lsquic_di_hash.c revision 2d296031
1/* Copyright (c) 2017 - 2019 LiteSpeed Technologies Inc. See LICENSE. */ 2/* 3 * lsquic_di_hash.c -- Copy incoming data into a hash 4 * 5 * While this implementation copies the data, its memory use is limited, 6 * which makes it a good choice when we have a lot of stream frames 7 * coming in. 8 * 9 * Another difference is that it does not check for frame overlap, which 10 * is something that is present in Chrome, but it is not required by QUIC. 11 */ 12 13 14#include <assert.h> 15#include <inttypes.h> 16#include <stddef.h> 17#include <stdint.h> 18#include <stdlib.h> 19#include <string.h> 20#include <sys/queue.h> 21 22#include "lsquic.h" 23#include "lsquic_int_types.h" 24#include "lsquic_types.h" 25#include "lsquic_conn_flow.h" 26#include "lsquic_packet_common.h" 27#include "lsquic_packet_in.h" 28#include "lsquic_rtt.h" 29#include "lsquic_sfcw.h" 30#include "lsquic_stream.h" 31#include "lsquic_mm.h" 32#include "lsquic_malo.h" 33#include "lsquic_conn.h" 34#include "lsquic_conn_public.h" 35#include "lsquic_data_in_if.h" 36 37 38#define LSQUIC_LOGGER_MODULE LSQLM_DI 39#define LSQUIC_LOG_CONN_ID hdi->hdi_conn_pub->lconn->cn_cid 40#define LSQUIC_LOG_STREAM_ID hdi->hdi_stream_id 41#include "lsquic_logger.h" 42 43 44#define N_DB_SETS 57 45 46#define DB_DATA_SIZE (0x1000 - sizeof(TAILQ_ENTRY(data_block)) - \ 47 sizeof(uint64_t) - N_DB_SETS * sizeof(uint64_t)) 48 49struct data_block 50{ 51 TAILQ_ENTRY(data_block) db_next; 52 uint64_t db_off; 53 uint64_t db_set[N_DB_SETS]; /* bit for each valid byte */ 54 unsigned char db_data[DB_DATA_SIZE]; 55}; 56 57typedef char db_set_covers_all_db_data[(N_DB_SETS * 64 >= DB_DATA_SIZE) ?1: - 1]; 58typedef char db_set_no_waste[(N_DB_SETS * 64 - 64 <= DB_DATA_SIZE)?1: - 1]; 59typedef char db_block_is_4K[(sizeof(struct data_block) == 0x1000) ?1:- 1]; 60 61 62TAILQ_HEAD(dblock_head, data_block); 63 64 65static const struct data_in_iface *di_if_hash_ptr; 66 67 68struct hash_data_in 69{ 70 struct data_in hdi_data_in; 71 struct lsquic_conn_public *hdi_conn_pub; 72 uint64_t hdi_fin_off; 73 struct dblock_head *hdi_buckets; 74 struct data_block *hdi_last_block; 75 struct data_frame hdi_data_frame; 76 uint32_t hdi_stream_id; 77 unsigned hdi_count; 78 unsigned hdi_nbits; 79 enum { 80 HDI_FIN = (1 << 0), 81 } hdi_flags; 82}; 83 84 85#define HDI_PTR(data_in) (struct hash_data_in *) \ 86 ((unsigned char *) (data_in) - offsetof(struct hash_data_in, hdi_data_in)) 87 88 89#define N_BUCKETS(n_bits) (1U << (n_bits)) 90#define BUCKNO(n_bits, off) ((off / DB_DATA_SIZE) & (N_BUCKETS(n_bits) - 1)) 91 92 93static unsigned 94my_log2 /* silly name to suppress compiler warning */ (unsigned sz) 95{ 96#if __GNUC__ 97 unsigned clz = __builtin_clz(sz); 98 return 32 - clz; 99#else 100 unsigned clz; 101 size_t y; 102 clz = 32; 103 y = sz >> 16; if (y) { clz -= 16; sz = y; } 104 y = sz >> 8; if (y) { clz -= 8; sz = y; } 105 y = sz >> 4; if (y) { clz -= 4; sz = y; } 106 y = sz >> 2; if (y) { clz -= 2; sz = y; } 107 y = sz >> 1; if (y) return 32 - clz + 1; 108 return 32 - clz + sz; 109#endif 110} 111 112 113struct data_in * 114data_in_hash_new (struct lsquic_conn_public *conn_pub, uint32_t stream_id, 115 uint64_t byteage) 116{ 117 struct hash_data_in *hdi; 118 unsigned n; 119 120 hdi = malloc(sizeof(*hdi)); 121 if (!hdi) 122 return NULL; 123 124 hdi->hdi_data_in.di_if = di_if_hash_ptr; 125 hdi->hdi_data_in.di_flags = 0; 126 hdi->hdi_conn_pub = conn_pub; 127 hdi->hdi_stream_id = stream_id; 128 hdi->hdi_fin_off = 0; 129 hdi->hdi_flags = 0; 130 hdi->hdi_last_block = NULL; 131 if (byteage >= DB_DATA_SIZE /* __builtin_clz is undefined if 132 argument is 0 */) 133 hdi->hdi_nbits = my_log2(byteage / DB_DATA_SIZE) + 2; 134 else 135 hdi->hdi_nbits = 3; 136 hdi->hdi_count = 0; 137 hdi->hdi_buckets = malloc(sizeof(hdi->hdi_buckets[0]) * 138 N_BUCKETS(hdi->hdi_nbits)); 139 if (!hdi->hdi_buckets) 140 { 141 free(hdi); 142 return NULL; 143 } 144 145 for (n = 0; n < N_BUCKETS(hdi->hdi_nbits); ++n) 146 TAILQ_INIT(&hdi->hdi_buckets[n]); 147 148 return &hdi->hdi_data_in; 149} 150 151 152static void 153hash_di_destroy (struct data_in *data_in) 154{ 155 struct hash_data_in *const hdi = HDI_PTR(data_in); 156 struct data_block *block; 157 unsigned n; 158 159 for (n = 0; n < N_BUCKETS(hdi->hdi_nbits); ++n) 160 { 161 while ((block = TAILQ_FIRST(&hdi->hdi_buckets[n]))) 162 { 163 TAILQ_REMOVE(&hdi->hdi_buckets[n], block, db_next); 164 free(block); 165 } 166 } 167 free(hdi->hdi_buckets); 168 free(hdi); 169} 170 171 172static int 173hash_grow (struct hash_data_in *hdi) 174{ 175 struct dblock_head *new_buckets, *new[2]; 176 struct data_block *block; 177 unsigned n, old_nbits; 178 int idx; 179 180 old_nbits = hdi->hdi_nbits; 181 LSQ_DEBUG("doubling number of buckets to %u", N_BUCKETS(old_nbits + 1)); 182 new_buckets = malloc(sizeof(hdi->hdi_buckets[0]) 183 * N_BUCKETS(old_nbits + 1)); 184 if (!new_buckets) 185 { 186 LSQ_WARN("malloc failed: potential trouble ahead"); 187 return -1; 188 } 189 190 for (n = 0; n < N_BUCKETS(old_nbits); ++n) 191 { 192 new[0] = &new_buckets[n]; 193 new[1] = &new_buckets[n + N_BUCKETS(old_nbits)]; 194 TAILQ_INIT(new[0]); 195 TAILQ_INIT(new[1]); 196 while ((block = TAILQ_FIRST(&hdi->hdi_buckets[n]))) 197 { 198 TAILQ_REMOVE(&hdi->hdi_buckets[n], block, db_next); 199 idx = (BUCKNO(old_nbits + 1, block->db_off) >> old_nbits) & 1; 200 TAILQ_INSERT_TAIL(new[idx], block, db_next); 201 } 202 } 203 free(hdi->hdi_buckets); 204 hdi->hdi_nbits = old_nbits + 1; 205 hdi->hdi_buckets = new_buckets; 206 return 0; 207} 208 209 210static int 211hash_insert (struct hash_data_in *hdi, struct data_block *block) 212{ 213 unsigned buckno; 214 215 if (hdi->hdi_count >= N_BUCKETS(hdi->hdi_nbits) / 2 && 0 != hash_grow(hdi)) 216 return -1; 217 218 buckno = BUCKNO(hdi->hdi_nbits, block->db_off); 219 TAILQ_INSERT_TAIL(&hdi->hdi_buckets[buckno], block, db_next); 220 ++hdi->hdi_count; 221 return 0; 222} 223 224 225static struct data_block * 226hash_find (const struct hash_data_in *hdi, uint64_t off) 227{ 228 struct data_block *block; 229 unsigned buckno; 230 231 buckno = BUCKNO(hdi->hdi_nbits, off); 232 TAILQ_FOREACH(block, &hdi->hdi_buckets[buckno], db_next) 233 if (off == block->db_off) 234 return block; 235 return NULL; 236} 237 238 239static void 240hash_remove (struct hash_data_in *hdi, struct data_block *block) 241{ 242 unsigned buckno; 243 244 buckno = BUCKNO(hdi->hdi_nbits, block->db_off); 245 TAILQ_REMOVE(&hdi->hdi_buckets[buckno], block, db_next); 246 --hdi->hdi_count; 247} 248 249 250static struct data_block * 251new_block (struct hash_data_in *hdi, uint64_t off) 252{ 253 struct data_block *block; 254 255 assert(0 == off % DB_DATA_SIZE); 256 257 block = malloc(sizeof(*block)); 258 if (!block) 259 return NULL; 260 261 block->db_off = off; 262 if (0 != hash_insert(hdi, block)) 263 { 264 free(block); 265 return NULL; 266 } 267 268 memset(block->db_set, 0, sizeof(block->db_set)); 269 return block; 270} 271 272 273static unsigned 274block_write (struct data_block *block, unsigned block_off, 275 const unsigned char *data, unsigned data_sz) 276{ 277 const unsigned char *begin, *end; 278 unsigned set, bit, n_full_sets, n; 279 uint64_t mask; 280 281 if (data_sz > DB_DATA_SIZE - block_off) 282 data_sz = DB_DATA_SIZE - block_off; 283 284 begin = data; 285 end = begin + data_sz; 286 set = block_off >> 6; 287 bit = block_off & 0x3F; 288 289 assert(set < N_DB_SETS); 290 291 if (bit) 292 { 293 n = 64 - bit; 294 if (n > data_sz) 295 n = data_sz; 296 mask = ~((1ULL << bit ) - 1) 297 & ((1ULL << (bit + n - 1)) | ((1ULL << (bit + n - 1)) - 1)); 298 block->db_set[ set ] |= mask; 299 memcpy(block->db_data + block_off, data, n); 300 data += n; 301 block_off += n; 302 ++set; 303 } 304 305 n_full_sets = (end - data) >> 6; 306 if (n_full_sets) 307 { 308 memcpy(block->db_data + block_off, data, n_full_sets * 64); 309 data += n_full_sets * 64; 310 block_off += n_full_sets * 64; 311 memset(&block->db_set[ set ], 0xFF, n_full_sets * 8); 312 set += n_full_sets; 313 } 314 315 if (data < end) 316 { 317 assert(end - data < 64); 318 block->db_set[ set ] |= ((1ULL << (end - data)) - 1); 319 memcpy(block->db_data + block_off, data, end - data); 320 data = end; 321 } 322 323 assert(set <= N_DB_SETS); 324 325 return data - begin; 326} 327 328 329static int 330has_bytes_after (const struct data_block *block, unsigned off) 331{ 332 unsigned bit, set; 333 int has; 334 335 set = off >> 6; 336 bit = off & 0x3F; 337 338 has = 0 != (block->db_set[ set ] >> bit); 339 ++set; 340 341 for ( ; set < N_DB_SETS; ++set) 342 has += 0 != block->db_set[ set ]; 343 344 return has > 0; 345} 346 347 348enum ins_frame 349data_in_hash_insert_data_frame (struct data_in *data_in, 350 const struct data_frame *data_frame, uint64_t read_offset) 351{ 352 struct hash_data_in *const hdi = HDI_PTR(data_in); 353 struct data_block *block; 354 uint64_t key, off, diff, fin_off; 355 const unsigned char *data; 356 unsigned size, nw; 357 358 if (data_frame->df_offset + data_frame->df_size < read_offset) 359 { 360 if (data_frame->df_fin) 361 return INS_FRAME_ERR; 362 else 363 return INS_FRAME_DUP; 364 } 365 366 if ((hdi->hdi_flags & HDI_FIN) && 367 ( 368 (data_frame->df_fin && 369 data_frame->df_offset + data_frame->df_size != hdi->hdi_fin_off) 370 || 371 data_frame->df_offset + data_frame->df_size > hdi->hdi_fin_off 372 ) 373 ) 374 { 375 return INS_FRAME_ERR; 376 } 377 378 if (data_frame->df_offset < read_offset) 379 { 380 diff = read_offset - data_frame->df_offset; 381 assert(diff <= data_frame->df_size); 382 size = data_frame->df_size - diff; 383 off = data_frame->df_offset + diff; 384 data = data_frame->df_data + diff; 385 } 386 else 387 { 388 size = data_frame->df_size; 389 off = data_frame->df_offset; 390 data = data_frame->df_data; 391 } 392 393 key = off - (off % DB_DATA_SIZE); 394 do 395 { 396 block = hash_find(hdi, key); 397 if (!block) 398 { 399 block = new_block(hdi, key); 400 if (!block) 401 return INS_FRAME_ERR; 402 } 403 nw = block_write(block, off % DB_DATA_SIZE, data, size); 404 size -= nw; 405 off += nw; 406 data += nw; 407 key += DB_DATA_SIZE; 408 } 409 while (size > 0); 410 411 if (data_frame->df_fin) 412 { 413 fin_off = data_frame->df_offset + data_frame->df_size; 414 if (has_bytes_after(block, fin_off - block->db_off) || 415 hash_find(hdi, key)) 416 { 417 return INS_FRAME_ERR; 418 } 419 hdi->hdi_flags |= HDI_FIN; 420 hdi->hdi_fin_off = fin_off; 421 } 422 423 return INS_FRAME_OK; 424} 425 426 427static enum ins_frame 428hash_di_insert_frame (struct data_in *data_in, 429 struct stream_frame *new_frame, uint64_t read_offset) 430{ 431 struct hash_data_in *const hdi = HDI_PTR(data_in); 432 const struct data_frame *const data_frame = &new_frame->data_frame; 433 enum ins_frame ins; 434 435 ins = data_in_hash_insert_data_frame(data_in, data_frame, read_offset); 436 assert(ins != INS_FRAME_OVERLAP); 437 lsquic_packet_in_put(hdi->hdi_conn_pub->mm, new_frame->packet_in); 438 lsquic_malo_put(new_frame); 439 return ins; 440} 441 442 443#if __GNUC__ 444# define ctz __builtin_ctzll 445#else 446static unsigned 447ctz (unsigned long long x) 448{ 449 unsigned n = 0; 450 if (0 == (x & ((1ULL << 32) - 1))) { n += 32; x >>= 32; } 451 if (0 == (x & ((1ULL << 16) - 1))) { n += 16; x >>= 16; } 452 if (0 == (x & ((1ULL << 8) - 1))) { n += 8; x >>= 8; } 453 if (0 == (x & ((1ULL << 4) - 1))) { n += 4; x >>= 4; } 454 if (0 == (x & ((1ULL << 2) - 1))) { n += 2; x >>= 2; } 455 if (0 == (x & ((1ULL << 1) - 1))) { n += 1; x >>= 1; } 456 return n; 457} 458 459 460#endif 461 462 463static unsigned 464n_avail_bytes (const struct data_block *block, unsigned set, unsigned bit) 465{ 466 unsigned count; 467 uint64_t part; 468 469 part = ~(block->db_set[ set ] >> bit); 470 if (part) 471 { 472 count = ctz(part); 473 if (count < 64 - bit) 474 return count; 475 } 476 else 477 count = 64; 478 ++set; 479 480 for ( ; set < N_DB_SETS && ~0ULL == block->db_set[ set ]; ++set) 481 count += 64; 482 483 if (set < N_DB_SETS) 484 { 485 part = ~block->db_set[ set ]; 486 if (part) 487 count += ctz(part); 488 else 489 count += 64; 490 } 491 492 return count; 493} 494 495 496/* Data block is readable if there is at least one readable byte at 497 * `read_offset' or there is FIN at that offset. 498 */ 499static int 500setup_data_frame (struct hash_data_in *hdi, const uint64_t read_offset, 501 struct data_block *block) 502{ 503 unsigned set, bit; 504 uint64_t offset; 505 506 offset = read_offset % DB_DATA_SIZE; 507 set = offset >> 6; 508 bit = offset & 0x3F; 509 510 if (block->db_set[ set ] & (1ULL << bit)) 511 { 512 hdi->hdi_last_block = block; 513 hdi->hdi_data_frame.df_data = block->db_data; 514 hdi->hdi_data_frame.df_offset = block->db_off; 515 hdi->hdi_data_frame.df_read_off = offset; 516 hdi->hdi_data_frame.df_size = offset + 517 n_avail_bytes(block, set, bit); 518 hdi->hdi_data_frame.df_fin = 519 (hdi->hdi_flags & HDI_FIN) && 520 hdi->hdi_data_frame.df_read_off + 521 hdi->hdi_data_frame.df_size == hdi->hdi_fin_off; 522 return 1; 523 } 524 else if ((hdi->hdi_flags & HDI_FIN) && read_offset == hdi->hdi_fin_off) 525 { 526 hdi->hdi_last_block = block; 527 hdi->hdi_data_frame.df_data = NULL; 528 hdi->hdi_data_frame.df_offset = block->db_off; 529 hdi->hdi_data_frame.df_read_off = offset; 530 hdi->hdi_data_frame.df_size = offset; 531 hdi->hdi_data_frame.df_fin = 1; 532 return 1; 533 } 534 else 535 return 0; 536} 537 538 539static struct data_frame * 540hash_di_get_frame (struct data_in *data_in, uint64_t read_offset) 541{ 542 struct hash_data_in *const hdi = HDI_PTR(data_in); 543 struct data_block *block; 544 uint64_t key; 545 546 key = read_offset - (read_offset % DB_DATA_SIZE); 547 block = hash_find(hdi, key); 548 if (!block) 549 { 550 if ((hdi->hdi_flags & HDI_FIN) && read_offset == hdi->hdi_fin_off) 551 { 552 hdi->hdi_last_block = NULL; 553 hdi->hdi_data_frame.df_data = NULL; 554 hdi->hdi_data_frame.df_offset = read_offset - 555 read_offset % DB_DATA_SIZE; 556 hdi->hdi_data_frame.df_read_off = 0; 557 hdi->hdi_data_frame.df_size = 0; 558 hdi->hdi_data_frame.df_fin = 1; 559 return &hdi->hdi_data_frame; 560 } 561 else 562 return NULL; 563 } 564 565 if (setup_data_frame(hdi, read_offset, block)) 566 return &hdi->hdi_data_frame; 567 else 568 return NULL; 569} 570 571 572static void 573hash_di_frame_done (struct data_in *data_in, struct data_frame *data_frame) 574{ 575 struct hash_data_in *const hdi = HDI_PTR(data_in); 576 struct data_block *const block = hdi->hdi_last_block; 577 578 if (block) 579 { 580 if (data_frame->df_read_off == DB_DATA_SIZE || 581 !has_bytes_after(block, data_frame->df_read_off)) 582 { 583 hash_remove(hdi, block); 584 free(block); 585 if (0 == hdi->hdi_count && 0 == (hdi->hdi_flags & HDI_FIN)) 586 { 587 LSQ_DEBUG("hash empty, want to switch"); 588 hdi->hdi_data_in.di_flags |= DI_SWITCH_IMPL; 589 } 590 } 591 } 592 else 593 assert(data_frame->df_fin && data_frame->df_size == 0); 594} 595 596 597static int 598hash_di_empty (struct data_in *data_in) 599{ 600 struct hash_data_in *const hdi = HDI_PTR(data_in); 601 return hdi->hdi_count == 0; 602} 603 604 605struct data_in * 606hash_di_switch_impl (struct data_in *data_in, uint64_t read_offset) 607{ 608 struct hash_data_in *const hdi = HDI_PTR(data_in); 609 struct data_in *new_data_in; 610 611 assert(hdi->hdi_count == 0); 612 613 new_data_in = data_in_nocopy_new(hdi->hdi_conn_pub, hdi->hdi_stream_id); 614 data_in->di_if->di_destroy(data_in); 615 616 return new_data_in; 617} 618 619 620static size_t 621hash_di_mem_used (struct data_in *data_in) 622{ 623 struct hash_data_in *const hdi = HDI_PTR(data_in); 624 const struct data_block *block; 625 size_t size; 626 unsigned n; 627 628 size = sizeof(*data_in); 629 630 for (n = 0; n < N_BUCKETS(hdi->hdi_nbits); ++n) 631 TAILQ_FOREACH(block, &hdi->hdi_buckets[n], db_next) 632 size += sizeof(*block); 633 634 size += N_BUCKETS(hdi->hdi_nbits) * sizeof(hdi->hdi_buckets[0]); 635 636 return size; 637} 638 639 640static const struct data_in_iface di_if_hash = { 641 .di_destroy = hash_di_destroy, 642 .di_empty = hash_di_empty, 643 .di_frame_done = hash_di_frame_done, 644 .di_get_frame = hash_di_get_frame, 645 .di_insert_frame = hash_di_insert_frame, 646 .di_mem_used = hash_di_mem_used, 647 .di_switch_impl = hash_di_switch_impl, 648}; 649 650static const struct data_in_iface *di_if_hash_ptr = &di_if_hash; 651