2 * SPDX-License-Identifier: LGPL-2.1-only
4 * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 * Ring buffer backend (internal helpers).
9 #ifndef _LTTNG_RING_BUFFER_BACKEND_INTERNAL_H
10 #define _LTTNG_RING_BUFFER_BACKEND_INTERNAL_H
15 #include <urcu/compiler.h>
17 #include <lttng/ringbuffer-config.h>
18 #include "backend_types.h"
19 #include "frontend_types.h"
21 #include "ust-helper.h"
23 /* Ring buffer backend API presented to the frontend */
25 /* Ring buffer and channel backend create/free */
28 int lib_ring_buffer_backend_create(struct lttng_ust_lib_ring_buffer_backend
*bufb
,
29 struct channel_backend
*chan
, int cpu
,
30 struct lttng_ust_shm_handle
*handle
,
31 struct shm_object
*shmobj
);
33 void channel_backend_unregister_notifiers(struct channel_backend
*chanb
);
35 void lib_ring_buffer_backend_free(struct lttng_ust_lib_ring_buffer_backend
*bufb
);
37 int channel_backend_init(struct channel_backend
*chanb
,
39 const struct lttng_ust_lib_ring_buffer_config
*config
,
41 size_t num_subbuf
, struct lttng_ust_shm_handle
*handle
,
42 const int *stream_fds
);
44 void channel_backend_free(struct channel_backend
*chanb
,
45 struct lttng_ust_shm_handle
*handle
);
48 void lib_ring_buffer_backend_reset(struct lttng_ust_lib_ring_buffer_backend
*bufb
,
49 struct lttng_ust_shm_handle
*handle
);
51 void channel_backend_reset(struct channel_backend
*chanb
);
54 int lib_ring_buffer_backend_init(void);
56 void lib_ring_buffer_backend_exit(void);
59 extern void _lib_ring_buffer_write(struct lttng_ust_lib_ring_buffer_backend
*bufb
,
60 size_t offset
, const void *src
, size_t len
,
64 * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
65 * exchanged atomically.
67 * Top half word, except lowest bit, belongs to "offset", which is used to keep
68 * to count the produced buffers. For overwrite mode, this provides the
69 * consumer with the capacity to read subbuffers in order, handling the
70 * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
71 * systems) concurrently with a single execution of get_subbuf (between offset
72 * sampling and subbuffer ID exchange).
75 #define HALF_ULONG_BITS (CAA_BITS_PER_LONG >> 1)
77 #define SB_ID_OFFSET_SHIFT (HALF_ULONG_BITS + 1)
78 #define SB_ID_OFFSET_COUNT (1UL << SB_ID_OFFSET_SHIFT)
79 #define SB_ID_OFFSET_MASK (~(SB_ID_OFFSET_COUNT - 1))
81 * Lowest bit of top word half belongs to noref. Used only for overwrite mode.
83 #define SB_ID_NOREF_SHIFT (SB_ID_OFFSET_SHIFT - 1)
84 #define SB_ID_NOREF_COUNT (1UL << SB_ID_NOREF_SHIFT)
85 #define SB_ID_NOREF_MASK SB_ID_NOREF_COUNT
87 * In overwrite mode: lowest half of word is used for index.
88 * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
89 * In producer-consumer mode: whole word used for index.
91 #define SB_ID_INDEX_SHIFT 0
92 #define SB_ID_INDEX_COUNT (1UL << SB_ID_INDEX_SHIFT)
93 #define SB_ID_INDEX_MASK (SB_ID_NOREF_COUNT - 1)
96 * Construct the subbuffer id from offset, index and noref. Use only the index
97 * for producer-consumer mode (offset and noref are only used in overwrite
101 unsigned long subbuffer_id(const struct lttng_ust_lib_ring_buffer_config
*config
,
102 unsigned long offset
, unsigned long noref
,
105 if (config
->mode
== RING_BUFFER_OVERWRITE
)
106 return (offset
<< SB_ID_OFFSET_SHIFT
)
107 | (noref
<< SB_ID_NOREF_SHIFT
)
114 * Compare offset with the offset contained within id. Return 1 if the offset
115 * bits are identical, else 0.
118 int subbuffer_id_compare_offset(const struct lttng_ust_lib_ring_buffer_config
*config
,
119 unsigned long id
, unsigned long offset
)
121 return (id
& SB_ID_OFFSET_MASK
) == (offset
<< SB_ID_OFFSET_SHIFT
);
125 unsigned long subbuffer_id_get_index(const struct lttng_ust_lib_ring_buffer_config
*config
,
128 if (config
->mode
== RING_BUFFER_OVERWRITE
)
129 return id
& SB_ID_INDEX_MASK
;
135 unsigned long subbuffer_id_is_noref(const struct lttng_ust_lib_ring_buffer_config
*config
,
138 if (config
->mode
== RING_BUFFER_OVERWRITE
)
139 return !!(id
& SB_ID_NOREF_MASK
);
145 * Only used by reader on subbuffer ID it has exclusive access to. No volatile
149 void subbuffer_id_set_noref(const struct lttng_ust_lib_ring_buffer_config
*config
,
152 if (config
->mode
== RING_BUFFER_OVERWRITE
)
153 *id
|= SB_ID_NOREF_MASK
;
157 void subbuffer_id_set_noref_offset(const struct lttng_ust_lib_ring_buffer_config
*config
,
158 unsigned long *id
, unsigned long offset
)
162 if (config
->mode
== RING_BUFFER_OVERWRITE
) {
164 tmp
&= ~SB_ID_OFFSET_MASK
;
165 tmp
|= offset
<< SB_ID_OFFSET_SHIFT
;
166 tmp
|= SB_ID_NOREF_MASK
;
167 /* Volatile store, read concurrently by readers. */
168 CMM_ACCESS_ONCE(*id
) = tmp
;
172 /* No volatile access, since already used locally */
174 void subbuffer_id_clear_noref(const struct lttng_ust_lib_ring_buffer_config
*config
,
177 if (config
->mode
== RING_BUFFER_OVERWRITE
)
178 *id
&= ~SB_ID_NOREF_MASK
;
182 * For overwrite mode, cap the number of subbuffers per buffer to:
183 * 2^16 on 32-bit architectures
184 * 2^32 on 64-bit architectures
185 * This is required to fit in the index part of the ID. Return 0 on success,
189 int subbuffer_id_check_index(const struct lttng_ust_lib_ring_buffer_config
*config
,
190 unsigned long num_subbuf
)
192 if (config
->mode
== RING_BUFFER_OVERWRITE
)
193 return (num_subbuf
> (1UL << HALF_ULONG_BITS
)) ? -EPERM
: 0;
199 int lib_ring_buffer_backend_get_pages(const struct lttng_ust_lib_ring_buffer_config
*config
,
200 struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
201 struct lttng_ust_lib_ring_buffer_backend_pages
**backend_pages
)
203 struct lttng_ust_lib_ring_buffer_backend
*bufb
= &ctx
->buf
->backend
;
204 struct channel_backend
*chanb
= &ctx
->chan
->backend
;
205 struct lttng_ust_shm_handle
*handle
= ctx
->handle
;
207 size_t offset
= ctx
->buf_offset
;
208 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
209 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*rpages
;
210 unsigned long sb_bindex
, id
;
211 struct lttng_ust_lib_ring_buffer_backend_pages
*_backend_pages
;
213 offset
&= chanb
->buf_size
- 1;
214 sbidx
= offset
>> chanb
->subbuf_size_order
;
215 wsb
= shmp_index(handle
, bufb
->buf_wsb
, sbidx
);
216 if (caa_unlikely(!wsb
))
219 sb_bindex
= subbuffer_id_get_index(config
, id
);
220 rpages
= shmp_index(handle
, bufb
->array
, sb_bindex
);
221 if (caa_unlikely(!rpages
))
223 CHAN_WARN_ON(ctx
->chan
,
224 config
->mode
== RING_BUFFER_OVERWRITE
225 && subbuffer_id_is_noref(config
, id
));
226 _backend_pages
= shmp(handle
, rpages
->shmp
);
227 if (caa_unlikely(!_backend_pages
))
229 *backend_pages
= _backend_pages
;
233 /* Get backend pages from cache. */
235 struct lttng_ust_lib_ring_buffer_backend_pages
*
236 lib_ring_buffer_get_backend_pages_from_ctx(const struct lttng_ust_lib_ring_buffer_config
*config
,
237 struct lttng_ust_lib_ring_buffer_ctx
*ctx
)
239 return ctx
->backend_pages
;
243 * The ring buffer can count events recorded and overwritten per buffer,
244 * but it is disabled by default due to its performance overhead.
246 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
248 void subbuffer_count_record(const struct lttng_ust_lib_ring_buffer_config
*config
,
249 const struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
250 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
251 unsigned long idx
, struct lttng_ust_shm_handle
*handle
)
253 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
255 backend_pages
= lib_ring_buffer_get_backend_pages_from_ctx(config
, ctx
);
256 if (caa_unlikely(!backend_pages
)) {
257 if (lib_ring_buffer_backend_get_pages(config
, ctx
, &backend_pages
))
260 v_inc(config
, &backend_pages
->records_commit
);
262 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
264 void subbuffer_count_record(const struct lttng_ust_lib_ring_buffer_config
*config
,
265 const struct lttng_ust_lib_ring_buffer_ctx
*ctx
,
266 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
267 unsigned long idx
, struct lttng_ust_shm_handle
*handle
)
270 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
273 * Reader has exclusive subbuffer access for record consumption. No need to
274 * perform the decrement atomically.
277 void subbuffer_consume_record(const struct lttng_ust_lib_ring_buffer_config
*config
,
278 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
279 struct lttng_ust_shm_handle
*handle
)
281 unsigned long sb_bindex
;
282 struct channel
*chan
;
283 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*pages_shmp
;
284 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
286 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
287 chan
= shmp(handle
, bufb
->chan
);
290 pages_shmp
= shmp_index(handle
, bufb
->array
, sb_bindex
);
293 backend_pages
= shmp(handle
, pages_shmp
->shmp
);
296 CHAN_WARN_ON(chan
, !v_read(config
, &backend_pages
->records_unread
));
297 /* Non-atomic decrement protected by exclusive subbuffer access */
298 _v_dec(config
, &backend_pages
->records_unread
);
299 v_inc(config
, &bufb
->records_read
);
303 unsigned long subbuffer_get_records_count(
304 const struct lttng_ust_lib_ring_buffer_config
*config
,
305 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
307 struct lttng_ust_shm_handle
*handle
)
309 unsigned long sb_bindex
;
310 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
311 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*rpages
;
312 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
314 wsb
= shmp_index(handle
, bufb
->buf_wsb
, idx
);
317 sb_bindex
= subbuffer_id_get_index(config
, wsb
->id
);
318 rpages
= shmp_index(handle
, bufb
->array
, sb_bindex
);
321 backend_pages
= shmp(handle
, rpages
->shmp
);
324 return v_read(config
, &backend_pages
->records_commit
);
328 * Must be executed at subbuffer delivery when the writer has _exclusive_
329 * subbuffer access. See lib_ring_buffer_check_deliver() for details.
330 * lib_ring_buffer_get_records_count() must be called to get the records
331 * count before this function, because it resets the records_commit
335 unsigned long subbuffer_count_records_overrun(
336 const struct lttng_ust_lib_ring_buffer_config
*config
,
337 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
339 struct lttng_ust_shm_handle
*handle
)
341 unsigned long overruns
, sb_bindex
;
342 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
343 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*rpages
;
344 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
346 wsb
= shmp_index(handle
, bufb
->buf_wsb
, idx
);
349 sb_bindex
= subbuffer_id_get_index(config
, wsb
->id
);
350 rpages
= shmp_index(handle
, bufb
->array
, sb_bindex
);
353 backend_pages
= shmp(handle
, rpages
->shmp
);
356 overruns
= v_read(config
, &backend_pages
->records_unread
);
357 v_set(config
, &backend_pages
->records_unread
,
358 v_read(config
, &backend_pages
->records_commit
));
359 v_set(config
, &backend_pages
->records_commit
, 0);
365 void subbuffer_set_data_size(const struct lttng_ust_lib_ring_buffer_config
*config
,
366 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
368 unsigned long data_size
,
369 struct lttng_ust_shm_handle
*handle
)
371 unsigned long sb_bindex
;
372 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
373 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*rpages
;
374 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
376 wsb
= shmp_index(handle
, bufb
->buf_wsb
, idx
);
379 sb_bindex
= subbuffer_id_get_index(config
, wsb
->id
);
380 rpages
= shmp_index(handle
, bufb
->array
, sb_bindex
);
383 backend_pages
= shmp(handle
, rpages
->shmp
);
386 backend_pages
->data_size
= data_size
;
390 unsigned long subbuffer_get_read_data_size(
391 const struct lttng_ust_lib_ring_buffer_config
*config
,
392 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
393 struct lttng_ust_shm_handle
*handle
)
395 unsigned long sb_bindex
;
396 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*pages_shmp
;
397 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
399 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
400 pages_shmp
= shmp_index(handle
, bufb
->array
, sb_bindex
);
403 backend_pages
= shmp(handle
, pages_shmp
->shmp
);
406 return backend_pages
->data_size
;
410 unsigned long subbuffer_get_data_size(
411 const struct lttng_ust_lib_ring_buffer_config
*config
,
412 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
414 struct lttng_ust_shm_handle
*handle
)
416 unsigned long sb_bindex
;
417 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
418 struct lttng_ust_lib_ring_buffer_backend_pages_shmp
*rpages
;
419 struct lttng_ust_lib_ring_buffer_backend_pages
*backend_pages
;
421 wsb
= shmp_index(handle
, bufb
->buf_wsb
, idx
);
424 sb_bindex
= subbuffer_id_get_index(config
, wsb
->id
);
425 rpages
= shmp_index(handle
, bufb
->array
, sb_bindex
);
428 backend_pages
= shmp(handle
, rpages
->shmp
);
431 return backend_pages
->data_size
;
435 void subbuffer_inc_packet_count(const struct lttng_ust_lib_ring_buffer_config
*config
,
436 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
437 unsigned long idx
, struct lttng_ust_shm_handle
*handle
)
439 struct lttng_ust_lib_ring_buffer_backend_counts
*counts
;
441 counts
= shmp_index(handle
, bufb
->buf_cnt
, idx
);
448 * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
452 void lib_ring_buffer_clear_noref(const struct lttng_ust_lib_ring_buffer_config
*config
,
453 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
455 struct lttng_ust_shm_handle
*handle
)
457 unsigned long id
, new_id
;
458 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
460 if (config
->mode
!= RING_BUFFER_OVERWRITE
)
464 * Performing a volatile access to read the sb_pages, because we want to
465 * read a coherent version of the pointer and the associated noref flag.
467 wsb
= shmp_index(handle
, bufb
->buf_wsb
, idx
);
470 id
= CMM_ACCESS_ONCE(wsb
->id
);
472 /* This check is called on the fast path for each record. */
473 if (caa_likely(!subbuffer_id_is_noref(config
, id
))) {
475 * Store after load dependency ordering the writes to
476 * the subbuffer after load and test of the noref flag
477 * matches the memory barrier implied by the cmpxchg()
478 * in update_read_sb_index().
480 return; /* Already writing to this buffer */
483 subbuffer_id_clear_noref(config
, &new_id
);
484 new_id
= uatomic_cmpxchg(&wsb
->id
, id
, new_id
);
485 if (caa_likely(new_id
== id
))
492 * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
496 void lib_ring_buffer_set_noref_offset(const struct lttng_ust_lib_ring_buffer_config
*config
,
497 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
498 unsigned long idx
, unsigned long offset
,
499 struct lttng_ust_shm_handle
*handle
)
501 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
502 struct channel
*chan
;
504 if (config
->mode
!= RING_BUFFER_OVERWRITE
)
507 wsb
= shmp_index(handle
, bufb
->buf_wsb
, idx
);
511 * Because ring_buffer_set_noref() is only called by a single thread
512 * (the one which updated the cc_sb value), there are no concurrent
513 * updates to take care of: other writers have not updated cc_sb, so
514 * they cannot set the noref flag, and concurrent readers cannot modify
515 * the pointer because the noref flag is not set yet.
516 * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
517 * to the subbuffer before this set noref operation.
518 * subbuffer_set_noref() uses a volatile store to deal with concurrent
519 * readers of the noref flag.
521 chan
= shmp(handle
, bufb
->chan
);
524 CHAN_WARN_ON(chan
, subbuffer_id_is_noref(config
, wsb
->id
));
526 * Memory barrier that ensures counter stores are ordered before set
530 subbuffer_id_set_noref_offset(config
, &wsb
->id
, offset
);
534 * update_read_sb_index - Read-side subbuffer index update.
537 int update_read_sb_index(const struct lttng_ust_lib_ring_buffer_config
*config
,
538 struct lttng_ust_lib_ring_buffer_backend
*bufb
,
539 struct channel_backend
*chanb
,
540 unsigned long consumed_idx
,
541 unsigned long consumed_count
,
542 struct lttng_ust_shm_handle
*handle
)
544 struct lttng_ust_lib_ring_buffer_backend_subbuffer
*wsb
;
545 unsigned long old_id
, new_id
;
547 wsb
= shmp_index(handle
, bufb
->buf_wsb
, consumed_idx
);
548 if (caa_unlikely(!wsb
))
551 if (config
->mode
== RING_BUFFER_OVERWRITE
) {
552 struct channel
*chan
;
555 * Exchange the target writer subbuffer with our own unused
556 * subbuffer. No need to use CMM_ACCESS_ONCE() here to read the
557 * old_wpage, because the value read will be confirmed by the
558 * following cmpxchg().
561 if (caa_unlikely(!subbuffer_id_is_noref(config
, old_id
)))
564 * Make sure the offset count we are expecting matches the one
565 * indicated by the writer.
567 if (caa_unlikely(!subbuffer_id_compare_offset(config
, old_id
,
570 chan
= shmp(handle
, bufb
->chan
);
571 if (caa_unlikely(!chan
))
573 CHAN_WARN_ON(chan
, !subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
574 subbuffer_id_set_noref_offset(config
, &bufb
->buf_rsb
.id
,
576 new_id
= uatomic_cmpxchg(&wsb
->id
, old_id
, bufb
->buf_rsb
.id
);
577 if (caa_unlikely(old_id
!= new_id
))
579 bufb
->buf_rsb
.id
= new_id
;
581 /* No page exchange, use the writer page directly */
582 bufb
->buf_rsb
.id
= wsb
->id
;
587 #ifndef inline_memcpy
588 #define inline_memcpy(dest, src, n) memcpy(dest, src, n)
591 static inline __attribute__((always_inline
))
592 void lttng_inline_memcpy(void *dest
, const void *src
,
597 *(uint8_t *) dest
= *(const uint8_t *) src
;
600 *(uint16_t *) dest
= *(const uint16_t *) src
;
603 *(uint32_t *) dest
= *(const uint32_t *) src
;
606 *(uint64_t *) dest
= *(const uint64_t *) src
;
609 inline_memcpy(dest
, src
, len
);
614 * Use the architecture-specific memcpy implementation for constant-sized
615 * inputs, but rely on an inline memcpy for length statically unknown.
616 * The function call to memcpy is just way too expensive for a fast path.
618 #define lib_ring_buffer_do_copy(config, dest, src, len) \
620 size_t __len = (len); \
621 if (__builtin_constant_p(len)) \
622 memcpy(dest, src, __len); \
624 lttng_inline_memcpy(dest, src, __len); \
628 * write len bytes to dest with c
631 void lib_ring_buffer_do_memset(char *dest
, int c
, unsigned long len
)
635 for (i
= 0; i
< len
; i
++)
639 /* arch-agnostic implementation */
641 static inline int lttng_ust_fls(unsigned int x
)
647 if (!(x
& 0xFFFF0000U
)) {
651 if (!(x
& 0xFF000000U
)) {
655 if (!(x
& 0xF0000000U
)) {
659 if (!(x
& 0xC0000000U
)) {
663 if (!(x
& 0x80000000U
)) {
664 /* No need to bit shift on last operation */
670 static inline int get_count_order(unsigned int count
)
674 order
= lttng_ust_fls(count
) - 1;
675 if (count
& (count
- 1))
680 #endif /* _LTTNG_RING_BUFFER_BACKEND_INTERNAL_H */