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  * Except for the transaction init (which should normally not block),
  75  * we essentially run the strategy operation asynchronously via a XOP.
  76  *
  77  * XXX This isn't supposed to be able to deadlock against vfs_sync vfsync()
  78  *     calls but it has in the past when multiple flushes are queued.
  79  *
  80  * XXX We currently terminate the transaction once we get a quorum, otherwise
  81  *     the frontend can stall, but this can leave the remaining nodes with
  82  *     a potential flush conflict.  We need to delay flushes on those nodes
  83  *     until running transactions complete separately from the normal
  84  *     transaction sequencing.  FIXME TODO.
  85  */
  86 static void hammer2_strategy_xop_read(hammer2_thread_t *thr,
  87                                 hammer2_xop_t *arg);
  88 static void hammer2_strategy_xop_write(hammer2_thread_t *thr,
  89                                 hammer2_xop_t *arg);
  90 static int hammer2_strategy_read(struct vop_strategy_args *ap);
  91 static int hammer2_strategy_write(struct vop_strategy_args *ap);
  92 static void hammer2_strategy_read_completion(hammer2_chain_t *chain,
  93                                 char *data, struct bio *bio);
  94
  95 static hammer2_off_t hammer2_dedup_lookup(hammer2_dev_t *hmp,
  96                         char **datap, int pblksize);
  97
  98 int h2timer[32];
  99 int h2last;
 100 int h2lid;
 101
 102 #define TIMER(which)    do {                            \
 103         if (h2last)                                     \
 104                 h2timer[h2lid] += (int)(ticks - h2last);\
 105         h2last = ticks;                                 \
 106         h2lid = which;                                  \
 107 } while(0)
 108
 109 int
 110 hammer2_vop_strategy(struct vop_strategy_args *ap)
 111 {
 112         struct bio *biop;
 113         struct buf *bp;
 114         int error;
 115
 116         biop = ap->a_bio;
 117         bp = biop->bio_buf;
 118
 119         switch(bp->b_cmd) {
 120         case BUF_CMD_READ:
 121                 error = hammer2_strategy_read(ap);
 122                 ++hammer2_iod_file_read;
 123                 break;
 124         case BUF_CMD_WRITE:
 125                 error = hammer2_strategy_write(ap);
 126                 ++hammer2_iod_file_write;
 127                 break;
 128         default:
 129                 bp->b_error = error = EINVAL;
 130                 bp->b_flags |= B_ERROR;
 131                 biodone(biop);
 132                 break;
 133         }
 134         return (error);
 135 }
 136
 137 /*
 138  * Return the largest contiguous physical disk range for the logical
 139  * request, in bytes.
 140  *
 141  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
 142  *
 143  * Basically disabled, the logical buffer write thread has to deal with
 144  * buffers one-at-a-time.  Note that this should not prevent cluster_read()
 145  * from reading-ahead, it simply prevents it from trying form a single
 146  * cluster buffer for the logical request.  H2 already uses 64KB buffers!
 147  */
 148 int
 149 hammer2_vop_bmap(struct vop_bmap_args *ap)
 150 {
 151         *ap->a_doffsetp = NOOFFSET;
 152         if (ap->a_runp)
 153                 *ap->a_runp = 0;
 154         if (ap->a_runb)
 155                 *ap->a_runb = 0;
 156         return (EOPNOTSUPP);
 157 }
 158
 159 /****************************************************************************
 160  *                              READ SUPPORT                                *
 161  ****************************************************************************/
 162 /*
 163  * Callback used in read path in case that a block is compressed with LZ4.
 164  */
 165 static
 166 void
 167 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio)
 168 {
 169         struct buf *bp;
 170         char *compressed_buffer;
 171         int compressed_size;
 172         int result;
 173
 174         bp = bio->bio_buf;
 175
 176 #if 0
 177         if bio->bio_caller_info2.index &&
 178               bio->bio_caller_info1.uvalue32 !=
 179               crc32(bp->b_data, bp->b_bufsize) --- return error
 180 #endif
 181
 182         KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
 183         compressed_size = *(const int *)data;
 184         KKASSERT((uint32_t)compressed_size <= bytes - sizeof(int));
 185
 186         compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
 187         result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]),
 188                                      compressed_buffer,
 189                                      compressed_size,
 190                                      bp->b_bufsize);
 191         if (result < 0) {
 192                 kprintf("READ PATH: Error during decompression."
 193                         "bio %016jx/%d\n",
 194                         (intmax_t)bio->bio_offset, bytes);
 195                 /* make sure it isn't random garbage */
 196                 bzero(compressed_buffer, bp->b_bufsize);
 197         }
 198         KKASSERT(result <= bp->b_bufsize);
 199         bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
 200         if (result < bp->b_bufsize)
 201                 bzero(bp->b_data + result, bp->b_bufsize - result);
 202         objcache_put(cache_buffer_read, compressed_buffer);
 203         bp->b_resid = 0;
 204         bp->b_flags |= B_AGE;
 205 }
 206
 207 /*
 208  * Callback used in read path in case that a block is compressed with ZLIB.
 209  * It is almost identical to LZ4 callback, so in theory they can be unified,
 210  * but we didn't want to make changes in bio structure for that.
 211  */
 212 static
 213 void
 214 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio)
 215 {
 216         struct buf *bp;
 217         char *compressed_buffer;
 218         z_stream strm_decompress;
 219         int result;
 220         int ret;
 221
 222         bp = bio->bio_buf;
 223
 224         KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
 225         strm_decompress.avail_in = 0;
 226         strm_decompress.next_in = Z_NULL;
 227
 228         ret = inflateInit(&strm_decompress);
 229
 230         if (ret != Z_OK)
 231                 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n");
 232
 233         compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
 234         strm_decompress.next_in = __DECONST(char *, data);
 235
 236         /* XXX supply proper size, subset of device bp */
 237         strm_decompress.avail_in = bytes;
 238         strm_decompress.next_out = compressed_buffer;
 239         strm_decompress.avail_out = bp->b_bufsize;
 240
 241         ret = inflate(&strm_decompress, Z_FINISH);
 242         if (ret != Z_STREAM_END) {
 243                 kprintf("HAMMER2 ZLIB: Fatar error during decompression.\n");
 244                 bzero(compressed_buffer, bp->b_bufsize);
 245         }
 246         bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
 247         result = bp->b_bufsize - strm_decompress.avail_out;
 248         if (result < bp->b_bufsize)
 249                 bzero(bp->b_data + result, strm_decompress.avail_out);
 250         objcache_put(cache_buffer_read, compressed_buffer);
 251         ret = inflateEnd(&strm_decompress);
 252
 253         bp->b_resid = 0;
 254         bp->b_flags |= B_AGE;
 255 }
 256
 257 /*
 258  * Logical buffer I/O, async read.
 259  */
 260 static
 261 int
 262 hammer2_strategy_read(struct vop_strategy_args *ap)
 263 {
 264         hammer2_xop_strategy_t *xop;
 265         struct buf *bp;
 266         struct bio *bio;
 267         struct bio *nbio;
 268         hammer2_inode_t *ip;
 269         hammer2_key_t lbase;
 270
 271         bio = ap->a_bio;
 272         bp = bio->bio_buf;
 273         ip = VTOI(ap->a_vp);
 274         nbio = push_bio(bio);
 275
 276         lbase = bio->bio_offset;
 277         KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0);
 278
 279         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_STRATEGY);
 280         xop->finished = 0;
 281         xop->bio = bio;
 282         xop->lbase = lbase;
 283         hammer2_mtx_init(&xop->lock, "h2bior");
 284         hammer2_xop_start(&xop->head, hammer2_strategy_xop_read);
 285         /* asynchronous completion */
 286
 287         return(0);
 288 }
 289
 290 /*
 291  * Per-node XOP (threaded), do a synchronous lookup of the chain and
 292  * its data.  The frontend is asynchronous, so we are also responsible
 293  * for racing to terminate the frontend.
 294  */
 295 static
 296 void
 297 hammer2_strategy_xop_read(hammer2_thread_t *thr, hammer2_xop_t *arg)
 298 {
 299         hammer2_xop_strategy_t *xop = &arg->xop_strategy;
 300         hammer2_chain_t *parent;
 301         hammer2_chain_t *chain;
 302         hammer2_key_t key_dummy;
 303         hammer2_key_t lbase;
 304         struct bio *bio;
 305         struct buf *bp;
 306         int cache_index = -1;
 307         int error;
 308
 309         /*
 310          * Note that we can race completion of the bio supplied by
 311          * the front-end so we cannot access it until we determine
 312          * that we are the ones finishing it up.
 313          */
 314         TIMER(0);
 315         lbase = xop->lbase;
 316
 317         /*
 318          * This is difficult to optimize.  The logical buffer might be
 319          * partially dirty (contain dummy zero-fill pages), which would
 320          * mess up our crc calculation if we were to try a direct read.
 321          * So for now we always double-buffer through the underlying
 322          * storage.
 323          *
 324          * If not for the above problem we could conditionalize on
 325          * (1) 64KB buffer, (2) one chain (not multi-master) and
 326          * (3) !hammer2_double_buffer, and issue a direct read into the
 327          * logical buffer.
 328          */
 329         parent = hammer2_inode_chain(xop->head.ip1, thr->clindex,
 330                                      HAMMER2_RESOLVE_ALWAYS |
 331                                      HAMMER2_RESOLVE_SHARED);
 332         TIMER(1);
 333         if (parent) {
 334                 chain = hammer2_chain_lookup(&parent, &key_dummy,
 335                                              lbase, lbase,
 336                                              &cache_index,
 337                                              HAMMER2_LOOKUP_ALWAYS |
 338                                              HAMMER2_LOOKUP_SHARED);
 339                 error = chain ? chain->error : 0;
 340         } else {
 341                 error = EIO;
 342                 chain = NULL;
 343         }
 344         TIMER(2);
 345         error = hammer2_xop_feed(&xop->head, chain, thr->clindex, error);
 346         TIMER(3);
 347         if (chain) {
 348                 hammer2_chain_unlock(chain);
 349                 hammer2_chain_drop(chain);
 350         }
 351         if (parent) {
 352                 hammer2_chain_unlock(parent);
 353                 hammer2_chain_drop(parent);
 354         }
 355         chain = NULL;   /* safety */
 356         parent = NULL;  /* safety */
 357         TIMER(4);
 358
 359         /*
 360          * Race to finish the frontend.  First-to-complete.  bio is only
 361          * valid if we are determined to be the ones able to complete
 362          * the operation.
 363          */
 364         if (xop->finished)
 365                 return;
 366         hammer2_mtx_ex(&xop->lock);
 367         if (xop->finished) {
 368                 hammer2_mtx_unlock(&xop->lock);
 369                 return;
 370         }
 371         bio = xop->bio;
 372         bp = bio->bio_buf;
 373
 374         /*
 375          * Async operation has not completed and we now own the lock.
 376          * Determine if we can complete the operation by issuing the
 377          * frontend collection non-blocking.
 378          *
 379          * H2 double-buffers the data, setting B_NOTMETA on the logical
 380          * buffer hints to the OS that the logical buffer should not be
 381          * swapcached (since the device buffer can be).
 382          *
 383          * Also note that even for compressed data we would rather the
 384          * kernel cache/swapcache device buffers more and (decompressed)
 385          * logical buffers less, since that will significantly improve
 386          * the amount of end-user data that can be cached.
 387          */
 388         error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
 389         TIMER(5);
 390
 391         switch(error) {
 392         case 0:
 393                 xop->finished = 1;
 394                 hammer2_mtx_unlock(&xop->lock);
 395                 bp->b_flags |= B_NOTMETA;
 396                 chain = xop->head.cluster.focus;
 397                 hammer2_strategy_read_completion(chain, (char *)chain->data,
 398                                                  xop->bio);
 399                 biodone(bio);
 400                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 401                 break;
 402         case ENOENT:
 403                 xop->finished = 1;
 404                 hammer2_mtx_unlock(&xop->lock);
 405                 bp->b_flags |= B_NOTMETA;
 406                 bp->b_resid = 0;
 407                 bp->b_error = 0;
 408                 bzero(bp->b_data, bp->b_bcount);
 409                 biodone(bio);
 410                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 411                 break;
 412         case EINPROGRESS:
 413                 hammer2_mtx_unlock(&xop->lock);
 414                 break;
 415         default:
 416                 kprintf("strategy_xop_read: error %d loff=%016jx\n",
 417                         error, bp->b_loffset);
 418                 xop->finished = 1;
 419                 hammer2_mtx_unlock(&xop->lock);
 420                 bp->b_flags |= B_ERROR;
 421                 bp->b_error = EIO;
 422                 biodone(bio);
 423                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 424                 break;
 425         }
 426         TIMER(6);
 427 }
 428
 429 static
 430 void
 431 hammer2_strategy_read_completion(hammer2_chain_t *chain, char *data,
 432                                  struct bio *bio)
 433 {
 434         struct buf *bp = bio->bio_buf;
 435
 436         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
 437                 /*
 438                  * Copy from in-memory inode structure.
 439                  */
 440                 bcopy(((hammer2_inode_data_t *)data)->u.data,
 441                       bp->b_data, HAMMER2_EMBEDDED_BYTES);
 442                 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
 443                       bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
 444                 bp->b_resid = 0;
 445                 bp->b_error = 0;
 446         } else if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
 447                 /*
 448                  * Data is on-media, record for live dedup.  Release the
 449                  * chain (try to free it) when done.  The data is still
 450                  * cached by both the buffer cache in front and the
 451                  * block device behind us.  This leaves more room in the
 452                  * LRU chain cache for meta-data chains which we really
 453                  * want to retain.
 454                  *
 455                  * NOTE: Deduplication cannot be safely recorded for
 456                  *       records without a check code.
 457                  */
 458                 hammer2_dedup_record(chain, NULL, data);
 459                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 460
 461                 /*
 462                  * Decompression and copy.
 463                  */
 464                 switch (HAMMER2_DEC_COMP(chain->bref.methods)) {
 465                 case HAMMER2_COMP_LZ4:
 466                         hammer2_decompress_LZ4_callback(data, chain->bytes,
 467                                                         bio);
 468                         /* b_resid set by call */
 469                         break;
 470                 case HAMMER2_COMP_ZLIB:
 471                         hammer2_decompress_ZLIB_callback(data, chain->bytes,
 472                                                          bio);
 473                         /* b_resid set by call */
 474                         break;
 475                 case HAMMER2_COMP_NONE:
 476                         KKASSERT(chain->bytes <= bp->b_bcount);
 477                         bcopy(data, bp->b_data, chain->bytes);
 478                         if (chain->bytes < bp->b_bcount) {
 479                                 bzero(bp->b_data + chain->bytes,
 480                                       bp->b_bcount - chain->bytes);
 481                         }
 482                         bp->b_resid = 0;
 483                         bp->b_error = 0;
 484                         break;
 485                 default:
 486                         panic("hammer2_strategy_read: "
 487                               "unknown compression type");
 488                 }
 489         } else {
 490                 panic("hammer2_strategy_read: unknown bref type");
 491         }
 492 }
 493
 494 /****************************************************************************
 495  *                              WRITE SUPPORT                               *
 496  ****************************************************************************/
 497
 498 /*
 499  * Functions for compression in threads,
 500  * from hammer2_vnops.c
 501  */
 502 static void hammer2_write_file_core(char *data, hammer2_inode_t *ip,
 503                                 hammer2_chain_t **parentp,
 504                                 hammer2_key_t lbase, int ioflag, int pblksize,
 505                                 hammer2_tid_t mtid, int *errorp);
 506 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
 507                                 hammer2_chain_t **parentp,
 508                                 hammer2_key_t lbase, int ioflag, int pblksize,
 509                                 hammer2_tid_t mtid, int *errorp,
 510                                 int comp_algo, int check_algo);
 511 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
 512                                 hammer2_chain_t **parentp,
 513                                 hammer2_key_t lbase, int ioflag, int pblksize,
 514                                 hammer2_tid_t mtid, int *errorp,
 515                                 int check_algo);
 516 static int test_block_zeros(const char *buf, size_t bytes);
 517 static void zero_write(char *data, hammer2_inode_t *ip,
 518                                 hammer2_chain_t **parentp,
 519                                 hammer2_key_t lbase,
 520                                 hammer2_tid_t mtid, int *errorp);
 521 static void hammer2_write_bp(hammer2_chain_t *chain, char *data,
 522                                 int ioflag, int pblksize,
 523                                 hammer2_tid_t mtid, int *errorp,
 524                                 int check_algo);
 525
 526 static
 527 int
 528 hammer2_strategy_write(struct vop_strategy_args *ap)
 529 {
 530         hammer2_xop_strategy_t *xop;
 531         hammer2_pfs_t *pmp;
 532         struct bio *bio;
 533         struct buf *bp;
 534         hammer2_inode_t *ip;
 535
 536         bio = ap->a_bio;
 537         bp = bio->bio_buf;
 538         ip = VTOI(ap->a_vp);
 539         pmp = ip->pmp;
 540
 541         hammer2_lwinprog_ref(pmp);
 542         hammer2_trans_assert_strategy(pmp);
 543         hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE);
 544
 545         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
 546                                     HAMMER2_XOP_STRATEGY);
 547         xop->finished = 0;
 548         xop->bio = bio;
 549         xop->lbase = bio->bio_offset;
 550         hammer2_mtx_init(&xop->lock, "h2biow");
 551         hammer2_xop_start(&xop->head, hammer2_strategy_xop_write);
 552         /* asynchronous completion */
 553
 554         hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
 555
 556         return(0);
 557 }
 558
 559 /*
 560  * Per-node XOP (threaded).  Write the logical buffer to the media.
 561  *
 562  * This is a bit problematic because there may be multiple target and
 563  * any of them may be able to release the bp.  In addition, if our
 564  * particulr target is offline we don't want to block the bp (and thus
 565  * the frontend).  To accomplish this we copy the data to the per-thr
 566  * scratch buffer.
 567  */
 568 static
 569 void
 570 hammer2_strategy_xop_write(hammer2_thread_t *thr, hammer2_xop_t *arg)
 571 {
 572         hammer2_xop_strategy_t *xop = &arg->xop_strategy;
 573         hammer2_chain_t *parent;
 574         hammer2_key_t lbase;
 575         hammer2_inode_t *ip;
 576         struct bio *bio;
 577         struct buf *bp;
 578         int error;
 579         int lblksize;
 580         int pblksize;
 581         hammer2_off_t bio_offset;
 582         char *bio_data;
 583
 584         /*
 585          * We can only access the bp/bio if the frontend has not yet
 586          * completed.
 587          */
 588         if (xop->finished)
 589                 return;
 590         hammer2_mtx_sh(&xop->lock);
 591         if (xop->finished) {
 592                 hammer2_mtx_unlock(&xop->lock);
 593                 return;
 594         }
 595
 596         lbase = xop->lbase;
 597         bio = xop->bio;                 /* ephermal */
 598         bp = bio->bio_buf;              /* ephermal */
 599         ip = xop->head.ip1;             /* retained by ref */
 600         bio_offset = bio->bio_offset;
 601         bio_data = thr->scratch;
 602
 603         /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
 604
 605         lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
 606         pblksize = hammer2_calc_physical(ip, lbase);
 607         bcopy(bp->b_data, bio_data, lblksize);
 608
 609         hammer2_mtx_unlock(&xop->lock);
 610         bp = NULL;      /* safety, illegal to access after unlock */
 611         bio = NULL;     /* safety, illegal to access after unlock */
 612
 613         /*
 614          * Actual operation
 615          */
 616         parent = hammer2_inode_chain(ip, thr->clindex, HAMMER2_RESOLVE_ALWAYS);
 617         hammer2_write_file_core(bio_data, ip, &parent,
 618                                 lbase, IO_ASYNC, pblksize,
 619                                 xop->head.mtid, &error);
 620         if (parent) {
 621                 hammer2_chain_unlock(parent);
 622                 hammer2_chain_drop(parent);
 623                 parent = NULL;  /* safety */
 624         }
 625         hammer2_xop_feed(&xop->head, NULL, thr->clindex, error);
 626
 627         /*
 628          * Try to complete the operation on behalf of the front-end.
 629          */
 630         if (xop->finished)
 631                 return;
 632         hammer2_mtx_ex(&xop->lock);
 633         if (xop->finished) {
 634                 hammer2_mtx_unlock(&xop->lock);
 635                 return;
 636         }
 637
 638         /*
 639          * Async operation has not completed and we now own the lock.
 640          * Determine if we can complete the operation by issuing the
 641          * frontend collection non-blocking.
 642          *
 643          * H2 double-buffers the data, setting B_NOTMETA on the logical
 644          * buffer hints to the OS that the logical buffer should not be
 645          * swapcached (since the device buffer can be).
 646          */
 647         error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
 648
 649         if (error == EINPROGRESS) {
 650                 hammer2_mtx_unlock(&xop->lock);
 651                 return;
 652         }
 653
 654         /*
 655          * Async operation has completed.
 656          */
 657         xop->finished = 1;
 658         hammer2_mtx_unlock(&xop->lock);
 659
 660         bio = xop->bio;         /* now owned by us */
 661         bp = bio->bio_buf;      /* now owned by us */
 662
 663         if (error == ENOENT || error == 0) {
 664                 bp->b_flags |= B_NOTMETA;
 665                 bp->b_resid = 0;
 666                 bp->b_error = 0;
 667                 biodone(bio);
 668         } else {
 669                 kprintf("strategy_xop_write: error %d loff=%016jx\n",
 670                         error, bp->b_loffset);
 671                 bp->b_flags |= B_ERROR;
 672                 bp->b_error = EIO;
 673                 biodone(bio);
 674         }
 675         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 676         hammer2_trans_assert_strategy(ip->pmp);
 677         hammer2_lwinprog_drop(ip->pmp);
 678         hammer2_trans_done(ip->pmp);
 679 }
 680
 681 /*
 682  * Wait for pending I/O to complete
 683  */
 684 void
 685 hammer2_bioq_sync(hammer2_pfs_t *pmp)
 686 {
 687         hammer2_lwinprog_wait(pmp, 0);
 688 }
 689
 690 /*
 691  * Create a new cluster at (cparent, lbase) and assign physical storage,
 692  * returning a cluster suitable for I/O.  The cluster will be in a modified
 693  * state.
 694  *
 695  * cparent can wind up being anything.
 696  *
 697  * If datap is not NULL, *datap points to the real data we intend to write.
 698  * If we can dedup the storage location we set *datap to NULL to indicate
 699  * to the caller that a dedup occurred.
 700  *
 701  * NOTE: Special case for data embedded in inode.
 702  */
 703 static
 704 hammer2_chain_t *
 705 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
 706                         hammer2_key_t lbase, int pblksize,
 707                         hammer2_tid_t mtid, char **datap, int *errorp)
 708 {
 709         hammer2_chain_t *chain;
 710         hammer2_key_t key_dummy;
 711         hammer2_off_t dedup_off;
 712         int pradix = hammer2_getradix(pblksize);
 713         int cache_index = -1;
 714
 715         /*
 716          * Locate the chain associated with lbase, return a locked chain.
 717          * However, do not instantiate any data reference (which utilizes a
 718          * device buffer) because we will be using direct IO via the
 719          * logical buffer cache buffer.
 720          */
 721         *errorp = 0;
 722         KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
 723 retry:
 724         TIMER(30);
 725         chain = hammer2_chain_lookup(parentp, &key_dummy,
 726                                      lbase, lbase,
 727                                      &cache_index,
 728                                      HAMMER2_LOOKUP_NODATA);
 729
 730         /*
 731          * The lookup code should not return a DELETED chain to us, unless
 732          * its a short-file embedded in the inode.  Then it is possible for
 733          * the lookup to return a deleted inode.
 734          */
 735         if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) &&
 736             chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
 737                 kprintf("assign physical deleted chain @ "
 738                         "%016jx (%016jx.%02x) ip %016jx\n",
 739                         lbase, chain->bref.data_off, chain->bref.type,
 740                         ip->meta.inum);
 741                 Debugger("bleh");
 742         }
 743
 744         if (chain == NULL) {
 745                 /*
 746                  * We found a hole, create a new chain entry.
 747                  *
 748                  * NOTE: DATA chains are created without device backing
 749                  *       store (nor do we want any).
 750                  */
 751                 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap,
 752                                                  pblksize);
 753                 *errorp = hammer2_chain_create(parentp, &chain,
 754                                                ip->pmp,
 755                                        HAMMER2_ENC_CHECK(ip->meta.check_algo) |
 756                                        HAMMER2_ENC_COMP(HAMMER2_COMP_NONE),
 757                                                lbase, HAMMER2_PBUFRADIX,
 758                                                HAMMER2_BREF_TYPE_DATA,
 759                                                pblksize, mtid,
 760                                                dedup_off, 0);
 761                 if (chain == NULL) {
 762                         panic("hammer2_chain_create: par=%p error=%d\n",
 763                               *parentp, *errorp);
 764                         goto retry;
 765                 }
 766                 /*ip->delta_dcount += pblksize;*/
 767         } else {
 768                 switch (chain->bref.type) {
 769                 case HAMMER2_BREF_TYPE_INODE:
 770                         /*
 771                          * The data is embedded in the inode, which requires
 772                          * a bit more finess.
 773                          */
 774                         hammer2_chain_modify_ip(ip, chain, mtid, 0);
 775                         break;
 776                 case HAMMER2_BREF_TYPE_DATA:
 777                         dedup_off = hammer2_dedup_lookup(chain->hmp, datap,
 778                                                          pblksize);
 779                         if (chain->bytes != pblksize) {
 780                                 hammer2_chain_resize(chain,
 781                                                      mtid, dedup_off,
 782                                                      pradix,
 783                                                      HAMMER2_MODIFY_OPTDATA);
 784                         }
 785
 786                         /*
 787                          * DATA buffers must be marked modified whether the
 788                          * data is in a logical buffer or not.  We also have
 789                          * to make this call to fixup the chain data pointers
 790                          * after resizing in case this is an encrypted or
 791                          * compressed buffer.
 792                          */
 793                         hammer2_chain_modify(chain, mtid, dedup_off,
 794                                              HAMMER2_MODIFY_OPTDATA);
 795                         break;
 796                 default:
 797                         panic("hammer2_assign_physical: bad type");
 798                         /* NOT REACHED */
 799                         break;
 800                 }
 801         }
 802         TIMER(31);
 803         return (chain);
 804 }
 805
 806 /*
 807  * hammer2_write_file_core() - hammer2_write_thread() helper
 808  *
 809  * The core write function which determines which path to take
 810  * depending on compression settings.  We also have to locate the
 811  * related chains so we can calculate and set the check data for
 812  * the blockref.
 813  */
 814 static
 815 void
 816 hammer2_write_file_core(char *data, hammer2_inode_t *ip,
 817                         hammer2_chain_t **parentp,
 818                         hammer2_key_t lbase, int ioflag, int pblksize,
 819                         hammer2_tid_t mtid, int *errorp)
 820 {
 821         hammer2_chain_t *chain;
 822         char *bdata;
 823
 824         *errorp = 0;
 825
 826         switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
 827         case HAMMER2_COMP_NONE:
 828                 /*
 829                  * We have to assign physical storage to the buffer
 830                  * we intend to dirty or write now to avoid deadlocks
 831                  * in the strategy code later.
 832                  *
 833                  * This can return NOOFFSET for inode-embedded data.
 834                  * The strategy code will take care of it in that case.
 835                  */
 836                 bdata = data;
 837                 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
 838                                                 mtid, &bdata, errorp);
 839                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
 840                         hammer2_inode_data_t *wipdata;
 841
 842                         wipdata = &chain->data->ipdata;
 843                         KKASSERT(wipdata->meta.op_flags &
 844                                  HAMMER2_OPFLAG_DIRECTDATA);
 845                         bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
 846                         ++hammer2_iod_file_wembed;
 847                 } else if (bdata == NULL) {
 848                         /*
 849                          * Copy of data already present on-media.
 850                          */
 851                         chain->bref.methods =
 852                                 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
 853                                 HAMMER2_ENC_CHECK(ip->meta.check_algo);
 854                         hammer2_chain_setcheck(chain, data);
 855                 } else {
 856                         hammer2_write_bp(chain, data, ioflag, pblksize,
 857                                          mtid, errorp, ip->meta.check_algo);
 858                 }
 859                 if (chain) {
 860                         hammer2_chain_unlock(chain);
 861                         hammer2_chain_drop(chain);
 862                 }
 863                 break;
 864         case HAMMER2_COMP_AUTOZERO:
 865                 /*
 866                  * Check for zero-fill only
 867                  */
 868                 hammer2_zero_check_and_write(data, ip, parentp,
 869                                              lbase, ioflag, pblksize,
 870                                              mtid, errorp,
 871                                              ip->meta.check_algo);
 872                 break;
 873         case HAMMER2_COMP_LZ4:
 874         case HAMMER2_COMP_ZLIB:
 875         default:
 876                 /*
 877                  * Check for zero-fill and attempt compression.
 878                  */
 879                 hammer2_compress_and_write(data, ip, parentp,
 880                                            lbase, ioflag, pblksize,
 881                                            mtid, errorp,
 882                                            ip->meta.comp_algo,
 883                                            ip->meta.check_algo);
 884                 break;
 885         }
 886 }
 887
 888 /*
 889  * Helper
 890  *
 891  * Generic function that will perform the compression in compression
 892  * write path. The compression algorithm is determined by the settings
 893  * obtained from inode.
 894  */
 895 static
 896 void
 897 hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
 898         hammer2_chain_t **parentp,
 899         hammer2_key_t lbase, int ioflag, int pblksize,
 900         hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
 901 {
 902         hammer2_chain_t *chain;
 903         int comp_size;
 904         int comp_block_size;
 905         char *comp_buffer;
 906         char *bdata;
 907
 908         /*
 909          * An all-zeros write creates a hole unless the check code
 910          * is disabled.  When the check code is disabled all writes
 911          * are done in-place, including any all-zeros writes.
 912          *
 913          * NOTE: A snapshot will still force a copy-on-write
 914          *       (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
 915          */
 916         if (check_algo != HAMMER2_CHECK_NONE &&
 917             test_block_zeros(data, pblksize)) {
 918                 zero_write(data, ip, parentp, lbase, mtid, errorp);
 919                 return;
 920         }
 921
 922         /*
 923          * Compression requested.  Try to compress the block.  We store
 924          * the data normally if we cannot sufficiently compress it.
 925          *
 926          * We have a heuristic to detect files which are mostly
 927          * uncompressable and avoid the compression attempt in that
 928          * case.  If the compression heuristic is turned off, we always
 929          * try to compress.
 930          */
 931         comp_size = 0;
 932         comp_buffer = NULL;
 933
 934         KKASSERT(pblksize / 2 <= 32768);
 935
 936         if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 ||
 937             hammer2_always_compress) {
 938                 z_stream strm_compress;
 939                 int comp_level;
 940                 int ret;
 941
 942                 switch(HAMMER2_DEC_ALGO(comp_algo)) {
 943                 case HAMMER2_COMP_LZ4:
 944                         comp_buffer = objcache_get(cache_buffer_write,
 945                                                    M_INTWAIT);
 946                         comp_size = LZ4_compress_limitedOutput(
 947                                         data,
 948                                         &comp_buffer[sizeof(int)],
 949                                         pblksize,
 950                                         pblksize / 2 - sizeof(int));
 951                         /*
 952                          * We need to prefix with the size, LZ4
 953                          * doesn't do it for us.  Add the related
 954                          * overhead.
 955                          */
 956                         *(int *)comp_buffer = comp_size;
 957                         if (comp_size)
 958                                 comp_size += sizeof(int);
 959                         break;
 960                 case HAMMER2_COMP_ZLIB:
 961                         comp_level = HAMMER2_DEC_LEVEL(comp_algo);
 962                         if (comp_level == 0)
 963                                 comp_level = 6; /* default zlib compression */
 964                         else if (comp_level < 6)
 965                                 comp_level = 6;
 966                         else if (comp_level > 9)
 967                                 comp_level = 9;
 968                         ret = deflateInit(&strm_compress, comp_level);
 969                         if (ret != Z_OK) {
 970                                 kprintf("HAMMER2 ZLIB: fatal error "
 971                                         "on deflateInit.\n");
 972                         }
 973
 974                         comp_buffer = objcache_get(cache_buffer_write,
 975                                                    M_INTWAIT);
 976                         strm_compress.next_in = data;
 977                         strm_compress.avail_in = pblksize;
 978                         strm_compress.next_out = comp_buffer;
 979                         strm_compress.avail_out = pblksize / 2;
 980                         ret = deflate(&strm_compress, Z_FINISH);
 981                         if (ret == Z_STREAM_END) {
 982                                 comp_size = pblksize / 2 -
 983                                             strm_compress.avail_out;
 984                         } else {
 985                                 comp_size = 0;
 986                         }
 987                         ret = deflateEnd(&strm_compress);
 988                         break;
 989                 default:
 990                         kprintf("Error: Unknown compression method.\n");
 991                         kprintf("Comp_method = %d.\n", comp_algo);
 992                         break;
 993                 }
 994         }
 995
 996         if (comp_size == 0) {
 997                 /*
 998                  * compression failed or turned off
 999                  */
1000                 comp_block_size = pblksize;     /* safety */
1001                 if (++ip->comp_heuristic > 128)
1002                         ip->comp_heuristic = 8;
1003         } else {
1004                 /*
1005                  * compression succeeded
1006                  */
1007                 ip->comp_heuristic = 0;
1008                 if (comp_size <= 1024) {
1009                         comp_block_size = 1024;
1010                 } else if (comp_size <= 2048) {
1011                         comp_block_size = 2048;
1012                 } else if (comp_size <= 4096) {
1013                         comp_block_size = 4096;
1014                 } else if (comp_size <= 8192) {
1015                         comp_block_size = 8192;
1016                 } else if (comp_size <= 16384) {
1017                         comp_block_size = 16384;
1018                 } else if (comp_size <= 32768) {
1019                         comp_block_size = 32768;
1020                 } else {
1021                         panic("hammer2: WRITE PATH: "
1022                               "Weird comp_size value.");
1023                         /* NOT REACHED */
1024                         comp_block_size = pblksize;
1025                 }
1026
1027                 /*
1028                  * Must zero the remainder or dedup (which operates on a
1029                  * physical block basis) will not find matches.
1030                  */
1031                 if (comp_size < comp_block_size) {
1032                         bzero(comp_buffer + comp_size,
1033                               comp_block_size - comp_size);
1034                 }
1035         }
1036
1037         /*
1038          * Assign physical storage, data will be set to NULL if a live-dedup
1039          * was successful.
1040          */
1041         bdata = comp_size ? comp_buffer : data;
1042         chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
1043                                         mtid, &bdata, errorp);
1044
1045         if (*errorp) {
1046                 kprintf("WRITE PATH: An error occurred while "
1047                         "assigning physical space.\n");
1048                 KKASSERT(chain == NULL);
1049                 goto done;
1050         }
1051
1052         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1053                 hammer2_inode_data_t *wipdata;
1054
1055                 hammer2_chain_modify_ip(ip, chain, mtid, 0);
1056                 wipdata = &chain->data->ipdata;
1057                 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1058                 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1059                 ++hammer2_iod_file_wembed;
1060         } else if (bdata == NULL) {
1061                 /*
1062                  * Live deduplication, a copy of the data is already present
1063                  * on the media.
1064                  */
1065                 if (comp_size) {
1066                         chain->bref.methods =
1067                                 HAMMER2_ENC_COMP(comp_algo) +
1068                                 HAMMER2_ENC_CHECK(check_algo);
1069                 } else {
1070                         chain->bref.methods =
1071                                 HAMMER2_ENC_COMP(
1072                                         HAMMER2_COMP_NONE) +
1073                                 HAMMER2_ENC_CHECK(check_algo);
1074                 }
1075                 bdata = comp_size ? comp_buffer : data;
1076                 hammer2_chain_setcheck(chain, bdata);
1077                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1078         } else {
1079                 hammer2_io_t *dio;
1080
1081                 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1082
1083                 switch(chain->bref.type) {
1084                 case HAMMER2_BREF_TYPE_INODE:
1085                         panic("hammer2_write_bp: unexpected inode\n");
1086                         break;
1087                 case HAMMER2_BREF_TYPE_DATA:
1088                         /*
1089                          * Optimize out the read-before-write
1090                          * if possible.
1091                          */
1092                         *errorp = hammer2_io_newnz(chain->hmp,
1093                                                    chain->bref.type,
1094                                                    chain->bref.data_off,
1095                                                    chain->bytes,
1096                                                    &dio);
1097                         if (*errorp) {
1098                                 hammer2_io_brelse(&dio);
1099                                 kprintf("hammer2: WRITE PATH: "
1100                                         "dbp bread error\n");
1101                                 break;
1102                         }
1103                         bdata = hammer2_io_data(dio, chain->bref.data_off);
1104
1105                         /*
1106                          * When loading the block make sure we don't
1107                          * leave garbage after the compressed data.
1108                          */
1109                         if (comp_size) {
1110                                 chain->bref.methods =
1111                                         HAMMER2_ENC_COMP(comp_algo) +
1112                                         HAMMER2_ENC_CHECK(check_algo);
1113                                 bcopy(comp_buffer, bdata, comp_size);
1114                         } else {
1115                                 chain->bref.methods =
1116                                         HAMMER2_ENC_COMP(
1117                                                 HAMMER2_COMP_NONE) +
1118                                         HAMMER2_ENC_CHECK(check_algo);
1119                                 bcopy(data, bdata, pblksize);
1120                         }
1121
1122                         /*
1123                          * The flush code doesn't calculate check codes for
1124                          * file data (doing so can result in excessive I/O),
1125                          * so we do it here.
1126                          */
1127                         hammer2_chain_setcheck(chain, bdata);
1128
1129                         /*
1130                          * Device buffer is now valid, chain is no longer in
1131                          * the initial state.
1132                          *
1133                          * (No blockref table worries with file data)
1134                          */
1135                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1136                         hammer2_dedup_record(chain, dio, bdata);
1137
1138                         /* Now write the related bdp. */
1139                         if (ioflag & IO_SYNC) {
1140                                 /*
1141                                  * Synchronous I/O requested.
1142                                  */
1143                                 hammer2_io_bwrite(&dio);
1144                         /*
1145                         } else if ((ioflag & IO_DIRECT) &&
1146                                    loff + n == pblksize) {
1147                                 hammer2_io_bdwrite(&dio);
1148                         */
1149                         } else if (ioflag & IO_ASYNC) {
1150                                 hammer2_io_bawrite(&dio);
1151                         } else {
1152                                 hammer2_io_bdwrite(&dio);
1153                         }
1154                         break;
1155                 default:
1156                         panic("hammer2_write_bp: bad chain type %d\n",
1157                                 chain->bref.type);
1158                         /* NOT REACHED */
1159                         break;
1160                 }
1161         }
1162 done:
1163         if (chain) {
1164                 hammer2_chain_unlock(chain);
1165                 hammer2_chain_drop(chain);
1166         }
1167         if (comp_buffer)
1168                 objcache_put(cache_buffer_write, comp_buffer);
1169 }
1170
1171 /*
1172  * Helper
1173  *
1174  * Function that performs zero-checking and writing without compression,
1175  * it corresponds to default zero-checking path.
1176  */
1177 static
1178 void
1179 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
1180         hammer2_chain_t **parentp,
1181         hammer2_key_t lbase, int ioflag, int pblksize,
1182         hammer2_tid_t mtid, int *errorp,
1183         int check_algo)
1184 {
1185         hammer2_chain_t *chain;
1186         char *bdata;
1187
1188         if (check_algo != HAMMER2_CHECK_NONE &&
1189             test_block_zeros(data, pblksize)) {
1190                 /*
1191                  * An all-zeros write creates a hole unless the check code
1192                  * is disabled.  When the check code is disabled all writes
1193                  * are done in-place, including any all-zeros writes.
1194                  *
1195                  * NOTE: A snapshot will still force a copy-on-write
1196                  *       (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
1197                  */
1198                 zero_write(data, ip, parentp, lbase, mtid, errorp);
1199         } else {
1200                 /*
1201                  * Normal write
1202                  */
1203                 bdata = data;
1204                 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
1205                                                 mtid, &bdata, errorp);
1206                 if (bdata) {
1207                         hammer2_write_bp(chain, data, ioflag, pblksize,
1208                                          mtid, errorp, check_algo);
1209                 } else {
1210                         /* dedup occurred */
1211                         chain->bref.methods =
1212                                 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1213                                 HAMMER2_ENC_CHECK(check_algo);
1214                         hammer2_chain_setcheck(chain, data);
1215                 }
1216                 if (chain) {
1217                         hammer2_chain_unlock(chain);
1218                         hammer2_chain_drop(chain);
1219                 }
1220         }
1221 }
1222
1223 /*
1224  * Helper
1225  *
1226  * A function to test whether a block of data contains only zeros,
1227  * returns TRUE (non-zero) if the block is all zeros.
1228  */
1229 static
1230 int
1231 test_block_zeros(const char *buf, size_t bytes)
1232 {
1233         size_t i;
1234
1235         for (i = 0; i < bytes; i += sizeof(long)) {
1236                 if (*(const long *)(buf + i) != 0)
1237                         return (0);
1238         }
1239         return (1);
1240 }
1241
1242 /*
1243  * Helper
1244  *
1245  * Function to "write" a block that contains only zeros.
1246  */
1247 static
1248 void
1249 zero_write(char *data, hammer2_inode_t *ip,
1250            hammer2_chain_t **parentp,
1251            hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp)
1252 {
1253         hammer2_chain_t *chain;
1254         hammer2_key_t key_dummy;
1255         int cache_index = -1;
1256
1257         *errorp = 0;
1258         chain = hammer2_chain_lookup(parentp, &key_dummy,
1259                                      lbase, lbase,
1260                                      &cache_index,
1261                                      HAMMER2_LOOKUP_NODATA);
1262         if (chain) {
1263                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1264                         hammer2_inode_data_t *wipdata;
1265
1266                         hammer2_chain_modify_ip(ip, chain, mtid, 0);
1267                         wipdata = &chain->data->ipdata;
1268                         KKASSERT(wipdata->meta.op_flags &
1269                                  HAMMER2_OPFLAG_DIRECTDATA);
1270                         bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1271                         ++hammer2_iod_file_wembed;
1272                 } else {
1273                         hammer2_chain_delete(*parentp, chain,
1274                                              mtid, HAMMER2_DELETE_PERMANENT);
1275                         ++hammer2_iod_file_wzero;
1276                 }
1277                 hammer2_chain_unlock(chain);
1278                 hammer2_chain_drop(chain);
1279         } else {
1280                 ++hammer2_iod_file_wzero;
1281         }
1282 }
1283
1284 /*
1285  * Helper
1286  *
1287  * Function to write the data as it is, without performing any sort of
1288  * compression. This function is used in path without compression and
1289  * default zero-checking path.
1290  */
1291 static
1292 void
1293 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag,
1294                  int pblksize,
1295                  hammer2_tid_t mtid, int *errorp, int check_algo)
1296 {
1297         hammer2_inode_data_t *wipdata;
1298         hammer2_io_t *dio;
1299         char *bdata;
1300         int error;
1301
1302         error = 0;      /* XXX TODO below */
1303
1304         KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1305
1306         switch(chain->bref.type) {
1307         case HAMMER2_BREF_TYPE_INODE:
1308                 wipdata = &chain->data->ipdata;
1309                 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1310                 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1311                 error = 0;
1312                 ++hammer2_iod_file_wembed;
1313                 break;
1314         case HAMMER2_BREF_TYPE_DATA:
1315                 error = hammer2_io_newnz(chain->hmp,
1316                                          chain->bref.type,
1317                                          chain->bref.data_off,
1318                                          chain->bytes, &dio);
1319                 if (error) {
1320                         hammer2_io_bqrelse(&dio);
1321                         kprintf("hammer2: WRITE PATH: "
1322                                 "dbp bread error\n");
1323                         break;
1324                 }
1325                 bdata = hammer2_io_data(dio, chain->bref.data_off);
1326
1327                 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1328                                       HAMMER2_ENC_CHECK(check_algo);
1329                 bcopy(data, bdata, chain->bytes);
1330
1331                 /*
1332                  * The flush code doesn't calculate check codes for
1333                  * file data (doing so can result in excessive I/O),
1334                  * so we do it here.
1335                  */
1336                 hammer2_chain_setcheck(chain, bdata);
1337
1338                 /*
1339                  * Device buffer is now valid, chain is no longer in
1340                  * the initial state.
1341                  *
1342                  * (No blockref table worries with file data)
1343                  */
1344                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1345                 hammer2_dedup_record(chain, dio, bdata);
1346
1347                 if (ioflag & IO_SYNC) {
1348                         /*
1349                          * Synchronous I/O requested.
1350                          */
1351                         hammer2_io_bwrite(&dio);
1352                 /*
1353                 } else if ((ioflag & IO_DIRECT) &&
1354                            loff + n == pblksize) {
1355                         hammer2_io_bdwrite(&dio);
1356                 */
1357                 } else if (ioflag & IO_ASYNC) {
1358                         hammer2_io_bawrite(&dio);
1359                 } else {
1360                         hammer2_io_bdwrite(&dio);
1361                 }
1362                 break;
1363         default:
1364                 panic("hammer2_write_bp: bad chain type %d\n",
1365                       chain->bref.type);
1366                 /* NOT REACHED */
1367                 error = 0;
1368                 break;
1369         }
1370         KKASSERT(error == 0);   /* XXX TODO */
1371         *errorp = error;
1372 }
1373
1374 /*
1375  * LIVE DEDUP HEURISTICS
1376  *
1377  * Record media and crc information for possible dedup operation.  Note
1378  * that the dedup mask bits must also be set in the related DIO for a dedup
1379  * to be fully validated (which is handled in the freemap allocation code).
1380  *
1381  * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1382  *          All fields must be loaded into locals and validated.
1383  *
1384  * WARNING! Should only be used for file data and directory entries,
1385  *          hammer2_chain_modify() only checks for the dedup case on data
1386  *          chains.  Also, dedup data can only be recorded for committed
1387  *          chains (so NOT strategy writes which can undergo further
1388  *          modification after the fact!).
1389  */
1390 void
1391 hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio, char *data)
1392 {
1393         hammer2_dev_t *hmp;
1394         hammer2_dedup_t *dedup;
1395         uint64_t crc;
1396         uint64_t mask;
1397         int best = 0;
1398         int i;
1399         int dticks;
1400
1401         /*
1402          * We can only record a dedup if we have media data to test against.
1403          * If dedup is not enabled, return early, which allows a chain to
1404          * remain marked MODIFIED (which might have benefits in special
1405          * situations, though typically it does not).
1406          */
1407         if (hammer2_dedup_enable == 0)
1408                 return;
1409         if (dio == NULL) {
1410                 dio = chain->dio;
1411                 if (dio == NULL)
1412                         return;
1413         }
1414
1415 #if 0
1416         /*
1417          * Only committed data can be recorded for de-duplication, otherwise
1418          * the contents may change out from under us.  So, on read if the
1419          * chain is not modified, and on flush when the chain is committed.
1420          */
1421         if ((chain->flags &
1422             (HAMMER2_CHAIN_MODIFIED | HAMMER2_CHAIN_INITIAL)) == 0) {
1423                 return;
1424         }
1425 #endif
1426
1427         hmp = chain->hmp;
1428
1429         switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
1430         case HAMMER2_CHECK_ISCSI32:
1431                 /*
1432                  * XXX use the built-in crc (the dedup lookup sequencing
1433                  * needs to be fixed so the check code is already present
1434                  * when dedup_lookup is called)
1435                  */
1436 #if 0
1437                 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value;
1438 #endif
1439                 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1440                 break;
1441         case HAMMER2_CHECK_XXHASH64:
1442                 crc = chain->bref.check.xxhash64.value;
1443                 break;
1444         case HAMMER2_CHECK_SHA192:
1445                 /*
1446                  * XXX use the built-in crc (the dedup lookup sequencing
1447                  * needs to be fixed so the check code is already present
1448                  * when dedup_lookup is called)
1449                  */
1450 #if 0
1451                 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^
1452                       ((uint64_t *)chain->bref.check.sha192.data)[1] ^
1453                       ((uint64_t *)chain->bref.check.sha192.data)[2];
1454 #endif
1455                 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1456                 break;
1457         default:
1458                 /*
1459                  * Cannot dedup without a check code
1460                  *
1461                  * NOTE: In particular, CHECK_NONE allows a sector to be
1462                  *       overwritten without copy-on-write, recording
1463                  *       a dedup block for a CHECK_NONE object would be
1464                  *       a disaster!
1465                  */
1466                 return;
1467         }
1468         dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1469         for (i = 0; i < 4; ++i) {
1470                 if (dedup[i].data_crc == crc) {
1471                         best = i;
1472                         break;
1473                 }
1474                 dticks = (int)(dedup[i].ticks - dedup[best].ticks);
1475                 if (dticks < 0 || dticks > hz * 60 * 30)
1476                         best = i;
1477         }
1478         dedup += best;
1479         if (hammer2_debug & 0x40000) {
1480                 kprintf("REC %04x %016jx %016jx\n",
1481                         (int)(dedup - hmp->heur_dedup),
1482                         crc,
1483                         chain->bref.data_off);
1484         }
1485         dedup->ticks = ticks;
1486         dedup->data_off = chain->bref.data_off;
1487         dedup->data_crc = crc;
1488
1489         /*
1490          * Set the valid bits for the dedup only after we know the data
1491          * buffer has been updated.  The alloc bits were set (and the valid
1492          * bits cleared) when the media was allocated.
1493          *
1494          * This is done in two stages becuase the bulkfree code can race
1495          * the gap between allocation and data population.  Both masks must
1496          * be set before a bcmp/dedup operation is able to use the block.
1497          */
1498         mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes);
1499         atomic_set_64(&dio->dedup_valid, mask);
1500
1501         /*
1502          * Once we record the dedup the chain must be marked clean to
1503          * prevent reuse of the underlying block.   Remember that this
1504          * write occurs when the buffer cache is flushed (i.e. on sync(),
1505          * fsync(), filesystem periodic sync, or when the kernel needs to
1506          * flush a buffer), and not whenever the user write()s.
1507          */
1508         if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1509                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1510                 atomic_add_long(&hammer2_count_modified_chains, -1);
1511                 if (chain->pmp)
1512                         hammer2_pfs_memory_wakeup(chain->pmp);
1513         }
1514 }
1515
1516 static
1517 hammer2_off_t
1518 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize)
1519 {
1520         hammer2_dedup_t *dedup;
1521         hammer2_io_t *dio;
1522         hammer2_off_t off;
1523         uint64_t crc;
1524         uint64_t mask;
1525         char *data;
1526         char *dtmp;
1527         int i;
1528
1529         if (hammer2_dedup_enable == 0)
1530                 return 0;
1531         data = *datap;
1532         if (data == NULL)
1533                 return 0;
1534
1535         /*
1536          * XXX use the built-in crc (the dedup lookup sequencing
1537          * needs to be fixed so the check code is already present
1538          * when dedup_lookup is called)
1539          */
1540         crc = XXH64(data, pblksize, XXH_HAMMER2_SEED);
1541         dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1542
1543         if (hammer2_debug & 0x40000) {
1544                 kprintf("LOC %04x/4 %016jx\n",
1545                         (int)(dedup - hmp->heur_dedup),
1546                         crc);
1547         }
1548
1549         for (i = 0; i < 4; ++i) {
1550                 off = dedup[i].data_off;
1551                 cpu_ccfence();
1552                 if (dedup[i].data_crc != crc)
1553                         continue;
1554                 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize)
1555                         continue;
1556                 dio = hammer2_io_getquick(hmp, off, pblksize, 0);
1557                 if (dio) {
1558                         dtmp = hammer2_io_data(dio, off),
1559                         mask = hammer2_dedup_mask(dio, off, pblksize);
1560                         if ((dio->dedup_alloc & mask) == mask &&
1561                             (dio->dedup_valid & mask) == mask &&
1562                             bcmp(data, dtmp, pblksize) == 0) {
1563                                 if (hammer2_debug & 0x40000) {
1564                                         kprintf("DEDUP SUCCESS %016jx\n",
1565                                                 (intmax_t)off);
1566                                 }
1567                                 hammer2_io_putblk(&dio);
1568                                 *datap = NULL;
1569                                 dedup[i].ticks = ticks;   /* update use */
1570                                 atomic_add_long(&hammer2_iod_file_wdedup,
1571                                                 pblksize);
1572
1573                                 return off;             /* RETURN */
1574                         }
1575                         hammer2_io_putblk(&dio);
1576                 }
1577         }
1578         return 0;
1579 }
1580
1581 /*
1582  * Poof.  Races are ok, if someone gets in and reuses a dedup offset
1583  * before or while we are clearing it they will also recover the freemap
1584  * entry (set it to fully allocated), so a bulkfree race can only set it
1585  * to a possibly-free state.
1586  *
1587  * XXX ok, well, not really sure races are ok but going to run with it
1588  *     for the moment.
1589  */
1590 void
1591 hammer2_dedup_clear(hammer2_dev_t *hmp)
1592 {
1593         int i;
1594
1595         for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) {
1596                 hmp->heur_dedup[i].data_off = 0;
1597                 hmp->heur_dedup[i].ticks = ticks - 1;
1598         }
1599 }