sys/vfs/hammer2/hammer2_strategy.c

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