1 #ifndef _LIB_RING_BUFFER_BACKEND_INTERNAL_H
2 #define _LIB_RING_BUFFER_BACKEND_INTERNAL_H
5 * lib/ringbuffer/backend_internal.h
7 * Ring buffer backend (internal helpers).
9 * Copyright (C) 2008-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
11 * This library is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public
13 * License as published by the Free Software Foundation; only
14 * version 2.1 of the License.
16 * This library is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with this library; if not, write to the Free Software
23 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 #include <wrapper/ringbuffer/config.h>
27 #include <wrapper/ringbuffer/backend_types.h>
28 #include <wrapper/ringbuffer/frontend_types.h>
29 #include <linux/string.h>
30 #include <linux/uaccess.h>
32 /* Ring buffer backend API presented to the frontend */
34 /* Ring buffer and channel backend create/free */
36 int lib_ring_buffer_backend_create(struct lib_ring_buffer_backend
*bufb
,
37 struct channel_backend
*chan
, int cpu
);
38 void channel_backend_unregister_notifiers(struct channel_backend
*chanb
);
39 void lib_ring_buffer_backend_free(struct lib_ring_buffer_backend
*bufb
);
40 int channel_backend_init(struct channel_backend
*chanb
,
42 const struct lib_ring_buffer_config
*config
,
43 void *priv
, size_t subbuf_size
,
45 void channel_backend_free(struct channel_backend
*chanb
);
47 void lib_ring_buffer_backend_reset(struct lib_ring_buffer_backend
*bufb
);
48 void channel_backend_reset(struct channel_backend
*chanb
);
50 int lib_ring_buffer_backend_init(void);
51 void lib_ring_buffer_backend_exit(void);
53 extern void _lib_ring_buffer_write(struct lib_ring_buffer_backend
*bufb
,
54 size_t offset
, const void *src
, size_t len
,
56 extern void _lib_ring_buffer_memset(struct lib_ring_buffer_backend
*bufb
,
57 size_t offset
, int c
, size_t len
,
59 extern void _lib_ring_buffer_strcpy(struct lib_ring_buffer_backend
*bufb
,
60 size_t offset
, const char *src
, size_t len
,
61 size_t pagecpy
, int pad
);
62 extern void _lib_ring_buffer_copy_from_user_inatomic(struct lib_ring_buffer_backend
*bufb
,
63 size_t offset
, const void *src
,
64 size_t len
, size_t pagecpy
);
65 extern void _lib_ring_buffer_strcpy_from_user_inatomic(struct lib_ring_buffer_backend
*bufb
,
66 size_t offset
, const char __user
*src
, size_t len
,
67 size_t pagecpy
, int pad
);
70 * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
71 * exchanged atomically.
73 * Top half word, except lowest bit, belongs to "offset", which is used to keep
74 * to count the produced buffers. For overwrite mode, this provides the
75 * consumer with the capacity to read subbuffers in order, handling the
76 * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
77 * systems) concurrently with a single execution of get_subbuf (between offset
78 * sampling and subbuffer ID exchange).
81 #define HALF_ULONG_BITS (BITS_PER_LONG >> 1)
83 #define SB_ID_OFFSET_SHIFT (HALF_ULONG_BITS + 1)
84 #define SB_ID_OFFSET_COUNT (1UL << SB_ID_OFFSET_SHIFT)
85 #define SB_ID_OFFSET_MASK (~(SB_ID_OFFSET_COUNT - 1))
87 * Lowest bit of top word half belongs to noref. Used only for overwrite mode.
89 #define SB_ID_NOREF_SHIFT (SB_ID_OFFSET_SHIFT - 1)
90 #define SB_ID_NOREF_COUNT (1UL << SB_ID_NOREF_SHIFT)
91 #define SB_ID_NOREF_MASK SB_ID_NOREF_COUNT
93 * In overwrite mode: lowest half of word is used for index.
94 * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
95 * In producer-consumer mode: whole word used for index.
97 #define SB_ID_INDEX_SHIFT 0
98 #define SB_ID_INDEX_COUNT (1UL << SB_ID_INDEX_SHIFT)
99 #define SB_ID_INDEX_MASK (SB_ID_NOREF_COUNT - 1)
102 * Construct the subbuffer id from offset, index and noref. Use only the index
103 * for producer-consumer mode (offset and noref are only used in overwrite
107 unsigned long subbuffer_id(const struct lib_ring_buffer_config
*config
,
108 unsigned long offset
, unsigned long noref
,
111 if (config
->mode
== RING_BUFFER_OVERWRITE
)
112 return (offset
<< SB_ID_OFFSET_SHIFT
)
113 | (noref
<< SB_ID_NOREF_SHIFT
)
120 * Compare offset with the offset contained within id. Return 1 if the offset
121 * bits are identical, else 0.
124 int subbuffer_id_compare_offset(const struct lib_ring_buffer_config
*config
,
125 unsigned long id
, unsigned long offset
)
127 return (id
& SB_ID_OFFSET_MASK
) == (offset
<< SB_ID_OFFSET_SHIFT
);
131 unsigned long subbuffer_id_get_index(const struct lib_ring_buffer_config
*config
,
134 if (config
->mode
== RING_BUFFER_OVERWRITE
)
135 return id
& SB_ID_INDEX_MASK
;
141 unsigned long subbuffer_id_is_noref(const struct lib_ring_buffer_config
*config
,
144 if (config
->mode
== RING_BUFFER_OVERWRITE
)
145 return !!(id
& SB_ID_NOREF_MASK
);
151 * Only used by reader on subbuffer ID it has exclusive access to. No volatile
155 void subbuffer_id_set_noref(const struct lib_ring_buffer_config
*config
,
158 if (config
->mode
== RING_BUFFER_OVERWRITE
)
159 *id
|= SB_ID_NOREF_MASK
;
163 void subbuffer_id_set_noref_offset(const struct lib_ring_buffer_config
*config
,
164 unsigned long *id
, unsigned long offset
)
168 if (config
->mode
== RING_BUFFER_OVERWRITE
) {
170 tmp
&= ~SB_ID_OFFSET_MASK
;
171 tmp
|= offset
<< SB_ID_OFFSET_SHIFT
;
172 tmp
|= SB_ID_NOREF_MASK
;
173 /* Volatile store, read concurrently by readers. */
174 ACCESS_ONCE(*id
) = tmp
;
178 /* No volatile access, since already used locally */
180 void subbuffer_id_clear_noref(const struct lib_ring_buffer_config
*config
,
183 if (config
->mode
== RING_BUFFER_OVERWRITE
)
184 *id
&= ~SB_ID_NOREF_MASK
;
188 * For overwrite mode, cap the number of subbuffers per buffer to:
189 * 2^16 on 32-bit architectures
190 * 2^32 on 64-bit architectures
191 * This is required to fit in the index part of the ID. Return 0 on success,
195 int subbuffer_id_check_index(const struct lib_ring_buffer_config
*config
,
196 unsigned long num_subbuf
)
198 if (config
->mode
== RING_BUFFER_OVERWRITE
)
199 return (num_subbuf
> (1UL << HALF_ULONG_BITS
)) ? -EPERM
: 0;
205 void lib_ring_buffer_backend_get_pages(const struct lib_ring_buffer_config
*config
,
206 struct lib_ring_buffer_ctx
*ctx
,
207 struct lib_ring_buffer_backend_pages
**backend_pages
)
209 struct lib_ring_buffer_backend
*bufb
= &ctx
->buf
->backend
;
210 struct channel_backend
*chanb
= &ctx
->chan
->backend
;
211 size_t sbidx
, offset
= ctx
->buf_offset
;
212 unsigned long sb_bindex
, id
;
213 struct lib_ring_buffer_backend_pages
*rpages
;
215 offset
&= chanb
->buf_size
- 1;
216 sbidx
= offset
>> chanb
->subbuf_size_order
;
217 id
= bufb
->buf_wsb
[sbidx
].id
;
218 sb_bindex
= subbuffer_id_get_index(config
, id
);
219 rpages
= bufb
->array
[sb_bindex
];
220 CHAN_WARN_ON(ctx
->chan
,
221 config
->mode
== RING_BUFFER_OVERWRITE
222 && subbuffer_id_is_noref(config
, id
));
223 *backend_pages
= rpages
;
226 /* Get backend pages from cache. */
228 struct lib_ring_buffer_backend_pages
*
229 lib_ring_buffer_get_backend_pages_from_ctx(const struct lib_ring_buffer_config
*config
,
230 struct lib_ring_buffer_ctx
*ctx
)
232 return ctx
->backend_pages
;
236 * The ring buffer can count events recorded and overwritten per buffer,
237 * but it is disabled by default due to its performance overhead.
239 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
241 void subbuffer_count_record(const struct lib_ring_buffer_config
*config
,
242 struct lib_ring_buffer_backend
*bufb
,
245 unsigned long sb_bindex
;
247 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
248 v_inc(config
, &bufb
->array
[sb_bindex
]->records_commit
);
250 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
252 void subbuffer_count_record(const struct lib_ring_buffer_config
*config
,
253 struct lib_ring_buffer_backend
*bufb
,
257 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
260 * Reader has exclusive subbuffer access for record consumption. No need to
261 * perform the decrement atomically.
264 void subbuffer_consume_record(const struct lib_ring_buffer_config
*config
,
265 struct lib_ring_buffer_backend
*bufb
)
267 unsigned long sb_bindex
;
269 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
270 CHAN_WARN_ON(bufb
->chan
,
271 !v_read(config
, &bufb
->array
[sb_bindex
]->records_unread
));
272 /* Non-atomic decrement protected by exclusive subbuffer access */
273 _v_dec(config
, &bufb
->array
[sb_bindex
]->records_unread
);
274 v_inc(config
, &bufb
->records_read
);
278 unsigned long subbuffer_get_records_count(
279 const struct lib_ring_buffer_config
*config
,
280 struct lib_ring_buffer_backend
*bufb
,
283 unsigned long sb_bindex
;
285 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
286 return v_read(config
, &bufb
->array
[sb_bindex
]->records_commit
);
290 * Must be executed at subbuffer delivery when the writer has _exclusive_
291 * subbuffer access. See lib_ring_buffer_check_deliver() for details.
292 * lib_ring_buffer_get_records_count() must be called to get the records
293 * count before this function, because it resets the records_commit
297 unsigned long subbuffer_count_records_overrun(
298 const struct lib_ring_buffer_config
*config
,
299 struct lib_ring_buffer_backend
*bufb
,
302 struct lib_ring_buffer_backend_pages
*pages
;
303 unsigned long overruns
, sb_bindex
;
305 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
306 pages
= bufb
->array
[sb_bindex
];
307 overruns
= v_read(config
, &pages
->records_unread
);
308 v_set(config
, &pages
->records_unread
,
309 v_read(config
, &pages
->records_commit
));
310 v_set(config
, &pages
->records_commit
, 0);
316 void subbuffer_set_data_size(const struct lib_ring_buffer_config
*config
,
317 struct lib_ring_buffer_backend
*bufb
,
319 unsigned long data_size
)
321 struct lib_ring_buffer_backend_pages
*pages
;
322 unsigned long sb_bindex
;
324 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
325 pages
= bufb
->array
[sb_bindex
];
326 pages
->data_size
= data_size
;
330 unsigned long subbuffer_get_read_data_size(
331 const struct lib_ring_buffer_config
*config
,
332 struct lib_ring_buffer_backend
*bufb
)
334 struct lib_ring_buffer_backend_pages
*pages
;
335 unsigned long sb_bindex
;
337 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
338 pages
= bufb
->array
[sb_bindex
];
339 return pages
->data_size
;
343 unsigned long subbuffer_get_data_size(
344 const struct lib_ring_buffer_config
*config
,
345 struct lib_ring_buffer_backend
*bufb
,
348 struct lib_ring_buffer_backend_pages
*pages
;
349 unsigned long sb_bindex
;
351 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
352 pages
= bufb
->array
[sb_bindex
];
353 return pages
->data_size
;
357 void subbuffer_inc_packet_count(const struct lib_ring_buffer_config
*config
,
358 struct lib_ring_buffer_backend
*bufb
,
361 bufb
->buf_cnt
[idx
].seq_cnt
++;
365 * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
369 void lib_ring_buffer_clear_noref(const struct lib_ring_buffer_config
*config
,
370 struct lib_ring_buffer_backend
*bufb
,
373 unsigned long id
, new_id
;
375 if (config
->mode
!= RING_BUFFER_OVERWRITE
)
379 * Performing a volatile access to read the sb_pages, because we want to
380 * read a coherent version of the pointer and the associated noref flag.
382 id
= ACCESS_ONCE(bufb
->buf_wsb
[idx
].id
);
384 /* This check is called on the fast path for each record. */
385 if (likely(!subbuffer_id_is_noref(config
, id
))) {
387 * Store after load dependency ordering the writes to
388 * the subbuffer after load and test of the noref flag
389 * matches the memory barrier implied by the cmpxchg()
390 * in update_read_sb_index().
392 return; /* Already writing to this buffer */
395 subbuffer_id_clear_noref(config
, &new_id
);
396 new_id
= cmpxchg(&bufb
->buf_wsb
[idx
].id
, id
, new_id
);
397 if (likely(new_id
== id
))
404 * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
408 void lib_ring_buffer_set_noref_offset(const struct lib_ring_buffer_config
*config
,
409 struct lib_ring_buffer_backend
*bufb
,
410 unsigned long idx
, unsigned long offset
)
412 if (config
->mode
!= RING_BUFFER_OVERWRITE
)
416 * Because ring_buffer_set_noref() is only called by a single thread
417 * (the one which updated the cc_sb value), there are no concurrent
418 * updates to take care of: other writers have not updated cc_sb, so
419 * they cannot set the noref flag, and concurrent readers cannot modify
420 * the pointer because the noref flag is not set yet.
421 * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
422 * to the subbuffer before this set noref operation.
423 * subbuffer_set_noref() uses a volatile store to deal with concurrent
424 * readers of the noref flag.
426 CHAN_WARN_ON(bufb
->chan
,
427 subbuffer_id_is_noref(config
, bufb
->buf_wsb
[idx
].id
));
429 * Memory barrier that ensures counter stores are ordered before set
433 subbuffer_id_set_noref_offset(config
, &bufb
->buf_wsb
[idx
].id
, offset
);
437 * update_read_sb_index - Read-side subbuffer index update.
440 int update_read_sb_index(const struct lib_ring_buffer_config
*config
,
441 struct lib_ring_buffer_backend
*bufb
,
442 struct channel_backend
*chanb
,
443 unsigned long consumed_idx
,
444 unsigned long consumed_count
)
446 unsigned long old_id
, new_id
;
448 if (config
->mode
== RING_BUFFER_OVERWRITE
) {
450 * Exchange the target writer subbuffer with our own unused
451 * subbuffer. No need to use ACCESS_ONCE() here to read the
452 * old_wpage, because the value read will be confirmed by the
453 * following cmpxchg().
455 old_id
= bufb
->buf_wsb
[consumed_idx
].id
;
456 if (unlikely(!subbuffer_id_is_noref(config
, old_id
)))
459 * Make sure the offset count we are expecting matches the one
460 * indicated by the writer.
462 if (unlikely(!subbuffer_id_compare_offset(config
, old_id
,
465 CHAN_WARN_ON(bufb
->chan
,
466 !subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
467 subbuffer_id_set_noref_offset(config
, &bufb
->buf_rsb
.id
,
469 new_id
= cmpxchg(&bufb
->buf_wsb
[consumed_idx
].id
, old_id
,
471 if (unlikely(old_id
!= new_id
))
473 bufb
->buf_rsb
.id
= new_id
;
475 /* No page exchange, use the writer page directly */
476 bufb
->buf_rsb
.id
= bufb
->buf_wsb
[consumed_idx
].id
;
481 static inline __attribute__((always_inline
))
482 void lttng_inline_memcpy(void *dest
, const void *src
,
487 *(uint8_t *) dest
= *(const uint8_t *) src
;
490 *(uint16_t *) dest
= *(const uint16_t *) src
;
493 *(uint32_t *) dest
= *(const uint32_t *) src
;
496 *(uint64_t *) dest
= *(const uint64_t *) src
;
499 inline_memcpy(dest
, src
, len
);
504 * Use the architecture-specific memcpy implementation for constant-sized
505 * inputs, but rely on an inline memcpy for length statically unknown.
506 * The function call to memcpy is just way too expensive for a fast path.
508 #define lib_ring_buffer_do_copy(config, dest, src, len) \
510 size_t __len = (len); \
511 if (__builtin_constant_p(len)) \
512 memcpy(dest, src, __len); \
514 lttng_inline_memcpy(dest, src, __len); \
518 * We use __copy_from_user_inatomic to copy userspace data since we already
519 * did the access_ok for the whole range.
521 * Return 0 if OK, nonzero on error.
524 unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest
,
525 const void __user
*src
,
528 return __copy_from_user_inatomic(dest
, src
, len
);
532 * write len bytes to dest with c
535 void lib_ring_buffer_do_memset(char *dest
, int c
,
540 for (i
= 0; i
< len
; i
++)
544 #endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */