3 * LTTng userspace tracer buffering system
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)
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.
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.
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
30 #include <ust/clock.h>
35 #include "tracercore.h"
38 struct ltt_reserve_switch_offsets
{
40 long begin_switch
, end_switch_current
, end_switch_old
;
41 size_t before_hdr_pad
, size
;
45 static DEFINE_MUTEX(ust_buffers_channels_mutex
);
46 static CDS_LIST_HEAD(ust_buffers_channels
);
48 static int get_n_cpus(void)
51 static int n_cpus
= 0;
54 /* On Linux, when some processors are offline
55 * _SC_NPROCESSORS_CONF counts the offline
56 * processors, whereas _SC_NPROCESSORS_ONLN
57 * does not. If we used _SC_NPROCESSORS_ONLN,
58 * getcpu() could return a value greater than
59 * this sysconf, in which case the arrays
60 * indexed by processor would overflow.
62 result
= sysconf(_SC_NPROCESSORS_CONF
);
74 * _ust_buffers_strncpy_fixup - Fix an incomplete string in a ltt_relay buffer.
76 * @offset : offset within the buffer
77 * @len : length to write
78 * @copied: string actually copied
79 * @terminated: does string end with \0
81 * Fills string with "X" if incomplete.
83 void _ust_buffers_strncpy_fixup(struct ust_buffer
*buf
, size_t offset
,
84 size_t len
, size_t copied
, int terminated
)
86 size_t buf_offset
, cpy
;
90 * Deal with non-terminated string.
94 buf_offset
= BUFFER_OFFSET(offset
, buf
->chan
);
96 * Underlying layer should never ask for writes across
100 < buf
->chan
->subbuf_size
*buf
->chan
->subbuf_cnt
);
101 ust_buffers_do_memset(buf
->buf_data
+ buf_offset
, '\0', 1);
106 * Deal with incomplete string.
107 * Overwrite string's \0 with X too.
113 buf_offset
= BUFFER_OFFSET(offset
, buf
->chan
);
116 * Underlying layer should never ask for writes across subbuffers.
119 < buf
->chan
->subbuf_size
*buf
->chan
->subbuf_cnt
);
121 ust_buffers_do_memset(buf
->buf_data
+ buf_offset
,
125 * Overwrite last 'X' with '\0'.
128 buf_offset
= BUFFER_OFFSET(offset
, buf
->chan
);
130 * Underlying layer should never ask for writes across subbuffers.
133 < buf
->chan
->subbuf_size
*buf
->chan
->subbuf_cnt
);
134 ust_buffers_do_memset(buf
->buf_data
+ buf_offset
, '\0', 1);
137 static void ltt_buffer_begin(struct ust_buffer
*buf
,
138 u64 tsc
, unsigned int subbuf_idx
)
140 struct ust_channel
*channel
= buf
->chan
;
141 struct ltt_subbuffer_header
*header
=
142 (struct ltt_subbuffer_header
*)
143 ust_buffers_offset_address(buf
,
144 subbuf_idx
* buf
->chan
->subbuf_size
);
146 header
->cycle_count_begin
= tsc
;
147 header
->data_size
= 0xFFFFFFFF; /* for recognizing crashed buffers */
148 header
->sb_size
= 0xFFFFFFFF; /* for recognizing crashed buffers */
149 /* FIXME: add memory barrier? */
150 ltt_write_trace_header(channel
->trace
, header
);
153 static int map_buf_data(struct ust_buffer
*buf
, size_t *size
)
158 *size
= PAGE_ALIGN(*size
);
160 result
= buf
->shmid
= shmget(getpid(), *size
, IPC_CREAT
| IPC_EXCL
| 0700);
161 if (result
< 0 && errno
== EINVAL
) {
162 ERR("shmget() returned EINVAL; maybe /proc/sys/kernel/shmmax should be increased.");
164 } else if (result
< 0) {
169 ptr
= shmat(buf
->shmid
, NULL
, 0);
170 if (ptr
== (void *) -1) {
175 /* Already mark the shared memory for destruction. This will occur only
176 * when all users have detached.
178 result
= shmctl(buf
->shmid
, IPC_RMID
, NULL
);
185 buf
->buf_size
= *size
;
190 result
= shmctl(buf
->shmid
, IPC_RMID
, NULL
);
198 static int open_buf(struct ust_channel
*chan
, int cpu
)
202 struct ust_trace
*trace
= chan
->trace
;
203 struct ust_buffer
*buf
= chan
->buf
[cpu
];
204 unsigned int n_subbufs
= chan
->subbuf_cnt
;
207 result
= map_buf_data(buf
, &chan
->alloc_size
);
212 zmalloc(sizeof(*buf
->commit_count
) * n_subbufs
);
213 if (!buf
->commit_count
)
219 goto free_commit_count
;
221 buf
->data_ready_fd_read
= fds
[0];
222 buf
->data_ready_fd_write
= fds
[1];
227 uatomic_set(&buf
->offset
, ltt_subbuffer_header_size());
228 uatomic_set(&buf
->consumed
, 0);
229 uatomic_set(&buf
->active_readers
, 0);
230 for (j
= 0; j
< n_subbufs
; j
++) {
231 uatomic_set(&buf
->commit_count
[j
].cc
, 0);
232 uatomic_set(&buf
->commit_count
[j
].cc_sb
, 0);
235 ltt_buffer_begin(buf
, trace
->start_tsc
, 0);
237 uatomic_add(&buf
->commit_count
[0].cc
, ltt_subbuffer_header_size());
239 uatomic_set(&buf
->events_lost
, 0);
240 uatomic_set(&buf
->corrupted_subbuffers
, 0);
242 memset(buf
->commit_seq
, 0, sizeof(buf
->commit_seq
[0]) * n_subbufs
);
247 free(buf
->commit_count
);
250 if (shmdt(buf
->buf_data
) < 0) {
251 PERROR("shmdt failed");
257 static void close_buf(struct ust_buffer
*buf
)
261 result
= shmdt(buf
->buf_data
);
266 result
= close(buf
->data_ready_fd_read
);
271 result
= close(buf
->data_ready_fd_write
);
272 if (result
< 0 && errno
!= EBADF
) {
278 static int open_channel(struct ust_channel
*chan
, size_t subbuf_size
,
284 if(subbuf_size
== 0 || subbuf_cnt
== 0)
287 /* Check that the subbuffer size is larger than a page. */
288 WARN_ON_ONCE(subbuf_size
< PAGE_SIZE
);
291 * Make sure the number of subbuffers and subbuffer size are power of 2.
293 WARN_ON_ONCE(hweight32(subbuf_size
) != 1);
294 WARN_ON(hweight32(subbuf_cnt
) != 1);
296 chan
->version
= UST_CHANNEL_VERSION
;
297 chan
->subbuf_cnt
= subbuf_cnt
;
298 chan
->subbuf_size
= subbuf_size
;
299 chan
->subbuf_size_order
= get_count_order(subbuf_size
);
300 chan
->alloc_size
= subbuf_size
* subbuf_cnt
;
302 pthread_mutex_lock(&ust_buffers_channels_mutex
);
303 for (i
=0; i
< chan
->n_cpus
; i
++) {
304 result
= open_buf(chan
, i
);
308 cds_list_add(&chan
->list
, &ust_buffers_channels
);
309 pthread_mutex_unlock(&ust_buffers_channels_mutex
);
315 for(i
--; i
>= 0; i
--)
316 close_buf(chan
->buf
[i
]);
318 pthread_mutex_unlock(&ust_buffers_channels_mutex
);
322 static void close_channel(struct ust_channel
*chan
)
328 pthread_mutex_lock(&ust_buffers_channels_mutex
);
329 for(i
=0; i
<chan
->n_cpus
; i
++) {
330 /* FIXME: if we make it here, then all buffers were necessarily allocated. Moreover, we don't
331 * initialize to NULL so we cannot use this check. Should we? */
332 //ust// if (chan->buf[i])
333 close_buf(chan
->buf
[i
]);
336 cds_list_del(&chan
->list
);
338 pthread_mutex_unlock(&ust_buffers_channels_mutex
);
341 static void ltt_force_switch(struct ust_buffer
*buf
,
342 enum force_switch_mode mode
);
347 * offset is assumed to never be 0 here : never deliver a completely empty
348 * subbuffer. The lost size is between 0 and subbuf_size-1.
350 static notrace
void ltt_buffer_end(struct ust_buffer
*buf
,
351 u64 tsc
, unsigned int offset
, unsigned int subbuf_idx
)
353 struct ltt_subbuffer_header
*header
=
354 (struct ltt_subbuffer_header
*)
355 ust_buffers_offset_address(buf
,
356 subbuf_idx
* buf
->chan
->subbuf_size
);
357 u32 data_size
= SUBBUF_OFFSET(offset
- 1, buf
->chan
) + 1;
359 header
->data_size
= data_size
;
360 header
->sb_size
= PAGE_ALIGN(data_size
);
361 header
->cycle_count_end
= tsc
;
362 header
->events_lost
= uatomic_read(&buf
->events_lost
);
363 header
->subbuf_corrupt
= uatomic_read(&buf
->corrupted_subbuffers
);
364 if(unlikely(header
->events_lost
> 0)) {
365 DBG("Some events (%d) were lost in %s_%d", header
->events_lost
, buf
->chan
->channel_name
, buf
->cpu
);
370 * This function should not be called from NMI interrupt context
372 static notrace
void ltt_buf_unfull(struct ust_buffer
*buf
,
373 unsigned int subbuf_idx
,
379 * Promote compiler cmm_barrier to a smp_mb().
380 * For the specific LTTng case, this IPI call should be removed if the
381 * architecture does not reorder writes. This should eventually be provided by
382 * a separate architecture-specific infrastructure.
384 //ust// static void remote_mb(void *info)
389 int ust_buffers_get_subbuf(struct ust_buffer
*buf
, long *consumed
)
391 struct ust_channel
*channel
= buf
->chan
;
392 long consumed_old
, consumed_idx
, commit_count
, write_offset
;
395 consumed_old
= uatomic_read(&buf
->consumed
);
396 consumed_idx
= SUBBUF_INDEX(consumed_old
, buf
->chan
);
397 commit_count
= uatomic_read(&buf
->commit_count
[consumed_idx
].cc_sb
);
399 * Make sure we read the commit count before reading the buffer
400 * data and the write offset. Correct consumed offset ordering
401 * wrt commit count is insured by the use of cmpxchg to update
402 * the consumed offset.
403 * smp_call_function_single can fail if the remote CPU is offline,
404 * this is OK because then there is no wmb to execute there.
405 * If our thread is executing on the same CPU as the on the buffers
406 * belongs to, we don't have to synchronize it at all. If we are
407 * migrated, the scheduler will take care of the memory cmm_barriers.
408 * Normally, smp_call_function_single() should ensure program order when
409 * executing the remote function, which implies that it surrounds the
410 * function execution with :
421 * However, smp_call_function_single() does not seem to clearly execute
422 * such barriers. It depends on spinlock semantic to provide the barrier
423 * before executing the IPI and, when busy-looping, csd_lock_wait only
424 * executes smp_mb() when it has to wait for the other CPU.
426 * I don't trust this code. Therefore, let's add the smp_mb() sequence
427 * required ourself, even if duplicated. It has no performance impact
430 * smp_mb() is needed because cmm_smp_rmb() and cmm_smp_wmb() only order read vs
431 * read and write vs write. They do not ensure core synchronization. We
432 * really have to ensure total order between the 3 cmm_barriers running on
435 //ust// #ifdef LTT_NO_IPI_BARRIER
437 * Local rmb to match the remote wmb to read the commit count before the
438 * buffer data and the write offset.
442 //ust// if (raw_smp_processor_id() != buf->cpu) {
443 //ust// smp_mb(); /* Total order with IPI handler smp_mb() */
444 //ust// smp_call_function_single(buf->cpu, remote_mb, NULL, 1);
445 //ust// smp_mb(); /* Total order with IPI handler smp_mb() */
449 write_offset
= uatomic_read(&buf
->offset
);
451 * Check that the subbuffer we are trying to consume has been
452 * already fully committed.
454 if (((commit_count
- buf
->chan
->subbuf_size
)
455 & channel
->commit_count_mask
)
456 - (BUFFER_TRUNC(consumed_old
, buf
->chan
)
457 >> channel
->n_subbufs_order
)
462 * Check that we are not about to read the same subbuffer in
463 * which the writer head is.
465 if ((SUBBUF_TRUNC(write_offset
, buf
->chan
)
466 - SUBBUF_TRUNC(consumed_old
, buf
->chan
))
471 /* FIXME: is this ok to disable the reading feature? */
472 //ust// retval = update_read_sb_index(buf, consumed_idx);
474 //ust// return retval;
476 *consumed
= consumed_old
;
481 int ust_buffers_put_subbuf(struct ust_buffer
*buf
, unsigned long uconsumed_old
)
483 long consumed_new
, consumed_old
;
485 consumed_old
= uatomic_read(&buf
->consumed
);
486 consumed_old
= consumed_old
& (~0xFFFFFFFFL
);
487 consumed_old
= consumed_old
| uconsumed_old
;
488 consumed_new
= SUBBUF_ALIGN(consumed_old
, buf
->chan
);
490 //ust// spin_lock(<t_buf->full_lock);
491 if (uatomic_cmpxchg(&buf
->consumed
, consumed_old
,
494 /* We have been pushed by the writer : the last
495 * buffer read _is_ corrupted! It can also
496 * happen if this is a buffer we never got. */
497 //ust// spin_unlock(<t_buf->full_lock);
500 /* tell the client that buffer is now unfull */
503 index
= SUBBUF_INDEX(consumed_old
, buf
->chan
);
504 data
= BUFFER_OFFSET(consumed_old
, buf
->chan
);
505 ltt_buf_unfull(buf
, index
, data
);
506 //ust// spin_unlock(<t_buf->full_lock);
511 static int map_buf_structs(struct ust_channel
*chan
)
518 size
= PAGE_ALIGN(1);
520 for(i
=0; i
<chan
->n_cpus
; i
++) {
522 result
= chan
->buf_struct_shmids
[i
] = shmget(getpid(), size
, IPC_CREAT
| IPC_EXCL
| 0700);
525 goto destroy_previous
;
528 ptr
= shmat(chan
->buf_struct_shmids
[i
], NULL
, 0);
529 if(ptr
== (void *) -1) {
534 /* Already mark the shared memory for destruction. This will occur only
535 * when all users have detached.
537 result
= shmctl(chan
->buf_struct_shmids
[i
], IPC_RMID
, NULL
);
540 goto destroy_previous
;
548 /* Jumping inside this loop occurs from within the other loop above with i as
549 * counter, so it unallocates the structures for the cpu = current_i down to
553 result
= shmctl(chan
->buf_struct_shmids
[i
], IPC_RMID
, NULL
);
565 static int unmap_buf_structs(struct ust_channel
*chan
)
569 for (i
=0; i
< chan
->n_cpus
; i
++) {
570 if (shmdt(chan
->buf
[i
]) < 0) {
580 static int create_channel(const char *trace_name
, struct ust_trace
*trace
,
581 const char *channel_name
, struct ust_channel
*chan
,
582 unsigned int subbuf_size
, unsigned int n_subbufs
, int overwrite
)
587 chan
->overwrite
= overwrite
;
588 chan
->n_subbufs_order
= get_count_order(n_subbufs
);
589 chan
->commit_count_mask
= (~0UL >> chan
->n_subbufs_order
);
590 chan
->n_cpus
= get_n_cpus();
592 /* These mappings should ideall be per-cpu, if somebody can do that
593 * from userspace, that would be cool!
595 chan
->buf
= (void *) zmalloc(chan
->n_cpus
* sizeof(void *));
596 if(chan
->buf
== NULL
) {
599 chan
->buf_struct_shmids
= (int *) zmalloc(chan
->n_cpus
* sizeof(int));
600 if(chan
->buf_struct_shmids
== NULL
)
603 result
= map_buf_structs(chan
);
605 goto free_buf_struct_shmids
;
608 result
= open_channel(chan
, subbuf_size
, n_subbufs
);
610 ERR("Cannot open channel for trace %s", trace_name
);
611 goto unmap_buf_structs
;
617 for (i
=0; i
< chan
->n_cpus
; i
++) {
618 if (shmdt(chan
->buf
[i
]) < 0) {
619 PERROR("shmdt bufstruct");
623 free_buf_struct_shmids
:
624 free(chan
->buf_struct_shmids
);
634 static void remove_channel(struct ust_channel
*chan
)
638 unmap_buf_structs(chan
);
640 free(chan
->buf_struct_shmids
);
645 static void ltt_relay_async_wakeup_chan(struct ust_channel
*ltt_channel
)
647 //ust// unsigned int i;
648 //ust// struct rchan *rchan = ltt_channel->trans_channel_data;
650 //ust// for_each_possible_cpu(i) {
651 //ust// struct ltt_channel_buf_struct *ltt_buf =
652 //ust// percpu_ptr(ltt_channel->buf, i);
654 //ust// if (uatomic_read(<t_buf->wakeup_readers) == 1) {
655 //ust// uatomic_set(<t_buf->wakeup_readers, 0);
656 //ust// wake_up_interruptible(&rchan->buf[i]->read_wait);
661 static void ltt_relay_finish_buffer(struct ust_channel
*channel
, unsigned int cpu
)
665 if (channel
->buf
[cpu
]) {
666 struct ust_buffer
*buf
= channel
->buf
[cpu
];
667 ltt_force_switch(buf
, FORCE_FLUSH
);
669 /* closing the pipe tells the consumer the buffer is finished */
670 close(buf
->data_ready_fd_write
);
675 static void finish_channel(struct ust_channel
*channel
)
679 for(i
=0; i
<channel
->n_cpus
; i
++) {
680 ltt_relay_finish_buffer(channel
, i
);
686 * ltt_reserve_switch_old_subbuf: switch old subbuffer
688 * Concurrency safe because we are the last and only thread to alter this
689 * sub-buffer. As long as it is not delivered and read, no other thread can
690 * alter the offset, alter the reserve_count or call the
691 * client_buffer_end_callback on this sub-buffer.
693 * The only remaining threads could be the ones with pending commits. They will
694 * have to do the deliver themselves. Not concurrency safe in overwrite mode.
695 * We detect corrupted subbuffers with commit and reserve counts. We keep a
696 * corrupted sub-buffers count and push the readers across these sub-buffers.
698 * Not concurrency safe if a writer is stalled in a subbuffer and another writer
699 * switches in, finding out it's corrupted. The result will be than the old
700 * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer
701 * will be declared corrupted too because of the commit count adjustment.
703 * Note : offset_old should never be 0 here.
705 static void ltt_reserve_switch_old_subbuf(
706 struct ust_channel
*chan
, struct ust_buffer
*buf
,
707 struct ltt_reserve_switch_offsets
*offsets
, u64
*tsc
)
709 long oldidx
= SUBBUF_INDEX(offsets
->old
- 1, chan
);
710 long commit_count
, padding_size
;
712 padding_size
= chan
->subbuf_size
713 - (SUBBUF_OFFSET(offsets
->old
- 1, chan
) + 1);
714 ltt_buffer_end(buf
, *tsc
, offsets
->old
, oldidx
);
717 * Must write slot data before incrementing commit count.
718 * This compiler barrier is upgraded into a cmm_smp_wmb() by the IPI
719 * sent by get_subbuf() when it does its cmm_smp_rmb().
722 uatomic_add(&buf
->commit_count
[oldidx
].cc
, padding_size
);
723 commit_count
= uatomic_read(&buf
->commit_count
[oldidx
].cc
);
724 ltt_check_deliver(chan
, buf
, offsets
->old
- 1, commit_count
, oldidx
);
725 ltt_write_commit_counter(chan
, buf
, oldidx
,
726 offsets
->old
, commit_count
, padding_size
);
730 * ltt_reserve_switch_new_subbuf: Populate new subbuffer.
732 * This code can be executed unordered : writers may already have written to the
733 * sub-buffer before this code gets executed, caution. The commit makes sure
734 * that this code is executed before the deliver of this sub-buffer.
736 static void ltt_reserve_switch_new_subbuf(
737 struct ust_channel
*chan
, struct ust_buffer
*buf
,
738 struct ltt_reserve_switch_offsets
*offsets
, u64
*tsc
)
740 long beginidx
= SUBBUF_INDEX(offsets
->begin
, chan
);
743 ltt_buffer_begin(buf
, *tsc
, beginidx
);
746 * Must write slot data before incrementing commit count.
747 * This compiler barrier is upgraded into a cmm_smp_wmb() by the IPI
748 * sent by get_subbuf() when it does its cmm_smp_rmb().
751 uatomic_add(&buf
->commit_count
[beginidx
].cc
, ltt_subbuffer_header_size());
752 commit_count
= uatomic_read(&buf
->commit_count
[beginidx
].cc
);
753 /* Check if the written buffer has to be delivered */
754 ltt_check_deliver(chan
, buf
, offsets
->begin
, commit_count
, beginidx
);
755 ltt_write_commit_counter(chan
, buf
, beginidx
,
756 offsets
->begin
, commit_count
, ltt_subbuffer_header_size());
760 * ltt_reserve_end_switch_current: finish switching current subbuffer
762 * Concurrency safe because we are the last and only thread to alter this
763 * sub-buffer. As long as it is not delivered and read, no other thread can
764 * alter the offset, alter the reserve_count or call the
765 * client_buffer_end_callback on this sub-buffer.
767 * The only remaining threads could be the ones with pending commits. They will
768 * have to do the deliver themselves. Not concurrency safe in overwrite mode.
769 * We detect corrupted subbuffers with commit and reserve counts. We keep a
770 * corrupted sub-buffers count and push the readers across these sub-buffers.
772 * Not concurrency safe if a writer is stalled in a subbuffer and another writer
773 * switches in, finding out it's corrupted. The result will be than the old
774 * (uncommited) subbuffer will be declared corrupted, and that the new subbuffer
775 * will be declared corrupted too because of the commit count adjustment.
777 static void ltt_reserve_end_switch_current(
778 struct ust_channel
*chan
,
779 struct ust_buffer
*buf
,
780 struct ltt_reserve_switch_offsets
*offsets
, u64
*tsc
)
782 long endidx
= SUBBUF_INDEX(offsets
->end
- 1, chan
);
783 long commit_count
, padding_size
;
785 padding_size
= chan
->subbuf_size
786 - (SUBBUF_OFFSET(offsets
->end
- 1, chan
) + 1);
788 ltt_buffer_end(buf
, *tsc
, offsets
->end
, endidx
);
791 * Must write slot data before incrementing commit count.
792 * This compiler barrier is upgraded into a cmm_smp_wmb() by the IPI
793 * sent by get_subbuf() when it does its cmm_smp_rmb().
796 uatomic_add(&buf
->commit_count
[endidx
].cc
, padding_size
);
797 commit_count
= uatomic_read(&buf
->commit_count
[endidx
].cc
);
798 ltt_check_deliver(chan
, buf
,
799 offsets
->end
- 1, commit_count
, endidx
);
800 ltt_write_commit_counter(chan
, buf
, endidx
,
801 offsets
->end
, commit_count
, padding_size
);
807 * !0 if execution must be aborted.
809 static int ltt_relay_try_switch_slow(
810 enum force_switch_mode mode
,
811 struct ust_channel
*chan
,
812 struct ust_buffer
*buf
,
813 struct ltt_reserve_switch_offsets
*offsets
,
817 long reserve_commit_diff
;
819 offsets
->begin
= uatomic_read(&buf
->offset
);
820 offsets
->old
= offsets
->begin
;
821 offsets
->begin_switch
= 0;
822 offsets
->end_switch_old
= 0;
824 *tsc
= trace_clock_read64();
826 if (SUBBUF_OFFSET(offsets
->begin
, buf
->chan
) != 0) {
827 offsets
->begin
= SUBBUF_ALIGN(offsets
->begin
, buf
->chan
);
828 offsets
->end_switch_old
= 1;
830 /* we do not have to switch : buffer is empty */
833 if (mode
== FORCE_ACTIVE
)
834 offsets
->begin
+= ltt_subbuffer_header_size();
836 * Always begin_switch in FORCE_ACTIVE mode.
837 * Test new buffer integrity
839 subbuf_index
= SUBBUF_INDEX(offsets
->begin
, buf
->chan
);
840 reserve_commit_diff
=
841 (BUFFER_TRUNC(offsets
->begin
, buf
->chan
)
842 >> chan
->n_subbufs_order
)
843 - (uatomic_read(&buf
->commit_count
[subbuf_index
].cc_sb
)
844 & chan
->commit_count_mask
);
845 if (reserve_commit_diff
== 0) {
846 /* Next buffer not corrupted. */
847 if (mode
== FORCE_ACTIVE
849 && offsets
->begin
- uatomic_read(&buf
->consumed
)
850 >= chan
->alloc_size
) {
852 * We do not overwrite non consumed buffers and we are
853 * full : ignore switch while tracing is active.
859 * Next subbuffer corrupted. Force pushing reader even in normal
863 offsets
->end
= offsets
->begin
;
868 * Force a sub-buffer switch for a per-cpu buffer. This operation is
869 * completely reentrant : can be called while tracing is active with
870 * absolutely no lock held.
872 void ltt_force_switch_lockless_slow(struct ust_buffer
*buf
,
873 enum force_switch_mode mode
)
875 struct ust_channel
*chan
= buf
->chan
;
876 struct ltt_reserve_switch_offsets offsets
;
881 DBG("Switching (forced) %s_%d", chan
->channel_name
, buf
->cpu
);
883 * Perform retryable operations.
886 if (ltt_relay_try_switch_slow(mode
, chan
, buf
,
889 } while (uatomic_cmpxchg(&buf
->offset
, offsets
.old
,
890 offsets
.end
) != offsets
.old
);
893 * Atomically update last_tsc. This update races against concurrent
894 * atomic updates, but the race will always cause supplementary full TSC
895 * events, never the opposite (missing a full TSC event when it would be
898 save_last_tsc(buf
, tsc
);
901 * Push the reader if necessary
903 if (mode
== FORCE_ACTIVE
) {
904 ltt_reserve_push_reader(chan
, buf
, offsets
.end
- 1);
905 //ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.end - 1, chan));
909 * Switch old subbuffer if needed.
911 if (offsets
.end_switch_old
) {
912 //ust// ltt_clear_noref_flag(rchan, buf, SUBBUF_INDEX(offsets.old - 1, rchan));
913 ltt_reserve_switch_old_subbuf(chan
, buf
, &offsets
, &tsc
);
917 * Populate new subbuffer.
919 if (mode
== FORCE_ACTIVE
)
920 ltt_reserve_switch_new_subbuf(chan
, buf
, &offsets
, &tsc
);
926 * !0 if execution must be aborted.
928 static int ltt_relay_try_reserve_slow(struct ust_channel
*chan
, struct ust_buffer
*buf
,
929 struct ltt_reserve_switch_offsets
*offsets
, size_t data_size
,
930 u64
*tsc
, unsigned int *rflags
, int largest_align
)
932 long reserve_commit_diff
;
934 offsets
->begin
= uatomic_read(&buf
->offset
);
935 offsets
->old
= offsets
->begin
;
936 offsets
->begin_switch
= 0;
937 offsets
->end_switch_current
= 0;
938 offsets
->end_switch_old
= 0;
940 *tsc
= trace_clock_read64();
941 if (last_tsc_overflow(buf
, *tsc
))
942 *rflags
= LTT_RFLAG_ID_SIZE_TSC
;
944 if (unlikely(SUBBUF_OFFSET(offsets
->begin
, buf
->chan
) == 0)) {
945 offsets
->begin_switch
= 1; /* For offsets->begin */
947 offsets
->size
= ust_get_header_size(chan
,
948 offsets
->begin
, data_size
,
949 &offsets
->before_hdr_pad
, *rflags
);
950 offsets
->size
+= ltt_align(offsets
->begin
+ offsets
->size
,
953 if (unlikely((SUBBUF_OFFSET(offsets
->begin
, buf
->chan
) +
954 offsets
->size
) > buf
->chan
->subbuf_size
)) {
955 offsets
->end_switch_old
= 1; /* For offsets->old */
956 offsets
->begin_switch
= 1; /* For offsets->begin */
959 if (unlikely(offsets
->begin_switch
)) {
963 * We are typically not filling the previous buffer completely.
965 if (likely(offsets
->end_switch_old
))
966 offsets
->begin
= SUBBUF_ALIGN(offsets
->begin
,
968 offsets
->begin
= offsets
->begin
+ ltt_subbuffer_header_size();
969 /* Test new buffer integrity */
970 subbuf_index
= SUBBUF_INDEX(offsets
->begin
, buf
->chan
);
971 reserve_commit_diff
=
972 (BUFFER_TRUNC(offsets
->begin
, buf
->chan
)
973 >> chan
->n_subbufs_order
)
974 - (uatomic_read(&buf
->commit_count
[subbuf_index
].cc_sb
)
975 & chan
->commit_count_mask
);
976 if (likely(reserve_commit_diff
== 0)) {
977 /* Next buffer not corrupted. */
978 if (unlikely(!chan
->overwrite
&&
979 (SUBBUF_TRUNC(offsets
->begin
, buf
->chan
)
980 - SUBBUF_TRUNC(uatomic_read(
983 >= chan
->alloc_size
)) {
985 * We do not overwrite non consumed buffers
986 * and we are full : event is lost.
988 uatomic_inc(&buf
->events_lost
);
992 * next buffer not corrupted, we are either in
993 * overwrite mode or the buffer is not full.
994 * It's safe to write in this new subbuffer.
999 * Next subbuffer corrupted. Drop event in normal and
1000 * overwrite mode. Caused by either a writer OOPS or
1001 * too many nested writes over a reserve/commit pair.
1003 uatomic_inc(&buf
->events_lost
);
1006 offsets
->size
= ust_get_header_size(chan
,
1007 offsets
->begin
, data_size
,
1008 &offsets
->before_hdr_pad
, *rflags
);
1009 offsets
->size
+= ltt_align(offsets
->begin
+ offsets
->size
,
1012 if (unlikely((SUBBUF_OFFSET(offsets
->begin
, buf
->chan
)
1013 + offsets
->size
) > buf
->chan
->subbuf_size
)) {
1015 * Event too big for subbuffers, report error, don't
1016 * complete the sub-buffer switch.
1018 uatomic_inc(&buf
->events_lost
);
1022 * We just made a successful buffer switch and the event
1023 * fits in the new subbuffer. Let's write.
1028 * Event fits in the current buffer and we are not on a switch
1029 * boundary. It's safe to write.
1032 offsets
->end
= offsets
->begin
+ offsets
->size
;
1034 if (unlikely((SUBBUF_OFFSET(offsets
->end
, buf
->chan
)) == 0)) {
1036 * The offset_end will fall at the very beginning of the next
1039 offsets
->end_switch_current
= 1; /* For offsets->begin */
1045 * ltt_relay_reserve_slot_lockless_slow - Atomic slot reservation in a buffer.
1046 * @trace: the trace structure to log to.
1047 * @ltt_channel: channel structure
1048 * @transport_data: data structure specific to ltt relay
1049 * @data_size: size of the variable length data to log.
1050 * @slot_size: pointer to total size of the slot (out)
1051 * @buf_offset : pointer to reserved buffer offset (out)
1052 * @tsc: pointer to the tsc at the slot reservation (out)
1055 * Return : -ENOSPC if not enough space, else returns 0.
1056 * It will take care of sub-buffer switching.
1058 int ltt_reserve_slot_lockless_slow(struct ust_channel
*chan
,
1059 struct ust_trace
*trace
, size_t data_size
,
1060 int largest_align
, int cpu
,
1061 struct ust_buffer
**ret_buf
,
1062 size_t *slot_size
, long *buf_offset
,
1063 u64
*tsc
, unsigned int *rflags
)
1065 struct ust_buffer
*buf
= *ret_buf
= chan
->buf
[cpu
];
1066 struct ltt_reserve_switch_offsets offsets
;
1071 if (unlikely(ltt_relay_try_reserve_slow(chan
, buf
, &offsets
,
1072 data_size
, tsc
, rflags
, largest_align
)))
1074 } while (unlikely(uatomic_cmpxchg(&buf
->offset
, offsets
.old
,
1075 offsets
.end
) != offsets
.old
));
1078 * Atomically update last_tsc. This update races against concurrent
1079 * atomic updates, but the race will always cause supplementary full TSC
1080 * events, never the opposite (missing a full TSC event when it would be
1083 save_last_tsc(buf
, *tsc
);
1086 * Push the reader if necessary
1088 ltt_reserve_push_reader(chan
, buf
, offsets
.end
- 1);
1091 * Clear noref flag for this subbuffer.
1093 //ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.end - 1, chan));
1096 * Switch old subbuffer if needed.
1098 if (unlikely(offsets
.end_switch_old
)) {
1099 //ust// ltt_clear_noref_flag(chan, buf, SUBBUF_INDEX(offsets.old - 1, chan));
1100 ltt_reserve_switch_old_subbuf(chan
, buf
, &offsets
, tsc
);
1101 DBG("Switching %s_%d", chan
->channel_name
, cpu
);
1105 * Populate new subbuffer.
1107 if (unlikely(offsets
.begin_switch
))
1108 ltt_reserve_switch_new_subbuf(chan
, buf
, &offsets
, tsc
);
1110 if (unlikely(offsets
.end_switch_current
))
1111 ltt_reserve_end_switch_current(chan
, buf
, &offsets
, tsc
);
1113 *slot_size
= offsets
.size
;
1114 *buf_offset
= offsets
.begin
+ offsets
.before_hdr_pad
;
1118 static struct ltt_transport ust_relay_transport
= {
1121 .create_channel
= create_channel
,
1122 .finish_channel
= finish_channel
,
1123 .remove_channel
= remove_channel
,
1124 .wakeup_channel
= ltt_relay_async_wakeup_chan
,
1128 static char initialized
= 0;
1130 void __attribute__((constructor
)) init_ustrelay_transport(void)
1133 ltt_transport_register(&ust_relay_transport
);
1138 static void __attribute__((destructor
)) ust_buffers_exit(void)
1140 ltt_transport_unregister(&ust_relay_transport
);
1143 size_t ltt_write_event_header_slow(struct ust_channel
*channel
,
1144 struct ust_buffer
*buf
, long buf_offset
,
1145 u16 eID
, u32 event_size
,
1146 u64 tsc
, unsigned int rflags
)
1148 struct ltt_event_header header
;
1152 case LTT_RFLAG_ID_SIZE_TSC
:
1153 header
.id_time
= 29 << LTT_TSC_BITS
;
1155 case LTT_RFLAG_ID_SIZE
:
1156 header
.id_time
= 30 << LTT_TSC_BITS
;
1159 header
.id_time
= 31 << LTT_TSC_BITS
;
1167 header
.id_time
|= (u32
)tsc
& LTT_TSC_MASK
;
1168 ust_buffers_write(buf
, buf_offset
, &header
, sizeof(header
));
1169 buf_offset
+= sizeof(header
);
1172 case LTT_RFLAG_ID_SIZE_TSC
:
1173 small_size
= (u16
)min_t(u32
, event_size
, LTT_MAX_SMALL_SIZE
);
1174 ust_buffers_write(buf
, buf_offset
,
1176 buf_offset
+= sizeof(u16
);
1177 ust_buffers_write(buf
, buf_offset
,
1178 &small_size
, sizeof(u16
));
1179 buf_offset
+= sizeof(u16
);
1180 if (small_size
== LTT_MAX_SMALL_SIZE
) {
1181 ust_buffers_write(buf
, buf_offset
,
1182 &event_size
, sizeof(u32
));
1183 buf_offset
+= sizeof(u32
);
1185 buf_offset
+= ltt_align(buf_offset
, sizeof(u64
));
1186 ust_buffers_write(buf
, buf_offset
,
1188 buf_offset
+= sizeof(u64
);
1190 case LTT_RFLAG_ID_SIZE
:
1191 small_size
= (u16
)min_t(u32
, event_size
, LTT_MAX_SMALL_SIZE
);
1192 ust_buffers_write(buf
, buf_offset
,
1194 buf_offset
+= sizeof(u16
);
1195 ust_buffers_write(buf
, buf_offset
,
1196 &small_size
, sizeof(u16
));
1197 buf_offset
+= sizeof(u16
);
1198 if (small_size
== LTT_MAX_SMALL_SIZE
) {
1199 ust_buffers_write(buf
, buf_offset
,
1200 &event_size
, sizeof(u32
));
1201 buf_offset
+= sizeof(u32
);
1205 ust_buffers_write(buf
, buf_offset
,
1207 buf_offset
+= sizeof(u16
);