1 /* SPDX-License-Identifier: (GPL-2.0-only OR LGPL-2.1-only)
3 * ringbuffer/backend_internal.h
5 * Ring buffer backend (internal helpers).
7 * Copyright (C) 2008-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
10 #ifndef _LIB_RING_BUFFER_BACKEND_INTERNAL_H
11 #define _LIB_RING_BUFFER_BACKEND_INTERNAL_H
13 #include <wrapper/compiler.h>
14 #include <wrapper/inline_memcpy.h>
15 #include <ringbuffer/config.h>
16 #include <ringbuffer/backend_types.h>
17 #include <ringbuffer/frontend_types.h>
18 #include <linux/string.h>
19 #include <linux/uaccess.h>
21 /* Ring buffer backend API presented to the frontend */
23 /* Ring buffer and channel backend create/free */
25 int lib_ring_buffer_backend_create(struct lib_ring_buffer_backend
*bufb
,
26 struct channel_backend
*chan
, int cpu
);
27 void channel_backend_unregister_notifiers(struct channel_backend
*chanb
);
28 void lib_ring_buffer_backend_free(struct lib_ring_buffer_backend
*bufb
);
29 int channel_backend_init(struct channel_backend
*chanb
,
31 const struct lib_ring_buffer_config
*config
,
32 void *priv
, size_t subbuf_size
,
34 void channel_backend_free(struct channel_backend
*chanb
);
36 void lib_ring_buffer_backend_reset(struct lib_ring_buffer_backend
*bufb
);
37 void channel_backend_reset(struct channel_backend
*chanb
);
39 int lib_ring_buffer_backend_init(void);
40 void lib_ring_buffer_backend_exit(void);
42 extern void _lib_ring_buffer_write(struct lib_ring_buffer_backend
*bufb
,
43 size_t offset
, const void *src
, size_t len
,
45 extern void _lib_ring_buffer_memset(struct lib_ring_buffer_backend
*bufb
,
46 size_t offset
, int c
, size_t len
,
48 extern void _lib_ring_buffer_strcpy(struct lib_ring_buffer_backend
*bufb
,
49 size_t offset
, const char *src
, size_t len
,
50 size_t pagecpy
, int pad
);
51 extern void _lib_ring_buffer_copy_from_user_inatomic(struct lib_ring_buffer_backend
*bufb
,
52 size_t offset
, const void *src
,
53 size_t len
, size_t pagecpy
);
54 extern void _lib_ring_buffer_strcpy_from_user_inatomic(struct lib_ring_buffer_backend
*bufb
,
55 size_t offset
, const char __user
*src
, size_t len
,
56 size_t pagecpy
, int pad
);
59 * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
60 * exchanged atomically.
62 * Top half word, except lowest bit, belongs to "offset", which is used to keep
63 * to count the produced buffers. For overwrite mode, this provides the
64 * consumer with the capacity to read subbuffers in order, handling the
65 * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
66 * systems) concurrently with a single execution of get_subbuf (between offset
67 * sampling and subbuffer ID exchange).
70 #define HALF_ULONG_BITS (BITS_PER_LONG >> 1)
72 #define SB_ID_OFFSET_SHIFT (HALF_ULONG_BITS + 1)
73 #define SB_ID_OFFSET_COUNT (1UL << SB_ID_OFFSET_SHIFT)
74 #define SB_ID_OFFSET_MASK (~(SB_ID_OFFSET_COUNT - 1))
76 * Lowest bit of top word half belongs to noref. Used only for overwrite mode.
78 #define SB_ID_NOREF_SHIFT (SB_ID_OFFSET_SHIFT - 1)
79 #define SB_ID_NOREF_COUNT (1UL << SB_ID_NOREF_SHIFT)
80 #define SB_ID_NOREF_MASK SB_ID_NOREF_COUNT
82 * In overwrite mode: lowest half of word is used for index.
83 * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
84 * In producer-consumer mode: whole word used for index.
86 #define SB_ID_INDEX_SHIFT 0
87 #define SB_ID_INDEX_COUNT (1UL << SB_ID_INDEX_SHIFT)
88 #define SB_ID_INDEX_MASK (SB_ID_NOREF_COUNT - 1)
91 * Construct the subbuffer id from offset, index and noref. Use only the index
92 * for producer-consumer mode (offset and noref are only used in overwrite
96 unsigned long subbuffer_id(const struct lib_ring_buffer_config
*config
,
97 unsigned long offset
, unsigned long noref
,
100 if (config
->mode
== RING_BUFFER_OVERWRITE
)
101 return (offset
<< SB_ID_OFFSET_SHIFT
)
102 | (noref
<< SB_ID_NOREF_SHIFT
)
109 * Compare offset with the offset contained within id. Return 1 if the offset
110 * bits are identical, else 0.
113 int subbuffer_id_compare_offset(const struct lib_ring_buffer_config
*config
,
114 unsigned long id
, unsigned long offset
)
116 return (id
& SB_ID_OFFSET_MASK
) == (offset
<< SB_ID_OFFSET_SHIFT
);
120 unsigned long subbuffer_id_get_index(const struct lib_ring_buffer_config
*config
,
123 if (config
->mode
== RING_BUFFER_OVERWRITE
)
124 return id
& SB_ID_INDEX_MASK
;
130 unsigned long subbuffer_id_is_noref(const struct lib_ring_buffer_config
*config
,
133 if (config
->mode
== RING_BUFFER_OVERWRITE
)
134 return !!(id
& SB_ID_NOREF_MASK
);
140 * Only used by reader on subbuffer ID it has exclusive access to. No volatile
144 void subbuffer_id_set_noref(const struct lib_ring_buffer_config
*config
,
147 if (config
->mode
== RING_BUFFER_OVERWRITE
)
148 *id
|= SB_ID_NOREF_MASK
;
152 void subbuffer_id_set_noref_offset(const struct lib_ring_buffer_config
*config
,
153 unsigned long *id
, unsigned long offset
)
157 if (config
->mode
== RING_BUFFER_OVERWRITE
) {
159 tmp
&= ~SB_ID_OFFSET_MASK
;
160 tmp
|= offset
<< SB_ID_OFFSET_SHIFT
;
161 tmp
|= SB_ID_NOREF_MASK
;
162 /* Volatile store, read concurrently by readers. */
163 WRITE_ONCE(*id
, tmp
);
167 /* No volatile access, since already used locally */
169 void subbuffer_id_clear_noref(const struct lib_ring_buffer_config
*config
,
172 if (config
->mode
== RING_BUFFER_OVERWRITE
)
173 *id
&= ~SB_ID_NOREF_MASK
;
177 * For overwrite mode, cap the number of subbuffers per buffer to:
178 * 2^16 on 32-bit architectures
179 * 2^32 on 64-bit architectures
180 * This is required to fit in the index part of the ID. Return 0 on success,
184 int subbuffer_id_check_index(const struct lib_ring_buffer_config
*config
,
185 unsigned long num_subbuf
)
187 if (config
->mode
== RING_BUFFER_OVERWRITE
)
188 return (num_subbuf
> (1UL << HALF_ULONG_BITS
)) ? -EPERM
: 0;
194 void lib_ring_buffer_backend_get_pages(const struct lib_ring_buffer_config
*config
,
195 struct lib_ring_buffer_ctx
*ctx
,
196 struct lib_ring_buffer_backend_pages
**backend_pages
)
198 struct lib_ring_buffer_backend
*bufb
= &ctx
->buf
->backend
;
199 struct channel_backend
*chanb
= &ctx
->chan
->backend
;
200 size_t sbidx
, offset
= ctx
->buf_offset
;
201 unsigned long sb_bindex
, id
;
202 struct lib_ring_buffer_backend_pages
*rpages
;
204 offset
&= chanb
->buf_size
- 1;
205 sbidx
= offset
>> chanb
->subbuf_size_order
;
206 id
= bufb
->buf_wsb
[sbidx
].id
;
207 sb_bindex
= subbuffer_id_get_index(config
, id
);
208 rpages
= bufb
->array
[sb_bindex
];
209 CHAN_WARN_ON(ctx
->chan
,
210 config
->mode
== RING_BUFFER_OVERWRITE
211 && subbuffer_id_is_noref(config
, id
));
212 *backend_pages
= rpages
;
215 /* Get backend pages from cache. */
217 struct lib_ring_buffer_backend_pages
*
218 lib_ring_buffer_get_backend_pages_from_ctx(const struct lib_ring_buffer_config
*config
,
219 struct lib_ring_buffer_ctx
*ctx
)
221 return ctx
->backend_pages
;
225 * The ring buffer can count events recorded and overwritten per buffer,
226 * but it is disabled by default due to its performance overhead.
228 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
230 void subbuffer_count_record(const struct lib_ring_buffer_config
*config
,
231 struct lib_ring_buffer_backend
*bufb
,
234 unsigned long sb_bindex
;
236 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
237 v_inc(config
, &bufb
->array
[sb_bindex
]->records_commit
);
239 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
241 void subbuffer_count_record(const struct lib_ring_buffer_config
*config
,
242 struct lib_ring_buffer_backend
*bufb
,
246 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
249 * Reader has exclusive subbuffer access for record consumption. No need to
250 * perform the decrement atomically.
253 void subbuffer_consume_record(const struct lib_ring_buffer_config
*config
,
254 struct lib_ring_buffer_backend
*bufb
)
256 unsigned long sb_bindex
;
258 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
259 CHAN_WARN_ON(bufb
->chan
,
260 !v_read(config
, &bufb
->array
[sb_bindex
]->records_unread
));
261 /* Non-atomic decrement protected by exclusive subbuffer access */
262 _v_dec(config
, &bufb
->array
[sb_bindex
]->records_unread
);
263 v_inc(config
, &bufb
->records_read
);
267 unsigned long subbuffer_get_records_count(
268 const struct lib_ring_buffer_config
*config
,
269 struct lib_ring_buffer_backend
*bufb
,
272 unsigned long sb_bindex
;
274 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
275 return v_read(config
, &bufb
->array
[sb_bindex
]->records_commit
);
279 * Must be executed at subbuffer delivery when the writer has _exclusive_
280 * subbuffer access. See lib_ring_buffer_check_deliver() for details.
281 * lib_ring_buffer_get_records_count() must be called to get the records
282 * count before this function, because it resets the records_commit
286 unsigned long subbuffer_count_records_overrun(
287 const struct lib_ring_buffer_config
*config
,
288 struct lib_ring_buffer_backend
*bufb
,
291 struct lib_ring_buffer_backend_pages
*pages
;
292 unsigned long overruns
, sb_bindex
;
294 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
295 pages
= bufb
->array
[sb_bindex
];
296 overruns
= v_read(config
, &pages
->records_unread
);
297 v_set(config
, &pages
->records_unread
,
298 v_read(config
, &pages
->records_commit
));
299 v_set(config
, &pages
->records_commit
, 0);
305 void subbuffer_set_data_size(const struct lib_ring_buffer_config
*config
,
306 struct lib_ring_buffer_backend
*bufb
,
308 unsigned long data_size
)
310 struct lib_ring_buffer_backend_pages
*pages
;
311 unsigned long sb_bindex
;
313 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
314 pages
= bufb
->array
[sb_bindex
];
315 pages
->data_size
= data_size
;
319 unsigned long subbuffer_get_read_data_size(
320 const struct lib_ring_buffer_config
*config
,
321 struct lib_ring_buffer_backend
*bufb
)
323 struct lib_ring_buffer_backend_pages
*pages
;
324 unsigned long sb_bindex
;
326 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
327 pages
= bufb
->array
[sb_bindex
];
328 return pages
->data_size
;
332 unsigned long subbuffer_get_data_size(
333 const struct lib_ring_buffer_config
*config
,
334 struct lib_ring_buffer_backend
*bufb
,
337 struct lib_ring_buffer_backend_pages
*pages
;
338 unsigned long sb_bindex
;
340 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
341 pages
= bufb
->array
[sb_bindex
];
342 return pages
->data_size
;
346 void subbuffer_inc_packet_count(const struct lib_ring_buffer_config
*config
,
347 struct lib_ring_buffer_backend
*bufb
,
350 bufb
->buf_cnt
[idx
].seq_cnt
++;
354 * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
358 void lib_ring_buffer_clear_noref(const struct lib_ring_buffer_config
*config
,
359 struct lib_ring_buffer_backend
*bufb
,
362 unsigned long id
, new_id
;
364 if (config
->mode
!= RING_BUFFER_OVERWRITE
)
368 * Performing a volatile access to read the sb_pages, because we want to
369 * read a coherent version of the pointer and the associated noref flag.
371 id
= READ_ONCE(bufb
->buf_wsb
[idx
].id
);
373 /* This check is called on the fast path for each record. */
374 if (likely(!subbuffer_id_is_noref(config
, id
))) {
376 * Store after load dependency ordering the writes to
377 * the subbuffer after load and test of the noref flag
378 * matches the memory barrier implied by the cmpxchg()
379 * in update_read_sb_index().
381 return; /* Already writing to this buffer */
384 subbuffer_id_clear_noref(config
, &new_id
);
385 new_id
= cmpxchg(&bufb
->buf_wsb
[idx
].id
, id
, new_id
);
386 if (likely(new_id
== id
))
393 * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
397 void lib_ring_buffer_set_noref_offset(const struct lib_ring_buffer_config
*config
,
398 struct lib_ring_buffer_backend
*bufb
,
399 unsigned long idx
, unsigned long offset
)
401 if (config
->mode
!= RING_BUFFER_OVERWRITE
)
405 * Because ring_buffer_set_noref() is only called by a single thread
406 * (the one which updated the cc_sb value), there are no concurrent
407 * updates to take care of: other writers have not updated cc_sb, so
408 * they cannot set the noref flag, and concurrent readers cannot modify
409 * the pointer because the noref flag is not set yet.
410 * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
411 * to the subbuffer before this set noref operation.
412 * subbuffer_set_noref() uses a volatile store to deal with concurrent
413 * readers of the noref flag.
415 CHAN_WARN_ON(bufb
->chan
,
416 subbuffer_id_is_noref(config
, bufb
->buf_wsb
[idx
].id
));
418 * Memory barrier that ensures counter stores are ordered before set
422 subbuffer_id_set_noref_offset(config
, &bufb
->buf_wsb
[idx
].id
, offset
);
426 * update_read_sb_index - Read-side subbuffer index update.
429 int update_read_sb_index(const struct lib_ring_buffer_config
*config
,
430 struct lib_ring_buffer_backend
*bufb
,
431 struct channel_backend
*chanb
,
432 unsigned long consumed_idx
,
433 unsigned long consumed_count
)
435 unsigned long old_id
, new_id
;
437 if (config
->mode
== RING_BUFFER_OVERWRITE
) {
439 * Exchange the target writer subbuffer with our own unused
440 * subbuffer. No need to use READ_ONCE() here to read the
441 * old_wpage, because the value read will be confirmed by the
442 * following cmpxchg().
444 old_id
= bufb
->buf_wsb
[consumed_idx
].id
;
445 if (unlikely(!subbuffer_id_is_noref(config
, old_id
)))
448 * Make sure the offset count we are expecting matches the one
449 * indicated by the writer.
451 if (unlikely(!subbuffer_id_compare_offset(config
, old_id
,
454 CHAN_WARN_ON(bufb
->chan
,
455 !subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
456 subbuffer_id_set_noref_offset(config
, &bufb
->buf_rsb
.id
,
458 new_id
= cmpxchg(&bufb
->buf_wsb
[consumed_idx
].id
, old_id
,
460 if (unlikely(old_id
!= new_id
))
462 bufb
->buf_rsb
.id
= new_id
;
464 /* No page exchange, use the writer page directly */
465 bufb
->buf_rsb
.id
= bufb
->buf_wsb
[consumed_idx
].id
;
470 static inline __attribute__((always_inline
))
471 void lttng_inline_memcpy(void *dest
, const void *src
,
476 *(uint8_t *) dest
= *(const uint8_t *) src
;
479 *(uint16_t *) dest
= *(const uint16_t *) src
;
482 *(uint32_t *) dest
= *(const uint32_t *) src
;
485 *(uint64_t *) dest
= *(const uint64_t *) src
;
488 inline_memcpy(dest
, src
, len
);
493 * Use the architecture-specific memcpy implementation for constant-sized
494 * inputs, but rely on an inline memcpy for length statically unknown.
495 * The function call to memcpy is just way too expensive for a fast path.
497 #define lib_ring_buffer_do_copy(config, dest, src, len) \
499 size_t __len = (len); \
500 if (__builtin_constant_p(len)) \
501 memcpy(dest, src, __len); \
503 lttng_inline_memcpy(dest, src, __len); \
507 * We use __copy_from_user_inatomic to copy userspace data since we already
508 * did the access_ok for the whole range.
510 * Return 0 if OK, nonzero on error.
513 unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest
,
514 const void __user
*src
,
517 return __copy_from_user_inatomic(dest
, src
, len
);
521 * write len bytes to dest with c
524 void lib_ring_buffer_do_memset(char *dest
, int c
,
529 for (i
= 0; i
< len
; i
++)
533 #endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */