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  * $DragonFly: src/sys/vfs/hammer/hammer_vnops.c,v 1.102 2008/10/16 17:24:16 dillon Exp $
  35  */
  36
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/kernel.h>
  40 #include <sys/fcntl.h>
  41 #include <sys/namecache.h>
  42 #include <sys/vnode.h>
  43 #include <sys/lockf.h>
  44 #include <sys/event.h>
  45 #include <sys/stat.h>
  46 #include <sys/dirent.h>
  47 #include <sys/file.h>
  48 #include <vm/vm_extern.h>
  49 #include <vfs/fifofs/fifo.h>
  50
  51 #include <sys/mplock2.h>
  52
  53 #include "hammer.h"
  54
  55 /*
  56  * USERFS VNOPS
  57  */
  58 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
  59 static int hammer_vop_fsync(struct vop_fsync_args *);
  60 static int hammer_vop_read(struct vop_read_args *);
  61 static int hammer_vop_write(struct vop_write_args *);
  62 static int hammer_vop_access(struct vop_access_args *);
  63 static int hammer_vop_advlock(struct vop_advlock_args *);
  64 static int hammer_vop_close(struct vop_close_args *);
  65 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  66 static int hammer_vop_getattr(struct vop_getattr_args *);
  67 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  68 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  69 static int hammer_vop_nlink(struct vop_nlink_args *);
  70 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  71 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  72 static int hammer_vop_open(struct vop_open_args *);
  73 static int hammer_vop_print(struct vop_print_args *);
  74 static int hammer_vop_readdir(struct vop_readdir_args *);
  75 static int hammer_vop_readlink(struct vop_readlink_args *);
  76 static int hammer_vop_nremove(struct vop_nremove_args *);
  77 static int hammer_vop_nrename(struct vop_nrename_args *);
  78 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  79 static int hammer_vop_markatime(struct vop_markatime_args *);
  80 static int hammer_vop_setattr(struct vop_setattr_args *);
  81 static int hammer_vop_strategy(struct vop_strategy_args *);
  82 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  83 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  84 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  85 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  86 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  87 static int hammer_vop_kqfilter (struct vop_kqfilter_args *);
  88
  89 static int hammer_vop_fifoclose (struct vop_close_args *);
  90 static int hammer_vop_fiforead (struct vop_read_args *);
  91 static int hammer_vop_fifowrite (struct vop_write_args *);
  92 static int hammer_vop_fifokqfilter (struct vop_kqfilter_args *);
  93
  94 struct vop_ops hammer_vnode_vops = {
  95         .vop_default =          vop_defaultop,
  96         .vop_fsync =            hammer_vop_fsync,
  97         .vop_getpages =         vop_stdgetpages,
  98         .vop_putpages =         vop_stdputpages,
  99         .vop_read =             hammer_vop_read,
 100         .vop_write =            hammer_vop_write,
 101         .vop_access =           hammer_vop_access,
 102         .vop_advlock =          hammer_vop_advlock,
 103         .vop_close =            hammer_vop_close,
 104         .vop_ncreate =          hammer_vop_ncreate,
 105         .vop_getattr =          hammer_vop_getattr,
 106         .vop_inactive =         hammer_vop_inactive,
 107         .vop_reclaim =          hammer_vop_reclaim,
 108         .vop_nresolve =         hammer_vop_nresolve,
 109         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
 110         .vop_nlink =            hammer_vop_nlink,
 111         .vop_nmkdir =           hammer_vop_nmkdir,
 112         .vop_nmknod =           hammer_vop_nmknod,
 113         .vop_open =             hammer_vop_open,
 114         .vop_pathconf =         vop_stdpathconf,
 115         .vop_print =            hammer_vop_print,
 116         .vop_readdir =          hammer_vop_readdir,
 117         .vop_readlink =         hammer_vop_readlink,
 118         .vop_nremove =          hammer_vop_nremove,
 119         .vop_nrename =          hammer_vop_nrename,
 120         .vop_nrmdir =           hammer_vop_nrmdir,
 121         .vop_markatime =        hammer_vop_markatime,
 122         .vop_setattr =          hammer_vop_setattr,
 123         .vop_bmap =             hammer_vop_bmap,
 124         .vop_strategy =         hammer_vop_strategy,
 125         .vop_nsymlink =         hammer_vop_nsymlink,
 126         .vop_nwhiteout =        hammer_vop_nwhiteout,
 127         .vop_ioctl =            hammer_vop_ioctl,
 128         .vop_mountctl =         hammer_vop_mountctl,
 129         .vop_kqfilter =         hammer_vop_kqfilter
 130 };
 131
 132 struct vop_ops hammer_spec_vops = {
 133         .vop_default =          vop_defaultop,
 134         .vop_fsync =            hammer_vop_fsync,
 135         .vop_read =             vop_stdnoread,
 136         .vop_write =            vop_stdnowrite,
 137         .vop_access =           hammer_vop_access,
 138         .vop_close =            hammer_vop_close,
 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 };
 145
 146 struct vop_ops hammer_fifo_vops = {
 147         .vop_default =          fifo_vnoperate,
 148         .vop_fsync =            hammer_vop_fsync,
 149         .vop_read =             hammer_vop_fiforead,
 150         .vop_write =            hammer_vop_fifowrite,
 151         .vop_access =           hammer_vop_access,
 152         .vop_close =            hammer_vop_fifoclose,
 153         .vop_markatime =        hammer_vop_markatime,
 154         .vop_getattr =          hammer_vop_getattr,
 155         .vop_inactive =         hammer_vop_inactive,
 156         .vop_reclaim =          hammer_vop_reclaim,
 157         .vop_setattr =          hammer_vop_setattr,
 158         .vop_kqfilter =         hammer_vop_fifokqfilter
 159 };
 160
 161 static __inline
 162 void
 163 hammer_knote(struct vnode *vp, int flags)
 164 {
 165         if (flags)
 166                 KNOTE(&vp->v_pollinfo.vpi_selinfo.si_note, flags);
 167 }
 168
 169 #ifdef DEBUG_TRUNCATE
 170 struct hammer_inode *HammerTruncIp;
 171 #endif
 172
 173 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 174                            struct vnode *dvp, struct ucred *cred,
 175                            int flags, int isdir);
 176 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 177 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 178
 179 #if 0
 180 static
 181 int
 182 hammer_vop_vnoperate(struct vop_generic_args *)
 183 {
 184         return (VOCALL(&hammer_vnode_vops, ap));
 185 }
 186 #endif
 187
 188 /*
 189  * hammer_vop_fsync { vp, waitfor }
 190  *
 191  * fsync() an inode to disk and wait for it to be completely committed
 192  * such that the information would not be undone if a crash occured after
 193  * return.
 194  *
 195  * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
 196  *       a REDO log.  A sysctl is provided to relax HAMMER's fsync()
 197  *       operation.
 198  *
 199  *       Ultimately the combination of a REDO log and use of fast storage
 200  *       to front-end cluster caches will make fsync fast, but it aint
 201  *       here yet.  And, in anycase, we need real transactional
 202  *       all-or-nothing features which are not restricted to a single file.
 203  */
 204 static
 205 int
 206 hammer_vop_fsync(struct vop_fsync_args *ap)
 207 {
 208         hammer_inode_t ip = VTOI(ap->a_vp);
 209         hammer_mount_t hmp = ip->hmp;
 210         int waitfor = ap->a_waitfor;
 211         int mode;
 212
 213         /*
 214          * Fsync rule relaxation (default is either full synchronous flush
 215          * or REDO semantics with synchronous flush).
 216          */
 217         if (ap->a_flags & VOP_FSYNC_SYSCALL) {
 218                 switch(hammer_fsync_mode) {
 219                 case 0:
 220 mode0:
 221                         /* disable REDO, full synchronous flush */
 222                         ip->redo_count = SIZE_T_MAX;
 223                         goto skip;
 224                 case 1:
 225 mode1:
 226                         /* disable REDO, full asynchronous flush */
 227                         ip->redo_count = SIZE_T_MAX;
 228                         if (waitfor == MNT_WAIT)
 229                                 waitfor = MNT_NOWAIT;
 230                         goto skip;
 231                 case 2:
 232                         /* REDO semantics, synchronous flush */
 233                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 234                                 goto mode0;
 235                         mode = HAMMER_FLUSH_UNDOS_AUTO;
 236                         break;
 237                 case 3:
 238                         /* REDO semantics, relaxed asynchronous flush */
 239                         if (hmp->version < HAMMER_VOL_VERSION_FOUR)
 240                                 goto mode1;
 241                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 242                         if (waitfor == MNT_WAIT)
 243                                 waitfor = MNT_NOWAIT;
 244                         break;
 245                 case 4:
 246                         /* ignore the fsync() system call */
 247                         return(0);
 248                 default:
 249                         /* we have to do something */
 250                         mode = HAMMER_FLUSH_UNDOS_RELAXED;
 251                         if (waitfor == MNT_WAIT)
 252                                 waitfor = MNT_NOWAIT;
 253                         break;
 254                 }
 255
 256                 /*
 257                  * redo_count is initialized to a maximal value and set
 258                  * to 0 after the first fsync() on a file, which enables
 259                  * REDO logging on the inode unless the number of bytes
 260                  * written exceeds the limit.
 261                  */
 262                 if (ip->redo_count < hammer_limit_redo &&
 263                     (ip->flags & HAMMER_INODE_MODMASK_NOREDO) == 0
 264                 ) {
 265                         ++hammer_count_fsyncs;
 266                         hammer_flusher_flush_undos(hmp, mode);
 267                         ip->redo_count = 0;
 268                         return(0);
 269                 }
 270                 ip->redo_count = 0;
 271         }
 272 skip:
 273
 274         /*
 275          * Do a full flush sequence.
 276          */
 277         ++hammer_count_fsyncs;
 278         vfsync(ap->a_vp, waitfor, 1, NULL, NULL);
 279         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 280         if (waitfor == MNT_WAIT) {
 281                 vn_unlock(ap->a_vp);
 282                 hammer_wait_inode(ip);
 283                 vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
 284         }
 285         return (ip->error);
 286 }
 287
 288 /*
 289  * hammer_vop_read { vp, uio, ioflag, cred }
 290  *
 291  * MPALMOSTSAFE
 292  */
 293 static
 294 int
 295 hammer_vop_read(struct vop_read_args *ap)
 296 {
 297         struct hammer_transaction trans;
 298         hammer_inode_t ip;
 299         off_t offset;
 300         struct buf *bp;
 301         struct uio *uio;
 302         int error;
 303         int n;
 304         int seqcount;
 305         int ioseqcount;
 306         int blksize;
 307         int got_mplock;
 308         int bigread;
 309
 310         if (ap->a_vp->v_type != VREG)
 311                 return (EINVAL);
 312         ip = VTOI(ap->a_vp);
 313         error = 0;
 314         uio = ap->a_uio;
 315
 316         /*
 317          * Allow the UIO's size to override the sequential heuristic.
 318          */
 319         blksize = hammer_blocksize(uio->uio_offset);
 320         seqcount = (uio->uio_resid + (blksize - 1)) / blksize;
 321         ioseqcount = ap->a_ioflag >> 16;
 322         if (seqcount < ioseqcount)
 323                 seqcount = ioseqcount;
 324
 325         /*
 326          * Temporary hack until more of HAMMER can be made MPSAFE.
 327          */
 328 #ifdef SMP
 329         if (curthread->td_mpcount) {
 330                 got_mplock = -1;
 331                 hammer_start_transaction(&trans, ip->hmp);
 332         } else {
 333                 got_mplock = 0;
 334         }
 335 #else
 336         hammer_start_transaction(&trans, ip->hmp);
 337         got_mplock = -1;
 338 #endif
 339
 340         /*
 341          * If reading or writing a huge amount of data we have to break
 342          * atomicy and allow the operation to be interrupted by a signal
 343          * or it can DOS the machine.
 344          */
 345         bigread = (uio->uio_resid > 100 * 1024 * 1024);
 346
 347         /*
 348          * Access the data typically in HAMMER_BUFSIZE blocks via the
 349          * buffer cache, but HAMMER may use a variable block size based
 350          * on the offset.
 351          *
 352          * XXX Temporary hack, delay the start transaction while we remain
 353          *     MPSAFE.  NOTE: ino_data.size cannot change while vnode is
 354          *     locked-shared.
 355          */
 356         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 357                 int64_t base_offset;
 358                 int64_t file_limit;
 359
 360                 blksize = hammer_blocksize(uio->uio_offset);
 361                 offset = (int)uio->uio_offset & (blksize - 1);
 362                 base_offset = uio->uio_offset - offset;
 363
 364                 if (bigread && (error = hammer_signal_check(ip->hmp)) != 0)
 365                         break;
 366
 367                 /*
 368                  * MPSAFE
 369                  */
 370                 bp = getcacheblk(ap->a_vp, base_offset);
 371                 if (bp) {
 372                         error = 0;
 373                         goto skip;
 374                 }
 375
 376                 /*
 377                  * MPUNSAFE
 378                  */
 379                 if (got_mplock == 0) {
 380                         got_mplock = 1;
 381                         get_mplock();
 382                         hammer_start_transaction(&trans, ip->hmp);
 383                 }
 384
 385                 if (hammer_cluster_enable) {
 386                         /*
 387                          * Use file_limit to prevent cluster_read() from
 388                          * creating buffers of the wrong block size past
 389                          * the demarc.
 390                          */
 391                         file_limit = ip->ino_data.size;
 392                         if (base_offset < HAMMER_XDEMARC &&
 393                             file_limit > HAMMER_XDEMARC) {
 394                                 file_limit = HAMMER_XDEMARC;
 395                         }
 396                         error = cluster_read(ap->a_vp,
 397                                              file_limit, base_offset,
 398                                              blksize, MAXPHYS,
 399                                              seqcount, &bp);
 400                 } else {
 401                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 402                 }
 403                 if (error) {
 404                         brelse(bp);
 405                         break;
 406                 }
 407 skip:
 408
 409                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 410                 n = blksize - offset;
 411                 if (n > uio->uio_resid)
 412                         n = uio->uio_resid;
 413                 if (n > ip->ino_data.size - uio->uio_offset)
 414                         n = (int)(ip->ino_data.size - uio->uio_offset);
 415                 error = uiomove((char *)bp->b_data + offset, n, uio);
 416
 417                 /* data has a lower priority then meta-data */
 418                 bp->b_flags |= B_AGE;
 419                 bqrelse(bp);
 420                 if (error)
 421                         break;
 422                 hammer_stats_file_read += n;
 423         }
 424
 425         /*
 426          * XXX only update the atime if we had to get the MP lock.
 427          * XXX hack hack hack, fixme.
 428          */
 429         if (got_mplock) {
 430                 if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 431                     (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 432                         ip->ino_data.atime = trans.time;
 433                         hammer_modify_inode(ip, HAMMER_INODE_ATIME);
 434                 }
 435                 hammer_done_transaction(&trans);
 436                 if (got_mplock > 0)
 437                         rel_mplock();
 438         }
 439         return (error);
 440 }
 441
 442 /*
 443  * hammer_vop_write { vp, uio, ioflag, cred }
 444  */
 445 static
 446 int
 447 hammer_vop_write(struct vop_write_args *ap)
 448 {
 449         struct hammer_transaction trans;
 450         struct hammer_inode *ip;
 451         hammer_mount_t hmp;
 452         struct uio *uio;
 453         int offset;
 454         off_t base_offset;
 455         struct buf *bp;
 456         int kflags;
 457         int error;
 458         int n;
 459         int flags;
 460         int seqcount;
 461         int bigwrite;
 462
 463         if (ap->a_vp->v_type != VREG)
 464                 return (EINVAL);
 465         ip = VTOI(ap->a_vp);
 466         hmp = ip->hmp;
 467         error = 0;
 468         kflags = 0;
 469         seqcount = ap->a_ioflag >> 16;
 470
 471         if (ip->flags & HAMMER_INODE_RO)
 472                 return (EROFS);
 473
 474         /*
 475          * Create a transaction to cover the operations we perform.
 476          */
 477         hammer_start_transaction(&trans, hmp);
 478         uio = ap->a_uio;
 479
 480         /*
 481          * Check append mode
 482          */
 483         if (ap->a_ioflag & IO_APPEND)
 484                 uio->uio_offset = ip->ino_data.size;
 485
 486         /*
 487          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 488          *
 489          * NOTE: the base_off assignment is required to work around what
 490          * I consider to be a GCC-4 optimization bug.
 491          */
 492         if (uio->uio_offset < 0) {
 493                 hammer_done_transaction(&trans);
 494                 return (EFBIG);
 495         }
 496         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 497         if (uio->uio_resid > 0 && base_offset <= uio->uio_offset) {
 498                 hammer_done_transaction(&trans);
 499                 return (EFBIG);
 500         }
 501
 502         /*
 503          * If reading or writing a huge amount of data we have to break
 504          * atomicy and allow the operation to be interrupted by a signal
 505          * or it can DOS the machine.
 506          *
 507          * Adjust redo_count early to avoid generating unnecessary redos.
 508          */
 509         bigwrite = (uio->uio_resid > 100 * 1024 * 1024);
 510         if (ip->redo_count < hammer_limit_redo)
 511                 ip->redo_count += uio->uio_resid;
 512
 513         /*
 514          * Access the data typically in HAMMER_BUFSIZE blocks via the
 515          * buffer cache, but HAMMER may use a variable block size based
 516          * on the offset.
 517          */
 518         while (uio->uio_resid > 0) {
 519                 int fixsize = 0;
 520                 int blksize;
 521                 int blkmask;
 522
 523                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 524                         break;
 525                 if (bigwrite && (error = hammer_signal_check(hmp)) != 0)
 526                         break;
 527
 528                 blksize = hammer_blocksize(uio->uio_offset);
 529
 530                 /*
 531                  * Do not allow HAMMER to blow out the buffer cache.  Very
 532                  * large UIOs can lockout other processes due to bwillwrite()
 533                  * mechanics.
 534                  *
 535                  * The hammer inode is not locked during these operations.
 536                  * The vnode is locked which can interfere with the pageout
 537                  * daemon for non-UIO_NOCOPY writes but should not interfere
 538                  * with the buffer cache.  Even so, we cannot afford to
 539                  * allow the pageout daemon to build up too many dirty buffer
 540                  * cache buffers.
 541                  *
 542                  * Only call this if we aren't being recursively called from
 543                  * a virtual disk device (vn), else we may deadlock.
 544                  */
 545                 if ((ap->a_ioflag & IO_RECURSE) == 0)
 546                         bwillwrite(blksize);
 547
 548                 /*
 549                  * Control the number of pending records associated with
 550                  * this inode.  If too many have accumulated start a
 551                  * flush.  Try to maintain a pipeline with the flusher.
 552                  */
 553                 if (ip->rsv_recs >= hammer_limit_inode_recs) {
 554                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 555                 }
 556                 if (ip->rsv_recs >= hammer_limit_inode_recs * 2) {
 557                         while (ip->rsv_recs >= hammer_limit_inode_recs) {
 558                                 tsleep(&ip->rsv_recs, 0, "hmrwww", hz);
 559                         }
 560                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 561                 }
 562
 563 #if 0
 564                 /*
 565                  * Do not allow HAMMER to blow out system memory by
 566                  * accumulating too many records.   Records are so well
 567                  * decoupled from the buffer cache that it is possible
 568                  * for userland to push data out to the media via
 569                  * direct-write, but build up the records queued to the
 570                  * backend faster then the backend can flush them out.
 571                  * HAMMER has hit its write limit but the frontend has
 572                  * no pushback to slow it down.
 573                  */
 574                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
 575                         /*
 576                          * Get the inode on the flush list
 577                          */
 578                         if (ip->rsv_recs >= 64)
 579                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 580                         else if (ip->rsv_recs >= 16)
 581                                 hammer_flush_inode(ip, 0);
 582
 583                         /*
 584                          * Keep the flusher going if the system keeps
 585                          * queueing records.
 586                          */
 587                         delta = hmp->count_newrecords -
 588                                 hmp->last_newrecords;
 589                         if (delta < 0 || delta > hammer_limit_recs / 2) {
 590                                 hmp->last_newrecords = hmp->count_newrecords;
 591                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
 592                         }
 593
 594                         /*
 595                          * If we have gotten behind start slowing
 596                          * down the writers.
 597                          */
 598                         delta = (hmp->rsv_recs - hammer_limit_recs) *
 599                                 hz / hammer_limit_recs;
 600                         if (delta > 0)
 601                                 tsleep(&trans, 0, "hmrslo", delta);
 602                 }
 603 #endif
 604
 605                 /*
 606                  * Calculate the blocksize at the current offset and figure
 607                  * out how much we can actually write.
 608                  */
 609                 blkmask = blksize - 1;
 610                 offset = (int)uio->uio_offset & blkmask;
 611                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 612                 n = blksize - offset;
 613                 if (n > uio->uio_resid)
 614                         n = uio->uio_resid;
 615                 if (uio->uio_offset + n > ip->ino_data.size) {
 616                         vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
 617                         fixsize = 1;
 618                         kflags |= NOTE_EXTEND;
 619                 }
 620
 621                 if (uio->uio_segflg == UIO_NOCOPY) {
 622                         /*
 623                          * Issuing a write with the same data backing the
 624                          * buffer.  Instantiate the buffer to collect the
 625                          * backing vm pages, then read-in any missing bits.
 626                          *
 627                          * This case is used by vop_stdputpages().
 628                          */
 629                         bp = getblk(ap->a_vp, base_offset,
 630                                     blksize, GETBLK_BHEAVY, 0);
 631                         if ((bp->b_flags & B_CACHE) == 0) {
 632                                 bqrelse(bp);
 633                                 error = bread(ap->a_vp, base_offset,
 634                                               blksize, &bp);
 635                         }
 636                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 637                         /*
 638                          * Even though we are entirely overwriting the buffer
 639                          * we may still have to zero it out to avoid a
 640                          * mmap/write visibility issue.
 641                          */
 642                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 643                         if ((bp->b_flags & B_CACHE) == 0)
 644                                 vfs_bio_clrbuf(bp);
 645                 } else if (base_offset >= ip->ino_data.size) {
 646                         /*
 647                          * If the base offset of the buffer is beyond the
 648                          * file EOF, we don't have to issue a read.
 649                          */
 650                         bp = getblk(ap->a_vp, base_offset,
 651                                     blksize, GETBLK_BHEAVY, 0);
 652                         vfs_bio_clrbuf(bp);
 653                 } else {
 654                         /*
 655                          * Partial overwrite, read in any missing bits then
 656                          * replace the portion being written.
 657                          */
 658                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 659                         if (error == 0)
 660                                 bheavy(bp);
 661                 }
 662                 if (error == 0)
 663                         error = uiomove(bp->b_data + offset, n, uio);
 664
 665                 /*
 666                  * Generate REDO records while redo_count has not exceeded
 667                  * the limit.  Note that redo_count is initialized to a
 668                  * maximal value until the first fsync(), and zerod on every
 669                  * fsync().  Thus at least one fsync() is required before we
 670                  * start generating REDO records for the ip.
 671                  */
 672                 if (hmp->version >= HAMMER_VOL_VERSION_FOUR &&
 673                     ip->redo_count < hammer_limit_redo &&
 674                     error == 0) {
 675                         hammer_sync_lock_sh(&trans);
 676                         error = hammer_generate_redo(&trans, ip,
 677                                                      base_offset + offset,
 678                                                      bp->b_data + offset,
 679                                                      (size_t)n);
 680                         hammer_sync_unlock(&trans);
 681                 }
 682
 683                 /*
 684                  * If we screwed up we have to undo any VM size changes we
 685                  * made.
 686                  */
 687                 if (error) {
 688                         brelse(bp);
 689                         if (fixsize) {
 690                                 vtruncbuf(ap->a_vp, ip->ino_data.size,
 691                                           hammer_blocksize(ip->ino_data.size));
 692                         }
 693                         break;
 694                 }
 695                 kflags |= NOTE_WRITE;
 696                 hammer_stats_file_write += n;
 697                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 698                 if (ip->ino_data.size < uio->uio_offset) {
 699                         ip->ino_data.size = uio->uio_offset;
 700                         flags = HAMMER_INODE_SDIRTY;
 701                         vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
 702                 } else {
 703                         flags = 0;
 704                 }
 705                 ip->ino_data.mtime = trans.time;
 706                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 707                 hammer_modify_inode(ip, flags);
 708
 709                 /*
 710                  * Once we dirty the buffer any cached zone-X offset
 711                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 712                  * allow overwriting over the same data sector unless
 713                  * we provide UNDOs for the old data, which we don't.
 714                  */
 715                 bp->b_bio2.bio_offset = NOOFFSET;
 716
 717                 /*
 718                  * Final buffer disposition.
 719                  *
 720                  * Because meta-data updates are deferred, HAMMER is
 721                  * especially sensitive to excessive bdwrite()s because
 722                  * the I/O stream is not broken up by disk reads.  So the
 723                  * buffer cache simply cannot keep up.
 724                  *
 725                  * WARNING!  blksize is variable.  cluster_write() is
 726                  * expected to not blow up if it encounters buffers that
 727                  * do not match the passed blksize.
 728                  *
 729                  * NOTE!  Hammer shouldn't need to bawrite()/cluster_write().
 730                  *        The ip->rsv_recs check should burst-flush the data.
 731                  *        If we queue it immediately the buf could be left
 732                  *        locked on the device queue for a very long time.
 733                  */
 734                 bp->b_flags |= B_AGE;
 735                 if (ap->a_ioflag & IO_SYNC) {
 736                         bwrite(bp);
 737                 } else if (ap->a_ioflag & IO_DIRECT) {
 738                         bawrite(bp);
 739                 } else {
 740 #if 0
 741                 if (offset + n == blksize) {
 742                         if (hammer_cluster_enable == 0 ||
 743                             (ap->a_vp->v_mount->mnt_flag & MNT_NOCLUSTERW)) {
 744                                 bawrite(bp);
 745                         } else {
 746                                 cluster_write(bp, ip->ino_data.size,
 747                                               blksize, seqcount);
 748                         }
 749                 } else {
 750 #endif
 751                         bdwrite(bp);
 752                 }
 753         }
 754         hammer_done_transaction(&trans);
 755         hammer_knote(ap->a_vp, kflags);
 756         return (error);
 757 }
 758
 759 /*
 760  * hammer_vop_access { vp, mode, cred }
 761  */
 762 static
 763 int
 764 hammer_vop_access(struct vop_access_args *ap)
 765 {
 766         struct hammer_inode *ip = VTOI(ap->a_vp);
 767         uid_t uid;
 768         gid_t gid;
 769         int error;
 770
 771         ++hammer_stats_file_iopsr;
 772         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 773         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 774
 775         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 776                                   ip->ino_data.uflags);
 777         return (error);
 778 }
 779
 780 /*
 781  * hammer_vop_advlock { vp, id, op, fl, flags }
 782  */
 783 static
 784 int
 785 hammer_vop_advlock(struct vop_advlock_args *ap)
 786 {
 787         hammer_inode_t ip = VTOI(ap->a_vp);
 788
 789         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 790 }
 791
 792 /*
 793  * hammer_vop_close { vp, fflag }
 794  *
 795  * We can only sync-on-close for normal closes.
 796  */
 797 static
 798 int
 799 hammer_vop_close(struct vop_close_args *ap)
 800 {
 801 #if 0
 802         struct vnode *vp = ap->a_vp;
 803         hammer_inode_t ip = VTOI(vp);
 804         int waitfor;
 805         if (ip->flags & (HAMMER_INODE_CLOSESYNC|HAMMER_INODE_CLOSEASYNC)) {
 806                 if (vn_islocked(vp) == LK_EXCLUSIVE &&
 807                     (vp->v_flag & (VINACTIVE|VRECLAIMED)) == 0) {
 808                         if (ip->flags & HAMMER_INODE_CLOSESYNC)
 809                                 waitfor = MNT_WAIT;
 810                         else
 811                                 waitfor = MNT_NOWAIT;
 812                         ip->flags &= ~(HAMMER_INODE_CLOSESYNC |
 813                                        HAMMER_INODE_CLOSEASYNC);
 814                         VOP_FSYNC(vp, MNT_NOWAIT, waitfor);
 815                 }
 816         }
 817 #endif
 818         return (vop_stdclose(ap));
 819 }
 820
 821 /*
 822  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 823  *
 824  * The operating system has already ensured that the directory entry
 825  * does not exist and done all appropriate namespace locking.
 826  */
 827 static
 828 int
 829 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 830 {
 831         struct hammer_transaction trans;
 832         struct hammer_inode *dip;
 833         struct hammer_inode *nip;
 834         struct nchandle *nch;
 835         int error;
 836
 837         nch = ap->a_nch;
 838         dip = VTOI(ap->a_dvp);
 839
 840         if (dip->flags & HAMMER_INODE_RO)
 841                 return (EROFS);
 842         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
 843                 return (error);
 844
 845         /*
 846          * Create a transaction to cover the operations we perform.
 847          */
 848         hammer_start_transaction(&trans, dip->hmp);
 849         ++hammer_stats_file_iopsw;
 850
 851         /*
 852          * Create a new filesystem object of the requested type.  The
 853          * returned inode will be referenced and shared-locked to prevent
 854          * it from being moved to the flusher.
 855          */
 856         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 857                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
 858                                     NULL, &nip);
 859         if (error) {
 860                 hkprintf("hammer_create_inode error %d\n", error);
 861                 hammer_done_transaction(&trans);
 862                 *ap->a_vpp = NULL;
 863                 return (error);
 864         }
 865
 866         /*
 867          * Add the new filesystem object to the directory.  This will also
 868          * bump the inode's link count.
 869          */
 870         error = hammer_ip_add_directory(&trans, dip,
 871                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 872                                         nip);
 873         if (error)
 874                 hkprintf("hammer_ip_add_directory error %d\n", error);
 875
 876         /*
 877          * Finish up.
 878          */
 879         if (error) {
 880                 hammer_rel_inode(nip, 0);
 881                 hammer_done_transaction(&trans);
 882                 *ap->a_vpp = NULL;
 883         } else {
 884                 error = hammer_get_vnode(nip, ap->a_vpp);
 885                 hammer_done_transaction(&trans);
 886                 hammer_rel_inode(nip, 0);
 887                 if (error == 0) {
 888                         cache_setunresolved(ap->a_nch);
 889                         cache_setvp(ap->a_nch, *ap->a_vpp);
 890                 }
 891                 hammer_knote(ap->a_dvp, NOTE_WRITE);
 892         }
 893         return (error);
 894 }
 895
 896 /*
 897  * hammer_vop_getattr { vp, vap }
 898  *
 899  * Retrieve an inode's attribute information.  When accessing inodes
 900  * historically we fake the atime field to ensure consistent results.
 901  * The atime field is stored in the B-Tree element and allowed to be
 902  * updated without cycling the element.
 903  *
 904  * MPSAFE
 905  */
 906 static
 907 int
 908 hammer_vop_getattr(struct vop_getattr_args *ap)
 909 {
 910         struct hammer_inode *ip = VTOI(ap->a_vp);
 911         struct vattr *vap = ap->a_vap;
 912
 913         /*
 914          * We want the fsid to be different when accessing a filesystem
 915          * with different as-of's so programs like diff don't think
 916          * the files are the same.
 917          *
 918          * We also want the fsid to be the same when comparing snapshots,
 919          * or when comparing mirrors (which might be backed by different
 920          * physical devices).  HAMMER fsids are based on the PFS's
 921          * shared_uuid field.
 922          *
 923          * XXX there is a chance of collision here.  The va_fsid reported
 924          * by stat is different from the more involved fsid used in the
 925          * mount structure.
 926          */
 927         ++hammer_stats_file_iopsr;
 928         hammer_lock_sh(&ip->lock);
 929         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
 930                        (u_int32_t)(ip->obj_asof >> 32);
 931
 932         vap->va_fileid = ip->ino_leaf.base.obj_id;
 933         vap->va_mode = ip->ino_data.mode;
 934         vap->va_nlink = ip->ino_data.nlinks;
 935         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
 936         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
 937         vap->va_rmajor = 0;
 938         vap->va_rminor = 0;
 939         vap->va_size = ip->ino_data.size;
 940
 941         /*
 942          * Special case for @@PFS softlinks.  The actual size of the
 943          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
 944          * or for MAX_TID is    "@@-1:%05d" == 10 bytes.
 945          */
 946         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
 947             ip->ino_data.size == 10 &&
 948             ip->obj_asof == HAMMER_MAX_TID &&
 949             ip->obj_localization == 0 &&
 950             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
 951                     if (ip->pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE)
 952                             vap->va_size = 26;
 953                     else
 954                             vap->va_size = 10;
 955         }
 956
 957         /*
 958          * We must provide a consistent atime and mtime for snapshots
 959          * so people can do a 'tar cf - ... | md5' on them and get
 960          * consistent results.
 961          */
 962         if (ip->flags & HAMMER_INODE_RO) {
 963                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
 964                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
 965         } else {
 966                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
 967                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
 968         }
 969         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
 970         vap->va_flags = ip->ino_data.uflags;
 971         vap->va_gen = 1;        /* hammer inums are unique for all time */
 972         vap->va_blocksize = HAMMER_BUFSIZE;
 973         if (ip->ino_data.size >= HAMMER_XDEMARC) {
 974                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
 975                                 ~HAMMER_XBUFMASK64;
 976         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
 977                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
 978                                 ~HAMMER_BUFMASK64;
 979         } else {
 980                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
 981         }
 982
 983         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
 984         vap->va_filerev = 0;    /* XXX */
 985         vap->va_uid_uuid = ip->ino_data.uid;
 986         vap->va_gid_uuid = ip->ino_data.gid;
 987         vap->va_fsid_uuid = ip->hmp->fsid;
 988         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
 989                           VA_FSID_UUID_VALID;
 990
 991         switch (ip->ino_data.obj_type) {
 992         case HAMMER_OBJTYPE_CDEV:
 993         case HAMMER_OBJTYPE_BDEV:
 994                 vap->va_rmajor = ip->ino_data.rmajor;
 995                 vap->va_rminor = ip->ino_data.rminor;
 996                 break;
 997         default:
 998                 break;
 999         }
1000         hammer_unlock(&ip->lock);
1001         return(0);
1002 }
1003
1004 /*
1005  * hammer_vop_nresolve { nch, dvp, cred }
1006  *
1007  * Locate the requested directory entry.
1008  */
1009 static
1010 int
1011 hammer_vop_nresolve(struct vop_nresolve_args *ap)
1012 {
1013         struct hammer_transaction trans;
1014         struct namecache *ncp;
1015         hammer_inode_t dip;
1016         hammer_inode_t ip;
1017         hammer_tid_t asof;
1018         struct hammer_cursor cursor;
1019         struct vnode *vp;
1020         int64_t namekey;
1021         int error;
1022         int i;
1023         int nlen;
1024         int flags;
1025         int ispfs;
1026         int64_t obj_id;
1027         u_int32_t localization;
1028         u_int32_t max_iterations;
1029
1030         /*
1031          * Misc initialization, plus handle as-of name extensions.  Look for
1032          * the '@@' extension.  Note that as-of files and directories cannot
1033          * be modified.
1034          */
1035         dip = VTOI(ap->a_dvp);
1036         ncp = ap->a_nch->ncp;
1037         asof = dip->obj_asof;
1038         localization = dip->obj_localization;   /* for code consistency */
1039         nlen = ncp->nc_nlen;
1040         flags = dip->flags & HAMMER_INODE_RO;
1041         ispfs = 0;
1042
1043         hammer_simple_transaction(&trans, dip->hmp);
1044         ++hammer_stats_file_iopsr;
1045
1046         for (i = 0; i < nlen; ++i) {
1047                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
1048                         error = hammer_str_to_tid(ncp->nc_name + i + 2,
1049                                                   &ispfs, &asof, &localization);
1050                         if (error != 0) {
1051                                 i = nlen;
1052                                 break;
1053                         }
1054                         if (asof != HAMMER_MAX_TID)
1055                                 flags |= HAMMER_INODE_RO;
1056                         break;
1057                 }
1058         }
1059         nlen = i;
1060
1061         /*
1062          * If this is a PFS softlink we dive into the PFS
1063          */
1064         if (ispfs && nlen == 0) {
1065                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
1066                                       asof, localization,
1067                                       flags, &error);
1068                 if (error == 0) {
1069                         error = hammer_get_vnode(ip, &vp);
1070                         hammer_rel_inode(ip, 0);
1071                 } else {
1072                         vp = NULL;
1073                 }
1074                 if (error == 0) {
1075                         vn_unlock(vp);
1076                         cache_setvp(ap->a_nch, vp);
1077                         vrele(vp);
1078                 }
1079                 goto done;
1080         }
1081
1082         /*
1083          * If there is no path component the time extension is relative to dip.
1084          * e.g. "fubar/@@<snapshot>"
1085          *
1086          * "." is handled by the kernel, but ".@@<snapshot>" is not.
1087          * e.g. "fubar/.@@<snapshot>"
1088          *
1089          * ".." is handled by the kernel.  We do not currently handle
1090          * "..@<snapshot>".
1091          */
1092         if (nlen == 0 || (nlen == 1 && ncp->nc_name[0] == '.')) {
1093                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
1094                                       asof, dip->obj_localization,
1095                                       flags, &error);
1096                 if (error == 0) {
1097                         error = hammer_get_vnode(ip, &vp);
1098                         hammer_rel_inode(ip, 0);
1099                 } else {
1100                         vp = NULL;
1101                 }
1102                 if (error == 0) {
1103                         vn_unlock(vp);
1104                         cache_setvp(ap->a_nch, vp);
1105                         vrele(vp);
1106                 }
1107                 goto done;
1108         }
1109
1110         /*
1111          * Calculate the namekey and setup the key range for the scan.  This
1112          * works kinda like a chained hash table where the lower 32 bits
1113          * of the namekey synthesize the chain.
1114          *
1115          * The key range is inclusive of both key_beg and key_end.
1116          */
1117         namekey = hammer_directory_namekey(dip, ncp->nc_name, nlen,
1118                                            &max_iterations);
1119
1120         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
1121         cursor.key_beg.localization = dip->obj_localization +
1122                                       hammer_dir_localization(dip);
1123         cursor.key_beg.obj_id = dip->obj_id;
1124         cursor.key_beg.key = namekey;
1125         cursor.key_beg.create_tid = 0;
1126         cursor.key_beg.delete_tid = 0;
1127         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1128         cursor.key_beg.obj_type = 0;
1129
1130         cursor.key_end = cursor.key_beg;
1131         cursor.key_end.key += max_iterations;
1132         cursor.asof = asof;
1133         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1134
1135         /*
1136          * Scan all matching records (the chain), locate the one matching
1137          * the requested path component.
1138          *
1139          * The hammer_ip_*() functions merge in-memory records with on-disk
1140          * records for the purposes of the search.
1141          */
1142         obj_id = 0;
1143         localization = HAMMER_DEF_LOCALIZATION;
1144
1145         if (error == 0) {
1146                 error = hammer_ip_first(&cursor);
1147                 while (error == 0) {
1148                         error = hammer_ip_resolve_data(&cursor);
1149                         if (error)
1150                                 break;
1151                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
1152                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1153                                 obj_id = cursor.data->entry.obj_id;
1154                                 localization = cursor.data->entry.localization;
1155                                 break;
1156                         }
1157                         error = hammer_ip_next(&cursor);
1158                 }
1159         }
1160         hammer_done_cursor(&cursor);
1161
1162         /*
1163          * Lookup the obj_id.  This should always succeed.  If it does not
1164          * the filesystem may be damaged and we return a dummy inode.
1165          */
1166         if (error == 0) {
1167                 ip = hammer_get_inode(&trans, dip, obj_id,
1168                                       asof, localization,
1169                                       flags, &error);
1170                 if (error == ENOENT) {
1171                         kprintf("HAMMER: WARNING: Missing "
1172                                 "inode for dirent \"%s\"\n"
1173                                 "\tobj_id = %016llx, asof=%016llx, lo=%08x\n",
1174                                 ncp->nc_name,
1175                                 (long long)obj_id, (long long)asof,
1176                                 localization);
1177                         error = 0;
1178                         ip = hammer_get_dummy_inode(&trans, dip, obj_id,
1179                                                     asof, localization,
1180                                                     flags, &error);
1181                 }
1182                 if (error == 0) {
1183                         error = hammer_get_vnode(ip, &vp);
1184                         hammer_rel_inode(ip, 0);
1185                 } else {
1186                         vp = NULL;
1187                 }
1188                 if (error == 0) {
1189                         vn_unlock(vp);
1190                         cache_setvp(ap->a_nch, vp);
1191                         vrele(vp);
1192                 }
1193         } else if (error == ENOENT) {
1194                 cache_setvp(ap->a_nch, NULL);
1195         }
1196 done:
1197         hammer_done_transaction(&trans);
1198         return (error);
1199 }
1200
1201 /*
1202  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1203  *
1204  * Locate the parent directory of a directory vnode.
1205  *
1206  * dvp is referenced but not locked.  *vpp must be returned referenced and
1207  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
1208  * at the root, instead it could indicate that the directory we were in was
1209  * removed.
1210  *
1211  * NOTE: as-of sequences are not linked into the directory structure.  If
1212  * we are at the root with a different asof then the mount point, reload
1213  * the same directory with the mount point's asof.   I'm not sure what this
1214  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
1215  * get confused, but it hasn't been tested.
1216  */
1217 static
1218 int
1219 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
1220 {
1221         struct hammer_transaction trans;
1222         struct hammer_inode *dip;
1223         struct hammer_inode *ip;
1224         int64_t parent_obj_id;
1225         u_int32_t parent_obj_localization;
1226         hammer_tid_t asof;
1227         int error;
1228
1229         dip = VTOI(ap->a_dvp);
1230         asof = dip->obj_asof;
1231
1232         /*
1233          * Whos are parent?  This could be the root of a pseudo-filesystem
1234          * whos parent is in another localization domain.
1235          */
1236         parent_obj_id = dip->ino_data.parent_obj_id;
1237         if (dip->obj_id == HAMMER_OBJID_ROOT)
1238                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
1239         else
1240                 parent_obj_localization = dip->obj_localization;
1241
1242         if (parent_obj_id == 0) {
1243                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
1244                    asof != dip->hmp->asof) {
1245                         parent_obj_id = dip->obj_id;
1246                         asof = dip->hmp->asof;
1247                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
1248                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
1249                                   (long long)dip->obj_asof);
1250                 } else {
1251                         *ap->a_vpp = NULL;
1252                         return ENOENT;
1253                 }
1254         }
1255
1256         hammer_simple_transaction(&trans, dip->hmp);
1257         ++hammer_stats_file_iopsr;
1258
1259         ip = hammer_get_inode(&trans, dip, parent_obj_id,
1260                               asof, parent_obj_localization,
1261                               dip->flags, &error);
1262         if (ip) {
1263                 error = hammer_get_vnode(ip, ap->a_vpp);
1264                 hammer_rel_inode(ip, 0);
1265         } else {
1266                 *ap->a_vpp = NULL;
1267         }
1268         hammer_done_transaction(&trans);
1269         return (error);
1270 }
1271
1272 /*
1273  * hammer_vop_nlink { nch, dvp, vp, cred }
1274  */
1275 static
1276 int
1277 hammer_vop_nlink(struct vop_nlink_args *ap)
1278 {
1279         struct hammer_transaction trans;
1280         struct hammer_inode *dip;
1281         struct hammer_inode *ip;
1282         struct nchandle *nch;
1283         int error;
1284
1285         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1286                 return(EXDEV);
1287
1288         nch = ap->a_nch;
1289         dip = VTOI(ap->a_dvp);
1290         ip = VTOI(ap->a_vp);
1291
1292         if (dip->obj_localization != ip->obj_localization)
1293                 return(EXDEV);
1294
1295         if (dip->flags & HAMMER_INODE_RO)
1296                 return (EROFS);
1297         if (ip->flags & HAMMER_INODE_RO)
1298                 return (EROFS);
1299         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1300                 return (error);
1301
1302         /*
1303          * Create a transaction to cover the operations we perform.
1304          */
1305         hammer_start_transaction(&trans, dip->hmp);
1306         ++hammer_stats_file_iopsw;
1307
1308         /*
1309          * Add the filesystem object to the directory.  Note that neither
1310          * dip nor ip are referenced or locked, but their vnodes are
1311          * referenced.  This function will bump the inode's link count.
1312          */
1313         error = hammer_ip_add_directory(&trans, dip,
1314                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1315                                         ip);
1316
1317         /*
1318          * Finish up.
1319          */
1320         if (error == 0) {
1321                 cache_setunresolved(nch);
1322                 cache_setvp(nch, ap->a_vp);
1323         }
1324         hammer_done_transaction(&trans);
1325         hammer_knote(ap->a_vp, NOTE_LINK);
1326         hammer_knote(ap->a_dvp, NOTE_WRITE);
1327         return (error);
1328 }
1329
1330 /*
1331  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1332  *
1333  * The operating system has already ensured that the directory entry
1334  * does not exist and done all appropriate namespace locking.
1335  */
1336 static
1337 int
1338 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1339 {
1340         struct hammer_transaction trans;
1341         struct hammer_inode *dip;
1342         struct hammer_inode *nip;
1343         struct nchandle *nch;
1344         int error;
1345
1346         nch = ap->a_nch;
1347         dip = VTOI(ap->a_dvp);
1348
1349         if (dip->flags & HAMMER_INODE_RO)
1350                 return (EROFS);
1351         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1352                 return (error);
1353
1354         /*
1355          * Create a transaction to cover the operations we perform.
1356          */
1357         hammer_start_transaction(&trans, dip->hmp);
1358         ++hammer_stats_file_iopsw;
1359
1360         /*
1361          * Create a new filesystem object of the requested type.  The
1362          * returned inode will be referenced but not locked.
1363          */
1364         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1365                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1366                                     NULL, &nip);
1367         if (error) {
1368                 hkprintf("hammer_mkdir error %d\n", error);
1369                 hammer_done_transaction(&trans);
1370                 *ap->a_vpp = NULL;
1371                 return (error);
1372         }
1373         /*
1374          * Add the new filesystem object to the directory.  This will also
1375          * bump the inode's link count.
1376          */
1377         error = hammer_ip_add_directory(&trans, dip,
1378                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1379                                         nip);
1380         if (error)
1381                 hkprintf("hammer_mkdir (add) error %d\n", error);
1382
1383         /*
1384          * Finish up.
1385          */
1386         if (error) {
1387                 hammer_rel_inode(nip, 0);
1388                 *ap->a_vpp = NULL;
1389         } else {
1390                 error = hammer_get_vnode(nip, ap->a_vpp);
1391                 hammer_rel_inode(nip, 0);
1392                 if (error == 0) {
1393                         cache_setunresolved(ap->a_nch);
1394                         cache_setvp(ap->a_nch, *ap->a_vpp);
1395                 }
1396         }
1397         hammer_done_transaction(&trans);
1398         if (error == 0)
1399                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1400         return (error);
1401 }
1402
1403 /*
1404  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1405  *
1406  * The operating system has already ensured that the directory entry
1407  * does not exist and done all appropriate namespace locking.
1408  */
1409 static
1410 int
1411 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1412 {
1413         struct hammer_transaction trans;
1414         struct hammer_inode *dip;
1415         struct hammer_inode *nip;
1416         struct nchandle *nch;
1417         int error;
1418
1419         nch = ap->a_nch;
1420         dip = VTOI(ap->a_dvp);
1421
1422         if (dip->flags & HAMMER_INODE_RO)
1423                 return (EROFS);
1424         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1425                 return (error);
1426
1427         /*
1428          * Create a transaction to cover the operations we perform.
1429          */
1430         hammer_start_transaction(&trans, dip->hmp);
1431         ++hammer_stats_file_iopsw;
1432
1433         /*
1434          * Create a new filesystem object of the requested type.  The
1435          * returned inode will be referenced but not locked.
1436          *
1437          * If mknod specifies a directory a pseudo-fs is created.
1438          */
1439         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1440                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
1441                                     NULL, &nip);
1442         if (error) {
1443                 hammer_done_transaction(&trans);
1444                 *ap->a_vpp = NULL;
1445                 return (error);
1446         }
1447
1448         /*
1449          * Add the new filesystem object to the directory.  This will also
1450          * bump the inode's link count.
1451          */
1452         error = hammer_ip_add_directory(&trans, dip,
1453                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1454                                         nip);
1455
1456         /*
1457          * Finish up.
1458          */
1459         if (error) {
1460                 hammer_rel_inode(nip, 0);
1461                 *ap->a_vpp = NULL;
1462         } else {
1463                 error = hammer_get_vnode(nip, ap->a_vpp);
1464                 hammer_rel_inode(nip, 0);
1465                 if (error == 0) {
1466                         cache_setunresolved(ap->a_nch);
1467                         cache_setvp(ap->a_nch, *ap->a_vpp);
1468                 }
1469         }
1470         hammer_done_transaction(&trans);
1471         if (error == 0)
1472                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1473         return (error);
1474 }
1475
1476 /*
1477  * hammer_vop_open { vp, mode, cred, fp }
1478  */
1479 static
1480 int
1481 hammer_vop_open(struct vop_open_args *ap)
1482 {
1483         hammer_inode_t ip;
1484
1485         ++hammer_stats_file_iopsr;
1486         ip = VTOI(ap->a_vp);
1487
1488         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1489                 return (EROFS);
1490         return(vop_stdopen(ap));
1491 }
1492
1493 /*
1494  * hammer_vop_print { vp }
1495  */
1496 static
1497 int
1498 hammer_vop_print(struct vop_print_args *ap)
1499 {
1500         return EOPNOTSUPP;
1501 }
1502
1503 /*
1504  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1505  */
1506 static
1507 int
1508 hammer_vop_readdir(struct vop_readdir_args *ap)
1509 {
1510         struct hammer_transaction trans;
1511         struct hammer_cursor cursor;
1512         struct hammer_inode *ip;
1513         struct uio *uio;
1514         hammer_base_elm_t base;
1515         int error;
1516         int cookie_index;
1517         int ncookies;
1518         off_t *cookies;
1519         off_t saveoff;
1520         int r;
1521         int dtype;
1522
1523         ++hammer_stats_file_iopsr;
1524         ip = VTOI(ap->a_vp);
1525         uio = ap->a_uio;
1526         saveoff = uio->uio_offset;
1527
1528         if (ap->a_ncookies) {
1529                 ncookies = uio->uio_resid / 16 + 1;
1530                 if (ncookies > 1024)
1531                         ncookies = 1024;
1532                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1533                 cookie_index = 0;
1534         } else {
1535                 ncookies = -1;
1536                 cookies = NULL;
1537                 cookie_index = 0;
1538         }
1539
1540         hammer_simple_transaction(&trans, ip->hmp);
1541
1542         /*
1543          * Handle artificial entries
1544          *
1545          * It should be noted that the minimum value for a directory
1546          * hash key on-media is 0x0000000100000000, so we can use anything
1547          * less then that to represent our 'special' key space.
1548          */
1549         error = 0;
1550         if (saveoff == 0) {
1551                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1552                 if (r)
1553                         goto done;
1554                 if (cookies)
1555                         cookies[cookie_index] = saveoff;
1556                 ++saveoff;
1557                 ++cookie_index;
1558                 if (cookie_index == ncookies)
1559                         goto done;
1560         }
1561         if (saveoff == 1) {
1562                 if (ip->ino_data.parent_obj_id) {
1563                         r = vop_write_dirent(&error, uio,
1564                                              ip->ino_data.parent_obj_id,
1565                                              DT_DIR, 2, "..");
1566                 } else {
1567                         r = vop_write_dirent(&error, uio,
1568                                              ip->obj_id, DT_DIR, 2, "..");
1569                 }
1570                 if (r)
1571                         goto done;
1572                 if (cookies)
1573                         cookies[cookie_index] = saveoff;
1574                 ++saveoff;
1575                 ++cookie_index;
1576                 if (cookie_index == ncookies)
1577                         goto done;
1578         }
1579
1580         /*
1581          * Key range (begin and end inclusive) to scan.  Directory keys
1582          * directly translate to a 64 bit 'seek' position.
1583          */
1584         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1585         cursor.key_beg.localization = ip->obj_localization +
1586                                       hammer_dir_localization(ip);
1587         cursor.key_beg.obj_id = ip->obj_id;
1588         cursor.key_beg.create_tid = 0;
1589         cursor.key_beg.delete_tid = 0;
1590         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1591         cursor.key_beg.obj_type = 0;
1592         cursor.key_beg.key = saveoff;
1593
1594         cursor.key_end = cursor.key_beg;
1595         cursor.key_end.key = HAMMER_MAX_KEY;
1596         cursor.asof = ip->obj_asof;
1597         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1598
1599         error = hammer_ip_first(&cursor);
1600
1601         while (error == 0) {
1602                 error = hammer_ip_resolve_data(&cursor);
1603                 if (error)
1604                         break;
1605                 base = &cursor.leaf->base;
1606                 saveoff = base->key;
1607                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1608
1609                 if (base->obj_id != ip->obj_id)
1610                         panic("readdir: bad record at %p", cursor.node);
1611
1612                 /*
1613                  * Convert pseudo-filesystems into softlinks
1614                  */
1615                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1616                 r = vop_write_dirent(
1617                              &error, uio, cursor.data->entry.obj_id,
1618                              dtype,
1619                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1620                              (void *)cursor.data->entry.name);
1621                 if (r)
1622                         break;
1623                 ++saveoff;
1624                 if (cookies)
1625                         cookies[cookie_index] = base->key;
1626                 ++cookie_index;
1627                 if (cookie_index == ncookies)
1628                         break;
1629                 error = hammer_ip_next(&cursor);
1630         }
1631         hammer_done_cursor(&cursor);
1632
1633 done:
1634         hammer_done_transaction(&trans);
1635
1636         if (ap->a_eofflag)
1637                 *ap->a_eofflag = (error == ENOENT);
1638         uio->uio_offset = saveoff;
1639         if (error && cookie_index == 0) {
1640                 if (error == ENOENT)
1641                         error = 0;
1642                 if (cookies) {
1643                         kfree(cookies, M_TEMP);
1644                         *ap->a_ncookies = 0;
1645                         *ap->a_cookies = NULL;
1646                 }
1647         } else {
1648                 if (error == ENOENT)
1649                         error = 0;
1650                 if (cookies) {
1651                         *ap->a_ncookies = cookie_index;
1652                         *ap->a_cookies = cookies;
1653                 }
1654         }
1655         return(error);
1656 }
1657
1658 /*
1659  * hammer_vop_readlink { vp, uio, cred }
1660  */
1661 static
1662 int
1663 hammer_vop_readlink(struct vop_readlink_args *ap)
1664 {
1665         struct hammer_transaction trans;
1666         struct hammer_cursor cursor;
1667         struct hammer_inode *ip;
1668         char buf[32];
1669         u_int32_t localization;
1670         hammer_pseudofs_inmem_t pfsm;
1671         int error;
1672
1673         ip = VTOI(ap->a_vp);
1674
1675         /*
1676          * Shortcut if the symlink data was stuffed into ino_data.
1677          *
1678          * Also expand special "@@PFS%05d" softlinks (expansion only
1679          * occurs for non-historical (current) accesses made from the
1680          * primary filesystem).
1681          */
1682         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1683                 char *ptr;
1684                 int bytes;
1685
1686                 ptr = ip->ino_data.ext.symlink;
1687                 bytes = (int)ip->ino_data.size;
1688                 if (bytes == 10 &&
1689                     ip->obj_asof == HAMMER_MAX_TID &&
1690                     ip->obj_localization == 0 &&
1691                     strncmp(ptr, "@@PFS", 5) == 0) {
1692                         hammer_simple_transaction(&trans, ip->hmp);
1693                         bcopy(ptr + 5, buf, 5);
1694                         buf[5] = 0;
1695                         localization = strtoul(buf, NULL, 10) << 16;
1696                         pfsm = hammer_load_pseudofs(&trans, localization,
1697                                                     &error);
1698                         if (error == 0) {
1699                                 if (pfsm->pfsd.mirror_flags &
1700                                     HAMMER_PFSD_SLAVE) {
1701                                         /* vap->va_size == 26 */
1702                                         ksnprintf(buf, sizeof(buf),
1703                                                   "@@0x%016llx:%05d",
1704                                                   (long long)pfsm->pfsd.sync_end_tid,
1705                                                   localization >> 16);
1706                                 } else {
1707                                         /* vap->va_size == 10 */
1708                                         ksnprintf(buf, sizeof(buf),
1709                                                   "@@-1:%05d",
1710                                                   localization >> 16);
1711 #if 0
1712                                         ksnprintf(buf, sizeof(buf),
1713                                                   "@@0x%016llx:%05d",
1714                                                   (long long)HAMMER_MAX_TID,
1715                                                   localization >> 16);
1716 #endif
1717                                 }
1718                                 ptr = buf;
1719                                 bytes = strlen(buf);
1720                         }
1721                         if (pfsm)
1722                                 hammer_rel_pseudofs(trans.hmp, pfsm);
1723                         hammer_done_transaction(&trans);
1724                 }
1725                 error = uiomove(ptr, bytes, ap->a_uio);
1726                 return(error);
1727         }
1728
1729         /*
1730          * Long version
1731          */
1732         hammer_simple_transaction(&trans, ip->hmp);
1733         ++hammer_stats_file_iopsr;
1734         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1735
1736         /*
1737          * Key range (begin and end inclusive) to scan.  Directory keys
1738          * directly translate to a 64 bit 'seek' position.
1739          */
1740         cursor.key_beg.localization = ip->obj_localization +
1741                                       HAMMER_LOCALIZE_MISC;
1742         cursor.key_beg.obj_id = ip->obj_id;
1743         cursor.key_beg.create_tid = 0;
1744         cursor.key_beg.delete_tid = 0;
1745         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1746         cursor.key_beg.obj_type = 0;
1747         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1748         cursor.asof = ip->obj_asof;
1749         cursor.flags |= HAMMER_CURSOR_ASOF;
1750
1751         error = hammer_ip_lookup(&cursor);
1752         if (error == 0) {
1753                 error = hammer_ip_resolve_data(&cursor);
1754                 if (error == 0) {
1755                         KKASSERT(cursor.leaf->data_len >=
1756                                  HAMMER_SYMLINK_NAME_OFF);
1757                         error = uiomove(cursor.data->symlink.name,
1758                                         cursor.leaf->data_len -
1759                                                 HAMMER_SYMLINK_NAME_OFF,
1760                                         ap->a_uio);
1761                 }
1762         }
1763         hammer_done_cursor(&cursor);
1764         hammer_done_transaction(&trans);
1765         return(error);
1766 }
1767
1768 /*
1769  * hammer_vop_nremove { nch, dvp, cred }
1770  */
1771 static
1772 int
1773 hammer_vop_nremove(struct vop_nremove_args *ap)
1774 {
1775         struct hammer_transaction trans;
1776         struct hammer_inode *dip;
1777         int error;
1778
1779         dip = VTOI(ap->a_dvp);
1780
1781         if (hammer_nohistory(dip) == 0 &&
1782             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1783                 return (error);
1784         }
1785
1786         hammer_start_transaction(&trans, dip->hmp);
1787         ++hammer_stats_file_iopsw;
1788         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1789         hammer_done_transaction(&trans);
1790         if (error == 0)
1791                 hammer_knote(ap->a_dvp, NOTE_WRITE);
1792         return (error);
1793 }
1794
1795 /*
1796  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1797  */
1798 static
1799 int
1800 hammer_vop_nrename(struct vop_nrename_args *ap)
1801 {
1802         struct hammer_transaction trans;
1803         struct namecache *fncp;
1804         struct namecache *tncp;
1805         struct hammer_inode *fdip;
1806         struct hammer_inode *tdip;
1807         struct hammer_inode *ip;
1808         struct hammer_cursor cursor;
1809         int64_t namekey;
1810         u_int32_t max_iterations;
1811         int nlen, error;
1812
1813         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1814                 return(EXDEV);
1815         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1816                 return(EXDEV);
1817
1818         fdip = VTOI(ap->a_fdvp);
1819         tdip = VTOI(ap->a_tdvp);
1820         fncp = ap->a_fnch->ncp;
1821         tncp = ap->a_tnch->ncp;
1822         ip = VTOI(fncp->nc_vp);
1823         KKASSERT(ip != NULL);
1824
1825         if (fdip->obj_localization != tdip->obj_localization)
1826                 return(EXDEV);
1827         if (fdip->obj_localization != ip->obj_localization)
1828                 return(EXDEV);
1829
1830         if (fdip->flags & HAMMER_INODE_RO)
1831                 return (EROFS);
1832         if (tdip->flags & HAMMER_INODE_RO)
1833                 return (EROFS);
1834         if (ip->flags & HAMMER_INODE_RO)
1835                 return (EROFS);
1836         if ((error = hammer_checkspace(fdip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1837                 return (error);
1838
1839         hammer_start_transaction(&trans, fdip->hmp);
1840         ++hammer_stats_file_iopsw;
1841
1842         /*
1843          * Remove tncp from the target directory and then link ip as
1844          * tncp. XXX pass trans to dounlink
1845          *
1846          * Force the inode sync-time to match the transaction so it is
1847          * in-sync with the creation of the target directory entry.
1848          */
1849         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1850                                 ap->a_cred, 0, -1);
1851         if (error == 0 || error == ENOENT) {
1852                 error = hammer_ip_add_directory(&trans, tdip,
1853                                                 tncp->nc_name, tncp->nc_nlen,
1854                                                 ip);
1855                 if (error == 0) {
1856                         ip->ino_data.parent_obj_id = tdip->obj_id;
1857                         ip->ino_data.ctime = trans.time;
1858                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1859                 }
1860         }
1861         if (error)
1862                 goto failed; /* XXX */
1863
1864         /*
1865          * Locate the record in the originating directory and remove it.
1866          *
1867          * Calculate the namekey and setup the key range for the scan.  This
1868          * works kinda like a chained hash table where the lower 32 bits
1869          * of the namekey synthesize the chain.
1870          *
1871          * The key range is inclusive of both key_beg and key_end.
1872          */
1873         namekey = hammer_directory_namekey(fdip, fncp->nc_name, fncp->nc_nlen,
1874                                            &max_iterations);
1875 retry:
1876         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1877         cursor.key_beg.localization = fdip->obj_localization +
1878                                       hammer_dir_localization(fdip);
1879         cursor.key_beg.obj_id = fdip->obj_id;
1880         cursor.key_beg.key = namekey;
1881         cursor.key_beg.create_tid = 0;
1882         cursor.key_beg.delete_tid = 0;
1883         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1884         cursor.key_beg.obj_type = 0;
1885
1886         cursor.key_end = cursor.key_beg;
1887         cursor.key_end.key += max_iterations;
1888         cursor.asof = fdip->obj_asof;
1889         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1890
1891         /*
1892          * Scan all matching records (the chain), locate the one matching
1893          * the requested path component.
1894          *
1895          * The hammer_ip_*() functions merge in-memory records with on-disk
1896          * records for the purposes of the search.
1897          */
1898         error = hammer_ip_first(&cursor);
1899         while (error == 0) {
1900                 if (hammer_ip_resolve_data(&cursor) != 0)
1901                         break;
1902                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1903                 KKASSERT(nlen > 0);
1904                 if (fncp->nc_nlen == nlen &&
1905                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1906                         break;
1907                 }
1908                 error = hammer_ip_next(&cursor);
1909         }
1910
1911         /*
1912          * If all is ok we have to get the inode so we can adjust nlinks.
1913          *
1914          * WARNING: hammer_ip_del_directory() may have to terminate the
1915          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1916          * twice.
1917          */
1918         if (error == 0)
1919                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1920
1921         /*
1922          * XXX A deadlock here will break rename's atomicy for the purposes
1923          * of crash recovery.
1924          */
1925         if (error == EDEADLK) {
1926                 hammer_done_cursor(&cursor);
1927                 goto retry;
1928         }
1929
1930         /*
1931          * Cleanup and tell the kernel that the rename succeeded.
1932          */
1933         hammer_done_cursor(&cursor);
1934         if (error == 0) {
1935                 cache_rename(ap->a_fnch, ap->a_tnch);
1936                 hammer_knote(ap->a_fdvp, NOTE_WRITE);
1937                 hammer_knote(ap->a_tdvp, NOTE_WRITE);
1938                 if (ip->vp)
1939                         hammer_knote(ip->vp, NOTE_RENAME);
1940         }
1941
1942 failed:
1943         hammer_done_transaction(&trans);
1944         return (error);
1945 }
1946
1947 /*
1948  * hammer_vop_nrmdir { nch, dvp, cred }
1949  */
1950 static
1951 int
1952 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1953 {
1954         struct hammer_transaction trans;
1955         struct hammer_inode *dip;
1956         int error;
1957
1958         dip = VTOI(ap->a_dvp);
1959
1960         if (hammer_nohistory(dip) == 0 &&
1961             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1962                 return (error);
1963         }
1964
1965         hammer_start_transaction(&trans, dip->hmp);
1966         ++hammer_stats_file_iopsw;
1967         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
1968         hammer_done_transaction(&trans);
1969         if (error == 0)
1970                 hammer_knote(ap->a_dvp, NOTE_WRITE | NOTE_LINK);
1971         return (error);
1972 }
1973
1974 /*
1975  * hammer_vop_markatime { vp, cred }
1976  */
1977 static
1978 int
1979 hammer_vop_markatime(struct vop_markatime_args *ap)
1980 {
1981         struct hammer_transaction trans;
1982         struct hammer_inode *ip;
1983
1984         ip = VTOI(ap->a_vp);
1985         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1986                 return (EROFS);
1987         if (ip->flags & HAMMER_INODE_RO)
1988                 return (EROFS);
1989         if (ip->hmp->mp->mnt_flag & MNT_NOATIME)
1990                 return (0);
1991         hammer_start_transaction(&trans, ip->hmp);
1992         ++hammer_stats_file_iopsw;
1993
1994         ip->ino_data.atime = trans.time;
1995         hammer_modify_inode(ip, HAMMER_INODE_ATIME);
1996         hammer_done_transaction(&trans);
1997         hammer_knote(ap->a_vp, NOTE_ATTRIB);
1998         return (0);
1999 }
2000
2001 /*
2002  * hammer_vop_setattr { vp, vap, cred }
2003  */
2004 static
2005 int
2006 hammer_vop_setattr(struct vop_setattr_args *ap)
2007 {
2008         struct hammer_transaction trans;
2009         struct vattr *vap;
2010         struct hammer_inode *ip;
2011         int modflags;
2012         int error;
2013         int truncating;
2014         int blksize;
2015         int kflags;
2016         int64_t aligned_size;
2017         u_int32_t flags;
2018
2019         vap = ap->a_vap;
2020         ip = ap->a_vp->v_data;
2021         modflags = 0;
2022         kflags = 0;
2023
2024         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
2025                 return(EROFS);
2026         if (ip->flags & HAMMER_INODE_RO)
2027                 return (EROFS);
2028         if (hammer_nohistory(ip) == 0 &&
2029             (error = hammer_checkspace(ip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
2030                 return (error);
2031         }
2032
2033         hammer_start_transaction(&trans, ip->hmp);
2034         ++hammer_stats_file_iopsw;
2035         error = 0;
2036
2037         if (vap->va_flags != VNOVAL) {
2038                 flags = ip->ino_data.uflags;
2039                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
2040                                          hammer_to_unix_xid(&ip->ino_data.uid),
2041                                          ap->a_cred);
2042                 if (error == 0) {
2043                         if (ip->ino_data.uflags != flags) {
2044                                 ip->ino_data.uflags = flags;
2045                                 ip->ino_data.ctime = trans.time;
2046                                 modflags |= HAMMER_INODE_DDIRTY;
2047                                 kflags |= NOTE_ATTRIB;
2048                         }
2049                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2050                                 error = 0;
2051                                 goto done;
2052                         }
2053                 }
2054                 goto done;
2055         }
2056         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
2057                 error = EPERM;
2058                 goto done;
2059         }
2060         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
2061                 mode_t cur_mode = ip->ino_data.mode;
2062                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2063                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2064                 uuid_t uuid_uid;
2065                 uuid_t uuid_gid;
2066
2067                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
2068                                          ap->a_cred,
2069                                          &cur_uid, &cur_gid, &cur_mode);
2070                 if (error == 0) {
2071                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
2072                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
2073                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
2074                                  sizeof(uuid_uid)) ||
2075                             bcmp(&uuid_gid, &ip->ino_data.gid,
2076                                  sizeof(uuid_gid)) ||
2077                             ip->ino_data.mode != cur_mode
2078                         ) {
2079                                 ip->ino_data.uid = uuid_uid;
2080                                 ip->ino_data.gid = uuid_gid;
2081                                 ip->ino_data.mode = cur_mode;
2082                                 ip->ino_data.ctime = trans.time;
2083                                 modflags |= HAMMER_INODE_DDIRTY;
2084                         }
2085                         kflags |= NOTE_ATTRIB;
2086                 }
2087         }
2088         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
2089                 switch(ap->a_vp->v_type) {
2090                 case VREG:
2091                         if (vap->va_size == ip->ino_data.size)
2092                                 break;
2093                         /*
2094                          * XXX break atomicy, we can deadlock the backend
2095                          * if we do not release the lock.  Probably not a
2096                          * big deal here.
2097                          */
2098                         blksize = hammer_blocksize(vap->va_size);
2099                         if (vap->va_size < ip->ino_data.size) {
2100                                 vtruncbuf(ap->a_vp, vap->va_size, blksize);
2101                                 truncating = 1;
2102                                 kflags |= NOTE_WRITE;
2103                         } else {
2104                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
2105                                 truncating = 0;
2106                                 kflags |= NOTE_WRITE | NOTE_EXTEND;
2107                         }
2108                         ip->ino_data.size = vap->va_size;
2109                         ip->ino_data.mtime = trans.time;
2110                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2111
2112                         /*
2113                          * on-media truncation is cached in the inode until
2114                          * the inode is synchronized.
2115                          */
2116                         if (truncating) {
2117                                 hammer_ip_frontend_trunc(ip, vap->va_size);
2118 #ifdef DEBUG_TRUNCATE
2119                                 if (HammerTruncIp == NULL)
2120                                         HammerTruncIp = ip;
2121 #endif
2122                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2123                                         ip->flags |= HAMMER_INODE_TRUNCATED;
2124                                         ip->trunc_off = vap->va_size;
2125 #ifdef DEBUG_TRUNCATE
2126                                         if (ip == HammerTruncIp)
2127                                         kprintf("truncate1 %016llx\n",
2128                                                 (long long)ip->trunc_off);
2129 #endif
2130                                 } else if (ip->trunc_off > vap->va_size) {
2131                                         ip->trunc_off = vap->va_size;
2132 #ifdef DEBUG_TRUNCATE
2133                                         if (ip == HammerTruncIp)
2134                                         kprintf("truncate2 %016llx\n",
2135                                                 (long long)ip->trunc_off);
2136 #endif
2137                                 } else {
2138 #ifdef DEBUG_TRUNCATE
2139                                         if (ip == HammerTruncIp)
2140                                         kprintf("truncate3 %016llx (ignored)\n",
2141                                                 (long long)vap->va_size);
2142 #endif
2143                                 }
2144                         }
2145
2146                         /*
2147                          * If truncating we have to clean out a portion of
2148                          * the last block on-disk.  We do this in the
2149                          * front-end buffer cache.
2150                          */
2151                         aligned_size = (vap->va_size + (blksize - 1)) &
2152                                        ~(int64_t)(blksize - 1);
2153                         if (truncating && vap->va_size < aligned_size) {
2154                                 struct buf *bp;
2155                                 int offset;
2156
2157                                 aligned_size -= blksize;
2158
2159                                 offset = (int)vap->va_size & (blksize - 1);
2160                                 error = bread(ap->a_vp, aligned_size,
2161                                               blksize, &bp);
2162                                 hammer_ip_frontend_trunc(ip, aligned_size);
2163                                 if (error == 0) {
2164                                         bzero(bp->b_data + offset,
2165                                               blksize - offset);
2166                                         /* must de-cache direct-io offset */
2167                                         bp->b_bio2.bio_offset = NOOFFSET;
2168                                         bdwrite(bp);
2169                                 } else {
2170                                         kprintf("ERROR %d\n", error);
2171                                         brelse(bp);
2172                                 }
2173                         }
2174                         break;
2175                 case VDATABASE:
2176                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
2177                                 ip->flags |= HAMMER_INODE_TRUNCATED;
2178                                 ip->trunc_off = vap->va_size;
2179                         } else if (ip->trunc_off > vap->va_size) {
2180                                 ip->trunc_off = vap->va_size;
2181                         }
2182                         hammer_ip_frontend_trunc(ip, vap->va_size);
2183                         ip->ino_data.size = vap->va_size;
2184                         ip->ino_data.mtime = trans.time;
2185                         modflags |= HAMMER_INODE_MTIME | HAMMER_INODE_DDIRTY;
2186                         kflags |= NOTE_ATTRIB;
2187                         break;
2188                 default:
2189                         error = EINVAL;
2190                         goto done;
2191                 }
2192                 break;
2193         }
2194         if (vap->va_atime.tv_sec != VNOVAL) {
2195                 ip->ino_data.atime = hammer_timespec_to_time(&vap->va_atime);
2196                 modflags |= HAMMER_INODE_ATIME;
2197                 kflags |= NOTE_ATTRIB;
2198         }
2199         if (vap->va_mtime.tv_sec != VNOVAL) {
2200                 ip->ino_data.mtime = hammer_timespec_to_time(&vap->va_mtime);
2201                 modflags |= HAMMER_INODE_MTIME;
2202                 kflags |= NOTE_ATTRIB;
2203         }
2204         if (vap->va_mode != (mode_t)VNOVAL) {
2205                 mode_t   cur_mode = ip->ino_data.mode;
2206                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
2207                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
2208
2209                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
2210                                          cur_uid, cur_gid, &cur_mode);
2211                 if (error == 0 && ip->ino_data.mode != cur_mode) {
2212                         ip->ino_data.mode = cur_mode;
2213                         ip->ino_data.ctime = trans.time;
2214                         modflags |= HAMMER_INODE_DDIRTY;
2215                         kflags |= NOTE_ATTRIB;
2216                 }
2217         }
2218 done:
2219         if (error == 0)
2220                 hammer_modify_inode(ip, modflags);
2221         hammer_done_transaction(&trans);
2222         hammer_knote(ap->a_vp, kflags);
2223         return (error);
2224 }
2225
2226 /*
2227  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2228  */
2229 static
2230 int
2231 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
2232 {
2233         struct hammer_transaction trans;
2234         struct hammer_inode *dip;
2235         struct hammer_inode *nip;
2236         struct nchandle *nch;
2237         hammer_record_t record;
2238         int error;
2239         int bytes;
2240
2241         ap->a_vap->va_type = VLNK;
2242
2243         nch = ap->a_nch;
2244         dip = VTOI(ap->a_dvp);
2245
2246         if (dip->flags & HAMMER_INODE_RO)
2247                 return (EROFS);
2248         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
2249                 return (error);
2250
2251         /*
2252          * Create a transaction to cover the operations we perform.
2253          */
2254         hammer_start_transaction(&trans, dip->hmp);
2255         ++hammer_stats_file_iopsw;
2256
2257         /*
2258          * Create a new filesystem object of the requested type.  The
2259          * returned inode will be referenced but not locked.
2260          */
2261
2262         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
2263                                     dip, nch->ncp->nc_name, nch->ncp->nc_nlen,
2264                                     NULL, &nip);
2265         if (error) {
2266                 hammer_done_transaction(&trans);
2267                 *ap->a_vpp = NULL;
2268                 return (error);
2269         }
2270
2271         /*
2272          * Add a record representing the symlink.  symlink stores the link
2273          * as pure data, not a string, and is no \0 terminated.
2274          */
2275         if (error == 0) {
2276                 bytes = strlen(ap->a_target);
2277
2278                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
2279                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
2280                 } else {
2281                         record = hammer_alloc_mem_record(nip, bytes);
2282                         record->type = HAMMER_MEM_RECORD_GENERAL;
2283
2284                         record->leaf.base.localization = nip->obj_localization +
2285                                                          HAMMER_LOCALIZE_MISC;
2286                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
2287                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
2288                         record->leaf.data_len = bytes;
2289                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
2290                         bcopy(ap->a_target, record->data->symlink.name, bytes);
2291                         error = hammer_ip_add_record(&trans, record);
2292                 }
2293
2294                 /*
2295                  * Set the file size to the length of the link.
2296                  */
2297                 if (error == 0) {
2298                         nip->ino_data.size = bytes;
2299                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
2300                 }
2301         }
2302         if (error == 0)
2303                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
2304                                                 nch->ncp->nc_nlen, nip);
2305
2306         /*
2307          * Finish up.
2308          */
2309         if (error) {
2310                 hammer_rel_inode(nip, 0);
2311                 *ap->a_vpp = NULL;
2312         } else {
2313                 error = hammer_get_vnode(nip, ap->a_vpp);
2314                 hammer_rel_inode(nip, 0);
2315                 if (error == 0) {
2316                         cache_setunresolved(ap->a_nch);
2317                         cache_setvp(ap->a_nch, *ap->a_vpp);
2318                         hammer_knote(ap->a_dvp, NOTE_WRITE);
2319                 }
2320         }
2321         hammer_done_transaction(&trans);
2322         return (error);
2323 }
2324
2325 /*
2326  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2327  */
2328 static
2329 int
2330 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
2331 {
2332         struct hammer_transaction trans;
2333         struct hammer_inode *dip;
2334         int error;
2335
2336         dip = VTOI(ap->a_dvp);
2337
2338         if (hammer_nohistory(dip) == 0 &&
2339             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0) {
2340                 return (error);
2341         }
2342
2343         hammer_start_transaction(&trans, dip->hmp);
2344         ++hammer_stats_file_iopsw;
2345         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
2346                                 ap->a_cred, ap->a_flags, -1);
2347         hammer_done_transaction(&trans);
2348
2349         return (error);
2350 }
2351
2352 /*
2353  * hammer_vop_ioctl { vp, command, data, fflag, cred }
2354  */
2355 static
2356 int
2357 hammer_vop_ioctl(struct vop_ioctl_args *ap)
2358 {
2359         struct hammer_inode *ip = ap->a_vp->v_data;
2360
2361         ++hammer_stats_file_iopsr;
2362         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
2363                             ap->a_fflag, ap->a_cred));
2364 }
2365
2366 static
2367 int
2368 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2369 {
2370         static const struct mountctl_opt extraopt[] = {
2371                 { HMNT_NOHISTORY,       "nohistory" },
2372                 { HMNT_MASTERID,        "master" },
2373                 { 0, NULL}
2374
2375         };
2376         struct hammer_mount *hmp;
2377         struct mount *mp;
2378         int usedbytes;
2379         int error;
2380
2381         error = 0;
2382         usedbytes = 0;
2383         mp = ap->a_head.a_ops->head.vv_mount;
2384         KKASSERT(mp->mnt_data != NULL);
2385         hmp = (struct hammer_mount *)mp->mnt_data;
2386
2387         switch(ap->a_op) {
2388
2389         case MOUNTCTL_SET_EXPORT:
2390                 if (ap->a_ctllen != sizeof(struct export_args))
2391                         error = EINVAL;
2392                 else
2393                         error = hammer_vfs_export(mp, ap->a_op,
2394                                       (const struct export_args *)ap->a_ctl);
2395                 break;
2396         case MOUNTCTL_MOUNTFLAGS:
2397         {
2398                 /*
2399                  * Call standard mountctl VOP function
2400                  * so we get user mount flags.
2401                  */
2402                 error = vop_stdmountctl(ap);
2403                 if (error)
2404                         break;
2405
2406                 usedbytes = *ap->a_res;
2407
2408                 if (usedbytes > 0 && usedbytes < ap->a_buflen) {
2409                         usedbytes += vfs_flagstostr(hmp->hflags, extraopt, ap->a_buf,
2410                                                     ap->a_buflen - usedbytes,
2411                                                     &error);
2412                 }
2413
2414                 *ap->a_res += usedbytes;
2415                 break;
2416         }
2417         default:
2418                 error = vop_stdmountctl(ap);
2419                 break;
2420         }
2421         return(error);
2422 }
2423
2424 /*
2425  * hammer_vop_strategy { vp, bio }
2426  *
2427  * Strategy call, used for regular file read & write only.  Note that the
2428  * bp may represent a cluster.
2429  *
2430  * To simplify operation and allow better optimizations in the future,
2431  * this code does not make any assumptions with regards to buffer alignment
2432  * or size.
2433  */
2434 static
2435 int
2436 hammer_vop_strategy(struct vop_strategy_args *ap)
2437 {
2438         struct buf *bp;
2439         int error;
2440
2441         bp = ap->a_bio->bio_buf;
2442
2443         switch(bp->b_cmd) {
2444         case BUF_CMD_READ:
2445                 error = hammer_vop_strategy_read(ap);
2446                 break;
2447         case BUF_CMD_WRITE:
2448                 error = hammer_vop_strategy_write(ap);
2449                 break;
2450         default:
2451                 bp->b_error = error = EINVAL;
2452                 bp->b_flags |= B_ERROR;
2453                 biodone(ap->a_bio);
2454                 break;
2455         }
2456         return (error);
2457 }
2458
2459 /*
2460  * Read from a regular file.  Iterate the related records and fill in the
2461  * BIO/BUF.  Gaps are zero-filled.
2462  *
2463  * The support code in hammer_object.c should be used to deal with mixed
2464  * in-memory and on-disk records.
2465  *
2466  * NOTE: Can be called from the cluster code with an oversized buf.
2467  *
2468  * XXX atime update
2469  */
2470 static
2471 int
2472 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2473 {
2474         struct hammer_transaction trans;
2475         struct hammer_inode *ip;
2476         struct hammer_inode *dip;
2477         struct hammer_cursor cursor;
2478         hammer_base_elm_t base;
2479         hammer_off_t disk_offset;
2480         struct bio *bio;
2481         struct bio *nbio;
2482         struct buf *bp;
2483         int64_t rec_offset;
2484         int64_t ran_end;
2485         int64_t tmp64;
2486         int error;
2487         int boff;
2488         int roff;
2489         int n;
2490
2491         bio = ap->a_bio;
2492         bp = bio->bio_buf;
2493         ip = ap->a_vp->v_data;
2494
2495         /*
2496          * The zone-2 disk offset may have been set by the cluster code via
2497          * a BMAP operation, or else should be NOOFFSET.
2498          *
2499          * Checking the high bits for a match against zone-2 should suffice.
2500          */
2501         nbio = push_bio(bio);
2502         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2503             HAMMER_ZONE_LARGE_DATA) {
2504                 error = hammer_io_direct_read(ip->hmp, nbio, NULL);
2505                 return (error);
2506         }
2507
2508         /*
2509          * Well, that sucked.  Do it the hard way.  If all the stars are
2510          * aligned we may still be able to issue a direct-read.
2511          */
2512         hammer_simple_transaction(&trans, ip->hmp);
2513         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2514
2515         /*
2516          * Key range (begin and end inclusive) to scan.  Note that the key's
2517          * stored in the actual records represent BASE+LEN, not BASE.  The
2518          * first record containing bio_offset will have a key > bio_offset.
2519          */
2520         cursor.key_beg.localization = ip->obj_localization +
2521                                       HAMMER_LOCALIZE_MISC;
2522         cursor.key_beg.obj_id = ip->obj_id;
2523         cursor.key_beg.create_tid = 0;
2524         cursor.key_beg.delete_tid = 0;
2525         cursor.key_beg.obj_type = 0;
2526         cursor.key_beg.key = bio->bio_offset + 1;
2527         cursor.asof = ip->obj_asof;
2528         cursor.flags |= HAMMER_CURSOR_ASOF;
2529
2530         cursor.key_end = cursor.key_beg;
2531         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2532 #if 0
2533         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2534                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2535                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2536                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2537         } else
2538 #endif
2539         {
2540                 ran_end = bio->bio_offset + bp->b_bufsize;
2541                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2542                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2543                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2544                 if (tmp64 < ran_end)
2545                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2546                 else
2547                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2548         }
2549         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2550
2551         error = hammer_ip_first(&cursor);
2552         boff = 0;
2553
2554         while (error == 0) {
2555                 /*
2556                  * Get the base file offset of the record.  The key for
2557                  * data records is (base + bytes) rather then (base).
2558                  */
2559                 base = &cursor.leaf->base;
2560                 rec_offset = base->key - cursor.leaf->data_len;
2561
2562                 /*
2563                  * Calculate the gap, if any, and zero-fill it.
2564                  *
2565                  * n is the offset of the start of the record verses our
2566                  * current seek offset in the bio.
2567                  */
2568                 n = (int)(rec_offset - (bio->bio_offset + boff));
2569                 if (n > 0) {
2570                         if (n > bp->b_bufsize - boff)
2571                                 n = bp->b_bufsize - boff;
2572                         bzero((char *)bp->b_data + boff, n);
2573                         boff += n;
2574                         n = 0;
2575                 }
2576
2577                 /*
2578                  * Calculate the data offset in the record and the number
2579                  * of bytes we can copy.
2580                  *
2581                  * There are two degenerate cases.  First, boff may already
2582                  * be at bp->b_bufsize.  Secondly, the data offset within
2583                  * the record may exceed the record's size.
2584                  */
2585                 roff = -n;
2586                 rec_offset += roff;
2587                 n = cursor.leaf->data_len - roff;
2588                 if (n <= 0) {
2589                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2590                         n = 0;
2591                 } else if (n > bp->b_bufsize - boff) {
2592                         n = bp->b_bufsize - boff;
2593                 }
2594
2595                 /*
2596                  * Deal with cached truncations.  This cool bit of code
2597                  * allows truncate()/ftruncate() to avoid having to sync
2598                  * the file.
2599                  *
2600                  * If the frontend is truncated then all backend records are
2601                  * subject to the frontend's truncation.
2602                  *
2603                  * If the backend is truncated then backend records on-disk
2604                  * (but not in-memory) are subject to the backend's
2605                  * truncation.  In-memory records owned by the backend
2606                  * represent data written after the truncation point on the
2607                  * backend and must not be truncated.
2608                  *
2609                  * Truncate operations deal with frontend buffer cache
2610                  * buffers and frontend-owned in-memory records synchronously.
2611                  */
2612                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2613                         if (hammer_cursor_ondisk(&cursor) ||
2614                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2615                                 if (ip->trunc_off <= rec_offset)
2616                                         n = 0;
2617                                 else if (ip->trunc_off < rec_offset + n)
2618                                         n = (int)(ip->trunc_off - rec_offset);
2619                         }
2620                 }
2621                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2622                         if (hammer_cursor_ondisk(&cursor)) {
2623                                 if (ip->sync_trunc_off <= rec_offset)
2624                                         n = 0;
2625                                 else if (ip->sync_trunc_off < rec_offset + n)
2626                                         n = (int)(ip->sync_trunc_off - rec_offset);
2627                         }
2628                 }
2629
2630                 /*
2631                  * Try to issue a direct read into our bio if possible,
2632                  * otherwise resolve the element data into a hammer_buffer
2633                  * and copy.
2634                  *
2635                  * The buffer on-disk should be zerod past any real
2636                  * truncation point, but may not be for any synthesized
2637                  * truncation point from above.
2638                  */
2639                 disk_offset = cursor.leaf->data_offset + roff;
2640                 if (boff == 0 && n == bp->b_bufsize &&
2641                     hammer_cursor_ondisk(&cursor) &&
2642                     (disk_offset & HAMMER_BUFMASK) == 0) {
2643                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2644                                  HAMMER_ZONE_LARGE_DATA);
2645                         nbio->bio_offset = disk_offset;
2646                         error = hammer_io_direct_read(trans.hmp, nbio,
2647                                                       cursor.leaf);
2648                         goto done;
2649                 } else if (n) {
2650                         error = hammer_ip_resolve_data(&cursor);
2651                         if (error == 0) {
2652                                 bcopy((char *)cursor.data + roff,
2653                                       (char *)bp->b_data + boff, n);
2654                         }
2655                 }
2656                 if (error)
2657                         break;
2658
2659                 /*
2660                  * Iterate until we have filled the request.
2661                  */
2662                 boff += n;
2663                 if (boff == bp->b_bufsize)
2664                         break;
2665                 error = hammer_ip_next(&cursor);
2666         }
2667
2668         /*
2669          * There may have been a gap after the last record
2670          */
2671         if (error == ENOENT)
2672                 error = 0;
2673         if (error == 0 && boff != bp->b_bufsize) {
2674                 KKASSERT(boff < bp->b_bufsize);
2675                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2676                 /* boff = bp->b_bufsize; */
2677         }
2678         bp->b_resid = 0;
2679         bp->b_error = error;
2680         if (error)
2681                 bp->b_flags |= B_ERROR;
2682         biodone(ap->a_bio);
2683
2684 done:
2685         /*
2686          * Cache the b-tree node for the last data read in cache[1].
2687          *
2688          * If we hit the file EOF then also cache the node in the
2689          * governing director's cache[3], it will be used to initialize
2690          * the inode's cache[1] for any inodes looked up via the directory.
2691          *
2692          * This doesn't reduce disk accesses since the B-Tree chain is
2693          * likely cached, but it does reduce cpu overhead when looking
2694          * up file offsets for cpdup/tar/cpio style iterations.
2695          */
2696         if (cursor.node)
2697                 hammer_cache_node(&ip->cache[1], cursor.node);
2698         if (ran_end >= ip->ino_data.size) {
2699                 dip = hammer_find_inode(&trans, ip->ino_data.parent_obj_id,
2700                                         ip->obj_asof, ip->obj_localization);
2701                 if (dip) {
2702                         hammer_cache_node(&dip->cache[3], cursor.node);
2703                         hammer_rel_inode(dip, 0);
2704                 }
2705         }
2706         hammer_done_cursor(&cursor);
2707         hammer_done_transaction(&trans);
2708         return(error);
2709 }
2710
2711 /*
2712  * BMAP operation - used to support cluster_read() only.
2713  *
2714  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2715  *
2716  * This routine may return EOPNOTSUPP if the opration is not supported for
2717  * the specified offset.  The contents of the pointer arguments do not
2718  * need to be initialized in that case.
2719  *
2720  * If a disk address is available and properly aligned return 0 with
2721  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2722  * to the run-length relative to that offset.  Callers may assume that
2723  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2724  * large, so return EOPNOTSUPP if it is not sufficiently large.
2725  */
2726 static
2727 int
2728 hammer_vop_bmap(struct vop_bmap_args *ap)
2729 {
2730         struct hammer_transaction trans;
2731         struct hammer_inode *ip;
2732         struct hammer_cursor cursor;
2733         hammer_base_elm_t base;
2734         int64_t rec_offset;
2735         int64_t ran_end;
2736         int64_t tmp64;
2737         int64_t base_offset;
2738         int64_t base_disk_offset;
2739         int64_t last_offset;
2740         hammer_off_t last_disk_offset;
2741         hammer_off_t disk_offset;
2742         int     rec_len;
2743         int     error;
2744         int     blksize;
2745
2746         ++hammer_stats_file_iopsr;
2747         ip = ap->a_vp->v_data;
2748
2749         /*
2750          * We can only BMAP regular files.  We can't BMAP database files,
2751          * directories, etc.
2752          */
2753         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2754                 return(EOPNOTSUPP);
2755
2756         /*
2757          * bmap is typically called with runp/runb both NULL when used
2758          * for writing.  We do not support BMAP for writing atm.
2759          */
2760         if (ap->a_cmd != BUF_CMD_READ)
2761                 return(EOPNOTSUPP);
2762
2763         /*
2764          * Scan the B-Tree to acquire blockmap addresses, then translate
2765          * to raw addresses.
2766          */
2767         hammer_simple_transaction(&trans, ip->hmp);
2768 #if 0
2769         kprintf("bmap_beg %016llx ip->cache %p\n",
2770                 (long long)ap->a_loffset, ip->cache[1]);
2771 #endif
2772         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2773
2774         /*
2775          * Key range (begin and end inclusive) to scan.  Note that the key's
2776          * stored in the actual records represent BASE+LEN, not BASE.  The
2777          * first record containing bio_offset will have a key > bio_offset.
2778          */
2779         cursor.key_beg.localization = ip->obj_localization +
2780                                       HAMMER_LOCALIZE_MISC;
2781         cursor.key_beg.obj_id = ip->obj_id;
2782         cursor.key_beg.create_tid = 0;
2783         cursor.key_beg.delete_tid = 0;
2784         cursor.key_beg.obj_type = 0;
2785         if (ap->a_runb)
2786                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2787         else
2788                 cursor.key_beg.key = ap->a_loffset + 1;
2789         if (cursor.key_beg.key < 0)
2790                 cursor.key_beg.key = 0;
2791         cursor.asof = ip->obj_asof;
2792         cursor.flags |= HAMMER_CURSOR_ASOF;
2793
2794         cursor.key_end = cursor.key_beg;
2795         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2796
2797         ran_end = ap->a_loffset + MAXPHYS;
2798         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2799         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2800         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2801         if (tmp64 < ran_end)
2802                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2803         else
2804                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2805
2806         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2807
2808         error = hammer_ip_first(&cursor);
2809         base_offset = last_offset = 0;
2810         base_disk_offset = last_disk_offset = 0;
2811
2812         while (error == 0) {
2813                 /*
2814                  * Get the base file offset of the record.  The key for
2815                  * data records is (base + bytes) rather then (base).
2816                  *
2817                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2818                  * The extra bytes should be zero on-disk and the BMAP op
2819                  * should still be ok.
2820                  */
2821                 base = &cursor.leaf->base;
2822                 rec_offset = base->key - cursor.leaf->data_len;
2823                 rec_len    = cursor.leaf->data_len;
2824
2825                 /*
2826                  * Incorporate any cached truncation.
2827                  *
2828                  * NOTE: Modifications to rec_len based on synthesized
2829                  * truncation points remove the guarantee that any extended
2830                  * data on disk is zero (since the truncations may not have
2831                  * taken place on-media yet).
2832                  */
2833                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2834                         if (hammer_cursor_ondisk(&cursor) ||
2835                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2836                                 if (ip->trunc_off <= rec_offset)
2837                                         rec_len = 0;
2838                                 else if (ip->trunc_off < rec_offset + rec_len)
2839                                         rec_len = (int)(ip->trunc_off - rec_offset);
2840                         }
2841                 }
2842                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2843                         if (hammer_cursor_ondisk(&cursor)) {
2844                                 if (ip->sync_trunc_off <= rec_offset)
2845                                         rec_len = 0;
2846                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2847                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2848                         }
2849                 }
2850
2851                 /*
2852                  * Accumulate information.  If we have hit a discontiguous
2853                  * block reset base_offset unless we are already beyond the
2854                  * requested offset.  If we are, that's it, we stop.
2855                  */
2856                 if (error)
2857                         break;
2858                 if (hammer_cursor_ondisk(&cursor)) {
2859                         disk_offset = cursor.leaf->data_offset;
2860                         if (rec_offset != last_offset ||
2861                             disk_offset != last_disk_offset) {
2862                                 if (rec_offset > ap->a_loffset)
2863                                         break;
2864                                 base_offset = rec_offset;
2865                                 base_disk_offset = disk_offset;
2866                         }
2867                         last_offset = rec_offset + rec_len;
2868                         last_disk_offset = disk_offset + rec_len;
2869                 }
2870                 error = hammer_ip_next(&cursor);
2871         }
2872
2873 #if 0
2874         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2875                 (long long)ap->a_loffset,
2876                 (long long)base_offset,
2877                 (long long)last_offset);
2878         kprintf("BMAP %16s:  %016llx - %016llx\n", "",
2879                 (long long)base_disk_offset,
2880                 (long long)last_disk_offset);
2881 #endif
2882
2883         if (cursor.node) {
2884                 hammer_cache_node(&ip->cache[1], cursor.node);
2885 #if 0
2886                 kprintf("bmap_end2 %016llx ip->cache %p\n",
2887                         (long long)ap->a_loffset, ip->cache[1]);
2888 #endif
2889         }
2890         hammer_done_cursor(&cursor);
2891         hammer_done_transaction(&trans);
2892
2893         /*
2894          * If we couldn't find any records or the records we did find were
2895          * all behind the requested offset, return failure.  A forward
2896          * truncation can leave a hole w/ no on-disk records.
2897          */
2898         if (last_offset == 0 || last_offset < ap->a_loffset)
2899                 return (EOPNOTSUPP);
2900
2901         /*
2902          * Figure out the block size at the requested offset and adjust
2903          * our limits so the cluster_read() does not create inappropriately
2904          * sized buffer cache buffers.
2905          */
2906         blksize = hammer_blocksize(ap->a_loffset);
2907         if (hammer_blocksize(base_offset) != blksize) {
2908                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2909         }
2910         if (last_offset != ap->a_loffset &&
2911             hammer_blocksize(last_offset - 1) != blksize) {
2912                 last_offset = hammer_blockdemarc(ap->a_loffset,
2913                                                  last_offset - 1);
2914         }
2915
2916         /*
2917          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2918          * from occuring.
2919          */
2920         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2921
2922         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
2923                 /*
2924                  * Only large-data zones can be direct-IOd
2925                  */
2926                 error = EOPNOTSUPP;
2927         } else if ((disk_offset & HAMMER_BUFMASK) ||
2928                    (last_offset - ap->a_loffset) < blksize) {
2929                 /*
2930                  * doffsetp is not aligned or the forward run size does
2931                  * not cover a whole buffer, disallow the direct I/O.
2932                  */
2933                 error = EOPNOTSUPP;
2934         } else {
2935                 /*
2936                  * We're good.
2937                  */
2938                 *ap->a_doffsetp = disk_offset;
2939                 if (ap->a_runb) {
2940                         *ap->a_runb = ap->a_loffset - base_offset;
2941                         KKASSERT(*ap->a_runb >= 0);
2942                 }
2943                 if (ap->a_runp) {
2944                         *ap->a_runp = last_offset - ap->a_loffset;
2945                         KKASSERT(*ap->a_runp >= 0);
2946                 }
2947                 error = 0;
2948         }
2949         return(error);
2950 }
2951
2952 /*
2953  * Write to a regular file.   Because this is a strategy call the OS is
2954  * trying to actually get data onto the media.
2955  */
2956 static
2957 int
2958 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2959 {
2960         hammer_record_t record;
2961         hammer_mount_t hmp;
2962         hammer_inode_t ip;
2963         struct bio *bio;
2964         struct buf *bp;
2965         int blksize;
2966         int bytes;
2967         int error;
2968
2969         bio = ap->a_bio;
2970         bp = bio->bio_buf;
2971         ip = ap->a_vp->v_data;
2972         hmp = ip->hmp;
2973
2974         blksize = hammer_blocksize(bio->bio_offset);
2975         KKASSERT(bp->b_bufsize == blksize);
2976
2977         if (ip->flags & HAMMER_INODE_RO) {
2978                 bp->b_error = EROFS;
2979                 bp->b_flags |= B_ERROR;
2980                 biodone(ap->a_bio);
2981                 return(EROFS);
2982         }
2983
2984         /*
2985          * Interlock with inode destruction (no in-kernel or directory
2986          * topology visibility).  If we queue new IO while trying to
2987          * destroy the inode we can deadlock the vtrunc call in
2988          * hammer_inode_unloadable_check().
2989          *
2990          * Besides, there's no point flushing a bp associated with an
2991          * inode that is being destroyed on-media and has no kernel
2992          * references.
2993          */
2994         if ((ip->flags | ip->sync_flags) &
2995             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2996                 bp->b_resid = 0;
2997                 biodone(ap->a_bio);
2998                 return(0);
2999         }
3000
3001         /*
3002          * Reserve space and issue a direct-write from the front-end.
3003          * NOTE: The direct_io code will hammer_bread/bcopy smaller
3004          * allocations.
3005          *
3006          * An in-memory record will be installed to reference the storage
3007          * until the flusher can get to it.
3008          *
3009          * Since we own the high level bio the front-end will not try to
3010          * do a direct-read until the write completes.
3011          *
3012          * NOTE: The only time we do not reserve a full-sized buffers
3013          * worth of data is if the file is small.  We do not try to
3014          * allocate a fragment (from the small-data zone) at the end of
3015          * an otherwise large file as this can lead to wildly separated
3016          * data.
3017          */
3018         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
3019         KKASSERT(bio->bio_offset < ip->ino_data.size);
3020         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
3021                 bytes = bp->b_bufsize;
3022         else
3023                 bytes = ((int)ip->ino_data.size + 15) & ~15;
3024
3025         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
3026                                     bytes, &error);
3027         if (record) {
3028                 hammer_io_direct_write(hmp, record, bio);
3029                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
3030                         hammer_flush_inode(ip, 0);
3031         } else {
3032                 bp->b_bio2.bio_offset = NOOFFSET;
3033                 bp->b_error = error;
3034                 bp->b_flags |= B_ERROR;
3035                 biodone(ap->a_bio);
3036         }
3037         return(error);
3038 }
3039
3040 /*
3041  * dounlink - disconnect a directory entry
3042  *
3043  * XXX whiteout support not really in yet
3044  */
3045 static int
3046 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
3047                 struct vnode *dvp, struct ucred *cred,
3048                 int flags, int isdir)
3049 {
3050         struct namecache *ncp;
3051         hammer_inode_t dip;
3052         hammer_inode_t ip;
3053         struct hammer_cursor cursor;
3054         int64_t namekey;
3055         u_int32_t max_iterations;
3056         int nlen, error;
3057
3058         /*
3059          * Calculate the namekey and setup the key range for the scan.  This
3060          * works kinda like a chained hash table where the lower 32 bits
3061          * of the namekey synthesize the chain.
3062          *
3063          * The key range is inclusive of both key_beg and key_end.
3064          */
3065         dip = VTOI(dvp);
3066         ncp = nch->ncp;
3067
3068         if (dip->flags & HAMMER_INODE_RO)
3069                 return (EROFS);
3070
3071         namekey = hammer_directory_namekey(dip, ncp->nc_name, ncp->nc_nlen,
3072                                            &max_iterations);
3073 retry:
3074         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
3075         cursor.key_beg.localization = dip->obj_localization +
3076                                       hammer_dir_localization(dip);
3077         cursor.key_beg.obj_id = dip->obj_id;
3078         cursor.key_beg.key = namekey;
3079         cursor.key_beg.create_tid = 0;
3080         cursor.key_beg.delete_tid = 0;
3081         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
3082         cursor.key_beg.obj_type = 0;
3083
3084         cursor.key_end = cursor.key_beg;
3085         cursor.key_end.key += max_iterations;
3086         cursor.asof = dip->obj_asof;
3087         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
3088
3089         /*
3090          * Scan all matching records (the chain), locate the one matching
3091          * the requested path component.  info->last_error contains the
3092          * error code on search termination and could be 0, ENOENT, or
3093          * something else.
3094          *
3095          * The hammer_ip_*() functions merge in-memory records with on-disk
3096          * records for the purposes of the search.
3097          */
3098         error = hammer_ip_first(&cursor);
3099
3100         while (error == 0) {
3101                 error = hammer_ip_resolve_data(&cursor);
3102                 if (error)
3103                         break;
3104                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
3105                 KKASSERT(nlen > 0);
3106                 if (ncp->nc_nlen == nlen &&
3107                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
3108                         break;
3109                 }
3110                 error = hammer_ip_next(&cursor);
3111         }
3112
3113         /*
3114          * If all is ok we have to get the inode so we can adjust nlinks.
3115          * To avoid a deadlock with the flusher we must release the inode
3116          * lock on the directory when acquiring the inode for the entry.
3117          *
3118          * If the target is a directory, it must be empty.
3119          */
3120         if (error == 0) {
3121                 hammer_unlock(&cursor.ip->lock);
3122                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
3123                                       dip->hmp->asof,
3124                                       cursor.data->entry.localization,
3125                                       0, &error);
3126                 hammer_lock_sh(&cursor.ip->lock);
3127                 if (error == ENOENT) {
3128                         kprintf("HAMMER: WARNING: Removing "
3129                                 "dirent w/missing inode \"%s\"\n"
3130                                 "\tobj_id = %016llx\n",
3131                                 ncp->nc_name,
3132                                 (long long)cursor.data->entry.obj_id);
3133                         error = 0;
3134                 }
3135
3136                 /*
3137                  * If isdir >= 0 we validate that the entry is or is not a
3138                  * directory.  If isdir < 0 we don't care.
3139                  */
3140                 if (error == 0 && isdir >= 0 && ip) {
3141                         if (isdir &&
3142                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
3143                                 error = ENOTDIR;
3144                         } else if (isdir == 0 &&
3145                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
3146                                 error = EISDIR;
3147                         }
3148                 }
3149
3150                 /*
3151                  * If we are trying to remove a directory the directory must
3152                  * be empty.
3153                  *
3154                  * The check directory code can loop and deadlock/retry.  Our
3155                  * own cursor's node locks must be released to avoid a 3-way
3156                  * deadlock with the flusher if the check directory code
3157                  * blocks.
3158                  *
3159                  * If any changes whatsoever have been made to the cursor
3160                  * set EDEADLK and retry.
3161                  *
3162                  * WARNING: See warnings in hammer_unlock_cursor()
3163                  *          function.
3164                  */
3165                 if (error == 0 && ip && ip->ino_data.obj_type ==
3166                                         HAMMER_OBJTYPE_DIRECTORY) {
3167                         hammer_unlock_cursor(&cursor);
3168                         error = hammer_ip_check_directory_empty(trans, ip);
3169                         hammer_lock_cursor(&cursor);
3170                         if (cursor.flags & HAMMER_CURSOR_RETEST) {
3171                                 kprintf("HAMMER: Warning: avoided deadlock "
3172                                         "on rmdir '%s'\n",
3173                                         ncp->nc_name);
3174                                 error = EDEADLK;
3175                         }
3176                 }
3177
3178                 /*
3179                  * Delete the directory entry.
3180                  *
3181                  * WARNING: hammer_ip_del_directory() may have to terminate
3182                  * the cursor to avoid a deadlock.  It is ok to call
3183                  * hammer_done_cursor() twice.
3184                  */
3185                 if (error == 0) {
3186                         error = hammer_ip_del_directory(trans, &cursor,
3187                                                         dip, ip);
3188                 }
3189                 hammer_done_cursor(&cursor);
3190                 if (error == 0) {
3191                         cache_setunresolved(nch);
3192                         cache_setvp(nch, NULL);
3193                         /* XXX locking */
3194                         if (ip && ip->vp) {
3195                                 hammer_knote(ip->vp, NOTE_DELETE);
3196                                 cache_inval_vp(ip->vp, CINV_DESTROY);
3197                         }
3198                 }
3199                 if (ip)
3200                         hammer_rel_inode(ip, 0);
3201         } else {
3202                 hammer_done_cursor(&cursor);
3203         }
3204         if (error == EDEADLK)
3205                 goto retry;
3206
3207         return (error);
3208 }
3209
3210 /************************************************************************
3211  *                          FIFO AND SPECFS OPS                         *
3212  ************************************************************************
3213  *
3214  */
3215
3216 static int
3217 hammer_vop_fifoclose (struct vop_close_args *ap)
3218 {
3219         /* XXX update itimes */
3220         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
3221 }
3222
3223 static int
3224 hammer_vop_fiforead (struct vop_read_args *ap)
3225 {
3226         int error;
3227
3228         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3229         /* XXX update access time */
3230         return (error);
3231 }
3232
3233 static int
3234 hammer_vop_fifowrite (struct vop_write_args *ap)
3235 {
3236         int error;
3237
3238         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3239         /* XXX update access time */
3240         return (error);
3241 }
3242
3243 static
3244 int
3245 hammer_vop_fifokqfilter(struct vop_kqfilter_args *ap)
3246 {
3247         int error;
3248
3249         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
3250         if (error)
3251                 error = hammer_vop_kqfilter(ap);
3252         return(error);
3253 }
3254
3255 /************************************************************************
3256  *                          KQFILTER OPS                                *
3257  ************************************************************************
3258  *
3259  */
3260 static void filt_hammerdetach(struct knote *kn);
3261 static int filt_hammerread(struct knote *kn, long hint);
3262 static int filt_hammerwrite(struct knote *kn, long hint);
3263 static int filt_hammervnode(struct knote *kn, long hint);
3264
3265 static struct filterops hammerread_filtops =
3266         { 1, NULL, filt_hammerdetach, filt_hammerread };
3267 static struct filterops hammerwrite_filtops =
3268         { 1, NULL, filt_hammerdetach, filt_hammerwrite };
3269 static struct filterops hammervnode_filtops =
3270         { 1, NULL, filt_hammerdetach, filt_hammervnode };
3271
3272 static
3273 int
3274 hammer_vop_kqfilter(struct vop_kqfilter_args *ap)
3275 {
3276         struct vnode *vp = ap->a_vp;
3277         struct knote *kn = ap->a_kn;
3278         lwkt_tokref vlock;
3279
3280         switch (kn->kn_filter) {
3281         case EVFILT_READ:
3282                 kn->kn_fop = &hammerread_filtops;
3283                 break;
3284         case EVFILT_WRITE:
3285                 kn->kn_fop = &hammerwrite_filtops;
3286                 break;
3287         case EVFILT_VNODE:
3288                 kn->kn_fop = &hammervnode_filtops;
3289                 break;
3290         default:
3291                 return (1);
3292         }
3293
3294         kn->kn_hook = (caddr_t)vp;
3295
3296         lwkt_gettoken(&vlock, &vp->v_token);
3297         SLIST_INSERT_HEAD(&vp->v_pollinfo.vpi_selinfo.si_note, kn, kn_selnext);
3298         lwkt_reltoken(&vlock);
3299
3300         return(0);
3301 }
3302
3303 static void
3304 filt_hammerdetach(struct knote *kn)
3305 {
3306         struct vnode *vp = (void *)kn->kn_hook;
3307         lwkt_tokref vlock;
3308
3309         lwkt_gettoken(&vlock, &vp->v_token);
3310         SLIST_REMOVE(&vp->v_pollinfo.vpi_selinfo.si_note,
3311                      kn, knote, kn_selnext);
3312         lwkt_reltoken(&vlock);
3313 }
3314
3315 static int
3316 filt_hammerread(struct knote *kn, long hint)
3317 {
3318         struct vnode *vp = (void *)kn->kn_hook;
3319         hammer_inode_t ip = VTOI(vp);
3320
3321         if (hint == NOTE_REVOKE) {
3322                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3323                 return(1);
3324         }
3325         kn->kn_data = ip->ino_data.size - kn->kn_fp->f_offset;
3326         return (kn->kn_data != 0);
3327 }
3328
3329 static int
3330 filt_hammerwrite(struct knote *kn, long hint)
3331 {
3332         if (hint == NOTE_REVOKE)
3333                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3334         kn->kn_data = 0;
3335         return (1);
3336 }
3337
3338 static int
3339 filt_hammervnode(struct knote *kn, long hint)
3340 {
3341         if (kn->kn_sfflags & hint)
3342                 kn->kn_fflags |= hint;
3343         if (hint == NOTE_REVOKE) {
3344                 kn->kn_flags |= EV_EOF;
3345                 return (1);
3346         }
3347         return (kn->kn_fflags != 0);
3348 }
3349