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 = (ip->ino_data.size + HAMMER_XBUFMASK64) &
1059                                 ~HAMMER_XBUFMASK64;
1060         } else if (ip->ino_data.size > HAMMER_HBUFSIZE) {
1061                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
1062                                 ~HAMMER_BUFMASK64;
1063         } else {
1064                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
1065         }
1066
1067         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
1068         vap->va_filerev = 0;    /* XXX */
1069         vap->va_uid_uuid = ip->ino_data.uid;
1070         vap->va_gid_uuid = ip->ino_data.gid;
1071         vap->va_fsid_uuid = ip->hmp->fsid;
1072         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
1073                           VA_FSID_UUID_VALID;
1074
1075         switch (ip->ino_data.obj_type) {
1076         case HAMMER_OBJTYPE_CDEV:
1077         case HAMMER_OBJTYPE_BDEV:
1078                 vap->va_rmajor = ip->ino_data.rmajor;
1079                 vap->va_rminor = ip->ino_data.rminor;
1080                 break;
1081         default:
1082                 break;
1083         }
1084         hammer_unlock(&ip->lock);
1085         return(0);
1086 }
1087
1088 /*
1089  * hammer_vop_nresolve { nch, dvp, cred }
1090  *
1091  * Locate the requested directory entry.
1092  */
1093 static
1094 int
1095 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1096 {
1097         struct hammer_transaction trans;
1098         struct namecache *ncp;
1099         hammer_mount_t hmp;
1100         hammer_inode_t dip;
1101         hammer_inode_t ip;
1102         hammer_tid_t asof;
1103         struct hammer_cursor cursor;
1104         struct vnode *vp;
1105         int64_t namekey;
1106         int error;
1107         int i;
1108         int nlen;
1109         int flags;
1110         int ispfs;
1111         int64_t obj_id;
1112         uint32_t localization;
1113         uint32_t max_iterations;
1114
1115         /*
1116          * Misc initialization, plus handle as-of name extensions.  Look for
1117          * the '@@' extension.  Note that as-of files and directories cannot
1118          * be modified.
1119          */
1120         dip = VTOI(ap->a_dvp);
1121         ncp = ap->a_nch->ncp;
1122         asof = dip->obj_asof;
1123         localization = dip->obj_localization;   /* for code consistency */
1124         nlen = ncp->nc_nlen;
1125         flags = dip->flags & HAMMER_INODE_RO;
1126         ispfs = 0;
1127         hmp = dip->hmp;
1128
1129         lwkt_gettoken(&hmp->fs_token);
1130         hammer_simple_transaction(&trans, hmp);
1131         ++hammer_stats_file_iopsr;
1132
1133         for (i = 0; i < nlen; ++i) {
1134                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1135                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1136                                                   &ispfs, &asof, &localization);
1137                         if (error != 0) {
1138                                 i = nlen;
1139                                 break;
1140                         }
1141                         if (asof != HAMMER_MAX_TID)
1142                                 flags |= HAMMER_INODE_RO;
1143                         break;
1144                 }
1145         }
1146         nlen = i;
1147
1148         /*
1149          * If this is a PFS softlink we dive into the PFS
1150          */
1151         if (ispfs && nlen == 0) {
1152                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1153                                       asof, localization,
1154                                       flags, &error);
1155                 if (error == 0) {
1156                         error = hammer_get_vnode(ip, &vp);
1157                         hammer_rel_inode(ip, 0);
1158                 } else {
1159                         vp = NULL;
1160                 }
1161                 if (error == 0) {
1162                         vn_unlock(vp);
1163                         cache_setvp(ap->a_nch, vp);
1164                         vrele(vp);
1165                 }
1166                 goto done;
1167         }
1168
1169         /*
1170          * If there is no path component the time extension is relative to dip.
1171          * e.g. "fubar/@@<snapshot>"
1172          *
1173          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1174          * e.g. "fubar/.@@<snapshot>"
1175          *
1176          * ".." is handled by the kernel.  We do not currently handle
1177          * "..@<snapshot>".
1178          */
1179         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1180                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1181                                       asof, dip->obj_localization,
1182                                       flags, &error);
1183                 if (error == 0) {
1184                         error = hammer_get_vnode(ip, &vp);
1185                         hammer_rel_inode(ip, 0);
1186                 } else {
1187                         vp = NULL;
1188                 }
1189                 if (error == 0) {
1190                         vn_unlock(vp);
1191                         cache_setvp(ap->a_nch, vp);
1192                         vrele(vp);
1193                 }
1194                 goto done;
1195         }
1196
1197         /*
1198          * Calculate the namekey and setup the key range for the scan.  This
1199          * works kinda like a chained hash table where the lower 32 bits
1200          * of the namekey synthesize the chain.
1201          *
1202          * The key range is inclusive of both key_beg and key_end.
1203          */
1204         namekey = hammer_direntry_namekey(dip, ncp->nc_name, nlen,
1205                                            &max_iterations);
1206
1207         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1208         cursor.key_beg.localization = dip->obj_localization |
1209                                       hammer_dir_localization(dip);
1210         cursor.key_beg.obj_id = dip->obj_id;
1211         cursor.key_beg.key = namekey;
1212         cursor.key_beg.create_tid = 0;
1213         cursor.key_beg.delete_tid = 0;
1214         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1215         cursor.key_beg.obj_type = 0;
1216
1217         cursor.key_end = cursor.key_beg;
1218         cursor.key_end.key += max_iterations;
1219         cursor.asof = asof;
1220         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1221
1222         /*
1223          * Scan all matching records (the chain), locate the one matching
1224          * the requested path component.
1225          *
1226          * The hammer_ip_*() functions merge in-memory records with on-disk
1227          * records for the purposes of the search.
1228          */
1229         obj_id = 0;
1230         localization = HAMMER_DEF_LOCALIZATION;
1231
1232         if (error == 0) {
1233                 error = hammer_ip_first(&cursor);
1234                 while (error == 0) {
1235                         error = hammer_ip_resolve_data(&cursor);
1236                         if (error)
1237                                 break;
1238                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1239                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1240                                 obj_id = cursor.data->entry.obj_id;
1241                                 localization = cursor.data->entry.localization;
1242                                 break;
1243                         }
1244                         error = hammer_ip_next(&cursor);
1245                 }
1246         }
1247         hammer_done_cursor(&cursor);
1248
1249         /*
1250          * Lookup the obj_id.  This should always succeed.  If it does not
1251          * the filesystem may be damaged and we return a dummy inode.
1252          */
1253         if (error == 0) {
1254                 ip = hammer_get_inode(&trans, dip, obj_id,
1255                                       asof, localization,
1256                                       flags, &error);
1257                 if (error == ENOENT) {
1258                         hkprintf("WARNING: Missing inode for dirent \"%s\"\n"
1259                                 "\tobj_id = %016jx, asof=%016jx, lo=%08x\n",
1260                                 ncp->nc_name,
1261                                 (intmax_t)obj_id, (intmax_t)asof,
1262                                 localization);
1263                         error = 0;
1264                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1265                                                     asof, localization,
1266                                                     flags, &error);
1267                 }
1268                 if (error == 0) {
1269                         error = hammer_get_vnode(ip, &vp);
1270                         hammer_rel_inode(ip, 0);
1271                 } else {
1272                         vp = NULL;
1273                 }
1274                 if (error == 0) {
1275                         vn_unlock(vp);
1276                         cache_setvp(ap->a_nch, vp);
1277                         vrele(vp);
1278                 }
1279         } else if (error == ENOENT) {
1280                 cache_setvp(ap->a_nch, NULL);
1281         }
1282 done:
1283         hammer_done_transaction(&trans);
1284         lwkt_reltoken(&hmp->fs_token);
1285         return (error);
1286 }
1287
1288 /*
1289  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1290  *
1291  * Locate the parent directory of a directory vnode.
1292  *
1293  * dvp is referenced but not locked.  *vpp must be returned referenced and
1294  * locked.  A parent_obj_id of 0 indicates that we are at the root.
1295  *
1296  * NOTE: as-of sequences are not linked into the directory structure.  If
1297  * we are at the root with a different asof then the mount point, reload
1298  * the same directory with the mount point's asof.   I'm not sure what this
1299  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1300  * get confused, but it hasn't been tested.
1301  */
1302 static
1303 int
1304 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1305 {
1306         struct hammer_transaction trans;
1307         hammer_inode_t dip;
1308         hammer_inode_t ip;
1309         hammer_mount_t hmp;
1310         int64_t parent_obj_id;
1311         uint32_t parent_obj_localization;
1312         hammer_tid_t asof;
1313         int error;
1314
1315         dip = VTOI(ap->a_dvp);
1316         asof = dip->obj_asof;
1317         hmp = dip->hmp;
1318
1319         /*
1320          * Whos are parent?  This could be the root of a pseudo-filesystem
1321          * whos parent is in another localization domain.
1322          */
1323         lwkt_gettoken(&hmp->fs_token);
1324         parent_obj_id = dip->ino_data.parent_obj_id;
1325         if (dip->obj_id == HAMMER_OBJID_ROOT)
1326                 parent_obj_localization = HAMMER_DEF_LOCALIZATION;
1327         else
1328                 parent_obj_localization = dip->obj_localization;
1329
1330         /*
1331          * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1332          */
1333         if (parent_obj_id == 0) {
1334                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1335                    asof != hmp->asof) {
1336                         parent_obj_id = dip->obj_id;
1337                         asof = hmp->asof;
1338                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1339                         ksnprintf(*ap->a_fakename, 19, "0x%016jx",
1340                                   (intmax_t)dip->obj_asof);
1341                 } else {
1342                         *ap->a_vpp = NULL;
1343                         lwkt_reltoken(&hmp->fs_token);
1344                         return ENOENT;
1345                 }
1346         }
1347
1348         hammer_simple_transaction(&trans, hmp);
1349         ++hammer_stats_file_iopsr;
1350
1351         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1352                               asof, parent_obj_localization,
1353                               dip->flags, &error);
1354         if (ip) {
1355                 error = hammer_get_vnode(ip, ap->a_vpp);
1356                 hammer_rel_inode(ip, 0);
1357         } else {
1358                 *ap->a_vpp = NULL;
1359         }
1360         hammer_done_transaction(&trans);
1361         lwkt_reltoken(&hmp->fs_token);
1362         return (error);
1363 }
1364
1365 /*
1366  * hammer_vop_nlink { nch, dvp, vp, cred }
1367  */
1368 static
1369 int
1370 hammer_vop_nlink(struct vop_nlink_args *ap)
1371 {
1372         struct hammer_transaction trans;
1373         hammer_inode_t dip;
1374         hammer_inode_t ip;
1375         struct nchandle *nch;
1376         hammer_mount_t hmp;
1377         int error;
1378
1379         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1380                 return(EXDEV);
1381
1382         nch = ap->a_nch;
1383         dip = VTOI(ap->a_dvp);
1384         ip = VTOI(ap->a_vp);
1385         hmp = dip->hmp;
1386
1387         if (dip->obj_localization != ip->obj_localization)
1388                 return(EXDEV);
1389
1390         if (dip->flags & HAMMER_INODE_RO)
1391                 return (EROFS);
1392         if (ip->flags & HAMMER_INODE_RO)
1393                 return (EROFS);
1394         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1395                 return (error);
1396
1397         /*
1398          * Create a transaction to cover the operations we perform.
1399          */
1400         lwkt_gettoken(&hmp->fs_token);
1401         hammer_start_transaction(&trans, hmp);
1402         ++hammer_stats_file_iopsw;
1403
1404         /*
1405          * Add the filesystem object to the directory.  Note that neither
1406          * dip nor ip are referenced or locked, but their vnodes are
1407          * referenced.  This function will bump the inode's link count.
1408          */
1409         error = hammer_ip_add_direntry(&trans, dip,
1410                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1411                                         ip);
1412
1413         /*
1414          * Finish up.
1415          */
1416         if (error == 0) {
1417                 cache_setunresolved(nch);
1418                 cache_setvp(nch, ap->a_vp);
1419         }
1420         hammer_done_transaction(&trans);
1421         hammer_knote(ap->a_vp, NOTE_LINK);
1422         hammer_knote(ap->a_dvp, NOTE_WRITE);
1423         lwkt_reltoken(&hmp->fs_token);
1424         return (error);
1425 }
1426
1427 /*
1428  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1429  *
1430  * The operating system has already ensured that the directory entry
1431  * does not exist and done all appropriate namespace locking.
1432  */
1433 static
1434 int
1435 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1436 {
1437         struct hammer_transaction trans;
1438         hammer_inode_t dip;
1439         hammer_inode_t nip;
1440         struct nchandle *nch;
1441         hammer_mount_t hmp;
1442         int error;
1443
1444         nch = ap->a_nch;
1445         dip = VTOI(ap->a_dvp);
1446         hmp = dip->hmp;
1447
1448         if (dip->flags & HAMMER_INODE_RO)
1449                 return (EROFS);
1450         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1451                 return (error);
1452
1453         /*
1454          * Create a transaction to cover the operations we perform.
1455          */
1456         lwkt_gettoken(&hmp->fs_token);
1457         hammer_start_transaction(&trans, hmp);
1458         ++hammer_stats_file_iopsw;
1459
1460         /*
1461          * Create a new filesystem object of the requested type.  The
1462          * returned inode will be referenced but not locked.
1463          */
1464         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1465                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1466                                     NULL, &nip);
1467         if (error) {
1468                 hammer_done_transaction(&trans);
1469                 *ap->a_vpp = NULL;
1470                 lwkt_reltoken(&hmp->fs_token);
1471                 return (error);
1472         }
1473         /*
1474          * Add the new filesystem object to the directory.  This will also
1475          * bump the inode's link count.
1476          */
1477         error = hammer_ip_add_direntry(&trans, dip,
1478                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1479                                         nip);
1480         if (error)
1481                 hkprintf("hammer_mkdir (add) error %d\n", error);
1482
1483         /*
1484          * Finish up.
1485          */
1486         if (error) {
1487                 hammer_rel_inode(nip, 0);
1488                 *ap->a_vpp = NULL;
1489         } else {
1490                 error = hammer_get_vnode(nip, ap->a_vpp);
1491                 hammer_rel_inode(nip, 0);
1492                 if (error == 0) {
1493                         cache_setunresolved(ap->a_nch);
1494                         cache_setvp(ap->a_nch, *ap->a_vpp);
1495                 }
1496         }
1497         hammer_done_transaction(&trans);
1498         if (error == 0)
1499                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1500         lwkt_reltoken(&hmp->fs_token);
1501         return (error);
1502 }
1503
1504 /*
1505  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1506  *
1507  * The operating system has already ensured that the directory entry
1508  * does not exist and done all appropriate namespace locking.
1509  */
1510 static
1511 int
1512 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1513 {
1514         struct hammer_transaction trans;
1515         hammer_inode_t dip;
1516         hammer_inode_t nip;
1517         struct nchandle *nch;
1518         hammer_mount_t hmp;
1519         int error;
1520
1521         nch = ap->a_nch;
1522         dip = VTOI(ap->a_dvp);
1523         hmp = dip->hmp;
1524
1525         if (dip->flags & HAMMER_INODE_RO)
1526                 return (EROFS);
1527         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1528                 return (error);
1529
1530         /*
1531          * Create a transaction to cover the operations we perform.
1532          */
1533         lwkt_gettoken(&hmp->fs_token);
1534         hammer_start_transaction(&trans, hmp);
1535         ++hammer_stats_file_iopsw;
1536
1537         /*
1538          * Create a new filesystem object of the requested type.  The
1539          * returned inode will be referenced but not locked.
1540          *
1541          * If mknod specifies a directory a pseudo-fs is created.
1542          */
1543         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1544                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1545                                     NULL, &nip);
1546         if (error) {
1547                 hammer_done_transaction(&trans);
1548                 *ap->a_vpp = NULL;
1549                 lwkt_reltoken(&hmp->fs_token);
1550                 return (error);
1551         }
1552
1553         /*
1554          * Add the new filesystem object to the directory.  This will also
1555          * bump the inode's link count.
1556          */
1557         error = hammer_ip_add_direntry(&trans, dip,
1558                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1559                                         nip);
1560
1561         /*
1562          * Finish up.
1563          */
1564         if (error) {
1565                 hammer_rel_inode(nip, 0);
1566                 *ap->a_vpp = NULL;
1567         } else {
1568                 error = hammer_get_vnode(nip, ap->a_vpp);
1569                 hammer_rel_inode(nip, 0);
1570                 if (error == 0) {
1571                         cache_setunresolved(ap->a_nch);
1572                         cache_setvp(ap->a_nch, *ap->a_vpp);
1573                 }
1574         }
1575         hammer_done_transaction(&trans);
1576         if (error == 0)
1577                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1578         lwkt_reltoken(&hmp->fs_token);
1579         return (error);
1580 }
1581
1582 /*
1583  * hammer_vop_open { vp, mode, cred, fp }
1584  *
1585  * MPSAFE (does not require fs_token)
1586  */
1587 static
1588 int
1589 hammer_vop_open(struct vop_open_args *ap)
1590 {
1591         hammer_inode_t ip;
1592
1593         ++hammer_stats_file_iopsr;
1594         ip = VTOI(ap->a_vp);
1595
1596         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1597                 return (EROFS);
1598         return(vop_stdopen(ap));
1599 }
1600
1601 /*
1602  * hammer_vop_print { vp }
1603  */
1604 static
1605 int
1606 hammer_vop_print(struct vop_print_args *ap)
1607 {
1608         return EOPNOTSUPP;
1609 }
1610
1611 /*
1612  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1613  */
1614 static
1615 int
1616 hammer_vop_readdir(struct vop_readdir_args *ap)
1617 {
1618         struct hammer_transaction trans;
1619         struct hammer_cursor cursor;
1620         hammer_inode_t ip;
1621         hammer_mount_t hmp;
1622         struct uio *uio;
1623         hammer_base_elm_t base;
1624         int error;
1625         int cookie_index;
1626         int ncookies;
1627         off_t *cookies;
1628         off_t saveoff;
1629         int r;
1630         int dtype;
1631
1632         ++hammer_stats_file_iopsr;
1633         ip = VTOI(ap->a_vp);
1634         uio = ap->a_uio;
1635         saveoff = uio->uio_offset;
1636         hmp = ip->hmp;
1637
1638         if (ap->a_ncookies) {
1639                 ncookies = uio->uio_resid / 16 + 1;
1640                 if (ncookies > 1024)
1641                         ncookies = 1024;
1642                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1643                 cookie_index = 0;
1644         } else {
1645                 ncookies = -1;
1646                 cookies = NULL;
1647                 cookie_index = 0;
1648         }
1649
1650         lwkt_gettoken(&hmp->fs_token);
1651         hammer_simple_transaction(&trans, hmp);
1652
1653         /*
1654          * Handle artificial entries
1655          *
1656          * It should be noted that the minimum value for a directory
1657          * hash key on-media is 0x0000000100000000, so we can use anything
1658          * less then that to represent our 'special' key space.
1659          */
1660         error = 0;
1661         if (saveoff == 0) {
1662                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1663                 if (r)
1664                         goto done;
1665                 if (cookies)
1666                         cookies[cookie_index] = saveoff;
1667                 ++saveoff;
1668                 ++cookie_index;
1669                 if (cookie_index == ncookies)
1670                         goto done;
1671         }
1672         if (saveoff == 1) {
1673                 if (ip->ino_data.parent_obj_id) {
1674                         r = vop_write_dirent(&error, uio,
1675                                              ip->ino_data.parent_obj_id,
1676                                              DT_DIR, 2, "..");
1677                 } else {
1678                         r = vop_write_dirent(&error, uio,
1679                                              ip->obj_id, DT_DIR, 2, "..");
1680                 }
1681                 if (r)
1682                         goto done;
1683                 if (cookies)
1684                         cookies[cookie_index] = saveoff;
1685                 ++saveoff;
1686                 ++cookie_index;
1687                 if (cookie_index == ncookies)
1688                         goto done;
1689         }
1690
1691         /*
1692          * Key range (begin and end inclusive) to scan.  Directory keys
1693          * directly translate to a 64 bit 'seek' position.
1694          */
1695         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1696         cursor.key_beg.localization = ip->obj_localization |
1697                                       hammer_dir_localization(ip);
1698         cursor.key_beg.obj_id = ip->obj_id;
1699         cursor.key_beg.create_tid = 0;
1700         cursor.key_beg.delete_tid = 0;
1701         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1702         cursor.key_beg.obj_type = 0;
1703         cursor.key_beg.key = saveoff;
1704
1705         cursor.key_end = cursor.key_beg;
1706         cursor.key_end.key = HAMMER_MAX_KEY;
1707         cursor.asof = ip->obj_asof;
1708         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1709
1710         error = hammer_ip_first(&cursor);
1711
1712         while (error == 0) {
1713                 error = hammer_ip_resolve_data(&cursor);
1714                 if (error)
1715                         break;
1716                 base = &cursor.leaf->base;
1717                 saveoff = base->key;
1718                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1719
1720                 if (base->obj_id != ip->obj_id)
1721                         hpanic("bad record at %p", cursor.node);
1722
1723                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1724                 r = vop_write_dirent(
1725                              &error, uio, cursor.data->entry.obj_id,
1726                              dtype,
1727                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1728                              (void *)cursor.data->entry.name);
1729                 if (r)
1730                         break;
1731                 ++saveoff;
1732                 if (cookies)
1733                         cookies[cookie_index] = base->key;
1734                 ++cookie_index;
1735                 if (cookie_index == ncookies)
1736                         break;
1737                 error = hammer_ip_next(&cursor);
1738         }
1739         hammer_done_cursor(&cursor);
1740
1741 done:
1742         hammer_done_transaction(&trans);
1743
1744         if (ap->a_eofflag)
1745                 *ap->a_eofflag = (error == ENOENT);
1746         uio->uio_offset = saveoff;
1747         if (error && cookie_index == 0) {
1748                 if (error == ENOENT)
1749                         error = 0;
1750                 if (cookies) {
1751                         kfree(cookies, M_TEMP);
1752                         *ap->a_ncookies = 0;
1753                         *ap->a_cookies = NULL;
1754                 }
1755         } else {
1756                 if (error == ENOENT)
1757                         error = 0;
1758                 if (cookies) {
1759                         *ap->a_ncookies = cookie_index;
1760                         *ap->a_cookies = cookies;
1761                 }
1762         }
1763         lwkt_reltoken(&hmp->fs_token);
1764         return(error);
1765 }
1766
1767 /*
1768  * hammer_vop_readlink { vp, uio, cred }
1769  */
1770 static
1771 int
1772 hammer_vop_readlink(struct vop_readlink_args *ap)
1773 {
1774         struct hammer_transaction trans;
1775         struct hammer_cursor cursor;
1776         hammer_inode_t ip;
1777         hammer_mount_t hmp;
1778         char buf[32];
1779         uint32_t localization;
1780         hammer_pseudofs_inmem_t pfsm;
1781         int error;
1782
1783         ip = VTOI(ap->a_vp);
1784         hmp = ip->hmp;
1785
1786         lwkt_gettoken(&hmp->fs_token);
1787
1788         /*
1789          * Shortcut if the symlink data was stuffed into ino_data.
1790          *
1791          * Also expand special "@@PFS%05d" softlinks (expansion only
1792          * occurs for non-historical (current) accesses made from the
1793          * primary filesystem).
1794          *
1795          * Note that userspace hammer command does not allow users to
1796          * create a @@PFS softlink under an existing other PFS (id!=0)
1797          * so the ip localization here for @@PFS softlink is always 0.
1798          */
1799         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1800                 char *ptr;
1801                 int bytes;
1802
1803                 ptr = ip->ino_data.ext.symlink;
1804                 bytes = (int)ip->ino_data.size;
1805                 if (bytes == 10 &&
1806                     ip->obj_asof == HAMMER_MAX_TID &&
1807                     ip->obj_localization == HAMMER_DEF_LOCALIZATION &&
1808                     strncmp(ptr, "@@PFS", 5) == 0) {
1809                         hammer_simple_transaction(&trans, hmp);
1810                         bcopy(ptr + 5, buf, 5);
1811                         buf[5] = 0;
1812                         localization = pfs_to_lo(strtoul(buf, NULL, 10));
1813                         pfsm = hammer_load_pseudofs(&trans, localization,
1814                                                     &error);
1815                         if (error == 0) {
1816                                 if (hammer_is_pfs_slave(&pfsm->pfsd)) {
1817                                         /* vap->va_size == 26 */
1818                                         ksnprintf(buf, sizeof(buf),
1819                                                   "@@0x%016jx:%05d",
1820                                                   (intmax_t)pfsm->pfsd.sync_end_tid,
1821                                                   lo_to_pfs(localization));
1822                                 } else {
1823                                         /* vap->va_size == 10 */
1824                                         ksnprintf(buf, sizeof(buf),
1825                                                   "@@-1:%05d",
1826                                                   lo_to_pfs(localization));
1827                                 }
1828                                 ptr = buf;
1829                                 bytes = strlen(buf);
1830                         }
1831                         if (pfsm)
1832                                 hammer_rel_pseudofs(hmp, pfsm);
1833                         hammer_done_transaction(&trans);
1834                 }
1835                 error = uiomove(ptr, bytes, ap->a_uio);
1836                 lwkt_reltoken(&hmp->fs_token);
1837                 return(error);
1838         }
1839
1840         /*
1841          * Long version
1842          */
1843         hammer_simple_transaction(&trans, hmp);
1844         ++hammer_stats_file_iopsr;
1845         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1846
1847         /*
1848          * Key range (begin and end inclusive) to scan.  Directory keys
1849          * directly translate to a 64 bit 'seek' position.
1850          */
1851         cursor.key_beg.localization = ip->obj_localization |
1852                                       HAMMER_LOCALIZE_MISC;
1853         cursor.key_beg.obj_id = ip->obj_id;
1854         cursor.key_beg.create_tid = 0;
1855         cursor.key_beg.delete_tid = 0;
1856         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1857         cursor.key_beg.obj_type = 0;
1858         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1859         cursor.asof = ip->obj_asof;
1860         cursor.flags |= HAMMER_CURSOR_ASOF;
1861
1862         error = hammer_ip_lookup(&cursor);
1863         if (error == 0) {
1864                 error = hammer_ip_resolve_data(&cursor);
1865                 if (error == 0) {
1866                         KKASSERT(cursor.leaf->data_len >=
1867                                  HAMMER_SYMLINK_NAME_OFF);
1868                         error = uiomove(cursor.data->symlink.name,
1869                                         cursor.leaf->data_len -
1870                                                 HAMMER_SYMLINK_NAME_OFF,
1871                                         ap->a_uio);
1872                 }
1873         }
1874         hammer_done_cursor(&cursor);
1875         hammer_done_transaction(&trans);
1876         lwkt_reltoken(&hmp->fs_token);
1877         return(error);
1878 }
1879
1880 /*
1881  * hammer_vop_nremove { nch, dvp, cred }
1882  */
1883 static
1884 int
1885 hammer_vop_nremove(struct vop_nremove_args *ap)
1886 {
1887         struct hammer_transaction trans;
1888         hammer_inode_t dip;
1889         hammer_mount_t hmp;
1890         int error;
1891
1892         dip = VTOI(ap->a_dvp);
1893         hmp = dip->hmp;
1894
1895         if (hammer_nohistory(dip) == 0 &&
1896             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1897                 return (error);
1898         }
1899
1900         lwkt_gettoken(&hmp->fs_token);
1901         hammer_start_transaction(&trans, hmp);
1902         ++hammer_stats_file_iopsw;
1903         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1904         hammer_done_transaction(&trans);
1905         if (error == 0)
1906                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1907         lwkt_reltoken(&hmp->fs_token);
1908         return (error);
1909 }
1910
1911 /*
1912  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1913  */
1914 static
1915 int
1916 hammer_vop_nrename(struct vop_nrename_args *ap)
1917 {
1918         struct hammer_transaction trans;
1919         struct namecache *fncp;
1920         struct namecache *tncp;
1921         hammer_inode_t fdip;
1922         hammer_inode_t tdip;
1923         hammer_inode_t ip;
1924         hammer_mount_t hmp;
1925         struct hammer_cursor cursor;
1926         int64_t namekey;
1927         uint32_t max_iterations;
1928         int nlen, error;
1929
1930         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1931                 return(EXDEV);
1932         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1933                 return(EXDEV);
1934
1935         fdip = VTOI(ap->a_fdvp);
1936         tdip = VTOI(ap->a_tdvp);
1937         fncp = ap->a_fnch->ncp;
1938         tncp = ap->a_tnch->ncp;
1939         ip = VTOI(fncp->nc_vp);
1940         KKASSERT(ip != NULL);
1941
1942         hmp = ip->hmp;
1943
1944         if (fdip->obj_localization != tdip->obj_localization)
1945                 return(EXDEV);
1946         if (fdip->obj_localization != ip->obj_localization)
1947                 return(EXDEV);
1948
1949         if (fdip->flags & HAMMER_INODE_RO)
1950                 return (EROFS);
1951         if (tdip->flags & HAMMER_INODE_RO)
1952                 return (EROFS);
1953         if (ip->flags & HAMMER_INODE_RO)
1954                 return (EROFS);
1955         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
1956                 return (error);
1957
1958         lwkt_gettoken(&hmp->fs_token);
1959         hammer_start_transaction(&trans, hmp);
1960         ++hammer_stats_file_iopsw;
1961
1962         /*
1963          * Remove tncp from the target directory and then link ip as
1964          * tncp. XXX pass trans to dounlink
1965          *
1966          * Force the inode sync-time to match the transaction so it is
1967          * in-sync with the creation of the target directory entry.
1968          */
1969         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1970                                 ap->a_cred, 0, -1);
1971         if (error == 0 || error == ENOENT) {
1972                 error = hammer_ip_add_direntry(&trans, tdip,
1973                                                 tncp->nc_name, tncp->nc_nlen,
1974                                                 ip);
1975                 if (error == 0) {
1976                         ip->ino_data.parent_obj_id = tdip->obj_id;
1977                         ip->ino_data.ctime = trans.time;
1978                         hammer_modify_inode(&trans, ip, HAMMER_INODE_DDIRTY);
1979                 }
1980         }
1981         if (error)
1982                 goto failed; /* XXX */
1983
1984         /*
1985          * Locate the record in the originating directory and remove it.
1986          *
1987          * Calculate the namekey and setup the key range for the scan.  This
1988          * works kinda like a chained hash table where the lower 32 bits
1989          * of the namekey synthesize the chain.
1990          *
1991          * The key range is inclusive of both key_beg and key_end.
1992          */
1993         namekey = hammer_direntry_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1994                                            &max_iterations);
1995 retry:
1996         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1997         cursor.key_beg.localization = fdip->obj_localization |
1998                                       hammer_dir_localization(fdip);
1999         cursor.key_beg.obj_id = fdip->obj_id;
2000         cursor.key_beg.key = namekey;
2001         cursor.key_beg.create_tid = 0;
2002         cursor.key_beg.delete_tid = 0;
2003         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2004         cursor.key_beg.obj_type = 0;
2005
2006         cursor.key_end = cursor.key_beg;
2007         cursor.key_end.key += max_iterations;
2008         cursor.asof = fdip->obj_asof;
2009         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2010
2011         /*
2012          * Scan all matching records (the chain), locate the one matching
2013          * the requested path component.
2014          *
2015          * The hammer_ip_*() functions merge in-memory records with on-disk
2016          * records for the purposes of the search.
2017          */
2018         error = hammer_ip_first(&cursor);
2019         while (error == 0) {
2020                 if (hammer_ip_resolve_data(&cursor) != 0)
2021                         break;
2022                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2023                 KKASSERT(nlen > 0);
2024                 if (fncp->nc_nlen == nlen &&
2025                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2026                         break;
2027                 }
2028                 error = hammer_ip_next(&cursor);
2029         }
2030
2031         /*
2032          * If all is ok we have to get the inode so we can adjust nlinks.
2033          *
2034          * WARNING: hammer_ip_del_direntry() may have to terminate the
2035          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
2036          * twice.
2037          */
2038         if (error == 0)
2039                 error = hammer_ip_del_direntry(&trans, &cursor, fdip, ip);
2040
2041         /*
2042          * XXX A deadlock here will break rename's atomicy for the purposes
2043          * of crash recovery.
2044          */
2045         if (error == EDEADLK) {
2046                 hammer_done_cursor(&cursor);
2047                 goto retry;
2048         }
2049
2050         /*
2051          * Cleanup and tell the kernel that the rename succeeded.
2052          *
2053          * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2054          *       without formally acquiring the vp since the vp might
2055          *       have zero refs on it, or in the middle of a reclaim,
2056          *       etc.
2057          */
2058         hammer_done_cursor(&cursor);
2059         if (error == 0) {
2060                 cache_rename(ap->a_fnch, ap->a_tnch);
2061                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
2062                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
2063                 while (ip->vp) {
2064                         struct vnode *vp;
2065
2066                         error = hammer_get_vnode(ip, &vp);
2067                         if (error == 0 && vp) {
2068                                 vn_unlock(vp);
2069                                 hammer_knote(ip->vp, NOTE_RENAME);
2070                                 vrele(vp);
2071                                 break;
2072                         }
2073                         hdkprintf("ip/vp race2 avoided\n");
2074                 }
2075         }
2076
2077 failed:
2078         hammer_done_transaction(&trans);
2079         lwkt_reltoken(&hmp->fs_token);
2080         return (error);
2081 }
2082
2083 /*
2084  * hammer_vop_nrmdir { nch, dvp, cred }
2085  */
2086 static
2087 int
2088 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
2089 {
2090         struct hammer_transaction trans;
2091         hammer_inode_t dip;
2092         hammer_mount_t hmp;
2093         int error;
2094
2095         dip = VTOI(ap->a_dvp);
2096         hmp = dip->hmp;
2097
2098         if (hammer_nohistory(dip) == 0 &&
2099             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2100                 return (error);
2101         }
2102
2103         lwkt_gettoken(&hmp->fs_token);
2104         hammer_start_transaction(&trans, hmp);
2105         ++hammer_stats_file_iopsw;
2106         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
2107         hammer_done_transaction(&trans);
2108         if (error == 0)
2109                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
2110         lwkt_reltoken(&hmp->fs_token);
2111         return (error);
2112 }
2113
2114 /*
2115  * hammer_vop_markatime { vp, cred }
2116  */
2117 static
2118 int
2119 hammer_vop_markatime(struct vop_markatime_args *ap)
2120 {
2121         struct hammer_transaction trans;
2122         hammer_inode_t ip;
2123         hammer_mount_t hmp;
2124
2125         ip = VTOI(ap->a_vp);
2126         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2127                 return (EROFS);
2128         if (ip->flags & HAMMER_INODE_RO)
2129                 return (EROFS);
2130         hmp = ip->hmp;
2131         if (hmp->mp->mnt_flag & MNT_NOATIME)
2132                 return (0);
2133         lwkt_gettoken(&hmp->fs_token);
2134         hammer_start_transaction(&trans, hmp);
2135         ++hammer_stats_file_iopsw;
2136
2137         ip->ino_data.atime = trans.time;
2138         hammer_modify_inode(&trans, ip, HAMMER_INODE_ATIME);
2139         hammer_done_transaction(&trans);
2140         hammer_knote(ap->a_vp, NOTE_ATTRIB);
2141         lwkt_reltoken(&hmp->fs_token);
2142         return (0);
2143 }
2144
2145 /*
2146  * hammer_vop_setattr { vp, vap, cred }
2147  */
2148 static
2149 int
2150 hammer_vop_setattr(struct vop_setattr_args *ap)
2151 {
2152         struct hammer_transaction trans;
2153         hammer_inode_t ip;
2154         struct vattr *vap;
2155         hammer_mount_t hmp;
2156         int modflags;
2157         int error;
2158         int truncating;
2159         int blksize;
2160         int kflags;
2161 #if 0
2162         int64_t aligned_size;
2163 #endif
2164         uint32_t flags;
2165
2166         vap = ap->a_vap;
2167         ip = ap->a_vp->v_data;
2168         modflags = 0;
2169         kflags = 0;
2170         hmp = ip->hmp;
2171
2172         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2173                 return(EROFS);
2174         if (ip->flags & HAMMER_INODE_RO)
2175                 return (EROFS);
2176         if (hammer_nohistory(ip) == 0 &&
2177             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2178                 return (error);
2179         }
2180
2181         lwkt_gettoken(&hmp->fs_token);
2182         hammer_start_transaction(&trans, hmp);
2183         ++hammer_stats_file_iopsw;
2184         error = 0;
2185
2186         if (vap->va_flags != VNOVAL) {
2187                 flags = ip->ino_data.uflags;
2188                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2189                                          hammer_to_unix_xid(&ip->ino_data.uid),
2190                                          ap->a_cred);
2191                 if (error == 0) {
2192                         if (ip->ino_data.uflags != flags) {
2193                                 ip->ino_data.uflags = flags;
2194                                 ip->ino_data.ctime = trans.time;
2195                                 modflags |= HAMMER_INODE_DDIRTY;
2196                                 kflags |= NOTE_ATTRIB;
2197                         }
2198                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2199                                 error = 0;
2200                                 goto done;
2201                         }
2202                 }
2203                 goto done;
2204         }
2205         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2206                 error = EPERM;
2207                 goto done;
2208         }
2209         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2210                 mode_t cur_mode = ip->ino_data.mode;
2211                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2212                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2213                 uuid_t uuid_uid;
2214                 uuid_t uuid_gid;
2215
2216                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2217                                          ap->a_cred,
2218                                          &cur_uid, &cur_gid, &cur_mode);
2219                 if (error == 0) {
2220                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2221                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2222                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2223                                  sizeof(uuid_uid)) ||
2224                             bcmp(&uuid_gid, &ip->ino_data.gid,
2225                                  sizeof(uuid_gid)) ||
2226                             ip->ino_data.mode != cur_mode) {
2227                                 ip->ino_data.uid = uuid_uid;
2228                                 ip->ino_data.gid = uuid_gid;
2229                                 ip->ino_data.mode = cur_mode;
2230                                 ip->ino_data.ctime = trans.time;
2231                                 modflags |= HAMMER_INODE_DDIRTY;
2232                         }
2233                         kflags |= NOTE_ATTRIB;
2234                 }
2235         }
2236         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2237                 switch(ap->a_vp->v_type) {
2238                 case VREG:
2239                         if (vap->va_size == ip->ino_data.size)
2240                                 break;
2241
2242                         /*
2243                          * Log the operation if in fast-fsync mode or if
2244                          * there are unterminated redo write records present.
2245                          *
2246                          * The second check is needed so the recovery code
2247                          * properly truncates write redos even if nominal
2248                          * REDO operations is turned off due to excessive
2249                          * writes, because the related records might be
2250                          * destroyed and never lay down a TERM_WRITE.
2251                          */
2252                         if ((ip->flags & HAMMER_INODE_REDO) ||
2253                             (ip->flags & HAMMER_INODE_RDIRTY)) {
2254                                 error = hammer_generate_redo(&trans, ip,
2255                                                              vap->va_size,
2256                                                              HAMMER_REDO_TRUNC,
2257                                                              NULL, 0);
2258                         }
2259                         blksize = hammer_blocksize(vap->va_size);
2260
2261                         /*
2262                          * XXX break atomicy, we can deadlock the backend
2263                          * if we do not release the lock.  Probably not a
2264                          * big deal here.
2265                          */
2266                         if (vap->va_size < ip->ino_data.size) {
2267                                 nvtruncbuf(ap->a_vp, vap->va_size,
2268                                            blksize,
2269                                            hammer_blockoff(vap->va_size),
2270                                            0);
2271                                 truncating = 1;
2272                                 kflags |= NOTE_WRITE;
2273                         } else {
2274                                 nvextendbuf(ap->a_vp,
2275                                             ip->ino_data.size,
2276                                             vap->va_size,
2277                                             hammer_blocksize(ip->ino_data.size),
2278                                             hammer_blocksize(vap->va_size),
2279                                             hammer_blockoff(ip->ino_data.size),
2280                                             hammer_blockoff(vap->va_size),
2281                                             0);
2282                                 truncating = 0;
2283                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2284                         }
2285                         ip->ino_data.size = vap->va_size;
2286                         ip->ino_data.mtime = trans.time;
2287                         /* XXX safe to use SDIRTY instead of DDIRTY here? */
2288                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2289
2290                         /*
2291                          * On-media truncation is cached in the inode until
2292                          * the inode is synchronized.  We must immediately
2293                          * handle any frontend records.
2294                          */
2295                         if (truncating) {
2296                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2297                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2298                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2299                                         ip->trunc_off = vap->va_size;
2300                                         hammer_inode_dirty(ip);
2301                                 } else if (ip->trunc_off > vap->va_size) {
2302                                         ip->trunc_off = vap->va_size;
2303                                 }
2304                         }
2305
2306 #if 0
2307                         /*
2308                          * When truncating, nvtruncbuf() may have cleaned out
2309                          * a portion of the last block on-disk in the buffer
2310                          * cache.  We must clean out any frontend records
2311                          * for blocks beyond the new last block.
2312                          */
2313                         aligned_size = (vap->va_size + (blksize - 1)) &
2314                                        ~(int64_t)(blksize - 1);
2315                         if (truncating && vap->va_size < aligned_size) {
2316                                 aligned_size -= blksize;
2317                                 hammer_ip_frontend_trunc(ip, aligned_size);
2318                         }
2319 #endif
2320                         break;
2321                 case VDATABASE:
2322                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2323                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2324                                 ip->trunc_off = vap->va_size;
2325                                 hammer_inode_dirty(ip);
2326                         } else if (ip->trunc_off > vap->va_size) {
2327                                 ip->trunc_off = vap->va_size;
2328                         }
2329                         hammer_ip_frontend_trunc(ip, vap->va_size);
2330                         ip->ino_data.size = vap->va_size;
2331                         ip->ino_data.mtime = trans.time;
2332                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2333                         kflags |= NOTE_ATTRIB;
2334                         break;
2335                 default:
2336                         error = EINVAL;
2337                         goto done;
2338                 }
2339                 break;
2340         }
2341         if (vap->va_atime.tv_sec != VNOVAL) {
2342                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2343                 modflags |= HAMMER_INODE_ATIME;
2344                 kflags |= NOTE_ATTRIB;
2345         }
2346         if (vap->va_mtime.tv_sec != VNOVAL) {
2347                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2348                 modflags |= HAMMER_INODE_MTIME;
2349                 kflags |= NOTE_ATTRIB;
2350         }
2351         if (vap->va_mode != (mode_t)VNOVAL) {
2352                 mode_t   cur_mode = ip->ino_data.mode;
2353                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2354                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2355
2356                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2357                                          cur_uid, cur_gid, &cur_mode);
2358                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2359                         ip->ino_data.mode = cur_mode;
2360                         ip->ino_data.ctime = trans.time;
2361                         modflags |= HAMMER_INODE_DDIRTY;
2362                         kflags |= NOTE_ATTRIB;
2363                 }
2364         }
2365 done:
2366         if (error == 0)
2367                 hammer_modify_inode(&trans, ip, modflags);
2368         hammer_done_transaction(&trans);
2369         hammer_knote(ap->a_vp, kflags);
2370         lwkt_reltoken(&hmp->fs_token);
2371         return (error);
2372 }
2373
2374 /*
2375  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2376  */
2377 static
2378 int
2379 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2380 {
2381         struct hammer_transaction trans;
2382         hammer_inode_t dip;
2383         hammer_inode_t nip;
2384         hammer_record_t record;
2385         struct nchandle *nch;
2386         hammer_mount_t hmp;
2387         int error;
2388         int bytes;
2389
2390         ap->a_vap->va_type = VLNK;
2391
2392         nch = ap->a_nch;
2393         dip = VTOI(ap->a_dvp);
2394         hmp = dip->hmp;
2395
2396         if (dip->flags & HAMMER_INODE_RO)
2397                 return (EROFS);
2398         if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0)
2399                 return (error);
2400
2401         /*
2402          * Create a transaction to cover the operations we perform.
2403          */
2404         lwkt_gettoken(&hmp->fs_token);
2405         hammer_start_transaction(&trans, hmp);
2406         ++hammer_stats_file_iopsw;
2407
2408         /*
2409          * Create a new filesystem object of the requested type.  The
2410          * returned inode will be referenced but not locked.
2411          */
2412
2413         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2414                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2415                                     NULL, &nip);
2416         if (error) {
2417                 hammer_done_transaction(&trans);
2418                 *ap->a_vpp = NULL;
2419                 lwkt_reltoken(&hmp->fs_token);
2420                 return (error);
2421         }
2422
2423         /*
2424          * Add a record representing the symlink.  symlink stores the link
2425          * as pure data, not a string, and is no \0 terminated.
2426          */
2427         if (error == 0) {
2428                 bytes = strlen(ap->a_target);
2429
2430                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2431                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2432                 } else {
2433                         record = hammer_alloc_mem_record(nip, bytes);
2434                         record->type = HAMMER_MEM_RECORD_GENERAL;
2435
2436                         record->leaf.base.localization = nip->obj_localization |
2437                                                          HAMMER_LOCALIZE_MISC;
2438                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2439                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2440                         record->leaf.data_len = bytes;
2441                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2442                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2443                         error = hammer_ip_add_record(&trans, record);
2444                 }
2445
2446                 /*
2447                  * Set the file size to the length of the link.
2448                  */
2449                 if (error == 0) {
2450                         nip->ino_data.size = bytes;
2451                         hammer_modify_inode(&trans, nip, HAMMER_INODE_DDIRTY);
2452                 }
2453         }
2454         if (error == 0)
2455                 error = hammer_ip_add_direntry(&trans, dip, nch->ncp->nc_name,
2456                                                 nch->ncp->nc_nlen, nip);
2457
2458         /*
2459          * Finish up.
2460          */
2461         if (error) {
2462                 hammer_rel_inode(nip, 0);
2463                 *ap->a_vpp = NULL;
2464         } else {
2465                 error = hammer_get_vnode(nip, ap->a_vpp);
2466                 hammer_rel_inode(nip, 0);
2467                 if (error == 0) {
2468                         cache_setunresolved(ap->a_nch);
2469                         cache_setvp(ap->a_nch, *ap->a_vpp);
2470                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2471                 }
2472         }
2473         hammer_done_transaction(&trans);
2474         lwkt_reltoken(&hmp->fs_token);
2475         return (error);
2476 }
2477
2478 /*
2479  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2480  */
2481 static
2482 int
2483 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2484 {
2485         struct hammer_transaction trans;
2486         hammer_inode_t dip;
2487         hammer_mount_t hmp;
2488         int error;
2489
2490         dip = VTOI(ap->a_dvp);
2491         hmp = dip->hmp;
2492
2493         if (hammer_nohistory(dip) == 0 &&
2494             (error = hammer_checkspace(hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2495                 return (error);
2496         }
2497
2498         lwkt_gettoken(&hmp->fs_token);
2499         hammer_start_transaction(&trans, hmp);
2500         ++hammer_stats_file_iopsw;
2501         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2502                                 ap->a_cred, ap->a_flags, -1);
2503         hammer_done_transaction(&trans);
2504         lwkt_reltoken(&hmp->fs_token);
2505
2506         return (error);
2507 }
2508
2509 /*
2510  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2511  */
2512 static
2513 int
2514 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2515 {
2516         hammer_inode_t ip = ap->a_vp->v_data;
2517         hammer_mount_t hmp = ip->hmp;
2518         int error;
2519
2520         ++hammer_stats_file_iopsr;
2521         lwkt_gettoken(&hmp->fs_token);
2522         error = hammer_ioctl(ip, ap->a_command, ap->a_data,
2523                              ap->a_fflag, ap->a_cred);
2524         lwkt_reltoken(&hmp->fs_token);
2525         return (error);
2526 }
2527
2528 static
2529 int
2530 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2531 {
2532         static const struct mountctl_opt extraopt[] = {
2533                 { HMNT_NOHISTORY,       "nohistory" },
2534                 { HMNT_MASTERID,        "master" },
2535                 { HMNT_NOMIRROR,        "nomirror" },
2536                 { 0, NULL}
2537
2538         };
2539         hammer_mount_t hmp;
2540         struct mount *mp;
2541         int usedbytes;
2542         int error;
2543
2544         error = 0;
2545         usedbytes = 0;
2546         mp = ap->a_head.a_ops->head.vv_mount;
2547         KKASSERT(mp->mnt_data != NULL);
2548         hmp = (hammer_mount_t)mp->mnt_data;
2549
2550         lwkt_gettoken(&hmp->fs_token);
2551
2552         switch(ap->a_op) {
2553         case MOUNTCTL_SET_EXPORT:
2554                 if (ap->a_ctllen != sizeof(struct export_args))
2555                         error = EINVAL;
2556                 else
2557                         error = hammer_vfs_export(mp, ap->a_op,
2558                                       (const struct export_args *)ap->a_ctl);
2559                 break;
2560         case MOUNTCTL_MOUNTFLAGS:
2561                 /*
2562                  * Call standard mountctl VOP function
2563                  * so we get user mount flags.
2564                  */
2565                 error = vop_stdmountctl(ap);
2566                 if (error)
2567                         break;
2568
2569                 usedbytes = *ap->a_res;
2570
2571                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2572                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt,
2573                                                     ap->a_buf,
2574                                                     ap->a_buflen - usedbytes,
2575                                                     &error);
2576                 }
2577
2578                 *ap->a_res += usedbytes;
2579                 break;
2580         default:
2581                 error = vop_stdmountctl(ap);
2582                 break;
2583         }
2584         lwkt_reltoken(&hmp->fs_token);
2585         return(error);
2586 }
2587
2588 /*
2589  * hammer_vop_strategy { vp, bio }
2590  *
2591  * Strategy call, used for regular file read & write only.  Note that the
2592  * bp may represent a cluster.
2593  *
2594  * To simplify operation and allow better optimizations in the future,
2595  * this code does not make any assumptions with regards to buffer alignment
2596  * or size.
2597  */
2598 static
2599 int
2600 hammer_vop_strategy(struct vop_strategy_args *ap)
2601 {
2602         struct buf *bp;
2603         int error;
2604
2605         bp = ap->a_bio->bio_buf;
2606
2607         switch(bp->b_cmd) {
2608         case BUF_CMD_READ:
2609                 error = hammer_vop_strategy_read(ap);
2610                 break;
2611         case BUF_CMD_WRITE:
2612                 error = hammer_vop_strategy_write(ap);
2613                 break;
2614         default:
2615                 bp->b_error = error = EINVAL;
2616                 bp->b_flags |= B_ERROR;
2617                 biodone(ap->a_bio);
2618                 break;
2619         }
2620
2621         /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2622
2623         return (error);
2624 }
2625
2626 /*
2627  * Read from a regular file.  Iterate the related records and fill in the
2628  * BIO/BUF.  Gaps are zero-filled.
2629  *
2630  * The support code in hammer_object.c should be used to deal with mixed
2631  * in-memory and on-disk records.
2632  *
2633  * NOTE: Can be called from the cluster code with an oversized buf.
2634  *
2635  * XXX atime update
2636  */
2637 static
2638 int
2639 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2640 {
2641         struct hammer_transaction trans;
2642         hammer_inode_t ip;
2643         hammer_inode_t dip;
2644         hammer_mount_t hmp;
2645         struct hammer_cursor cursor;
2646         hammer_base_elm_t base;
2647         hammer_off_t disk_offset;
2648         struct bio *bio;
2649         struct bio *nbio;
2650         struct buf *bp;
2651         int64_t rec_offset;
2652         int64_t ran_end;
2653         int64_t tmp64;
2654         int error;
2655         int boff;
2656         int roff;
2657         int n;
2658         int isdedupable;
2659
2660         bio = ap->a_bio;
2661         bp = bio->bio_buf;
2662         ip = ap->a_vp->v_data;
2663         hmp = ip->hmp;
2664
2665         /*
2666          * The zone-2 disk offset may have been set by the cluster code via
2667          * a BMAP operation, or else should be NOOFFSET.
2668          *
2669          * Checking the high bits for a match against zone-2 should suffice.
2670          *
2671          * In cases where a lot of data duplication is present it may be
2672          * more beneficial to drop through and doubule-buffer through the
2673          * device.
2674          */
2675         nbio = push_bio(bio);
2676         if (hammer_is_zone_large_data(nbio->bio_offset)) {
2677                 if (hammer_double_buffer == 0) {
2678                         lwkt_gettoken(&hmp->fs_token);
2679                         error = hammer_io_direct_read(hmp, nbio, NULL);
2680                         lwkt_reltoken(&hmp->fs_token);
2681                         return (error);
2682                 }
2683
2684                 /*
2685                  * Try to shortcut requests for double_buffer mode too.
2686                  * Since this mode runs through the device buffer cache
2687                  * only compatible buffer sizes (meaning those generated
2688                  * by normal filesystem buffers) are legal.
2689                  */
2690                 if (hammer_live_dedup == 0 && (bp->b_flags & B_PAGING) == 0) {
2691                         lwkt_gettoken(&hmp->fs_token);
2692                         error = hammer_io_indirect_read(hmp, nbio, NULL);
2693                         lwkt_reltoken(&hmp->fs_token);
2694                         return (error);
2695                 }
2696         }
2697
2698         /*
2699          * Well, that sucked.  Do it the hard way.  If all the stars are
2700          * aligned we may still be able to issue a direct-read.
2701          */
2702         lwkt_gettoken(&hmp->fs_token);
2703         hammer_simple_transaction(&trans, hmp);
2704         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2705
2706         /*
2707          * Key range (begin and end inclusive) to scan.  Note that the key's
2708          * stored in the actual records represent BASE+LEN, not BASE.  The
2709          * first record containing bio_offset will have a key > bio_offset.
2710          */
2711         cursor.key_beg.localization = ip->obj_localization |
2712                                       HAMMER_LOCALIZE_MISC;
2713         cursor.key_beg.obj_id = ip->obj_id;
2714         cursor.key_beg.create_tid = 0;
2715         cursor.key_beg.delete_tid = 0;
2716         cursor.key_beg.obj_type = 0;
2717         cursor.key_beg.key = bio->bio_offset + 1;
2718         cursor.asof = ip->obj_asof;
2719         cursor.flags |= HAMMER_CURSOR_ASOF;
2720
2721         cursor.key_end = cursor.key_beg;
2722         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2723 #if 0
2724         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2725                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2726                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2727                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2728         } else
2729 #endif
2730         {
2731                 ran_end = bio->bio_offset + bp->b_bufsize;
2732                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2733                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2734                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2735                 if (tmp64 < ran_end)
2736                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2737                 else
2738                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2739         }
2740         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2741
2742         /*
2743          * Set NOSWAPCACHE for cursor data extraction if double buffering
2744          * is disabled or (if the file is not marked cacheable via chflags
2745          * and vm.swapcache_use_chflags is enabled).
2746          */
2747         if (hammer_double_buffer == 0 ||
2748             ((ap->a_vp->v_flag & VSWAPCACHE) == 0 &&
2749              vm_swapcache_use_chflags)) {
2750                 cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
2751         }
2752
2753         error = hammer_ip_first(&cursor);
2754         boff = 0;
2755
2756         while (error == 0) {
2757                 /*
2758                  * Get the base file offset of the record.  The key for
2759                  * data records is (base + bytes) rather then (base).
2760                  */
2761                 base = &cursor.leaf->base;
2762                 rec_offset = base->key - cursor.leaf->data_len;
2763
2764                 /*
2765                  * Calculate the gap, if any, and zero-fill it.
2766                  *
2767                  * n is the offset of the start of the record verses our
2768                  * current seek offset in the bio.
2769                  */
2770                 n = (int)(rec_offset - (bio->bio_offset + boff));
2771                 if (n > 0) {
2772                         if (n > bp->b_bufsize - boff)
2773                                 n = bp->b_bufsize - boff;
2774                         bzero((char *)bp->b_data + boff, n);
2775                         boff += n;
2776                         n = 0;
2777                 }
2778
2779                 /*
2780                  * Calculate the data offset in the record and the number
2781                  * of bytes we can copy.
2782                  *
2783                  * There are two degenerate cases.  First, boff may already
2784                  * be at bp->b_bufsize.  Secondly, the data offset within
2785                  * the record may exceed the record's size.
2786                  */
2787                 roff = -n;
2788                 rec_offset += roff;
2789                 n = cursor.leaf->data_len - roff;
2790                 if (n <= 0) {
2791                         hdkprintf("bad n=%d roff=%d\n", n, roff);
2792                         n = 0;
2793                 } else if (n > bp->b_bufsize - boff) {
2794                         n = bp->b_bufsize - boff;
2795                 }
2796
2797                 /*
2798                  * Deal with cached truncations.  This cool bit of code
2799                  * allows truncate()/ftruncate() to avoid having to sync
2800                  * the file.
2801                  *
2802                  * If the frontend is truncated then all backend records are
2803                  * subject to the frontend's truncation.
2804                  *
2805                  * If the backend is truncated then backend records on-disk
2806                  * (but not in-memory) are subject to the backend's
2807                  * truncation.  In-memory records owned by the backend
2808                  * represent data written after the truncation point on the
2809                  * backend and must not be truncated.
2810                  *
2811                  * Truncate operations deal with frontend buffer cache
2812                  * buffers and frontend-owned in-memory records synchronously.
2813                  */
2814                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2815                         if (hammer_cursor_ondisk(&cursor)/* ||
2816                             cursor.iprec->flush_state == HAMMER_FST_FLUSH*/) {
2817                                 if (ip->trunc_off <= rec_offset)
2818                                         n = 0;
2819                                 else if (ip->trunc_off < rec_offset + n)
2820                                         n = (int)(ip->trunc_off - rec_offset);
2821                         }
2822                 }
2823                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2824                         if (hammer_cursor_ondisk(&cursor)) {
2825                                 if (ip->sync_trunc_off <= rec_offset)
2826                                         n = 0;
2827                                 else if (ip->sync_trunc_off < rec_offset + n)
2828                                         n = (int)(ip->sync_trunc_off - rec_offset);
2829                         }
2830                 }
2831
2832                 /*
2833                  * Try to issue a direct read into our bio if possible,
2834                  * otherwise resolve the element data into a hammer_buffer
2835                  * and copy.
2836                  *
2837                  * The buffer on-disk should be zerod past any real
2838                  * truncation point, but may not be for any synthesized
2839                  * truncation point from above.
2840                  *
2841                  * NOTE: disk_offset is only valid if the cursor data is
2842                  *       on-disk.
2843                  */
2844                 disk_offset = cursor.leaf->data_offset + roff;
2845                 isdedupable = (boff == 0 && n == bp->b_bufsize &&
2846                                hammer_cursor_ondisk(&cursor) &&
2847                                ((int)disk_offset & HAMMER_BUFMASK) == 0);
2848
2849                 if (isdedupable && hammer_double_buffer == 0) {
2850                         /*
2851                          * Direct read case
2852                          */
2853                         KKASSERT(hammer_is_zone_large_data(disk_offset));
2854                         nbio->bio_offset = disk_offset;
2855                         error = hammer_io_direct_read(hmp, nbio, cursor.leaf);
2856                         if (hammer_live_dedup && error == 0)
2857                                 hammer_dedup_cache_add(ip, cursor.leaf);
2858                         goto done;
2859                 } else if (isdedupable) {
2860                         /*
2861                          * Async I/O case for reading from backing store
2862                          * and copying the data to the filesystem buffer.
2863                          * live-dedup has to verify the data anyway if it
2864                          * gets a hit later so we can just add the entry
2865                          * now.
2866                          */
2867                         KKASSERT(hammer_is_zone_large_data(disk_offset));
2868                         nbio->bio_offset = disk_offset;
2869                         if (hammer_live_dedup)
2870                                 hammer_dedup_cache_add(ip, cursor.leaf);
2871                         error = hammer_io_indirect_read(hmp, nbio, cursor.leaf);
2872                         goto done;
2873                 } else if (n) {
2874                         error = hammer_ip_resolve_data(&cursor);
2875                         if (error == 0) {
2876                                 if (hammer_live_dedup && isdedupable)
2877                                         hammer_dedup_cache_add(ip, cursor.leaf);
2878                                 bcopy((char *)cursor.data + roff,
2879                                       (char *)bp->b_data + boff, n);
2880                         }
2881                 }
2882                 if (error)
2883                         break;
2884
2885                 /*
2886                  * We have to be sure that the only elements added to the
2887                  * dedup cache are those which are already on-media.
2888                  */
2889                 if (hammer_live_dedup && hammer_cursor_ondisk(&cursor))
2890                         hammer_dedup_cache_add(ip, cursor.leaf);
2891
2892                 /*
2893                  * Iterate until we have filled the request.
2894                  */
2895                 boff += n;
2896                 if (boff == bp->b_bufsize)
2897                         break;
2898                 error = hammer_ip_next(&cursor);
2899         }
2900
2901         /*
2902          * There may have been a gap after the last record
2903          */
2904         if (error == ENOENT)
2905                 error = 0;
2906         if (error == 0 && boff != bp->b_bufsize) {
2907                 KKASSERT(boff < bp->b_bufsize);
2908                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2909                 /* boff = bp->b_bufsize; */
2910         }
2911
2912         /*
2913          * Disallow swapcache operation on the vnode buffer if double
2914          * buffering is enabled, the swapcache will get the data via
2915          * the block device buffer.
2916          */
2917         if (hammer_double_buffer)
2918                 bp->b_flags |= B_NOTMETA;
2919
2920         /*
2921          * Cleanup
2922          */
2923         bp->b_resid = 0;
2924         bp->b_error = error;
2925         if (error)
2926                 bp->b_flags |= B_ERROR;
2927         biodone(ap->a_bio);
2928
2929 done:
2930         /*
2931          * Cache the b-tree node for the last data read in cache[1].
2932          *
2933          * If we hit the file EOF then also cache the node in the
2934          * governing directory's cache[3], it will be used to initialize
2935          * the new inode's cache[1] for any inodes looked up via the directory.
2936          *
2937          * This doesn't reduce disk accesses since the B-Tree chain is
2938          * likely cached, but it does reduce cpu overhead when looking
2939          * up file offsets for cpdup/tar/cpio style iterations.
2940          */
2941         if (cursor.node)
2942                 hammer_cache_node(&ip->cache[1], cursor.node);
2943         if (ran_end >= ip->ino_data.size) {
2944                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2945                                         ip->obj_asof, ip->obj_localization);
2946                 if (dip) {
2947                         hammer_cache_node(&dip->cache[3], cursor.node);
2948                         hammer_rel_inode(dip, 0);
2949                 }
2950         }
2951         hammer_done_cursor(&cursor);
2952         hammer_done_transaction(&trans);
2953         lwkt_reltoken(&hmp->fs_token);
2954         return(error);
2955 }
2956
2957 /*
2958  * BMAP operation - used to support cluster_read() only.
2959  *
2960  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2961  *
2962  * This routine may return EOPNOTSUPP if the opration is not supported for
2963  * the specified offset.  The contents of the pointer arguments do not
2964  * need to be initialized in that case.
2965  *
2966  * If a disk address is available and properly aligned return 0 with
2967  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2968  * to the run-length relative to that offset.  Callers may assume that
2969  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2970  * large, so return EOPNOTSUPP if it is not sufficiently large.
2971  */
2972 static
2973 int
2974 hammer_vop_bmap(struct vop_bmap_args *ap)
2975 {
2976         struct hammer_transaction trans;
2977         hammer_inode_t ip;
2978         hammer_mount_t hmp;
2979         struct hammer_cursor cursor;
2980         hammer_base_elm_t base;
2981         int64_t rec_offset;
2982         int64_t ran_end;
2983         int64_t tmp64;
2984         int64_t base_offset;
2985         int64_t base_disk_offset;
2986         int64_t last_offset;
2987         hammer_off_t last_disk_offset;
2988         hammer_off_t disk_offset;
2989         int     rec_len;
2990         int     error;
2991         int     blksize;
2992
2993         ++hammer_stats_file_iopsr;
2994         ip = ap->a_vp->v_data;
2995         hmp = ip->hmp;
2996
2997         /*
2998          * We can only BMAP regular files.  We can't BMAP database files,
2999          * directories, etc.
3000          */
3001         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
3002                 return(EOPNOTSUPP);
3003
3004         /*
3005          * bmap is typically called with runp/runb both NULL when used
3006          * for writing.  We do not support BMAP for writing atm.
3007          */
3008         if (ap->a_cmd != BUF_CMD_READ)
3009                 return(EOPNOTSUPP);
3010
3011         /*
3012          * Scan the B-Tree to acquire blockmap addresses, then translate
3013          * to raw addresses.
3014          */
3015         lwkt_gettoken(&hmp->fs_token);
3016         hammer_simple_transaction(&trans, hmp);
3017
3018         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
3019
3020         /*
3021          * Key range (begin and end inclusive) to scan.  Note that the key's
3022          * stored in the actual records represent BASE+LEN, not BASE.  The
3023          * first record containing bio_offset will have a key > bio_offset.
3024          */
3025         cursor.key_beg.localization = ip->obj_localization |
3026                                       HAMMER_LOCALIZE_MISC;
3027         cursor.key_beg.obj_id = ip->obj_id;
3028         cursor.key_beg.create_tid = 0;
3029         cursor.key_beg.delete_tid = 0;
3030         cursor.key_beg.obj_type = 0;
3031         if (ap->a_runb)
3032                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
3033         else
3034                 cursor.key_beg.key = ap->a_loffset + 1;
3035         if (cursor.key_beg.key < 0)
3036                 cursor.key_beg.key = 0;
3037         cursor.asof = ip->obj_asof;
3038         cursor.flags |= HAMMER_CURSOR_ASOF;
3039
3040         cursor.key_end = cursor.key_beg;
3041         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
3042
3043         ran_end = ap->a_loffset + MAXPHYS;
3044         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
3045         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
3046         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
3047         if (tmp64 < ran_end)
3048                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
3049         else
3050                 cursor.key_end.key = ran_end + MAXPHYS + 1;
3051
3052         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
3053
3054         error = hammer_ip_first(&cursor);
3055         base_offset = last_offset = 0;
3056         base_disk_offset = last_disk_offset = 0;
3057
3058         while (error == 0) {
3059                 /*
3060                  * Get the base file offset of the record.  The key for
3061                  * data records is (base + bytes) rather then (base).
3062                  *
3063                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
3064                  * The extra bytes should be zero on-disk and the BMAP op
3065                  * should still be ok.
3066                  */
3067                 base = &cursor.leaf->base;
3068                 rec_offset = base->key - cursor.leaf->data_len;
3069                 rec_len    = cursor.leaf->data_len;
3070
3071                 /*
3072                  * Incorporate any cached truncation.
3073                  *
3074                  * NOTE: Modifications to rec_len based on synthesized
3075                  * truncation points remove the guarantee that any extended
3076                  * data on disk is zero (since the truncations may not have
3077                  * taken place on-media yet).
3078                  */
3079                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
3080                         if (hammer_cursor_ondisk(&cursor) ||
3081                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
3082                                 if (ip->trunc_off <= rec_offset)
3083                                         rec_len = 0;
3084                                 else if (ip->trunc_off < rec_offset + rec_len)
3085                                         rec_len = (int)(ip->trunc_off - rec_offset);
3086                         }
3087                 }
3088                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
3089                         if (hammer_cursor_ondisk(&cursor)) {
3090                                 if (ip->sync_trunc_off <= rec_offset)
3091                                         rec_len = 0;
3092                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
3093                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
3094                         }
3095                 }
3096
3097                 /*
3098                  * Accumulate information.  If we have hit a discontiguous
3099                  * block reset base_offset unless we are already beyond the
3100                  * requested offset.  If we are, that's it, we stop.
3101                  */
3102                 if (error)
3103                         break;
3104                 if (hammer_cursor_ondisk(&cursor)) {
3105                         disk_offset = cursor.leaf->data_offset;
3106                         if (rec_offset != last_offset ||
3107                             disk_offset != last_disk_offset) {
3108                                 if (rec_offset > ap->a_loffset)
3109                                         break;
3110                                 base_offset = rec_offset;
3111                                 base_disk_offset = disk_offset;
3112                         }
3113                         last_offset = rec_offset + rec_len;
3114                         last_disk_offset = disk_offset + rec_len;
3115
3116                         if (hammer_live_dedup)
3117                                 hammer_dedup_cache_add(ip, cursor.leaf);
3118                 }
3119
3120                 error = hammer_ip_next(&cursor);
3121         }
3122
3123         if (cursor.node)
3124                 hammer_cache_node(&ip->cache[1], cursor.node);
3125
3126         hammer_done_cursor(&cursor);
3127         hammer_done_transaction(&trans);
3128         lwkt_reltoken(&hmp->fs_token);
3129
3130         /*
3131          * If we couldn't find any records or the records we did find were
3132          * all behind the requested offset, return failure.  A forward
3133          * truncation can leave a hole w/ no on-disk records.
3134          */
3135         if (last_offset == 0 || last_offset < ap->a_loffset)
3136                 return (EOPNOTSUPP);
3137
3138         /*
3139          * Figure out the block size at the requested offset and adjust
3140          * our limits so the cluster_read() does not create inappropriately
3141          * sized buffer cache buffers.
3142          */
3143         blksize = hammer_blocksize(ap->a_loffset);
3144         if (hammer_blocksize(base_offset) != blksize) {
3145                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
3146         }
3147         if (last_offset != ap->a_loffset &&
3148             hammer_blocksize(last_offset - 1) != blksize) {
3149                 last_offset = hammer_blockdemarc(ap->a_loffset,
3150                                                  last_offset - 1);
3151         }
3152
3153         /*
3154          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3155          * from occuring.
3156          */
3157         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
3158
3159         if (!hammer_is_zone_large_data(disk_offset)) {
3160                 /*
3161                  * Only large-data zones can be direct-IOd
3162                  */
3163                 error = EOPNOTSUPP;
3164         } else if ((disk_offset & HAMMER_BUFMASK) ||
3165                    (last_offset - ap->a_loffset) < blksize) {
3166                 /*
3167                  * doffsetp is not aligned or the forward run size does
3168                  * not cover a whole buffer, disallow the direct I/O.
3169                  */
3170                 error = EOPNOTSUPP;
3171         } else {
3172                 /*
3173                  * We're good.
3174                  */
3175                 *ap->a_doffsetp = disk_offset;
3176                 if (ap->a_runb) {
3177                         *ap->a_runb = ap->a_loffset - base_offset;
3178                         KKASSERT(*ap->a_runb >= 0);
3179                 }
3180                 if (ap->a_runp) {
3181                         *ap->a_runp = last_offset - ap->a_loffset;
3182                         KKASSERT(*ap->a_runp >= 0);
3183                 }
3184                 error = 0;
3185         }
3186         return(error);
3187 }
3188
3189 /*
3190  * Write to a regular file.   Because this is a strategy call the OS is
3191  * trying to actually get data onto the media.
3192  */
3193 static
3194 int
3195 hammer_vop_strategy_write(struct vop_strategy_args *ap)
3196 {
3197         hammer_record_t record;
3198         hammer_mount_t hmp;
3199         hammer_inode_t ip;
3200         struct bio *bio;
3201         struct buf *bp;
3202         int blksize __debugvar;
3203         int bytes;
3204         int error;
3205
3206         bio = ap->a_bio;
3207         bp = bio->bio_buf;
3208         ip = ap->a_vp->v_data;
3209         hmp = ip->hmp;
3210
3211         blksize = hammer_blocksize(bio->bio_offset);
3212         KKASSERT(bp->b_bufsize == blksize);
3213
3214         if (ip->flags & HAMMER_INODE_RO) {
3215                 bp->b_error = EROFS;
3216                 bp->b_flags |= B_ERROR;
3217                 biodone(ap->a_bio);
3218                 return(EROFS);
3219         }
3220
3221         lwkt_gettoken(&hmp->fs_token);
3222
3223         /*
3224          * Disallow swapcache operation on the vnode buffer if double
3225          * buffering is enabled, the swapcache will get the data via
3226          * the block device buffer.
3227          */
3228         if (hammer_double_buffer)
3229                 bp->b_flags |= B_NOTMETA;
3230
3231         /*
3232          * Interlock with inode destruction (no in-kernel or directory
3233          * topology visibility).  If we queue new IO while trying to
3234          * destroy the inode we can deadlock the vtrunc call in
3235          * hammer_inode_unloadable_check().
3236          *
3237          * Besides, there's no point flushing a bp associated with an
3238          * inode that is being destroyed on-media and has no kernel
3239          * references.
3240          */
3241         if ((ip->flags | ip->sync_flags) &
3242             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
3243                 bp->b_resid = 0;
3244                 biodone(ap->a_bio);
3245                 lwkt_reltoken(&hmp->fs_token);
3246                 return(0);
3247         }
3248
3249         /*
3250          * Reserve space and issue a direct-write from the front-end.
3251          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3252          * allocations.
3253          *
3254          * An in-memory record will be installed to reference the storage
3255          * until the flusher can get to it.
3256          *
3257          * Since we own the high level bio the front-end will not try to
3258          * do a direct-read until the write completes.
3259          *
3260          * NOTE: The only time we do not reserve a full-sized buffers
3261          * worth of data is if the file is small.  We do not try to
3262          * allocate a fragment (from the small-data zone) at the end of
3263          * an otherwise large file as this can lead to wildly separated
3264          * data.
3265          */
3266         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3267         KKASSERT(bio->bio_offset < ip->ino_data.size);
3268         if (bio->bio_offset || ip->ino_data.size > HAMMER_HBUFSIZE)
3269                 bytes = bp->b_bufsize;
3270         else
3271                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3272
3273         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3274                                     bytes, &error);
3275
3276         /*
3277          * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3278          * in hammer_vop_write().  We must flag the record so the proper
3279          * REDO_TERM_WRITE entry is generated during the flush.
3280          */
3281         if (record) {
3282                 if (bp->b_flags & B_VFSFLAG1) {
3283                         record->flags |= HAMMER_RECF_REDO;
3284                         bp->b_flags &= ~B_VFSFLAG1;
3285                 }
3286                 if (record->flags & HAMMER_RECF_DEDUPED) {
3287                         bp->b_resid = 0;
3288                         hammer_ip_replace_bulk(hmp, record);
3289                         biodone(ap->a_bio);
3290                 } else {
3291                         hammer_io_direct_write(hmp, bio, record);
3292                 }
3293                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3294                         hammer_flush_inode(ip, 0);
3295         } else {
3296                 bp->b_bio2.bio_offset = NOOFFSET;
3297                 bp->b_error = error;
3298                 bp->b_flags |= B_ERROR;
3299                 biodone(ap->a_bio);
3300         }
3301         lwkt_reltoken(&hmp->fs_token);
3302         return(error);
3303 }
3304
3305 /*
3306  * dounlink - disconnect a directory entry
3307  *
3308  * XXX whiteout support not really in yet
3309  */
3310 static int
3311 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3312                 struct vnode *dvp, struct ucred *cred,
3313                 int flags, int isdir)
3314 {
3315         struct namecache *ncp;
3316         hammer_inode_t dip;
3317         hammer_inode_t ip;
3318         hammer_mount_t hmp;
3319         struct hammer_cursor cursor;
3320         int64_t namekey;
3321         uint32_t max_iterations;
3322         int nlen, error;
3323
3324         /*
3325          * Calculate the namekey and setup the key range for the scan.  This
3326          * works kinda like a chained hash table where the lower 32 bits
3327          * of the namekey synthesize the chain.
3328          *
3329          * The key range is inclusive of both key_beg and key_end.
3330          */
3331         dip = VTOI(dvp);
3332         ncp = nch->ncp;
3333         hmp = dip->hmp;
3334
3335         if (dip->flags & HAMMER_INODE_RO)
3336                 return (EROFS);
3337
3338         namekey = hammer_direntry_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3339                                            &max_iterations);
3340 retry:
3341         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3342         cursor.key_beg.localization = dip->obj_localization |
3343                                       hammer_dir_localization(dip);
3344         cursor.key_beg.obj_id = dip->obj_id;
3345         cursor.key_beg.key = namekey;
3346         cursor.key_beg.create_tid = 0;
3347         cursor.key_beg.delete_tid = 0;
3348         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3349         cursor.key_beg.obj_type = 0;
3350
3351         cursor.key_end = cursor.key_beg;
3352         cursor.key_end.key += max_iterations;
3353         cursor.asof = dip->obj_asof;
3354         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3355
3356         /*
3357          * Scan all matching records (the chain), locate the one matching
3358          * the requested path component.  info->last_error contains the
3359          * error code on search termination and could be 0, ENOENT, or
3360          * something else.
3361          *
3362          * The hammer_ip_*() functions merge in-memory records with on-disk
3363          * records for the purposes of the search.
3364          */
3365         error = hammer_ip_first(&cursor);
3366
3367         while (error == 0) {
3368                 error = hammer_ip_resolve_data(&cursor);
3369                 if (error)
3370                         break;
3371                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3372                 KKASSERT(nlen > 0);
3373                 if (ncp->nc_nlen == nlen &&
3374                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3375                         break;
3376                 }
3377                 error = hammer_ip_next(&cursor);
3378         }
3379
3380         /*
3381          * If all is ok we have to get the inode so we can adjust nlinks.
3382          * To avoid a deadlock with the flusher we must release the inode
3383          * lock on the directory when acquiring the inode for the entry.
3384          *
3385          * If the target is a directory, it must be empty.
3386          */
3387         if (error == 0) {
3388                 hammer_unlock(&cursor.ip->lock);
3389                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3390                                       hmp->asof,
3391                                       cursor.data->entry.localization,
3392                                       0, &error);
3393                 hammer_lock_sh(&cursor.ip->lock);
3394                 if (error == ENOENT) {
3395                         hkprintf("WARNING: Removing dirent w/missing inode "
3396                                 "\"%s\"\n"
3397                                 "\tobj_id = %016jx\n",
3398                                 ncp->nc_name,
3399                                 (intmax_t)cursor.data->entry.obj_id);
3400                         error = 0;
3401                 }
3402
3403                 /*
3404                  * If isdir >= 0 we validate that the entry is or is not a
3405                  * directory.  If isdir < 0 we don't care.
3406                  */
3407                 if (error == 0 && isdir >= 0 && ip) {
3408                         if (isdir &&
3409                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3410                                 error = ENOTDIR;
3411                         } else if (isdir == 0 &&
3412                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3413                                 error = EISDIR;
3414                         }
3415                 }
3416
3417                 /*
3418                  * If we are trying to remove a directory the directory must
3419                  * be empty.
3420                  *
3421                  * The check directory code can loop and deadlock/retry.  Our
3422                  * own cursor's node locks must be released to avoid a 3-way
3423                  * deadlock with the flusher if the check directory code
3424                  * blocks.
3425                  *
3426                  * If any changes whatsoever have been made to the cursor
3427                  * set EDEADLK and retry.
3428                  *
3429                  * WARNING: See warnings in hammer_unlock_cursor()
3430                  *          function.
3431                  */
3432                 if (error == 0 && ip && ip->ino_data.obj_type ==
3433                                         HAMMER_OBJTYPE_DIRECTORY) {
3434                         hammer_unlock_cursor(&cursor);
3435                         error = hammer_ip_check_directory_empty(trans, ip);
3436                         hammer_lock_cursor(&cursor);
3437                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3438                                 hkprintf("Warning: avoided deadlock "
3439                                         "on rmdir '%s'\n",
3440                                         ncp->nc_name);
3441                                 error = EDEADLK;
3442                         }
3443                 }
3444
3445                 /*
3446                  * Delete the directory entry.
3447                  *
3448                  * WARNING: hammer_ip_del_direntry() may have to terminate
3449                  * the cursor to avoid a deadlock.  It is ok to call
3450                  * hammer_done_cursor() twice.
3451                  */
3452                 if (error == 0) {
3453                         error = hammer_ip_del_direntry(trans, &cursor,
3454                                                         dip, ip);
3455                 }
3456                 hammer_done_cursor(&cursor);
3457                 if (error == 0) {
3458                         /*
3459                          * Tell the namecache that we are now unlinked.
3460                          */
3461                         cache_unlink(nch);
3462
3463                         /*
3464                          * NOTE: ip->vp, if non-NULL, cannot be directly
3465                          *       referenced without formally acquiring the
3466                          *       vp since the vp might have zero refs on it,
3467                          *       or in the middle of a reclaim, etc.
3468                          *
3469                          * NOTE: The cache_setunresolved() can rip the vp
3470                          *       out from under us since the vp may not have
3471                          *       any refs, in which case ip->vp will be NULL
3472                          *       from the outset.
3473                          */
3474                         while (ip && ip->vp) {
3475                                 struct vnode *vp;
3476
3477                                 error = hammer_get_vnode(ip, &vp);
3478                                 if (error == 0 && vp) {
3479                                         vn_unlock(vp);
3480                                         hammer_knote(ip->vp, NOTE_DELETE);
3481 #if 0
3482                                         /*
3483                                          * Don't do this, it can deadlock
3484                                          * on concurrent rm's of hardlinks.
3485                                          * Shouldn't be needed any more.
3486                                          */
3487                                         cache_inval_vp(ip->vp, CINV_DESTROY);
3488 #endif
3489                                         vrele(vp);
3490                                         break;
3491                                 }
3492                                 hdkprintf("ip/vp race1 avoided\n");
3493                         }
3494                 }
3495                 if (ip)
3496                         hammer_rel_inode(ip, 0);
3497         } else {
3498                 hammer_done_cursor(&cursor);
3499         }
3500         if (error == EDEADLK)
3501                 goto retry;
3502
3503         return (error);
3504 }
3505
3506 /************************************************************************
3507  *                          FIFO AND SPECFS OPS                         *
3508  ************************************************************************
3509  *
3510  */
3511 static int
3512 hammer_vop_fifoclose (struct vop_close_args *ap)
3513 {
3514         /* XXX update itimes */
3515         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3516 }
3517
3518 static int
3519 hammer_vop_fiforead (struct vop_read_args *ap)
3520 {
3521         int error;
3522
3523         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3524         /* XXX update access time */
3525         return (error);
3526 }
3527
3528 static int
3529 hammer_vop_fifowrite (struct vop_write_args *ap)
3530 {
3531         int error;
3532
3533         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3534         /* XXX update access time */
3535         return (error);
3536 }
3537
3538 static
3539 int
3540 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3541 {
3542         int error;
3543
3544         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3545         if (error)
3546                 error = hammer_vop_kqfilter(ap);
3547         return(error);
3548 }
3549
3550 /************************************************************************
3551  *                          KQFILTER OPS                                *
3552  ************************************************************************
3553  *
3554  */
3555 static void filt_hammerdetach(struct knote *kn);
3556 static int filt_hammerread(struct knote *kn, long hint);
3557 static int filt_hammerwrite(struct knote *kn, long hint);
3558 static int filt_hammervnode(struct knote *kn, long hint);
3559
3560 static struct filterops hammerread_filtops =
3561         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3562           NULL, filt_hammerdetach, filt_hammerread };
3563 static struct filterops hammerwrite_filtops =
3564         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3565           NULL, filt_hammerdetach, filt_hammerwrite };
3566 static struct filterops hammervnode_filtops =
3567         { FILTEROP_ISFD | FILTEROP_MPSAFE,
3568           NULL, filt_hammerdetach, filt_hammervnode };
3569
3570 static
3571 int
3572 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3573 {
3574         struct vnode *vp = ap->a_vp;
3575         struct knote *kn = ap->a_kn;
3576
3577         switch (kn->kn_filter) {
3578         case EVFILT_READ:
3579                 kn->kn_fop = &hammerread_filtops;
3580                 break;
3581         case EVFILT_WRITE:
3582                 kn->kn_fop = &hammerwrite_filtops;
3583                 break;
3584         case EVFILT_VNODE:
3585                 kn->kn_fop = &hammervnode_filtops;
3586                 break;
3587         default:
3588                 return (EOPNOTSUPP);
3589         }
3590
3591         kn->kn_hook = (caddr_t)vp;
3592
3593         knote_insert(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3594
3595         return(0);
3596 }
3597
3598 static void
3599 filt_hammerdetach(struct knote *kn)
3600 {
3601         struct vnode *vp = (void *)kn->kn_hook;
3602
3603         knote_remove(&vp->v_pollinfo.vpi_kqinfo.ki_note, kn);
3604 }
3605
3606 static int
3607 filt_hammerread(struct knote *kn, long hint)
3608 {
3609         struct vnode *vp = (void *)kn->kn_hook;
3610         hammer_inode_t ip = VTOI(vp);
3611         hammer_mount_t hmp = ip->hmp;
3612         off_t off;
3613
3614         if (hint == NOTE_REVOKE) {
3615                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3616                 return(1);
3617         }
3618         lwkt_gettoken(&hmp->fs_token);  /* XXX use per-ip-token */
3619         off = ip->ino_data.size - kn->kn_fp->f_offset;
3620         kn->kn_data = (off < INTPTR_MAX) ? off : INTPTR_MAX;
3621         lwkt_reltoken(&hmp->fs_token);
3622         if (kn->kn_sfflags & NOTE_OLDAPI)
3623                 return(1);
3624         return (kn->kn_data != 0);
3625 }
3626
3627 static int
3628 filt_hammerwrite(struct knote *kn, long hint)
3629 {
3630         if (hint == NOTE_REVOKE)
3631                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
3632         kn->kn_data = 0;
3633         return (1);
3634 }
3635
3636 static int
3637 filt_hammervnode(struct knote *kn, long hint)
3638 {
3639         if (kn->kn_sfflags & hint)
3640                 kn->kn_fflags |= hint;
3641         if (hint == NOTE_REVOKE) {
3642                 kn->kn_flags |= (EV_EOF | EV_NODATA);
3643                 return (1);
3644         }
3645         return (kn->kn_fflags != 0);
3646 }
3647