1 |
/*************************************************************************** |
2 |
* * |
3 |
* LinuxSampler - modular, streaming capable sampler * |
4 |
* * |
5 |
* Copyright (C) 2003, 2004 by Benno Senoner and Christian Schoenebeck * |
6 |
* Copyright (C) 2005 - 2008 Christian Schoenebeck * |
7 |
* * |
8 |
* This program is free software; you can redistribute it and/or modify * |
9 |
* it under the terms of the GNU General Public License as published by * |
10 |
* the Free Software Foundation; either version 2 of the License, or * |
11 |
* (at your option) any later version. * |
12 |
* * |
13 |
* This program 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 * |
16 |
* GNU General Public License for more details. * |
17 |
* * |
18 |
* You should have received a copy of the GNU General Public License * |
19 |
* along with this program; if not, write to the Free Software * |
20 |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, * |
21 |
* MA 02111-1307 USA * |
22 |
***************************************************************************/ |
23 |
|
24 |
#ifndef RINGBUFFER_H |
25 |
#define RINGBUFFER_H |
26 |
|
27 |
#define DEFAULT_WRAP_ELEMENTS 0 |
28 |
|
29 |
#include <string.h> |
30 |
|
31 |
#include "lsatomic.h" |
32 |
|
33 |
using LinuxSampler::atomic; |
34 |
using LinuxSampler::memory_order_relaxed; |
35 |
using LinuxSampler::memory_order_acquire; |
36 |
using LinuxSampler::memory_order_release; |
37 |
|
38 |
|
39 |
/** @brief Real-time safe and type safe RingBuffer implementation. |
40 |
* |
41 |
* This constant size buffer can be used to send data from exactly one |
42 |
* sender / writing thread to exactly one receiver / reading thread. It is |
43 |
* real-time safe due to the fact that data is only allocated when this |
44 |
* RingBuffer is created and no system level mechanisms are used for |
45 |
* ensuring thread safety of this class. |
46 |
* |
47 |
* <b>Important:</b> There are two distinct behaviors of this RingBuffer |
48 |
* which has to be given as template argument @c T_DEEP_COPY, which is a |
49 |
* boolean flag: |
50 |
* |
51 |
* - @c true: The RingBuffer will copy elements of type @c T by using type |
52 |
* @c T's assignment operator. This behavior is mandatory for all data |
53 |
* structures (classes) which additionally allocate memory on the heap. |
54 |
* Type @c T's needs to have an assignment operator implementation though, |
55 |
* otherwise this will cause a compilation error. This behavior is more |
56 |
* safe, but usually slower (except for very small buffer sizes, where it |
57 |
* might be even faster). |
58 |
* - @c false: The RingBuffer will copy elements of type @c T by flatly |
59 |
* copying their structural data ( i.e. with @c memcpy() ) in one piece. |
60 |
* This will only work if class @c T (and all of its subelements) does not |
61 |
* allocate any additional data on the heap by itself. So use this option |
62 |
* with great care, because otherwise it will result in very ugly behavior |
63 |
* and crashes! For larger buffer sizes, this behavior will most probably |
64 |
* be faster. |
65 |
*/ |
66 |
template<class T, bool T_DEEP_COPY> |
67 |
class RingBuffer |
68 |
{ |
69 |
public: |
70 |
RingBuffer (int sz, int wrap_elements = DEFAULT_WRAP_ELEMENTS) : |
71 |
write_ptr(0), read_ptr(0) { |
72 |
int power_of_two; |
73 |
|
74 |
this->wrap_elements = wrap_elements; |
75 |
|
76 |
// the write-with-wrap functions need wrap_elements extra |
77 |
// space in the buffer to be able to copy the wrap space |
78 |
sz += wrap_elements; |
79 |
|
80 |
for (power_of_two = 1; |
81 |
1<<power_of_two < sz; |
82 |
power_of_two++); |
83 |
|
84 |
size = 1<<power_of_two; |
85 |
size_mask = size; |
86 |
size_mask -= 1; |
87 |
buf = new T[size + wrap_elements]; |
88 |
}; |
89 |
|
90 |
virtual ~RingBuffer() { |
91 |
delete [] buf; |
92 |
} |
93 |
|
94 |
/** |
95 |
* Sets all remaining write space elements to zero. The write pointer |
96 |
* will currently not be incremented after, but that might change in |
97 |
* future. |
98 |
* |
99 |
* @e Caution: for @c T_DEEP_COPY=true you might probably @e NOT want |
100 |
* to call this method at all, at least not in case type @c T allocates |
101 |
* any additional data on the heap by itself. |
102 |
*/ |
103 |
inline void fill_write_space_with_null() { |
104 |
int w = write_ptr.load(memory_order_relaxed), |
105 |
r = read_ptr.load(memory_order_acquire); |
106 |
memset(get_write_ptr(), 0, sizeof(T)*write_space_to_end()); |
107 |
if (r && w >= r) { |
108 |
memset(get_buffer_begin(), 0, sizeof(T)*(r - 1)); |
109 |
} |
110 |
|
111 |
// set the wrap space elements to null |
112 |
if (wrap_elements) memset(&buf[size], 0, sizeof(T)*wrap_elements); |
113 |
} |
114 |
|
115 |
__inline int read (T *dest, int cnt); |
116 |
__inline int write (T *src, int cnt); |
117 |
|
118 |
inline int push(T* src) { return write(src,1); } |
119 |
inline int pop(T* dst) { return read(dst,1); } |
120 |
|
121 |
__inline T *get_buffer_begin(); |
122 |
|
123 |
__inline T *get_read_ptr(void) { |
124 |
return(&buf[read_ptr.load(memory_order_relaxed)]); |
125 |
} |
126 |
|
127 |
/** |
128 |
* Returns a pointer to the element from the current read position, |
129 |
* advanced by \a offset elements. |
130 |
*/ |
131 |
/*inline T* get_read_ptr(int offset) { |
132 |
int r = read_ptr.load(memory_order_relaxed); |
133 |
r += offset; |
134 |
r &= size_mask; |
135 |
return &buf[r]; |
136 |
}*/ |
137 |
|
138 |
__inline T *get_write_ptr(); |
139 |
__inline void increment_read_ptr(int cnt) { |
140 |
read_ptr.store((read_ptr.load(memory_order_relaxed) + cnt) & size_mask, memory_order_release); |
141 |
} |
142 |
__inline void set_read_ptr(int val) { |
143 |
read_ptr.store(val, memory_order_release); |
144 |
} |
145 |
|
146 |
__inline void increment_write_ptr(int cnt) { |
147 |
write_ptr.store((write_ptr.load(memory_order_relaxed) + cnt) & size_mask, memory_order_release); |
148 |
} |
149 |
|
150 |
/* this function increments the write_ptr by cnt, if the buffer wraps then |
151 |
subtract size from the write_ptr value so that it stays within 0<write_ptr<size |
152 |
use this function to increment the write ptr after you filled the buffer |
153 |
with a number of elements given by write_space_to_end_with_wrap(). |
154 |
This ensures that the data that is written to the buffer fills up all |
155 |
the wrap space that resides past the regular buffer. The wrap_space is needed for |
156 |
interpolation. So that the audio thread sees the ringbuffer like a linear space |
157 |
which allows us to use faster routines. |
158 |
When the buffer wraps the wrap part is memcpy()ied to the beginning of the buffer |
159 |
and the write ptr incremented accordingly. |
160 |
*/ |
161 |
__inline void increment_write_ptr_with_wrap(int cnt) { |
162 |
int w = write_ptr.load(memory_order_relaxed); |
163 |
w += cnt; |
164 |
if(w >= size) { |
165 |
w -= size; |
166 |
copy(&buf[0], &buf[size], w); |
167 |
//printf("DEBUG !!!! increment_write_ptr_with_wrap: buffer wrapped, elements wrapped = %d (wrap_elements %d)\n",w,wrap_elements); |
168 |
} |
169 |
write_ptr.store(w, memory_order_release); |
170 |
} |
171 |
|
172 |
/* this function returns the available write space in the buffer |
173 |
when the read_ptr > write_ptr it returns the space inbetween, otherwise |
174 |
when the read_ptr < write_ptr it returns the space between write_ptr and |
175 |
the buffer end, including the wrap_space. |
176 |
There is an exception to it. When read_ptr <= wrap_elements. In that |
177 |
case we return the write space to buffer end (-1) without the wrap_elements, |
178 |
this is needed because a subsequent increment_write_ptr which copies the |
179 |
data that resides into the wrap space to the beginning of the buffer and increments |
180 |
the write_ptr would cause the write_ptr overstepping the read_ptr which would be an error. |
181 |
So basically the if(r<=wrap_elements) we return the buffer space to end - 1 which |
182 |
ensures that at the next call there will be one element free to write before the buffer wraps |
183 |
and usually (except in EOF situations) the next write_space_to_end_with_wrap() will return |
184 |
1 + wrap_elements which ensures that the wrap part gets fully filled with data |
185 |
*/ |
186 |
__inline int write_space_to_end_with_wrap() { |
187 |
int w, r; |
188 |
|
189 |
w = write_ptr.load(memory_order_relaxed); |
190 |
r = read_ptr.load(memory_order_acquire); |
191 |
//printf("write_space_to_end: w=%d r=%d\n",w,r); |
192 |
if(r > w) { |
193 |
//printf("DEBUG: write_space_to_end_with_wrap: r>w r=%d w=%d val=%d\n",r,w,r - w - 1); |
194 |
return(r - w - 1); |
195 |
} |
196 |
if(r <= wrap_elements) { |
197 |
//printf("DEBUG: write_space_to_end_with_wrap: ATTENTION r <= wrap_elements: r=%d w=%d val=%d\n",r,w,size - w -1); |
198 |
return(size - w -1); |
199 |
} |
200 |
if(r) { |
201 |
//printf("DEBUG: write_space_to_end_with_wrap: r=%d w=%d val=%d\n",r,w,size - w + wrap_elements); |
202 |
return(size - w + wrap_elements); |
203 |
} |
204 |
//printf("DEBUG: write_space_to_end_with_wrap: r=0 w=%d val=%d\n",w,size - w - 1 + wrap_elements); |
205 |
return(size - w - 1 + wrap_elements); |
206 |
} |
207 |
|
208 |
/* this function adjusts the number of elements to write into the ringbuffer |
209 |
in a way that the size boundary is avoided and that the wrap space always gets |
210 |
entirely filled. |
211 |
cnt contains the write_space_to_end_with_wrap() amount while |
212 |
capped_cnt contains a capped amount of samples to read. |
213 |
normally capped_cnt == cnt but in some cases eg when the disk thread needs |
214 |
to refill tracks with smaller blocks because lots of streams require immediate |
215 |
refill because lots of notes were started simultaneously. |
216 |
In that case we set for example capped_cnt to a fixed amount (< cnt, eg 64k), |
217 |
which helps to reduce the buffer refill latencies that occur between streams. |
218 |
the first if() checks if the current write_ptr + capped_cnt resides within |
219 |
the wrap area but is < size+wrap_elements. in that case we cannot return |
220 |
capped_cnt because it would lead to a write_ptr wrapping and only a partial fill |
221 |
of wrap space which would lead to errors. So we simply return cnt which ensures |
222 |
that the the entire wrap space will get filled correctly. |
223 |
In all other cases (which are not problematic because no write_ptr wrapping |
224 |
occurs) we simply return capped_cnt. |
225 |
*/ |
226 |
__inline int adjust_write_space_to_avoid_boundary(int cnt, int capped_cnt) { |
227 |
int w; |
228 |
w = write_ptr.load(memory_order_relaxed); |
229 |
if((w+capped_cnt) >= size && (w+capped_cnt) < (size+wrap_elements)) { |
230 |
//printf("adjust_write_space_to_avoid_boundary returning cnt = %d\n",cnt); |
231 |
return(cnt); |
232 |
} |
233 |
//printf("adjust_write_space_to_avoid_boundary returning capped_cnt = %d\n",capped_cnt); |
234 |
return(capped_cnt); |
235 |
} |
236 |
|
237 |
__inline int write_space_to_end() { |
238 |
int w, r; |
239 |
|
240 |
w = write_ptr.load(memory_order_relaxed); |
241 |
r = read_ptr.load(memory_order_acquire); |
242 |
//printf("write_space_to_end: w=%d r=%d\n",w,r); |
243 |
if(r > w) return(r - w - 1); |
244 |
if(r) return(size - w); |
245 |
return(size - w - 1); |
246 |
} |
247 |
|
248 |
__inline int read_space_to_end() { |
249 |
int w, r; |
250 |
|
251 |
w = write_ptr.load(memory_order_acquire); |
252 |
r = read_ptr.load(memory_order_relaxed); |
253 |
if(w >= r) return(w - r); |
254 |
return(size - r); |
255 |
} |
256 |
__inline void init() { |
257 |
write_ptr.store(0, memory_order_relaxed); |
258 |
read_ptr.store(0, memory_order_relaxed); |
259 |
// wrap=0; |
260 |
} |
261 |
|
262 |
int write_space () { |
263 |
int w, r; |
264 |
|
265 |
w = write_ptr.load(memory_order_relaxed); |
266 |
r = read_ptr.load(memory_order_acquire); |
267 |
|
268 |
if (w > r) { |
269 |
return ((r - w + size) & size_mask) - 1; |
270 |
} else if (w < r) { |
271 |
return (r - w) - 1; |
272 |
} else { |
273 |
return size - 1; |
274 |
} |
275 |
} |
276 |
|
277 |
int read_space () { |
278 |
int w, r; |
279 |
|
280 |
w = write_ptr.load(memory_order_acquire); |
281 |
r = read_ptr.load(memory_order_relaxed); |
282 |
|
283 |
if (w >= r) { |
284 |
return w - r; |
285 |
} else { |
286 |
return (w - r + size) & size_mask; |
287 |
} |
288 |
} |
289 |
|
290 |
int size; |
291 |
int wrap_elements; |
292 |
|
293 |
/** |
294 |
* Independent, random access reading from a RingBuffer. This class |
295 |
* allows to read from a RingBuffer without being forced to free read |
296 |
* data while reading / positioning. |
297 |
*/ |
298 |
template<class T1, bool T1_DEEP_COPY> |
299 |
class _NonVolatileReader { |
300 |
public: |
301 |
int read_space() { |
302 |
int r = read_ptr; |
303 |
int w = pBuf->write_ptr.load(memory_order_acquire); |
304 |
return (w >= r) ? w - r : (w - r + pBuf->size) & pBuf->size_mask; |
305 |
} |
306 |
|
307 |
/** |
308 |
* Prefix decrement operator, for reducing NonVolatileReader's |
309 |
* read position by one. |
310 |
*/ |
311 |
inline void operator--() { |
312 |
if (read_ptr == pBuf->read_ptr.load(memory_order_relaxed)) return; //TODO: or should we react oh this case (e.g. force segfault), as this is a very odd case? |
313 |
read_ptr = (read_ptr-1) & pBuf->size_mask; |
314 |
} |
315 |
|
316 |
/** |
317 |
* Postfix decrement operator, for reducing NonVolatileReader's |
318 |
* read position by one. |
319 |
*/ |
320 |
inline void operator--(int) { |
321 |
--*this; |
322 |
} |
323 |
|
324 |
/** |
325 |
* Returns pointer to the RingBuffer data of current |
326 |
* NonVolatileReader's read position and increments |
327 |
* NonVolatileReader's read position by one. |
328 |
* |
329 |
* @returns pointer to element of current read position |
330 |
*/ |
331 |
T* pop() { |
332 |
if (!read_space()) return NULL; |
333 |
T* pData = &pBuf->buf[read_ptr]; |
334 |
read_ptr++; |
335 |
read_ptr &= pBuf->size_mask; |
336 |
return pData; |
337 |
} |
338 |
|
339 |
/** |
340 |
* Reads one element from the NonVolatileReader's current read |
341 |
* position and copies it to the variable pointed by \a dst and |
342 |
* finally increments the NonVolatileReader's read position by |
343 |
* one. |
344 |
* |
345 |
* @param dst - where the element is copied to |
346 |
* @returns 1 on success, 0 otherwise |
347 |
*/ |
348 |
int pop(T* dst) { return read(dst,1); } |
349 |
|
350 |
/** |
351 |
* Reads \a cnt elements from the NonVolatileReader's current |
352 |
* read position and copies it to the buffer pointed by \a dest |
353 |
* and finally increments the NonVolatileReader's read position |
354 |
* by the number of read elements. |
355 |
* |
356 |
* @param dest - destination buffer |
357 |
* @param cnt - number of elements to read |
358 |
* @returns number of read elements |
359 |
*/ |
360 |
int read(T* dest, int cnt) { |
361 |
int free_cnt; |
362 |
int cnt2; |
363 |
int to_read; |
364 |
int n1, n2; |
365 |
int priv_read_ptr; |
366 |
|
367 |
priv_read_ptr = read_ptr; |
368 |
|
369 |
if ((free_cnt = read_space()) == 0) return 0; |
370 |
|
371 |
to_read = cnt > free_cnt ? free_cnt : cnt; |
372 |
|
373 |
cnt2 = priv_read_ptr + to_read; |
374 |
|
375 |
if (cnt2 > pBuf->size) { |
376 |
n1 = pBuf->size - priv_read_ptr; |
377 |
n2 = cnt2 & pBuf->size_mask; |
378 |
} else { |
379 |
n1 = to_read; |
380 |
n2 = 0; |
381 |
} |
382 |
|
383 |
copy(dest, &pBuf->buf[priv_read_ptr], n1); |
384 |
priv_read_ptr = (priv_read_ptr + n1) & pBuf->size_mask; |
385 |
|
386 |
if (n2) { |
387 |
copy(dest+n1, pBuf->buf, n2); |
388 |
priv_read_ptr = n2; |
389 |
} |
390 |
|
391 |
this->read_ptr = priv_read_ptr; |
392 |
return to_read; |
393 |
} |
394 |
|
395 |
/** |
396 |
* Finally when the read data is not needed anymore, this method |
397 |
* should be called to free the data in the RingBuffer up to the |
398 |
* current read position of this NonVolatileReader. |
399 |
* |
400 |
* @see RingBuffer::increment_read_ptr() |
401 |
*/ |
402 |
void free() { |
403 |
pBuf->read_ptr.store(read_ptr, memory_order_release); |
404 |
} |
405 |
|
406 |
protected: |
407 |
_NonVolatileReader(RingBuffer<T1,T1_DEEP_COPY>* pBuf) { |
408 |
this->pBuf = pBuf; |
409 |
this->read_ptr = pBuf->read_ptr.load(memory_order_relaxed); |
410 |
} |
411 |
|
412 |
RingBuffer<T1,T1_DEEP_COPY>* pBuf; |
413 |
int read_ptr; |
414 |
|
415 |
friend class RingBuffer<T1,T1_DEEP_COPY>; |
416 |
}; |
417 |
|
418 |
typedef _NonVolatileReader<T,T_DEEP_COPY> NonVolatileReader; |
419 |
|
420 |
NonVolatileReader get_non_volatile_reader() { return NonVolatileReader(this); } |
421 |
|
422 |
protected: |
423 |
T *buf; |
424 |
atomic<int> write_ptr; |
425 |
atomic<int> read_ptr; |
426 |
int size_mask; |
427 |
|
428 |
/** |
429 |
* Copies \a n amount of elements from the buffer given by |
430 |
* \a pSrc to the buffer given by \a pDst. |
431 |
*/ |
432 |
inline static void copy(T* pDst, T* pSrc, int n); |
433 |
|
434 |
friend class _NonVolatileReader<T,T_DEEP_COPY>; |
435 |
}; |
436 |
|
437 |
template<class T, bool T_DEEP_COPY> |
438 |
T* RingBuffer<T,T_DEEP_COPY>::get_write_ptr (void) { |
439 |
return(&buf[write_ptr.load(memory_order_relaxed)]); |
440 |
} |
441 |
|
442 |
template<class T, bool T_DEEP_COPY> |
443 |
T* RingBuffer<T,T_DEEP_COPY>::get_buffer_begin (void) { |
444 |
return(buf); |
445 |
} |
446 |
|
447 |
|
448 |
|
449 |
template<class T, bool T_DEEP_COPY> |
450 |
int RingBuffer<T,T_DEEP_COPY>::read(T* dest, int cnt) |
451 |
{ |
452 |
int free_cnt; |
453 |
int cnt2; |
454 |
int to_read; |
455 |
int n1, n2; |
456 |
int priv_read_ptr; |
457 |
|
458 |
priv_read_ptr = read_ptr.load(memory_order_relaxed); |
459 |
|
460 |
if ((free_cnt = read_space ()) == 0) { |
461 |
return 0; |
462 |
} |
463 |
|
464 |
to_read = cnt > free_cnt ? free_cnt : cnt; |
465 |
|
466 |
cnt2 = priv_read_ptr + to_read; |
467 |
|
468 |
if (cnt2 > size) { |
469 |
n1 = size - priv_read_ptr; |
470 |
n2 = cnt2 & size_mask; |
471 |
} else { |
472 |
n1 = to_read; |
473 |
n2 = 0; |
474 |
} |
475 |
|
476 |
copy(dest, &buf[priv_read_ptr], n1); |
477 |
priv_read_ptr = (priv_read_ptr + n1) & size_mask; |
478 |
|
479 |
if (n2) { |
480 |
copy(dest+n1, buf, n2); |
481 |
priv_read_ptr = n2; |
482 |
} |
483 |
|
484 |
read_ptr.store(priv_read_ptr, memory_order_release); |
485 |
return to_read; |
486 |
} |
487 |
|
488 |
template<class T, bool T_DEEP_COPY> |
489 |
int RingBuffer<T,T_DEEP_COPY>::write(T* src, int cnt) |
490 |
{ |
491 |
int free_cnt; |
492 |
int cnt2; |
493 |
int to_write; |
494 |
int n1, n2; |
495 |
int priv_write_ptr; |
496 |
|
497 |
priv_write_ptr = write_ptr.load(memory_order_relaxed); |
498 |
|
499 |
if ((free_cnt = write_space ()) == 0) { |
500 |
return 0; |
501 |
} |
502 |
|
503 |
to_write = cnt > free_cnt ? free_cnt : cnt; |
504 |
|
505 |
cnt2 = priv_write_ptr + to_write; |
506 |
|
507 |
if (cnt2 > size) { |
508 |
n1 = size - priv_write_ptr; |
509 |
n2 = cnt2 & size_mask; |
510 |
} else { |
511 |
n1 = to_write; |
512 |
n2 = 0; |
513 |
} |
514 |
|
515 |
copy(&buf[priv_write_ptr], src, n1); |
516 |
priv_write_ptr = (priv_write_ptr + n1) & size_mask; |
517 |
|
518 |
if (n2) { |
519 |
copy(buf, src+n1, n2); |
520 |
priv_write_ptr = n2; |
521 |
} |
522 |
write_ptr.store(priv_write_ptr, memory_order_release); |
523 |
return to_write; |
524 |
} |
525 |
|
526 |
template<class T, bool T_DEEP_COPY> |
527 |
void RingBuffer<T,T_DEEP_COPY>::copy(T* pDst, T* pSrc, int n) { |
528 |
if (T_DEEP_COPY) { // deep copy - won't work for data structures without assignment operator implementation |
529 |
for (int i = 0; i < n; i++) pDst[i] = pSrc[i]; |
530 |
} else { // flat copy - won't work for complex data structures ! |
531 |
memcpy(pDst, pSrc, n * sizeof(T)); |
532 |
} |
533 |
} |
534 |
|
535 |
#endif /* RINGBUFFER_H */ |