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
  41 #include <sys/param.h>
  42 #include <sys/systm.h>
  43 #include <sys/kernel.h>
  44 #include <sys/fcntl.h>
  45 #include <sys/buf.h>
  46 #include <sys/proc.h>
  47 #include <sys/namei.h>
  48 #include <sys/mount.h>
  49 #include <sys/vnode.h>
  50 #include <sys/mountctl.h>
  51 #include <sys/dirent.h>
  52 #include <sys/uio.h>
  53 #include <sys/objcache.h>
  54 #include <sys/event.h>
  55 #include <sys/file.h>
  56 #include <vfs/fifofs/fifo.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_lz4.h"
  60
  61 #include "zlib/hammer2_zlib.h"
  62
  63 struct objcache *cache_buffer_read;
  64 struct objcache *cache_buffer_write;
  65
  66 /*
  67  * Strategy code (async logical file buffer I/O from system)
  68  *
  69  * WARNING: The strategy code cannot safely use hammer2 transactions
  70  *          as this can deadlock against vfs_sync's vfsync() call
  71  *          if multiple flushes are queued.  All H2 structures must
  72  *          already be present and ready for the DIO.
  73  *
  74  *          Reads can be initiated asynchronously, writes have to be
  75  *          spooled to a separate thread for action to avoid deadlocks.
  76  */
  77 static void hammer2_strategy_xop_read(hammer2_xop_t *arg, int clindex);
  78 static void hammer2_strategy_xop_write(hammer2_xop_t *arg, int clindex);
  79 static int hammer2_strategy_read(struct vop_strategy_args *ap);
  80 static int hammer2_strategy_write(struct vop_strategy_args *ap);
  81 static void hammer2_strategy_read_completion(hammer2_chain_t *chain,
  82                                 char *data, struct bio *bio);
  83
  84 int
  85 hammer2_vop_strategy(struct vop_strategy_args *ap)
  86 {
  87         struct bio *biop;
  88         struct buf *bp;
  89         int error;
  90
  91         biop = ap->a_bio;
  92         bp = biop->bio_buf;
  93
  94         switch(bp->b_cmd) {
  95         case BUF_CMD_READ:
  96                 error = hammer2_strategy_read(ap);
  97                 ++hammer2_iod_file_read;
  98                 break;
  99         case BUF_CMD_WRITE:
 100                 error = hammer2_strategy_write(ap);
 101                 ++hammer2_iod_file_write;
 102                 break;
 103         default:
 104                 bp->b_error = error = EINVAL;
 105                 bp->b_flags |= B_ERROR;
 106                 biodone(biop);
 107                 break;
 108         }
 109         return (error);
 110 }
 111
 112 /*
 113  * Return the largest contiguous physical disk range for the logical
 114  * request, in bytes.
 115  *
 116  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
 117  *
 118  * Basically disabled, the logical buffer write thread has to deal with
 119  * buffers one-at-a-time.
 120  */
 121 int
 122 hammer2_vop_bmap(struct vop_bmap_args *ap)
 123 {
 124         *ap->a_doffsetp = NOOFFSET;
 125         if (ap->a_runp)
 126                 *ap->a_runp = 0;
 127         if (ap->a_runb)
 128                 *ap->a_runb = 0;
 129         return (EOPNOTSUPP);
 130 }
 131
 132 /****************************************************************************
 133  *                              READ SUPPORT                                *
 134  ****************************************************************************/
 135 /*
 136  * Callback used in read path in case that a block is compressed with LZ4.
 137  */
 138 static
 139 void
 140 hammer2_decompress_LZ4_callback(const char *data, u_int bytes, struct bio *bio)
 141 {
 142         struct buf *bp;
 143         char *compressed_buffer;
 144         int compressed_size;
 145         int result;
 146
 147         bp = bio->bio_buf;
 148
 149 #if 0
 150         if bio->bio_caller_info2.index &&
 151               bio->bio_caller_info1.uvalue32 !=
 152               crc32(bp->b_data, bp->b_bufsize) --- return error
 153 #endif
 154
 155         KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
 156         compressed_size = *(const int *)data;
 157         KKASSERT(compressed_size <= bytes - sizeof(int));
 158
 159         compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
 160         result = LZ4_decompress_safe(__DECONST(char *, &data[sizeof(int)]),
 161                                      compressed_buffer,
 162                                      compressed_size,
 163                                      bp->b_bufsize);
 164         if (result < 0) {
 165                 kprintf("READ PATH: Error during decompression."
 166                         "bio %016jx/%d\n",
 167                         (intmax_t)bio->bio_offset, bytes);
 168                 /* make sure it isn't random garbage */
 169                 bzero(compressed_buffer, bp->b_bufsize);
 170         }
 171         KKASSERT(result <= bp->b_bufsize);
 172         bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
 173         if (result < bp->b_bufsize)
 174                 bzero(bp->b_data + result, bp->b_bufsize - result);
 175         objcache_put(cache_buffer_read, compressed_buffer);
 176         bp->b_resid = 0;
 177         bp->b_flags |= B_AGE;
 178 }
 179
 180 /*
 181  * Callback used in read path in case that a block is compressed with ZLIB.
 182  * It is almost identical to LZ4 callback, so in theory they can be unified,
 183  * but we didn't want to make changes in bio structure for that.
 184  */
 185 static
 186 void
 187 hammer2_decompress_ZLIB_callback(const char *data, u_int bytes, struct bio *bio)
 188 {
 189         struct buf *bp;
 190         char *compressed_buffer;
 191         z_stream strm_decompress;
 192         int result;
 193         int ret;
 194
 195         bp = bio->bio_buf;
 196
 197         KKASSERT(bp->b_bufsize <= HAMMER2_PBUFSIZE);
 198         strm_decompress.avail_in = 0;
 199         strm_decompress.next_in = Z_NULL;
 200
 201         ret = inflateInit(&strm_decompress);
 202
 203         if (ret != Z_OK)
 204                 kprintf("HAMMER2 ZLIB: Fatal error in inflateInit.\n");
 205
 206         compressed_buffer = objcache_get(cache_buffer_read, M_INTWAIT);
 207         strm_decompress.next_in = __DECONST(char *, data);
 208
 209         /* XXX supply proper size, subset of device bp */
 210         strm_decompress.avail_in = bytes;
 211         strm_decompress.next_out = compressed_buffer;
 212         strm_decompress.avail_out = bp->b_bufsize;
 213
 214         ret = inflate(&strm_decompress, Z_FINISH);
 215         if (ret != Z_STREAM_END) {
 216                 kprintf("HAMMER2 ZLIB: Fatar error during decompression.\n");
 217                 bzero(compressed_buffer, bp->b_bufsize);
 218         }
 219         bcopy(compressed_buffer, bp->b_data, bp->b_bufsize);
 220         result = bp->b_bufsize - strm_decompress.avail_out;
 221         if (result < bp->b_bufsize)
 222                 bzero(bp->b_data + result, strm_decompress.avail_out);
 223         objcache_put(cache_buffer_read, compressed_buffer);
 224         ret = inflateEnd(&strm_decompress);
 225
 226         bp->b_resid = 0;
 227         bp->b_flags |= B_AGE;
 228 }
 229
 230 /*
 231  * Logical buffer I/O, async read.
 232  */
 233 static
 234 int
 235 hammer2_strategy_read(struct vop_strategy_args *ap)
 236 {
 237         hammer2_xop_strategy_t *xop;
 238         struct buf *bp;
 239         struct bio *bio;
 240         struct bio *nbio;
 241         hammer2_inode_t *ip;
 242         hammer2_key_t lbase;
 243
 244         bio = ap->a_bio;
 245         bp = bio->bio_buf;
 246         ip = VTOI(ap->a_vp);
 247         nbio = push_bio(bio);
 248
 249         lbase = bio->bio_offset;
 250         KKASSERT(((int)lbase & HAMMER2_PBUFMASK) == 0);
 251
 252         xop = hammer2_xop_alloc(ip, 0);
 253         xop->finished = 0;
 254         xop->bio = bio;
 255         xop->lbase = lbase;
 256         hammer2_xop_start(&xop->head, hammer2_strategy_xop_read);
 257
 258         return(0);
 259 }
 260
 261 /*
 262  * Per-node XOP (threaded), do a synchronous lookup of the chain and
 263  * its data.  The frontend is asynchronous, so we are also responsible
 264  * for racing to terminate the frontend.
 265  */
 266 static
 267 void
 268 hammer2_strategy_xop_read(hammer2_xop_t *arg, int clindex)
 269 {
 270         hammer2_xop_strategy_t *xop = &arg->xop_strategy;
 271         hammer2_chain_t *parent;
 272         hammer2_chain_t *chain;
 273         hammer2_key_t key_dummy;
 274         hammer2_key_t lbase;
 275         struct bio *bio;
 276         struct buf *bp;
 277         int cache_index = -1;
 278         int error;
 279
 280         lbase = xop->lbase;
 281         bio = xop->bio;
 282         bp = bio->bio_buf;
 283
 284         parent = hammer2_inode_chain(xop->head.ip, clindex,
 285                                      HAMMER2_RESOLVE_ALWAYS |
 286                                      HAMMER2_RESOLVE_SHARED);
 287         if (parent) {
 288                 chain = hammer2_chain_lookup(&parent, &key_dummy,
 289                                              lbase, lbase,
 290                                              &cache_index,
 291                                              HAMMER2_LOOKUP_ALWAYS |
 292                                              HAMMER2_LOOKUP_SHARED);
 293                 error = chain ? chain->error : 0;
 294         } else {
 295                 error = EIO;
 296                 chain = NULL;
 297         }
 298         error = hammer2_xop_feed(&xop->head, chain, clindex, error);
 299         if (chain)
 300                 hammer2_chain_drop(chain);
 301         if (parent) {
 302                 hammer2_chain_unlock(parent);
 303                 hammer2_chain_drop(parent);
 304         }
 305         chain = NULL;   /* safety */
 306         parent = NULL;  /* safety */
 307
 308         /*
 309          * Race to finish the frontend
 310          */
 311         if (xop->finished)
 312                 return;
 313         hammer2_mtx_ex(&xop->head.xgrp->mtx2);
 314         if (xop->finished) {
 315                 hammer2_mtx_unlock(&xop->head.xgrp->mtx2);
 316                 return;
 317         }
 318
 319         /*
 320          * Async operation has not completed and we now own the lock.
 321          * Determine if we can complete the operation by issuing the
 322          * frontend collection non-blocking.
 323          */
 324         error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
 325
 326         switch(error) {
 327         case 0:
 328                 xop->finished = 1;
 329                 hammer2_mtx_unlock(&xop->head.xgrp->mtx2);
 330                 chain = xop->head.cluster.focus;
 331                 hammer2_strategy_read_completion(chain, (char *)chain->data,
 332                                                  xop->bio);
 333                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 334                 biodone(bio);
 335                 break;
 336         case ENOENT:
 337                 xop->finished = 1;
 338                 hammer2_mtx_unlock(&xop->head.xgrp->mtx2);
 339                 bp->b_resid = 0;
 340                 bp->b_error = 0;
 341                 bzero(bp->b_data, bp->b_bcount);
 342                 biodone(bio);
 343                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 344                 break;
 345         case EINPROGRESS:
 346                 hammer2_mtx_unlock(&xop->head.xgrp->mtx2);
 347                 break;
 348         default:
 349                 xop->finished = 1;
 350                 hammer2_mtx_unlock(&xop->head.xgrp->mtx2);
 351                 bp->b_flags |= B_ERROR;
 352                 bp->b_error = EIO;
 353                 biodone(bio);
 354                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 355                 break;
 356         }
 357 }
 358
 359 static
 360 void
 361 hammer2_strategy_read_completion(hammer2_chain_t *chain, char *data,
 362                                  struct bio *bio)
 363 {
 364         struct buf *bp = bio->bio_buf;
 365
 366         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
 367                 /*
 368                  * Data is embedded in the inode (copy from inode).
 369                  */
 370                 bcopy(((hammer2_inode_data_t *)data)->u.data,
 371                       bp->b_data, HAMMER2_EMBEDDED_BYTES);
 372                 bzero(bp->b_data + HAMMER2_EMBEDDED_BYTES,
 373                       bp->b_bcount - HAMMER2_EMBEDDED_BYTES);
 374                 bp->b_resid = 0;
 375                 bp->b_error = 0;
 376         } else if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
 377                 /*
 378                  * Data is on-media, issue device I/O and copy.
 379                  *
 380                  * XXX direct-IO shortcut could go here XXX.
 381                  */
 382                 switch (HAMMER2_DEC_COMP(chain->bref.methods)) {
 383                 case HAMMER2_COMP_LZ4:
 384                         hammer2_decompress_LZ4_callback(data, chain->bytes,
 385                                                         bio);
 386                         break;
 387                 case HAMMER2_COMP_ZLIB:
 388                         hammer2_decompress_ZLIB_callback(data, chain->bytes,
 389                                                          bio);
 390                         break;
 391                 case HAMMER2_COMP_NONE:
 392                         KKASSERT(chain->bytes <= bp->b_bcount);
 393                         bcopy(data, bp->b_data, chain->bytes);
 394                         if (chain->bytes < bp->b_bcount) {
 395                                 bzero(bp->b_data + chain->bytes,
 396                                       bp->b_bcount - chain->bytes);
 397                         }
 398                         bp->b_flags |= B_NOTMETA;
 399                         bp->b_resid = 0;
 400                         bp->b_error = 0;
 401                         break;
 402                 default:
 403                         panic("hammer2_strategy_read: "
 404                               "unknown compression type");
 405                 }
 406         } else {
 407                 panic("hammer2_strategy_read: unknown bref type");
 408         }
 409 }
 410
 411 /****************************************************************************
 412  *                              WRITE SUPPORT                               *
 413  ****************************************************************************/
 414
 415 /*
 416  * Functions for compression in threads,
 417  * from hammer2_vnops.c
 418  */
 419 static void hammer2_write_file_core(struct buf *bp, hammer2_inode_t *ip,
 420                                 hammer2_chain_t **parentp,
 421                                 hammer2_key_t lbase, int ioflag, int pblksize,
 422                                 hammer2_tid_t mtid, int *errorp);
 423 static void hammer2_compress_and_write(struct buf *bp, hammer2_inode_t *ip,
 424                                 hammer2_chain_t **parentp,
 425                                 hammer2_key_t lbase, int ioflag, int pblksize,
 426                                 hammer2_tid_t mtid, int *errorp,
 427                                 int comp_algo, int check_algo);
 428 static void hammer2_zero_check_and_write(struct buf *bp, hammer2_inode_t *ip,
 429                                 hammer2_chain_t **parentp,
 430                                 hammer2_key_t lbase, int ioflag, int pblksize,
 431                                 hammer2_tid_t mtid, int *errorp,
 432                                 int check_algo);
 433 static int test_block_zeros(const char *buf, size_t bytes);
 434 static void zero_write(struct buf *bp, hammer2_inode_t *ip,
 435                                 hammer2_chain_t **parentp,
 436                                 hammer2_key_t lbase,
 437                                 hammer2_tid_t mtid, int *errorp);
 438 static void hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp,
 439                                 int ioflag, int pblksize,
 440                                 hammer2_tid_t mtid, int *errorp,
 441                                 int check_algo);
 442
 443 static
 444 int
 445 hammer2_strategy_write(struct vop_strategy_args *ap)
 446 {
 447         hammer2_xop_strategy_t *xop;
 448         hammer2_pfs_t *pmp;
 449         struct bio *bio;
 450         struct buf *bp;
 451         hammer2_inode_t *ip;
 452
 453         bio = ap->a_bio;
 454         bp = bio->bio_buf;
 455         ip = VTOI(ap->a_vp);
 456         pmp = ip->pmp;
 457
 458         hammer2_lwinprog_ref(pmp);
 459         hammer2_trans_assert_strategy(pmp);
 460
 461         xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
 462         xop->finished = 0;
 463         xop->bio = bio;
 464         xop->lbase = bio->bio_offset;
 465         hammer2_xop_start(&xop->head, hammer2_strategy_xop_write);
 466         /* asynchronous completion */
 467
 468         hammer2_lwinprog_wait(pmp, hammer2_flush_pipe);
 469
 470         return(0);
 471 }
 472
 473 /*
 474  * Per-node XOP (threaded).  Write the logical buffer to the media.
 475  */
 476 static
 477 void
 478 hammer2_strategy_xop_write(hammer2_xop_t *arg, int clindex)
 479 {
 480         hammer2_xop_strategy_t *xop = &arg->xop_strategy;
 481         hammer2_chain_t *parent;
 482         hammer2_key_t lbase;
 483         hammer2_inode_t *ip;
 484         struct bio *bio;
 485         struct buf *bp;
 486         int error;
 487         int lblksize;
 488         int pblksize;
 489
 490         lbase = xop->lbase;
 491         bio = xop->bio;
 492         bp = bio->bio_buf;
 493         ip = xop->head.ip;
 494
 495         /* hammer2_trans_init(parent->hmp->spmp, HAMMER2_TRANS_BUFCACHE); */
 496
 497         lblksize = hammer2_calc_logical(ip, bio->bio_offset, &lbase, NULL);
 498         pblksize = hammer2_calc_physical(ip, lbase);
 499         parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
 500         hammer2_write_file_core(bp, ip, &parent,
 501                                 lbase, IO_ASYNC, pblksize,
 502                                 xop->head.mtid, &error);
 503         if (parent) {
 504                 hammer2_chain_unlock(parent);
 505                 hammer2_chain_drop(parent);
 506                 parent = NULL;  /* safety */
 507         }
 508         error = hammer2_xop_feed(&xop->head, NULL, clindex, error);
 509
 510         /*
 511          * Race to finish the frontend
 512          */
 513         if (xop->finished)
 514                 return;
 515         hammer2_mtx_ex(&xop->head.xgrp->mtx2);
 516         if (xop->finished) {
 517                 hammer2_mtx_unlock(&xop->head.xgrp->mtx2);
 518                 return;
 519         }
 520
 521         /*
 522          * Async operation has not completed and we now own the lock.
 523          * Determine if we can complete the operation by issuing the
 524          * frontend collection non-blocking.
 525          */
 526         error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_NOWAIT);
 527
 528         switch(error) {
 529         case ENOENT:
 530         case 0:
 531                 xop->finished = 1;
 532                 hammer2_mtx_unlock(&xop->head.xgrp->mtx2);
 533                 bp->b_resid = 0;
 534                 bp->b_error = 0;
 535                 biodone(bio);
 536                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 537                 hammer2_lwinprog_drop(ip->pmp);
 538                 break;
 539         case EINPROGRESS:
 540                 hammer2_mtx_unlock(&xop->head.xgrp->mtx2);
 541                 break;
 542         default:
 543                 xop->finished = 1;
 544                 hammer2_mtx_unlock(&xop->head.xgrp->mtx2);
 545                 bp->b_flags |= B_ERROR;
 546                 bp->b_error = EIO;
 547                 biodone(bio);
 548                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 549                 hammer2_lwinprog_drop(ip->pmp);
 550                 break;
 551         }
 552 }
 553
 554 /*
 555  * Wait for pending I/O to complete
 556  */
 557 void
 558 hammer2_bioq_sync(hammer2_pfs_t *pmp)
 559 {
 560         hammer2_lwinprog_wait(pmp, 0);
 561 }
 562
 563 /*
 564  * Create a new cluster at (cparent, lbase) and assign physical storage,
 565  * returning a cluster suitable for I/O.  The cluster will be in a modified
 566  * state.
 567  *
 568  * cparent can wind up being anything.
 569  *
 570  * NOTE: Special case for data embedded in inode.
 571  */
 572 static
 573 hammer2_chain_t *
 574 hammer2_assign_physical(hammer2_inode_t *ip, hammer2_chain_t **parentp,
 575                         hammer2_key_t lbase, int pblksize,
 576                         hammer2_tid_t mtid, int *errorp)
 577 {
 578         hammer2_chain_t *chain;
 579         hammer2_key_t key_dummy;
 580         int pradix = hammer2_getradix(pblksize);
 581         int cache_index = -1;
 582
 583         /*
 584          * Locate the chain associated with lbase, return a locked chain.
 585          * However, do not instantiate any data reference (which utilizes a
 586          * device buffer) because we will be using direct IO via the
 587          * logical buffer cache buffer.
 588          */
 589         *errorp = 0;
 590         KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
 591 retry:
 592         chain = hammer2_chain_lookup(parentp, &key_dummy,
 593                                      lbase, lbase,
 594                                      &cache_index,
 595                                      HAMMER2_LOOKUP_NODATA);
 596         if (chain == NULL) {
 597                 /*
 598                  * We found a hole, create a new chain entry.
 599                  *
 600                  * NOTE: DATA chains are created without device backing
 601                  *       store (nor do we want any).
 602                  */
 603                 *errorp = hammer2_chain_create(parentp, &chain, ip->pmp,
 604                                                lbase, HAMMER2_PBUFRADIX,
 605                                                HAMMER2_BREF_TYPE_DATA,
 606                                                pblksize,
 607                                                mtid, 0);
 608                 if (chain == NULL) {
 609                         panic("hammer2_chain_create: par=%p error=%d\n",
 610                               *parentp, *errorp);
 611                         goto retry;
 612                 }
 613                 /*ip->delta_dcount += pblksize;*/
 614         } else {
 615                 switch (chain->bref.type) {
 616                 case HAMMER2_BREF_TYPE_INODE:
 617                         /*
 618                          * The data is embedded in the inode, which requires
 619                          * a bit more finess.
 620                          */
 621                         hammer2_chain_modify_ip(ip, chain, mtid, 0);
 622                         break;
 623                 case HAMMER2_BREF_TYPE_DATA:
 624                         if (chain->bytes != pblksize) {
 625                                 hammer2_chain_resize(ip, *parentp, chain,
 626                                                      mtid, pradix,
 627                                                      HAMMER2_MODIFY_OPTDATA);
 628                         }
 629
 630                         /*
 631                          * DATA buffers must be marked modified whether the
 632                          * data is in a logical buffer or not.  We also have
 633                          * to make this call to fixup the chain data pointers
 634                          * after resizing in case this is an encrypted or
 635                          * compressed buffer.
 636                          */
 637                         hammer2_chain_modify(chain, mtid,
 638                                              HAMMER2_MODIFY_OPTDATA);
 639                         break;
 640                 default:
 641                         panic("hammer2_assign_physical: bad type");
 642                         /* NOT REACHED */
 643                         break;
 644                 }
 645         }
 646         return (chain);
 647 }
 648
 649 /*
 650  * hammer2_write_file_core() - hammer2_write_thread() helper
 651  *
 652  * The core write function which determines which path to take
 653  * depending on compression settings.  We also have to locate the
 654  * related chains so we can calculate and set the check data for
 655  * the blockref.
 656  */
 657 static
 658 void
 659 hammer2_write_file_core(struct buf *bp, hammer2_inode_t *ip,
 660                         hammer2_chain_t **parentp,
 661                         hammer2_key_t lbase, int ioflag, int pblksize,
 662                         hammer2_tid_t mtid, int *errorp)
 663 {
 664         hammer2_chain_t *chain;
 665
 666         switch(HAMMER2_DEC_ALGO(ip->meta.comp_algo)) {
 667         case HAMMER2_COMP_NONE:
 668                 /*
 669                  * We have to assign physical storage to the buffer
 670                  * we intend to dirty or write now to avoid deadlocks
 671                  * in the strategy code later.
 672                  *
 673                  * This can return NOOFFSET for inode-embedded data.
 674                  * The strategy code will take care of it in that case.
 675                  */
 676                 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
 677                                                 mtid, errorp);
 678                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
 679                         hammer2_inode_data_t *wipdata;
 680
 681                         wipdata = &chain->data->ipdata;
 682                         KKASSERT(wipdata->meta.op_flags &
 683                                  HAMMER2_OPFLAG_DIRECTDATA);
 684                         KKASSERT(bp->b_loffset == 0);
 685                         bcopy(bp->b_data, wipdata->u.data,
 686                               HAMMER2_EMBEDDED_BYTES);
 687                 } else {
 688                         hammer2_write_bp(chain, bp, ioflag, pblksize,
 689                                          mtid, errorp, ip->meta.check_algo);
 690                 }
 691                 if (chain) {
 692                         hammer2_chain_unlock(chain);
 693                         hammer2_chain_drop(chain);
 694                 }
 695                 break;
 696         case HAMMER2_COMP_AUTOZERO:
 697                 /*
 698                  * Check for zero-fill only
 699                  */
 700                 hammer2_zero_check_and_write(bp, ip, parentp,
 701                                              lbase, ioflag, pblksize,
 702                                              mtid, errorp,
 703                                              ip->meta.check_algo);
 704                 break;
 705         case HAMMER2_COMP_LZ4:
 706         case HAMMER2_COMP_ZLIB:
 707         default:
 708                 /*
 709                  * Check for zero-fill and attempt compression.
 710                  */
 711                 hammer2_compress_and_write(bp, ip, parentp,
 712                                            lbase, ioflag, pblksize,
 713                                            mtid, errorp,
 714                                            ip->meta.comp_algo,
 715                                            ip->meta.check_algo);
 716                 break;
 717         }
 718 }
 719
 720 /*
 721  * Helper
 722  *
 723  * Generic function that will perform the compression in compression
 724  * write path. The compression algorithm is determined by the settings
 725  * obtained from inode.
 726  */
 727 static
 728 void
 729 hammer2_compress_and_write(struct buf *bp, hammer2_inode_t *ip,
 730         hammer2_chain_t **parentp,
 731         hammer2_key_t lbase, int ioflag, int pblksize,
 732         hammer2_tid_t mtid, int *errorp, int comp_algo, int check_algo)
 733 {
 734         hammer2_chain_t *chain;
 735         int comp_size;
 736         int comp_block_size;
 737         char *comp_buffer;
 738
 739         if (test_block_zeros(bp->b_data, pblksize)) {
 740                 zero_write(bp, ip, parentp, lbase, mtid, errorp);
 741                 return;
 742         }
 743
 744         comp_size = 0;
 745         comp_buffer = NULL;
 746
 747         KKASSERT(pblksize / 2 <= 32768);
 748
 749         if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
 750                 z_stream strm_compress;
 751                 int comp_level;
 752                 int ret;
 753
 754                 switch(HAMMER2_DEC_ALGO(comp_algo)) {
 755                 case HAMMER2_COMP_LZ4:
 756                         comp_buffer = objcache_get(cache_buffer_write,
 757                                                    M_INTWAIT);
 758                         comp_size = LZ4_compress_limitedOutput(
 759                                         bp->b_data,
 760                                         &comp_buffer[sizeof(int)],
 761                                         pblksize,
 762                                         pblksize / 2 - sizeof(int));
 763                         /*
 764                          * We need to prefix with the size, LZ4
 765                          * doesn't do it for us.  Add the related
 766                          * overhead.
 767                          */
 768                         *(int *)comp_buffer = comp_size;
 769                         if (comp_size)
 770                                 comp_size += sizeof(int);
 771                         break;
 772                 case HAMMER2_COMP_ZLIB:
 773                         comp_level = HAMMER2_DEC_LEVEL(comp_algo);
 774                         if (comp_level == 0)
 775                                 comp_level = 6; /* default zlib compression */
 776                         else if (comp_level < 6)
 777                                 comp_level = 6;
 778                         else if (comp_level > 9)
 779                                 comp_level = 9;
 780                         ret = deflateInit(&strm_compress, comp_level);
 781                         if (ret != Z_OK) {
 782                                 kprintf("HAMMER2 ZLIB: fatal error "
 783                                         "on deflateInit.\n");
 784                         }
 785
 786                         comp_buffer = objcache_get(cache_buffer_write,
 787                                                    M_INTWAIT);
 788                         strm_compress.next_in = bp->b_data;
 789                         strm_compress.avail_in = pblksize;
 790                         strm_compress.next_out = comp_buffer;
 791                         strm_compress.avail_out = pblksize / 2;
 792                         ret = deflate(&strm_compress, Z_FINISH);
 793                         if (ret == Z_STREAM_END) {
 794                                 comp_size = pblksize / 2 -
 795                                             strm_compress.avail_out;
 796                         } else {
 797                                 comp_size = 0;
 798                         }
 799                         ret = deflateEnd(&strm_compress);
 800                         break;
 801                 default:
 802                         kprintf("Error: Unknown compression method.\n");
 803                         kprintf("Comp_method = %d.\n", comp_algo);
 804                         break;
 805                 }
 806         }
 807
 808         if (comp_size == 0) {
 809                 /*
 810                  * compression failed or turned off
 811                  */
 812                 comp_block_size = pblksize;     /* safety */
 813                 if (++ip->comp_heuristic > 128)
 814                         ip->comp_heuristic = 8;
 815         } else {
 816                 /*
 817                  * compression succeeded
 818                  */
 819                 ip->comp_heuristic = 0;
 820                 if (comp_size <= 1024) {
 821                         comp_block_size = 1024;
 822                 } else if (comp_size <= 2048) {
 823                         comp_block_size = 2048;
 824                 } else if (comp_size <= 4096) {
 825                         comp_block_size = 4096;
 826                 } else if (comp_size <= 8192) {
 827                         comp_block_size = 8192;
 828                 } else if (comp_size <= 16384) {
 829                         comp_block_size = 16384;
 830                 } else if (comp_size <= 32768) {
 831                         comp_block_size = 32768;
 832                 } else {
 833                         panic("hammer2: WRITE PATH: "
 834                               "Weird comp_size value.");
 835                         /* NOT REACHED */
 836                         comp_block_size = pblksize;
 837                 }
 838         }
 839
 840         chain = hammer2_assign_physical(ip, parentp, lbase, comp_block_size,
 841                                         mtid, errorp);
 842         if (*errorp) {
 843                 kprintf("WRITE PATH: An error occurred while "
 844                         "assigning physical space.\n");
 845                 KKASSERT(chain == NULL);
 846                 goto done;
 847         }
 848
 849         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
 850                 hammer2_inode_data_t *wipdata;
 851
 852                 hammer2_chain_modify_ip(ip, chain, mtid, 0);
 853                 wipdata = &chain->data->ipdata;
 854                 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
 855                 KKASSERT(bp->b_loffset == 0);
 856                 bcopy(bp->b_data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
 857         } else {
 858                 hammer2_io_t *dio;
 859                 char *bdata;
 860
 861                 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
 862
 863                 switch(chain->bref.type) {
 864                 case HAMMER2_BREF_TYPE_INODE:
 865                         panic("hammer2_write_bp: unexpected inode\n");
 866                         break;
 867                 case HAMMER2_BREF_TYPE_DATA:
 868                         /*
 869                          * Optimize out the read-before-write
 870                          * if possible.
 871                          */
 872                         *errorp = hammer2_io_newnz(chain->hmp,
 873                                                    chain->bref.data_off,
 874                                                    chain->bytes,
 875                                                    &dio);
 876                         if (*errorp) {
 877                                 hammer2_io_brelse(&dio);
 878                                 kprintf("hammer2: WRITE PATH: "
 879                                         "dbp bread error\n");
 880                                 break;
 881                         }
 882                         bdata = hammer2_io_data(dio, chain->bref.data_off);
 883
 884                         /*
 885                          * When loading the block make sure we don't
 886                          * leave garbage after the compressed data.
 887                          */
 888                         if (comp_size) {
 889                                 chain->bref.methods =
 890                                         HAMMER2_ENC_COMP(comp_algo) +
 891                                         HAMMER2_ENC_CHECK(check_algo);
 892                                 bcopy(comp_buffer, bdata, comp_size);
 893                                 if (comp_size != comp_block_size) {
 894                                         bzero(bdata + comp_size,
 895                                               comp_block_size - comp_size);
 896                                 }
 897                         } else {
 898                                 chain->bref.methods =
 899                                         HAMMER2_ENC_COMP(
 900                                                 HAMMER2_COMP_NONE) +
 901                                         HAMMER2_ENC_CHECK(check_algo);
 902                                 bcopy(bp->b_data, bdata, pblksize);
 903                         }
 904
 905                         /*
 906                          * The flush code doesn't calculate check codes for
 907                          * file data (doing so can result in excessive I/O),
 908                          * so we do it here.
 909                          */
 910                         hammer2_chain_setcheck(chain, bdata);
 911
 912                         /*
 913                          * Device buffer is now valid, chain is no longer in
 914                          * the initial state.
 915                          *
 916                          * (No blockref table worries with file data)
 917                          */
 918                         atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
 919
 920                         /* Now write the related bdp. */
 921                         if (ioflag & IO_SYNC) {
 922                                 /*
 923                                  * Synchronous I/O requested.
 924                                  */
 925                                 hammer2_io_bwrite(&dio);
 926                         /*
 927                         } else if ((ioflag & IO_DIRECT) &&
 928                                    loff + n == pblksize) {
 929                                 hammer2_io_bdwrite(&dio);
 930                         */
 931                         } else if (ioflag & IO_ASYNC) {
 932                                 hammer2_io_bawrite(&dio);
 933                         } else {
 934                                 hammer2_io_bdwrite(&dio);
 935                         }
 936                         break;
 937                 default:
 938                         panic("hammer2_write_bp: bad chain type %d\n",
 939                                 chain->bref.type);
 940                         /* NOT REACHED */
 941                         break;
 942                 }
 943         }
 944 done:
 945         if (chain) {
 946                 hammer2_chain_unlock(chain);
 947                 hammer2_chain_drop(chain);
 948         }
 949         if (comp_buffer)
 950                 objcache_put(cache_buffer_write, comp_buffer);
 951 }
 952
 953 /*
 954  * Helper
 955  *
 956  * Function that performs zero-checking and writing without compression,
 957  * it corresponds to default zero-checking path.
 958  */
 959 static
 960 void
 961 hammer2_zero_check_and_write(struct buf *bp, hammer2_inode_t *ip,
 962         hammer2_chain_t **parentp,
 963         hammer2_key_t lbase, int ioflag, int pblksize,
 964         hammer2_tid_t mtid, int *errorp,
 965         int check_algo)
 966 {
 967         hammer2_chain_t *chain;
 968
 969         if (test_block_zeros(bp->b_data, pblksize)) {
 970                 zero_write(bp, ip, parentp, lbase, mtid, errorp);
 971         } else {
 972                 chain = hammer2_assign_physical(ip, parentp, lbase, pblksize,
 973                                                 mtid, errorp);
 974                 hammer2_write_bp(chain, bp, ioflag, pblksize,
 975                                  mtid, errorp, check_algo);
 976                 if (chain) {
 977                         hammer2_chain_unlock(chain);
 978                         hammer2_chain_drop(chain);
 979                 }
 980         }
 981 }
 982
 983 /*
 984  * Helper
 985  *
 986  * A function to test whether a block of data contains only zeros,
 987  * returns TRUE (non-zero) if the block is all zeros.
 988  */
 989 static
 990 int
 991 test_block_zeros(const char *buf, size_t bytes)
 992 {
 993         size_t i;
 994
 995         for (i = 0; i < bytes; i += sizeof(long)) {
 996                 if (*(const long *)(buf + i) != 0)
 997                         return (0);
 998         }
 999         return (1);
1000 }
1001
1002 /*
1003  * Helper
1004  *
1005  * Function to "write" a block that contains only zeros.
1006  */
1007 static
1008 void
1009 zero_write(struct buf *bp, hammer2_inode_t *ip,
1010            hammer2_chain_t **parentp,
1011            hammer2_key_t lbase, hammer2_tid_t mtid, int *errorp __unused)
1012 {
1013         hammer2_chain_t *chain;
1014         hammer2_key_t key_dummy;
1015         int cache_index = -1;
1016
1017         chain = hammer2_chain_lookup(parentp, &key_dummy,
1018                                      lbase, lbase,
1019                                      &cache_index,
1020                                      HAMMER2_LOOKUP_NODATA);
1021         if (chain) {
1022                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1023                         hammer2_inode_data_t *wipdata;
1024
1025                         hammer2_chain_modify_ip(ip, chain, mtid, 0);
1026                         wipdata = &chain->data->ipdata;
1027                         KKASSERT(wipdata->meta.op_flags &
1028                                  HAMMER2_OPFLAG_DIRECTDATA);
1029                         KKASSERT(bp->b_loffset == 0);
1030                         bzero(wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1031                 } else {
1032                         hammer2_chain_delete(*parentp, chain,
1033                                              mtid, HAMMER2_DELETE_PERMANENT);
1034                 }
1035                 hammer2_chain_unlock(chain);
1036                 hammer2_chain_drop(chain);
1037         }
1038 }
1039
1040 /*
1041  * Helper
1042  *
1043  * Function to write the data as it is, without performing any sort of
1044  * compression. This function is used in path without compression and
1045  * default zero-checking path.
1046  */
1047 static
1048 void
1049 hammer2_write_bp(hammer2_chain_t *chain, struct buf *bp, int ioflag,
1050                  int pblksize,
1051                  hammer2_tid_t mtid, int *errorp, int check_algo)
1052 {
1053         hammer2_inode_data_t *wipdata;
1054         hammer2_io_t *dio;
1055         char *bdata;
1056         int error;
1057
1058         error = 0;      /* XXX TODO below */
1059
1060         KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1061
1062         switch(chain->bref.type) {
1063         case HAMMER2_BREF_TYPE_INODE:
1064                 wipdata = &chain->data->ipdata;
1065                 KKASSERT(wipdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA);
1066                 KKASSERT(bp->b_loffset == 0);
1067                 bcopy(bp->b_data, wipdata->u.data, HAMMER2_EMBEDDED_BYTES);
1068                 error = 0;
1069                 break;
1070         case HAMMER2_BREF_TYPE_DATA:
1071                 error = hammer2_io_newnz(chain->hmp,
1072                                          chain->bref.data_off,
1073                                          chain->bytes, &dio);
1074                 if (error) {
1075                         hammer2_io_bqrelse(&dio);
1076                         kprintf("hammer2: WRITE PATH: "
1077                                 "dbp bread error\n");
1078                         break;
1079                 }
1080                 bdata = hammer2_io_data(dio, chain->bref.data_off);
1081
1082                 chain->bref.methods = HAMMER2_ENC_COMP(
1083                                                 HAMMER2_COMP_NONE) +
1084                                       HAMMER2_ENC_CHECK(check_algo);
1085                 bcopy(bp->b_data, bdata, chain->bytes);
1086
1087                 /*
1088                  * The flush code doesn't calculate check codes for
1089                  * file data (doing so can result in excessive I/O),
1090                  * so we do it here.
1091                  */
1092                 hammer2_chain_setcheck(chain, bdata);
1093
1094                 /*
1095                  * Device buffer is now valid, chain is no longer in
1096                  * the initial state.
1097                  *
1098                  * (No blockref table worries with file data)
1099                  */
1100                 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1101
1102                 if (ioflag & IO_SYNC) {
1103                         /*
1104                          * Synchronous I/O requested.
1105                          */
1106                         hammer2_io_bwrite(&dio);
1107                 /*
1108                 } else if ((ioflag & IO_DIRECT) &&
1109                            loff + n == pblksize) {
1110                         hammer2_io_bdwrite(&dio);
1111                 */
1112                 } else if (ioflag & IO_ASYNC) {
1113                         hammer2_io_bawrite(&dio);
1114                 } else {
1115                         hammer2_io_bdwrite(&dio);
1116                 }
1117                 break;
1118         default:
1119                 panic("hammer2_write_bp: bad chain type %d\n",
1120                       chain->bref.type);
1121                 /* NOT REACHED */
1122                 error = 0;
1123                 break;
1124         }
1125         KKASSERT(error == 0);   /* XXX TODO */
1126         *errorp = error;
1127 }