1 /* 2 * Copyright (c) 2011-2018 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@dragonflybsd.org> 6 * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org> 7 * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression) 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in 17 * the documentation and/or other materials provided with the 18 * distribution. 19 * 3. Neither the name of The DragonFly Project nor the names of its 20 * contributors may be used to endorse or promote products derived 21 * from this software without specific, prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 */ 36 /* 37 * This module handles low level logical file I/O (strategy) which backs 38 * the logical buffer cache. 39 * 40 * [De]compression, zero-block, check codes, and buffer cache operations 41 * for file data is handled here. 42 * 43 * Live dedup makes its home here as well. 44 */ 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/kernel.h> 49 #include <sys/buf.h> 50 #include <sys/proc.h> 51 #include <sys/mount.h> 52 #include <sys/vnode.h> 53 #include <sys/objcache.h> 54 55 #include "hammer2.h" 56 #include "hammer2_lz4.h" 57 58 #include "zlib/hammer2_zlib.h" 59 60 struct objcache *cache_buffer_read; 61 struct objcache *cache_buffer_write; 62 63 /* 64 * Strategy code (async logical file buffer I/O from system) 65 * 66 * Except for the transaction init (which should normally not block), 67 * we essentially run the strategy operation asynchronously via a XOP. 68 * 69 * WARNING! The XOP deals with buffer synchronization. It is not synchronized 70 * to the current cpu. 71 * 72 * XXX This isn't supposed to be able to deadlock against vfs_sync vfsync() 73 * calls but it has in the past when multiple flushes are queued. 74 * 75 * XXX We currently terminate the transaction once we get a quorum, otherwise 76 * the frontend can stall, but this can leave the remaining nodes with 77 * a potential flush conflict. We need to delay flushes on those nodes 78 * until running transactions complete separately from the normal 79 * transaction sequencing. FIXME TODO. 80 */ 81 static int hammer2_strategy_read(struct vop_strategy_args *ap); 82 static int hammer2_strategy_write(struct vop_strategy_args *ap); 83 static void hammer2_strategy_read_completion(hammer2_chain_t *focus, 84 const char *data, struct bio *bio); 85 86 static hammer2_off_t hammer2_dedup_lookup(hammer2_dev_t *hmp, 87 char **datap, int pblksize); 88 89 int 90 hammer2_vop_strategy(struct vop_strategy_args *ap) 91 { 92 struct bio *biop; 93 struct buf *bp; 94 int error; 95 96 biop = ap->a_bio; 97 bp = biop->bio_buf; 98 99 switch(bp->b_cmd) { 100 case BUF_CMD_READ: 101 error = hammer2_strategy_read(ap); 102 ++hammer2_iod_file_read; 103 break; 104 case BUF_CMD_WRITE: 105 error = hammer2_strategy_write(ap); 106 ++hammer2_iod_file_write; 107 break; 108 default: 109 bp->b_error = error = EINVAL; 110 bp->b_flags |= B_ERROR; 111 biodone(biop); 112 break; 113 } 114 return (error); 115 } 116 117 /* 118 * Return the largest contiguous physical disk range for the logical 119 * request, in bytes. 120 * 121 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb) 122 * 123 * Basically disabled, the logical buffer write thread has to deal with 124 * buffers one-at-a-time. Note that this should not prevent cluster_read() 125 * from reading-ahead, it simply prevents it from trying form a single 126 * cluster buffer for the logical request. H2 already uses 64KB buffers! 127 */ 128 int 129 hammer2_vop_bmap(struct vop_bmap_args *ap) 130 { 131 *ap->a_doffsetp = NOOFFSET; 132 if (ap->a_runp) 133 *ap->a_runp = 0; 134 if (ap->a_runb) 135 *ap->a_runb = 0; 136 return (EOPNOTSUPP); 137 } 138 139 /**************************************************************************** 140 * READ SUPPORT * 141 ****************************************************************************/ 142 /* 143 * Callback used in read path in case that a block is compressed with LZ4. 144 */ 145 static 146 void 147 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio) 148 { 149 struct buf *bp; 150 char *compressed_buffer; 151 int compressed_size; 152 int result; 153 154 bp = bio->bio_buf; 155 156 #if 0 157 if bio->bio_caller_info2.index && 158 bio->bio_caller_info1.uvalue32 != 159 crc32(bp->b_data, bp->b_bufsize) --- return error 160 #endif 161 162 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE); 163 compressed_size = *(const int *)data; 164 KKASSERT((uint32_t)compressed_size <= bytes - sizeof(int)); 165 166 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT); 167 result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]), 168 compressed_buffer, 169 compressed_size, 170 bp->b_bufsize); 171 if (result < 0) { 172 kprintf("READ PATH: Error during decompression." 173 "bio %016jx/%d\n", 174 (intmax_t)bio->bio_offset, bytes); 175 /* make sure it isn't random garbage */ 176 bzero(compressed_buffer, bp->b_bufsize); 177 } 178 KKASSERT(result <= bp->b_bufsize); 179 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize); 180 if (result < bp->b_bufsize) 181 bzero((char *)bp->b_data + result, bp->b_bufsize - result); 182 objcache_put(cache_buffer_read, compressed_buffer); 183 bp->b_resid = 0; 184 bp->b_flags |= B_AGE; 185 } 186 187 /* 188 * Callback used in read path in case that a block is compressed with ZLIB. 189 * It is almost identical to LZ4 callback, so in theory they can be unified, 190 * but we didn't want to make changes in bio structure for that. 191 */ 192 static 193 void 194 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio) 195 { 196 struct buf *bp; 197 char *compressed_buffer; 198 z_stream strm_decompress; 199 int result; 200 int ret; 201 202 bp = bio->bio_buf; 203 204 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE); 205 strm_decompress.avail_in = 0; 206 strm_decompress.next_in = Z_NULL; 207 208 ret = inflateInit(&strm_decompress); 209 210 if (ret != Z_OK) 211 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n"); 212 213 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT); 214 strm_decompress.next_in = __DECONST(char *, data); 215 216 /* XXX supply proper size, subset of device bp */ 217 strm_decompress.avail_in = bytes; 218 strm_decompress.next_out = compressed_buffer; 219 strm_decompress.avail_out = bp->b_bufsize; 220 221 ret = inflate(&strm_decompress, Z_FINISH); 222 if (ret != Z_STREAM_END) { 223 kprintf("HAMMER2 ZLIB: Fatal error during decompression.\n"); 224 bzero(compressed_buffer, bp->b_bufsize); 225 } 226 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize); 227 result = bp->b_bufsize - strm_decompress.avail_out; 228 if (result < bp->b_bufsize) 229 bzero((char *)bp->b_data + result, strm_decompress.avail_out); 230 objcache_put(cache_buffer_read, compressed_buffer); 231 ret = inflateEnd(&strm_decompress); 232 233 bp->b_resid = 0; 234 bp->b_flags |= B_AGE; 235 } 236 237 /* 238 * Logical buffer I/O, async read. 239 */ 240 static 241 int 242 hammer2_strategy_read(struct vop_strategy_args *ap) 243 { 244 hammer2_xop_strategy_t *xop; 245 struct bio *bio; 246 hammer2_inode_t *ip; 247 hammer2_key_t lbase; 248 249 bio = ap->a_bio; 250 ip = VTOI(ap->a_vp); 251 252 lbase = bio->bio_offset; 253 KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0); 254 255 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_STRATEGY); 256 xop->finished = 0; 257 xop->bio = bio; 258 xop->lbase = lbase; 259 hammer2_mtx_init(&xop->lock, "h2bior"); 260 hammer2_xop_start(&xop->head, &hammer2_strategy_read_desc); 261 /* asynchronous completion */ 262 263 return(0); 264 } 265 266 /* 267 * Per-node XOP (threaded), do a synchronous lookup of the chain and 268 * its data. The frontend is asynchronous, so we are also responsible 269 * for racing to terminate the frontend. 270 */ 271 void 272 hammer2_xop_strategy_read(hammer2_xop_t *arg, void *scratch, int clindex) 273 { 274 hammer2_xop_strategy_t *xop = &arg->xop_strategy; 275 hammer2_chain_t *parent; 276 hammer2_chain_t *chain; 277 hammer2_chain_t *focus; 278 hammer2_key_t key_dummy; 279 hammer2_key_t lbase; 280 struct bio *bio; 281 struct buf *bp; 282 const char *data; 283 int error; 284 285 /* 286 * Note that we can race completion of the bio supplied by 287 * the front-end so we cannot access it until we determine 288 * that we are the ones finishing it up. 289 */ 290 lbase = xop->lbase; 291 292 /* 293 * This is difficult to optimize. The logical buffer might be 294 * partially dirty (contain dummy zero-fill pages), which would 295 * mess up our crc calculation if we were to try a direct read. 296 * So for now we always double-buffer through the underlying 297 * storage. 298 * 299 * If not for the above problem we could conditionalize on 300 * (1) 64KB buffer, (2) one chain (not multi-master) and 301 * (3) !hammer2_double_buffer, and issue a direct read into the 302 * logical buffer. 303 */ 304 parent = hammer2_inode_chain(xop->head.ip1, clindex, 305 HAMMER2_RESOLVE_ALWAYS | 306 HAMMER2_RESOLVE_SHARED); 307 if (parent) { 308 chain = hammer2_chain_lookup(&parent, &key_dummy, 309 lbase, lbase, 310 &error, 311 HAMMER2_LOOKUP_ALWAYS | 312 HAMMER2_LOOKUP_SHARED); 313 if (chain) 314 error = chain->error; 315 } else { 316 error = HAMMER2_ERROR_EIO; 317 chain = NULL; 318 } 319 error = hammer2_xop_feed(&xop->head, chain, clindex, error); 320 if (chain) { 321 hammer2_chain_unlock(chain); 322 hammer2_chain_drop(chain); 323 } 324 if (parent) { 325 hammer2_chain_unlock(parent); 326 hammer2_chain_drop(parent); 327 } 328 chain = NULL; /* safety */ 329 parent = NULL; /* safety */ 330 331 /* 332 * Race to finish the frontend. First-to-complete. bio is only 333 * valid if we are determined to be the ones able to complete 334 * the operation. 335 */ 336 if (xop->finished) 337 return; 338 hammer2_mtx_ex(&xop->lock); 339 if (xop->finished) { 340 hammer2_mtx_unlock(&xop->lock); 341 return; 342 } 343 bio = xop->bio; 344 bp = bio->bio_buf; 345 bkvasync(bp); 346 347 /* 348 * Async operation has not completed and we now own the lock. 349 * Determine if we can complete the operation by issuing the 350 * frontend collection non-blocking. 351 * 352 * H2 double-buffers the data, setting B_NOTMETA on the logical 353 * buffer hints to the OS that the logical buffer should not be 354 * swapcached (since the device buffer can be). 355 * 356 * Also note that even for compressed data we would rather the 357 * kernel cache/swapcache device buffers more and (decompressed) 358 * logical buffers less, since that will significantly improve 359 * the amount of end-user data that can be cached. 360 * 361 * NOTE: The chain->data for xop->head.cluster.focus will be 362 * synchronized to the current cpu by xop_collect(), 363 * but other chains in the cluster might not be. 364 */ 365 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT); 366 367 switch(error) { 368 case 0: 369 xop->finished = 1; 370 hammer2_mtx_unlock(&xop->lock); 371 bp->b_flags |= B_NOTMETA; 372 focus = xop->head.cluster.focus; 373 data = hammer2_xop_gdata(&xop->head)->buf; 374 hammer2_strategy_read_completion(focus, data, xop->bio); 375 hammer2_xop_pdata(&xop->head); 376 biodone(bio); 377 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 378 break; 379 case HAMMER2_ERROR_ENOENT: 380 xop->finished = 1; 381 hammer2_mtx_unlock(&xop->lock); 382 bp->b_flags |= B_NOTMETA; 383 bp->b_resid = 0; 384 bp->b_error = 0; 385 bzero(bp->b_data, bp->b_bcount); 386 biodone(bio); 387 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 388 break; 389 case HAMMER2_ERROR_EINPROGRESS: 390 hammer2_mtx_unlock(&xop->lock); 391 break; 392 default: 393 kprintf("xop_strategy_read: error %08x loff=%016jx\n", 394 error, (intmax_t)bp->b_loffset); 395 xop->finished = 1; 396 hammer2_mtx_unlock(&xop->lock); 397 bp->b_flags |= B_ERROR; 398 bp->b_error = EIO; 399 biodone(bio); 400 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 401 break; 402 } 403 } 404 405 static 406 void 407 hammer2_strategy_read_completion(hammer2_chain_t *focus, const char *data, 408 struct bio *bio) 409 { 410 struct buf *bp = bio->bio_buf; 411 412 if (focus->bref.type == HAMMER2_BREF_TYPE_INODE) { 413 /* 414 * Copy from in-memory inode structure. 415 */ 416 bcopy(((const hammer2_inode_data_t *)data)->u.data, 417 bp->b_data, HAMMER2_EMBEDDED_BYTES); 418 bzero((char *)bp->b_data + HAMMER2_EMBEDDED_BYTES, 419 bp->b_bcount - HAMMER2_EMBEDDED_BYTES); 420 bp->b_resid = 0; 421 bp->b_error = 0; 422 } else if (focus->bref.type == HAMMER2_BREF_TYPE_DATA) { 423 /* 424 * Data is on-media, record for live dedup. Release the 425 * chain (try to free it) when done. The data is still 426 * cached by both the buffer cache in front and the 427 * block device behind us. This leaves more room in the 428 * LRU chain cache for meta-data chains which we really 429 * want to retain. 430 * 431 * NOTE: Deduplication cannot be safely recorded for 432 * records without a check code. 433 */ 434 hammer2_dedup_record(focus, NULL, data); 435 atomic_set_int(&focus->flags, HAMMER2_CHAIN_RELEASE); 436 437 /* 438 * Decompression and copy. 439 */ 440 switch (HAMMER2_DEC_COMP(focus->bref.methods)) { 441 case HAMMER2_COMP_LZ4: 442 hammer2_decompress_LZ4_callback(data, focus->bytes, 443 bio); 444 /* b_resid set by call */ 445 break; 446 case HAMMER2_COMP_ZLIB: 447 hammer2_decompress_ZLIB_callback(data, focus->bytes, 448 bio); 449 /* b_resid set by call */ 450 break; 451 case HAMMER2_COMP_NONE: 452 KKASSERT(focus->bytes <= bp->b_bcount); 453 bcopy(data, bp->b_data, focus->bytes); 454 if (focus->bytes < bp->b_bcount) { 455 bzero((char *)bp->b_data + focus->bytes, 456 bp->b_bcount - focus->bytes); 457 } 458 bp->b_resid = 0; 459 bp->b_error = 0; 460 break; 461 default: 462 panic("hammer2_strategy_read_completion: " 463 "unknown compression type"); 464 } 465 } else { 466 panic("hammer2_strategy_read_completion: unknown bref type"); 467 } 468 } 469 470 /**************************************************************************** 471 * WRITE SUPPORT * 472 ****************************************************************************/ 473 474 /* 475 * Functions for compression in threads, 476 * from hammer2_vnops.c 477 */ 478 static void hammer2_write_file_core(char *data, hammer2_inode_t *ip, 479 hammer2_chain_t **parentp, 480 hammer2_key_t lbase, int ioflag, int pblksize, 481 hammer2_tid_t mtid, int *errorp); 482 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip, 483 hammer2_chain_t **parentp, 484 hammer2_key_t lbase, int ioflag, int pblksize, 485 hammer2_tid_t mtid, int *errorp, 486 int comp_algo, int check_algo); 487 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip, 488 hammer2_chain_t **parentp, 489 hammer2_key_t lbase, int ioflag, int pblksize, 490 hammer2_tid_t mtid, int *errorp, 491 int check_algo); 492 static int test_block_zeros(const char *buf, size_t bytes); 493 static void zero_write(char *data, hammer2_inode_t *ip, 494 hammer2_chain_t **parentp, 495 hammer2_key_t lbase, 496 hammer2_tid_t mtid, int *errorp); 497 static void hammer2_write_bp(hammer2_chain_t *chain, char *data, 498 int ioflag, int pblksize, 499 hammer2_tid_t mtid, int *errorp, 500 int check_algo); 501 502 int 503 hammer2_strategy_write(struct vop_strategy_args *ap) 504 { 505 hammer2_xop_strategy_t *xop; 506 hammer2_pfs_t *pmp; 507 struct bio *bio; 508 hammer2_inode_t *ip; 509 510 bio = ap->a_bio; 511 ip = VTOI(ap->a_vp); 512 pmp = ip->pmp; 513 514 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 515 hammer2_lwinprog_ref(pmp); 516 hammer2_trans_assert_strategy(pmp); 517 hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE); 518 519 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 520 HAMMER2_XOP_STRATEGY); 521 xop->finished = 0; 522 xop->bio = bio; 523 xop->lbase = bio->bio_offset; 524 hammer2_mtx_init(&xop->lock, "h2biow"); 525 hammer2_xop_start(&xop->head, &hammer2_strategy_write_desc); 526 /* asynchronous completion */ 527 528 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe); 529 530 return(0); 531 } 532 533 /* 534 * Per-node XOP (threaded). Write the logical buffer to the media. 535 * 536 * This is a bit problematic because there may be multiple target and 537 * any of them may be able to release the bp. In addition, if our 538 * particulr target is offline we don't want to block the bp (and thus 539 * the frontend). To accomplish this we copy the data to the per-thr 540 * scratch buffer. 541 */ 542 void 543 hammer2_xop_strategy_write(hammer2_xop_t *arg, void *scratch, int clindex) 544 { 545 hammer2_xop_strategy_t *xop = &arg->xop_strategy; 546 hammer2_chain_t *parent; 547 hammer2_key_t lbase; 548 hammer2_inode_t *ip; 549 struct bio *bio; 550 struct buf *bp; 551 int error; 552 int lblksize; 553 int pblksize; 554 char *bio_data; 555 556 /* 557 * We can only access the bp/bio if the frontend has not yet 558 * completed. 559 */ 560 if (xop->finished) 561 return; 562 hammer2_mtx_sh(&xop->lock); 563 if (xop->finished) { 564 hammer2_mtx_unlock(&xop->lock); 565 return; 566 } 567 568 lbase = xop->lbase; 569 bio = xop->bio; /* ephermal */ 570 bp = bio->bio_buf; /* ephermal */ 571 ip = xop->head.ip1; /* retained by ref */ 572 bio_data = scratch; 573 574 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */ 575 576 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL); 577 pblksize = hammer2_calc_physical(ip, lbase); 578 bkvasync(bp); 579 KKASSERT(lblksize <= MAXPHYS); 580 bcopy(bp->b_data, bio_data, lblksize); 581 582 hammer2_mtx_unlock(&xop->lock); 583 bp = NULL; /* safety, illegal to access after unlock */ 584 bio = NULL; /* safety, illegal to access after unlock */ 585 586 /* 587 * Actual operation 588 */ 589 parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS); 590 hammer2_write_file_core(bio_data, ip, &parent, 591 lbase, IO_ASYNC, pblksize, 592 xop->head.mtid, &error); 593 if (parent) { 594 hammer2_chain_unlock(parent); 595 hammer2_chain_drop(parent); 596 parent = NULL; /* safety */ 597 } 598 hammer2_xop_feed(&xop->head, NULL, clindex, error); 599 600 /* 601 * Try to complete the operation on behalf of the front-end. 602 */ 603 if (xop->finished) 604 return; 605 hammer2_mtx_ex(&xop->lock); 606 if (xop->finished) { 607 hammer2_mtx_unlock(&xop->lock); 608 return; 609 } 610 611 /* 612 * Async operation has not completed and we now own the lock. 613 * Determine if we can complete the operation by issuing the 614 * frontend collection non-blocking. 615 * 616 * H2 double-buffers the data, setting B_NOTMETA on the logical 617 * buffer hints to the OS that the logical buffer should not be 618 * swapcached (since the device buffer can be). 619 */ 620 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT); 621 622 if (error == HAMMER2_ERROR_EINPROGRESS) { 623 hammer2_mtx_unlock(&xop->lock); 624 return; 625 } 626 627 /* 628 * Async operation has completed. 629 */ 630 xop->finished = 1; 631 hammer2_mtx_unlock(&xop->lock); 632 633 bio = xop->bio; /* now owned by us */ 634 bp = bio->bio_buf; /* now owned by us */ 635 636 if (error == HAMMER2_ERROR_ENOENT || error == 0) { 637 bp->b_flags |= B_NOTMETA; 638 bp->b_resid = 0; 639 bp->b_error = 0; 640 biodone(bio); 641 } else { 642 kprintf("xop_strategy_write: error %d loff=%016jx\n", 643 error, (intmax_t)bp->b_loffset); 644 bp->b_flags |= B_ERROR; 645 bp->b_error = EIO; 646 biodone(bio); 647 } 648 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 649 hammer2_trans_assert_strategy(ip->pmp); 650 hammer2_lwinprog_drop(ip->pmp); 651 hammer2_trans_done(ip->pmp, HAMMER2_TRANS_BUFCACHE); 652 } 653 654 /* 655 * Wait for pending I/O to complete 656 */ 657 void 658 hammer2_bioq_sync(hammer2_pfs_t *pmp) 659 { 660 hammer2_lwinprog_wait(pmp, 0); 661 } 662 663 /* 664 * Assign physical storage at (cparent, lbase), returning a suitable chain 665 * and setting *errorp appropriately. 666 * 667 * If no error occurs, the returned chain will be in a modified state. 668 * 669 * If an error occurs, the returned chain may or may not be NULL. If 670 * not-null any chain->error (if not 0) will also be rolled up into *errorp. 671 * So the caller only needs to test *errorp. 672 * 673 * cparent can wind up being anything. 674 * 675 * If datap is not NULL, *datap points to the real data we intend to write. 676 * If we can dedup the storage location we set *datap to NULL to indicate 677 * to the caller that a dedup occurred. 678 * 679 * NOTE: Special case for data embedded in inode. 680 */ 681 static 682 hammer2_chain_t * 683 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp, 684 hammer2_key_t lbase, int pblksize, 685 hammer2_tid_t mtid, char **datap, int *errorp) 686 { 687 hammer2_chain_t *chain; 688 hammer2_key_t key_dummy; 689 hammer2_off_t dedup_off; 690 int pradix = hammer2_getradix(pblksize); 691 692 /* 693 * Locate the chain associated with lbase, return a locked chain. 694 * However, do not instantiate any data reference (which utilizes a 695 * device buffer) because we will be using direct IO via the 696 * logical buffer cache buffer. 697 */ 698 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN); 699 700 chain = hammer2_chain_lookup(parentp, &key_dummy, 701 lbase, lbase, 702 errorp, 703 HAMMER2_LOOKUP_NODATA); 704 705 /* 706 * The lookup code should not return a DELETED chain to us, unless 707 * its a short-file embedded in the inode. Then it is possible for 708 * the lookup to return a deleted inode. 709 */ 710 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) && 711 chain->bref.type != HAMMER2_BREF_TYPE_INODE) { 712 kprintf("assign physical deleted chain @ " 713 "%016jx (%016jx.%02x) ip %016jx\n", 714 lbase, chain->bref.data_off, chain->bref.type, 715 ip->meta.inum); 716 Debugger("bleh"); 717 } 718 719 if (chain == NULL) { 720 /* 721 * We found a hole, create a new chain entry. 722 * 723 * NOTE: DATA chains are created without device backing 724 * store (nor do we want any). 725 */ 726 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap, 727 pblksize); 728 *errorp |= hammer2_chain_create(parentp, &chain, NULL, ip->pmp, 729 HAMMER2_ENC_CHECK(ip->meta.check_algo) | 730 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE), 731 lbase, HAMMER2_PBUFRADIX, 732 HAMMER2_BREF_TYPE_DATA, 733 pblksize, mtid, 734 dedup_off, 0); 735 if (chain == NULL) 736 goto failed; 737 /*ip->delta_dcount += pblksize;*/ 738 } else if (chain->error == 0) { 739 switch (chain->bref.type) { 740 case HAMMER2_BREF_TYPE_INODE: 741 /* 742 * The data is embedded in the inode, which requires 743 * a bit more finess. 744 */ 745 *errorp |= hammer2_chain_modify_ip(ip, chain, mtid, 0); 746 break; 747 case HAMMER2_BREF_TYPE_DATA: 748 dedup_off = hammer2_dedup_lookup(chain->hmp, datap, 749 pblksize); 750 if (chain->bytes != pblksize) { 751 *errorp |= hammer2_chain_resize(chain, 752 mtid, dedup_off, 753 pradix, 754 HAMMER2_MODIFY_OPTDATA); 755 if (*errorp) 756 break; 757 } 758 759 /* 760 * DATA buffers must be marked modified whether the 761 * data is in a logical buffer or not. We also have 762 * to make this call to fixup the chain data pointers 763 * after resizing in case this is an encrypted or 764 * compressed buffer. 765 */ 766 *errorp |= hammer2_chain_modify(chain, mtid, dedup_off, 767 HAMMER2_MODIFY_OPTDATA); 768 break; 769 default: 770 panic("hammer2_assign_physical: bad type"); 771 /* NOT REACHED */ 772 break; 773 } 774 } else { 775 *errorp = chain->error; 776 } 777 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 778 failed: 779 return (chain); 780 } 781 782 /* 783 * hammer2_write_file_core() 784 * 785 * The core write function which determines which path to take 786 * depending on compression settings. We also have to locate the 787 * related chains so we can calculate and set the check data for 788 * the blockref. 789 */ 790 static 791 void 792 hammer2_write_file_core(char *data, hammer2_inode_t *ip, 793 hammer2_chain_t **parentp, 794 hammer2_key_t lbase, int ioflag, int pblksize, 795 hammer2_tid_t mtid, int *errorp) 796 { 797 hammer2_chain_t *chain; 798 char *bdata; 799 800 *errorp = 0; 801 802 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) { 803 case HAMMER2_COMP_NONE: 804 /* 805 * We have to assign physical storage to the buffer 806 * we intend to dirty or write now to avoid deadlocks 807 * in the strategy code later. 808 * 809 * This can return NOOFFSET for inode-embedded data. 810 * The strategy code will take care of it in that case. 811 */ 812 bdata = data; 813 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize, 814 mtid, &bdata, errorp); 815 if (*errorp) { 816 /* skip modifications */ 817 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 818 hammer2_inode_data_t *wipdata; 819 820 wipdata = &chain->data->ipdata; 821 KKASSERT(wipdata->meta.op_flags & 822 HAMMER2_OPFLAG_DIRECTDATA); 823 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 824 ++hammer2_iod_file_wembed; 825 } else if (bdata == NULL) { 826 /* 827 * Copy of data already present on-media. 828 */ 829 chain->bref.methods = 830 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 831 HAMMER2_ENC_CHECK(ip->meta.check_algo); 832 hammer2_chain_setcheck(chain, data); 833 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 834 } else { 835 hammer2_write_bp(chain, data, ioflag, pblksize, 836 mtid, errorp, ip->meta.check_algo); 837 } 838 if (chain) { 839 hammer2_chain_unlock(chain); 840 hammer2_chain_drop(chain); 841 } 842 break; 843 case HAMMER2_COMP_AUTOZERO: 844 /* 845 * Check for zero-fill only 846 */ 847 hammer2_zero_check_and_write(data, ip, parentp, 848 lbase, ioflag, pblksize, 849 mtid, errorp, 850 ip->meta.check_algo); 851 break; 852 case HAMMER2_COMP_LZ4: 853 case HAMMER2_COMP_ZLIB: 854 default: 855 /* 856 * Check for zero-fill and attempt compression. 857 */ 858 hammer2_compress_and_write(data, ip, parentp, 859 lbase, ioflag, pblksize, 860 mtid, errorp, 861 ip->meta.comp_algo, 862 ip->meta.check_algo); 863 break; 864 } 865 } 866 867 /* 868 * Helper 869 * 870 * Generic function that will perform the compression in compression 871 * write path. The compression algorithm is determined by the settings 872 * obtained from inode. 873 */ 874 static 875 void 876 hammer2_compress_and_write(char *data, hammer2_inode_t *ip, 877 hammer2_chain_t **parentp, 878 hammer2_key_t lbase, int ioflag, int pblksize, 879 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo) 880 { 881 hammer2_chain_t *chain; 882 int comp_size; 883 int comp_block_size; 884 char *comp_buffer; 885 char *bdata; 886 887 /* 888 * An all-zeros write creates a hole unless the check code 889 * is disabled. When the check code is disabled all writes 890 * are done in-place, including any all-zeros writes. 891 * 892 * NOTE: A snapshot will still force a copy-on-write 893 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c). 894 */ 895 if (check_algo != HAMMER2_CHECK_NONE && 896 test_block_zeros(data, pblksize)) { 897 zero_write(data, ip, parentp, lbase, mtid, errorp); 898 return; 899 } 900 901 /* 902 * Compression requested. Try to compress the block. We store 903 * the data normally if we cannot sufficiently compress it. 904 * 905 * We have a heuristic to detect files which are mostly 906 * uncompressable and avoid the compression attempt in that 907 * case. If the compression heuristic is turned off, we always 908 * try to compress. 909 */ 910 comp_size = 0; 911 comp_buffer = NULL; 912 913 KKASSERT(pblksize / 2 <= 32768); 914 915 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 || 916 hammer2_always_compress) { 917 z_stream strm_compress; 918 int comp_level; 919 int ret; 920 921 switch(HAMMER2_DEC_ALGO(comp_algo)) { 922 case HAMMER2_COMP_LZ4: 923 /* 924 * We need to prefix with the size, LZ4 925 * doesn't do it for us. Add the related 926 * overhead. 927 * 928 * NOTE: The LZ4 code seems to assume at least an 929 * 8-byte buffer size granularity and may 930 * overrun the buffer if given a 4-byte 931 * granularity. 932 */ 933 comp_buffer = objcache_get(cache_buffer_write, 934 M_INTWAIT); 935 comp_size = LZ4_compress_limitedOutput( 936 data, 937 &comp_buffer[sizeof(int)], 938 pblksize, 939 pblksize / 2 - sizeof(int64_t)); 940 *(int *)comp_buffer = comp_size; 941 if (comp_size) 942 comp_size += sizeof(int); 943 break; 944 case HAMMER2_COMP_ZLIB: 945 comp_level = HAMMER2_DEC_LEVEL(comp_algo); 946 if (comp_level == 0) 947 comp_level = 6; /* default zlib compression */ 948 else if (comp_level < 6) 949 comp_level = 6; 950 else if (comp_level > 9) 951 comp_level = 9; 952 ret = deflateInit(&strm_compress, comp_level); 953 if (ret != Z_OK) { 954 kprintf("HAMMER2 ZLIB: fatal error " 955 "on deflateInit.\n"); 956 } 957 958 comp_buffer = objcache_get(cache_buffer_write, 959 M_INTWAIT); 960 strm_compress.next_in = data; 961 strm_compress.avail_in = pblksize; 962 strm_compress.next_out = comp_buffer; 963 strm_compress.avail_out = pblksize / 2; 964 ret = deflate(&strm_compress, Z_FINISH); 965 if (ret == Z_STREAM_END) { 966 comp_size = pblksize / 2 - 967 strm_compress.avail_out; 968 } else { 969 comp_size = 0; 970 } 971 ret = deflateEnd(&strm_compress); 972 break; 973 default: 974 kprintf("Error: Unknown compression method.\n"); 975 kprintf("Comp_method = %d.\n", comp_algo); 976 break; 977 } 978 } 979 980 if (comp_size == 0) { 981 /* 982 * compression failed or turned off 983 */ 984 comp_block_size = pblksize; /* safety */ 985 if (++ip->comp_heuristic > 128) 986 ip->comp_heuristic = 8; 987 } else { 988 /* 989 * compression succeeded 990 */ 991 ip->comp_heuristic = 0; 992 if (comp_size <= 1024) { 993 comp_block_size = 1024; 994 } else if (comp_size <= 2048) { 995 comp_block_size = 2048; 996 } else if (comp_size <= 4096) { 997 comp_block_size = 4096; 998 } else if (comp_size <= 8192) { 999 comp_block_size = 8192; 1000 } else if (comp_size <= 16384) { 1001 comp_block_size = 16384; 1002 } else if (comp_size <= 32768) { 1003 comp_block_size = 32768; 1004 } else { 1005 panic("hammer2: WRITE PATH: " 1006 "Weird comp_size value."); 1007 /* NOT REACHED */ 1008 comp_block_size = pblksize; 1009 } 1010 1011 /* 1012 * Must zero the remainder or dedup (which operates on a 1013 * physical block basis) will not find matches. 1014 */ 1015 if (comp_size < comp_block_size) { 1016 bzero(comp_buffer + comp_size, 1017 comp_block_size - comp_size); 1018 } 1019 } 1020 1021 /* 1022 * Assign physical storage, bdata will be set to NULL if a live-dedup 1023 * was successful. 1024 */ 1025 bdata = comp_size ? comp_buffer : data; 1026 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size, 1027 mtid, &bdata, errorp); 1028 1029 if (*errorp) { 1030 goto done; 1031 } 1032 1033 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 1034 hammer2_inode_data_t *wipdata; 1035 1036 *errorp = hammer2_chain_modify_ip(ip, chain, mtid, 0); 1037 if (*errorp == 0) { 1038 wipdata = &chain->data->ipdata; 1039 KKASSERT(wipdata->meta.op_flags & 1040 HAMMER2_OPFLAG_DIRECTDATA); 1041 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1042 ++hammer2_iod_file_wembed; 1043 } 1044 } else if (bdata == NULL) { 1045 /* 1046 * Live deduplication, a copy of the data is already present 1047 * on the media. 1048 */ 1049 if (comp_size) { 1050 chain->bref.methods = 1051 HAMMER2_ENC_COMP(comp_algo) + 1052 HAMMER2_ENC_CHECK(check_algo); 1053 } else { 1054 chain->bref.methods = 1055 HAMMER2_ENC_COMP( 1056 HAMMER2_COMP_NONE) + 1057 HAMMER2_ENC_CHECK(check_algo); 1058 } 1059 bdata = comp_size ? comp_buffer : data; 1060 hammer2_chain_setcheck(chain, bdata); 1061 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1062 } else { 1063 hammer2_io_t *dio; 1064 1065 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED); 1066 1067 switch(chain->bref.type) { 1068 case HAMMER2_BREF_TYPE_INODE: 1069 panic("hammer2_compress_and_write: unexpected inode\n"); 1070 break; 1071 case HAMMER2_BREF_TYPE_DATA: 1072 /* 1073 * Optimize out the read-before-write 1074 * if possible. 1075 */ 1076 *errorp = hammer2_io_newnz(chain->hmp, 1077 chain->bref.type, 1078 chain->bref.data_off, 1079 chain->bytes, 1080 &dio); 1081 if (*errorp) { 1082 hammer2_io_brelse(&dio); 1083 kprintf("hammer2: WRITE PATH: " 1084 "dbp bread error\n"); 1085 break; 1086 } 1087 bdata = hammer2_io_data(dio, chain->bref.data_off); 1088 1089 /* 1090 * When loading the block make sure we don't 1091 * leave garbage after the compressed data. 1092 */ 1093 if (comp_size) { 1094 chain->bref.methods = 1095 HAMMER2_ENC_COMP(comp_algo) + 1096 HAMMER2_ENC_CHECK(check_algo); 1097 bcopy(comp_buffer, bdata, comp_block_size); 1098 } else { 1099 chain->bref.methods = 1100 HAMMER2_ENC_COMP( 1101 HAMMER2_COMP_NONE) + 1102 HAMMER2_ENC_CHECK(check_algo); 1103 bcopy(data, bdata, pblksize); 1104 } 1105 1106 /* 1107 * The flush code doesn't calculate check codes for 1108 * file data (doing so can result in excessive I/O), 1109 * so we do it here. 1110 */ 1111 hammer2_chain_setcheck(chain, bdata); 1112 1113 /* 1114 * Device buffer is now valid, chain is no longer in 1115 * the initial state. 1116 * 1117 * (No blockref table worries with file data) 1118 */ 1119 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1120 hammer2_dedup_record(chain, dio, bdata); 1121 1122 /* Now write the related bdp. */ 1123 if (ioflag & IO_SYNC) { 1124 /* 1125 * Synchronous I/O requested. 1126 */ 1127 hammer2_io_bwrite(&dio); 1128 /* 1129 } else if ((ioflag & IO_DIRECT) && 1130 loff + n == pblksize) { 1131 hammer2_io_bdwrite(&dio); 1132 */ 1133 } else if (ioflag & IO_ASYNC) { 1134 hammer2_io_bawrite(&dio); 1135 } else { 1136 hammer2_io_bdwrite(&dio); 1137 } 1138 break; 1139 default: 1140 panic("hammer2_compress_and_write: bad chain type %d\n", 1141 chain->bref.type); 1142 /* NOT REACHED */ 1143 break; 1144 } 1145 } 1146 done: 1147 if (chain) { 1148 hammer2_chain_unlock(chain); 1149 hammer2_chain_drop(chain); 1150 } 1151 if (comp_buffer) 1152 objcache_put(cache_buffer_write, comp_buffer); 1153 } 1154 1155 /* 1156 * Helper 1157 * 1158 * Function that performs zero-checking and writing without compression, 1159 * it corresponds to default zero-checking path. 1160 */ 1161 static 1162 void 1163 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip, 1164 hammer2_chain_t **parentp, 1165 hammer2_key_t lbase, int ioflag, int pblksize, 1166 hammer2_tid_t mtid, int *errorp, 1167 int check_algo) 1168 { 1169 hammer2_chain_t *chain; 1170 char *bdata; 1171 1172 if (check_algo != HAMMER2_CHECK_NONE && 1173 test_block_zeros(data, pblksize)) { 1174 /* 1175 * An all-zeros write creates a hole unless the check code 1176 * is disabled. When the check code is disabled all writes 1177 * are done in-place, including any all-zeros writes. 1178 * 1179 * NOTE: A snapshot will still force a copy-on-write 1180 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c). 1181 */ 1182 zero_write(data, ip, parentp, lbase, mtid, errorp); 1183 } else { 1184 /* 1185 * Normal write (bdata set to NULL if de-duplicated) 1186 */ 1187 bdata = data; 1188 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize, 1189 mtid, &bdata, errorp); 1190 if (*errorp) { 1191 /* do nothing */ 1192 } else if (bdata) { 1193 hammer2_write_bp(chain, data, ioflag, pblksize, 1194 mtid, errorp, check_algo); 1195 } else { 1196 /* dedup occurred */ 1197 chain->bref.methods = 1198 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 1199 HAMMER2_ENC_CHECK(check_algo); 1200 hammer2_chain_setcheck(chain, data); 1201 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1202 } 1203 if (chain) { 1204 hammer2_chain_unlock(chain); 1205 hammer2_chain_drop(chain); 1206 } 1207 } 1208 } 1209 1210 /* 1211 * Helper 1212 * 1213 * A function to test whether a block of data contains only zeros, 1214 * returns TRUE (non-zero) if the block is all zeros. 1215 */ 1216 static 1217 int 1218 test_block_zeros(const char *buf, size_t bytes) 1219 { 1220 size_t i; 1221 1222 for (i = 0; i < bytes; i += sizeof(long)) { 1223 if (*(const long *)(buf + i) != 0) 1224 return (0); 1225 } 1226 return (1); 1227 } 1228 1229 /* 1230 * Helper 1231 * 1232 * Function to "write" a block that contains only zeros. 1233 */ 1234 static 1235 void 1236 zero_write(char *data, hammer2_inode_t *ip, 1237 hammer2_chain_t **parentp, 1238 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp) 1239 { 1240 hammer2_chain_t *chain; 1241 hammer2_key_t key_dummy; 1242 1243 chain = hammer2_chain_lookup(parentp, &key_dummy, 1244 lbase, lbase, 1245 errorp, 1246 HAMMER2_LOOKUP_NODATA); 1247 if (chain) { 1248 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 1249 hammer2_inode_data_t *wipdata; 1250 1251 if (*errorp == 0) { 1252 *errorp = hammer2_chain_modify_ip(ip, chain, 1253 mtid, 0); 1254 } 1255 if (*errorp == 0) { 1256 wipdata = &chain->data->ipdata; 1257 KKASSERT(wipdata->meta.op_flags & 1258 HAMMER2_OPFLAG_DIRECTDATA); 1259 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1260 ++hammer2_iod_file_wembed; 1261 } 1262 } else { 1263 /* chain->error ok for deletion */ 1264 hammer2_chain_delete(*parentp, chain, 1265 mtid, HAMMER2_DELETE_PERMANENT); 1266 ++hammer2_iod_file_wzero; 1267 } 1268 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 1269 hammer2_chain_unlock(chain); 1270 hammer2_chain_drop(chain); 1271 } else { 1272 ++hammer2_iod_file_wzero; 1273 } 1274 } 1275 1276 /* 1277 * Helper 1278 * 1279 * Function to write the data as it is, without performing any sort of 1280 * compression. This function is used in path without compression and 1281 * default zero-checking path. 1282 */ 1283 static 1284 void 1285 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag, 1286 int pblksize, 1287 hammer2_tid_t mtid, int *errorp, int check_algo) 1288 { 1289 hammer2_inode_data_t *wipdata; 1290 hammer2_io_t *dio; 1291 char *bdata; 1292 int error; 1293 1294 error = 0; /* XXX TODO below */ 1295 1296 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED); 1297 1298 switch(chain->bref.type) { 1299 case HAMMER2_BREF_TYPE_INODE: 1300 wipdata = &chain->data->ipdata; 1301 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA); 1302 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1303 error = 0; 1304 ++hammer2_iod_file_wembed; 1305 break; 1306 case HAMMER2_BREF_TYPE_DATA: 1307 error = hammer2_io_newnz(chain->hmp, 1308 chain->bref.type, 1309 chain->bref.data_off, 1310 chain->bytes, &dio); 1311 if (error) { 1312 hammer2_io_bqrelse(&dio); 1313 kprintf("hammer2: WRITE PATH: " 1314 "dbp bread error\n"); 1315 break; 1316 } 1317 bdata = hammer2_io_data(dio, chain->bref.data_off); 1318 1319 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 1320 HAMMER2_ENC_CHECK(check_algo); 1321 bcopy(data, bdata, chain->bytes); 1322 1323 /* 1324 * The flush code doesn't calculate check codes for 1325 * file data (doing so can result in excessive I/O), 1326 * so we do it here. 1327 */ 1328 hammer2_chain_setcheck(chain, bdata); 1329 1330 /* 1331 * Device buffer is now valid, chain is no longer in 1332 * the initial state. 1333 * 1334 * (No blockref table worries with file data) 1335 */ 1336 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1337 hammer2_dedup_record(chain, dio, bdata); 1338 1339 if (ioflag & IO_SYNC) { 1340 /* 1341 * Synchronous I/O requested. 1342 */ 1343 hammer2_io_bwrite(&dio); 1344 /* 1345 } else if ((ioflag & IO_DIRECT) && 1346 loff + n == pblksize) { 1347 hammer2_io_bdwrite(&dio); 1348 */ 1349 } else if (ioflag & IO_ASYNC) { 1350 hammer2_io_bawrite(&dio); 1351 } else { 1352 hammer2_io_bdwrite(&dio); 1353 } 1354 break; 1355 default: 1356 panic("hammer2_write_bp: bad chain type %d\n", 1357 chain->bref.type); 1358 /* NOT REACHED */ 1359 error = 0; 1360 break; 1361 } 1362 *errorp = error; 1363 } 1364 1365 /* 1366 * LIVE DEDUP HEURISTICS 1367 * 1368 * Record media and crc information for possible dedup operation. Note 1369 * that the dedup mask bits must also be set in the related DIO for a dedup 1370 * to be fully validated (which is handled in the freemap allocation code). 1371 * 1372 * WARNING! This code is SMP safe but the heuristic allows SMP collisions. 1373 * All fields must be loaded into locals and validated. 1374 * 1375 * WARNING! Should only be used for file data and directory entries, 1376 * hammer2_chain_modify() only checks for the dedup case on data 1377 * chains. Also, dedup data can only be recorded for committed 1378 * chains (so NOT strategy writes which can undergo further 1379 * modification after the fact!). 1380 */ 1381 void 1382 hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio, 1383 const char *data) 1384 { 1385 hammer2_dev_t *hmp; 1386 hammer2_dedup_t *dedup; 1387 uint64_t crc; 1388 uint64_t mask; 1389 int best = 0; 1390 int i; 1391 int dticks; 1392 1393 /* 1394 * We can only record a dedup if we have media data to test against. 1395 * If dedup is not enabled, return early, which allows a chain to 1396 * remain marked MODIFIED (which might have benefits in special 1397 * situations, though typically it does not). 1398 */ 1399 if (hammer2_dedup_enable == 0) 1400 return; 1401 if (dio == NULL) { 1402 dio = chain->dio; 1403 if (dio == NULL) 1404 return; 1405 } 1406 1407 hmp = chain->hmp; 1408 1409 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) { 1410 case HAMMER2_CHECK_ISCSI32: 1411 /* 1412 * XXX use the built-in crc (the dedup lookup sequencing 1413 * needs to be fixed so the check code is already present 1414 * when dedup_lookup is called) 1415 */ 1416 #if 0 1417 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value; 1418 #endif 1419 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED); 1420 break; 1421 case HAMMER2_CHECK_XXHASH64: 1422 crc = chain->bref.check.xxhash64.value; 1423 break; 1424 case HAMMER2_CHECK_SHA192: 1425 /* 1426 * XXX use the built-in crc (the dedup lookup sequencing 1427 * needs to be fixed so the check code is already present 1428 * when dedup_lookup is called) 1429 */ 1430 #if 0 1431 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^ 1432 ((uint64_t *)chain->bref.check.sha192.data)[1] ^ 1433 ((uint64_t *)chain->bref.check.sha192.data)[2]; 1434 #endif 1435 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED); 1436 break; 1437 default: 1438 /* 1439 * Cannot dedup without a check code 1440 * 1441 * NOTE: In particular, CHECK_NONE allows a sector to be 1442 * overwritten without copy-on-write, recording 1443 * a dedup block for a CHECK_NONE object would be 1444 * a disaster! 1445 */ 1446 return; 1447 } 1448 1449 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUPABLE); 1450 1451 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)]; 1452 for (i = 0; i < 4; ++i) { 1453 if (dedup[i].data_crc == crc) { 1454 best = i; 1455 break; 1456 } 1457 dticks = (int)(dedup[i].ticks - dedup[best].ticks); 1458 if (dticks < 0 || dticks > hz * 60 * 30) 1459 best = i; 1460 } 1461 dedup += best; 1462 if (hammer2_debug & 0x40000) { 1463 kprintf("REC %04x %016jx %016jx\n", 1464 (int)(dedup - hmp->heur_dedup), 1465 crc, 1466 chain->bref.data_off); 1467 } 1468 dedup->ticks = ticks; 1469 dedup->data_off = chain->bref.data_off; 1470 dedup->data_crc = crc; 1471 1472 /* 1473 * Set the valid bits for the dedup only after we know the data 1474 * buffer has been updated. The alloc bits were set (and the valid 1475 * bits cleared) when the media was allocated. 1476 * 1477 * This is done in two stages becuase the bulkfree code can race 1478 * the gap between allocation and data population. Both masks must 1479 * be set before a bcmp/dedup operation is able to use the block. 1480 */ 1481 mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes); 1482 atomic_set_64(&dio->dedup_valid, mask); 1483 1484 #if 0 1485 /* 1486 * XXX removed. MODIFIED is an integral part of the flush code, 1487 * lets not just clear it 1488 */ 1489 /* 1490 * Once we record the dedup the chain must be marked clean to 1491 * prevent reuse of the underlying block. Remember that this 1492 * write occurs when the buffer cache is flushed (i.e. on sync(), 1493 * fsync(), filesystem periodic sync, or when the kernel needs to 1494 * flush a buffer), and not whenever the user write()s. 1495 */ 1496 if (chain->flags & HAMMER2_CHAIN_MODIFIED) { 1497 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED); 1498 atomic_add_long(&hammer2_count_modified_chains, -1); 1499 if (chain->pmp) 1500 hammer2_pfs_memory_wakeup(chain->pmp, -1); 1501 } 1502 #endif 1503 } 1504 1505 static 1506 hammer2_off_t 1507 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize) 1508 { 1509 hammer2_dedup_t *dedup; 1510 hammer2_io_t *dio; 1511 hammer2_off_t off; 1512 uint64_t crc; 1513 uint64_t mask; 1514 char *data; 1515 char *dtmp; 1516 int i; 1517 1518 if (hammer2_dedup_enable == 0) 1519 return 0; 1520 data = *datap; 1521 if (data == NULL) 1522 return 0; 1523 1524 /* 1525 * XXX use the built-in crc (the dedup lookup sequencing 1526 * needs to be fixed so the check code is already present 1527 * when dedup_lookup is called) 1528 */ 1529 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED); 1530 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)]; 1531 1532 if (hammer2_debug & 0x40000) { 1533 kprintf("LOC %04x/4 %016jx\n", 1534 (int)(dedup - hmp->heur_dedup), 1535 crc); 1536 } 1537 1538 for (i = 0; i < 4; ++i) { 1539 off = dedup[i].data_off; 1540 cpu_ccfence(); 1541 if (dedup[i].data_crc != crc) 1542 continue; 1543 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize) 1544 continue; 1545 dio = hammer2_io_getquick(hmp, off, pblksize); 1546 if (dio) { 1547 dtmp = hammer2_io_data(dio, off), 1548 mask = hammer2_dedup_mask(dio, off, pblksize); 1549 if ((dio->dedup_alloc & mask) == mask && 1550 (dio->dedup_valid & mask) == mask && 1551 bcmp(data, dtmp, pblksize) == 0) { 1552 if (hammer2_debug & 0x40000) { 1553 kprintf("DEDUP SUCCESS %016jx\n", 1554 (intmax_t)off); 1555 } 1556 hammer2_io_putblk(&dio); 1557 *datap = NULL; 1558 dedup[i].ticks = ticks; /* update use */ 1559 atomic_add_long(&hammer2_iod_file_wdedup, 1560 pblksize); 1561 1562 return off; /* RETURN */ 1563 } 1564 hammer2_io_putblk(&dio); 1565 } 1566 } 1567 return 0; 1568 } 1569 1570 /* 1571 * Poof. Races are ok, if someone gets in and reuses a dedup offset 1572 * before or while we are clearing it they will also recover the freemap 1573 * entry (set it to fully allocated), so a bulkfree race can only set it 1574 * to a possibly-free state. 1575 * 1576 * XXX ok, well, not really sure races are ok but going to run with it 1577 * for the moment. 1578 */ 1579 void 1580 hammer2_dedup_clear(hammer2_dev_t *hmp) 1581 { 1582 int i; 1583 1584 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) { 1585 hmp->heur_dedup[i].data_off = 0; 1586 hmp->heur_dedup[i].ticks = ticks - 1; 1587 } 1588 } 1589