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, uio->uio_resid,
 416                                              seqcount * MAXBSIZE, &bp);
 417                 } else {
 418                         error = breadnx(ap->a_vp, base_offset, blksize,
 419                                         NULL, NULL, 0, &bp);
 420                 }
 421                 if (error) {
 422                         brelse(bp);
 423                         break;
 424                 }
 425 skip:
 426                 if ((hammer_debug_io & 0x0001) && (bp->b_flags & B_IOISSUED)) {
 427                         hdkprintf("zone2_offset %016jx read file %016jx@%016jx\n",
 428                                 (intmax_t)bp->b_bio2.bio_offset,
 429                                 (intmax_t)ip->obj_id,
 430                                 (intmax_t)bp->b_loffset);
 431                 }
 432                 bp->b_flags &= ~B_IOISSUED;
 433                 if (blksize == HAMMER_XBUFSIZE)
 434                         bp->b_flags |= B_CLUSTEROK;
 435
 436                 n = blksize - offset;
 437                 if (n > uio->uio_resid)
 438                         n = uio->uio_resid;
 439                 if (n > ip->ino_data.size - uio->uio_offset)
 440                         n = (int)(ip->ino_data.size - uio->uio_offset);
 441
 442                 /*
 443                  * Set B_AGE, data has a lower priority than meta-data.
 444                  *
 445                  * Use a hold/unlock/drop sequence to run the uiomove
 446                  * with the buffer unlocked, avoiding deadlocks against
 447                  * read()s on mmap()'d spaces.
 448                  */
 449                 bp->b_flags |= B_AGE;
 450                 error = uiomovebp(bp, (char *)bp->b_data + offset, n, uio);
 451                 bqrelse(bp);
 452
 453                 if (error)
 454                         break;
 455                 hammer_stats_file_read += n;
 456         }
 457
 458 finished:
 459
 460         /*
 461          * Try to update the atime with just the inode lock for maximum
 462          * concurrency.  If we can't shortcut it we have to get the full
 463          * blown transaction.
 464          */
 465         if (got_trans == 0 && hammer_update_atime_quick(ip) < 0) {
 466                 hammer_start_transaction(&trans, ip->hmp);
 467                 got_trans = 1;
 468         }
 469
 470         if (got_trans) {
 471                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 472                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 473                         lwkt_gettoken(&hmp->fs_token);
 474                         ip->ino_data.atime = trans.time;
 475                         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
 476                         hammer_done_transaction(&trans);
 477                         lwkt_reltoken(&hmp->fs_token);
 478                 } else {
 479                         hammer_done_transaction(&trans);
 480                 }
 481         }
 482         return (error);
 483 }
 484
 485 /*
 486  * hammer_vop_write { vp, uio, ioflag, cred }
 487  */
 488 static
 489 int
 490 hammer_vop_write(struct vop_write_args *ap)
 491 {
 492         struct hammer_transaction trans;
 493         hammer_inode_t ip;
 494         hammer_mount_t hmp;
 495         thread_t td;
 496         struct uio *uio;
 497         int offset;
 498         off_t base_offset;
 499         int64_t cluster_eof;
 500         struct buf *bp;
 501         int kflags;
 502         int error;
 503         int n;
 504         int flags;
 505         int seqcount;
 506         int bigwrite;
 507
 508         if (ap->a_vp->v_type != VREG)
 509                 return (EINVAL);
 510         ip = VTOI(ap->a_vp);
 511         hmp = ip->hmp;
 512         error = 0;
 513         kflags = 0;
 514         seqcount = ap->a_ioflag >> 16;
 515
 516         if (ip->flags & HAMMER_INODE_RO)
 517                 return (EROFS);
 518
 519         /*
 520          * Create a transaction to cover the operations we perform.
 521          */
 522         hammer_start_transaction(&trans, hmp);
 523         uio = ap->a_uio;
 524
 525         /*
 526          * Check append mode
 527          */
 528         if (ap->a_ioflag & IO_APPEND)
 529                 uio->uio_offset = ip->ino_data.size;
 530
 531         /*
 532          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 533          *
 534          * NOTE: the base_off assignment is required to work around what
 535          * I consider to be a GCC-4 optimization bug.
 536          */
 537         if (uio->uio_offset < 0) {
 538                 hammer_done_transaction(&trans);
 539                 return (EFBIG);
 540         }
 541         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 542         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 543                 hammer_done_transaction(&trans);
 544                 return (EFBIG);
 545         }
 546
 547         if (uio->uio_resid > 0 && (td = uio->uio_td) != NULL && td->td_proc &&
 548             base_offset > td->td_proc->p_rlimit[RLIMIT_FSIZE].rlim_cur) {
 549                 hammer_done_transaction(&trans);
 550                 lwpsignal(td->td_proc, td->td_lwp, SIGXFSZ);
 551                 return (EFBIG);
 552         }
 553
 554         /*
 555          * If reading or writing a huge amount of data we have to break
 556          * atomicy and allow the operation to be interrupted by a signal
 557          * or it can DOS the machine.
 558          *
 559          * Preset redo_count so we stop generating REDOs earlier if the
 560          * limit is exceeded.
 561          *
 562          * redo_count is heuristical, SMP races are ok
 563          */
 564         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 565         if ((ip->flags & HAMMER_INODE_REDO) &&
 566             ip->redo_count < hammer_limit_redo) {
 567                 ip->redo_count += uio->uio_resid;
 568         }
 569
 570         /*
 571          * Access the data typically in HAMMER_BUFSIZE blocks via the
 572          * buffer cache, but HAMMER may use a variable block size based
 573          * on the offset.
 574          */
 575         while (uio->uio_resid > 0) {
 576                 int fixsize = 0;
 577                 int blksize;
 578                 int blkmask;
 579                 int trivial;
 580                 int endofblk;
 581                 off_t nsize;
 582
 583                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 584                         break;
 585                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 586                         break;
 587
 588                 blksize = hammer_blocksize(uio->uio_offset);
 589
 590                 /*
 591                  * Control the number of pending records associated with
 592                  * this inode.  If too many have accumulated start a
 593                  * flush.  Try to maintain a pipeline with the flusher.
 594                  *
 595                  * NOTE: It is possible for other sources to grow the
 596                  *       records but not necessarily issue another flush,
 597                  *       so use a timeout and ensure that a re-flush occurs.
 598                  */
 599                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 600                         lwkt_gettoken(&hmp->fs_token);
 601                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 602                         while (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 603                                 ip->flags |= HAMMER_INODE_RECSW;
 604                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 605                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 606                         }
 607                         lwkt_reltoken(&hmp->fs_token);
 608                 }
 609
 610                 /*
 611                  * Do not allow HAMMER to blow out the buffer cache.  Very
 612                  * large UIOs can lockout other processes due to bwillwrite()
 613                  * mechanics.
 614                  *
 615                  * The hammer inode is not locked during these operations.
 616                  * The vnode is locked which can interfere with the pageout
 617                  * daemon for non-UIO_NOCOPY writes but should not interfere
 618                  * with the buffer cache.  Even so, we cannot afford to
 619                  * allow the pageout daemon to build up too many dirty buffer
 620                  * cache buffers.
 621                  *
 622                  * Only call this if we aren't being recursively called from
 623                  * a virtual disk device (vn), else we may deadlock.
 624                  */
 625                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 626                         bwillwrite(blksize);
 627
 628                 /*
 629                  * Calculate the blocksize at the current offset and figure
 630                  * out how much we can actually write.
 631                  */
 632                 blkmask = blksize - 1;
 633                 offset = (int)uio->uio_offset & blkmask;
 634                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 635                 n = blksize - offset;
 636                 if (n > uio->uio_resid) {
 637                         n = uio->uio_resid;
 638                         endofblk = 0;
 639                 } else {
 640                         endofblk = 1;
 641                 }
 642                 nsize = uio->uio_offset + n;
 643                 if (nsize > ip->ino_data.size) {
 644                         if (uio->uio_offset > ip->ino_data.size)
 645                                 trivial = 0;
 646                         else
 647                                 trivial = 1;
 648                         nvextendbuf(ap->a_vp,
 649                                     ip->ino_data.size,
 650                                     nsize,
 651                                     hammer_blocksize(ip->ino_data.size),
 652                                     hammer_blocksize(nsize),
 653                                     hammer_blockoff(ip->ino_data.size),
 654                                     hammer_blockoff(nsize),
 655                                     trivial);
 656                         fixsize = 1;
 657                         kflags |= NOTE_EXTEND;
 658                 }
 659
 660                 if (uio->uio_segflg == UIO_NOCOPY) {
 661                         /*
 662                          * Issuing a write with the same data backing the
 663                          * buffer.  Instantiate the buffer to collect the
 664                          * backing vm pages, then read-in any missing bits.
 665                          *
 666                          * This case is used by vop_stdputpages().
 667                          */
 668                         bp = getblk(ap->a_vp, base_offset,
 669                                     blksize, GETBLK_BHEAVY, 0);
 670                         if ((bp->b_flags & B_CACHE) == 0) {
 671                                 bqrelse(bp);
 672                                 error = bread(ap->a_vp, base_offset,
 673                                               blksize, &bp);
 674                         }
 675                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 676                         /*
 677                          * Even though we are entirely overwriting the buffer
 678                          * we may still have to zero it out to avoid a
 679                          * mmap/write visibility issue.
 680                          */
 681                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 682                         if ((bp->b_flags & B_CACHE) == 0)
 683                                 vfs_bio_clrbuf(bp);
 684                 } else if (base_offset >= ip->ino_data.size) {
 685                         /*
 686                          * If the base offset of the buffer is beyond the
 687                          * file EOF, we don't have to issue a read.
 688                          */
 689                         bp = getblk(ap->a_vp, base_offset,
 690                                     blksize, GETBLK_BHEAVY, 0);
 691                         vfs_bio_clrbuf(bp);
 692                 } else {
 693                         /*
 694                          * Partial overwrite, read in any missing bits then
 695                          * replace the portion being written.
 696                          */
 697                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 698                         if (error == 0)
 699                                 bheavy(bp);
 700                 }
 701                 if (error == 0)
 702                         error = uiomovebp(bp, bp->b_data + offset, n, uio);
 703
 704                 lwkt_gettoken(&hmp->fs_token);
 705
 706                 /*
 707                  * Generate REDO records if enabled and redo_count will not
 708                  * exceeded the limit.
 709                  *
 710                  * If redo_count exceeds the limit we stop generating records
 711                  * and clear HAMMER_INODE_REDO.  This will cause the next
 712                  * fsync() to do a full meta-data sync instead of just an
 713                  * UNDO/REDO fifo update.
 714                  *
 715                  * When clearing HAMMER_INODE_REDO any pre-existing REDOs
 716                  * will still be tracked.  The tracks will be terminated
 717                  * when the related meta-data (including possible data
 718                  * modifications which are not tracked via REDO) is
 719                  * flushed.
 720                  */
 721                 if ((ip->flags & HAMMER_INODE_REDO) && error == 0) {
 722                         if (ip->redo_count < hammer_limit_redo) {
 723                                 bp->b_flags |= B_VFSFLAG1;
 724                                 error = hammer_generate_redo(&trans, ip,
 725                                                      base_offset + offset,
 726                                                      HAMMER_REDO_WRITE,
 727                                                      bp->b_data + offset,
 728                                                      (size_t)n);
 729                         } else {
 730                                 ip->flags &= ~HAMMER_INODE_REDO;
 731                         }
 732                 }
 733
 734                 /*
 735                  * If we screwed up we have to undo any VM size changes we
 736                  * made.
 737                  */
 738                 if (error) {
 739                         brelse(bp);
 740                         if (fixsize) {
 741                                 nvtruncbuf(ap->a_vp, ip->ino_data.size,
 742                                           hammer_blocksize(ip->ino_data.size),
 743                                           hammer_blockoff(ip->ino_data.size),
 744                                           0);
 745                         }
 746                         lwkt_reltoken(&hmp->fs_token);
 747                         break;
 748                 }
 749                 kflags |= NOTE_WRITE;
 750                 hammer_stats_file_write += n;
 751                 if (blksize == HAMMER_XBUFSIZE)
 752                         bp->b_flags |= B_CLUSTEROK;
 753                 if (ip->ino_data.size < uio->uio_offset) {
 754                         ip->ino_data.size = uio->uio_offset;
 755                         flags = HAMMER_INODE_SDIRTY;
 756                 } else {
 757                         flags = 0;
 758                 }
 759                 ip->ino_data.mtime = trans.time;
 760                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 761                 hammer_modify_inode(&trans, ip, flags);
 762
 763                 /*
 764                  * Once we dirty the buffer any cached zone-X offset
 765                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 766                  * allow overwriting over the same data sector unless
 767                  * we provide UNDOs for the old data, which we don't.
 768                  */
 769                 bp->b_bio2.bio_offset = NOOFFSET;
 770
 771                 lwkt_reltoken(&hmp->fs_token);
 772
 773                 /*
 774                  * Final buffer disposition.
 775                  *
 776                  * Because meta-data updates are deferred, HAMMER is
 777                  * especially sensitive to excessive bdwrite()s because
 778                  * the I/O stream is not broken up by disk reads.  So the
 779                  * buffer cache simply cannot keep up.
 780                  *
 781                  * WARNING!  blksize is variable.  cluster_write() is
 782                  *           expected to not blow up if it encounters
 783                  *           buffers that do not match the passed blksize.
 784                  *
 785                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 786                  *        The ip->rsv_recs check should burst-flush the data.
 787                  *        If we queue it immediately the buf could be left
 788                  *        locked on the device queue for a very long time.
 789                  *
 790                  *        However, failing to flush a dirty buffer out when
 791                  *        issued from the pageout daemon can result in a low
 792                  *        memory deadlock against bio_page_alloc(), so we
 793                  *        have to bawrite() on IO_ASYNC as well.
 794                  *
 795                  * NOTE!  To avoid degenerate stalls due to mismatched block
 796                  *        sizes we only honor IO_DIRECT on the write which
 797                  *        abuts the end of the buffer.  However, we must
 798                  *        honor IO_SYNC in case someone is silly enough to
 799                  *        configure a HAMMER file as swap, or when HAMMER
 800                  *        is serving NFS (for commits).  Ick ick.
 801                  */
 802                 bp->b_flags |= B_AGE;
 803                 if (blksize == HAMMER_XBUFSIZE)
 804                         bp->b_flags |= B_CLUSTEROK;
 805
 806                 if (ap->a_ioflag & IO_SYNC) {
 807                         bwrite(bp);
 808                 } else if ((ap->a_ioflag & IO_DIRECT) && endofblk) {
 809                         bawrite(bp);
 810                 } else if (ap->a_ioflag & IO_ASYNC) {
 811                         bawrite(bp);
 812                 } else if (hammer_cluster_enable &&
 813                            !(ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 814                         if (base_offset < HAMMER_XDEMARC)
 815                                 cluster_eof = hammer_blockdemarc(base_offset,
 816                                                          ip->ino_data.size);
 817                         else
 818                                 cluster_eof = ip->ino_data.size;
 819                         cluster_write(bp, cluster_eof, blksize, seqcount);
 820                 } else {
 821                         bdwrite(bp);
 822                 }
 823         }
 824         hammer_done_transaction(&trans);
 825         hammer_knote(ap->a_vp, kflags);
 826
 827         return (error);
 828 }
 829
 830 /*
 831  * hammer_vop_access { vp, mode, cred }
 832  *
 833  * MPSAFE - does not require fs_token
 834  */
 835 static
 836 int
 837 hammer_vop_access(struct vop_access_args *ap)
 838 {
 839         hammer_inode_t ip = VTOI(ap->a_vp);
 840         uid_t uid;
 841         gid_t gid;
 842         int error;
 843
 844         ++hammer_stats_file_iopsr;
 845         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 846         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 847
 848         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 849                                   ip->ino_data.uflags);
 850         return (error);
 851 }
 852
 853 /*
 854  * hammer_vop_advlock { vp, id, op, fl, flags }
 855  *
 856  * MPSAFE - does not require fs_token
 857  */
 858 static
 859 int
 860 hammer_vop_advlock(struct vop_advlock_args *ap)
 861 {
 862         hammer_inode_t ip = VTOI(ap->a_vp);
 863
 864         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 865 }
 866
 867 /*
 868  * hammer_vop_close { vp, fflag }
 869  *
 870  * We can only sync-on-close for normal closes.  XXX disabled for now.
 871  */
 872 static
 873 int
 874 hammer_vop_close(struct vop_close_args *ap)
 875 {
 876 #if 0
 877         struct vnode *vp = ap->a_vp;
 878         hammer_inode_t ip = VTOI(vp);
 879         int waitfor;
 880         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 881                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 882                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 883                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 884                                 waitfor = MNT_WAIT;
 885                         else
 886                                 waitfor = MNT_NOWAIT;
 887                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 888                                        HAMMER_INODE_CLOSEASYNC);
 889                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 890                 }
 891         }
 892 #endif
 893         return (vop_stdclose(ap));
 894 }
 895
 896 /*
 897  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 898  *
 899  * The operating system has already ensured that the directory entry
 900  * does not exist and done all appropriate namespace locking.
 901  */
 902 static
 903 int
 904 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 905 {
 906         struct hammer_transaction trans;
 907         hammer_inode_t dip;
 908         hammer_inode_t nip;
 909         struct nchandle *nch;
 910         hammer_mount_t hmp;
 911         int error;
 912
 913         nch = ap->a_nch;
 914         dip = VTOI(ap->a_dvp);
 915         hmp = dip->hmp;
 916
 917         if (dip->flags & HAMMER_INODE_RO)
 918                 return (EROFS);
 919         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
 920                 return (error);
 921
 922         /*
 923          * Create a transaction to cover the operations we perform.
 924          */
 925         lwkt_gettoken(&hmp->fs_token);
 926         hammer_start_transaction(&trans, hmp);
 927         ++hammer_stats_file_iopsw;
 928
 929         /*
 930          * Create a new filesystem object of the requested type.  The
 931          * returned inode will be referenced and shared-locked to prevent
 932          * it from being moved to the flusher.
 933          */
 934         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 935                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 936                                     NULL, &nip);
 937         if (error) {
 938                 hkprintf("hammer_create_inode error %d\n", error);
 939                 hammer_done_transaction(&trans);
 940                 *ap->a_vpp = NULL;
 941                 lwkt_reltoken(&hmp->fs_token);
 942                 return (error);
 943         }
 944
 945         /*
 946          * Add the new filesystem object to the directory.  This will also
 947          * bump the inode's link count.
 948          */
 949         error = hammer_ip_add_direntry(&trans, dip,
 950                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 951                                         nip);
 952         if (error)
 953                 hkprintf("hammer_ip_add_direntry error %d\n", error);
 954
 955         /*
 956          * Finish up.
 957          */
 958         if (error) {
 959                 hammer_rel_inode(nip, 0);
 960                 hammer_done_transaction(&trans);
 961                 *ap->a_vpp = NULL;
 962         } else {
 963                 error = hammer_get_vnode(nip, ap->a_vpp);
 964                 hammer_done_transaction(&trans);
 965                 hammer_rel_inode(nip, 0);
 966                 if (error == 0) {
 967                         cache_setunresolved(ap->a_nch);
 968                         cache_setvp(ap->a_nch, *ap->a_vpp);
 969                 }
 970                 hammer_knote(ap->a_dvp, NOTE_WRITE);
 971         }
 972         lwkt_reltoken(&hmp->fs_token);
 973         return (error);
 974 }
 975
 976 /*
 977  * hammer_vop_getattr { vp, vap }
 978  *
 979  * Retrieve an inode's attribute information.  When accessing inodes
 980  * historically we fake the atime field to ensure consistent results.
 981  * The atime field is stored in the B-Tree element and allowed to be
 982  * updated without cycling the element.
 983  *
 984  * MPSAFE - does not require fs_token
 985  */
 986 static
 987 int
 988 hammer_vop_getattr(struct vop_getattr_args *ap)
 989 {
 990         hammer_inode_t ip = VTOI(ap->a_vp);
 991         struct vattr *vap = ap->a_vap;
 992
 993         /*
 994          * We want the fsid to be different when accessing a filesystem
 995          * with different as-of's so programs like diff don't think
 996          * the files are the same.
 997          *
 998          * We also want the fsid to be the same when comparing snapshots,
 999          * or when comparing mirrors (which might be backed by different
1000          * physical devices).  HAMMER fsids are based on the PFS's
1001          * shared_uuid field.
1002          *
1003          * XXX there is a chance of collision here.  The va_fsid reported
1004          * by stat is different from the more involved fsid used in the
1005          * mount structure.
1006          */
1007         ++hammer_stats_file_iopsr;
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         ++hammer_stats_file_iopsr;
1130
1131         for (i = 0; i < nlen; ++i) {
1132                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1133                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1134                                                   &ispfs, &asof, &localization);
1135                         if (error != 0) {
1136                                 i = nlen;
1137                                 break;
1138                         }
1139                         if (asof != HAMMER_MAX_TID)
1140                                 flags |= HAMMER_INODE_RO;
1141                         break;
1142                 }
1143         }
1144         nlen = i;
1145
1146         /*
1147          * If this is a PFS softlink we dive into the PFS
1148          */
1149         if (ispfs && nlen == 0) {
1150                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1151                                       asof, localization,
1152                                       flags, &error);
1153                 if (error == 0) {
1154                         error = hammer_get_vnode(ip, &vp);
1155                         hammer_rel_inode(ip, 0);
1156                 } else {
1157                         vp = NULL;
1158                 }
1159                 if (error == 0) {
1160                         vn_unlock(vp);
1161                         cache_setvp(ap->a_nch, vp);
1162                         vrele(vp);
1163                 }
1164                 goto done;
1165         }
1166
1167         /*
1168          * If there is no path component the time extension is relative to dip.
1169          * e.g. "fubar/@@<snapshot>"
1170          *
1171          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1172          * e.g. "fubar/.@@<snapshot>"
1173          *
1174          * ".." is handled by the kernel.  We do not currently handle
1175          * "..@<snapshot>".
1176          */
1177         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1178                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1179                                       asof, dip->obj_localization,
1180                                       flags, &error);
1181                 if (error == 0) {
1182                         error = hammer_get_vnode(ip, &vp);
1183                         hammer_rel_inode(ip, 0);
1184                 } else {
1185                         vp = NULL;
1186                 }
1187                 if (error == 0) {
1188                         vn_unlock(vp);
1189                         cache_setvp(ap->a_nch, vp);
1190                         vrele(vp);
1191                 }
1192                 goto done;
1193         }
1194
1195         /*
1196          * Calculate the namekey and setup the key range for the scan.  This
1197          * works kinda like a chained hash table where the lower 32 bits
1198          * of the namekey synthesize the chain.
1199          *
1200          * The key range is inclusive of both key_beg and key_end.
1201          */
1202         namekey = hammer_direntry_namekey(dip, ncp->nc_name, nlen,
1203                                            &max_iterations);
1204
1205         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1206         cursor.key_beg.localization = dip->obj_localization |
1207                                       hammer_dir_localization(dip);
1208         cursor.key_beg.obj_id = dip->obj_id;
1209         cursor.key_beg.key = namekey;
1210         cursor.key_beg.create_tid = 0;
1211         cursor.key_beg.delete_tid = 0;
1212         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1213         cursor.key_beg.obj_type = 0;
1214
1215         cursor.key_end = cursor.key_beg;
1216         cursor.key_end.key += max_iterations;
1217         cursor.asof = asof;
1218         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1219
1220         /*
1221          * Scan all matching records (the chain), locate the one matching
1222          * the requested path component.
1223          *
1224          * The hammer_ip_*() functions merge in-memory records with on-disk
1225          * records for the purposes of the search.
1226          */
1227         obj_id = 0;
1228         localization = HAMMER_DEF_LOCALIZATION;
1229
1230         if (error == 0) {
1231                 error = hammer_ip_first(&cursor);
1232                 while (error == 0) {
1233                         error = hammer_ip_resolve_data(&cursor);
1234                         if (error)
1235                                 break;
1236                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1237                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1238                                 obj_id = cursor.data->entry.obj_id;
1239                                 localization = cursor.data->entry.localization;
1240                                 break;
1241                         }
1242                         error = hammer_ip_next(&cursor);
1243                 }
1244         }
1245         hammer_done_cursor(&cursor);
1246
1247         /*
1248          * Lookup the obj_id.  This should always succeed.  If it does not
1249          * the filesystem may be damaged and we return a dummy inode.
1250          */
1251         if (error == 0) {
1252                 ip = hammer_get_inode(&trans, dip, obj_id,
1253                                       asof, localization,
1254                                       flags, &error);
1255                 if (error == ENOENT) {
1256                         hkprintf("WARNING: Missing inode for dirent \"%s\"\n"
1257                                 "\tobj_id = %016jx, asof=%016jx, lo=%08x\n",
1258                                 ncp->nc_name,
1259                                 (intmax_t)obj_id, (intmax_t)asof,
1260                                 localization);
1261                         error = 0;
1262                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1263                                                     asof, localization,
1264                                                     flags, &error);
1265                 }
1266                 if (error == 0) {
1267                         error = hammer_get_vnode(ip, &vp);
1268                         hammer_rel_inode(ip, 0);
1269                 } else {
1270                         vp = NULL;
1271                 }
1272                 if (error == 0) {
1273                         vn_unlock(vp);
1274                         cache_setvp(ap->a_nch, vp);
1275                         vrele(vp);
1276                 }
1277         } else if (error == ENOENT) {
1278                 cache_setvp(ap->a_nch, NULL);
1279         }
1280 done:
1281         hammer_done_transaction(&trans);
1282         lwkt_reltoken(&hmp->fs_token);
1283         return (error);
1284 }
1285
1286 /*
1287  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1288  *
1289  * Locate the parent directory of a directory vnode.
1290  *
1291  * dvp is referenced but not locked.  *vpp must be returned referenced and
1292  * locked.  A parent_obj_id of 0 indicates that we are at the root.
1293  *
1294  * NOTE: as-of sequences are not linked into the directory structure.  If
1295  * we are at the root with a different asof then the mount point, reload
1296  * the same directory with the mount point's asof.   I'm not sure what this
1297  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1298  * get confused, but it hasn't been tested.
1299  */
1300 static
1301 int
1302 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1303 {
1304         struct hammer_transaction trans;
1305         hammer_inode_t dip;
1306         hammer_inode_t ip;
1307         hammer_mount_t hmp;
1308         int64_t parent_obj_id;
1309         uint32_t parent_obj_localization;
1310         hammer_tid_t asof;
1311         int error;
1312
1313         dip = VTOI(ap->a_dvp);
1314         asof = dip->obj_asof;
1315         hmp = dip->hmp;
1316
1317         /*
1318          * Whos are parent?  This could be the root of a pseudo-filesystem
1319          * whos parent is in another localization domain.
1320          */
1321         lwkt_gettoken(&hmp->fs_token);
1322         parent_obj_id = dip->ino_data.parent_obj_id;
1323         if (dip->obj_id == HAMMER_OBJID_ROOT)
1324                 parent_obj_localization = HAMMER_DEF_LOCALIZATION;
1325         else
1326                 parent_obj_localization = dip->obj_localization;
1327
1328         /*
1329          * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1330          */
1331         if (parent_obj_id == 0) {
1332                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1333                    asof != hmp->asof) {
1334                         parent_obj_id = dip->obj_id;
1335                         asof = hmp->asof;
1336                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1337                         ksnprintf(*ap->a_fakename, 19, "0x%016jx",
1338                                   (intmax_t)dip->obj_asof);
1339                 } else {
1340                         *ap->a_vpp = NULL;
1341                         lwkt_reltoken(&hmp->fs_token);
1342                         return ENOENT;
1343                 }
1344         }
1345
1346         hammer_simple_transaction(&trans, hmp);
1347         ++hammer_stats_file_iopsr;
1348
1349         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1350                               asof, parent_obj_localization,
1351                               dip->flags, &error);
1352         if (ip) {
1353                 error = hammer_get_vnode(ip, ap->a_vpp);
1354                 hammer_rel_inode(ip, 0);
1355         } else {
1356                 *ap->a_vpp = NULL;
1357         }
1358         hammer_done_transaction(&trans);
1359         lwkt_reltoken(&hmp->fs_token);
1360         return (error);
1361 }
1362
1363 /*
1364  * hammer_vop_nlink { nch, dvp, vp, cred }
1365  */
1366 static
1367 int
1368 hammer_vop_nlink(struct vop_nlink_args *ap)
1369 {
1370         struct hammer_transaction trans;
1371         hammer_inode_t dip;
1372         hammer_inode_t ip;
1373         struct nchandle *nch;
1374         hammer_mount_t hmp;
1375         int error;
1376
1377         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1378                 return(EXDEV);
1379
1380         nch = ap->a_nch;
1381         dip = VTOI(ap->a_dvp);
1382         ip = VTOI(ap->a_vp);
1383         hmp = dip->hmp;
1384
1385         if (dip->obj_localization != ip->obj_localization)
1386                 return(EXDEV);
1387
1388         if (dip->flags & HAMMER_INODE_RO)
1389                 return (EROFS);
1390         if (ip->flags & HAMMER_INODE_RO)
1391                 return (EROFS);
1392         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1393                 return (error);
1394
1395         /*
1396          * Create a transaction to cover the operations we perform.
1397          */
1398         lwkt_gettoken(&hmp->fs_token);
1399         hammer_start_transaction(&trans, hmp);
1400         ++hammer_stats_file_iopsw;
1401
1402         /*
1403          * Add the filesystem object to the directory.  Note that neither
1404          * dip nor ip are referenced or locked, but their vnodes are
1405          * referenced.  This function will bump the inode's link count.
1406          */
1407         error = hammer_ip_add_direntry(&trans, dip,
1408                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1409                                         ip);
1410
1411         /*
1412          * Finish up.
1413          */
1414         if (error == 0) {
1415                 cache_setunresolved(nch);
1416                 cache_setvp(nch, ap->a_vp);
1417         }
1418         hammer_done_transaction(&trans);
1419         hammer_knote(ap->a_vp, NOTE_LINK);
1420         hammer_knote(ap->a_dvp, NOTE_WRITE);
1421         lwkt_reltoken(&hmp->fs_token);
1422         return (error);
1423 }
1424
1425 /*
1426  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1427  *
1428  * The operating system has already ensured that the directory entry
1429  * does not exist and done all appropriate namespace locking.
1430  */
1431 static
1432 int
1433 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1434 {
1435         struct hammer_transaction trans;
1436         hammer_inode_t dip;
1437         hammer_inode_t nip;
1438         struct nchandle *nch;
1439         hammer_mount_t hmp;
1440         int error;
1441
1442         nch = ap->a_nch;
1443         dip = VTOI(ap->a_dvp);
1444         hmp = dip->hmp;
1445
1446         if (dip->flags & HAMMER_INODE_RO)
1447                 return (EROFS);
1448         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1449                 return (error);
1450
1451         /*
1452          * Create a transaction to cover the operations we perform.
1453          */
1454         lwkt_gettoken(&hmp->fs_token);
1455         hammer_start_transaction(&trans, hmp);
1456         ++hammer_stats_file_iopsw;
1457
1458         /*
1459          * Create a new filesystem object of the requested type.  The
1460          * returned inode will be referenced but not locked.
1461          */
1462         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1463                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1464                                     NULL, &nip);
1465         if (error) {
1466                 hammer_done_transaction(&trans);
1467                 *ap->a_vpp = NULL;
1468                 lwkt_reltoken(&hmp->fs_token);
1469                 return (error);
1470         }
1471         /*
1472          * Add the new filesystem object to the directory.  This will also
1473          * bump the inode's link count.
1474          */
1475         error = hammer_ip_add_direntry(&trans, dip,
1476                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1477                                         nip);
1478         if (error)
1479                 hkprintf("hammer_mkdir (add) error %d\n", error);
1480
1481         /*
1482          * Finish up.
1483          */
1484         if (error) {
1485                 hammer_rel_inode(nip, 0);
1486                 *ap->a_vpp = NULL;
1487         } else {
1488                 error = hammer_get_vnode(nip, ap->a_vpp);
1489                 hammer_rel_inode(nip, 0);
1490                 if (error == 0) {
1491                         cache_setunresolved(ap->a_nch);
1492                         cache_setvp(ap->a_nch, *ap->a_vpp);
1493                 }
1494         }
1495         hammer_done_transaction(&trans);
1496         if (error == 0)
1497                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1498         lwkt_reltoken(&hmp->fs_token);
1499         return (error);
1500 }
1501
1502 /*
1503  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1504  *
1505  * The operating system has already ensured that the directory entry
1506  * does not exist and done all appropriate namespace locking.
1507  */
1508 static
1509 int
1510 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1511 {
1512         struct hammer_transaction trans;
1513         hammer_inode_t dip;
1514         hammer_inode_t nip;
1515         struct nchandle *nch;
1516         hammer_mount_t hmp;
1517         int error;
1518
1519         nch = ap->a_nch;
1520         dip = VTOI(ap->a_dvp);
1521         hmp = dip->hmp;
1522
1523         if (dip->flags & HAMMER_INODE_RO)
1524                 return (EROFS);
1525         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1526                 return (error);
1527
1528         /*
1529          * Create a transaction to cover the operations we perform.
1530          */
1531         lwkt_gettoken(&hmp->fs_token);
1532         hammer_start_transaction(&trans, hmp);
1533         ++hammer_stats_file_iopsw;
1534
1535         /*
1536          * Create a new filesystem object of the requested type.  The
1537          * returned inode will be referenced but not locked.
1538          *
1539          * If mknod specifies a directory a pseudo-fs is created.
1540          */
1541         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1542                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1543                                     NULL, &nip);
1544         if (error) {
1545                 hammer_done_transaction(&trans);
1546                 *ap->a_vpp = NULL;
1547                 lwkt_reltoken(&hmp->fs_token);
1548                 return (error);
1549         }
1550
1551         /*
1552          * Add the new filesystem object to the directory.  This will also
1553          * bump the inode's link count.
1554          */
1555         error = hammer_ip_add_direntry(&trans, dip,
1556                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1557                                         nip);
1558
1559         /*
1560          * Finish up.
1561          */
1562         if (error) {
1563                 hammer_rel_inode(nip, 0);
1564                 *ap->a_vpp = NULL;
1565         } else {
1566                 error = hammer_get_vnode(nip, ap->a_vpp);
1567                 hammer_rel_inode(nip, 0);
1568                 if (error == 0) {
1569                         cache_setunresolved(ap->a_nch);
1570                         cache_setvp(ap->a_nch, *ap->a_vpp);
1571                 }
1572         }
1573         hammer_done_transaction(&trans);
1574         if (error == 0)
1575                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1576         lwkt_reltoken(&hmp->fs_token);
1577         return (error);
1578 }
1579
1580 /*
1581  * hammer_vop_open { vp, mode, cred, fp }
1582  *
1583  * MPSAFE (does not require fs_token)
1584  */
1585 static
1586 int
1587 hammer_vop_open(struct vop_open_args *ap)
1588 {
1589         hammer_inode_t ip;
1590
1591         ++hammer_stats_file_iopsr;
1592         ip = VTOI(ap->a_vp);
1593
1594         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1595                 return (EROFS);
1596         return(vop_stdopen(ap));
1597 }
1598
1599 /*
1600  * hammer_vop_print { vp }
1601  */
1602 static
1603 int
1604 hammer_vop_print(struct vop_print_args *ap)
1605 {
1606         return EOPNOTSUPP;
1607 }
1608
1609 /*
1610  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1611  */
1612 static
1613 int
1614 hammer_vop_readdir(struct vop_readdir_args *ap)
1615 {
1616         struct hammer_transaction trans;
1617         struct hammer_cursor cursor;
1618         hammer_inode_t ip;
1619         hammer_mount_t hmp;
1620         struct uio *uio;
1621         hammer_base_elm_t base;
1622         int error;
1623         int cookie_index;
1624         int ncookies;
1625         off_t *cookies;
1626         off_t saveoff;
1627         int r;
1628         int dtype;
1629
1630         ++hammer_stats_file_iopsr;
1631         ip = VTOI(ap->a_vp);
1632         uio = ap->a_uio;
1633         saveoff = uio->uio_offset;
1634         hmp = ip->hmp;
1635
1636         if (ap->a_ncookies) {
1637                 ncookies = uio->uio_resid / 16 + 1;
1638                 if (ncookies > 1024)
1639                         ncookies = 1024;
1640                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1641                 cookie_index = 0;
1642         } else {
1643                 ncookies = -1;
1644                 cookies = NULL;
1645                 cookie_index = 0;
1646         }
1647
1648         lwkt_gettoken(&hmp->fs_token);
1649         hammer_simple_transaction(&trans, hmp);
1650
1651         /*
1652          * Handle artificial entries
1653          *
1654          * It should be noted that the minimum value for a directory
1655          * hash key on-media is 0x0000000100000000, so we can use anything
1656          * less then that to represent our 'special' key space.
1657          */
1658         error = 0;
1659         if (saveoff == 0) {
1660                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1661                 if (r)
1662                         goto done;
1663                 if (cookies)
1664                         cookies[cookie_index] = saveoff;
1665                 ++saveoff;
1666                 ++cookie_index;
1667                 if (cookie_index == ncookies)
1668                         goto done;
1669         }
1670         if (saveoff == 1) {
1671                 if (ip->ino_data.parent_obj_id) {
1672                         r = vop_write_dirent(&error, uio,
1673                                              ip->ino_data.parent_obj_id,
1674                                              DT_DIR, 2, "..");
1675                 } else {
1676                         r = vop_write_dirent(&error, uio,
1677                                              ip->obj_id, DT_DIR, 2, "..");
1678                 }
1679                 if (r)
1680                         goto done;
1681                 if (cookies)
1682                         cookies[cookie_index] = saveoff;
1683                 ++saveoff;
1684                 ++cookie_index;
1685                 if (cookie_index == ncookies)
1686                         goto done;
1687         }
1688
1689         /*
1690          * Key range (begin and end inclusive) to scan.  Directory keys
1691          * directly translate to a 64 bit 'seek' position.
1692          */
1693         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1694         cursor.key_beg.localization = ip->obj_localization |
1695                                       hammer_dir_localization(ip);
1696         cursor.key_beg.obj_id = ip->obj_id;
1697         cursor.key_beg.create_tid = 0;
1698         cursor.key_beg.delete_tid = 0;
1699         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1700         cursor.key_beg.obj_type = 0;
1701         cursor.key_beg.key = saveoff;
1702
1703         cursor.key_end = cursor.key_beg;
1704         cursor.key_end.key = HAMMER_MAX_KEY;
1705         cursor.asof = ip->obj_asof;
1706         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1707
1708         error = hammer_ip_first(&cursor);
1709
1710         while (error == 0) {
1711                 error = hammer_ip_resolve_data(&cursor);
1712                 if (error)
1713                         break;
1714                 base = &cursor.leaf->base;
1715                 saveoff = base->key;
1716                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1717
1718                 if (base->obj_id != ip->obj_id)
1719                         hpanic("bad record at %p", cursor.node);
1720
1721                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1722                 r = vop_write_dirent(
1723                              &error, uio, cursor.data->entry.obj_id,
1724                              dtype,
1725                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1726                              (void *)cursor.data->entry.name);
1727                 if (r)
1728                         break;
1729                 ++saveoff;
1730                 if (cookies)
1731                         cookies[cookie_index] = base->key;
1732                 ++cookie_index;
1733                 if (cookie_index == ncookies)
1734                         break;
1735                 error = hammer_ip_next(&cursor);
1736         }
1737         hammer_done_cursor(&cursor);
1738
1739 done:
1740         hammer_done_transaction(&trans);
1741
1742         if (ap->a_eofflag)
1743                 *ap->a_eofflag = (error == ENOENT);
1744         uio->uio_offset = saveoff;
1745         if (error && cookie_index == 0) {
1746                 if (error == ENOENT)
1747                         error = 0;
1748                 if (cookies) {
1749                         kfree(cookies, M_TEMP);
1750                         *ap->a_ncookies = 0;
1751                         *ap->a_cookies = NULL;
1752                 }
1753         } else {
1754                 if (error == ENOENT)
1755                         error = 0;
1756                 if (cookies) {
1757                         *ap->a_ncookies = cookie_index;
1758                         *ap->a_cookies = cookies;
1759                 }
1760         }
1761         lwkt_reltoken(&hmp->fs_token);
1762         return(error);
1763 }
1764
1765 /*
1766  * hammer_vop_readlink { vp, uio, cred }
1767  */
1768 static
1769 int
1770 hammer_vop_readlink(struct vop_readlink_args *ap)
1771 {
1772         struct hammer_transaction trans;
1773         struct hammer_cursor cursor;
1774         hammer_inode_t ip;
1775         hammer_mount_t hmp;
1776         char buf[32];
1777         uint32_t localization;
1778         hammer_pseudofs_inmem_t pfsm;
1779         int error;
1780
1781         ip = VTOI(ap->a_vp);
1782         hmp = ip->hmp;
1783
1784         lwkt_gettoken(&hmp->fs_token);
1785
1786         /*
1787          * Shortcut if the symlink data was stuffed into ino_data.
1788          *
1789          * Also expand special "@@PFS%05d" softlinks (expansion only
1790          * occurs for non-historical (current) accesses made from the
1791          * primary filesystem).
1792          *
1793          * Note that userspace hammer command does not allow users to
1794          * create a @@PFS softlink under an existing other PFS (id!=0)
1795          * so the ip localization here for @@PFS softlink is always 0.
1796          */
1797         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1798                 char *ptr;
1799                 int bytes;
1800
1801                 ptr = ip->ino_data.ext.symlink;
1802                 bytes = (int)ip->ino_data.size;
1803                 if (bytes == 10 &&
1804                     ip->obj_asof == HAMMER_MAX_TID &&
1805                     ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1806                     strncmp(ptr, "@@PFS", 5) == 0) {
1807                         hammer_simple_transaction(&trans, hmp);
1808                         bcopy(ptr + 5, buf, 5);
1809                         buf[5] = 0;
1810                         localization = pfs_to_lo(strtoul(buf, NULL, 10));
1811                         pfsm = hammer_load_pseudofs(&trans, localization,
1812                                                     &error);
1813                         if (error == 0) {
1814                                 if (hammer_is_pfs_slave(&pfsm->pfsd)) {
1815                                         /* vap->va_size == 26 */
1816                                         ksnprintf(buf, sizeof(buf),
1817                                                   "@@0x%016jx:%05d",
1818                                                   (intmax_t)pfsm->pfsd.sync_end_tid,
1819                                                   lo_to_pfs(localization));
1820                                 } else {
1821                                         /* vap->va_size == 10 */
1822                                         ksnprintf(buf, sizeof(buf),
1823                                                   "@@-1:%05d",
1824                                                   lo_to_pfs(localization));
1825                                 }
1826                                 ptr = buf;
1827                                 bytes = strlen(buf);
1828                         }
1829                         if (pfsm)
1830                                 hammer_rel_pseudofs(hmp, pfsm);
1831                         hammer_done_transaction(&trans);
1832                 }
1833                 error = uiomove(ptr, bytes, ap->a_uio);
1834                 lwkt_reltoken(&hmp->fs_token);
1835                 return(error);
1836         }
1837
1838         /*
1839          * Long version
1840          */
1841         hammer_simple_transaction(&trans, hmp);
1842         ++hammer_stats_file_iopsr;
1843         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1844
1845         /*
1846          * Key range (begin and end inclusive) to scan.  Directory keys
1847          * directly translate to a 64 bit 'seek' position.
1848          */
1849         cursor.key_beg.localization = ip->obj_localization |
1850                                       HAMMER_LOCALIZE_MISC;
1851         cursor.key_beg.obj_id = ip->obj_id;
1852         cursor.key_beg.create_tid = 0;
1853         cursor.key_beg.delete_tid = 0;
1854         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1855         cursor.key_beg.obj_type = 0;
1856         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1857         cursor.asof = ip->obj_asof;
1858         cursor.flags |= HAMMER_CURSOR_ASOF;
1859
1860         error = hammer_ip_lookup(&cursor);
1861         if (error == 0) {
1862                 error = hammer_ip_resolve_data(&cursor);
1863                 if (error == 0) {
1864                         KKASSERT(cursor.leaf->data_len >=
1865                                  HAMMER_SYMLINK_NAME_OFF);
1866                         error = uiomove(cursor.data->symlink.name,
1867                                         cursor.leaf->data_len -
1868                                                 HAMMER_SYMLINK_NAME_OFF,
1869                                         ap->a_uio);
1870                 }
1871         }
1872         hammer_done_cursor(&cursor);
1873         hammer_done_transaction(&trans);
1874         lwkt_reltoken(&hmp->fs_token);
1875         return(error);
1876 }
1877
1878 /*
1879  * hammer_vop_nremove { nch, dvp, cred }
1880  */
1881 static
1882 int
1883 hammer_vop_nremove(struct vop_nremove_args *ap)
1884 {
1885         struct hammer_transaction trans;
1886         hammer_inode_t dip;
1887         hammer_mount_t hmp;
1888         int error;
1889
1890         dip = VTOI(ap->a_dvp);
1891         hmp = dip->hmp;
1892
1893         if (hammer_nohistory(dip) == 0 &&
1894             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1895                 return (error);
1896         }
1897
1898         lwkt_gettoken(&hmp->fs_token);
1899         hammer_start_transaction(&trans, hmp);
1900         ++hammer_stats_file_iopsw;
1901         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1902         hammer_done_transaction(&trans);
1903         if (error == 0)
1904                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1905         lwkt_reltoken(&hmp->fs_token);
1906         return (error);
1907 }
1908
1909 /*
1910  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1911  */
1912 static
1913 int
1914 hammer_vop_nrename(struct vop_nrename_args *ap)
1915 {
1916         struct hammer_transaction trans;
1917         struct namecache *fncp;
1918         struct namecache *tncp;
1919         hammer_inode_t fdip;
1920         hammer_inode_t tdip;
1921         hammer_inode_t ip;
1922         hammer_mount_t hmp;
1923         struct hammer_cursor cursor;
1924         int64_t namekey;
1925         uint32_t max_iterations;
1926         int nlen, error;
1927
1928         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1929                 return(EXDEV);
1930         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1931                 return(EXDEV);
1932
1933         fdip = VTOI(ap->a_fdvp);
1934         tdip = VTOI(ap->a_tdvp);
1935         fncp = ap->a_fnch->ncp;
1936         tncp = ap->a_tnch->ncp;
1937         ip = VTOI(fncp->nc_vp);
1938         KKASSERT(ip != NULL);
1939
1940         hmp = ip->hmp;
1941
1942         if (fdip->obj_localization != tdip->obj_localization)
1943                 return(EXDEV);
1944         if (fdip->obj_localization != ip->obj_localization)
1945                 return(EXDEV);
1946
1947         if (fdip->flags & HAMMER_INODE_RO)
1948                 return (EROFS);
1949         if (tdip->flags & HAMMER_INODE_RO)
1950                 return (EROFS);
1951         if (ip->flags & HAMMER_INODE_RO)
1952                 return (EROFS);
1953         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1954                 return (error);
1955
1956         lwkt_gettoken(&hmp->fs_token);
1957         hammer_start_transaction(&trans, hmp);
1958         ++hammer_stats_file_iopsw;
1959
1960         /*
1961          * Remove tncp from the target directory and then link ip as
1962          * tncp. XXX pass trans to dounlink
1963          *
1964          * Force the inode sync-time to match the transaction so it is
1965          * in-sync with the creation of the target directory entry.
1966          */
1967         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1968                                 ap->a_cred, 0, -1);
1969         if (error == 0 || error == ENOENT) {
1970                 error = hammer_ip_add_direntry(&trans, tdip,
1971                                                 tncp->nc_name, tncp->nc_nlen,
1972                                                 ip);
1973                 if (error == 0) {
1974                         ip->ino_data.parent_obj_id = tdip->obj_id;
1975                         ip->ino_data.ctime = trans.time;
1976                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1977                 }
1978         }
1979         if (error)
1980                 goto failed; /* XXX */
1981
1982         /*
1983          * Locate the record in the originating directory and remove it.
1984          *
1985          * Calculate the namekey and setup the key range for the scan.  This
1986          * works kinda like a chained hash table where the lower 32 bits
1987          * of the namekey synthesize the chain.
1988          *
1989          * The key range is inclusive of both key_beg and key_end.
1990          */
1991         namekey = hammer_direntry_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1992                                            &max_iterations);
1993 retry:
1994         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1995         cursor.key_beg.localization = fdip->obj_localization |
1996                                       hammer_dir_localization(fdip);
1997         cursor.key_beg.obj_id = fdip->obj_id;
1998         cursor.key_beg.key = namekey;
1999         cursor.key_beg.create_tid = 0;
2000         cursor.key_beg.delete_tid = 0;
2001         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2002         cursor.key_beg.obj_type = 0;
2003
2004         cursor.key_end = cursor.key_beg;
2005         cursor.key_end.key += max_iterations;
2006         cursor.asof = fdip->obj_asof;
2007         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2008
2009         /*
2010          * Scan all matching records (the chain), locate the one matching
2011          * the requested path component.
2012          *
2013          * The hammer_ip_*() functions merge in-memory records with on-disk
2014          * records for the purposes of the search.
2015          */
2016         error = hammer_ip_first(&cursor);
2017         while (error == 0) {
2018                 if (hammer_ip_resolve_data(&cursor) != 0)
2019                         break;
2020                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2021                 KKASSERT(nlen > 0);
2022                 if (fncp->nc_nlen == nlen &&
2023                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2024                         break;
2025                 }
2026                 error = hammer_ip_next(&cursor);
2027         }
2028
2029         /*
2030          * If all is ok we have to get the inode so we can adjust nlinks.
2031          *
2032          * WARNING: hammer_ip_del_direntry() may have to terminate the
2033          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2034          * twice.
2035          */
2036         if (error == 0)
2037                 error = hammer_ip_del_direntry(&trans, &cursor, fdip, ip);
2038
2039         /*
2040          * XXX A deadlock here will break rename's atomicy for the purposes
2041          * of crash recovery.
2042          */
2043         if (error == EDEADLK) {
2044                 hammer_done_cursor(&cursor);
2045                 goto retry;
2046         }
2047
2048         /*
2049          * Cleanup and tell the kernel that the rename succeeded.
2050          *
2051          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2052          *       without formally acquiring the vp since the vp might
2053          *       have zero refs on it, or in the middle of a reclaim,
2054          *       etc.
2055          */
2056         hammer_done_cursor(&cursor);
2057         if (error == 0) {
2058                 cache_rename(ap->a_fnch, ap->a_tnch);
2059                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2060                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2061                 while (ip->vp) {
2062                         struct vnode *vp;
2063
2064                         error = hammer_get_vnode(ip, &vp);
2065                         if (error == 0 && vp) {
2066                                 vn_unlock(vp);
2067                                 hammer_knote(ip->vp, NOTE_RENAME);
2068                                 vrele(vp);
2069                                 break;
2070                         }
2071                         hdkprintf("ip/vp race2 avoided\n");
2072                 }
2073         }
2074
2075 failed:
2076         hammer_done_transaction(&trans);
2077         lwkt_reltoken(&hmp->fs_token);
2078         return (error);
2079 }
2080
2081 /*
2082  * hammer_vop_nrmdir { nch, dvp, cred }
2083  */
2084 static
2085 int
2086 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2087 {
2088         struct hammer_transaction trans;
2089         hammer_inode_t dip;
2090         hammer_mount_t hmp;
2091         int error;
2092
2093         dip = VTOI(ap->a_dvp);
2094         hmp = dip->hmp;
2095
2096         if (hammer_nohistory(dip) == 0 &&
2097             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2098                 return (error);
2099         }
2100
2101         lwkt_gettoken(&hmp->fs_token);
2102         hammer_start_transaction(&trans, hmp);
2103         ++hammer_stats_file_iopsw;
2104         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2105         hammer_done_transaction(&trans);
2106         if (error == 0)
2107                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2108         lwkt_reltoken(&hmp->fs_token);
2109         return (error);
2110 }
2111
2112 /*
2113  * hammer_vop_markatime { vp, cred }
2114  */
2115 static
2116 int
2117 hammer_vop_markatime(struct vop_markatime_args *ap)
2118 {
2119         struct hammer_transaction trans;
2120         hammer_inode_t ip;
2121         hammer_mount_t hmp;
2122
2123         ip = VTOI(ap->a_vp);
2124         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2125                 return (EROFS);
2126         if (ip->flags & HAMMER_INODE_RO)
2127                 return (EROFS);
2128         hmp = ip->hmp;
2129         if (hmp->mp->mnt_flag & MNT_NOATIME)
2130                 return (0);
2131         lwkt_gettoken(&hmp->fs_token);
2132         hammer_start_transaction(&trans, hmp);
2133         ++hammer_stats_file_iopsw;
2134
2135         ip->ino_data.atime = trans.time;
2136         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2137         hammer_done_transaction(&trans);
2138         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2139         lwkt_reltoken(&hmp->fs_token);
2140         return (0);
2141 }
2142
2143 /*
2144  * hammer_vop_setattr { vp, vap, cred }
2145  */
2146 static
2147 int
2148 hammer_vop_setattr(struct vop_setattr_args *ap)
2149 {
2150         struct hammer_transaction trans;
2151         hammer_inode_t ip;
2152         struct vattr *vap;
2153         hammer_mount_t hmp;
2154         int modflags;
2155         int error;
2156         int truncating;
2157         int blksize;
2158         int kflags;
2159 #if 0
2160         int64_t aligned_size;
2161 #endif
2162         uint32_t flags;
2163
2164         vap = ap->a_vap;
2165         ip = ap->a_vp->v_data;
2166         modflags = 0;
2167         kflags = 0;
2168         hmp = ip->hmp;
2169
2170         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2171                 return(EROFS);
2172         if (ip->flags & HAMMER_INODE_RO)
2173                 return (EROFS);
2174         if (hammer_nohistory(ip) == 0 &&
2175             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2176                 return (error);
2177         }
2178
2179         lwkt_gettoken(&hmp->fs_token);
2180         hammer_start_transaction(&trans, hmp);
2181         ++hammer_stats_file_iopsw;
2182         error = 0;
2183
2184         if (vap->va_flags != VNOVAL) {
2185                 flags = ip->ino_data.uflags;
2186                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2187                                          hammer_to_unix_xid(&ip->ino_data.uid),
2188                                          ap->a_cred);
2189                 if (error == 0) {
2190                         if (ip->ino_data.uflags != flags) {
2191                                 ip->ino_data.uflags = flags;
2192                                 ip->ino_data.ctime = trans.time;
2193                                 modflags |= HAMMER_INODE_DDIRTY;
2194                                 kflags |= NOTE_ATTRIB;
2195                         }
2196                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2197                                 error = 0;
2198                                 goto done;
2199                         }
2200                 }
2201                 goto done;
2202         }
2203         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2204                 error = EPERM;
2205                 goto done;
2206         }
2207         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2208                 mode_t cur_mode = ip->ino_data.mode;
2209                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2210                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2211                 uuid_t uuid_uid;
2212                 uuid_t uuid_gid;
2213
2214                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2215                                          ap->a_cred,
2216                                          &cur_uid, &cur_gid, &cur_mode);
2217                 if (error == 0) {
2218                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2219                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2220                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2221                                  sizeof(uuid_uid)) ||
2222                             bcmp(&uuid_gid, &ip->ino_data.gid,
2223                                  sizeof(uuid_gid)) ||
2224                             ip->ino_data.mode != cur_mode) {
2225                                 ip->ino_data.uid = uuid_uid;
2226                                 ip->ino_data.gid = uuid_gid;
2227                                 ip->ino_data.mode = cur_mode;
2228                                 ip->ino_data.ctime = trans.time;
2229                                 modflags |= HAMMER_INODE_DDIRTY;
2230                         }
2231                         kflags |= NOTE_ATTRIB;
2232                 }
2233         }
2234         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2235                 switch(ap->a_vp->v_type) {
2236                 case VREG:
2237                         if (vap->va_size == ip->ino_data.size)
2238                                 break;
2239
2240                         /*
2241                          * Log the operation if in fast-fsync mode or if
2242                          * there are unterminated redo write records present.
2243                          *
2244                          * The second check is needed so the recovery code
2245                          * properly truncates write redos even if nominal
2246                          * REDO operations is turned off due to excessive
2247                          * writes, because the related records might be
2248                          * destroyed and never lay down a TERM_WRITE.
2249                          */
2250                         if ((ip->flags & HAMMER_INODE_REDO) ||
2251                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2252                                 error = hammer_generate_redo(&trans, ip,
2253                                                              vap->va_size,
2254                                                              HAMMER_REDO_TRUNC,
2255                                                              NULL, 0);
2256                         }
2257                         blksize = hammer_blocksize(vap->va_size);
2258
2259                         /*
2260                          * XXX break atomicy, we can deadlock the backend
2261                          * if we do not release the lock.  Probably not a
2262                          * big deal here.
2263                          */
2264                         if (vap->va_size < ip->ino_data.size) {
2265                                 nvtruncbuf(ap->a_vp, vap->va_size,
2266                                            blksize,
2267                                            hammer_blockoff(vap->va_size),
2268                                            0);
2269                                 truncating = 1;
2270                                 kflags |= NOTE_WRITE;
2271                         } else {
2272                                 nvextendbuf(ap->a_vp,
2273                                             ip->ino_data.size,
2274                                             vap->va_size,
2275                                             hammer_blocksize(ip->ino_data.size),
2276                                             hammer_blocksize(vap->va_size),
2277                                             hammer_blockoff(ip->ino_data.size),
2278                                             hammer_blockoff(vap->va_size),
2279                                             0);
2280                                 truncating = 0;
2281                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2282                         }
2283                         ip->ino_data.size = vap->va_size;
2284                         ip->ino_data.mtime = trans.time;
2285                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2286                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2287
2288                         /*
2289                          * On-media truncation is cached in the inode until
2290                          * the inode is synchronized.  We must immediately
2291                          * handle any frontend records.
2292                          */
2293                         if (truncating) {
2294                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2295                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2296                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2297                                         ip->trunc_off = vap->va_size;
2298                                         hammer_inode_dirty(ip);
2299                                 } else if (ip->trunc_off > vap->va_size) {
2300                                         ip->trunc_off = vap->va_size;
2301                                 }
2302                         }
2303
2304 #if 0
2305                         /*
2306                          * When truncating, nvtruncbuf() may have cleaned out
2307                          * a portion of the last block on-disk in the buffer
2308                          * cache.  We must clean out any frontend records
2309                          * for blocks beyond the new last block.
2310                          */
2311                         aligned_size = (vap->va_size + (blksize - 1)) &
2312                                        ~(int64_t)(blksize - 1);
2313                         if (truncating && vap->va_size < aligned_size) {
2314                                 aligned_size -= blksize;
2315                                 hammer_ip_frontend_trunc(ip, aligned_size);
2316                         }
2317 #endif
2318                         break;
2319                 case VDATABASE:
2320                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2321                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2322                                 ip->trunc_off = vap->va_size;
2323                                 hammer_inode_dirty(ip);
2324                         } else if (ip->trunc_off > vap->va_size) {
2325                                 ip->trunc_off = vap->va_size;
2326                         }
2327                         hammer_ip_frontend_trunc(ip, vap->va_size);
2328                         ip->ino_data.size = vap->va_size;
2329                         ip->ino_data.mtime = trans.time;
2330                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2331                         kflags |= NOTE_ATTRIB;
2332                         break;
2333                 default:
2334                         error = EINVAL;
2335                         goto done;
2336                 }
2337                 break;
2338         }
2339         if (vap->va_atime.tv_sec != VNOVAL) {
2340                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2341                 modflags |= HAMMER_INODE_ATIME;
2342                 kflags |= NOTE_ATTRIB;
2343         }
2344         if (vap->va_mtime.tv_sec != VNOVAL) {
2345                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2346                 modflags |= HAMMER_INODE_MTIME;
2347                 kflags |= NOTE_ATTRIB;
2348         }
2349         if (vap->va_mode != (mode_t)VNOVAL) {
2350                 mode_t   cur_mode = ip->ino_data.mode;
2351                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2352                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2353
2354                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2355                                          cur_uid, cur_gid, &cur_mode);
2356                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2357                         ip->ino_data.mode = cur_mode;
2358                         ip->ino_data.ctime = trans.time;
2359                         modflags |= HAMMER_INODE_DDIRTY;
2360                         kflags |= NOTE_ATTRIB;
2361                 }
2362         }
2363 done:
2364         if (error == 0)
2365                 hammer_modify_inode(&trans, ip, modflags);
2366         hammer_done_transaction(&trans);
2367         hammer_knote(ap->a_vp, kflags);
2368         lwkt_reltoken(&hmp->fs_token);
2369         return (error);
2370 }
2371
2372 /*
2373  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2374  */
2375 static
2376 int
2377 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2378 {
2379         struct hammer_transaction trans;
2380         hammer_inode_t dip;
2381         hammer_inode_t nip;
2382         hammer_record_t record;
2383         struct nchandle *nch;
2384         hammer_mount_t hmp;
2385         int error;
2386         int bytes;
2387
2388         ap->a_vap->va_type = VLNK;
2389
2390         nch = ap->a_nch;
2391         dip = VTOI(ap->a_dvp);
2392         hmp = dip->hmp;
2393
2394         if (dip->flags & HAMMER_INODE_RO)
2395                 return (EROFS);
2396         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2397                 return (error);
2398
2399         /*
2400          * Create a transaction to cover the operations we perform.
2401          */
2402         lwkt_gettoken(&hmp->fs_token);
2403         hammer_start_transaction(&trans, hmp);
2404         ++hammer_stats_file_iopsw;
2405
2406         /*
2407          * Create a new filesystem object of the requested type.  The
2408          * returned inode will be referenced but not locked.
2409          */
2410
2411         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2412                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2413                                     NULL, &nip);
2414         if (error) {
2415                 hammer_done_transaction(&trans);
2416                 *ap->a_vpp = NULL;
2417                 lwkt_reltoken(&hmp->fs_token);
2418                 return (error);
2419         }
2420
2421         /*
2422          * Add a record representing the symlink.  symlink stores the link
2423          * as pure data, not a string, and is no \0 terminated.
2424          */
2425         if (error == 0) {
2426                 bytes = strlen(ap->a_target);
2427
2428                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2429                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2430                 } else {
2431                         record = hammer_alloc_mem_record(nip, bytes);
2432                         record->type = HAMMER_MEM_RECORD_GENERAL;
2433
2434                         record->leaf.base.localization = nip->obj_localization |
2435                                                          HAMMER_LOCALIZE_MISC;
2436                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2437                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2438                         record->leaf.data_len = bytes;
2439                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2440                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2441                         error = hammer_ip_add_record(&trans, record);
2442                 }
2443
2444                 /*
2445                  * Set the file size to the length of the link.
2446                  */
2447                 if (error == 0) {
2448                         nip->ino_data.size = bytes;
2449                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2450                 }
2451         }
2452         if (error == 0)
2453                 error = hammer_ip_add_direntry(&trans, dip, nch->ncp->nc_name,
2454                                                 nch->ncp->nc_nlen, nip);
2455
2456         /*
2457          * Finish up.
2458          */
2459         if (error) {
2460                 hammer_rel_inode(nip, 0);
2461                 *ap->a_vpp = NULL;
2462         } else {
2463                 error = hammer_get_vnode(nip, ap->a_vpp);
2464                 hammer_rel_inode(nip, 0);
2465                 if (error == 0) {
2466                         cache_setunresolved(ap->a_nch);
2467                         cache_setvp(ap->a_nch, *ap->a_vpp);
2468                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2469                 }
2470         }
2471         hammer_done_transaction(&trans);
2472         lwkt_reltoken(&hmp->fs_token);
2473         return (error);
2474 }
2475
2476 /*
2477  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2478  */
2479 static
2480 int
2481 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2482 {
2483         struct hammer_transaction trans;
2484         hammer_inode_t dip;
2485         hammer_mount_t hmp;
2486         int error;
2487
2488         dip = VTOI(ap->a_dvp);
2489         hmp = dip->hmp;
2490
2491         if (hammer_nohistory(dip) == 0 &&
2492             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2493                 return (error);
2494         }
2495
2496         lwkt_gettoken(&hmp->fs_token);
2497         hammer_start_transaction(&trans, hmp);
2498         ++hammer_stats_file_iopsw;
2499         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2500                                 ap->a_cred, ap->a_flags, -1);
2501         hammer_done_transaction(&trans);
2502         lwkt_reltoken(&hmp->fs_token);
2503
2504         return (error);
2505 }
2506
2507 /*
2508  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2509  */
2510 static
2511 int
2512 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2513 {
2514         hammer_inode_t ip = ap->a_vp->v_data;
2515         hammer_mount_t hmp = ip->hmp;
2516         int error;
2517
2518         ++hammer_stats_file_iopsr;
2519         lwkt_gettoken(&hmp->fs_token);
2520         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2521                              ap->a_fflag, ap->a_cred);
2522         lwkt_reltoken(&hmp->fs_token);
2523         return (error);
2524 }
2525
2526 static
2527 int
2528 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2529 {
2530         static const struct mountctl_opt extraopt[] = {
2531                 { HMNT_NOHISTORY,       "nohistory" },
2532                 { HMNT_MASTERID,        "master" },
2533                 { HMNT_NOMIRROR,        "nomirror" },
2534                 { 0, NULL}
2535
2536         };
2537         hammer_mount_t hmp;
2538         struct mount *mp;
2539         int usedbytes;
2540         int error;
2541
2542         error = 0;
2543         usedbytes = 0;
2544         mp = ap->a_head.a_ops->head.vv_mount;
2545         KKASSERT(mp->mnt_data != NULL);
2546         hmp = (hammer_mount_t)mp->mnt_data;
2547
2548         lwkt_gettoken(&hmp->fs_token);
2549
2550         switch(ap->a_op) {
2551         case MOUNTCTL_SET_EXPORT:
2552                 if (ap->a_ctllen != sizeof(struct export_args))
2553                         error = EINVAL;
2554                 else
2555                         error = hammer_vfs_export(mp, ap->a_op,
2556                                       (const struct export_args *)ap->a_ctl);
2557                 break;
2558         case MOUNTCTL_MOUNTFLAGS:
2559                 /*
2560                  * Call standard mountctl VOP function
2561                  * so we get user mount flags.
2562                  */
2563                 error = vop_stdmountctl(ap);
2564                 if (error)
2565                         break;
2566
2567                 usedbytes = *ap->a_res;
2568
2569                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2570                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2571                                                     ap->a_buf,
2572                                                     ap->a_buflen - usedbytes,
2573                                                     &error);
2574                 }
2575
2576                 *ap->a_res += usedbytes;
2577                 break;
2578         default:
2579                 error = vop_stdmountctl(ap);
2580                 break;
2581         }
2582         lwkt_reltoken(&hmp->fs_token);
2583         return(error);
2584 }
2585
2586 /*
2587  * hammer_vop_strategy { vp, bio }
2588  *
2589  * Strategy call, used for regular file read & write only.  Note that the
2590  * bp may represent a cluster.
2591  *
2592  * To simplify operation and allow better optimizations in the future,
2593  * this code does not make any assumptions with regards to buffer alignment
2594  * or size.
2595  */
2596 static
2597 int
2598 hammer_vop_strategy(struct vop_strategy_args *ap)
2599 {
2600         struct buf *bp;
2601         int error;
2602
2603         bp = ap->a_bio->bio_buf;
2604
2605         switch(bp->b_cmd) {
2606         case BUF_CMD_READ:
2607                 error = hammer_vop_strategy_read(ap);
2608                 break;
2609         case BUF_CMD_WRITE:
2610                 error = hammer_vop_strategy_write(ap);
2611                 break;
2612         default:
2613                 bp->b_error = error = EINVAL;
2614                 bp->b_flags |= B_ERROR;
2615                 biodone(ap->a_bio);
2616                 break;
2617         }
2618
2619         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2620
2621         return (error);
2622 }
2623
2624 /*
2625  * Read from a regular file.  Iterate the related records and fill in the
2626  * BIO/BUF.  Gaps are zero-filled.
2627  *
2628  * The support code in hammer_object.c should be used to deal with mixed
2629  * in-memory and on-disk records.
2630  *
2631  * NOTE: Can be called from the cluster code with an oversized buf.
2632  *
2633  * XXX atime update
2634  */
2635 static
2636 int
2637 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2638 {
2639         struct hammer_transaction trans;
2640         hammer_inode_t ip;
2641         hammer_inode_t dip;
2642         hammer_mount_t hmp;
2643         struct hammer_cursor cursor;
2644         hammer_base_elm_t base;
2645         hammer_off_t disk_offset;
2646         struct bio *bio;
2647         struct bio *nbio;
2648         struct buf *bp;
2649         int64_t rec_offset;
2650         int64_t ran_end;
2651         int64_t tmp64;
2652         int error;
2653         int boff;
2654         int roff;
2655         int n;
2656         int isdedupable;
2657
2658         bio = ap->a_bio;
2659         bp = bio->bio_buf;
2660         ip = ap->a_vp->v_data;
2661         hmp = ip->hmp;
2662
2663         /*
2664          * The zone-2 disk offset may have been set by the cluster code via
2665          * a BMAP operation, or else should be NOOFFSET.
2666          *
2667          * Checking the high bits for a match against zone-2 should suffice.
2668          *
2669          * In cases where a lot of data duplication is present it may be
2670          * more beneficial to drop through and doubule-buffer through the
2671          * device.
2672          */
2673         nbio = push_bio(bio);
2674         if (hammer_is_zone_large_data(nbio->bio_offset)) {
2675                 if (hammer_double_buffer == 0) {
2676                         lwkt_gettoken(&hmp->fs_token);
2677                         error = hammer_io_direct_read(hmp, nbio, NULL);
2678                         lwkt_reltoken(&hmp->fs_token);
2679                         return (error);
2680                 }
2681
2682                 /*
2683                  * Try to shortcut requests for double_buffer mode too.
2684                  * Since this mode runs through the device buffer cache
2685                  * only compatible buffer sizes (meaning those generated
2686                  * by normal filesystem buffers) are legal.
2687                  */
2688                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2689                         lwkt_gettoken(&hmp->fs_token);
2690                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2691                         lwkt_reltoken(&hmp->fs_token);
2692                         return (error);
2693                 }
2694         }
2695
2696         /*
2697          * Well, that sucked.  Do it the hard way.  If all the stars are
2698          * aligned we may still be able to issue a direct-read.
2699          */
2700         lwkt_gettoken(&hmp->fs_token);
2701         hammer_simple_transaction(&trans, hmp);
2702         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2703
2704         /*
2705          * Key range (begin and end inclusive) to scan.  Note that the key's
2706          * stored in the actual records represent BASE+LEN, not BASE.  The
2707          * first record containing bio_offset will have a key > bio_offset.
2708          */
2709         cursor.key_beg.localization = ip->obj_localization |
2710                                       HAMMER_LOCALIZE_MISC;
2711         cursor.key_beg.obj_id = ip->obj_id;
2712         cursor.key_beg.create_tid = 0;
2713         cursor.key_beg.delete_tid = 0;
2714         cursor.key_beg.obj_type = 0;
2715         cursor.key_beg.key = bio->bio_offset + 1;
2716         cursor.asof = ip->obj_asof;
2717         cursor.flags |= HAMMER_CURSOR_ASOF;
2718
2719         cursor.key_end = cursor.key_beg;
2720         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2721 #if 0
2722         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2723                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2724                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2725                 cursor.key_end.key = HAMMER_MAX_KEY;
2726         } else
2727 #endif
2728         {
2729                 ran_end = bio->bio_offset + bp->b_bufsize;
2730                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2731                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2732                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2733                 if (tmp64 < ran_end)
2734                         cursor.key_end.key = HAMMER_MAX_KEY;
2735                 else
2736                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2737         }
2738         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2739
2740         /*
2741          * Set NOSWAPCACHE for cursor data extraction if double buffering
2742          * is disabled or (if the file is not marked cacheable via chflags
2743          * and vm.swapcache_use_chflags is enabled).
2744          */
2745         if (hammer_double_buffer == 0 ||
2746             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2747              vm_swapcache_use_chflags)) {
2748                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2749         }
2750
2751         error = hammer_ip_first(&cursor);
2752         boff = 0;
2753
2754         while (error == 0) {
2755                 /*
2756                  * Get the base file offset of the record.  The key for
2757                  * data records is (base + bytes) rather then (base).
2758                  */
2759                 base = &cursor.leaf->base;
2760                 rec_offset = base->key - cursor.leaf->data_len;
2761
2762                 /*
2763                  * Calculate the gap, if any, and zero-fill it.
2764                  *
2765                  * n is the offset of the start of the record verses our
2766                  * current seek offset in the bio.
2767                  */
2768                 n = (int)(rec_offset - (bio->bio_offset + boff));
2769                 if (n > 0) {
2770                         if (n > bp->b_bufsize - boff)
2771                                 n = bp->b_bufsize - boff;
2772                         bzero((char *)bp->b_data + boff, n);
2773                         boff += n;
2774                         n = 0;
2775                 }
2776
2777                 /*
2778                  * Calculate the data offset in the record and the number
2779                  * of bytes we can copy.
2780                  *
2781                  * There are two degenerate cases.  First, boff may already
2782                  * be at bp->b_bufsize.  Secondly, the data offset within
2783                  * the record may exceed the record's size.
2784                  */
2785                 roff = -n;
2786                 rec_offset += roff;
2787                 n = cursor.leaf->data_len - roff;
2788                 if (n <= 0) {
2789                         hdkprintf("bad n=%d roff=%d\n", n, roff);
2790                         n = 0;
2791                 } else if (n > bp->b_bufsize - boff) {
2792                         n = bp->b_bufsize - boff;
2793                 }
2794
2795                 /*
2796                  * Deal with cached truncations.  This cool bit of code
2797                  * allows truncate()/ftruncate() to avoid having to sync
2798                  * the file.
2799                  *
2800                  * If the frontend is truncated then all backend records are
2801                  * subject to the frontend's truncation.
2802                  *
2803                  * If the backend is truncated then backend records on-disk
2804                  * (but not in-memory) are subject to the backend's
2805                  * truncation.  In-memory records owned by the backend
2806                  * represent data written after the truncation point on the
2807                  * backend and must not be truncated.
2808                  *
2809                  * Truncate operations deal with frontend buffer cache
2810                  * buffers and frontend-owned in-memory records synchronously.
2811                  */
2812                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2813                         if (hammer_cursor_ondisk(&cursor)/* ||
2814                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2815                                 if (ip->trunc_off <= rec_offset)
2816                                         n = 0;
2817                                 else if (ip->trunc_off < rec_offset + n)
2818                                         n = (int)(ip->trunc_off - rec_offset);
2819                         }
2820                 }
2821                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2822                         if (hammer_cursor_ondisk(&cursor)) {
2823                                 if (ip->sync_trunc_off <= rec_offset)
2824                                         n = 0;
2825                                 else if (ip->sync_trunc_off < rec_offset + n)
2826                                         n = (int)(ip->sync_trunc_off - rec_offset);
2827                         }
2828                 }
2829
2830                 /*
2831                  * Try to issue a direct read into our bio if possible,
2832                  * otherwise resolve the element data into a hammer_buffer
2833                  * and copy.
2834                  *
2835                  * The buffer on-disk should be zerod past any real
2836                  * truncation point, but may not be for any synthesized
2837                  * truncation point from above.
2838                  *
2839                  * NOTE: disk_offset is only valid if the cursor data is
2840                  *       on-disk.
2841                  */
2842                 disk_offset = cursor.leaf->data_offset + roff;
2843                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2844                                hammer_cursor_ondisk(&cursor) &&
2845                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2846
2847                 if (isdedupable && hammer_double_buffer == 0) {
2848                         /*
2849                          * Direct read case
2850                          */
2851                         KKASSERT(hammer_is_zone_large_data(disk_offset));
2852                         nbio->bio_offset = disk_offset;
2853                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2854                         if (hammer_live_dedup && error == 0)
2855                                 hammer_dedup_cache_add(ip, cursor.leaf);
2856                         goto done;
2857                 } else if (isdedupable) {
2858                         /*
2859                          * Async I/O case for reading from backing store
2860                          * and copying the data to the filesystem buffer.
2861                          * live-dedup has to verify the data anyway if it
2862                          * gets a hit later so we can just add the entry
2863                          * now.
2864                          */
2865                         KKASSERT(hammer_is_zone_large_data(disk_offset));
2866                         nbio->bio_offset = disk_offset;
2867                         if (hammer_live_dedup)
2868                                 hammer_dedup_cache_add(ip, cursor.leaf);
2869                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2870                         goto done;
2871                 } else if (n) {
2872                         error = hammer_ip_resolve_data(&cursor);
2873                         if (error == 0) {
2874                                 if (hammer_live_dedup && isdedupable)
2875                                         hammer_dedup_cache_add(ip, cursor.leaf);
2876                                 bcopy((char *)cursor.data + roff,
2877                                       (char *)bp->b_data + boff, n);
2878                         }
2879                 }
2880                 if (error)
2881                         break;
2882
2883                 /*
2884                  * We have to be sure that the only elements added to the
2885                  * dedup cache are those which are already on-media.
2886                  */
2887                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2888                         hammer_dedup_cache_add(ip, cursor.leaf);
2889
2890                 /*
2891                  * Iterate until we have filled the request.
2892                  */
2893                 boff += n;
2894                 if (boff == bp->b_bufsize)
2895                         break;
2896                 error = hammer_ip_next(&cursor);
2897         }
2898
2899         /*
2900          * There may have been a gap after the last record
2901          */
2902         if (error == ENOENT)
2903                 error = 0;
2904         if (error == 0 && boff != bp->b_bufsize) {
2905                 KKASSERT(boff < bp->b_bufsize);
2906                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2907                 /* boff = bp->b_bufsize; */
2908         }
2909
2910         /*
2911          * Disallow swapcache operation on the vnode buffer if double
2912          * buffering is enabled, the swapcache will get the data via
2913          * the block device buffer.
2914          */
2915         if (hammer_double_buffer)
2916                 bp->b_flags |= B_NOTMETA;
2917
2918         /*
2919          * Cleanup
2920          */
2921         bp->b_resid = 0;
2922         bp->b_error = error;
2923         if (error)
2924                 bp->b_flags |= B_ERROR;
2925         biodone(ap->a_bio);
2926
2927 done:
2928         /*
2929          * Cache the b-tree node for the last data read in cache[1].
2930          *
2931          * If we hit the file EOF then also cache the node in the
2932          * governing directory's cache[3], it will be used to initialize
2933          * the new inode's cache[1] for any inodes looked up via the directory.
2934          *
2935          * This doesn't reduce disk accesses since the B-Tree chain is
2936          * likely cached, but it does reduce cpu overhead when looking
2937          * up file offsets for cpdup/tar/cpio style iterations.
2938          */
2939         if (cursor.node)
2940                 hammer_cache_node(&ip->cache[1], cursor.node);
2941         if (ran_end >= ip->ino_data.size) {
2942                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2943                                         ip->obj_asof, ip->obj_localization);
2944                 if (dip) {
2945                         hammer_cache_node(&dip->cache[3], cursor.node);
2946                         hammer_rel_inode(dip, 0);
2947                 }
2948         }
2949         hammer_done_cursor(&cursor);
2950         hammer_done_transaction(&trans);
2951         lwkt_reltoken(&hmp->fs_token);
2952         return(error);
2953 }
2954
2955 /*
2956  * BMAP operation - used to support cluster_read() only.
2957  *
2958  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2959  *
2960  * This routine may return EOPNOTSUPP if the opration is not supported for
2961  * the specified offset.  The contents of the pointer arguments do not
2962  * need to be initialized in that case.
2963  *
2964  * If a disk address is available and properly aligned return 0 with
2965  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2966  * to the run-length relative to that offset.  Callers may assume that
2967  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2968  * large, so return EOPNOTSUPP if it is not sufficiently large.
2969  */
2970 static
2971 int
2972 hammer_vop_bmap(struct vop_bmap_args *ap)
2973 {
2974         struct hammer_transaction trans;
2975         hammer_inode_t ip;
2976         hammer_mount_t hmp;
2977         struct hammer_cursor cursor;
2978         hammer_base_elm_t base;
2979         int64_t rec_offset;
2980         int64_t ran_end;
2981         int64_t tmp64;
2982         int64_t base_offset;
2983         int64_t base_disk_offset;
2984         int64_t last_offset;
2985         hammer_off_t last_disk_offset;
2986         hammer_off_t disk_offset;
2987         int     rec_len;
2988         int     error;
2989         int     blksize;
2990
2991         ++hammer_stats_file_iopsr;
2992         ip = ap->a_vp->v_data;
2993         hmp = ip->hmp;
2994
2995         /*
2996          * We can only BMAP regular files.  We can't BMAP database files,
2997          * directories, etc.
2998          */
2999         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3000                 return(EOPNOTSUPP);
3001
3002         /*
3003          * bmap is typically called with runp/runb both NULL when used
3004          * for writing.  We do not support BMAP for writing atm.
3005          */
3006         if (ap->a_cmd != BUF_CMD_READ)
3007                 return(EOPNOTSUPP);
3008
3009         /*
3010          * Scan the B-Tree to acquire blockmap addresses, then translate
3011          * to raw addresses.
3012          */
3013         lwkt_gettoken(&hmp->fs_token);
3014         hammer_simple_transaction(&trans, hmp);
3015
3016         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3017
3018         /*
3019          * Key range (begin and end inclusive) to scan.  Note that the key's
3020          * stored in the actual records represent BASE+LEN, not BASE.  The
3021          * first record containing bio_offset will have a key > bio_offset.
3022          */
3023         cursor.key_beg.localization = ip->obj_localization |
3024                                       HAMMER_LOCALIZE_MISC;
3025         cursor.key_beg.obj_id = ip->obj_id;
3026         cursor.key_beg.create_tid = 0;
3027         cursor.key_beg.delete_tid = 0;
3028         cursor.key_beg.obj_type = 0;
3029         if (ap->a_runb)
3030                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3031         else
3032                 cursor.key_beg.key = ap->a_loffset + 1;
3033         if (cursor.key_beg.key < 0)
3034                 cursor.key_beg.key = 0;
3035         cursor.asof = ip->obj_asof;
3036         cursor.flags |= HAMMER_CURSOR_ASOF;
3037
3038         cursor.key_end = cursor.key_beg;
3039         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3040
3041         ran_end = ap->a_loffset + MAXPHYS;
3042         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3043         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3044         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3045         if (tmp64 < ran_end)
3046                 cursor.key_end.key = HAMMER_MAX_KEY;
3047         else
3048                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3049
3050         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3051
3052         error = hammer_ip_first(&cursor);
3053         base_offset = last_offset = 0;
3054         base_disk_offset = last_disk_offset = 0;
3055
3056         while (error == 0) {
3057                 /*
3058                  * Get the base file offset of the record.  The key for
3059                  * data records is (base + bytes) rather then (base).
3060                  *
3061                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3062                  * The extra bytes should be zero on-disk and the BMAP op
3063                  * should still be ok.
3064                  */
3065                 base = &cursor.leaf->base;
3066                 rec_offset = base->key - cursor.leaf->data_len;
3067                 rec_len    = cursor.leaf->data_len;
3068
3069                 /*
3070                  * Incorporate any cached truncation.
3071                  *
3072                  * NOTE: Modifications to rec_len based on synthesized
3073                  * truncation points remove the guarantee that any extended
3074                  * data on disk is zero (since the truncations may not have
3075                  * taken place on-media yet).
3076                  */
3077                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3078                         if (hammer_cursor_ondisk(&cursor) ||
3079                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3080                                 if (ip->trunc_off <= rec_offset)
3081                                         rec_len = 0;
3082                                 else if (ip->trunc_off < rec_offset + rec_len)
3083                                         rec_len = (int)(ip->trunc_off - rec_offset);
3084                         }
3085                 }
3086                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3087                         if (hammer_cursor_ondisk(&cursor)) {
3088                                 if (ip->sync_trunc_off <= rec_offset)
3089                                         rec_len = 0;
3090                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3091                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3092                         }
3093                 }
3094
3095                 /*
3096                  * Accumulate information.  If we have hit a discontiguous
3097                  * block reset base_offset unless we are already beyond the
3098                  * requested offset.  If we are, that's it, we stop.
3099                  */
3100                 if (error)
3101                         break;
3102                 if (hammer_cursor_ondisk(&cursor)) {
3103                         disk_offset = cursor.leaf->data_offset;
3104                         if (rec_offset != last_offset ||
3105                             disk_offset != last_disk_offset) {
3106                                 if (rec_offset > ap->a_loffset)
3107                                         break;
3108                                 base_offset = rec_offset;
3109                                 base_disk_offset = disk_offset;
3110                         }
3111                         last_offset = rec_offset + rec_len;
3112                         last_disk_offset = disk_offset + rec_len;
3113
3114                         if (hammer_live_dedup)
3115                                 hammer_dedup_cache_add(ip, cursor.leaf);
3116                 }
3117
3118                 error = hammer_ip_next(&cursor);
3119         }
3120
3121         if (cursor.node)
3122                 hammer_cache_node(&ip->cache[1], cursor.node);
3123
3124         hammer_done_cursor(&cursor);
3125         hammer_done_transaction(&trans);
3126         lwkt_reltoken(&hmp->fs_token);
3127
3128         /*
3129          * If we couldn't find any records or the records we did find were
3130          * all behind the requested offset, return failure.  A forward
3131          * truncation can leave a hole w/ no on-disk records.
3132          */
3133         if (last_offset == 0 || last_offset < ap->a_loffset)
3134                 return (EOPNOTSUPP);
3135
3136         /*
3137          * Figure out the block size at the requested offset and adjust
3138          * our limits so the cluster_read() does not create inappropriately
3139          * sized buffer cache buffers.
3140          */
3141         blksize = hammer_blocksize(ap->a_loffset);
3142         if (hammer_blocksize(base_offset) != blksize) {
3143                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3144         }
3145         if (last_offset != ap->a_loffset &&
3146             hammer_blocksize(last_offset - 1) != blksize) {
3147                 last_offset = hammer_blockdemarc(ap->a_loffset,
3148                                                  last_offset - 1);
3149         }
3150
3151         /*
3152          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3153          * from occuring.
3154          */
3155         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3156
3157         if (!hammer_is_zone_large_data(disk_offset)) {
3158                 /*
3159                  * Only large-data zones can be direct-IOd
3160                  */
3161                 error = EOPNOTSUPP;
3162         } else if ((disk_offset & HAMMER_BUFMASK) ||
3163                    (last_offset - ap->a_loffset) < blksize) {
3164                 /*
3165                  * doffsetp is not aligned or the forward run size does
3166                  * not cover a whole buffer, disallow the direct I/O.
3167                  */
3168                 error = EOPNOTSUPP;
3169         } else {
3170                 /*
3171                  * We're good.
3172                  */
3173                 *ap->a_doffsetp = disk_offset;
3174                 if (ap->a_runb) {
3175                         *ap->a_runb = ap->a_loffset - base_offset;
3176                         KKASSERT(*ap->a_runb >= 0);
3177                 }
3178                 if (ap->a_runp) {
3179                         *ap->a_runp = last_offset - ap->a_loffset;
3180                         KKASSERT(*ap->a_runp >= 0);
3181                 }
3182                 error = 0;
3183         }
3184         return(error);
3185 }
3186
3187 /*
3188  * Write to a regular file.   Because this is a strategy call the OS is
3189  * trying to actually get data onto the media.
3190  */
3191 static
3192 int
3193 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3194 {
3195         hammer_record_t record;
3196         hammer_mount_t hmp;
3197         hammer_inode_t ip;
3198         struct bio *bio;
3199         struct buf *bp;
3200         int blksize __debugvar;
3201         int bytes;
3202         int error;
3203
3204         bio = ap->a_bio;
3205         bp = bio->bio_buf;
3206         ip = ap->a_vp->v_data;
3207         hmp = ip->hmp;
3208
3209         blksize = hammer_blocksize(bio->bio_offset);
3210         KKASSERT(bp->b_bufsize == blksize);
3211
3212         if (ip->flags & HAMMER_INODE_RO) {
3213                 bp->b_error = EROFS;
3214                 bp->b_flags |= B_ERROR;
3215                 biodone(ap->a_bio);
3216                 return(EROFS);
3217         }
3218
3219         lwkt_gettoken(&hmp->fs_token);
3220
3221         /*
3222          * Disallow swapcache operation on the vnode buffer if double
3223          * buffering is enabled, the swapcache will get the data via
3224          * the block device buffer.
3225          */
3226         if (hammer_double_buffer)
3227                 bp->b_flags |= B_NOTMETA;
3228
3229         /*
3230          * Interlock with inode destruction (no in-kernel or directory
3231          * topology visibility).  If we queue new IO while trying to
3232          * destroy the inode we can deadlock the vtrunc call in
3233          * hammer_inode_unloadable_check().
3234          *
3235          * Besides, there's no point flushing a bp associated with an
3236          * inode that is being destroyed on-media and has no kernel
3237          * references.
3238          */
3239         if ((ip->flags | ip->sync_flags) &
3240             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3241                 bp->b_resid = 0;
3242                 biodone(ap->a_bio);
3243                 lwkt_reltoken(&hmp->fs_token);
3244                 return(0);
3245         }
3246
3247         /*
3248          * Reserve space and issue a direct-write from the front-end.
3249          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3250          * allocations.
3251          *
3252          * An in-memory record will be installed to reference the storage
3253          * until the flusher can get to it.
3254          *
3255          * Since we own the high level bio the front-end will not try to
3256          * do a direct-read until the write completes.
3257          *
3258          * NOTE: The only time we do not reserve a full-sized buffers
3259          * worth of data is if the file is small.  We do not try to
3260          * allocate a fragment (from the small-data zone) at the end of
3261          * an otherwise large file as this can lead to wildly separated
3262          * data.
3263          */
3264         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3265         KKASSERT(bio->bio_offset < ip->ino_data.size);
3266         if (bio->bio_offset || ip->ino_data.size > HAMMER_HBUFSIZE)
3267                 bytes = bp->b_bufsize;
3268         else
3269                 bytes = HAMMER_DATA_DOALIGN_WITH(int, ip->ino_data.size);
3270
3271         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3272                                     bytes, &error);
3273
3274         /*
3275          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3276          * in hammer_vop_write().  We must flag the record so the proper
3277          * REDO_TERM_WRITE entry is generated during the flush.
3278          */
3279         if (record) {
3280                 if (bp->b_flags & B_VFSFLAG1) {
3281                         record->flags |= HAMMER_RECF_REDO;
3282                         bp->b_flags &= ~B_VFSFLAG1;
3283                 }
3284                 if (record->flags & HAMMER_RECF_DEDUPED) {
3285                         bp->b_resid = 0;
3286                         hammer_ip_replace_bulk(hmp, record);
3287                         biodone(ap->a_bio);
3288                 } else {
3289                         hammer_io_direct_write(hmp, bio, record);
3290                 }
3291                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3292                         hammer_flush_inode(ip, 0);
3293         } else {
3294                 bp->b_bio2.bio_offset = NOOFFSET;
3295                 bp->b_error = error;
3296                 bp->b_flags |= B_ERROR;
3297                 biodone(ap->a_bio);
3298         }
3299         lwkt_reltoken(&hmp->fs_token);
3300         return(error);
3301 }
3302
3303 /*
3304  * dounlink - disconnect a directory entry
3305  *
3306  * XXX whiteout support not really in yet
3307  */
3308 static int
3309 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3310                 struct vnode *dvp, struct ucred *cred,
3311                 int flags, int isdir)
3312 {
3313         struct namecache *ncp;
3314         hammer_inode_t dip;
3315         hammer_inode_t ip;
3316         hammer_mount_t hmp;
3317         struct hammer_cursor cursor;
3318         int64_t namekey;
3319         uint32_t max_iterations;
3320         int nlen, error;
3321
3322         /*
3323          * Calculate the namekey and setup the key range for the scan.  This
3324          * works kinda like a chained hash table where the lower 32 bits
3325          * of the namekey synthesize the chain.
3326          *
3327          * The key range is inclusive of both key_beg and key_end.
3328          */
3329         dip = VTOI(dvp);
3330         ncp = nch->ncp;
3331         hmp = dip->hmp;
3332
3333         if (dip->flags & HAMMER_INODE_RO)
3334                 return (EROFS);
3335
3336         namekey = hammer_direntry_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3337                                            &max_iterations);
3338 retry:
3339         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3340         cursor.key_beg.localization = dip->obj_localization |
3341                                       hammer_dir_localization(dip);
3342         cursor.key_beg.obj_id = dip->obj_id;
3343         cursor.key_beg.key = namekey;
3344         cursor.key_beg.create_tid = 0;
3345         cursor.key_beg.delete_tid = 0;
3346         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3347         cursor.key_beg.obj_type = 0;
3348
3349         cursor.key_end = cursor.key_beg;
3350         cursor.key_end.key += max_iterations;
3351         cursor.asof = dip->obj_asof;
3352         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3353
3354         /*
3355          * Scan all matching records (the chain), locate the one matching
3356          * the requested path component.  info->last_error contains the
3357          * error code on search termination and could be 0, ENOENT, or
3358          * something else.
3359          *
3360          * The hammer_ip_*() functions merge in-memory records with on-disk
3361          * records for the purposes of the search.
3362          */
3363         error = hammer_ip_first(&cursor);
3364
3365         while (error == 0) {
3366                 error = hammer_ip_resolve_data(&cursor);
3367                 if (error)
3368                         break;
3369                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3370                 KKASSERT(nlen > 0);
3371                 if (ncp->nc_nlen == nlen &&
3372                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3373                         break;
3374                 }
3375                 error = hammer_ip_next(&cursor);
3376         }
3377
3378         /*
3379          * If all is ok we have to get the inode so we can adjust nlinks.
3380          * To avoid a deadlock with the flusher we must release the inode
3381          * lock on the directory when acquiring the inode for the entry.
3382          *
3383          * If the target is a directory, it must be empty.
3384          */
3385         if (error == 0) {
3386                 hammer_unlock(&cursor.ip->lock);
3387                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3388                                       hmp->asof,
3389                                       cursor.data->entry.localization,
3390                                       0, &error);
3391                 hammer_lock_sh(&cursor.ip->lock);
3392                 if (error == ENOENT) {
3393                         hkprintf("WARNING: Removing dirent w/missing inode "
3394                                 "\"%s\"\n"
3395                                 "\tobj_id = %016jx\n",
3396                                 ncp->nc_name,
3397                                 (intmax_t)cursor.data->entry.obj_id);
3398                         error = 0;
3399                 }
3400
3401                 /*
3402                  * If isdir >= 0 we validate that the entry is or is not a
3403                  * directory.  If isdir < 0 we don't care.
3404                  */
3405                 if (error == 0 && isdir >= 0 && ip) {
3406                         if (isdir &&
3407                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3408                                 error = ENOTDIR;
3409                         } else if (isdir == 0 &&
3410                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3411                                 error = EISDIR;
3412                         }
3413                 }
3414
3415                 /*
3416                  * If we are trying to remove a directory the directory must
3417                  * be empty.
3418                  *
3419                  * The check directory code can loop and deadlock/retry.  Our
3420                  * own cursor's node locks must be released to avoid a 3-way
3421                  * deadlock with the flusher if the check directory code
3422                  * blocks.
3423                  *
3424                  * If any changes whatsoever have been made to the cursor
3425                  * set EDEADLK and retry.
3426                  *
3427                  * WARNING: See warnings in hammer_unlock_cursor()
3428                  *          function.
3429                  */
3430                 if (error == 0 && ip && ip->ino_data.obj_type ==
3431                                         HAMMER_OBJTYPE_DIRECTORY) {
3432                         hammer_unlock_cursor(&cursor);
3433                         error = hammer_ip_check_directory_empty(trans, ip);
3434                         hammer_lock_cursor(&cursor);
3435                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3436                                 hkprintf("Warning: avoided deadlock "
3437                                         "on rmdir '%s'\n",
3438                                         ncp->nc_name);
3439                                 error = EDEADLK;
3440                         }
3441                 }
3442
3443                 /*
3444                  * Delete the directory entry.
3445                  *
3446                  * WARNING: hammer_ip_del_direntry() may have to terminate
3447                  * the cursor to avoid a deadlock.  It is ok to call
3448                  * hammer_done_cursor() twice.
3449                  */
3450                 if (error == 0) {
3451                         error = hammer_ip_del_direntry(trans, &cursor,
3452                                                         dip, ip);
3453                 }
3454                 hammer_done_cursor(&cursor);
3455                 if (error == 0) {
3456                         /*
3457                          * Tell the namecache that we are now unlinked.
3458                          */
3459                         cache_unlink(nch);
3460
3461                         /*
3462                          * NOTE: ip->vp, if non-NULL, cannot be directly
3463                          *       referenced without formally acquiring the
3464                          *       vp since the vp might have zero refs on it,
3465                          *       or in the middle of a reclaim, etc.
3466                          *
3467                          * NOTE: The cache_setunresolved() can rip the vp
3468                          *       out from under us since the vp may not have
3469                          *       any refs, in which case ip->vp will be NULL
3470                          *       from the outset.
3471                          */
3472                         while (ip && ip->vp) {
3473                                 struct vnode *vp;
3474
3475                                 error = hammer_get_vnode(ip, &vp);
3476                                 if (error == 0 && vp) {
3477                                         vn_unlock(vp);
3478                                         hammer_knote(ip->vp, NOTE_DELETE);
3479 #if 0
3480                                         /*
3481                                          * Don't do this, it can deadlock
3482                                          * on concurrent rm's of hardlinks.
3483                                          * Shouldn't be needed any more.
3484                                          */
3485                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3486 #endif
3487                                         vrele(vp);
3488                                         break;
3489                                 }
3490                                 hdkprintf("ip/vp race1 avoided\n");
3491                         }
3492                 }
3493                 if (ip)
3494                         hammer_rel_inode(ip, 0);
3495         } else {
3496                 hammer_done_cursor(&cursor);
3497         }
3498         if (error == EDEADLK)
3499                 goto retry;
3500
3501         return (error);
3502 }
3503
3504 /************************************************************************
3505  *                          FIFO AND SPECFS OPS                         *
3506  ************************************************************************
3507  *
3508  */
3509 static int
3510 hammer_vop_fifoclose (struct vop_close_args *ap)
3511 {
3512         /* XXX update itimes */
3513         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3514 }
3515
3516 static int
3517 hammer_vop_fiforead (struct vop_read_args *ap)
3518 {
3519         int error;
3520
3521         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3522         /* XXX update access time */
3523         return (error);
3524 }
3525
3526 static int
3527 hammer_vop_fifowrite (struct vop_write_args *ap)
3528 {
3529         int error;
3530
3531         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3532         /* XXX update access time */
3533         return (error);
3534 }
3535
3536 static
3537 int
3538 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3539 {
3540         int error;
3541
3542         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3543         if (error)
3544                 error = hammer_vop_kqfilter(ap);
3545         return(error);
3546 }
3547
3548 /************************************************************************
3549  *                          KQFILTER OPS                                *
3550  ************************************************************************
3551  *
3552  */
3553 static void filt_hammerdetach(struct knote *kn);
3554 static int filt_hammerread(struct knote *kn, long hint);
3555 static int filt_hammerwrite(struct knote *kn, long hint);
3556 static int filt_hammervnode(struct knote *kn, long hint);
3557
3558 static struct filterops hammerread_filtops =
3559         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3560           NULL, filt_hammerdetach, filt_hammerread };
3561 static struct filterops hammerwrite_filtops =
3562         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3563           NULL, filt_hammerdetach, filt_hammerwrite };
3564 static struct filterops hammervnode_filtops =
3565         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3566           NULL, filt_hammerdetach, filt_hammervnode };
3567
3568 static
3569 int
3570 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3571 {
3572         struct vnode *vp = ap->a_vp;
3573         struct knote *kn = ap->a_kn;
3574
3575         switch (kn->kn_filter) {
3576         case EVFILT_READ:
3577                 kn->kn_fop = &hammerread_filtops;
3578                 break;
3579         case EVFILT_WRITE:
3580                 kn->kn_fop = &hammerwrite_filtops;
3581                 break;
3582         case EVFILT_VNODE:
3583                 kn->kn_fop = &hammervnode_filtops;
3584                 break;
3585         default:
3586                 return (EOPNOTSUPP);
3587         }
3588
3589         kn->kn_hook = (caddr_t)vp;
3590
3591         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3592
3593         return(0);
3594 }
3595
3596 static void
3597 filt_hammerdetach(struct knote *kn)
3598 {
3599         struct vnode *vp = (void *)kn->kn_hook;
3600
3601         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3602 }
3603
3604 static int
3605 filt_hammerread(struct knote *kn, long hint)
3606 {
3607         struct vnode *vp = (void *)kn->kn_hook;
3608         hammer_inode_t ip = VTOI(vp);
3609         hammer_mount_t hmp = ip->hmp;
3610         off_t off;
3611
3612         if (hint == NOTE_REVOKE) {
3613                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3614                 return(1);
3615         }
3616         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3617         off = ip->ino_data.size - kn->kn_fp->f_offset;
3618         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3619         lwkt_reltoken(&hmp->fs_token);
3620         if (kn->kn_sfflags & NOTE_OLDAPI)
3621                 return(1);
3622         return (kn->kn_data != 0);
3623 }
3624
3625 static int
3626 filt_hammerwrite(struct knote *kn, long hint)
3627 {
3628         if (hint == NOTE_REVOKE)
3629                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3630         kn->kn_data = 0;
3631         return (1);
3632 }
3633
3634 static int
3635 filt_hammervnode(struct knote *kn, long hint)
3636 {
3637         if (kn->kn_sfflags & hint)
3638                 kn->kn_fflags |= hint;
3639         if (hint == NOTE_REVOKE) {
3640                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3641                 return (1);
3642         }
3643         return (kn->kn_fflags != 0);
3644 }
3645