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 #include <string.h> 38 #include <rte_compat.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 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 85 __rte_bitmap_mask1_get(struct rte_bitmap *bmp) 86 { 87 return (~1llu) << bmp->offset1; 88 } 89 90 static inline void 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 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 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 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 * 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 /* Setup bitmap */ 192 memset(mem, 0, size); 193 bmp = (struct rte_bitmap *) mem; 194 195 bmp->array1 = (uint64_t *) &mem[array1_byte_offset]; 196 bmp->array1_size = array1_slabs; 197 bmp->array2 = (uint64_t *) &mem[array2_byte_offset]; 198 bmp->array2_size = array2_slabs; 199 200 __rte_bitmap_scan_init(bmp); 201 202 return bmp; 203 } 204 205 /** 206 * @warning 207 * @b EXPERIMENTAL: this API may change without prior notice. 208 * 209 * Bitmap clear slab overhead bits. 210 * 211 * @param slabs 212 * Slab array. 213 * @param slab_size 214 * Number of 64-bit slabs in the slabs array. 215 * @param pos 216 * The start bit position in the slabs to be cleared. 217 */ 218 __rte_experimental 219 static inline void 220 __rte_bitmap_clear_slab_overhead_bits(uint64_t *slabs, uint32_t slab_size, 221 uint32_t pos) 222 { 223 uint32_t i; 224 uint32_t index = pos / RTE_BITMAP_SLAB_BIT_SIZE; 225 uint32_t offset = pos & RTE_BITMAP_SLAB_BIT_MASK; 226 227 if (offset) { 228 for (i = offset; i < RTE_BITMAP_SLAB_BIT_SIZE; i++) 229 slabs[index] &= ~(1llu << i); 230 index++; 231 } 232 if (index < slab_size) 233 memset(&slabs[index], 0, sizeof(slabs[0]) * 234 (slab_size - index)); 235 } 236 237 /** 238 * @warning 239 * @b EXPERIMENTAL: this API may change without prior notice. 240 * 241 * Bitmap initialization with all bits set 242 * 243 * @param n_bits 244 * Number of pre-allocated bits in array2. 245 * @param mem 246 * Base address of array1 and array2. 247 * @param mem_size 248 * Minimum expected size of bitmap. 249 * @return 250 * Handle to bitmap instance. 251 */ 252 __rte_experimental 253 static inline struct rte_bitmap * 254 rte_bitmap_init_with_all_set(uint32_t n_bits, uint8_t *mem, uint32_t mem_size) 255 { 256 struct rte_bitmap *bmp; 257 uint32_t array1_byte_offset, array1_slabs; 258 uint32_t array2_byte_offset, array2_slabs; 259 uint32_t size; 260 261 /* Check input arguments */ 262 if (!n_bits || !mem || (((uintptr_t) mem) & RTE_CACHE_LINE_MASK)) 263 return NULL; 264 265 size = __rte_bitmap_get_memory_footprint(n_bits, 266 &array1_byte_offset, &array1_slabs, 267 &array2_byte_offset, &array2_slabs); 268 if (size < mem_size) 269 return NULL; 270 271 /* Setup bitmap */ 272 bmp = (struct rte_bitmap *) mem; 273 bmp->array1 = (uint64_t *) &mem[array1_byte_offset]; 274 bmp->array1_size = array1_slabs; 275 bmp->array2 = (uint64_t *) &mem[array2_byte_offset]; 276 bmp->array2_size = array2_slabs; 277 278 __rte_bitmap_scan_init(bmp); 279 280 memset(bmp->array1, 0xff, bmp->array1_size * sizeof(bmp->array1[0])); 281 memset(bmp->array2, 0xff, bmp->array2_size * sizeof(bmp->array2[0])); 282 /* Clear overhead bits. */ 283 __rte_bitmap_clear_slab_overhead_bits(bmp->array1, bmp->array1_size, 284 bmp->array2_size >> RTE_BITMAP_CL_SLAB_SIZE_LOG2); 285 __rte_bitmap_clear_slab_overhead_bits(bmp->array2, bmp->array2_size, 286 n_bits); 287 return bmp; 288 } 289 290 /** 291 * Bitmap free 292 * 293 * @param bmp 294 * Handle to bitmap instance 295 * @return 296 * 0 upon success, error code otherwise 297 */ 298 static inline int 299 rte_bitmap_free(struct rte_bitmap *bmp) 300 { 301 /* Check input arguments */ 302 if (bmp == NULL) { 303 return -1; 304 } 305 306 return 0; 307 } 308 309 /** 310 * Bitmap reset 311 * 312 * @param bmp 313 * Handle to bitmap instance 314 */ 315 static inline void 316 rte_bitmap_reset(struct rte_bitmap *bmp) 317 { 318 memset(bmp->array1, 0, bmp->array1_size * sizeof(uint64_t)); 319 memset(bmp->array2, 0, bmp->array2_size * sizeof(uint64_t)); 320 __rte_bitmap_scan_init(bmp); 321 } 322 323 /** 324 * Bitmap location prefetch into CPU L1 cache 325 * 326 * @param bmp 327 * Handle to bitmap instance 328 * @param pos 329 * Bit position 330 */ 331 static inline void 332 rte_bitmap_prefetch0(struct rte_bitmap *bmp, uint32_t pos) 333 { 334 uint64_t *slab2; 335 uint32_t index2; 336 337 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 338 slab2 = bmp->array2 + index2; 339 rte_prefetch0((void *) slab2); 340 } 341 342 /** 343 * Bitmap bit get 344 * 345 * @param bmp 346 * Handle to bitmap instance 347 * @param pos 348 * Bit position 349 * @return 350 * 0 when bit is cleared, non-zero when bit is set 351 */ 352 static inline uint64_t 353 rte_bitmap_get(struct rte_bitmap *bmp, uint32_t pos) 354 { 355 uint64_t *slab2; 356 uint32_t index2, offset2; 357 358 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 359 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK; 360 slab2 = bmp->array2 + index2; 361 return (*slab2) & (1llu << offset2); 362 } 363 364 /** 365 * Bitmap bit set 366 * 367 * @param bmp 368 * Handle to bitmap instance 369 * @param pos 370 * Bit position 371 */ 372 static inline void 373 rte_bitmap_set(struct rte_bitmap *bmp, uint32_t pos) 374 { 375 uint64_t *slab1, *slab2; 376 uint32_t index1, index2, offset1, offset2; 377 378 /* Set bit in array2 slab and set bit in array1 slab */ 379 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 380 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK; 381 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2); 382 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK; 383 slab2 = bmp->array2 + index2; 384 slab1 = bmp->array1 + index1; 385 386 *slab2 |= 1llu << offset2; 387 *slab1 |= 1llu << offset1; 388 } 389 390 /** 391 * Bitmap slab set 392 * 393 * @param bmp 394 * Handle to bitmap instance 395 * @param pos 396 * Bit position identifying the array2 slab 397 * @param slab 398 * Value to be assigned to the 64-bit slab in array2 399 */ 400 static inline void 401 rte_bitmap_set_slab(struct rte_bitmap *bmp, uint32_t pos, uint64_t slab) 402 { 403 uint64_t *slab1, *slab2; 404 uint32_t index1, index2, offset1; 405 406 /* Set bits in array2 slab and set bit in array1 slab */ 407 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 408 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2); 409 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK; 410 slab2 = bmp->array2 + index2; 411 slab1 = bmp->array1 + index1; 412 413 *slab2 |= slab; 414 *slab1 |= 1llu << offset1; 415 } 416 417 #if RTE_BITMAP_CL_SLAB_SIZE == 8 418 static inline uint64_t 419 __rte_bitmap_line_not_empty(uint64_t *slab2) 420 { 421 uint64_t v1, v2, v3, v4; 422 423 v1 = slab2[0] | slab2[1]; 424 v2 = slab2[2] | slab2[3]; 425 v3 = slab2[4] | slab2[5]; 426 v4 = slab2[6] | slab2[7]; 427 v1 |= v2; 428 v3 |= v4; 429 430 return v1 | v3; 431 } 432 433 #elif RTE_BITMAP_CL_SLAB_SIZE == 16 434 static inline uint64_t 435 __rte_bitmap_line_not_empty(uint64_t *slab2) 436 { 437 uint64_t v1, v2, v3, v4, v5, v6, v7, v8; 438 439 v1 = slab2[0] | slab2[1]; 440 v2 = slab2[2] | slab2[3]; 441 v3 = slab2[4] | slab2[5]; 442 v4 = slab2[6] | slab2[7]; 443 v5 = slab2[8] | slab2[9]; 444 v6 = slab2[10] | slab2[11]; 445 v7 = slab2[12] | slab2[13]; 446 v8 = slab2[14] | slab2[15]; 447 v1 |= v2; 448 v3 |= v4; 449 v5 |= v6; 450 v7 |= v8; 451 452 return v1 | v3 | v5 | v7; 453 } 454 455 #endif /* RTE_BITMAP_CL_SLAB_SIZE */ 456 457 /** 458 * Bitmap bit clear 459 * 460 * @param bmp 461 * Handle to bitmap instance 462 * @param pos 463 * Bit position 464 */ 465 static inline void 466 rte_bitmap_clear(struct rte_bitmap *bmp, uint32_t pos) 467 { 468 uint64_t *slab1, *slab2; 469 uint32_t index1, index2, offset1, offset2; 470 471 /* Clear bit in array2 slab */ 472 index2 = pos >> RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 473 offset2 = pos & RTE_BITMAP_SLAB_BIT_MASK; 474 slab2 = bmp->array2 + index2; 475 476 /* Return if array2 slab is not all-zeros */ 477 *slab2 &= ~(1llu << offset2); 478 if (*slab2){ 479 return; 480 } 481 482 /* Check the entire cache line of array2 for all-zeros */ 483 index2 &= ~ RTE_BITMAP_CL_SLAB_MASK; 484 slab2 = bmp->array2 + index2; 485 if (__rte_bitmap_line_not_empty(slab2)) { 486 return; 487 } 488 489 /* The array2 cache line is all-zeros, so clear bit in array1 slab */ 490 index1 = pos >> (RTE_BITMAP_SLAB_BIT_SIZE_LOG2 + RTE_BITMAP_CL_BIT_SIZE_LOG2); 491 offset1 = (pos >> RTE_BITMAP_CL_BIT_SIZE_LOG2) & RTE_BITMAP_SLAB_BIT_MASK; 492 slab1 = bmp->array1 + index1; 493 *slab1 &= ~(1llu << offset1); 494 495 return; 496 } 497 498 static inline int 499 __rte_bitmap_scan_search(struct rte_bitmap *bmp) 500 { 501 uint64_t value1; 502 uint32_t i; 503 504 /* Check current array1 slab */ 505 value1 = bmp->array1[bmp->index1]; 506 value1 &= __rte_bitmap_mask1_get(bmp); 507 508 if (rte_bsf64_safe(value1, &bmp->offset1)) 509 return 1; 510 511 __rte_bitmap_index1_inc(bmp); 512 bmp->offset1 = 0; 513 514 /* Look for another array1 slab */ 515 for (i = 0; i < bmp->array1_size; i ++, __rte_bitmap_index1_inc(bmp)) { 516 value1 = bmp->array1[bmp->index1]; 517 518 if (rte_bsf64_safe(value1, &bmp->offset1)) 519 return 1; 520 } 521 522 return 0; 523 } 524 525 static inline void 526 __rte_bitmap_scan_read_init(struct rte_bitmap *bmp) 527 { 528 __rte_bitmap_index2_set(bmp); 529 bmp->go2 = 1; 530 rte_prefetch1((void *)(bmp->array2 + bmp->index2 + 8)); 531 } 532 533 static inline int 534 __rte_bitmap_scan_read(struct rte_bitmap *bmp, uint32_t *pos, uint64_t *slab) 535 { 536 uint64_t *slab2; 537 538 slab2 = bmp->array2 + bmp->index2; 539 for ( ; bmp->go2 ; bmp->index2 ++, slab2 ++, bmp->go2 = bmp->index2 & RTE_BITMAP_CL_SLAB_MASK) { 540 if (*slab2) { 541 *pos = bmp->index2 << RTE_BITMAP_SLAB_BIT_SIZE_LOG2; 542 *slab = *slab2; 543 544 bmp->index2 ++; 545 slab2 ++; 546 bmp->go2 = bmp->index2 & RTE_BITMAP_CL_SLAB_MASK; 547 return 1; 548 } 549 } 550 551 return 0; 552 } 553 554 /** 555 * Bitmap scan (with automatic wrap-around) 556 * 557 * @param bmp 558 * Handle to bitmap instance 559 * @param pos 560 * When function call returns 1, pos contains the position of the next set 561 * bit, otherwise not modified 562 * @param slab 563 * When function call returns 1, slab contains the value of the entire 64-bit 564 * slab where the bit indicated by pos is located. Slabs are always 64-bit 565 * aligned, so the position of the first bit of the slab (this bit is not 566 * necessarily set) is pos / 64. Once a slab has been returned by the bitmap 567 * scan operation, the internal pointers of the bitmap are updated to point 568 * after this slab, so the same slab will not be returned again if it 569 * contains more than one bit which is set. When function call returns 0, 570 * slab is not modified. 571 * @return 572 * 0 if there is no bit set in the bitmap, 1 otherwise 573 */ 574 static inline int 575 rte_bitmap_scan(struct rte_bitmap *bmp, uint32_t *pos, uint64_t *slab) 576 { 577 /* Return data from current array2 line if available */ 578 if (__rte_bitmap_scan_read(bmp, pos, slab)) { 579 return 1; 580 } 581 582 /* Look for non-empty array2 line */ 583 if (__rte_bitmap_scan_search(bmp)) { 584 __rte_bitmap_scan_read_init(bmp); 585 __rte_bitmap_scan_read(bmp, pos, slab); 586 return 1; 587 } 588 589 /* Empty bitmap */ 590 return 0; 591 } 592 593 #ifdef __cplusplus 594 } 595 #endif 596 597 #endif /* __INCLUDE_RTE_BITMAP_H__ */ 598