1 /* 2 * Copyright (c) 2013-2018 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@dragonflybsd.org> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include "hammer2.h" 36 37 #define HAMMER2_DOP_READ 1 38 #define HAMMER2_DOP_NEW 2 39 #define HAMMER2_DOP_NEWNZ 3 40 #define HAMMER2_DOP_READQ 4 41 42 /* 43 * Implements an abstraction layer for synchronous and asynchronous 44 * buffered device I/O. Can be used as an OS-abstraction but the main 45 * purpose is to allow larger buffers to be used against hammer2_chain's 46 * using smaller allocations, without causing deadlocks. 47 * 48 * The DIOs also record temporary state with limited persistence. This 49 * feature is used to keep track of dedupable blocks. 50 */ 51 static int hammer2_io_cleanup_callback(hammer2_io_t *dio, void *arg); 52 static void dio_write_stats_update(hammer2_io_t *dio, struct buf *bp); 53 54 static int 55 hammer2_io_cmp(hammer2_io_t *io1, hammer2_io_t *io2) 56 { 57 if (io1->pbase < io2->pbase) 58 return(-1); 59 if (io1->pbase > io2->pbase) 60 return(1); 61 return(0); 62 } 63 64 RB_PROTOTYPE2(hammer2_io_tree, hammer2_io, rbnode, hammer2_io_cmp, off_t); 65 RB_GENERATE2(hammer2_io_tree, hammer2_io, rbnode, hammer2_io_cmp, 66 off_t, pbase); 67 68 struct hammer2_cleanupcb_info { 69 struct hammer2_io_tree tmptree; 70 int count; 71 }; 72 73 #if 0 74 static __inline 75 uint64_t 76 hammer2_io_mask(hammer2_io_t *dio, hammer2_off_t off, u_int bytes) 77 { 78 uint64_t mask; 79 int i; 80 81 if (bytes < 1024) /* smaller chunks not supported */ 82 return 0; 83 84 /* 85 * Calculate crc check mask for larger chunks 86 */ 87 i = (((off & ~HAMMER2_OFF_MASK_RADIX) - dio->pbase) & 88 HAMMER2_PBUFMASK) >> 10; 89 if (i == 0 && bytes == HAMMER2_PBUFSIZE) 90 return((uint64_t)-1); 91 mask = ((uint64_t)1U << (bytes >> 10)) - 1; 92 mask <<= i; 93 94 return mask; 95 } 96 #endif 97 98 #ifdef HAMMER2_IO_DEBUG 99 100 static __inline void 101 DIO_RECORD(hammer2_io_t *dio HAMMER2_IO_DEBUG_ARGS) 102 { 103 int i; 104 105 i = atomic_fetchadd_int(&dio->debug_index, 1) & HAMMER2_IO_DEBUG_MASK; 106 107 dio->debug_file[i] = file; 108 dio->debug_line[i] = line; 109 dio->debug_refs[i] = dio->refs; 110 dio->debug_td[i] = curthread; 111 } 112 113 #else 114 115 #define DIO_RECORD(dio) 116 117 #endif 118 119 /* 120 * Returns the DIO corresponding to the data|radix, creating it if necessary. 121 * 122 * If createit is 0, NULL can be returned indicating that the DIO does not 123 * exist. (btype) is ignored when createit is 0. 124 */ 125 static __inline 126 hammer2_io_t * 127 hammer2_io_alloc(hammer2_dev_t *hmp, hammer2_key_t data_off, uint8_t btype, 128 int createit, int *isgoodp) 129 { 130 hammer2_io_t *dio; 131 hammer2_io_t *xio; 132 hammer2_key_t lbase; 133 hammer2_key_t pbase; 134 hammer2_key_t pmask; 135 hammer2_volume_t *vol; 136 uint64_t refs; 137 int lsize; 138 int psize; 139 140 psize = HAMMER2_PBUFSIZE; 141 pmask = ~(hammer2_off_t)(psize - 1); 142 if ((int)(data_off & HAMMER2_OFF_MASK_RADIX)) 143 lsize = 1 << (int)(data_off & HAMMER2_OFF_MASK_RADIX); 144 else 145 lsize = 0; 146 lbase = data_off & ~HAMMER2_OFF_MASK_RADIX; 147 pbase = lbase & pmask; 148 149 if (pbase == 0 || ((lbase + lsize - 1) & pmask) != pbase) { 150 kprintf("Illegal: %016jx %016jx+%08x / %016jx\n", 151 pbase, lbase, lsize, pmask); 152 } 153 KKASSERT(pbase != 0 && ((lbase + lsize - 1) & pmask) == pbase); 154 *isgoodp = 0; 155 156 /* 157 * Access/Allocate the DIO, bump dio->refs to prevent destruction. 158 * 159 * If DIO_GOOD is set the ref should prevent it from being cleared 160 * out from under us, we can set *isgoodp, and the caller can operate 161 * on the buffer without any further interaction. 162 */ 163 hammer2_spin_sh(&hmp->io_spin); 164 dio = RB_LOOKUP(hammer2_io_tree, &hmp->iotree, pbase); 165 if (dio) { 166 refs = atomic_fetchadd_64(&dio->refs, 1); 167 if ((refs & HAMMER2_DIO_MASK) == 0) { 168 atomic_add_int(&dio->hmp->iofree_count, -1); 169 } 170 if (refs & HAMMER2_DIO_GOOD) 171 *isgoodp = 1; 172 hammer2_spin_unsh(&hmp->io_spin); 173 } else if (createit) { 174 refs = 0; 175 hammer2_spin_unsh(&hmp->io_spin); 176 vol = hammer2_get_volume(hmp, pbase); 177 dio = kmalloc(sizeof(*dio), M_HAMMER2, M_INTWAIT | M_ZERO); 178 dio->hmp = hmp; 179 dio->devvp = vol->dev->devvp; 180 dio->dbase = vol->offset; 181 KKASSERT((dio->dbase & HAMMER2_FREEMAP_LEVEL1_MASK) == 0); 182 dio->pbase = pbase; 183 dio->psize = psize; 184 dio->btype = btype; 185 dio->refs = refs + 1; 186 dio->act = 5; 187 hammer2_spin_ex(&hmp->io_spin); 188 xio = RB_INSERT(hammer2_io_tree, &hmp->iotree, dio); 189 if (xio == NULL) { 190 atomic_add_int(&hammer2_dio_count, 1); 191 hammer2_spin_unex(&hmp->io_spin); 192 } else { 193 refs = atomic_fetchadd_64(&xio->refs, 1); 194 if ((refs & HAMMER2_DIO_MASK) == 0) 195 atomic_add_int(&xio->hmp->iofree_count, -1); 196 if (refs & HAMMER2_DIO_GOOD) 197 *isgoodp = 1; 198 hammer2_spin_unex(&hmp->io_spin); 199 kfree(dio, M_HAMMER2); 200 dio = xio; 201 } 202 } else { 203 hammer2_spin_unsh(&hmp->io_spin); 204 return NULL; 205 } 206 dio->ticks = ticks; 207 if (dio->act < 10) 208 ++dio->act; 209 210 return dio; 211 } 212 213 /* 214 * Acquire the requested dio. If DIO_GOOD is not set we must instantiate 215 * a buffer. If set the buffer already exists and is good to go. 216 */ 217 hammer2_io_t * 218 _hammer2_io_getblk(hammer2_dev_t *hmp, int btype, off_t lbase, 219 int lsize, int op HAMMER2_IO_DEBUG_ARGS) 220 { 221 hammer2_io_t *dio; 222 hammer2_off_t dev_pbase; 223 off_t peof; 224 uint64_t orefs; 225 uint64_t nrefs; 226 int isgood; 227 int error; 228 int hce; 229 int bflags; 230 231 bflags = ((btype == HAMMER2_BREF_TYPE_DATA) ? B_NOTMETA : 0); 232 bflags |= B_KVABIO; 233 234 KKASSERT((1 << (int)(lbase & HAMMER2_OFF_MASK_RADIX)) == lsize); 235 236 if (op == HAMMER2_DOP_READQ) { 237 dio = hammer2_io_alloc(hmp, lbase, btype, 0, &isgood); 238 if (dio == NULL) 239 return NULL; 240 op = HAMMER2_DOP_READ; 241 } else { 242 dio = hammer2_io_alloc(hmp, lbase, btype, 1, &isgood); 243 } 244 245 for (;;) { 246 orefs = dio->refs; 247 cpu_ccfence(); 248 249 /* 250 * Buffer is already good, handle the op and return. 251 */ 252 if (orefs & HAMMER2_DIO_GOOD) { 253 if (isgood == 0) 254 cpu_mfence(); 255 bkvasync(dio->bp); 256 257 switch(op) { 258 case HAMMER2_DOP_NEW: 259 bzero(hammer2_io_data(dio, lbase), lsize); 260 /* fall through */ 261 case HAMMER2_DOP_NEWNZ: 262 atomic_set_long(&dio->refs, HAMMER2_DIO_DIRTY); 263 break; 264 case HAMMER2_DOP_READ: 265 default: 266 /* nothing to do */ 267 break; 268 } 269 DIO_RECORD(dio HAMMER2_IO_DEBUG_CALL); 270 return (dio); 271 } 272 273 /* 274 * Try to own the DIO 275 */ 276 if (orefs & HAMMER2_DIO_INPROG) { 277 nrefs = orefs | HAMMER2_DIO_WAITING; 278 tsleep_interlock(dio, 0); 279 if (atomic_cmpset_64(&dio->refs, orefs, nrefs)) { 280 tsleep(dio, PINTERLOCKED, "h2dio", hz); 281 } 282 /* retry */ 283 } else { 284 nrefs = orefs | HAMMER2_DIO_INPROG; 285 if (atomic_cmpset_64(&dio->refs, orefs, nrefs)) { 286 break; 287 } 288 } 289 } 290 291 /* 292 * We break to here if GOOD is not set and we acquired INPROG for 293 * the I/O. 294 */ 295 KKASSERT(dio->bp == NULL); 296 if (btype == HAMMER2_BREF_TYPE_DATA) 297 hce = hammer2_cluster_data_read; 298 else 299 hce = hammer2_cluster_meta_read; 300 301 error = 0; 302 dev_pbase = dio->pbase - dio->dbase; 303 if (dio->pbase == (lbase & ~HAMMER2_OFF_MASK_RADIX) && 304 dio->psize == lsize) { 305 switch(op) { 306 case HAMMER2_DOP_NEW: 307 case HAMMER2_DOP_NEWNZ: 308 dio->bp = getblk(dio->devvp, 309 dev_pbase, dio->psize, 310 GETBLK_KVABIO, 0); 311 if (op == HAMMER2_DOP_NEW) { 312 bkvasync(dio->bp); 313 bzero(dio->bp->b_data, dio->psize); 314 } 315 atomic_set_long(&dio->refs, HAMMER2_DIO_DIRTY); 316 break; 317 case HAMMER2_DOP_READ: 318 default: 319 KKASSERT(dio->bp == NULL); 320 if (hce > 0) { 321 /* 322 * Synchronous cluster I/O for now. 323 */ 324 peof = (dio->pbase + HAMMER2_SEGMASK64) & 325 ~HAMMER2_SEGMASK64; 326 peof -= dio->dbase; 327 error = cluster_readx(dio->devvp, 328 peof, dev_pbase, 329 dio->psize, bflags, 330 dio->psize, 331 HAMMER2_PBUFSIZE*hce, 332 &dio->bp); 333 } else { 334 error = breadnx(dio->devvp, dev_pbase, 335 dio->psize, bflags, 336 NULL, NULL, 0, &dio->bp); 337 } 338 } 339 } else { 340 if (hce > 0) { 341 /* 342 * Synchronous cluster I/O for now. 343 */ 344 peof = (dio->pbase + HAMMER2_SEGMASK64) & 345 ~HAMMER2_SEGMASK64; 346 peof -= dio->dbase; 347 error = cluster_readx(dio->devvp, 348 peof, dev_pbase, dio->psize, 349 bflags, 350 dio->psize, HAMMER2_PBUFSIZE*hce, 351 &dio->bp); 352 } else { 353 error = breadnx(dio->devvp, dev_pbase, 354 dio->psize, bflags, 355 NULL, NULL, 0, &dio->bp); 356 } 357 if (dio->bp) { 358 /* 359 * Handle NEW flags 360 */ 361 switch(op) { 362 case HAMMER2_DOP_NEW: 363 bkvasync(dio->bp); 364 bzero(hammer2_io_data(dio, lbase), lsize); 365 /* fall through */ 366 case HAMMER2_DOP_NEWNZ: 367 atomic_set_long(&dio->refs, HAMMER2_DIO_DIRTY); 368 break; 369 case HAMMER2_DOP_READ: 370 default: 371 break; 372 } 373 374 /* 375 * Tell the kernel that the buffer cache is not 376 * meta-data based on the btype. This allows 377 * swapcache to distinguish between data and 378 * meta-data. 379 */ 380 switch(btype) { 381 case HAMMER2_BREF_TYPE_DATA: 382 dio->bp->b_flags |= B_NOTMETA; 383 break; 384 default: 385 break; 386 } 387 } 388 } 389 390 if (dio->bp) { 391 bkvasync(dio->bp); 392 BUF_KERNPROC(dio->bp); 393 dio->bp->b_flags &= ~B_AGE; 394 /* dio->bp->b_debug_info2 = dio; */ 395 } 396 dio->error = error; 397 398 /* 399 * Clear INPROG and WAITING, set GOOD wake up anyone waiting. 400 */ 401 for (;;) { 402 orefs = dio->refs; 403 cpu_ccfence(); 404 nrefs = orefs & ~(HAMMER2_DIO_INPROG | HAMMER2_DIO_WAITING); 405 if (error == 0) 406 nrefs |= HAMMER2_DIO_GOOD; 407 if (atomic_cmpset_64(&dio->refs, orefs, nrefs)) { 408 if (orefs & HAMMER2_DIO_WAITING) 409 wakeup(dio); 410 break; 411 } 412 cpu_pause(); 413 } 414 415 /* XXX error handling */ 416 DIO_RECORD(dio HAMMER2_IO_DEBUG_CALL); 417 418 return dio; 419 } 420 421 /* 422 * Release our ref on *diop. 423 * 424 * On the 1->0 transition we clear DIO_GOOD, set DIO_INPROG, and dispose 425 * of dio->bp. Then we clean up DIO_INPROG and DIO_WAITING. 426 */ 427 void 428 _hammer2_io_putblk(hammer2_io_t **diop HAMMER2_IO_DEBUG_ARGS) 429 { 430 hammer2_dev_t *hmp; 431 hammer2_io_t *dio; 432 struct buf *bp; 433 off_t pbase; 434 int psize; 435 int dio_limit; 436 uint64_t orefs; 437 uint64_t nrefs; 438 439 dio = *diop; 440 *diop = NULL; 441 hmp = dio->hmp; 442 DIO_RECORD(dio HAMMER2_IO_DEBUG_CALL); 443 444 KKASSERT((dio->refs & HAMMER2_DIO_MASK) != 0); 445 446 /* 447 * Drop refs. 448 * 449 * On the 1->0 transition clear GOOD and set INPROG, and break. 450 * On any other transition we can return early. 451 */ 452 for (;;) { 453 orefs = dio->refs; 454 cpu_ccfence(); 455 456 if ((orefs & HAMMER2_DIO_MASK) == 1 && 457 (orefs & HAMMER2_DIO_INPROG) == 0) { 458 /* 459 * Lastdrop case, INPROG can be set. GOOD must be 460 * cleared to prevent the getblk shortcut. 461 */ 462 nrefs = orefs - 1; 463 nrefs &= ~(HAMMER2_DIO_GOOD | HAMMER2_DIO_DIRTY); 464 nrefs |= HAMMER2_DIO_INPROG; 465 if (atomic_cmpset_64(&dio->refs, orefs, nrefs)) 466 break; 467 } else if ((orefs & HAMMER2_DIO_MASK) == 1) { 468 /* 469 * Lastdrop case, INPROG already set. We must 470 * wait for INPROG to clear. 471 */ 472 nrefs = orefs | HAMMER2_DIO_WAITING; 473 tsleep_interlock(dio, 0); 474 if (atomic_cmpset_64(&dio->refs, orefs, nrefs)) { 475 tsleep(dio, PINTERLOCKED, "h2dio", hz); 476 } 477 /* retry */ 478 } else { 479 /* 480 * Normal drop case. 481 */ 482 nrefs = orefs - 1; 483 if (atomic_cmpset_64(&dio->refs, orefs, nrefs)) 484 return; 485 /* retry */ 486 } 487 cpu_pause(); 488 /* retry */ 489 } 490 491 /* 492 * Lastdrop (1->0 transition). INPROG has been set, GOOD and DIRTY 493 * have been cleared. iofree_count has not yet been incremented, 494 * note that another accessor race will decrement iofree_count so 495 * we have to increment it regardless. 496 * 497 * We can now dispose of the buffer, and should do it before calling 498 * io_complete() in case there's a race against a new reference 499 * which causes io_complete() to chain and instantiate the bp again. 500 */ 501 pbase = dio->pbase; 502 psize = dio->psize; 503 bp = dio->bp; 504 dio->bp = NULL; 505 506 if ((orefs & HAMMER2_DIO_GOOD) && bp) { 507 /* 508 * Non-errored disposal of bp 509 */ 510 if (orefs & HAMMER2_DIO_DIRTY) { 511 dio_write_stats_update(dio, bp); 512 513 /* 514 * Allows dirty buffers to accumulate and 515 * possibly be canceled (e.g. by a 'rm'), 516 * by default we will burst-write later. 517 * 518 * We generally do NOT want to issue an actual 519 * b[a]write() or cluster_write() here. Due to 520 * the way chains are locked, buffers may be cycled 521 * in and out quite often and disposal here can cause 522 * multiple writes or write-read stalls. 523 * 524 * If FLUSH is set we do want to issue the actual 525 * write. This typically occurs in the write-behind 526 * case when writing to large files. 527 */ 528 off_t peof; 529 int hce; 530 if (dio->refs & HAMMER2_DIO_FLUSH) { 531 if ((hce = hammer2_cluster_write) != 0) { 532 peof = (pbase + HAMMER2_SEGMASK64) & 533 ~HAMMER2_SEGMASK64; 534 peof -= dio->dbase; 535 bp->b_flags |= B_CLUSTEROK; 536 cluster_write(bp, peof, psize, hce); 537 } else { 538 bp->b_flags &= ~B_CLUSTEROK; 539 bawrite(bp); 540 } 541 } else { 542 bp->b_flags &= ~B_CLUSTEROK; 543 bdwrite(bp); 544 } 545 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_RELBUF)) { 546 brelse(bp); 547 } else { 548 bqrelse(bp); 549 } 550 } else if (bp) { 551 /* 552 * Errored disposal of bp 553 */ 554 brelse(bp); 555 } 556 557 /* 558 * Update iofree_count before disposing of the dio 559 */ 560 hmp = dio->hmp; 561 atomic_add_int(&hmp->iofree_count, 1); 562 563 /* 564 * Clear INPROG, GOOD, and WAITING (GOOD should already be clear). 565 * 566 * Also clear FLUSH as it was handled above. 567 */ 568 for (;;) { 569 orefs = dio->refs; 570 cpu_ccfence(); 571 nrefs = orefs & ~(HAMMER2_DIO_INPROG | HAMMER2_DIO_GOOD | 572 HAMMER2_DIO_WAITING | HAMMER2_DIO_FLUSH); 573 if (atomic_cmpset_64(&dio->refs, orefs, nrefs)) { 574 if (orefs & HAMMER2_DIO_WAITING) 575 wakeup(dio); 576 break; 577 } 578 cpu_pause(); 579 } 580 581 /* 582 * We cache free buffers so re-use cases can use a shared lock, but 583 * if too many build up we have to clean them out. 584 */ 585 dio_limit = hammer2_dio_limit; 586 if (dio_limit < 256) 587 dio_limit = 256; 588 if (dio_limit > 1024*1024) 589 dio_limit = 1024*1024; 590 if (hmp->iofree_count > dio_limit) { 591 struct hammer2_cleanupcb_info info; 592 593 RB_INIT(&info.tmptree); 594 hammer2_spin_ex(&hmp->io_spin); 595 if (hmp->iofree_count > dio_limit) { 596 info.count = hmp->iofree_count / 5; 597 RB_SCAN(hammer2_io_tree, &hmp->iotree, NULL, 598 hammer2_io_cleanup_callback, &info); 599 } 600 hammer2_spin_unex(&hmp->io_spin); 601 hammer2_io_cleanup(hmp, &info.tmptree); 602 } 603 } 604 605 /* 606 * Cleanup any dio's with (INPROG | refs) == 0. 607 * 608 * Called to clean up cached DIOs on umount after all activity has been 609 * flushed. 610 */ 611 static 612 int 613 hammer2_io_cleanup_callback(hammer2_io_t *dio, void *arg) 614 { 615 struct hammer2_cleanupcb_info *info = arg; 616 hammer2_io_t *xio; 617 618 if ((dio->refs & (HAMMER2_DIO_MASK | HAMMER2_DIO_INPROG)) == 0) { 619 if (dio->act > 0) { 620 int act; 621 622 act = dio->act - (ticks - dio->ticks) / hz - 1; 623 if (act > 0) { 624 dio->act = act; 625 return 0; 626 } 627 dio->act = 0; 628 } 629 KKASSERT(dio->bp == NULL); 630 if (info->count > 0) { 631 RB_REMOVE(hammer2_io_tree, &dio->hmp->iotree, dio); 632 xio = RB_INSERT(hammer2_io_tree, &info->tmptree, dio); 633 KKASSERT(xio == NULL); 634 --info->count; 635 } 636 } 637 return 0; 638 } 639 640 void 641 hammer2_io_cleanup(hammer2_dev_t *hmp, struct hammer2_io_tree *tree) 642 { 643 hammer2_io_t *dio; 644 645 while ((dio = RB_ROOT(tree)) != NULL) { 646 RB_REMOVE(hammer2_io_tree, tree, dio); 647 KKASSERT(dio->bp == NULL && 648 (dio->refs & (HAMMER2_DIO_MASK | HAMMER2_DIO_INPROG)) == 0); 649 if (dio->refs & HAMMER2_DIO_DIRTY) { 650 kprintf("hammer2_io_cleanup: Dirty buffer " 651 "%016jx/%d (bp=%p)\n", 652 dio->pbase, dio->psize, dio->bp); 653 } 654 kfree(dio, M_HAMMER2); 655 atomic_add_int(&hammer2_dio_count, -1); 656 atomic_add_int(&hmp->iofree_count, -1); 657 } 658 } 659 660 /* 661 * Returns a pointer to the requested data. 662 */ 663 char * 664 hammer2_io_data(hammer2_io_t *dio, off_t lbase) 665 { 666 struct buf *bp; 667 int off; 668 669 bp = dio->bp; 670 KKASSERT(bp != NULL); 671 bkvasync(bp); 672 lbase -= dio->dbase; 673 off = (lbase & ~HAMMER2_OFF_MASK_RADIX) - bp->b_loffset; 674 KKASSERT(off >= 0 && off < bp->b_bufsize); 675 return(bp->b_data + off); 676 } 677 678 int 679 hammer2_io_new(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize, 680 hammer2_io_t **diop) 681 { 682 *diop = hammer2_io_getblk(hmp, btype, lbase, lsize, HAMMER2_DOP_NEW); 683 return ((*diop)->error); 684 } 685 686 int 687 hammer2_io_newnz(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize, 688 hammer2_io_t **diop) 689 { 690 *diop = hammer2_io_getblk(hmp, btype, lbase, lsize, HAMMER2_DOP_NEWNZ); 691 return ((*diop)->error); 692 } 693 694 int 695 _hammer2_io_bread(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize, 696 hammer2_io_t **diop HAMMER2_IO_DEBUG_ARGS) 697 { 698 #ifdef HAMMER2_IO_DEBUG 699 hammer2_io_t *dio; 700 #endif 701 702 *diop = _hammer2_io_getblk(hmp, btype, lbase, lsize, 703 HAMMER2_DOP_READ HAMMER2_IO_DEBUG_CALL); 704 #ifdef HAMMER2_IO_DEBUG 705 if ((dio = *diop) != NULL) { 706 int i = (dio->debug_index - 1) & HAMMER2_IO_DEBUG_MASK; 707 dio->debug_data[i] = debug_data; 708 } 709 #endif 710 return ((*diop)->error); 711 } 712 713 hammer2_io_t * 714 _hammer2_io_getquick(hammer2_dev_t *hmp, off_t lbase, 715 int lsize HAMMER2_IO_DEBUG_ARGS) 716 { 717 hammer2_io_t *dio; 718 719 dio = _hammer2_io_getblk(hmp, 0, lbase, lsize, 720 HAMMER2_DOP_READQ HAMMER2_IO_DEBUG_CALL); 721 return dio; 722 } 723 724 void 725 _hammer2_io_bawrite(hammer2_io_t **diop HAMMER2_IO_DEBUG_ARGS) 726 { 727 atomic_set_64(&(*diop)->refs, HAMMER2_DIO_DIRTY | 728 HAMMER2_DIO_FLUSH); 729 _hammer2_io_putblk(diop HAMMER2_IO_DEBUG_CALL); 730 } 731 732 void 733 _hammer2_io_bdwrite(hammer2_io_t **diop HAMMER2_IO_DEBUG_ARGS) 734 { 735 atomic_set_64(&(*diop)->refs, HAMMER2_DIO_DIRTY); 736 _hammer2_io_putblk(diop HAMMER2_IO_DEBUG_CALL); 737 } 738 739 int 740 _hammer2_io_bwrite(hammer2_io_t **diop HAMMER2_IO_DEBUG_ARGS) 741 { 742 atomic_set_64(&(*diop)->refs, HAMMER2_DIO_DIRTY | 743 HAMMER2_DIO_FLUSH); 744 _hammer2_io_putblk(diop HAMMER2_IO_DEBUG_CALL); 745 return (0); /* XXX */ 746 } 747 748 void 749 hammer2_io_setdirty(hammer2_io_t *dio) 750 { 751 atomic_set_64(&dio->refs, HAMMER2_DIO_DIRTY); 752 } 753 754 /* 755 * This routine is called when a MODIFIED chain is being DESTROYED, 756 * in an attempt to allow the related buffer cache buffer to be 757 * invalidated and discarded instead of flushing it to disk. 758 * 759 * At the moment this case is only really useful for file meta-data. 760 * File data is already handled via the logical buffer cache associated 761 * with the vnode, and will be discarded if it was never flushed to disk. 762 * File meta-data may include inodes, directory entries, and indirect blocks. 763 * 764 * XXX 765 * However, our DIO buffers are PBUFSIZE'd (64KB), and the area being 766 * invalidated might be smaller. Most of the meta-data structures above 767 * are in the 'smaller' category. For now, don't try to invalidate the 768 * data areas. 769 */ 770 void 771 hammer2_io_inval(hammer2_io_t *dio, hammer2_off_t data_off, u_int bytes) 772 { 773 /* NOP */ 774 } 775 776 void 777 _hammer2_io_brelse(hammer2_io_t **diop HAMMER2_IO_DEBUG_ARGS) 778 { 779 _hammer2_io_putblk(diop HAMMER2_IO_DEBUG_CALL); 780 } 781 782 void 783 _hammer2_io_bqrelse(hammer2_io_t **diop HAMMER2_IO_DEBUG_ARGS) 784 { 785 _hammer2_io_putblk(diop HAMMER2_IO_DEBUG_CALL); 786 } 787 788 /* 789 * Set dedup validation bits in a DIO. We do not need the buffer cache 790 * buffer for this. This must be done concurrent with setting bits in 791 * the freemap so as to interlock with bulkfree's clearing of those bits. 792 */ 793 void 794 hammer2_io_dedup_set(hammer2_dev_t *hmp, hammer2_blockref_t *bref) 795 { 796 hammer2_io_t *dio; 797 uint64_t mask; 798 int lsize; 799 int isgood; 800 801 dio = hammer2_io_alloc(hmp, bref->data_off, bref->type, 1, &isgood); 802 if ((int)(bref->data_off & HAMMER2_OFF_MASK_RADIX)) 803 lsize = 1 << (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX); 804 else 805 lsize = 0; 806 mask = hammer2_dedup_mask(dio, bref->data_off, lsize); 807 atomic_clear_64(&dio->dedup_valid, mask); 808 atomic_set_64(&dio->dedup_alloc, mask); 809 hammer2_io_putblk(&dio); 810 } 811 812 /* 813 * Clear dedup validation bits in a DIO. This is typically done when 814 * a modified chain is destroyed or by the bulkfree code. No buffer 815 * is needed for this operation. If the DIO no longer exists it is 816 * equivalent to the bits not being set. 817 */ 818 void 819 hammer2_io_dedup_delete(hammer2_dev_t *hmp, uint8_t btype, 820 hammer2_off_t data_off, u_int bytes) 821 { 822 hammer2_io_t *dio; 823 uint64_t mask; 824 int isgood; 825 826 if ((data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) 827 return; 828 if (btype != HAMMER2_BREF_TYPE_DATA) 829 return; 830 dio = hammer2_io_alloc(hmp, data_off, btype, 0, &isgood); 831 if (dio) { 832 if (data_off < dio->pbase || 833 (data_off & ~HAMMER2_OFF_MASK_RADIX) + bytes > 834 dio->pbase + dio->psize) { 835 panic("hammer2_io_dedup_delete: DATAOFF BAD " 836 "%016jx/%d %016jx\n", 837 data_off, bytes, dio->pbase); 838 } 839 mask = hammer2_dedup_mask(dio, data_off, bytes); 840 atomic_clear_64(&dio->dedup_alloc, mask); 841 atomic_clear_64(&dio->dedup_valid, mask); 842 hammer2_io_putblk(&dio); 843 } 844 } 845 846 /* 847 * Assert that dedup allocation bits in a DIO are not set. This operation 848 * does not require a buffer. The DIO does not need to exist. 849 */ 850 void 851 hammer2_io_dedup_assert(hammer2_dev_t *hmp, hammer2_off_t data_off, u_int bytes) 852 { 853 hammer2_io_t *dio; 854 int isgood; 855 856 dio = hammer2_io_alloc(hmp, data_off, HAMMER2_BREF_TYPE_DATA, 857 0, &isgood); 858 if (dio) { 859 KASSERT((dio->dedup_alloc & 860 hammer2_dedup_mask(dio, data_off, bytes)) == 0, 861 ("hammer2_dedup_assert: %016jx/%d %016jx/%016jx", 862 data_off, 863 bytes, 864 hammer2_dedup_mask(dio, data_off, bytes), 865 dio->dedup_alloc)); 866 hammer2_io_putblk(&dio); 867 } 868 } 869 870 static 871 void 872 dio_write_stats_update(hammer2_io_t *dio, struct buf *bp) 873 { 874 if (bp->b_flags & B_DELWRI) 875 return; 876 hammer2_adjwritecounter(dio->btype, dio->psize); 877 } 878 879 void 880 hammer2_io_bkvasync(hammer2_io_t *dio) 881 { 882 KKASSERT(dio->bp != NULL); 883 bkvasync(dio->bp); 884 } 885 886 /* 887 * Ref a dio that is already owned 888 */ 889 void 890 _hammer2_io_ref(hammer2_io_t *dio HAMMER2_IO_DEBUG_ARGS) 891 { 892 DIO_RECORD(dio HAMMER2_IO_DEBUG_CALL); 893 atomic_add_64(&dio->refs, 1); 894 } 895