1 /* 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 2022 Tomohiro Kusumi <tkusumi@netbsd.org> 5 * Copyright (c) 2011-2022 The DragonFly Project. All rights reserved. 6 * 7 * This code is derived from software contributed to The DragonFly Project 8 * by Matthew Dillon <dillon@dragonflybsd.org> 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in 18 * the documentation and/or other materials provided with the 19 * distribution. 20 * 3. Neither the name of The DragonFly Project nor the names of its 21 * contributors may be used to endorse or promote products derived 22 * from this software without specific, prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 26 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 27 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 28 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 29 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 30 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 31 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 32 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 33 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 34 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 35 * SUCH DAMAGE. 36 */ 37 /* 38 * This module handles low level logical file I/O (strategy) which backs 39 * the logical buffer cache. 40 * 41 * [De]compression, zero-block, check codes, and buffer cache operations 42 * for file data is handled here. 43 * 44 * Live dedup makes its home here as well. 45 */ 46 47 /* 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/kernel.h> 51 #include <sys/buf.h> 52 #include <sys/proc.h> 53 #include <sys/mount.h> 54 #include <sys/vnode.h> 55 #include <sys/objcache.h> 56 */ 57 58 #include "hammer2.h" 59 #include "hammer2_lz4.h" 60 61 #include "zlib/hammer2_zlib.h" 62 63 /* 64 struct objcache *cache_buffer_read; 65 struct objcache *cache_buffer_write; 66 */ 67 68 /* 69 * Strategy code (async logical file buffer I/O from system) 70 * 71 * Except for the transaction init (which should normally not block), 72 * we essentially run the strategy operation asynchronously via a XOP. 73 * 74 * WARNING! The XOP deals with buffer synchronization. It is not synchronized 75 * to the current cpu. 76 * 77 * XXX This isn't supposed to be able to deadlock against vfs_sync vfsync() 78 * calls but it has in the past when multiple flushes are queued. 79 * 80 * XXX We currently terminate the transaction once we get a quorum, otherwise 81 * the frontend can stall, but this can leave the remaining nodes with 82 * a potential flush conflict. We need to delay flushes on those nodes 83 * until running transactions complete separately from the normal 84 * transaction sequencing. FIXME TODO. 85 */ 86 static int hammer2_strategy_read(struct vop_strategy_args *ap); 87 static int hammer2_strategy_write(struct vop_strategy_args *ap); 88 static void hammer2_strategy_read_completion(hammer2_chain_t *focus, 89 const char *data, struct bio *bio); 90 91 static hammer2_off_t hammer2_dedup_lookup(hammer2_dev_t *hmp, 92 char **datap, int pblksize); 93 94 int 95 hammer2_vop_strategy(struct vop_strategy_args *ap) 96 { 97 struct bio *biop; 98 struct m_buf *bp; 99 int error; 100 101 biop = ap->a_bio; 102 bp = biop->bio_buf; 103 104 switch(bp->b_cmd) { 105 case BUF_CMD_READ: 106 error = hammer2_strategy_read(ap); 107 break; 108 case BUF_CMD_WRITE: 109 error = hammer2_strategy_write(ap); 110 break; 111 default: 112 assert(0); 113 /* 114 bp->b_error = error = EINVAL; 115 bp->b_flags |= B_ERROR; 116 biodone(biop); 117 */ 118 break; 119 } 120 return (error); 121 } 122 123 /* 124 * Return the largest contiguous physical disk range for the logical 125 * request, in bytes. 126 * 127 * (struct m_vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb) 128 * 129 * Basically disabled, the logical buffer write thread has to deal with 130 * buffers one-at-a-time. Note that this should not prevent cluster_read() 131 * from reading-ahead, it simply prevents it from trying form a single 132 * cluster buffer for the logical request. H2 already uses 64KB buffers! 133 */ 134 int 135 hammer2_vop_bmap(struct vop_bmap_args *ap) 136 { 137 *ap->a_doffsetp = NOOFFSET; 138 if (ap->a_runp) 139 *ap->a_runp = 0; 140 if (ap->a_runb) 141 *ap->a_runb = 0; 142 return (EOPNOTSUPP); 143 } 144 145 /**************************************************************************** 146 * READ SUPPORT * 147 ****************************************************************************/ 148 /* 149 * Callback used in read path in case that a block is compressed with LZ4. 150 */ 151 static 152 void 153 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio) 154 { 155 struct m_buf *bp; 156 char *compressed_buffer; 157 int compressed_size; 158 int result; 159 160 bp = bio->bio_buf; 161 162 #if 0 163 if bio->bio_caller_info2.index && 164 bio->bio_caller_info1.uvalue32 != 165 crc32(bp->b_data, bp->b_bufsize) --- return error 166 #endif 167 168 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE); 169 compressed_size = *(const int *)data; 170 KKASSERT((uint32_t)compressed_size <= bytes - sizeof(int)); 171 172 compressed_buffer = ecalloc(1, 65536); 173 result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]), 174 compressed_buffer, 175 compressed_size, 176 bp->b_bufsize); 177 if (result < 0) { 178 kprintf("READ PATH: Error during decompression." 179 "bio %016jx/%d\n", 180 (intmax_t)bio->bio_offset, bytes); 181 /* make sure it isn't random garbage */ 182 bzero(compressed_buffer, bp->b_bufsize); 183 } 184 KKASSERT(result <= bp->b_bufsize); 185 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize); 186 if (result < bp->b_bufsize) 187 bzero(bp->b_data + result, bp->b_bufsize - result); 188 free(compressed_buffer); 189 /* 190 bp->b_resid = 0; 191 bp->b_flags |= B_AGE; 192 */ 193 } 194 195 /* 196 * Callback used in read path in case that a block is compressed with ZLIB. 197 * It is almost identical to LZ4 callback, so in theory they can be unified, 198 * but we didn't want to make changes in bio structure for that. 199 */ 200 static 201 void 202 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio) 203 { 204 struct m_buf *bp; 205 char *compressed_buffer; 206 z_stream strm_decompress; 207 int result; 208 int ret; 209 210 bp = bio->bio_buf; 211 212 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE); 213 strm_decompress.avail_in = 0; 214 strm_decompress.next_in = Z_NULL; 215 216 ret = inflateInit(&strm_decompress); 217 218 if (ret != Z_OK) 219 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n"); 220 221 compressed_buffer = ecalloc(1, 65536); 222 strm_decompress.next_in = __DECONST(char *, data); 223 224 /* XXX supply proper size, subset of device bp */ 225 strm_decompress.avail_in = bytes; 226 strm_decompress.next_out = compressed_buffer; 227 strm_decompress.avail_out = bp->b_bufsize; 228 229 ret = inflate(&strm_decompress, Z_FINISH); 230 if (ret != Z_STREAM_END) { 231 kprintf("HAMMER2 ZLIB: Fatal error during decompression.\n"); 232 bzero(compressed_buffer, bp->b_bufsize); 233 } 234 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize); 235 result = bp->b_bufsize - strm_decompress.avail_out; 236 if (result < bp->b_bufsize) 237 bzero(bp->b_data + result, strm_decompress.avail_out); 238 free(compressed_buffer); 239 ret = inflateEnd(&strm_decompress); 240 241 /* 242 bp->b_resid = 0; 243 bp->b_flags |= B_AGE; 244 */ 245 } 246 247 /* 248 * Logical buffer I/O, async read. 249 */ 250 static 251 int 252 hammer2_strategy_read(struct vop_strategy_args *ap) 253 { 254 hammer2_xop_strategy_t *xop; 255 struct bio *bio; 256 hammer2_inode_t *ip; 257 hammer2_key_t lbase; 258 259 bio = ap->a_bio; 260 ip = VTOI(ap->a_vp); 261 262 lbase = bio->bio_offset; 263 KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0); 264 265 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_STRATEGY); 266 xop->finished = 0; 267 xop->bio = bio; 268 xop->lbase = lbase; 269 hammer2_mtx_init(&xop->lock, "h2bior"); 270 hammer2_xop_start(&xop->head, &hammer2_strategy_read_desc); 271 /* asynchronous completion */ 272 273 return(0); 274 } 275 276 /* 277 * Per-node XOP (threaded), do a synchronous lookup of the chain and 278 * its data. The frontend is asynchronous, so we are also responsible 279 * for racing to terminate the frontend. 280 */ 281 void 282 hammer2_xop_strategy_read(hammer2_xop_t *arg, void *scratch, int clindex) 283 { 284 hammer2_xop_strategy_t *xop = &arg->xop_strategy; 285 hammer2_chain_t *parent; 286 hammer2_chain_t *chain; 287 hammer2_chain_t *focus; 288 hammer2_key_t key_dummy; 289 hammer2_key_t lbase; 290 struct bio *bio; 291 struct m_buf *bp; 292 const char *data; 293 int error; 294 295 /* 296 * Note that we can race completion of the bio supplied by 297 * the front-end so we cannot access it until we determine 298 * that we are the ones finishing it up. 299 */ 300 lbase = xop->lbase; 301 302 /* 303 * This is difficult to optimize. The logical buffer might be 304 * partially dirty (contain dummy zero-fill pages), which would 305 * mess up our crc calculation if we were to try a direct read. 306 * So for now we always double-buffer through the underlying 307 * storage. 308 * 309 * If not for the above problem we could conditionalize on 310 * (1) 64KB buffer, (2) one chain (not multi-master) and 311 * (3) !hammer2_double_buffer, and issue a direct read into the 312 * logical buffer. 313 */ 314 parent = hammer2_inode_chain(xop->head.ip1, clindex, 315 HAMMER2_RESOLVE_ALWAYS | 316 HAMMER2_RESOLVE_SHARED); 317 if (parent) { 318 chain = hammer2_chain_lookup(&parent, &key_dummy, 319 lbase, lbase, 320 &error, 321 HAMMER2_LOOKUP_ALWAYS | 322 HAMMER2_LOOKUP_SHARED); 323 if (chain) 324 error = chain->error; 325 } else { 326 error = HAMMER2_ERROR_EIO; 327 chain = NULL; 328 } 329 error = hammer2_xop_feed(&xop->head, chain, clindex, error); 330 if (chain) { 331 hammer2_chain_unlock(chain); 332 hammer2_chain_drop(chain); 333 } 334 if (parent) { 335 hammer2_chain_unlock(parent); 336 hammer2_chain_drop(parent); 337 } 338 chain = NULL; /* safety */ 339 parent = NULL; /* safety */ 340 341 /* 342 * Race to finish the frontend. First-to-complete. bio is only 343 * valid if we are determined to be the ones able to complete 344 * the operation. 345 */ 346 if (xop->finished) 347 return; 348 hammer2_mtx_ex(&xop->lock); 349 if (xop->finished) { 350 hammer2_mtx_unlock(&xop->lock); 351 return; 352 } 353 bio = xop->bio; 354 bp = bio->bio_buf; 355 bkvasync(bp); 356 357 /* 358 * Async operation has not completed and we now own the lock. 359 * Determine if we can complete the operation by issuing the 360 * frontend collection non-blocking. 361 * 362 * H2 double-buffers the data, setting B_NOTMETA on the logical 363 * buffer hints to the OS that the logical buffer should not be 364 * swapcached (since the device buffer can be). 365 * 366 * Also note that even for compressed data we would rather the 367 * kernel cache/swapcache device buffers more and (decompressed) 368 * logical buffers less, since that will significantly improve 369 * the amount of end-user data that can be cached. 370 * 371 * NOTE: The chain->data for xop->head.cluster.focus will be 372 * synchronized to the current cpu by xop_collect(), 373 * but other chains in the cluster might not be. 374 */ 375 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT); 376 377 switch(error) { 378 case 0: 379 xop->finished = 1; 380 hammer2_mtx_unlock(&xop->lock); 381 //bp->b_flags |= B_NOTMETA; 382 focus = xop->head.cluster.focus; 383 data = hammer2_xop_gdata(&xop->head)->buf; 384 hammer2_strategy_read_completion(focus, data, xop->bio); 385 hammer2_xop_pdata(&xop->head); 386 //biodone(bio); 387 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 388 break; 389 case HAMMER2_ERROR_ENOENT: 390 xop->finished = 1; 391 hammer2_mtx_unlock(&xop->lock); 392 /* 393 bp->b_flags |= B_NOTMETA; 394 bp->b_resid = 0; 395 bp->b_error = 0; 396 */ 397 bzero(bp->b_data, bp->b_bcount); 398 //biodone(bio); 399 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 400 break; 401 case HAMMER2_ERROR_EINPROGRESS: 402 hammer2_mtx_unlock(&xop->lock); 403 break; 404 default: 405 kprintf("xop_strategy_read: error %08x loff=%016jx\n", 406 error, (intmax_t)bp->b_loffset); 407 xop->finished = 1; 408 hammer2_mtx_unlock(&xop->lock); 409 assert(0); 410 /* 411 bp->b_flags |= B_ERROR; 412 bp->b_error = EIO; 413 biodone(bio); 414 */ 415 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 416 break; 417 } 418 } 419 420 static 421 void 422 hammer2_strategy_read_completion(hammer2_chain_t *focus, const char *data, 423 struct bio *bio) 424 { 425 struct m_buf *bp = bio->bio_buf; 426 427 if (focus->bref.type == HAMMER2_BREF_TYPE_INODE) { 428 /* 429 * Copy from in-memory inode structure. 430 */ 431 bcopy(((const hammer2_inode_data_t *)data)->u.data, 432 bp->b_data, HAMMER2_EMBEDDED_BYTES); 433 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES, 434 bp->b_bcount - HAMMER2_EMBEDDED_BYTES); 435 /* 436 bp->b_resid = 0; 437 bp->b_error = 0; 438 */ 439 } else if (focus->bref.type == HAMMER2_BREF_TYPE_DATA) { 440 /* 441 * Data is on-media, record for live dedup. Release the 442 * chain (try to free it) when done. The data is still 443 * cached by both the buffer cache in front and the 444 * block device behind us. 445 * 446 * NOTE: Deduplication cannot be safely recorded for 447 * records without a check code. 448 */ 449 hammer2_dedup_record(focus, NULL, data); 450 atomic_set_int(&focus->flags, HAMMER2_CHAIN_RELEASE); 451 452 /* 453 * Decompression and copy. 454 */ 455 switch (HAMMER2_DEC_COMP(focus->bref.methods)) { 456 case HAMMER2_COMP_LZ4: 457 hammer2_decompress_LZ4_callback(data, focus->bytes, 458 bio); 459 /* b_resid set by call */ 460 break; 461 case HAMMER2_COMP_ZLIB: 462 hammer2_decompress_ZLIB_callback(data, focus->bytes, 463 bio); 464 /* b_resid set by call */ 465 break; 466 case HAMMER2_COMP_NONE: 467 KKASSERT(focus->bytes <= bp->b_bcount); 468 bcopy(data, bp->b_data, focus->bytes); 469 if (focus->bytes < bp->b_bcount) { 470 bzero(bp->b_data + focus->bytes, 471 bp->b_bcount - focus->bytes); 472 } 473 /* 474 bp->b_resid = 0; 475 bp->b_error = 0; 476 */ 477 break; 478 default: 479 panic("hammer2_strategy_read_completion: " 480 "unknown compression type"); 481 } 482 } else { 483 panic("hammer2_strategy_read_completion: unknown bref type"); 484 } 485 } 486 487 /**************************************************************************** 488 * WRITE SUPPORT * 489 ****************************************************************************/ 490 491 /* 492 * Functions for compression in threads, 493 * from hammer2_vnops.c 494 */ 495 static void hammer2_write_file_core(char *data, hammer2_inode_t *ip, 496 hammer2_chain_t **parentp, 497 hammer2_key_t lbase, int ioflag, int pblksize, 498 hammer2_tid_t mtid, int *errorp); 499 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip, 500 hammer2_chain_t **parentp, 501 hammer2_key_t lbase, int ioflag, int pblksize, 502 hammer2_tid_t mtid, int *errorp, 503 int comp_algo, int check_algo); 504 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip, 505 hammer2_chain_t **parentp, 506 hammer2_key_t lbase, int ioflag, int pblksize, 507 hammer2_tid_t mtid, int *errorp, 508 int check_algo); 509 static int test_block_zeros(const char *buf, size_t bytes); 510 static void zero_write(char *data, hammer2_inode_t *ip, 511 hammer2_chain_t **parentp, 512 hammer2_key_t lbase, 513 hammer2_tid_t mtid, int *errorp); 514 static void hammer2_write_bp(hammer2_chain_t *chain, char *data, 515 int ioflag, int pblksize, 516 hammer2_tid_t mtid, int *errorp, 517 int check_algo); 518 519 int 520 hammer2_strategy_write(struct vop_strategy_args *ap) 521 { 522 hammer2_xop_strategy_t *xop; 523 hammer2_pfs_t *pmp; 524 struct bio *bio; 525 hammer2_inode_t *ip; 526 527 bio = ap->a_bio; 528 ip = VTOI(ap->a_vp); 529 pmp = ip->pmp; 530 531 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 532 hammer2_lwinprog_ref(pmp); 533 hammer2_trans_assert_strategy(pmp); 534 hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE); 535 536 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 537 HAMMER2_XOP_STRATEGY); 538 xop->finished = 0; 539 xop->bio = bio; 540 xop->lbase = bio->bio_offset; 541 hammer2_mtx_init(&xop->lock, "h2biow"); 542 hammer2_xop_start(&xop->head, &hammer2_strategy_write_desc); 543 /* asynchronous completion */ 544 545 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe); 546 547 return(0); 548 } 549 550 /* 551 * Per-node XOP (threaded). Write the logical buffer to the media. 552 * 553 * This is a bit problematic because there may be multiple target and 554 * any of them may be able to release the bp. In addition, if our 555 * particulr target is offline we don't want to block the bp (and thus 556 * the frontend). To accomplish this we copy the data to the per-thr 557 * scratch buffer. 558 */ 559 void 560 hammer2_xop_strategy_write(hammer2_xop_t *arg, void *scratch, int clindex) 561 { 562 hammer2_xop_strategy_t *xop = &arg->xop_strategy; 563 hammer2_chain_t *parent; 564 hammer2_key_t lbase; 565 hammer2_inode_t *ip; 566 struct bio *bio; 567 struct m_buf *bp; 568 int error; 569 int lblksize; 570 int pblksize; 571 char *bio_data; 572 573 /* 574 * We can only access the bp/bio if the frontend has not yet 575 * completed. 576 */ 577 if (xop->finished) 578 return; 579 hammer2_mtx_sh(&xop->lock); 580 if (xop->finished) { 581 hammer2_mtx_unlock(&xop->lock); 582 return; 583 } 584 585 lbase = xop->lbase; 586 bio = xop->bio; /* ephermal */ 587 bp = bio->bio_buf; /* ephermal */ 588 ip = xop->head.ip1; /* retained by ref */ 589 bio_data = scratch; 590 591 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */ 592 593 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL); 594 pblksize = hammer2_calc_physical(ip, lbase); 595 bkvasync(bp); 596 KKASSERT(lblksize <= MAXPHYS); 597 bcopy(bp->b_data, bio_data, lblksize); 598 599 hammer2_mtx_unlock(&xop->lock); 600 bp = NULL; /* safety, illegal to access after unlock */ 601 bio = NULL; /* safety, illegal to access after unlock */ 602 603 /* 604 * Actual operation 605 */ 606 parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS); 607 hammer2_write_file_core(bio_data, ip, &parent, 608 lbase, IO_ASYNC, pblksize, 609 xop->head.mtid, &error); 610 if (parent) { 611 hammer2_chain_unlock(parent); 612 hammer2_chain_drop(parent); 613 parent = NULL; /* safety */ 614 } 615 hammer2_xop_feed(&xop->head, NULL, clindex, error); 616 617 /* 618 * Try to complete the operation on behalf of the front-end. 619 */ 620 if (xop->finished) 621 return; 622 hammer2_mtx_ex(&xop->lock); 623 if (xop->finished) { 624 hammer2_mtx_unlock(&xop->lock); 625 return; 626 } 627 628 /* 629 * Async operation has not completed and we now own the lock. 630 * Determine if we can complete the operation by issuing the 631 * frontend collection non-blocking. 632 * 633 * H2 double-buffers the data, setting B_NOTMETA on the logical 634 * buffer hints to the OS that the logical buffer should not be 635 * swapcached (since the device buffer can be). 636 */ 637 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT); 638 639 if (error == HAMMER2_ERROR_EINPROGRESS) { 640 hammer2_mtx_unlock(&xop->lock); 641 return; 642 } 643 644 /* 645 * Async operation has completed. 646 */ 647 xop->finished = 1; 648 hammer2_mtx_unlock(&xop->lock); 649 650 bio = xop->bio; /* now owned by us */ 651 bp = bio->bio_buf; /* now owned by us */ 652 653 if (error == HAMMER2_ERROR_ENOENT || error == 0) { 654 /* 655 bp->b_flags |= B_NOTMETA; 656 bp->b_resid = 0; 657 bp->b_error = 0; 658 biodone(bio); 659 */ 660 } else { 661 kprintf("xop_strategy_write: error %d loff=%016jx\n", 662 error, (intmax_t)bp->b_loffset); 663 assert(0); 664 /* 665 bp->b_flags |= B_ERROR; 666 bp->b_error = EIO; 667 biodone(bio); 668 */ 669 } 670 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 671 hammer2_trans_assert_strategy(ip->pmp); 672 hammer2_lwinprog_drop(ip->pmp); 673 hammer2_trans_done(ip->pmp, HAMMER2_TRANS_BUFCACHE); 674 } 675 676 /* 677 * Wait for pending I/O to complete 678 */ 679 void 680 hammer2_bioq_sync(hammer2_pfs_t *pmp) 681 { 682 hammer2_lwinprog_wait(pmp, 0); 683 } 684 685 /* 686 * Assign physical storage at (cparent, lbase), returning a suitable chain 687 * and setting *errorp appropriately. 688 * 689 * If no error occurs, the returned chain will be in a modified state. 690 * 691 * If an error occurs, the returned chain may or may not be NULL. If 692 * not-null any chain->error (if not 0) will also be rolled up into *errorp. 693 * So the caller only needs to test *errorp. 694 * 695 * cparent can wind up being anything. 696 * 697 * If datap is not NULL, *datap points to the real data we intend to write. 698 * If we can dedup the storage location we set *datap to NULL to indicate 699 * to the caller that a dedup occurred. 700 * 701 * NOTE: Special case for data embedded in inode. 702 */ 703 static 704 hammer2_chain_t * 705 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp, 706 hammer2_key_t lbase, int pblksize, 707 hammer2_tid_t mtid, char **datap, int *errorp) 708 { 709 hammer2_chain_t *chain; 710 hammer2_key_t key_dummy; 711 hammer2_off_t dedup_off; 712 int pradix = hammer2_getradix(pblksize); 713 714 /* 715 * Locate the chain associated with lbase, return a locked chain. 716 * However, do not instantiate any data reference (which utilizes a 717 * device buffer) because we will be using direct IO via the 718 * logical buffer cache buffer. 719 */ 720 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN); 721 722 chain = hammer2_chain_lookup(parentp, &key_dummy, 723 lbase, lbase, 724 errorp, 725 HAMMER2_LOOKUP_NODATA); 726 727 /* 728 * The lookup code should not return a DELETED chain to us, unless 729 * its a short-file embedded in the inode. Then it is possible for 730 * the lookup to return a deleted inode. 731 */ 732 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) && 733 chain->bref.type != HAMMER2_BREF_TYPE_INODE) { 734 kprintf("assign physical deleted chain @ " 735 "%016jx (%016jx.%02x) ip %016jx\n", 736 lbase, chain->bref.data_off, chain->bref.type, 737 ip->meta.inum); 738 Debugger("bleh"); 739 } 740 741 if (chain == NULL) { 742 /* 743 * We found a hole, create a new chain entry. 744 * 745 * NOTE: DATA chains are created without device backing 746 * store (nor do we want any). 747 */ 748 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap, 749 pblksize); 750 *errorp |= hammer2_chain_create(parentp, &chain, NULL, ip->pmp, 751 HAMMER2_ENC_CHECK(ip->meta.check_algo) | 752 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE), 753 lbase, HAMMER2_PBUFRADIX, 754 HAMMER2_BREF_TYPE_DATA, 755 pblksize, mtid, 756 dedup_off, 0); 757 if (chain == NULL) 758 goto failed; 759 /*ip->delta_dcount += pblksize;*/ 760 } else if (chain->error == 0) { 761 switch (chain->bref.type) { 762 case HAMMER2_BREF_TYPE_INODE: 763 /* 764 * The data is embedded in the inode, which requires 765 * a bit more finess. 766 */ 767 *errorp |= hammer2_chain_modify_ip(ip, chain, mtid, 0); 768 break; 769 case HAMMER2_BREF_TYPE_DATA: 770 dedup_off = hammer2_dedup_lookup(chain->hmp, datap, 771 pblksize); 772 if (chain->bytes != pblksize) { 773 *errorp |= hammer2_chain_resize(chain, 774 mtid, dedup_off, 775 pradix, 776 HAMMER2_MODIFY_OPTDATA); 777 if (*errorp) 778 break; 779 } 780 781 /* 782 * DATA buffers must be marked modified whether the 783 * data is in a logical buffer or not. We also have 784 * to make this call to fixup the chain data pointers 785 * after resizing in case this is an encrypted or 786 * compressed buffer. 787 */ 788 *errorp |= hammer2_chain_modify(chain, mtid, dedup_off, 789 HAMMER2_MODIFY_OPTDATA); 790 break; 791 default: 792 panic("hammer2_assign_physical: bad type"); 793 /* NOT REACHED */ 794 break; 795 } 796 } else { 797 *errorp = chain->error; 798 } 799 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 800 failed: 801 return (chain); 802 } 803 804 /* 805 * hammer2_write_file_core() 806 * 807 * The core write function which determines which path to take 808 * depending on compression settings. We also have to locate the 809 * related chains so we can calculate and set the check data for 810 * the blockref. 811 */ 812 static 813 void 814 hammer2_write_file_core(char *data, hammer2_inode_t *ip, 815 hammer2_chain_t **parentp, 816 hammer2_key_t lbase, int ioflag, int pblksize, 817 hammer2_tid_t mtid, int *errorp) 818 { 819 hammer2_chain_t *chain; 820 char *bdata; 821 822 *errorp = 0; 823 824 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) { 825 case HAMMER2_COMP_NONE: 826 /* 827 * We have to assign physical storage to the buffer 828 * we intend to dirty or write now to avoid deadlocks 829 * in the strategy code later. 830 * 831 * This can return NOOFFSET for inode-embedded data. 832 * The strategy code will take care of it in that case. 833 */ 834 bdata = data; 835 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize, 836 mtid, &bdata, errorp); 837 if (*errorp) { 838 /* skip modifications */ 839 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 840 hammer2_inode_data_t *wipdata; 841 842 wipdata = &chain->data->ipdata; 843 KKASSERT(wipdata->meta.op_flags & 844 HAMMER2_OPFLAG_DIRECTDATA); 845 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 846 ++hammer2_iod_file_wembed; 847 } else if (bdata == NULL) { 848 /* 849 * Copy of data already present on-media. 850 */ 851 chain->bref.methods = 852 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 853 HAMMER2_ENC_CHECK(ip->meta.check_algo); 854 hammer2_chain_setcheck(chain, data); 855 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 856 } else { 857 hammer2_write_bp(chain, data, ioflag, pblksize, 858 mtid, errorp, ip->meta.check_algo); 859 } 860 if (chain) { 861 hammer2_chain_unlock(chain); 862 hammer2_chain_drop(chain); 863 } 864 break; 865 case HAMMER2_COMP_AUTOZERO: 866 /* 867 * Check for zero-fill only 868 */ 869 hammer2_zero_check_and_write(data, ip, parentp, 870 lbase, ioflag, pblksize, 871 mtid, errorp, 872 ip->meta.check_algo); 873 break; 874 case HAMMER2_COMP_LZ4: 875 case HAMMER2_COMP_ZLIB: 876 default: 877 /* 878 * Check for zero-fill and attempt compression. 879 */ 880 hammer2_compress_and_write(data, ip, parentp, 881 lbase, ioflag, pblksize, 882 mtid, errorp, 883 ip->meta.comp_algo, 884 ip->meta.check_algo); 885 break; 886 } 887 } 888 889 /* 890 * Helper 891 * 892 * Generic function that will perform the compression in compression 893 * write path. The compression algorithm is determined by the settings 894 * obtained from inode. 895 */ 896 static 897 void 898 hammer2_compress_and_write(char *data, hammer2_inode_t *ip, 899 hammer2_chain_t **parentp, 900 hammer2_key_t lbase, int ioflag, int pblksize, 901 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo) 902 { 903 hammer2_chain_t *chain; 904 int comp_size; 905 int comp_block_size; 906 char *comp_buffer; 907 char *bdata; 908 909 /* 910 * An all-zeros write creates a hole unless the check code 911 * is disabled. When the check code is disabled all writes 912 * are done in-place, including any all-zeros writes. 913 * 914 * NOTE: A snapshot will still force a copy-on-write 915 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c). 916 */ 917 if (check_algo != HAMMER2_CHECK_NONE && 918 test_block_zeros(data, pblksize)) { 919 zero_write(data, ip, parentp, lbase, mtid, errorp); 920 return; 921 } 922 923 /* 924 * Compression requested. Try to compress the block. We store 925 * the data normally if we cannot sufficiently compress it. 926 * 927 * We have a heuristic to detect files which are mostly 928 * uncompressable and avoid the compression attempt in that 929 * case. If the compression heuristic is turned off, we always 930 * try to compress. 931 */ 932 comp_size = 0; 933 comp_buffer = NULL; 934 935 KKASSERT(pblksize / 2 <= 32768); 936 937 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 || 938 hammer2_always_compress) { 939 z_stream strm_compress; 940 int comp_level; 941 int ret; 942 943 switch(HAMMER2_DEC_ALGO(comp_algo)) { 944 case HAMMER2_COMP_LZ4: 945 /* 946 * We need to prefix with the size, LZ4 947 * doesn't do it for us. Add the related 948 * overhead. 949 * 950 * NOTE: The LZ4 code seems to assume at least an 951 * 8-byte buffer size granularity and may 952 * overrun the buffer if given a 4-byte 953 * granularity. 954 */ 955 comp_buffer = ecalloc(1, 32768); 956 comp_size = LZ4_compress_limitedOutput( 957 data, 958 &comp_buffer[sizeof(int)], 959 pblksize, 960 pblksize / 2 - sizeof(int64_t)); 961 *(int *)comp_buffer = comp_size; 962 if (comp_size) 963 comp_size += sizeof(int); 964 break; 965 case HAMMER2_COMP_ZLIB: 966 comp_level = HAMMER2_DEC_LEVEL(comp_algo); 967 if (comp_level == 0) 968 comp_level = 6; /* default zlib compression */ 969 else if (comp_level < 6) 970 comp_level = 6; 971 else if (comp_level > 9) 972 comp_level = 9; 973 ret = deflateInit(&strm_compress, comp_level); 974 if (ret != Z_OK) { 975 kprintf("HAMMER2 ZLIB: fatal error " 976 "on deflateInit.\n"); 977 } 978 979 comp_buffer = ecalloc(1, 32768); 980 strm_compress.next_in = data; 981 strm_compress.avail_in = pblksize; 982 strm_compress.next_out = comp_buffer; 983 strm_compress.avail_out = pblksize / 2; 984 ret = deflate(&strm_compress, Z_FINISH); 985 if (ret == Z_STREAM_END) { 986 comp_size = pblksize / 2 - 987 strm_compress.avail_out; 988 } else { 989 comp_size = 0; 990 } 991 ret = deflateEnd(&strm_compress); 992 break; 993 default: 994 kprintf("Error: Unknown compression method.\n"); 995 kprintf("Comp_method = %d.\n", comp_algo); 996 break; 997 } 998 } 999 1000 if (comp_size == 0) { 1001 /* 1002 * compression failed or turned off 1003 */ 1004 comp_block_size = pblksize; /* safety */ 1005 if (++ip->comp_heuristic > 128) 1006 ip->comp_heuristic = 8; 1007 } else { 1008 /* 1009 * compression succeeded 1010 */ 1011 ip->comp_heuristic = 0; 1012 if (comp_size <= 1024) { 1013 comp_block_size = 1024; 1014 } else if (comp_size <= 2048) { 1015 comp_block_size = 2048; 1016 } else if (comp_size <= 4096) { 1017 comp_block_size = 4096; 1018 } else if (comp_size <= 8192) { 1019 comp_block_size = 8192; 1020 } else if (comp_size <= 16384) { 1021 comp_block_size = 16384; 1022 } else if (comp_size <= 32768) { 1023 comp_block_size = 32768; 1024 } else { 1025 panic("hammer2: WRITE PATH: " 1026 "Weird comp_size value."); 1027 /* NOT REACHED */ 1028 comp_block_size = pblksize; 1029 } 1030 1031 /* 1032 * Must zero the remainder or dedup (which operates on a 1033 * physical block basis) will not find matches. 1034 */ 1035 if (comp_size < comp_block_size) { 1036 bzero(comp_buffer + comp_size, 1037 comp_block_size - comp_size); 1038 } 1039 } 1040 1041 /* 1042 * Assign physical storage, bdata will be set to NULL if a live-dedup 1043 * was successful. 1044 */ 1045 bdata = comp_size ? comp_buffer : data; 1046 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size, 1047 mtid, &bdata, errorp); 1048 1049 if (*errorp) { 1050 goto done; 1051 } 1052 1053 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 1054 hammer2_inode_data_t *wipdata; 1055 1056 *errorp = hammer2_chain_modify_ip(ip, chain, mtid, 0); 1057 if (*errorp == 0) { 1058 wipdata = &chain->data->ipdata; 1059 KKASSERT(wipdata->meta.op_flags & 1060 HAMMER2_OPFLAG_DIRECTDATA); 1061 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1062 ++hammer2_iod_file_wembed; 1063 } 1064 } else if (bdata == NULL) { 1065 /* 1066 * Live deduplication, a copy of the data is already present 1067 * on the media. 1068 */ 1069 if (comp_size) { 1070 chain->bref.methods = 1071 HAMMER2_ENC_COMP(comp_algo) + 1072 HAMMER2_ENC_CHECK(check_algo); 1073 } else { 1074 chain->bref.methods = 1075 HAMMER2_ENC_COMP( 1076 HAMMER2_COMP_NONE) + 1077 HAMMER2_ENC_CHECK(check_algo); 1078 } 1079 bdata = comp_size ? comp_buffer : data; 1080 hammer2_chain_setcheck(chain, bdata); 1081 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1082 } else { 1083 hammer2_io_t *dio; 1084 1085 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED); 1086 1087 switch(chain->bref.type) { 1088 case HAMMER2_BREF_TYPE_INODE: 1089 panic("hammer2_compress_and_write: unexpected inode\n"); 1090 break; 1091 case HAMMER2_BREF_TYPE_DATA: 1092 /* 1093 * Optimize out the read-before-write 1094 * if possible. 1095 */ 1096 *errorp = hammer2_io_newnz(chain->hmp, 1097 chain->bref.type, 1098 chain->bref.data_off, 1099 chain->bytes, 1100 &dio); 1101 if (*errorp) { 1102 hammer2_io_brelse(&dio); 1103 kprintf("hammer2: WRITE PATH: " 1104 "dbp bread error\n"); 1105 break; 1106 } 1107 bdata = hammer2_io_data(dio, chain->bref.data_off); 1108 1109 /* 1110 * When loading the block make sure we don't 1111 * leave garbage after the compressed data. 1112 */ 1113 if (comp_size) { 1114 chain->bref.methods = 1115 HAMMER2_ENC_COMP(comp_algo) + 1116 HAMMER2_ENC_CHECK(check_algo); 1117 bcopy(comp_buffer, bdata, comp_block_size); 1118 } else { 1119 chain->bref.methods = 1120 HAMMER2_ENC_COMP( 1121 HAMMER2_COMP_NONE) + 1122 HAMMER2_ENC_CHECK(check_algo); 1123 bcopy(data, bdata, pblksize); 1124 } 1125 1126 /* 1127 * The flush code doesn't calculate check codes for 1128 * file data (doing so can result in excessive I/O), 1129 * so we do it here. 1130 */ 1131 hammer2_chain_setcheck(chain, bdata); 1132 1133 /* 1134 * Device buffer is now valid, chain is no longer in 1135 * the initial state. 1136 * 1137 * (No blockref table worries with file data) 1138 */ 1139 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1140 hammer2_dedup_record(chain, dio, bdata); 1141 1142 /* Now write the related bdp. */ 1143 if (ioflag & IO_SYNC) { 1144 /* 1145 * Synchronous I/O requested. 1146 */ 1147 hammer2_io_bwrite(&dio); 1148 /* 1149 } else if ((ioflag & IO_DIRECT) && 1150 loff + n == pblksize) { 1151 hammer2_io_bdwrite(&dio); 1152 */ 1153 } else if (ioflag & IO_ASYNC) { 1154 hammer2_io_bawrite(&dio); 1155 } else { 1156 hammer2_io_bdwrite(&dio); 1157 } 1158 break; 1159 default: 1160 panic("hammer2_compress_and_write: bad chain type %d\n", 1161 chain->bref.type); 1162 /* NOT REACHED */ 1163 break; 1164 } 1165 } 1166 done: 1167 if (chain) { 1168 hammer2_chain_unlock(chain); 1169 hammer2_chain_drop(chain); 1170 } 1171 if (comp_buffer) 1172 free(comp_buffer); 1173 } 1174 1175 /* 1176 * Helper 1177 * 1178 * Function that performs zero-checking and writing without compression, 1179 * it corresponds to default zero-checking path. 1180 */ 1181 static 1182 void 1183 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip, 1184 hammer2_chain_t **parentp, 1185 hammer2_key_t lbase, int ioflag, int pblksize, 1186 hammer2_tid_t mtid, int *errorp, 1187 int check_algo) 1188 { 1189 hammer2_chain_t *chain; 1190 char *bdata; 1191 1192 if (check_algo != HAMMER2_CHECK_NONE && 1193 test_block_zeros(data, pblksize)) { 1194 /* 1195 * An all-zeros write creates a hole unless the check code 1196 * is disabled. When the check code is disabled all writes 1197 * are done in-place, including any all-zeros writes. 1198 * 1199 * NOTE: A snapshot will still force a copy-on-write 1200 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c). 1201 */ 1202 zero_write(data, ip, parentp, lbase, mtid, errorp); 1203 } else { 1204 /* 1205 * Normal write (bdata set to NULL if de-duplicated) 1206 */ 1207 bdata = data; 1208 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize, 1209 mtid, &bdata, errorp); 1210 if (*errorp) { 1211 /* do nothing */ 1212 } else if (bdata) { 1213 hammer2_write_bp(chain, data, ioflag, pblksize, 1214 mtid, errorp, check_algo); 1215 } else { 1216 /* dedup occurred */ 1217 chain->bref.methods = 1218 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 1219 HAMMER2_ENC_CHECK(check_algo); 1220 hammer2_chain_setcheck(chain, data); 1221 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1222 } 1223 if (chain) { 1224 hammer2_chain_unlock(chain); 1225 hammer2_chain_drop(chain); 1226 } 1227 } 1228 } 1229 1230 /* 1231 * Helper 1232 * 1233 * A function to test whether a block of data contains only zeros, 1234 * returns TRUE (non-zero) if the block is all zeros. 1235 */ 1236 static 1237 int 1238 test_block_zeros(const char *buf, size_t bytes) 1239 { 1240 size_t i; 1241 1242 for (i = 0; i < bytes; i += sizeof(long)) { 1243 if (*(const long *)(buf + i) != 0) 1244 return (0); 1245 } 1246 return (1); 1247 } 1248 1249 /* 1250 * Helper 1251 * 1252 * Function to "write" a block that contains only zeros. 1253 */ 1254 static 1255 void 1256 zero_write(char *data, hammer2_inode_t *ip, 1257 hammer2_chain_t **parentp, 1258 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp) 1259 { 1260 hammer2_chain_t *chain; 1261 hammer2_key_t key_dummy; 1262 1263 chain = hammer2_chain_lookup(parentp, &key_dummy, 1264 lbase, lbase, 1265 errorp, 1266 HAMMER2_LOOKUP_NODATA); 1267 if (chain) { 1268 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 1269 hammer2_inode_data_t *wipdata; 1270 1271 if (*errorp == 0) { 1272 *errorp = hammer2_chain_modify_ip(ip, chain, 1273 mtid, 0); 1274 } 1275 if (*errorp == 0) { 1276 wipdata = &chain->data->ipdata; 1277 KKASSERT(wipdata->meta.op_flags & 1278 HAMMER2_OPFLAG_DIRECTDATA); 1279 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1280 ++hammer2_iod_file_wembed; 1281 } 1282 } else { 1283 /* chain->error ok for deletion */ 1284 hammer2_chain_delete(*parentp, chain, 1285 mtid, HAMMER2_DELETE_PERMANENT); 1286 ++hammer2_iod_file_wzero; 1287 } 1288 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 1289 hammer2_chain_unlock(chain); 1290 hammer2_chain_drop(chain); 1291 } else { 1292 ++hammer2_iod_file_wzero; 1293 } 1294 } 1295 1296 /* 1297 * Helper 1298 * 1299 * Function to write the data as it is, without performing any sort of 1300 * compression. This function is used in path without compression and 1301 * default zero-checking path. 1302 */ 1303 static 1304 void 1305 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag, 1306 int pblksize, 1307 hammer2_tid_t mtid, int *errorp, int check_algo) 1308 { 1309 hammer2_inode_data_t *wipdata; 1310 hammer2_io_t *dio; 1311 char *bdata; 1312 int error; 1313 1314 error = 0; /* XXX TODO below */ 1315 1316 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED); 1317 1318 switch(chain->bref.type) { 1319 case HAMMER2_BREF_TYPE_INODE: 1320 wipdata = &chain->data->ipdata; 1321 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA); 1322 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1323 error = 0; 1324 ++hammer2_iod_file_wembed; 1325 break; 1326 case HAMMER2_BREF_TYPE_DATA: 1327 error = hammer2_io_newnz(chain->hmp, 1328 chain->bref.type, 1329 chain->bref.data_off, 1330 chain->bytes, &dio); 1331 if (error) { 1332 hammer2_io_bqrelse(&dio); 1333 kprintf("hammer2: WRITE PATH: " 1334 "dbp bread error\n"); 1335 break; 1336 } 1337 bdata = hammer2_io_data(dio, chain->bref.data_off); 1338 1339 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 1340 HAMMER2_ENC_CHECK(check_algo); 1341 bcopy(data, bdata, chain->bytes); 1342 1343 /* 1344 * The flush code doesn't calculate check codes for 1345 * file data (doing so can result in excessive I/O), 1346 * so we do it here. 1347 */ 1348 hammer2_chain_setcheck(chain, bdata); 1349 1350 /* 1351 * Device buffer is now valid, chain is no longer in 1352 * the initial state. 1353 * 1354 * (No blockref table worries with file data) 1355 */ 1356 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1357 hammer2_dedup_record(chain, dio, bdata); 1358 1359 if (ioflag & IO_SYNC) { 1360 /* 1361 * Synchronous I/O requested. 1362 */ 1363 hammer2_io_bwrite(&dio); 1364 /* 1365 } else if ((ioflag & IO_DIRECT) && 1366 loff + n == pblksize) { 1367 hammer2_io_bdwrite(&dio); 1368 */ 1369 } else if (ioflag & IO_ASYNC) { 1370 hammer2_io_bawrite(&dio); 1371 } else { 1372 hammer2_io_bdwrite(&dio); 1373 } 1374 break; 1375 default: 1376 panic("hammer2_write_bp: bad chain type %d\n", 1377 chain->bref.type); 1378 /* NOT REACHED */ 1379 error = 0; 1380 break; 1381 } 1382 *errorp = error; 1383 } 1384 1385 /* 1386 * LIVE DEDUP HEURISTICS 1387 * 1388 * Record media and crc information for possible dedup operation. Note 1389 * that the dedup mask bits must also be set in the related DIO for a dedup 1390 * to be fully validated (which is handled in the freemap allocation code). 1391 * 1392 * WARNING! This code is SMP safe but the heuristic allows SMP collisions. 1393 * All fields must be loaded into locals and validated. 1394 * 1395 * WARNING! Should only be used for file data and directory entries, 1396 * hammer2_chain_modify() only checks for the dedup case on data 1397 * chains. Also, dedup data can only be recorded for committed 1398 * chains (so NOT strategy writes which can undergo further 1399 * modification after the fact!). 1400 */ 1401 void 1402 hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio, 1403 const char *data) 1404 { 1405 hammer2_dev_t *hmp; 1406 hammer2_dedup_t *dedup; 1407 uint64_t crc; 1408 uint64_t mask; 1409 int best = 0; 1410 int i; 1411 int dticks; 1412 1413 /* 1414 * We can only record a dedup if we have media data to test against. 1415 * If dedup is not enabled, return early, which allows a chain to 1416 * remain marked MODIFIED (which might have benefits in special 1417 * situations, though typically it does not). 1418 */ 1419 if (hammer2_dedup_enable == 0) 1420 return; 1421 if (dio == NULL) { 1422 dio = chain->dio; 1423 if (dio == NULL) 1424 return; 1425 } 1426 1427 hmp = chain->hmp; 1428 1429 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) { 1430 case HAMMER2_CHECK_ISCSI32: 1431 /* 1432 * XXX use the built-in crc (the dedup lookup sequencing 1433 * needs to be fixed so the check code is already present 1434 * when dedup_lookup is called) 1435 */ 1436 #if 0 1437 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value; 1438 #endif 1439 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED); 1440 break; 1441 case HAMMER2_CHECK_XXHASH64: 1442 crc = chain->bref.check.xxhash64.value; 1443 break; 1444 case HAMMER2_CHECK_SHA192: 1445 /* 1446 * XXX use the built-in crc (the dedup lookup sequencing 1447 * needs to be fixed so the check code is already present 1448 * when dedup_lookup is called) 1449 */ 1450 #if 0 1451 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^ 1452 ((uint64_t *)chain->bref.check.sha192.data)[1] ^ 1453 ((uint64_t *)chain->bref.check.sha192.data)[2]; 1454 #endif 1455 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED); 1456 break; 1457 default: 1458 /* 1459 * Cannot dedup without a check code 1460 * 1461 * NOTE: In particular, CHECK_NONE allows a sector to be 1462 * overwritten without copy-on-write, recording 1463 * a dedup block for a CHECK_NONE object would be 1464 * a disaster! 1465 */ 1466 return; 1467 } 1468 1469 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUPABLE); 1470 1471 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)]; 1472 for (i = 0; i < 4; ++i) { 1473 if (dedup[i].data_crc == crc) { 1474 best = i; 1475 break; 1476 } 1477 dticks = (int)(dedup[i].ticks - dedup[best].ticks); 1478 if (dticks < 0 || dticks > hz * 60 * 30) 1479 best = i; 1480 } 1481 dedup += best; 1482 if (hammer2_debug & 0x40000) { 1483 kprintf("REC %04x %016jx %016jx\n", 1484 (int)(dedup - hmp->heur_dedup), 1485 crc, 1486 chain->bref.data_off); 1487 } 1488 dedup->ticks = ticks; 1489 dedup->data_off = chain->bref.data_off; 1490 dedup->data_crc = crc; 1491 1492 /* 1493 * Set the valid bits for the dedup only after we know the data 1494 * buffer has been updated. The alloc bits were set (and the valid 1495 * bits cleared) when the media was allocated. 1496 * 1497 * This is done in two stages becuase the bulkfree code can race 1498 * the gap between allocation and data population. Both masks must 1499 * be set before a bcmp/dedup operation is able to use the block. 1500 */ 1501 mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes); 1502 atomic_set_64(&dio->dedup_valid, mask); 1503 1504 #if 0 1505 /* 1506 * XXX removed. MODIFIED is an integral part of the flush code, 1507 * lets not just clear it 1508 */ 1509 /* 1510 * Once we record the dedup the chain must be marked clean to 1511 * prevent reuse of the underlying block. Remember that this 1512 * write occurs when the buffer cache is flushed (i.e. on sync(), 1513 * fsync(), filesystem periodic sync, or when the kernel needs to 1514 * flush a buffer), and not whenever the user write()s. 1515 */ 1516 if (chain->flags & HAMMER2_CHAIN_MODIFIED) { 1517 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED); 1518 atomic_add_long(&hammer2_count_modified_chains, -1); 1519 if (chain->pmp) 1520 hammer2_pfs_memory_wakeup(chain->pmp, -1); 1521 } 1522 #endif 1523 } 1524 1525 static 1526 hammer2_off_t 1527 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize) 1528 { 1529 hammer2_dedup_t *dedup; 1530 hammer2_io_t *dio; 1531 hammer2_off_t off; 1532 uint64_t crc; 1533 uint64_t mask; 1534 char *data; 1535 char *dtmp; 1536 int i; 1537 1538 if (hammer2_dedup_enable == 0) 1539 return 0; 1540 data = *datap; 1541 if (data == NULL) 1542 return 0; 1543 1544 /* 1545 * XXX use the built-in crc (the dedup lookup sequencing 1546 * needs to be fixed so the check code is already present 1547 * when dedup_lookup is called) 1548 */ 1549 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED); 1550 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)]; 1551 1552 if (hammer2_debug & 0x40000) { 1553 kprintf("LOC %04x/4 %016jx\n", 1554 (int)(dedup - hmp->heur_dedup), 1555 crc); 1556 } 1557 1558 for (i = 0; i < 4; ++i) { 1559 off = dedup[i].data_off; 1560 cpu_ccfence(); 1561 if (dedup[i].data_crc != crc) 1562 continue; 1563 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize) 1564 continue; 1565 dio = hammer2_io_getquick(hmp, off, pblksize); 1566 if (dio) { 1567 dtmp = hammer2_io_data(dio, off), 1568 mask = hammer2_dedup_mask(dio, off, pblksize); 1569 if ((dio->dedup_alloc & mask) == mask && 1570 (dio->dedup_valid & mask) == mask && 1571 bcmp(data, dtmp, pblksize) == 0) { 1572 if (hammer2_debug & 0x40000) { 1573 kprintf("DEDUP SUCCESS %016jx\n", 1574 (intmax_t)off); 1575 } 1576 hammer2_io_putblk(&dio); 1577 *datap = NULL; 1578 dedup[i].ticks = ticks; /* update use */ 1579 atomic_add_long(&hammer2_iod_file_wdedup, 1580 pblksize); 1581 1582 return off; /* RETURN */ 1583 } 1584 hammer2_io_putblk(&dio); 1585 } 1586 } 1587 return 0; 1588 } 1589 1590 /* 1591 * Poof. Races are ok, if someone gets in and reuses a dedup offset 1592 * before or while we are clearing it they will also recover the freemap 1593 * entry (set it to fully allocated), so a bulkfree race can only set it 1594 * to a possibly-free state. 1595 * 1596 * XXX ok, well, not really sure races are ok but going to run with it 1597 * for the moment. 1598 */ 1599 void 1600 hammer2_dedup_clear(hammer2_dev_t *hmp) 1601 { 1602 int i; 1603 1604 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) { 1605 hmp->heur_dedup[i].data_off = 0; 1606 hmp->heur_dedup[i].ticks = ticks - 1; 1607 } 1608 } 1609