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 ++hammer2_iod_file_read; 108 break; 109 case BUF_CMD_WRITE: 110 error = hammer2_strategy_write(ap); 111 ++hammer2_iod_file_write; 112 break; 113 default: 114 assert(0); 115 /* 116 bp->b_error = error = EINVAL; 117 bp->b_flags |= B_ERROR; 118 biodone(biop); 119 */ 120 break; 121 } 122 return (error); 123 } 124 125 /* 126 * Return the largest contiguous physical disk range for the logical 127 * request, in bytes. 128 * 129 * (struct m_vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb) 130 * 131 * Basically disabled, the logical buffer write thread has to deal with 132 * buffers one-at-a-time. Note that this should not prevent cluster_read() 133 * from reading-ahead, it simply prevents it from trying form a single 134 * cluster buffer for the logical request. H2 already uses 64KB buffers! 135 */ 136 int 137 hammer2_vop_bmap(struct vop_bmap_args *ap) 138 { 139 *ap->a_doffsetp = NOOFFSET; 140 if (ap->a_runp) 141 *ap->a_runp = 0; 142 if (ap->a_runb) 143 *ap->a_runb = 0; 144 return (EOPNOTSUPP); 145 } 146 147 /**************************************************************************** 148 * READ SUPPORT * 149 ****************************************************************************/ 150 /* 151 * Callback used in read path in case that a block is compressed with LZ4. 152 */ 153 static 154 void 155 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio) 156 { 157 struct m_buf *bp; 158 char *compressed_buffer; 159 int compressed_size; 160 int result; 161 162 bp = bio->bio_buf; 163 164 #if 0 165 if bio->bio_caller_info2.index && 166 bio->bio_caller_info1.uvalue32 != 167 crc32(bp->b_data, bp->b_bufsize) --- return error 168 #endif 169 170 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE); 171 compressed_size = *(const int *)data; 172 KKASSERT((uint32_t)compressed_size <= bytes - sizeof(int)); 173 174 compressed_buffer = ecalloc(1, 65536); 175 result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]), 176 compressed_buffer, 177 compressed_size, 178 bp->b_bufsize); 179 if (result < 0) { 180 kprintf("READ PATH: Error during decompression." 181 "bio %016jx/%d\n", 182 (intmax_t)bio->bio_offset, bytes); 183 /* make sure it isn't random garbage */ 184 bzero(compressed_buffer, bp->b_bufsize); 185 } 186 KKASSERT(result <= bp->b_bufsize); 187 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize); 188 if (result < bp->b_bufsize) 189 bzero(bp->b_data + result, bp->b_bufsize - result); 190 free(compressed_buffer); 191 /* 192 bp->b_resid = 0; 193 bp->b_flags |= B_AGE; 194 */ 195 } 196 197 /* 198 * Callback used in read path in case that a block is compressed with ZLIB. 199 * It is almost identical to LZ4 callback, so in theory they can be unified, 200 * but we didn't want to make changes in bio structure for that. 201 */ 202 static 203 void 204 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio) 205 { 206 struct m_buf *bp; 207 char *compressed_buffer; 208 z_stream strm_decompress; 209 int result; 210 int ret; 211 212 bp = bio->bio_buf; 213 214 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE); 215 strm_decompress.avail_in = 0; 216 strm_decompress.next_in = Z_NULL; 217 218 ret = inflateInit(&strm_decompress); 219 220 if (ret != Z_OK) 221 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n"); 222 223 compressed_buffer = ecalloc(1, 65536); 224 strm_decompress.next_in = __DECONST(char *, data); 225 226 /* XXX supply proper size, subset of device bp */ 227 strm_decompress.avail_in = bytes; 228 strm_decompress.next_out = compressed_buffer; 229 strm_decompress.avail_out = bp->b_bufsize; 230 231 ret = inflate(&strm_decompress, Z_FINISH); 232 if (ret != Z_STREAM_END) { 233 kprintf("HAMMER2 ZLIB: Fatal error during decompression.\n"); 234 bzero(compressed_buffer, bp->b_bufsize); 235 } 236 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize); 237 result = bp->b_bufsize - strm_decompress.avail_out; 238 if (result < bp->b_bufsize) 239 bzero(bp->b_data + result, strm_decompress.avail_out); 240 free(compressed_buffer); 241 ret = inflateEnd(&strm_decompress); 242 243 /* 244 bp->b_resid = 0; 245 bp->b_flags |= B_AGE; 246 */ 247 } 248 249 /* 250 * Logical buffer I/O, async read. 251 */ 252 static 253 int 254 hammer2_strategy_read(struct vop_strategy_args *ap) 255 { 256 hammer2_xop_strategy_t *xop; 257 struct bio *bio; 258 hammer2_inode_t *ip; 259 hammer2_key_t lbase; 260 261 bio = ap->a_bio; 262 ip = VTOI(ap->a_vp); 263 264 lbase = bio->bio_offset; 265 KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0); 266 267 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_STRATEGY); 268 xop->finished = 0; 269 xop->bio = bio; 270 xop->lbase = lbase; 271 hammer2_mtx_init(&xop->lock, "h2bior"); 272 hammer2_xop_start(&xop->head, &hammer2_strategy_read_desc); 273 /* asynchronous completion */ 274 275 return(0); 276 } 277 278 /* 279 * Per-node XOP (threaded), do a synchronous lookup of the chain and 280 * its data. The frontend is asynchronous, so we are also responsible 281 * for racing to terminate the frontend. 282 */ 283 void 284 hammer2_xop_strategy_read(hammer2_xop_t *arg, void *scratch, int clindex) 285 { 286 hammer2_xop_strategy_t *xop = &arg->xop_strategy; 287 hammer2_chain_t *parent; 288 hammer2_chain_t *chain; 289 hammer2_chain_t *focus; 290 hammer2_key_t key_dummy; 291 hammer2_key_t lbase; 292 struct bio *bio; 293 struct m_buf *bp; 294 const char *data; 295 int error; 296 297 /* 298 * Note that we can race completion of the bio supplied by 299 * the front-end so we cannot access it until we determine 300 * that we are the ones finishing it up. 301 */ 302 lbase = xop->lbase; 303 304 /* 305 * This is difficult to optimize. The logical buffer might be 306 * partially dirty (contain dummy zero-fill pages), which would 307 * mess up our crc calculation if we were to try a direct read. 308 * So for now we always double-buffer through the underlying 309 * storage. 310 * 311 * If not for the above problem we could conditionalize on 312 * (1) 64KB buffer, (2) one chain (not multi-master) and 313 * (3) !hammer2_double_buffer, and issue a direct read into the 314 * logical buffer. 315 */ 316 parent = hammer2_inode_chain(xop->head.ip1, clindex, 317 HAMMER2_RESOLVE_ALWAYS | 318 HAMMER2_RESOLVE_SHARED); 319 if (parent) { 320 chain = hammer2_chain_lookup(&parent, &key_dummy, 321 lbase, lbase, 322 &error, 323 HAMMER2_LOOKUP_ALWAYS | 324 HAMMER2_LOOKUP_SHARED); 325 if (chain) 326 error = chain->error; 327 } else { 328 error = HAMMER2_ERROR_EIO; 329 chain = NULL; 330 } 331 error = hammer2_xop_feed(&xop->head, chain, clindex, error); 332 if (chain) { 333 hammer2_chain_unlock(chain); 334 hammer2_chain_drop(chain); 335 } 336 if (parent) { 337 hammer2_chain_unlock(parent); 338 hammer2_chain_drop(parent); 339 } 340 chain = NULL; /* safety */ 341 parent = NULL; /* safety */ 342 343 /* 344 * Race to finish the frontend. First-to-complete. bio is only 345 * valid if we are determined to be the ones able to complete 346 * the operation. 347 */ 348 if (xop->finished) 349 return; 350 hammer2_mtx_ex(&xop->lock); 351 if (xop->finished) { 352 hammer2_mtx_unlock(&xop->lock); 353 return; 354 } 355 bio = xop->bio; 356 bp = bio->bio_buf; 357 bkvasync(bp); 358 359 /* 360 * Async operation has not completed and we now own the lock. 361 * Determine if we can complete the operation by issuing the 362 * frontend collection non-blocking. 363 * 364 * H2 double-buffers the data, setting B_NOTMETA on the logical 365 * buffer hints to the OS that the logical buffer should not be 366 * swapcached (since the device buffer can be). 367 * 368 * Also note that even for compressed data we would rather the 369 * kernel cache/swapcache device buffers more and (decompressed) 370 * logical buffers less, since that will significantly improve 371 * the amount of end-user data that can be cached. 372 * 373 * NOTE: The chain->data for xop->head.cluster.focus will be 374 * synchronized to the current cpu by xop_collect(), 375 * but other chains in the cluster might not be. 376 */ 377 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT); 378 379 switch(error) { 380 case 0: 381 xop->finished = 1; 382 hammer2_mtx_unlock(&xop->lock); 383 //bp->b_flags |= B_NOTMETA; 384 focus = xop->head.cluster.focus; 385 data = hammer2_xop_gdata(&xop->head)->buf; 386 hammer2_strategy_read_completion(focus, data, xop->bio); 387 hammer2_xop_pdata(&xop->head); 388 //biodone(bio); 389 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 390 break; 391 case HAMMER2_ERROR_ENOENT: 392 xop->finished = 1; 393 hammer2_mtx_unlock(&xop->lock); 394 /* 395 bp->b_flags |= B_NOTMETA; 396 bp->b_resid = 0; 397 bp->b_error = 0; 398 */ 399 bzero(bp->b_data, bp->b_bcount); 400 //biodone(bio); 401 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 402 break; 403 case HAMMER2_ERROR_EINPROGRESS: 404 hammer2_mtx_unlock(&xop->lock); 405 break; 406 default: 407 kprintf("xop_strategy_read: error %08x loff=%016jx\n", 408 error, (intmax_t)bp->b_loffset); 409 xop->finished = 1; 410 hammer2_mtx_unlock(&xop->lock); 411 assert(0); 412 /* 413 bp->b_flags |= B_ERROR; 414 bp->b_error = EIO; 415 biodone(bio); 416 */ 417 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 418 break; 419 } 420 } 421 422 static 423 void 424 hammer2_strategy_read_completion(hammer2_chain_t *focus, const char *data, 425 struct bio *bio) 426 { 427 struct m_buf *bp = bio->bio_buf; 428 429 if (focus->bref.type == HAMMER2_BREF_TYPE_INODE) { 430 /* 431 * Copy from in-memory inode structure. 432 */ 433 bcopy(((const hammer2_inode_data_t *)data)->u.data, 434 bp->b_data, HAMMER2_EMBEDDED_BYTES); 435 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES, 436 bp->b_bcount - HAMMER2_EMBEDDED_BYTES); 437 /* 438 bp->b_resid = 0; 439 bp->b_error = 0; 440 */ 441 } else if (focus->bref.type == HAMMER2_BREF_TYPE_DATA) { 442 /* 443 * Data is on-media, record for live dedup. Release the 444 * chain (try to free it) when done. The data is still 445 * cached by both the buffer cache in front and the 446 * block device behind us. This leaves more room in the 447 * LRU chain cache for meta-data chains which we really 448 * want to retain. 449 * 450 * NOTE: Deduplication cannot be safely recorded for 451 * records without a check code. 452 */ 453 hammer2_dedup_record(focus, NULL, data); 454 atomic_set_int(&focus->flags, HAMMER2_CHAIN_RELEASE); 455 456 /* 457 * Decompression and copy. 458 */ 459 switch (HAMMER2_DEC_COMP(focus->bref.methods)) { 460 case HAMMER2_COMP_LZ4: 461 hammer2_decompress_LZ4_callback(data, focus->bytes, 462 bio); 463 /* b_resid set by call */ 464 break; 465 case HAMMER2_COMP_ZLIB: 466 hammer2_decompress_ZLIB_callback(data, focus->bytes, 467 bio); 468 /* b_resid set by call */ 469 break; 470 case HAMMER2_COMP_NONE: 471 KKASSERT(focus->bytes <= bp->b_bcount); 472 bcopy(data, bp->b_data, focus->bytes); 473 if (focus->bytes < bp->b_bcount) { 474 bzero(bp->b_data + focus->bytes, 475 bp->b_bcount - focus->bytes); 476 } 477 /* 478 bp->b_resid = 0; 479 bp->b_error = 0; 480 */ 481 break; 482 default: 483 panic("hammer2_strategy_read_completion: " 484 "unknown compression type"); 485 } 486 } else { 487 panic("hammer2_strategy_read_completion: unknown bref type"); 488 } 489 } 490 491 /**************************************************************************** 492 * WRITE SUPPORT * 493 ****************************************************************************/ 494 495 /* 496 * Functions for compression in threads, 497 * from hammer2_vnops.c 498 */ 499 static void hammer2_write_file_core(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 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip, 504 hammer2_chain_t **parentp, 505 hammer2_key_t lbase, int ioflag, int pblksize, 506 hammer2_tid_t mtid, int *errorp, 507 int comp_algo, int check_algo); 508 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip, 509 hammer2_chain_t **parentp, 510 hammer2_key_t lbase, int ioflag, int pblksize, 511 hammer2_tid_t mtid, int *errorp, 512 int check_algo); 513 static int test_block_zeros(const char *buf, size_t bytes); 514 static void zero_write(char *data, hammer2_inode_t *ip, 515 hammer2_chain_t **parentp, 516 hammer2_key_t lbase, 517 hammer2_tid_t mtid, int *errorp); 518 static void hammer2_write_bp(hammer2_chain_t *chain, char *data, 519 int ioflag, int pblksize, 520 hammer2_tid_t mtid, int *errorp, 521 int check_algo); 522 523 int 524 hammer2_strategy_write(struct vop_strategy_args *ap) 525 { 526 hammer2_xop_strategy_t *xop; 527 hammer2_pfs_t *pmp; 528 struct bio *bio; 529 hammer2_inode_t *ip; 530 531 bio = ap->a_bio; 532 ip = VTOI(ap->a_vp); 533 pmp = ip->pmp; 534 535 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 536 hammer2_lwinprog_ref(pmp); 537 hammer2_trans_assert_strategy(pmp); 538 hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE); 539 540 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 541 HAMMER2_XOP_STRATEGY); 542 xop->finished = 0; 543 xop->bio = bio; 544 xop->lbase = bio->bio_offset; 545 hammer2_mtx_init(&xop->lock, "h2biow"); 546 hammer2_xop_start(&xop->head, &hammer2_strategy_write_desc); 547 /* asynchronous completion */ 548 549 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe); 550 551 return(0); 552 } 553 554 /* 555 * Per-node XOP (threaded). Write the logical buffer to the media. 556 * 557 * This is a bit problematic because there may be multiple target and 558 * any of them may be able to release the bp. In addition, if our 559 * particulr target is offline we don't want to block the bp (and thus 560 * the frontend). To accomplish this we copy the data to the per-thr 561 * scratch buffer. 562 */ 563 void 564 hammer2_xop_strategy_write(hammer2_xop_t *arg, void *scratch, int clindex) 565 { 566 hammer2_xop_strategy_t *xop = &arg->xop_strategy; 567 hammer2_chain_t *parent; 568 hammer2_key_t lbase; 569 hammer2_inode_t *ip; 570 struct bio *bio; 571 struct m_buf *bp; 572 int error; 573 int lblksize; 574 int pblksize; 575 char *bio_data; 576 577 /* 578 * We can only access the bp/bio if the frontend has not yet 579 * completed. 580 */ 581 if (xop->finished) 582 return; 583 hammer2_mtx_sh(&xop->lock); 584 if (xop->finished) { 585 hammer2_mtx_unlock(&xop->lock); 586 return; 587 } 588 589 lbase = xop->lbase; 590 bio = xop->bio; /* ephermal */ 591 bp = bio->bio_buf; /* ephermal */ 592 ip = xop->head.ip1; /* retained by ref */ 593 bio_data = scratch; 594 595 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */ 596 597 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL); 598 pblksize = hammer2_calc_physical(ip, lbase); 599 bkvasync(bp); 600 KKASSERT(lblksize <= MAXPHYS); 601 bcopy(bp->b_data, bio_data, lblksize); 602 603 hammer2_mtx_unlock(&xop->lock); 604 bp = NULL; /* safety, illegal to access after unlock */ 605 bio = NULL; /* safety, illegal to access after unlock */ 606 607 /* 608 * Actual operation 609 */ 610 parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS); 611 hammer2_write_file_core(bio_data, ip, &parent, 612 lbase, IO_ASYNC, pblksize, 613 xop->head.mtid, &error); 614 if (parent) { 615 hammer2_chain_unlock(parent); 616 hammer2_chain_drop(parent); 617 parent = NULL; /* safety */ 618 } 619 hammer2_xop_feed(&xop->head, NULL, clindex, error); 620 621 /* 622 * Try to complete the operation on behalf of the front-end. 623 */ 624 if (xop->finished) 625 return; 626 hammer2_mtx_ex(&xop->lock); 627 if (xop->finished) { 628 hammer2_mtx_unlock(&xop->lock); 629 return; 630 } 631 632 /* 633 * Async operation has not completed and we now own the lock. 634 * Determine if we can complete the operation by issuing the 635 * frontend collection non-blocking. 636 * 637 * H2 double-buffers the data, setting B_NOTMETA on the logical 638 * buffer hints to the OS that the logical buffer should not be 639 * swapcached (since the device buffer can be). 640 */ 641 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT); 642 643 if (error == HAMMER2_ERROR_EINPROGRESS) { 644 hammer2_mtx_unlock(&xop->lock); 645 return; 646 } 647 648 /* 649 * Async operation has completed. 650 */ 651 xop->finished = 1; 652 hammer2_mtx_unlock(&xop->lock); 653 654 bio = xop->bio; /* now owned by us */ 655 bp = bio->bio_buf; /* now owned by us */ 656 657 if (error == HAMMER2_ERROR_ENOENT || error == 0) { 658 /* 659 bp->b_flags |= B_NOTMETA; 660 bp->b_resid = 0; 661 bp->b_error = 0; 662 biodone(bio); 663 */ 664 } else { 665 kprintf("xop_strategy_write: error %d loff=%016jx\n", 666 error, (intmax_t)bp->b_loffset); 667 assert(0); 668 /* 669 bp->b_flags |= B_ERROR; 670 bp->b_error = EIO; 671 biodone(bio); 672 */ 673 } 674 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 675 hammer2_trans_assert_strategy(ip->pmp); 676 hammer2_lwinprog_drop(ip->pmp); 677 hammer2_trans_done(ip->pmp, HAMMER2_TRANS_BUFCACHE); 678 } 679 680 /* 681 * Wait for pending I/O to complete 682 */ 683 void 684 hammer2_bioq_sync(hammer2_pfs_t *pmp) 685 { 686 hammer2_lwinprog_wait(pmp, 0); 687 } 688 689 /* 690 * Assign physical storage at (cparent, lbase), returning a suitable chain 691 * and setting *errorp appropriately. 692 * 693 * If no error occurs, the returned chain will be in a modified state. 694 * 695 * If an error occurs, the returned chain may or may not be NULL. If 696 * not-null any chain->error (if not 0) will also be rolled up into *errorp. 697 * So the caller only needs to test *errorp. 698 * 699 * cparent can wind up being anything. 700 * 701 * If datap is not NULL, *datap points to the real data we intend to write. 702 * If we can dedup the storage location we set *datap to NULL to indicate 703 * to the caller that a dedup occurred. 704 * 705 * NOTE: Special case for data embedded in inode. 706 */ 707 static 708 hammer2_chain_t * 709 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp, 710 hammer2_key_t lbase, int pblksize, 711 hammer2_tid_t mtid, char **datap, int *errorp) 712 { 713 hammer2_chain_t *chain; 714 hammer2_key_t key_dummy; 715 hammer2_off_t dedup_off; 716 int pradix = hammer2_getradix(pblksize); 717 718 /* 719 * Locate the chain associated with lbase, return a locked chain. 720 * However, do not instantiate any data reference (which utilizes a 721 * device buffer) because we will be using direct IO via the 722 * logical buffer cache buffer. 723 */ 724 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN); 725 726 chain = hammer2_chain_lookup(parentp, &key_dummy, 727 lbase, lbase, 728 errorp, 729 HAMMER2_LOOKUP_NODATA); 730 731 /* 732 * The lookup code should not return a DELETED chain to us, unless 733 * its a short-file embedded in the inode. Then it is possible for 734 * the lookup to return a deleted inode. 735 */ 736 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) && 737 chain->bref.type != HAMMER2_BREF_TYPE_INODE) { 738 kprintf("assign physical deleted chain @ " 739 "%016jx (%016jx.%02x) ip %016jx\n", 740 lbase, chain->bref.data_off, chain->bref.type, 741 ip->meta.inum); 742 Debugger("bleh"); 743 } 744 745 if (chain == NULL) { 746 /* 747 * We found a hole, create a new chain entry. 748 * 749 * NOTE: DATA chains are created without device backing 750 * store (nor do we want any). 751 */ 752 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap, 753 pblksize); 754 *errorp |= hammer2_chain_create(parentp, &chain, NULL, ip->pmp, 755 HAMMER2_ENC_CHECK(ip->meta.check_algo) | 756 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE), 757 lbase, HAMMER2_PBUFRADIX, 758 HAMMER2_BREF_TYPE_DATA, 759 pblksize, mtid, 760 dedup_off, 0); 761 if (chain == NULL) 762 goto failed; 763 /*ip->delta_dcount += pblksize;*/ 764 } else if (chain->error == 0) { 765 switch (chain->bref.type) { 766 case HAMMER2_BREF_TYPE_INODE: 767 /* 768 * The data is embedded in the inode, which requires 769 * a bit more finess. 770 */ 771 *errorp |= hammer2_chain_modify_ip(ip, chain, mtid, 0); 772 break; 773 case HAMMER2_BREF_TYPE_DATA: 774 dedup_off = hammer2_dedup_lookup(chain->hmp, datap, 775 pblksize); 776 if (chain->bytes != pblksize) { 777 *errorp |= hammer2_chain_resize(chain, 778 mtid, dedup_off, 779 pradix, 780 HAMMER2_MODIFY_OPTDATA); 781 if (*errorp) 782 break; 783 } 784 785 /* 786 * DATA buffers must be marked modified whether the 787 * data is in a logical buffer or not. We also have 788 * to make this call to fixup the chain data pointers 789 * after resizing in case this is an encrypted or 790 * compressed buffer. 791 */ 792 *errorp |= hammer2_chain_modify(chain, mtid, dedup_off, 793 HAMMER2_MODIFY_OPTDATA); 794 break; 795 default: 796 panic("hammer2_assign_physical: bad type"); 797 /* NOT REACHED */ 798 break; 799 } 800 } else { 801 *errorp = chain->error; 802 } 803 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 804 failed: 805 return (chain); 806 } 807 808 /* 809 * hammer2_write_file_core() 810 * 811 * The core write function which determines which path to take 812 * depending on compression settings. We also have to locate the 813 * related chains so we can calculate and set the check data for 814 * the blockref. 815 */ 816 static 817 void 818 hammer2_write_file_core(char *data, hammer2_inode_t *ip, 819 hammer2_chain_t **parentp, 820 hammer2_key_t lbase, int ioflag, int pblksize, 821 hammer2_tid_t mtid, int *errorp) 822 { 823 hammer2_chain_t *chain; 824 char *bdata; 825 826 *errorp = 0; 827 828 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) { 829 case HAMMER2_COMP_NONE: 830 /* 831 * We have to assign physical storage to the buffer 832 * we intend to dirty or write now to avoid deadlocks 833 * in the strategy code later. 834 * 835 * This can return NOOFFSET for inode-embedded data. 836 * The strategy code will take care of it in that case. 837 */ 838 bdata = data; 839 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize, 840 mtid, &bdata, errorp); 841 if (*errorp) { 842 /* skip modifications */ 843 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 844 hammer2_inode_data_t *wipdata; 845 846 wipdata = &chain->data->ipdata; 847 KKASSERT(wipdata->meta.op_flags & 848 HAMMER2_OPFLAG_DIRECTDATA); 849 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 850 ++hammer2_iod_file_wembed; 851 } else if (bdata == NULL) { 852 /* 853 * Copy of data already present on-media. 854 */ 855 chain->bref.methods = 856 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 857 HAMMER2_ENC_CHECK(ip->meta.check_algo); 858 hammer2_chain_setcheck(chain, data); 859 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 860 } else { 861 hammer2_write_bp(chain, data, ioflag, pblksize, 862 mtid, errorp, ip->meta.check_algo); 863 } 864 if (chain) { 865 hammer2_chain_unlock(chain); 866 hammer2_chain_drop(chain); 867 } 868 break; 869 case HAMMER2_COMP_AUTOZERO: 870 /* 871 * Check for zero-fill only 872 */ 873 hammer2_zero_check_and_write(data, ip, parentp, 874 lbase, ioflag, pblksize, 875 mtid, errorp, 876 ip->meta.check_algo); 877 break; 878 case HAMMER2_COMP_LZ4: 879 case HAMMER2_COMP_ZLIB: 880 default: 881 /* 882 * Check for zero-fill and attempt compression. 883 */ 884 hammer2_compress_and_write(data, ip, parentp, 885 lbase, ioflag, pblksize, 886 mtid, errorp, 887 ip->meta.comp_algo, 888 ip->meta.check_algo); 889 break; 890 } 891 } 892 893 /* 894 * Helper 895 * 896 * Generic function that will perform the compression in compression 897 * write path. The compression algorithm is determined by the settings 898 * obtained from inode. 899 */ 900 static 901 void 902 hammer2_compress_and_write(char *data, hammer2_inode_t *ip, 903 hammer2_chain_t **parentp, 904 hammer2_key_t lbase, int ioflag, int pblksize, 905 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo) 906 { 907 hammer2_chain_t *chain; 908 int comp_size; 909 int comp_block_size; 910 char *comp_buffer; 911 char *bdata; 912 913 /* 914 * An all-zeros write creates a hole unless the check code 915 * is disabled. When the check code is disabled all writes 916 * are done in-place, including any all-zeros writes. 917 * 918 * NOTE: A snapshot will still force a copy-on-write 919 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c). 920 */ 921 if (check_algo != HAMMER2_CHECK_NONE && 922 test_block_zeros(data, pblksize)) { 923 zero_write(data, ip, parentp, lbase, mtid, errorp); 924 return; 925 } 926 927 /* 928 * Compression requested. Try to compress the block. We store 929 * the data normally if we cannot sufficiently compress it. 930 * 931 * We have a heuristic to detect files which are mostly 932 * uncompressable and avoid the compression attempt in that 933 * case. If the compression heuristic is turned off, we always 934 * try to compress. 935 */ 936 comp_size = 0; 937 comp_buffer = NULL; 938 939 KKASSERT(pblksize / 2 <= 32768); 940 941 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 || 942 hammer2_always_compress) { 943 z_stream strm_compress; 944 int comp_level; 945 int ret; 946 947 switch(HAMMER2_DEC_ALGO(comp_algo)) { 948 case HAMMER2_COMP_LZ4: 949 /* 950 * We need to prefix with the size, LZ4 951 * doesn't do it for us. Add the related 952 * overhead. 953 * 954 * NOTE: The LZ4 code seems to assume at least an 955 * 8-byte buffer size granularity and may 956 * overrun the buffer if given a 4-byte 957 * granularity. 958 */ 959 comp_buffer = ecalloc(1, 32768); 960 comp_size = LZ4_compress_limitedOutput( 961 data, 962 &comp_buffer[sizeof(int)], 963 pblksize, 964 pblksize / 2 - sizeof(int64_t)); 965 *(int *)comp_buffer = comp_size; 966 if (comp_size) 967 comp_size += sizeof(int); 968 break; 969 case HAMMER2_COMP_ZLIB: 970 comp_level = HAMMER2_DEC_LEVEL(comp_algo); 971 if (comp_level == 0) 972 comp_level = 6; /* default zlib compression */ 973 else if (comp_level < 6) 974 comp_level = 6; 975 else if (comp_level > 9) 976 comp_level = 9; 977 ret = deflateInit(&strm_compress, comp_level); 978 if (ret != Z_OK) { 979 kprintf("HAMMER2 ZLIB: fatal error " 980 "on deflateInit.\n"); 981 } 982 983 comp_buffer = ecalloc(1, 32768); 984 strm_compress.next_in = data; 985 strm_compress.avail_in = pblksize; 986 strm_compress.next_out = comp_buffer; 987 strm_compress.avail_out = pblksize / 2; 988 ret = deflate(&strm_compress, Z_FINISH); 989 if (ret == Z_STREAM_END) { 990 comp_size = pblksize / 2 - 991 strm_compress.avail_out; 992 } else { 993 comp_size = 0; 994 } 995 ret = deflateEnd(&strm_compress); 996 break; 997 default: 998 kprintf("Error: Unknown compression method.\n"); 999 kprintf("Comp_method = %d.\n", comp_algo); 1000 break; 1001 } 1002 } 1003 1004 if (comp_size == 0) { 1005 /* 1006 * compression failed or turned off 1007 */ 1008 comp_block_size = pblksize; /* safety */ 1009 if (++ip->comp_heuristic > 128) 1010 ip->comp_heuristic = 8; 1011 } else { 1012 /* 1013 * compression succeeded 1014 */ 1015 ip->comp_heuristic = 0; 1016 if (comp_size <= 1024) { 1017 comp_block_size = 1024; 1018 } else if (comp_size <= 2048) { 1019 comp_block_size = 2048; 1020 } else if (comp_size <= 4096) { 1021 comp_block_size = 4096; 1022 } else if (comp_size <= 8192) { 1023 comp_block_size = 8192; 1024 } else if (comp_size <= 16384) { 1025 comp_block_size = 16384; 1026 } else if (comp_size <= 32768) { 1027 comp_block_size = 32768; 1028 } else { 1029 panic("hammer2: WRITE PATH: " 1030 "Weird comp_size value."); 1031 /* NOT REACHED */ 1032 comp_block_size = pblksize; 1033 } 1034 1035 /* 1036 * Must zero the remainder or dedup (which operates on a 1037 * physical block basis) will not find matches. 1038 */ 1039 if (comp_size < comp_block_size) { 1040 bzero(comp_buffer + comp_size, 1041 comp_block_size - comp_size); 1042 } 1043 } 1044 1045 /* 1046 * Assign physical storage, bdata will be set to NULL if a live-dedup 1047 * was successful. 1048 */ 1049 bdata = comp_size ? comp_buffer : data; 1050 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size, 1051 mtid, &bdata, errorp); 1052 1053 if (*errorp) { 1054 goto done; 1055 } 1056 1057 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 1058 hammer2_inode_data_t *wipdata; 1059 1060 *errorp = hammer2_chain_modify_ip(ip, chain, mtid, 0); 1061 if (*errorp == 0) { 1062 wipdata = &chain->data->ipdata; 1063 KKASSERT(wipdata->meta.op_flags & 1064 HAMMER2_OPFLAG_DIRECTDATA); 1065 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1066 ++hammer2_iod_file_wembed; 1067 } 1068 } else if (bdata == NULL) { 1069 /* 1070 * Live deduplication, a copy of the data is already present 1071 * on the media. 1072 */ 1073 if (comp_size) { 1074 chain->bref.methods = 1075 HAMMER2_ENC_COMP(comp_algo) + 1076 HAMMER2_ENC_CHECK(check_algo); 1077 } else { 1078 chain->bref.methods = 1079 HAMMER2_ENC_COMP( 1080 HAMMER2_COMP_NONE) + 1081 HAMMER2_ENC_CHECK(check_algo); 1082 } 1083 bdata = comp_size ? comp_buffer : data; 1084 hammer2_chain_setcheck(chain, bdata); 1085 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1086 } else { 1087 hammer2_io_t *dio; 1088 1089 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED); 1090 1091 switch(chain->bref.type) { 1092 case HAMMER2_BREF_TYPE_INODE: 1093 panic("hammer2_compress_and_write: unexpected inode\n"); 1094 break; 1095 case HAMMER2_BREF_TYPE_DATA: 1096 /* 1097 * Optimize out the read-before-write 1098 * if possible. 1099 */ 1100 *errorp = hammer2_io_newnz(chain->hmp, 1101 chain->bref.type, 1102 chain->bref.data_off, 1103 chain->bytes, 1104 &dio); 1105 if (*errorp) { 1106 hammer2_io_brelse(&dio); 1107 kprintf("hammer2: WRITE PATH: " 1108 "dbp bread error\n"); 1109 break; 1110 } 1111 bdata = hammer2_io_data(dio, chain->bref.data_off); 1112 1113 /* 1114 * When loading the block make sure we don't 1115 * leave garbage after the compressed data. 1116 */ 1117 if (comp_size) { 1118 chain->bref.methods = 1119 HAMMER2_ENC_COMP(comp_algo) + 1120 HAMMER2_ENC_CHECK(check_algo); 1121 bcopy(comp_buffer, bdata, comp_block_size); 1122 } else { 1123 chain->bref.methods = 1124 HAMMER2_ENC_COMP( 1125 HAMMER2_COMP_NONE) + 1126 HAMMER2_ENC_CHECK(check_algo); 1127 bcopy(data, bdata, pblksize); 1128 } 1129 1130 /* 1131 * The flush code doesn't calculate check codes for 1132 * file data (doing so can result in excessive I/O), 1133 * so we do it here. 1134 */ 1135 hammer2_chain_setcheck(chain, bdata); 1136 1137 /* 1138 * Device buffer is now valid, chain is no longer in 1139 * the initial state. 1140 * 1141 * (No blockref table worries with file data) 1142 */ 1143 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1144 hammer2_dedup_record(chain, dio, bdata); 1145 1146 /* Now write the related bdp. */ 1147 if (ioflag & IO_SYNC) { 1148 /* 1149 * Synchronous I/O requested. 1150 */ 1151 hammer2_io_bwrite(&dio); 1152 /* 1153 } else if ((ioflag & IO_DIRECT) && 1154 loff + n == pblksize) { 1155 hammer2_io_bdwrite(&dio); 1156 */ 1157 } else if (ioflag & IO_ASYNC) { 1158 hammer2_io_bawrite(&dio); 1159 } else { 1160 hammer2_io_bdwrite(&dio); 1161 } 1162 break; 1163 default: 1164 panic("hammer2_compress_and_write: bad chain type %d\n", 1165 chain->bref.type); 1166 /* NOT REACHED */ 1167 break; 1168 } 1169 } 1170 done: 1171 if (chain) { 1172 hammer2_chain_unlock(chain); 1173 hammer2_chain_drop(chain); 1174 } 1175 if (comp_buffer) 1176 free(comp_buffer); 1177 } 1178 1179 /* 1180 * Helper 1181 * 1182 * Function that performs zero-checking and writing without compression, 1183 * it corresponds to default zero-checking path. 1184 */ 1185 static 1186 void 1187 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip, 1188 hammer2_chain_t **parentp, 1189 hammer2_key_t lbase, int ioflag, int pblksize, 1190 hammer2_tid_t mtid, int *errorp, 1191 int check_algo) 1192 { 1193 hammer2_chain_t *chain; 1194 char *bdata; 1195 1196 if (check_algo != HAMMER2_CHECK_NONE && 1197 test_block_zeros(data, pblksize)) { 1198 /* 1199 * An all-zeros write creates a hole unless the check code 1200 * is disabled. When the check code is disabled all writes 1201 * are done in-place, including any all-zeros writes. 1202 * 1203 * NOTE: A snapshot will still force a copy-on-write 1204 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c). 1205 */ 1206 zero_write(data, ip, parentp, lbase, mtid, errorp); 1207 } else { 1208 /* 1209 * Normal write (bdata set to NULL if de-duplicated) 1210 */ 1211 bdata = data; 1212 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize, 1213 mtid, &bdata, errorp); 1214 if (*errorp) { 1215 /* do nothing */ 1216 } else if (bdata) { 1217 hammer2_write_bp(chain, data, ioflag, pblksize, 1218 mtid, errorp, check_algo); 1219 } else { 1220 /* dedup occurred */ 1221 chain->bref.methods = 1222 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 1223 HAMMER2_ENC_CHECK(check_algo); 1224 hammer2_chain_setcheck(chain, data); 1225 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1226 } 1227 if (chain) { 1228 hammer2_chain_unlock(chain); 1229 hammer2_chain_drop(chain); 1230 } 1231 } 1232 } 1233 1234 /* 1235 * Helper 1236 * 1237 * A function to test whether a block of data contains only zeros, 1238 * returns TRUE (non-zero) if the block is all zeros. 1239 */ 1240 static 1241 int 1242 test_block_zeros(const char *buf, size_t bytes) 1243 { 1244 size_t i; 1245 1246 for (i = 0; i < bytes; i += sizeof(long)) { 1247 if (*(const long *)(buf + i) != 0) 1248 return (0); 1249 } 1250 return (1); 1251 } 1252 1253 /* 1254 * Helper 1255 * 1256 * Function to "write" a block that contains only zeros. 1257 */ 1258 static 1259 void 1260 zero_write(char *data, hammer2_inode_t *ip, 1261 hammer2_chain_t **parentp, 1262 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp) 1263 { 1264 hammer2_chain_t *chain; 1265 hammer2_key_t key_dummy; 1266 1267 chain = hammer2_chain_lookup(parentp, &key_dummy, 1268 lbase, lbase, 1269 errorp, 1270 HAMMER2_LOOKUP_NODATA); 1271 if (chain) { 1272 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 1273 hammer2_inode_data_t *wipdata; 1274 1275 if (*errorp == 0) { 1276 *errorp = hammer2_chain_modify_ip(ip, chain, 1277 mtid, 0); 1278 } 1279 if (*errorp == 0) { 1280 wipdata = &chain->data->ipdata; 1281 KKASSERT(wipdata->meta.op_flags & 1282 HAMMER2_OPFLAG_DIRECTDATA); 1283 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1284 ++hammer2_iod_file_wembed; 1285 } 1286 } else { 1287 /* chain->error ok for deletion */ 1288 hammer2_chain_delete(*parentp, chain, 1289 mtid, HAMMER2_DELETE_PERMANENT); 1290 ++hammer2_iod_file_wzero; 1291 } 1292 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 1293 hammer2_chain_unlock(chain); 1294 hammer2_chain_drop(chain); 1295 } else { 1296 ++hammer2_iod_file_wzero; 1297 } 1298 } 1299 1300 /* 1301 * Helper 1302 * 1303 * Function to write the data as it is, without performing any sort of 1304 * compression. This function is used in path without compression and 1305 * default zero-checking path. 1306 */ 1307 static 1308 void 1309 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag, 1310 int pblksize, 1311 hammer2_tid_t mtid, int *errorp, int check_algo) 1312 { 1313 hammer2_inode_data_t *wipdata; 1314 hammer2_io_t *dio; 1315 char *bdata; 1316 int error; 1317 1318 error = 0; /* XXX TODO below */ 1319 1320 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED); 1321 1322 switch(chain->bref.type) { 1323 case HAMMER2_BREF_TYPE_INODE: 1324 wipdata = &chain->data->ipdata; 1325 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA); 1326 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1327 error = 0; 1328 ++hammer2_iod_file_wembed; 1329 break; 1330 case HAMMER2_BREF_TYPE_DATA: 1331 error = hammer2_io_newnz(chain->hmp, 1332 chain->bref.type, 1333 chain->bref.data_off, 1334 chain->bytes, &dio); 1335 if (error) { 1336 hammer2_io_bqrelse(&dio); 1337 kprintf("hammer2: WRITE PATH: " 1338 "dbp bread error\n"); 1339 break; 1340 } 1341 bdata = hammer2_io_data(dio, chain->bref.data_off); 1342 1343 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 1344 HAMMER2_ENC_CHECK(check_algo); 1345 bcopy(data, bdata, chain->bytes); 1346 1347 /* 1348 * The flush code doesn't calculate check codes for 1349 * file data (doing so can result in excessive I/O), 1350 * so we do it here. 1351 */ 1352 hammer2_chain_setcheck(chain, bdata); 1353 1354 /* 1355 * Device buffer is now valid, chain is no longer in 1356 * the initial state. 1357 * 1358 * (No blockref table worries with file data) 1359 */ 1360 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1361 hammer2_dedup_record(chain, dio, bdata); 1362 1363 if (ioflag & IO_SYNC) { 1364 /* 1365 * Synchronous I/O requested. 1366 */ 1367 hammer2_io_bwrite(&dio); 1368 /* 1369 } else if ((ioflag & IO_DIRECT) && 1370 loff + n == pblksize) { 1371 hammer2_io_bdwrite(&dio); 1372 */ 1373 } else if (ioflag & IO_ASYNC) { 1374 hammer2_io_bawrite(&dio); 1375 } else { 1376 hammer2_io_bdwrite(&dio); 1377 } 1378 break; 1379 default: 1380 panic("hammer2_write_bp: bad chain type %d\n", 1381 chain->bref.type); 1382 /* NOT REACHED */ 1383 error = 0; 1384 break; 1385 } 1386 *errorp = error; 1387 } 1388 1389 /* 1390 * LIVE DEDUP HEURISTICS 1391 * 1392 * Record media and crc information for possible dedup operation. Note 1393 * that the dedup mask bits must also be set in the related DIO for a dedup 1394 * to be fully validated (which is handled in the freemap allocation code). 1395 * 1396 * WARNING! This code is SMP safe but the heuristic allows SMP collisions. 1397 * All fields must be loaded into locals and validated. 1398 * 1399 * WARNING! Should only be used for file data and directory entries, 1400 * hammer2_chain_modify() only checks for the dedup case on data 1401 * chains. Also, dedup data can only be recorded for committed 1402 * chains (so NOT strategy writes which can undergo further 1403 * modification after the fact!). 1404 */ 1405 void 1406 hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio, 1407 const char *data) 1408 { 1409 hammer2_dev_t *hmp; 1410 hammer2_dedup_t *dedup; 1411 uint64_t crc; 1412 uint64_t mask; 1413 int best = 0; 1414 int i; 1415 int dticks; 1416 1417 /* 1418 * We can only record a dedup if we have media data to test against. 1419 * If dedup is not enabled, return early, which allows a chain to 1420 * remain marked MODIFIED (which might have benefits in special 1421 * situations, though typically it does not). 1422 */ 1423 if (hammer2_dedup_enable == 0) 1424 return; 1425 if (dio == NULL) { 1426 dio = chain->dio; 1427 if (dio == NULL) 1428 return; 1429 } 1430 1431 hmp = chain->hmp; 1432 1433 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) { 1434 case HAMMER2_CHECK_ISCSI32: 1435 /* 1436 * XXX use the built-in crc (the dedup lookup sequencing 1437 * needs to be fixed so the check code is already present 1438 * when dedup_lookup is called) 1439 */ 1440 #if 0 1441 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value; 1442 #endif 1443 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED); 1444 break; 1445 case HAMMER2_CHECK_XXHASH64: 1446 crc = chain->bref.check.xxhash64.value; 1447 break; 1448 case HAMMER2_CHECK_SHA192: 1449 /* 1450 * XXX use the built-in crc (the dedup lookup sequencing 1451 * needs to be fixed so the check code is already present 1452 * when dedup_lookup is called) 1453 */ 1454 #if 0 1455 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^ 1456 ((uint64_t *)chain->bref.check.sha192.data)[1] ^ 1457 ((uint64_t *)chain->bref.check.sha192.data)[2]; 1458 #endif 1459 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED); 1460 break; 1461 default: 1462 /* 1463 * Cannot dedup without a check code 1464 * 1465 * NOTE: In particular, CHECK_NONE allows a sector to be 1466 * overwritten without copy-on-write, recording 1467 * a dedup block for a CHECK_NONE object would be 1468 * a disaster! 1469 */ 1470 return; 1471 } 1472 1473 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUPABLE); 1474 1475 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)]; 1476 for (i = 0; i < 4; ++i) { 1477 if (dedup[i].data_crc == crc) { 1478 best = i; 1479 break; 1480 } 1481 dticks = (int)(dedup[i].ticks - dedup[best].ticks); 1482 if (dticks < 0 || dticks > hz * 60 * 30) 1483 best = i; 1484 } 1485 dedup += best; 1486 if (hammer2_debug & 0x40000) { 1487 kprintf("REC %04x %016jx %016jx\n", 1488 (int)(dedup - hmp->heur_dedup), 1489 crc, 1490 chain->bref.data_off); 1491 } 1492 dedup->ticks = ticks; 1493 dedup->data_off = chain->bref.data_off; 1494 dedup->data_crc = crc; 1495 1496 /* 1497 * Set the valid bits for the dedup only after we know the data 1498 * buffer has been updated. The alloc bits were set (and the valid 1499 * bits cleared) when the media was allocated. 1500 * 1501 * This is done in two stages becuase the bulkfree code can race 1502 * the gap between allocation and data population. Both masks must 1503 * be set before a bcmp/dedup operation is able to use the block. 1504 */ 1505 mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes); 1506 atomic_set_64(&dio->dedup_valid, mask); 1507 1508 #if 0 1509 /* 1510 * XXX removed. MODIFIED is an integral part of the flush code, 1511 * lets not just clear it 1512 */ 1513 /* 1514 * Once we record the dedup the chain must be marked clean to 1515 * prevent reuse of the underlying block. Remember that this 1516 * write occurs when the buffer cache is flushed (i.e. on sync(), 1517 * fsync(), filesystem periodic sync, or when the kernel needs to 1518 * flush a buffer), and not whenever the user write()s. 1519 */ 1520 if (chain->flags & HAMMER2_CHAIN_MODIFIED) { 1521 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED); 1522 atomic_add_long(&hammer2_count_modified_chains, -1); 1523 if (chain->pmp) 1524 hammer2_pfs_memory_wakeup(chain->pmp, -1); 1525 } 1526 #endif 1527 } 1528 1529 static 1530 hammer2_off_t 1531 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize) 1532 { 1533 hammer2_dedup_t *dedup; 1534 hammer2_io_t *dio; 1535 hammer2_off_t off; 1536 uint64_t crc; 1537 uint64_t mask; 1538 char *data; 1539 char *dtmp; 1540 int i; 1541 1542 if (hammer2_dedup_enable == 0) 1543 return 0; 1544 data = *datap; 1545 if (data == NULL) 1546 return 0; 1547 1548 /* 1549 * XXX use the built-in crc (the dedup lookup sequencing 1550 * needs to be fixed so the check code is already present 1551 * when dedup_lookup is called) 1552 */ 1553 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED); 1554 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)]; 1555 1556 if (hammer2_debug & 0x40000) { 1557 kprintf("LOC %04x/4 %016jx\n", 1558 (int)(dedup - hmp->heur_dedup), 1559 crc); 1560 } 1561 1562 for (i = 0; i < 4; ++i) { 1563 off = dedup[i].data_off; 1564 cpu_ccfence(); 1565 if (dedup[i].data_crc != crc) 1566 continue; 1567 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize) 1568 continue; 1569 dio = hammer2_io_getquick(hmp, off, pblksize); 1570 if (dio) { 1571 dtmp = hammer2_io_data(dio, off), 1572 mask = hammer2_dedup_mask(dio, off, pblksize); 1573 if ((dio->dedup_alloc & mask) == mask && 1574 (dio->dedup_valid & mask) == mask && 1575 bcmp(data, dtmp, pblksize) == 0) { 1576 if (hammer2_debug & 0x40000) { 1577 kprintf("DEDUP SUCCESS %016jx\n", 1578 (intmax_t)off); 1579 } 1580 hammer2_io_putblk(&dio); 1581 *datap = NULL; 1582 dedup[i].ticks = ticks; /* update use */ 1583 atomic_add_long(&hammer2_iod_file_wdedup, 1584 pblksize); 1585 1586 return off; /* RETURN */ 1587 } 1588 hammer2_io_putblk(&dio); 1589 } 1590 } 1591 return 0; 1592 } 1593 1594 /* 1595 * Poof. Races are ok, if someone gets in and reuses a dedup offset 1596 * before or while we are clearing it they will also recover the freemap 1597 * entry (set it to fully allocated), so a bulkfree race can only set it 1598 * to a possibly-free state. 1599 * 1600 * XXX ok, well, not really sure races are ok but going to run with it 1601 * for the moment. 1602 */ 1603 void 1604 hammer2_dedup_clear(hammer2_dev_t *hmp) 1605 { 1606 int i; 1607 1608 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) { 1609 hmp->heur_dedup[i].data_off = 0; 1610 hmp->heur_dedup[i].ticks = ticks - 1; 1611 } 1612 } 1613