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 lttng_kernel_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 lttng_kernel_ring_buffer_backend
*bufb
);
29 int channel_backend_init(struct channel_backend
*chanb
,
31 const struct lttng_kernel_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 lttng_kernel_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 lttng_kernel_ring_buffer_backend
*bufb
,
43 size_t offset
, const void *src
, size_t len
,
45 extern void _lib_ring_buffer_memset(struct lttng_kernel_ring_buffer_backend
*bufb
,
46 size_t offset
, int c
, size_t len
,
48 extern void _lib_ring_buffer_strcpy(struct lttng_kernel_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_pstrcpy(struct lttng_kernel_ring_buffer_backend
*bufb
,
52 size_t offset
, const char *src
, size_t len
, int pad
);
53 extern void _lib_ring_buffer_copy_from_user_inatomic(struct lttng_kernel_ring_buffer_backend
*bufb
,
54 size_t offset
, const void *src
,
55 size_t len
, size_t pagecpy
);
56 extern void _lib_ring_buffer_strcpy_from_user_inatomic(struct lttng_kernel_ring_buffer_backend
*bufb
,
57 size_t offset
, const char __user
*src
, size_t len
,
58 size_t pagecpy
, int pad
);
59 extern void _lib_ring_buffer_pstrcpy_from_user_inatomic(struct lttng_kernel_ring_buffer_backend
*bufb
,
60 size_t offset
, const char __user
*src
, size_t len
, int pad
);
63 * Subbuffer ID bits for overwrite mode. Need to fit within a single word to be
64 * exchanged atomically.
66 * Top half word, except lowest bit, belongs to "offset", which is used to keep
67 * to count the produced buffers. For overwrite mode, this provides the
68 * consumer with the capacity to read subbuffers in order, handling the
69 * situation where producers would write up to 2^15 buffers (or 2^31 for 64-bit
70 * systems) concurrently with a single execution of get_subbuf (between offset
71 * sampling and subbuffer ID exchange).
74 #define HALF_ULONG_BITS (BITS_PER_LONG >> 1)
76 #define SB_ID_OFFSET_SHIFT (HALF_ULONG_BITS + 1)
77 #define SB_ID_OFFSET_COUNT (1UL << SB_ID_OFFSET_SHIFT)
78 #define SB_ID_OFFSET_MASK (~(SB_ID_OFFSET_COUNT - 1))
80 * Lowest bit of top word half belongs to noref. Used only for overwrite mode.
82 #define SB_ID_NOREF_SHIFT (SB_ID_OFFSET_SHIFT - 1)
83 #define SB_ID_NOREF_COUNT (1UL << SB_ID_NOREF_SHIFT)
84 #define SB_ID_NOREF_MASK SB_ID_NOREF_COUNT
86 * In overwrite mode: lowest half of word is used for index.
87 * Limit of 2^16 subbuffers per buffer on 32-bit, 2^32 on 64-bit.
88 * In producer-consumer mode: whole word used for index.
90 #define SB_ID_INDEX_SHIFT 0
91 #define SB_ID_INDEX_COUNT (1UL << SB_ID_INDEX_SHIFT)
92 #define SB_ID_INDEX_MASK (SB_ID_NOREF_COUNT - 1)
95 * Construct the subbuffer id from offset, index and noref. Use only the index
96 * for producer-consumer mode (offset and noref are only used in overwrite
100 unsigned long subbuffer_id(const struct lttng_kernel_ring_buffer_config
*config
,
101 unsigned long offset
, unsigned long noref
,
104 if (config
->mode
== RING_BUFFER_OVERWRITE
)
105 return (offset
<< SB_ID_OFFSET_SHIFT
)
106 | (noref
<< SB_ID_NOREF_SHIFT
)
113 * Compare offset with the offset contained within id. Return 1 if the offset
114 * bits are identical, else 0.
117 int subbuffer_id_compare_offset(const struct lttng_kernel_ring_buffer_config
*config
,
118 unsigned long id
, unsigned long offset
)
120 return (id
& SB_ID_OFFSET_MASK
) == (offset
<< SB_ID_OFFSET_SHIFT
);
124 unsigned long subbuffer_id_get_index(const struct lttng_kernel_ring_buffer_config
*config
,
127 if (config
->mode
== RING_BUFFER_OVERWRITE
)
128 return id
& SB_ID_INDEX_MASK
;
134 unsigned long subbuffer_id_is_noref(const struct lttng_kernel_ring_buffer_config
*config
,
137 if (config
->mode
== RING_BUFFER_OVERWRITE
)
138 return !!(id
& SB_ID_NOREF_MASK
);
144 * Only used by reader on subbuffer ID it has exclusive access to. No volatile
148 void subbuffer_id_set_noref(const struct lttng_kernel_ring_buffer_config
*config
,
151 if (config
->mode
== RING_BUFFER_OVERWRITE
)
152 *id
|= SB_ID_NOREF_MASK
;
156 void subbuffer_id_set_noref_offset(const struct lttng_kernel_ring_buffer_config
*config
,
157 unsigned long *id
, unsigned long offset
)
161 if (config
->mode
== RING_BUFFER_OVERWRITE
) {
163 tmp
&= ~SB_ID_OFFSET_MASK
;
164 tmp
|= offset
<< SB_ID_OFFSET_SHIFT
;
165 tmp
|= SB_ID_NOREF_MASK
;
166 /* Volatile store, read concurrently by readers. */
167 WRITE_ONCE(*id
, tmp
);
171 /* No volatile access, since already used locally */
173 void subbuffer_id_clear_noref(const struct lttng_kernel_ring_buffer_config
*config
,
176 if (config
->mode
== RING_BUFFER_OVERWRITE
)
177 *id
&= ~SB_ID_NOREF_MASK
;
181 * For overwrite mode, cap the number of subbuffers per buffer to:
182 * 2^16 on 32-bit architectures
183 * 2^32 on 64-bit architectures
184 * This is required to fit in the index part of the ID. Return 0 on success,
188 int subbuffer_id_check_index(const struct lttng_kernel_ring_buffer_config
*config
,
189 unsigned long num_subbuf
)
191 if (config
->mode
== RING_BUFFER_OVERWRITE
)
192 return (num_subbuf
> (1UL << HALF_ULONG_BITS
)) ? -EPERM
: 0;
198 void lib_ring_buffer_backend_get_pages(const struct lttng_kernel_ring_buffer_config
*config
,
199 struct lttng_kernel_ring_buffer_ctx
*ctx
,
200 struct lttng_kernel_ring_buffer_backend_pages
**backend_pages
)
202 struct lttng_kernel_ring_buffer_backend
*bufb
= &ctx
->priv
.buf
->backend
;
203 struct channel_backend
*chanb
= &ctx
->priv
.chan
->backend
;
204 size_t sbidx
, offset
= ctx
->priv
.buf_offset
;
205 unsigned long sb_bindex
, id
;
206 struct lttng_kernel_ring_buffer_backend_pages
*rpages
;
208 offset
&= chanb
->buf_size
- 1;
209 sbidx
= offset
>> chanb
->subbuf_size_order
;
210 id
= bufb
->buf_wsb
[sbidx
].id
;
211 sb_bindex
= subbuffer_id_get_index(config
, id
);
212 rpages
= bufb
->array
[sb_bindex
];
213 CHAN_WARN_ON(ctx
->priv
.chan
,
214 config
->mode
== RING_BUFFER_OVERWRITE
215 && subbuffer_id_is_noref(config
, id
));
216 *backend_pages
= rpages
;
219 /* Get backend pages from cache. */
221 struct lttng_kernel_ring_buffer_backend_pages
*
222 lib_ring_buffer_get_backend_pages_from_ctx(const struct lttng_kernel_ring_buffer_config
*config
,
223 struct lttng_kernel_ring_buffer_ctx
*ctx
)
225 return ctx
->priv
.backend_pages
;
229 * The ring buffer can count events recorded and overwritten per buffer,
230 * but it is disabled by default due to its performance overhead.
232 #ifdef LTTNG_RING_BUFFER_COUNT_EVENTS
234 void subbuffer_count_record(const struct lttng_kernel_ring_buffer_config
*config
,
235 struct lttng_kernel_ring_buffer_backend
*bufb
,
238 unsigned long sb_bindex
;
240 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
241 v_inc(config
, &bufb
->array
[sb_bindex
]->records_commit
);
245 * Reader has exclusive subbuffer access for record consumption. No need to
246 * perform the decrement atomically.
249 void subbuffer_consume_record(const struct lttng_kernel_ring_buffer_config
*config
,
250 struct lttng_kernel_ring_buffer_backend
*bufb
)
252 unsigned long sb_bindex
;
254 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
255 CHAN_WARN_ON(bufb
->chan
,
256 !v_read(config
, &bufb
->array
[sb_bindex
]->records_unread
));
257 /* Non-atomic decrement protected by exclusive subbuffer access */
258 _v_dec(config
, &bufb
->array
[sb_bindex
]->records_unread
);
259 v_inc(config
, &bufb
->records_read
);
261 #else /* LTTNG_RING_BUFFER_COUNT_EVENTS */
263 void subbuffer_count_record(const struct lttng_kernel_ring_buffer_config
*config
,
264 struct lttng_kernel_ring_buffer_backend
*bufb
,
269 void subbuffer_consume_record(const struct lttng_kernel_ring_buffer_config
*config
,
270 struct lttng_kernel_ring_buffer_backend
*bufb
)
273 #endif /* #else LTTNG_RING_BUFFER_COUNT_EVENTS */
276 unsigned long subbuffer_get_records_count(
277 const struct lttng_kernel_ring_buffer_config
*config
,
278 struct lttng_kernel_ring_buffer_backend
*bufb
,
281 unsigned long sb_bindex
;
283 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
284 return v_read(config
, &bufb
->array
[sb_bindex
]->records_commit
);
288 * Must be executed at subbuffer delivery when the writer has _exclusive_
289 * subbuffer access. See lib_ring_buffer_check_deliver() for details.
290 * lib_ring_buffer_get_records_count() must be called to get the records
291 * count before this function, because it resets the records_commit
295 unsigned long subbuffer_count_records_overrun(
296 const struct lttng_kernel_ring_buffer_config
*config
,
297 struct lttng_kernel_ring_buffer_backend
*bufb
,
300 struct lttng_kernel_ring_buffer_backend_pages
*pages
;
301 unsigned long overruns
, sb_bindex
;
303 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
304 pages
= bufb
->array
[sb_bindex
];
305 overruns
= v_read(config
, &pages
->records_unread
);
306 v_set(config
, &pages
->records_unread
,
307 v_read(config
, &pages
->records_commit
));
308 v_set(config
, &pages
->records_commit
, 0);
314 void subbuffer_set_data_size(const struct lttng_kernel_ring_buffer_config
*config
,
315 struct lttng_kernel_ring_buffer_backend
*bufb
,
317 unsigned long data_size
)
319 struct lttng_kernel_ring_buffer_backend_pages
*pages
;
320 unsigned long sb_bindex
;
322 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
323 pages
= bufb
->array
[sb_bindex
];
324 pages
->data_size
= data_size
;
328 unsigned long subbuffer_get_read_data_size(
329 const struct lttng_kernel_ring_buffer_config
*config
,
330 struct lttng_kernel_ring_buffer_backend
*bufb
)
332 struct lttng_kernel_ring_buffer_backend_pages
*pages
;
333 unsigned long sb_bindex
;
335 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_rsb
.id
);
336 pages
= bufb
->array
[sb_bindex
];
337 return pages
->data_size
;
341 unsigned long subbuffer_get_data_size(
342 const struct lttng_kernel_ring_buffer_config
*config
,
343 struct lttng_kernel_ring_buffer_backend
*bufb
,
346 struct lttng_kernel_ring_buffer_backend_pages
*pages
;
347 unsigned long sb_bindex
;
349 sb_bindex
= subbuffer_id_get_index(config
, bufb
->buf_wsb
[idx
].id
);
350 pages
= bufb
->array
[sb_bindex
];
351 return pages
->data_size
;
355 void subbuffer_inc_packet_count(const struct lttng_kernel_ring_buffer_config
*config
,
356 struct lttng_kernel_ring_buffer_backend
*bufb
,
359 bufb
->buf_cnt
[idx
].seq_cnt
++;
363 * lib_ring_buffer_clear_noref - Clear the noref subbuffer flag, called by
367 void lib_ring_buffer_clear_noref(const struct lttng_kernel_ring_buffer_config
*config
,
368 struct lttng_kernel_ring_buffer_backend
*bufb
,
371 unsigned long id
, new_id
;
373 if (config
->mode
!= RING_BUFFER_OVERWRITE
)
377 * Performing a volatile access to read the sb_pages, because we want to
378 * read a coherent version of the pointer and the associated noref flag.
380 id
= LTTNG_READ_ONCE(bufb
->buf_wsb
[idx
].id
);
382 /* This check is called on the fast path for each record. */
383 if (likely(!subbuffer_id_is_noref(config
, id
))) {
385 * Store after load dependency ordering the writes to
386 * the subbuffer after load and test of the noref flag
387 * matches the memory barrier implied by the cmpxchg()
388 * in update_read_sb_index().
390 return; /* Already writing to this buffer */
393 subbuffer_id_clear_noref(config
, &new_id
);
394 new_id
= cmpxchg(&bufb
->buf_wsb
[idx
].id
, id
, new_id
);
395 if (likely(new_id
== id
))
402 * lib_ring_buffer_set_noref_offset - Set the noref subbuffer flag and offset,
406 void lib_ring_buffer_set_noref_offset(const struct lttng_kernel_ring_buffer_config
*config
,
407 struct lttng_kernel_ring_buffer_backend
*bufb
,
408 unsigned long idx
, unsigned long offset
)
410 if (config
->mode
!= RING_BUFFER_OVERWRITE
)
414 * Because ring_buffer_set_noref() is only called by a single thread
415 * (the one which updated the cc_sb value), there are no concurrent
416 * updates to take care of: other writers have not updated cc_sb, so
417 * they cannot set the noref flag, and concurrent readers cannot modify
418 * the pointer because the noref flag is not set yet.
419 * The smp_wmb() in ring_buffer_commit() takes care of ordering writes
420 * to the subbuffer before this set noref operation.
421 * subbuffer_set_noref() uses a volatile store to deal with concurrent
422 * readers of the noref flag.
424 CHAN_WARN_ON(bufb
->chan
,
425 subbuffer_id_is_noref(config
, bufb
->buf_wsb
[idx
].id
));
427 * Memory barrier that ensures counter stores are ordered before set
431 subbuffer_id_set_noref_offset(config
, &bufb
->buf_wsb
[idx
].id
, offset
);
435 * update_read_sb_index - Read-side subbuffer index update.
438 int update_read_sb_index(const struct lttng_kernel_ring_buffer_config
*config
,
439 struct lttng_kernel_ring_buffer_backend
*bufb
,
440 struct channel_backend
*chanb
,
441 unsigned long consumed_idx
,
442 unsigned long consumed_count
)
444 unsigned long old_id
, new_id
;
446 if (config
->mode
== RING_BUFFER_OVERWRITE
) {
448 * Exchange the target writer subbuffer with our own unused
449 * subbuffer. No need to use READ_ONCE() here to read the
450 * old_wpage, because the value read will be confirmed by the
451 * following cmpxchg().
453 old_id
= bufb
->buf_wsb
[consumed_idx
].id
;
454 if (unlikely(!subbuffer_id_is_noref(config
, old_id
)))
457 * Make sure the offset count we are expecting matches the one
458 * indicated by the writer.
460 if (unlikely(!subbuffer_id_compare_offset(config
, old_id
,
463 CHAN_WARN_ON(bufb
->chan
,
464 !subbuffer_id_is_noref(config
, bufb
->buf_rsb
.id
));
465 subbuffer_id_set_noref_offset(config
, &bufb
->buf_rsb
.id
,
467 new_id
= cmpxchg(&bufb
->buf_wsb
[consumed_idx
].id
, old_id
,
469 if (unlikely(old_id
!= new_id
))
471 bufb
->buf_rsb
.id
= new_id
;
473 /* No page exchange, use the writer page directly */
474 bufb
->buf_rsb
.id
= bufb
->buf_wsb
[consumed_idx
].id
;
479 static inline __attribute__((always_inline
))
480 void lttng_inline_memcpy(void *dest
, const void *src
,
485 *(uint8_t *) dest
= *(const uint8_t *) src
;
488 *(uint16_t *) dest
= *(const uint16_t *) src
;
491 *(uint32_t *) dest
= *(const uint32_t *) src
;
494 *(uint64_t *) dest
= *(const uint64_t *) src
;
497 inline_memcpy(dest
, src
, len
);
502 * Use the architecture-specific memcpy implementation for constant-sized
503 * inputs, but rely on an inline memcpy for length statically unknown.
504 * The function call to memcpy is just way too expensive for a fast path.
506 #define lib_ring_buffer_do_copy(config, dest, src, len) \
508 size_t __len = (len); \
509 if (__builtin_constant_p(len)) \
510 memcpy(dest, src, __len); \
512 lttng_inline_memcpy(dest, src, __len); \
516 * We use __copy_from_user_inatomic to copy userspace data since we already
517 * did the access_ok for the whole range.
519 * Return 0 if OK, nonzero on error.
522 unsigned long lib_ring_buffer_do_copy_from_user_inatomic(void *dest
,
523 const void __user
*src
,
526 return __copy_from_user_inatomic(dest
, src
, len
);
530 * write len bytes to dest with c
533 void lib_ring_buffer_do_memset(char *dest
, int c
,
538 for (i
= 0; i
< len
; i
++)
542 #endif /* _LIB_RING_BUFFER_BACKEND_INTERNAL_H */