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