sys/vfs/hammer2/hammer2_strategy.c

   1 /*
   2  * Copyright (c) 2011-2018 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@dragonflybsd.org>
   6  * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
   7  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  *
  13  * 1. Redistributions of source code must retain the above copyright
  14  *    notice, this list of conditions and the following disclaimer.
  15  * 2. Redistributions in binary form must reproduce the above copyright
  16  *    notice, this list of conditions and the following disclaimer in
  17  *    the documentation and/or other materials provided with the
  18  *    distribution.
  19  * 3. Neither the name of The DragonFly Project nor the names of its
  20  *    contributors may be used to endorse or promote products derived
  21  *    from this software without specific, prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  24  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  25  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  26  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  27  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  28  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  29  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  30  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  31  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  32  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  33  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  34  * SUCH DAMAGE.
  35  */
  36 /*
  37  * This module handles low level logical file I/O (strategy) which backs
  38  * the logical buffer cache.
  39  *
  40  * [De]compression, zero-block, check codes, and buffer cache operations
  41  * for file data is handled here.
  42  *
  43  * Live dedup makes its home here as well.
  44  */
  45
  46 #include <sys/param.h>
  47 #include <sys/systm.h>
  48 #include <sys/kernel.h>
  49 #include <sys/fcntl.h>
  50 #include <sys/buf.h>
  51 #include <sys/proc.h>
  52 #include <sys/mount.h>
  53 #include <sys/vnode.h>
  54 #include <sys/mountctl.h>
  55 #include <sys/dirent.h>
  56 #include <sys/uio.h>
  57 #include <sys/objcache.h>
  58 #include <sys/event.h>
  59 #include <sys/file.h>
  60 #include <vfs/fifofs/fifo.h>
  61
  62 #include "hammer2.h"
  63 #include "hammer2_lz4.h"
  64
  65 #include "zlib/hammer2_zlib.h"
  66
  67 struct objcache *cache_buffer_read;
  68 struct objcache *cache_buffer_write;
  69
  70 /*
  71  * Strategy code (async logical file buffer I/O from system)
  72  *
  73  * Except for the transaction init (which should normally not block),
  74  * we essentially run the strategy operation asynchronously via a XOP.
  75  *
  76  * WARNING! The XOP deals with buffer synchronization.  It is not synchronized
  77  *          to the current cpu.
  78  *
  79  * XXX This isn't supposed to be able to deadlock against vfs_sync vfsync()
  80  *     calls but it has in the past when multiple flushes are queued.
  81  *
  82  * XXX We currently terminate the transaction once we get a quorum, otherwise
  83  *     the frontend can stall, but this can leave the remaining nodes with
  84  *     a potential flush conflict.  We need to delay flushes on those nodes
  85  *     until running transactions complete separately from the normal
  86  *     transaction sequencing.  FIXME TODO.
  87  */
  88 static int hammer2_strategy_read(struct vop_strategy_args *ap);
  89 static int hammer2_strategy_write(struct vop_strategy_args *ap);
  90 static void hammer2_strategy_read_completion(hammer2_chain_t *focus,
  91                                 const char *data, struct bio *bio);
  92
  93 static hammer2_off_t hammer2_dedup_lookup(hammer2_dev_t *hmp,
  94                         char **datap, int pblksize);
  95
  96 int
  97 hammer2_vop_strategy(struct vop_strategy_args *ap)
  98 {
  99         struct bio *biop;
 100         struct buf *bp;
 101         int error;
 102
 103         biop = ap->a_bio;
 104         bp = biop->bio_buf;
 105
 106         switch(bp->b_cmd) {
 107         case BUF_CMD_READ:
 108                 error = hammer2_strategy_read(ap);
 109                 ++hammer2_iod_file_read;
 110                 break;
 111         case BUF_CMD_WRITE:
 112                 error = hammer2_strategy_write(ap);
 113                 ++hammer2_iod_file_write;
 114                 break;
 115         default:
 116                 bp->b_error = error = EINVAL;
 117                 bp->b_flags |= B_ERROR;
 118                 biodone(biop);
 119                 break;
 120         }
 121         return (error);
 122 }
 123
 124 /*
 125  * Return the largest contiguous physical disk range for the logical
 126  * request, in bytes.
 127  *
 128  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
 129  *
 130  * Basically disabled, the logical buffer write thread has to deal with
 131  * buffers one-at-a-time.  Note that this should not prevent cluster_read()
 132  * from reading-ahead, it simply prevents it from trying form a single
 133  * cluster buffer for the logical request.  H2 already uses 64KB buffers!
 134  */
 135 int
 136 hammer2_vop_bmap(struct vop_bmap_args *ap)
 137 {
 138         *ap->a_doffsetp = NOOFFSET;
 139         if (ap->a_runp)
 140                 *ap->a_runp = 0;
 141         if (ap->a_runb)
 142                 *ap->a_runb = 0;
 143         return (EOPNOTSUPP);
 144 }
 145
 146 /****************************************************************************
 147  *                              READ SUPPORT                                *
 148  ****************************************************************************/
 149 /*
 150  * Callback used in read path in case that a block is compressed with LZ4.
 151  */
 152 static
 153 void
 154 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio)
 155 {
 156         struct buf *bp;
 157         char *compressed_buffer;
 158         int compressed_size;
 159         int result;
 160
 161         bp = bio->bio_buf;
 162
 163 #if 0
 164         if bio->bio_caller_info2.index &&
 165               bio->bio_caller_info1.uvalue32 !=
 166               crc32(bp->b_data, bp->b_bufsize) --- return error
 167 #endif
 168
 169         KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
 170         compressed_size = *(const int *)data;
 171         KKASSERT((uint32_t)compressed_size <= bytes - sizeof(int));
 172
 173         compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
 174         result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]),
 175                                      compressed_buffer,
 176                                      compressed_size,
 177                                      bp->b_bufsize);
 178         if (result < 0) {
 179                 kprintf("READ PATH: Error during decompression."
 180                         "bio %016jx/%d\n",
 181                         (intmax_t)bio->bio_offset, bytes);
 182                 /* make sure it isn't random garbage */
 183                 bzero(compressed_buffer, bp->b_bufsize);
 184         }
 185         KKASSERT(result <= bp->b_bufsize);
 186         bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
 187         if (result < bp->b_bufsize)
 188                 bzero(bp->b_data + result, bp->b_bufsize - result);
 189         objcache_put(cache_buffer_read, compressed_buffer);
 190         bp->b_resid = 0;
 191         bp->b_flags |= B_AGE;
 192 }
 193
 194 /*
 195  * Callback used in read path in case that a block is compressed with ZLIB.
 196  * It is almost identical to LZ4 callback, so in theory they can be unified,
 197  * but we didn't want to make changes in bio structure for that.
 198  */
 199 static
 200 void
 201 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio)
 202 {
 203         struct buf *bp;
 204         char *compressed_buffer;
 205         z_stream strm_decompress;
 206         int result;
 207         int ret;
 208
 209         bp = bio->bio_buf;
 210
 211         KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
 212         strm_decompress.avail_in = 0;
 213         strm_decompress.next_in = Z_NULL;
 214
 215         ret = inflateInit(&strm_decompress);
 216
 217         if (ret != Z_OK)
 218                 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n");
 219
 220         compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
 221         strm_decompress.next_in = __DECONST(char *, data);
 222
 223         /* XXX supply proper size, subset of device bp */
 224         strm_decompress.avail_in = bytes;
 225         strm_decompress.next_out = compressed_buffer;
 226         strm_decompress.avail_out = bp->b_bufsize;
 227
 228         ret = inflate(&strm_decompress, Z_FINISH);
 229         if (ret != Z_STREAM_END) {
 230                 kprintf("HAMMER2 ZLIB: Fatar error during decompression.\n");
 231                 bzero(compressed_buffer, bp->b_bufsize);
 232         }
 233         bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
 234         result = bp->b_bufsize - strm_decompress.avail_out;
 235         if (result < bp->b_bufsize)
 236                 bzero(bp->b_data + result, strm_decompress.avail_out);
 237         objcache_put(cache_buffer_read, compressed_buffer);
 238         ret = inflateEnd(&strm_decompress);
 239
 240         bp->b_resid = 0;
 241         bp->b_flags |= B_AGE;
 242 }
 243
 244 /*
 245  * Logical buffer I/O, async read.
 246  */
 247 static
 248 int
 249 hammer2_strategy_read(struct vop_strategy_args *ap)
 250 {
 251         hammer2_xop_strategy_t *xop;
 252         struct bio *bio;
 253         hammer2_inode_t *ip;
 254         hammer2_key_t lbase;
 255
 256         bio = ap->a_bio;
 257         ip = VTOI(ap->a_vp);
 258
 259         lbase = bio->bio_offset;
 260         KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0);
 261
 262         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_STRATEGY);
 263         xop->finished = 0;
 264         xop->bio = bio;
 265         xop->lbase = lbase;
 266         hammer2_mtx_init(&xop->lock, "h2bior");
 267         hammer2_xop_start(&xop->head, &hammer2_strategy_read_desc);
 268         /* asynchronous completion */
 269
 270         return(0);
 271 }
 272
 273 /*
 274  * Per-node XOP (threaded), do a synchronous lookup of the chain and
 275  * its data.  The frontend is asynchronous, so we are also responsible
 276  * for racing to terminate the frontend.
 277  */
 278 void
 279 hammer2_xop_strategy_read(hammer2_xop_t *arg, void *scratch, int clindex)
 280 {
 281         hammer2_xop_strategy_t *xop = &arg->xop_strategy;
 282         hammer2_chain_t *parent;
 283         hammer2_chain_t *chain;
 284         hammer2_chain_t *focus;
 285         hammer2_key_t key_dummy;
 286         hammer2_key_t lbase;
 287         struct bio *bio;
 288         struct buf *bp;
 289         const char *data;
 290         int error;
 291
 292         /*
 293          * Note that we can race completion of the bio supplied by
 294          * the front-end so we cannot access it until we determine
 295          * that we are the ones finishing it up.
 296          */
 297         lbase = xop->lbase;
 298
 299         /*
 300          * This is difficult to optimize.  The logical buffer might be
 301          * partially dirty (contain dummy zero-fill pages), which would
 302          * mess up our crc calculation if we were to try a direct read.
 303          * So for now we always double-buffer through the underlying
 304          * storage.
 305          *
 306          * If not for the above problem we could conditionalize on
 307          * (1) 64KB buffer, (2) one chain (not multi-master) and
 308          * (3) !hammer2_double_buffer, and issue a direct read into the
 309          * logical buffer.
 310          */
 311         parent = hammer2_inode_chain(xop->head.ip1, clindex,
 312                                      HAMMER2_RESOLVE_ALWAYS |
 313                                      HAMMER2_RESOLVE_SHARED);
 314         if (parent) {
 315                 chain = hammer2_chain_lookup(&parent, &key_dummy,
 316                                              lbase, lbase,
 317                                              &error,
 318                                              HAMMER2_LOOKUP_ALWAYS |
 319                                              HAMMER2_LOOKUP_SHARED);
 320                 if (chain)
 321                         error = chain->error;
 322         } else {
 323                 error = HAMMER2_ERROR_EIO;
 324                 chain = NULL;
 325         }
 326         error = hammer2_xop_feed(&xop->head, chain, clindex, error);
 327         if (chain) {
 328                 hammer2_chain_unlock(chain);
 329                 hammer2_chain_drop(chain);
 330         }
 331         if (parent) {
 332                 hammer2_chain_unlock(parent);
 333                 hammer2_chain_drop(parent);
 334         }
 335         chain = NULL;   /* safety */
 336         parent = NULL;  /* safety */
 337
 338         /*
 339          * Race to finish the frontend.  First-to-complete.  bio is only
 340          * valid if we are determined to be the ones able to complete
 341          * the operation.
 342          */
 343         if (xop->finished)
 344                 return;
 345         hammer2_mtx_ex(&xop->lock);
 346         if (xop->finished) {
 347                 hammer2_mtx_unlock(&xop->lock);
 348                 return;
 349         }
 350         bio = xop->bio;
 351         bp = bio->bio_buf;
 352         bkvasync(bp);
 353
 354         /*
 355          * Async operation has not completed and we now own the lock.
 356          * Determine if we can complete the operation by issuing the
 357          * frontend collection non-blocking.
 358          *
 359          * H2 double-buffers the data, setting B_NOTMETA on the logical
 360          * buffer hints to the OS that the logical buffer should not be
 361          * swapcached (since the device buffer can be).
 362          *
 363          * Also note that even for compressed data we would rather the
 364          * kernel cache/swapcache device buffers more and (decompressed)
 365          * logical buffers less, since that will significantly improve
 366          * the amount of end-user data that can be cached.
 367          *
 368          * NOTE: The chain->data for xop->head.cluster.focus will be
 369          *       synchronized to the current cpu by xop_collect(),
 370          *       but other chains in the cluster might not be.
 371          */
 372         error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
 373
 374         switch(error) {
 375         case 0:
 376                 xop->finished = 1;
 377                 hammer2_mtx_unlock(&xop->lock);
 378                 bp->b_flags |= B_NOTMETA;
 379                 focus = xop->head.cluster.focus;
 380                 data = hammer2_xop_gdata(&xop->head)->buf;
 381                 hammer2_strategy_read_completion(focus, data, xop->bio);
 382                 hammer2_xop_pdata(&xop->head);
 383                 biodone(bio);
 384                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 385                 break;
 386         case HAMMER2_ERROR_ENOENT:
 387                 xop->finished = 1;
 388                 hammer2_mtx_unlock(&xop->lock);
 389                 bp->b_flags |= B_NOTMETA;
 390                 bp->b_resid = 0;
 391                 bp->b_error = 0;
 392                 bzero(bp->b_data, bp->b_bcount);
 393                 biodone(bio);
 394                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 395                 break;
 396         case HAMMER2_ERROR_EINPROGRESS:
 397                 hammer2_mtx_unlock(&xop->lock);
 398                 break;
 399         default:
 400                 kprintf("xop_strategy_read: error %08x loff=%016jx\n",
 401                         error, bp->b_loffset);
 402                 xop->finished = 1;
 403                 hammer2_mtx_unlock(&xop->lock);
 404                 bp->b_flags |= B_ERROR;
 405                 bp->b_error = EIO;
 406                 biodone(bio);
 407                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 408                 break;
 409         }
 410 }
 411
 412 static
 413 void
 414 hammer2_strategy_read_completion(hammer2_chain_t *focus, const char *data,
 415                                  struct bio *bio)
 416 {
 417         struct buf *bp = bio->bio_buf;
 418
 419         if (focus->bref.type == HAMMER2_BREF_TYPE_INODE) {
 420                 /*
 421                  * Copy from in-memory inode structure.
 422                  */
 423                 bcopy(((const hammer2_inode_data_t *)data)->u.data,
 424                       bp->b_data, HAMMER2_EMBEDDED_BYTES);
 425                 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
 426                       bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
 427                 bp->b_resid = 0;
 428                 bp->b_error = 0;
 429         } else if (focus->bref.type == HAMMER2_BREF_TYPE_DATA) {
 430                 /*
 431                  * Data is on-media, record for live dedup.  Release the
 432                  * chain (try to free it) when done.  The data is still
 433                  * cached by both the buffer cache in front and the
 434                  * block device behind us.  This leaves more room in the
 435                  * LRU chain cache for meta-data chains which we really
 436                  * want to retain.
 437                  *
 438                  * NOTE: Deduplication cannot be safely recorded for
 439                  *       records without a check code.
 440                  */
 441                 hammer2_dedup_record(focus, NULL, data);
 442                 atomic_set_int(&focus->flags, HAMMER2_CHAIN_RELEASE);
 443
 444                 /*
 445                  * Decompression and copy.
 446                  */
 447                 switch (HAMMER2_DEC_COMP(focus->bref.methods)) {
 448                 case HAMMER2_COMP_LZ4:
 449                         hammer2_decompress_LZ4_callback(data, focus->bytes,
 450                                                         bio);
 451                         /* b_resid set by call */
 452                         break;
 453                 case HAMMER2_COMP_ZLIB:
 454                         hammer2_decompress_ZLIB_callback(data, focus->bytes,
 455                                                          bio);
 456                         /* b_resid set by call */
 457                         break;
 458                 case HAMMER2_COMP_NONE:
 459                         KKASSERT(focus->bytes <= bp->b_bcount);
 460                         bcopy(data, bp->b_data, focus->bytes);
 461                         if (focus->bytes < bp->b_bcount) {
 462                                 bzero(bp->b_data + focus->bytes,
 463                                       bp->b_bcount - focus->bytes);
 464                         }
 465                         bp->b_resid = 0;
 466                         bp->b_error = 0;
 467                         break;
 468                 default:
 469                         panic("hammer2_strategy_read_completion: "
 470                               "unknown compression type");
 471                 }
 472         } else {
 473                 panic("hammer2_strategy_read_completion: unknown bref type");
 474         }
 475 }
 476
 477 /****************************************************************************
 478  *                              WRITE SUPPORT                               *
 479  ****************************************************************************/
 480
 481 /*
 482  * Functions for compression in threads,
 483  * from hammer2_vnops.c
 484  */
 485 static void hammer2_write_file_core(char *data, hammer2_inode_t *ip,
 486                                 hammer2_chain_t **parentp,
 487                                 hammer2_key_t lbase, int ioflag, int pblksize,
 488                                 hammer2_tid_t mtid, int *errorp);
 489 static void hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
 490                                 hammer2_chain_t **parentp,
 491                                 hammer2_key_t lbase, int ioflag, int pblksize,
 492                                 hammer2_tid_t mtid, int *errorp,
 493                                 int comp_algo, int check_algo);
 494 static void hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
 495                                 hammer2_chain_t **parentp,
 496                                 hammer2_key_t lbase, int ioflag, int pblksize,
 497                                 hammer2_tid_t mtid, int *errorp,
 498                                 int check_algo);
 499 static int test_block_zeros(const char *buf, size_t bytes);
 500 static void zero_write(char *data, hammer2_inode_t *ip,
 501                                 hammer2_chain_t **parentp,
 502                                 hammer2_key_t lbase,
 503                                 hammer2_tid_t mtid, int *errorp);
 504 static void hammer2_write_bp(hammer2_chain_t *chain, char *data,
 505                                 int ioflag, int pblksize,
 506                                 hammer2_tid_t mtid, int *errorp,
 507                                 int check_algo);
 508
 509 int
 510 hammer2_strategy_write(struct vop_strategy_args *ap)
 511 {
 512         hammer2_xop_strategy_t *xop;
 513         hammer2_pfs_t *pmp;
 514         struct bio *bio;
 515         hammer2_inode_t *ip;
 516
 517         bio = ap->a_bio;
 518         ip = VTOI(ap->a_vp);
 519         pmp = ip->pmp;
 520
 521         atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA);
 522         hammer2_lwinprog_ref(pmp);
 523         hammer2_trans_assert_strategy(pmp);
 524         hammer2_trans_init(pmp, HAMMER2_TRANS_BUFCACHE);
 525
 526         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING |
 527                                     HAMMER2_XOP_STRATEGY);
 528         xop->finished = 0;
 529         xop->bio = bio;
 530         xop->lbase = bio->bio_offset;
 531         hammer2_mtx_init(&xop->lock, "h2biow");
 532         hammer2_xop_start(&xop->head, &hammer2_strategy_write_desc);
 533         /* asynchronous completion */
 534
 535         hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
 536
 537         return(0);
 538 }
 539
 540 /*
 541  * Per-node XOP (threaded).  Write the logical buffer to the media.
 542  *
 543  * This is a bit problematic because there may be multiple target and
 544  * any of them may be able to release the bp.  In addition, if our
 545  * particulr target is offline we don't want to block the bp (and thus
 546  * the frontend).  To accomplish this we copy the data to the per-thr
 547  * scratch buffer.
 548  */
 549 void
 550 hammer2_xop_strategy_write(hammer2_xop_t *arg, void *scratch, int clindex)
 551 {
 552         hammer2_xop_strategy_t *xop = &arg->xop_strategy;
 553         hammer2_chain_t *parent;
 554         hammer2_key_t lbase;
 555         hammer2_inode_t *ip;
 556         struct bio *bio;
 557         struct buf *bp;
 558         int error;
 559         int lblksize;
 560         int pblksize;
 561         hammer2_off_t bio_offset;
 562         char *bio_data;
 563
 564         /*
 565          * We can only access the bp/bio if the frontend has not yet
 566          * completed.
 567          */
 568         if (xop->finished)
 569                 return;
 570         hammer2_mtx_sh(&xop->lock);
 571         if (xop->finished) {
 572                 hammer2_mtx_unlock(&xop->lock);
 573                 return;
 574         }
 575
 576         lbase = xop->lbase;
 577         bio = xop->bio;                 /* ephermal */
 578         bp = bio->bio_buf;              /* ephermal */
 579         ip = xop->head.ip1;             /* retained by ref */
 580         bio_offset = bio->bio_offset;
 581         bio_data = scratch;
 582
 583         /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
 584
 585         lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
 586         pblksize = hammer2_calc_physical(ip, lbase);
 587         bkvasync(bp);
 588         KKASSERT(lblksize <= MAXPHYS);
 589         bcopy(bp->b_data, bio_data, lblksize);
 590
 591         hammer2_mtx_unlock(&xop->lock);
 592         bp = NULL;      /* safety, illegal to access after unlock */
 593         bio = NULL;     /* safety, illegal to access after unlock */
 594
 595         /*
 596          * Actual operation
 597          */
 598         parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
 599         hammer2_write_file_core(bio_data, ip, &parent,
 600                                 lbase, IO_ASYNC, pblksize,
 601                                 xop->head.mtid, &error);
 602         if (parent) {
 603                 hammer2_chain_unlock(parent);
 604                 hammer2_chain_drop(parent);
 605                 parent = NULL;  /* safety */
 606         }
 607         hammer2_xop_feed(&xop->head, NULL, clindex, error);
 608
 609         /*
 610          * Try to complete the operation on behalf of the front-end.
 611          */
 612         if (xop->finished)
 613                 return;
 614         hammer2_mtx_ex(&xop->lock);
 615         if (xop->finished) {
 616                 hammer2_mtx_unlock(&xop->lock);
 617                 return;
 618         }
 619
 620         /*
 621          * Async operation has not completed and we now own the lock.
 622          * Determine if we can complete the operation by issuing the
 623          * frontend collection non-blocking.
 624          *
 625          * H2 double-buffers the data, setting B_NOTMETA on the logical
 626          * buffer hints to the OS that the logical buffer should not be
 627          * swapcached (since the device buffer can be).
 628          */
 629         error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
 630
 631         if (error == HAMMER2_ERROR_EINPROGRESS) {
 632                 hammer2_mtx_unlock(&xop->lock);
 633                 return;
 634         }
 635
 636         /*
 637          * Async operation has completed.
 638          */
 639         xop->finished = 1;
 640         hammer2_mtx_unlock(&xop->lock);
 641
 642         bio = xop->bio;         /* now owned by us */
 643         bp = bio->bio_buf;      /* now owned by us */
 644
 645         if (error == HAMMER2_ERROR_ENOENT || error == 0) {
 646                 bp->b_flags |= B_NOTMETA;
 647                 bp->b_resid = 0;
 648                 bp->b_error = 0;
 649                 biodone(bio);
 650         } else {
 651                 kprintf("xop_strategy_write: error %d loff=%016jx\n",
 652                         error, bp->b_loffset);
 653                 bp->b_flags |= B_ERROR;
 654                 bp->b_error = EIO;
 655                 biodone(bio);
 656         }
 657         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 658         hammer2_trans_assert_strategy(ip->pmp);
 659         hammer2_lwinprog_drop(ip->pmp);
 660         hammer2_trans_done(ip->pmp, HAMMER2_TRANS_BUFCACHE);
 661 }
 662
 663 /*
 664  * Wait for pending I/O to complete
 665  */
 666 void
 667 hammer2_bioq_sync(hammer2_pfs_t *pmp)
 668 {
 669         hammer2_lwinprog_wait(pmp, 0);
 670 }
 671
 672 /*
 673  * Assign physical storage at (cparent, lbase), returning a suitable chain
 674  * and setting *errorp appropriately.
 675  *
 676  * If no error occurs, the returned chain will be in a modified state.
 677  *
 678  * If an error occurs, the returned chain may or may not be NULL.  If
 679  * not-null any chain->error (if not 0) will also be rolled up into *errorp.
 680  * So the caller only needs to test *errorp.
 681  *
 682  * cparent can wind up being anything.
 683  *
 684  * If datap is not NULL, *datap points to the real data we intend to write.
 685  * If we can dedup the storage location we set *datap to NULL to indicate
 686  * to the caller that a dedup occurred.
 687  *
 688  * NOTE: Special case for data embedded in inode.
 689  */
 690 static
 691 hammer2_chain_t *
 692 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
 693                         hammer2_key_t lbase, int pblksize,
 694                         hammer2_tid_t mtid, char **datap, int *errorp)
 695 {
 696         hammer2_chain_t *chain;
 697         hammer2_key_t key_dummy;
 698         hammer2_off_t dedup_off;
 699         int pradix = hammer2_getradix(pblksize);
 700
 701         /*
 702          * Locate the chain associated with lbase, return a locked chain.
 703          * However, do not instantiate any data reference (which utilizes a
 704          * device buffer) because we will be using direct IO via the
 705          * logical buffer cache buffer.
 706          */
 707         KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
 708
 709         chain = hammer2_chain_lookup(parentp, &key_dummy,
 710                                      lbase, lbase,
 711                                      errorp,
 712                                      HAMMER2_LOOKUP_NODATA);
 713
 714         /*
 715          * The lookup code should not return a DELETED chain to us, unless
 716          * its a short-file embedded in the inode.  Then it is possible for
 717          * the lookup to return a deleted inode.
 718          */
 719         if (chain && (chain->flags & HAMMER2_CHAIN_DELETED) &&
 720             chain->bref.type != HAMMER2_BREF_TYPE_INODE) {
 721                 kprintf("assign physical deleted chain @ "
 722                         "%016jx (%016jx.%02x) ip %016jx\n",
 723                         lbase, chain->bref.data_off, chain->bref.type,
 724                         ip->meta.inum);
 725                 Debugger("bleh");
 726         }
 727
 728         if (chain == NULL) {
 729                 /*
 730                  * We found a hole, create a new chain entry.
 731                  *
 732                  * NOTE: DATA chains are created without device backing
 733                  *       store (nor do we want any).
 734                  */
 735                 dedup_off = hammer2_dedup_lookup((*parentp)->hmp, datap,
 736                                                  pblksize);
 737                 *errorp |= hammer2_chain_create(parentp, &chain, NULL, ip->pmp,
 738                                        HAMMER2_ENC_CHECK(ip->meta.check_algo) |
 739                                        HAMMER2_ENC_COMP(HAMMER2_COMP_NONE),
 740                                                 lbase, HAMMER2_PBUFRADIX,
 741                                                 HAMMER2_BREF_TYPE_DATA,
 742                                                 pblksize, mtid,
 743                                                 dedup_off, 0);
 744                 if (chain == NULL)
 745                         goto failed;
 746                 /*ip->delta_dcount += pblksize;*/
 747         } else if (chain->error == 0) {
 748                 switch (chain->bref.type) {
 749                 case HAMMER2_BREF_TYPE_INODE:
 750                         /*
 751                          * The data is embedded in the inode, which requires
 752                          * a bit more finess.
 753                          */
 754                         *errorp |= hammer2_chain_modify_ip(ip, chain, mtid, 0);
 755                         break;
 756                 case HAMMER2_BREF_TYPE_DATA:
 757                         dedup_off = hammer2_dedup_lookup(chain->hmp, datap,
 758                                                          pblksize);
 759                         if (chain->bytes != pblksize) {
 760                                 *errorp |= hammer2_chain_resize(chain,
 761                                                      mtid, dedup_off,
 762                                                      pradix,
 763                                                      HAMMER2_MODIFY_OPTDATA);
 764                                 if (*errorp)
 765                                         break;
 766                         }
 767
 768                         /*
 769                          * DATA buffers must be marked modified whether the
 770                          * data is in a logical buffer or not.  We also have
 771                          * to make this call to fixup the chain data pointers
 772                          * after resizing in case this is an encrypted or
 773                          * compressed buffer.
 774                          */
 775                         *errorp |= hammer2_chain_modify(chain, mtid, dedup_off,
 776                                                         HAMMER2_MODIFY_OPTDATA);
 777                         break;
 778                 default:
 779                         panic("hammer2_assign_physical: bad type");
 780                         /* NOT REACHED */
 781                         break;
 782                 }
 783         } else {
 784                 *errorp = chain->error;
 785         }
 786         atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA);
 787 failed:
 788         return (chain);
 789 }
 790
 791 /*
 792  * hammer2_write_file_core()
 793  *
 794  * The core write function which determines which path to take
 795  * depending on compression settings.  We also have to locate the
 796  * related chains so we can calculate and set the check data for
 797  * the blockref.
 798  */
 799 static
 800 void
 801 hammer2_write_file_core(char *data, hammer2_inode_t *ip,
 802                         hammer2_chain_t **parentp,
 803                         hammer2_key_t lbase, int ioflag, int pblksize,
 804                         hammer2_tid_t mtid, int *errorp)
 805 {
 806         hammer2_chain_t *chain;
 807         char *bdata;
 808
 809         *errorp = 0;
 810
 811         switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
 812         case HAMMER2_COMP_NONE:
 813                 /*
 814                  * We have to assign physical storage to the buffer
 815                  * we intend to dirty or write now to avoid deadlocks
 816                  * in the strategy code later.
 817                  *
 818                  * This can return NOOFFSET for inode-embedded data.
 819                  * The strategy code will take care of it in that case.
 820                  */
 821                 bdata = data;
 822                 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
 823                                                 mtid, &bdata, errorp);
 824                 if (*errorp) {
 825                         /* skip modifications */
 826                 } else if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
 827                         hammer2_inode_data_t *wipdata;
 828
 829                         wipdata = &chain->data->ipdata;
 830                         KKASSERT(wipdata->meta.op_flags &
 831                                  HAMMER2_OPFLAG_DIRECTDATA);
 832                         bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
 833                         ++hammer2_iod_file_wembed;
 834                 } else if (bdata == NULL) {
 835                         /*
 836                          * Copy of data already present on-media.
 837                          */
 838                         chain->bref.methods =
 839                                 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
 840                                 HAMMER2_ENC_CHECK(ip->meta.check_algo);
 841                         hammer2_chain_setcheck(chain, data);
 842                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
 843                 } else {
 844                         hammer2_write_bp(chain, data, ioflag, pblksize,
 845                                          mtid, errorp, ip->meta.check_algo);
 846                 }
 847                 if (chain) {
 848                         hammer2_chain_unlock(chain);
 849                         hammer2_chain_drop(chain);
 850                 }
 851                 break;
 852         case HAMMER2_COMP_AUTOZERO:
 853                 /*
 854                  * Check for zero-fill only
 855                  */
 856                 hammer2_zero_check_and_write(data, ip, parentp,
 857                                              lbase, ioflag, pblksize,
 858                                              mtid, errorp,
 859                                              ip->meta.check_algo);
 860                 break;
 861         case HAMMER2_COMP_LZ4:
 862         case HAMMER2_COMP_ZLIB:
 863         default:
 864                 /*
 865                  * Check for zero-fill and attempt compression.
 866                  */
 867                 hammer2_compress_and_write(data, ip, parentp,
 868                                            lbase, ioflag, pblksize,
 869                                            mtid, errorp,
 870                                            ip->meta.comp_algo,
 871                                            ip->meta.check_algo);
 872                 break;
 873         }
 874 }
 875
 876 /*
 877  * Helper
 878  *
 879  * Generic function that will perform the compression in compression
 880  * write path. The compression algorithm is determined by the settings
 881  * obtained from inode.
 882  */
 883 static
 884 void
 885 hammer2_compress_and_write(char *data, hammer2_inode_t *ip,
 886         hammer2_chain_t **parentp,
 887         hammer2_key_t lbase, int ioflag, int pblksize,
 888         hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
 889 {
 890         hammer2_chain_t *chain;
 891         int comp_size;
 892         int comp_block_size;
 893         char *comp_buffer;
 894         char *bdata;
 895
 896         /*
 897          * An all-zeros write creates a hole unless the check code
 898          * is disabled.  When the check code is disabled all writes
 899          * are done in-place, including any all-zeros writes.
 900          *
 901          * NOTE: A snapshot will still force a copy-on-write
 902          *       (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
 903          */
 904         if (check_algo != HAMMER2_CHECK_NONE &&
 905             test_block_zeros(data, pblksize)) {
 906                 zero_write(data, ip, parentp, lbase, mtid, errorp);
 907                 return;
 908         }
 909
 910         /*
 911          * Compression requested.  Try to compress the block.  We store
 912          * the data normally if we cannot sufficiently compress it.
 913          *
 914          * We have a heuristic to detect files which are mostly
 915          * uncompressable and avoid the compression attempt in that
 916          * case.  If the compression heuristic is turned off, we always
 917          * try to compress.
 918          */
 919         comp_size = 0;
 920         comp_buffer = NULL;
 921
 922         KKASSERT(pblksize / 2 <= 32768);
 923
 924         if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0 ||
 925             hammer2_always_compress) {
 926                 z_stream strm_compress;
 927                 int comp_level;
 928                 int ret;
 929
 930                 switch(HAMMER2_DEC_ALGO(comp_algo)) {
 931                 case HAMMER2_COMP_LZ4:
 932                         /*
 933                          * We need to prefix with the size, LZ4
 934                          * doesn't do it for us.  Add the related
 935                          * overhead.
 936                          *
 937                          * NOTE: The LZ4 code seems to assume at least an
 938                          *       8-byte buffer size granularity and may
 939                          *       overrun the buffer if given a 4-byte
 940                          *       granularity.
 941                          */
 942                         comp_buffer = objcache_get(cache_buffer_write,
 943                                                    M_INTWAIT);
 944                         comp_size = LZ4_compress_limitedOutput(
 945                                         data,
 946                                         &comp_buffer[sizeof(int)],
 947                                         pblksize,
 948                                         pblksize / 2 - sizeof(int64_t));
 949                         *(int *)comp_buffer = comp_size;
 950                         if (comp_size)
 951                                 comp_size += sizeof(int);
 952                         break;
 953                 case HAMMER2_COMP_ZLIB:
 954                         comp_level = HAMMER2_DEC_LEVEL(comp_algo);
 955                         if (comp_level == 0)
 956                                 comp_level = 6; /* default zlib compression */
 957                         else if (comp_level < 6)
 958                                 comp_level = 6;
 959                         else if (comp_level > 9)
 960                                 comp_level = 9;
 961                         ret = deflateInit(&strm_compress, comp_level);
 962                         if (ret != Z_OK) {
 963                                 kprintf("HAMMER2 ZLIB: fatal error "
 964                                         "on deflateInit.\n");
 965                         }
 966
 967                         comp_buffer = objcache_get(cache_buffer_write,
 968                                                    M_INTWAIT);
 969                         strm_compress.next_in = data;
 970                         strm_compress.avail_in = pblksize;
 971                         strm_compress.next_out = comp_buffer;
 972                         strm_compress.avail_out = pblksize / 2;
 973                         ret = deflate(&strm_compress, Z_FINISH);
 974                         if (ret == Z_STREAM_END) {
 975                                 comp_size = pblksize / 2 -
 976                                             strm_compress.avail_out;
 977                         } else {
 978                                 comp_size = 0;
 979                         }
 980                         ret = deflateEnd(&strm_compress);
 981                         break;
 982                 default:
 983                         kprintf("Error: Unknown compression method.\n");
 984                         kprintf("Comp_method = %d.\n", comp_algo);
 985                         break;
 986                 }
 987         }
 988
 989         if (comp_size == 0) {
 990                 /*
 991                  * compression failed or turned off
 992                  */
 993                 comp_block_size = pblksize;     /* safety */
 994                 if (++ip->comp_heuristic > 128)
 995                         ip->comp_heuristic = 8;
 996         } else {
 997                 /*
 998                  * compression succeeded
 999                  */
1000                 ip->comp_heuristic = 0;
1001                 if (comp_size <= 1024) {
1002                         comp_block_size = 1024;
1003                 } else if (comp_size <= 2048) {
1004                         comp_block_size = 2048;
1005                 } else if (comp_size <= 4096) {
1006                         comp_block_size = 4096;
1007                 } else if (comp_size <= 8192) {
1008                         comp_block_size = 8192;
1009                 } else if (comp_size <= 16384) {
1010                         comp_block_size = 16384;
1011                 } else if (comp_size <= 32768) {
1012                         comp_block_size = 32768;
1013                 } else {
1014                         panic("hammer2: WRITE PATH: "
1015                               "Weird comp_size value.");
1016                         /* NOT REACHED */
1017                         comp_block_size = pblksize;
1018                 }
1019
1020                 /*
1021                  * Must zero the remainder or dedup (which operates on a
1022                  * physical block basis) will not find matches.
1023                  */
1024                 if (comp_size < comp_block_size) {
1025                         bzero(comp_buffer + comp_size,
1026                               comp_block_size - comp_size);
1027                 }
1028         }
1029
1030         /*
1031          * Assign physical storage, bdata will be set to NULL if a live-dedup
1032          * was successful.
1033          */
1034         bdata = comp_size ? comp_buffer : data;
1035         chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
1036                                         mtid, &bdata, errorp);
1037
1038         if (*errorp) {
1039                 goto done;
1040         }
1041
1042         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1043                 hammer2_inode_data_t *wipdata;
1044
1045                 *errorp = hammer2_chain_modify_ip(ip, chain, mtid, 0);
1046                 if (*errorp == 0) {
1047                         wipdata = &chain->data->ipdata;
1048                         KKASSERT(wipdata->meta.op_flags &
1049                                  HAMMER2_OPFLAG_DIRECTDATA);
1050                         bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1051                         ++hammer2_iod_file_wembed;
1052                 }
1053         } else if (bdata == NULL) {
1054                 /*
1055                  * Live deduplication, a copy of the data is already present
1056                  * on the media.
1057                  */
1058                 if (comp_size) {
1059                         chain->bref.methods =
1060                                 HAMMER2_ENC_COMP(comp_algo) +
1061                                 HAMMER2_ENC_CHECK(check_algo);
1062                 } else {
1063                         chain->bref.methods =
1064                                 HAMMER2_ENC_COMP(
1065                                         HAMMER2_COMP_NONE) +
1066                                 HAMMER2_ENC_CHECK(check_algo);
1067                 }
1068                 bdata = comp_size ? comp_buffer : data;
1069                 hammer2_chain_setcheck(chain, bdata);
1070                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1071         } else {
1072                 hammer2_io_t *dio;
1073
1074                 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1075
1076                 switch(chain->bref.type) {
1077                 case HAMMER2_BREF_TYPE_INODE:
1078                         panic("hammer2_compress_and_write: unexpected inode\n");
1079                         break;
1080                 case HAMMER2_BREF_TYPE_DATA:
1081                         /*
1082                          * Optimize out the read-before-write
1083                          * if possible.
1084                          */
1085                         *errorp = hammer2_io_newnz(chain->hmp,
1086                                                    chain->bref.type,
1087                                                    chain->bref.data_off,
1088                                                    chain->bytes,
1089                                                    &dio);
1090                         if (*errorp) {
1091                                 hammer2_io_brelse(&dio);
1092                                 kprintf("hammer2: WRITE PATH: "
1093                                         "dbp bread error\n");
1094                                 break;
1095                         }
1096                         bdata = hammer2_io_data(dio, chain->bref.data_off);
1097
1098                         /*
1099                          * When loading the block make sure we don't
1100                          * leave garbage after the compressed data.
1101                          */
1102                         if (comp_size) {
1103                                 chain->bref.methods =
1104                                         HAMMER2_ENC_COMP(comp_algo) +
1105                                         HAMMER2_ENC_CHECK(check_algo);
1106                                 bcopy(comp_buffer, bdata, comp_block_size);
1107                         } else {
1108                                 chain->bref.methods =
1109                                         HAMMER2_ENC_COMP(
1110                                                 HAMMER2_COMP_NONE) +
1111                                         HAMMER2_ENC_CHECK(check_algo);
1112                                 bcopy(data, bdata, pblksize);
1113                         }
1114
1115                         /*
1116                          * The flush code doesn't calculate check codes for
1117                          * file data (doing so can result in excessive I/O),
1118                          * so we do it here.
1119                          */
1120                         hammer2_chain_setcheck(chain, bdata);
1121
1122                         /*
1123                          * Device buffer is now valid, chain is no longer in
1124                          * the initial state.
1125                          *
1126                          * (No blockref table worries with file data)
1127                          */
1128                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1129                         hammer2_dedup_record(chain, dio, bdata);
1130
1131                         /* Now write the related bdp. */
1132                         if (ioflag & IO_SYNC) {
1133                                 /*
1134                                  * Synchronous I/O requested.
1135                                  */
1136                                 hammer2_io_bwrite(&dio);
1137                         /*
1138                         } else if ((ioflag & IO_DIRECT) &&
1139                                    loff + n == pblksize) {
1140                                 hammer2_io_bdwrite(&dio);
1141                         */
1142                         } else if (ioflag & IO_ASYNC) {
1143                                 hammer2_io_bawrite(&dio);
1144                         } else {
1145                                 hammer2_io_bdwrite(&dio);
1146                         }
1147                         break;
1148                 default:
1149                         panic("hammer2_compress_and_write: bad chain type %d\n",
1150                                 chain->bref.type);
1151                         /* NOT REACHED */
1152                         break;
1153                 }
1154         }
1155 done:
1156         if (chain) {
1157                 hammer2_chain_unlock(chain);
1158                 hammer2_chain_drop(chain);
1159         }
1160         if (comp_buffer)
1161                 objcache_put(cache_buffer_write, comp_buffer);
1162 }
1163
1164 /*
1165  * Helper
1166  *
1167  * Function that performs zero-checking and writing without compression,
1168  * it corresponds to default zero-checking path.
1169  */
1170 static
1171 void
1172 hammer2_zero_check_and_write(char *data, hammer2_inode_t *ip,
1173         hammer2_chain_t **parentp,
1174         hammer2_key_t lbase, int ioflag, int pblksize,
1175         hammer2_tid_t mtid, int *errorp,
1176         int check_algo)
1177 {
1178         hammer2_chain_t *chain;
1179         char *bdata;
1180
1181         if (check_algo != HAMMER2_CHECK_NONE &&
1182             test_block_zeros(data, pblksize)) {
1183                 /*
1184                  * An all-zeros write creates a hole unless the check code
1185                  * is disabled.  When the check code is disabled all writes
1186                  * are done in-place, including any all-zeros writes.
1187                  *
1188                  * NOTE: A snapshot will still force a copy-on-write
1189                  *       (see the HAMMER2_CHECK_NONE in hammer2_chain.c).
1190                  */
1191                 zero_write(data, ip, parentp, lbase, mtid, errorp);
1192         } else {
1193                 /*
1194                  * Normal write (bdata set to NULL if de-duplicated)
1195                  */
1196                 bdata = data;
1197                 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
1198                                                 mtid, &bdata, errorp);
1199                 if (*errorp) {
1200                         /* do nothing */
1201                 } else if (bdata) {
1202                         hammer2_write_bp(chain, data, ioflag, pblksize,
1203                                          mtid, errorp, check_algo);
1204                 } else {
1205                         /* dedup occurred */
1206                         chain->bref.methods =
1207                                 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1208                                 HAMMER2_ENC_CHECK(check_algo);
1209                         hammer2_chain_setcheck(chain, data);
1210                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1211                 }
1212                 if (chain) {
1213                         hammer2_chain_unlock(chain);
1214                         hammer2_chain_drop(chain);
1215                 }
1216         }
1217 }
1218
1219 /*
1220  * Helper
1221  *
1222  * A function to test whether a block of data contains only zeros,
1223  * returns TRUE (non-zero) if the block is all zeros.
1224  */
1225 static
1226 int
1227 test_block_zeros(const char *buf, size_t bytes)
1228 {
1229         size_t i;
1230
1231         for (i = 0; i < bytes; i += sizeof(long)) {
1232                 if (*(const long *)(buf + i) != 0)
1233                         return (0);
1234         }
1235         return (1);
1236 }
1237
1238 /*
1239  * Helper
1240  *
1241  * Function to "write" a block that contains only zeros.
1242  */
1243 static
1244 void
1245 zero_write(char *data, hammer2_inode_t *ip,
1246            hammer2_chain_t **parentp,
1247            hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp)
1248 {
1249         hammer2_chain_t *chain;
1250         hammer2_key_t key_dummy;
1251
1252         chain = hammer2_chain_lookup(parentp, &key_dummy,
1253                                      lbase, lbase,
1254                                      errorp,
1255                                      HAMMER2_LOOKUP_NODATA);
1256         if (chain) {
1257                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1258                         hammer2_inode_data_t *wipdata;
1259
1260                         if (*errorp == 0) {
1261                                 *errorp = hammer2_chain_modify_ip(ip, chain,
1262                                                                   mtid, 0);
1263                         }
1264                         if (*errorp == 0) {
1265                                 wipdata = &chain->data->ipdata;
1266                                 KKASSERT(wipdata->meta.op_flags &
1267                                          HAMMER2_OPFLAG_DIRECTDATA);
1268                                 bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1269                                 ++hammer2_iod_file_wembed;
1270                         }
1271                 } else {
1272                         /* chain->error ok for deletion */
1273                         hammer2_chain_delete(*parentp, chain,
1274                                              mtid, HAMMER2_DELETE_PERMANENT);
1275                         ++hammer2_iod_file_wzero;
1276                 }
1277                 atomic_set_int(&ip->flags, HAMMER2_INODE_DIRTYDATA);
1278                 hammer2_chain_unlock(chain);
1279                 hammer2_chain_drop(chain);
1280         } else {
1281                 ++hammer2_iod_file_wzero;
1282         }
1283 }
1284
1285 /*
1286  * Helper
1287  *
1288  * Function to write the data as it is, without performing any sort of
1289  * compression. This function is used in path without compression and
1290  * default zero-checking path.
1291  */
1292 static
1293 void
1294 hammer2_write_bp(hammer2_chain_t *chain, char *data, int ioflag,
1295                  int pblksize,
1296                  hammer2_tid_t mtid, int *errorp, int check_algo)
1297 {
1298         hammer2_inode_data_t *wipdata;
1299         hammer2_io_t *dio;
1300         char *bdata;
1301         int error;
1302
1303         error = 0;      /* XXX TODO below */
1304
1305         KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1306
1307         switch(chain->bref.type) {
1308         case HAMMER2_BREF_TYPE_INODE:
1309                 wipdata = &chain->data->ipdata;
1310                 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1311                 bcopy(data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1312                 error = 0;
1313                 ++hammer2_iod_file_wembed;
1314                 break;
1315         case HAMMER2_BREF_TYPE_DATA:
1316                 error = hammer2_io_newnz(chain->hmp,
1317                                          chain->bref.type,
1318                                          chain->bref.data_off,
1319                                          chain->bytes, &dio);
1320                 if (error) {
1321                         hammer2_io_bqrelse(&dio);
1322                         kprintf("hammer2: WRITE PATH: "
1323                                 "dbp bread error\n");
1324                         break;
1325                 }
1326                 bdata = hammer2_io_data(dio, chain->bref.data_off);
1327
1328                 chain->bref.methods = HAMMER2_ENC_COMP(HAMMER2_COMP_NONE) +
1329                                       HAMMER2_ENC_CHECK(check_algo);
1330                 bcopy(data, bdata, chain->bytes);
1331
1332                 /*
1333                  * The flush code doesn't calculate check codes for
1334                  * file data (doing so can result in excessive I/O),
1335                  * so we do it here.
1336                  */
1337                 hammer2_chain_setcheck(chain, bdata);
1338
1339                 /*
1340                  * Device buffer is now valid, chain is no longer in
1341                  * the initial state.
1342                  *
1343                  * (No blockref table worries with file data)
1344                  */
1345                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1346                 hammer2_dedup_record(chain, dio, bdata);
1347
1348                 if (ioflag & IO_SYNC) {
1349                         /*
1350                          * Synchronous I/O requested.
1351                          */
1352                         hammer2_io_bwrite(&dio);
1353                 /*
1354                 } else if ((ioflag & IO_DIRECT) &&
1355                            loff + n == pblksize) {
1356                         hammer2_io_bdwrite(&dio);
1357                 */
1358                 } else if (ioflag & IO_ASYNC) {
1359                         hammer2_io_bawrite(&dio);
1360                 } else {
1361                         hammer2_io_bdwrite(&dio);
1362                 }
1363                 break;
1364         default:
1365                 panic("hammer2_write_bp: bad chain type %d\n",
1366                       chain->bref.type);
1367                 /* NOT REACHED */
1368                 error = 0;
1369                 break;
1370         }
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,
1392                      const char *data)
1393 {
1394         hammer2_dev_t *hmp;
1395         hammer2_dedup_t *dedup;
1396         uint64_t crc;
1397         uint64_t mask;
1398         int best = 0;
1399         int i;
1400         int dticks;
1401
1402         /*
1403          * We can only record a dedup if we have media data to test against.
1404          * If dedup is not enabled, return early, which allows a chain to
1405          * remain marked MODIFIED (which might have benefits in special
1406          * situations, though typically it does not).
1407          */
1408         if (hammer2_dedup_enable == 0)
1409                 return;
1410         if (dio == NULL) {
1411                 dio = chain->dio;
1412                 if (dio == NULL)
1413                         return;
1414         }
1415
1416         hmp = chain->hmp;
1417
1418         switch(HAMMER2_DEC_CHECK(chain->bref.methods)) {
1419         case HAMMER2_CHECK_ISCSI32:
1420                 /*
1421                  * XXX use the built-in crc (the dedup lookup sequencing
1422                  * needs to be fixed so the check code is already present
1423                  * when dedup_lookup is called)
1424                  */
1425 #if 0
1426                 crc = (uint64_t)(uint32_t)chain->bref.check.iscsi32.value;
1427 #endif
1428                 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1429                 break;
1430         case HAMMER2_CHECK_XXHASH64:
1431                 crc = chain->bref.check.xxhash64.value;
1432                 break;
1433         case HAMMER2_CHECK_SHA192:
1434                 /*
1435                  * XXX use the built-in crc (the dedup lookup sequencing
1436                  * needs to be fixed so the check code is already present
1437                  * when dedup_lookup is called)
1438                  */
1439 #if 0
1440                 crc = ((uint64_t *)chain->bref.check.sha192.data)[0] ^
1441                       ((uint64_t *)chain->bref.check.sha192.data)[1] ^
1442                       ((uint64_t *)chain->bref.check.sha192.data)[2];
1443 #endif
1444                 crc = XXH64(data, chain->bytes, XXH_HAMMER2_SEED);
1445                 break;
1446         default:
1447                 /*
1448                  * Cannot dedup without a check code
1449                  *
1450                  * NOTE: In particular, CHECK_NONE allows a sector to be
1451                  *       overwritten without copy-on-write, recording
1452                  *       a dedup block for a CHECK_NONE object would be
1453                  *       a disaster!
1454                  */
1455                 return;
1456         }
1457
1458         atomic_set_int(&chain->flags, HAMMER2_CHAIN_DEDUPABLE);
1459
1460         dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1461         for (i = 0; i < 4; ++i) {
1462                 if (dedup[i].data_crc == crc) {
1463                         best = i;
1464                         break;
1465                 }
1466                 dticks = (int)(dedup[i].ticks - dedup[best].ticks);
1467                 if (dticks < 0 || dticks > hz * 60 * 30)
1468                         best = i;
1469         }
1470         dedup += best;
1471         if (hammer2_debug & 0x40000) {
1472                 kprintf("REC %04x %016jx %016jx\n",
1473                         (int)(dedup - hmp->heur_dedup),
1474                         crc,
1475                         chain->bref.data_off);
1476         }
1477         dedup->ticks = ticks;
1478         dedup->data_off = chain->bref.data_off;
1479         dedup->data_crc = crc;
1480
1481         /*
1482          * Set the valid bits for the dedup only after we know the data
1483          * buffer has been updated.  The alloc bits were set (and the valid
1484          * bits cleared) when the media was allocated.
1485          *
1486          * This is done in two stages becuase the bulkfree code can race
1487          * the gap between allocation and data population.  Both masks must
1488          * be set before a bcmp/dedup operation is able to use the block.
1489          */
1490         mask = hammer2_dedup_mask(dio, chain->bref.data_off, chain->bytes);
1491         atomic_set_64(&dio->dedup_valid, mask);
1492
1493 #if 0
1494         /*
1495          * XXX removed. MODIFIED is an integral part of the flush code,
1496          * lets not just clear it
1497          */
1498         /*
1499          * Once we record the dedup the chain must be marked clean to
1500          * prevent reuse of the underlying block.   Remember that this
1501          * write occurs when the buffer cache is flushed (i.e. on sync(),
1502          * fsync(), filesystem periodic sync, or when the kernel needs to
1503          * flush a buffer), and not whenever the user write()s.
1504          */
1505         if (chain->flags & HAMMER2_CHAIN_MODIFIED) {
1506                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED);
1507                 atomic_add_long(&hammer2_count_modified_chains, -1);
1508                 if (chain->pmp)
1509                         hammer2_pfs_memory_wakeup(chain->pmp, -1);
1510         }
1511 #endif
1512 }
1513
1514 static
1515 hammer2_off_t
1516 hammer2_dedup_lookup(hammer2_dev_t *hmp, char **datap, int pblksize)
1517 {
1518         hammer2_dedup_t *dedup;
1519         hammer2_io_t *dio;
1520         hammer2_off_t off;
1521         uint64_t crc;
1522         uint64_t mask;
1523         char *data;
1524         char *dtmp;
1525         int i;
1526
1527         if (hammer2_dedup_enable == 0)
1528                 return 0;
1529         data = *datap;
1530         if (data == NULL)
1531                 return 0;
1532
1533         /*
1534          * XXX use the built-in crc (the dedup lookup sequencing
1535          * needs to be fixed so the check code is already present
1536          * when dedup_lookup is called)
1537          */
1538         crc = XXH64(data, pblksize, XXH_HAMMER2_SEED);
1539         dedup = &hmp->heur_dedup[crc & (HAMMER2_DEDUP_HEUR_MASK & ~3)];
1540
1541         if (hammer2_debug & 0x40000) {
1542                 kprintf("LOC %04x/4 %016jx\n",
1543                         (int)(dedup - hmp->heur_dedup),
1544                         crc);
1545         }
1546
1547         for (i = 0; i < 4; ++i) {
1548                 off = dedup[i].data_off;
1549                 cpu_ccfence();
1550                 if (dedup[i].data_crc != crc)
1551                         continue;
1552                 if ((1 << (int)(off & HAMMER2_OFF_MASK_RADIX)) != pblksize)
1553                         continue;
1554                 dio = hammer2_io_getquick(hmp, off, pblksize);
1555                 if (dio) {
1556                         dtmp = hammer2_io_data(dio, off),
1557                         mask = hammer2_dedup_mask(dio, off, pblksize);
1558                         if ((dio->dedup_alloc & mask) == mask &&
1559                             (dio->dedup_valid & mask) == mask &&
1560                             bcmp(data, dtmp, pblksize) == 0) {
1561                                 if (hammer2_debug & 0x40000) {
1562                                         kprintf("DEDUP SUCCESS %016jx\n",
1563                                                 (intmax_t)off);
1564                                 }
1565                                 hammer2_io_putblk(&dio);
1566                                 *datap = NULL;
1567                                 dedup[i].ticks = ticks;   /* update use */
1568                                 atomic_add_long(&hammer2_iod_file_wdedup,
1569                                                 pblksize);
1570
1571                                 return off;             /* RETURN */
1572                         }
1573                         hammer2_io_putblk(&dio);
1574                 }
1575         }
1576         return 0;
1577 }
1578
1579 /*
1580  * Poof.  Races are ok, if someone gets in and reuses a dedup offset
1581  * before or while we are clearing it they will also recover the freemap
1582  * entry (set it to fully allocated), so a bulkfree race can only set it
1583  * to a possibly-free state.
1584  *
1585  * XXX ok, well, not really sure races are ok but going to run with it
1586  *     for the moment.
1587  */
1588 void
1589 hammer2_dedup_clear(hammer2_dev_t *hmp)
1590 {
1591         int i;
1592
1593         for (i = 0; i < HAMMER2_DEDUP_HEUR_SIZE; ++i) {
1594                 hmp->heur_dedup[i].data_off = 0;
1595                 hmp->heur_dedup[i].ticks = ticks - 1;
1596         }
1597 }