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