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/fcntl.h> 50 #include <sys/buf.h> 51 #include <sys/proc.h> 52 #include <sys/namei.h> 53 #include <sys/mount.h> 54 #include <sys/vnode.h> 55 #include <sys/mountctl.h> 56 #include <sys/dirent.h> 57 #include <sys/uio.h> 58 #include <sys/objcache.h> 59 #include <sys/event.h> 60 #include <sys/file.h> 61 #include <vfs/fifofs/fifo.h> 62 63 #include "hammer2.h" 64 #include "hammer2_lz4.h" 65 66 #include "zlib/hammer2_zlib.h" 67 68 struct objcache *cache_buffer_read; 69 struct objcache *cache_buffer_write; 70 71 /* 72 * Strategy code (async logical file buffer I/O from system) 73 * 74 * Except for the transaction init (which should normally not block), 75 * we essentially run the strategy operation asynchronously via a XOP. 76 * 77 * WARNING! The XOP deals with buffer synchronization. It is not synchronized 78 * to the current cpu. 79 * 80 * XXX This isn't supposed to be able to deadlock against vfs_sync vfsync() 81 * calls but it has in the past when multiple flushes are queued. 82 * 83 * XXX We currently terminate the transaction once we get a quorum, otherwise 84 * the frontend can stall, but this can leave the remaining nodes with 85 * a potential flush conflict. We need to delay flushes on those nodes 86 * until running transactions complete separately from the normal 87 * transaction sequencing. FIXME TODO. 88 */ 89 static void hammer2_strategy_xop_read(hammer2_thread_t *thr, 90 hammer2_xop_t *arg); 91 static void hammer2_strategy_xop_write(hammer2_thread_t *thr, 92 hammer2_xop_t *arg); 93 static int hammer2_strategy_read(struct vop_strategy_args *ap); 94 static int hammer2_strategy_write(struct vop_strategy_args *ap); 95 static void hammer2_strategy_read_completion(hammer2_chain_t *focus, 96 const char *data, struct bio *bio); 97 98 static hammer2_off_t hammer2_dedup_lookup(hammer2_dev_t *hmp, 99 char **datap, int pblksize); 100 101 int 102 hammer2_vop_strategy(struct vop_strategy_args *ap) 103 { 104 struct bio *biop; 105 struct buf *bp; 106 int error; 107 108 biop = ap->a_bio; 109 bp = biop->bio_buf; 110 111 switch(bp->b_cmd) { 112 case BUF_CMD_READ: 113 error = hammer2_strategy_read(ap); 114 ++hammer2_iod_file_read; 115 break; 116 case BUF_CMD_WRITE: 117 error = hammer2_strategy_write(ap); 118 ++hammer2_iod_file_write; 119 break; 120 default: 121 bp->b_error = error = EINVAL; 122 bp->b_flags |= B_ERROR; 123 biodone(biop); 124 break; 125 } 126 return (error); 127 } 128 129 /* 130 * Return the largest contiguous physical disk range for the logical 131 * request, in bytes. 132 * 133 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb) 134 * 135 * Basically disabled, the logical buffer write thread has to deal with 136 * buffers one-at-a-time. Note that this should not prevent cluster_read() 137 * from reading-ahead, it simply prevents it from trying form a single 138 * cluster buffer for the logical request. H2 already uses 64KB buffers! 139 */ 140 int 141 hammer2_vop_bmap(struct vop_bmap_args *ap) 142 { 143 *ap->a_doffsetp = NOOFFSET; 144 if (ap->a_runp) 145 *ap->a_runp = 0; 146 if (ap->a_runb) 147 *ap->a_runb = 0; 148 return (EOPNOTSUPP); 149 } 150 151 /**************************************************************************** 152 * READ SUPPORT * 153 ****************************************************************************/ 154 /* 155 * Callback used in read path in case that a block is compressed with LZ4. 156 */ 157 static 158 void 159 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio) 160 { 161 struct buf *bp; 162 char *compressed_buffer; 163 int compressed_size; 164 int result; 165 166 bp = bio->bio_buf; 167 168 #if 0 169 if bio->bio_caller_info2.index && 170 bio->bio_caller_info1.uvalue32 != 171 crc32(bp->b_data, bp->b_bufsize) --- return error 172 #endif 173 174 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE); 175 compressed_size = *(const int *)data; 176 KKASSERT((uint32_t)compressed_size <= bytes - sizeof(int)); 177 178 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT); 179 result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]), 180 compressed_buffer, 181 compressed_size, 182 bp->b_bufsize); 183 if (result < 0) { 184 kprintf("READ PATH: Error during decompression." 185 "bio %016jx/%d\n", 186 (intmax_t)bio->bio_offset, bytes); 187 /* make sure it isn't random garbage */ 188 bzero(compressed_buffer, bp->b_bufsize); 189 } 190 KKASSERT(result <= bp->b_bufsize); 191 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize); 192 if (result < bp->b_bufsize) 193 bzero(bp->b_data + result, bp->b_bufsize - result); 194 objcache_put(cache_buffer_read, compressed_buffer); 195 bp->b_resid = 0; 196 bp->b_flags |= B_AGE; 197 } 198 199 /* 200 * Callback used in read path in case that a block is compressed with ZLIB. 201 * It is almost identical to LZ4 callback, so in theory they can be unified, 202 * but we didn't want to make changes in bio structure for that. 203 */ 204 static 205 void 206 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio) 207 { 208 struct buf *bp; 209 char *compressed_buffer; 210 z_stream strm_decompress; 211 int result; 212 int ret; 213 214 bp = bio->bio_buf; 215 216 KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE); 217 strm_decompress.avail_in = 0; 218 strm_decompress.next_in = Z_NULL; 219 220 ret = inflateInit(&strm_decompress); 221 222 if (ret != Z_OK) 223 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n"); 224 225 compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT); 226 strm_decompress.next_in = __DECONST(char *, data); 227 228 /* XXX supply proper size, subset of device bp */ 229 strm_decompress.avail_in = bytes; 230 strm_decompress.next_out = compressed_buffer; 231 strm_decompress.avail_out = bp->b_bufsize; 232 233 ret = inflate(&strm_decompress, Z_FINISH); 234 if (ret != Z_STREAM_END) { 235 kprintf("HAMMER2 ZLIB: Fatar error during decompression.\n"); 236 bzero(compressed_buffer, bp->b_bufsize); 237 } 238 bcopy(compressed_buffer, bp->b_data, bp->b_bufsize); 239 result = bp->b_bufsize - strm_decompress.avail_out; 240 if (result < bp->b_bufsize) 241 bzero(bp->b_data + result, strm_decompress.avail_out); 242 objcache_put(cache_buffer_read, compressed_buffer); 243 ret = inflateEnd(&strm_decompress); 244 245 bp->b_resid = 0; 246 bp->b_flags |= B_AGE; 247 } 248 249 /* 250 * Logical buffer I/O, async read. 251 */ 252 static 253 int 254 hammer2_strategy_read(struct vop_strategy_args *ap) 255 { 256 hammer2_xop_strategy_t *xop; 257 struct buf *bp; 258 struct bio *bio; 259 struct bio *nbio; 260 hammer2_inode_t *ip; 261 hammer2_key_t lbase; 262 263 bio = ap->a_bio; 264 bp = bio->bio_buf; 265 ip = VTOI(ap->a_vp); 266 nbio = push_bio(bio); 267 268 lbase = bio->bio_offset; 269 KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0); 270 271 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_STRATEGY); 272 xop->finished = 0; 273 xop->bio = bio; 274 xop->lbase = lbase; 275 hammer2_mtx_init(&xop->lock, "h2bior"); 276 hammer2_xop_start(&xop->head, hammer2_strategy_xop_read); 277 /* asynchronous completion */ 278 279 return(0); 280 } 281 282 /* 283 * Per-node XOP (threaded), do a synchronous lookup of the chain and 284 * its data. The frontend is asynchronous, so we are also responsible 285 * for racing to terminate the frontend. 286 */ 287 static 288 void 289 hammer2_strategy_xop_read(hammer2_thread_t *thr, hammer2_xop_t *arg) 290 { 291 hammer2_xop_strategy_t *xop = &arg->xop_strategy; 292 hammer2_chain_t *parent; 293 hammer2_chain_t *chain; 294 hammer2_chain_t *focus; 295 hammer2_key_t key_dummy; 296 hammer2_key_t lbase; 297 struct bio *bio; 298 struct buf *bp; 299 const char *data; 300 int error; 301 302 /* 303 * Note that we can race completion of the bio supplied by 304 * the front-end so we cannot access it until we determine 305 * that we are the ones finishing it up. 306 */ 307 lbase = xop->lbase; 308 309 /* 310 * This is difficult to optimize. The logical buffer might be 311 * partially dirty (contain dummy zero-fill pages), which would 312 * mess up our crc calculation if we were to try a direct read. 313 * So for now we always double-buffer through the underlying 314 * storage. 315 * 316 * If not for the above problem we could conditionalize on 317 * (1) 64KB buffer, (2) one chain (not multi-master) and 318 * (3) !hammer2_double_buffer, and issue a direct read into the 319 * logical buffer. 320 */ 321 parent = hammer2_inode_chain(xop->head.ip1, thr->clindex, 322 HAMMER2_RESOLVE_ALWAYS | 323 HAMMER2_RESOLVE_SHARED); 324 if (parent) { 325 chain = hammer2_chain_lookup(&parent, &key_dummy, 326 lbase, lbase, 327 &error, 328 HAMMER2_LOOKUP_ALWAYS | 329 HAMMER2_LOOKUP_SHARED); 330 if (chain) 331 error = chain->error; 332 } else { 333 error = HAMMER2_ERROR_EIO; 334 chain = NULL; 335 } 336 error = hammer2_xop_feed(&xop->head, chain, thr->clindex, error); 337 if (chain) { 338 hammer2_chain_unlock(chain); 339 hammer2_chain_drop(chain); 340 } 341 if (parent) { 342 hammer2_chain_unlock(parent); 343 hammer2_chain_drop(parent); 344 } 345 chain = NULL; /* safety */ 346 parent = NULL; /* safety */ 347 348 /* 349 * Race to finish the frontend. First-to-complete. bio is only 350 * valid if we are determined to be the ones able to complete 351 * the operation. 352 */ 353 if (xop->finished) 354 return; 355 hammer2_mtx_ex(&xop->lock); 356 if (xop->finished) { 357 hammer2_mtx_unlock(&xop->lock); 358 return; 359 } 360 bio = xop->bio; 361 bp = bio->bio_buf; 362 bkvasync(bp); 363 364 /* 365 * Async operation has not completed and we now own the lock. 366 * Determine if we can complete the operation by issuing the 367 * frontend collection non-blocking. 368 * 369 * H2 double-buffers the data, setting B_NOTMETA on the logical 370 * buffer hints to the OS that the logical buffer should not be 371 * swapcached (since the device buffer can be). 372 * 373 * Also note that even for compressed data we would rather the 374 * kernel cache/swapcache device buffers more and (decompressed) 375 * logical buffers less, since that will significantly improve 376 * the amount of end-user data that can be cached. 377 * 378 * NOTE: The chain->data for xop->head.cluster.focus will be 379 * synchronized to the current cpu by xop_collect(), 380 * but other chains in the cluster might not be. 381 */ 382 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT); 383 384 switch(error) { 385 case 0: 386 xop->finished = 1; 387 hammer2_mtx_unlock(&xop->lock); 388 bp->b_flags |= B_NOTMETA; 389 focus = xop->head.cluster.focus; 390 data = hammer2_xop_gdata(&xop->head)->buf; 391 hammer2_strategy_read_completion(focus, data, xop->bio); 392 hammer2_xop_pdata(&xop->head); 393 biodone(bio); 394 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 395 break; 396 case HAMMER2_ERROR_ENOENT: 397 xop->finished = 1; 398 hammer2_mtx_unlock(&xop->lock); 399 bp->b_flags |= B_NOTMETA; 400 bp->b_resid = 0; 401 bp->b_error = 0; 402 bzero(bp->b_data, bp->b_bcount); 403 biodone(bio); 404 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 405 break; 406 case HAMMER2_ERROR_EINPROGRESS: 407 hammer2_mtx_unlock(&xop->lock); 408 break; 409 default: 410 kprintf("strategy_xop_read: error %08x loff=%016jx\n", 411 error, bp->b_loffset); 412 xop->finished = 1; 413 hammer2_mtx_unlock(&xop->lock); 414 bp->b_flags |= B_ERROR; 415 bp->b_error = EIO; 416 biodone(bio); 417 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 418 break; 419 } 420 } 421 422 static 423 void 424 hammer2_strategy_read_completion(hammer2_chain_t *focus, const char *data, 425 struct bio *bio) 426 { 427 struct buf *bp = bio->bio_buf; 428 429 if (focus->bref.type == HAMMER2_BREF_TYPE_INODE) { 430 /* 431 * Copy from in-memory inode structure. 432 */ 433 bcopy(((const hammer2_inode_data_t *)data)->u.data, 434 bp->b_data, HAMMER2_EMBEDDED_BYTES); 435 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES, 436 bp->b_bcount - HAMMER2_EMBEDDED_BYTES); 437 bp->b_resid = 0; 438 bp->b_error = 0; 439 } else if (focus->bref.type == HAMMER2_BREF_TYPE_DATA) { 440 /* 441 * Data is on-media, record for live dedup. Release the 442 * chain (try to free it) when done. The data is still 443 * cached by both the buffer cache in front and the 444 * block device behind us. This leaves more room in the 445 * LRU chain cache for meta-data chains which we really 446 * want to retain. 447 * 448 * NOTE: Deduplication cannot be safely recorded for 449 * records without a check code. 450 */ 451 hammer2_dedup_record(focus, NULL, data); 452 atomic_set_int(&focus->flags, HAMMER2_CHAIN_RELEASE); 453 454 /* 455 * Decompression and copy. 456 */ 457 switch (HAMMER2_DEC_COMP(focus->bref.methods)) { 458 case HAMMER2_COMP_LZ4: 459 hammer2_decompress_LZ4_callback(data, focus->bytes, 460 bio); 461 /* b_resid set by call */ 462 break; 463 case HAMMER2_COMP_ZLIB: 464 hammer2_decompress_ZLIB_callback(data, focus->bytes, 465 bio); 466 /* b_resid set by call */ 467 break; 468 case HAMMER2_COMP_NONE: 469 KKASSERT(focus->bytes <= bp->b_bcount); 470 bcopy(data, bp->b_data, focus->bytes); 471 if (focus->bytes < bp->b_bcount) { 472 bzero(bp->b_data + focus->bytes, 473 bp->b_bcount - focus->bytes); 474 } 475 bp->b_resid = 0; 476 bp->b_error = 0; 477 break; 478 default: 479 panic("hammer2_strategy_read: " 480 "unknown compression type"); 481 } 482 } else { 483 panic("hammer2_strategy_read: unknown bref type"); 484 } 485 } 486 487 /**************************************************************************** 488 * WRITE SUPPORT * 489 ****************************************************************************/ 490 491 /* 492 * Functions for compression in threads, 493 * from hammer2_vnops.c 494 */ 495 static void hammer2_write_file_core(char *data, hammer2_inode_t *ip, 496 hammer2_chain_t **parentp, 497 hammer2_key_t lbase, int ioflag, int pblksize, 498 hammer2_tid_t mtid, int *errorp); 499 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip, 500 hammer2_chain_t **parentp, 501 hammer2_key_t lbase, int ioflag, int pblksize, 502 hammer2_tid_t mtid, int *errorp, 503 int comp_algo, int check_algo); 504 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip, 505 hammer2_chain_t **parentp, 506 hammer2_key_t lbase, int ioflag, int pblksize, 507 hammer2_tid_t mtid, int *errorp, 508 int check_algo); 509 static int test_block_zeros(const char *buf, size_t bytes); 510 static void zero_write(char *data, hammer2_inode_t *ip, 511 hammer2_chain_t **parentp, 512 hammer2_key_t lbase, 513 hammer2_tid_t mtid, int *errorp); 514 static void hammer2_write_bp(hammer2_chain_t *chain, char *data, 515 int ioflag, int pblksize, 516 hammer2_tid_t mtid, int *errorp, 517 int check_algo); 518 519 static 520 int 521 hammer2_strategy_write(struct vop_strategy_args *ap) 522 { 523 hammer2_xop_strategy_t *xop; 524 hammer2_pfs_t *pmp; 525 struct bio *bio; 526 struct buf *bp; 527 hammer2_inode_t *ip; 528 529 bio = ap->a_bio; 530 bp = bio->bio_buf; 531 ip = VTOI(ap->a_vp); 532 pmp = ip->pmp; 533 534 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 535 hammer2_lwinprog_ref(pmp); 536 hammer2_trans_assert_strategy(pmp); 537 hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE); 538 539 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | 540 HAMMER2_XOP_STRATEGY); 541 xop->finished = 0; 542 xop->bio = bio; 543 xop->lbase = bio->bio_offset; 544 hammer2_mtx_init(&xop->lock, "h2biow"); 545 hammer2_xop_start(&xop->head, hammer2_strategy_xop_write); 546 /* asynchronous completion */ 547 548 hammer2_lwinprog_wait(pmp, hammer2_flush_pipe); 549 550 return(0); 551 } 552 553 /* 554 * Per-node XOP (threaded). Write the logical buffer to the media. 555 * 556 * This is a bit problematic because there may be multiple target and 557 * any of them may be able to release the bp. In addition, if our 558 * particulr target is offline we don't want to block the bp (and thus 559 * the frontend). To accomplish this we copy the data to the per-thr 560 * scratch buffer. 561 */ 562 static 563 void 564 hammer2_strategy_xop_write(hammer2_thread_t *thr, hammer2_xop_t *arg) 565 { 566 hammer2_xop_strategy_t *xop = &arg->xop_strategy; 567 hammer2_chain_t *parent; 568 hammer2_key_t lbase; 569 hammer2_inode_t *ip; 570 struct bio *bio; 571 struct buf *bp; 572 int error; 573 int lblksize; 574 int pblksize; 575 hammer2_off_t bio_offset; 576 char *bio_data; 577 578 /* 579 * We can only access the bp/bio if the frontend has not yet 580 * completed. 581 */ 582 if (xop->finished) 583 return; 584 hammer2_mtx_sh(&xop->lock); 585 if (xop->finished) { 586 hammer2_mtx_unlock(&xop->lock); 587 return; 588 } 589 590 lbase = xop->lbase; 591 bio = xop->bio; /* ephermal */ 592 bp = bio->bio_buf; /* ephermal */ 593 ip = xop->head.ip1; /* retained by ref */ 594 bio_offset = bio->bio_offset; 595 bio_data = thr->scratch; 596 597 /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */ 598 599 lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL); 600 pblksize = hammer2_calc_physical(ip, lbase); 601 bkvasync(bp); 602 KKASSERT(lblksize <= MAXPHYS); 603 bcopy(bp->b_data, bio_data, lblksize); 604 605 hammer2_mtx_unlock(&xop->lock); 606 bp = NULL; /* safety, illegal to access after unlock */ 607 bio = NULL; /* safety, illegal to access after unlock */ 608 609 /* 610 * Actual operation 611 */ 612 parent = hammer2_inode_chain(ip, thr->clindex, HAMMER2_RESOLVE_ALWAYS); 613 hammer2_write_file_core(bio_data, ip, &parent, 614 lbase, IO_ASYNC, pblksize, 615 xop->head.mtid, &error); 616 if (parent) { 617 hammer2_chain_unlock(parent); 618 hammer2_chain_drop(parent); 619 parent = NULL; /* safety */ 620 } 621 hammer2_xop_feed(&xop->head, NULL, thr->clindex, error); 622 623 /* 624 * Try to complete the operation on behalf of the front-end. 625 */ 626 if (xop->finished) 627 return; 628 hammer2_mtx_ex(&xop->lock); 629 if (xop->finished) { 630 hammer2_mtx_unlock(&xop->lock); 631 return; 632 } 633 634 /* 635 * Async operation has not completed and we now own the lock. 636 * Determine if we can complete the operation by issuing the 637 * frontend collection non-blocking. 638 * 639 * H2 double-buffers the data, setting B_NOTMETA on the logical 640 * buffer hints to the OS that the logical buffer should not be 641 * swapcached (since the device buffer can be). 642 */ 643 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT); 644 645 if (error == HAMMER2_ERROR_EINPROGRESS) { 646 hammer2_mtx_unlock(&xop->lock); 647 return; 648 } 649 650 /* 651 * Async operation has completed. 652 */ 653 xop->finished = 1; 654 hammer2_mtx_unlock(&xop->lock); 655 656 bio = xop->bio; /* now owned by us */ 657 bp = bio->bio_buf; /* now owned by us */ 658 659 if (error == HAMMER2_ERROR_ENOENT || error == 0) { 660 bp->b_flags |= B_NOTMETA; 661 bp->b_resid = 0; 662 bp->b_error = 0; 663 biodone(bio); 664 } else { 665 kprintf("strategy_xop_write: error %d loff=%016jx\n", 666 error, bp->b_loffset); 667 bp->b_flags |= B_ERROR; 668 bp->b_error = EIO; 669 biodone(bio); 670 } 671 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 672 hammer2_trans_assert_strategy(ip->pmp); 673 hammer2_lwinprog_drop(ip->pmp); 674 hammer2_trans_done(ip->pmp, 0); 675 } 676 677 /* 678 * Wait for pending I/O to complete 679 */ 680 void 681 hammer2_bioq_sync(hammer2_pfs_t *pmp) 682 { 683 hammer2_lwinprog_wait(pmp, 0); 684 } 685 686 /* 687 * Assign physical storage at (cparent, lbase), returning a suitable chain 688 * and setting *errorp appropriately. 689 * 690 * If no error occurs, the returned chain will be in a modified state. 691 * 692 * If an error occurs, the returned chain may or may not be NULL. If 693 * not-null any chain->error (if not 0) will also be rolled up into *errorp. 694 * So the caller only needs to test *errorp. 695 * 696 * cparent can wind up being anything. 697 * 698 * If datap is not NULL, *datap points to the real data we intend to write. 699 * If we can dedup the storage location we set *datap to NULL to indicate 700 * to the caller that a dedup occurred. 701 * 702 * NOTE: Special case for data embedded in inode. 703 */ 704 static 705 hammer2_chain_t * 706 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp, 707 hammer2_key_t lbase, int pblksize, 708 hammer2_tid_t mtid, char **datap, int *errorp) 709 { 710 hammer2_chain_t *chain; 711 hammer2_key_t key_dummy; 712 hammer2_off_t dedup_off; 713 int pradix = hammer2_getradix(pblksize); 714 715 /* 716 * Locate the chain associated with lbase, return a locked chain. 717 * However, do not instantiate any data reference (which utilizes a 718 * device buffer) because we will be using direct IO via the 719 * logical buffer cache buffer. 720 */ 721 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN); 722 723 chain = hammer2_chain_lookup(parentp, &key_dummy, 724 lbase, lbase, 725 errorp, 726 HAMMER2_LOOKUP_NODATA); 727 728 /* 729 * The lookup code should not return a DELETED chain to us, unless 730 * its a short-file embedded in the inode. Then it is possible for 731 * the lookup to return a deleted inode. 732 */ 733 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) && 734 chain->bref.type != HAMMER2_BREF_TYPE_INODE) { 735 kprintf("assign physical deleted chain @ " 736 "%016jx (%016jx.%02x) ip %016jx\n", 737 lbase, chain->bref.data_off, chain->bref.type, 738 ip->meta.inum); 739 Debugger("bleh"); 740 } 741 742 if (chain == NULL) { 743 /* 744 * We found a hole, create a new chain entry. 745 * 746 * NOTE: DATA chains are created without device backing 747 * store (nor do we want any). 748 */ 749 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap, 750 pblksize); 751 *errorp |= hammer2_chain_create(parentp, &chain, 752 ip->pmp, 753 HAMMER2_ENC_CHECK(ip->meta.check_algo) | 754 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE), 755 lbase, HAMMER2_PBUFRADIX, 756 HAMMER2_BREF_TYPE_DATA, 757 pblksize, mtid, 758 dedup_off, 0); 759 if (chain == NULL) 760 goto failed; 761 /*ip->delta_dcount += pblksize;*/ 762 } else if (chain->error == 0) { 763 switch (chain->bref.type) { 764 case HAMMER2_BREF_TYPE_INODE: 765 /* 766 * The data is embedded in the inode, which requires 767 * a bit more finess. 768 */ 769 *errorp |= hammer2_chain_modify_ip(ip, chain, mtid, 0); 770 break; 771 case HAMMER2_BREF_TYPE_DATA: 772 dedup_off = hammer2_dedup_lookup(chain->hmp, datap, 773 pblksize); 774 if (chain->bytes != pblksize) { 775 *errorp |= hammer2_chain_resize(chain, 776 mtid, dedup_off, 777 pradix, 778 HAMMER2_MODIFY_OPTDATA); 779 if (*errorp) 780 break; 781 } 782 783 /* 784 * DATA buffers must be marked modified whether the 785 * data is in a logical buffer or not. We also have 786 * to make this call to fixup the chain data pointers 787 * after resizing in case this is an encrypted or 788 * compressed buffer. 789 */ 790 *errorp |= hammer2_chain_modify(chain, mtid, dedup_off, 791 HAMMER2_MODIFY_OPTDATA); 792 break; 793 default: 794 panic("hammer2_assign_physical: bad type"); 795 /* NOT REACHED */ 796 break; 797 } 798 } else { 799 *errorp = chain->error; 800 } 801 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 802 failed: 803 return (chain); 804 } 805 806 /* 807 * hammer2_write_file_core() - hammer2_write_thread() helper 808 * 809 * The core write function which determines which path to take 810 * depending on compression settings. We also have to locate the 811 * related chains so we can calculate and set the check data for 812 * the blockref. 813 */ 814 static 815 void 816 hammer2_write_file_core(char *data, hammer2_inode_t *ip, 817 hammer2_chain_t **parentp, 818 hammer2_key_t lbase, int ioflag, int pblksize, 819 hammer2_tid_t mtid, int *errorp) 820 { 821 hammer2_chain_t *chain; 822 char *bdata; 823 824 *errorp = 0; 825 826 switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) { 827 case HAMMER2_COMP_NONE: 828 /* 829 * We have to assign physical storage to the buffer 830 * we intend to dirty or write now to avoid deadlocks 831 * in the strategy code later. 832 * 833 * This can return NOOFFSET for inode-embedded data. 834 * The strategy code will take care of it in that case. 835 */ 836 bdata = data; 837 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize, 838 mtid, &bdata, errorp); 839 if (*errorp) { 840 /* skip modifications */ 841 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 842 hammer2_inode_data_t *wipdata; 843 844 wipdata = &chain->data->ipdata; 845 KKASSERT(wipdata->meta.op_flags & 846 HAMMER2_OPFLAG_DIRECTDATA); 847 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 848 ++hammer2_iod_file_wembed; 849 } else if (bdata == NULL) { 850 /* 851 * Copy of data already present on-media. 852 */ 853 chain->bref.methods = 854 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 855 HAMMER2_ENC_CHECK(ip->meta.check_algo); 856 hammer2_chain_setcheck(chain, data); 857 } else { 858 hammer2_write_bp(chain, data, ioflag, pblksize, 859 mtid, errorp, ip->meta.check_algo); 860 } 861 if (chain) { 862 hammer2_chain_unlock(chain); 863 hammer2_chain_drop(chain); 864 } 865 break; 866 case HAMMER2_COMP_AUTOZERO: 867 /* 868 * Check for zero-fill only 869 */ 870 hammer2_zero_check_and_write(data, ip, parentp, 871 lbase, ioflag, pblksize, 872 mtid, errorp, 873 ip->meta.check_algo); 874 break; 875 case HAMMER2_COMP_LZ4: 876 case HAMMER2_COMP_ZLIB: 877 default: 878 /* 879 * Check for zero-fill and attempt compression. 880 */ 881 hammer2_compress_and_write(data, ip, parentp, 882 lbase, ioflag, pblksize, 883 mtid, errorp, 884 ip->meta.comp_algo, 885 ip->meta.check_algo); 886 break; 887 } 888 } 889 890 /* 891 * Helper 892 * 893 * Generic function that will perform the compression in compression 894 * write path. The compression algorithm is determined by the settings 895 * obtained from inode. 896 */ 897 static 898 void 899 hammer2_compress_and_write(char *data, hammer2_inode_t *ip, 900 hammer2_chain_t **parentp, 901 hammer2_key_t lbase, int ioflag, int pblksize, 902 hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo) 903 { 904 hammer2_chain_t *chain; 905 int comp_size; 906 int comp_block_size; 907 char *comp_buffer; 908 char *bdata; 909 910 /* 911 * An all-zeros write creates a hole unless the check code 912 * is disabled. When the check code is disabled all writes 913 * are done in-place, including any all-zeros writes. 914 * 915 * NOTE: A snapshot will still force a copy-on-write 916 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c). 917 */ 918 if (check_algo != HAMMER2_CHECK_NONE && 919 test_block_zeros(data, pblksize)) { 920 zero_write(data, ip, parentp, lbase, mtid, errorp); 921 return; 922 } 923 924 /* 925 * Compression requested. Try to compress the block. We store 926 * the data normally if we cannot sufficiently compress it. 927 * 928 * We have a heuristic to detect files which are mostly 929 * uncompressable and avoid the compression attempt in that 930 * case. If the compression heuristic is turned off, we always 931 * try to compress. 932 */ 933 comp_size = 0; 934 comp_buffer = NULL; 935 936 KKASSERT(pblksize / 2 <= 32768); 937 938 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 || 939 hammer2_always_compress) { 940 z_stream strm_compress; 941 int comp_level; 942 int ret; 943 944 switch(HAMMER2_DEC_ALGO(comp_algo)) { 945 case HAMMER2_COMP_LZ4: 946 /* 947 * We need to prefix with the size, LZ4 948 * doesn't do it for us. Add the related 949 * overhead. 950 * 951 * NOTE: The LZ4 code seems to assume at least an 952 * 8-byte buffer size granularity and may 953 * overrun the buffer if given a 4-byte 954 * granularity. 955 */ 956 comp_buffer = objcache_get(cache_buffer_write, 957 M_INTWAIT); 958 comp_size = LZ4_compress_limitedOutput( 959 data, 960 &comp_buffer[sizeof(int)], 961 pblksize, 962 pblksize / 2 - sizeof(int64_t)); 963 *(int *)comp_buffer = comp_size; 964 if (comp_size) 965 comp_size += sizeof(int); 966 break; 967 case HAMMER2_COMP_ZLIB: 968 comp_level = HAMMER2_DEC_LEVEL(comp_algo); 969 if (comp_level == 0) 970 comp_level = 6; /* default zlib compression */ 971 else if (comp_level < 6) 972 comp_level = 6; 973 else if (comp_level > 9) 974 comp_level = 9; 975 ret = deflateInit(&strm_compress, comp_level); 976 if (ret != Z_OK) { 977 kprintf("HAMMER2 ZLIB: fatal error " 978 "on deflateInit.\n"); 979 } 980 981 comp_buffer = objcache_get(cache_buffer_write, 982 M_INTWAIT); 983 strm_compress.next_in = data; 984 strm_compress.avail_in = pblksize; 985 strm_compress.next_out = comp_buffer; 986 strm_compress.avail_out = pblksize / 2; 987 ret = deflate(&strm_compress, Z_FINISH); 988 if (ret == Z_STREAM_END) { 989 comp_size = pblksize / 2 - 990 strm_compress.avail_out; 991 } else { 992 comp_size = 0; 993 } 994 ret = deflateEnd(&strm_compress); 995 break; 996 default: 997 kprintf("Error: Unknown compression method.\n"); 998 kprintf("Comp_method = %d.\n", comp_algo); 999 break; 1000 } 1001 } 1002 1003 if (comp_size == 0) { 1004 /* 1005 * compression failed or turned off 1006 */ 1007 comp_block_size = pblksize; /* safety */ 1008 if (++ip->comp_heuristic > 128) 1009 ip->comp_heuristic = 8; 1010 } else { 1011 /* 1012 * compression succeeded 1013 */ 1014 ip->comp_heuristic = 0; 1015 if (comp_size <= 1024) { 1016 comp_block_size = 1024; 1017 } else if (comp_size <= 2048) { 1018 comp_block_size = 2048; 1019 } else if (comp_size <= 4096) { 1020 comp_block_size = 4096; 1021 } else if (comp_size <= 8192) { 1022 comp_block_size = 8192; 1023 } else if (comp_size <= 16384) { 1024 comp_block_size = 16384; 1025 } else if (comp_size <= 32768) { 1026 comp_block_size = 32768; 1027 } else { 1028 panic("hammer2: WRITE PATH: " 1029 "Weird comp_size value."); 1030 /* NOT REACHED */ 1031 comp_block_size = pblksize; 1032 } 1033 1034 /* 1035 * Must zero the remainder or dedup (which operates on a 1036 * physical block basis) will not find matches. 1037 */ 1038 if (comp_size < comp_block_size) { 1039 bzero(comp_buffer + comp_size, 1040 comp_block_size - comp_size); 1041 } 1042 } 1043 1044 /* 1045 * Assign physical storage, data will be set to NULL if a live-dedup 1046 * was successful. 1047 */ 1048 bdata = comp_size ? comp_buffer : data; 1049 chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size, 1050 mtid, &bdata, errorp); 1051 1052 if (*errorp) { 1053 goto done; 1054 } 1055 1056 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 1057 hammer2_inode_data_t *wipdata; 1058 1059 *errorp = hammer2_chain_modify_ip(ip, chain, mtid, 0); 1060 if (*errorp == 0) { 1061 wipdata = &chain->data->ipdata; 1062 KKASSERT(wipdata->meta.op_flags & 1063 HAMMER2_OPFLAG_DIRECTDATA); 1064 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1065 ++hammer2_iod_file_wembed; 1066 } 1067 } else if (bdata == NULL) { 1068 /* 1069 * Live deduplication, a copy of the data is already present 1070 * on the media. 1071 */ 1072 if (comp_size) { 1073 chain->bref.methods = 1074 HAMMER2_ENC_COMP(comp_algo) + 1075 HAMMER2_ENC_CHECK(check_algo); 1076 } else { 1077 chain->bref.methods = 1078 HAMMER2_ENC_COMP( 1079 HAMMER2_COMP_NONE) + 1080 HAMMER2_ENC_CHECK(check_algo); 1081 } 1082 bdata = comp_size ? comp_buffer : data; 1083 hammer2_chain_setcheck(chain, bdata); 1084 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1085 } else { 1086 hammer2_io_t *dio; 1087 1088 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED); 1089 1090 switch(chain->bref.type) { 1091 case HAMMER2_BREF_TYPE_INODE: 1092 panic("hammer2_write_bp: unexpected inode\n"); 1093 break; 1094 case HAMMER2_BREF_TYPE_DATA: 1095 /* 1096 * Optimize out the read-before-write 1097 * if possible. 1098 */ 1099 *errorp = hammer2_io_newnz(chain->hmp, 1100 chain->bref.type, 1101 chain->bref.data_off, 1102 chain->bytes, 1103 &dio); 1104 if (*errorp) { 1105 hammer2_io_brelse(&dio); 1106 kprintf("hammer2: WRITE PATH: " 1107 "dbp bread error\n"); 1108 break; 1109 } 1110 bdata = hammer2_io_data(dio, chain->bref.data_off); 1111 1112 /* 1113 * When loading the block make sure we don't 1114 * leave garbage after the compressed data. 1115 */ 1116 if (comp_size) { 1117 chain->bref.methods = 1118 HAMMER2_ENC_COMP(comp_algo) + 1119 HAMMER2_ENC_CHECK(check_algo); 1120 bcopy(comp_buffer, bdata, comp_size); 1121 } else { 1122 chain->bref.methods = 1123 HAMMER2_ENC_COMP( 1124 HAMMER2_COMP_NONE) + 1125 HAMMER2_ENC_CHECK(check_algo); 1126 bcopy(data, bdata, pblksize); 1127 } 1128 1129 /* 1130 * The flush code doesn't calculate check codes for 1131 * file data (doing so can result in excessive I/O), 1132 * so we do it here. 1133 */ 1134 hammer2_chain_setcheck(chain, bdata); 1135 1136 /* 1137 * Device buffer is now valid, chain is no longer in 1138 * the initial state. 1139 * 1140 * (No blockref table worries with file data) 1141 */ 1142 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1143 hammer2_dedup_record(chain, dio, bdata); 1144 1145 /* Now write the related bdp. */ 1146 if (ioflag & IO_SYNC) { 1147 /* 1148 * Synchronous I/O requested. 1149 */ 1150 hammer2_io_bwrite(&dio); 1151 /* 1152 } else if ((ioflag & IO_DIRECT) && 1153 loff + n == pblksize) { 1154 hammer2_io_bdwrite(&dio); 1155 */ 1156 } else if (ioflag & IO_ASYNC) { 1157 hammer2_io_bawrite(&dio); 1158 } else { 1159 hammer2_io_bdwrite(&dio); 1160 } 1161 break; 1162 default: 1163 panic("hammer2_write_bp: bad chain type %d\n", 1164 chain->bref.type); 1165 /* NOT REACHED */ 1166 break; 1167 } 1168 } 1169 done: 1170 if (chain) { 1171 hammer2_chain_unlock(chain); 1172 hammer2_chain_drop(chain); 1173 } 1174 if (comp_buffer) 1175 objcache_put(cache_buffer_write, comp_buffer); 1176 } 1177 1178 /* 1179 * Helper 1180 * 1181 * Function that performs zero-checking and writing without compression, 1182 * it corresponds to default zero-checking path. 1183 */ 1184 static 1185 void 1186 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip, 1187 hammer2_chain_t **parentp, 1188 hammer2_key_t lbase, int ioflag, int pblksize, 1189 hammer2_tid_t mtid, int *errorp, 1190 int check_algo) 1191 { 1192 hammer2_chain_t *chain; 1193 char *bdata; 1194 1195 if (check_algo != HAMMER2_CHECK_NONE && 1196 test_block_zeros(data, pblksize)) { 1197 /* 1198 * An all-zeros write creates a hole unless the check code 1199 * is disabled. When the check code is disabled all writes 1200 * are done in-place, including any all-zeros writes. 1201 * 1202 * NOTE: A snapshot will still force a copy-on-write 1203 * (see the HAMMER2_CHECK_NONE in hammer2_chain.c). 1204 */ 1205 zero_write(data, ip, parentp, lbase, mtid, errorp); 1206 } else { 1207 /* 1208 * Normal write 1209 */ 1210 bdata = data; 1211 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize, 1212 mtid, &bdata, errorp); 1213 if (*errorp) { 1214 /* do nothing */ 1215 } else if (bdata) { 1216 hammer2_write_bp(chain, data, ioflag, pblksize, 1217 mtid, errorp, check_algo); 1218 } else { 1219 /* dedup occurred */ 1220 chain->bref.methods = 1221 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 1222 HAMMER2_ENC_CHECK(check_algo); 1223 hammer2_chain_setcheck(chain, data); 1224 } 1225 if (chain) { 1226 hammer2_chain_unlock(chain); 1227 hammer2_chain_drop(chain); 1228 } 1229 } 1230 } 1231 1232 /* 1233 * Helper 1234 * 1235 * A function to test whether a block of data contains only zeros, 1236 * returns TRUE (non-zero) if the block is all zeros. 1237 */ 1238 static 1239 int 1240 test_block_zeros(const char *buf, size_t bytes) 1241 { 1242 size_t i; 1243 1244 for (i = 0; i < bytes; i += sizeof(long)) { 1245 if (*(const long *)(buf + i) != 0) 1246 return (0); 1247 } 1248 return (1); 1249 } 1250 1251 /* 1252 * Helper 1253 * 1254 * Function to "write" a block that contains only zeros. 1255 */ 1256 static 1257 void 1258 zero_write(char *data, hammer2_inode_t *ip, 1259 hammer2_chain_t **parentp, 1260 hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp) 1261 { 1262 hammer2_chain_t *chain; 1263 hammer2_key_t key_dummy; 1264 1265 chain = hammer2_chain_lookup(parentp, &key_dummy, 1266 lbase, lbase, 1267 errorp, 1268 HAMMER2_LOOKUP_NODATA); 1269 if (chain) { 1270 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 1271 hammer2_inode_data_t *wipdata; 1272 1273 if (*errorp == 0) { 1274 *errorp = hammer2_chain_modify_ip(ip, chain, 1275 mtid, 0); 1276 } 1277 if (*errorp == 0) { 1278 wipdata = &chain->data->ipdata; 1279 KKASSERT(wipdata->meta.op_flags & 1280 HAMMER2_OPFLAG_DIRECTDATA); 1281 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1282 ++hammer2_iod_file_wembed; 1283 } 1284 } else { 1285 /* chain->error ok for deletion */ 1286 hammer2_chain_delete(*parentp, chain, 1287 mtid, HAMMER2_DELETE_PERMANENT); 1288 ++hammer2_iod_file_wzero; 1289 } 1290 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 1291 hammer2_chain_unlock(chain); 1292 hammer2_chain_drop(chain); 1293 } else { 1294 ++hammer2_iod_file_wzero; 1295 } 1296 } 1297 1298 /* 1299 * Helper 1300 * 1301 * Function to write the data as it is, without performing any sort of 1302 * compression. This function is used in path without compression and 1303 * default zero-checking path. 1304 */ 1305 static 1306 void 1307 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag, 1308 int pblksize, 1309 hammer2_tid_t mtid, int *errorp, int check_algo) 1310 { 1311 hammer2_inode_data_t *wipdata; 1312 hammer2_io_t *dio; 1313 char *bdata; 1314 int error; 1315 1316 error = 0; /* XXX TODO below */ 1317 1318 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED); 1319 1320 switch(chain->bref.type) { 1321 case HAMMER2_BREF_TYPE_INODE: 1322 wipdata = &chain->data->ipdata; 1323 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA); 1324 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES); 1325 error = 0; 1326 ++hammer2_iod_file_wembed; 1327 break; 1328 case HAMMER2_BREF_TYPE_DATA: 1329 error = hammer2_io_newnz(chain->hmp, 1330 chain->bref.type, 1331 chain->bref.data_off, 1332 chain->bytes, &dio); 1333 if (error) { 1334 hammer2_io_bqrelse(&dio); 1335 kprintf("hammer2: WRITE PATH: " 1336 "dbp bread error\n"); 1337 break; 1338 } 1339 bdata = hammer2_io_data(dio, chain->bref.data_off); 1340 1341 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) + 1342 HAMMER2_ENC_CHECK(check_algo); 1343 bcopy(data, bdata, chain->bytes); 1344 1345 /* 1346 * The flush code doesn't calculate check codes for 1347 * file data (doing so can result in excessive I/O), 1348 * so we do it here. 1349 */ 1350 hammer2_chain_setcheck(chain, bdata); 1351 1352 /* 1353 * Device buffer is now valid, chain is no longer in 1354 * the initial state. 1355 * 1356 * (No blockref table worries with file data) 1357 */ 1358 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1359 hammer2_dedup_record(chain, dio, bdata); 1360 1361 if (ioflag & IO_SYNC) { 1362 /* 1363 * Synchronous I/O requested. 1364 */ 1365 hammer2_io_bwrite(&dio); 1366 /* 1367 } else if ((ioflag & IO_DIRECT) && 1368 loff + n == pblksize) { 1369 hammer2_io_bdwrite(&dio); 1370 */ 1371 } else if (ioflag & IO_ASYNC) { 1372 hammer2_io_bawrite(&dio); 1373 } else { 1374 hammer2_io_bdwrite(&dio); 1375 } 1376 break; 1377 default: 1378 panic("hammer2_write_bp: bad chain type %d\n", 1379 chain->bref.type); 1380 /* NOT REACHED */ 1381 error = 0; 1382 break; 1383 } 1384 *errorp = error; 1385 } 1386 1387 /* 1388 * LIVE DEDUP HEURISTICS 1389 * 1390 * Record media and crc information for possible dedup operation. Note 1391 * that the dedup mask bits must also be set in the related DIO for a dedup 1392 * to be fully validated (which is handled in the freemap allocation code). 1393 * 1394 * WARNING! This code is SMP safe but the heuristic allows SMP collisions. 1395 * All fields must be loaded into locals and validated. 1396 * 1397 * WARNING! Should only be used for file data and directory entries, 1398 * hammer2_chain_modify() only checks for the dedup case on data 1399 * chains. Also, dedup data can only be recorded for committed 1400 * chains (so NOT strategy writes which can undergo further 1401 * modification after the fact!). 1402 */ 1403 void 1404 hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio, 1405 const char *data) 1406 { 1407 hammer2_dev_t *hmp; 1408 hammer2_dedup_t *dedup; 1409 uint64_t crc; 1410 uint64_t mask; 1411 int best = 0; 1412 int i; 1413 int dticks; 1414 1415 /* 1416 * We can only record a dedup if we have media data to test against. 1417 * If dedup is not enabled, return early, which allows a chain to 1418 * remain marked MODIFIED (which might have benefits in special 1419 * situations, though typically it does not). 1420 */ 1421 if (hammer2_dedup_enable == 0) 1422 return; 1423 if (dio == NULL) { 1424 dio = chain->dio; 1425 if (dio == NULL) 1426 return; 1427 } 1428 1429 hmp = chain->hmp; 1430 1431 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) { 1432 case HAMMER2_CHECK_ISCSI32: 1433 /* 1434 * XXX use the built-in crc (the dedup lookup sequencing 1435 * needs to be fixed so the check code is already present 1436 * when dedup_lookup is called) 1437 */ 1438 #if 0 1439 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value; 1440 #endif 1441 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED); 1442 break; 1443 case HAMMER2_CHECK_XXHASH64: 1444 crc = chain->bref.check.xxhash64.value; 1445 break; 1446 case HAMMER2_CHECK_SHA192: 1447 /* 1448 * XXX use the built-in crc (the dedup lookup sequencing 1449 * needs to be fixed so the check code is already present 1450 * when dedup_lookup is called) 1451 */ 1452 #if 0 1453 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^ 1454 ((uint64_t *)chain->bref.check.sha192.data)[1] ^ 1455 ((uint64_t *)chain->bref.check.sha192.data)[2]; 1456 #endif 1457 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED); 1458 break; 1459 default: 1460 /* 1461 * Cannot dedup without a check code 1462 * 1463 * NOTE: In particular, CHECK_NONE allows a sector to be 1464 * overwritten without copy-on-write, recording 1465 * a dedup block for a CHECK_NONE object would be 1466 * a disaster! 1467 */ 1468 return; 1469 } 1470 1471 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUPABLE); 1472 1473 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)]; 1474 for (i = 0; i < 4; ++i) { 1475 if (dedup[i].data_crc == crc) { 1476 best = i; 1477 break; 1478 } 1479 dticks = (int)(dedup[i].ticks - dedup[best].ticks); 1480 if (dticks < 0 || dticks > hz * 60 * 30) 1481 best = i; 1482 } 1483 dedup += best; 1484 if (hammer2_debug & 0x40000) { 1485 kprintf("REC %04x %016jx %016jx\n", 1486 (int)(dedup - hmp->heur_dedup), 1487 crc, 1488 chain->bref.data_off); 1489 } 1490 dedup->ticks = ticks; 1491 dedup->data_off = chain->bref.data_off; 1492 dedup->data_crc = crc; 1493 1494 /* 1495 * Set the valid bits for the dedup only after we know the data 1496 * buffer has been updated. The alloc bits were set (and the valid 1497 * bits cleared) when the media was allocated. 1498 * 1499 * This is done in two stages becuase the bulkfree code can race 1500 * the gap between allocation and data population. Both masks must 1501 * be set before a bcmp/dedup operation is able to use the block. 1502 */ 1503 mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes); 1504 atomic_set_64(&dio->dedup_valid, mask); 1505 1506 #if 0 1507 /* 1508 * XXX removed. MODIFIED is an integral part of the flush code, 1509 * lets not just clear it 1510 */ 1511 /* 1512 * Once we record the dedup the chain must be marked clean to 1513 * prevent reuse of the underlying block. Remember that this 1514 * write occurs when the buffer cache is flushed (i.e. on sync(), 1515 * fsync(), filesystem periodic sync, or when the kernel needs to 1516 * flush a buffer), and not whenever the user write()s. 1517 */ 1518 if (chain->flags & HAMMER2_CHAIN_MODIFIED) { 1519 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED); 1520 atomic_add_long(&hammer2_count_modified_chains, -1); 1521 if (chain->pmp) 1522 hammer2_pfs_memory_wakeup(chain->pmp); 1523 } 1524 #endif 1525 } 1526 1527 static 1528 hammer2_off_t 1529 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize) 1530 { 1531 hammer2_dedup_t *dedup; 1532 hammer2_io_t *dio; 1533 hammer2_off_t off; 1534 uint64_t crc; 1535 uint64_t mask; 1536 char *data; 1537 char *dtmp; 1538 int i; 1539 1540 if (hammer2_dedup_enable == 0) 1541 return 0; 1542 data = *datap; 1543 if (data == NULL) 1544 return 0; 1545 1546 /* 1547 * XXX use the built-in crc (the dedup lookup sequencing 1548 * needs to be fixed so the check code is already present 1549 * when dedup_lookup is called) 1550 */ 1551 crc = XXH64(data, pblksize, XXH_HAMMER2_SEED); 1552 dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)]; 1553 1554 if (hammer2_debug & 0x40000) { 1555 kprintf("LOC %04x/4 %016jx\n", 1556 (int)(dedup - hmp->heur_dedup), 1557 crc); 1558 } 1559 1560 for (i = 0; i < 4; ++i) { 1561 off = dedup[i].data_off; 1562 cpu_ccfence(); 1563 if (dedup[i].data_crc != crc) 1564 continue; 1565 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize) 1566 continue; 1567 dio = hammer2_io_getquick(hmp, off, pblksize); 1568 if (dio) { 1569 dtmp = hammer2_io_data(dio, off), 1570 mask = hammer2_dedup_mask(dio, off, pblksize); 1571 if ((dio->dedup_alloc & mask) == mask && 1572 (dio->dedup_valid & mask) == mask && 1573 bcmp(data, dtmp, pblksize) == 0) { 1574 if (hammer2_debug & 0x40000) { 1575 kprintf("DEDUP SUCCESS %016jx\n", 1576 (intmax_t)off); 1577 } 1578 hammer2_io_putblk(&dio); 1579 *datap = NULL; 1580 dedup[i].ticks = ticks; /* update use */ 1581 atomic_add_long(&hammer2_iod_file_wdedup, 1582 pblksize); 1583 1584 return off; /* RETURN */ 1585 } 1586 hammer2_io_putblk(&dio); 1587 } 1588 } 1589 return 0; 1590 } 1591 1592 /* 1593 * Poof. Races are ok, if someone gets in and reuses a dedup offset 1594 * before or while we are clearing it they will also recover the freemap 1595 * entry (set it to fully allocated), so a bulkfree race can only set it 1596 * to a possibly-free state. 1597 * 1598 * XXX ok, well, not really sure races are ok but going to run with it 1599 * for the moment. 1600 */ 1601 void 1602 hammer2_dedup_clear(hammer2_dev_t *hmp) 1603 { 1604 int i; 1605 1606 for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) { 1607 hmp->heur_dedup[i].data_off = 0; 1608 hmp->heur_dedup[i].ticks = ticks - 1; 1609 } 1610 } 1611