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 1.100 2008/09/28 05:04:21 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 <vm/vm_extern.h>
  48 #include <vfs/fifofs/fifo.h>
  49 #include "hammer.h"
  50
  51 /*
  52  * USERFS VNOPS
  53  */
  54 /*static int hammer_vop_vnoperate(struct vop_generic_args *);*/
  55 static int hammer_vop_fsync(struct vop_fsync_args *);
  56 static int hammer_vop_read(struct vop_read_args *);
  57 static int hammer_vop_write(struct vop_write_args *);
  58 static int hammer_vop_access(struct vop_access_args *);
  59 static int hammer_vop_advlock(struct vop_advlock_args *);
  60 static int hammer_vop_close(struct vop_close_args *);
  61 static int hammer_vop_ncreate(struct vop_ncreate_args *);
  62 static int hammer_vop_getattr(struct vop_getattr_args *);
  63 static int hammer_vop_nresolve(struct vop_nresolve_args *);
  64 static int hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *);
  65 static int hammer_vop_nlink(struct vop_nlink_args *);
  66 static int hammer_vop_nmkdir(struct vop_nmkdir_args *);
  67 static int hammer_vop_nmknod(struct vop_nmknod_args *);
  68 static int hammer_vop_open(struct vop_open_args *);
  69 static int hammer_vop_print(struct vop_print_args *);
  70 static int hammer_vop_readdir(struct vop_readdir_args *);
  71 static int hammer_vop_readlink(struct vop_readlink_args *);
  72 static int hammer_vop_nremove(struct vop_nremove_args *);
  73 static int hammer_vop_nrename(struct vop_nrename_args *);
  74 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  75 static int hammer_vop_setattr(struct vop_setattr_args *);
  76 static int hammer_vop_strategy(struct vop_strategy_args *);
  77 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  78 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  79 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  80 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  81 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  82
  83 static int hammer_vop_fifoclose (struct vop_close_args *);
  84 static int hammer_vop_fiforead (struct vop_read_args *);
  85 static int hammer_vop_fifowrite (struct vop_write_args *);
  86
  87 static int hammer_vop_specclose (struct vop_close_args *);
  88 static int hammer_vop_specread (struct vop_read_args *);
  89 static int hammer_vop_specwrite (struct vop_write_args *);
  90
  91 struct vop_ops hammer_vnode_vops = {
  92         .vop_default =          vop_defaultop,
  93         .vop_fsync =            hammer_vop_fsync,
  94         .vop_getpages =         vop_stdgetpages,
  95         .vop_putpages =         vop_stdputpages,
  96         .vop_read =             hammer_vop_read,
  97         .vop_write =            hammer_vop_write,
  98         .vop_access =           hammer_vop_access,
  99         .vop_advlock =          hammer_vop_advlock,
 100         .vop_close =            hammer_vop_close,
 101         .vop_ncreate =          hammer_vop_ncreate,
 102         .vop_getattr =          hammer_vop_getattr,
 103         .vop_inactive =         hammer_vop_inactive,
 104         .vop_reclaim =          hammer_vop_reclaim,
 105         .vop_nresolve =         hammer_vop_nresolve,
 106         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
 107         .vop_nlink =            hammer_vop_nlink,
 108         .vop_nmkdir =           hammer_vop_nmkdir,
 109         .vop_nmknod =           hammer_vop_nmknod,
 110         .vop_open =             hammer_vop_open,
 111         .vop_pathconf =         vop_stdpathconf,
 112         .vop_print =            hammer_vop_print,
 113         .vop_readdir =          hammer_vop_readdir,
 114         .vop_readlink =         hammer_vop_readlink,
 115         .vop_nremove =          hammer_vop_nremove,
 116         .vop_nrename =          hammer_vop_nrename,
 117         .vop_nrmdir =           hammer_vop_nrmdir,
 118         .vop_setattr =          hammer_vop_setattr,
 119         .vop_bmap =             hammer_vop_bmap,
 120         .vop_strategy =         hammer_vop_strategy,
 121         .vop_nsymlink =         hammer_vop_nsymlink,
 122         .vop_nwhiteout =        hammer_vop_nwhiteout,
 123         .vop_ioctl =            hammer_vop_ioctl,
 124         .vop_mountctl =         hammer_vop_mountctl
 125 };
 126
 127 struct vop_ops hammer_spec_vops = {
 128         .vop_default =          spec_vnoperate,
 129         .vop_fsync =            hammer_vop_fsync,
 130         .vop_read =             hammer_vop_specread,
 131         .vop_write =            hammer_vop_specwrite,
 132         .vop_access =           hammer_vop_access,
 133         .vop_close =            hammer_vop_specclose,
 134         .vop_getattr =          hammer_vop_getattr,
 135         .vop_inactive =         hammer_vop_inactive,
 136         .vop_reclaim =          hammer_vop_reclaim,
 137         .vop_setattr =          hammer_vop_setattr
 138 };
 139
 140 struct vop_ops hammer_fifo_vops = {
 141         .vop_default =          fifo_vnoperate,
 142         .vop_fsync =            hammer_vop_fsync,
 143         .vop_read =             hammer_vop_fiforead,
 144         .vop_write =            hammer_vop_fifowrite,
 145         .vop_access =           hammer_vop_access,
 146         .vop_close =            hammer_vop_fifoclose,
 147         .vop_getattr =          hammer_vop_getattr,
 148         .vop_inactive =         hammer_vop_inactive,
 149         .vop_reclaim =          hammer_vop_reclaim,
 150         .vop_setattr =          hammer_vop_setattr
 151 };
 152
 153 #ifdef DEBUG_TRUNCATE
 154 struct hammer_inode *HammerTruncIp;
 155 #endif
 156
 157 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 158                            struct vnode *dvp, struct ucred *cred,
 159                            int flags, int isdir);
 160 static int hammer_vop_strategy_read(struct vop_strategy_args *ap);
 161 static int hammer_vop_strategy_write(struct vop_strategy_args *ap);
 162
 163 #if 0
 164 static
 165 int
 166 hammer_vop_vnoperate(struct vop_generic_args *)
 167 {
 168         return (VOCALL(&hammer_vnode_vops, ap));
 169 }
 170 #endif
 171
 172 /*
 173  * hammer_vop_fsync { vp, waitfor }
 174  *
 175  * fsync() an inode to disk and wait for it to be completely committed
 176  * such that the information would not be undone if a crash occured after
 177  * return.
 178  */
 179 static
 180 int
 181 hammer_vop_fsync(struct vop_fsync_args *ap)
 182 {
 183         hammer_inode_t ip = VTOI(ap->a_vp);
 184
 185         ++hammer_count_fsyncs;
 186         vfsync(ap->a_vp, ap->a_waitfor, 1, NULL, NULL);
 187         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 188         if (ap->a_waitfor == MNT_WAIT) {
 189                 vn_unlock(ap->a_vp);
 190                 hammer_wait_inode(ip);
 191                 vn_lock(ap->a_vp, LK_EXCLUSIVE | LK_RETRY);
 192         }
 193         return (ip->error);
 194 }
 195
 196 /*
 197  * hammer_vop_read { vp, uio, ioflag, cred }
 198  */
 199 static
 200 int
 201 hammer_vop_read(struct vop_read_args *ap)
 202 {
 203         struct hammer_transaction trans;
 204         hammer_inode_t ip;
 205         off_t offset;
 206         struct buf *bp;
 207         struct uio *uio;
 208         int error;
 209         int n;
 210         int seqcount;
 211         int ioseqcount;
 212         int blksize;
 213
 214         if (ap->a_vp->v_type != VREG)
 215                 return (EINVAL);
 216         ip = VTOI(ap->a_vp);
 217         error = 0;
 218         uio = ap->a_uio;
 219
 220         /*
 221          * Allow the UIO's size to override the sequential heuristic.
 222          */
 223         blksize = hammer_blocksize(uio->uio_offset);
 224         seqcount = (uio->uio_resid + (blksize - 1)) / blksize;
 225         ioseqcount = ap->a_ioflag >> 16;
 226         if (seqcount < ioseqcount)
 227                 seqcount = ioseqcount;
 228
 229         hammer_start_transaction(&trans, ip->hmp);
 230
 231         /*
 232          * Access the data typically in HAMMER_BUFSIZE blocks via the
 233          * buffer cache, but HAMMER may use a variable block size based
 234          * on the offset.
 235          */
 236         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 237                 int64_t base_offset;
 238                 int64_t file_limit;
 239
 240                 blksize = hammer_blocksize(uio->uio_offset);
 241                 offset = (int)uio->uio_offset & (blksize - 1);
 242                 base_offset = uio->uio_offset - offset;
 243
 244                 if (hammer_cluster_enable) {
 245                         /*
 246                          * Use file_limit to prevent cluster_read() from
 247                          * creating buffers of the wrong block size past
 248                          * the demarc.
 249                          */
 250                         file_limit = ip->ino_data.size;
 251                         if (base_offset < HAMMER_XDEMARC &&
 252                             file_limit > HAMMER_XDEMARC) {
 253                                 file_limit = HAMMER_XDEMARC;
 254                         }
 255                         error = cluster_read(ap->a_vp,
 256                                              file_limit, base_offset,
 257                                              blksize, MAXPHYS,
 258                                              seqcount, &bp);
 259                 } else {
 260                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 261                 }
 262                 if (error) {
 263                         kprintf("error %d\n", error);
 264                         brelse(bp);
 265                         break;
 266                 }
 267
 268                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 269                 n = blksize - offset;
 270                 if (n > uio->uio_resid)
 271                         n = uio->uio_resid;
 272                 if (n > ip->ino_data.size - uio->uio_offset)
 273                         n = (int)(ip->ino_data.size - uio->uio_offset);
 274                 error = uiomove((char *)bp->b_data + offset, n, uio);
 275
 276                 /* data has a lower priority then meta-data */
 277                 bp->b_flags |= B_AGE;
 278                 bqrelse(bp);
 279                 if (error)
 280                         break;
 281                 hammer_stats_file_read += n;
 282         }
 283         if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 284             (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 285                 ip->ino_data.atime = trans.time;
 286                 hammer_modify_inode(ip, HAMMER_INODE_ATIME);
 287         }
 288         hammer_done_transaction(&trans);
 289         return (error);
 290 }
 291
 292 /*
 293  * hammer_vop_write { vp, uio, ioflag, cred }
 294  */
 295 static
 296 int
 297 hammer_vop_write(struct vop_write_args *ap)
 298 {
 299         struct hammer_transaction trans;
 300         struct hammer_inode *ip;
 301         hammer_mount_t hmp;
 302         struct uio *uio;
 303         int offset;
 304         off_t base_offset;
 305         struct buf *bp;
 306         int error;
 307         int n;
 308         int flags;
 309         int delta;
 310         int seqcount;
 311
 312         if (ap->a_vp->v_type != VREG)
 313                 return (EINVAL);
 314         ip = VTOI(ap->a_vp);
 315         hmp = ip->hmp;
 316         error = 0;
 317         seqcount = ap->a_ioflag >> 16;
 318
 319         if (ip->flags & HAMMER_INODE_RO)
 320                 return (EROFS);
 321
 322         /*
 323          * Create a transaction to cover the operations we perform.
 324          */
 325         hammer_start_transaction(&trans, hmp);
 326         uio = ap->a_uio;
 327
 328         /*
 329          * Check append mode
 330          */
 331         if (ap->a_ioflag & IO_APPEND)
 332                 uio->uio_offset = ip->ino_data.size;
 333
 334         /*
 335          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 336          *
 337          * NOTE: the base_off assignment is required to work around what
 338          * I consider to be a GCC-4 optimization bug.
 339          */
 340         if (uio->uio_offset < 0) {
 341                 hammer_done_transaction(&trans);
 342                 return (EFBIG);
 343         }
 344         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 345         if (uio->uio_resid > 0 && base_offset <= 0) {
 346                 hammer_done_transaction(&trans);
 347                 return (EFBIG);
 348         }
 349
 350         /*
 351          * Access the data typically in HAMMER_BUFSIZE blocks via the
 352          * buffer cache, but HAMMER may use a variable block size based
 353          * on the offset.
 354          */
 355         while (uio->uio_resid > 0) {
 356                 int fixsize = 0;
 357                 int blksize;
 358                 int blkmask;
 359
 360                 if ((error = hammer_checkspace(hmp, HAMMER_CHKSPC_WRITE)) != 0)
 361                         break;
 362
 363                 blksize = hammer_blocksize(uio->uio_offset);
 364
 365                 /*
 366                  * Do not allow HAMMER to blow out the buffer cache.  Very
 367                  * large UIOs can lockout other processes due to bwillwrite()
 368                  * mechanics.
 369                  *
 370                  * The hammer inode is not locked during these operations.
 371                  * The vnode is locked which can interfere with the pageout
 372                  * daemon for non-UIO_NOCOPY writes but should not interfere
 373                  * with the buffer cache.  Even so, we cannot afford to
 374                  * allow the pageout daemon to build up too many dirty buffer
 375                  * cache buffers.
 376                  */
 377                 /*if (((int)uio->uio_offset & (blksize - 1)) == 0)*/
 378                 bwillwrite(blksize);
 379
 380                 /*
 381                  * Do not allow HAMMER to blow out system memory by
 382                  * accumulating too many records.   Records are so well
 383                  * decoupled from the buffer cache that it is possible
 384                  * for userland to push data out to the media via
 385                  * direct-write, but build up the records queued to the
 386                  * backend faster then the backend can flush them out.
 387                  * HAMMER has hit its write limit but the frontend has
 388                  * no pushback to slow it down.
 389                  */
 390                 if (hmp->rsv_recs > hammer_limit_recs / 2) {
 391                         /*
 392                          * Get the inode on the flush list
 393                          */
 394                         if (ip->rsv_recs >= 64)
 395                                 hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 396                         else if (ip->rsv_recs >= 16)
 397                                 hammer_flush_inode(ip, 0);
 398
 399                         /*
 400                          * Keep the flusher going if the system keeps
 401                          * queueing records.
 402                          */
 403                         delta = hmp->count_newrecords -
 404                                 hmp->last_newrecords;
 405                         if (delta < 0 || delta > hammer_limit_recs / 2) {
 406                                 hmp->last_newrecords = hmp->count_newrecords;
 407                                 hammer_sync_hmp(hmp, MNT_NOWAIT);
 408                         }
 409
 410                         /*
 411                          * If we have gotten behind start slowing
 412                          * down the writers.
 413                          */
 414                         delta = (hmp->rsv_recs - hammer_limit_recs) *
 415                                 hz / hammer_limit_recs;
 416                         if (delta > 0)
 417                                 tsleep(&trans, 0, "hmrslo", delta);
 418                 }
 419
 420                 /*
 421                  * Calculate the blocksize at the current offset and figure
 422                  * out how much we can actually write.
 423                  */
 424                 blkmask = blksize - 1;
 425                 offset = (int)uio->uio_offset & blkmask;
 426                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 427                 n = blksize - offset;
 428                 if (n > uio->uio_resid)
 429                         n = uio->uio_resid;
 430                 if (uio->uio_offset + n > ip->ino_data.size) {
 431                         vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
 432                         fixsize = 1;
 433                 }
 434
 435                 if (uio->uio_segflg == UIO_NOCOPY) {
 436                         /*
 437                          * Issuing a write with the same data backing the
 438                          * buffer.  Instantiate the buffer to collect the
 439                          * backing vm pages, then read-in any missing bits.
 440                          *
 441                          * This case is used by vop_stdputpages().
 442                          */
 443                         bp = getblk(ap->a_vp, base_offset,
 444                                     blksize, GETBLK_BHEAVY, 0);
 445                         if ((bp->b_flags & B_CACHE) == 0) {
 446                                 bqrelse(bp);
 447                                 error = bread(ap->a_vp, base_offset,
 448                                               blksize, &bp);
 449                         }
 450                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 451                         /*
 452                          * Even though we are entirely overwriting the buffer
 453                          * we may still have to zero it out to avoid a
 454                          * mmap/write visibility issue.
 455                          */
 456                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 457                         if ((bp->b_flags & B_CACHE) == 0)
 458                                 vfs_bio_clrbuf(bp);
 459                 } else if (base_offset >= ip->ino_data.size) {
 460                         /*
 461                          * If the base offset of the buffer is beyond the
 462                          * file EOF, we don't have to issue a read.
 463                          */
 464                         bp = getblk(ap->a_vp, base_offset,
 465                                     blksize, GETBLK_BHEAVY, 0);
 466                         vfs_bio_clrbuf(bp);
 467                 } else {
 468                         /*
 469                          * Partial overwrite, read in any missing bits then
 470                          * replace the portion being written.
 471                          */
 472                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 473                         if (error == 0)
 474                                 bheavy(bp);
 475                 }
 476                 if (error == 0) {
 477                         error = uiomove((char *)bp->b_data + offset,
 478                                         n, uio);
 479                 }
 480
 481                 /*
 482                  * If we screwed up we have to undo any VM size changes we
 483                  * made.
 484                  */
 485                 if (error) {
 486                         brelse(bp);
 487                         if (fixsize) {
 488                                 vtruncbuf(ap->a_vp, ip->ino_data.size,
 489                                           hammer_blocksize(ip->ino_data.size));
 490                         }
 491                         break;
 492                 }
 493                 hammer_stats_file_write += n;
 494                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 495                 if (ip->ino_data.size < uio->uio_offset) {
 496                         ip->ino_data.size = uio->uio_offset;
 497                         flags = HAMMER_INODE_DDIRTY;
 498                         vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
 499                 } else {
 500                         flags = 0;
 501                 }
 502                 ip->ino_data.mtime = trans.time;
 503                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 504                 hammer_modify_inode(ip, flags);
 505
 506                 /*
 507                  * Once we dirty the buffer any cached zone-X offset
 508                  * becomes invalid.  HAMMER NOTE: no-history mode cannot
 509                  * allow overwriting over the same data sector unless
 510                  * we provide UNDOs for the old data, which we don't.
 511                  */
 512                 bp->b_bio2.bio_offset = NOOFFSET;
 513
 514                 /*
 515                  * Final buffer disposition.
 516                  */
 517                 bp->b_flags |= B_AGE;
 518                 if (ap->a_ioflag & IO_SYNC) {
 519                         bwrite(bp);
 520                 } else if (ap->a_ioflag & IO_DIRECT) {
 521                         bawrite(bp);
 522                 } else {
 523                         bdwrite(bp);
 524                 }
 525         }
 526         hammer_done_transaction(&trans);
 527         return (error);
 528 }
 529
 530 /*
 531  * hammer_vop_access { vp, mode, cred }
 532  */
 533 static
 534 int
 535 hammer_vop_access(struct vop_access_args *ap)
 536 {
 537         struct hammer_inode *ip = VTOI(ap->a_vp);
 538         uid_t uid;
 539         gid_t gid;
 540         int error;
 541
 542         ++hammer_stats_file_iopsr;
 543         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 544         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 545
 546         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 547                                   ip->ino_data.uflags);
 548         return (error);
 549 }
 550
 551 /*
 552  * hammer_vop_advlock { vp, id, op, fl, flags }
 553  */
 554 static
 555 int
 556 hammer_vop_advlock(struct vop_advlock_args *ap)
 557 {
 558         hammer_inode_t ip = VTOI(ap->a_vp);
 559
 560         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 561 }
 562
 563 /*
 564  * hammer_vop_close { vp, fflag }
 565  */
 566 static
 567 int
 568 hammer_vop_close(struct vop_close_args *ap)
 569 {
 570         /*hammer_inode_t ip = VTOI(ap->a_vp);*/
 571         return (vop_stdclose(ap));
 572 }
 573
 574 /*
 575  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 576  *
 577  * The operating system has already ensured that the directory entry
 578  * does not exist and done all appropriate namespace locking.
 579  */
 580 static
 581 int
 582 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 583 {
 584         struct hammer_transaction trans;
 585         struct hammer_inode *dip;
 586         struct hammer_inode *nip;
 587         struct nchandle *nch;
 588         int error;
 589
 590         nch = ap->a_nch;
 591         dip = VTOI(ap->a_dvp);
 592
 593         if (dip->flags & HAMMER_INODE_RO)
 594                 return (EROFS);
 595         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
 596                 return (error);
 597
 598         /*
 599          * Create a transaction to cover the operations we perform.
 600          */
 601         hammer_start_transaction(&trans, dip->hmp);
 602         ++hammer_stats_file_iopsw;
 603
 604         /*
 605          * Create a new filesystem object of the requested type.  The
 606          * returned inode will be referenced and shared-locked to prevent
 607          * it from being moved to the flusher.
 608          */
 609
 610         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 611                                     dip, NULL, &nip);
 612         if (error) {
 613                 hkprintf("hammer_create_inode error %d\n", error);
 614                 hammer_done_transaction(&trans);
 615                 *ap->a_vpp = NULL;
 616                 return (error);
 617         }
 618
 619         /*
 620          * Add the new filesystem object to the directory.  This will also
 621          * bump the inode's link count.
 622          */
 623         error = hammer_ip_add_directory(&trans, dip,
 624                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 625                                         nip);
 626         if (error)
 627                 hkprintf("hammer_ip_add_directory error %d\n", error);
 628
 629         /*
 630          * Finish up.
 631          */
 632         if (error) {
 633                 hammer_rel_inode(nip, 0);
 634                 hammer_done_transaction(&trans);
 635                 *ap->a_vpp = NULL;
 636         } else {
 637                 error = hammer_get_vnode(nip, ap->a_vpp);
 638                 hammer_done_transaction(&trans);
 639                 hammer_rel_inode(nip, 0);
 640                 if (error == 0) {
 641                         cache_setunresolved(ap->a_nch);
 642                         cache_setvp(ap->a_nch, *ap->a_vpp);
 643                 }
 644         }
 645         return (error);
 646 }
 647
 648 /*
 649  * hammer_vop_getattr { vp, vap }
 650  *
 651  * Retrieve an inode's attribute information.  When accessing inodes
 652  * historically we fake the atime field to ensure consistent results.
 653  * The atime field is stored in the B-Tree element and allowed to be
 654  * updated without cycling the element.
 655  */
 656 static
 657 int
 658 hammer_vop_getattr(struct vop_getattr_args *ap)
 659 {
 660         struct hammer_inode *ip = VTOI(ap->a_vp);
 661         struct vattr *vap = ap->a_vap;
 662
 663         /*
 664          * We want the fsid to be different when accessing a filesystem
 665          * with different as-of's so programs like diff don't think
 666          * the files are the same.
 667          *
 668          * We also want the fsid to be the same when comparing snapshots,
 669          * or when comparing mirrors (which might be backed by different
 670          * physical devices).  HAMMER fsids are based on the PFS's
 671          * shared_uuid field.
 672          *
 673          * XXX there is a chance of collision here.  The va_fsid reported
 674          * by stat is different from the more involved fsid used in the
 675          * mount structure.
 676          */
 677         ++hammer_stats_file_iopsr;
 678         vap->va_fsid = ip->pfsm->fsid_udev ^ (u_int32_t)ip->obj_asof ^
 679                        (u_int32_t)(ip->obj_asof >> 32);
 680
 681         vap->va_fileid = ip->ino_leaf.base.obj_id;
 682         vap->va_mode = ip->ino_data.mode;
 683         vap->va_nlink = ip->ino_data.nlinks;
 684         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
 685         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
 686         vap->va_rmajor = 0;
 687         vap->va_rminor = 0;
 688         vap->va_size = ip->ino_data.size;
 689
 690         /*
 691          * Special case for @@PFS softlinks.  The actual size of the
 692          * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
 693          */
 694         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_SOFTLINK &&
 695             ip->ino_data.size == 10 &&
 696             ip->obj_asof == HAMMER_MAX_TID &&
 697             ip->obj_localization == 0 &&
 698             strncmp(ip->ino_data.ext.symlink, "@@PFS", 5) == 0) {
 699                     vap->va_size = 26;
 700         }
 701
 702         /*
 703          * We must provide a consistent atime and mtime for snapshots
 704          * so people can do a 'tar cf - ... | md5' on them and get
 705          * consistent results.
 706          */
 707         if (ip->flags & HAMMER_INODE_RO) {
 708                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
 709                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
 710         } else {
 711                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
 712                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
 713         }
 714         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
 715         vap->va_flags = ip->ino_data.uflags;
 716         vap->va_gen = 1;        /* hammer inums are unique for all time */
 717         vap->va_blocksize = HAMMER_BUFSIZE;
 718         if (ip->ino_data.size >= HAMMER_XDEMARC) {
 719                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
 720                                 ~HAMMER_XBUFMASK64;
 721         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
 722                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
 723                                 ~HAMMER_BUFMASK64;
 724         } else {
 725                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
 726         }
 727
 728         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
 729         vap->va_filerev = 0;    /* XXX */
 730         /* mtime uniquely identifies any adjustments made to the file XXX */
 731         vap->va_fsmid = ip->ino_data.mtime;
 732         vap->va_uid_uuid = ip->ino_data.uid;
 733         vap->va_gid_uuid = ip->ino_data.gid;
 734         vap->va_fsid_uuid = ip->hmp->fsid;
 735         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
 736                           VA_FSID_UUID_VALID;
 737
 738         switch (ip->ino_data.obj_type) {
 739         case HAMMER_OBJTYPE_CDEV:
 740         case HAMMER_OBJTYPE_BDEV:
 741                 vap->va_rmajor = ip->ino_data.rmajor;
 742                 vap->va_rminor = ip->ino_data.rminor;
 743                 break;
 744         default:
 745                 break;
 746         }
 747         return(0);
 748 }
 749
 750 /*
 751  * hammer_vop_nresolve { nch, dvp, cred }
 752  *
 753  * Locate the requested directory entry.
 754  */
 755 static
 756 int
 757 hammer_vop_nresolve(struct vop_nresolve_args *ap)
 758 {
 759         struct hammer_transaction trans;
 760         struct namecache *ncp;
 761         hammer_inode_t dip;
 762         hammer_inode_t ip;
 763         hammer_tid_t asof;
 764         struct hammer_cursor cursor;
 765         struct vnode *vp;
 766         int64_t namekey;
 767         int error;
 768         int i;
 769         int nlen;
 770         int flags;
 771         int ispfs;
 772         int64_t obj_id;
 773         u_int32_t localization;
 774
 775         /*
 776          * Misc initialization, plus handle as-of name extensions.  Look for
 777          * the '@@' extension.  Note that as-of files and directories cannot
 778          * be modified.
 779          */
 780         dip = VTOI(ap->a_dvp);
 781         ncp = ap->a_nch->ncp;
 782         asof = dip->obj_asof;
 783         nlen = ncp->nc_nlen;
 784         flags = dip->flags & HAMMER_INODE_RO;
 785         ispfs = 0;
 786
 787         hammer_simple_transaction(&trans, dip->hmp);
 788         ++hammer_stats_file_iopsr;
 789
 790         for (i = 0; i < nlen; ++i) {
 791                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
 792                         asof = hammer_str_to_tid(ncp->nc_name + i + 2,
 793                                                  &ispfs, &localization);
 794                         if (asof != HAMMER_MAX_TID)
 795                                 flags |= HAMMER_INODE_RO;
 796                         break;
 797                 }
 798         }
 799         nlen = i;
 800
 801         /*
 802          * If this is a PFS softlink we dive into the PFS
 803          */
 804         if (ispfs && nlen == 0) {
 805                 ip = hammer_get_inode(&trans, dip, HAMMER_OBJID_ROOT,
 806                                       asof, localization,
 807                                       flags, &error);
 808                 if (error == 0) {
 809                         error = hammer_get_vnode(ip, &vp);
 810                         hammer_rel_inode(ip, 0);
 811                 } else {
 812                         vp = NULL;
 813                 }
 814                 if (error == 0) {
 815                         vn_unlock(vp);
 816                         cache_setvp(ap->a_nch, vp);
 817                         vrele(vp);
 818                 }
 819                 goto done;
 820         }
 821
 822         /*
 823          * If there is no path component the time extension is relative to
 824          * dip.
 825          */
 826         if (nlen == 0) {
 827                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
 828                                       asof, dip->obj_localization,
 829                                       flags, &error);
 830                 if (error == 0) {
 831                         error = hammer_get_vnode(ip, &vp);
 832                         hammer_rel_inode(ip, 0);
 833                 } else {
 834                         vp = NULL;
 835                 }
 836                 if (error == 0) {
 837                         vn_unlock(vp);
 838                         cache_setvp(ap->a_nch, vp);
 839                         vrele(vp);
 840                 }
 841                 goto done;
 842         }
 843
 844         /*
 845          * Calculate the namekey and setup the key range for the scan.  This
 846          * works kinda like a chained hash table where the lower 32 bits
 847          * of the namekey synthesize the chain.
 848          *
 849          * The key range is inclusive of both key_beg and key_end.
 850          */
 851         namekey = hammer_directory_namekey(ncp->nc_name, nlen);
 852
 853         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
 854         cursor.key_beg.localization = dip->obj_localization +
 855                                       HAMMER_LOCALIZE_MISC;
 856         cursor.key_beg.obj_id = dip->obj_id;
 857         cursor.key_beg.key = namekey;
 858         cursor.key_beg.create_tid = 0;
 859         cursor.key_beg.delete_tid = 0;
 860         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
 861         cursor.key_beg.obj_type = 0;
 862
 863         cursor.key_end = cursor.key_beg;
 864         cursor.key_end.key |= 0xFFFFFFFFULL;
 865         cursor.asof = asof;
 866         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
 867
 868         /*
 869          * Scan all matching records (the chain), locate the one matching
 870          * the requested path component.
 871          *
 872          * The hammer_ip_*() functions merge in-memory records with on-disk
 873          * records for the purposes of the search.
 874          */
 875         obj_id = 0;
 876         localization = HAMMER_DEF_LOCALIZATION;
 877
 878         if (error == 0) {
 879                 error = hammer_ip_first(&cursor);
 880                 while (error == 0) {
 881                         error = hammer_ip_resolve_data(&cursor);
 882                         if (error)
 883                                 break;
 884                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
 885                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
 886                                 obj_id = cursor.data->entry.obj_id;
 887                                 localization = cursor.data->entry.localization;
 888                                 break;
 889                         }
 890                         error = hammer_ip_next(&cursor);
 891                 }
 892         }
 893         hammer_done_cursor(&cursor);
 894         if (error == 0) {
 895                 ip = hammer_get_inode(&trans, dip, obj_id,
 896                                       asof, localization,
 897                                       flags, &error);
 898                 if (error == 0) {
 899                         error = hammer_get_vnode(ip, &vp);
 900                         hammer_rel_inode(ip, 0);
 901                 } else {
 902                         vp = NULL;
 903                 }
 904                 if (error == 0) {
 905                         vn_unlock(vp);
 906                         cache_setvp(ap->a_nch, vp);
 907                         vrele(vp);
 908                 }
 909         } else if (error == ENOENT) {
 910                 cache_setvp(ap->a_nch, NULL);
 911         }
 912 done:
 913         hammer_done_transaction(&trans);
 914         return (error);
 915 }
 916
 917 /*
 918  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
 919  *
 920  * Locate the parent directory of a directory vnode.
 921  *
 922  * dvp is referenced but not locked.  *vpp must be returned referenced and
 923  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
 924  * at the root, instead it could indicate that the directory we were in was
 925  * removed.
 926  *
 927  * NOTE: as-of sequences are not linked into the directory structure.  If
 928  * we are at the root with a different asof then the mount point, reload
 929  * the same directory with the mount point's asof.   I'm not sure what this
 930  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
 931  * get confused, but it hasn't been tested.
 932  */
 933 static
 934 int
 935 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
 936 {
 937         struct hammer_transaction trans;
 938         struct hammer_inode *dip;
 939         struct hammer_inode *ip;
 940         int64_t parent_obj_id;
 941         u_int32_t parent_obj_localization;
 942         hammer_tid_t asof;
 943         int error;
 944
 945         dip = VTOI(ap->a_dvp);
 946         asof = dip->obj_asof;
 947
 948         /*
 949          * Whos are parent?  This could be the root of a pseudo-filesystem
 950          * whos parent is in another localization domain.
 951          */
 952         parent_obj_id = dip->ino_data.parent_obj_id;
 953         if (dip->obj_id == HAMMER_OBJID_ROOT)
 954                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
 955         else
 956                 parent_obj_localization = dip->obj_localization;
 957
 958         if (parent_obj_id == 0) {
 959                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
 960                    asof != dip->hmp->asof) {
 961                         parent_obj_id = dip->obj_id;
 962                         asof = dip->hmp->asof;
 963                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
 964                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
 965                                    dip->obj_asof);
 966                 } else {
 967                         *ap->a_vpp = NULL;
 968                         return ENOENT;
 969                 }
 970         }
 971
 972         hammer_simple_transaction(&trans, dip->hmp);
 973         ++hammer_stats_file_iopsr;
 974
 975         ip = hammer_get_inode(&trans, dip, parent_obj_id,
 976                               asof, parent_obj_localization,
 977                               dip->flags, &error);
 978         if (ip) {
 979                 error = hammer_get_vnode(ip, ap->a_vpp);
 980                 hammer_rel_inode(ip, 0);
 981         } else {
 982                 *ap->a_vpp = NULL;
 983         }
 984         hammer_done_transaction(&trans);
 985         return (error);
 986 }
 987
 988 /*
 989  * hammer_vop_nlink { nch, dvp, vp, cred }
 990  */
 991 static
 992 int
 993 hammer_vop_nlink(struct vop_nlink_args *ap)
 994 {
 995         struct hammer_transaction trans;
 996         struct hammer_inode *dip;
 997         struct hammer_inode *ip;
 998         struct nchandle *nch;
 999         int error;
1000
1001         if (ap->a_dvp->v_mount != ap->a_vp->v_mount)
1002                 return(EXDEV);
1003
1004         nch = ap->a_nch;
1005         dip = VTOI(ap->a_dvp);
1006         ip = VTOI(ap->a_vp);
1007
1008         if (dip->obj_localization != ip->obj_localization)
1009                 return(EXDEV);
1010
1011         if (dip->flags & HAMMER_INODE_RO)
1012                 return (EROFS);
1013         if (ip->flags & HAMMER_INODE_RO)
1014                 return (EROFS);
1015         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1016                 return (error);
1017
1018         /*
1019          * Create a transaction to cover the operations we perform.
1020          */
1021         hammer_start_transaction(&trans, dip->hmp);
1022         ++hammer_stats_file_iopsw;
1023
1024         /*
1025          * Add the filesystem object to the directory.  Note that neither
1026          * dip nor ip are referenced or locked, but their vnodes are
1027          * referenced.  This function will bump the inode's link count.
1028          */
1029         error = hammer_ip_add_directory(&trans, dip,
1030                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1031                                         ip);
1032
1033         /*
1034          * Finish up.
1035          */
1036         if (error == 0) {
1037                 cache_setunresolved(nch);
1038                 cache_setvp(nch, ap->a_vp);
1039         }
1040         hammer_done_transaction(&trans);
1041         return (error);
1042 }
1043
1044 /*
1045  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1046  *
1047  * The operating system has already ensured that the directory entry
1048  * does not exist and done all appropriate namespace locking.
1049  */
1050 static
1051 int
1052 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1053 {
1054         struct hammer_transaction trans;
1055         struct hammer_inode *dip;
1056         struct hammer_inode *nip;
1057         struct nchandle *nch;
1058         int error;
1059
1060         nch = ap->a_nch;
1061         dip = VTOI(ap->a_dvp);
1062
1063         if (dip->flags & HAMMER_INODE_RO)
1064                 return (EROFS);
1065         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1066                 return (error);
1067
1068         /*
1069          * Create a transaction to cover the operations we perform.
1070          */
1071         hammer_start_transaction(&trans, dip->hmp);
1072         ++hammer_stats_file_iopsw;
1073
1074         /*
1075          * Create a new filesystem object of the requested type.  The
1076          * returned inode will be referenced but not locked.
1077          */
1078         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1079                                     dip, NULL, &nip);
1080         if (error) {
1081                 hkprintf("hammer_mkdir error %d\n", error);
1082                 hammer_done_transaction(&trans);
1083                 *ap->a_vpp = NULL;
1084                 return (error);
1085         }
1086         /*
1087          * Add the new filesystem object to the directory.  This will also
1088          * bump the inode's link count.
1089          */
1090         error = hammer_ip_add_directory(&trans, dip,
1091                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1092                                         nip);
1093         if (error)
1094                 hkprintf("hammer_mkdir (add) error %d\n", error);
1095
1096         /*
1097          * Finish up.
1098          */
1099         if (error) {
1100                 hammer_rel_inode(nip, 0);
1101                 *ap->a_vpp = NULL;
1102         } else {
1103                 error = hammer_get_vnode(nip, ap->a_vpp);
1104                 hammer_rel_inode(nip, 0);
1105                 if (error == 0) {
1106                         cache_setunresolved(ap->a_nch);
1107                         cache_setvp(ap->a_nch, *ap->a_vpp);
1108                 }
1109         }
1110         hammer_done_transaction(&trans);
1111         return (error);
1112 }
1113
1114 /*
1115  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1116  *
1117  * The operating system has already ensured that the directory entry
1118  * does not exist and done all appropriate namespace locking.
1119  */
1120 static
1121 int
1122 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1123 {
1124         struct hammer_transaction trans;
1125         struct hammer_inode *dip;
1126         struct hammer_inode *nip;
1127         struct nchandle *nch;
1128         int error;
1129
1130         nch = ap->a_nch;
1131         dip = VTOI(ap->a_dvp);
1132
1133         if (dip->flags & HAMMER_INODE_RO)
1134                 return (EROFS);
1135         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1136                 return (error);
1137
1138         /*
1139          * Create a transaction to cover the operations we perform.
1140          */
1141         hammer_start_transaction(&trans, dip->hmp);
1142         ++hammer_stats_file_iopsw;
1143
1144         /*
1145          * Create a new filesystem object of the requested type.  The
1146          * returned inode will be referenced but not locked.
1147          *
1148          * If mknod specifies a directory a pseudo-fs is created.
1149          */
1150         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1151                                     dip, NULL, &nip);
1152         if (error) {
1153                 hammer_done_transaction(&trans);
1154                 *ap->a_vpp = NULL;
1155                 return (error);
1156         }
1157
1158         /*
1159          * Add the new filesystem object to the directory.  This will also
1160          * bump the inode's link count.
1161          */
1162         error = hammer_ip_add_directory(&trans, dip,
1163                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1164                                         nip);
1165
1166         /*
1167          * Finish up.
1168          */
1169         if (error) {
1170                 hammer_rel_inode(nip, 0);
1171                 *ap->a_vpp = NULL;
1172         } else {
1173                 error = hammer_get_vnode(nip, ap->a_vpp);
1174                 hammer_rel_inode(nip, 0);
1175                 if (error == 0) {
1176                         cache_setunresolved(ap->a_nch);
1177                         cache_setvp(ap->a_nch, *ap->a_vpp);
1178                 }
1179         }
1180         hammer_done_transaction(&trans);
1181         return (error);
1182 }
1183
1184 /*
1185  * hammer_vop_open { vp, mode, cred, fp }
1186  */
1187 static
1188 int
1189 hammer_vop_open(struct vop_open_args *ap)
1190 {
1191         hammer_inode_t ip;
1192
1193         ++hammer_stats_file_iopsr;
1194         ip = VTOI(ap->a_vp);
1195
1196         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1197                 return (EROFS);
1198         return(vop_stdopen(ap));
1199 }
1200
1201 /*
1202  * hammer_vop_print { vp }
1203  */
1204 static
1205 int
1206 hammer_vop_print(struct vop_print_args *ap)
1207 {
1208         return EOPNOTSUPP;
1209 }
1210
1211 /*
1212  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1213  */
1214 static
1215 int
1216 hammer_vop_readdir(struct vop_readdir_args *ap)
1217 {
1218         struct hammer_transaction trans;
1219         struct hammer_cursor cursor;
1220         struct hammer_inode *ip;
1221         struct uio *uio;
1222         hammer_base_elm_t base;
1223         int error;
1224         int cookie_index;
1225         int ncookies;
1226         off_t *cookies;
1227         off_t saveoff;
1228         int r;
1229         int dtype;
1230
1231         ++hammer_stats_file_iopsr;
1232         ip = VTOI(ap->a_vp);
1233         uio = ap->a_uio;
1234         saveoff = uio->uio_offset;
1235
1236         if (ap->a_ncookies) {
1237                 ncookies = uio->uio_resid / 16 + 1;
1238                 if (ncookies > 1024)
1239                         ncookies = 1024;
1240                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1241                 cookie_index = 0;
1242         } else {
1243                 ncookies = -1;
1244                 cookies = NULL;
1245                 cookie_index = 0;
1246         }
1247
1248         hammer_simple_transaction(&trans, ip->hmp);
1249
1250         /*
1251          * Handle artificial entries
1252          */
1253         error = 0;
1254         if (saveoff == 0) {
1255                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1256                 if (r)
1257                         goto done;
1258                 if (cookies)
1259                         cookies[cookie_index] = saveoff;
1260                 ++saveoff;
1261                 ++cookie_index;
1262                 if (cookie_index == ncookies)
1263                         goto done;
1264         }
1265         if (saveoff == 1) {
1266                 if (ip->ino_data.parent_obj_id) {
1267                         r = vop_write_dirent(&error, uio,
1268                                              ip->ino_data.parent_obj_id,
1269                                              DT_DIR, 2, "..");
1270                 } else {
1271                         r = vop_write_dirent(&error, uio,
1272                                              ip->obj_id, DT_DIR, 2, "..");
1273                 }
1274                 if (r)
1275                         goto done;
1276                 if (cookies)
1277                         cookies[cookie_index] = saveoff;
1278                 ++saveoff;
1279                 ++cookie_index;
1280                 if (cookie_index == ncookies)
1281                         goto done;
1282         }
1283
1284         /*
1285          * Key range (begin and end inclusive) to scan.  Directory keys
1286          * directly translate to a 64 bit 'seek' position.
1287          */
1288         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1289         cursor.key_beg.localization = ip->obj_localization +
1290                                       HAMMER_LOCALIZE_MISC;
1291         cursor.key_beg.obj_id = ip->obj_id;
1292         cursor.key_beg.create_tid = 0;
1293         cursor.key_beg.delete_tid = 0;
1294         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1295         cursor.key_beg.obj_type = 0;
1296         cursor.key_beg.key = saveoff;
1297
1298         cursor.key_end = cursor.key_beg;
1299         cursor.key_end.key = HAMMER_MAX_KEY;
1300         cursor.asof = ip->obj_asof;
1301         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1302
1303         error = hammer_ip_first(&cursor);
1304
1305         while (error == 0) {
1306                 error = hammer_ip_resolve_data(&cursor);
1307                 if (error)
1308                         break;
1309                 base = &cursor.leaf->base;
1310                 saveoff = base->key;
1311                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1312
1313                 if (base->obj_id != ip->obj_id)
1314                         panic("readdir: bad record at %p", cursor.node);
1315
1316                 /*
1317                  * Convert pseudo-filesystems into softlinks
1318                  */
1319                 dtype = hammer_get_dtype(cursor.leaf->base.obj_type);
1320                 r = vop_write_dirent(
1321                              &error, uio, cursor.data->entry.obj_id,
1322                              dtype,
1323                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1324                              (void *)cursor.data->entry.name);
1325                 if (r)
1326                         break;
1327                 ++saveoff;
1328                 if (cookies)
1329                         cookies[cookie_index] = base->key;
1330                 ++cookie_index;
1331                 if (cookie_index == ncookies)
1332                         break;
1333                 error = hammer_ip_next(&cursor);
1334         }
1335         hammer_done_cursor(&cursor);
1336
1337 done:
1338         hammer_done_transaction(&trans);
1339
1340         if (ap->a_eofflag)
1341                 *ap->a_eofflag = (error == ENOENT);
1342         uio->uio_offset = saveoff;
1343         if (error && cookie_index == 0) {
1344                 if (error == ENOENT)
1345                         error = 0;
1346                 if (cookies) {
1347                         kfree(cookies, M_TEMP);
1348                         *ap->a_ncookies = 0;
1349                         *ap->a_cookies = NULL;
1350                 }
1351         } else {
1352                 if (error == ENOENT)
1353                         error = 0;
1354                 if (cookies) {
1355                         *ap->a_ncookies = cookie_index;
1356                         *ap->a_cookies = cookies;
1357                 }
1358         }
1359         return(error);
1360 }
1361
1362 /*
1363  * hammer_vop_readlink { vp, uio, cred }
1364  */
1365 static
1366 int
1367 hammer_vop_readlink(struct vop_readlink_args *ap)
1368 {
1369         struct hammer_transaction trans;
1370         struct hammer_cursor cursor;
1371         struct hammer_inode *ip;
1372         char buf[32];
1373         u_int32_t localization;
1374         hammer_pseudofs_inmem_t pfsm;
1375         int error;
1376
1377         ip = VTOI(ap->a_vp);
1378
1379         /*
1380          * Shortcut if the symlink data was stuffed into ino_data.
1381          *
1382          * Also expand special "@@PFS%05d" softlinks (expansion only
1383          * occurs for non-historical (current) accesses made from the
1384          * primary filesystem).
1385          */
1386         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1387                 char *ptr;
1388                 int bytes;
1389
1390                 ptr = ip->ino_data.ext.symlink;
1391                 bytes = (int)ip->ino_data.size;
1392                 if (bytes == 10 &&
1393                     ip->obj_asof == HAMMER_MAX_TID &&
1394                     ip->obj_localization == 0 &&
1395                     strncmp(ptr, "@@PFS", 5) == 0) {
1396                         hammer_simple_transaction(&trans, ip->hmp);
1397                         bcopy(ptr + 5, buf, 5);
1398                         buf[5] = 0;
1399                         localization = strtoul(buf, NULL, 10) << 16;
1400                         pfsm = hammer_load_pseudofs(&trans, localization,
1401                                                     &error);
1402                         if (error == 0) {
1403                                 if (pfsm->pfsd.mirror_flags &
1404                                     HAMMER_PFSD_SLAVE) {
1405                                         ksnprintf(buf, sizeof(buf),
1406                                                   "@@0x%016llx:%05d",
1407                                                   pfsm->pfsd.sync_end_tid,
1408                                                   localization >> 16);
1409                                 } else {
1410                                         ksnprintf(buf, sizeof(buf),
1411                                                   "@@0x%016llx:%05d",
1412                                                   HAMMER_MAX_TID,
1413                                                   localization >> 16);
1414                                 }
1415                                 ptr = buf;
1416                                 bytes = strlen(buf);
1417                         }
1418                         if (pfsm)
1419                                 hammer_rel_pseudofs(trans.hmp, pfsm);
1420                         hammer_done_transaction(&trans);
1421                 }
1422                 error = uiomove(ptr, bytes, ap->a_uio);
1423                 return(error);
1424         }
1425
1426         /*
1427          * Long version
1428          */
1429         hammer_simple_transaction(&trans, ip->hmp);
1430         ++hammer_stats_file_iopsr;
1431         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1432
1433         /*
1434          * Key range (begin and end inclusive) to scan.  Directory keys
1435          * directly translate to a 64 bit 'seek' position.
1436          */
1437         cursor.key_beg.localization = ip->obj_localization +
1438                                       HAMMER_LOCALIZE_MISC;
1439         cursor.key_beg.obj_id = ip->obj_id;
1440         cursor.key_beg.create_tid = 0;
1441         cursor.key_beg.delete_tid = 0;
1442         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1443         cursor.key_beg.obj_type = 0;
1444         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1445         cursor.asof = ip->obj_asof;
1446         cursor.flags |= HAMMER_CURSOR_ASOF;
1447
1448         error = hammer_ip_lookup(&cursor);
1449         if (error == 0) {
1450                 error = hammer_ip_resolve_data(&cursor);
1451                 if (error == 0) {
1452                         KKASSERT(cursor.leaf->data_len >=
1453                                  HAMMER_SYMLINK_NAME_OFF);
1454                         error = uiomove(cursor.data->symlink.name,
1455                                         cursor.leaf->data_len -
1456                                                 HAMMER_SYMLINK_NAME_OFF,
1457                                         ap->a_uio);
1458                 }
1459         }
1460         hammer_done_cursor(&cursor);
1461         hammer_done_transaction(&trans);
1462         return(error);
1463 }
1464
1465 /*
1466  * hammer_vop_nremove { nch, dvp, cred }
1467  */
1468 static
1469 int
1470 hammer_vop_nremove(struct vop_nremove_args *ap)
1471 {
1472         struct hammer_transaction trans;
1473         struct hammer_inode *dip;
1474         int error;
1475
1476         dip = VTOI(ap->a_dvp);
1477
1478         if (hammer_nohistory(dip) == 0 &&
1479             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1480                 return (error);
1481         }
1482
1483         hammer_start_transaction(&trans, dip->hmp);
1484         ++hammer_stats_file_iopsw;
1485         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 0);
1486         hammer_done_transaction(&trans);
1487
1488         return (error);
1489 }
1490
1491 /*
1492  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1493  */
1494 static
1495 int
1496 hammer_vop_nrename(struct vop_nrename_args *ap)
1497 {
1498         struct hammer_transaction trans;
1499         struct namecache *fncp;
1500         struct namecache *tncp;
1501         struct hammer_inode *fdip;
1502         struct hammer_inode *tdip;
1503         struct hammer_inode *ip;
1504         struct hammer_cursor cursor;
1505         int64_t namekey;
1506         int nlen, error;
1507
1508         if (ap->a_fdvp->v_mount != ap->a_tdvp->v_mount)
1509                 return(EXDEV);
1510         if (ap->a_fdvp->v_mount != ap->a_fnch->ncp->nc_vp->v_mount)
1511                 return(EXDEV);
1512
1513         fdip = VTOI(ap->a_fdvp);
1514         tdip = VTOI(ap->a_tdvp);
1515         fncp = ap->a_fnch->ncp;
1516         tncp = ap->a_tnch->ncp;
1517         ip = VTOI(fncp->nc_vp);
1518         KKASSERT(ip != NULL);
1519
1520         if (fdip->obj_localization != tdip->obj_localization)
1521                 return(EXDEV);
1522         if (fdip->obj_localization != ip->obj_localization)
1523                 return(EXDEV);
1524
1525         if (fdip->flags & HAMMER_INODE_RO)
1526                 return (EROFS);
1527         if (tdip->flags & HAMMER_INODE_RO)
1528                 return (EROFS);
1529         if (ip->flags & HAMMER_INODE_RO)
1530                 return (EROFS);
1531         if ((error = hammer_checkspace(fdip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1532                 return (error);
1533
1534         hammer_start_transaction(&trans, fdip->hmp);
1535         ++hammer_stats_file_iopsw;
1536
1537         /*
1538          * Remove tncp from the target directory and then link ip as
1539          * tncp. XXX pass trans to dounlink
1540          *
1541          * Force the inode sync-time to match the transaction so it is
1542          * in-sync with the creation of the target directory entry.
1543          */
1544         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp,
1545                                 ap->a_cred, 0, -1);
1546         if (error == 0 || error == ENOENT) {
1547                 error = hammer_ip_add_directory(&trans, tdip,
1548                                                 tncp->nc_name, tncp->nc_nlen,
1549                                                 ip);
1550                 if (error == 0) {
1551                         ip->ino_data.parent_obj_id = tdip->obj_id;
1552                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1553                 }
1554         }
1555         if (error)
1556                 goto failed; /* XXX */
1557
1558         /*
1559          * Locate the record in the originating directory and remove it.
1560          *
1561          * Calculate the namekey and setup the key range for the scan.  This
1562          * works kinda like a chained hash table where the lower 32 bits
1563          * of the namekey synthesize the chain.
1564          *
1565          * The key range is inclusive of both key_beg and key_end.
1566          */
1567         namekey = hammer_directory_namekey(fncp->nc_name, fncp->nc_nlen);
1568 retry:
1569         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1570         cursor.key_beg.localization = fdip->obj_localization +
1571                                       HAMMER_LOCALIZE_MISC;
1572         cursor.key_beg.obj_id = fdip->obj_id;
1573         cursor.key_beg.key = namekey;
1574         cursor.key_beg.create_tid = 0;
1575         cursor.key_beg.delete_tid = 0;
1576         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1577         cursor.key_beg.obj_type = 0;
1578
1579         cursor.key_end = cursor.key_beg;
1580         cursor.key_end.key |= 0xFFFFFFFFULL;
1581         cursor.asof = fdip->obj_asof;
1582         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1583
1584         /*
1585          * Scan all matching records (the chain), locate the one matching
1586          * the requested path component.
1587          *
1588          * The hammer_ip_*() functions merge in-memory records with on-disk
1589          * records for the purposes of the search.
1590          */
1591         error = hammer_ip_first(&cursor);
1592         while (error == 0) {
1593                 if (hammer_ip_resolve_data(&cursor) != 0)
1594                         break;
1595                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1596                 KKASSERT(nlen > 0);
1597                 if (fncp->nc_nlen == nlen &&
1598                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1599                         break;
1600                 }
1601                 error = hammer_ip_next(&cursor);
1602         }
1603
1604         /*
1605          * If all is ok we have to get the inode so we can adjust nlinks.
1606          *
1607          * WARNING: hammer_ip_del_directory() may have to terminate the
1608          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1609          * twice.
1610          */
1611         if (error == 0)
1612                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1613
1614         /*
1615          * XXX A deadlock here will break rename's atomicy for the purposes
1616          * of crash recovery.
1617          */
1618         if (error == EDEADLK) {
1619                 hammer_done_cursor(&cursor);
1620                 goto retry;
1621         }
1622
1623         /*
1624          * Cleanup and tell the kernel that the rename succeeded.
1625          */
1626         hammer_done_cursor(&cursor);
1627         if (error == 0)
1628                 cache_rename(ap->a_fnch, ap->a_tnch);
1629
1630 failed:
1631         hammer_done_transaction(&trans);
1632         return (error);
1633 }
1634
1635 /*
1636  * hammer_vop_nrmdir { nch, dvp, cred }
1637  */
1638 static
1639 int
1640 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1641 {
1642         struct hammer_transaction trans;
1643         struct hammer_inode *dip;
1644         int error;
1645
1646         dip = VTOI(ap->a_dvp);
1647
1648         if (hammer_nohistory(dip) == 0 &&
1649             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1650                 return (error);
1651         }
1652
1653         hammer_start_transaction(&trans, dip->hmp);
1654         ++hammer_stats_file_iopsw;
1655         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0, 1);
1656         hammer_done_transaction(&trans);
1657
1658         return (error);
1659 }
1660
1661 /*
1662  * hammer_vop_setattr { vp, vap, cred }
1663  */
1664 static
1665 int
1666 hammer_vop_setattr(struct vop_setattr_args *ap)
1667 {
1668         struct hammer_transaction trans;
1669         struct vattr *vap;
1670         struct hammer_inode *ip;
1671         int modflags;
1672         int error;
1673         int truncating;
1674         int blksize;
1675         int64_t aligned_size;
1676         u_int32_t flags;
1677
1678         vap = ap->a_vap;
1679         ip = ap->a_vp->v_data;
1680         modflags = 0;
1681
1682         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1683                 return(EROFS);
1684         if (ip->flags & HAMMER_INODE_RO)
1685                 return (EROFS);
1686         if (hammer_nohistory(ip) == 0 &&
1687             (error = hammer_checkspace(ip->hmp, HAMMER_CHKSPC_REMOVE)) != 0) {
1688                 return (error);
1689         }
1690
1691         hammer_start_transaction(&trans, ip->hmp);
1692         ++hammer_stats_file_iopsw;
1693         error = 0;
1694
1695         if (vap->va_flags != VNOVAL) {
1696                 flags = ip->ino_data.uflags;
1697                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1698                                          hammer_to_unix_xid(&ip->ino_data.uid),
1699                                          ap->a_cred);
1700                 if (error == 0) {
1701                         if (ip->ino_data.uflags != flags) {
1702                                 ip->ino_data.uflags = flags;
1703                                 modflags |= HAMMER_INODE_DDIRTY;
1704                         }
1705                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1706                                 error = 0;
1707                                 goto done;
1708                         }
1709                 }
1710                 goto done;
1711         }
1712         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1713                 error = EPERM;
1714                 goto done;
1715         }
1716         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
1717                 mode_t cur_mode = ip->ino_data.mode;
1718                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1719                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1720                 uuid_t uuid_uid;
1721                 uuid_t uuid_gid;
1722
1723                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
1724                                          ap->a_cred,
1725                                          &cur_uid, &cur_gid, &cur_mode);
1726                 if (error == 0) {
1727                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
1728                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
1729                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
1730                                  sizeof(uuid_uid)) ||
1731                             bcmp(&uuid_gid, &ip->ino_data.gid,
1732                                  sizeof(uuid_gid)) ||
1733                             ip->ino_data.mode != cur_mode
1734                         ) {
1735                                 ip->ino_data.uid = uuid_uid;
1736                                 ip->ino_data.gid = uuid_gid;
1737                                 ip->ino_data.mode = cur_mode;
1738                         }
1739                         modflags |= HAMMER_INODE_DDIRTY;
1740                 }
1741         }
1742         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
1743                 switch(ap->a_vp->v_type) {
1744                 case VREG:
1745                         if (vap->va_size == ip->ino_data.size)
1746                                 break;
1747                         /*
1748                          * XXX break atomicy, we can deadlock the backend
1749                          * if we do not release the lock.  Probably not a
1750                          * big deal here.
1751                          */
1752                         blksize = hammer_blocksize(vap->va_size);
1753                         if (vap->va_size < ip->ino_data.size) {
1754                                 vtruncbuf(ap->a_vp, vap->va_size, blksize);
1755                                 truncating = 1;
1756                         } else {
1757                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
1758                                 truncating = 0;
1759                         }
1760                         ip->ino_data.size = vap->va_size;
1761                         modflags |= HAMMER_INODE_DDIRTY;
1762
1763                         /*
1764                          * on-media truncation is cached in the inode until
1765                          * the inode is synchronized.
1766                          */
1767                         if (truncating) {
1768                                 hammer_ip_frontend_trunc(ip, vap->va_size);
1769 #ifdef DEBUG_TRUNCATE
1770                                 if (HammerTruncIp == NULL)
1771                                         HammerTruncIp = ip;
1772 #endif
1773                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1774                                         ip->flags |= HAMMER_INODE_TRUNCATED;
1775                                         ip->trunc_off = vap->va_size;
1776 #ifdef DEBUG_TRUNCATE
1777                                         if (ip == HammerTruncIp)
1778                                         kprintf("truncate1 %016llx\n", ip->trunc_off);
1779 #endif
1780                                 } else if (ip->trunc_off > vap->va_size) {
1781                                         ip->trunc_off = vap->va_size;
1782 #ifdef DEBUG_TRUNCATE
1783                                         if (ip == HammerTruncIp)
1784                                         kprintf("truncate2 %016llx\n", ip->trunc_off);
1785 #endif
1786                                 } else {
1787 #ifdef DEBUG_TRUNCATE
1788                                         if (ip == HammerTruncIp)
1789                                         kprintf("truncate3 %016llx (ignored)\n", vap->va_size);
1790 #endif
1791                                 }
1792                         }
1793
1794                         /*
1795                          * If truncating we have to clean out a portion of
1796                          * the last block on-disk.  We do this in the
1797                          * front-end buffer cache.
1798                          */
1799                         aligned_size = (vap->va_size + (blksize - 1)) &
1800                                        ~(int64_t)(blksize - 1);
1801                         if (truncating && vap->va_size < aligned_size) {
1802                                 struct buf *bp;
1803                                 int offset;
1804
1805                                 aligned_size -= blksize;
1806
1807                                 offset = (int)vap->va_size & (blksize - 1);
1808                                 error = bread(ap->a_vp, aligned_size,
1809                                               blksize, &bp);
1810                                 hammer_ip_frontend_trunc(ip, aligned_size);
1811                                 if (error == 0) {
1812                                         bzero(bp->b_data + offset,
1813                                               blksize - offset);
1814                                         /* must de-cache direct-io offset */
1815                                         bp->b_bio2.bio_offset = NOOFFSET;
1816                                         bdwrite(bp);
1817                                 } else {
1818                                         kprintf("ERROR %d\n", error);
1819                                         brelse(bp);
1820                                 }
1821                         }
1822                         break;
1823                 case VDATABASE:
1824                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1825                                 ip->flags |= HAMMER_INODE_TRUNCATED;
1826                                 ip->trunc_off = vap->va_size;
1827                         } else if (ip->trunc_off > vap->va_size) {
1828                                 ip->trunc_off = vap->va_size;
1829                         }
1830                         hammer_ip_frontend_trunc(ip, vap->va_size);
1831                         ip->ino_data.size = vap->va_size;
1832                         modflags |= HAMMER_INODE_DDIRTY;
1833                         break;
1834                 default:
1835                         error = EINVAL;
1836                         goto done;
1837                 }
1838                 break;
1839         }
1840         if (vap->va_atime.tv_sec != VNOVAL) {
1841                 ip->ino_data.atime =
1842                         hammer_timespec_to_time(&vap->va_atime);
1843                 modflags |= HAMMER_INODE_ATIME;
1844         }
1845         if (vap->va_mtime.tv_sec != VNOVAL) {
1846                 ip->ino_data.mtime =
1847                         hammer_timespec_to_time(&vap->va_mtime);
1848                 modflags |= HAMMER_INODE_MTIME;
1849         }
1850         if (vap->va_mode != (mode_t)VNOVAL) {
1851                 mode_t   cur_mode = ip->ino_data.mode;
1852                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1853                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1854
1855                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
1856                                          cur_uid, cur_gid, &cur_mode);
1857                 if (error == 0 && ip->ino_data.mode != cur_mode) {
1858                         ip->ino_data.mode = cur_mode;
1859                         modflags |= HAMMER_INODE_DDIRTY;
1860                 }
1861         }
1862 done:
1863         if (error == 0)
1864                 hammer_modify_inode(ip, modflags);
1865         hammer_done_transaction(&trans);
1866         return (error);
1867 }
1868
1869 /*
1870  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
1871  */
1872 static
1873 int
1874 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
1875 {
1876         struct hammer_transaction trans;
1877         struct hammer_inode *dip;
1878         struct hammer_inode *nip;
1879         struct nchandle *nch;
1880         hammer_record_t record;
1881         int error;
1882         int bytes;
1883
1884         ap->a_vap->va_type = VLNK;
1885
1886         nch = ap->a_nch;
1887         dip = VTOI(ap->a_dvp);
1888
1889         if (dip->flags & HAMMER_INODE_RO)
1890                 return (EROFS);
1891         if ((error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0)
1892                 return (error);
1893
1894         /*
1895          * Create a transaction to cover the operations we perform.
1896          */
1897         hammer_start_transaction(&trans, dip->hmp);
1898         ++hammer_stats_file_iopsw;
1899
1900         /*
1901          * Create a new filesystem object of the requested type.  The
1902          * returned inode will be referenced but not locked.
1903          */
1904
1905         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1906                                     dip, NULL, &nip);
1907         if (error) {
1908                 hammer_done_transaction(&trans);
1909                 *ap->a_vpp = NULL;
1910                 return (error);
1911         }
1912
1913         /*
1914          * Add a record representing the symlink.  symlink stores the link
1915          * as pure data, not a string, and is no \0 terminated.
1916          */
1917         if (error == 0) {
1918                 bytes = strlen(ap->a_target);
1919
1920                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
1921                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
1922                 } else {
1923                         record = hammer_alloc_mem_record(nip, bytes);
1924                         record->type = HAMMER_MEM_RECORD_GENERAL;
1925
1926                         record->leaf.base.localization = nip->obj_localization +
1927                                                          HAMMER_LOCALIZE_MISC;
1928                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
1929                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
1930                         record->leaf.data_len = bytes;
1931                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
1932                         bcopy(ap->a_target, record->data->symlink.name, bytes);
1933                         error = hammer_ip_add_record(&trans, record);
1934                 }
1935
1936                 /*
1937                  * Set the file size to the length of the link.
1938                  */
1939                 if (error == 0) {
1940                         nip->ino_data.size = bytes;
1941                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
1942                 }
1943         }
1944         if (error == 0)
1945                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
1946                                                 nch->ncp->nc_nlen, nip);
1947
1948         /*
1949          * Finish up.
1950          */
1951         if (error) {
1952                 hammer_rel_inode(nip, 0);
1953                 *ap->a_vpp = NULL;
1954         } else {
1955                 error = hammer_get_vnode(nip, ap->a_vpp);
1956                 hammer_rel_inode(nip, 0);
1957                 if (error == 0) {
1958                         cache_setunresolved(ap->a_nch);
1959                         cache_setvp(ap->a_nch, *ap->a_vpp);
1960                 }
1961         }
1962         hammer_done_transaction(&trans);
1963         return (error);
1964 }
1965
1966 /*
1967  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
1968  */
1969 static
1970 int
1971 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
1972 {
1973         struct hammer_transaction trans;
1974         struct hammer_inode *dip;
1975         int error;
1976
1977         dip = VTOI(ap->a_dvp);
1978
1979         if (hammer_nohistory(dip) == 0 &&
1980             (error = hammer_checkspace(dip->hmp, HAMMER_CHKSPC_CREATE)) != 0) {
1981                 return (error);
1982         }
1983
1984         hammer_start_transaction(&trans, dip->hmp);
1985         ++hammer_stats_file_iopsw;
1986         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
1987                                 ap->a_cred, ap->a_flags, -1);
1988         hammer_done_transaction(&trans);
1989
1990         return (error);
1991 }
1992
1993 /*
1994  * hammer_vop_ioctl { vp, command, data, fflag, cred }
1995  */
1996 static
1997 int
1998 hammer_vop_ioctl(struct vop_ioctl_args *ap)
1999 {
2000         struct hammer_inode *ip = ap->a_vp->v_data;
2001
2002         ++hammer_stats_file_iopsr;
2003         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
2004                             ap->a_fflag, ap->a_cred));
2005 }
2006
2007 static
2008 int
2009 hammer_vop_mountctl(struct vop_mountctl_args *ap)
2010 {
2011         struct mount *mp;
2012         int error;
2013
2014         mp = ap->a_head.a_ops->head.vv_mount;
2015
2016         switch(ap->a_op) {
2017         case MOUNTCTL_SET_EXPORT:
2018                 if (ap->a_ctllen != sizeof(struct export_args))
2019                         error = EINVAL;
2020                 else
2021                         error = hammer_vfs_export(mp, ap->a_op,
2022                                       (const struct export_args *)ap->a_ctl);
2023                 break;
2024         default:
2025                 error = journal_mountctl(ap);
2026                 break;
2027         }
2028         return(error);
2029 }
2030
2031 /*
2032  * hammer_vop_strategy { vp, bio }
2033  *
2034  * Strategy call, used for regular file read & write only.  Note that the
2035  * bp may represent a cluster.
2036  *
2037  * To simplify operation and allow better optimizations in the future,
2038  * this code does not make any assumptions with regards to buffer alignment
2039  * or size.
2040  */
2041 static
2042 int
2043 hammer_vop_strategy(struct vop_strategy_args *ap)
2044 {
2045         struct buf *bp;
2046         int error;
2047
2048         bp = ap->a_bio->bio_buf;
2049
2050         switch(bp->b_cmd) {
2051         case BUF_CMD_READ:
2052                 error = hammer_vop_strategy_read(ap);
2053                 break;
2054         case BUF_CMD_WRITE:
2055                 error = hammer_vop_strategy_write(ap);
2056                 break;
2057         default:
2058                 bp->b_error = error = EINVAL;
2059                 bp->b_flags |= B_ERROR;
2060                 biodone(ap->a_bio);
2061                 break;
2062         }
2063         return (error);
2064 }
2065
2066 /*
2067  * Read from a regular file.  Iterate the related records and fill in the
2068  * BIO/BUF.  Gaps are zero-filled.
2069  *
2070  * The support code in hammer_object.c should be used to deal with mixed
2071  * in-memory and on-disk records.
2072  *
2073  * NOTE: Can be called from the cluster code with an oversized buf.
2074  *
2075  * XXX atime update
2076  */
2077 static
2078 int
2079 hammer_vop_strategy_read(struct vop_strategy_args *ap)
2080 {
2081         struct hammer_transaction trans;
2082         struct hammer_inode *ip;
2083         struct hammer_cursor cursor;
2084         hammer_base_elm_t base;
2085         hammer_off_t disk_offset;
2086         struct bio *bio;
2087         struct bio *nbio;
2088         struct buf *bp;
2089         int64_t rec_offset;
2090         int64_t ran_end;
2091         int64_t tmp64;
2092         int error;
2093         int boff;
2094         int roff;
2095         int n;
2096
2097         bio = ap->a_bio;
2098         bp = bio->bio_buf;
2099         ip = ap->a_vp->v_data;
2100
2101         /*
2102          * The zone-2 disk offset may have been set by the cluster code via
2103          * a BMAP operation, or else should be NOOFFSET.
2104          *
2105          * Checking the high bits for a match against zone-2 should suffice.
2106          */
2107         nbio = push_bio(bio);
2108         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2109             HAMMER_ZONE_LARGE_DATA) {
2110                 error = hammer_io_direct_read(ip->hmp, nbio, NULL);
2111                 return (error);
2112         }
2113
2114         /*
2115          * Well, that sucked.  Do it the hard way.  If all the stars are
2116          * aligned we may still be able to issue a direct-read.
2117          */
2118         hammer_simple_transaction(&trans, ip->hmp);
2119         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2120
2121         /*
2122          * Key range (begin and end inclusive) to scan.  Note that the key's
2123          * stored in the actual records represent BASE+LEN, not BASE.  The
2124          * first record containing bio_offset will have a key > bio_offset.
2125          */
2126         cursor.key_beg.localization = ip->obj_localization +
2127                                       HAMMER_LOCALIZE_MISC;
2128         cursor.key_beg.obj_id = ip->obj_id;
2129         cursor.key_beg.create_tid = 0;
2130         cursor.key_beg.delete_tid = 0;
2131         cursor.key_beg.obj_type = 0;
2132         cursor.key_beg.key = bio->bio_offset + 1;
2133         cursor.asof = ip->obj_asof;
2134         cursor.flags |= HAMMER_CURSOR_ASOF;
2135
2136         cursor.key_end = cursor.key_beg;
2137         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2138 #if 0
2139         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2140                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2141                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2142                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2143         } else
2144 #endif
2145         {
2146                 ran_end = bio->bio_offset + bp->b_bufsize;
2147                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2148                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2149                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2150                 if (tmp64 < ran_end)
2151                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2152                 else
2153                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2154         }
2155         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2156
2157         error = hammer_ip_first(&cursor);
2158         boff = 0;
2159
2160         while (error == 0) {
2161                 /*
2162                  * Get the base file offset of the record.  The key for
2163                  * data records is (base + bytes) rather then (base).
2164                  */
2165                 base = &cursor.leaf->base;
2166                 rec_offset = base->key - cursor.leaf->data_len;
2167
2168                 /*
2169                  * Calculate the gap, if any, and zero-fill it.
2170                  *
2171                  * n is the offset of the start of the record verses our
2172                  * current seek offset in the bio.
2173                  */
2174                 n = (int)(rec_offset - (bio->bio_offset + boff));
2175                 if (n > 0) {
2176                         if (n > bp->b_bufsize - boff)
2177                                 n = bp->b_bufsize - boff;
2178                         bzero((char *)bp->b_data + boff, n);
2179                         boff += n;
2180                         n = 0;
2181                 }
2182
2183                 /*
2184                  * Calculate the data offset in the record and the number
2185                  * of bytes we can copy.
2186                  *
2187                  * There are two degenerate cases.  First, boff may already
2188                  * be at bp->b_bufsize.  Secondly, the data offset within
2189                  * the record may exceed the record's size.
2190                  */
2191                 roff = -n;
2192                 rec_offset += roff;
2193                 n = cursor.leaf->data_len - roff;
2194                 if (n <= 0) {
2195                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2196                         n = 0;
2197                 } else if (n > bp->b_bufsize - boff) {
2198                         n = bp->b_bufsize - boff;
2199                 }
2200
2201                 /*
2202                  * Deal with cached truncations.  This cool bit of code
2203                  * allows truncate()/ftruncate() to avoid having to sync
2204                  * the file.
2205                  *
2206                  * If the frontend is truncated then all backend records are
2207                  * subject to the frontend's truncation.
2208                  *
2209                  * If the backend is truncated then backend records on-disk
2210                  * (but not in-memory) are subject to the backend's
2211                  * truncation.  In-memory records owned by the backend
2212                  * represent data written after the truncation point on the
2213                  * backend and must not be truncated.
2214                  *
2215                  * Truncate operations deal with frontend buffer cache
2216                  * buffers and frontend-owned in-memory records synchronously.
2217                  */
2218                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2219                         if (hammer_cursor_ondisk(&cursor) ||
2220                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2221                                 if (ip->trunc_off <= rec_offset)
2222                                         n = 0;
2223                                 else if (ip->trunc_off < rec_offset + n)
2224                                         n = (int)(ip->trunc_off - rec_offset);
2225                         }
2226                 }
2227                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2228                         if (hammer_cursor_ondisk(&cursor)) {
2229                                 if (ip->sync_trunc_off <= rec_offset)
2230                                         n = 0;
2231                                 else if (ip->sync_trunc_off < rec_offset + n)
2232                                         n = (int)(ip->sync_trunc_off - rec_offset);
2233                         }
2234                 }
2235
2236                 /*
2237                  * Try to issue a direct read into our bio if possible,
2238                  * otherwise resolve the element data into a hammer_buffer
2239                  * and copy.
2240                  *
2241                  * The buffer on-disk should be zerod past any real
2242                  * truncation point, but may not be for any synthesized
2243                  * truncation point from above.
2244                  */
2245                 disk_offset = cursor.leaf->data_offset + roff;
2246                 if (boff == 0 && n == bp->b_bufsize &&
2247                     hammer_cursor_ondisk(&cursor) &&
2248                     (disk_offset & HAMMER_BUFMASK) == 0) {
2249                         KKASSERT((disk_offset & HAMMER_OFF_ZONE_MASK) ==
2250                                  HAMMER_ZONE_LARGE_DATA);
2251                         nbio->bio_offset = disk_offset;
2252                         error = hammer_io_direct_read(trans.hmp, nbio,
2253                                                       cursor.leaf);
2254                         goto done;
2255                 } else if (n) {
2256                         error = hammer_ip_resolve_data(&cursor);
2257                         if (error == 0) {
2258                                 bcopy((char *)cursor.data + roff,
2259                                       (char *)bp->b_data + boff, n);
2260                         }
2261                 }
2262                 if (error)
2263                         break;
2264
2265                 /*
2266                  * Iterate until we have filled the request.
2267                  */
2268                 boff += n;
2269                 if (boff == bp->b_bufsize)
2270                         break;
2271                 error = hammer_ip_next(&cursor);
2272         }
2273
2274         /*
2275          * There may have been a gap after the last record
2276          */
2277         if (error == ENOENT)
2278                 error = 0;
2279         if (error == 0 && boff != bp->b_bufsize) {
2280                 KKASSERT(boff < bp->b_bufsize);
2281                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2282                 /* boff = bp->b_bufsize; */
2283         }
2284         bp->b_resid = 0;
2285         bp->b_error = error;
2286         if (error)
2287                 bp->b_flags |= B_ERROR;
2288         biodone(ap->a_bio);
2289
2290 done:
2291         if (cursor.node)
2292                 hammer_cache_node(&ip->cache[1], cursor.node);
2293         hammer_done_cursor(&cursor);
2294         hammer_done_transaction(&trans);
2295         return(error);
2296 }
2297
2298 /*
2299  * BMAP operation - used to support cluster_read() only.
2300  *
2301  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2302  *
2303  * This routine may return EOPNOTSUPP if the opration is not supported for
2304  * the specified offset.  The contents of the pointer arguments do not
2305  * need to be initialized in that case.
2306  *
2307  * If a disk address is available and properly aligned return 0 with
2308  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2309  * to the run-length relative to that offset.  Callers may assume that
2310  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2311  * large, so return EOPNOTSUPP if it is not sufficiently large.
2312  */
2313 static
2314 int
2315 hammer_vop_bmap(struct vop_bmap_args *ap)
2316 {
2317         struct hammer_transaction trans;
2318         struct hammer_inode *ip;
2319         struct hammer_cursor cursor;
2320         hammer_base_elm_t base;
2321         int64_t rec_offset;
2322         int64_t ran_end;
2323         int64_t tmp64;
2324         int64_t base_offset;
2325         int64_t base_disk_offset;
2326         int64_t last_offset;
2327         hammer_off_t last_disk_offset;
2328         hammer_off_t disk_offset;
2329         int     rec_len;
2330         int     error;
2331         int     blksize;
2332
2333         ++hammer_stats_file_iopsr;
2334         ip = ap->a_vp->v_data;
2335
2336         /*
2337          * We can only BMAP regular files.  We can't BMAP database files,
2338          * directories, etc.
2339          */
2340         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2341                 return(EOPNOTSUPP);
2342
2343         /*
2344          * bmap is typically called with runp/runb both NULL when used
2345          * for writing.  We do not support BMAP for writing atm.
2346          */
2347         if (ap->a_cmd != BUF_CMD_READ)
2348                 return(EOPNOTSUPP);
2349
2350         /*
2351          * Scan the B-Tree to acquire blockmap addresses, then translate
2352          * to raw addresses.
2353          */
2354         hammer_simple_transaction(&trans, ip->hmp);
2355 #if 0
2356         kprintf("bmap_beg %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2357 #endif
2358         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2359
2360         /*
2361          * Key range (begin and end inclusive) to scan.  Note that the key's
2362          * stored in the actual records represent BASE+LEN, not BASE.  The
2363          * first record containing bio_offset will have a key > bio_offset.
2364          */
2365         cursor.key_beg.localization = ip->obj_localization +
2366                                       HAMMER_LOCALIZE_MISC;
2367         cursor.key_beg.obj_id = ip->obj_id;
2368         cursor.key_beg.create_tid = 0;
2369         cursor.key_beg.delete_tid = 0;
2370         cursor.key_beg.obj_type = 0;
2371         if (ap->a_runb)
2372                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2373         else
2374                 cursor.key_beg.key = ap->a_loffset + 1;
2375         if (cursor.key_beg.key < 0)
2376                 cursor.key_beg.key = 0;
2377         cursor.asof = ip->obj_asof;
2378         cursor.flags |= HAMMER_CURSOR_ASOF;
2379
2380         cursor.key_end = cursor.key_beg;
2381         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2382
2383         ran_end = ap->a_loffset + MAXPHYS;
2384         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2385         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2386         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2387         if (tmp64 < ran_end)
2388                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2389         else
2390                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2391
2392         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2393
2394         error = hammer_ip_first(&cursor);
2395         base_offset = last_offset = 0;
2396         base_disk_offset = last_disk_offset = 0;
2397
2398         while (error == 0) {
2399                 /*
2400                  * Get the base file offset of the record.  The key for
2401                  * data records is (base + bytes) rather then (base).
2402                  *
2403                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2404                  * The extra bytes should be zero on-disk and the BMAP op
2405                  * should still be ok.
2406                  */
2407                 base = &cursor.leaf->base;
2408                 rec_offset = base->key - cursor.leaf->data_len;
2409                 rec_len    = cursor.leaf->data_len;
2410
2411                 /*
2412                  * Incorporate any cached truncation.
2413                  *
2414                  * NOTE: Modifications to rec_len based on synthesized
2415                  * truncation points remove the guarantee that any extended
2416                  * data on disk is zero (since the truncations may not have
2417                  * taken place on-media yet).
2418                  */
2419                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2420                         if (hammer_cursor_ondisk(&cursor) ||
2421                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2422                                 if (ip->trunc_off <= rec_offset)
2423                                         rec_len = 0;
2424                                 else if (ip->trunc_off < rec_offset + rec_len)
2425                                         rec_len = (int)(ip->trunc_off - rec_offset);
2426                         }
2427                 }
2428                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2429                         if (hammer_cursor_ondisk(&cursor)) {
2430                                 if (ip->sync_trunc_off <= rec_offset)
2431                                         rec_len = 0;
2432                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2433                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2434                         }
2435                 }
2436
2437                 /*
2438                  * Accumulate information.  If we have hit a discontiguous
2439                  * block reset base_offset unless we are already beyond the
2440                  * requested offset.  If we are, that's it, we stop.
2441                  */
2442                 if (error)
2443                         break;
2444                 if (hammer_cursor_ondisk(&cursor)) {
2445                         disk_offset = cursor.leaf->data_offset;
2446                         if (rec_offset != last_offset ||
2447                             disk_offset != last_disk_offset) {
2448                                 if (rec_offset > ap->a_loffset)
2449                                         break;
2450                                 base_offset = rec_offset;
2451                                 base_disk_offset = disk_offset;
2452                         }
2453                         last_offset = rec_offset + rec_len;
2454                         last_disk_offset = disk_offset + rec_len;
2455                 }
2456                 error = hammer_ip_next(&cursor);
2457         }
2458
2459 #if 0
2460         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2461                 ap->a_loffset, base_offset, last_offset);
2462         kprintf("BMAP %16s:  %016llx - %016llx\n",
2463                 "", base_disk_offset, last_disk_offset);
2464 #endif
2465
2466         if (cursor.node) {
2467                 hammer_cache_node(&ip->cache[1], cursor.node);
2468 #if 0
2469                 kprintf("bmap_end2 %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2470 #endif
2471         }
2472         hammer_done_cursor(&cursor);
2473         hammer_done_transaction(&trans);
2474
2475         /*
2476          * If we couldn't find any records or the records we did find were
2477          * all behind the requested offset, return failure.  A forward
2478          * truncation can leave a hole w/ no on-disk records.
2479          */
2480         if (last_offset == 0 || last_offset < ap->a_loffset)
2481                 return (EOPNOTSUPP);
2482
2483         /*
2484          * Figure out the block size at the requested offset and adjust
2485          * our limits so the cluster_read() does not create inappropriately
2486          * sized buffer cache buffers.
2487          */
2488         blksize = hammer_blocksize(ap->a_loffset);
2489         if (hammer_blocksize(base_offset) != blksize) {
2490                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2491         }
2492         if (last_offset != ap->a_loffset &&
2493             hammer_blocksize(last_offset - 1) != blksize) {
2494                 last_offset = hammer_blockdemarc(ap->a_loffset,
2495                                                  last_offset - 1);
2496         }
2497
2498         /*
2499          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2500          * from occuring.
2501          */
2502         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2503
2504         if ((disk_offset & HAMMER_OFF_ZONE_MASK) != HAMMER_ZONE_LARGE_DATA) {
2505                 /*
2506                  * Only large-data zones can be direct-IOd
2507                  */
2508                 error = EOPNOTSUPP;
2509         } else if ((disk_offset & HAMMER_BUFMASK) ||
2510                    (last_offset - ap->a_loffset) < blksize) {
2511                 /*
2512                  * doffsetp is not aligned or the forward run size does
2513                  * not cover a whole buffer, disallow the direct I/O.
2514                  */
2515                 error = EOPNOTSUPP;
2516         } else {
2517                 /*
2518                  * We're good.
2519                  */
2520                 *ap->a_doffsetp = disk_offset;
2521                 if (ap->a_runb) {
2522                         *ap->a_runb = ap->a_loffset - base_offset;
2523                         KKASSERT(*ap->a_runb >= 0);
2524                 }
2525                 if (ap->a_runp) {
2526                         *ap->a_runp = last_offset - ap->a_loffset;
2527                         KKASSERT(*ap->a_runp >= 0);
2528                 }
2529                 error = 0;
2530         }
2531         return(error);
2532 }
2533
2534 /*
2535  * Write to a regular file.   Because this is a strategy call the OS is
2536  * trying to actually get data onto the media.
2537  */
2538 static
2539 int
2540 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2541 {
2542         hammer_record_t record;
2543         hammer_mount_t hmp;
2544         hammer_inode_t ip;
2545         struct bio *bio;
2546         struct buf *bp;
2547         int blksize;
2548         int bytes;
2549         int error;
2550
2551         bio = ap->a_bio;
2552         bp = bio->bio_buf;
2553         ip = ap->a_vp->v_data;
2554         hmp = ip->hmp;
2555
2556         blksize = hammer_blocksize(bio->bio_offset);
2557         KKASSERT(bp->b_bufsize == blksize);
2558
2559         if (ip->flags & HAMMER_INODE_RO) {
2560                 bp->b_error = EROFS;
2561                 bp->b_flags |= B_ERROR;
2562                 biodone(ap->a_bio);
2563                 return(EROFS);
2564         }
2565
2566         /*
2567          * Interlock with inode destruction (no in-kernel or directory
2568          * topology visibility).  If we queue new IO while trying to
2569          * destroy the inode we can deadlock the vtrunc call in
2570          * hammer_inode_unloadable_check().
2571          *
2572          * Besides, there's no point flushing a bp associated with an
2573          * inode that is being destroyed on-media and has no kernel
2574          * references.
2575          */
2576         if ((ip->flags | ip->sync_flags) &
2577             (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2578                 bp->b_resid = 0;
2579                 biodone(ap->a_bio);
2580                 return(0);
2581         }
2582
2583         /*
2584          * Reserve space and issue a direct-write from the front-end.
2585          * NOTE: The direct_io code will hammer_bread/bcopy smaller
2586          * allocations.
2587          *
2588          * An in-memory record will be installed to reference the storage
2589          * until the flusher can get to it.
2590          *
2591          * Since we own the high level bio the front-end will not try to
2592          * do a direct-read until the write completes.
2593          *
2594          * NOTE: The only time we do not reserve a full-sized buffers
2595          * worth of data is if the file is small.  We do not try to
2596          * allocate a fragment (from the small-data zone) at the end of
2597          * an otherwise large file as this can lead to wildly separated
2598          * data.
2599          */
2600         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2601         KKASSERT(bio->bio_offset < ip->ino_data.size);
2602         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2603                 bytes = bp->b_bufsize;
2604         else
2605                 bytes = ((int)ip->ino_data.size + 15) & ~15;
2606
2607         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2608                                     bytes, &error);
2609         if (record) {
2610                 hammer_io_direct_write(hmp, record, bio);
2611                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
2612                         hammer_flush_inode(ip, 0);
2613         } else {
2614                 bp->b_bio2.bio_offset = NOOFFSET;
2615                 bp->b_error = error;
2616                 bp->b_flags |= B_ERROR;
2617                 biodone(ap->a_bio);
2618         }
2619         return(error);
2620 }
2621
2622 /*
2623  * dounlink - disconnect a directory entry
2624  *
2625  * XXX whiteout support not really in yet
2626  */
2627 static int
2628 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2629                 struct vnode *dvp, struct ucred *cred,
2630                 int flags, int isdir)
2631 {
2632         struct namecache *ncp;
2633         hammer_inode_t dip;
2634         hammer_inode_t ip;
2635         struct hammer_cursor cursor;
2636         int64_t namekey;
2637         int nlen, error;
2638
2639         /*
2640          * Calculate the namekey and setup the key range for the scan.  This
2641          * works kinda like a chained hash table where the lower 32 bits
2642          * of the namekey synthesize the chain.
2643          *
2644          * The key range is inclusive of both key_beg and key_end.
2645          */
2646         dip = VTOI(dvp);
2647         ncp = nch->ncp;
2648
2649         if (dip->flags & HAMMER_INODE_RO)
2650                 return (EROFS);
2651
2652         namekey = hammer_directory_namekey(ncp->nc_name, ncp->nc_nlen);
2653 retry:
2654         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
2655         cursor.key_beg.localization = dip->obj_localization +
2656                                       HAMMER_LOCALIZE_MISC;
2657         cursor.key_beg.obj_id = dip->obj_id;
2658         cursor.key_beg.key = namekey;
2659         cursor.key_beg.create_tid = 0;
2660         cursor.key_beg.delete_tid = 0;
2661         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2662         cursor.key_beg.obj_type = 0;
2663
2664         cursor.key_end = cursor.key_beg;
2665         cursor.key_end.key |= 0xFFFFFFFFULL;
2666         cursor.asof = dip->obj_asof;
2667         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2668
2669         /*
2670          * Scan all matching records (the chain), locate the one matching
2671          * the requested path component.  info->last_error contains the
2672          * error code on search termination and could be 0, ENOENT, or
2673          * something else.
2674          *
2675          * The hammer_ip_*() functions merge in-memory records with on-disk
2676          * records for the purposes of the search.
2677          */
2678         error = hammer_ip_first(&cursor);
2679
2680         while (error == 0) {
2681                 error = hammer_ip_resolve_data(&cursor);
2682                 if (error)
2683                         break;
2684                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2685                 KKASSERT(nlen > 0);
2686                 if (ncp->nc_nlen == nlen &&
2687                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2688                         break;
2689                 }
2690                 error = hammer_ip_next(&cursor);
2691         }
2692
2693         /*
2694          * If all is ok we have to get the inode so we can adjust nlinks.
2695          * To avoid a deadlock with the flusher we must release the inode
2696          * lock on the directory when acquiring the inode for the entry.
2697          *
2698          * If the target is a directory, it must be empty.
2699          */
2700         if (error == 0) {
2701                 hammer_unlock(&cursor.ip->lock);
2702                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
2703                                       dip->hmp->asof,
2704                                       cursor.data->entry.localization,
2705                                       0, &error);
2706                 hammer_lock_sh(&cursor.ip->lock);
2707                 if (error == ENOENT) {
2708                         kprintf("obj_id %016llx\n", cursor.data->entry.obj_id);
2709                         Debugger("ENOENT unlinking object that should exist");
2710                 }
2711
2712                 /*
2713                  * If isdir >= 0 we validate that the entry is or is not a
2714                  * directory.  If isdir < 0 we don't care.
2715                  */
2716                 if (error == 0 && isdir >= 0) {
2717                         if (isdir &&
2718                             ip->ino_data.obj_type != HAMMER_OBJTYPE_DIRECTORY) {
2719                                 error = ENOTDIR;
2720                         } else if (isdir == 0 &&
2721                             ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
2722                                 error = EISDIR;
2723                         }
2724                 }
2725
2726                 /*
2727                  * If we are trying to remove a directory the directory must
2728                  * be empty.
2729                  *
2730                  * WARNING: hammer_ip_check_directory_empty() may have to
2731                  * terminate the cursor to avoid a deadlock.  It is ok to
2732                  * call hammer_done_cursor() twice.
2733                  */
2734                 if (error == 0 && ip->ino_data.obj_type ==
2735                                   HAMMER_OBJTYPE_DIRECTORY) {
2736                         error = hammer_ip_check_directory_empty(trans, ip);
2737                 }
2738
2739                 /*
2740                  * Delete the directory entry.
2741                  *
2742                  * WARNING: hammer_ip_del_directory() may have to terminate
2743                  * the cursor to avoid a deadlock.  It is ok to call
2744                  * hammer_done_cursor() twice.
2745                  */
2746                 if (error == 0) {
2747                         error = hammer_ip_del_directory(trans, &cursor,
2748                                                         dip, ip);
2749                 }
2750                 hammer_done_cursor(&cursor);
2751                 if (error == 0) {
2752                         cache_setunresolved(nch);
2753                         cache_setvp(nch, NULL);
2754                         /* XXX locking */
2755                         if (ip->vp)
2756                                 cache_inval_vp(ip->vp, CINV_DESTROY);
2757                 }
2758                 if (ip)
2759                         hammer_rel_inode(ip, 0);
2760         } else {
2761                 hammer_done_cursor(&cursor);
2762         }
2763         if (error == EDEADLK)
2764                 goto retry;
2765
2766         return (error);
2767 }
2768
2769 /************************************************************************
2770  *                          FIFO AND SPECFS OPS                         *
2771  ************************************************************************
2772  *
2773  */
2774
2775 static int
2776 hammer_vop_fifoclose (struct vop_close_args *ap)
2777 {
2778         /* XXX update itimes */
2779         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
2780 }
2781
2782 static int
2783 hammer_vop_fiforead (struct vop_read_args *ap)
2784 {
2785         int error;
2786
2787         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2788         /* XXX update access time */
2789         return (error);
2790 }
2791
2792 static int
2793 hammer_vop_fifowrite (struct vop_write_args *ap)
2794 {
2795         int error;
2796
2797         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2798         /* XXX update access time */
2799         return (error);
2800 }
2801
2802 static int
2803 hammer_vop_specclose (struct vop_close_args *ap)
2804 {
2805         /* XXX update itimes */
2806         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2807 }
2808
2809 static int
2810 hammer_vop_specread (struct vop_read_args *ap)
2811 {
2812         /* XXX update access time */
2813         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2814 }
2815
2816 static int
2817 hammer_vop_specwrite (struct vop_write_args *ap)
2818 {
2819         /* XXX update last change time */
2820         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2821 }
2822