sys/vfs/hammer2/hammer2_strategy.c

   1 /*
   2  * Copyright (c) 2011-2015 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  * WARNING: The strategy code cannot safely use hammer2 transactions
  75  *          as this can deadlock against vfs_sync's vfsync() call
  76  *          if multiple flushes are queued.  All H2 structures must
  77  *          already be present and ready for the DIO.
  78  *
  79  *          Reads can be initiated asynchronously, writes have to be
  80  *          spooled to a separate thread for action to avoid deadlocks.
  81  */
  82 static void hammer2_strategy_xop_read(hammer2_xop_t *arg, int clindex);
  83 static void hammer2_strategy_xop_write(hammer2_xop_t *arg, int clindex);
  84 static int hammer2_strategy_read(struct vop_strategy_args *ap);
  85 static int hammer2_strategy_write(struct vop_strategy_args *ap);
  86 static void hammer2_strategy_read_completion(hammer2_chain_t *chain,
  87                                 char *data, struct bio *bio);
  88
  89 static void hammer2_dedup_record(hammer2_chain_t *chain, char *data);
  90 static hammer2_off_t hammer2_dedup_lookup(hammer2_dev_t *hmp,
  91                         char **datap, int pblksize);
  92
  93 int h2timer[32];
  94 int h2last;
  95 int h2lid;
  96
  97 #define TIMER(which)    do {                            \
  98         if (h2last)                                     \
  99                 h2timer[h2lid] += (int)(ticks - h2last);\
 100         h2last = ticks;                                 \
 101         h2lid = which;                                  \
 102 } while(0)
 103
 104 int
 105 hammer2_vop_strategy(struct vop_strategy_args *ap)
 106 {
 107         struct bio *biop;
 108         struct buf *bp;
 109         int error;
 110
 111         biop = ap->a_bio;
 112         bp = biop->bio_buf;
 113
 114         switch(bp->b_cmd) {
 115         case BUF_CMD_READ:
 116                 error = hammer2_strategy_read(ap);
 117                 ++hammer2_iod_file_read;
 118                 break;
 119         case BUF_CMD_WRITE:
 120                 error = hammer2_strategy_write(ap);
 121                 ++hammer2_iod_file_write;
 122                 break;
 123         default:
 124                 bp->b_error = error = EINVAL;
 125                 bp->b_flags |= B_ERROR;
 126                 biodone(biop);
 127                 break;
 128         }
 129         return (error);
 130 }
 131
 132 /*
 133  * Return the largest contiguous physical disk range for the logical
 134  * request, in bytes.
 135  *
 136  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
 137  *
 138  * Basically disabled, the logical buffer write thread has to deal with
 139  * buffers one-at-a-time.  Note that this should not prevent cluster_read()
 140  * from reading-ahead, it simply prevents it from trying form a single
 141  * cluster buffer for the logical request.  H2 already uses 64KB buffers!
 142  */
 143 int
 144 hammer2_vop_bmap(struct vop_bmap_args *ap)
 145 {
 146         *ap->a_doffsetp = NOOFFSET;
 147         if (ap->a_runp)
 148                 *ap->a_runp = 0;
 149         if (ap->a_runb)
 150                 *ap->a_runb = 0;
 151         return (EOPNOTSUPP);
 152 }
 153
 154 /****************************************************************************
 155  *                              READ SUPPORT                                *
 156  ****************************************************************************/
 157 /*
 158  * Callback used in read path in case that a block is compressed with LZ4.
 159  */
 160 static
 161 void
 162 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio)
 163 {
 164         struct buf *bp;
 165         char *compressed_buffer;
 166         int compressed_size;
 167         int result;
 168
 169         bp = bio->bio_buf;
 170
 171 #if 0
 172         if bio->bio_caller_info2.index &&
 173               bio->bio_caller_info1.uvalue32 !=
 174               crc32(bp->b_data, bp->b_bufsize) --- return error
 175 #endif
 176
 177         KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
 178         compressed_size = *(const int *)data;
 179         KKASSERT((uint32_t)compressed_size <= bytes - sizeof(int));
 180
 181         compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
 182         result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]),
 183                                      compressed_buffer,
 184                                      compressed_size,
 185                                      bp->b_bufsize);
 186         if (result < 0) {
 187                 kprintf("READ PATH: Error during decompression."
 188                         "bio %016jx/%d\n",
 189                         (intmax_t)bio->bio_offset, bytes);
 190                 /* make sure it isn't random garbage */
 191                 bzero(compressed_buffer, bp->b_bufsize);
 192         }
 193         KKASSERT(result <= bp->b_bufsize);
 194         bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
 195         if (result < bp->b_bufsize)
 196                 bzero(bp->b_data + result, bp->b_bufsize - result);
 197         objcache_put(cache_buffer_read, compressed_buffer);
 198         bp->b_resid = 0;
 199         bp->b_flags |= B_AGE;
 200 }
 201
 202 /*
 203  * Callback used in read path in case that a block is compressed with ZLIB.
 204  * It is almost identical to LZ4 callback, so in theory they can be unified,
 205  * but we didn't want to make changes in bio structure for that.
 206  */
 207 static
 208 void
 209 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio)
 210 {
 211         struct buf *bp;
 212         char *compressed_buffer;
 213         z_stream strm_decompress;
 214         int result;
 215         int ret;
 216
 217         bp = bio->bio_buf;
 218
 219         KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
 220         strm_decompress.avail_in = 0;
 221         strm_decompress.next_in = Z_NULL;
 222
 223         ret = inflateInit(&strm_decompress);
 224
 225         if (ret != Z_OK)
 226                 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n");
 227
 228         compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
 229         strm_decompress.next_in = __DECONST(char *, data);
 230
 231         /* XXX supply proper size, subset of device bp */
 232         strm_decompress.avail_in = bytes;
 233         strm_decompress.next_out = compressed_buffer;
 234         strm_decompress.avail_out = bp->b_bufsize;
 235
 236         ret = inflate(&strm_decompress, Z_FINISH);
 237         if (ret != Z_STREAM_END) {
 238                 kprintf("HAMMER2 ZLIB: Fatar error during decompression.\n");
 239                 bzero(compressed_buffer, bp->b_bufsize);
 240         }
 241         bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
 242         result = bp->b_bufsize - strm_decompress.avail_out;
 243         if (result < bp->b_bufsize)
 244                 bzero(bp->b_data + result, strm_decompress.avail_out);
 245         objcache_put(cache_buffer_read, compressed_buffer);
 246         ret = inflateEnd(&strm_decompress);
 247
 248         bp->b_resid = 0;
 249         bp->b_flags |= B_AGE;
 250 }
 251
 252 /*
 253  * Logical buffer I/O, async read.
 254  */
 255 static
 256 int
 257 hammer2_strategy_read(struct vop_strategy_args *ap)
 258 {
 259         hammer2_xop_strategy_t *xop;
 260         struct buf *bp;
 261         struct bio *bio;
 262         struct bio *nbio;
 263         hammer2_inode_t *ip;
 264         hammer2_key_t lbase;
 265
 266         bio = ap->a_bio;
 267         bp = bio->bio_buf;
 268         ip = VTOI(ap->a_vp);
 269         nbio = push_bio(bio);
 270
 271         lbase = bio->bio_offset;
 272         KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0);
 273
 274         if (bp->b_bio1.bio_flags & BIO_SYNC) {
 275                 xop = hammer2_xop_alloc(ip, 0);
 276         } else {
 277                 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_ITERATOR);
 278         }
 279         xop->finished = 0;
 280         xop->bio = bio;
 281         xop->lbase = lbase;
 282         hammer2_mtx_init(&xop->lock, "h2bior");
 283         hammer2_xop_start(&xop->head, hammer2_strategy_xop_read);
 284         /* asynchronous completion */
 285
 286         return(0);
 287 }
 288
 289 /*
 290  * Per-node XOP (threaded), do a synchronous lookup of the chain and
 291  * its data.  The frontend is asynchronous, so we are also responsible
 292  * for racing to terminate the frontend.
 293  */
 294 static
 295 void
 296 hammer2_strategy_xop_read(hammer2_xop_t *arg, int clindex)
 297 {
 298         hammer2_xop_strategy_t *xop = &arg->xop_strategy;
 299         hammer2_chain_t *parent;
 300         hammer2_chain_t *chain;
 301         hammer2_key_t key_dummy;
 302         hammer2_key_t lbase;
 303         struct bio *bio;
 304         struct buf *bp;
 305         int cache_index = -1;
 306         int error;
 307
 308         TIMER(0);
 309         lbase = xop->lbase;
 310         bio = xop->bio;
 311         bp = bio->bio_buf;
 312
 313         /*
 314          * This is difficult to optimize.  The logical buffer might be
 315          * partially dirty (contain dummy zero-fill pages), which would
 316          * mess up our crc calculation if we were to try a direct read.
 317          * So for now we always double-buffer through the underlying
 318          * storage.
 319          *
 320          * If not for the above problem we could conditionalize on
 321          * (1) 64KB buffer, (2) one chain (not multi-master) and
 322          * (3) !hammer2_double_buffer, and issue a direct read into the
 323          * logical buffer.
 324          */
 325         parent = hammer2_inode_chain(xop->head.ip1, clindex,
 326                                      HAMMER2_RESOLVE_ALWAYS |
 327                                      HAMMER2_RESOLVE_SHARED);
 328         TIMER(1);
 329         if (parent) {
 330                 chain = hammer2_chain_lookup(&parent, &key_dummy,
 331                                              lbase, lbase,
 332                                              &cache_index,
 333                                              HAMMER2_LOOKUP_ALWAYS |
 334                                              HAMMER2_LOOKUP_SHARED);
 335                 error = chain ? chain->error : 0;
 336         } else {
 337                 error = EIO;
 338                 chain = NULL;
 339         }
 340         TIMER(2);
 341         error = hammer2_xop_feed(&xop->head, chain, clindex, error);
 342         TIMER(3);
 343         if (chain) {
 344                 hammer2_chain_unlock(chain);
 345                 hammer2_chain_drop(chain);
 346         }
 347         if (parent) {
 348                 hammer2_chain_unlock(parent);
 349                 hammer2_chain_drop(parent);
 350         }
 351         chain = NULL;   /* safety */
 352         parent = NULL;  /* safety */
 353         TIMER(4);
 354
 355         /*
 356          * Race to finish the frontend
 357          */
 358         if (xop->finished)
 359                 return;
 360         hammer2_mtx_ex(&xop->lock);
 361         if (xop->finished) {
 362                 hammer2_mtx_unlock(&xop->lock);
 363                 return;
 364         }
 365
 366         /*
 367          * Async operation has not completed and we now own the lock.
 368          * Determine if we can complete the operation by issuing the
 369          * frontend collection non-blocking.
 370          *
 371          * H2 double-buffers the data, setting B_NOTMETA on the logical
 372          * buffer hints to the OS that the logical buffer should not be
 373          * swapcached (since the device buffer can be).
 374          *
 375          * Also note that even for compressed data we would rather the
 376          * kernel cache/swapcache device buffers more and (decompressed)
 377          * logical buffers less, since that will significantly improve
 378          * the amount of end-user data that can be cached.
 379          */
 380         error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
 381         TIMER(5);
 382
 383         switch(error) {
 384         case 0:
 385                 xop->finished = 1;
 386                 hammer2_mtx_unlock(&xop->lock);
 387                 bp->b_flags |= B_NOTMETA;
 388                 chain = xop->head.cluster.focus;
 389                 hammer2_strategy_read_completion(chain, (char *)chain->data,
 390                                                  xop->bio);
 391                 biodone(bio);
 392                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 393                 break;
 394         case ENOENT:
 395                 xop->finished = 1;
 396                 hammer2_mtx_unlock(&xop->lock);
 397                 bp->b_flags |= B_NOTMETA;
 398                 bp->b_resid = 0;
 399                 bp->b_error = 0;
 400                 bzero(bp->b_data, bp->b_bcount);
 401                 biodone(bio);
 402                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 403                 break;
 404         case EINPROGRESS:
 405                 hammer2_mtx_unlock(&xop->lock);
 406                 break;
 407         default:
 408                 kprintf("strategy_xop_read: error %d loff=%016jx\n",
 409                         error, bp->b_loffset);
 410                 xop->finished = 1;
 411                 hammer2_mtx_unlock(&xop->lock);
 412                 bp->b_flags |= B_ERROR;
 413                 bp->b_error = EIO;
 414                 biodone(bio);
 415                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 416                 break;
 417         }
 418         TIMER(6);
 419 }
 420
 421 static
 422 void
 423 hammer2_strategy_read_completion(hammer2_chain_t *chain, char *data,
 424                                  struct bio *bio)
 425 {
 426         struct buf *bp = bio->bio_buf;
 427
 428         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
 429                 /*
 430                  * Data is embedded in the inode (copy from inode).
 431                  */
 432                 bcopy(((hammer2_inode_data_t *)data)->u.data,
 433                       bp->b_data, HAMMER2_EMBEDDED_BYTES);
 434                 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
 435                       bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
 436                 bp->b_resid = 0;
 437                 bp->b_error = 0;
 438         } else if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
 439                 /*
 440                  * Data is on-media, record for live dedup.  Release the
 441                  * chain (try to free it) when done.  The data is still
 442                  * cached by both the buffer cache in front and the
 443                  * block device behind us.  This leaves more room in the
 444                  * LRU chain cache for meta-data chains which we really
 445                  * want to retain.
 446                  */
 447                 hammer2_dedup_record(chain, data);
 448                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 449
 450                 /*
 451                  * Decompression and copy.
 452                  */
 453                 switch (HAMMER2_DEC_COMP(chain->bref.methods)) {
 454                 case HAMMER2_COMP_LZ4:
 455                         hammer2_decompress_LZ4_callback(data, chain->bytes,
 456                                                         bio);
 457                         /* b_resid set by call */
 458                         break;
 459                 case HAMMER2_COMP_ZLIB:
 460                         hammer2_decompress_ZLIB_callback(data, chain->bytes,
 461                                                          bio);
 462                         /* b_resid set by call */
 463                         break;
 464                 case HAMMER2_COMP_NONE:
 465                         KKASSERT(chain->bytes <= bp->b_bcount);
 466                         bcopy(data, bp->b_data, chain->bytes);
 467                         if (chain->bytes < bp->b_bcount) {
 468                                 bzero(bp->b_data + chain->bytes,
 469                                       bp->b_bcount - chain->bytes);
 470                         }
 471                         bp->b_resid = 0;
 472                         bp->b_error = 0;
 473                         break;
 474                 default:
 475                         panic("hammer2_strategy_read: "
 476                               "unknown compression type");
 477                 }
 478         } else {
 479                 panic("hammer2_strategy_read: unknown bref type");
 480         }
 481 }
 482
 483 /****************************************************************************
 484  *                              WRITE SUPPORT                               *
 485  ****************************************************************************/
 486
 487 /*
 488  * Functions for compression in threads,
 489  * from hammer2_vnops.c
 490  */
 491 static void hammer2_write_file_core(struct buf *bp, hammer2_inode_t *ip,
 492                                 hammer2_chain_t **parentp,
 493                                 hammer2_key_t lbase, int ioflag, int pblksize,
 494                                 hammer2_tid_t mtid, int *errorp);
 495 static void hammer2_compress_and_write(struct buf *bp, 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                                 int comp_algo, int check_algo);
 500 static void hammer2_zero_check_and_write(struct buf *bp, hammer2_inode_t *ip,
 501                                 hammer2_chain_t **parentp,
 502                                 hammer2_key_t lbase, int ioflag, int pblksize,
 503                                 hammer2_tid_t mtid, int *errorp,
 504                                 int check_algo);
 505 static int test_block_zeros(const char *buf, size_t bytes);
 506 static void zero_write(struct buf *bp, hammer2_inode_t *ip,
 507                                 hammer2_chain_t **parentp,
 508                                 hammer2_key_t lbase,
 509                                 hammer2_tid_t mtid, int *errorp);
 510 static void hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp,
 511                                 int ioflag, int pblksize,
 512                                 hammer2_tid_t mtid, int *errorp,
 513                                 int check_algo);
 514
 515 static
 516 int
 517 hammer2_strategy_write(struct vop_strategy_args *ap)
 518 {
 519         hammer2_xop_strategy_t *xop;
 520         hammer2_pfs_t *pmp;
 521         struct bio *bio;
 522         struct buf *bp;
 523         hammer2_inode_t *ip;
 524
 525         bio = ap->a_bio;
 526         bp = bio->bio_buf;
 527         ip = VTOI(ap->a_vp);
 528         pmp = ip->pmp;
 529
 530         hammer2_lwinprog_ref(pmp);
 531         hammer2_trans_assert_strategy(pmp);
 532
 533         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
 534         xop->finished = 0;
 535         xop->bio = bio;
 536         xop->lbase = bio->bio_offset;
 537         hammer2_mtx_init(&xop->lock, "h2biow");
 538         hammer2_xop_start(&xop->head, hammer2_strategy_xop_write);
 539         /* asynchronous completion */
 540
 541         hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
 542
 543         return(0);
 544 }
 545
 546 /*
 547  * Per-node XOP (threaded).  Write the logical buffer to the media.
 548  */
 549 static
 550 void
 551 hammer2_strategy_xop_write(hammer2_xop_t *arg, int clindex)
 552 {
 553         hammer2_xop_strategy_t *xop = &arg->xop_strategy;
 554         hammer2_chain_t *parent;
 555         hammer2_key_t lbase;
 556         hammer2_inode_t *ip;
 557         struct bio *bio;
 558         struct buf *bp;
 559         int error;
 560         int lblksize;
 561         int pblksize;
 562
 563         lbase = xop->lbase;
 564         bio = xop->bio;
 565         bp = bio->bio_buf;
 566         ip = xop->head.ip1;
 567
 568         /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
 569
 570         lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
 571         pblksize = hammer2_calc_physical(ip, lbase);
 572         parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
 573         hammer2_write_file_core(bp, ip, &parent,
 574                                 lbase, IO_ASYNC, pblksize,
 575                                 xop->head.mtid, &error);
 576         if (parent) {
 577                 hammer2_chain_unlock(parent);
 578                 hammer2_chain_drop(parent);
 579                 parent = NULL;  /* safety */
 580         }
 581         hammer2_xop_feed(&xop->head, NULL, clindex, error);
 582
 583         /*
 584          * Race to finish the frontend
 585          */
 586         if (xop->finished)
 587                 return;
 588         hammer2_mtx_ex(&xop->lock);
 589         if (xop->finished) {
 590                 hammer2_mtx_unlock(&xop->lock);
 591                 return;
 592         }
 593
 594         /*
 595          * Async operation has not completed and we now own the lock.
 596          * Determine if we can complete the operation by issuing the
 597          * frontend collection non-blocking.
 598          *
 599          * H2 double-buffers the data, setting B_NOTMETA on the logical
 600          * buffer hints to the OS that the logical buffer should not be
 601          * swapcached (since the device buffer can be).
 602          */
 603         error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
 604
 605         switch(error) {
 606         case ENOENT:
 607         case 0:
 608                 xop->finished = 1;
 609                 hammer2_mtx_unlock(&xop->lock);
 610                 bp->b_flags |= B_NOTMETA;
 611                 bp->b_resid = 0;
 612                 bp->b_error = 0;
 613                 biodone(bio);
 614                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 615                 hammer2_lwinprog_drop(ip->pmp);
 616                 break;
 617         case EINPROGRESS:
 618                 hammer2_mtx_unlock(&xop->lock);
 619                 break;
 620         default:
 621                 kprintf("strategy_xop_write: error %d loff=%016jx\n",
 622                         error, bp->b_loffset);
 623                 xop->finished = 1;
 624                 hammer2_mtx_unlock(&xop->lock);
 625                 bp->b_flags |= B_ERROR;
 626                 bp->b_error = EIO;
 627                 biodone(bio);
 628                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 629                 hammer2_lwinprog_drop(ip->pmp);
 630                 break;
 631         }
 632 }
 633
 634 /*
 635  * Wait for pending I/O to complete
 636  */
 637 void
 638 hammer2_bioq_sync(hammer2_pfs_t *pmp)
 639 {
 640         hammer2_lwinprog_wait(pmp, 0);
 641 }
 642
 643 /*
 644  * Create a new cluster at (cparent, lbase) and assign physical storage,
 645  * returning a cluster suitable for I/O.  The cluster will be in a modified
 646  * state.
 647  *
 648  * cparent can wind up being anything.
 649  *
 650  * If datap is not NULL, *datap points to the real data we intend to write.
 651  * If we can dedup the storage location we set *datap to NULL to indicate
 652  * to the caller that a dedup occurred.
 653  *
 654  * NOTE: Special case for data embedded in inode.
 655  */
 656 static
 657 hammer2_chain_t *
 658 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
 659                         hammer2_key_t lbase, int pblksize,
 660                         hammer2_tid_t mtid, char **datap, int *errorp)
 661 {
 662         hammer2_chain_t *chain;
 663         hammer2_key_t key_dummy;
 664         hammer2_off_t dedup_off;
 665         int pradix = hammer2_getradix(pblksize);
 666         int cache_index = -1;
 667
 668         /*
 669          * Locate the chain associated with lbase, return a locked chain.
 670          * However, do not instantiate any data reference (which utilizes a
 671          * device buffer) because we will be using direct IO via the
 672          * logical buffer cache buffer.
 673          */
 674         *errorp = 0;
 675         KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
 676 retry:
 677         TIMER(30);
 678         chain = hammer2_chain_lookup(parentp, &key_dummy,
 679                                      lbase, lbase,
 680                                      &cache_index,
 681                                      HAMMER2_LOOKUP_NODATA);
 682
 683         if (chain && (chain->flags & HAMMER2_CHAIN_DELETED))
 684                 kprintf("assign physical deleted chain @ %016jx\n",
 685                         lbase);
 686
 687         if (chain == NULL) {
 688                 /*
 689                  * We found a hole, create a new chain entry.
 690                  *
 691                  * NOTE: DATA chains are created without device backing
 692                  *       store (nor do we want any).
 693                  */
 694                 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap,
 695                                                  pblksize);
 696                 *errorp = hammer2_chain_create(parentp, &chain, ip->pmp,
 697                                                lbase, HAMMER2_PBUFRADIX,
 698                                                HAMMER2_BREF_TYPE_DATA,
 699                                                pblksize, mtid,
 700                                                dedup_off, 0);
 701                 if (chain == NULL) {
 702                         panic("hammer2_chain_create: par=%p error=%d\n",
 703                               *parentp, *errorp);
 704                         goto retry;
 705                 }
 706                 /*ip->delta_dcount += pblksize;*/
 707         } else {
 708                 switch (chain->bref.type) {
 709                 case HAMMER2_BREF_TYPE_INODE:
 710                         /*
 711                          * The data is embedded in the inode, which requires
 712                          * a bit more finess.
 713                          */
 714                         hammer2_chain_modify_ip(ip, chain, mtid, 0);
 715                         break;
 716                 case HAMMER2_BREF_TYPE_DATA:
 717                         dedup_off = hammer2_dedup_lookup(chain->hmp, datap,
 718                                                          pblksize);
 719                         if (chain->bytes != pblksize) {
 720                                 hammer2_chain_resize(ip, *parentp, chain,
 721                                                      mtid, dedup_off,
 722                                                      pradix,
 723                                                      HAMMER2_MODIFY_OPTDATA);
 724                         }
 725
 726                         /*
 727                          * DATA buffers must be marked modified whether the
 728                          * data is in a logical buffer or not.  We also have
 729                          * to make this call to fixup the chain data pointers
 730                          * after resizing in case this is an encrypted or
 731                          * compressed buffer.
 732                          */
 733                         hammer2_chain_modify(chain, mtid, dedup_off,
 734                                              HAMMER2_MODIFY_OPTDATA);
 735                         break;
 736                 default:
 737                         panic("hammer2_assign_physical: bad type");
 738                         /* NOT REACHED */
 739                         break;
 740                 }
 741         }
 742         TIMER(31);
 743         return (chain);
 744 }
 745
 746 /*
 747  * hammer2_write_file_core() - hammer2_write_thread() helper
 748  *
 749  * The core write function which determines which path to take
 750  * depending on compression settings.  We also have to locate the
 751  * related chains so we can calculate and set the check data for
 752  * the blockref.
 753  */
 754 static
 755 void
 756 hammer2_write_file_core(struct buf *bp, hammer2_inode_t *ip,
 757                         hammer2_chain_t **parentp,
 758                         hammer2_key_t lbase, int ioflag, int pblksize,
 759                         hammer2_tid_t mtid, int *errorp)
 760 {
 761         hammer2_chain_t *chain;
 762         char *data = bp->b_data;
 763
 764         *errorp = 0;
 765
 766         switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
 767         case HAMMER2_COMP_NONE:
 768                 /*
 769                  * We have to assign physical storage to the buffer
 770                  * we intend to dirty or write now to avoid deadlocks
 771                  * in the strategy code later.
 772                  *
 773                  * This can return NOOFFSET for inode-embedded data.
 774                  * The strategy code will take care of it in that case.
 775                  */
 776                 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
 777                                                 mtid, &data, errorp);
 778                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
 779                         hammer2_inode_data_t *wipdata;
 780
 781                         wipdata = &chain->data->ipdata;
 782                         KKASSERT(wipdata->meta.op_flags &
 783                                  HAMMER2_OPFLAG_DIRECTDATA);
 784                         KKASSERT(bp->b_loffset == 0);
 785                         bcopy(bp->b_data, wipdata->u.data,
 786                               HAMMER2_EMBEDDED_BYTES);
 787                         ++hammer2_iod_file_wembed;
 788                 } else if (data == NULL) {
 789                         /*
 790                          * Copy of data already present on-media.
 791                          */
 792                         chain->bref.methods =
 793                                 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
 794                                 HAMMER2_ENC_CHECK(ip->meta.check_algo);
 795                         hammer2_chain_setcheck(chain, bp->b_data);
 796                 } else {
 797                         hammer2_write_bp(chain, bp, ioflag, pblksize,
 798                                          mtid, errorp, ip->meta.check_algo);
 799                 }
 800                 if (chain) {
 801                         hammer2_chain_unlock(chain);
 802                         hammer2_chain_drop(chain);
 803                 }
 804                 break;
 805         case HAMMER2_COMP_AUTOZERO:
 806                 /*
 807                  * Check for zero-fill only
 808                  */
 809                 hammer2_zero_check_and_write(bp, ip, parentp,
 810                                              lbase, ioflag, pblksize,
 811                                              mtid, errorp,
 812                                              ip->meta.check_algo);
 813                 break;
 814         case HAMMER2_COMP_LZ4:
 815         case HAMMER2_COMP_ZLIB:
 816         default:
 817                 /*
 818                  * Check for zero-fill and attempt compression.
 819                  */
 820                 hammer2_compress_and_write(bp, ip, parentp,
 821                                            lbase, ioflag, pblksize,
 822                                            mtid, errorp,
 823                                            ip->meta.comp_algo,
 824                                            ip->meta.check_algo);
 825                 break;
 826         }
 827 }
 828
 829 /*
 830  * Helper
 831  *
 832  * Generic function that will perform the compression in compression
 833  * write path. The compression algorithm is determined by the settings
 834  * obtained from inode.
 835  */
 836 static
 837 void
 838 hammer2_compress_and_write(struct buf *bp, hammer2_inode_t *ip,
 839         hammer2_chain_t **parentp,
 840         hammer2_key_t lbase, int ioflag, int pblksize,
 841         hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
 842 {
 843         hammer2_chain_t *chain;
 844         int comp_size;
 845         int comp_block_size;
 846         char *comp_buffer;
 847         char *data;
 848
 849         if (test_block_zeros(bp->b_data, pblksize)) {
 850                 zero_write(bp, ip, parentp, lbase, mtid, errorp);
 851                 return;
 852         }
 853
 854         comp_size = 0;
 855         comp_buffer = NULL;
 856
 857         KKASSERT(pblksize / 2 <= 32768);
 858
 859         if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
 860                 z_stream strm_compress;
 861                 int comp_level;
 862                 int ret;
 863
 864                 switch(HAMMER2_DEC_ALGO(comp_algo)) {
 865                 case HAMMER2_COMP_LZ4:
 866                         comp_buffer = objcache_get(cache_buffer_write,
 867                                                    M_INTWAIT);
 868                         comp_size = LZ4_compress_limitedOutput(
 869                                         bp->b_data,
 870                                         &comp_buffer[sizeof(int)],
 871                                         pblksize,
 872                                         pblksize / 2 - sizeof(int));
 873                         /*
 874                          * We need to prefix with the size, LZ4
 875                          * doesn't do it for us.  Add the related
 876                          * overhead.
 877                          */
 878                         *(int *)comp_buffer = comp_size;
 879                         if (comp_size)
 880                                 comp_size += sizeof(int);
 881                         break;
 882                 case HAMMER2_COMP_ZLIB:
 883                         comp_level = HAMMER2_DEC_LEVEL(comp_algo);
 884                         if (comp_level == 0)
 885                                 comp_level = 6; /* default zlib compression */
 886                         else if (comp_level < 6)
 887                                 comp_level = 6;
 888                         else if (comp_level > 9)
 889                                 comp_level = 9;
 890                         ret = deflateInit(&strm_compress, comp_level);
 891                         if (ret != Z_OK) {
 892                                 kprintf("HAMMER2 ZLIB: fatal error "
 893                                         "on deflateInit.\n");
 894                         }
 895
 896                         comp_buffer = objcache_get(cache_buffer_write,
 897                                                    M_INTWAIT);
 898                         strm_compress.next_in = bp->b_data;
 899                         strm_compress.avail_in = pblksize;
 900                         strm_compress.next_out = comp_buffer;
 901                         strm_compress.avail_out = pblksize / 2;
 902                         ret = deflate(&strm_compress, Z_FINISH);
 903                         if (ret == Z_STREAM_END) {
 904                                 comp_size = pblksize / 2 -
 905                                             strm_compress.avail_out;
 906                         } else {
 907                                 comp_size = 0;
 908                         }
 909                         ret = deflateEnd(&strm_compress);
 910                         break;
 911                 default:
 912                         kprintf("Error: Unknown compression method.\n");
 913                         kprintf("Comp_method = %d.\n", comp_algo);
 914                         break;
 915                 }
 916         }
 917
 918         if (comp_size == 0) {
 919                 /*
 920                  * compression failed or turned off
 921                  */
 922                 comp_block_size = pblksize;     /* safety */
 923                 if (++ip->comp_heuristic > 128)
 924                         ip->comp_heuristic = 8;
 925         } else {
 926                 /*
 927                  * compression succeeded
 928                  */
 929                 ip->comp_heuristic = 0;
 930                 if (comp_size <= 1024) {
 931                         comp_block_size = 1024;
 932                 } else if (comp_size <= 2048) {
 933                         comp_block_size = 2048;
 934                 } else if (comp_size <= 4096) {
 935                         comp_block_size = 4096;
 936                 } else if (comp_size <= 8192) {
 937                         comp_block_size = 8192;
 938                 } else if (comp_size <= 16384) {
 939                         comp_block_size = 16384;
 940                 } else if (comp_size <= 32768) {
 941                         comp_block_size = 32768;
 942                 } else {
 943                         panic("hammer2: WRITE PATH: "
 944                               "Weird comp_size value.");
 945                         /* NOT REACHED */
 946                         comp_block_size = pblksize;
 947                 }
 948
 949                 /*
 950                  * Must zero the remainder or dedup (which operates on a
 951                  * physical block basis) will not find matches.
 952                  */
 953                 if (comp_size < comp_block_size) {
 954                         bzero(comp_buffer + comp_size,
 955                               comp_block_size - comp_size);
 956                 }
 957         }
 958
 959         /*
 960          * Assign physical storage, data will be set to NULL if a live-dedup
 961          * was successful.
 962          */
 963         data = comp_size ? comp_buffer : bp->b_data;
 964         chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
 965                                         mtid, &data, errorp);
 966
 967         if (*errorp) {
 968                 kprintf("WRITE PATH: An error occurred while "
 969                         "assigning physical space.\n");
 970                 KKASSERT(chain == NULL);
 971                 goto done;
 972         }
 973
 974         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
 975                 hammer2_inode_data_t *wipdata;
 976
 977                 hammer2_chain_modify_ip(ip, chain, mtid, 0);
 978                 wipdata = &chain->data->ipdata;
 979                 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
 980                 KKASSERT(bp->b_loffset == 0);
 981                 bcopy(bp->b_data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
 982                 ++hammer2_iod_file_wembed;
 983         } else if (data == NULL) {
 984                 /*
 985                  * Live deduplication, a copy of the data is already present
 986                  * on the media.
 987                  */
 988                 char *bdata;
 989
 990                 if (comp_size) {
 991                         chain->bref.methods =
 992                                 HAMMER2_ENC_COMP(comp_algo) +
 993                                 HAMMER2_ENC_CHECK(check_algo);
 994                 } else {
 995                         chain->bref.methods =
 996                                 HAMMER2_ENC_COMP(
 997                                         HAMMER2_COMP_NONE) +
 998                                 HAMMER2_ENC_CHECK(check_algo);
 999                 }
1000                 bdata = comp_size ? comp_buffer : bp->b_data;
1001                 hammer2_chain_setcheck(chain, bdata);
1002                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1003         } else {
1004                 hammer2_io_t *dio;
1005                 char *bdata;
1006
1007                 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1008
1009                 switch(chain->bref.type) {
1010                 case HAMMER2_BREF_TYPE_INODE:
1011                         panic("hammer2_write_bp: unexpected inode\n");
1012                         break;
1013                 case HAMMER2_BREF_TYPE_DATA:
1014                         /*
1015                          * Optimize out the read-before-write
1016                          * if possible.
1017                          */
1018                         *errorp = hammer2_io_newnz(chain->hmp,
1019                                                    chain->bref.type,
1020                                                    chain->bref.data_off,
1021                                                    chain->bytes,
1022                                                    &dio);
1023                         if (*errorp) {
1024                                 hammer2_io_brelse(&dio);
1025                                 kprintf("hammer2: WRITE PATH: "
1026                                         "dbp bread error\n");
1027                                 break;
1028                         }
1029                         bdata = hammer2_io_data(dio, chain->bref.data_off);
1030
1031                         /*
1032                          * When loading the block make sure we don't
1033                          * leave garbage after the compressed data.
1034                          */
1035                         if (comp_size) {
1036                                 chain->bref.methods =
1037                                         HAMMER2_ENC_COMP(comp_algo) +
1038                                         HAMMER2_ENC_CHECK(check_algo);
1039                                 bcopy(comp_buffer, bdata, comp_size);
1040                         } else {
1041                                 chain->bref.methods =
1042                                         HAMMER2_ENC_COMP(
1043                                                 HAMMER2_COMP_NONE) +
1044                                         HAMMER2_ENC_CHECK(check_algo);
1045                                 bcopy(bp->b_data, bdata, pblksize);
1046                         }
1047
1048                         /*
1049                          * The flush code doesn't calculate check codes for
1050                          * file data (doing so can result in excessive I/O),
1051                          * so we do it here.
1052                          *
1053                          * Record for dedup only after the DIO's buffer cache
1054                          * buffer has been updated.
1055                          */
1056                         hammer2_chain_setcheck(chain, bdata);
1057                         hammer2_dedup_record(chain, bdata);
1058
1059                         /*
1060                          * Device buffer is now valid, chain is no longer in
1061                          * the initial state.
1062                          *
1063                          * (No blockref table worries with file data)
1064                          */
1065                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1066
1067                         /* Now write the related bdp. */
1068                         if (ioflag & IO_SYNC) {
1069                                 /*
1070                                  * Synchronous I/O requested.
1071                                  */
1072                                 hammer2_io_bwrite(&dio);
1073                         /*
1074                         } else if ((ioflag & IO_DIRECT) &&
1075                                    loff + n == pblksize) {
1076                                 hammer2_io_bdwrite(&dio);
1077                         */
1078                         } else if (ioflag & IO_ASYNC) {
1079                                 hammer2_io_bawrite(&dio);
1080                         } else {
1081                                 hammer2_io_bdwrite(&dio);
1082                         }
1083                         break;
1084                 default:
1085                         panic("hammer2_write_bp: bad chain type %d\n",
1086                                 chain->bref.type);
1087                         /* NOT REACHED */
1088                         break;
1089                 }
1090         }
1091 done:
1092         if (chain) {
1093                 hammer2_chain_unlock(chain);
1094                 hammer2_chain_drop(chain);
1095         }
1096         if (comp_buffer)
1097                 objcache_put(cache_buffer_write, comp_buffer);
1098 }
1099
1100 /*
1101  * Helper
1102  *
1103  * Function that performs zero-checking and writing without compression,
1104  * it corresponds to default zero-checking path.
1105  */
1106 static
1107 void
1108 hammer2_zero_check_and_write(struct buf *bp, hammer2_inode_t *ip,
1109         hammer2_chain_t **parentp,
1110         hammer2_key_t lbase, int ioflag, int pblksize,
1111         hammer2_tid_t mtid, int *errorp,
1112         int check_algo)
1113 {
1114         hammer2_chain_t *chain;
1115         char *data = bp->b_data;
1116
1117         if (test_block_zeros(bp->b_data, pblksize)) {
1118                 zero_write(bp, ip, parentp, lbase, mtid, errorp);
1119         } else {
1120                 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
1121                                                 mtid, &data, errorp);
1122                 if (data) {
1123                         hammer2_write_bp(chain, bp, ioflag, pblksize,
1124                                          mtid, errorp, check_algo);
1125                 } /* else dedup occurred */
1126                 if (chain) {
1127                         hammer2_chain_unlock(chain);
1128                         hammer2_chain_drop(chain);
1129                 }
1130         }
1131 }
1132
1133 /*
1134  * Helper
1135  *
1136  * A function to test whether a block of data contains only zeros,
1137  * returns TRUE (non-zero) if the block is all zeros.
1138  */
1139 static
1140 int
1141 test_block_zeros(const char *buf, size_t bytes)
1142 {
1143         size_t i;
1144
1145         for (i = 0; i < bytes; i += sizeof(long)) {
1146                 if (*(const long *)(buf + i) != 0)
1147                         return (0);
1148         }
1149         return (1);
1150 }
1151
1152 /*
1153  * Helper
1154  *
1155  * Function to "write" a block that contains only zeros.
1156  */
1157 static
1158 void
1159 zero_write(struct buf *bp, hammer2_inode_t *ip,
1160            hammer2_chain_t **parentp,
1161            hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp)
1162 {
1163         hammer2_chain_t *chain;
1164         hammer2_key_t key_dummy;
1165         int cache_index = -1;
1166
1167         *errorp = 0;
1168         chain = hammer2_chain_lookup(parentp, &key_dummy,
1169                                      lbase, lbase,
1170                                      &cache_index,
1171                                      HAMMER2_LOOKUP_NODATA);
1172         if (chain) {
1173                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1174                         hammer2_inode_data_t *wipdata;
1175
1176                         hammer2_chain_modify_ip(ip, chain, mtid, 0);
1177                         wipdata = &chain->data->ipdata;
1178                         KKASSERT(wipdata->meta.op_flags &
1179                                  HAMMER2_OPFLAG_DIRECTDATA);
1180                         KKASSERT(bp->b_loffset == 0);
1181                         bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1182                         ++hammer2_iod_file_wembed;
1183                 } else {
1184                         hammer2_chain_delete(*parentp, chain,
1185                                              mtid, HAMMER2_DELETE_PERMANENT);
1186                         ++hammer2_iod_file_wzero;
1187                 }
1188                 hammer2_chain_unlock(chain);
1189                 hammer2_chain_drop(chain);
1190         } else {
1191                 ++hammer2_iod_file_wzero;
1192         }
1193 }
1194
1195 /*
1196  * Helper
1197  *
1198  * Function to write the data as it is, without performing any sort of
1199  * compression. This function is used in path without compression and
1200  * default zero-checking path.
1201  */
1202 static
1203 void
1204 hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp, int ioflag,
1205                  int pblksize,
1206                  hammer2_tid_t mtid, int *errorp, int check_algo)
1207 {
1208         hammer2_inode_data_t *wipdata;
1209         hammer2_io_t *dio;
1210         char *bdata;
1211         int error;
1212
1213         error = 0;      /* XXX TODO below */
1214
1215         KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1216
1217         switch(chain->bref.type) {
1218         case HAMMER2_BREF_TYPE_INODE:
1219                 wipdata = &chain->data->ipdata;
1220                 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1221                 KKASSERT(bp->b_loffset == 0);
1222                 bcopy(bp->b_data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1223                 error = 0;
1224                 ++hammer2_iod_file_wembed;
1225                 break;
1226         case HAMMER2_BREF_TYPE_DATA:
1227                 error = hammer2_io_newnz(chain->hmp,
1228                                          chain->bref.type,
1229                                          chain->bref.data_off,
1230                                          chain->bytes, &dio);
1231                 if (error) {
1232                         hammer2_io_bqrelse(&dio);
1233                         kprintf("hammer2: WRITE PATH: "
1234                                 "dbp bread error\n");
1235                         break;
1236                 }
1237                 bdata = hammer2_io_data(dio, chain->bref.data_off);
1238
1239                 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1240                                       HAMMER2_ENC_CHECK(check_algo);
1241                 bcopy(bp->b_data, bdata, chain->bytes);
1242
1243                 /*
1244                  * The flush code doesn't calculate check codes for
1245                  * file data (doing so can result in excessive I/O),
1246                  * so we do it here.
1247                  *
1248                  * Record for dedup only after the DIO's buffer cache
1249                  * buffer has been updated.
1250                  */
1251                 hammer2_chain_setcheck(chain, bdata);
1252                 hammer2_dedup_record(chain, bdata);
1253
1254                 /*
1255                  * Device buffer is now valid, chain is no longer in
1256                  * the initial state.
1257                  *
1258                  * (No blockref table worries with file data)
1259                  */
1260                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1261
1262                 if (ioflag & IO_SYNC) {
1263                         /*
1264                          * Synchronous I/O requested.
1265                          */
1266                         hammer2_io_bwrite(&dio);
1267                 /*
1268                 } else if ((ioflag & IO_DIRECT) &&
1269                            loff + n == pblksize) {
1270                         hammer2_io_bdwrite(&dio);
1271                 */
1272                 } else if (ioflag & IO_ASYNC) {
1273                         hammer2_io_bawrite(&dio);
1274                 } else {
1275                         hammer2_io_bdwrite(&dio);
1276                 }
1277                 break;
1278         default:
1279                 panic("hammer2_write_bp: bad chain type %d\n",
1280                       chain->bref.type);
1281                 /* NOT REACHED */
1282                 error = 0;
1283                 break;
1284         }
1285         KKASSERT(error == 0);   /* XXX TODO */
1286         *errorp = error;
1287 }
1288
1289 /*
1290  * LIVE DEDUP HEURISTIC
1291  *
1292  * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1293  *          All fields must be loaded into locals and validated.
1294  */
1295 static
1296 void
1297 hammer2_dedup_record(hammer2_chain_t *chain, char *data)
1298 {
1299         hammer2_dev_t *hmp;
1300         hammer2_dedup_t *dedup;
1301         uint64_t crc;
1302         int best = 0;
1303         int i;
1304         int dticks;
1305
1306         hmp = chain->hmp;
1307
1308         switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
1309         case HAMMER2_CHECK_ISCSI32:
1310                 /*
1311                  * XXX use the built-in crc (the dedup lookup sequencing
1312                  * needs to be fixed so the check code is already present
1313                  * when dedup_lookup is called)
1314                  */
1315 #if 0
1316                 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value;
1317 #endif
1318                 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1319                 break;
1320         case HAMMER2_CHECK_XXHASH64:
1321                 crc = chain->bref.check.xxhash64.value;
1322                 break;
1323         case HAMMER2_CHECK_SHA192:
1324                 /*
1325                  * XXX use the built-in crc (the dedup lookup sequencing
1326                  * needs to be fixed so the check code is already present
1327                  * when dedup_lookup is called)
1328                  */
1329 #if 0
1330                 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^
1331                       ((uint64_t *)chain->bref.check.sha192.data)[1] ^
1332                       ((uint64_t *)chain->bref.check.sha192.data)[2];
1333 #endif
1334                 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1335                 break;
1336         default:
1337                 /*
1338                  * Cannot dedup without a check code
1339                  */
1340                 return;
1341         }
1342         dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1343         for (i = 0; i < 4; ++i) {
1344                 if (dedup[i].data_crc == crc) {
1345                         best = i;
1346                         break;
1347                 }
1348                 dticks = (int)(dedup[i].ticks - dedup[best].ticks);
1349                 if (dticks < 0 || dticks > hz * 60 * 30)
1350                         best = i;
1351         }
1352         dedup += best;
1353         if (hammer2_debug & 0x40000) {
1354                 kprintf("REC %04x %016jx %016jx\n",
1355                         (int)(dedup - hmp->heur_dedup),
1356                         crc,
1357                         chain->bref.data_off);
1358         }
1359         dedup->ticks = ticks;
1360         dedup->data_off = chain->bref.data_off;
1361         dedup->data_crc = crc;
1362         atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUP);
1363 }
1364
1365 static
1366 hammer2_off_t
1367 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize)
1368 {
1369         hammer2_dedup_t *dedup;
1370         hammer2_io_t *dio;
1371         hammer2_off_t off;
1372         uint64_t crc;
1373         char *data;
1374         int i;
1375
1376         data = *datap;
1377         if (data == NULL)
1378                 return 0;
1379
1380         /*
1381          * XXX use the built-in crc (the dedup lookup sequencing
1382          * needs to be fixed so the check code is already present
1383          * when dedup_lookup is called)
1384          */
1385         crc = XXH64(data, pblksize, XXH_HAMMER2_SEED);
1386         dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1387
1388         if (hammer2_debug & 0x40000) {
1389                 kprintf("LOC %04x/4 %016jx\n",
1390                         (int)(dedup - hmp->heur_dedup),
1391                         crc);
1392         }
1393
1394         for (i = 0; i < 4; ++i) {
1395                 off = dedup[i].data_off;
1396                 cpu_ccfence();
1397                 if (dedup[i].data_crc != crc)
1398                         continue;
1399                 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize)
1400                         continue;
1401                 dio = hammer2_io_getquick(hmp, off, pblksize);
1402                 if (dio &&
1403                     bcmp(data, hammer2_io_data(dio, off), pblksize) == 0) {
1404                         if (hammer2_debug & 0x40000) {
1405                                 kprintf("DEDUP SUCCESS %016jx\n",
1406                                         (intmax_t)off);
1407                         }
1408                         hammer2_io_putblk(&dio);
1409                         *datap = NULL;
1410                         dedup[i].ticks = ticks; /* update use */
1411                         ++hammer2_iod_file_wdedup;
1412                         return off;             /* RETURN */
1413                 }
1414                 if (dio)
1415                         hammer2_io_putblk(&dio);
1416         }
1417         return 0;
1418 }
1419
1420 /*
1421  * Poof.  Races are ok, if someone gets in and reuses a dedup offset
1422  * before or while we are clearing it they will also recover the freemap
1423  * entry (set it to fully allocated), so a bulkfree race can only set it
1424  * to a possibly-free state.
1425  *
1426  * XXX ok, well, not really sure races are ok but going to run with it
1427  *     for the moment.
1428  */
1429 void
1430 hammer2_dedup_clear(hammer2_dev_t *hmp)
1431 {
1432         int i;
1433
1434         for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) {
1435                 hmp->heur_dedup[i].data_off = 0;
1436                 hmp->heur_dedup[i].ticks = ticks - 1;
1437         }
1438 }