projects / ust.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
fix support for trace format 2.5
[ust.git] / libust / buffers.c
... / ...
CommitLineData
1/*
2 * buffers.c
3 * LTTng userspace tracer buffering system
4 *
5 * Copyright (C) 2009 - Pierre-Marc Fournier (pierre-marc dot fournier at polymtl dot ca)
6 * Copyright (C) 2008 - Mathieu Desnoyers (mathieu.desnoyers@polymtl.ca)
7 *
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23#include <unistd.h>
24#include <sys/mman.h>
25#include <sys/ipc.h>
26#include <sys/shm.h>
27#include <fcntl.h>
28#include <ust/kernelcompat.h>
29#include <kcompat/kref.h>
30#include "buffers.h"
31#include "channels.h"
32#include "tracer.h"
33#include "tracercore.h"
34#include "usterr.h"
35
36struct ltt_reserve_switch_offsets {
37 long begin, end, old;
38 long begin_switch, end_switch_current, end_switch_old;
39 size_t before_hdr_pad, size;
40};
41
42
43static DEFINE_MUTEX(ust_buffers_channels_mutex);
44static LIST_HEAD(ust_buffers_channels);
45
46static int get_n_cpus(void)
47{
48 int result;
49 static int n_cpus = 0;
50
51 if(n_cpus) {
52 return n_cpus;
53 }
54
55 /* On Linux, when some processors are offline
56 * _SC_NPROCESSORS_CONF counts the offline
57 * processors, whereas _SC_NPROCESSORS_ONLN
58 * does not. If we used _SC_NPROCESSORS_ONLN,
59 * getcpu() could return a value greater than
60 * this sysconf, in which case the arrays
61 * indexed by processor would overflow.
62 */
63 result = sysconf(_SC_NPROCESSORS_CONF);
64 if(result == -1) {
65 return -1;
66 }
67
68 n_cpus = result;
69
70 return result;
71}
72
73/* _ust_buffers_write()
74 *
75 * @buf: destination buffer
76 * @offset: offset in destination
77 * @src: source buffer
78 * @len: length of source
79 * @cpy: already copied
80 */
81
82void _ust_buffers_write(struct ust_buffer *buf, size_t offset,
83 const void *src, size_t len, ssize_t cpy)
84{
85 do {
86 len -= cpy;
87 src += cpy;
88 offset += cpy;
89
90 WARN_ON(offset >= buf->buf_size);
91
92 cpy = min_t(size_t, len, buf->buf_size - offset);
93 ust_buffers_do_copy(buf->buf_data + offset, src, cpy);
94 } while (unlikely(len != cpy));
95}
96
97static int ust_buffers_init_buffer(struct ust_trace *trace,
98 struct ust_channel *ltt_chan,
99 struct ust_buffer *buf,
100 unsigned int n_subbufs);
101
102static int ust_buffers_alloc_buf(struct ust_buffer *buf, size_t *size)
103{
104 void *ptr;
105 int result;
106
107 *size = PAGE_ALIGN(*size);
108
109 result = buf->shmid = shmget(getpid(), *size, IPC_CREAT | IPC_EXCL | 0700);
110 if(result == -1 && errno == EINVAL) {
111 ERR("shmget() returned EINVAL; maybe /proc/sys/kernel/shmmax should be increased.");
112 return -1;
113 }
114 else if(result == -1) {
115 PERROR("shmget");
116 return -1;
117 }
118
119 /* FIXME: should have matching call to shmdt */
120 ptr = shmat(buf->shmid, NULL, 0);
121 if(ptr == (void *) -1) {
122 perror("shmat");
123 goto destroy_shmem;
124 }
125
126 /* Already mark the shared memory for destruction. This will occur only
127 * when all users have detached.
128 */
129 result = shmctl(buf->shmid, IPC_RMID, NULL);
130 if(result == -1) {
131 perror("shmctl");
132 return -1;
133 }
134
135 buf->buf_data = ptr;
136 buf->buf_size = *size;
137
138 return 0;
139
140 destroy_shmem:
141 result = shmctl(buf->shmid, IPC_RMID, NULL);
142 if(result == -1) {
143 perror("shmctl");
144 }
145
146 return -1;
147}
148
149int ust_buffers_create_buf(struct ust_channel *channel, int cpu)
150{
151 int result;
152 struct ust_buffer *buf = channel->buf[cpu];
153
154 buf->cpu = cpu;
155 result = ust_buffers_alloc_buf(buf, &channel->alloc_size);
156 if(result)
157 return -1;
158
159 buf->chan = channel;
160 kref_get(&channel->kref);
161 return 0;
162}
163
164static void ust_buffers_destroy_channel(struct kref *kref)
165{
166 struct ust_channel *chan = container_of(kref, struct ust_channel, kref);
167 free(chan);
168}
169
170static void ust_buffers_destroy_buf(struct ust_buffer *buf)
171{
172 struct ust_channel *chan = buf->chan;
173 int result;
174
175 result = munmap(buf->buf_data, buf->buf_size);
176 if(result == -1) {
177 PERROR("munmap");
178 }
179
180//ust// chan->buf[buf->cpu] = NULL;
181 free(buf);
182 kref_put(&chan->kref, ust_buffers_destroy_channel);
183}
184
185/* called from kref_put */
186static void ust_buffers_remove_buf(struct kref *kref)
187{
188 struct ust_buffer *buf = container_of(kref, struct ust_buffer, kref);
189 ust_buffers_destroy_buf(buf);
190}
191
192int ust_buffers_open_buf(struct ust_channel *chan, int cpu)
193{
194 int result;
195
196 result = ust_buffers_create_buf(chan, cpu);
197 if (result == -1)
198 return -1;
199
200 kref_init(&chan->buf[cpu]->kref);
201
202 result = ust_buffers_init_buffer(chan->trace, chan, chan->buf[cpu], chan->subbuf_cnt);
203 if(result == -1)
204 return -1;
205
206 return 0;
207
208 /* FIXME: decrementally destroy on error? */
209}
210
211/**
212 * ust_buffers_close_buf - close a channel buffer
213 * @buf: buffer
214 */
215static void ust_buffers_close_buf(struct ust_buffer *buf)
216{
217 kref_put(&buf->kref, ust_buffers_remove_buf);
218}
219
220int ust_buffers_channel_open(struct ust_channel *chan, size_t subbuf_size, size_t subbuf_cnt)
221{
222 int i;
223 int result;
224
225 if(subbuf_size == 0 || subbuf_cnt == 0)
226 return -1;
227
228 /* Check that the subbuffer size is larger than a page. */
229 WARN_ON_ONCE(subbuf_size < PAGE_SIZE);
230
231 /*
232 * Make sure the number of subbuffers and subbuffer size are power of 2.
233 */
234 WARN_ON_ONCE(hweight32(subbuf_size) != 1);
235 WARN_ON(hweight32(subbuf_cnt) != 1);
236
237 chan->version = UST_CHANNEL_VERSION;
238 chan->subbuf_cnt = subbuf_cnt;
239 chan->subbuf_size = subbuf_size;
240 chan->subbuf_size_order = get_count_order(subbuf_size);
241 chan->alloc_size = subbuf_size * subbuf_cnt;
242
243 kref_init(&chan->kref);
244
245 mutex_lock(&ust_buffers_channels_mutex);
246 for(i=0; i<chan->n_cpus; i++) {
247 result = ust_buffers_open_buf(chan, i);
248 if (result == -1)
249 goto error;
250 }
251 list_add(&chan->list, &ust_buffers_channels);
252 mutex_unlock(&ust_buffers_channels_mutex);
253
254 return 0;
255
256 /* Jump directly inside the loop to close the buffers that were already
257 * opened. */
258 for(; i>=0; i--) {
259 ust_buffers_close_buf(chan->buf[i]);
260error:
261 do {} while(0);
262 }
263
264 kref_put(&chan->kref, ust_buffers_destroy_channel);
265 mutex_unlock(&ust_buffers_channels_mutex);
266 return -1;
267}
268
269void ust_buffers_channel_close(struct ust_channel *chan)
270{
271 int i;
272 if(!chan)
273 return;
274
275 mutex_lock(&ust_buffers_channels_mutex);
276 for(i=0; i<chan->n_cpus; i++) {
277 /* FIXME: if we make it here, then all buffers were necessarily allocated. Moreover, we don't
278 * initialize to NULL so we cannot use this check. Should we? */
279//ust// if (chan->buf[i])
280 ust_buffers_close_buf(chan->buf[i]);
281 }
282
283 list_del(&chan->list);
284 kref_put(&chan->kref, ust_buffers_destroy_channel);
285 mutex_unlock(&ust_buffers_channels_mutex);
286}
287
288/*
289 * -------
290 */
291
292static void ust_buffers_destroy_buffer(struct ust_channel *ltt_chan, int cpu);
293
294static void ltt_force_switch(struct ust_buffer *buf,
295 enum force_switch_mode mode);
296
297/*
298 * Trace callbacks
299 */
300static void ltt_buffer_begin(struct ust_buffer *buf,
301 u64 tsc, unsigned int subbuf_idx)
302{
303 struct ust_channel *channel = buf->chan;
304 struct ltt_subbuffer_header *header =
305 (struct ltt_subbuffer_header *)
306 ust_buffers_offset_address(buf,
307 subbuf_idx * buf->chan->subbuf_size);
308
309 header->cycle_count_begin = tsc;
310 header->data_size = 0xFFFFFFFF; /* for recognizing crashed buffers */
311 header->sb_size = 0xFFFFFFFF; /* for recognizing crashed buffers */
312 /* FIXME: add memory barrier? */
313 ltt_write_trace_header(channel->trace, header);
314}
315
316/*
317 * offset is assumed to never be 0 here : never deliver a completely empty
318 * subbuffer. The lost size is between 0 and subbuf_size-1.
319 */
320static notrace void ltt_buffer_end(struct ust_buffer *buf,
321 u64 tsc, unsigned int offset, unsigned int subbuf_idx)
322{
323 struct ltt_subbuffer_header *header =
324 (struct ltt_subbuffer_header *)
325 ust_buffers_offset_address(buf,
326 subbuf_idx * buf->chan->subbuf_size);
327 u32 data_size = SUBBUF_OFFSET(offset - 1, buf->chan) + 1;
328
329 header->data_size = data_size;
330 header->sb_size = PAGE_ALIGN(data_size);
331 header->cycle_count_end = tsc;
332 header->events_lost = local_read(&buf->events_lost);
333 header->subbuf_corrupt = local_read(&buf->corrupted_subbuffers);
334}
335
336/*
337 * This function should not be called from NMI interrupt context
338 */
339static notrace void ltt_buf_unfull(struct ust_buffer *buf,
340 unsigned int subbuf_idx,
341 long offset)
342{
343}
344
345/*
346 * Promote compiler barrier to a smp_mb().
347 * For the specific LTTng case, this IPI call should be removed if the
348 * architecture does not reorder writes. This should eventually be provided by
349 * a separate architecture-specific infrastructure.
350 */
351//ust// static void remote_mb(void *info)
352//ust// {
353//ust// smp_mb();
354//ust// }
355
356int ust_buffers_get_subbuf(struct ust_buffer *buf, long *consumed)
357{
358 struct ust_channel *channel = buf->chan;
359 long consumed_old, consumed_idx, commit_count, write_offset;
360//ust// int retval;
361
362 consumed_old = atomic_long_read(&buf->consumed);
363 consumed_idx = SUBBUF_INDEX(consumed_old, buf->chan);
364 commit_count = local_read(&buf->commit_count[consumed_idx].cc_sb);
365 /*
366 * Make sure we read the commit count before reading the buffer
367 * data and the write offset. Correct consumed offset ordering
368 * wrt commit count is insured by the use of cmpxchg to update
369 * the consumed offset.
370 * smp_call_function_single can fail if the remote CPU is offline,
371 * this is OK because then there is no wmb to execute there.
372 * If our thread is executing on the same CPU as the on the buffers
373 * belongs to, we don't have to synchronize it at all. If we are
374 * migrated, the scheduler will take care of the memory barriers.
375 * Normally, smp_call_function_single() should ensure program order when
376 * executing the remote function, which implies that it surrounds the
377 * function execution with :
378 * smp_mb()
379 * send IPI
380 * csd_lock_wait
381 * recv IPI
382 * smp_mb()
383 * exec. function
384 * smp_mb()
385 * csd unlock
386 * smp_mb()
387 *
388 * However, smp_call_function_single() does not seem to clearly execute
389 * such barriers. It depends on spinlock semantic to provide the barrier
390 * before executing the IPI and, when busy-looping, csd_lock_wait only
391 * executes smp_mb() when it has to wait for the other CPU.
392 *
393 * I don't trust this code. Therefore, let's add the smp_mb() sequence
394 * required ourself, even if duplicated. It has no performance impact
395 * anyway.
396 *
397 * smp_mb() is needed because smp_rmb() and smp_wmb() only order read vs
398 * read and write vs write. They do not ensure core synchronization. We
399 * really have to ensure total order between the 3 barriers running on
400 * the 2 CPUs.
401 */
402//ust// #ifdef LTT_NO_IPI_BARRIER
403 /*
404 * Local rmb to match the remote wmb to read the commit count before the
405 * buffer data and the write offset.
406 */
407 smp_rmb();
408//ust// #else
409//ust// if (raw_smp_processor_id() != buf->cpu) {
410//ust// smp_mb(); /* Total order with IPI handler smp_mb() */
411//ust// smp_call_function_single(buf->cpu, remote_mb, NULL, 1);
412//ust// smp_mb(); /* Total order with IPI handler smp_mb() */
413//ust// }
414//ust// #endif
415
416 write_offset = local_read(&buf->offset);
417 /*
418 * Check that the subbuffer we are trying to consume has been
419 * already fully committed.
420 */
421 if (((commit_count - buf->chan->subbuf_size)
422 & channel->commit_count_mask)
423 - (BUFFER_TRUNC(consumed_old, buf->chan)
424 >> channel->n_subbufs_order)
425 != 0) {
426 return -EAGAIN;
427 }
428 /*
429 * Check that we are not about to read the same subbuffer in
430 * which the writer head is.
431 */
432 if ((SUBBUF_TRUNC(write_offset, buf->chan)
433 - SUBBUF_TRUNC(consumed_old, buf->chan))
434 == 0) {
435 return -EAGAIN;
436 }
437
438 /* FIXME: is this ok to disable the reading feature? */
439//ust// retval = update_read_sb_index(buf, consumed_idx);
440//ust// if (retval)
441//ust// return retval;
442
443 *consumed = consumed_old;
444
445 return 0;
446}
447
448int ust_buffers_put_subbuf(struct ust_buffer *buf, unsigned long uconsumed_old)
449{
450 long consumed_new, consumed_old;
451
452 consumed_old = atomic_long_read(&buf->consumed);
453 consumed_old = consumed_old & (~0xFFFFFFFFL);
454 consumed_old = consumed_old | uconsumed_old;
455 consumed_new = SUBBUF_ALIGN(consumed_old, buf->chan);
456
457//ust// spin_lock(&ltt_buf->full_lock);
458 if (atomic_long_cmpxchg(&buf->consumed, consumed_old,
459 consumed_new)
460 != consumed_old) {
461 /* We have been pushed by the writer : the last
462 * buffer read _is_ corrupted! It can also
463 * happen if this is a buffer we never got. */
464//ust// spin_unlock(&ltt_buf->full_lock);
465 return -EIO;
466 } else {
467 /* tell the client that buffer is now unfull */
468 int index;
469 long data;
470 index = SUBBUF_INDEX(consumed_old, buf->chan);
471 data = BUFFER_OFFSET(consumed_old, buf->chan);
472 ltt_buf_unfull(buf, index, data);
473//ust// spin_unlock(&ltt_buf->full_lock);
474 }
475 return 0;
476}
477
478//ust// static void switch_buffer(unsigned long data)
479//ust// {
480//ust// struct ltt_channel_buf_struct *ltt_buf =
481//ust// (struct ltt_channel_buf_struct *)data;
482//ust// struct rchan_buf *buf = ltt_buf->rbuf;
483//ust//
484//ust// if (buf)
485//ust// ltt_force_switch(buf, FORCE_ACTIVE);
486//ust//
487//ust// ltt_buf->switch_timer.expires += ltt_buf->switch_timer_interval;
488//ust// add_timer_on(&ltt_buf->switch_timer, smp_processor_id());
489//ust// }
490//ust//
491//ust// static void start_switch_timer(struct ltt_channel_struct *ltt_channel)
492//ust// {
493//ust// struct rchan *rchan = ltt_channel->trans_channel_data;
494//ust// int cpu;
495//ust//
496//ust// if (!ltt_channel->switch_timer_interval)
497//ust// return;
498//ust//
499//ust// // TODO : hotplug
500//ust// for_each_online_cpu(cpu) {
501//ust// struct ltt_channel_buf_struct *ltt_buf;
502//ust// struct rchan_buf *buf;
503//ust//
504//ust// buf = rchan->buf[cpu];
505//ust// ltt_buf = buf->chan_private;
506//ust// buf->random_access = 1;
507//ust// ltt_buf->switch_timer_interval =
508//ust// ltt_channel->switch_timer_interval;
509//ust// init_timer(&ltt_buf->switch_timer);
510//ust// ltt_buf->switch_timer.function = switch_buffer;
511//ust// ltt_buf->switch_timer.expires = jiffies +
512//ust// ltt_buf->switch_timer_interval;
513//ust// ltt_buf->switch_timer.data = (unsigned long)ltt_buf;
514//ust// add_timer_on(&ltt_buf->switch_timer, cpu);
515//ust// }
516//ust// }
517//ust//
518//ust// /*
519//ust// * Cannot use del_timer_sync with add_timer_on, so use an IPI to locally
520//ust// * delete the timer.
521//ust// */
522//ust// static void stop_switch_timer_ipi(void *info)
523//ust// {
524//ust// struct ltt_channel_buf_struct *ltt_buf =
525//ust// (struct ltt_channel_buf_struct *)info;
526//ust//
527//ust// del_timer(&ltt_buf->switch_timer);
528//ust// }
529//ust//
530//ust// static void stop_switch_timer(struct ltt_channel_struct *ltt_channel)
531//ust// {
532//ust// struct rchan *rchan = ltt_channel->trans_channel_data;
533//ust// int cpu;
534//ust//
535//ust// if (!ltt_channel->switch_timer_interval)
536//ust// return;
537//ust//
538//ust// // TODO : hotplug
539//ust// for_each_online_cpu(cpu) {
540//ust// struct ltt_channel_buf_struct *ltt_buf;
541//ust// struct rchan_buf *buf;
542//ust//
543//ust// buf = rchan->buf[cpu];
544//ust// ltt_buf = buf->chan_private;
545//ust// smp_call_function(stop_switch_timer_ipi, ltt_buf, 1);
546//ust// buf->random_access = 0;
547//ust// }
548//ust// }
549
550//ust// static void ust_buffers_print_written(struct ust_channel *chan,
551//ust// long cons_off, unsigned int cpu)
552//ust// {
553//ust// struct ust_buffer *buf = chan->buf[cpu];
554//ust// long cons_idx, events_count;
555//ust//
556//ust// cons_idx = SUBBUF_INDEX(cons_off, chan);
557//ust// events_count = local_read(&buf->commit_count[cons_idx].events);
558//ust//
559//ust// if (events_count)
560//ust// printk(KERN_INFO
561//ust// "channel %s: %lu events written (cpu %u, index %lu)\n",
562//ust// chan->channel_name, events_count, cpu, cons_idx);
563//ust// }
564
565static void ltt_relay_print_subbuffer_errors(
566 struct ust_channel *channel,
567 long cons_off, int cpu)
568{
569 struct ust_buffer *ltt_buf = channel->buf[cpu];
570 long cons_idx, commit_count, commit_count_sb, write_offset;
571
572 cons_idx = SUBBUF_INDEX(cons_off, channel);
573 commit_count = local_read(&ltt_buf->commit_count[cons_idx].cc);
574 commit_count_sb = local_read(&ltt_buf->commit_count[cons_idx].cc_sb);
575
576 /*
577 * No need to order commit_count and write_offset reads because we
578 * execute after trace is stopped when there are no readers left.
579 */
580 write_offset = local_read(&ltt_buf->offset);
581 WARN( "LTT : unread channel %s offset is %ld "
582 "and cons_off : %ld (cpu %d)\n",
583 channel->channel_name, write_offset, cons_off, cpu);
584 /* Check each sub-buffer for non filled commit count */
585 if (((commit_count - channel->subbuf_size) & channel->commit_count_mask)
586 - (BUFFER_TRUNC(cons_off, channel) >> channel->n_subbufs_order) != 0) {
587 ERR("LTT : %s : subbuffer %lu has non filled "
588 "commit count [cc, cc_sb] [%lu,%lu].\n",
589 channel->channel_name, cons_idx, commit_count, commit_count_sb);
590 }
591 ERR("LTT : %s : commit count : %lu, subbuf size %zd\n",
592 channel->channel_name, commit_count,
593 channel->subbuf_size);
594}
595
596static void ltt_relay_print_errors(struct ust_trace *trace,
597 struct ust_channel *channel, int cpu)
598{
599 struct ust_buffer *ltt_buf = channel->buf[cpu];
600 long cons_off;
601
602 /*
603 * Can be called in the error path of allocation when
604 * trans_channel_data is not yet set.
605 */
606 if (!channel)
607 return;
608
609//ust// for (cons_off = 0; cons_off < rchan->alloc_size;
610//ust// cons_off = SUBBUF_ALIGN(cons_off, rchan))
611//ust// ust_buffers_print_written(ltt_chan, cons_off, cpu);
612 for (cons_off = atomic_long_read(&ltt_buf->consumed);
613 (SUBBUF_TRUNC(local_read(&ltt_buf->offset),
614 channel)
615 - cons_off) > 0;
616 cons_off = SUBBUF_ALIGN(cons_off, channel))
617 ltt_relay_print_subbuffer_errors(channel, cons_off, cpu);
618}
619
620static void ltt_relay_print_buffer_errors(struct ust_channel *channel, int cpu)
621{
622 struct ust_trace *trace = channel->trace;
623 struct ust_buffer *ltt_buf = channel->buf[cpu];
624
625 if (local_read(&ltt_buf->events_lost))
626 ERR("channel %s: %ld events lost (cpu %d)",
627 channel->channel_name,
628 local_read(&ltt_buf->events_lost), cpu);
629 if (local_read(&ltt_buf->corrupted_subbuffers))
630 ERR("channel %s : %ld corrupted subbuffers (cpu %d)",
631 channel->channel_name,
632 local_read(&ltt_buf->corrupted_subbuffers), cpu);
633
634 ltt_relay_print_errors(trace, channel, cpu);
635}
636
637static void ltt_relay_release_channel(struct kref *kref)
638{
639 struct ust_channel *ltt_chan = container_of(kref,
640 struct ust_channel, kref);
641 free(ltt_chan->buf);
642}
643
644/*
645 * Create ltt buffer.
646 */
647//ust// static int ltt_relay_create_buffer(struct ust_trace *trace,
648//ust// struct ltt_channel_struct *ltt_chan, struct rchan_buf *buf,
649//ust// unsigned int cpu, unsigned int n_subbufs)
650//ust// {
651//ust// struct ltt_channel_buf_struct *ltt_buf =
652//ust// percpu_ptr(ltt_chan->buf, cpu);
653//ust// unsigned int j;
654//ust//
655//ust// ltt_buf->commit_count =
656//ust// kzalloc_node(sizeof(ltt_buf->commit_count) * n_subbufs,
657//ust// GFP_KERNEL, cpu_to_node(cpu));
658//ust// if (!ltt_buf->commit_count)
659//ust// return -ENOMEM;
660//ust// kref_get(&trace->kref);
661//ust// kref_get(&trace->ltt_transport_kref);
662//ust// kref_get(&ltt_chan->kref);
663//ust// local_set(&ltt_buf->offset, ltt_subbuffer_header_size());
664//ust// atomic_long_set(&ltt_buf->consumed, 0);
665//ust// atomic_long_set(&ltt_buf->active_readers, 0);
666//ust// for (j = 0; j < n_subbufs; j++)
667//ust// local_set(&ltt_buf->commit_count[j], 0);
668//ust// init_waitqueue_head(&ltt_buf->write_wait);
669//ust// atomic_set(&ltt_buf->wakeup_readers, 0);
670//ust// spin_lock_init(&ltt_buf->full_lock);
671//ust//
672//ust// ltt_buffer_begin_callback(buf, trace->start_tsc, 0);
673//ust// /* atomic_add made on local variable on data that belongs to
674//ust// * various CPUs : ok because tracing not started (for this cpu). */
675//ust// local_add(ltt_subbuffer_header_size(), &ltt_buf->commit_count[0]);
676//ust//
677//ust// local_set(&ltt_buf->events_lost, 0);
678//ust// local_set(&ltt_buf->corrupted_subbuffers, 0);
679//ust//
680//ust// return 0;
681//ust// }
682
683static int ust_buffers_init_buffer(struct ust_trace *trace,
684 struct ust_channel *ltt_chan, struct ust_buffer *buf,
685 unsigned int n_subbufs)
686{
687 unsigned int j;
688 int fds[2];
689 int result;
690
691 buf->commit_count =
692 zmalloc(sizeof(*buf->commit_count) * n_subbufs);
693 if (!buf->commit_count)
694 return -ENOMEM;
695 kref_get(&trace->kref);
696 kref_get(&trace->ltt_transport_kref);
697 kref_get(&ltt_chan->kref);
698 local_set(&buf->offset, ltt_subbuffer_header_size());
699 atomic_long_set(&buf->consumed, 0);
700 atomic_long_set(&buf->active_readers, 0);
701 for (j = 0; j < n_subbufs; j++) {
702 local_set(&buf->commit_count[j].cc, 0);
703 local_set(&buf->commit_count[j].cc_sb, 0);
704 }
705//ust// init_waitqueue_head(&buf->write_wait);
706//ust// atomic_set(&buf->wakeup_readers, 0);
707//ust// spin_lock_init(&buf->full_lock);
708
709 ltt_buffer_begin(buf, trace->start_tsc, 0);
710
711 local_add(ltt_subbuffer_header_size(), &buf->commit_count[0].cc);
712
713 local_set(&buf->events_lost, 0);
714 local_set(&buf->corrupted_subbuffers, 0);
715
716 result = pipe(fds);
717 if(result == -1) {
718 PERROR("pipe");
719 return -1;
720 }
721 buf->data_ready_fd_read = fds[0];
722 buf->data_ready_fd_write = fds[1];
723
724 /* FIXME: do we actually need this? */
725 result = fcntl(fds[0], F_SETFL, O_NONBLOCK);
726 if(result == -1) {
727 PERROR("fcntl");
728 }
729
730//ust// buf->commit_seq = malloc(sizeof(buf->commit_seq) * n_subbufs);
731//ust// if(!ltt_buf->commit_seq) {
732//ust// return -1;
733//ust// }
734 memset(buf->commit_seq, 0, sizeof(buf->commit_seq[0]) * n_subbufs);
735
736 /* FIXME: decrementally destroy on error */
737
738 return 0;
739}
740
741/* FIXME: use this function */
742static void ust_buffers_destroy_buffer(struct ust_channel *ltt_chan, int cpu)
743{
744 struct ust_trace *trace = ltt_chan->trace;
745 struct ust_buffer *ltt_buf = ltt_chan->buf[cpu];
746
747 kref_put(&ltt_chan->trace->ltt_transport_kref,
748 ltt_release_transport);
749 ltt_relay_print_buffer_errors(ltt_chan, cpu);
750//ust// free(ltt_buf->commit_seq);
751 kfree(ltt_buf->commit_count);
752 ltt_buf->commit_count = NULL;
753 kref_put(&ltt_chan->kref, ltt_relay_release_channel);
754 kref_put(&trace->kref, ltt_release_trace);
755//ust// wake_up_interruptible(&trace->kref_wq);
756}
757
758static int ust_buffers_alloc_channel_buf_structs(struct ust_channel *chan)
759{
760 void *ptr;
761 int result;
762 size_t size;
763 int i;
764
765 size = PAGE_ALIGN(1);
766
767 for(i=0; i<chan->n_cpus; i++) {
768
769 result = chan->buf_struct_shmids[i] = shmget(getpid(), size, IPC_CREAT | IPC_EXCL | 0700);
770 if(result == -1) {
771 PERROR("shmget");
772 goto destroy_previous;
773 }
774
775 /* FIXME: should have matching call to shmdt */
776 ptr = shmat(chan->buf_struct_shmids[i], NULL, 0);
777 if(ptr == (void *) -1) {
778 perror("shmat");
779 goto destroy_shm;
780 }
781
782 /* Already mark the shared memory for destruction. This will occur only
783 * when all users have detached.
784 */
785 result = shmctl(chan->buf_struct_shmids[i], IPC_RMID, NULL);
786 if(result == -1) {
787 perror("shmctl");
788 goto destroy_previous;
789 }
790
791 chan->buf[i] = ptr;
792 }
793
794 return 0;
795
796 /* Jumping inside this loop occurs from within the other loop above with i as
797 * counter, so it unallocates the structures for the cpu = current_i down to
798 * zero. */
799 for(; i>=0; i--) {
800 destroy_shm:
801 result = shmctl(chan->buf_struct_shmids[i], IPC_RMID, NULL);
802 if(result == -1) {
803 perror("shmctl");
804 }
805
806 destroy_previous:
807 continue;
808 }
809
810 return -1;
811}
812
813/*
814 * Create channel.
815 */
816static int ust_buffers_create_channel(const char *trace_name, struct ust_trace *trace,
817 const char *channel_name, struct ust_channel *ltt_chan,
818 unsigned int subbuf_size, unsigned int n_subbufs, int overwrite)
819{
820 int result;
821
822 kref_init(&ltt_chan->kref);
823
824 ltt_chan->trace = trace;
825 ltt_chan->overwrite = overwrite;
826 ltt_chan->n_subbufs_order = get_count_order(n_subbufs);
827 ltt_chan->commit_count_mask = (~0UL >> ltt_chan->n_subbufs_order);
828 ltt_chan->n_cpus = get_n_cpus();
829//ust// ltt_chan->buf = percpu_alloc_mask(sizeof(struct ltt_channel_buf_struct), GFP_KERNEL, cpu_possible_map);
830 ltt_chan->buf = (void *) malloc(ltt_chan->n_cpus * sizeof(void *));
831 if(ltt_chan->buf == NULL) {
832 goto error;
833 }
834 ltt_chan->buf_struct_shmids = (int *) malloc(ltt_chan->n_cpus * sizeof(int));
835 if(ltt_chan->buf_struct_shmids == NULL)
836 goto free_buf;
837
838 result = ust_buffers_alloc_channel_buf_structs(ltt_chan);
839 if(result != 0) {
840 goto free_buf_struct_shmids;
841 }
842
843 result = ust_buffers_channel_open(ltt_chan, subbuf_size, n_subbufs);
844 if (result != 0) {
845 ERR("Cannot open channel for trace %s", trace_name);
846 goto unalloc_buf_structs;
847 }
848
849 return 0;
850
851unalloc_buf_structs:
852 /* FIXME: put a call here to unalloc the buf structs! */
853
854free_buf_struct_shmids:
855 free(ltt_chan->buf_struct_shmids);
856
857free_buf:
858 free(ltt_chan->buf);
859
860error:
861 return -1;
862}
863
864/*
865 * LTTng channel flush function.
866 *
867 * Must be called when no tracing is active in the channel, because of
868 * accesses across CPUs.
869 */
870static notrace void ltt_relay_buffer_flush(struct ust_buffer *buf)
871{
872 int result;
873
874//ust// buf->finalized = 1;
875 ltt_force_switch(buf, FORCE_FLUSH);
876
877 result = write(buf->data_ready_fd_write, "1", 1);
878 if(result == -1) {
879 PERROR("write (in ltt_relay_buffer_flush)");
880 ERR("this should never happen!");
881 }
882}
883
884static void ltt_relay_async_wakeup_chan(struct ust_channel *ltt_channel)
885{
886//ust// unsigned int i;
887//ust// struct rchan *rchan = ltt_channel->trans_channel_data;
888//ust//
889//ust// for_each_possible_cpu(i) {
890//ust// struct ltt_channel_buf_struct *ltt_buf =
891//ust// percpu_ptr(ltt_channel->buf, i);
892//ust//
893//ust// if (atomic_read(&ltt_buf->wakeup_readers) == 1) {
894//ust// atomic_set(&ltt_buf->wakeup_readers, 0);
895//ust// wake_up_interruptible(&rchan->buf[i]->read_wait);
896//ust// }
897//ust// }
898}
899
900static void ltt_relay_finish_buffer(struct ust_channel *channel, unsigned int cpu)
901{
902// int result;
903
904 if (channel->buf[cpu]) {
905 struct ust_buffer *buf = channel->buf[cpu];
906 ltt_relay_buffer_flush(buf);
907//ust// ltt_relay_wake_writers(ltt_buf);
908 /* closing the pipe tells the consumer the buffer is finished */
909
910 //result = write(ltt_buf->data_ready_fd_write, "D", 1);
911 //if(result == -1) {
912 // PERROR("write (in ltt_relay_finish_buffer)");
913 // ERR("this should never happen!");
914 //}
915 close(buf->data_ready_fd_write);
916 }
917}
918
919
920static void ltt_relay_finish_channel(struct ust_channel *channel)
921{
922 unsigned int i;
923
924 for(i=0; i<channel->n_cpus; i++) {
925 ltt_relay_finish_buffer(channel, i);
926 }
927}
928
929static void ltt_relay_remove_channel(struct ust_channel *channel)
930{
931 ust_buffers_channel_close(channel);
932 kref_put(&channel->kref, ltt_relay_release_channel);
933}
934
935//ust// /*
936//ust// * Returns :
937//ust// * 0 if ok
938//ust// * !0 if execution must be aborted.
939//ust// */
940//ust// static inline int ltt_relay_try_reserve(
941//ust// struct ust_channel *channel, struct ust_buffer *buf,
942//ust// struct ltt_reserve_switch_offsets *offsets, size_t data_size,
943//ust// u64 *tsc, unsigned int *rflags, int largest_align)
944//ust// {
945//ust// offsets->begin = local_read(&buf->offset);
946//ust// offsets->old = offsets->begin;
947//ust// offsets->begin_switch = 0;
948//ust// offsets->end_switch_current = 0;
949//ust// offsets->end_switch_old = 0;
950//ust//
951//ust// *tsc = trace_clock_read64();
952//ust// if (last_tsc_overflow(buf, *tsc))
953//ust// *rflags = LTT_RFLAG_ID_SIZE_TSC;
954//ust//
955//ust// if (SUBBUF_OFFSET(offsets->begin, buf->chan) == 0) {
956//ust// offsets->begin_switch = 1; /* For offsets->begin */
957//ust// } else {
958//ust// offsets->size = ust_get_header_size(channel,
959//ust// offsets->begin, data_size,
960//ust// &offsets->before_hdr_pad, *rflags);
961//ust// offsets->size += ltt_align(offsets->begin + offsets->size,
962//ust// largest_align)
963//ust// + data_size;
964//ust// if ((SUBBUF_OFFSET(offsets->begin, buf->chan) + offsets->size)
965//ust// > buf->chan->subbuf_size) {
966//ust// offsets->end_switch_old = 1; /* For offsets->old */
967//ust// offsets->begin_switch = 1; /* For offsets->begin */
968//ust// }
969//ust// }
970//ust// if (offsets->begin_switch) {
971//ust// long subbuf_index;
972//ust//
973//ust// if (offsets->end_switch_old)
974//ust// offsets->begin = SUBBUF_ALIGN(offsets->begin,
975//ust// buf->chan);
976//ust// offsets->begin = offsets->begin + ltt_subbuffer_header_size();
977//ust// /* Test new buffer integrity */
978//ust// subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan);
979//ust// offsets->reserve_commit_diff =
980//ust// (BUFFER_TRUNC(offsets->begin, buf->chan)
981//ust// >> channel->n_subbufs_order)
982//ust// - (local_read(&buf->commit_count[subbuf_index])
983//ust// & channel->commit_count_mask);
984//ust// if (offsets->reserve_commit_diff == 0) {
985//ust// long consumed;
986//ust//
987//ust// consumed = atomic_long_read(&buf->consumed);
988//ust//
989//ust// /* Next buffer not corrupted. */
990//ust// if (!channel->overwrite &&
991//ust// (SUBBUF_TRUNC(offsets->begin, buf->chan)
992//ust// - SUBBUF_TRUNC(consumed, buf->chan))
993//ust// >= channel->alloc_size) {
994//ust//
995//ust// long consumed_idx = SUBBUF_INDEX(consumed, buf->chan);
996//ust// long commit_count = local_read(&buf->commit_count[consumed_idx]);
997//ust// if(((commit_count - buf->chan->subbuf_size) & channel->commit_count_mask) - (BUFFER_TRUNC(consumed, buf->chan) >> channel->n_subbufs_order) != 0) {
998//ust// WARN("Event dropped. Caused by non-committed event.");
999//ust// }
1000//ust// else {
1001//ust// WARN("Event dropped. Caused by non-consumed buffer.");
1002//ust// }
1003//ust// /*
1004//ust// * We do not overwrite non consumed buffers
1005//ust// * and we are full : event is lost.
1006//ust// */
1007//ust// local_inc(&buf->events_lost);
1008//ust// return -1;
1009//ust// } else {
1010//ust// /*
1011//ust// * next buffer not corrupted, we are either in
1012//ust// * overwrite mode or the buffer is not full.
1013//ust// * It's safe to write in this new subbuffer.
1014//ust// */
1015//ust// }
1016//ust// } else {
1017//ust// /*
1018//ust// * Next subbuffer corrupted. Force pushing reader even
1019//ust// * in normal mode. It's safe to write in this new
1020//ust// * subbuffer.
1021//ust// */
1022//ust// }
1023//ust// offsets->size = ust_get_header_size(channel,
1024//ust// offsets->begin, data_size,
1025//ust// &offsets->before_hdr_pad, *rflags);
1026//ust// offsets->size += ltt_align(offsets->begin + offsets->size,
1027//ust// largest_align)
1028//ust// + data_size;
1029//ust// if ((SUBBUF_OFFSET(offsets->begin, buf->chan) + offsets->size)
1030//ust// > buf->chan->subbuf_size) {
1031//ust// /*
1032//ust// * Event too big for subbuffers, report error, don't
1033//ust// * complete the sub-buffer switch.
1034//ust// */
1035//ust// local_inc(&buf->events_lost);
1036//ust// return -1;
1037//ust// } else {
1038//ust// /*
1039//ust// * We just made a successful buffer switch and the event
1040//ust// * fits in the new subbuffer. Let's write.
1041//ust// */
1042//ust// }
1043//ust// } else {
1044//ust// /*
1045//ust// * Event fits in the current buffer and we are not on a switch
1046//ust// * boundary. It's safe to write.
1047//ust// */
1048//ust// }
1049//ust// offsets->end = offsets->begin + offsets->size;
1050//ust//
1051//ust// if ((SUBBUF_OFFSET(offsets->end, buf->chan)) == 0) {
1052//ust// /*
1053//ust// * The offset_end will fall at the very beginning of the next
1054//ust// * subbuffer.
1055//ust// */
1056//ust// offsets->end_switch_current = 1; /* For offsets->begin */
1057//ust// }
1058//ust// return 0;
1059//ust// }
1060//ust//
1061//ust// /*
1062//ust// * Returns :
1063//ust// * 0 if ok
1064//ust// * !0 if execution must be aborted.
1065//ust// */
1066//ust// static inline int ltt_relay_try_switch(
1067//ust// enum force_switch_mode mode,
1068//ust// struct ust_channel *channel,
1069//ust// struct ust_buffer *buf,
1070//ust// struct ltt_reserve_switch_offsets *offsets,
1071//ust// u64 *tsc)
1072//ust// {
1073//ust// long subbuf_index;
1074//ust//
1075//ust// offsets->begin = local_read(&buf->offset);
1076//ust// offsets->old = offsets->begin;
1077//ust// offsets->begin_switch = 0;
1078//ust// offsets->end_switch_old = 0;
1079//ust//
1080//ust// *tsc = trace_clock_read64();
1081//ust//
1082//ust// if (SUBBUF_OFFSET(offsets->begin, buf->chan) != 0) {
1083//ust// offsets->begin = SUBBUF_ALIGN(offsets->begin, buf->chan);
1084//ust// offsets->end_switch_old = 1;
1085//ust// } else {
1086//ust// /* we do not have to switch : buffer is empty */
1087//ust// return -1;
1088//ust// }
1089//ust// if (mode == FORCE_ACTIVE)
1090//ust// offsets->begin += ltt_subbuffer_header_size();
1091//ust// /*
1092//ust// * Always begin_switch in FORCE_ACTIVE mode.
1093//ust// * Test new buffer integrity
1094//ust// */
1095//ust// subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan);
1096//ust// offsets->reserve_commit_diff =
1097//ust// (BUFFER_TRUNC(offsets->begin, buf->chan)
1098//ust// >> channel->n_subbufs_order)
1099//ust// - (local_read(&buf->commit_count[subbuf_index])
1100//ust// & channel->commit_count_mask);
1101//ust// if (offsets->reserve_commit_diff == 0) {
1102//ust// /* Next buffer not corrupted. */
1103//ust// if (mode == FORCE_ACTIVE
1104//ust// && !channel->overwrite
1105//ust// && offsets->begin - atomic_long_read(&buf->consumed)
1106//ust// >= channel->alloc_size) {
1107//ust// /*
1108//ust// * We do not overwrite non consumed buffers and we are
1109//ust// * full : ignore switch while tracing is active.
1110//ust// */
1111//ust// return -1;
1112//ust// }
1113//ust// } else {
1114//ust// /*
1115//ust// * Next subbuffer corrupted. Force pushing reader even in normal
1116//ust// * mode
1117//ust// */
1118//ust// }
1119//ust// offsets->end = offsets->begin;
1120//ust// return 0;
1121//ust// }
1122//ust//
1123//ust// static inline void ltt_reserve_push_reader(
1124//ust// struct ust_channel *channel,
1125//ust// struct ust_buffer *buf,
1126//ust// struct ltt_reserve_switch_offsets *offsets)
1127//ust// {
1128//ust// long consumed_old, consumed_new;
1129//ust//
1130//ust// do {
1131//ust// consumed_old = atomic_long_read(&buf->consumed);
1132//ust// /*
1133//ust// * If buffer is in overwrite mode, push the reader consumed
1134//ust// * count if the write position has reached it and we are not
1135//ust// * at the first iteration (don't push the reader farther than
1136//ust// * the writer). This operation can be done concurrently by many
1137//ust// * writers in the same buffer, the writer being at the farthest
1138//ust// * write position sub-buffer index in the buffer being the one
1139//ust// * which will win this loop.
1140//ust// * If the buffer is not in overwrite mode, pushing the reader
1141//ust// * only happens if a sub-buffer is corrupted.
1142//ust// */
1143//ust// if ((SUBBUF_TRUNC(offsets->end-1, buf->chan)
1144//ust// - SUBBUF_TRUNC(consumed_old, buf->chan))
1145//ust// >= channel->alloc_size)
1146//ust// consumed_new = SUBBUF_ALIGN(consumed_old, buf->chan);
1147//ust// else {
1148//ust// consumed_new = consumed_old;
1149//ust// break;
1150//ust// }
1151//ust// } while (atomic_long_cmpxchg(&buf->consumed, consumed_old,
1152//ust// consumed_new) != consumed_old);
1153//ust//
1154//ust// if (consumed_old != consumed_new) {
1155//ust// /*
1156//ust// * Reader pushed : we are the winner of the push, we can
1157//ust// * therefore reequilibrate reserve and commit. Atomic increment
1158//ust// * of the commit count permits other writers to play around
1159//ust// * with this variable before us. We keep track of
1160//ust// * corrupted_subbuffers even in overwrite mode :
1161//ust// * we never want to write over a non completely committed
1162//ust// * sub-buffer : possible causes : the buffer size is too low
1163//ust// * compared to the unordered data input, or there is a writer
1164//ust// * that died between the reserve and the commit.
1165//ust// */
1166//ust// if (offsets->reserve_commit_diff) {
1167//ust// /*
1168//ust// * We have to alter the sub-buffer commit count.
1169//ust// * We do not deliver the previous subbuffer, given it
1170//ust// * was either corrupted or not consumed (overwrite
1171//ust// * mode).
1172//ust// */
1173//ust// local_add(offsets->reserve_commit_diff,
1174//ust// &buf->commit_count[
1175//ust// SUBBUF_INDEX(offsets->begin,
1176//ust// buf->chan)]);
1177//ust// if (!channel->overwrite
1178//ust// || offsets->reserve_commit_diff
1179//ust// != channel->subbuf_size) {
1180//ust// /*
1181//ust// * The reserve commit diff was not subbuf_size :
1182//ust// * it means the subbuffer was partly written to
1183//ust// * and is therefore corrupted. If it is multiple
1184//ust// * of subbuffer size and we are in flight
1185//ust// * recorder mode, we are skipping over a whole
1186//ust// * subbuffer.
1187//ust// */
1188//ust// local_inc(&buf->corrupted_subbuffers);
1189//ust// }
1190//ust// }
1191//ust// }
1192//ust// }
1193//ust//
1194//ust// /**
1195//ust// * ltt_relay_reserve_slot - Atomic slot reservation in a LTTng buffer.
1196//ust// * @trace: the trace structure to log to.
1197//ust// * @ltt_channel: channel structure
1198//ust// * @transport_data: data structure specific to ltt relay
1199//ust// * @data_size: size of the variable length data to log.
1200//ust// * @slot_size: pointer to total size of the slot (out)
1201//ust// * @buf_offset : pointer to reserved buffer offset (out)
1202//ust// * @tsc: pointer to the tsc at the slot reservation (out)
1203//ust// * @cpu: cpuid
1204//ust// *
1205//ust// * Return : -ENOSPC if not enough space, else returns 0.
1206//ust// * It will take care of sub-buffer switching.
1207//ust// */
1208//ust// static notrace int ltt_relay_reserve_slot(struct ust_trace *trace,
1209//ust// struct ust_channel *channel, void **transport_data,
1210//ust// size_t data_size, size_t *slot_size, long *buf_offset, u64 *tsc,
1211//ust// unsigned int *rflags, int largest_align, int cpu)
1212//ust// {
1213//ust// struct ust_buffer *buf = *transport_data = channel->buf[cpu];
1214//ust// struct ltt_reserve_switch_offsets offsets;
1215//ust//
1216//ust// offsets.reserve_commit_diff = 0;
1217//ust// offsets.size = 0;
1218//ust//
1219//ust// /*
1220//ust// * Perform retryable operations.
1221//ust// */
1222//ust// if (ltt_nesting > 4) {
1223//ust// local_inc(&buf->events_lost);
1224//ust// return -EPERM;
1225//ust// }
1226//ust// do {
1227//ust// if (ltt_relay_try_reserve(channel, buf, &offsets, data_size, tsc, rflags,
1228//ust// largest_align))
1229//ust// return -ENOSPC;
1230//ust// } while (local_cmpxchg(&buf->offset, offsets.old,
1231//ust// offsets.end) != offsets.old);
1232//ust//
1233//ust// /*
1234//ust// * Atomically update last_tsc. This update races against concurrent
1235//ust// * atomic updates, but the race will always cause supplementary full TSC
1236//ust// * events, never the opposite (missing a full TSC event when it would be
1237//ust// * needed).
1238//ust// */
1239//ust// save_last_tsc(buf, *tsc);
1240//ust//
1241//ust// /*
1242//ust// * Push the reader if necessary
1243//ust// */
1244//ust// ltt_reserve_push_reader(channel, buf, &offsets);
1245//ust//
1246//ust// /*
1247//ust// * Switch old subbuffer if needed.
1248//ust// */
1249//ust// if (offsets.end_switch_old)
1250//ust// ltt_reserve_switch_old_subbuf(channel, buf, &offsets, tsc);
1251//ust//
1252//ust// /*
1253//ust// * Populate new subbuffer.
1254//ust// */
1255//ust// if (offsets.begin_switch)
1256//ust// ltt_reserve_switch_new_subbuf(channel, buf, &offsets, tsc);
1257//ust//
1258//ust// if (offsets.end_switch_current)
1259//ust// ltt_reserve_end_switch_current(channel, buf, &offsets, tsc);
1260//ust//
1261//ust// *slot_size = offsets.size;
1262//ust// *buf_offset = offsets.begin + offsets.before_hdr_pad;
1263//ust// return 0;
1264//ust// }
1265//ust//
1266//ust// /*
1267//ust// * Force a sub-buffer switch for a per-cpu buffer. This operation is
1268//ust// * completely reentrant : can be called while tracing is active with
1269//ust// * absolutely no lock held.
1270//ust// *
1271//ust// * Note, however, that as a local_cmpxchg is used for some atomic
1272//ust// * operations, this function must be called from the CPU which owns the buffer
1273//ust// * for a ACTIVE flush.
1274//ust// */
1275//ust// static notrace void ltt_force_switch(struct ust_buffer *buf,
1276//ust// enum force_switch_mode mode)
1277//ust// {
1278//ust// struct ust_channel *channel = buf->chan;
1279//ust// struct ltt_reserve_switch_offsets offsets;
1280//ust// u64 tsc;
1281//ust//
1282//ust// offsets.reserve_commit_diff = 0;
1283//ust// offsets.size = 0;
1284//ust//
1285//ust// /*
1286//ust// * Perform retryable operations.
1287//ust// */
1288//ust// do {
1289//ust// if (ltt_relay_try_switch(mode, channel, buf, &offsets, &tsc))
1290//ust// return;
1291//ust// } while (local_cmpxchg(&buf->offset, offsets.old,
1292//ust// offsets.end) != offsets.old);
1293//ust//
1294//ust// /*
1295//ust// * Atomically update last_tsc. This update races against concurrent
1296//ust// * atomic updates, but the race will always cause supplementary full TSC
1297//ust// * events, never the opposite (missing a full TSC event when it would be
1298//ust// * needed).
1299//ust// */
1300//ust// save_last_tsc(buf, tsc);
1301//ust//
1302//ust// /*
1303//ust// * Push the reader if necessary
1304//ust// */
1305//ust// if (mode == FORCE_ACTIVE)
1306//ust// ltt_reserve_push_reader(channel, buf, &offsets);
1307//ust//
1308//ust// /*
1309//ust// * Switch old subbuffer if needed.
1310//ust// */
1311//ust// if (offsets.end_switch_old)
1312//ust// ltt_reserve_switch_old_subbuf(channel, buf, &offsets, &tsc);
1313//ust//
1314//ust// /*
1315//ust// * Populate new subbuffer.
1316//ust// */
1317//ust// if (mode == FORCE_ACTIVE)
1318//ust// ltt_reserve_switch_new_subbuf(channel, buf, &offsets, &tsc);
1319//ust// }
1320
1321/*
1322 * ltt_reserve_switch_old_subbuf: switch old subbuffer
1323 *
1324 * Concurrency safe because we are the last and only thread to alter this
1325 * sub-buffer. As long as it is not delivered and read, no other thread can
1326 * alter the offset, alter the reserve_count or call the
1327 * client_buffer_end_callback on this sub-buffer.
1328 *
1329 * The only remaining threads could be the ones with pending commits. They will
1330 * have to do the deliver themselves. Not concurrency safe in overwrite mode.
1331 * We detect corrupted subbuffers with commit and reserve counts. We keep a
1332 * corrupted sub-buffers count and push the readers across these sub-buffers.
1333 *
1334 * Not concurrency safe if a writer is stalled in a subbuffer and another writer
1335 * switches in, finding out it's corrupted. The result will be than the old
1336 * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer
1337 * will be declared corrupted too because of the commit count adjustment.
1338 *
1339 * Note : offset_old should never be 0 here.
1340 */
1341static void ltt_reserve_switch_old_subbuf(
1342 struct ust_channel *chan, struct ust_buffer *buf,
1343 struct ltt_reserve_switch_offsets *offsets, u64 *tsc)
1344{
1345 long oldidx = SUBBUF_INDEX(offsets->old - 1, chan);
1346 long commit_count, padding_size;
1347
1348 padding_size = chan->subbuf_size
1349 - (SUBBUF_OFFSET(offsets->old - 1, chan) + 1);
1350 ltt_buffer_end(buf, *tsc, offsets->old, oldidx);
1351
1352 /*
1353 * Must write slot data before incrementing commit count.
1354 * This compiler barrier is upgraded into a smp_wmb() by the IPI
1355 * sent by get_subbuf() when it does its smp_rmb().
1356 */
1357 barrier();
1358 local_add(padding_size,
1359 &buf->commit_count[oldidx].cc);
1360 commit_count = local_read(&buf->commit_count[oldidx].cc);
1361 ltt_check_deliver(chan, buf, offsets->old - 1, commit_count, oldidx);
1362 ltt_write_commit_counter(buf, oldidx,
1363 offsets->old, commit_count, padding_size);
1364}
1365
1366/*
1367 * ltt_reserve_switch_new_subbuf: Populate new subbuffer.
1368 *
1369 * This code can be executed unordered : writers may already have written to the
1370 * sub-buffer before this code gets executed, caution. The commit makes sure
1371 * that this code is executed before the deliver of this sub-buffer.
1372 */
1373static void ltt_reserve_switch_new_subbuf(
1374 struct ust_channel *chan, struct ust_buffer *buf,
1375 struct ltt_reserve_switch_offsets *offsets, u64 *tsc)
1376{
1377 long beginidx = SUBBUF_INDEX(offsets->begin, chan);
1378 long commit_count;
1379
1380 ltt_buffer_begin(buf, *tsc, beginidx);
1381
1382 /*
1383 * Must write slot data before incrementing commit count.
1384 * This compiler barrier is upgraded into a smp_wmb() by the IPI
1385 * sent by get_subbuf() when it does its smp_rmb().
1386 */
1387 barrier();
1388 local_add(ltt_subbuffer_header_size(),
1389 &buf->commit_count[beginidx].cc);
1390 commit_count = local_read(&buf->commit_count[beginidx].cc);
1391 /* Check if the written buffer has to be delivered */
1392 ltt_check_deliver(chan, buf, offsets->begin, commit_count, beginidx);
1393 ltt_write_commit_counter(buf, beginidx,
1394 offsets->begin, commit_count, ltt_subbuffer_header_size());
1395}
1396
1397/*
1398 * ltt_reserve_end_switch_current: finish switching current subbuffer
1399 *
1400 * Concurrency safe because we are the last and only thread to alter this
1401 * sub-buffer. As long as it is not delivered and read, no other thread can
1402 * alter the offset, alter the reserve_count or call the
1403 * client_buffer_end_callback on this sub-buffer.
1404 *
1405 * The only remaining threads could be the ones with pending commits. They will
1406 * have to do the deliver themselves. Not concurrency safe in overwrite mode.
1407 * We detect corrupted subbuffers with commit and reserve counts. We keep a
1408 * corrupted sub-buffers count and push the readers across these sub-buffers.
1409 *
1410 * Not concurrency safe if a writer is stalled in a subbuffer and another writer
1411 * switches in, finding out it's corrupted. The result will be than the old
1412 * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer
1413 * will be declared corrupted too because of the commit count adjustment.
1414 */
1415static void ltt_reserve_end_switch_current(
1416 struct ust_channel *chan,
1417 struct ust_buffer *buf,
1418 struct ltt_reserve_switch_offsets *offsets, u64 *tsc)
1419{
1420 long endidx = SUBBUF_INDEX(offsets->end - 1, chan);
1421 long commit_count, padding_size;
1422
1423 padding_size = chan->subbuf_size
1424 - (SUBBUF_OFFSET(offsets->end - 1, chan) + 1);
1425
1426 ltt_buffer_end(buf, *tsc, offsets->end, endidx);
1427
1428 /*
1429 * Must write slot data before incrementing commit count.
1430 * This compiler barrier is upgraded into a smp_wmb() by the IPI
1431 * sent by get_subbuf() when it does its smp_rmb().
1432 */
1433 barrier();
1434 local_add(padding_size,
1435 &buf->commit_count[endidx].cc);
1436 commit_count = local_read(&buf->commit_count[endidx].cc);
1437 ltt_check_deliver(chan, buf,
1438 offsets->end - 1, commit_count, endidx);
1439 ltt_write_commit_counter(buf, endidx,
1440 offsets->end, commit_count, padding_size);
1441}
1442
1443/*
1444 * Returns :
1445 * 0 if ok
1446 * !0 if execution must be aborted.
1447 */
1448static int ltt_relay_try_switch_slow(
1449 enum force_switch_mode mode,
1450 struct ust_channel *chan,
1451 struct ust_buffer *buf,
1452 struct ltt_reserve_switch_offsets *offsets,
1453 u64 *tsc)
1454{
1455 long subbuf_index;
1456 long reserve_commit_diff;
1457
1458 offsets->begin = local_read(&buf->offset);
1459 offsets->old = offsets->begin;
1460 offsets->begin_switch = 0;
1461 offsets->end_switch_old = 0;
1462
1463 *tsc = trace_clock_read64();
1464
1465 if (SUBBUF_OFFSET(offsets->begin, buf->chan) != 0) {
1466 offsets->begin = SUBBUF_ALIGN(offsets->begin, buf->chan);
1467 offsets->end_switch_old = 1;
1468 } else {
1469 /* we do not have to switch : buffer is empty */
1470 return -1;
1471 }
1472 if (mode == FORCE_ACTIVE)
1473 offsets->begin += ltt_subbuffer_header_size();
1474 /*
1475 * Always begin_switch in FORCE_ACTIVE mode.
1476 * Test new buffer integrity
1477 */
1478 subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan);
1479 reserve_commit_diff =
1480 (BUFFER_TRUNC(offsets->begin, buf->chan)
1481 >> chan->n_subbufs_order)
1482 - (local_read(&buf->commit_count[subbuf_index].cc_sb)
1483 & chan->commit_count_mask);
1484 if (reserve_commit_diff == 0) {
1485 /* Next buffer not corrupted. */
1486 if (mode == FORCE_ACTIVE
1487 && !chan->overwrite
1488 && offsets->begin - atomic_long_read(&buf->consumed)
1489 >= chan->alloc_size) {
1490 /*
1491 * We do not overwrite non consumed buffers and we are
1492 * full : ignore switch while tracing is active.
1493 */
1494 return -1;
1495 }
1496 } else {
1497 /*
1498 * Next subbuffer corrupted. Force pushing reader even in normal
1499 * mode
1500 */
1501 }
1502 offsets->end = offsets->begin;
1503 return 0;
1504}
1505
1506/*
1507 * Force a sub-buffer switch for a per-cpu buffer. This operation is
1508 * completely reentrant : can be called while tracing is active with
1509 * absolutely no lock held.
1510 *
1511 * Note, however, that as a local_cmpxchg is used for some atomic
1512 * operations, this function must be called from the CPU which owns the buffer
1513 * for a ACTIVE flush.
1514 */
1515void ltt_force_switch_lockless_slow(struct ust_buffer *buf,
1516 enum force_switch_mode mode)
1517{
1518 struct ust_channel *chan = buf->chan;
1519 struct ltt_reserve_switch_offsets offsets;
1520 u64 tsc;
1521
1522 offsets.size = 0;
1523
1524 /*
1525 * Perform retryable operations.
1526 */
1527 do {
1528 if (ltt_relay_try_switch_slow(mode, chan, buf,
1529 &offsets, &tsc))
1530 return;
1531 } while (local_cmpxchg(&buf->offset, offsets.old,
1532 offsets.end) != offsets.old);
1533
1534 /*
1535 * Atomically update last_tsc. This update races against concurrent
1536 * atomic updates, but the race will always cause supplementary full TSC
1537 * events, never the opposite (missing a full TSC event when it would be
1538 * needed).
1539 */
1540 save_last_tsc(buf, tsc);
1541
1542 /*
1543 * Push the reader if necessary
1544 */
1545 if (mode == FORCE_ACTIVE) {
1546 ltt_reserve_push_reader(chan, buf, offsets.end - 1);
1547//ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.end - 1, chan));
1548 }
1549
1550 /*
1551 * Switch old subbuffer if needed.
1552 */
1553 if (offsets.end_switch_old) {
1554//ust// ltt_clear_noref_flag(rchan, buf, SUBBUF_INDEX(offsets.old - 1, rchan));
1555 ltt_reserve_switch_old_subbuf(chan, buf, &offsets, &tsc);
1556 }
1557
1558 /*
1559 * Populate new subbuffer.
1560 */
1561 if (mode == FORCE_ACTIVE)
1562 ltt_reserve_switch_new_subbuf(chan, buf, &offsets, &tsc);
1563}
1564
1565/*
1566 * Returns :
1567 * 0 if ok
1568 * !0 if execution must be aborted.
1569 */
1570static int ltt_relay_try_reserve_slow(struct ust_channel *chan, struct ust_buffer *buf,
1571 struct ltt_reserve_switch_offsets *offsets, size_t data_size,
1572 u64 *tsc, unsigned int *rflags, int largest_align)
1573{
1574 long reserve_commit_diff;
1575
1576 offsets->begin = local_read(&buf->offset);
1577 offsets->old = offsets->begin;
1578 offsets->begin_switch = 0;
1579 offsets->end_switch_current = 0;
1580 offsets->end_switch_old = 0;
1581
1582 *tsc = trace_clock_read64();
1583 if (last_tsc_overflow(buf, *tsc))
1584 *rflags = LTT_RFLAG_ID_SIZE_TSC;
1585
1586 if (unlikely(SUBBUF_OFFSET(offsets->begin, buf->chan) == 0)) {
1587 offsets->begin_switch = 1; /* For offsets->begin */
1588 } else {
1589 offsets->size = ust_get_header_size(chan,
1590 offsets->begin, data_size,
1591 &offsets->before_hdr_pad, *rflags);
1592 offsets->size += ltt_align(offsets->begin + offsets->size,
1593 largest_align)
1594 + data_size;
1595 if (unlikely((SUBBUF_OFFSET(offsets->begin, buf->chan) +
1596 offsets->size) > buf->chan->subbuf_size)) {
1597 offsets->end_switch_old = 1; /* For offsets->old */
1598 offsets->begin_switch = 1; /* For offsets->begin */
1599 }
1600 }
1601 if (unlikely(offsets->begin_switch)) {
1602 long subbuf_index;
1603
1604 /*
1605 * We are typically not filling the previous buffer completely.
1606 */
1607 if (likely(offsets->end_switch_old))
1608 offsets->begin = SUBBUF_ALIGN(offsets->begin,
1609 buf->chan);
1610 offsets->begin = offsets->begin + ltt_subbuffer_header_size();
1611 /* Test new buffer integrity */
1612 subbuf_index = SUBBUF_INDEX(offsets->begin, buf->chan);
1613 reserve_commit_diff =
1614 (BUFFER_TRUNC(offsets->begin, buf->chan)
1615 >> chan->n_subbufs_order)
1616 - (local_read(&buf->commit_count[subbuf_index].cc_sb)
1617 & chan->commit_count_mask);
1618 if (likely(reserve_commit_diff == 0)) {
1619 /* Next buffer not corrupted. */
1620 if (unlikely(!chan->overwrite &&
1621 (SUBBUF_TRUNC(offsets->begin, buf->chan)
1622 - SUBBUF_TRUNC(atomic_long_read(
1623 &buf->consumed),
1624 buf->chan))
1625 >= chan->alloc_size)) {
1626 /*
1627 * We do not overwrite non consumed buffers
1628 * and we are full : event is lost.
1629 */
1630 local_inc(&buf->events_lost);
1631 return -1;
1632 } else {
1633 /*
1634 * next buffer not corrupted, we are either in
1635 * overwrite mode or the buffer is not full.
1636 * It's safe to write in this new subbuffer.
1637 */
1638 }
1639 } else {
1640 /*
1641 * Next subbuffer corrupted. Drop event in normal and
1642 * overwrite mode. Caused by either a writer OOPS or
1643 * too many nested writes over a reserve/commit pair.
1644 */
1645 local_inc(&buf->events_lost);
1646 return -1;
1647 }
1648 offsets->size = ust_get_header_size(chan,
1649 offsets->begin, data_size,
1650 &offsets->before_hdr_pad, *rflags);
1651 offsets->size += ltt_align(offsets->begin + offsets->size,
1652 largest_align)
1653 + data_size;
1654 if (unlikely((SUBBUF_OFFSET(offsets->begin, buf->chan)
1655 + offsets->size) > buf->chan->subbuf_size)) {
1656 /*
1657 * Event too big for subbuffers, report error, don't
1658 * complete the sub-buffer switch.
1659 */
1660 local_inc(&buf->events_lost);
1661 return -1;
1662 } else {
1663 /*
1664 * We just made a successful buffer switch and the event
1665 * fits in the new subbuffer. Let's write.
1666 */
1667 }
1668 } else {
1669 /*
1670 * Event fits in the current buffer and we are not on a switch
1671 * boundary. It's safe to write.
1672 */
1673 }
1674 offsets->end = offsets->begin + offsets->size;
1675
1676 if (unlikely((SUBBUF_OFFSET(offsets->end, buf->chan)) == 0)) {
1677 /*
1678 * The offset_end will fall at the very beginning of the next
1679 * subbuffer.
1680 */
1681 offsets->end_switch_current = 1; /* For offsets->begin */
1682 }
1683 return 0;
1684}
1685
1686/**
1687 * ltt_relay_reserve_slot_lockless_slow - Atomic slot reservation in a buffer.
1688 * @trace: the trace structure to log to.
1689 * @ltt_channel: channel structure
1690 * @transport_data: data structure specific to ltt relay
1691 * @data_size: size of the variable length data to log.
1692 * @slot_size: pointer to total size of the slot (out)
1693 * @buf_offset : pointer to reserved buffer offset (out)
1694 * @tsc: pointer to the tsc at the slot reservation (out)
1695 * @cpu: cpuid
1696 *
1697 * Return : -ENOSPC if not enough space, else returns 0.
1698 * It will take care of sub-buffer switching.
1699 */
1700int ltt_reserve_slot_lockless_slow(struct ust_trace *trace,
1701 struct ust_channel *chan, void **transport_data,
1702 size_t data_size, size_t *slot_size, long *buf_offset, u64 *tsc,
1703 unsigned int *rflags, int largest_align, int cpu)
1704{
1705 struct ust_buffer *buf = chan->buf[cpu];
1706 struct ltt_reserve_switch_offsets offsets;
1707
1708 offsets.size = 0;
1709
1710 do {
1711 if (unlikely(ltt_relay_try_reserve_slow(chan, buf, &offsets,
1712 data_size, tsc, rflags, largest_align)))
1713 return -ENOSPC;
1714 } while (unlikely(local_cmpxchg(&buf->offset, offsets.old,
1715 offsets.end) != offsets.old));
1716
1717 /*
1718 * Atomically update last_tsc. This update races against concurrent
1719 * atomic updates, but the race will always cause supplementary full TSC
1720 * events, never the opposite (missing a full TSC event when it would be
1721 * needed).
1722 */
1723 save_last_tsc(buf, *tsc);
1724
1725 /*
1726 * Push the reader if necessary
1727 */
1728 ltt_reserve_push_reader(chan, buf, offsets.end - 1);
1729
1730 /*
1731 * Clear noref flag for this subbuffer.
1732 */
1733//ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.end - 1, chan));
1734
1735 /*
1736 * Switch old subbuffer if needed.
1737 */
1738 if (unlikely(offsets.end_switch_old)) {
1739//ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.old - 1, chan));
1740 ltt_reserve_switch_old_subbuf(chan, buf, &offsets, tsc);
1741 }
1742
1743 /*
1744 * Populate new subbuffer.
1745 */
1746 if (unlikely(offsets.begin_switch))
1747 ltt_reserve_switch_new_subbuf(chan, buf, &offsets, tsc);
1748
1749 if (unlikely(offsets.end_switch_current))
1750 ltt_reserve_end_switch_current(chan, buf, &offsets, tsc);
1751
1752 *slot_size = offsets.size;
1753 *buf_offset = offsets.begin + offsets.before_hdr_pad;
1754 return 0;
1755}
1756
1757static struct ltt_transport ust_relay_transport = {
1758 .name = "ustrelay",
1759 .ops = {
1760 .create_channel = ust_buffers_create_channel,
1761 .finish_channel = ltt_relay_finish_channel,
1762 .remove_channel = ltt_relay_remove_channel,
1763 .wakeup_channel = ltt_relay_async_wakeup_chan,
1764 },
1765};
1766
1767static char initialized = 0;
1768
1769void __attribute__((constructor)) init_ustrelay_transport(void)
1770{
1771 if(!initialized) {
1772 ltt_transport_register(&ust_relay_transport);
1773 initialized = 1;
1774 }
1775}
1776
1777static void __attribute__((destructor)) ust_buffers_exit(void)
1778{
1779 ltt_transport_unregister(&ust_relay_transport);
1780}
1781
1782size_t ltt_write_event_header_slow(struct ust_trace *trace,
1783 struct ust_channel *channel,
1784 struct ust_buffer *buf, long buf_offset,
1785 u16 eID, u32 event_size,
1786 u64 tsc, unsigned int rflags)
1787{
1788 struct ltt_event_header header;
1789 u16 small_size;
1790
1791 switch (rflags) {
1792 case LTT_RFLAG_ID_SIZE_TSC:
1793 header.id_time = 29 << LTT_TSC_BITS;
1794 break;
1795 case LTT_RFLAG_ID_SIZE:
1796 header.id_time = 30 << LTT_TSC_BITS;
1797 break;
1798 case LTT_RFLAG_ID:
1799 header.id_time = 31 << LTT_TSC_BITS;
1800 break;
1801 }
1802
1803 header.id_time |= (u32)tsc & LTT_TSC_MASK;
1804 ust_buffers_write(buf, buf_offset, &header, sizeof(header));
1805 buf_offset += sizeof(header);
1806
1807 switch (rflags) {
1808 case LTT_RFLAG_ID_SIZE_TSC:
1809 small_size = (u16)min_t(u32, event_size, LTT_MAX_SMALL_SIZE);
1810 ust_buffers_write(buf, buf_offset,
1811 &eID, sizeof(u16));
1812 buf_offset += sizeof(u16);
1813 ust_buffers_write(buf, buf_offset,
1814 &small_size, sizeof(u16));
1815 buf_offset += sizeof(u16);
1816 if (small_size == LTT_MAX_SMALL_SIZE) {
1817 ust_buffers_write(buf, buf_offset,
1818 &event_size, sizeof(u32));
1819 buf_offset += sizeof(u32);
1820 }
1821 buf_offset += ltt_align(buf_offset, sizeof(u64));
1822 ust_buffers_write(buf, buf_offset,
1823 &tsc, sizeof(u64));
1824 buf_offset += sizeof(u64);
1825 break;
1826 case LTT_RFLAG_ID_SIZE:
1827 small_size = (u16)min_t(u32, event_size, LTT_MAX_SMALL_SIZE);
1828 ust_buffers_write(buf, buf_offset,
1829 &eID, sizeof(u16));
1830 buf_offset += sizeof(u16);
1831 ust_buffers_write(buf, buf_offset,
1832 &small_size, sizeof(u16));
1833 buf_offset += sizeof(u16);
1834 if (small_size == LTT_MAX_SMALL_SIZE) {
1835 ust_buffers_write(buf, buf_offset,
1836 &event_size, sizeof(u32));
1837 buf_offset += sizeof(u32);
1838 }
1839 break;
1840 case LTT_RFLAG_ID:
1841 ust_buffers_write(buf, buf_offset,
1842 &eID, sizeof(u16));
1843 buf_offset += sizeof(u16);
1844 break;
1845 }
1846
1847 return buf_offset;
1848}
UST 0.x - Userspace Tracer
This page took 0.03097 seconds and 4 git commands to generate.