1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
3 */
4
5 #ifndef __INCLUDE_RTE_BITMAP_H__
6 #define __INCLUDE_RTE_BITMAP_H__
7
8 #ifdef __cplusplus
9 extern "C" {
10 #endif
11
12 /**
13 * @file
14 * RTE Bitmap
15 *
16 * The bitmap component provides a mechanism to manage large arrays of bits
17 * through bit get/set/clear and bit array scan operations.
18 *
19 * The bitmap scan operation is optimized for 64-bit CPUs using 64/128 byte cache
20 * lines. The bitmap is hierarchically organized using two arrays (array1 and
21 * array2), with each bit in array1 being associated with a full cache line
22 * (512/1024 bits) of bitmap bits, which are stored in array2: the bit in array1
23 * is set only when there is at least one bit set within its associated array2
24 * bits, otherwise the bit in array1 is cleared. The read and write operations
25 * for array1 and array2 are always done in slabs of 64 bits.
26 *
27 * This bitmap is not thread safe. For lock free operation on a specific bitmap
28 * instance, a single writer thread performing bit set/clear operations is
29 * allowed, only the writer thread can do bitmap scan operations, while there
30 * can be several reader threads performing bit get operations in parallel with
31 * the writer thread. When the use of locking primitives is acceptable, the
32 * serialization of the bit set/clear and bitmap scan operations needs to be
33 * enforced by the caller, while the bit get operation does not require locking
34 * the bitmap.
35 *
36 ***/
37
38 #include <string.h>
39 #include <rte_common.h>
40 #include <rte_config.h>
41 #include <rte_debug.h>
42 #include <rte_memory.h>
43 #include <rte_branch_prediction.h>
44 #include <rte_prefetch.h>
45
46 /* Slab */
47 #define RTE_BITMAP_SLAB_BIT_SIZE 64
48 #define RTE_BITMAP_SLAB_BIT_SIZE_LOG2 6
49 #define RTE_BITMAP_SLAB_BIT_MASK (RTE_BITMAP_SLAB_BIT_SIZE - 1)
50
51 /* Cache line (CL) */
52 #define RTE_BITMAP_CL_BIT_SIZE (RTE_CACHE_LINE_SIZE * 8)
53 #define RTE_BITMAP_CL_BIT_SIZE_LOG2 (RTE_CACHE_LINE_SIZE_LOG2 + 3)
54 #define RTE_BITMAP_CL_BIT_MASK (RTE_BITMAP_CL_BIT_SIZE - 1)
55
56 #define RTE_BITMAP_CL_SLAB_SIZE (RTE_BITMAP_CL_BIT_SIZE / RTE_BITMAP_SLAB_BIT_SIZE)
57 #define RTE_BITMAP_CL_SLAB_SIZE_LOG2 (RTE_BITMAP_CL_BIT_SIZE_LOG2 - RTE_BITMAP_SLAB_BIT_SIZE_LOG2)
58 #define RTE_BITMAP_CL_SLAB_MASK (RTE_BITMAP_CL_SLAB_SIZE - 1)
59
60 /** Bitmap data structure */
61 struct rte_bitmap {
62 /* Context for array1 and array2 */
63 uint64_t *array1; /**< Bitmap array1 */
64 uint64_t *array2; /**< Bitmap array2 */
65 uint32_t array1_size; /**< Number of 64-bit slabs in array1 that are actually used */
66 uint32_t array2_size; /**< Number of 64-bit slabs in array2 */
67
68 /* Context for the "scan next" operation */
69 uint32_t index1; /**< Bitmap scan: Index of current array1 slab */
70 uint32_t offset1; /**< Bitmap scan: Offset of current bit within current array1 slab */
71 uint32_t index2; /**< Bitmap scan: Index of current array2 slab */
72 uint32_t go2; /**< Bitmap scan: Go/stop condition for current array2 cache line */
73
74 /* Storage space for array1 and array2 */
75 uint8_t memory[];
76 };
77
78 static inline void
__rte_bitmap_index1_inc(struct rte_bitmap * bmp)79 __rte_bitmap_index1_inc(struct rte_bitmap *bmp)
80 {
81 bmp->index1 = (bmp->index1 + 1) & (bmp->array1_size - 1);
82 }
83
84 static inline uint64_t
__rte_bitmap_mask1_get(struct rte_bitmap * bmp)85 __rte_bitmap_mask1_get(struct rte_bitmap *bmp)
86 {
87 return (~1llu) << bmp->offset1;
88 }
89
90 static inline void
__rte_bitmap_index2_set(struct rte_bitmap * bmp)91 __rte_bitmap_index2_set(struct rte_bitmap *bmp)
92 {
93 bmp->index2 = (((bmp->index1 << RTE_BITMAP_SLAB_BIT_SIZE_LOG2) + bmp->offset1) << RTE_BITMAP_CL_SLAB_SIZE_LOG2);
94 }
95
96 static inline uint32_t
__rte_bitmap_get_memory_footprint(uint32_t n_bits,uint32_t * array1_byte_offset,uint32_t * array1_slabs,uint32_t * array2_byte_offset,uint32_t * array2_slabs)97 __rte_bitmap_get_memory_footprint(uint32_t n_bits,
98 uint32_t *array1_byte_offset, uint32_t *array1_slabs,
99 uint32_t *array2_byte_offset, uint32_t *array2_slabs)
100 {
101 uint32_t n_slabs_context, n_slabs_array1, n_cache_lines_context_and_array1;
102 uint32_t n_cache_lines_array2;
103 uint32_t n_bytes_total;
104
105 n_cache_lines_array2 = (n_bits + RTE_BITMAP_CL_BIT_SIZE - 1) / RTE_BITMAP_CL_BIT_SIZE;
106 n_slabs_array1 = (n_cache_lines_array2 + RTE_BITMAP_SLAB_BIT_SIZE - 1) / RTE_BITMAP_SLAB_BIT_SIZE;
107 n_slabs_array1 = rte_align32pow2(n_slabs_array1);
108 n_slabs_context = (sizeof(struct rte_bitmap) + (RTE_BITMAP_SLAB_BIT_SIZE / 8) - 1) / (RTE_BITMAP_SLAB_BIT_SIZE / 8);
109 n_cache_lines_context_and_array1 = (n_slabs_context + n_slabs_array1 + RTE_BITMAP_CL_SLAB_SIZE - 1) / RTE_BITMAP_CL_SLAB_SIZE;
110 n_bytes_total = (n_cache_lines_context_and_array1 + n_cache_lines_array2) * RTE_CACHE_LINE_SIZE;
111
112 if (array1_byte_offset) {
113 *array1_byte_offset = n_slabs_context * (RTE_BITMAP_SLAB_BIT_SIZE / 8);
114 }
115 if (array1_slabs) {
116 *array1_slabs = n_slabs_array1;
117 }
118 if (array2_byte_offset) {
119 *array2_byte_offset = n_cache_lines_context_and_array1 * RTE_CACHE_LINE_SIZE;
120 }
121 if (array2_slabs) {
122 *array2_slabs = n_cache_lines_array2 * RTE_BITMAP_CL_SLAB_SIZE;
123 }
124
125 return n_bytes_total;
126 }
127
128 static inline void
__rte_bitmap_scan_init(struct rte_bitmap * bmp)129 __rte_bitmap_scan_init(struct rte_bitmap *bmp)
130 {
131 bmp->index1 = bmp->array1_size - 1;
132 bmp->offset1 = RTE_BITMAP_SLAB_BIT_SIZE - 1;
133 __rte_bitmap_index2_set(bmp);
134 bmp->index2 += RTE_BITMAP_CL_SLAB_SIZE;
135
136 bmp->go2 = 0;
137 }
138
139 /**
140 * Bitmap memory footprint calculation
141 *
142 * @param n_bits
143 * Number of bits in the bitmap
144 * @return
145 * Bitmap memory footprint measured in bytes on success, 0 on error
146 */
147 static inline uint32_t
rte_bitmap_get_memory_footprint(uint32_t n_bits)148 rte_bitmap_get_memory_footprint(uint32_t n_bits) {
149 /* Check input arguments */
150 if (n_bits == 0) {
151 return 0;
152 }
153
154 return __rte_bitmap_get_memory_footprint(n_bits, NULL, NULL, NULL, NULL);
155 }
156
157 /**
158 * Bitmap initialization
159 *
160 * @param n_bits
161 * Number of pre-allocated bits in array2.
162 * @param mem
163 * Base address of array1 and array2.
164 * @param mem_size
165 * Minimum expected size of bitmap.
166 * @return
167 * Handle to bitmap instance.
168 */
169 static inline struct rte_bitmap *
rte_bitmap_init(uint32_t n_bits,uint8_t * mem,uint32_t mem_size)170 rte_bitmap_init(uint32_t n_bits, uint8_t *mem, uint32_t mem_size)
171 {
172 struct rte_bitmap *bmp;
173 uint32_t array1_byte_offset, array1_slabs, array2_byte_offset, array2_slabs;
174 uint32_t size;
175
176 /* Check input arguments */
177 if (n_bits == 0) {
178 return NULL;
179 }
180
181 if ((mem == NULL) || (((uintptr_t) mem) & RTE_CACHE_LINE_MASK)) {
182 return NULL;
183 }
184
185 size = __rte_bitmap_get_memory_footprint(n_bits,
186 &array1_byte_offset, &array1_slabs,
187 &array2_byte_offset, &array2_slabs);
188 if (size < mem_size) {
189 return NULL;
190 }
191
192 /* Setup bitmap */
193 memset(mem, 0, size);
194 bmp = (struct rte_bitmap *) mem;
195
196 bmp->array1 = (uint64_t *) &mem[array1_byte_offset];
197 bmp->array1_size = array1_slabs;
198 bmp->array2 = (uint64_t *) &mem[array2_byte_offset];
199 bmp->array2_size = array2_slabs;
200
201 __rte_bitmap_scan_init(bmp);
202
203 return bmp;
204 }
205
206 /**
207 * @warning
208 * @b EXPERIMENTAL: this API may change without prior notice.
209 *
210 * Bitmap clear slab overhead bits.
211 *
212 * @param slabs
213 * Slab array.
214 * @param slab_size
215 * Number of 64-bit slabs in the slabs array.
216 * @param pos
217 * The start bit position in the slabs to be cleared.
218 */
219 __rte_experimental
220 static inline void
__rte_bitmap_clear_slab_overhead_bits(uint64_t * slabs,uint32_t slab_size,uint32_t pos)221 __rte_bitmap_clear_slab_overhead_bits(uint64_t *slabs, uint32_t slab_size,
222 uint32_t pos)
223 {
224 uint32_t i;
225 uint32_t index = pos / RTE_BITMAP_SLAB_BIT_SIZE;
226 uint32_t offset = pos & RTE_BITMAP_SLAB_BIT_MASK;
227
228 if (offset) {
229 for (i = offset; i < RTE_BITMAP_SLAB_BIT_SIZE; i++)
230 slabs[index] &= ~(1llu << i);
231 index++;
232 }
233 if (index < slab_size)
234 memset(&slabs[index], 0, sizeof(slabs[0]) *
235 (slab_size - index));
236 }
237
238 /**
239 * @warning
240 * @b EXPERIMENTAL: this API may change without prior notice.
241 *
242 * Bitmap initialization with all bits set
243 *
244 * @param n_bits
245 * Number of pre-allocated bits in array2.
246 * @param mem
247 * Base address of array1 and array2.
248 * @param mem_size
249 * Minimum expected size of bitmap.
250 * @return
251 * Handle to bitmap instance.
252 */
253 __rte_experimental
254 static inline struct rte_bitmap *
rte_bitmap_init_with_all_set(uint32_t n_bits,uint8_t * mem,uint32_t mem_size)255 rte_bitmap_init_with_all_set(uint32_t n_bits, uint8_t *mem, uint32_t mem_size)
256 {
257 struct rte_bitmap *bmp;
258 uint32_t array1_byte_offset, array1_slabs;
259 uint32_t array2_byte_offset, array2_slabs;
260 uint32_t size;
261
262 /* Check input arguments */
263 if (!n_bits || !mem || (((uintptr_t) mem) & RTE_CACHE_LINE_MASK))
264 return NULL;
265
266 size = __rte_bitmap_get_memory_footprint(n_bits,
267 &array1_byte_offset, &array1_slabs,
268 &array2_byte_offset, &array2_slabs);
269 if (size < mem_size)
270 return NULL;
271
272 /* Setup bitmap */
273 bmp = (struct rte_bitmap *) mem;
274 bmp->array1 = (uint64_t *) &mem[array1_byte_offset];
275 bmp->array1_size = array1_slabs;
276 bmp->array2 = (uint64_t *) &mem[array2_byte_offset];
277 bmp->array2_size = array2_slabs;
278
279 __rte_bitmap_scan_init(bmp);
280
281 memset(bmp->array1, 0xff, bmp->array1_size * sizeof(bmp->array1[0]));
282 memset(bmp->array2, 0xff, bmp->array2_size * sizeof(bmp->array2[0]));
283 /* Clear overhead bits. */
284 __rte_bitmap_clear_slab_overhead_bits(bmp->array1, bmp->array1_size,
285 bmp->array2_size >> RTE_BITMAP_CL_SLAB_SIZE_LOG2);
286 __rte_bitmap_clear_slab_overhead_bits(bmp->array2, bmp->array2_size,
287 n_bits);
288 return bmp;
289 }
290
291 /**
292 * Bitmap free
293 *
294 * @param bmp
295 * Handle to bitmap instance
296 * @return
297 * 0 upon success, error code otherwise
298 */
299 static inline int
rte_bitmap_free(struct rte_bitmap * bmp)300 rte_bitmap_free(struct rte_bitmap *bmp)
301 {
302 /* Check input arguments */
303 if (bmp == NULL) {
304 return -1;
305 }
306
307 return 0;
308 }
309
310 /**
311 * Bitmap reset
312 *
313 * @param bmp
314 * Handle to bitmap instance
315 */
316 static inline void
rte_bitmap_reset(struct rte_bitmap * bmp)317 rte_bitmap_reset(struct rte_bitmap *bmp)
318 {
319 memset(bmp->array1, 0, bmp->array1_size * sizeof(uint64_t));
320 memset(bmp->array2, 0, bmp->array2_size * sizeof(uint64_t));
321 __rte_bitmap_scan_init(bmp);
322 }
323
324 /**
325 * Bitmap location prefetch into CPU L1 cache
326 *
327 * @param bmp
328 * Handle to bitmap instance
329 * @param pos
330 * Bit position
331 * @return
332 * 0 upon success, error code otherwise
333 */
334 static inline void
rte_bitmap_prefetch0(struct rte_bitmap * bmp,uint32_t pos)335 rte_bitmap_prefetch0(struct rte_bitmap *bmp, uint32_t pos)
336 {
337 uint64_t *slab2;
338 uint32_t index2;
339
340 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
341 slab2 = bmp->array2 + index2;
342 rte_prefetch0((void *) slab2);
343 }
344
345 /**
346 * Bitmap bit get
347 *
348 * @param bmp
349 * Handle to bitmap instance
350 * @param pos
351 * Bit position
352 * @return
353 * 0 when bit is cleared, non-zero when bit is set
354 */
355 static inline uint64_t
rte_bitmap_get(struct rte_bitmap * bmp,uint32_t pos)356 rte_bitmap_get(struct rte_bitmap *bmp, uint32_t pos)
357 {
358 uint64_t *slab2;
359 uint32_t index2, offset2;
360
361 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
362 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK;
363 slab2 = bmp->array2 + index2;
364 return (*slab2) & (1llu << offset2);
365 }
366
367 /**
368 * Bitmap bit set
369 *
370 * @param bmp
371 * Handle to bitmap instance
372 * @param pos
373 * Bit position
374 */
375 static inline void
rte_bitmap_set(struct rte_bitmap * bmp,uint32_t pos)376 rte_bitmap_set(struct rte_bitmap *bmp, uint32_t pos)
377 {
378 uint64_t *slab1, *slab2;
379 uint32_t index1, index2, offset1, offset2;
380
381 /* Set bit in array2 slab and set bit in array1 slab */
382 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
383 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK;
384 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2);
385 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK;
386 slab2 = bmp->array2 + index2;
387 slab1 = bmp->array1 + index1;
388
389 *slab2 |= 1llu << offset2;
390 *slab1 |= 1llu << offset1;
391 }
392
393 /**
394 * Bitmap slab set
395 *
396 * @param bmp
397 * Handle to bitmap instance
398 * @param pos
399 * Bit position identifying the array2 slab
400 * @param slab
401 * Value to be assigned to the 64-bit slab in array2
402 */
403 static inline void
rte_bitmap_set_slab(struct rte_bitmap * bmp,uint32_t pos,uint64_t slab)404 rte_bitmap_set_slab(struct rte_bitmap *bmp, uint32_t pos, uint64_t slab)
405 {
406 uint64_t *slab1, *slab2;
407 uint32_t index1, index2, offset1;
408
409 /* Set bits in array2 slab and set bit in array1 slab */
410 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
411 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2);
412 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK;
413 slab2 = bmp->array2 + index2;
414 slab1 = bmp->array1 + index1;
415
416 *slab2 |= slab;
417 *slab1 |= 1llu << offset1;
418 }
419
420 static inline uint64_t
__rte_bitmap_line_not_empty(uint64_t * slab2)421 __rte_bitmap_line_not_empty(uint64_t *slab2)
422 {
423 uint64_t v1, v2, v3, v4;
424
425 v1 = slab2[0] | slab2[1];
426 v2 = slab2[2] | slab2[3];
427 v3 = slab2[4] | slab2[5];
428 v4 = slab2[6] | slab2[7];
429 v1 |= v2;
430 v3 |= v4;
431
432 return v1 | v3;
433 }
434
435 /**
436 * Bitmap bit clear
437 *
438 * @param bmp
439 * Handle to bitmap instance
440 * @param pos
441 * Bit position
442 */
443 static inline void
rte_bitmap_clear(struct rte_bitmap * bmp,uint32_t pos)444 rte_bitmap_clear(struct rte_bitmap *bmp, uint32_t pos)
445 {
446 uint64_t *slab1, *slab2;
447 uint32_t index1, index2, offset1, offset2;
448
449 /* Clear bit in array2 slab */
450 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
451 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK;
452 slab2 = bmp->array2 + index2;
453
454 /* Return if array2 slab is not all-zeros */
455 *slab2 &= ~(1llu << offset2);
456 if (*slab2){
457 return;
458 }
459
460 /* Check the entire cache line of array2 for all-zeros */
461 index2 &= ~ RTE_BITMAP_CL_SLAB_MASK;
462 slab2 = bmp->array2 + index2;
463 if (__rte_bitmap_line_not_empty(slab2)) {
464 return;
465 }
466
467 /* The array2 cache line is all-zeros, so clear bit in array1 slab */
468 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2);
469 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK;
470 slab1 = bmp->array1 + index1;
471 *slab1 &= ~(1llu << offset1);
472
473 return;
474 }
475
476 static inline int
__rte_bitmap_scan_search(struct rte_bitmap * bmp)477 __rte_bitmap_scan_search(struct rte_bitmap *bmp)
478 {
479 uint64_t value1;
480 uint32_t i;
481
482 /* Check current array1 slab */
483 value1 = bmp->array1[bmp->index1];
484 value1 &= __rte_bitmap_mask1_get(bmp);
485
486 if (rte_bsf64_safe(value1, &bmp->offset1))
487 return 1;
488
489 __rte_bitmap_index1_inc(bmp);
490 bmp->offset1 = 0;
491
492 /* Look for another array1 slab */
493 for (i = 0; i < bmp->array1_size; i ++, __rte_bitmap_index1_inc(bmp)) {
494 value1 = bmp->array1[bmp->index1];
495
496 if (rte_bsf64_safe(value1, &bmp->offset1))
497 return 1;
498 }
499
500 return 0;
501 }
502
503 static inline void
__rte_bitmap_scan_read_init(struct rte_bitmap * bmp)504 __rte_bitmap_scan_read_init(struct rte_bitmap *bmp)
505 {
506 __rte_bitmap_index2_set(bmp);
507 bmp->go2 = 1;
508 rte_prefetch1((void *)(bmp->array2 + bmp->index2 + 8));
509 }
510
511 static inline int
__rte_bitmap_scan_read(struct rte_bitmap * bmp,uint32_t * pos,uint64_t * slab)512 __rte_bitmap_scan_read(struct rte_bitmap *bmp, uint32_t *pos, uint64_t *slab)
513 {
514 uint64_t *slab2;
515
516 slab2 = bmp->array2 + bmp->index2;
517 for ( ; bmp->go2 ; bmp->index2 ++, slab2 ++, bmp->go2 = bmp->index2 & RTE_BITMAP_CL_SLAB_MASK) {
518 if (*slab2) {
519 *pos = bmp->index2 << RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
520 *slab = *slab2;
521
522 bmp->index2 ++;
523 slab2 ++;
524 bmp->go2 = bmp->index2 & RTE_BITMAP_CL_SLAB_MASK;
525 return 1;
526 }
527 }
528
529 return 0;
530 }
531
532 /**
533 * Bitmap scan (with automatic wrap-around)
534 *
535 * @param bmp
536 * Handle to bitmap instance
537 * @param pos
538 * When function call returns 1, pos contains the position of the next set
539 * bit, otherwise not modified
540 * @param slab
541 * When function call returns 1, slab contains the value of the entire 64-bit
542 * slab where the bit indicated by pos is located. Slabs are always 64-bit
543 * aligned, so the position of the first bit of the slab (this bit is not
544 * necessarily set) is pos / 64. Once a slab has been returned by the bitmap
545 * scan operation, the internal pointers of the bitmap are updated to point
546 * after this slab, so the same slab will not be returned again if it
547 * contains more than one bit which is set. When function call returns 0,
548 * slab is not modified.
549 * @return
550 * 0 if there is no bit set in the bitmap, 1 otherwise
551 */
552 static inline int
rte_bitmap_scan(struct rte_bitmap * bmp,uint32_t * pos,uint64_t * slab)553 rte_bitmap_scan(struct rte_bitmap *bmp, uint32_t *pos, uint64_t *slab)
554 {
555 /* Return data from current array2 line if available */
556 if (__rte_bitmap_scan_read(bmp, pos, slab)) {
557 return 1;
558 }
559
560 /* Look for non-empty array2 line */
561 if (__rte_bitmap_scan_search(bmp)) {
562 __rte_bitmap_scan_read_init(bmp);
563 __rte_bitmap_scan_read(bmp, pos, slab);
564 return 1;
565 }
566
567 /* Empty bitmap */
568 return 0;
569 }
570
571 #ifdef __cplusplus
572 }
573 #endif
574
575 #endif /* __INCLUDE_RTE_BITMAP_H__ */
576