf, param->mode); |
5e391252 MD |
1922 | } |
1923 | ||
64af2437 MD |
1924 | static void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf, |
1925 | enum switch_mode mode) | |
5e391252 MD |
1926 | { |
1927 | struct channel *chan = buf->backend.chan; | |
1928 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1929 | int ret; | |
9dded431 | 1930 | struct switch_param param; |
5e391252 MD |
1931 | |
1932 | /* | |
1933 | * With global synchronization we don't need to use the IPI scheme. | |
1934 | */ | |
1935 | if (config->sync == RING_BUFFER_SYNC_GLOBAL) { | |
64af2437 | 1936 | lib_ring_buffer_switch_slow(buf, mode); |
5e391252 MD |
1937 | return; |
1938 | } | |
1939 | ||
1940 | /* | |
8e3b6653 | 1941 | * Disabling preemption ensures two things: first, that the |
5e391252 | 1942 | * target cpu is not taken concurrently offline while we are within |
8e3b6653 MD |
1943 | * smp_call_function_single(). Secondly, if it happens that the |
1944 | * CPU is not online, our own call to lib_ring_buffer_switch_slow() | |
1945 | * needs to be protected from CPU hotplug handlers, which can | |
1946 | * also perform a remote subbuffer switch. | |
5e391252 | 1947 | */ |
8e3b6653 | 1948 | preempt_disable(); |
9dded431 MD |
1949 | param.buf = buf; |
1950 | param.mode = mode; | |
5e391252 | 1951 | ret = smp_call_function_single(buf->backend.cpu, |
9dded431 | 1952 | remote_switch, ¶m, 1); |
5e391252 MD |
1953 | if (ret) { |
1954 | /* Remote CPU is offline, do it ourself. */ | |
64af2437 | 1955 | lib_ring_buffer_switch_slow(buf, mode); |
5e391252 | 1956 | } |
8e3b6653 | 1957 | preempt_enable(); |
5e391252 | 1958 | } |
64af2437 | 1959 | |
e1cdd6e9 | 1960 | /* Switch sub-buffer if current sub-buffer is non-empty. */ |
64af2437 MD |
1961 | void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf) |
1962 | { | |
1963 | _lib_ring_buffer_switch_remote(buf, SWITCH_ACTIVE); | |
1964 | } | |
5e391252 MD |
1965 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote); |
1966 | ||
e1cdd6e9 MD |
1967 | /* Switch sub-buffer even if current sub-buffer is empty. */ |
1968 | void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer *buf) | |
1969 | { | |
1970 | _lib_ring_buffer_switch_remote(buf, SWITCH_FLUSH); | |
1971 | } | |
1972 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote_empty); | |
1973 | ||
f3bc08c5 MD |
1974 | /* |
1975 | * Returns : | |
1976 | * 0 if ok | |
97ca2c54 MD |
1977 | * -ENOSPC if event size is too large for packet. |
1978 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1979 | * -EIO if data cannot be written into the buffer for any other reason. | |
f3bc08c5 MD |
1980 | */ |
1981 | static | |
1982 | int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, | |
1983 | struct channel *chan, | |
1984 | struct switch_offsets *offsets, | |
1985 | struct lib_ring_buffer_ctx *ctx) | |
1986 | { | |
5a8fd222 | 1987 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
0fdec686 | 1988 | unsigned long reserve_commit_diff, offset_cmp; |
f3bc08c5 | 1989 | |
0fdec686 MD |
1990 | retry: |
1991 | offsets->begin = offset_cmp = v_read(config, &buf->offset); | |
f3bc08c5 MD |
1992 | offsets->old = offsets->begin; |
1993 | offsets->switch_new_start = 0; | |
f5ea5800 | 1994 | offsets->switch_new_end = 0; |
f3bc08c5 MD |
1995 | offsets->switch_old_end = 0; |
1996 | offsets->pre_header_padding = 0; | |
1997 | ||
1998 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
97ca2c54 MD |
1999 | if ((int64_t) ctx->tsc == -EIO) |
2000 | return -EIO; | |
f3bc08c5 MD |
2001 | |
2002 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
64c796d8 | 2003 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; |
f3bc08c5 MD |
2004 | |
2005 | if (unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { | |
2006 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
2007 | } else { | |
2008 | offsets->size = config->cb.record_header_size(config, chan, | |
2009 | offsets->begin, | |
f3bc08c5 | 2010 | &offsets->pre_header_padding, |
64c796d8 | 2011 | ctx); |
f3bc08c5 MD |
2012 | offsets->size += |
2013 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
2014 | ctx->largest_align) | |
2015 | + ctx->data_size; | |
2016 | if (unlikely(subbuf_offset(offsets->begin, chan) + | |
2017 | offsets->size > chan->backend.subbuf_size)) { | |
2018 | offsets->switch_old_end = 1; /* For offsets->old */ | |
2019 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
2020 | } | |
2021 | } | |
2022 | if (unlikely(offsets->switch_new_start)) { | |
0fdec686 | 2023 | unsigned long sb_index, commit_count; |
f3bc08c5 MD |
2024 | |
2025 | /* | |
2026 | * We are typically not filling the previous buffer completely. | |
2027 | */ | |
2028 | if (likely(offsets->switch_old_end)) | |
2029 | offsets->begin = subbuf_align(offsets->begin, chan); | |
2030 | offsets->begin = offsets->begin | |
2031 | + config->cb.subbuffer_header_size(); | |
2032 | /* Test new buffer integrity */ | |
2033 | sb_index = subbuf_index(offsets->begin, chan); | |
0fdec686 MD |
2034 | /* |
2035 | * Read buf->offset before buf->commit_cold[sb_index].cc_sb. | |
2036 | * lib_ring_buffer_check_deliver() has the matching | |
2037 | * memory barriers required around commit_cold cc_sb | |
2038 | * updates to ensure reserve and commit counter updates | |
2039 | * are not seen reordered when updated by another CPU. | |
2040 | */ | |
2041 | smp_rmb(); | |
2042 | commit_count = v_read(config, | |
2043 | &buf->commit_cold[sb_index].cc_sb); | |
2044 | /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */ | |
2045 | smp_rmb(); | |
2046 | if (unlikely(offset_cmp != v_read(config, &buf->offset))) { | |
2047 | /* | |
2048 | * The reserve counter have been concurrently updated | |
2049 | * while we read the commit counter. This means the | |
2050 | * commit counter we read might not match buf->offset | |
2051 | * due to concurrent update. We therefore need to retry. | |
2052 | */ | |
2053 | goto retry; | |
2054 | } | |
f3bc08c5 MD |
2055 | reserve_commit_diff = |
2056 | (buf_trunc(offsets->begin, chan) | |
2057 | >> chan->backend.num_subbuf_order) | |
0fdec686 | 2058 | - (commit_count & chan->commit_count_mask); |
f3bc08c5 MD |
2059 | if (likely(reserve_commit_diff == 0)) { |
2060 | /* Next subbuffer not being written to. */ | |
2061 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
2062 | subbuf_trunc(offsets->begin, chan) | |
2063 | - subbuf_trunc((unsigned long) | |
2064 | atomic_long_read(&buf->consumed), chan) | |
2065 | >= chan->backend.buf_size)) { | |
2066 | /* | |
2067 | * We do not overwrite non consumed buffers | |
2068 | * and we are full : record is lost. | |
2069 | */ | |
2070 | v_inc(config, &buf->records_lost_full); | |
97ca2c54 | 2071 | return -ENOBUFS; |
f3bc08c5 MD |
2072 | } else { |
2073 | /* | |
2074 | * Next subbuffer not being written to, and we | |
2075 | * are either in overwrite mode or the buffer is | |
2076 | * not full. It's safe to write in this new | |
2077 | * subbuffer. | |
2078 | */ | |
2079 | } | |
2080 | } else { | |
2081 | /* | |
2082 | * Next subbuffer reserve offset does not match the | |
0fdec686 MD |
2083 | * commit offset, and this did not involve update to the |
2084 | * reserve counter. Drop record in producer-consumer and | |
2085 | * overwrite mode. Caused by either a writer OOPS or | |
2086 | * too many nested writes over a reserve/commit pair. | |
f3bc08c5 MD |
2087 | */ |
2088 | v_inc(config, &buf->records_lost_wrap); | |
97ca2c54 | 2089 | return -EIO; |
f3bc08c5 MD |
2090 | } |
2091 | offsets->size = | |
2092 | config->cb.record_header_size(config, chan, | |
2093 | offsets->begin, | |
f3bc08c5 | 2094 | &offsets->pre_header_padding, |
64c796d8 | 2095 | ctx); |
f3bc08c5 MD |
2096 | offsets->size += |
2097 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
2098 | ctx->largest_align) | |
2099 | + ctx->data_size; | |
2100 | if (unlikely(subbuf_offset(offsets->begin, chan) | |
2101 | + offsets->size > chan->backend.subbuf_size)) { | |
2102 | /* | |
2103 | * Record too big for subbuffers, report error, don't | |
2104 | * complete the sub-buffer switch. | |
2105 | */ | |
2106 | v_inc(config, &buf->records_lost_big); | |
97ca2c54 | 2107 | return -ENOSPC; |
f3bc08c5 MD |
2108 | } else { |
2109 | /* | |
2110 | * We just made a successful buffer switch and the | |
2111 | * record fits in the new subbuffer. Let's write. | |
2112 | */ | |
2113 | } | |
2114 | } else { | |
2115 | /* | |
2116 | * Record fits in the current buffer and we are not on a switch | |
2117 | * boundary. It's safe to write. | |
2118 | */ | |
2119 | } | |
2120 | offsets->end = offsets->begin + offsets->size; | |
f5ea5800 MD |
2121 | |
2122 | if (unlikely(subbuf_offset(offsets->end, chan) == 0)) { | |
2123 | /* | |
2124 | * The offset_end will fall at the very beginning of the next | |
2125 | * subbuffer. | |
2126 | */ | |
2127 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
2128 | } | |
f3bc08c5 MD |
2129 | return 0; |
2130 | } | |
2131 | ||
d7e74017 MD |
2132 | static struct lib_ring_buffer *get_current_buf(struct channel *chan, int cpu) |
2133 | { | |
2134 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
2135 | ||
2136 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
2137 | return per_cpu_ptr(chan->backend.buf, cpu); | |
2138 | else | |
2139 | return chan->backend.buf; | |
2140 | } | |
2141 | ||
2142 | void lib_ring_buffer_lost_event_too_big(struct channel *chan) | |
2143 | { | |
2144 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
2145 | struct lib_ring_buffer *buf = get_current_buf(chan, smp_processor_id()); | |
2146 | ||
2147 | v_inc(config, &buf->records_lost_big); | |
2148 | } | |
2149 | EXPORT_SYMBOL_GPL(lib_ring_buffer_lost_event_too_big); | |
2150 | ||
f3bc08c5 MD |
2151 | /** |
2152 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
2153 | * @ctx: ring buffer context. | |
2154 | * | |
97ca2c54 MD |
2155 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, |
2156 | * -EIO for other errors, else returns 0. | |
f3bc08c5 MD |
2157 | * It will take care of sub-buffer switching. |
2158 | */ | |
2159 | int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx) | |
2160 | { | |
2161 | struct channel *chan = ctx->chan; | |
5a8fd222 | 2162 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
2163 | struct lib_ring_buffer *buf; |
2164 | struct switch_offsets offsets; | |
c099397a | 2165 | int ret; |
f3bc08c5 | 2166 | |
d7e74017 | 2167 | ctx->buf = buf = get_current_buf(chan, ctx->cpu); |
f3bc08c5 MD |
2168 | offsets.size = 0; |
2169 | ||
2170 | do { | |
97ca2c54 MD |
2171 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, |
2172 | ctx); | |
2173 | if (unlikely(ret)) | |
2174 | return ret; | |
f3bc08c5 MD |
2175 | } while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
2176 | offsets.end) | |
2177 | != offsets.old)); | |
2178 | ||
2179 | /* | |
2180 | * Atomically update last_tsc. This update races against concurrent | |
2181 | * atomic updates, but the race will always cause supplementary full TSC | |
2182 | * records, never the opposite (missing a full TSC record when it would | |
2183 | * be needed). | |
2184 | */ | |
2185 | save_last_tsc(config, buf, ctx->tsc); | |
2186 | ||
2187 | /* | |
2188 | * Push the reader if necessary | |
2189 | */ | |
2190 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
2191 | ||
2192 | /* | |
2193 | * Clear noref flag for this subbuffer. | |
2194 | */ | |
2195 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
2196 | subbuf_index(offsets.end - 1, chan)); | |
2197 | ||
2198 | /* | |
2199 | * Switch old subbuffer if needed. | |
2200 | */ | |
2201 | if (unlikely(offsets.switch_old_end)) { | |
2202 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
2203 | subbuf_index(offsets.old - 1, chan)); | |
2204 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc); | |
2205 | } | |
2206 | ||
2207 | /* | |
2208 | * Populate new subbuffer. | |
2209 | */ | |
2210 | if (unlikely(offsets.switch_new_start)) | |
2211 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc); | |
2212 | ||
f5ea5800 MD |
2213 | if (unlikely(offsets.switch_new_end)) |
2214 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc); | |
2215 | ||
f3bc08c5 MD |
2216 | ctx->slot_size = offsets.size; |
2217 | ctx->pre_offset = offsets.begin; | |
2218 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
2219 | return 0; | |
2220 | } | |
2221 | EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow); | |
6fb8de4b | 2222 | |
aece661f MD |
2223 | static |
2224 | void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config, | |
2225 | struct lib_ring_buffer *buf, | |
2226 | unsigned long commit_count, | |
2227 | unsigned long idx) | |
2228 | { | |
2229 | if (config->oops == RING_BUFFER_OOPS_CONSISTENCY) | |
2230 | v_set(config, &buf->commit_hot[idx].seq, commit_count); | |
2231 | } | |
2232 | ||
25337cb5 MD |
2233 | /* |
2234 | * The ring buffer can count events recorded and overwritten per buffer, | |
2235 | * but it is disabled by default due to its performance overhead. | |
2236 | */ | |
2237 | #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS | |
2238 | static | |
2239 | void deliver_count_events(const struct lib_ring_buffer_config *config, | |
2240 | struct lib_ring_buffer *buf, | |
2241 | unsigned long idx) | |
2242 | { | |
2243 | v_add(config, subbuffer_get_records_count(config, | |
2244 | &buf->backend, idx), | |
2245 | &buf->records_count); | |
2246 | v_add(config, subbuffer_count_records_overrun(config, | |
2247 | &buf->backend, idx), | |
2248 | &buf->records_overrun); | |
2249 | } | |
2250 | #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */ | |
2251 | static | |
2252 | void deliver_count_events(const struct lib_ring_buffer_config *config, | |
2253 | struct lib_ring_buffer *buf, | |
2254 | unsigned long idx) | |
2255 | { | |
2256 | } | |
2257 | #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */ | |
2258 | ||
2259 | ||
aece661f MD |
2260 | void lib_ring_buffer_check_deliver_slow(const struct lib_ring_buffer_config *config, |
2261 | struct lib_ring_buffer *buf, | |
2262 | struct channel *chan, | |
2263 | unsigned long offset, | |
2264 | unsigned long commit_count, | |
2265 | unsigned long idx, | |
2266 | u64 tsc) | |
2267 | { | |
2268 | unsigned long old_commit_count = commit_count | |
2269 | - chan->backend.subbuf_size; | |
2270 | ||
2271 | /* | |
2272 | * If we succeeded at updating cc_sb below, we are the subbuffer | |
2273 | * writer delivering the subbuffer. Deals with concurrent | |
2274 | * updates of the "cc" value without adding a add_return atomic | |
2275 | * operation to the fast path. | |
2276 | * | |
2277 | * We are doing the delivery in two steps: | |
2278 | * - First, we cmpxchg() cc_sb to the new value | |
2279 | * old_commit_count + 1. This ensures that we are the only | |
2280 | * subbuffer user successfully filling the subbuffer, but we | |
2281 | * do _not_ set the cc_sb value to "commit_count" yet. | |
2282 | * Therefore, other writers that would wrap around the ring | |
2283 | * buffer and try to start writing to our subbuffer would | |
2284 | * have to drop records, because it would appear as | |
2285 | * non-filled. | |
2286 | * We therefore have exclusive access to the subbuffer control | |
2287 | * structures. This mutual exclusion with other writers is | |
2288 | * crucially important to perform record overruns count in | |
2289 | * flight recorder mode locklessly. | |
2290 | * - When we are ready to release the subbuffer (either for | |
2291 | * reading or for overrun by other writers), we simply set the | |
2292 | * cc_sb value to "commit_count" and perform delivery. | |
2293 | * | |
2294 | * The subbuffer size is least 2 bytes (minimum size: 1 page). | |
2295 | * This guarantees that old_commit_count + 1 != commit_count. | |
2296 | */ | |
2297 | ||
2298 | /* | |
2299 | * Order prior updates to reserve count prior to the | |
2300 | * commit_cold cc_sb update. | |
2301 | */ | |
2302 | smp_wmb(); | |
2303 | if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb, | |
2304 | old_commit_count, old_commit_count + 1) | |
2305 | == old_commit_count)) { | |
a1817dcb MD |
2306 | u64 *ts_end; |
2307 | ||
aece661f MD |
2308 | /* |
2309 | * Start of exclusive subbuffer access. We are | |
2310 | * guaranteed to be the last writer in this subbuffer | |
2311 | * and any other writer trying to access this subbuffer | |
2312 | * in this state is required to drop records. | |
a1817dcb MD |
2313 | * |
2314 | * We can read the ts_end for the current sub-buffer | |
2315 | * which has been saved by the very last space | |
2316 | * reservation for the current sub-buffer. | |
2317 | * | |
2318 | * Order increment of commit counter before reading ts_end. | |
aece661f | 2319 | */ |
a1817dcb MD |
2320 | smp_mb(); |
2321 | ts_end = &buf->ts_end[idx]; | |
25337cb5 | 2322 | deliver_count_events(config, buf, idx); |
a1817dcb | 2323 | config->cb.buffer_end(buf, *ts_end, idx, |
aece661f MD |
2324 | lib_ring_buffer_get_data_size(config, |
2325 | buf, | |
2326 | idx)); | |
2327 | ||
2328 | /* | |
2329 | * Increment the packet counter while we have exclusive | |
2330 | * access. | |
2331 | */ | |
2332 | subbuffer_inc_packet_count(config, &buf->backend, idx); | |
2333 | ||
2334 | /* | |
2335 | * Set noref flag and offset for this subbuffer id. | |
2336 | * Contains a memory barrier that ensures counter stores | |
2337 | * are ordered before set noref and offset. | |
2338 | */ | |
2339 | lib_ring_buffer_set_noref_offset(config, &buf->backend, idx, | |
2340 | buf_trunc_val(offset, chan)); | |
2341 | ||
2342 | /* | |
2343 | * Order set_noref and record counter updates before the | |
2344 | * end of subbuffer exclusive access. Orders with | |
2345 | * respect to writers coming into the subbuffer after | |
2346 | * wrap around, and also order wrt concurrent readers. | |
2347 | */ | |
2348 | smp_mb(); | |
2349 | /* End of exclusive subbuffer access */ | |
2350 | v_set(config, &buf->commit_cold[idx].cc_sb, | |
2351 | commit_count); | |
2352 | /* | |
2353 | * Order later updates to reserve count after | |
2354 | * the commit_cold cc_sb update. | |
2355 | */ | |
2356 | smp_wmb(); | |
2357 | lib_ring_buffer_vmcore_check_deliver(config, buf, | |
2358 | commit_count, idx); | |
2359 | ||
2360 | /* | |
2361 | * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free. | |
2362 | */ | |
2363 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER | |
2364 | && atomic_long_read(&buf->active_readers) | |
2365 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
2366 | wake_up_interruptible(&buf->read_wait); | |
2367 | wake_up_interruptible(&chan->read_wait); | |
2368 | } | |
2369 | ||
2370 | } | |
2371 | } | |
2372 | EXPORT_SYMBOL_GPL(lib_ring_buffer_check_deliver_slow); | |
2373 | ||
02a766bb | 2374 | int __init init_lib_ring_buffer_frontend(void) |
6fb8de4b MD |
2375 | { |
2376 | int cpu; | |
2377 | ||
2378 | for_each_possible_cpu(cpu) | |
2379 | spin_lock_init(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
02a766bb | 2380 | return 0; |
6fb8de4b | 2381 | } |
02a766bb MD |
2382 | |
2383 | module_init(init_lib_ring_buffer_frontend); | |
1a5db82d MD |
2384 | |
2385 | void __exit exit_lib_ring_buffer_frontend(void) | |
2386 | { | |
2387 | } | |
2388 | ||
2389 | module_exit(exit_lib_ring_buffer_frontend); |