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. This leaves more room in the 445 * LRU chain cache for meta-data chains which we really 446 * want to retain. 447 * 448 * NOTE: Deduplication cannot be safely recorded for 449 * records without a check code. 450 */ 451 hammer2_dedup_record(focus, NULL, data); 452 atomic_set_int(&focus->flags, HAMMER2_CHAIN_RELEASE); 453 454 /* 455 * Decompression and copy. 456 */ 457 switch (HAMMER2_DEC_COMP(focus->bref.methods)) { 458 case HAMMER2_COMP_LZ4: 459 hammer2_decompress_LZ4_callback(data, focus->bytes, 460 bio); 461 /* b_resid set by call */ 462 break; 463 case HAMMER2_COMP_ZLIB: 464 hammer2_decompress_ZLIB_callback(data, focus->bytes, 465 bio); 466 /* b_resid set by call */ 467 break; 468 case HAMMER2_COMP_NONE: 469 KKASSERT(focus->bytes <= bp->b_bcount); 470 bcopy(data, bp->b_data, focus->bytes); 471 if (focus->bytes < bp->b_bcount) { 472 bzero(bp->b_data + focus->bytes, 473 bp->b_bcount - focus->bytes); 474 } 475 /* 476 bp->b_resid = 0; 477 bp->b_error = 0; 478 */ 479 break; 480 default: 481 panic("hammer2_strategy_read_completion: " 482 "unknown compression type"); 483 } 484 } else { 485 panic("hammer2_strategy_read_completion: unknown bref type"); 486 } 487 } 488 489 /**************************************************************************** 490 * WRITE SUPPORT * 491 ****************************************************************************/ 492 493 /* 494 * Functions for compression in threads, 495 * from hammer2_vnops.c 496 */ 497 static void hammer2_write_file_core(char *data, hammer2_inode_t *ip, 498 hammer2_chain_t **parentp, 499 hammer2_key_t lbase, int ioflag, int pblksize, 500 hammer2_tid_t mtid, int *errorp); 501 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip, 502 hammer2_chain_t **parentp, 503 hammer2_key_t lbase, int ioflag, int pblksize, 504 hammer2_tid_t mtid, int *errorp, 505 int comp_algo, int check_algo); 506 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip, 507 hammer2_chain_t **parentp, 508 hammer2_key_t lbase, int ioflag, int pblksize, 509 hammer2_tid_t mtid, int *errorp, 510 int check_algo); 511 static int test_block_zeros(const char *buf, size_t bytes); 512 static void zero_write(char *data, hammer2_inode_t *ip, 513 hammer2_chain_t **parentp, 514 hammer2_key_t lbase, 515 hammer2_tid_t mtid, int *errorp); 516 static void hammer2_write_bp(hammer2_chain_t *chain, char *data, 517 int ioflag, int pblksize, 518 hammer2_tid_t mtid, int *errorp, 519 int check_algo); 520 521 int 522 hammer2_strategy_write(struct vop_strategy_args *ap) 523 { 524 hammer2_xop_strategy_t *xop; 525 hammer2_pfs_t *pmp; 526 struct bio *bio; 527 hammer2_inode_t *ip; 528 529 bio = ap->a_bio; 530 ip = VTOI(ap->a_vp); 531 pmp = ip->pmp; 532 533 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 534 hammer2_lwinprog_ref(pmp); 535 hammer2_trans_assert_strategy(pmp); 536 hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE); 537 538 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 539 HAMMER2_XOP_STRATEGY); 540 xop->finished = 0; 541 xop->bio = bio; 542 xop->lbase = bio->bio_offset; 543 hammer2_mtx_init(&xop->lock, "h2biow"); 544 hammer2_xop_start(&xop->head, &hammer2_strategy_write_desc); 545 /* asynchronous completion */ 546 547 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe); 548 549 return(0); 550 } 551 552 /* 553 * Per-node XOP (threaded). Write the logical buffer to the media. 554 * 555 * This is a bit problematic because there may be multiple target and 556 * any of them may be able to release the bp. In addition, if our 557 * particulr target is offline we don't want to block the bp (and thus 558 * the frontend). To accomplish this we copy the data to the per-thr 559 * scratch buffer. 560 */ 561 void 562 hammer2_xop_strategy_write(hammer2_xop_t *arg, void *scratch, int clindex) 563 { 564 hammer2_xop_strategy_t *xop = &arg->xop_strategy; 565 hammer2_chain_t *parent; 566 hammer2_key_t lbase; 567 hammer2_inode_t *ip; 568 struct bio *bio; 569 struct m_buf *bp; 570 int error; 571 int lblksize; 572 int pblksize; 573 char *bio_data; 574 575 /* 576 * We can only access the bp/bio if the frontend has not yet 577 * completed. 578 */ 579 if (xop->finished) 580 return; 581 hammer2_mtx_sh(&xop->lock); 582 if (xop->finished) { 583 hammer2_mtx_unlock(&xop->lock); 584 return; 585 } 586 587 lbase = xop->lbase; 588 bio = xop->bio; /* ephermal */ 589 bp = bio->bio_buf; /* ephermal */ 590 ip = xop->head.ip1; /* retained by ref */ 591 bio_data = scratch; 592 593 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */ 594 595 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL); 596 pblksize = hammer2_calc_physical(ip, lbase); 597 bkvasync(bp); 598 KKASSERT(lblksize <= MAXPHYS); 599 bcopy(bp->b_data, bio_data, lblksize); 600 601 hammer2_mtx_unlock(&xop->lock); 602 bp = NULL; /* safety, illegal to access after unlock */ 603 bio = NULL; /* safety, illegal to access after unlock */ 604 605 /* 606 * Actual operation 607 */ 608 parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS); 609 hammer2_write_file_core(bio_data, ip, &parent, 610 lbase, IO_ASYNC, pblksize, 611 xop->head.mtid, &error); 612 if (parent) { 613 hammer2_chain_unlock(parent); 614 hammer2_chain_drop(parent); 615 parent = NULL; /* safety */ 616 } 617 hammer2_xop_feed(&xop->head, NULL, clindex, error); 618 619 /* 620 * Try to complete the operation on behalf of the front-end. 621 */ 622 if (xop->finished) 623 return; 624 hammer2_mtx_ex(&xop->lock); 625 if (xop->finished) { 626 hammer2_mtx_unlock(&xop->lock); 627 return; 628 } 629 630 /* 631 * Async operation has not completed and we now own the lock. 632 * Determine if we can complete the operation by issuing the 633 * frontend collection non-blocking. 634 * 635 * H2 double-buffers the data, setting B_NOTMETA on the logical 636 * buffer hints to the OS that the logical buffer should not be 637 * swapcached (since the device buffer can be). 638 */ 639 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT); 640 641 if (error == HAMMER2_ERROR_EINPROGRESS) { 642 hammer2_mtx_unlock(&xop->lock); 643 return; 644 } 645 646 /* 647 * Async operation has completed. 648 */ 649 xop->finished = 1; 650 hammer2_mtx_unlock(&xop->lock); 651 652 bio = xop->bio; /* now owned by us */ 653 bp = bio->bio_buf; /* now owned by us */ 654 655 if (error == HAMMER2_ERROR_ENOENT || error == 0) { 656 /* 657 bp->b_flags |= B_NOTMETA; 658 bp->b_resid = 0; 659 bp->b_error = 0; 660 biodone(bio); 661 */ 662 } else { 663 kprintf("xop_strategy_write: error %d loff=%016jx\n", 664 error, (intmax_t)bp->b_loffset); 665 assert(0); 666 /* 667 bp->b_flags |= B_ERROR; 668 bp->b_error = EIO; 669 biodone(bio); 670 */ 671 } 672 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 673 hammer2_trans_assert_strategy(ip->pmp); 674 hammer2_lwinprog_drop(ip->pmp); 675 hammer2_trans_done(ip->pmp, HAMMER2_TRANS_BUFCACHE); 676 } 677 678 /* 679 * Wait for pending I/O to complete 680 */ 681 void 682 hammer2_bioq_sync(hammer2_pfs_t *pmp) 683 { 684 hammer2_lwinprog_wait(pmp, 0); 685 } 686 687 /* 688 * Assign physical storage at (cparent, lbase), returning a suitable chain 689 * and setting *errorp appropriately. 690 * 691 * If no error occurs, the returned chain will be in a modified state. 692 * 693 * If an error occurs, the returned chain may or may not be NULL. If 694 * not-null any chain->error (if not 0) will also be rolled up into *errorp. 695 * So the caller only needs to test *errorp. 696 * 697 * cparent can wind up being anything. 698 * 699 * If datap is not NULL, *datap points to the real data we intend to write. 700 * If we can dedup the storage location we set *datap to NULL to indicate 701 * to the caller that a dedup occurred. 702 * 703 * NOTE: Special case for data embedded in inode. 704 */ 705 static 706 hammer2_chain_t * 707 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp, 708 hammer2_key_t lbase, int pblksize, 709 hammer2_tid_t mtid, char **datap, int *errorp) 710 { 711 hammer2_chain_t *chain; 712 hammer2_key_t key_dummy; 713 hammer2_off_t dedup_off; 714 int pradix = hammer2_getradix(pblksize); 715 716 /* 717 * Locate the chain associated with lbase, return a locked chain. 718 * However, do not instantiate any data reference (which utilizes a 719 * device buffer) because we will be using direct IO via the 720 * logical buffer cache buffer. 721 */ 722 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN); 723 724 chain = hammer2_chain_lookup(parentp, &key_dummy, 725 lbase, lbase, 726 errorp, 727 HAMMER2_LOOKUP_NODATA); 728 729 /* 730 * The lookup code should not return a DELETED chain to us, unless 731 * its a short-file embedded in the inode. Then it is possible for 732 * the lookup to return a deleted inode. 733 */ 734 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) && 735 chain->bref.type != HAMMER2_BREF_TYPE_INODE) { 736 kprintf("assign physical deleted chain @ " 737 "%016jx (%016jx.%02x) ip %016jx\n", 738 lbase, chain->bref.data_off, chain->bref.type, 739 ip->meta.inum); 740 Debugger("bleh"); 741 } 742 743 if (chain == NULL) { 744 /* 745 * We found a hole, create a new chain entry. 746 * 747 * NOTE: DATA chains are created without device backing 748 * store (nor do we want any). 749 */ 750 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap, 751 pblksize); 752 *errorp |= hammer2_chain_create(parentp, &chain, NULL, ip->pmp, 753 HAMMER2_ENC_CHECK(ip->meta.check_algo) | 754 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE), 755 lbase, HAMMER2_PBUFRADIX, 756 HAMMER2_BREF_TYPE_DATA, 757 pblksize, mtid, 758 dedup_off, 0); 759 if (chain == NULL) 760 goto failed; 761 /*ip->delta_dcount += pblksize;*/ 762 } else if (chain->error == 0) { 763 switch (chain->bref.type) { 764 case HAMMER2_BREF_TYPE_INODE: 765 /* 766 * The data is embedded in the inode, which requires 767 * a bit more finess. 768 */ 769 *errorp |= hammer2_chain_modify_ip(ip, chain, mtid, 0); 770 break; 771 case HAMMER2_BREF_TYPE_DATA: 772 dedup_off = hammer2_dedup_lookup(chain->hmp, datap, 773 pblksize); 774 if (chain->bytes != pblksize) { 775 *errorp |= hammer2_chain_resize(chain, 776 mtid, dedup_off, 777 pradix, 778 HAMMER2_MODIFY_OPTDATA); 779 if (*errorp) 780 break; 781 } 782 783 /* 784 * DATA buffers must be marked modified whether the 785 * data is in a logical buffer or not. We also have 786 * to make this call to fixup the chain data pointers 787 * after resizing in case this is an encrypted or 788 * compressed buffer. 789 */ 790 *errorp |= hammer2_chain_modify(chain, mtid, dedup_off, 791 HAMMER2_MODIFY_OPTDATA); 792 break; 793 default: 794 panic("hammer2_assign_physical: bad type"); 795 /* NOT REACHED */ 796 break; 797 } 798 } else { 799 *errorp = chain->error; 800 } 801 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 802 failed: 803 return (chain); 804 } 805 806 /* 807 * hammer2_write_file_core() 808 * 809 * The core write function which determines which path to take 810 * depending on compression settings. We also have to locate the 811 * related chains so we can calculate and set the check data for 812 * the blockref. 813 */ 814 static 815 void 816 hammer2_write_file_core(char *data, hammer2_inode_t *ip, 817 hammer2_chain_t **parentp, 818 hammer2_key_t lbase, int ioflag, int pblksize, 819 hammer2_tid_t mtid, int *errorp) 820 { 821 hammer2_chain_t *chain; 822 char *bdata; 823 824 *errorp = 0; 825 826 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) { 827 case HAMMER2_COMP_NONE: 828 /* 829 * We have to assign physical storage to the buffer 830 * we intend to dirty or write now to avoid deadlocks 831 * in the strategy code later. 832 * 833 * This can return NOOFFSET for inode-embedded data. 834 * The strategy code will take care of it in that case. 835 */ 836 bdata = data; 837 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize, 838 mtid, &bdata, errorp); 839 if (*errorp) { 840 /* skip modifications */ 841 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 842 hammer2_inode_data_t *wipdata; 843 844 wipdata = &chain->data->ipdata; 845 KKASSERT(wipdata->meta.op_flags & 846 HAMMER2_OPFLAG_DIRECTDATA); 847 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 848 ++hammer2_iod_file_wembed; 849 } else if (bdata == NULL) { 850 /* 851 * Copy of data already present on-media. 852 */ 853 chain->bref.methods = 854 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 855 HAMMER2_ENC_CHECK(ip->meta.check_algo); 856 hammer2_chain_setcheck(chain, data); 857 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 858 } else { 859 hammer2_write_bp(chain, data, ioflag, pblksize, 860 mtid, errorp, ip->meta.check_algo); 861 } 862 if (chain) { 863 hammer2_chain_unlock(chain); 864 hammer2_chain_drop(chain); 865 } 866 break; 867 case HAMMER2_COMP_AUTOZERO: 868 /* 869 * Check for zero-fill only 870 */ 871 hammer2_zero_check_and_write(data, ip, parentp, 872 lbase, ioflag, pblksize, 873 mtid, errorp, 874 ip->meta.check_algo); 875 break; 876 case HAMMER2_COMP_LZ4: 877 case HAMMER2_COMP_ZLIB: 878 default: 879 /* 880 * Check for zero-fill and attempt compression. 881 */ 882 hammer2_compress_and_write(data, ip, parentp, 883 lbase, ioflag, pblksize, 884 mtid, errorp, 885 ip->meta.comp_algo, 886 ip->meta.check_algo); 887 break; 888 } 889 } 890 891 /* 892 * Helper 893 * 894 * Generic function that will perform the compression in compression 895 * write path. The compression algorithm is determined by the settings 896 * obtained from inode. 897 */ 898 static 899 void 900 hammer2_compress_and_write(char *data, hammer2_inode_t *ip, 901 hammer2_chain_t **parentp, 902 hammer2_key_t lbase, int ioflag, int pblksize, 903 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo) 904 { 905 hammer2_chain_t *chain; 906 int comp_size; 907 int comp_block_size; 908 char *comp_buffer; 909 char *bdata; 910 911 /* 912 * An all-zeros write creates a hole unless the check code 913 * is disabled. When the check code is disabled all writes 914 * are done in-place, including any all-zeros writes. 915 * 916 * NOTE: A snapshot will still force a copy-on-write 917 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c). 918 */ 919 if (check_algo != HAMMER2_CHECK_NONE && 920 test_block_zeros(data, pblksize)) { 921 zero_write(data, ip, parentp, lbase, mtid, errorp); 922 return; 923 } 924 925 /* 926 * Compression requested. Try to compress the block. We store 927 * the data normally if we cannot sufficiently compress it. 928 * 929 * We have a heuristic to detect files which are mostly 930 * uncompressable and avoid the compression attempt in that 931 * case. If the compression heuristic is turned off, we always 932 * try to compress. 933 */ 934 comp_size = 0; 935 comp_buffer = NULL; 936 937 KKASSERT(pblksize / 2 <= 32768); 938 939 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 || 940 hammer2_always_compress) { 941 z_stream strm_compress; 942 int comp_level; 943 int ret; 944 945 switch(HAMMER2_DEC_ALGO(comp_algo)) { 946 case HAMMER2_COMP_LZ4: 947 /* 948 * We need to prefix with the size, LZ4 949 * doesn't do it for us. Add the related 950 * overhead. 951 * 952 * NOTE: The LZ4 code seems to assume at least an 953 * 8-byte buffer size granularity and may 954 * overrun the buffer if given a 4-byte 955 * granularity. 956 */ 957 comp_buffer = ecalloc(1, 32768); 958 comp_size = LZ4_compress_limitedOutput( 959 data, 960 &comp_buffer[sizeof(int)], 961 pblksize, 962 pblksize / 2 - sizeof(int64_t)); 963 *(int *)comp_buffer = comp_size; 964 if (comp_size) 965 comp_size += sizeof(int); 966 break; 967 case HAMMER2_COMP_ZLIB: 968 comp_level = HAMMER2_DEC_LEVEL(comp_algo); 969 if (comp_level == 0) 970 comp_level = 6; /* default zlib compression */ 971 else if (comp_level < 6) 972 comp_level = 6; 973 else if (comp_level > 9) 974 comp_level = 9; 975 ret = deflateInit(&strm_compress, comp_level); 976 if (ret != Z_OK) { 977 kprintf("HAMMER2 ZLIB: fatal error " 978 "on deflateInit.\n"); 979 } 980 981 comp_buffer = ecalloc(1, 32768); 982 strm_compress.next_in = data; 983 strm_compress.avail_in = pblksize; 984 strm_compress.next_out = comp_buffer; 985 strm_compress.avail_out = pblksize / 2; 986 ret = deflate(&strm_compress, Z_FINISH); 987 if (ret == Z_STREAM_END) { 988 comp_size = pblksize / 2 - 989 strm_compress.avail_out; 990 } else { 991 comp_size = 0; 992 } 993 ret = deflateEnd(&strm_compress); 994 break; 995 default: 996 kprintf("Error: Unknown compression method.\n"); 997 kprintf("Comp_method = %d.\n", comp_algo); 998 break; 999 } 1000 } 1001 1002 if (comp_size == 0) { 1003 /* 1004 * compression failed or turned off 1005 */ 1006 comp_block_size = pblksize; /* safety */ 1007 if (++ip->comp_heuristic > 128) 1008 ip->comp_heuristic = 8; 1009 } else { 1010 /* 1011 * compression succeeded 1012 */ 1013 ip->comp_heuristic = 0; 1014 if (comp_size <= 1024) { 1015 comp_block_size = 1024; 1016 } else if (comp_size <= 2048) { 1017 comp_block_size = 2048; 1018 } else if (comp_size <= 4096) { 1019 comp_block_size = 4096; 1020 } else if (comp_size <= 8192) { 1021 comp_block_size = 8192; 1022 } else if (comp_size <= 16384) { 1023 comp_block_size = 16384; 1024 } else if (comp_size <= 32768) { 1025 comp_block_size = 32768; 1026 } else { 1027 panic("hammer2: WRITE PATH: " 1028 "Weird comp_size value."); 1029 /* NOT REACHED */ 1030 comp_block_size = pblksize; 1031 } 1032 1033 /* 1034 * Must zero the remainder or dedup (which operates on a 1035 * physical block basis) will not find matches. 1036 */ 1037 if (comp_size < comp_block_size) { 1038 bzero(comp_buffer + comp_size, 1039 comp_block_size - comp_size); 1040 } 1041 } 1042 1043 /* 1044 * Assign physical storage, bdata will be set to NULL if a live-dedup 1045 * was successful. 1046 */ 1047 bdata = comp_size ? comp_buffer : data; 1048 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size, 1049 mtid, &bdata, errorp); 1050 1051 if (*errorp) { 1052 goto done; 1053 } 1054 1055 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 1056 hammer2_inode_data_t *wipdata; 1057 1058 *errorp = hammer2_chain_modify_ip(ip, chain, mtid, 0); 1059 if (*errorp == 0) { 1060 wipdata = &chain->data->ipdata; 1061 KKASSERT(wipdata->meta.op_flags & 1062 HAMMER2_OPFLAG_DIRECTDATA); 1063 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1064 ++hammer2_iod_file_wembed; 1065 } 1066 } else if (bdata == NULL) { 1067 /* 1068 * Live deduplication, a copy of the data is already present 1069 * on the media. 1070 */ 1071 if (comp_size) { 1072 chain->bref.methods = 1073 HAMMER2_ENC_COMP(comp_algo) + 1074 HAMMER2_ENC_CHECK(check_algo); 1075 } else { 1076 chain->bref.methods = 1077 HAMMER2_ENC_COMP( 1078 HAMMER2_COMP_NONE) + 1079 HAMMER2_ENC_CHECK(check_algo); 1080 } 1081 bdata = comp_size ? comp_buffer : data; 1082 hammer2_chain_setcheck(chain, bdata); 1083 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1084 } else { 1085 hammer2_io_t *dio; 1086 1087 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED); 1088 1089 switch(chain->bref.type) { 1090 case HAMMER2_BREF_TYPE_INODE: 1091 panic("hammer2_compress_and_write: unexpected inode\n"); 1092 break; 1093 case HAMMER2_BREF_TYPE_DATA: 1094 /* 1095 * Optimize out the read-before-write 1096 * if possible. 1097 */ 1098 *errorp = hammer2_io_newnz(chain->hmp, 1099 chain->bref.type, 1100 chain->bref.data_off, 1101 chain->bytes, 1102 &dio); 1103 if (*errorp) { 1104 hammer2_io_brelse(&dio); 1105 kprintf("hammer2: WRITE PATH: " 1106 "dbp bread error\n"); 1107 break; 1108 } 1109 bdata = hammer2_io_data(dio, chain->bref.data_off); 1110 1111 /* 1112 * When loading the block make sure we don't 1113 * leave garbage after the compressed data. 1114 */ 1115 if (comp_size) { 1116 chain->bref.methods = 1117 HAMMER2_ENC_COMP(comp_algo) + 1118 HAMMER2_ENC_CHECK(check_algo); 1119 bcopy(comp_buffer, bdata, comp_block_size); 1120 } else { 1121 chain->bref.methods = 1122 HAMMER2_ENC_COMP( 1123 HAMMER2_COMP_NONE) + 1124 HAMMER2_ENC_CHECK(check_algo); 1125 bcopy(data, bdata, pblksize); 1126 } 1127 1128 /* 1129 * The flush code doesn't calculate check codes for 1130 * file data (doing so can result in excessive I/O), 1131 * so we do it here. 1132 */ 1133 hammer2_chain_setcheck(chain, bdata); 1134 1135 /* 1136 * Device buffer is now valid, chain is no longer in 1137 * the initial state. 1138 * 1139 * (No blockref table worries with file data) 1140 */ 1141 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1142 hammer2_dedup_record(chain, dio, bdata); 1143 1144 /* Now write the related bdp. */ 1145 if (ioflag & IO_SYNC) { 1146 /* 1147 * Synchronous I/O requested. 1148 */ 1149 hammer2_io_bwrite(&dio); 1150 /* 1151 } else if ((ioflag & IO_DIRECT) && 1152 loff + n == pblksize) { 1153 hammer2_io_bdwrite(&dio); 1154 */ 1155 } else if (ioflag & IO_ASYNC) { 1156 hammer2_io_bawrite(&dio); 1157 } else { 1158 hammer2_io_bdwrite(&dio); 1159 } 1160 break; 1161 default: 1162 panic("hammer2_compress_and_write: bad chain type %d\n", 1163 chain->bref.type); 1164 /* NOT REACHED */ 1165 break; 1166 } 1167 } 1168 done: 1169 if (chain) { 1170 hammer2_chain_unlock(chain); 1171 hammer2_chain_drop(chain); 1172 } 1173 if (comp_buffer) 1174 free(comp_buffer); 1175 } 1176 1177 /* 1178 * Helper 1179 * 1180 * Function that performs zero-checking and writing without compression, 1181 * it corresponds to default zero-checking path. 1182 */ 1183 static 1184 void 1185 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip, 1186 hammer2_chain_t **parentp, 1187 hammer2_key_t lbase, int ioflag, int pblksize, 1188 hammer2_tid_t mtid, int *errorp, 1189 int check_algo) 1190 { 1191 hammer2_chain_t *chain; 1192 char *bdata; 1193 1194 if (check_algo != HAMMER2_CHECK_NONE && 1195 test_block_zeros(data, pblksize)) { 1196 /* 1197 * An all-zeros write creates a hole unless the check code 1198 * is disabled. When the check code is disabled all writes 1199 * are done in-place, including any all-zeros writes. 1200 * 1201 * NOTE: A snapshot will still force a copy-on-write 1202 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c). 1203 */ 1204 zero_write(data, ip, parentp, lbase, mtid, errorp); 1205 } else { 1206 /* 1207 * Normal write (bdata set to NULL if de-duplicated) 1208 */ 1209 bdata = data; 1210 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize, 1211 mtid, &bdata, errorp); 1212 if (*errorp) { 1213 /* do nothing */ 1214 } else if (bdata) { 1215 hammer2_write_bp(chain, data, ioflag, pblksize, 1216 mtid, errorp, check_algo); 1217 } else { 1218 /* dedup occurred */ 1219 chain->bref.methods = 1220 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 1221 HAMMER2_ENC_CHECK(check_algo); 1222 hammer2_chain_setcheck(chain, data); 1223 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1224 } 1225 if (chain) { 1226 hammer2_chain_unlock(chain); 1227 hammer2_chain_drop(chain); 1228 } 1229 } 1230 } 1231 1232 /* 1233 * Helper 1234 * 1235 * A function to test whether a block of data contains only zeros, 1236 * returns TRUE (non-zero) if the block is all zeros. 1237 */ 1238 static 1239 int 1240 test_block_zeros(const char *buf, size_t bytes) 1241 { 1242 size_t i; 1243 1244 for (i = 0; i < bytes; i += sizeof(long)) { 1245 if (*(const long *)(buf + i) != 0) 1246 return (0); 1247 } 1248 return (1); 1249 } 1250 1251 /* 1252 * Helper 1253 * 1254 * Function to "write" a block that contains only zeros. 1255 */ 1256 static 1257 void 1258 zero_write(char *data, hammer2_inode_t *ip, 1259 hammer2_chain_t **parentp, 1260 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp) 1261 { 1262 hammer2_chain_t *chain; 1263 hammer2_key_t key_dummy; 1264 1265 chain = hammer2_chain_lookup(parentp, &key_dummy, 1266 lbase, lbase, 1267 errorp, 1268 HAMMER2_LOOKUP_NODATA); 1269 if (chain) { 1270 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 1271 hammer2_inode_data_t *wipdata; 1272 1273 if (*errorp == 0) { 1274 *errorp = hammer2_chain_modify_ip(ip, chain, 1275 mtid, 0); 1276 } 1277 if (*errorp == 0) { 1278 wipdata = &chain->data->ipdata; 1279 KKASSERT(wipdata->meta.op_flags & 1280 HAMMER2_OPFLAG_DIRECTDATA); 1281 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1282 ++hammer2_iod_file_wembed; 1283 } 1284 } else { 1285 /* chain->error ok for deletion */ 1286 hammer2_chain_delete(*parentp, chain, 1287 mtid, HAMMER2_DELETE_PERMANENT); 1288 ++hammer2_iod_file_wzero; 1289 } 1290 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 1291 hammer2_chain_unlock(chain); 1292 hammer2_chain_drop(chain); 1293 } else { 1294 ++hammer2_iod_file_wzero; 1295 } 1296 } 1297 1298 /* 1299 * Helper 1300 * 1301 * Function to write the data as it is, without performing any sort of 1302 * compression. This function is used in path without compression and 1303 * default zero-checking path. 1304 */ 1305 static 1306 void 1307 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag, 1308 int pblksize, 1309 hammer2_tid_t mtid, int *errorp, int check_algo) 1310 { 1311 hammer2_inode_data_t *wipdata; 1312 hammer2_io_t *dio; 1313 char *bdata; 1314 int error; 1315 1316 error = 0; /* XXX TODO below */ 1317 1318 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED); 1319 1320 switch(chain->bref.type) { 1321 case HAMMER2_BREF_TYPE_INODE: 1322 wipdata = &chain->data->ipdata; 1323 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA); 1324 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1325 error = 0; 1326 ++hammer2_iod_file_wembed; 1327 break; 1328 case HAMMER2_BREF_TYPE_DATA: 1329 error = hammer2_io_newnz(chain->hmp, 1330 chain->bref.type, 1331 chain->bref.data_off, 1332 chain->bytes, &dio); 1333 if (error) { 1334 hammer2_io_bqrelse(&dio); 1335 kprintf("hammer2: WRITE PATH: " 1336 "dbp bread error\n"); 1337 break; 1338 } 1339 bdata = hammer2_io_data(dio, chain->bref.data_off); 1340 1341 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 1342 HAMMER2_ENC_CHECK(check_algo); 1343 bcopy(data, bdata, chain->bytes); 1344 1345 /* 1346 * The flush code doesn't calculate check codes for 1347 * file data (doing so can result in excessive I/O), 1348 * so we do it here. 1349 */ 1350 hammer2_chain_setcheck(chain, bdata); 1351 1352 /* 1353 * Device buffer is now valid, chain is no longer in 1354 * the initial state. 1355 * 1356 * (No blockref table worries with file data) 1357 */ 1358 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1359 hammer2_dedup_record(chain, dio, bdata); 1360 1361 if (ioflag & IO_SYNC) { 1362 /* 1363 * Synchronous I/O requested. 1364 */ 1365 hammer2_io_bwrite(&dio); 1366 /* 1367 } else if ((ioflag & IO_DIRECT) && 1368 loff + n == pblksize) { 1369 hammer2_io_bdwrite(&dio); 1370 */ 1371 } else if (ioflag & IO_ASYNC) { 1372 hammer2_io_bawrite(&dio); 1373 } else { 1374 hammer2_io_bdwrite(&dio); 1375 } 1376 break; 1377 default: 1378 panic("hammer2_write_bp: bad chain type %d\n", 1379 chain->bref.type); 1380 /* NOT REACHED */ 1381 error = 0; 1382 break; 1383 } 1384 *errorp = error; 1385 } 1386 1387 /* 1388 * LIVE DEDUP HEURISTICS 1389 * 1390 * Record media and crc information for possible dedup operation. Note 1391 * that the dedup mask bits must also be set in the related DIO for a dedup 1392 * to be fully validated (which is handled in the freemap allocation code). 1393 * 1394 * WARNING! This code is SMP safe but the heuristic allows SMP collisions. 1395 * All fields must be loaded into locals and validated. 1396 * 1397 * WARNING! Should only be used for file data and directory entries, 1398 * hammer2_chain_modify() only checks for the dedup case on data 1399 * chains. Also, dedup data can only be recorded for committed 1400 * chains (so NOT strategy writes which can undergo further 1401 * modification after the fact!). 1402 */ 1403 void 1404 hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio, 1405 const char *data) 1406 { 1407 hammer2_dev_t *hmp; 1408 hammer2_dedup_t *dedup; 1409 uint64_t crc; 1410 uint64_t mask; 1411 int best = 0; 1412 int i; 1413 int dticks; 1414 1415 /* 1416 * We can only record a dedup if we have media data to test against. 1417 * If dedup is not enabled, return early, which allows a chain to 1418 * remain marked MODIFIED (which might have benefits in special 1419 * situations, though typically it does not). 1420 */ 1421 if (hammer2_dedup_enable == 0) 1422 return; 1423 if (dio == NULL) { 1424 dio = chain->dio; 1425 if (dio == NULL) 1426 return; 1427 } 1428 1429 hmp = chain->hmp; 1430 1431 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) { 1432 case HAMMER2_CHECK_ISCSI32: 1433 /* 1434 * XXX use the built-in crc (the dedup lookup sequencing 1435 * needs to be fixed so the check code is already present 1436 * when dedup_lookup is called) 1437 */ 1438 #if 0 1439 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value; 1440 #endif 1441 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED); 1442 break; 1443 case HAMMER2_CHECK_XXHASH64: 1444 crc = chain->bref.check.xxhash64.value; 1445 break; 1446 case HAMMER2_CHECK_SHA192: 1447 /* 1448 * XXX use the built-in crc (the dedup lookup sequencing 1449 * needs to be fixed so the check code is already present 1450 * when dedup_lookup is called) 1451 */ 1452 #if 0 1453 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^ 1454 ((uint64_t *)chain->bref.check.sha192.data)[1] ^ 1455 ((uint64_t *)chain->bref.check.sha192.data)[2]; 1456 #endif 1457 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED); 1458 break; 1459 default: 1460 /* 1461 * Cannot dedup without a check code 1462 * 1463 * NOTE: In particular, CHECK_NONE allows a sector to be 1464 * overwritten without copy-on-write, recording 1465 * a dedup block for a CHECK_NONE object would be 1466 * a disaster! 1467 */ 1468 return; 1469 } 1470 1471 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUPABLE); 1472 1473 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)]; 1474 for (i = 0; i < 4; ++i) { 1475 if (dedup[i].data_crc == crc) { 1476 best = i; 1477 break; 1478 } 1479 dticks = (int)(dedup[i].ticks - dedup[best].ticks); 1480 if (dticks < 0 || dticks > hz * 60 * 30) 1481 best = i; 1482 } 1483 dedup += best; 1484 if (hammer2_debug & 0x40000) { 1485 kprintf("REC %04x %016jx %016jx\n", 1486 (int)(dedup - hmp->heur_dedup), 1487 crc, 1488 chain->bref.data_off); 1489 } 1490 dedup->ticks = ticks; 1491 dedup->data_off = chain->bref.data_off; 1492 dedup->data_crc = crc; 1493 1494 /* 1495 * Set the valid bits for the dedup only after we know the data 1496 * buffer has been updated. The alloc bits were set (and the valid 1497 * bits cleared) when the media was allocated. 1498 * 1499 * This is done in two stages becuase the bulkfree code can race 1500 * the gap between allocation and data population. Both masks must 1501 * be set before a bcmp/dedup operation is able to use the block. 1502 */ 1503 mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes); 1504 atomic_set_64(&dio->dedup_valid, mask); 1505 1506 #if 0 1507 /* 1508 * XXX removed. MODIFIED is an integral part of the flush code, 1509 * lets not just clear it 1510 */ 1511 /* 1512 * Once we record the dedup the chain must be marked clean to 1513 * prevent reuse of the underlying block. Remember that this 1514 * write occurs when the buffer cache is flushed (i.e. on sync(), 1515 * fsync(), filesystem periodic sync, or when the kernel needs to 1516 * flush a buffer), and not whenever the user write()s. 1517 */ 1518 if (chain->flags & HAMMER2_CHAIN_MODIFIED) { 1519 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED); 1520 atomic_add_long(&hammer2_count_modified_chains, -1); 1521 if (chain->pmp) 1522 hammer2_pfs_memory_wakeup(chain->pmp, -1); 1523 } 1524 #endif 1525 } 1526 1527 static 1528 hammer2_off_t 1529 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize) 1530 { 1531 hammer2_dedup_t *dedup; 1532 hammer2_io_t *dio; 1533 hammer2_off_t off; 1534 uint64_t crc; 1535 uint64_t mask; 1536 char *data; 1537 char *dtmp; 1538 int i; 1539 1540 if (hammer2_dedup_enable == 0) 1541 return 0; 1542 data = *datap; 1543 if (data == NULL) 1544 return 0; 1545 1546 /* 1547 * XXX use the built-in crc (the dedup lookup sequencing 1548 * needs to be fixed so the check code is already present 1549 * when dedup_lookup is called) 1550 */ 1551 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED); 1552 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)]; 1553 1554 if (hammer2_debug & 0x40000) { 1555 kprintf("LOC %04x/4 %016jx\n", 1556 (int)(dedup - hmp->heur_dedup), 1557 crc); 1558 } 1559 1560 for (i = 0; i < 4; ++i) { 1561 off = dedup[i].data_off; 1562 cpu_ccfence(); 1563 if (dedup[i].data_crc != crc) 1564 continue; 1565 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize) 1566 continue; 1567 dio = hammer2_io_getquick(hmp, off, pblksize); 1568 if (dio) { 1569 dtmp = hammer2_io_data(dio, off), 1570 mask = hammer2_dedup_mask(dio, off, pblksize); 1571 if ((dio->dedup_alloc & mask) == mask && 1572 (dio->dedup_valid & mask) == mask && 1573 bcmp(data, dtmp, pblksize) == 0) { 1574 if (hammer2_debug & 0x40000) { 1575 kprintf("DEDUP SUCCESS %016jx\n", 1576 (intmax_t)off); 1577 } 1578 hammer2_io_putblk(&dio); 1579 *datap = NULL; 1580 dedup[i].ticks = ticks; /* update use */ 1581 atomic_add_long(&hammer2_iod_file_wdedup, 1582 pblksize); 1583 1584 return off; /* RETURN */ 1585 } 1586 hammer2_io_putblk(&dio); 1587 } 1588 } 1589 return 0; 1590 } 1591 1592 /* 1593 * Poof. Races are ok, if someone gets in and reuses a dedup offset 1594 * before or while we are clearing it they will also recover the freemap 1595 * entry (set it to fully allocated), so a bulkfree race can only set it 1596 * to a possibly-free state. 1597 * 1598 * XXX ok, well, not really sure races are ok but going to run with it 1599 * for the moment. 1600 */ 1601 void 1602 hammer2_dedup_clear(hammer2_dev_t *hmp) 1603 { 1604 int i; 1605 1606 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) { 1607 hmp->heur_dedup[i].data_off = 0; 1608 hmp->heur_dedup[i].ticks = ticks - 1; 1609 } 1610 } 1611