1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
5 #ifndef __INCLUDE_RTE_BITMAP_H__
6 #define __INCLUDE_RTE_BITMAP_H__
16 * The bitmap component provides a mechanism to manage large arrays of bits
17 * through bit get/set/clear and bit array scan operations.
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.
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
39 #include <rte_common.h>
40 #include <rte_debug.h>
41 #include <rte_memory.h>
42 #include <rte_branch_prediction.h>
43 #include <rte_prefetch.h>
45 #ifndef RTE_BITMAP_OPTIMIZATIONS
46 #define RTE_BITMAP_OPTIMIZATIONS 1
50 #define RTE_BITMAP_SLAB_BIT_SIZE 64
51 #define RTE_BITMAP_SLAB_BIT_SIZE_LOG2 6
52 #define RTE_BITMAP_SLAB_BIT_MASK (RTE_BITMAP_SLAB_BIT_SIZE - 1)
55 #define RTE_BITMAP_CL_BIT_SIZE (RTE_CACHE_LINE_SIZE * 8)
56 #define RTE_BITMAP_CL_BIT_SIZE_LOG2 (RTE_CACHE_LINE_SIZE_LOG2 + 3)
57 #define RTE_BITMAP_CL_BIT_MASK (RTE_BITMAP_CL_BIT_SIZE - 1)
59 #define RTE_BITMAP_CL_SLAB_SIZE (RTE_BITMAP_CL_BIT_SIZE / RTE_BITMAP_SLAB_BIT_SIZE)
60 #define RTE_BITMAP_CL_SLAB_SIZE_LOG2 (RTE_BITMAP_CL_BIT_SIZE_LOG2 - RTE_BITMAP_SLAB_BIT_SIZE_LOG2)
61 #define RTE_BITMAP_CL_SLAB_MASK (RTE_BITMAP_CL_SLAB_SIZE - 1)
63 /** Bitmap data structure */
65 /* Context for array1 and array2 */
66 uint64_t *array1; /**< Bitmap array1 */
67 uint64_t *array2; /**< Bitmap array2 */
68 uint32_t array1_size; /**< Number of 64-bit slabs in array1 that are actually used */
69 uint32_t array2_size; /**< Number of 64-bit slabs in array2 */
71 /* Context for the "scan next" operation */
72 uint32_t index1; /**< Bitmap scan: Index of current array1 slab */
73 uint32_t offset1; /**< Bitmap scan: Offset of current bit within current array1 slab */
74 uint32_t index2; /**< Bitmap scan: Index of current array2 slab */
75 uint32_t go2; /**< Bitmap scan: Go/stop condition for current array2 cache line */
77 /* Storage space for array1 and array2 */
82 __rte_bitmap_index1_inc(struct rte_bitmap *bmp)
84 bmp->index1 = (bmp->index1 + 1) & (bmp->array1_size - 1);
87 static inline uint64_t
88 __rte_bitmap_mask1_get(struct rte_bitmap *bmp)
90 return (~1lu) << bmp->offset1;
94 __rte_bitmap_index2_set(struct rte_bitmap *bmp)
96 bmp->index2 = (((bmp->index1 << RTE_BITMAP_SLAB_BIT_SIZE_LOG2) + bmp->offset1) << RTE_BITMAP_CL_SLAB_SIZE_LOG2);
99 #if RTE_BITMAP_OPTIMIZATIONS
102 rte_bsf64(uint64_t slab, uint32_t *pos)
104 if (likely(slab == 0)) {
108 *pos = __builtin_ctzll(slab);
115 rte_bsf64(uint64_t slab, uint32_t *pos)
120 if (likely(slab == 0)) {
124 for (i = 0, mask = 1; i < RTE_BITMAP_SLAB_BIT_SIZE; i ++, mask <<= 1) {
125 if (unlikely(slab & mask)) {
136 static inline uint32_t
137 __rte_bitmap_get_memory_footprint(uint32_t n_bits,
138 uint32_t *array1_byte_offset, uint32_t *array1_slabs,
139 uint32_t *array2_byte_offset, uint32_t *array2_slabs)
141 uint32_t n_slabs_context, n_slabs_array1, n_cache_lines_context_and_array1;
142 uint32_t n_cache_lines_array2;
143 uint32_t n_bytes_total;
145 n_cache_lines_array2 = (n_bits + RTE_BITMAP_CL_BIT_SIZE - 1) / RTE_BITMAP_CL_BIT_SIZE;
146 n_slabs_array1 = (n_cache_lines_array2 + RTE_BITMAP_SLAB_BIT_SIZE - 1) / RTE_BITMAP_SLAB_BIT_SIZE;
147 n_slabs_array1 = rte_align32pow2(n_slabs_array1);
148 n_slabs_context = (sizeof(struct rte_bitmap) + (RTE_BITMAP_SLAB_BIT_SIZE / 8) - 1) / (RTE_BITMAP_SLAB_BIT_SIZE / 8);
149 n_cache_lines_context_and_array1 = (n_slabs_context + n_slabs_array1 + RTE_BITMAP_CL_SLAB_SIZE - 1) / RTE_BITMAP_CL_SLAB_SIZE;
150 n_bytes_total = (n_cache_lines_context_and_array1 + n_cache_lines_array2) * RTE_CACHE_LINE_SIZE;
152 if (array1_byte_offset) {
153 *array1_byte_offset = n_slabs_context * (RTE_BITMAP_SLAB_BIT_SIZE / 8);
156 *array1_slabs = n_slabs_array1;
158 if (array2_byte_offset) {
159 *array2_byte_offset = n_cache_lines_context_and_array1 * RTE_CACHE_LINE_SIZE;
162 *array2_slabs = n_cache_lines_array2 * RTE_BITMAP_CL_SLAB_SIZE;
165 return n_bytes_total;
169 __rte_bitmap_scan_init(struct rte_bitmap *bmp)
171 bmp->index1 = bmp->array1_size - 1;
172 bmp->offset1 = RTE_BITMAP_SLAB_BIT_SIZE - 1;
173 __rte_bitmap_index2_set(bmp);
174 bmp->index2 += RTE_BITMAP_CL_SLAB_SIZE;
180 * Bitmap memory footprint calculation
183 * Number of bits in the bitmap
185 * Bitmap memory footprint measured in bytes on success, 0 on error
187 static inline uint32_t
188 rte_bitmap_get_memory_footprint(uint32_t n_bits) {
189 /* Check input arguments */
194 return __rte_bitmap_get_memory_footprint(n_bits, NULL, NULL, NULL, NULL);
198 * Bitmap initialization
201 * Minimum expected size of bitmap.
203 * Base address of array1 and array2.
205 * Number of pre-allocated bits in array2. Must be non-zero and multiple of 512.
207 * Handle to bitmap instance.
209 static inline struct rte_bitmap *
210 rte_bitmap_init(uint32_t n_bits, uint8_t *mem, uint32_t mem_size)
212 struct rte_bitmap *bmp;
213 uint32_t array1_byte_offset, array1_slabs, array2_byte_offset, array2_slabs;
216 /* Check input arguments */
221 if ((mem == NULL) || (((uintptr_t) mem) & RTE_CACHE_LINE_MASK)) {
225 size = __rte_bitmap_get_memory_footprint(n_bits,
226 &array1_byte_offset, &array1_slabs,
227 &array2_byte_offset, &array2_slabs);
228 if (size < mem_size) {
233 memset(mem, 0, size);
234 bmp = (struct rte_bitmap *) mem;
236 bmp->array1 = (uint64_t *) &mem[array1_byte_offset];
237 bmp->array1_size = array1_slabs;
238 bmp->array2 = (uint64_t *) &mem[array2_byte_offset];
239 bmp->array2_size = array2_slabs;
241 __rte_bitmap_scan_init(bmp);
250 * Handle to bitmap instance
252 * 0 upon success, error code otherwise
255 rte_bitmap_free(struct rte_bitmap *bmp)
257 /* Check input arguments */
269 * Handle to bitmap instance
272 rte_bitmap_reset(struct rte_bitmap *bmp)
274 memset(bmp->array1, 0, bmp->array1_size * sizeof(uint64_t));
275 memset(bmp->array2, 0, bmp->array2_size * sizeof(uint64_t));
276 __rte_bitmap_scan_init(bmp);
280 * Bitmap location prefetch into CPU L1 cache
283 * Handle to bitmap instance
287 * 0 upon success, error code otherwise
290 rte_bitmap_prefetch0(struct rte_bitmap *bmp, uint32_t pos)
295 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
296 slab2 = bmp->array2 + index2;
297 rte_prefetch0((void *) slab2);
304 * Handle to bitmap instance
308 * 0 when bit is cleared, non-zero when bit is set
310 static inline uint64_t
311 rte_bitmap_get(struct rte_bitmap *bmp, uint32_t pos)
314 uint32_t index2, offset2;
316 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
317 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK;
318 slab2 = bmp->array2 + index2;
319 return (*slab2) & (1lu << offset2);
326 * Handle to bitmap instance
331 rte_bitmap_set(struct rte_bitmap *bmp, uint32_t pos)
333 uint64_t *slab1, *slab2;
334 uint32_t index1, index2, offset1, offset2;
336 /* Set bit in array2 slab and set bit in array1 slab */
337 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
338 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK;
339 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2);
340 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK;
341 slab2 = bmp->array2 + index2;
342 slab1 = bmp->array1 + index1;
344 *slab2 |= 1lu << offset2;
345 *slab1 |= 1lu << offset1;
352 * Handle to bitmap instance
354 * Bit position identifying the array2 slab
356 * Value to be assigned to the 64-bit slab in array2
359 rte_bitmap_set_slab(struct rte_bitmap *bmp, uint32_t pos, uint64_t slab)
361 uint64_t *slab1, *slab2;
362 uint32_t index1, index2, offset1;
364 /* Set bits in array2 slab and set bit in array1 slab */
365 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
366 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2);
367 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK;
368 slab2 = bmp->array2 + index2;
369 slab1 = bmp->array1 + index1;
372 *slab1 |= 1lu << offset1;
375 static inline uint64_t
376 __rte_bitmap_line_not_empty(uint64_t *slab2)
378 uint64_t v1, v2, v3, v4;
380 v1 = slab2[0] | slab2[1];
381 v2 = slab2[2] | slab2[3];
382 v3 = slab2[4] | slab2[5];
383 v4 = slab2[6] | slab2[7];
394 * Handle to bitmap instance
399 rte_bitmap_clear(struct rte_bitmap *bmp, uint32_t pos)
401 uint64_t *slab1, *slab2;
402 uint32_t index1, index2, offset1, offset2;
404 /* Clear bit in array2 slab */
405 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
406 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK;
407 slab2 = bmp->array2 + index2;
409 /* Return if array2 slab is not all-zeros */
410 *slab2 &= ~(1lu << offset2);
415 /* Check the entire cache line of array2 for all-zeros */
416 index2 &= ~ RTE_BITMAP_CL_SLAB_MASK;
417 slab2 = bmp->array2 + index2;
418 if (__rte_bitmap_line_not_empty(slab2)) {
422 /* The array2 cache line is all-zeros, so clear bit in array1 slab */
423 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2);
424 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK;
425 slab1 = bmp->array1 + index1;
426 *slab1 &= ~(1lu << offset1);
432 __rte_bitmap_scan_search(struct rte_bitmap *bmp)
437 /* Check current array1 slab */
438 value1 = bmp->array1[bmp->index1];
439 value1 &= __rte_bitmap_mask1_get(bmp);
441 if (rte_bsf64(value1, &bmp->offset1)) {
445 __rte_bitmap_index1_inc(bmp);
448 /* Look for another array1 slab */
449 for (i = 0; i < bmp->array1_size; i ++, __rte_bitmap_index1_inc(bmp)) {
450 value1 = bmp->array1[bmp->index1];
452 if (rte_bsf64(value1, &bmp->offset1)) {
461 __rte_bitmap_scan_read_init(struct rte_bitmap *bmp)
463 __rte_bitmap_index2_set(bmp);
465 rte_prefetch1((void *)(bmp->array2 + bmp->index2 + 8));
469 __rte_bitmap_scan_read(struct rte_bitmap *bmp, uint32_t *pos, uint64_t *slab)
473 slab2 = bmp->array2 + bmp->index2;
474 for ( ; bmp->go2 ; bmp->index2 ++, slab2 ++, bmp->go2 = bmp->index2 & RTE_BITMAP_CL_SLAB_MASK) {
476 *pos = bmp->index2 << RTE_BITMAP_SLAB_BIT_SIZE_LOG2;
481 bmp->go2 = bmp->index2 & RTE_BITMAP_CL_SLAB_MASK;
490 * Bitmap scan (with automatic wrap-around)
493 * Handle to bitmap instance
495 * When function call returns 1, pos contains the position of the next set
496 * bit, otherwise not modified
498 * When function call returns 1, slab contains the value of the entire 64-bit
499 * slab where the bit indicated by pos is located. Slabs are always 64-bit
500 * aligned, so the position of the first bit of the slab (this bit is not
501 * necessarily set) is pos / 64. Once a slab has been returned by the bitmap
502 * scan operation, the internal pointers of the bitmap are updated to point
503 * after this slab, so the same slab will not be returned again if it
504 * contains more than one bit which is set. When function call returns 0,
505 * slab is not modified.
507 * 0 if there is no bit set in the bitmap, 1 otherwise
510 rte_bitmap_scan(struct rte_bitmap *bmp, uint32_t *pos, uint64_t *slab)
512 /* Return data from current array2 line if available */
513 if (__rte_bitmap_scan_read(bmp, pos, slab)) {
517 /* Look for non-empty array2 line */
518 if (__rte_bitmap_scan_search(bmp)) {
519 __rte_bitmap_scan_read_init(bmp);
520 __rte_bitmap_scan_read(bmp, pos, slab);
532 #endif /* __INCLUDE_RTE_BITMAP_H__ */