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_compat.h> 40 #include <rte_common.h> 41 #include <rte_config.h> 42 #include <rte_debug.h> 43 #include <rte_memory.h> 44 #include <rte_branch_prediction.h> 45 #include <rte_prefetch.h> 46 47 /* Slab */ 48 #define RTE_BITMAP_SLAB_BIT_SIZE 64 49 #define RTE_BITMAP_SLAB_BIT_SIZE_LOG2 6 50 #define RTE_BITMAP_SLAB_BIT_MASK (RTE_BITMAP_SLAB_BIT_SIZE - 1) 51 52 /* Cache line (CL) */ 53 #define RTE_BITMAP_CL_BIT_SIZE (RTE_CACHE_LINE_SIZE * 8) 54 #define RTE_BITMAP_CL_BIT_SIZE_LOG2 (RTE_CACHE_LINE_SIZE_LOG2 + 3) 55 #define RTE_BITMAP_CL_BIT_MASK (RTE_BITMAP_CL_BIT_SIZE - 1) 56 57 #define RTE_BITMAP_CL_SLAB_SIZE (RTE_BITMAP_CL_BIT_SIZE / RTE_BITMAP_SLAB_BIT_SIZE) 58 #define RTE_BITMAP_CL_SLAB_SIZE_LOG2 (RTE_BITMAP_CL_BIT_SIZE_LOG2 - RTE_BITMAP_SLAB_BIT_SIZE_LOG2) 59 #define RTE_BITMAP_CL_SLAB_MASK (RTE_BITMAP_CL_SLAB_SIZE - 1) 60 61 /** Bitmap data structure */ 62 struct rte_bitmap { 63 /* Context for array1 and array2 */ 64 uint64_t *array1; /**< Bitmap array1 */ 65 uint64_t *array2; /**< Bitmap array2 */ 66 uint32_t array1_size; /**< Number of 64-bit slabs in array1 that are actually used */ 67 uint32_t array2_size; /**< Number of 64-bit slabs in array2 */ 68 69 /* Context for the "scan next" operation */ 70 uint32_t index1; /**< Bitmap scan: Index of current array1 slab */ 71 uint32_t offset1; /**< Bitmap scan: Offset of current bit within current array1 slab */ 72 uint32_t index2; /**< Bitmap scan: Index of current array2 slab */ 73 uint32_t go2; /**< Bitmap scan: Go/stop condition for current array2 cache line */ 74 75 /* Storage space for array1 and array2 */ 76 uint8_t memory[]; 77 }; 78 79 static inline void 80 __rte_bitmap_index1_inc(struct rte_bitmap *bmp) 81 { 82 bmp->index1 = (bmp->index1 + 1) & (bmp->array1_size - 1); 83 } 84 85 static inline uint64_t 86 __rte_bitmap_mask1_get(struct rte_bitmap *bmp) 87 { 88 return (~1llu) << bmp->offset1; 89 } 90 91 static inline void 92 __rte_bitmap_index2_set(struct rte_bitmap *bmp) 93 { 94 bmp->index2 = (((bmp->index1 << RTE_BITMAP_SLAB_BIT_SIZE_LOG2) + bmp->offset1) << RTE_BITMAP_CL_SLAB_SIZE_LOG2); 95 } 96 97 static inline uint32_t 98 __rte_bitmap_get_memory_footprint(uint32_t n_bits, 99 uint32_t *array1_byte_offset, uint32_t *array1_slabs, 100 uint32_t *array2_byte_offset, uint32_t *array2_slabs) 101 { 102 uint32_t n_slabs_context, n_slabs_array1, n_cache_lines_context_and_array1; 103 uint32_t n_cache_lines_array2; 104 uint32_t n_bytes_total; 105 106 n_cache_lines_array2 = (n_bits + RTE_BITMAP_CL_BIT_SIZE - 1) / RTE_BITMAP_CL_BIT_SIZE; 107 n_slabs_array1 = (n_cache_lines_array2 + RTE_BITMAP_SLAB_BIT_SIZE - 1) / RTE_BITMAP_SLAB_BIT_SIZE; 108 n_slabs_array1 = rte_align32pow2(n_slabs_array1); 109 n_slabs_context = (sizeof(struct rte_bitmap) + (RTE_BITMAP_SLAB_BIT_SIZE / 8) - 1) / (RTE_BITMAP_SLAB_BIT_SIZE / 8); 110 n_cache_lines_context_and_array1 = (n_slabs_context + n_slabs_array1 + RTE_BITMAP_CL_SLAB_SIZE - 1) / RTE_BITMAP_CL_SLAB_SIZE; 111 n_bytes_total = (n_cache_lines_context_and_array1 + n_cache_lines_array2) * RTE_CACHE_LINE_SIZE; 112 113 if (array1_byte_offset) { 114 *array1_byte_offset = n_slabs_context * (RTE_BITMAP_SLAB_BIT_SIZE / 8); 115 } 116 if (array1_slabs) { 117 *array1_slabs = n_slabs_array1; 118 } 119 if (array2_byte_offset) { 120 *array2_byte_offset = n_cache_lines_context_and_array1 * RTE_CACHE_LINE_SIZE; 121 } 122 if (array2_slabs) { 123 *array2_slabs = n_cache_lines_array2 * RTE_BITMAP_CL_SLAB_SIZE; 124 } 125 126 return n_bytes_total; 127 } 128 129 static inline void 130 __rte_bitmap_scan_init(struct rte_bitmap *bmp) 131 { 132 bmp->index1 = bmp->array1_size - 1; 133 bmp->offset1 = RTE_BITMAP_SLAB_BIT_SIZE - 1; 134 __rte_bitmap_index2_set(bmp); 135 bmp->index2 += RTE_BITMAP_CL_SLAB_SIZE; 136 137 bmp->go2 = 0; 138 } 139 140 /** 141 * Bitmap memory footprint calculation 142 * 143 * @param n_bits 144 * Number of bits in the bitmap 145 * @return 146 * Bitmap memory footprint measured in bytes on success, 0 on error 147 */ 148 static inline uint32_t 149 rte_bitmap_get_memory_footprint(uint32_t n_bits) { 150 /* Check input arguments */ 151 if (n_bits == 0) { 152 return 0; 153 } 154 155 return __rte_bitmap_get_memory_footprint(n_bits, NULL, NULL, NULL, NULL); 156 } 157 158 /** 159 * Bitmap initialization 160 * 161 * @param n_bits 162 * Number of pre-allocated bits in array2. 163 * @param mem 164 * Base address of array1 and array2. 165 * @param mem_size 166 * Minimum expected size of bitmap. 167 * @return 168 * Handle to bitmap instance. 169 */ 170 static inline struct rte_bitmap * 171 rte_bitmap_init(uint32_t n_bits, uint8_t *mem, uint32_t mem_size) 172 { 173 struct rte_bitmap *bmp; 174 uint32_t array1_byte_offset, array1_slabs, array2_byte_offset, array2_slabs; 175 uint32_t size; 176 177 /* Check input arguments */ 178 if (n_bits == 0) { 179 return NULL; 180 } 181 182 if ((mem == NULL) || (((uintptr_t) mem) & RTE_CACHE_LINE_MASK)) { 183 return NULL; 184 } 185 186 size = __rte_bitmap_get_memory_footprint(n_bits, 187 &array1_byte_offset, &array1_slabs, 188 &array2_byte_offset, &array2_slabs); 189 if (size > mem_size) 190 return NULL; 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 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 * 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 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 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 */ 332 static inline void 333 rte_bitmap_prefetch0(struct rte_bitmap *bmp, uint32_t pos) 334 { 335 uint64_t *slab2; 336 uint32_t index2; 337 338 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 339 slab2 = bmp->array2 + index2; 340 rte_prefetch0((void *) slab2); 341 } 342 343 /** 344 * Bitmap bit get 345 * 346 * @param bmp 347 * Handle to bitmap instance 348 * @param pos 349 * Bit position 350 * @return 351 * 0 when bit is cleared, non-zero when bit is set 352 */ 353 static inline uint64_t 354 rte_bitmap_get(struct rte_bitmap *bmp, uint32_t pos) 355 { 356 uint64_t *slab2; 357 uint32_t index2, offset2; 358 359 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 360 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK; 361 slab2 = bmp->array2 + index2; 362 return (*slab2) & (1llu << offset2); 363 } 364 365 /** 366 * Bitmap bit set 367 * 368 * @param bmp 369 * Handle to bitmap instance 370 * @param pos 371 * Bit position 372 */ 373 static inline void 374 rte_bitmap_set(struct rte_bitmap *bmp, uint32_t pos) 375 { 376 uint64_t *slab1, *slab2; 377 uint32_t index1, index2, offset1, offset2; 378 379 /* Set bit in array2 slab and set bit in array1 slab */ 380 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 381 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK; 382 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2); 383 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK; 384 slab2 = bmp->array2 + index2; 385 slab1 = bmp->array1 + index1; 386 387 *slab2 |= 1llu << offset2; 388 *slab1 |= 1llu << offset1; 389 } 390 391 /** 392 * Bitmap slab set 393 * 394 * @param bmp 395 * Handle to bitmap instance 396 * @param pos 397 * Bit position identifying the array2 slab 398 * @param slab 399 * Value to be assigned to the 64-bit slab in array2 400 */ 401 static inline void 402 rte_bitmap_set_slab(struct rte_bitmap *bmp, uint32_t pos, uint64_t slab) 403 { 404 uint64_t *slab1, *slab2; 405 uint32_t index1, index2, offset1; 406 407 /* Set bits in array2 slab and set bit in array1 slab */ 408 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 409 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2); 410 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK; 411 slab2 = bmp->array2 + index2; 412 slab1 = bmp->array1 + index1; 413 414 *slab2 |= slab; 415 *slab1 |= 1llu << offset1; 416 } 417 418 #if RTE_BITMAP_CL_SLAB_SIZE == 8 419 static inline uint64_t 420 __rte_bitmap_line_not_empty(uint64_t *slab2) 421 { 422 uint64_t v1, v2, v3, v4; 423 424 v1 = slab2[0] | slab2[1]; 425 v2 = slab2[2] | slab2[3]; 426 v3 = slab2[4] | slab2[5]; 427 v4 = slab2[6] | slab2[7]; 428 v1 |= v2; 429 v3 |= v4; 430 431 return v1 | v3; 432 } 433 434 #elif RTE_BITMAP_CL_SLAB_SIZE == 16 435 static inline uint64_t 436 __rte_bitmap_line_not_empty(uint64_t *slab2) 437 { 438 uint64_t v1, v2, v3, v4, v5, v6, v7, v8; 439 440 v1 = slab2[0] | slab2[1]; 441 v2 = slab2[2] | slab2[3]; 442 v3 = slab2[4] | slab2[5]; 443 v4 = slab2[6] | slab2[7]; 444 v5 = slab2[8] | slab2[9]; 445 v6 = slab2[10] | slab2[11]; 446 v7 = slab2[12] | slab2[13]; 447 v8 = slab2[14] | slab2[15]; 448 v1 |= v2; 449 v3 |= v4; 450 v5 |= v6; 451 v7 |= v8; 452 453 return v1 | v3 | v5 | v7; 454 } 455 456 #endif /* RTE_BITMAP_CL_SLAB_SIZE */ 457 458 /** 459 * Bitmap bit clear 460 * 461 * @param bmp 462 * Handle to bitmap instance 463 * @param pos 464 * Bit position 465 */ 466 static inline void 467 rte_bitmap_clear(struct rte_bitmap *bmp, uint32_t pos) 468 { 469 uint64_t *slab1, *slab2; 470 uint32_t index1, index2, offset1, offset2; 471 472 /* Clear bit in array2 slab */ 473 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 474 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK; 475 slab2 = bmp->array2 + index2; 476 477 /* Return if array2 slab is not all-zeros */ 478 *slab2 &= ~(1llu << offset2); 479 if (*slab2){ 480 return; 481 } 482 483 /* Check the entire cache line of array2 for all-zeros */ 484 index2 &= ~ RTE_BITMAP_CL_SLAB_MASK; 485 slab2 = bmp->array2 + index2; 486 if (__rte_bitmap_line_not_empty(slab2)) { 487 return; 488 } 489 490 /* The array2 cache line is all-zeros, so clear bit in array1 slab */ 491 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2); 492 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK; 493 slab1 = bmp->array1 + index1; 494 *slab1 &= ~(1llu << offset1); 495 496 return; 497 } 498 499 static inline int 500 __rte_bitmap_scan_search(struct rte_bitmap *bmp) 501 { 502 uint64_t value1; 503 uint32_t i; 504 505 /* Check current array1 slab */ 506 value1 = bmp->array1[bmp->index1]; 507 value1 &= __rte_bitmap_mask1_get(bmp); 508 509 if (rte_bsf64_safe(value1, &bmp->offset1)) 510 return 1; 511 512 __rte_bitmap_index1_inc(bmp); 513 bmp->offset1 = 0; 514 515 /* Look for another array1 slab */ 516 for (i = 0; i < bmp->array1_size; i ++, __rte_bitmap_index1_inc(bmp)) { 517 value1 = bmp->array1[bmp->index1]; 518 519 if (rte_bsf64_safe(value1, &bmp->offset1)) 520 return 1; 521 } 522 523 return 0; 524 } 525 526 static inline void 527 __rte_bitmap_scan_read_init(struct rte_bitmap *bmp) 528 { 529 __rte_bitmap_index2_set(bmp); 530 bmp->go2 = 1; 531 rte_prefetch1((void *)(bmp->array2 + bmp->index2 + 8)); 532 } 533 534 static inline int 535 __rte_bitmap_scan_read(struct rte_bitmap *bmp, uint32_t *pos, uint64_t *slab) 536 { 537 uint64_t *slab2; 538 539 slab2 = bmp->array2 + bmp->index2; 540 for ( ; bmp->go2 ; bmp->index2 ++, slab2 ++, bmp->go2 = bmp->index2 & RTE_BITMAP_CL_SLAB_MASK) { 541 if (*slab2) { 542 *pos = bmp->index2 << RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 543 *slab = *slab2; 544 545 bmp->index2 ++; 546 slab2 ++; 547 bmp->go2 = bmp->index2 & RTE_BITMAP_CL_SLAB_MASK; 548 return 1; 549 } 550 } 551 552 return 0; 553 } 554 555 /** 556 * Bitmap scan (with automatic wrap-around) 557 * 558 * @param bmp 559 * Handle to bitmap instance 560 * @param pos 561 * When function call returns 1, pos contains the position of the next set 562 * bit, otherwise not modified 563 * @param slab 564 * When function call returns 1, slab contains the value of the entire 64-bit 565 * slab where the bit indicated by pos is located. Slabs are always 64-bit 566 * aligned, so the position of the first bit of the slab (this bit is not 567 * necessarily set) is pos / 64. Once a slab has been returned by the bitmap 568 * scan operation, the internal pointers of the bitmap are updated to point 569 * after this slab, so the same slab will not be returned again if it 570 * contains more than one bit which is set. When function call returns 0, 571 * slab is not modified. 572 * @return 573 * 0 if there is no bit set in the bitmap, 1 otherwise 574 */ 575 static inline int 576 rte_bitmap_scan(struct rte_bitmap *bmp, uint32_t *pos, uint64_t *slab) 577 { 578 /* Return data from current array2 line if available */ 579 if (__rte_bitmap_scan_read(bmp, pos, slab)) { 580 return 1; 581 } 582 583 /* Look for non-empty array2 line */ 584 if (__rte_bitmap_scan_search(bmp)) { 585 __rte_bitmap_scan_read_init(bmp); 586 __rte_bitmap_scan_read(bmp, pos, slab); 587 return 1; 588 } 589 590 /* Empty bitmap */ 591 return 0; 592 } 593 594 #ifdef __cplusplus 595 } 596 #endif 597 598 #endif /* __INCLUDE_RTE_BITMAP_H__ */ 599