Commit | Line | Data |
---|---|---|
b7cdc182 | 1 | /* SPDX-License-Identifier: (GPL-2.0-only OR LGPL-2.1-only) |
9f36eaed | 2 | * |
f3bc08c5 MD |
3 | * ring_buffer_frontend.c |
4 | * | |
886d51a3 MD |
5 | * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
6 | * | |
f3bc08c5 MD |
7 | * Ring buffer wait-free buffer synchronization. Producer-consumer and flight |
8 | * recorder (overwrite) modes. See thesis: | |
9 | * | |
10 | * Desnoyers, Mathieu (2009), "Low-Impact Operating System Tracing", Ph.D. | |
11 | * dissertation, Ecole Polytechnique de Montreal. | |
12 | * http://www.lttng.org/pub/thesis/desnoyers-dissertation-2009-12.pdf | |
13 | * | |
14 | * - Algorithm presentation in Chapter 5: | |
15 | * "Lockless Multi-Core High-Throughput Buffering". | |
16 | * - Algorithm formal verification in Section 8.6: | |
17 | * "Formal verification of LTTng" | |
18 | * | |
19 | * Author: | |
20 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
21 | * | |
22 | * Inspired from LTT and RelayFS: | |
23 | * Karim Yaghmour <karim@opersys.com> | |
24 | * Tom Zanussi <zanussi@us.ibm.com> | |
25 | * Bob Wisniewski <bob@watson.ibm.com> | |
26 | * And from K42 : | |
27 | * Bob Wisniewski <bob@watson.ibm.com> | |
28 | * | |
29 | * Buffer reader semantic : | |
30 | * | |
31 | * - get_subbuf_size | |
32 | * while buffer is not finalized and empty | |
33 | * - get_subbuf | |
34 | * - if return value != 0, continue | |
35 | * - splice one subbuffer worth of data to a pipe | |
36 | * - splice the data from pipe to disk/network | |
37 | * - put_subbuf | |
f3bc08c5 MD |
38 | */ |
39 | ||
40 | #include <linux/delay.h> | |
41 | #include <linux/module.h> | |
42 | #include <linux/percpu.h> | |
2d70eed1 | 43 | #include <linux/kref.h> |
46e25482 | 44 | #include <linux/percpu-defs.h> |
be37f991 | 45 | #include <linux/timer.h> |
155b8d9b | 46 | #include <asm/cacheflush.h> |
f3bc08c5 | 47 | |
5760f3f4 MD |
48 | #include <include/ringbuffer/config.h> |
49 | #include <include/ringbuffer/backend.h> | |
50 | #include <include/ringbuffer/frontend.h> | |
51 | #include <include/ringbuffer/iterator.h> | |
52 | #include <include/ringbuffer/nohz.h> | |
f3bc08c5 MD |
53 | |
54 | /* | |
55 | * Internal structure representing offsets to use at a sub-buffer switch. | |
56 | */ | |
57 | struct switch_offsets { | |
58 | unsigned long begin, end, old; | |
59 | size_t pre_header_padding, size; | |
f5ea5800 MD |
60 | unsigned int switch_new_start:1, switch_new_end:1, switch_old_start:1, |
61 | switch_old_end:1; | |
f3bc08c5 MD |
62 | }; |
63 | ||
64 | #ifdef CONFIG_NO_HZ | |
65 | enum tick_nohz_val { | |
66 | TICK_NOHZ_STOP, | |
67 | TICK_NOHZ_FLUSH, | |
68 | TICK_NOHZ_RESTART, | |
69 | }; | |
70 | ||
71 | static ATOMIC_NOTIFIER_HEAD(tick_nohz_notifier); | |
72 | #endif /* CONFIG_NO_HZ */ | |
73 | ||
74 | static DEFINE_PER_CPU(spinlock_t, ring_buffer_nohz_lock); | |
75 | ||
76 | DEFINE_PER_CPU(unsigned int, lib_ring_buffer_nesting); | |
77 | EXPORT_PER_CPU_SYMBOL(lib_ring_buffer_nesting); | |
78 | ||
79 | static | |
80 | void lib_ring_buffer_print_errors(struct channel *chan, | |
81 | struct lib_ring_buffer *buf, int cpu); | |
64af2437 MD |
82 | static |
83 | void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf, | |
84 | enum switch_mode mode); | |
f3bc08c5 | 85 | |
aece661f MD |
86 | static |
87 | int lib_ring_buffer_poll_deliver(const struct lib_ring_buffer_config *config, | |
88 | struct lib_ring_buffer *buf, | |
89 | struct channel *chan) | |
90 | { | |
91 | unsigned long consumed_old, consumed_idx, commit_count, write_offset; | |
92 | ||
93 | consumed_old = atomic_long_read(&buf->consumed); | |
94 | consumed_idx = subbuf_index(consumed_old, chan); | |
95 | commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); | |
96 | /* | |
97 | * No memory barrier here, since we are only interested | |
98 | * in a statistically correct polling result. The next poll will | |
99 | * get the data is we are racing. The mb() that ensures correct | |
100 | * memory order is in get_subbuf. | |
101 | */ | |
102 | write_offset = v_read(config, &buf->offset); | |
103 | ||
104 | /* | |
105 | * Check that the subbuffer we are trying to consume has been | |
106 | * already fully committed. | |
107 | */ | |
108 | ||
109 | if (((commit_count - chan->backend.subbuf_size) | |
110 | & chan->commit_count_mask) | |
111 | - (buf_trunc(consumed_old, chan) | |
112 | >> chan->backend.num_subbuf_order) | |
113 | != 0) | |
114 | return 0; | |
115 | ||
116 | /* | |
117 | * Check that we are not about to read the same subbuffer in | |
118 | * which the writer head is. | |
119 | */ | |
120 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_old, chan) | |
121 | == 0) | |
122 | return 0; | |
123 | ||
124 | return 1; | |
125 | } | |
126 | ||
f3bc08c5 MD |
127 | /* |
128 | * Must be called under cpu hotplug protection. | |
129 | */ | |
130 | void lib_ring_buffer_free(struct lib_ring_buffer *buf) | |
131 | { | |
132 | struct channel *chan = buf->backend.chan; | |
133 | ||
134 | lib_ring_buffer_print_errors(chan, buf, buf->backend.cpu); | |
c50bdba2 MD |
135 | kvfree(buf->commit_hot); |
136 | kvfree(buf->commit_cold); | |
137 | kvfree(buf->ts_end); | |
f3bc08c5 MD |
138 | |
139 | lib_ring_buffer_backend_free(&buf->backend); | |
140 | } | |
141 | ||
142 | /** | |
143 | * lib_ring_buffer_reset - Reset ring buffer to initial values. | |
144 | * @buf: Ring buffer. | |
145 | * | |
146 | * Effectively empty the ring buffer. Should be called when the buffer is not | |
147 | * used for writing. The ring buffer can be opened for reading, but the reader | |
148 | * should not be using the iterator concurrently with reset. The previous | |
149 | * current iterator record is reset. | |
150 | */ | |
151 | void lib_ring_buffer_reset(struct lib_ring_buffer *buf) | |
152 | { | |
153 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 154 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
155 | unsigned int i; |
156 | ||
157 | /* | |
158 | * Reset iterator first. It will put the subbuffer if it currently holds | |
159 | * it. | |
160 | */ | |
161 | lib_ring_buffer_iterator_reset(buf); | |
162 | v_set(config, &buf->offset, 0); | |
163 | for (i = 0; i < chan->backend.num_subbuf; i++) { | |
164 | v_set(config, &buf->commit_hot[i].cc, 0); | |
165 | v_set(config, &buf->commit_hot[i].seq, 0); | |
166 | v_set(config, &buf->commit_cold[i].cc_sb, 0); | |
2485a430 | 167 | buf->ts_end[i] = 0; |
f3bc08c5 MD |
168 | } |
169 | atomic_long_set(&buf->consumed, 0); | |
170 | atomic_set(&buf->record_disabled, 0); | |
171 | v_set(config, &buf->last_tsc, 0); | |
172 | lib_ring_buffer_backend_reset(&buf->backend); | |
173 | /* Don't reset number of active readers */ | |
174 | v_set(config, &buf->records_lost_full, 0); | |
175 | v_set(config, &buf->records_lost_wrap, 0); | |
176 | v_set(config, &buf->records_lost_big, 0); | |
177 | v_set(config, &buf->records_count, 0); | |
178 | v_set(config, &buf->records_overrun, 0); | |
179 | buf->finalized = 0; | |
180 | } | |
181 | EXPORT_SYMBOL_GPL(lib_ring_buffer_reset); | |
182 | ||
183 | /** | |
184 | * channel_reset - Reset channel to initial values. | |
185 | * @chan: Channel. | |
186 | * | |
187 | * Effectively empty the channel. Should be called when the channel is not used | |
188 | * for writing. The channel can be opened for reading, but the reader should not | |
189 | * be using the iterator concurrently with reset. The previous current iterator | |
190 | * record is reset. | |
191 | */ | |
192 | void channel_reset(struct channel *chan) | |
193 | { | |
194 | /* | |
195 | * Reset iterators first. Will put the subbuffer if held for reading. | |
196 | */ | |
197 | channel_iterator_reset(chan); | |
198 | atomic_set(&chan->record_disabled, 0); | |
199 | /* Don't reset commit_count_mask, still valid */ | |
200 | channel_backend_reset(&chan->backend); | |
201 | /* Don't reset switch/read timer interval */ | |
202 | /* Don't reset notifiers and notifier enable bits */ | |
203 | /* Don't reset reader reference count */ | |
204 | } | |
205 | EXPORT_SYMBOL_GPL(channel_reset); | |
206 | ||
207 | /* | |
208 | * Must be called under cpu hotplug protection. | |
209 | */ | |
210 | int lib_ring_buffer_create(struct lib_ring_buffer *buf, | |
211 | struct channel_backend *chanb, int cpu) | |
212 | { | |
5a8fd222 | 213 | const struct lib_ring_buffer_config *config = &chanb->config; |
f3bc08c5 MD |
214 | struct channel *chan = container_of(chanb, struct channel, backend); |
215 | void *priv = chanb->priv; | |
f3bc08c5 MD |
216 | size_t subbuf_header_size; |
217 | u64 tsc; | |
218 | int ret; | |
219 | ||
220 | /* Test for cpu hotplug */ | |
221 | if (buf->backend.allocated) | |
222 | return 0; | |
223 | ||
224 | /* | |
225 | * Paranoia: per cpu dynamic allocation is not officially documented as | |
226 | * zeroing the memory, so let's do it here too, just in case. | |
227 | */ | |
228 | memset(buf, 0, sizeof(*buf)); | |
229 | ||
230 | ret = lib_ring_buffer_backend_create(&buf->backend, &chan->backend, cpu); | |
231 | if (ret) | |
232 | return ret; | |
233 | ||
234 | buf->commit_hot = | |
c50bdba2 | 235 | kvzalloc_node(ALIGN(sizeof(*buf->commit_hot) |
f3bc08c5 MD |
236 | * chan->backend.num_subbuf, |
237 | 1 << INTERNODE_CACHE_SHIFT), | |
df388b78 MD |
238 | GFP_KERNEL | __GFP_NOWARN, |
239 | cpu_to_node(max(cpu, 0))); | |
f3bc08c5 MD |
240 | if (!buf->commit_hot) { |
241 | ret = -ENOMEM; | |
242 | goto free_chanbuf; | |
243 | } | |
244 | ||
245 | buf->commit_cold = | |
c50bdba2 | 246 | kvzalloc_node(ALIGN(sizeof(*buf->commit_cold) |
f3bc08c5 MD |
247 | * chan->backend.num_subbuf, |
248 | 1 << INTERNODE_CACHE_SHIFT), | |
df388b78 MD |
249 | GFP_KERNEL | __GFP_NOWARN, |
250 | cpu_to_node(max(cpu, 0))); | |
f3bc08c5 MD |
251 | if (!buf->commit_cold) { |
252 | ret = -ENOMEM; | |
253 | goto free_commit; | |
254 | } | |
255 | ||
2485a430 | 256 | buf->ts_end = |
c50bdba2 | 257 | kvzalloc_node(ALIGN(sizeof(*buf->ts_end) |
2485a430 MD |
258 | * chan->backend.num_subbuf, |
259 | 1 << INTERNODE_CACHE_SHIFT), | |
260 | GFP_KERNEL | __GFP_NOWARN, | |
261 | cpu_to_node(max(cpu, 0))); | |
262 | if (!buf->ts_end) { | |
263 | ret = -ENOMEM; | |
264 | goto free_commit_cold; | |
265 | } | |
266 | ||
f3bc08c5 | 267 | init_waitqueue_head(&buf->read_wait); |
71c1d843 | 268 | init_waitqueue_head(&buf->write_wait); |
f3bc08c5 MD |
269 | raw_spin_lock_init(&buf->raw_tick_nohz_spinlock); |
270 | ||
271 | /* | |
272 | * Write the subbuffer header for first subbuffer so we know the total | |
273 | * duration of data gathering. | |
274 | */ | |
275 | subbuf_header_size = config->cb.subbuffer_header_size(); | |
276 | v_set(config, &buf->offset, subbuf_header_size); | |
277 | subbuffer_id_clear_noref(config, &buf->backend.buf_wsb[0].id); | |
278 | tsc = config->cb.ring_buffer_clock_read(buf->backend.chan); | |
279 | config->cb.buffer_begin(buf, tsc, 0); | |
280 | v_add(config, subbuf_header_size, &buf->commit_hot[0].cc); | |
281 | ||
282 | if (config->cb.buffer_create) { | |
283 | ret = config->cb.buffer_create(buf, priv, cpu, chanb->name); | |
284 | if (ret) | |
285 | goto free_init; | |
286 | } | |
287 | ||
288 | /* | |
289 | * Ensure the buffer is ready before setting it to allocated and setting | |
290 | * the cpumask. | |
291 | * Used for cpu hotplug vs cpumask iteration. | |
292 | */ | |
293 | smp_wmb(); | |
294 | buf->backend.allocated = 1; | |
295 | ||
296 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
297 | CHAN_WARN_ON(chan, cpumask_test_cpu(cpu, | |
298 | chan->backend.cpumask)); | |
299 | cpumask_set_cpu(cpu, chan->backend.cpumask); | |
300 | } | |
f3bc08c5 MD |
301 | return 0; |
302 | ||
303 | /* Error handling */ | |
304 | free_init: | |
c50bdba2 | 305 | kvfree(buf->ts_end); |
2485a430 | 306 | free_commit_cold: |
c50bdba2 | 307 | kvfree(buf->commit_cold); |
f3bc08c5 | 308 | free_commit: |
c50bdba2 | 309 | kvfree(buf->commit_hot); |
f3bc08c5 MD |
310 | free_chanbuf: |
311 | lib_ring_buffer_backend_free(&buf->backend); | |
312 | return ret; | |
313 | } | |
314 | ||
be37f991 | 315 | static void switch_buffer_timer(struct timer_list *t) |
f3bc08c5 | 316 | { |
be37f991 | 317 | struct lib_ring_buffer *buf = from_timer(buf, t, switch_timer); |
f3bc08c5 | 318 | struct channel *chan = buf->backend.chan; |
f3bc08c5 MD |
319 | |
320 | /* | |
321 | * Only flush buffers periodically if readers are active. | |
322 | */ | |
323 | if (atomic_long_read(&buf->active_readers)) | |
324 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
325 | ||
be37f991 MD |
326 | mod_timer(&buf->switch_timer, |
327 | jiffies + chan->switch_timer_interval); | |
f3bc08c5 MD |
328 | } |
329 | ||
330 | /* | |
331 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
332 | */ | |
333 | static void lib_ring_buffer_start_switch_timer(struct lib_ring_buffer *buf) | |
334 | { | |
335 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 336 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
1fd97f9f | 337 | unsigned int flags = 0; |
f3bc08c5 MD |
338 | |
339 | if (!chan->switch_timer_interval || buf->switch_timer_enabled) | |
340 | return; | |
152fe7fc MJ |
341 | |
342 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
be37f991 | 343 | flags = TIMER_PINNED; |
152fe7fc | 344 | |
be37f991 | 345 | timer_setup(&buf->switch_timer, switch_buffer_timer, flags); |
f3bc08c5 | 346 | buf->switch_timer.expires = jiffies + chan->switch_timer_interval; |
1fd97f9f | 347 | |
f3bc08c5 MD |
348 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
349 | add_timer_on(&buf->switch_timer, buf->backend.cpu); | |
350 | else | |
351 | add_timer(&buf->switch_timer); | |
1fd97f9f | 352 | |
f3bc08c5 MD |
353 | buf->switch_timer_enabled = 1; |
354 | } | |
355 | ||
356 | /* | |
357 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
358 | */ | |
359 | static void lib_ring_buffer_stop_switch_timer(struct lib_ring_buffer *buf) | |
360 | { | |
361 | struct channel *chan = buf->backend.chan; | |
362 | ||
363 | if (!chan->switch_timer_interval || !buf->switch_timer_enabled) | |
364 | return; | |
365 | ||
366 | del_timer_sync(&buf->switch_timer); | |
367 | buf->switch_timer_enabled = 0; | |
368 | } | |
369 | ||
370 | /* | |
371 | * Polling timer to check the channels for data. | |
372 | */ | |
be37f991 | 373 | static void read_buffer_timer(struct timer_list *t) |
f3bc08c5 | 374 | { |
be37f991 | 375 | struct lib_ring_buffer *buf = from_timer(buf, t, read_timer); |
f3bc08c5 | 376 | struct channel *chan = buf->backend.chan; |
5a8fd222 | 377 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
378 | |
379 | CHAN_WARN_ON(chan, !buf->backend.allocated); | |
380 | ||
381 | if (atomic_long_read(&buf->active_readers) | |
382 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
383 | wake_up_interruptible(&buf->read_wait); | |
384 | wake_up_interruptible(&chan->read_wait); | |
385 | } | |
386 | ||
be37f991 MD |
387 | mod_timer(&buf->read_timer, |
388 | jiffies + chan->read_timer_interval); | |
f3bc08c5 MD |
389 | } |
390 | ||
391 | /* | |
392 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
393 | */ | |
394 | static void lib_ring_buffer_start_read_timer(struct lib_ring_buffer *buf) | |
395 | { | |
396 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 397 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
b1ad8713 | 398 | unsigned int flags = 0; |
f3bc08c5 MD |
399 | |
400 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
401 | || !chan->read_timer_interval | |
402 | || buf->read_timer_enabled) | |
403 | return; | |
404 | ||
152fe7fc | 405 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) |
be37f991 | 406 | flags = TIMER_PINNED; |
152fe7fc | 407 | |
be37f991 | 408 | timer_setup(&buf->read_timer, read_buffer_timer, flags); |
f3bc08c5 | 409 | buf->read_timer.expires = jiffies + chan->read_timer_interval; |
f3bc08c5 MD |
410 | |
411 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
412 | add_timer_on(&buf->read_timer, buf->backend.cpu); | |
413 | else | |
414 | add_timer(&buf->read_timer); | |
1fd97f9f | 415 | |
f3bc08c5 MD |
416 | buf->read_timer_enabled = 1; |
417 | } | |
418 | ||
419 | /* | |
420 | * Called with ring_buffer_nohz_lock held for per-cpu buffers. | |
421 | */ | |
422 | static void lib_ring_buffer_stop_read_timer(struct lib_ring_buffer *buf) | |
423 | { | |
424 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 425 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
426 | |
427 | if (config->wakeup != RING_BUFFER_WAKEUP_BY_TIMER | |
428 | || !chan->read_timer_interval | |
429 | || !buf->read_timer_enabled) | |
430 | return; | |
431 | ||
432 | del_timer_sync(&buf->read_timer); | |
433 | /* | |
434 | * do one more check to catch data that has been written in the last | |
435 | * timer period. | |
436 | */ | |
437 | if (lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
438 | wake_up_interruptible(&buf->read_wait); | |
439 | wake_up_interruptible(&chan->read_wait); | |
440 | } | |
441 | buf->read_timer_enabled = 0; | |
442 | } | |
443 | ||
1e367326 MD |
444 | enum cpuhp_state lttng_rb_hp_prepare; |
445 | enum cpuhp_state lttng_rb_hp_online; | |
446 | ||
447 | void lttng_rb_set_hp_prepare(enum cpuhp_state val) | |
448 | { | |
449 | lttng_rb_hp_prepare = val; | |
450 | } | |
451 | EXPORT_SYMBOL_GPL(lttng_rb_set_hp_prepare); | |
452 | ||
453 | void lttng_rb_set_hp_online(enum cpuhp_state val) | |
454 | { | |
455 | lttng_rb_hp_online = val; | |
456 | } | |
457 | EXPORT_SYMBOL_GPL(lttng_rb_set_hp_online); | |
458 | ||
459 | int lttng_cpuhp_rb_frontend_dead(unsigned int cpu, | |
460 | struct lttng_cpuhp_node *node) | |
461 | { | |
462 | struct channel *chan = container_of(node, struct channel, | |
463 | cpuhp_prepare); | |
464 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
465 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
466 | ||
467 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
468 | ||
469 | /* | |
470 | * Performing a buffer switch on a remote CPU. Performed by | |
471 | * the CPU responsible for doing the hotunplug after the target | |
472 | * CPU stopped running completely. Ensures that all data | |
473 | * from that remote CPU is flushed. | |
474 | */ | |
475 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
476 | return 0; | |
477 | } | |
478 | EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_dead); | |
479 | ||
480 | int lttng_cpuhp_rb_frontend_online(unsigned int cpu, | |
481 | struct lttng_cpuhp_node *node) | |
482 | { | |
483 | struct channel *chan = container_of(node, struct channel, | |
484 | cpuhp_online); | |
485 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
486 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
487 | ||
488 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
489 | ||
490 | wake_up_interruptible(&chan->hp_wait); | |
491 | lib_ring_buffer_start_switch_timer(buf); | |
492 | lib_ring_buffer_start_read_timer(buf); | |
493 | return 0; | |
494 | } | |
495 | EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_online); | |
496 | ||
497 | int lttng_cpuhp_rb_frontend_offline(unsigned int cpu, | |
498 | struct lttng_cpuhp_node *node) | |
499 | { | |
500 | struct channel *chan = container_of(node, struct channel, | |
501 | cpuhp_online); | |
502 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, cpu); | |
503 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
504 | ||
505 | CHAN_WARN_ON(chan, config->alloc == RING_BUFFER_ALLOC_GLOBAL); | |
506 | ||
507 | lib_ring_buffer_stop_switch_timer(buf); | |
508 | lib_ring_buffer_stop_read_timer(buf); | |
509 | return 0; | |
510 | } | |
511 | EXPORT_SYMBOL_GPL(lttng_cpuhp_rb_frontend_offline); | |
512 | ||
23b908b0 | 513 | #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) |
f3bc08c5 MD |
514 | /* |
515 | * For per-cpu buffers, call the reader wakeups before switching the buffer, so | |
516 | * that wake-up-tracing generated events are flushed before going idle (in | |
517 | * tick_nohz). We test if the spinlock is locked to deal with the race where | |
518 | * readers try to sample the ring buffer before we perform the switch. We let | |
519 | * the readers retry in that case. If there is data in the buffer, the wake up | |
520 | * is going to forbid the CPU running the reader thread from going idle. | |
521 | */ | |
522 | static int notrace ring_buffer_tick_nohz_callback(struct notifier_block *nb, | |
523 | unsigned long val, | |
524 | void *data) | |
525 | { | |
526 | struct channel *chan = container_of(nb, struct channel, | |
527 | tick_nohz_notifier); | |
5a8fd222 | 528 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
529 | struct lib_ring_buffer *buf; |
530 | int cpu = smp_processor_id(); | |
531 | ||
532 | if (config->alloc != RING_BUFFER_ALLOC_PER_CPU) { | |
533 | /* | |
534 | * We don't support keeping the system idle with global buffers | |
535 | * and streaming active. In order to do so, we would need to | |
536 | * sample a non-nohz-cpumask racelessly with the nohz updates | |
537 | * without adding synchronization overhead to nohz. Leave this | |
538 | * use-case out for now. | |
539 | */ | |
540 | return 0; | |
541 | } | |
542 | ||
543 | buf = channel_get_ring_buffer(config, chan, cpu); | |
544 | switch (val) { | |
545 | case TICK_NOHZ_FLUSH: | |
546 | raw_spin_lock(&buf->raw_tick_nohz_spinlock); | |
547 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_TIMER | |
548 | && chan->read_timer_interval | |
549 | && atomic_long_read(&buf->active_readers) | |
550 | && (lib_ring_buffer_poll_deliver(config, buf, chan) | |
551 | || lib_ring_buffer_pending_data(config, buf, chan))) { | |
552 | wake_up_interruptible(&buf->read_wait); | |
553 | wake_up_interruptible(&chan->read_wait); | |
554 | } | |
555 | if (chan->switch_timer_interval) | |
556 | lib_ring_buffer_switch_slow(buf, SWITCH_ACTIVE); | |
557 | raw_spin_unlock(&buf->raw_tick_nohz_spinlock); | |
558 | break; | |
559 | case TICK_NOHZ_STOP: | |
46e25482 | 560 | spin_lock(this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
561 | lib_ring_buffer_stop_switch_timer(buf); |
562 | lib_ring_buffer_stop_read_timer(buf); | |
46e25482 | 563 | spin_unlock(this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
564 | break; |
565 | case TICK_NOHZ_RESTART: | |
46e25482 | 566 | spin_lock(this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
567 | lib_ring_buffer_start_read_timer(buf); |
568 | lib_ring_buffer_start_switch_timer(buf); | |
46e25482 | 569 | spin_unlock(this_cpu_ptr(&ring_buffer_nohz_lock)); |
f3bc08c5 MD |
570 | break; |
571 | } | |
572 | ||
573 | return 0; | |
574 | } | |
575 | ||
576 | void notrace lib_ring_buffer_tick_nohz_flush(void) | |
577 | { | |
578 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_FLUSH, | |
579 | NULL); | |
580 | } | |
581 | ||
582 | void notrace lib_ring_buffer_tick_nohz_stop(void) | |
583 | { | |
584 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_STOP, | |
585 | NULL); | |
586 | } | |
587 | ||
588 | void notrace lib_ring_buffer_tick_nohz_restart(void) | |
589 | { | |
590 | atomic_notifier_call_chain(&tick_nohz_notifier, TICK_NOHZ_RESTART, | |
591 | NULL); | |
592 | } | |
23b908b0 | 593 | #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */ |
f3bc08c5 MD |
594 | |
595 | /* | |
596 | * Holds CPU hotplug. | |
597 | */ | |
598 | static void channel_unregister_notifiers(struct channel *chan) | |
599 | { | |
5a8fd222 | 600 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
601 | |
602 | channel_iterator_unregister_notifiers(chan); | |
603 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
604 | #ifdef CONFIG_NO_HZ | |
605 | /* | |
606 | * Remove the nohz notifier first, so we are certain we stop | |
607 | * the timers. | |
608 | */ | |
609 | atomic_notifier_chain_unregister(&tick_nohz_notifier, | |
610 | &chan->tick_nohz_notifier); | |
611 | /* | |
612 | * ring_buffer_nohz_lock will not be needed below, because | |
613 | * we just removed the notifiers, which were the only source of | |
614 | * concurrency. | |
615 | */ | |
616 | #endif /* CONFIG_NO_HZ */ | |
1e367326 MD |
617 | { |
618 | int ret; | |
619 | ||
620 | ret = cpuhp_state_remove_instance(lttng_rb_hp_online, | |
621 | &chan->cpuhp_online.node); | |
622 | WARN_ON(ret); | |
623 | ret = cpuhp_state_remove_instance_nocalls(lttng_rb_hp_prepare, | |
624 | &chan->cpuhp_prepare.node); | |
625 | WARN_ON(ret); | |
f3bc08c5 | 626 | } |
f3bc08c5 MD |
627 | } else { |
628 | struct lib_ring_buffer *buf = chan->backend.buf; | |
629 | ||
630 | lib_ring_buffer_stop_switch_timer(buf); | |
631 | lib_ring_buffer_stop_read_timer(buf); | |
632 | } | |
633 | channel_backend_unregister_notifiers(&chan->backend); | |
634 | } | |
635 | ||
64af2437 MD |
636 | static void lib_ring_buffer_set_quiescent(struct lib_ring_buffer *buf) |
637 | { | |
638 | if (!buf->quiescent) { | |
639 | buf->quiescent = true; | |
640 | _lib_ring_buffer_switch_remote(buf, SWITCH_FLUSH); | |
641 | } | |
642 | } | |
643 | ||
644 | static void lib_ring_buffer_clear_quiescent(struct lib_ring_buffer *buf) | |
645 | { | |
646 | buf->quiescent = false; | |
647 | } | |
648 | ||
649 | void lib_ring_buffer_set_quiescent_channel(struct channel *chan) | |
650 | { | |
651 | int cpu; | |
652 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
653 | ||
654 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
655 | get_online_cpus(); | |
656 | for_each_channel_cpu(cpu, chan) { | |
657 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
658 | cpu); | |
659 | ||
660 | lib_ring_buffer_set_quiescent(buf); | |
661 | } | |
662 | put_online_cpus(); | |
663 | } else { | |
664 | struct lib_ring_buffer *buf = chan->backend.buf; | |
665 | ||
666 | lib_ring_buffer_set_quiescent(buf); | |
667 | } | |
668 | } | |
669 | EXPORT_SYMBOL_GPL(lib_ring_buffer_set_quiescent_channel); | |
670 | ||
671 | void lib_ring_buffer_clear_quiescent_channel(struct channel *chan) | |
672 | { | |
673 | int cpu; | |
674 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
675 | ||
676 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
677 | get_online_cpus(); | |
678 | for_each_channel_cpu(cpu, chan) { | |
679 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
680 | cpu); | |
681 | ||
682 | lib_ring_buffer_clear_quiescent(buf); | |
683 | } | |
684 | put_online_cpus(); | |
685 | } else { | |
686 | struct lib_ring_buffer *buf = chan->backend.buf; | |
687 | ||
688 | lib_ring_buffer_clear_quiescent(buf); | |
689 | } | |
690 | } | |
691 | EXPORT_SYMBOL_GPL(lib_ring_buffer_clear_quiescent_channel); | |
692 | ||
f3bc08c5 MD |
693 | static void channel_free(struct channel *chan) |
694 | { | |
dd5a0db3 MD |
695 | if (chan->backend.release_priv_ops) { |
696 | chan->backend.release_priv_ops(chan->backend.priv_ops); | |
697 | } | |
f3bc08c5 MD |
698 | channel_iterator_free(chan); |
699 | channel_backend_free(&chan->backend); | |
700 | kfree(chan); | |
701 | } | |
702 | ||
703 | /** | |
704 | * channel_create - Create channel. | |
705 | * @config: ring buffer instance configuration | |
706 | * @name: name of the channel | |
707 | * @priv: ring buffer client private data | |
708 | * @buf_addr: pointer the the beginning of the preallocated buffer contiguous | |
709 | * address mapping. It is used only by RING_BUFFER_STATIC | |
710 | * configuration. It can be set to NULL for other backends. | |
711 | * @subbuf_size: subbuffer size | |
712 | * @num_subbuf: number of subbuffers | |
713 | * @switch_timer_interval: Time interval (in us) to fill sub-buffers with | |
714 | * padding to let readers get those sub-buffers. | |
715 | * Used for live streaming. | |
716 | * @read_timer_interval: Time interval (in us) to wake up pending readers. | |
717 | * | |
718 | * Holds cpu hotplug. | |
719 | * Returns NULL on failure. | |
720 | */ | |
721 | struct channel *channel_create(const struct lib_ring_buffer_config *config, | |
722 | const char *name, void *priv, void *buf_addr, | |
723 | size_t subbuf_size, | |
724 | size_t num_subbuf, unsigned int switch_timer_interval, | |
725 | unsigned int read_timer_interval) | |
726 | { | |
1e367326 | 727 | int ret; |
f3bc08c5 MD |
728 | struct channel *chan; |
729 | ||
730 | if (lib_ring_buffer_check_config(config, switch_timer_interval, | |
731 | read_timer_interval)) | |
732 | return NULL; | |
733 | ||
734 | chan = kzalloc(sizeof(struct channel), GFP_KERNEL); | |
735 | if (!chan) | |
736 | return NULL; | |
737 | ||
738 | ret = channel_backend_init(&chan->backend, name, config, priv, | |
739 | subbuf_size, num_subbuf); | |
740 | if (ret) | |
741 | goto error; | |
742 | ||
743 | ret = channel_iterator_init(chan); | |
744 | if (ret) | |
745 | goto error_free_backend; | |
746 | ||
747 | chan->commit_count_mask = (~0UL >> chan->backend.num_subbuf_order); | |
748 | chan->switch_timer_interval = usecs_to_jiffies(switch_timer_interval); | |
749 | chan->read_timer_interval = usecs_to_jiffies(read_timer_interval); | |
f40270ad | 750 | kref_init(&chan->ref); |
f3bc08c5 | 751 | init_waitqueue_head(&chan->read_wait); |
24cedcfe | 752 | init_waitqueue_head(&chan->hp_wait); |
f3bc08c5 MD |
753 | |
754 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
1e367326 MD |
755 | chan->cpuhp_prepare.component = LTTNG_RING_BUFFER_FRONTEND; |
756 | ret = cpuhp_state_add_instance_nocalls(lttng_rb_hp_prepare, | |
757 | &chan->cpuhp_prepare.node); | |
758 | if (ret) | |
759 | goto cpuhp_prepare_error; | |
760 | ||
761 | chan->cpuhp_online.component = LTTNG_RING_BUFFER_FRONTEND; | |
762 | ret = cpuhp_state_add_instance(lttng_rb_hp_online, | |
763 | &chan->cpuhp_online.node); | |
764 | if (ret) | |
765 | goto cpuhp_online_error; | |
1e367326 | 766 | |
23b908b0 | 767 | #if defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) |
f3bc08c5 MD |
768 | /* Only benefit from NO_HZ idle with per-cpu buffers for now. */ |
769 | chan->tick_nohz_notifier.notifier_call = | |
770 | ring_buffer_tick_nohz_callback; | |
771 | chan->tick_nohz_notifier.priority = ~0U; | |
772 | atomic_notifier_chain_register(&tick_nohz_notifier, | |
773 | &chan->tick_nohz_notifier); | |
23b908b0 | 774 | #endif /* defined(CONFIG_NO_HZ) && defined(CONFIG_LIB_RING_BUFFER) */ |
f3bc08c5 | 775 | |
f3bc08c5 MD |
776 | } else { |
777 | struct lib_ring_buffer *buf = chan->backend.buf; | |
778 | ||
779 | lib_ring_buffer_start_switch_timer(buf); | |
780 | lib_ring_buffer_start_read_timer(buf); | |
781 | } | |
782 | ||
783 | return chan; | |
784 | ||
1e367326 MD |
785 | cpuhp_online_error: |
786 | ret = cpuhp_state_remove_instance_nocalls(lttng_rb_hp_prepare, | |
787 | &chan->cpuhp_prepare.node); | |
788 | WARN_ON(ret); | |
789 | cpuhp_prepare_error: | |
f3bc08c5 MD |
790 | error_free_backend: |
791 | channel_backend_free(&chan->backend); | |
792 | error: | |
793 | kfree(chan); | |
794 | return NULL; | |
795 | } | |
796 | EXPORT_SYMBOL_GPL(channel_create); | |
797 | ||
f40270ad MD |
798 | static |
799 | void channel_release(struct kref *kref) | |
800 | { | |
801 | struct channel *chan = container_of(kref, struct channel, ref); | |
802 | channel_free(chan); | |
803 | } | |
804 | ||
f3bc08c5 MD |
805 | /** |
806 | * channel_destroy - Finalize, wait for q.s. and destroy channel. | |
807 | * @chan: channel to destroy | |
808 | * | |
809 | * Holds cpu hotplug. | |
9a0df743 MD |
810 | * Call "destroy" callback, finalize channels, and then decrement the |
811 | * channel reference count. Note that when readers have completed data | |
812 | * consumption of finalized channels, get_subbuf() will return -ENODATA. | |
813 | * They should release their handle at that point. Returns the private | |
814 | * data pointer. | |
f3bc08c5 MD |
815 | */ |
816 | void *channel_destroy(struct channel *chan) | |
817 | { | |
818 | int cpu; | |
5a8fd222 | 819 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
820 | void *priv; |
821 | ||
822 | channel_unregister_notifiers(chan); | |
823 | ||
824 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
825 | /* | |
826 | * No need to hold cpu hotplug, because all notifiers have been | |
827 | * unregistered. | |
828 | */ | |
829 | for_each_channel_cpu(cpu, chan) { | |
830 | struct lib_ring_buffer *buf = per_cpu_ptr(chan->backend.buf, | |
831 | cpu); | |
832 | ||
833 | if (config->cb.buffer_finalize) | |
834 | config->cb.buffer_finalize(buf, | |
835 | chan->backend.priv, | |
836 | cpu); | |
f3bc08c5 MD |
837 | /* |
838 | * Perform flush before writing to finalized. | |
839 | */ | |
840 | smp_wmb(); | |
a8f2d0c7 | 841 | WRITE_ONCE(buf->finalized, 1); |
f3bc08c5 MD |
842 | wake_up_interruptible(&buf->read_wait); |
843 | } | |
844 | } else { | |
845 | struct lib_ring_buffer *buf = chan->backend.buf; | |
846 | ||
847 | if (config->cb.buffer_finalize) | |
848 | config->cb.buffer_finalize(buf, chan->backend.priv, -1); | |
f3bc08c5 MD |
849 | /* |
850 | * Perform flush before writing to finalized. | |
851 | */ | |
852 | smp_wmb(); | |
a8f2d0c7 | 853 | WRITE_ONCE(buf->finalized, 1); |
f3bc08c5 MD |
854 | wake_up_interruptible(&buf->read_wait); |
855 | } | |
a8f2d0c7 | 856 | WRITE_ONCE(chan->finalized, 1); |
24cedcfe | 857 | wake_up_interruptible(&chan->hp_wait); |
f3bc08c5 | 858 | wake_up_interruptible(&chan->read_wait); |
f3bc08c5 | 859 | priv = chan->backend.priv; |
ba1d61bc | 860 | kref_put(&chan->ref, channel_release); |
f3bc08c5 MD |
861 | return priv; |
862 | } | |
863 | EXPORT_SYMBOL_GPL(channel_destroy); | |
864 | ||
865 | struct lib_ring_buffer *channel_get_ring_buffer( | |
866 | const struct lib_ring_buffer_config *config, | |
867 | struct channel *chan, int cpu) | |
868 | { | |
869 | if (config->alloc == RING_BUFFER_ALLOC_GLOBAL) | |
870 | return chan->backend.buf; | |
871 | else | |
872 | return per_cpu_ptr(chan->backend.buf, cpu); | |
873 | } | |
874 | EXPORT_SYMBOL_GPL(channel_get_ring_buffer); | |
875 | ||
876 | int lib_ring_buffer_open_read(struct lib_ring_buffer *buf) | |
877 | { | |
878 | struct channel *chan = buf->backend.chan; | |
879 | ||
880 | if (!atomic_long_add_unless(&buf->active_readers, 1, 1)) | |
881 | return -EBUSY; | |
2d70eed1 | 882 | kref_get(&chan->ref); |
f1f2e114 | 883 | smp_mb__after_atomic(); |
f3bc08c5 MD |
884 | return 0; |
885 | } | |
886 | EXPORT_SYMBOL_GPL(lib_ring_buffer_open_read); | |
887 | ||
888 | void lib_ring_buffer_release_read(struct lib_ring_buffer *buf) | |
889 | { | |
890 | struct channel *chan = buf->backend.chan; | |
891 | ||
892 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
f1f2e114 | 893 | smp_mb__before_atomic(); |
f3bc08c5 | 894 | atomic_long_dec(&buf->active_readers); |
f40270ad | 895 | kref_put(&chan->ref, channel_release); |
f3bc08c5 MD |
896 | } |
897 | EXPORT_SYMBOL_GPL(lib_ring_buffer_release_read); | |
898 | ||
899 | /* | |
900 | * Promote compiler barrier to a smp_mb(). | |
901 | * For the specific ring buffer case, this IPI call should be removed if the | |
902 | * architecture does not reorder writes. This should eventually be provided by | |
903 | * a separate architecture-specific infrastructure. | |
904 | */ | |
905 | static void remote_mb(void *info) | |
906 | { | |
907 | smp_mb(); | |
908 | } | |
909 | ||
910 | /** | |
911 | * lib_ring_buffer_snapshot - save subbuffer position snapshot (for read) | |
912 | * @buf: ring buffer | |
913 | * @consumed: consumed count indicating the position where to read | |
914 | * @produced: produced count, indicates position when to stop reading | |
915 | * | |
916 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
917 | * data to read at consumed position, or 0 if the get operation succeeds. | |
918 | * Busy-loop trying to get data if the tick_nohz sequence lock is held. | |
919 | */ | |
920 | ||
921 | int lib_ring_buffer_snapshot(struct lib_ring_buffer *buf, | |
922 | unsigned long *consumed, unsigned long *produced) | |
923 | { | |
924 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 925 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
926 | unsigned long consumed_cur, write_offset; |
927 | int finalized; | |
928 | ||
929 | retry: | |
a8f2d0c7 | 930 | finalized = READ_ONCE(buf->finalized); |
f3bc08c5 MD |
931 | /* |
932 | * Read finalized before counters. | |
933 | */ | |
934 | smp_rmb(); | |
935 | consumed_cur = atomic_long_read(&buf->consumed); | |
936 | /* | |
937 | * No need to issue a memory barrier between consumed count read and | |
938 | * write offset read, because consumed count can only change | |
939 | * concurrently in overwrite mode, and we keep a sequence counter | |
940 | * identifier derived from the write offset to check we are getting | |
941 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
942 | * "tagged" upon writes, tags are checked upon read). | |
943 | */ | |
944 | write_offset = v_read(config, &buf->offset); | |
945 | ||
946 | /* | |
947 | * Check that we are not about to read the same subbuffer in | |
948 | * which the writer head is. | |
949 | */ | |
950 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed_cur, chan) | |
951 | == 0) | |
952 | goto nodata; | |
953 | ||
954 | *consumed = consumed_cur; | |
955 | *produced = subbuf_trunc(write_offset, chan); | |
956 | ||
957 | return 0; | |
958 | ||
959 | nodata: | |
960 | /* | |
961 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
962 | * of "raw_spin_is_locked" memory ordering. | |
963 | */ | |
964 | if (finalized) | |
965 | return -ENODATA; | |
966 | else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock)) | |
967 | goto retry; | |
968 | else | |
969 | return -EAGAIN; | |
970 | } | |
971 | EXPORT_SYMBOL_GPL(lib_ring_buffer_snapshot); | |
972 | ||
4dce5a48 JG |
973 | /** |
974 | * Performs the same function as lib_ring_buffer_snapshot(), but the positions | |
975 | * are saved regardless of whether the consumed and produced positions are | |
976 | * in the same subbuffer. | |
977 | * @buf: ring buffer | |
978 | * @consumed: consumed byte count indicating the last position read | |
979 | * @produced: produced byte count indicating the last position written | |
980 | * | |
981 | * This function is meant to provide information on the exact producer and | |
982 | * consumer positions without regard for the "snapshot" feature. | |
983 | */ | |
984 | int lib_ring_buffer_snapshot_sample_positions(struct lib_ring_buffer *buf, | |
985 | unsigned long *consumed, unsigned long *produced) | |
986 | { | |
987 | struct channel *chan = buf->backend.chan; | |
988 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
989 | ||
990 | smp_rmb(); | |
991 | *consumed = atomic_long_read(&buf->consumed); | |
992 | /* | |
993 | * No need to issue a memory barrier between consumed count read and | |
994 | * write offset read, because consumed count can only change | |
995 | * concurrently in overwrite mode, and we keep a sequence counter | |
996 | * identifier derived from the write offset to check we are getting | |
997 | * the same sub-buffer we are expecting (the sub-buffers are atomically | |
998 | * "tagged" upon writes, tags are checked upon read). | |
999 | */ | |
1000 | *produced = v_read(config, &buf->offset); | |
1001 | return 0; | |
1002 | } | |
1003 | ||
f3bc08c5 MD |
1004 | /** |
1005 | * lib_ring_buffer_put_snapshot - move consumed counter forward | |
71c1d843 MD |
1006 | * |
1007 | * Should only be called from consumer context. | |
f3bc08c5 MD |
1008 | * @buf: ring buffer |
1009 | * @consumed_new: new consumed count value | |
1010 | */ | |
1011 | void lib_ring_buffer_move_consumer(struct lib_ring_buffer *buf, | |
1012 | unsigned long consumed_new) | |
1013 | { | |
1014 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1015 | struct channel *chan = bufb->chan; | |
1016 | unsigned long consumed; | |
1017 | ||
1018 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
1019 | ||
1020 | /* | |
1021 | * Only push the consumed value forward. | |
1022 | * If the consumed cmpxchg fails, this is because we have been pushed by | |
1023 | * the writer in flight recorder mode. | |
1024 | */ | |
1025 | consumed = atomic_long_read(&buf->consumed); | |
1026 | while ((long) consumed - (long) consumed_new < 0) | |
1027 | consumed = atomic_long_cmpxchg(&buf->consumed, consumed, | |
1028 | consumed_new); | |
71c1d843 MD |
1029 | /* Wake-up the metadata producer */ |
1030 | wake_up_interruptible(&buf->write_wait); | |
f3bc08c5 MD |
1031 | } |
1032 | EXPORT_SYMBOL_GPL(lib_ring_buffer_move_consumer); | |
1033 | ||
155b8d9b MD |
1034 | #if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE |
1035 | static void lib_ring_buffer_flush_read_subbuf_dcache( | |
1036 | const struct lib_ring_buffer_config *config, | |
1037 | struct channel *chan, | |
1038 | struct lib_ring_buffer *buf) | |
1039 | { | |
1040 | struct lib_ring_buffer_backend_pages *pages; | |
1041 | unsigned long sb_bindex, id, i, nr_pages; | |
1042 | ||
1043 | if (config->output != RING_BUFFER_MMAP) | |
1044 | return; | |
1045 | ||
1046 | /* | |
1047 | * Architectures with caches aliased on virtual addresses may | |
1048 | * use different cache lines for the linear mapping vs | |
1049 | * user-space memory mapping. Given that the ring buffer is | |
1050 | * based on the kernel linear mapping, aligning it with the | |
1051 | * user-space mapping is not straightforward, and would require | |
1052 | * extra TLB entries. Therefore, simply flush the dcache for the | |
1053 | * entire sub-buffer before reading it. | |
1054 | */ | |
1055 | id = buf->backend.buf_rsb.id; | |
1056 | sb_bindex = subbuffer_id_get_index(config, id); | |
1057 | pages = buf->backend.array[sb_bindex]; | |
1058 | nr_pages = buf->backend.num_pages_per_subbuf; | |
1059 | for (i = 0; i < nr_pages; i++) { | |
1060 | struct lib_ring_buffer_backend_page *backend_page; | |
1061 | ||
1062 | backend_page = &pages->p[i]; | |
1063 | flush_dcache_page(pfn_to_page(backend_page->pfn)); | |
1064 | } | |
1065 | } | |
1066 | #else | |
1067 | static void lib_ring_buffer_flush_read_subbuf_dcache( | |
1068 | const struct lib_ring_buffer_config *config, | |
1069 | struct channel *chan, | |
1070 | struct lib_ring_buffer *buf) | |
1071 | { | |
1072 | } | |
1073 | #endif | |
1074 | ||
f3bc08c5 MD |
1075 | /** |
1076 | * lib_ring_buffer_get_subbuf - get exclusive access to subbuffer for reading | |
1077 | * @buf: ring buffer | |
1078 | * @consumed: consumed count indicating the position where to read | |
1079 | * | |
1080 | * Returns -ENODATA if buffer is finalized, -EAGAIN if there is currently no | |
1081 | * data to read at consumed position, or 0 if the get operation succeeds. | |
1082 | * Busy-loop trying to get data if the tick_nohz sequence lock is held. | |
1083 | */ | |
1084 | int lib_ring_buffer_get_subbuf(struct lib_ring_buffer *buf, | |
1085 | unsigned long consumed) | |
1086 | { | |
1087 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1088 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1089 | unsigned long consumed_cur, consumed_idx, commit_count, write_offset; |
1090 | int ret; | |
1091 | int finalized; | |
1092 | ||
8202b8a0 MD |
1093 | if (buf->get_subbuf) { |
1094 | /* | |
1095 | * Reader is trying to get a subbuffer twice. | |
1096 | */ | |
1097 | CHAN_WARN_ON(chan, 1); | |
1098 | return -EBUSY; | |
1099 | } | |
f3bc08c5 | 1100 | retry: |
a8f2d0c7 | 1101 | finalized = READ_ONCE(buf->finalized); |
f3bc08c5 MD |
1102 | /* |
1103 | * Read finalized before counters. | |
1104 | */ | |
1105 | smp_rmb(); | |
1106 | consumed_cur = atomic_long_read(&buf->consumed); | |
1107 | consumed_idx = subbuf_index(consumed, chan); | |
1108 | commit_count = v_read(config, &buf->commit_cold[consumed_idx].cc_sb); | |
1109 | /* | |
1110 | * Make sure we read the commit count before reading the buffer | |
1111 | * data and the write offset. Correct consumed offset ordering | |
1112 | * wrt commit count is insured by the use of cmpxchg to update | |
1113 | * the consumed offset. | |
1114 | * smp_call_function_single can fail if the remote CPU is offline, | |
1115 | * this is OK because then there is no wmb to execute there. | |
1116 | * If our thread is executing on the same CPU as the on the buffers | |
1117 | * belongs to, we don't have to synchronize it at all. If we are | |
1118 | * migrated, the scheduler will take care of the memory barriers. | |
1119 | * Normally, smp_call_function_single() should ensure program order when | |
1120 | * executing the remote function, which implies that it surrounds the | |
1121 | * function execution with : | |
1122 | * smp_mb() | |
1123 | * send IPI | |
1124 | * csd_lock_wait | |
1125 | * recv IPI | |
1126 | * smp_mb() | |
1127 | * exec. function | |
1128 | * smp_mb() | |
1129 | * csd unlock | |
1130 | * smp_mb() | |
1131 | * | |
1132 | * However, smp_call_function_single() does not seem to clearly execute | |
1133 | * such barriers. It depends on spinlock semantic to provide the barrier | |
1134 | * before executing the IPI and, when busy-looping, csd_lock_wait only | |
1135 | * executes smp_mb() when it has to wait for the other CPU. | |
1136 | * | |
1137 | * I don't trust this code. Therefore, let's add the smp_mb() sequence | |
1138 | * required ourself, even if duplicated. It has no performance impact | |
1139 | * anyway. | |
1140 | * | |
1141 | * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs | |
1142 | * read and write vs write. They do not ensure core synchronization. We | |
1143 | * really have to ensure total order between the 3 barriers running on | |
1144 | * the 2 CPUs. | |
1145 | */ | |
1146 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1147 | if (config->sync == RING_BUFFER_SYNC_PER_CPU | |
1148 | && config->alloc == RING_BUFFER_ALLOC_PER_CPU) { | |
1149 | if (raw_smp_processor_id() != buf->backend.cpu) { | |
1150 | /* Total order with IPI handler smp_mb() */ | |
1151 | smp_mb(); | |
1152 | smp_call_function_single(buf->backend.cpu, | |
1153 | remote_mb, NULL, 1); | |
1154 | /* Total order with IPI handler smp_mb() */ | |
1155 | smp_mb(); | |
1156 | } | |
1157 | } else { | |
1158 | /* Total order with IPI handler smp_mb() */ | |
1159 | smp_mb(); | |
1160 | smp_call_function(remote_mb, NULL, 1); | |
1161 | /* Total order with IPI handler smp_mb() */ | |
1162 | smp_mb(); | |
1163 | } | |
1164 | } else { | |
1165 | /* | |
1166 | * Local rmb to match the remote wmb to read the commit count | |
1167 | * before the buffer data and the write offset. | |
1168 | */ | |
1169 | smp_rmb(); | |
1170 | } | |
1171 | ||
1172 | write_offset = v_read(config, &buf->offset); | |
1173 | ||
1174 | /* | |
1175 | * Check that the buffer we are getting is after or at consumed_cur | |
1176 | * position. | |
1177 | */ | |
1178 | if ((long) subbuf_trunc(consumed, chan) | |
1179 | - (long) subbuf_trunc(consumed_cur, chan) < 0) | |
1180 | goto nodata; | |
1181 | ||
1182 | /* | |
1183 | * Check that the subbuffer we are trying to consume has been | |
1184 | * already fully committed. | |
1185 | */ | |
1186 | if (((commit_count - chan->backend.subbuf_size) | |
1187 | & chan->commit_count_mask) | |
c9b3b5e2 | 1188 | - (buf_trunc(consumed, chan) |
f3bc08c5 MD |
1189 | >> chan->backend.num_subbuf_order) |
1190 | != 0) | |
1191 | goto nodata; | |
1192 | ||
1193 | /* | |
1194 | * Check that we are not about to read the same subbuffer in | |
1195 | * which the writer head is. | |
1196 | */ | |
c9b3b5e2 | 1197 | if (subbuf_trunc(write_offset, chan) - subbuf_trunc(consumed, chan) |
f3bc08c5 MD |
1198 | == 0) |
1199 | goto nodata; | |
1200 | ||
1201 | /* | |
1202 | * Failure to get the subbuffer causes a busy-loop retry without going | |
1203 | * to a wait queue. These are caused by short-lived race windows where | |
1204 | * the writer is getting access to a subbuffer we were trying to get | |
1205 | * access to. Also checks that the "consumed" buffer count we are | |
1206 | * looking for matches the one contained in the subbuffer id. | |
1207 | */ | |
1208 | ret = update_read_sb_index(config, &buf->backend, &chan->backend, | |
1209 | consumed_idx, buf_trunc_val(consumed, chan)); | |
1210 | if (ret) | |
1211 | goto retry; | |
1212 | subbuffer_id_clear_noref(config, &buf->backend.buf_rsb.id); | |
1213 | ||
1214 | buf->get_subbuf_consumed = consumed; | |
1215 | buf->get_subbuf = 1; | |
1216 | ||
155b8d9b MD |
1217 | lib_ring_buffer_flush_read_subbuf_dcache(config, chan, buf); |
1218 | ||
f3bc08c5 MD |
1219 | return 0; |
1220 | ||
1221 | nodata: | |
1222 | /* | |
1223 | * The memory barriers __wait_event()/wake_up_interruptible() take care | |
1224 | * of "raw_spin_is_locked" memory ordering. | |
1225 | */ | |
1226 | if (finalized) | |
1227 | return -ENODATA; | |
1228 | else if (raw_spin_is_locked(&buf->raw_tick_nohz_spinlock)) | |
1229 | goto retry; | |
1230 | else | |
1231 | return -EAGAIN; | |
1232 | } | |
1233 | EXPORT_SYMBOL_GPL(lib_ring_buffer_get_subbuf); | |
1234 | ||
1235 | /** | |
1236 | * lib_ring_buffer_put_subbuf - release exclusive subbuffer access | |
1237 | * @buf: ring buffer | |
1238 | */ | |
1239 | void lib_ring_buffer_put_subbuf(struct lib_ring_buffer *buf) | |
1240 | { | |
1241 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1242 | struct channel *chan = bufb->chan; | |
5a8fd222 | 1243 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1244 | unsigned long read_sb_bindex, consumed_idx, consumed; |
1245 | ||
1246 | CHAN_WARN_ON(chan, atomic_long_read(&buf->active_readers) != 1); | |
1247 | ||
1248 | if (!buf->get_subbuf) { | |
1249 | /* | |
1250 | * Reader puts a subbuffer it did not get. | |
1251 | */ | |
1252 | CHAN_WARN_ON(chan, 1); | |
1253 | return; | |
1254 | } | |
1255 | consumed = buf->get_subbuf_consumed; | |
1256 | buf->get_subbuf = 0; | |
1257 | ||
1258 | /* | |
1259 | * Clear the records_unread counter. (overruns counter) | |
1260 | * Can still be non-zero if a file reader simply grabbed the data | |
1261 | * without using iterators. | |
1262 | * Can be below zero if an iterator is used on a snapshot more than | |
1263 | * once. | |
1264 | */ | |
1265 | read_sb_bindex = subbuffer_id_get_index(config, bufb->buf_rsb.id); | |
1266 | v_add(config, v_read(config, | |
1267 | &bufb->array[read_sb_bindex]->records_unread), | |
1268 | &bufb->records_read); | |
1269 | v_set(config, &bufb->array[read_sb_bindex]->records_unread, 0); | |
1270 | CHAN_WARN_ON(chan, config->mode == RING_BUFFER_OVERWRITE | |
1271 | && subbuffer_id_is_noref(config, bufb->buf_rsb.id)); | |
1272 | subbuffer_id_set_noref(config, &bufb->buf_rsb.id); | |
1273 | ||
1274 | /* | |
1275 | * Exchange the reader subbuffer with the one we put in its place in the | |
1276 | * writer subbuffer table. Expect the original consumed count. If | |
1277 | * update_read_sb_index fails, this is because the writer updated the | |
1278 | * subbuffer concurrently. We should therefore keep the subbuffer we | |
1279 | * currently have: it has become invalid to try reading this sub-buffer | |
1280 | * consumed count value anyway. | |
1281 | */ | |
1282 | consumed_idx = subbuf_index(consumed, chan); | |
1283 | update_read_sb_index(config, &buf->backend, &chan->backend, | |
1284 | consumed_idx, buf_trunc_val(consumed, chan)); | |
1285 | /* | |
1286 | * update_read_sb_index return value ignored. Don't exchange sub-buffer | |
1287 | * if the writer concurrently updated it. | |
1288 | */ | |
1289 | } | |
1290 | EXPORT_SYMBOL_GPL(lib_ring_buffer_put_subbuf); | |
1291 | ||
1292 | /* | |
1293 | * cons_offset is an iterator on all subbuffer offsets between the reader | |
1294 | * position and the writer position. (inclusive) | |
1295 | */ | |
1296 | static | |
1297 | void lib_ring_buffer_print_subbuffer_errors(struct lib_ring_buffer *buf, | |
1298 | struct channel *chan, | |
1299 | unsigned long cons_offset, | |
1300 | int cpu) | |
1301 | { | |
5a8fd222 | 1302 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1303 | unsigned long cons_idx, commit_count, commit_count_sb; |
1304 | ||
1305 | cons_idx = subbuf_index(cons_offset, chan); | |
1306 | commit_count = v_read(config, &buf->commit_hot[cons_idx].cc); | |
1307 | commit_count_sb = v_read(config, &buf->commit_cold[cons_idx].cc_sb); | |
1308 | ||
1309 | if (subbuf_offset(commit_count, chan) != 0) | |
1310 | printk(KERN_WARNING | |
1311 | "ring buffer %s, cpu %d: " | |
1312 | "commit count in subbuffer %lu,\n" | |
1313 | "expecting multiples of %lu bytes\n" | |
1314 | " [ %lu bytes committed, %lu bytes reader-visible ]\n", | |
1315 | chan->backend.name, cpu, cons_idx, | |
1316 | chan->backend.subbuf_size, | |
1317 | commit_count, commit_count_sb); | |
1318 | ||
1319 | printk(KERN_DEBUG "ring buffer: %s, cpu %d: %lu bytes committed\n", | |
1320 | chan->backend.name, cpu, commit_count); | |
1321 | } | |
1322 | ||
1323 | static | |
1324 | void lib_ring_buffer_print_buffer_errors(struct lib_ring_buffer *buf, | |
1325 | struct channel *chan, | |
1326 | void *priv, int cpu) | |
1327 | { | |
5a8fd222 | 1328 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1329 | unsigned long write_offset, cons_offset; |
1330 | ||
f3bc08c5 MD |
1331 | /* |
1332 | * No need to order commit_count, write_offset and cons_offset reads | |
1333 | * because we execute at teardown when no more writer nor reader | |
1334 | * references are left. | |
1335 | */ | |
1336 | write_offset = v_read(config, &buf->offset); | |
1337 | cons_offset = atomic_long_read(&buf->consumed); | |
1338 | if (write_offset != cons_offset) | |
05aad775 | 1339 | printk(KERN_DEBUG |
f3bc08c5 MD |
1340 | "ring buffer %s, cpu %d: " |
1341 | "non-consumed data\n" | |
1342 | " [ %lu bytes written, %lu bytes read ]\n", | |
1343 | chan->backend.name, cpu, write_offset, cons_offset); | |
1344 | ||
1345 | for (cons_offset = atomic_long_read(&buf->consumed); | |
1346 | (long) (subbuf_trunc((unsigned long) v_read(config, &buf->offset), | |
1347 | chan) | |
1348 | - cons_offset) > 0; | |
1349 | cons_offset = subbuf_align(cons_offset, chan)) | |
1350 | lib_ring_buffer_print_subbuffer_errors(buf, chan, cons_offset, | |
1351 | cpu); | |
1352 | } | |
1353 | ||
97df3a7f | 1354 | #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS |
f3bc08c5 | 1355 | static |
97df3a7f JG |
1356 | void lib_ring_buffer_print_records_count(struct channel *chan, |
1357 | struct lib_ring_buffer *buf, | |
1358 | int cpu) | |
f3bc08c5 | 1359 | { |
5a8fd222 | 1360 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 | 1361 | |
ec01ec93 MD |
1362 | if (!strcmp(chan->backend.name, "relay-metadata")) { |
1363 | printk(KERN_DEBUG "ring buffer %s: %lu records written, " | |
1364 | "%lu records overrun\n", | |
1365 | chan->backend.name, | |
1366 | v_read(config, &buf->records_count), | |
1367 | v_read(config, &buf->records_overrun)); | |
1368 | } else { | |
1369 | printk(KERN_DEBUG "ring buffer %s, cpu %d: %lu records written, " | |
1370 | "%lu records overrun\n", | |
1371 | chan->backend.name, cpu, | |
1372 | v_read(config, &buf->records_count), | |
1373 | v_read(config, &buf->records_overrun)); | |
97df3a7f JG |
1374 | } |
1375 | } | |
1376 | #else | |
1377 | static | |
1378 | void lib_ring_buffer_print_records_count(struct channel *chan, | |
1379 | struct lib_ring_buffer *buf, | |
1380 | int cpu) | |
1381 | { | |
1382 | } | |
1383 | #endif | |
1384 | ||
1385 | static | |
1386 | void lib_ring_buffer_print_errors(struct channel *chan, | |
1387 | struct lib_ring_buffer *buf, int cpu) | |
1388 | { | |
1389 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1390 | void *priv = chan->backend.priv; | |
ec01ec93 | 1391 | |
97df3a7f JG |
1392 | lib_ring_buffer_print_records_count(chan, buf, cpu); |
1393 | if (strcmp(chan->backend.name, "relay-metadata")) { | |
ec01ec93 MD |
1394 | if (v_read(config, &buf->records_lost_full) |
1395 | || v_read(config, &buf->records_lost_wrap) | |
1396 | || v_read(config, &buf->records_lost_big)) | |
1397 | printk(KERN_WARNING | |
1398 | "ring buffer %s, cpu %d: records were lost. Caused by:\n" | |
1399 | " [ %lu buffer full, %lu nest buffer wrap-around, " | |
1400 | "%lu event too big ]\n", | |
1401 | chan->backend.name, cpu, | |
1402 | v_read(config, &buf->records_lost_full), | |
1403 | v_read(config, &buf->records_lost_wrap), | |
1404 | v_read(config, &buf->records_lost_big)); | |
1405 | } | |
f3bc08c5 MD |
1406 | lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu); |
1407 | } | |
1408 | ||
1409 | /* | |
1410 | * lib_ring_buffer_switch_old_start: Populate old subbuffer header. | |
1411 | * | |
6140ad92 | 1412 | * Only executed when the buffer is finalized, in SWITCH_FLUSH. |
f3bc08c5 MD |
1413 | */ |
1414 | static | |
1415 | void lib_ring_buffer_switch_old_start(struct lib_ring_buffer *buf, | |
1416 | struct channel *chan, | |
1417 | struct switch_offsets *offsets, | |
1418 | u64 tsc) | |
1419 | { | |
5a8fd222 | 1420 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1421 | unsigned long oldidx = subbuf_index(offsets->old, chan); |
1422 | unsigned long commit_count; | |
fbe84fd7 | 1423 | struct commit_counters_hot *cc_hot; |
f3bc08c5 MD |
1424 | |
1425 | config->cb.buffer_begin(buf, tsc, oldidx); | |
1426 | ||
1427 | /* | |
1428 | * Order all writes to buffer before the commit count update that will | |
1429 | * determine that the subbuffer is full. | |
1430 | */ | |
1431 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1432 | /* | |
1433 | * Must write slot data before incrementing commit count. This | |
1434 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1435 | * by get_subbuf(). | |
1436 | */ | |
1437 | barrier(); | |
1438 | } else | |
1439 | smp_wmb(); | |
fbe84fd7 MD |
1440 | cc_hot = &buf->commit_hot[oldidx]; |
1441 | v_add(config, config->cb.subbuffer_header_size(), &cc_hot->cc); | |
1442 | commit_count = v_read(config, &cc_hot->cc); | |
f3bc08c5 MD |
1443 | /* Check if the written buffer has to be delivered */ |
1444 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old, | |
635e457c | 1445 | commit_count, oldidx, tsc); |
8ec496cf | 1446 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
7915e163 | 1447 | offsets->old + config->cb.subbuffer_header_size(), |
fbe84fd7 | 1448 | commit_count, cc_hot); |
f3bc08c5 MD |
1449 | } |
1450 | ||
1451 | /* | |
1452 | * lib_ring_buffer_switch_old_end: switch old subbuffer | |
1453 | * | |
1454 | * Note : offset_old should never be 0 here. It is ok, because we never perform | |
1455 | * buffer switch on an empty subbuffer in SWITCH_ACTIVE mode. The caller | |
1456 | * increments the offset_old value when doing a SWITCH_FLUSH on an empty | |
1457 | * subbuffer. | |
1458 | */ | |
1459 | static | |
1460 | void lib_ring_buffer_switch_old_end(struct lib_ring_buffer *buf, | |
1461 | struct channel *chan, | |
1462 | struct switch_offsets *offsets, | |
1463 | u64 tsc) | |
1464 | { | |
5a8fd222 | 1465 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1466 | unsigned long oldidx = subbuf_index(offsets->old - 1, chan); |
1467 | unsigned long commit_count, padding_size, data_size; | |
fbe84fd7 | 1468 | struct commit_counters_hot *cc_hot; |
2485a430 | 1469 | u64 *ts_end; |
f3bc08c5 MD |
1470 | |
1471 | data_size = subbuf_offset(offsets->old - 1, chan) + 1; | |
1472 | padding_size = chan->backend.subbuf_size - data_size; | |
1473 | subbuffer_set_data_size(config, &buf->backend, oldidx, data_size); | |
1474 | ||
2485a430 | 1475 | ts_end = &buf->ts_end[oldidx]; |
f3bc08c5 | 1476 | /* |
2485a430 MD |
1477 | * This is the last space reservation in that sub-buffer before |
1478 | * it gets delivered. This provides exclusive access to write to | |
1479 | * this sub-buffer's ts_end. There are also no concurrent | |
1480 | * readers of that ts_end because delivery of that sub-buffer is | |
1481 | * postponed until the commit counter is incremented for the | |
1482 | * current space reservation. | |
1483 | */ | |
1484 | *ts_end = tsc; | |
1485 | ||
1486 | /* | |
1487 | * Order all writes to buffer and store to ts_end before the commit | |
1488 | * count update that will determine that the subbuffer is full. | |
f3bc08c5 MD |
1489 | */ |
1490 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1491 | /* | |
1492 | * Must write slot data before incrementing commit count. This | |
1493 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1494 | * by get_subbuf(). | |
1495 | */ | |
1496 | barrier(); | |
1497 | } else | |
1498 | smp_wmb(); | |
fbe84fd7 MD |
1499 | cc_hot = &buf->commit_hot[oldidx]; |
1500 | v_add(config, padding_size, &cc_hot->cc); | |
1501 | commit_count = v_read(config, &cc_hot->cc); | |
f3bc08c5 | 1502 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->old - 1, |
635e457c | 1503 | commit_count, oldidx, tsc); |
8ec496cf MD |
1504 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
1505 | offsets->old + padding_size, commit_count, | |
fbe84fd7 | 1506 | cc_hot); |
f3bc08c5 MD |
1507 | } |
1508 | ||
1509 | /* | |
1510 | * lib_ring_buffer_switch_new_start: Populate new subbuffer. | |
1511 | * | |
1512 | * This code can be executed unordered : writers may already have written to the | |
1513 | * sub-buffer before this code gets executed, caution. The commit makes sure | |
1514 | * that this code is executed before the deliver of this sub-buffer. | |
1515 | */ | |
1516 | static | |
1517 | void lib_ring_buffer_switch_new_start(struct lib_ring_buffer *buf, | |
1518 | struct channel *chan, | |
1519 | struct switch_offsets *offsets, | |
1520 | u64 tsc) | |
1521 | { | |
5a8fd222 | 1522 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1523 | unsigned long beginidx = subbuf_index(offsets->begin, chan); |
1524 | unsigned long commit_count; | |
fbe84fd7 | 1525 | struct commit_counters_hot *cc_hot; |
f3bc08c5 MD |
1526 | |
1527 | config->cb.buffer_begin(buf, tsc, beginidx); | |
1528 | ||
1529 | /* | |
1530 | * Order all writes to buffer before the commit count update that will | |
1531 | * determine that the subbuffer is full. | |
1532 | */ | |
1533 | if (config->ipi == RING_BUFFER_IPI_BARRIER) { | |
1534 | /* | |
1535 | * Must write slot data before incrementing commit count. This | |
1536 | * compiler barrier is upgraded into a smp_mb() by the IPI sent | |
1537 | * by get_subbuf(). | |
1538 | */ | |
1539 | barrier(); | |
1540 | } else | |
1541 | smp_wmb(); | |
fbe84fd7 MD |
1542 | cc_hot = &buf->commit_hot[beginidx]; |
1543 | v_add(config, config->cb.subbuffer_header_size(), &cc_hot->cc); | |
1544 | commit_count = v_read(config, &cc_hot->cc); | |
f3bc08c5 MD |
1545 | /* Check if the written buffer has to be delivered */ |
1546 | lib_ring_buffer_check_deliver(config, buf, chan, offsets->begin, | |
635e457c | 1547 | commit_count, beginidx, tsc); |
8ec496cf | 1548 | lib_ring_buffer_write_commit_counter(config, buf, chan, |
7915e163 | 1549 | offsets->begin + config->cb.subbuffer_header_size(), |
fbe84fd7 | 1550 | commit_count, cc_hot); |
f3bc08c5 MD |
1551 | } |
1552 | ||
f5ea5800 MD |
1553 | /* |
1554 | * lib_ring_buffer_switch_new_end: finish switching current subbuffer | |
1555 | * | |
768b05c9 MD |
1556 | * Calls subbuffer_set_data_size() to set the data size of the current |
1557 | * sub-buffer. We do not need to perform check_deliver nor commit here, | |
1558 | * since this task will be done by the "commit" of the event for which | |
1559 | * we are currently doing the space reservation. | |
f5ea5800 MD |
1560 | */ |
1561 | static | |
1562 | void lib_ring_buffer_switch_new_end(struct lib_ring_buffer *buf, | |
1563 | struct channel *chan, | |
1564 | struct switch_offsets *offsets, | |
1565 | u64 tsc) | |
1566 | { | |
1567 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
768b05c9 | 1568 | unsigned long endidx, data_size; |
2485a430 | 1569 | u64 *ts_end; |
f5ea5800 | 1570 | |
768b05c9 | 1571 | endidx = subbuf_index(offsets->end - 1, chan); |
f5ea5800 | 1572 | data_size = subbuf_offset(offsets->end - 1, chan) + 1; |
f5ea5800 | 1573 | subbuffer_set_data_size(config, &buf->backend, endidx, data_size); |
2485a430 MD |
1574 | ts_end = &buf->ts_end[endidx]; |
1575 | /* | |
1576 | * This is the last space reservation in that sub-buffer before | |
1577 | * it gets delivered. This provides exclusive access to write to | |
1578 | * this sub-buffer's ts_end. There are also no concurrent | |
1579 | * readers of that ts_end because delivery of that sub-buffer is | |
1580 | * postponed until the commit counter is incremented for the | |
1581 | * current space reservation. | |
1582 | */ | |
1583 | *ts_end = tsc; | |
f5ea5800 MD |
1584 | } |
1585 | ||
f3bc08c5 MD |
1586 | /* |
1587 | * Returns : | |
1588 | * 0 if ok | |
1589 | * !0 if execution must be aborted. | |
1590 | */ | |
1591 | static | |
1592 | int lib_ring_buffer_try_switch_slow(enum switch_mode mode, | |
1593 | struct lib_ring_buffer *buf, | |
1594 | struct channel *chan, | |
1595 | struct switch_offsets *offsets, | |
1596 | u64 *tsc) | |
1597 | { | |
5a8fd222 | 1598 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
5334a2c5 | 1599 | unsigned long off, reserve_commit_diff; |
f3bc08c5 MD |
1600 | |
1601 | offsets->begin = v_read(config, &buf->offset); | |
1602 | offsets->old = offsets->begin; | |
1603 | offsets->switch_old_start = 0; | |
1604 | off = subbuf_offset(offsets->begin, chan); | |
1605 | ||
1606 | *tsc = config->cb.ring_buffer_clock_read(chan); | |
1607 | ||
1608 | /* | |
1609 | * Ensure we flush the header of an empty subbuffer when doing the | |
1610 | * finalize (SWITCH_FLUSH). This ensures that we end up knowing the | |
1611 | * total data gathering duration even if there were no records saved | |
1612 | * after the last buffer switch. | |
1613 | * In SWITCH_ACTIVE mode, switch the buffer when it contains events. | |
1614 | * SWITCH_ACTIVE only flushes the current subbuffer, dealing with end of | |
1615 | * subbuffer header as appropriate. | |
1616 | * The next record that reserves space will be responsible for | |
1617 | * populating the following subbuffer header. We choose not to populate | |
1618 | * the next subbuffer header here because we want to be able to use | |
1619 | * SWITCH_ACTIVE for periodical buffer flush and CPU tick_nohz stop | |
1620 | * buffer flush, which must guarantee that all the buffer content | |
1621 | * (records and header timestamps) are visible to the reader. This is | |
1622 | * required for quiescence guarantees for the fusion merge. | |
1623 | */ | |
5334a2c5 MD |
1624 | if (mode != SWITCH_FLUSH && !off) |
1625 | return -1; /* we do not have to switch : buffer is empty */ | |
1626 | ||
1627 | if (unlikely(off == 0)) { | |
1628 | unsigned long sb_index, commit_count; | |
1629 | ||
1630 | /* | |
6140ad92 MD |
1631 | * We are performing a SWITCH_FLUSH. At this stage, there are no |
1632 | * concurrent writes into the buffer. | |
5334a2c5 | 1633 | * |
6140ad92 MD |
1634 | * The client does not save any header information. Don't |
1635 | * switch empty subbuffer on finalize, because it is invalid to | |
1636 | * deliver a completely empty subbuffer. | |
5334a2c5 MD |
1637 | */ |
1638 | if (!config->cb.subbuffer_header_size()) | |
1639 | return -1; | |
1640 | ||
1641 | /* Test new buffer integrity */ | |
1642 | sb_index = subbuf_index(offsets->begin, chan); | |
1643 | commit_count = v_read(config, | |
1644 | &buf->commit_cold[sb_index].cc_sb); | |
1645 | reserve_commit_diff = | |
1646 | (buf_trunc(offsets->begin, chan) | |
1647 | >> chan->backend.num_subbuf_order) | |
1648 | - (commit_count & chan->commit_count_mask); | |
1649 | if (likely(reserve_commit_diff == 0)) { | |
1650 | /* Next subbuffer not being written to. */ | |
1651 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1652 | subbuf_trunc(offsets->begin, chan) | |
1653 | - subbuf_trunc((unsigned long) | |
1654 | atomic_long_read(&buf->consumed), chan) | |
1655 | >= chan->backend.buf_size)) { | |
1656 | /* | |
1657 | * We do not overwrite non consumed buffers | |
1658 | * and we are full : don't switch. | |
1659 | */ | |
f3bc08c5 | 1660 | return -1; |
5334a2c5 MD |
1661 | } else { |
1662 | /* | |
1663 | * Next subbuffer not being written to, and we | |
1664 | * are either in overwrite mode or the buffer is | |
1665 | * not full. It's safe to write in this new | |
1666 | * subbuffer. | |
1667 | */ | |
1668 | } | |
1669 | } else { | |
f3bc08c5 | 1670 | /* |
5334a2c5 MD |
1671 | * Next subbuffer reserve offset does not match the |
1672 | * commit offset. Don't perform switch in | |
1673 | * producer-consumer and overwrite mode. Caused by | |
1674 | * either a writer OOPS or too many nested writes over a | |
1675 | * reserve/commit pair. | |
f3bc08c5 | 1676 | */ |
5334a2c5 | 1677 | return -1; |
f3bc08c5 | 1678 | } |
5334a2c5 MD |
1679 | |
1680 | /* | |
1681 | * Need to write the subbuffer start header on finalize. | |
1682 | */ | |
1683 | offsets->switch_old_start = 1; | |
1684 | } | |
1685 | offsets->begin = subbuf_align(offsets->begin, chan); | |
f3bc08c5 MD |
1686 | /* Note: old points to the next subbuf at offset 0 */ |
1687 | offsets->end = offsets->begin; | |
1688 | return 0; | |
1689 | } | |
1690 | ||
1691 | /* | |
1692 | * Force a sub-buffer switch. This operation is completely reentrant : can be | |
1693 | * called while tracing is active with absolutely no lock held. | |
1694 | * | |
1695 | * Note, however, that as a v_cmpxchg is used for some atomic | |
1696 | * operations, this function must be called from the CPU which owns the buffer | |
1697 | * for a ACTIVE flush. | |
1698 | */ | |
1699 | void lib_ring_buffer_switch_slow(struct lib_ring_buffer *buf, enum switch_mode mode) | |
1700 | { | |
1701 | struct channel *chan = buf->backend.chan; | |
5a8fd222 | 1702 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
1703 | struct switch_offsets offsets; |
1704 | unsigned long oldidx; | |
1705 | u64 tsc; | |
1706 | ||
1707 | offsets.size = 0; | |
1708 | ||
1709 | /* | |
1710 | * Perform retryable operations. | |
1711 | */ | |
1712 | do { | |
1713 | if (lib_ring_buffer_try_switch_slow(mode, buf, chan, &offsets, | |
1714 | &tsc)) | |
1715 | return; /* Switch not needed */ | |
1716 | } while (v_cmpxchg(config, &buf->offset, offsets.old, offsets.end) | |
1717 | != offsets.old); | |
1718 | ||
1719 | /* | |
1720 | * Atomically update last_tsc. This update races against concurrent | |
1721 | * atomic updates, but the race will always cause supplementary full TSC | |
1722 | * records, never the opposite (missing a full TSC record when it would | |
1723 | * be needed). | |
1724 | */ | |
1725 | save_last_tsc(config, buf, tsc); | |
1726 | ||
1727 | /* | |
1728 | * Push the reader if necessary | |
1729 | */ | |
1730 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.old); | |
1731 | ||
1732 | oldidx = subbuf_index(offsets.old, chan); | |
1733 | lib_ring_buffer_clear_noref(config, &buf->backend, oldidx); | |
1734 | ||
1735 | /* | |
1736 | * May need to populate header start on SWITCH_FLUSH. | |
1737 | */ | |
1738 | if (offsets.switch_old_start) { | |
1739 | lib_ring_buffer_switch_old_start(buf, chan, &offsets, tsc); | |
1740 | offsets.old += config->cb.subbuffer_header_size(); | |
1741 | } | |
1742 | ||
1743 | /* | |
1744 | * Switch old subbuffer. | |
1745 | */ | |
1746 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, tsc); | |
1747 | } | |
1748 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_slow); | |
1749 | ||
9dded431 MD |
1750 | struct switch_param { |
1751 | struct lib_ring_buffer *buf; | |
1752 | enum switch_mode mode; | |
1753 | }; | |
1754 | ||
5e391252 MD |
1755 | static void remote_switch(void *info) |
1756 | { | |
9dded431 MD |
1757 | struct switch_param *param = info; |
1758 | struct lib_ring_buffer *buf = param->buf; | |
5e391252 | 1759 | |
9dded431 | 1760 | lib_ring_buffer_switch_slow(buf, param->mode); |
5e391252 MD |
1761 | } |
1762 | ||
64af2437 MD |
1763 | static void _lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf, |
1764 | enum switch_mode mode) | |
5e391252 MD |
1765 | { |
1766 | struct channel *chan = buf->backend.chan; | |
1767 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1768 | int ret; | |
9dded431 | 1769 | struct switch_param param; |
5e391252 MD |
1770 | |
1771 | /* | |
1772 | * With global synchronization we don't need to use the IPI scheme. | |
1773 | */ | |
1774 | if (config->sync == RING_BUFFER_SYNC_GLOBAL) { | |
64af2437 | 1775 | lib_ring_buffer_switch_slow(buf, mode); |
5e391252 MD |
1776 | return; |
1777 | } | |
1778 | ||
1779 | /* | |
f78cce4b | 1780 | * Disabling preemption ensures two things: first, that the |
5e391252 | 1781 | * target cpu is not taken concurrently offline while we are within |
f78cce4b MD |
1782 | * smp_call_function_single(). Secondly, if it happens that the |
1783 | * CPU is not online, our own call to lib_ring_buffer_switch_slow() | |
1784 | * needs to be protected from CPU hotplug handlers, which can | |
1785 | * also perform a remote subbuffer switch. | |
5e391252 | 1786 | */ |
f78cce4b | 1787 | preempt_disable(); |
9dded431 MD |
1788 | param.buf = buf; |
1789 | param.mode = mode; | |
5e391252 | 1790 | ret = smp_call_function_single(buf->backend.cpu, |
9dded431 | 1791 | remote_switch, ¶m, 1); |
5e391252 MD |
1792 | if (ret) { |
1793 | /* Remote CPU is offline, do it ourself. */ | |
64af2437 | 1794 | lib_ring_buffer_switch_slow(buf, mode); |
5e391252 | 1795 | } |
f78cce4b | 1796 | preempt_enable(); |
5e391252 | 1797 | } |
64af2437 | 1798 | |
c6f05468 | 1799 | /* Switch sub-buffer if current sub-buffer is non-empty. */ |
64af2437 MD |
1800 | void lib_ring_buffer_switch_remote(struct lib_ring_buffer *buf) |
1801 | { | |
1802 | _lib_ring_buffer_switch_remote(buf, SWITCH_ACTIVE); | |
1803 | } | |
5e391252 MD |
1804 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote); |
1805 | ||
c6f05468 MD |
1806 | /* Switch sub-buffer even if current sub-buffer is empty. */ |
1807 | void lib_ring_buffer_switch_remote_empty(struct lib_ring_buffer *buf) | |
1808 | { | |
1809 | _lib_ring_buffer_switch_remote(buf, SWITCH_FLUSH); | |
1810 | } | |
1811 | EXPORT_SYMBOL_GPL(lib_ring_buffer_switch_remote_empty); | |
1812 | ||
c245d0d3 MD |
1813 | void lib_ring_buffer_clear(struct lib_ring_buffer *buf) |
1814 | { | |
1815 | struct lib_ring_buffer_backend *bufb = &buf->backend; | |
1816 | struct channel *chan = bufb->chan; | |
1817 | ||
1818 | lib_ring_buffer_switch_remote(buf); | |
1819 | lib_ring_buffer_clear_reader(buf, chan); | |
1820 | } | |
1821 | EXPORT_SYMBOL_GPL(lib_ring_buffer_clear); | |
1822 | ||
f3bc08c5 MD |
1823 | /* |
1824 | * Returns : | |
1825 | * 0 if ok | |
97ca2c54 MD |
1826 | * -ENOSPC if event size is too large for packet. |
1827 | * -ENOBUFS if there is currently not enough space in buffer for the event. | |
1828 | * -EIO if data cannot be written into the buffer for any other reason. | |
f3bc08c5 MD |
1829 | */ |
1830 | static | |
1831 | int lib_ring_buffer_try_reserve_slow(struct lib_ring_buffer *buf, | |
1832 | struct channel *chan, | |
1833 | struct switch_offsets *offsets, | |
cc62f29e MD |
1834 | struct lib_ring_buffer_ctx *ctx, |
1835 | void *client_ctx) | |
f3bc08c5 | 1836 | { |
5a8fd222 | 1837 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
0fdec686 | 1838 | unsigned long reserve_commit_diff, offset_cmp; |
f3bc08c5 | 1839 | |
0fdec686 MD |
1840 | retry: |
1841 | offsets->begin = offset_cmp = v_read(config, &buf->offset); | |
f3bc08c5 MD |
1842 | offsets->old = offsets->begin; |
1843 | offsets->switch_new_start = 0; | |
f5ea5800 | 1844 | offsets->switch_new_end = 0; |
f3bc08c5 MD |
1845 | offsets->switch_old_end = 0; |
1846 | offsets->pre_header_padding = 0; | |
1847 | ||
1848 | ctx->tsc = config->cb.ring_buffer_clock_read(chan); | |
97ca2c54 MD |
1849 | if ((int64_t) ctx->tsc == -EIO) |
1850 | return -EIO; | |
f3bc08c5 MD |
1851 | |
1852 | if (last_tsc_overflow(config, buf, ctx->tsc)) | |
64c796d8 | 1853 | ctx->rflags |= RING_BUFFER_RFLAG_FULL_TSC; |
f3bc08c5 MD |
1854 | |
1855 | if (unlikely(subbuf_offset(offsets->begin, ctx->chan) == 0)) { | |
1856 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1857 | } else { | |
1858 | offsets->size = config->cb.record_header_size(config, chan, | |
1859 | offsets->begin, | |
f3bc08c5 | 1860 | &offsets->pre_header_padding, |
cc62f29e | 1861 | ctx, client_ctx); |
f3bc08c5 MD |
1862 | offsets->size += |
1863 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1864 | ctx->largest_align) | |
1865 | + ctx->data_size; | |
1866 | if (unlikely(subbuf_offset(offsets->begin, chan) + | |
1867 | offsets->size > chan->backend.subbuf_size)) { | |
1868 | offsets->switch_old_end = 1; /* For offsets->old */ | |
1869 | offsets->switch_new_start = 1; /* For offsets->begin */ | |
1870 | } | |
1871 | } | |
1872 | if (unlikely(offsets->switch_new_start)) { | |
0fdec686 | 1873 | unsigned long sb_index, commit_count; |
f3bc08c5 MD |
1874 | |
1875 | /* | |
1876 | * We are typically not filling the previous buffer completely. | |
1877 | */ | |
1878 | if (likely(offsets->switch_old_end)) | |
1879 | offsets->begin = subbuf_align(offsets->begin, chan); | |
1880 | offsets->begin = offsets->begin | |
1881 | + config->cb.subbuffer_header_size(); | |
1882 | /* Test new buffer integrity */ | |
1883 | sb_index = subbuf_index(offsets->begin, chan); | |
0fdec686 MD |
1884 | /* |
1885 | * Read buf->offset before buf->commit_cold[sb_index].cc_sb. | |
1886 | * lib_ring_buffer_check_deliver() has the matching | |
1887 | * memory barriers required around commit_cold cc_sb | |
1888 | * updates to ensure reserve and commit counter updates | |
1889 | * are not seen reordered when updated by another CPU. | |
1890 | */ | |
1891 | smp_rmb(); | |
1892 | commit_count = v_read(config, | |
1893 | &buf->commit_cold[sb_index].cc_sb); | |
1894 | /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */ | |
1895 | smp_rmb(); | |
1896 | if (unlikely(offset_cmp != v_read(config, &buf->offset))) { | |
1897 | /* | |
1898 | * The reserve counter have been concurrently updated | |
1899 | * while we read the commit counter. This means the | |
1900 | * commit counter we read might not match buf->offset | |
1901 | * due to concurrent update. We therefore need to retry. | |
1902 | */ | |
1903 | goto retry; | |
1904 | } | |
f3bc08c5 MD |
1905 | reserve_commit_diff = |
1906 | (buf_trunc(offsets->begin, chan) | |
1907 | >> chan->backend.num_subbuf_order) | |
0fdec686 | 1908 | - (commit_count & chan->commit_count_mask); |
f3bc08c5 MD |
1909 | if (likely(reserve_commit_diff == 0)) { |
1910 | /* Next subbuffer not being written to. */ | |
1911 | if (unlikely(config->mode != RING_BUFFER_OVERWRITE && | |
1912 | subbuf_trunc(offsets->begin, chan) | |
1913 | - subbuf_trunc((unsigned long) | |
1914 | atomic_long_read(&buf->consumed), chan) | |
1915 | >= chan->backend.buf_size)) { | |
1916 | /* | |
1917 | * We do not overwrite non consumed buffers | |
1918 | * and we are full : record is lost. | |
1919 | */ | |
1920 | v_inc(config, &buf->records_lost_full); | |
97ca2c54 | 1921 | return -ENOBUFS; |
f3bc08c5 MD |
1922 | } else { |
1923 | /* | |
1924 | * Next subbuffer not being written to, and we | |
1925 | * are either in overwrite mode or the buffer is | |
1926 | * not full. It's safe to write in this new | |
1927 | * subbuffer. | |
1928 | */ | |
1929 | } | |
1930 | } else { | |
1931 | /* | |
1932 | * Next subbuffer reserve offset does not match the | |
0fdec686 MD |
1933 | * commit offset, and this did not involve update to the |
1934 | * reserve counter. Drop record in producer-consumer and | |
1935 | * overwrite mode. Caused by either a writer OOPS or | |
1936 | * too many nested writes over a reserve/commit pair. | |
f3bc08c5 MD |
1937 | */ |
1938 | v_inc(config, &buf->records_lost_wrap); | |
97ca2c54 | 1939 | return -EIO; |
f3bc08c5 MD |
1940 | } |
1941 | offsets->size = | |
1942 | config->cb.record_header_size(config, chan, | |
1943 | offsets->begin, | |
f3bc08c5 | 1944 | &offsets->pre_header_padding, |
cc62f29e | 1945 | ctx, client_ctx); |
f3bc08c5 MD |
1946 | offsets->size += |
1947 | lib_ring_buffer_align(offsets->begin + offsets->size, | |
1948 | ctx->largest_align) | |
1949 | + ctx->data_size; | |
1950 | if (unlikely(subbuf_offset(offsets->begin, chan) | |
1951 | + offsets->size > chan->backend.subbuf_size)) { | |
1952 | /* | |
1953 | * Record too big for subbuffers, report error, don't | |
1954 | * complete the sub-buffer switch. | |
1955 | */ | |
1956 | v_inc(config, &buf->records_lost_big); | |
97ca2c54 | 1957 | return -ENOSPC; |
f3bc08c5 MD |
1958 | } else { |
1959 | /* | |
1960 | * We just made a successful buffer switch and the | |
1961 | * record fits in the new subbuffer. Let's write. | |
1962 | */ | |
1963 | } | |
1964 | } else { | |
1965 | /* | |
1966 | * Record fits in the current buffer and we are not on a switch | |
1967 | * boundary. It's safe to write. | |
1968 | */ | |
1969 | } | |
1970 | offsets->end = offsets->begin + offsets->size; | |
f5ea5800 MD |
1971 | |
1972 | if (unlikely(subbuf_offset(offsets->end, chan) == 0)) { | |
1973 | /* | |
1974 | * The offset_end will fall at the very beginning of the next | |
1975 | * subbuffer. | |
1976 | */ | |
1977 | offsets->switch_new_end = 1; /* For offsets->begin */ | |
1978 | } | |
f3bc08c5 MD |
1979 | return 0; |
1980 | } | |
1981 | ||
d7e74017 MD |
1982 | static struct lib_ring_buffer *get_current_buf(struct channel *chan, int cpu) |
1983 | { | |
1984 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1985 | ||
1986 | if (config->alloc == RING_BUFFER_ALLOC_PER_CPU) | |
1987 | return per_cpu_ptr(chan->backend.buf, cpu); | |
1988 | else | |
1989 | return chan->backend.buf; | |
1990 | } | |
1991 | ||
1992 | void lib_ring_buffer_lost_event_too_big(struct channel *chan) | |
1993 | { | |
1994 | const struct lib_ring_buffer_config *config = &chan->backend.config; | |
1995 | struct lib_ring_buffer *buf = get_current_buf(chan, smp_processor_id()); | |
1996 | ||
1997 | v_inc(config, &buf->records_lost_big); | |
1998 | } | |
1999 | EXPORT_SYMBOL_GPL(lib_ring_buffer_lost_event_too_big); | |
2000 | ||
f3bc08c5 MD |
2001 | /** |
2002 | * lib_ring_buffer_reserve_slow - Atomic slot reservation in a buffer. | |
2003 | * @ctx: ring buffer context. | |
2004 | * | |
97ca2c54 MD |
2005 | * Return : -NOBUFS if not enough space, -ENOSPC if event size too large, |
2006 | * -EIO for other errors, else returns 0. | |
f3bc08c5 MD |
2007 | * It will take care of sub-buffer switching. |
2008 | */ | |
cc62f29e MD |
2009 | int lib_ring_buffer_reserve_slow(struct lib_ring_buffer_ctx *ctx, |
2010 | void *client_ctx) | |
f3bc08c5 MD |
2011 | { |
2012 | struct channel *chan = ctx->chan; | |
5a8fd222 | 2013 | const struct lib_ring_buffer_config *config = &chan->backend.config; |
f3bc08c5 MD |
2014 | struct lib_ring_buffer *buf; |
2015 | struct switch_offsets offsets; | |
c099397a | 2016 | int ret; |
f3bc08c5 | 2017 | |
d7e74017 | 2018 | ctx->buf = buf = get_current_buf(chan, ctx->cpu); |
f3bc08c5 MD |
2019 | offsets.size = 0; |
2020 | ||
2021 | do { | |
97ca2c54 | 2022 | ret = lib_ring_buffer_try_reserve_slow(buf, chan, &offsets, |
cc62f29e | 2023 | ctx, client_ctx); |
97ca2c54 MD |
2024 | if (unlikely(ret)) |
2025 | return ret; | |
f3bc08c5 MD |
2026 | } while (unlikely(v_cmpxchg(config, &buf->offset, offsets.old, |
2027 | offsets.end) | |
2028 | != offsets.old)); | |
2029 | ||
2030 | /* | |
2031 | * Atomically update last_tsc. This update races against concurrent | |
2032 | * atomic updates, but the race will always cause supplementary full TSC | |
2033 | * records, never the opposite (missing a full TSC record when it would | |
2034 | * be needed). | |
2035 | */ | |
2036 | save_last_tsc(config, buf, ctx->tsc); | |
2037 | ||
2038 | /* | |
2039 | * Push the reader if necessary | |
2040 | */ | |
2041 | lib_ring_buffer_reserve_push_reader(buf, chan, offsets.end - 1); | |
2042 | ||
2043 | /* | |
2044 | * Clear noref flag for this subbuffer. | |
2045 | */ | |
2046 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
2047 | subbuf_index(offsets.end - 1, chan)); | |
2048 | ||
2049 | /* | |
2050 | * Switch old subbuffer if needed. | |
2051 | */ | |
2052 | if (unlikely(offsets.switch_old_end)) { | |
2053 | lib_ring_buffer_clear_noref(config, &buf->backend, | |
2054 | subbuf_index(offsets.old - 1, chan)); | |
2055 | lib_ring_buffer_switch_old_end(buf, chan, &offsets, ctx->tsc); | |
2056 | } | |
2057 | ||
2058 | /* | |
2059 | * Populate new subbuffer. | |
2060 | */ | |
2061 | if (unlikely(offsets.switch_new_start)) | |
2062 | lib_ring_buffer_switch_new_start(buf, chan, &offsets, ctx->tsc); | |
2063 | ||
f5ea5800 MD |
2064 | if (unlikely(offsets.switch_new_end)) |
2065 | lib_ring_buffer_switch_new_end(buf, chan, &offsets, ctx->tsc); | |
2066 | ||
f3bc08c5 MD |
2067 | ctx->slot_size = offsets.size; |
2068 | ctx->pre_offset = offsets.begin; | |
2069 | ctx->buf_offset = offsets.begin + offsets.pre_header_padding; | |
2070 | return 0; | |
2071 | } | |
2072 | EXPORT_SYMBOL_GPL(lib_ring_buffer_reserve_slow); | |
6fb8de4b | 2073 | |
aece661f MD |
2074 | static |
2075 | void lib_ring_buffer_vmcore_check_deliver(const struct lib_ring_buffer_config *config, | |
2076 | struct lib_ring_buffer *buf, | |
2077 | unsigned long commit_count, | |
2078 | unsigned long idx) | |
2079 | { | |
2080 | if (config->oops == RING_BUFFER_OOPS_CONSISTENCY) | |
2081 | v_set(config, &buf->commit_hot[idx].seq, commit_count); | |
2082 | } | |
2083 | ||
25337cb5 MD |
2084 | /* |
2085 | * The ring buffer can count events recorded and overwritten per buffer, | |
2086 | * but it is disabled by default due to its performance overhead. | |
2087 | */ | |
2088 | #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS | |
2089 | static | |
2090 | void deliver_count_events(const struct lib_ring_buffer_config *config, | |
2091 | struct lib_ring_buffer *buf, | |
2092 | unsigned long idx) | |
2093 | { | |
2094 | v_add(config, subbuffer_get_records_count(config, | |
2095 | &buf->backend, idx), | |
2096 | &buf->records_count); | |
2097 | v_add(config, subbuffer_count_records_overrun(config, | |
2098 | &buf->backend, idx), | |
2099 | &buf->records_overrun); | |
2100 | } | |
2101 | #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */ | |
2102 | static | |
2103 | void deliver_count_events(const struct lib_ring_buffer_config *config, | |
2104 | struct lib_ring_buffer *buf, | |
2105 | unsigned long idx) | |
2106 | { | |
2107 | } | |
2108 | #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */ | |
2109 | ||
2110 | ||
aece661f MD |
2111 | void lib_ring_buffer_check_deliver_slow(const struct lib_ring_buffer_config *config, |
2112 | struct lib_ring_buffer *buf, | |
2113 | struct channel *chan, | |
2114 | unsigned long offset, | |
2115 | unsigned long commit_count, | |
2116 | unsigned long idx, | |
2117 | u64 tsc) | |
2118 | { | |
2119 | unsigned long old_commit_count = commit_count | |
2120 | - chan->backend.subbuf_size; | |
2121 | ||
2122 | /* | |
2123 | * If we succeeded at updating cc_sb below, we are the subbuffer | |
2124 | * writer delivering the subbuffer. Deals with concurrent | |
2125 | * updates of the "cc" value without adding a add_return atomic | |
2126 | * operation to the fast path. | |
2127 | * | |
2128 | * We are doing the delivery in two steps: | |
2129 | * - First, we cmpxchg() cc_sb to the new value | |
2130 | * old_commit_count + 1. This ensures that we are the only | |
2131 | * subbuffer user successfully filling the subbuffer, but we | |
2132 | * do _not_ set the cc_sb value to "commit_count" yet. | |
2133 | * Therefore, other writers that would wrap around the ring | |
2134 | * buffer and try to start writing to our subbuffer would | |
2135 | * have to drop records, because it would appear as | |
2136 | * non-filled. | |
2137 | * We therefore have exclusive access to the subbuffer control | |
2138 | * structures. This mutual exclusion with other writers is | |
2139 | * crucially important to perform record overruns count in | |
2140 | * flight recorder mode locklessly. | |
2141 | * - When we are ready to release the subbuffer (either for | |
2142 | * reading or for overrun by other writers), we simply set the | |
2143 | * cc_sb value to "commit_count" and perform delivery. | |
2144 | * | |
2145 | * The subbuffer size is least 2 bytes (minimum size: 1 page). | |
2146 | * This guarantees that old_commit_count + 1 != commit_count. | |
2147 | */ | |
2148 | ||
2149 | /* | |
2150 | * Order prior updates to reserve count prior to the | |
2151 | * commit_cold cc_sb update. | |
2152 | */ | |
2153 | smp_wmb(); | |
2154 | if (likely(v_cmpxchg(config, &buf->commit_cold[idx].cc_sb, | |
2155 | old_commit_count, old_commit_count + 1) | |
2156 | == old_commit_count)) { | |
2485a430 MD |
2157 | u64 *ts_end; |
2158 | ||
aece661f MD |
2159 | /* |
2160 | * Start of exclusive subbuffer access. We are | |
2161 | * guaranteed to be the last writer in this subbuffer | |
2162 | * and any other writer trying to access this subbuffer | |
2163 | * in this state is required to drop records. | |
2485a430 MD |
2164 | * |
2165 | * We can read the ts_end for the current sub-buffer | |
2166 | * which has been saved by the very last space | |
2167 | * reservation for the current sub-buffer. | |
2168 | * | |
2169 | * Order increment of commit counter before reading ts_end. | |
aece661f | 2170 | */ |
2485a430 MD |
2171 | smp_mb(); |
2172 | ts_end = &buf->ts_end[idx]; | |
25337cb5 | 2173 | deliver_count_events(config, buf, idx); |
2485a430 | 2174 | config->cb.buffer_end(buf, *ts_end, idx, |
aece661f MD |
2175 | lib_ring_buffer_get_data_size(config, |
2176 | buf, | |
2177 | idx)); | |
2178 | ||
2179 | /* | |
2180 | * Increment the packet counter while we have exclusive | |
2181 | * access. | |
2182 | */ | |
2183 | subbuffer_inc_packet_count(config, &buf->backend, idx); | |
2184 | ||
2185 | /* | |
2186 | * Set noref flag and offset for this subbuffer id. | |
2187 | * Contains a memory barrier that ensures counter stores | |
2188 | * are ordered before set noref and offset. | |
2189 | */ | |
2190 | lib_ring_buffer_set_noref_offset(config, &buf->backend, idx, | |
2191 | buf_trunc_val(offset, chan)); | |
2192 | ||
2193 | /* | |
2194 | * Order set_noref and record counter updates before the | |
2195 | * end of subbuffer exclusive access. Orders with | |
2196 | * respect to writers coming into the subbuffer after | |
2197 | * wrap around, and also order wrt concurrent readers. | |
2198 | */ | |
2199 | smp_mb(); | |
2200 | /* End of exclusive subbuffer access */ | |
2201 | v_set(config, &buf->commit_cold[idx].cc_sb, | |
2202 | commit_count); | |
2203 | /* | |
2204 | * Order later updates to reserve count after | |
2205 | * the commit_cold cc_sb update. | |
2206 | */ | |
2207 | smp_wmb(); | |
2208 | lib_ring_buffer_vmcore_check_deliver(config, buf, | |
2209 | commit_count, idx); | |
2210 | ||
2211 | /* | |
2212 | * RING_BUFFER_WAKEUP_BY_WRITER wakeup is not lock-free. | |
2213 | */ | |
2214 | if (config->wakeup == RING_BUFFER_WAKEUP_BY_WRITER | |
2215 | && atomic_long_read(&buf->active_readers) | |
2216 | && lib_ring_buffer_poll_deliver(config, buf, chan)) { | |
2217 | wake_up_interruptible(&buf->read_wait); | |
2218 | wake_up_interruptible(&chan->read_wait); | |
2219 | } | |
2220 | ||
2221 | } | |
2222 | } | |
2223 | EXPORT_SYMBOL_GPL(lib_ring_buffer_check_deliver_slow); | |
2224 | ||
02a766bb | 2225 | int __init init_lib_ring_buffer_frontend(void) |
6fb8de4b MD |
2226 | { |
2227 | int cpu; | |
2228 | ||
2229 | for_each_possible_cpu(cpu) | |
2230 | spin_lock_init(&per_cpu(ring_buffer_nohz_lock, cpu)); | |
02a766bb | 2231 | return 0; |
6fb8de4b | 2232 | } |
02a766bb MD |
2233 | |
2234 | module_init(init_lib_ring_buffer_frontend); | |
1a5db82d MD |
2235 | |
2236 | void __exit exit_lib_ring_buffer_frontend(void) | |
2237 | { | |
2238 | } | |
2239 | ||
2240 | module_exit(exit_lib_ring_buffer_frontend); |