sys/vfs/hammer/hammer_vnops.c

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34
  35 #include <sys/mountctl.h>
  36 #include <sys/namecache.h>
  37 #include <sys/buf2.h>
  38 #include <vfs/fifofs/fifo.h>
  39
  40 #include "hammer.h"
  41
  42 /*
  43  * USERFS VNOPS
  44  */
  45 static int hammer_vop_fsync(struct vop_fsync_args *);
  46 static int hammer_vop_read(struct vop_read_args *);
  47 static int hammer_vop_write(struct vop_write_args *);
  48 static int hammer_vop_access(struct vop_access_args *);
  49 static int hammer_vop_advlock(struct vop_advlock_args *);
  50 static int hammer_vop_close(struct vop_close_args *);
  51 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  52 static int hammer_vop_getattr(struct vop_getattr_args *);
  53 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  54 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  55 static int hammer_vop_nlink(struct vop_nlink_args *);
  56 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  57 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  58 static int hammer_vop_open(struct vop_open_args *);
  59 static int hammer_vop_print(struct vop_print_args *);
  60 static int hammer_vop_readdir(struct vop_readdir_args *);
  61 static int hammer_vop_readlink(struct vop_readlink_args *);
  62 static int hammer_vop_nremove(struct vop_nremove_args *);
  63 static int hammer_vop_nrename(struct vop_nrename_args *);
  64 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  65 static int hammer_vop_markatime(struct vop_markatime_args *);
  66 static int hammer_vop_setattr(struct vop_setattr_args *);
  67 static int hammer_vop_strategy(struct vop_strategy_args *);
  68 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  69 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  70 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  71 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  72 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  73 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
  74
  75 static int hammer_vop_fifoclose (struct vop_close_args *);
  76 static int hammer_vop_fiforead (struct vop_read_args *);
  77 static int hammer_vop_fifowrite (struct vop_write_args *);
  78 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
  79
  80 struct vop_ops hammer_vnode_vops = {
  81         .vop_default =          vop_defaultop,
  82         .vop_fsync =            hammer_vop_fsync,
  83         .vop_getpages =         vop_stdgetpages,
  84         .vop_putpages =         vop_stdputpages,
  85         .vop_read =             hammer_vop_read,
  86         .vop_write =            hammer_vop_write,
  87         .vop_access =           hammer_vop_access,
  88         .vop_advlock =          hammer_vop_advlock,
  89         .vop_close =            hammer_vop_close,
  90         .vop_ncreate =          hammer_vop_ncreate,
  91         .vop_getattr =          hammer_vop_getattr,
  92         .vop_inactive =         hammer_vop_inactive,
  93         .vop_reclaim =          hammer_vop_reclaim,
  94         .vop_nresolve =         hammer_vop_nresolve,
  95         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
  96         .vop_nlink =            hammer_vop_nlink,
  97         .vop_nmkdir =           hammer_vop_nmkdir,
  98         .vop_nmknod =           hammer_vop_nmknod,
  99         .vop_open =             hammer_vop_open,
 100         .vop_pathconf =         vop_stdpathconf,
 101         .vop_print =            hammer_vop_print,
 102         .vop_readdir =          hammer_vop_readdir,
 103         .vop_readlink =         hammer_vop_readlink,
 104         .vop_nremove =          hammer_vop_nremove,
 105         .vop_nrename =          hammer_vop_nrename,
 106         .vop_nrmdir =           hammer_vop_nrmdir,
 107         .vop_markatime =        hammer_vop_markatime,
 108         .vop_setattr =          hammer_vop_setattr,
 109         .vop_bmap =             hammer_vop_bmap,
 110         .vop_strategy =         hammer_vop_strategy,
 111         .vop_nsymlink =         hammer_vop_nsymlink,
 112         .vop_nwhiteout =        hammer_vop_nwhiteout,
 113         .vop_ioctl =            hammer_vop_ioctl,
 114         .vop_mountctl =         hammer_vop_mountctl,
 115         .vop_kqfilter =         hammer_vop_kqfilter
 116 };
 117
 118 struct vop_ops hammer_spec_vops = {
 119         .vop_default =          vop_defaultop,
 120         .vop_fsync =            hammer_vop_fsync,
 121         .vop_read =             vop_stdnoread,
 122         .vop_write =            vop_stdnowrite,
 123         .vop_access =           hammer_vop_access,
 124         .vop_close =            hammer_vop_close,
 125         .vop_markatime =        hammer_vop_markatime,
 126         .vop_getattr =          hammer_vop_getattr,
 127         .vop_inactive =         hammer_vop_inactive,
 128         .vop_reclaim =          hammer_vop_reclaim,
 129         .vop_setattr =          hammer_vop_setattr
 130 };
 131
 132 struct vop_ops hammer_fifo_vops = {
 133         .vop_default =          fifo_vnoperate,
 134         .vop_fsync =            hammer_vop_fsync,
 135         .vop_read =             hammer_vop_fiforead,
 136         .vop_write =            hammer_vop_fifowrite,
 137         .vop_access =           hammer_vop_access,
 138         .vop_close =            hammer_vop_fifoclose,
 139         .vop_markatime =        hammer_vop_markatime,
 140         .vop_getattr =          hammer_vop_getattr,
 141         .vop_inactive =         hammer_vop_inactive,
 142         .vop_reclaim =          hammer_vop_reclaim,
 143         .vop_setattr =          hammer_vop_setattr,
 144         .vop_kqfilter =         hammer_vop_fifokqfilter
 145 };
 146
 147 static __inline
 148 void
 149 hammer_knote(struct vnode *vp, int flags)
 150 {
 151         if (flags)
 152                 KNOTE(&vp->v_pollinfo.vpi_kqinfo.ki_note, flags);
 153 }
 154
 155 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 156                            struct vnode *dvp, struct ucred *cred,
 157                            int flags, int isdir);
 158 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 159 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 160
 161 /*
 162  * hammer_vop_fsync { vp, waitfor }
 163  *
 164  * fsync() an inode to disk and wait for it to be completely committed
 165  * such that the information would not be undone if a crash occured after
 166  * return.
 167  *
 168  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
 169  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
 170  *       operation.
 171  *
 172  *       Ultimately the combination of a REDO log and use of fast storage
 173  *       to front-end cluster caches will make fsync fast, but it aint
 174  *       here yet.  And, in anycase, we need real transactional
 175  *       all-or-nothing features which are not restricted to a single file.
 176  */
 177 static
 178 int
 179 hammer_vop_fsync(struct vop_fsync_args *ap)
 180 {
 181         hammer_inode_t ip = VTOI(ap->a_vp);
 182         hammer_mount_t hmp = ip->hmp;
 183         int waitfor = ap->a_waitfor;
 184         int mode;
 185
 186         lwkt_gettoken(&hmp->fs_token);
 187
 188         /*
 189          * Fsync rule relaxation (default is either full synchronous flush
 190          * or REDO semantics with synchronous flush).
 191          */
 192         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
 193                 switch(hammer_fsync_mode) {
 194                 case 0:
 195 mode0:
 196                         /* no REDO, full synchronous flush */
 197                         goto skip;
 198                 case 1:
 199 mode1:
 200                         /* no REDO, full asynchronous flush */
 201                         if (waitfor == MNT_WAIT)
 202                                 waitfor = MNT_NOWAIT;
 203                         goto skip;
 204                 case 2:
 205                         /* REDO semantics, synchronous flush */
 206                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 207                                 goto mode0;
 208                         mode = HAMMER_FLUSH_UNDOS_AUTO;
 209                         break;
 210                 case 3:
 211                         /* REDO semantics, relaxed asynchronous flush */
 212                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 213                                 goto mode1;
 214                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 215                         if (waitfor == MNT_WAIT)
 216                                 waitfor = MNT_NOWAIT;
 217                         break;
 218                 case 4:
 219                         /* ignore the fsync() system call */
 220                         lwkt_reltoken(&hmp->fs_token);
 221                         return(0);
 222                 default:
 223                         /* we have to do something */
 224                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 225                         if (waitfor == MNT_WAIT)
 226                                 waitfor = MNT_NOWAIT;
 227                         break;
 228                 }
 229
 230                 /*
 231                  * Fast fsync only needs to flush the UNDO/REDO fifo if
 232                  * HAMMER_INODE_REDO is non-zero and the only modifications
 233                  * made to the file are write or write-extends.
 234                  */
 235                 if ((ip->flags & HAMMER_INODE_REDO) &&
 236                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0) {
 237                         ++hammer_count_fsyncs;
 238                         hammer_flusher_flush_undos(hmp, mode);
 239                         ip->redo_count = 0;
 240                         if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
 241                                 vclrisdirty(ip->vp);
 242                         lwkt_reltoken(&hmp->fs_token);
 243                         return(0);
 244                 }
 245
 246                 /*
 247                  * REDO is enabled by fsync(), the idea being we really only
 248                  * want to lay down REDO records when programs are using
 249                  * fsync() heavily.  The first fsync() on the file starts
 250                  * the gravy train going and later fsync()s keep it hot by
 251                  * resetting the redo_count.
 252                  *
 253                  * We weren't running REDOs before now so we have to fall
 254                  * through and do a full fsync of what we have.
 255                  */
 256                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
 257                     (hmp->flags & HAMMER_MOUNT_REDO_RECOVERY_RUN) == 0) {
 258                         ip->flags |= HAMMER_INODE_REDO;
 259                         ip->redo_count = 0;
 260                 }
 261         }
 262 skip:
 263
 264         /*
 265          * Do a full flush sequence.
 266          *
 267          * Attempt to release the vnode while waiting for the inode to
 268          * finish flushing.  This can really mess up inactive->reclaim
 269          * sequences so only do it if the vnode is active.
 270          *
 271          * WARNING! The VX lock functions must be used.  vn_lock() will
 272          *          fail when this is part of a VOP_RECLAIM sequence.
 273          */
 274         ++hammer_count_fsyncs;
 275         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
 276         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 277         if (waitfor == MNT_WAIT) {
 278                 int dorelock;
 279
 280                 if ((ap->a_vp->v_flag & VRECLAIMED) == 0) {
 281                         vx_unlock(ap->a_vp);
 282                         dorelock = 1;
 283                 } else {
 284                         dorelock = 0;
 285                 }
 286                 hammer_wait_inode(ip);
 287                 if (dorelock)
 288                         vx_lock(ap->a_vp);
 289         }
 290         if (ip->vp && (ip->flags & HAMMER_INODE_MODMASK) == 0)
 291                 vclrisdirty(ip->vp);
 292         lwkt_reltoken(&hmp->fs_token);
 293         return (ip->error);
 294 }
 295
 296 /*
 297  * hammer_vop_read { vp, uio, ioflag, cred }
 298  *
 299  * MPSAFE (for the cache safe does not require fs_token)
 300  */
 301 static
 302 int
 303 hammer_vop_read(struct vop_read_args *ap)
 304 {
 305         struct hammer_transaction trans;
 306         hammer_inode_t ip;
 307         hammer_mount_t hmp;
 308         off_t offset;
 309         struct buf *bp;
 310         struct uio *uio;
 311         int error;
 312         int n;
 313         int seqcount;
 314         int ioseqcount;
 315         int blksize;
 316         int bigread;
 317         int got_trans;
 318         size_t resid;
 319
 320         if (ap->a_vp->v_type != VREG)
 321                 return (EINVAL);
 322         ip = VTOI(ap->a_vp);
 323         hmp = ip->hmp;
 324         error = 0;
 325         got_trans = 0;
 326         uio = ap->a_uio;
 327
 328         /*
 329          * Attempt to shortcut directly to the VM object using lwbufs.
 330          * This is much faster than instantiating buffer cache buffers.
 331          */
 332         resid = uio->uio_resid;
 333         error = vop_helper_read_shortcut(ap);
 334         hammer_stats_file_read += resid - uio->uio_resid;
 335         if (error)
 336                 return (error);
 337         if (uio->uio_resid == 0)
 338                 goto finished;
 339
 340         /*
 341          * Allow the UIO's size to override the sequential heuristic.
 342          */
 343         blksize = hammer_blocksize(uio->uio_offset);
 344         seqcount = (uio->uio_resid + (MAXBSIZE - 1)) / MAXBSIZE;
 345         ioseqcount = (ap->a_ioflag >> 16);
 346         if (seqcount < ioseqcount)
 347                 seqcount = ioseqcount;
 348
 349         /*
 350          * If reading or writing a huge amount of data we have to break
 351          * atomicy and allow the operation to be interrupted by a signal
 352          * or it can DOS the machine.
 353          */
 354         bigread = (uio->uio_resid > 100 * 1024 * 1024);
 355
 356         /*
 357          * Access the data typically in HAMMER_BUFSIZE blocks via the
 358          * buffer cache, but HAMMER may use a variable block size based
 359          * on the offset.
 360          *
 361          * XXX Temporary hack, delay the start transaction while we remain
 362          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
 363          *     locked-shared.
 364          */
 365         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 366                 int64_t base_offset;
 367                 int64_t file_limit;
 368
 369                 blksize = hammer_blocksize(uio->uio_offset);
 370                 offset = (int)uio->uio_offset & (blksize - 1);
 371                 base_offset = uio->uio_offset - offset;
 372
 373                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
 374                         break;
 375
 376                 /*
 377                  * MPSAFE
 378                  */
 379                 bp = getblk(ap->a_vp, base_offset, blksize, 0, 0);
 380                 if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) == B_CACHE) {
 381                         bp->b_flags &= ~B_AGE;
 382                         error = 0;
 383                         goto skip;
 384                 }
 385                 if (ap->a_ioflag & IO_NRDELAY) {
 386                         bqrelse(bp);
 387                         return (EWOULDBLOCK);
 388                 }
 389
 390                 /*
 391                  * MPUNSAFE
 392                  */
 393                 if (got_trans == 0) {
 394                         hammer_start_transaction(&trans, ip->hmp);
 395                         got_trans = 1;
 396                 }
 397
 398                 /*
 399                  * NOTE: A valid bp has already been acquired, but was not
 400                  *       B_CACHE.
 401                  */
 402                 if (hammer_cluster_enable) {
 403                         /*
 404                          * Use file_limit to prevent cluster_read() from
 405                          * creating buffers of the wrong block size past
 406                          * the demarc.
 407                          */
 408                         file_limit = ip->ino_data.size;
 409                         if (base_offset < HAMMER_XDEMARC &&
 410                             file_limit > HAMMER_XDEMARC) {
 411                                 file_limit = HAMMER_XDEMARC;
 412                         }
 413                         error = cluster_readx(ap->a_vp,
 414                                              file_limit, base_offset,
 415                                              blksize, B_NOTMETA,
 416                                              uio->uio_resid,
 417                                              seqcount * MAXBSIZE,
 418                                              &bp);
 419                 } else {
 420                         error = breadnx(ap->a_vp, base_offset,
 421                                         blksize, B_NOTMETA,
 422                                         NULL, NULL, 0, &bp);
 423                 }
 424                 if (error) {
 425                         brelse(bp);
 426                         break;
 427                 }
 428 skip:
 429                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IOISSUED)) {
 430                         hdkprintf("zone2_offset %016jx read file %016jx@%016jx\n",
 431                                 (intmax_t)bp->b_bio2.bio_offset,
 432                                 (intmax_t)ip->obj_id,
 433                                 (intmax_t)bp->b_loffset);
 434                 }
 435                 bp->b_flags &= ~B_IOISSUED;
 436                 if (blksize == HAMMER_XBUFSIZE)
 437                         bp->b_flags |= B_CLUSTEROK;
 438
 439                 n = blksize - offset;
 440                 if (n > uio->uio_resid)
 441                         n = uio->uio_resid;
 442                 if (n > ip->ino_data.size - uio->uio_offset)
 443                         n = (int)(ip->ino_data.size - uio->uio_offset);
 444
 445                 /*
 446                  * Set B_AGE, data has a lower priority than meta-data.
 447                  *
 448                  * Use a hold/unlock/drop sequence to run the uiomove
 449                  * with the buffer unlocked, avoiding deadlocks against
 450                  * read()s on mmap()'d spaces.
 451                  */
 452                 bp->b_flags |= B_AGE;
 453                 error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
 454                 bqrelse(bp);
 455
 456                 if (error)
 457                         break;
 458                 hammer_stats_file_read += n;
 459         }
 460
 461 finished:
 462
 463         /*
 464          * Try to update the atime with just the inode lock for maximum
 465          * concurrency.  If we can't shortcut it we have to get the full
 466          * blown transaction.
 467          */
 468         if (got_trans == 0 && hammer_update_atime_quick(ip) < 0) {
 469                 hammer_start_transaction(&trans, ip->hmp);
 470                 got_trans = 1;
 471         }
 472
 473         if (got_trans) {
 474                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 475                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 476                         lwkt_gettoken(&hmp->fs_token);
 477                         ip->ino_data.atime = trans.time;
 478                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
 479                         hammer_done_transaction(&trans);
 480                         lwkt_reltoken(&hmp->fs_token);
 481                 } else {
 482                         hammer_done_transaction(&trans);
 483                 }
 484         }
 485         return (error);
 486 }
 487
 488 /*
 489  * hammer_vop_write { vp, uio, ioflag, cred }
 490  */
 491 static
 492 int
 493 hammer_vop_write(struct vop_write_args *ap)
 494 {
 495         struct hammer_transaction trans;
 496         hammer_inode_t ip;
 497         hammer_mount_t hmp;
 498         thread_t td;
 499         struct uio *uio;
 500         int offset;
 501         off_t base_offset;
 502         int64_t cluster_eof;
 503         struct buf *bp;
 504         int kflags;
 505         int error;
 506         int n;
 507         int flags;
 508         int seqcount;
 509         int bigwrite;
 510
 511         if (ap->a_vp->v_type != VREG)
 512                 return (EINVAL);
 513         ip = VTOI(ap->a_vp);
 514         hmp = ip->hmp;
 515         error = 0;
 516         kflags = 0;
 517         seqcount = ap->a_ioflag >> 16;
 518
 519         if (ip->flags & HAMMER_INODE_RO)
 520                 return (EROFS);
 521
 522         /*
 523          * Create a transaction to cover the operations we perform.
 524          */
 525         hammer_start_transaction(&trans, hmp);
 526         uio = ap->a_uio;
 527
 528         /*
 529          * Check append mode
 530          */
 531         if (ap->a_ioflag & IO_APPEND)
 532                 uio->uio_offset = ip->ino_data.size;
 533
 534         /*
 535          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 536          *
 537          * NOTE: the base_off assignment is required to work around what
 538          * I consider to be a GCC-4 optimization bug.
 539          */
 540         if (uio->uio_offset < 0) {
 541                 hammer_done_transaction(&trans);
 542                 return (EFBIG);
 543         }
 544         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 545         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 546                 hammer_done_transaction(&trans);
 547                 return (EFBIG);
 548         }
 549
 550         if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
 551             base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
 552                 hammer_done_transaction(&trans);
 553                 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
 554                 return (EFBIG);
 555         }
 556
 557         /*
 558          * If reading or writing a huge amount of data we have to break
 559          * atomicy and allow the operation to be interrupted by a signal
 560          * or it can DOS the machine.
 561          *
 562          * Preset redo_count so we stop generating REDOs earlier if the
 563          * limit is exceeded.
 564          *
 565          * redo_count is heuristical, SMP races are ok
 566          */
 567         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 568         if ((ip->flags & HAMMER_INODE_REDO) &&
 569             ip->redo_count < hammer_limit_redo) {
 570                 ip->redo_count += uio->uio_resid;
 571         }
 572
 573         /*
 574          * Access the data typically in HAMMER_BUFSIZE blocks via the
 575          * buffer cache, but HAMMER may use a variable block size based
 576          * on the offset.
 577          */
 578         while (uio->uio_resid > 0) {
 579                 int fixsize = 0;
 580                 int blksize;
 581                 int blkmask;
 582                 int trivial;
 583                 int endofblk;
 584                 off_t nsize;
 585
 586                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 587                         break;
 588                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 589                         break;
 590
 591                 blksize = hammer_blocksize(uio->uio_offset);
 592
 593                 /*
 594                  * Control the number of pending records associated with
 595                  * this inode.  If too many have accumulated start a
 596                  * flush.  Try to maintain a pipeline with the flusher.
 597                  *
 598                  * NOTE: It is possible for other sources to grow the
 599                  *       records but not necessarily issue another flush,
 600                  *       so use a timeout and ensure that a re-flush occurs.
 601                  */
 602                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 603                         lwkt_gettoken(&hmp->fs_token);
 604                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 605                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 606                                 ip->flags |= HAMMER_INODE_RECSW;
 607                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 608                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 609                         }
 610                         lwkt_reltoken(&hmp->fs_token);
 611                 }
 612
 613                 /*
 614                  * Do not allow HAMMER to blow out the buffer cache.  Very
 615                  * large UIOs can lockout other processes due to bwillwrite()
 616                  * mechanics.
 617                  *
 618                  * The hammer inode is not locked during these operations.
 619                  * The vnode is locked which can interfere with the pageout
 620                  * daemon for non-UIO_NOCOPY writes but should not interfere
 621                  * with the buffer cache.  Even so, we cannot afford to
 622                  * allow the pageout daemon to build up too many dirty buffer
 623                  * cache buffers.
 624                  *
 625                  * Only call this if we aren't being recursively called from
 626                  * a virtual disk device (vn), else we may deadlock.
 627                  */
 628                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 629                         bwillwrite(blksize);
 630
 631                 /*
 632                  * Calculate the blocksize at the current offset and figure
 633                  * out how much we can actually write.
 634                  */
 635                 blkmask = blksize - 1;
 636                 offset = (int)uio->uio_offset & blkmask;
 637                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 638                 n = blksize - offset;
 639                 if (n > uio->uio_resid) {
 640                         n = uio->uio_resid;
 641                         endofblk = 0;
 642                 } else {
 643                         endofblk = 1;
 644                 }
 645                 nsize = uio->uio_offset + n;
 646                 if (nsize > ip->ino_data.size) {
 647                         if (uio->uio_offset > ip->ino_data.size)
 648                                 trivial = 0;
 649                         else
 650                                 trivial = 1;
 651                         nvextendbuf(ap->a_vp,
 652                                     ip->ino_data.size,
 653                                     nsize,
 654                                     hammer_blocksize(ip->ino_data.size),
 655                                     hammer_blocksize(nsize),
 656                                     hammer_blockoff(ip->ino_data.size),
 657                                     hammer_blockoff(nsize),
 658                                     trivial);
 659                         fixsize = 1;
 660                         kflags |= NOTE_EXTEND;
 661                 }
 662
 663                 if (uio->uio_segflg == UIO_NOCOPY) {
 664                         /*
 665                          * Issuing a write with the same data backing the
 666                          * buffer.  Instantiate the buffer to collect the
 667                          * backing vm pages, then read-in any missing bits.
 668                          *
 669                          * This case is used by vop_stdputpages().
 670                          */
 671                         bp = getblk(ap->a_vp, base_offset,
 672                                     blksize, GETBLK_BHEAVY, 0);
 673                         if ((bp->b_flags & B_CACHE) == 0) {
 674                                 bqrelse(bp);
 675                                 error = bread(ap->a_vp, base_offset,
 676                                               blksize, &bp);
 677                         }
 678                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 679                         /*
 680                          * Even though we are entirely overwriting the buffer
 681                          * we may still have to zero it out to avoid a
 682                          * mmap/write visibility issue.
 683                          */
 684                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 685                         if ((bp->b_flags & B_CACHE) == 0)
 686                                 vfs_bio_clrbuf(bp);
 687                 } else if (base_offset >= ip->ino_data.size) {
 688                         /*
 689                          * If the base offset of the buffer is beyond the
 690                          * file EOF, we don't have to issue a read.
 691                          */
 692                         bp = getblk(ap->a_vp, base_offset,
 693                                     blksize, GETBLK_BHEAVY, 0);
 694                         vfs_bio_clrbuf(bp);
 695                 } else {
 696                         /*
 697                          * Partial overwrite, read in any missing bits then
 698                          * replace the portion being written.
 699                          */
 700                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 701                         if (error == 0)
 702                                 bheavy(bp);
 703                 }
 704                 if (error == 0)
 705                         error = uiomovebp(bp, bp->b_data + offset, n, uio);
 706
 707                 lwkt_gettoken(&hmp->fs_token);
 708
 709                 /*
 710                  * Generate REDO records if enabled and redo_count will not
 711                  * exceeded the limit.
 712                  *
 713                  * If redo_count exceeds the limit we stop generating records
 714                  * and clear HAMMER_INODE_REDO.  This will cause the next
 715                  * fsync() to do a full meta-data sync instead of just an
 716                  * UNDO/REDO fifo update.
 717                  *
 718                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
 719                  * will still be tracked.  The tracks will be terminated
 720                  * when the related meta-data (including possible data
 721                  * modifications which are not tracked via REDO) is
 722                  * flushed.
 723                  */
 724                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
 725                         if (ip->redo_count < hammer_limit_redo) {
 726                                 bp->b_flags |= B_VFSFLAG1;
 727                                 error = hammer_generate_redo(&trans, ip,
 728                                                      base_offset + offset,
 729                                                      HAMMER_REDO_WRITE,
 730                                                      bp->b_data + offset,
 731                                                      (size_t)n);
 732                         } else {
 733                                 ip->flags &= ~HAMMER_INODE_REDO;
 734                         }
 735                 }
 736
 737                 /*
 738                  * If we screwed up we have to undo any VM size changes we
 739                  * made.
 740                  */
 741                 if (error) {
 742                         brelse(bp);
 743                         if (fixsize) {
 744                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
 745                                           hammer_blocksize(ip->ino_data.size),
 746                                           hammer_blockoff(ip->ino_data.size),
 747                                           0);
 748                         }
 749                         lwkt_reltoken(&hmp->fs_token);
 750                         break;
 751                 }
 752                 kflags |= NOTE_WRITE;
 753                 hammer_stats_file_write += n;
 754                 if (blksize == HAMMER_XBUFSIZE)
 755                         bp->b_flags |= B_CLUSTEROK;
 756                 if (ip->ino_data.size < uio->uio_offset) {
 757                         ip->ino_data.size = uio->uio_offset;
 758                         flags = HAMMER_INODE_SDIRTY;
 759                 } else {
 760                         flags = 0;
 761                 }
 762                 ip->ino_data.mtime = trans.time;
 763                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 764                 hammer_modify_inode(&trans, ip, flags);
 765
 766                 /*
 767                  * Once we dirty the buffer any cached zone-X offset
 768                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 769                  * allow overwriting over the same data sector unless
 770                  * we provide UNDOs for the old data, which we don't.
 771                  */
 772                 bp->b_bio2.bio_offset = NOOFFSET;
 773
 774                 lwkt_reltoken(&hmp->fs_token);
 775
 776                 /*
 777                  * Final buffer disposition.
 778                  *
 779                  * Because meta-data updates are deferred, HAMMER is
 780                  * especially sensitive to excessive bdwrite()s because
 781                  * the I/O stream is not broken up by disk reads.  So the
 782                  * buffer cache simply cannot keep up.
 783                  *
 784                  * WARNING!  blksize is variable.  cluster_write() is
 785                  *           expected to not blow up if it encounters
 786                  *           buffers that do not match the passed blksize.
 787                  *
 788                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 789                  *        The ip->rsv_recs check should burst-flush the data.
 790                  *        If we queue it immediately the buf could be left
 791                  *        locked on the device queue for a very long time.
 792                  *
 793                  *        However, failing to flush a dirty buffer out when
 794                  *        issued from the pageout daemon can result in a low
 795                  *        memory deadlock against bio_page_alloc(), so we
 796                  *        have to bawrite() on IO_ASYNC as well.
 797                  *
 798                  * NOTE!  To avoid degenerate stalls due to mismatched block
 799                  *        sizes we only honor IO_DIRECT on the write which
 800                  *        abuts the end of the buffer.  However, we must
 801                  *        honor IO_SYNC in case someone is silly enough to
 802                  *        configure a HAMMER file as swap, or when HAMMER
 803                  *        is serving NFS (for commits).  Ick ick.
 804                  */
 805                 bp->b_flags |= B_AGE;
 806                 if (blksize == HAMMER_XBUFSIZE)
 807                         bp->b_flags |= B_CLUSTEROK;
 808
 809                 if (ap->a_ioflag & IO_SYNC) {
 810                         bwrite(bp);
 811                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
 812                         bawrite(bp);
 813                 } else if (ap->a_ioflag & IO_ASYNC) {
 814                         bawrite(bp);
 815                 } else if (hammer_cluster_enable &&
 816                            !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 817                         if (base_offset < HAMMER_XDEMARC)
 818                                 cluster_eof = hammer_blockdemarc(base_offset,
 819                                                          ip->ino_data.size);
 820                         else
 821                                 cluster_eof = ip->ino_data.size;
 822                         cluster_write(bp, cluster_eof, blksize, seqcount);
 823                 } else {
 824                         bdwrite(bp);
 825                 }
 826         }
 827         hammer_done_transaction(&trans);
 828         hammer_knote(ap->a_vp, kflags);
 829
 830         return (error);
 831 }
 832
 833 /*
 834  * hammer_vop_access { vp, mode, cred }
 835  *
 836  * MPSAFE - does not require fs_token
 837  */
 838 static
 839 int
 840 hammer_vop_access(struct vop_access_args *ap)
 841 {
 842         hammer_inode_t ip = VTOI(ap->a_vp);
 843         uid_t uid;
 844         gid_t gid;
 845         int error;
 846
 847         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 848         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 849
 850         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 851                                   ip->ino_data.uflags);
 852         return (error);
 853 }
 854
 855 /*
 856  * hammer_vop_advlock { vp, id, op, fl, flags }
 857  *
 858  * MPSAFE - does not require fs_token
 859  */
 860 static
 861 int
 862 hammer_vop_advlock(struct vop_advlock_args *ap)
 863 {
 864         hammer_inode_t ip = VTOI(ap->a_vp);
 865
 866         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 867 }
 868
 869 /*
 870  * hammer_vop_close { vp, fflag }
 871  *
 872  * We can only sync-on-close for normal closes.  XXX disabled for now.
 873  */
 874 static
 875 int
 876 hammer_vop_close(struct vop_close_args *ap)
 877 {
 878 #if 0
 879         struct vnode *vp = ap->a_vp;
 880         hammer_inode_t ip = VTOI(vp);
 881         int waitfor;
 882         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 883                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 884                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 885                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 886                                 waitfor = MNT_WAIT;
 887                         else
 888                                 waitfor = MNT_NOWAIT;
 889                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 890                                        HAMMER_INODE_CLOSEASYNC);
 891                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 892                 }
 893         }
 894 #endif
 895         return (vop_stdclose(ap));
 896 }
 897
 898 /*
 899  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 900  *
 901  * The operating system has already ensured that the directory entry
 902  * does not exist and done all appropriate namespace locking.
 903  */
 904 static
 905 int
 906 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 907 {
 908         struct hammer_transaction trans;
 909         hammer_inode_t dip;
 910         hammer_inode_t nip;
 911         struct nchandle *nch;
 912         hammer_mount_t hmp;
 913         int error;
 914
 915         nch = ap->a_nch;
 916         dip = VTOI(ap->a_dvp);
 917         hmp = dip->hmp;
 918
 919         if (dip->flags & HAMMER_INODE_RO)
 920                 return (EROFS);
 921         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
 922                 return (error);
 923
 924         /*
 925          * Create a transaction to cover the operations we perform.
 926          */
 927         lwkt_gettoken(&hmp->fs_token);
 928         hammer_start_transaction(&trans, hmp);
 929
 930         /*
 931          * Create a new filesystem object of the requested type.  The
 932          * returned inode will be referenced and shared-locked to prevent
 933          * it from being moved to the flusher.
 934          */
 935         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 936                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 937                                     NULL, &nip);
 938         if (error) {
 939                 hkprintf("hammer_create_inode error %d\n", error);
 940                 hammer_done_transaction(&trans);
 941                 *ap->a_vpp = NULL;
 942                 lwkt_reltoken(&hmp->fs_token);
 943                 return (error);
 944         }
 945
 946         /*
 947          * Add the new filesystem object to the directory.  This will also
 948          * bump the inode's link count.
 949          */
 950         error = hammer_ip_add_direntry(&trans, dip,
 951                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 952                                         nip);
 953         if (error)
 954                 hkprintf("hammer_ip_add_direntry error %d\n", error);
 955
 956         /*
 957          * Finish up.
 958          */
 959         if (error) {
 960                 hammer_rel_inode(nip, 0);
 961                 hammer_done_transaction(&trans);
 962                 *ap->a_vpp = NULL;
 963         } else {
 964                 error = hammer_get_vnode(nip, ap->a_vpp);
 965                 hammer_done_transaction(&trans);
 966                 hammer_rel_inode(nip, 0);
 967                 if (error == 0) {
 968                         cache_setunresolved(ap->a_nch);
 969                         cache_setvp(ap->a_nch, *ap->a_vpp);
 970                 }
 971                 hammer_knote(ap->a_dvp, NOTE_WRITE);
 972         }
 973         lwkt_reltoken(&hmp->fs_token);
 974         return (error);
 975 }
 976
 977 /*
 978  * hammer_vop_getattr { vp, vap }
 979  *
 980  * Retrieve an inode's attribute information.  When accessing inodes
 981  * historically we fake the atime field to ensure consistent results.
 982  * The atime field is stored in the B-Tree element and allowed to be
 983  * updated without cycling the element.
 984  *
 985  * MPSAFE - does not require fs_token
 986  */
 987 static
 988 int
 989 hammer_vop_getattr(struct vop_getattr_args *ap)
 990 {
 991         hammer_inode_t ip = VTOI(ap->a_vp);
 992         struct vattr *vap = ap->a_vap;
 993
 994         /*
 995          * We want the fsid to be different when accessing a filesystem
 996          * with different as-of's so programs like diff don't think
 997          * the files are the same.
 998          *
 999          * We also want the fsid to be the same when comparing snapshots,
1000          * or when comparing mirrors (which might be backed by different
1001          * physical devices).  HAMMER fsids are based on the PFS's
1002          * shared_uuid field.
1003          *
1004          * XXX there is a chance of collision here.  The va_fsid reported
1005          * by stat is different from the more involved fsid used in the
1006          * mount structure.
1007          */
1008         hammer_lock_sh(&ip->lock);
1009         vap->va_fsid = ip->pfsm->fsid_udev ^ (uint32_t)ip->obj_asof ^
1010                        (uint32_t)(ip->obj_asof >> 32);
1011
1012         vap->va_fileid = ip->ino_leaf.base.obj_id;
1013         vap->va_mode = ip->ino_data.mode;
1014         vap->va_nlink = ip->ino_data.nlinks;
1015         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1016         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1017         vap->va_rmajor = 0;
1018         vap->va_rminor = 0;
1019         vap->va_size = ip->ino_data.size;
1020
1021         /*
1022          * Special case for @@PFS softlinks.  The actual size of the
1023          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1024          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
1025          *
1026          * Note that userspace hammer command does not allow users to
1027          * create a @@PFS softlink under an existing other PFS (id!=0)
1028          * so the ip localization here for @@PFS softlink is always 0.
1029          */
1030         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
1031             ip->ino_data.size == 10 &&
1032             ip->obj_asof == HAMMER_MAX_TID &&
1033             ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1034             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
1035                 if (hammer_is_pfs_slave(&ip->pfsm->pfsd))
1036                         vap->va_size = 26;
1037                 else
1038                         vap->va_size = 10;
1039         }
1040
1041         /*
1042          * We must provide a consistent atime and mtime for snapshots
1043          * so people can do a 'tar cf - ... | md5' on them and get
1044          * consistent results.
1045          */
1046         if (ip->flags & HAMMER_INODE_RO) {
1047                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
1048                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
1049         } else {
1050                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
1051                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
1052         }
1053         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
1054         vap->va_flags = ip->ino_data.uflags;
1055         vap->va_gen = 1;        /* hammer inums are unique for all time */
1056         vap->va_blocksize = HAMMER_BUFSIZE;
1057         if (ip->ino_data.size >= HAMMER_XDEMARC) {
1058                 vap->va_bytes = HAMMER_XBUFSIZE64_DOALIGN(ip->ino_data.size);
1059         } else if (ip->ino_data.size > HAMMER_HBUFSIZE) {
1060                 vap->va_bytes = HAMMER_BUFSIZE64_DOALIGN(ip->ino_data.size);
1061         } else {
1062                 vap->va_bytes = HAMMER_DATA_DOALIGN(ip->ino_data.size);
1063         }
1064
1065         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1066         vap->va_filerev = 0;    /* XXX */
1067         vap->va_uid_uuid = ip->ino_data.uid;
1068         vap->va_gid_uuid = ip->ino_data.gid;
1069         vap->va_fsid_uuid = ip->hmp->fsid;
1070         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1071                           VA_FSID_UUID_VALID;
1072
1073         switch (ip->ino_data.obj_type) {
1074         case HAMMER_OBJTYPE_CDEV:
1075         case HAMMER_OBJTYPE_BDEV:
1076                 vap->va_rmajor = ip->ino_data.rmajor;
1077                 vap->va_rminor = ip->ino_data.rminor;
1078                 break;
1079         default:
1080                 break;
1081         }
1082         hammer_unlock(&ip->lock);
1083         return(0);
1084 }
1085
1086 /*
1087  * hammer_vop_nresolve { nch, dvp, cred }
1088  *
1089  * Locate the requested directory entry.
1090  */
1091 static
1092 int
1093 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1094 {
1095         struct hammer_transaction trans;
1096         struct namecache *ncp;
1097         hammer_mount_t hmp;
1098         hammer_inode_t dip;
1099         hammer_inode_t ip;
1100         hammer_tid_t asof;
1101         struct hammer_cursor cursor;
1102         struct vnode *vp;
1103         int64_t namekey;
1104         int error;
1105         int i;
1106         int nlen;
1107         int flags;
1108         int ispfs;
1109         int64_t obj_id;
1110         uint32_t localization;
1111         uint32_t max_iterations;
1112
1113         /*
1114          * Misc initialization, plus handle as-of name extensions.  Look for
1115          * the '@@' extension.  Note that as-of files and directories cannot
1116          * be modified.
1117          */
1118         dip = VTOI(ap->a_dvp);
1119         ncp = ap->a_nch->ncp;
1120         asof = dip->obj_asof;
1121         localization = dip->obj_localization;   /* for code consistency */
1122         nlen = ncp->nc_nlen;
1123         flags = dip->flags & HAMMER_INODE_RO;
1124         ispfs = 0;
1125         hmp = dip->hmp;
1126
1127         lwkt_gettoken(&hmp->fs_token);
1128         hammer_simple_transaction(&trans, hmp);
1129
1130         for (i = 0; i < nlen; ++i) {
1131                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1132                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1133                                                   &ispfs, &asof, &localization);
1134                         if (error != 0) {
1135                                 i = nlen;
1136                                 break;
1137                         }
1138                         if (asof != HAMMER_MAX_TID)
1139                                 flags |= HAMMER_INODE_RO;
1140                         break;
1141                 }
1142         }
1143         nlen = i;
1144
1145         /*
1146          * If this is a PFS we dive into the PFS root inode
1147          */
1148         if (ispfs && nlen == 0) {
1149                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1150                                       asof, localization,
1151                                       flags, &error);
1152                 if (error == 0) {
1153                         error = hammer_get_vnode(ip, &vp);
1154                         hammer_rel_inode(ip, 0);
1155                 } else {
1156                         vp = NULL;
1157                 }
1158                 if (error == 0) {
1159                         vn_unlock(vp);
1160                         cache_setvp(ap->a_nch, vp);
1161                         vrele(vp);
1162                 }
1163                 goto done;
1164         }
1165
1166         /*
1167          * If there is no path component the time extension is relative to dip.
1168          * e.g. "fubar/@@<snapshot>"
1169          *
1170          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1171          * e.g. "fubar/.@@<snapshot>"
1172          *
1173          * ".." is handled by the kernel.  We do not currently handle
1174          * "..@<snapshot>".
1175          */
1176         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1177                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1178                                       asof, dip->obj_localization,
1179                                       flags, &error);
1180                 if (error == 0) {
1181                         error = hammer_get_vnode(ip, &vp);
1182                         hammer_rel_inode(ip, 0);
1183                 } else {
1184                         vp = NULL;
1185                 }
1186                 if (error == 0) {
1187                         vn_unlock(vp);
1188                         cache_setvp(ap->a_nch, vp);
1189                         vrele(vp);
1190                 }
1191                 goto done;
1192         }
1193
1194         /*
1195          * Calculate the namekey and setup the key range for the scan.  This
1196          * works kinda like a chained hash table where the lower 32 bits
1197          * of the namekey synthesize the chain.
1198          *
1199          * The key range is inclusive of both key_beg and key_end.
1200          */
1201         namekey = hammer_direntry_namekey(dip, ncp->nc_name, nlen,
1202                                            &max_iterations);
1203
1204         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1205         cursor.key_beg.localization = dip->obj_localization |
1206                                       hammer_dir_localization(dip);
1207         cursor.key_beg.obj_id = dip->obj_id;
1208         cursor.key_beg.key = namekey;
1209         cursor.key_beg.create_tid = 0;
1210         cursor.key_beg.delete_tid = 0;
1211         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1212         cursor.key_beg.obj_type = 0;
1213
1214         cursor.key_end = cursor.key_beg;
1215         cursor.key_end.key += max_iterations;
1216         cursor.asof = asof;
1217         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1218
1219         /*
1220          * Scan all matching records (the chain), locate the one matching
1221          * the requested path component.
1222          *
1223          * The hammer_ip_*() functions merge in-memory records with on-disk
1224          * records for the purposes of the search.
1225          */
1226         obj_id = 0;
1227         localization = HAMMER_DEF_LOCALIZATION;
1228
1229         if (error == 0) {
1230                 error = hammer_ip_first(&cursor);
1231                 while (error == 0) {
1232                         error = hammer_ip_resolve_data(&cursor);
1233                         if (error)
1234                                 break;
1235                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1236                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1237                                 obj_id = cursor.data->entry.obj_id;
1238                                 localization = cursor.data->entry.localization;
1239                                 break;
1240                         }
1241                         error = hammer_ip_next(&cursor);
1242                 }
1243         }
1244         hammer_done_cursor(&cursor);
1245
1246         /*
1247          * Lookup the obj_id.  This should always succeed.  If it does not
1248          * the filesystem may be damaged and we return a dummy inode.
1249          */
1250         if (error == 0) {
1251                 ip = hammer_get_inode(&trans, dip, obj_id,
1252                                       asof, localization,
1253                                       flags, &error);
1254                 if (error == ENOENT) {
1255                         hkprintf("WARNING: Missing inode for dirent \"%s\"\n"
1256                                 "\tobj_id = %016jx, asof=%016jx, lo=%08x\n",
1257                                 ncp->nc_name,
1258                                 (intmax_t)obj_id, (intmax_t)asof,
1259                                 localization);
1260                         error = 0;
1261                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1262                                                     asof, localization,
1263                                                     flags, &error);
1264                 }
1265                 if (error == 0) {
1266                         error = hammer_get_vnode(ip, &vp);
1267                         hammer_rel_inode(ip, 0);
1268                 } else {
1269                         vp = NULL;
1270                 }
1271                 if (error == 0) {
1272                         vn_unlock(vp);
1273                         cache_setvp(ap->a_nch, vp);
1274                         vrele(vp);
1275                 }
1276         } else if (error == ENOENT) {
1277                 cache_setvp(ap->a_nch, NULL);
1278         }
1279 done:
1280         hammer_done_transaction(&trans);
1281         lwkt_reltoken(&hmp->fs_token);
1282         return (error);
1283 }
1284
1285 /*
1286  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1287  *
1288  * Locate the parent directory of a directory vnode.
1289  *
1290  * dvp is referenced but not locked.  *vpp must be returned referenced and
1291  * locked.  A parent_obj_id of 0 indicates that we are at the root.
1292  *
1293  * NOTE: as-of sequences are not linked into the directory structure.  If
1294  * we are at the root with a different asof then the mount point, reload
1295  * the same directory with the mount point's asof.   I'm not sure what this
1296  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1297  * get confused, but it hasn't been tested.
1298  */
1299 static
1300 int
1301 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1302 {
1303         struct hammer_transaction trans;
1304         hammer_inode_t dip;
1305         hammer_inode_t ip;
1306         hammer_mount_t hmp;
1307         int64_t parent_obj_id;
1308         uint32_t parent_obj_localization;
1309         hammer_tid_t asof;
1310         int error;
1311
1312         dip = VTOI(ap->a_dvp);
1313         asof = dip->obj_asof;
1314         hmp = dip->hmp;
1315
1316         /*
1317          * Whos are parent?  This could be the root of a pseudo-filesystem
1318          * whos parent is in another localization domain.
1319          */
1320         lwkt_gettoken(&hmp->fs_token);
1321         parent_obj_id = dip->ino_data.parent_obj_id;
1322         if (dip->obj_id == HAMMER_OBJID_ROOT)
1323                 parent_obj_localization = HAMMER_DEF_LOCALIZATION;
1324         else
1325                 parent_obj_localization = dip->obj_localization;
1326
1327         /*
1328          * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1329          */
1330         if (parent_obj_id == 0) {
1331                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1332                    asof != hmp->asof) {
1333                         parent_obj_id = dip->obj_id;
1334                         asof = hmp->asof;
1335                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1336                         ksnprintf(*ap->a_fakename, 19, "0x%016jx",
1337                                   (intmax_t)dip->obj_asof);
1338                 } else {
1339                         *ap->a_vpp = NULL;
1340                         lwkt_reltoken(&hmp->fs_token);
1341                         return ENOENT;
1342                 }
1343         }
1344
1345         hammer_simple_transaction(&trans, hmp);
1346
1347         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1348                               asof, parent_obj_localization,
1349                               dip->flags, &error);
1350         if (ip) {
1351                 error = hammer_get_vnode(ip, ap->a_vpp);
1352                 hammer_rel_inode(ip, 0);
1353         } else {
1354                 *ap->a_vpp = NULL;
1355         }
1356         hammer_done_transaction(&trans);
1357         lwkt_reltoken(&hmp->fs_token);
1358         return (error);
1359 }
1360
1361 /*
1362  * hammer_vop_nlink { nch, dvp, vp, cred }
1363  */
1364 static
1365 int
1366 hammer_vop_nlink(struct vop_nlink_args *ap)
1367 {
1368         struct hammer_transaction trans;
1369         hammer_inode_t dip;
1370         hammer_inode_t ip;
1371         struct nchandle *nch;
1372         hammer_mount_t hmp;
1373         int error;
1374
1375         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1376                 return(EXDEV);
1377
1378         nch = ap->a_nch;
1379         dip = VTOI(ap->a_dvp);
1380         ip = VTOI(ap->a_vp);
1381         hmp = dip->hmp;
1382
1383         if (dip->obj_localization != ip->obj_localization)
1384                 return(EXDEV);
1385
1386         if (dip->flags & HAMMER_INODE_RO)
1387                 return (EROFS);
1388         if (ip->flags & HAMMER_INODE_RO)
1389                 return (EROFS);
1390         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1391                 return (error);
1392
1393         /*
1394          * Create a transaction to cover the operations we perform.
1395          */
1396         lwkt_gettoken(&hmp->fs_token);
1397         hammer_start_transaction(&trans, hmp);
1398
1399         /*
1400          * Add the filesystem object to the directory.  Note that neither
1401          * dip nor ip are referenced or locked, but their vnodes are
1402          * referenced.  This function will bump the inode's link count.
1403          */
1404         error = hammer_ip_add_direntry(&trans, dip,
1405                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1406                                         ip);
1407
1408         /*
1409          * Finish up.
1410          */
1411         if (error == 0) {
1412                 cache_setunresolved(nch);
1413                 cache_setvp(nch, ap->a_vp);
1414         }
1415         hammer_done_transaction(&trans);
1416         hammer_knote(ap->a_vp, NOTE_LINK);
1417         hammer_knote(ap->a_dvp, NOTE_WRITE);
1418         lwkt_reltoken(&hmp->fs_token);
1419         return (error);
1420 }
1421
1422 /*
1423  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1424  *
1425  * The operating system has already ensured that the directory entry
1426  * does not exist and done all appropriate namespace locking.
1427  */
1428 static
1429 int
1430 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1431 {
1432         struct hammer_transaction trans;
1433         hammer_inode_t dip;
1434         hammer_inode_t nip;
1435         struct nchandle *nch;
1436         hammer_mount_t hmp;
1437         int error;
1438
1439         nch = ap->a_nch;
1440         dip = VTOI(ap->a_dvp);
1441         hmp = dip->hmp;
1442
1443         if (dip->flags & HAMMER_INODE_RO)
1444                 return (EROFS);
1445         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1446                 return (error);
1447
1448         /*
1449          * Create a transaction to cover the operations we perform.
1450          */
1451         lwkt_gettoken(&hmp->fs_token);
1452         hammer_start_transaction(&trans, hmp);
1453
1454         /*
1455          * Create a new filesystem object of the requested type.  The
1456          * returned inode will be referenced but not locked.
1457          */
1458         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1459                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1460                                     NULL, &nip);
1461         if (error) {
1462                 hammer_done_transaction(&trans);
1463                 *ap->a_vpp = NULL;
1464                 lwkt_reltoken(&hmp->fs_token);
1465                 return (error);
1466         }
1467         /*
1468          * Add the new filesystem object to the directory.  This will also
1469          * bump the inode's link count.
1470          */
1471         error = hammer_ip_add_direntry(&trans, dip,
1472                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1473                                         nip);
1474         if (error)
1475                 hkprintf("hammer_mkdir (add) error %d\n", error);
1476
1477         /*
1478          * Finish up.
1479          */
1480         if (error) {
1481                 hammer_rel_inode(nip, 0);
1482                 *ap->a_vpp = NULL;
1483         } else {
1484                 error = hammer_get_vnode(nip, ap->a_vpp);
1485                 hammer_rel_inode(nip, 0);
1486                 if (error == 0) {
1487                         cache_setunresolved(ap->a_nch);
1488                         cache_setvp(ap->a_nch, *ap->a_vpp);
1489                 }
1490         }
1491         hammer_done_transaction(&trans);
1492         if (error == 0)
1493                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1494         lwkt_reltoken(&hmp->fs_token);
1495         return (error);
1496 }
1497
1498 /*
1499  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1500  *
1501  * The operating system has already ensured that the directory entry
1502  * does not exist and done all appropriate namespace locking.
1503  */
1504 static
1505 int
1506 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1507 {
1508         struct hammer_transaction trans;
1509         hammer_inode_t dip;
1510         hammer_inode_t nip;
1511         struct nchandle *nch;
1512         hammer_mount_t hmp;
1513         int error;
1514
1515         nch = ap->a_nch;
1516         dip = VTOI(ap->a_dvp);
1517         hmp = dip->hmp;
1518
1519         if (dip->flags & HAMMER_INODE_RO)
1520                 return (EROFS);
1521         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1522                 return (error);
1523
1524         /*
1525          * Create a transaction to cover the operations we perform.
1526          */
1527         lwkt_gettoken(&hmp->fs_token);
1528         hammer_start_transaction(&trans, hmp);
1529
1530         /*
1531          * Create a new filesystem object of the requested type.  The
1532          * returned inode will be referenced but not locked.
1533          *
1534          * If mknod specifies a directory a pseudo-fs is created.
1535          */
1536         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1537                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1538                                     NULL, &nip);
1539         if (error) {
1540                 hammer_done_transaction(&trans);
1541                 *ap->a_vpp = NULL;
1542                 lwkt_reltoken(&hmp->fs_token);
1543                 return (error);
1544         }
1545
1546         /*
1547          * Add the new filesystem object to the directory.  This will also
1548          * bump the inode's link count.
1549          */
1550         error = hammer_ip_add_direntry(&trans, dip,
1551                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1552                                         nip);
1553
1554         /*
1555          * Finish up.
1556          */
1557         if (error) {
1558                 hammer_rel_inode(nip, 0);
1559                 *ap->a_vpp = NULL;
1560         } else {
1561                 error = hammer_get_vnode(nip, ap->a_vpp);
1562                 hammer_rel_inode(nip, 0);
1563                 if (error == 0) {
1564                         cache_setunresolved(ap->a_nch);
1565                         cache_setvp(ap->a_nch, *ap->a_vpp);
1566                 }
1567         }
1568         hammer_done_transaction(&trans);
1569         if (error == 0)
1570                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1571         lwkt_reltoken(&hmp->fs_token);
1572         return (error);
1573 }
1574
1575 /*
1576  * hammer_vop_open { vp, mode, cred, fp }
1577  *
1578  * MPSAFE (does not require fs_token)
1579  */
1580 static
1581 int
1582 hammer_vop_open(struct vop_open_args *ap)
1583 {
1584         hammer_inode_t ip;
1585
1586         ip = VTOI(ap->a_vp);
1587
1588         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1589                 return (EROFS);
1590         return(vop_stdopen(ap));
1591 }
1592
1593 /*
1594  * hammer_vop_print { vp }
1595  */
1596 static
1597 int
1598 hammer_vop_print(struct vop_print_args *ap)
1599 {
1600         return EOPNOTSUPP;
1601 }
1602
1603 /*
1604  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1605  */
1606 static
1607 int
1608 hammer_vop_readdir(struct vop_readdir_args *ap)
1609 {
1610         struct hammer_transaction trans;
1611         struct hammer_cursor cursor;
1612         hammer_inode_t ip;
1613         hammer_mount_t hmp;
1614         struct uio *uio;
1615         hammer_base_elm_t base;
1616         int error;
1617         int cookie_index;
1618         int ncookies;
1619         off_t *cookies;
1620         off_t saveoff;
1621         int r;
1622         int dtype;
1623
1624         ip = VTOI(ap->a_vp);
1625         uio = ap->a_uio;
1626         saveoff = uio->uio_offset;
1627         hmp = ip->hmp;
1628
1629         if (ap->a_ncookies) {
1630                 ncookies = uio->uio_resid / 16 + 1;
1631                 if (ncookies > 1024)
1632                         ncookies = 1024;
1633                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1634                 cookie_index = 0;
1635         } else {
1636                 ncookies = -1;
1637                 cookies = NULL;
1638                 cookie_index = 0;
1639         }
1640
1641         lwkt_gettoken(&hmp->fs_token);
1642         hammer_simple_transaction(&trans, hmp);
1643
1644         /*
1645          * Handle artificial entries
1646          *
1647          * It should be noted that the minimum value for a directory
1648          * hash key on-media is 0x0000000100000000, so we can use anything
1649          * less then that to represent our 'special' key space.
1650          */
1651         error = 0;
1652         if (saveoff == 0) {
1653                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1654                 if (r)
1655                         goto done;
1656                 if (cookies)
1657                         cookies[cookie_index] = saveoff;
1658                 ++saveoff;
1659                 ++cookie_index;
1660                 if (cookie_index == ncookies)
1661                         goto done;
1662         }
1663         if (saveoff == 1) {
1664                 if (ip->ino_data.parent_obj_id) {
1665                         r = vop_write_dirent(&error, uio,
1666                                              ip->ino_data.parent_obj_id,
1667                                              DT_DIR, 2, "..");
1668                 } else {
1669                         r = vop_write_dirent(&error, uio,
1670                                              ip->obj_id, DT_DIR, 2, "..");
1671                 }
1672                 if (r)
1673                         goto done;
1674                 if (cookies)
1675                         cookies[cookie_index] = saveoff;
1676                 ++saveoff;
1677                 ++cookie_index;
1678                 if (cookie_index == ncookies)
1679                         goto done;
1680         }
1681
1682         /*
1683          * Key range (begin and end inclusive) to scan.  Directory keys
1684          * directly translate to a 64 bit 'seek' position.
1685          */
1686         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1687         cursor.key_beg.localization = ip->obj_localization |
1688                                       hammer_dir_localization(ip);
1689         cursor.key_beg.obj_id = ip->obj_id;
1690         cursor.key_beg.create_tid = 0;
1691         cursor.key_beg.delete_tid = 0;
1692         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1693         cursor.key_beg.obj_type = 0;
1694         cursor.key_beg.key = saveoff;
1695
1696         cursor.key_end = cursor.key_beg;
1697         cursor.key_end.key = HAMMER_MAX_KEY;
1698         cursor.asof = ip->obj_asof;
1699         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1700
1701         error = hammer_ip_first(&cursor);
1702
1703         while (error == 0) {
1704                 error = hammer_ip_resolve_data(&cursor);
1705                 if (error)
1706                         break;
1707                 base = &cursor.leaf->base;
1708                 saveoff = base->key;
1709                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1710
1711                 if (base->obj_id != ip->obj_id)
1712                         hpanic("bad record at %p", cursor.node);
1713
1714                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1715                 r = vop_write_dirent(
1716                              &error, uio, cursor.data->entry.obj_id,
1717                              dtype,
1718                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1719                              (void *)cursor.data->entry.name);
1720                 if (r)
1721                         break;
1722                 ++saveoff;
1723                 if (cookies)
1724                         cookies[cookie_index] = base->key;
1725                 ++cookie_index;
1726                 if (cookie_index == ncookies)
1727                         break;
1728                 error = hammer_ip_next(&cursor);
1729         }
1730         hammer_done_cursor(&cursor);
1731
1732 done:
1733         hammer_done_transaction(&trans);
1734
1735         if (ap->a_eofflag)
1736                 *ap->a_eofflag = (error == ENOENT);
1737         uio->uio_offset = saveoff;
1738         if (error && cookie_index == 0) {
1739                 if (error == ENOENT)
1740                         error = 0;
1741                 if (cookies) {
1742                         kfree(cookies, M_TEMP);
1743                         *ap->a_ncookies = 0;
1744                         *ap->a_cookies = NULL;
1745                 }
1746         } else {
1747                 if (error == ENOENT)
1748                         error = 0;
1749                 if (cookies) {
1750                         *ap->a_ncookies = cookie_index;
1751                         *ap->a_cookies = cookies;
1752                 }
1753         }
1754         lwkt_reltoken(&hmp->fs_token);
1755         return(error);
1756 }
1757
1758 /*
1759  * hammer_vop_readlink { vp, uio, cred }
1760  */
1761 static
1762 int
1763 hammer_vop_readlink(struct vop_readlink_args *ap)
1764 {
1765         struct hammer_transaction trans;
1766         struct hammer_cursor cursor;
1767         hammer_inode_t ip;
1768         hammer_mount_t hmp;
1769         char buf[32];
1770         uint32_t localization;
1771         hammer_pseudofs_inmem_t pfsm;
1772         int error;
1773
1774         ip = VTOI(ap->a_vp);
1775         hmp = ip->hmp;
1776
1777         lwkt_gettoken(&hmp->fs_token);
1778
1779         /*
1780          * Shortcut if the symlink data was stuffed into ino_data.
1781          *
1782          * Also expand special "@@PFS%05d" softlinks (expansion only
1783          * occurs for non-historical (current) accesses made from the
1784          * primary filesystem).
1785          *
1786          * Note that userspace hammer command does not allow users to
1787          * create a @@PFS softlink under an existing other PFS (id!=0)
1788          * so the ip localization here for @@PFS softlink is always 0.
1789          */
1790         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1791                 char *ptr;
1792                 int bytes;
1793
1794                 ptr = ip->ino_data.ext.symlink;
1795                 bytes = (int)ip->ino_data.size;
1796                 if (bytes == 10 &&
1797                     ip->obj_asof == HAMMER_MAX_TID &&
1798                     ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1799                     strncmp(ptr, "@@PFS", 5) == 0) {
1800                         hammer_simple_transaction(&trans, hmp);
1801                         bcopy(ptr + 5, buf, 5);
1802                         buf[5] = 0;
1803                         localization = pfs_to_lo(strtoul(buf, NULL, 10));
1804                         pfsm = hammer_load_pseudofs(&trans, localization,
1805                                                     &error);
1806                         if (error == 0) {
1807                                 if (hammer_is_pfs_slave(&pfsm->pfsd)) {
1808                                         /* vap->va_size == 26 */
1809                                         ksnprintf(buf, sizeof(buf),
1810                                                   "@@0x%016jx:%05d",
1811                                                   (intmax_t)pfsm->pfsd.sync_end_tid,
1812                                                   lo_to_pfs(localization));
1813                                 } else {
1814                                         /* vap->va_size == 10 */
1815                                         ksnprintf(buf, sizeof(buf),
1816                                                   "@@-1:%05d",
1817                                                   lo_to_pfs(localization));
1818                                 }
1819                                 ptr = buf;
1820                                 bytes = strlen(buf);
1821                         }
1822                         if (pfsm)
1823                                 hammer_rel_pseudofs(hmp, pfsm);
1824                         hammer_done_transaction(&trans);
1825                 }
1826                 error = uiomove(ptr, bytes, ap->a_uio);
1827                 lwkt_reltoken(&hmp->fs_token);
1828                 return(error);
1829         }
1830
1831         /*
1832          * Long version
1833          */
1834         hammer_simple_transaction(&trans, hmp);
1835         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1836
1837         /*
1838          * Key range (begin and end inclusive) to scan.  Directory keys
1839          * directly translate to a 64 bit 'seek' position.
1840          */
1841         cursor.key_beg.localization = ip->obj_localization |
1842                                       HAMMER_LOCALIZE_MISC;
1843         cursor.key_beg.obj_id = ip->obj_id;
1844         cursor.key_beg.create_tid = 0;
1845         cursor.key_beg.delete_tid = 0;
1846         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1847         cursor.key_beg.obj_type = 0;
1848         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1849         cursor.asof = ip->obj_asof;
1850         cursor.flags |= HAMMER_CURSOR_ASOF;
1851
1852         error = hammer_ip_lookup(&cursor);
1853         if (error == 0) {
1854                 error = hammer_ip_resolve_data(&cursor);
1855                 if (error == 0) {
1856                         KKASSERT(cursor.leaf->data_len >=
1857                                  HAMMER_SYMLINK_NAME_OFF);
1858                         error = uiomove(cursor.data->symlink.name,
1859                                         cursor.leaf->data_len -
1860                                                 HAMMER_SYMLINK_NAME_OFF,
1861                                         ap->a_uio);
1862                 }
1863         }
1864         hammer_done_cursor(&cursor);
1865         hammer_done_transaction(&trans);
1866         lwkt_reltoken(&hmp->fs_token);
1867         return(error);
1868 }
1869
1870 /*
1871  * hammer_vop_nremove { nch, dvp, cred }
1872  */
1873 static
1874 int
1875 hammer_vop_nremove(struct vop_nremove_args *ap)
1876 {
1877         struct hammer_transaction trans;
1878         hammer_inode_t dip;
1879         hammer_mount_t hmp;
1880         int error;
1881
1882         dip = VTOI(ap->a_dvp);
1883         hmp = dip->hmp;
1884
1885         if (hammer_nohistory(dip) == 0 &&
1886             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1887                 return (error);
1888         }
1889
1890         lwkt_gettoken(&hmp->fs_token);
1891         hammer_start_transaction(&trans, hmp);
1892         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1893         hammer_done_transaction(&trans);
1894         if (error == 0)
1895                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1896         lwkt_reltoken(&hmp->fs_token);
1897         return (error);
1898 }
1899
1900 /*
1901  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1902  */
1903 static
1904 int
1905 hammer_vop_nrename(struct vop_nrename_args *ap)
1906 {
1907         struct hammer_transaction trans;
1908         struct namecache *fncp;
1909         struct namecache *tncp;
1910         hammer_inode_t fdip;
1911         hammer_inode_t tdip;
1912         hammer_inode_t ip;
1913         hammer_mount_t hmp;
1914         struct hammer_cursor cursor;
1915         int64_t namekey;
1916         uint32_t max_iterations;
1917         int nlen, error;
1918
1919         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1920                 return(EXDEV);
1921         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1922                 return(EXDEV);
1923
1924         fdip = VTOI(ap->a_fdvp);
1925         tdip = VTOI(ap->a_tdvp);
1926         fncp = ap->a_fnch->ncp;
1927         tncp = ap->a_tnch->ncp;
1928         ip = VTOI(fncp->nc_vp);
1929         KKASSERT(ip != NULL);
1930
1931         hmp = ip->hmp;
1932
1933         if (fdip->obj_localization != tdip->obj_localization)
1934                 return(EXDEV);
1935         if (fdip->obj_localization != ip->obj_localization)
1936                 return(EXDEV);
1937
1938         if (fdip->flags & HAMMER_INODE_RO)
1939                 return (EROFS);
1940         if (tdip->flags & HAMMER_INODE_RO)
1941                 return (EROFS);
1942         if (ip->flags & HAMMER_INODE_RO)
1943                 return (EROFS);
1944         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1945                 return (error);
1946
1947         lwkt_gettoken(&hmp->fs_token);
1948         hammer_start_transaction(&trans, hmp);
1949
1950         /*
1951          * Remove tncp from the target directory and then link ip as
1952          * tncp. XXX pass trans to dounlink
1953          *
1954          * Force the inode sync-time to match the transaction so it is
1955          * in-sync with the creation of the target directory entry.
1956          */
1957         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1958                                 ap->a_cred, 0, -1);
1959         if (error == 0 || error == ENOENT) {
1960                 error = hammer_ip_add_direntry(&trans, tdip,
1961                                                 tncp->nc_name, tncp->nc_nlen,
1962                                                 ip);
1963                 if (error == 0) {
1964                         ip->ino_data.parent_obj_id = tdip->obj_id;
1965                         ip->ino_data.ctime = trans.time;
1966                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1967                 }
1968         }
1969         if (error)
1970                 goto failed; /* XXX */
1971
1972         /*
1973          * Locate the record in the originating directory and remove it.
1974          *
1975          * Calculate the namekey and setup the key range for the scan.  This
1976          * works kinda like a chained hash table where the lower 32 bits
1977          * of the namekey synthesize the chain.
1978          *
1979          * The key range is inclusive of both key_beg and key_end.
1980          */
1981         namekey = hammer_direntry_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1982                                            &max_iterations);
1983 retry:
1984         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1985         cursor.key_beg.localization = fdip->obj_localization |
1986                                       hammer_dir_localization(fdip);
1987         cursor.key_beg.obj_id = fdip->obj_id;
1988         cursor.key_beg.key = namekey;
1989         cursor.key_beg.create_tid = 0;
1990         cursor.key_beg.delete_tid = 0;
1991         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1992         cursor.key_beg.obj_type = 0;
1993
1994         cursor.key_end = cursor.key_beg;
1995         cursor.key_end.key += max_iterations;
1996         cursor.asof = fdip->obj_asof;
1997         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1998
1999         /*
2000          * Scan all matching records (the chain), locate the one matching
2001          * the requested path component.
2002          *
2003          * The hammer_ip_*() functions merge in-memory records with on-disk
2004          * records for the purposes of the search.
2005          */
2006         error = hammer_ip_first(&cursor);
2007         while (error == 0) {
2008                 if (hammer_ip_resolve_data(&cursor) != 0)
2009                         break;
2010                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2011                 KKASSERT(nlen > 0);
2012                 if (fncp->nc_nlen == nlen &&
2013                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2014                         break;
2015                 }
2016                 error = hammer_ip_next(&cursor);
2017         }
2018
2019         /*
2020          * If all is ok we have to get the inode so we can adjust nlinks.
2021          *
2022          * WARNING: hammer_ip_del_direntry() may have to terminate the
2023          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2024          * twice.
2025          */
2026         if (error == 0)
2027                 error = hammer_ip_del_direntry(&trans, &cursor, fdip, ip);
2028
2029         /*
2030          * XXX A deadlock here will break rename's atomicy for the purposes
2031          * of crash recovery.
2032          */
2033         if (error == EDEADLK) {
2034                 hammer_done_cursor(&cursor);
2035                 goto retry;
2036         }
2037
2038         /*
2039          * Cleanup and tell the kernel that the rename succeeded.
2040          *
2041          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2042          *       without formally acquiring the vp since the vp might
2043          *       have zero refs on it, or in the middle of a reclaim,
2044          *       etc.
2045          */
2046         hammer_done_cursor(&cursor);
2047         if (error == 0) {
2048                 cache_rename(ap->a_fnch, ap->a_tnch);
2049                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2050                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2051                 while (ip->vp) {
2052                         struct vnode *vp;
2053
2054                         error = hammer_get_vnode(ip, &vp);
2055                         if (error == 0 && vp) {
2056                                 vn_unlock(vp);
2057                                 hammer_knote(ip->vp, NOTE_RENAME);
2058                                 vrele(vp);
2059                                 break;
2060                         }
2061                         hdkprintf("ip/vp race2 avoided\n");
2062                 }
2063         }
2064
2065 failed:
2066         hammer_done_transaction(&trans);
2067         lwkt_reltoken(&hmp->fs_token);
2068         return (error);
2069 }
2070
2071 /*
2072  * hammer_vop_nrmdir { nch, dvp, cred }
2073  */
2074 static
2075 int
2076 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2077 {
2078         struct hammer_transaction trans;
2079         hammer_inode_t dip;
2080         hammer_mount_t hmp;
2081         int error;
2082
2083         dip = VTOI(ap->a_dvp);
2084         hmp = dip->hmp;
2085
2086         if (hammer_nohistory(dip) == 0 &&
2087             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2088                 return (error);
2089         }
2090
2091         lwkt_gettoken(&hmp->fs_token);
2092         hammer_start_transaction(&trans, hmp);
2093         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2094         hammer_done_transaction(&trans);
2095         if (error == 0)
2096                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2097         lwkt_reltoken(&hmp->fs_token);
2098         return (error);
2099 }
2100
2101 /*
2102  * hammer_vop_markatime { vp, cred }
2103  */
2104 static
2105 int
2106 hammer_vop_markatime(struct vop_markatime_args *ap)
2107 {
2108         struct hammer_transaction trans;
2109         hammer_inode_t ip;
2110         hammer_mount_t hmp;
2111
2112         ip = VTOI(ap->a_vp);
2113         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2114                 return (EROFS);
2115         if (ip->flags & HAMMER_INODE_RO)
2116                 return (EROFS);
2117         hmp = ip->hmp;
2118         if (hmp->mp->mnt_flag & MNT_NOATIME)
2119                 return (0);
2120         lwkt_gettoken(&hmp->fs_token);
2121         hammer_start_transaction(&trans, hmp);
2122
2123         ip->ino_data.atime = trans.time;
2124         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2125         hammer_done_transaction(&trans);
2126         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2127         lwkt_reltoken(&hmp->fs_token);
2128         return (0);
2129 }
2130
2131 /*
2132  * hammer_vop_setattr { vp, vap, cred }
2133  */
2134 static
2135 int
2136 hammer_vop_setattr(struct vop_setattr_args *ap)
2137 {
2138         struct hammer_transaction trans;
2139         hammer_inode_t ip;
2140         struct vattr *vap;
2141         hammer_mount_t hmp;
2142         int modflags;
2143         int error;
2144         int truncating;
2145         int blksize;
2146         int kflags;
2147 #if 0
2148         int64_t aligned_size;
2149 #endif
2150         uint32_t flags;
2151
2152         vap = ap->a_vap;
2153         ip = ap->a_vp->v_data;
2154         modflags = 0;
2155         kflags = 0;
2156         hmp = ip->hmp;
2157
2158         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2159                 return(EROFS);
2160         if (ip->flags & HAMMER_INODE_RO)
2161                 return (EROFS);
2162         if (hammer_nohistory(ip) == 0 &&
2163             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2164                 return (error);
2165         }
2166
2167         lwkt_gettoken(&hmp->fs_token);
2168         hammer_start_transaction(&trans, hmp);
2169         error = 0;
2170
2171         if (vap->va_flags != VNOVAL) {
2172                 flags = ip->ino_data.uflags;
2173                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2174                                          hammer_to_unix_xid(&ip->ino_data.uid),
2175                                          ap->a_cred);
2176                 if (error == 0) {
2177                         if (ip->ino_data.uflags != flags) {
2178                                 ip->ino_data.uflags = flags;
2179                                 ip->ino_data.ctime = trans.time;
2180                                 modflags |= HAMMER_INODE_DDIRTY;
2181                                 kflags |= NOTE_ATTRIB;
2182                         }
2183                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2184                                 error = 0;
2185                                 goto done;
2186                         }
2187                 }
2188                 goto done;
2189         }
2190         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2191                 error = EPERM;
2192                 goto done;
2193         }
2194         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2195                 mode_t cur_mode = ip->ino_data.mode;
2196                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2197                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2198                 hammer_uuid_t uuid_uid;
2199                 hammer_uuid_t uuid_gid;
2200
2201                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2202                                          ap->a_cred,
2203                                          &cur_uid, &cur_gid, &cur_mode);
2204                 if (error == 0) {
2205                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2206                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2207                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2208                                  sizeof(uuid_uid)) ||
2209                             bcmp(&uuid_gid, &ip->ino_data.gid,
2210                                  sizeof(uuid_gid)) ||
2211                             ip->ino_data.mode != cur_mode) {
2212                                 ip->ino_data.uid = uuid_uid;
2213                                 ip->ino_data.gid = uuid_gid;
2214                                 ip->ino_data.mode = cur_mode;
2215                                 ip->ino_data.ctime = trans.time;
2216                                 modflags |= HAMMER_INODE_DDIRTY;
2217                         }
2218                         kflags |= NOTE_ATTRIB;
2219                 }
2220         }
2221         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2222                 switch(ap->a_vp->v_type) {
2223                 case VREG:
2224                         if (vap->va_size == ip->ino_data.size)
2225                                 break;
2226
2227                         /*
2228                          * Log the operation if in fast-fsync mode or if
2229                          * there are unterminated redo write records present.
2230                          *
2231                          * The second check is needed so the recovery code
2232                          * properly truncates write redos even if nominal
2233                          * REDO operations is turned off due to excessive
2234                          * writes, because the related records might be
2235                          * destroyed and never lay down a TERM_WRITE.
2236                          */
2237                         if ((ip->flags & HAMMER_INODE_REDO) ||
2238                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2239                                 error = hammer_generate_redo(&trans, ip,
2240                                                              vap->va_size,
2241                                                              HAMMER_REDO_TRUNC,
2242                                                              NULL, 0);
2243                         }
2244                         blksize = hammer_blocksize(vap->va_size);
2245
2246                         /*
2247                          * XXX break atomicy, we can deadlock the backend
2248                          * if we do not release the lock.  Probably not a
2249                          * big deal here.
2250                          */
2251                         if (vap->va_size < ip->ino_data.size) {
2252                                 nvtruncbuf(ap->a_vp, vap->va_size,
2253                                            blksize,
2254                                            hammer_blockoff(vap->va_size),
2255                                            0);
2256                                 truncating = 1;
2257                                 kflags |= NOTE_WRITE;
2258                         } else {
2259                                 nvextendbuf(ap->a_vp,
2260                                             ip->ino_data.size,
2261                                             vap->va_size,
2262                                             hammer_blocksize(ip->ino_data.size),
2263                                             hammer_blocksize(vap->va_size),
2264                                             hammer_blockoff(ip->ino_data.size),
2265                                             hammer_blockoff(vap->va_size),
2266                                             0);
2267                                 truncating = 0;
2268                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2269                         }
2270                         ip->ino_data.size = vap->va_size;
2271                         ip->ino_data.mtime = trans.time;
2272                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2273                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2274
2275                         /*
2276                          * On-media truncation is cached in the inode until
2277                          * the inode is synchronized.  We must immediately
2278                          * handle any frontend records.
2279                          */
2280                         if (truncating) {
2281                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2282                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2283                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2284                                         ip->trunc_off = vap->va_size;
2285                                         hammer_inode_dirty(ip);
2286                                 } else if (ip->trunc_off > vap->va_size) {
2287                                         ip->trunc_off = vap->va_size;
2288                                 }
2289                         }
2290
2291 #if 0
2292                         /*
2293                          * When truncating, nvtruncbuf() may have cleaned out
2294                          * a portion of the last block on-disk in the buffer
2295                          * cache.  We must clean out any frontend records
2296                          * for blocks beyond the new last block.
2297                          */
2298                         aligned_size = (vap->va_size + (blksize - 1)) &
2299                                        ~(int64_t)(blksize - 1);
2300                         if (truncating && vap->va_size < aligned_size) {
2301                                 aligned_size -= blksize;
2302                                 hammer_ip_frontend_trunc(ip, aligned_size);
2303                         }
2304 #endif
2305                         break;
2306                 case VDATABASE:
2307                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2308                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2309                                 ip->trunc_off = vap->va_size;
2310                                 hammer_inode_dirty(ip);
2311                         } else if (ip->trunc_off > vap->va_size) {
2312                                 ip->trunc_off = vap->va_size;
2313                         }
2314                         hammer_ip_frontend_trunc(ip, vap->va_size);
2315                         ip->ino_data.size = vap->va_size;
2316                         ip->ino_data.mtime = trans.time;
2317                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2318                         kflags |= NOTE_ATTRIB;
2319                         break;
2320                 default:
2321                         error = EINVAL;
2322                         goto done;
2323                 }
2324                 break;
2325         }
2326         if (vap->va_atime.tv_sec != VNOVAL) {
2327                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2328                 modflags |= HAMMER_INODE_ATIME;
2329                 kflags |= NOTE_ATTRIB;
2330         }
2331         if (vap->va_mtime.tv_sec != VNOVAL) {
2332                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2333                 modflags |= HAMMER_INODE_MTIME;
2334                 kflags |= NOTE_ATTRIB;
2335         }
2336         if (vap->va_mode != (mode_t)VNOVAL) {
2337                 mode_t   cur_mode = ip->ino_data.mode;
2338                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2339                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2340
2341                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2342                                          cur_uid, cur_gid, &cur_mode);
2343                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2344                         ip->ino_data.mode = cur_mode;
2345                         ip->ino_data.ctime = trans.time;
2346                         modflags |= HAMMER_INODE_DDIRTY;
2347                         kflags |= NOTE_ATTRIB;
2348                 }
2349         }
2350 done:
2351         if (error == 0)
2352                 hammer_modify_inode(&trans, ip, modflags);
2353         hammer_done_transaction(&trans);
2354         hammer_knote(ap->a_vp, kflags);
2355         lwkt_reltoken(&hmp->fs_token);
2356         return (error);
2357 }
2358
2359 /*
2360  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2361  */
2362 static
2363 int
2364 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2365 {
2366         struct hammer_transaction trans;
2367         hammer_inode_t dip;
2368         hammer_inode_t nip;
2369         hammer_record_t record;
2370         struct nchandle *nch;
2371         hammer_mount_t hmp;
2372         int error;
2373         int bytes;
2374
2375         ap->a_vap->va_type = VLNK;
2376
2377         nch = ap->a_nch;
2378         dip = VTOI(ap->a_dvp);
2379         hmp = dip->hmp;
2380
2381         if (dip->flags & HAMMER_INODE_RO)
2382                 return (EROFS);
2383         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2384                 return (error);
2385
2386         /*
2387          * Create a transaction to cover the operations we perform.
2388          */
2389         lwkt_gettoken(&hmp->fs_token);
2390         hammer_start_transaction(&trans, hmp);
2391
2392         /*
2393          * Create a new filesystem object of the requested type.  The
2394          * returned inode will be referenced but not locked.
2395          */
2396
2397         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2398                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2399                                     NULL, &nip);
2400         if (error) {
2401                 hammer_done_transaction(&trans);
2402                 *ap->a_vpp = NULL;
2403                 lwkt_reltoken(&hmp->fs_token);
2404                 return (error);
2405         }
2406
2407         /*
2408          * Add a record representing the symlink.  symlink stores the link
2409          * as pure data, not a string, and is no \0 terminated.
2410          */
2411         if (error == 0) {
2412                 bytes = strlen(ap->a_target);
2413
2414                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2415                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2416                 } else {
2417                         record = hammer_alloc_mem_record(nip, bytes);
2418                         record->type = HAMMER_MEM_RECORD_GENERAL;
2419
2420                         record->leaf.base.localization = nip->obj_localization |
2421                                                          HAMMER_LOCALIZE_MISC;
2422                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2423                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2424                         record->leaf.data_len = bytes;
2425                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2426                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2427                         error = hammer_ip_add_record(&trans, record);
2428                 }
2429
2430                 /*
2431                  * Set the file size to the length of the link.
2432                  */
2433                 if (error == 0) {
2434                         nip->ino_data.size = bytes;
2435                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2436                 }
2437         }
2438         if (error == 0)
2439                 error = hammer_ip_add_direntry(&trans, dip, nch->ncp->nc_name,
2440                                                 nch->ncp->nc_nlen, nip);
2441
2442         /*
2443          * Finish up.
2444          */
2445         if (error) {
2446                 hammer_rel_inode(nip, 0);
2447                 *ap->a_vpp = NULL;
2448         } else {
2449                 error = hammer_get_vnode(nip, ap->a_vpp);
2450                 hammer_rel_inode(nip, 0);
2451                 if (error == 0) {
2452                         cache_setunresolved(ap->a_nch);
2453                         cache_setvp(ap->a_nch, *ap->a_vpp);
2454                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2455                 }
2456         }
2457         hammer_done_transaction(&trans);
2458         lwkt_reltoken(&hmp->fs_token);
2459         return (error);
2460 }
2461
2462 /*
2463  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2464  */
2465 static
2466 int
2467 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2468 {
2469         struct hammer_transaction trans;
2470         hammer_inode_t dip;
2471         hammer_mount_t hmp;
2472         int error;
2473
2474         dip = VTOI(ap->a_dvp);
2475         hmp = dip->hmp;
2476
2477         if (hammer_nohistory(dip) == 0 &&
2478             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2479                 return (error);
2480         }
2481
2482         lwkt_gettoken(&hmp->fs_token);
2483         hammer_start_transaction(&trans, hmp);
2484         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2485                                 ap->a_cred, ap->a_flags, -1);
2486         hammer_done_transaction(&trans);
2487         lwkt_reltoken(&hmp->fs_token);
2488
2489         return (error);
2490 }
2491
2492 /*
2493  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2494  */
2495 static
2496 int
2497 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2498 {
2499         hammer_inode_t ip = ap->a_vp->v_data;
2500         hammer_mount_t hmp = ip->hmp;
2501         int error;
2502
2503         lwkt_gettoken(&hmp->fs_token);
2504         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2505                              ap->a_fflag, ap->a_cred);
2506         lwkt_reltoken(&hmp->fs_token);
2507         return (error);
2508 }
2509
2510 static
2511 int
2512 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2513 {
2514         static const struct mountctl_opt extraopt[] = {
2515                 { HMNT_NOHISTORY,       "nohistory" },
2516                 { HMNT_MASTERID,        "master" },
2517                 { HMNT_NOMIRROR,        "nomirror" },
2518                 { 0, NULL}
2519
2520         };
2521         hammer_mount_t hmp;
2522         struct mount *mp;
2523         int usedbytes;
2524         int error;
2525
2526         error = 0;
2527         usedbytes = 0;
2528         mp = ap->a_head.a_ops->head.vv_mount;
2529         KKASSERT(mp->mnt_data != NULL);
2530         hmp = (hammer_mount_t)mp->mnt_data;
2531
2532         lwkt_gettoken(&hmp->fs_token);
2533
2534         switch(ap->a_op) {
2535         case MOUNTCTL_SET_EXPORT:
2536                 if (ap->a_ctllen != sizeof(struct export_args))
2537                         error = EINVAL;
2538                 else
2539                         error = hammer_vfs_export(mp, ap->a_op,
2540                                       (const struct export_args *)ap->a_ctl);
2541                 break;
2542         case MOUNTCTL_MOUNTFLAGS:
2543                 /*
2544                  * Call standard mountctl VOP function
2545                  * so we get user mount flags.
2546                  */
2547                 error = vop_stdmountctl(ap);
2548                 if (error)
2549                         break;
2550
2551                 usedbytes = *ap->a_res;
2552
2553                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2554                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2555                                                     ap->a_buf,
2556                                                     ap->a_buflen - usedbytes,
2557                                                     &error);
2558                 }
2559
2560                 *ap->a_res += usedbytes;
2561                 break;
2562         default:
2563                 error = vop_stdmountctl(ap);
2564                 break;
2565         }
2566         lwkt_reltoken(&hmp->fs_token);
2567         return(error);
2568 }
2569
2570 /*
2571  * hammer_vop_strategy { vp, bio }
2572  *
2573  * Strategy call, used for regular file read & write only.  Note that the
2574  * bp may represent a cluster.
2575  *
2576  * To simplify operation and allow better optimizations in the future,
2577  * this code does not make any assumptions with regards to buffer alignment
2578  * or size.
2579  */
2580 static
2581 int
2582 hammer_vop_strategy(struct vop_strategy_args *ap)
2583 {
2584         struct buf *bp;
2585         int error;
2586
2587         bp = ap->a_bio->bio_buf;
2588
2589         switch(bp->b_cmd) {
2590         case BUF_CMD_READ:
2591                 error = hammer_vop_strategy_read(ap);
2592                 break;
2593         case BUF_CMD_WRITE:
2594                 error = hammer_vop_strategy_write(ap);
2595                 break;
2596         default:
2597                 bp->b_error = error = EINVAL;
2598                 bp->b_flags |= B_ERROR;
2599                 biodone(ap->a_bio);
2600                 break;
2601         }
2602
2603         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2604
2605         return (error);
2606 }
2607
2608 /*
2609  * Read from a regular file.  Iterate the related records and fill in the
2610  * BIO/BUF.  Gaps are zero-filled.
2611  *
2612  * The support code in hammer_object.c should be used to deal with mixed
2613  * in-memory and on-disk records.
2614  *
2615  * NOTE: Can be called from the cluster code with an oversized buf.
2616  *
2617  * XXX atime update
2618  */
2619 static
2620 int
2621 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2622 {
2623         struct hammer_transaction trans;
2624         hammer_inode_t ip;
2625         hammer_inode_t dip;
2626         hammer_mount_t hmp;
2627         struct hammer_cursor cursor;
2628         hammer_base_elm_t base;
2629         hammer_off_t disk_offset;
2630         struct bio *bio;
2631         struct bio *nbio;
2632         struct buf *bp;
2633         int64_t rec_offset;
2634         int64_t ran_end;
2635         int64_t tmp64;
2636         int error;
2637         int boff;
2638         int roff;
2639         int n;
2640         int isdedupable;
2641
2642         bio = ap->a_bio;
2643         bp = bio->bio_buf;
2644         ip = ap->a_vp->v_data;
2645         hmp = ip->hmp;
2646
2647         /*
2648          * The zone-2 disk offset may have been set by the cluster code via
2649          * a BMAP operation, or else should be NOOFFSET.
2650          *
2651          * Checking the high bits for a match against zone-2 should suffice.
2652          *
2653          * In cases where a lot of data duplication is present it may be
2654          * more beneficial to drop through and doubule-buffer through the
2655          * device.
2656          */
2657         nbio = push_bio(bio);
2658         if (hammer_is_zone_large_data(nbio->bio_offset)) {
2659                 if (hammer_double_buffer == 0) {
2660                         lwkt_gettoken(&hmp->fs_token);
2661                         error = hammer_io_direct_read(hmp, nbio, NULL);
2662                         lwkt_reltoken(&hmp->fs_token);
2663                         return (error);
2664                 }
2665
2666                 /*
2667                  * Try to shortcut requests for double_buffer mode too.
2668                  * Since this mode runs through the device buffer cache
2669                  * only compatible buffer sizes (meaning those generated
2670                  * by normal filesystem buffers) are legal.
2671                  */
2672                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2673                         lwkt_gettoken(&hmp->fs_token);
2674                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2675                         lwkt_reltoken(&hmp->fs_token);
2676                         return (error);
2677                 }
2678         }
2679
2680         /*
2681          * Well, that sucked.  Do it the hard way.  If all the stars are
2682          * aligned we may still be able to issue a direct-read.
2683          */
2684         lwkt_gettoken(&hmp->fs_token);
2685         hammer_simple_transaction(&trans, hmp);
2686         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2687
2688         /*
2689          * Key range (begin and end inclusive) to scan.  Note that the key's
2690          * stored in the actual records represent BASE+LEN, not BASE.  The
2691          * first record containing bio_offset will have a key > bio_offset.
2692          */
2693         cursor.key_beg.localization = ip->obj_localization |
2694                                       HAMMER_LOCALIZE_MISC;
2695         cursor.key_beg.obj_id = ip->obj_id;
2696         cursor.key_beg.create_tid = 0;
2697         cursor.key_beg.delete_tid = 0;
2698         cursor.key_beg.obj_type = 0;
2699         cursor.key_beg.key = bio->bio_offset + 1;
2700         cursor.asof = ip->obj_asof;
2701         cursor.flags |= HAMMER_CURSOR_ASOF;
2702
2703         cursor.key_end = cursor.key_beg;
2704         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2705 #if 0
2706         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2707                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2708                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2709                 cursor.key_end.key = HAMMER_MAX_KEY;
2710         } else
2711 #endif
2712         {
2713                 ran_end = bio->bio_offset + bp->b_bufsize;
2714                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2715                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2716                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2717                 if (tmp64 < ran_end)
2718                         cursor.key_end.key = HAMMER_MAX_KEY;
2719                 else
2720                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2721         }
2722         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2723
2724         /*
2725          * Set NOSWAPCACHE for cursor data extraction if double buffering
2726          * is disabled or (if the file is not marked cacheable via chflags
2727          * and vm.swapcache_use_chflags is enabled).
2728          */
2729         if (hammer_double_buffer == 0 ||
2730             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2731              vm_swapcache_use_chflags)) {
2732                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2733         }
2734
2735         error = hammer_ip_first(&cursor);
2736         boff = 0;
2737
2738         while (error == 0) {
2739                 /*
2740                  * Get the base file offset of the record.  The key for
2741                  * data records is (base + bytes) rather then (base).
2742                  */
2743                 base = &cursor.leaf->base;
2744                 rec_offset = base->key - cursor.leaf->data_len;
2745
2746                 /*
2747                  * Calculate the gap, if any, and zero-fill it.
2748                  *
2749                  * n is the offset of the start of the record verses our
2750                  * current seek offset in the bio.
2751                  */
2752                 n = (int)(rec_offset - (bio->bio_offset + boff));
2753                 if (n > 0) {
2754                         if (n > bp->b_bufsize - boff)
2755                                 n = bp->b_bufsize - boff;
2756                         bzero((char *)bp->b_data + boff, n);
2757                         boff += n;
2758                         n = 0;
2759                 }
2760
2761                 /*
2762                  * Calculate the data offset in the record and the number
2763                  * of bytes we can copy.
2764                  *
2765                  * There are two degenerate cases.  First, boff may already
2766                  * be at bp->b_bufsize.  Secondly, the data offset within
2767                  * the record may exceed the record's size.
2768                  */
2769                 roff = -n;
2770                 rec_offset += roff;
2771                 n = cursor.leaf->data_len - roff;
2772                 if (n <= 0) {
2773                         hdkprintf("bad n=%d roff=%d\n", n, roff);
2774                         n = 0;
2775                 } else if (n > bp->b_bufsize - boff) {
2776                         n = bp->b_bufsize - boff;
2777                 }
2778
2779                 /*
2780                  * Deal with cached truncations.  This cool bit of code
2781                  * allows truncate()/ftruncate() to avoid having to sync
2782                  * the file.
2783                  *
2784                  * If the frontend is truncated then all backend records are
2785                  * subject to the frontend's truncation.
2786                  *
2787                  * If the backend is truncated then backend records on-disk
2788                  * (but not in-memory) are subject to the backend's
2789                  * truncation.  In-memory records owned by the backend
2790                  * represent data written after the truncation point on the
2791                  * backend and must not be truncated.
2792                  *
2793                  * Truncate operations deal with frontend buffer cache
2794                  * buffers and frontend-owned in-memory records synchronously.
2795                  */
2796                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2797                         if (hammer_cursor_ondisk(&cursor)/* ||
2798                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2799                                 if (ip->trunc_off <= rec_offset)
2800                                         n = 0;
2801                                 else if (ip->trunc_off < rec_offset + n)
2802                                         n = (int)(ip->trunc_off - rec_offset);
2803                         }
2804                 }
2805                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2806                         if (hammer_cursor_ondisk(&cursor)) {
2807                                 if (ip->sync_trunc_off <= rec_offset)
2808                                         n = 0;
2809                                 else if (ip->sync_trunc_off < rec_offset + n)
2810                                         n = (int)(ip->sync_trunc_off - rec_offset);
2811                         }
2812                 }
2813
2814                 /*
2815                  * Try to issue a direct read into our bio if possible,
2816                  * otherwise resolve the element data into a hammer_buffer
2817                  * and copy.
2818                  *
2819                  * The buffer on-disk should be zerod past any real
2820                  * truncation point, but may not be for any synthesized
2821                  * truncation point from above.
2822                  *
2823                  * NOTE: disk_offset is only valid if the cursor data is
2824                  *       on-disk.
2825                  */
2826                 disk_offset = cursor.leaf->data_offset + roff;
2827                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2828                                hammer_cursor_ondisk(&cursor) &&
2829                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2830
2831                 if (isdedupable && hammer_double_buffer == 0) {
2832                         /*
2833                          * Direct read case
2834                          */
2835                         KKASSERT(hammer_is_zone_large_data(disk_offset));
2836                         nbio->bio_offset = disk_offset;
2837                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2838                         if (hammer_live_dedup && error == 0)
2839                                 hammer_dedup_cache_add(ip, cursor.leaf);
2840                         goto done;
2841                 } else if (isdedupable) {
2842                         /*
2843                          * Async I/O case for reading from backing store
2844                          * and copying the data to the filesystem buffer.
2845                          * live-dedup has to verify the data anyway if it
2846                          * gets a hit later so we can just add the entry
2847                          * now.
2848                          */
2849                         KKASSERT(hammer_is_zone_large_data(disk_offset));
2850                         nbio->bio_offset = disk_offset;
2851                         if (hammer_live_dedup)
2852                                 hammer_dedup_cache_add(ip, cursor.leaf);
2853                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2854                         goto done;
2855                 } else if (n) {
2856                         error = hammer_ip_resolve_data(&cursor);
2857                         if (error == 0) {
2858                                 if (hammer_live_dedup && isdedupable)
2859                                         hammer_dedup_cache_add(ip, cursor.leaf);
2860                                 bcopy((char *)cursor.data + roff,
2861                                       (char *)bp->b_data + boff, n);
2862                         }
2863                 }
2864                 if (error)
2865                         break;
2866
2867                 /*
2868                  * We have to be sure that the only elements added to the
2869                  * dedup cache are those which are already on-media.
2870                  */
2871                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2872                         hammer_dedup_cache_add(ip, cursor.leaf);
2873
2874                 /*
2875                  * Iterate until we have filled the request.
2876                  */
2877                 boff += n;
2878                 if (boff == bp->b_bufsize)
2879                         break;
2880                 error = hammer_ip_next(&cursor);
2881         }
2882
2883         /*
2884          * There may have been a gap after the last record
2885          */
2886         if (error == ENOENT)
2887                 error = 0;
2888         if (error == 0 && boff != bp->b_bufsize) {
2889                 KKASSERT(boff < bp->b_bufsize);
2890                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2891                 /* boff = bp->b_bufsize; */
2892         }
2893
2894         /*
2895          * Disallow swapcache operation on the vnode buffer if double
2896          * buffering is enabled, the swapcache will get the data via
2897          * the block device buffer.
2898          */
2899         if (hammer_double_buffer)
2900                 bp->b_flags |= B_NOTMETA;
2901
2902         /*
2903          * Cleanup
2904          */
2905         bp->b_resid = 0;
2906         bp->b_error = error;
2907         if (error)
2908                 bp->b_flags |= B_ERROR;
2909         biodone(ap->a_bio);
2910
2911 done:
2912         /*
2913          * Cache the b-tree node for the last data read in cache[1].
2914          *
2915          * If we hit the file EOF then also cache the node in the
2916          * governing directory's cache[3], it will be used to initialize
2917          * the new inode's cache[1] for any inodes looked up via the directory.
2918          *
2919          * This doesn't reduce disk accesses since the B-Tree chain is
2920          * likely cached, but it does reduce cpu overhead when looking
2921          * up file offsets for cpdup/tar/cpio style iterations.
2922          */
2923         if (cursor.node)
2924                 hammer_cache_node(&ip->cache[1], cursor.node);
2925         if (ran_end >= ip->ino_data.size) {
2926                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2927                                         ip->obj_asof, ip->obj_localization);
2928                 if (dip) {
2929                         hammer_cache_node(&dip->cache[3], cursor.node);
2930                         hammer_rel_inode(dip, 0);
2931                 }
2932         }
2933         hammer_done_cursor(&cursor);
2934         hammer_done_transaction(&trans);
2935         lwkt_reltoken(&hmp->fs_token);
2936         return(error);
2937 }
2938
2939 /*
2940  * BMAP operation - used to support cluster_read() only.
2941  *
2942  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2943  *
2944  * This routine may return EOPNOTSUPP if the opration is not supported for
2945  * the specified offset.  The contents of the pointer arguments do not
2946  * need to be initialized in that case.
2947  *
2948  * If a disk address is available and properly aligned return 0 with
2949  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2950  * to the run-length relative to that offset.  Callers may assume that
2951  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2952  * large, so return EOPNOTSUPP if it is not sufficiently large.
2953  */
2954 static
2955 int
2956 hammer_vop_bmap(struct vop_bmap_args *ap)
2957 {
2958         struct hammer_transaction trans;
2959         hammer_inode_t ip;
2960         hammer_mount_t hmp;
2961         struct hammer_cursor cursor;
2962         hammer_base_elm_t base;
2963         int64_t rec_offset;
2964         int64_t ran_end;
2965         int64_t tmp64;
2966         int64_t base_offset;
2967         int64_t base_disk_offset;
2968         int64_t last_offset;
2969         hammer_off_t last_disk_offset;
2970         hammer_off_t disk_offset;
2971         int     rec_len;
2972         int     error;
2973         int     blksize;
2974
2975         ip = ap->a_vp->v_data;
2976         hmp = ip->hmp;
2977
2978         /*
2979          * We can only BMAP regular files.  We can't BMAP database files,
2980          * directories, etc.
2981          */
2982         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2983                 return(EOPNOTSUPP);
2984
2985         /*
2986          * bmap is typically called with runp/runb both NULL when used
2987          * for writing.  We do not support BMAP for writing atm.
2988          */
2989         if (ap->a_cmd != BUF_CMD_READ)
2990                 return(EOPNOTSUPP);
2991
2992         /*
2993          * Scan the B-Tree to acquire blockmap addresses, then translate
2994          * to raw addresses.
2995          */
2996         lwkt_gettoken(&hmp->fs_token);
2997         hammer_simple_transaction(&trans, hmp);
2998
2999         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3000
3001         /*
3002          * Key range (begin and end inclusive) to scan.  Note that the key's
3003          * stored in the actual records represent BASE+LEN, not BASE.  The
3004          * first record containing bio_offset will have a key > bio_offset.
3005          */
3006         cursor.key_beg.localization = ip->obj_localization |
3007                                       HAMMER_LOCALIZE_MISC;
3008         cursor.key_beg.obj_id = ip->obj_id;
3009         cursor.key_beg.create_tid = 0;
3010         cursor.key_beg.delete_tid = 0;
3011         cursor.key_beg.obj_type = 0;
3012         if (ap->a_runb)
3013                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3014         else
3015                 cursor.key_beg.key = ap->a_loffset + 1;
3016         if (cursor.key_beg.key < 0)
3017                 cursor.key_beg.key = 0;
3018         cursor.asof = ip->obj_asof;
3019         cursor.flags |= HAMMER_CURSOR_ASOF;
3020
3021         cursor.key_end = cursor.key_beg;
3022         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3023
3024         ran_end = ap->a_loffset + MAXPHYS;
3025         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3026         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3027         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3028         if (tmp64 < ran_end)
3029                 cursor.key_end.key = HAMMER_MAX_KEY;
3030         else
3031                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3032
3033         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3034
3035         error = hammer_ip_first(&cursor);
3036         base_offset = last_offset = 0;
3037         base_disk_offset = last_disk_offset = 0;
3038
3039         while (error == 0) {
3040                 /*
3041                  * Get the base file offset of the record.  The key for
3042                  * data records is (base + bytes) rather then (base).
3043                  *
3044                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3045                  * The extra bytes should be zero on-disk and the BMAP op
3046                  * should still be ok.
3047                  */
3048                 base = &cursor.leaf->base;
3049                 rec_offset = base->key - cursor.leaf->data_len;
3050                 rec_len    = cursor.leaf->data_len;
3051
3052                 /*
3053                  * Incorporate any cached truncation.
3054                  *
3055                  * NOTE: Modifications to rec_len based on synthesized
3056                  * truncation points remove the guarantee that any extended
3057                  * data on disk is zero (since the truncations may not have
3058                  * taken place on-media yet).
3059                  */
3060                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3061                         if (hammer_cursor_ondisk(&cursor) ||
3062                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3063                                 if (ip->trunc_off <= rec_offset)
3064                                         rec_len = 0;
3065                                 else if (ip->trunc_off < rec_offset + rec_len)
3066                                         rec_len = (int)(ip->trunc_off - rec_offset);
3067                         }
3068                 }
3069                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3070                         if (hammer_cursor_ondisk(&cursor)) {
3071                                 if (ip->sync_trunc_off <= rec_offset)
3072                                         rec_len = 0;
3073                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3074                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3075                         }
3076                 }
3077
3078                 /*
3079                  * Accumulate information.  If we have hit a discontiguous
3080                  * block reset base_offset unless we are already beyond the
3081                  * requested offset.  If we are, that's it, we stop.
3082                  */
3083                 if (error)
3084                         break;
3085                 if (hammer_cursor_ondisk(&cursor)) {
3086                         disk_offset = cursor.leaf->data_offset;
3087                         if (rec_offset != last_offset ||
3088                             disk_offset != last_disk_offset) {
3089                                 if (rec_offset > ap->a_loffset)
3090                                         break;
3091                                 base_offset = rec_offset;
3092                                 base_disk_offset = disk_offset;
3093                         }
3094                         last_offset = rec_offset + rec_len;
3095                         last_disk_offset = disk_offset + rec_len;
3096
3097                         if (hammer_live_dedup)
3098                                 hammer_dedup_cache_add(ip, cursor.leaf);
3099                 }
3100
3101                 error = hammer_ip_next(&cursor);
3102         }
3103
3104         if (cursor.node)
3105                 hammer_cache_node(&ip->cache[1], cursor.node);
3106
3107         hammer_done_cursor(&cursor);
3108         hammer_done_transaction(&trans);
3109         lwkt_reltoken(&hmp->fs_token);
3110
3111         /*
3112          * If we couldn't find any records or the records we did find were
3113          * all behind the requested offset, return failure.  A forward
3114          * truncation can leave a hole w/ no on-disk records.
3115          */
3116         if (last_offset == 0 || last_offset < ap->a_loffset)
3117                 return (EOPNOTSUPP);
3118
3119         /*
3120          * Figure out the block size at the requested offset and adjust
3121          * our limits so the cluster_read() does not create inappropriately
3122          * sized buffer cache buffers.
3123          */
3124         blksize = hammer_blocksize(ap->a_loffset);
3125         if (hammer_blocksize(base_offset) != blksize) {
3126                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3127         }
3128         if (last_offset != ap->a_loffset &&
3129             hammer_blocksize(last_offset - 1) != blksize) {
3130                 last_offset = hammer_blockdemarc(ap->a_loffset,
3131                                                  last_offset - 1);
3132         }
3133
3134         /*
3135          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3136          * from occuring.
3137          */
3138         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3139
3140         if (!hammer_is_zone_large_data(disk_offset)) {
3141                 /*
3142                  * Only large-data zones can be direct-IOd
3143                  */
3144                 error = EOPNOTSUPP;
3145         } else if ((disk_offset & HAMMER_BUFMASK) ||
3146                    (last_offset - ap->a_loffset) < blksize) {
3147                 /*
3148                  * doffsetp is not aligned or the forward run size does
3149                  * not cover a whole buffer, disallow the direct I/O.
3150                  */
3151                 error = EOPNOTSUPP;
3152         } else {
3153                 /*
3154                  * We're good.
3155                  */
3156                 *ap->a_doffsetp = disk_offset;
3157                 if (ap->a_runb) {
3158                         *ap->a_runb = ap->a_loffset - base_offset;
3159                         KKASSERT(*ap->a_runb >= 0);
3160                 }
3161                 if (ap->a_runp) {
3162                         *ap->a_runp = last_offset - ap->a_loffset;
3163                         KKASSERT(*ap->a_runp >= 0);
3164                 }
3165                 error = 0;
3166         }
3167         return(error);
3168 }
3169
3170 /*
3171  * Write to a regular file.   Because this is a strategy call the OS is
3172  * trying to actually get data onto the media.
3173  */
3174 static
3175 int
3176 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3177 {
3178         hammer_record_t record;
3179         hammer_mount_t hmp;
3180         hammer_inode_t ip;
3181         struct bio *bio;
3182         struct buf *bp;
3183         int blksize __debugvar;
3184         int bytes;
3185         int error;
3186
3187         bio = ap->a_bio;
3188         bp = bio->bio_buf;
3189         ip = ap->a_vp->v_data;
3190         hmp = ip->hmp;
3191
3192         blksize = hammer_blocksize(bio->bio_offset);
3193         KKASSERT(bp->b_bufsize == blksize);
3194
3195         if (ip->flags & HAMMER_INODE_RO) {
3196                 bp->b_error = EROFS;
3197                 bp->b_flags |= B_ERROR;
3198                 biodone(ap->a_bio);
3199                 return(EROFS);
3200         }
3201
3202         lwkt_gettoken(&hmp->fs_token);
3203
3204         /*
3205          * Disallow swapcache operation on the vnode buffer if double
3206          * buffering is enabled, the swapcache will get the data via
3207          * the block device buffer.
3208          */
3209         if (hammer_double_buffer)
3210                 bp->b_flags |= B_NOTMETA;
3211
3212         /*
3213          * Interlock with inode destruction (no in-kernel or directory
3214          * topology visibility).  If we queue new IO while trying to
3215          * destroy the inode we can deadlock the vtrunc call in
3216          * hammer_inode_unloadable_check().
3217          *
3218          * Besides, there's no point flushing a bp associated with an
3219          * inode that is being destroyed on-media and has no kernel
3220          * references.
3221          */
3222         if ((ip->flags | ip->sync_flags) &
3223             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3224                 bp->b_resid = 0;
3225                 biodone(ap->a_bio);
3226                 lwkt_reltoken(&hmp->fs_token);
3227                 return(0);
3228         }
3229
3230         /*
3231          * Reserve space and issue a direct-write from the front-end.
3232          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3233          * allocations.
3234          *
3235          * An in-memory record will be installed to reference the storage
3236          * until the flusher can get to it.
3237          *
3238          * Since we own the high level bio the front-end will not try to
3239          * do a direct-read until the write completes.
3240          *
3241          * NOTE: The only time we do not reserve a full-sized buffers
3242          * worth of data is if the file is small.  We do not try to
3243          * allocate a fragment (from the small-data zone) at the end of
3244          * an otherwise large file as this can lead to wildly separated
3245          * data.
3246          */
3247         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3248         KKASSERT(bio->bio_offset < ip->ino_data.size);
3249         if (bio->bio_offset || ip->ino_data.size > HAMMER_HBUFSIZE)
3250                 bytes = bp->b_bufsize;
3251         else
3252                 bytes = HAMMER_DATA_DOALIGN_WITH(int, ip->ino_data.size);
3253
3254         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3255                                     bytes, &error);
3256
3257         /*
3258          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3259          * in hammer_vop_write().  We must flag the record so the proper
3260          * REDO_TERM_WRITE entry is generated during the flush.
3261          */
3262         if (record) {
3263                 if (bp->b_flags & B_VFSFLAG1) {
3264                         record->flags |= HAMMER_RECF_REDO;
3265                         bp->b_flags &= ~B_VFSFLAG1;
3266                 }
3267                 if (record->flags & HAMMER_RECF_DEDUPED) {
3268                         bp->b_resid = 0;
3269                         hammer_ip_replace_bulk(hmp, record);
3270                         biodone(ap->a_bio);
3271                 } else {
3272                         hammer_io_direct_write(hmp, bio, record);
3273                 }
3274                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3275                         hammer_flush_inode(ip, 0);
3276         } else {
3277                 bp->b_bio2.bio_offset = NOOFFSET;
3278                 bp->b_error = error;
3279                 bp->b_flags |= B_ERROR;
3280                 biodone(ap->a_bio);
3281         }
3282         lwkt_reltoken(&hmp->fs_token);
3283         return(error);
3284 }
3285
3286 /*
3287  * dounlink - disconnect a directory entry
3288  *
3289  * XXX whiteout support not really in yet
3290  */
3291 static int
3292 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3293                 struct vnode *dvp, struct ucred *cred,
3294                 int flags, int isdir)
3295 {
3296         struct namecache *ncp;
3297         hammer_inode_t dip;
3298         hammer_inode_t ip;
3299         hammer_mount_t hmp;
3300         struct hammer_cursor cursor;
3301         int64_t namekey;
3302         uint32_t max_iterations;
3303         int nlen, error;
3304
3305         /*
3306          * Calculate the namekey and setup the key range for the scan.  This
3307          * works kinda like a chained hash table where the lower 32 bits
3308          * of the namekey synthesize the chain.
3309          *
3310          * The key range is inclusive of both key_beg and key_end.
3311          */
3312         dip = VTOI(dvp);
3313         ncp = nch->ncp;
3314         hmp = dip->hmp;
3315
3316         if (dip->flags & HAMMER_INODE_RO)
3317                 return (EROFS);
3318
3319         namekey = hammer_direntry_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3320                                            &max_iterations);
3321 retry:
3322         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3323         cursor.key_beg.localization = dip->obj_localization |
3324                                       hammer_dir_localization(dip);
3325         cursor.key_beg.obj_id = dip->obj_id;
3326         cursor.key_beg.key = namekey;
3327         cursor.key_beg.create_tid = 0;
3328         cursor.key_beg.delete_tid = 0;
3329         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3330         cursor.key_beg.obj_type = 0;
3331
3332         cursor.key_end = cursor.key_beg;
3333         cursor.key_end.key += max_iterations;
3334         cursor.asof = dip->obj_asof;
3335         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3336
3337         /*
3338          * Scan all matching records (the chain), locate the one matching
3339          * the requested path component.  info->last_error contains the
3340          * error code on search termination and could be 0, ENOENT, or
3341          * something else.
3342          *
3343          * The hammer_ip_*() functions merge in-memory records with on-disk
3344          * records for the purposes of the search.
3345          */
3346         error = hammer_ip_first(&cursor);
3347
3348         while (error == 0) {
3349                 error = hammer_ip_resolve_data(&cursor);
3350                 if (error)
3351                         break;
3352                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3353                 KKASSERT(nlen > 0);
3354                 if (ncp->nc_nlen == nlen &&
3355                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3356                         break;
3357                 }
3358                 error = hammer_ip_next(&cursor);
3359         }
3360
3361         /*
3362          * If all is ok we have to get the inode so we can adjust nlinks.
3363          * To avoid a deadlock with the flusher we must release the inode
3364          * lock on the directory when acquiring the inode for the entry.
3365          *
3366          * If the target is a directory, it must be empty.
3367          */
3368         if (error == 0) {
3369                 hammer_unlock(&cursor.ip->lock);
3370                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3371                                       hmp->asof,
3372                                       cursor.data->entry.localization,
3373                                       0, &error);
3374                 hammer_lock_sh(&cursor.ip->lock);
3375                 if (error == ENOENT) {
3376                         hkprintf("WARNING: Removing dirent w/missing inode "
3377                                 "\"%s\"\n"
3378                                 "\tobj_id = %016jx\n",
3379                                 ncp->nc_name,
3380                                 (intmax_t)cursor.data->entry.obj_id);
3381                         error = 0;
3382                 }
3383
3384                 /*
3385                  * If isdir >= 0 we validate that the entry is or is not a
3386                  * directory.  If isdir < 0 we don't care.
3387                  */
3388                 if (error == 0 && isdir >= 0 && ip) {
3389                         if (isdir &&
3390                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3391                                 error = ENOTDIR;
3392                         } else if (isdir == 0 &&
3393                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3394                                 error = EISDIR;
3395                         }
3396                 }
3397
3398                 /*
3399                  * If we are trying to remove a directory the directory must
3400                  * be empty.
3401                  *
3402                  * The check directory code can loop and deadlock/retry.  Our
3403                  * own cursor's node locks must be released to avoid a 3-way
3404                  * deadlock with the flusher if the check directory code
3405                  * blocks.
3406                  *
3407                  * If any changes whatsoever have been made to the cursor
3408                  * set EDEADLK and retry.
3409                  *
3410                  * WARNING: See warnings in hammer_unlock_cursor()
3411                  *          function.
3412                  */
3413                 if (error == 0 && ip && ip->ino_data.obj_type ==
3414                                         HAMMER_OBJTYPE_DIRECTORY) {
3415                         hammer_unlock_cursor(&cursor);
3416                         error = hammer_ip_check_directory_empty(trans, ip);
3417                         hammer_lock_cursor(&cursor);
3418                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3419                                 hkprintf("Warning: avoided deadlock "
3420                                         "on rmdir '%s'\n",
3421                                         ncp->nc_name);
3422                                 error = EDEADLK;
3423                         }
3424                 }
3425
3426                 /*
3427                  * Delete the directory entry.
3428                  *
3429                  * WARNING: hammer_ip_del_direntry() may have to terminate
3430                  * the cursor to avoid a deadlock.  It is ok to call
3431                  * hammer_done_cursor() twice.
3432                  */
3433                 if (error == 0) {
3434                         error = hammer_ip_del_direntry(trans, &cursor,
3435                                                         dip, ip);
3436                 }
3437                 hammer_done_cursor(&cursor);
3438                 if (error == 0) {
3439                         /*
3440                          * Tell the namecache that we are now unlinked.
3441                          */
3442                         cache_unlink(nch);
3443
3444                         /*
3445                          * NOTE: ip->vp, if non-NULL, cannot be directly
3446                          *       referenced without formally acquiring the
3447                          *       vp since the vp might have zero refs on it,
3448                          *       or in the middle of a reclaim, etc.
3449                          *
3450                          * NOTE: The cache_setunresolved() can rip the vp
3451                          *       out from under us since the vp may not have
3452                          *       any refs, in which case ip->vp will be NULL
3453                          *       from the outset.
3454                          */
3455                         while (ip && ip->vp) {
3456                                 struct vnode *vp;
3457
3458                                 error = hammer_get_vnode(ip, &vp);
3459                                 if (error == 0 && vp) {
3460                                         vn_unlock(vp);
3461                                         hammer_knote(ip->vp, NOTE_DELETE);
3462 #if 0
3463                                         /*
3464                                          * Don't do this, it can deadlock
3465                                          * on concurrent rm's of hardlinks.
3466                                          * Shouldn't be needed any more.
3467                                          */
3468                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3469 #endif
3470                                         vrele(vp);
3471                                         break;
3472                                 }
3473                                 hdkprintf("ip/vp race1 avoided\n");
3474                         }
3475                 }
3476                 if (ip)
3477                         hammer_rel_inode(ip, 0);
3478         } else {
3479                 hammer_done_cursor(&cursor);
3480         }
3481         if (error == EDEADLK)
3482                 goto retry;
3483
3484         return (error);
3485 }
3486
3487 /************************************************************************
3488  *                          FIFO AND SPECFS OPS                         *
3489  ************************************************************************
3490  *
3491  */
3492 static int
3493 hammer_vop_fifoclose (struct vop_close_args *ap)
3494 {
3495         /* XXX update itimes */
3496         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3497 }
3498
3499 static int
3500 hammer_vop_fiforead (struct vop_read_args *ap)
3501 {
3502         int error;
3503
3504         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3505         /* XXX update access time */
3506         return (error);
3507 }
3508
3509 static int
3510 hammer_vop_fifowrite (struct vop_write_args *ap)
3511 {
3512         int error;
3513
3514         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3515         /* XXX update access time */
3516         return (error);
3517 }
3518
3519 static
3520 int
3521 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3522 {
3523         int error;
3524
3525         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3526         if (error)
3527                 error = hammer_vop_kqfilter(ap);
3528         return(error);
3529 }
3530
3531 /************************************************************************
3532  *                          KQFILTER OPS                                *
3533  ************************************************************************
3534  *
3535  */
3536 static void filt_hammerdetach(struct knote *kn);
3537 static int filt_hammerread(struct knote *kn, long hint);
3538 static int filt_hammerwrite(struct knote *kn, long hint);
3539 static int filt_hammervnode(struct knote *kn, long hint);
3540
3541 static struct filterops hammerread_filtops =
3542         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3543           NULL, filt_hammerdetach, filt_hammerread };
3544 static struct filterops hammerwrite_filtops =
3545         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3546           NULL, filt_hammerdetach, filt_hammerwrite };
3547 static struct filterops hammervnode_filtops =
3548         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3549           NULL, filt_hammerdetach, filt_hammervnode };
3550
3551 static
3552 int
3553 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3554 {
3555         struct vnode *vp = ap->a_vp;
3556         struct knote *kn = ap->a_kn;
3557
3558         switch (kn->kn_filter) {
3559         case EVFILT_READ:
3560                 kn->kn_fop = &hammerread_filtops;
3561                 break;
3562         case EVFILT_WRITE:
3563                 kn->kn_fop = &hammerwrite_filtops;
3564                 break;
3565         case EVFILT_VNODE:
3566                 kn->kn_fop = &hammervnode_filtops;
3567                 break;
3568         default:
3569                 return (EOPNOTSUPP);
3570         }
3571
3572         kn->kn_hook = (caddr_t)vp;
3573
3574         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3575
3576         return(0);
3577 }
3578
3579 static void
3580 filt_hammerdetach(struct knote *kn)
3581 {
3582         struct vnode *vp = (void *)kn->kn_hook;
3583
3584         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3585 }
3586
3587 static int
3588 filt_hammerread(struct knote *kn, long hint)
3589 {
3590         struct vnode *vp = (void *)kn->kn_hook;
3591         hammer_inode_t ip = VTOI(vp);
3592         hammer_mount_t hmp = ip->hmp;
3593         off_t off;
3594
3595         if (hint == NOTE_REVOKE) {
3596                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3597                 return(1);
3598         }
3599         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3600         off = ip->ino_data.size - kn->kn_fp->f_offset;
3601         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3602         lwkt_reltoken(&hmp->fs_token);
3603         if (kn->kn_sfflags & NOTE_OLDAPI)
3604                 return(1);
3605         return (kn->kn_data != 0);
3606 }
3607
3608 static int
3609 filt_hammerwrite(struct knote *kn, long hint)
3610 {
3611         if (hint == NOTE_REVOKE)
3612                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3613         kn->kn_data = 0;
3614         return (1);
3615 }
3616
3617 static int
3618 filt_hammervnode(struct knote *kn, long hint)
3619 {
3620         if (kn->kn_sfflags & hint)
3621                 kn->kn_fflags |= hint;
3622         if (hint == NOTE_REVOKE) {
3623                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3624                 return (1);
3625         }
3626         return (kn->kn_fflags != 0);
3627 }
3628