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 1.76 2008/06/23 07:31:14 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_pathconf(struct vop_pathconf_args *);
  70 static int hammer_vop_print(struct vop_print_args *);
  71 static int hammer_vop_readdir(struct vop_readdir_args *);
  72 static int hammer_vop_readlink(struct vop_readlink_args *);
  73 static int hammer_vop_nremove(struct vop_nremove_args *);
  74 static int hammer_vop_nrename(struct vop_nrename_args *);
  75 static int hammer_vop_nrmdir(struct vop_nrmdir_args *);
  76 static int hammer_vop_setattr(struct vop_setattr_args *);
  77 static int hammer_vop_strategy(struct vop_strategy_args *);
  78 static int hammer_vop_bmap(struct vop_bmap_args *ap);
  79 static int hammer_vop_nsymlink(struct vop_nsymlink_args *);
  80 static int hammer_vop_nwhiteout(struct vop_nwhiteout_args *);
  81 static int hammer_vop_ioctl(struct vop_ioctl_args *);
  82 static int hammer_vop_mountctl(struct vop_mountctl_args *);
  83
  84 static int hammer_vop_fifoclose (struct vop_close_args *);
  85 static int hammer_vop_fiforead (struct vop_read_args *);
  86 static int hammer_vop_fifowrite (struct vop_write_args *);
  87
  88 static int hammer_vop_specclose (struct vop_close_args *);
  89 static int hammer_vop_specread (struct vop_read_args *);
  90 static int hammer_vop_specwrite (struct vop_write_args *);
  91
  92 struct vop_ops hammer_vnode_vops = {
  93         .vop_default =          vop_defaultop,
  94         .vop_fsync =            hammer_vop_fsync,
  95         .vop_getpages =         vop_stdgetpages,
  96         .vop_putpages =         vop_stdputpages,
  97         .vop_read =             hammer_vop_read,
  98         .vop_write =            hammer_vop_write,
  99         .vop_access =           hammer_vop_access,
 100         .vop_advlock =          hammer_vop_advlock,
 101         .vop_close =            hammer_vop_close,
 102         .vop_ncreate =          hammer_vop_ncreate,
 103         .vop_getattr =          hammer_vop_getattr,
 104         .vop_inactive =         hammer_vop_inactive,
 105         .vop_reclaim =          hammer_vop_reclaim,
 106         .vop_nresolve =         hammer_vop_nresolve,
 107         .vop_nlookupdotdot =    hammer_vop_nlookupdotdot,
 108         .vop_nlink =            hammer_vop_nlink,
 109         .vop_nmkdir =           hammer_vop_nmkdir,
 110         .vop_nmknod =           hammer_vop_nmknod,
 111         .vop_open =             hammer_vop_open,
 112         .vop_pathconf =         hammer_vop_pathconf,
 113         .vop_print =            hammer_vop_print,
 114         .vop_readdir =          hammer_vop_readdir,
 115         .vop_readlink =         hammer_vop_readlink,
 116         .vop_nremove =          hammer_vop_nremove,
 117         .vop_nrename =          hammer_vop_nrename,
 118         .vop_nrmdir =           hammer_vop_nrmdir,
 119         .vop_setattr =          hammer_vop_setattr,
 120         .vop_bmap =             hammer_vop_bmap,
 121         .vop_strategy =         hammer_vop_strategy,
 122         .vop_nsymlink =         hammer_vop_nsymlink,
 123         .vop_nwhiteout =        hammer_vop_nwhiteout,
 124         .vop_ioctl =            hammer_vop_ioctl,
 125         .vop_mountctl =         hammer_vop_mountctl
 126 };
 127
 128 struct vop_ops hammer_spec_vops = {
 129         .vop_default =          spec_vnoperate,
 130         .vop_fsync =            hammer_vop_fsync,
 131         .vop_read =             hammer_vop_specread,
 132         .vop_write =            hammer_vop_specwrite,
 133         .vop_access =           hammer_vop_access,
 134         .vop_close =            hammer_vop_specclose,
 135         .vop_getattr =          hammer_vop_getattr,
 136         .vop_inactive =         hammer_vop_inactive,
 137         .vop_reclaim =          hammer_vop_reclaim,
 138         .vop_setattr =          hammer_vop_setattr
 139 };
 140
 141 struct vop_ops hammer_fifo_vops = {
 142         .vop_default =          fifo_vnoperate,
 143         .vop_fsync =            hammer_vop_fsync,
 144         .vop_read =             hammer_vop_fiforead,
 145         .vop_write =            hammer_vop_fifowrite,
 146         .vop_access =           hammer_vop_access,
 147         .vop_close =            hammer_vop_fifoclose,
 148         .vop_getattr =          hammer_vop_getattr,
 149         .vop_inactive =         hammer_vop_inactive,
 150         .vop_reclaim =          hammer_vop_reclaim,
 151         .vop_setattr =          hammer_vop_setattr
 152 };
 153
 154 #ifdef DEBUG_TRUNCATE
 155 struct hammer_inode *HammerTruncIp;
 156 #endif
 157
 158 static int hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
 159                            struct vnode *dvp, struct ucred *cred, int flags);
 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 static void hammer_cleanup_write_io(hammer_inode_t ip);
 163 static void hammer_update_rsv_databufs(hammer_inode_t ip);
 164
 165 #if 0
 166 static
 167 int
 168 hammer_vop_vnoperate(struct vop_generic_args *)
 169 {
 170         return (VOCALL(&hammer_vnode_vops, ap));
 171 }
 172 #endif
 173
 174 /*
 175  * hammer_vop_fsync { vp, waitfor }
 176  *
 177  * fsync() an inode to disk and wait for it to be completely committed
 178  * such that the information would not be undone if a crash occured after
 179  * return.
 180  */
 181 static
 182 int
 183 hammer_vop_fsync(struct vop_fsync_args *ap)
 184 {
 185         hammer_inode_t ip = VTOI(ap->a_vp);
 186
 187         vfsync(ap->a_vp, ap->a_waitfor, 1, NULL, NULL);
 188         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 189         if (ap->a_waitfor == MNT_WAIT)
 190                 hammer_wait_inode(ip);
 191         return (ip->error);
 192 }
 193
 194 /*
 195  * hammer_vop_read { vp, uio, ioflag, cred }
 196  */
 197 static
 198 int
 199 hammer_vop_read(struct vop_read_args *ap)
 200 {
 201         struct hammer_transaction trans;
 202         hammer_inode_t ip;
 203         off_t offset;
 204         struct buf *bp;
 205         struct uio *uio;
 206         int error;
 207         int n;
 208         int seqcount;
 209         int ioseqcount;
 210         int blksize;
 211
 212         if (ap->a_vp->v_type != VREG)
 213                 return (EINVAL);
 214         ip = VTOI(ap->a_vp);
 215         error = 0;
 216         uio = ap->a_uio;
 217
 218         /*
 219          * Allow the UIO's size to override the sequential heuristic.
 220          */
 221         blksize = hammer_blocksize(uio->uio_offset);
 222         seqcount = (uio->uio_resid + (blksize - 1)) / blksize;
 223         ioseqcount = ap->a_ioflag >> 16;
 224         if (seqcount < ioseqcount)
 225                 seqcount = ioseqcount;
 226
 227         hammer_start_transaction(&trans, ip->hmp);
 228
 229         /*
 230          * Access the data typically in HAMMER_BUFSIZE blocks via the
 231          * buffer cache, but HAMMER may use a variable block size based
 232          * on the offset.
 233          */
 234         while (uio->uio_resid > 0 && uio->uio_offset < ip->ino_data.size) {
 235                 int64_t base_offset;
 236                 int64_t file_limit;
 237
 238                 blksize = hammer_blocksize(uio->uio_offset);
 239                 offset = (int)uio->uio_offset & (blksize - 1);
 240                 base_offset = uio->uio_offset - offset;
 241
 242                 if (hammer_debug_cluster_enable) {
 243                         /*
 244                          * Use file_limit to prevent cluster_read() from
 245                          * creating buffers of the wrong block size past
 246                          * the demarc.
 247                          */
 248                         file_limit = ip->ino_data.size;
 249                         if (base_offset < HAMMER_XDEMARC &&
 250                             file_limit > HAMMER_XDEMARC) {
 251                                 file_limit = HAMMER_XDEMARC;
 252                         }
 253                         error = cluster_read(ap->a_vp,
 254                                              file_limit, base_offset,
 255                                              blksize, MAXPHYS,
 256                                              seqcount, &bp);
 257                 } else {
 258                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 259                 }
 260                 if (error) {
 261                         kprintf("error %d\n", error);
 262                         brelse(bp);
 263                         break;
 264                 }
 265
 266                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 267                 n = blksize - offset;
 268                 if (n > uio->uio_resid)
 269                         n = uio->uio_resid;
 270                 if (n > ip->ino_data.size - uio->uio_offset)
 271                         n = (int)(ip->ino_data.size - uio->uio_offset);
 272                 error = uiomove((char *)bp->b_data + offset, n, uio);
 273
 274                 /* data has a lower priority then meta-data */
 275                 bp->b_flags |= B_AGE;
 276                 bqrelse(bp);
 277                 if (error)
 278                         break;
 279         }
 280         if ((ip->flags & HAMMER_INODE_RO) == 0 &&
 281             (ip->hmp->mp->mnt_flag & MNT_NOATIME) == 0) {
 282                 ip->ino_data.atime = trans.time;
 283                 hammer_modify_inode(ip, HAMMER_INODE_ATIME);
 284         }
 285         hammer_done_transaction(&trans);
 286         return (error);
 287 }
 288
 289 /*
 290  * hammer_vop_write { vp, uio, ioflag, cred }
 291  */
 292 static
 293 int
 294 hammer_vop_write(struct vop_write_args *ap)
 295 {
 296         struct hammer_transaction trans;
 297         struct hammer_inode *ip;
 298         hammer_mount_t hmp;
 299         struct uio *uio;
 300         int offset;
 301         off_t base_offset;
 302         struct buf *bp;
 303         int error;
 304         int n;
 305         int flags;
 306         int count;
 307         int delta;
 308         int seqcount;
 309
 310         if (ap->a_vp->v_type != VREG)
 311                 return (EINVAL);
 312         ip = VTOI(ap->a_vp);
 313         hmp = ip->hmp;
 314         error = 0;
 315         seqcount = ap->a_ioflag >> 16;
 316
 317         if (ip->flags & HAMMER_INODE_RO)
 318                 return (EROFS);
 319
 320         /*
 321          * Create a transaction to cover the operations we perform.
 322          */
 323         hammer_start_transaction(&trans, hmp);
 324         uio = ap->a_uio;
 325
 326         /*
 327          * Check append mode
 328          */
 329         if (ap->a_ioflag & IO_APPEND)
 330                 uio->uio_offset = ip->ino_data.size;
 331
 332         /*
 333          * Check for illegal write offsets.  Valid range is 0...2^63-1.
 334          *
 335          * NOTE: the base_off assignment is required to work around what
 336          * I consider to be a GCC-4 optimization bug.
 337          */
 338         if (uio->uio_offset < 0) {
 339                 hammer_done_transaction(&trans);
 340                 return (EFBIG);
 341         }
 342         base_offset = uio->uio_offset + uio->uio_resid; /* work around gcc-4 */
 343         if (uio->uio_resid > 0 && base_offset <= 0) {
 344                 hammer_done_transaction(&trans);
 345                 return (EFBIG);
 346         }
 347
 348         /*
 349          * Access the data typically in HAMMER_BUFSIZE blocks via the
 350          * buffer cache, but HAMMER may use a variable block size based
 351          * on the offset.
 352          */
 353         count = 0;
 354         while (uio->uio_resid > 0) {
 355                 int fixsize = 0;
 356                 int blksize;
 357                 int blkmask;
 358
 359                 if ((error = hammer_checkspace(hmp)) != 0)
 360                         break;
 361
 362                 /*
 363                  * Do not allow HAMMER to blow out the buffer cache.  Very
 364                  * large UIOs can lockout other processes due to bwillwrite()
 365                  * mechanics.
 366                  *
 367                  * Do not allow HAMMER to blow out system memory by
 368                  * accumulating too many records.   Records are so well
 369                  * decoupled from the buffer cache that it is possible
 370                  * for userland to push data out to the media via
 371                  * direct-write, but build up the records queued to the
 372                  * backend faster then the backend can flush them out.
 373                  * HAMMER has hit its write limit but the frontend has
 374                  * no pushback to slow it down.
 375                  *
 376                  * Always check at the beginning so separate writes are
 377                  * not able to bypass this code (count++).
 378                  *
 379                  * WARNING: Cannot unlock vp when doing a NOCOPY write as
 380                  * part of a putpages operation.  Doing so could cause us
 381                  * to deadlock against the VM system when we try to re-lock.
 382                  */
 383                 if ((count++ & 15) == 0 || count > 64) {
 384                         /*
 385                          * Buffer cache check
 386                          */
 387                         if (uio->uio_segflg != UIO_NOCOPY) {
 388                                 vn_unlock(ap->a_vp);
 389                                 if ((ap->a_ioflag & IO_NOBWILL) == 0)
 390                                         bwillwrite();
 391                         }
 392
 393                         /*
 394                          * Pending record flush check.
 395                          */
 396                         if (hmp->rsv_recs > hammer_limit_recs / 2) {
 397                                 /*
 398                                  * Get the inode on the flush list
 399                                  */
 400                                 if (ip->rsv_recs >= 64)
 401                                         hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
 402                                 else if (ip->rsv_recs >= 16)
 403                                         hammer_flush_inode(ip, 0);
 404
 405                                 /*
 406                                  * Keep the flusher going if the system keeps
 407                                  * queueing records.
 408                                  */
 409                                 delta = hmp->count_newrecords -
 410                                         hmp->last_newrecords;
 411                                 if (delta < 0 || delta > hammer_limit_recs / 2) {
 412                                         hmp->last_newrecords = hmp->count_newrecords;
 413                                         hammer_sync_hmp(hmp, MNT_NOWAIT);
 414                                 }
 415
 416                                 /*
 417                                  * If we have gotten behind start slowing
 418                                  * down the writers.
 419                                  */
 420                                 delta = (hmp->rsv_recs - hammer_limit_recs) *
 421                                         hz / hammer_limit_recs;
 422                                 if (delta > 0)
 423                                         tsleep(&trans, 0, "hmrslo", delta);
 424                         }
 425
 426                         if (uio->uio_segflg != UIO_NOCOPY)
 427                                 vn_lock(ap->a_vp, LK_EXCLUSIVE|LK_RETRY);
 428                 }
 429
 430                 /*
 431                  * Calculate the blocksize at the current offset and figure
 432                  * out how much we can actually write.
 433                  */
 434                 blksize = hammer_blocksize(uio->uio_offset);
 435                 blkmask = blksize - 1;
 436                 offset = (int)uio->uio_offset & blkmask;
 437                 base_offset = uio->uio_offset & ~(int64_t)blkmask;
 438                 n = blksize - offset;
 439                 if (n > uio->uio_resid)
 440                         n = uio->uio_resid;
 441                 if (uio->uio_offset + n > ip->ino_data.size) {
 442                         vnode_pager_setsize(ap->a_vp, uio->uio_offset + n);
 443                         fixsize = 1;
 444                 }
 445
 446                 if (uio->uio_segflg == UIO_NOCOPY) {
 447                         /*
 448                          * Issuing a write with the same data backing the
 449                          * buffer.  Instantiate the buffer to collect the
 450                          * backing vm pages, then read-in any missing bits.
 451                          *
 452                          * This case is used by vop_stdputpages().
 453                          */
 454                         bp = getblk(ap->a_vp, base_offset,
 455                                     blksize, GETBLK_BHEAVY, 0);
 456                         if ((bp->b_flags & B_CACHE) == 0) {
 457                                 bqrelse(bp);
 458                                 error = bread(ap->a_vp, base_offset,
 459                                               blksize, &bp);
 460                         }
 461                 } else if (offset == 0 && uio->uio_resid >= blksize) {
 462                         /*
 463                          * Even though we are entirely overwriting the buffer
 464                          * we may still have to zero it out to avoid a
 465                          * mmap/write visibility issue.
 466                          */
 467                         bp = getblk(ap->a_vp, base_offset, blksize, GETBLK_BHEAVY, 0);
 468                         if ((bp->b_flags & B_CACHE) == 0)
 469                                 vfs_bio_clrbuf(bp);
 470                 } else if (base_offset >= ip->ino_data.size) {
 471                         /*
 472                          * If the base offset of the buffer is beyond the
 473                          * file EOF, we don't have to issue a read.
 474                          */
 475                         bp = getblk(ap->a_vp, base_offset,
 476                                     blksize, GETBLK_BHEAVY, 0);
 477                         vfs_bio_clrbuf(bp);
 478                 } else {
 479                         /*
 480                          * Partial overwrite, read in any missing bits then
 481                          * replace the portion being written.
 482                          */
 483                         error = bread(ap->a_vp, base_offset, blksize, &bp);
 484                         if (error == 0)
 485                                 bheavy(bp);
 486                 }
 487                 if (error == 0) {
 488                         error = uiomove((char *)bp->b_data + offset,
 489                                         n, uio);
 490                 }
 491
 492                 /*
 493                  * If we screwed up we have to undo any VM size changes we
 494                  * made.
 495                  */
 496                 if (error) {
 497                         brelse(bp);
 498                         if (fixsize) {
 499                                 vtruncbuf(ap->a_vp, ip->ino_data.size,
 500                                           hammer_blocksize(ip->ino_data.size));
 501                         }
 502                         break;
 503                 }
 504                 /* bp->b_flags |= B_CLUSTEROK; temporarily disabled */
 505                 if (ip->ino_data.size < uio->uio_offset) {
 506                         ip->ino_data.size = uio->uio_offset;
 507                         flags = HAMMER_INODE_DDIRTY;
 508                         vnode_pager_setsize(ap->a_vp, ip->ino_data.size);
 509                 } else {
 510                         flags = 0;
 511                 }
 512                 ip->ino_data.mtime = trans.time;
 513                 flags |= HAMMER_INODE_MTIME | HAMMER_INODE_BUFS;
 514                 hammer_modify_inode(ip, flags);
 515
 516                 /*
 517                  * Try to keep track of cached dirty data.
 518                  */
 519                 if ((bp->b_flags & B_DIRTY) == 0) {
 520                         ++ip->rsv_databufs;
 521                         ++hmp->rsv_databufs;
 522                 }
 523
 524                 /*
 525                  * Final buffer disposition.
 526                  *
 527                  * If write_mode is non-zero we call bawrite()
 528                  * unconditionally.  Otherwise we only use bawrite()
 529                  * if the writes are clearly sequential.
 530                  */
 531                 bp->b_flags |= B_AGE;
 532                 if (ap->a_ioflag & IO_SYNC) {
 533                         bwrite(bp);
 534                 } else if (ap->a_ioflag & IO_DIRECT) {
 535                         bawrite(bp);
 536                 } else {
 537                         bdwrite(bp);
 538                 }
 539 #if 0
 540                 else if (hammer_write_mode &&
 541                            ((int)uio->uio_offset & blkmask) == 0) {
 542 #if 0
 543                         bp->b_flags |= B_CLUSTEROK;
 544                         cluster_write(bp, ip->ino_data.size, XXX seqcount);
 545 #else
 546                         bdwrite(bp);
 547 #endif
 548                 } else if ((ap->a_ioflag >> 16) == IO_SEQMAX &&
 549                            ((int)uio->uio_offset & blkmask) == 0) {
 550                         /*
 551                          * If seqcount indicates sequential operation and
 552                          * we just finished filling a buffer, push it out
 553                          * now to prevent the buffer cache from becoming
 554                          * too full, which would trigger non-optimal
 555                          * flushes.
 556                          */
 557                         bawrite(bp);
 558                 } else {
 559                         bdwrite(bp);
 560                 }
 561 #endif
 562         }
 563         hammer_done_transaction(&trans);
 564         return (error);
 565 }
 566
 567 /*
 568  * hammer_vop_access { vp, mode, cred }
 569  */
 570 static
 571 int
 572 hammer_vop_access(struct vop_access_args *ap)
 573 {
 574         struct hammer_inode *ip = VTOI(ap->a_vp);
 575         uid_t uid;
 576         gid_t gid;
 577         int error;
 578
 579         uid = hammer_to_unix_xid(&ip->ino_data.uid);
 580         gid = hammer_to_unix_xid(&ip->ino_data.gid);
 581
 582         error = vop_helper_access(ap, uid, gid, ip->ino_data.mode,
 583                                   ip->ino_data.uflags);
 584         return (error);
 585 }
 586
 587 /*
 588  * hammer_vop_advlock { vp, id, op, fl, flags }
 589  */
 590 static
 591 int
 592 hammer_vop_advlock(struct vop_advlock_args *ap)
 593 {
 594         hammer_inode_t ip = VTOI(ap->a_vp);
 595
 596         return (lf_advlock(ap, &ip->advlock, ip->ino_data.size));
 597 }
 598
 599 /*
 600  * hammer_vop_close { vp, fflag }
 601  */
 602 static
 603 int
 604 hammer_vop_close(struct vop_close_args *ap)
 605 {
 606         hammer_inode_t ip = VTOI(ap->a_vp);
 607
 608         if ((ip->flags | ip->sync_flags) & HAMMER_INODE_MODMASK)
 609                 hammer_inode_waitreclaims(ip->hmp);
 610         return (vop_stdclose(ap));
 611 }
 612
 613 /*
 614  * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
 615  *
 616  * The operating system has already ensured that the directory entry
 617  * does not exist and done all appropriate namespace locking.
 618  */
 619 static
 620 int
 621 hammer_vop_ncreate(struct vop_ncreate_args *ap)
 622 {
 623         struct hammer_transaction trans;
 624         struct hammer_inode *dip;
 625         struct hammer_inode *nip;
 626         struct nchandle *nch;
 627         int error;
 628
 629         nch = ap->a_nch;
 630         dip = VTOI(ap->a_dvp);
 631
 632         if (dip->flags & HAMMER_INODE_RO)
 633                 return (EROFS);
 634         if ((error = hammer_checkspace(dip->hmp)) != 0)
 635                 return (error);
 636
 637         /*
 638          * Create a transaction to cover the operations we perform.
 639          */
 640         hammer_start_transaction(&trans, dip->hmp);
 641
 642         /*
 643          * Create a new filesystem object of the requested type.  The
 644          * returned inode will be referenced and shared-locked to prevent
 645          * it from being moved to the flusher.
 646          */
 647
 648         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
 649                                     dip, 0, &nip);
 650         if (error) {
 651                 hkprintf("hammer_create_inode error %d\n", error);
 652                 hammer_done_transaction(&trans);
 653                 *ap->a_vpp = NULL;
 654                 return (error);
 655         }
 656
 657         /*
 658          * Add the new filesystem object to the directory.  This will also
 659          * bump the inode's link count.
 660          */
 661         error = hammer_ip_add_directory(&trans, dip,
 662                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
 663                                         nip);
 664         if (error)
 665                 hkprintf("hammer_ip_add_directory error %d\n", error);
 666
 667         /*
 668          * Finish up.
 669          */
 670         if (error) {
 671                 hammer_rel_inode(nip, 0);
 672                 hammer_done_transaction(&trans);
 673                 *ap->a_vpp = NULL;
 674         } else {
 675                 error = hammer_get_vnode(nip, ap->a_vpp);
 676                 hammer_done_transaction(&trans);
 677                 hammer_rel_inode(nip, 0);
 678                 if (error == 0) {
 679                         cache_setunresolved(ap->a_nch);
 680                         cache_setvp(ap->a_nch, *ap->a_vpp);
 681                 }
 682         }
 683         return (error);
 684 }
 685
 686 /*
 687  * hammer_vop_getattr { vp, vap }
 688  *
 689  * Retrieve an inode's attribute information.  When accessing inodes
 690  * historically we fake the atime field to ensure consistent results.
 691  * The atime field is stored in the B-Tree element and allowed to be
 692  * updated without cycling the element.
 693  */
 694 static
 695 int
 696 hammer_vop_getattr(struct vop_getattr_args *ap)
 697 {
 698         struct hammer_inode *ip = VTOI(ap->a_vp);
 699         struct vattr *vap = ap->a_vap;
 700
 701         vap->va_fsid = ip->hmp->fsid_udev;
 702         vap->va_fileid = ip->ino_leaf.base.obj_id;
 703         vap->va_mode = ip->ino_data.mode;
 704         vap->va_nlink = ip->ino_data.nlinks;
 705         vap->va_uid = hammer_to_unix_xid(&ip->ino_data.uid);
 706         vap->va_gid = hammer_to_unix_xid(&ip->ino_data.gid);
 707         vap->va_rmajor = 0;
 708         vap->va_rminor = 0;
 709         vap->va_size = ip->ino_data.size;
 710
 711         /*
 712          * We must provide a consistent atime and mtime for snapshots
 713          * so people can do a 'tar cf - ... | md5' on them and get
 714          * consistent results.
 715          */
 716         if (ip->flags & HAMMER_INODE_RO) {
 717                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_atime);
 718                 hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_mtime);
 719         } else {
 720                 hammer_time_to_timespec(ip->ino_data.atime, &vap->va_atime);
 721                 hammer_time_to_timespec(ip->ino_data.mtime, &vap->va_mtime);
 722         }
 723         hammer_time_to_timespec(ip->ino_data.ctime, &vap->va_ctime);
 724         vap->va_flags = ip->ino_data.uflags;
 725         vap->va_gen = 1;        /* hammer inums are unique for all time */
 726         vap->va_blocksize = HAMMER_BUFSIZE;
 727         if (ip->ino_data.size >= HAMMER_XDEMARC) {
 728                 vap->va_bytes = (ip->ino_data.size + HAMMER_XBUFMASK64) &
 729                                 ~HAMMER_XBUFMASK64;
 730         } else if (ip->ino_data.size > HAMMER_BUFSIZE / 2) {
 731                 vap->va_bytes = (ip->ino_data.size + HAMMER_BUFMASK64) &
 732                                 ~HAMMER_BUFMASK64;
 733         } else {
 734                 vap->va_bytes = (ip->ino_data.size + 15) & ~15;
 735         }
 736         vap->va_type = hammer_get_vnode_type(ip->ino_data.obj_type);
 737         vap->va_filerev = 0;    /* XXX */
 738         /* mtime uniquely identifies any adjustments made to the file XXX */
 739         vap->va_fsmid = ip->ino_data.mtime;
 740         vap->va_uid_uuid = ip->ino_data.uid;
 741         vap->va_gid_uuid = ip->ino_data.gid;
 742         vap->va_fsid_uuid = ip->hmp->fsid;
 743         vap->va_vaflags = VA_UID_UUID_VALID | VA_GID_UUID_VALID |
 744                           VA_FSID_UUID_VALID;
 745
 746         switch (ip->ino_data.obj_type) {
 747         case HAMMER_OBJTYPE_CDEV:
 748         case HAMMER_OBJTYPE_BDEV:
 749                 vap->va_rmajor = ip->ino_data.rmajor;
 750                 vap->va_rminor = ip->ino_data.rminor;
 751                 break;
 752         default:
 753                 break;
 754         }
 755
 756         return(0);
 757 }
 758
 759 /*
 760  * hammer_vop_nresolve { nch, dvp, cred }
 761  *
 762  * Locate the requested directory entry.
 763  */
 764 static
 765 int
 766 hammer_vop_nresolve(struct vop_nresolve_args *ap)
 767 {
 768         struct hammer_transaction trans;
 769         struct namecache *ncp;
 770         hammer_inode_t dip;
 771         hammer_inode_t ip;
 772         hammer_tid_t asof;
 773         struct hammer_cursor cursor;
 774         struct vnode *vp;
 775         int64_t namekey;
 776         int error;
 777         int i;
 778         int nlen;
 779         int flags;
 780         u_int64_t obj_id;
 781         u_int32_t localization;
 782
 783         /*
 784          * Misc initialization, plus handle as-of name extensions.  Look for
 785          * the '@@' extension.  Note that as-of files and directories cannot
 786          * be modified.
 787          */
 788         dip = VTOI(ap->a_dvp);
 789         ncp = ap->a_nch->ncp;
 790         asof = dip->obj_asof;
 791         nlen = ncp->nc_nlen;
 792         flags = dip->flags;
 793
 794         hammer_simple_transaction(&trans, dip->hmp);
 795
 796         for (i = 0; i < nlen; ++i) {
 797                 if (ncp->nc_name[i] == '@' && ncp->nc_name[i+1] == '@') {
 798                         asof = hammer_str_to_tid(ncp->nc_name + i + 2);
 799                         flags |= HAMMER_INODE_RO;
 800                         break;
 801                 }
 802         }
 803         nlen = i;
 804
 805         /*
 806          * If there is no path component the time extension is relative to
 807          * dip.
 808          */
 809         if (nlen == 0) {
 810                 ip = hammer_get_inode(&trans, dip, dip->obj_id,
 811                                       asof, dip->obj_localization,
 812                                       flags, &error);
 813                 if (error == 0) {
 814                         error = hammer_get_vnode(ip, &vp);
 815                         hammer_rel_inode(ip, 0);
 816                 } else {
 817                         vp = NULL;
 818                 }
 819                 if (error == 0) {
 820                         vn_unlock(vp);
 821                         cache_setvp(ap->a_nch, vp);
 822                         vrele(vp);
 823                 }
 824                 goto done;
 825         }
 826
 827         /*
 828          * Calculate the namekey and setup the key range for the scan.  This
 829          * works kinda like a chained hash table where the lower 32 bits
 830          * of the namekey synthesize the chain.
 831          *
 832          * The key range is inclusive of both key_beg and key_end.
 833          */
 834         namekey = hammer_directory_namekey(ncp->nc_name, nlen);
 835
 836         error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
 837         cursor.key_beg.localization = dip->obj_localization +
 838                                       HAMMER_LOCALIZE_MISC;
 839         cursor.key_beg.obj_id = dip->obj_id;
 840         cursor.key_beg.key = namekey;
 841         cursor.key_beg.create_tid = 0;
 842         cursor.key_beg.delete_tid = 0;
 843         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
 844         cursor.key_beg.obj_type = 0;
 845
 846         cursor.key_end = cursor.key_beg;
 847         cursor.key_end.key |= 0xFFFFFFFFULL;
 848         cursor.asof = asof;
 849         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
 850
 851         /*
 852          * Scan all matching records (the chain), locate the one matching
 853          * the requested path component.
 854          *
 855          * The hammer_ip_*() functions merge in-memory records with on-disk
 856          * records for the purposes of the search.
 857          */
 858         obj_id = 0;
 859         localization = HAMMER_DEF_LOCALIZATION;
 860
 861         if (error == 0) {
 862                 error = hammer_ip_first(&cursor);
 863                 while (error == 0) {
 864                         error = hammer_ip_resolve_data(&cursor);
 865                         if (error)
 866                                 break;
 867                         if (nlen == cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF &&
 868                             bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
 869                                 obj_id = cursor.data->entry.obj_id;
 870                                 localization = cursor.data->entry.localization;
 871                                 break;
 872                         }
 873                         error = hammer_ip_next(&cursor);
 874                 }
 875         }
 876         hammer_done_cursor(&cursor);
 877         if (error == 0) {
 878                 ip = hammer_get_inode(&trans, dip, obj_id,
 879                                       asof, localization,
 880                                       flags, &error);
 881                 if (error == 0) {
 882                         error = hammer_get_vnode(ip, &vp);
 883                         hammer_rel_inode(ip, 0);
 884                 } else {
 885                         vp = NULL;
 886                 }
 887                 if (error == 0) {
 888                         vn_unlock(vp);
 889                         cache_setvp(ap->a_nch, vp);
 890                         vrele(vp);
 891                 }
 892         } else if (error == ENOENT) {
 893                 cache_setvp(ap->a_nch, NULL);
 894         }
 895 done:
 896         hammer_done_transaction(&trans);
 897         return (error);
 898 }
 899
 900 /*
 901  * hammer_vop_nlookupdotdot { dvp, vpp, cred }
 902  *
 903  * Locate the parent directory of a directory vnode.
 904  *
 905  * dvp is referenced but not locked.  *vpp must be returned referenced and
 906  * locked.  A parent_obj_id of 0 does not necessarily indicate that we are
 907  * at the root, instead it could indicate that the directory we were in was
 908  * removed.
 909  *
 910  * NOTE: as-of sequences are not linked into the directory structure.  If
 911  * we are at the root with a different asof then the mount point, reload
 912  * the same directory with the mount point's asof.   I'm not sure what this
 913  * will do to NFS.  We encode ASOF stamps in NFS file handles so it might not
 914  * get confused, but it hasn't been tested.
 915  */
 916 static
 917 int
 918 hammer_vop_nlookupdotdot(struct vop_nlookupdotdot_args *ap)
 919 {
 920         struct hammer_transaction trans;
 921         struct hammer_inode *dip;
 922         struct hammer_inode *ip;
 923         int64_t parent_obj_id;
 924         u_int32_t parent_obj_localization;
 925         hammer_tid_t asof;
 926         int error;
 927
 928         dip = VTOI(ap->a_dvp);
 929         asof = dip->obj_asof;
 930
 931         /*
 932          * Whos are parent?  This could be the root of a pseudo-filesystem
 933          * whos parent is in another localization domain.
 934          */
 935         parent_obj_id = dip->ino_data.parent_obj_id;
 936         if (dip->obj_id == HAMMER_OBJID_ROOT)
 937                 parent_obj_localization = dip->ino_data.ext.obj.parent_obj_localization;
 938         else
 939                 parent_obj_localization = dip->obj_localization;
 940
 941         if (parent_obj_id == 0) {
 942                 if (dip->obj_id == HAMMER_OBJID_ROOT &&
 943                    asof != dip->hmp->asof) {
 944                         parent_obj_id = dip->obj_id;
 945                         asof = dip->hmp->asof;
 946                         *ap->a_fakename = kmalloc(19, M_TEMP, M_WAITOK);
 947                         ksnprintf(*ap->a_fakename, 19, "0x%016llx",
 948                                    dip->obj_asof);
 949                 } else {
 950                         *ap->a_vpp = NULL;
 951                         return ENOENT;
 952                 }
 953         }
 954
 955         hammer_simple_transaction(&trans, dip->hmp);
 956
 957         ip = hammer_get_inode(&trans, dip, parent_obj_id,
 958                               asof, parent_obj_localization,
 959                               dip->flags, &error);
 960         if (ip) {
 961                 error = hammer_get_vnode(ip, ap->a_vpp);
 962                 hammer_rel_inode(ip, 0);
 963         } else {
 964                 *ap->a_vpp = NULL;
 965         }
 966         hammer_done_transaction(&trans);
 967         return (error);
 968 }
 969
 970 /*
 971  * hammer_vop_nlink { nch, dvp, vp, cred }
 972  */
 973 static
 974 int
 975 hammer_vop_nlink(struct vop_nlink_args *ap)
 976 {
 977         struct hammer_transaction trans;
 978         struct hammer_inode *dip;
 979         struct hammer_inode *ip;
 980         struct nchandle *nch;
 981         int error;
 982
 983         nch = ap->a_nch;
 984         dip = VTOI(ap->a_dvp);
 985         ip = VTOI(ap->a_vp);
 986
 987         if (dip->flags & HAMMER_INODE_RO)
 988                 return (EROFS);
 989         if (ip->flags & HAMMER_INODE_RO)
 990                 return (EROFS);
 991         if ((error = hammer_checkspace(dip->hmp)) != 0)
 992                 return (error);
 993
 994         /*
 995          * Create a transaction to cover the operations we perform.
 996          */
 997         hammer_start_transaction(&trans, dip->hmp);
 998
 999         /*
1000          * Add the filesystem object to the directory.  Note that neither
1001          * dip nor ip are referenced or locked, but their vnodes are
1002          * referenced.  This function will bump the inode's link count.
1003          */
1004         error = hammer_ip_add_directory(&trans, dip,
1005                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1006                                         ip);
1007
1008         /*
1009          * Finish up.
1010          */
1011         if (error == 0) {
1012                 cache_setunresolved(nch);
1013                 cache_setvp(nch, ap->a_vp);
1014         }
1015         hammer_done_transaction(&trans);
1016         return (error);
1017 }
1018
1019 /*
1020  * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1021  *
1022  * The operating system has already ensured that the directory entry
1023  * does not exist and done all appropriate namespace locking.
1024  */
1025 static
1026 int
1027 hammer_vop_nmkdir(struct vop_nmkdir_args *ap)
1028 {
1029         struct hammer_transaction trans;
1030         struct hammer_inode *dip;
1031         struct hammer_inode *nip;
1032         struct nchandle *nch;
1033         int error;
1034
1035         nch = ap->a_nch;
1036         dip = VTOI(ap->a_dvp);
1037
1038         if (dip->flags & HAMMER_INODE_RO)
1039                 return (EROFS);
1040         if ((error = hammer_checkspace(dip->hmp)) != 0)
1041                 return (error);
1042
1043         /*
1044          * Create a transaction to cover the operations we perform.
1045          */
1046         hammer_start_transaction(&trans, dip->hmp);
1047
1048         /*
1049          * Create a new filesystem object of the requested type.  The
1050          * returned inode will be referenced but not locked.
1051          */
1052         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1053                                     dip, 0, &nip);
1054         if (error) {
1055                 hkprintf("hammer_mkdir error %d\n", error);
1056                 hammer_done_transaction(&trans);
1057                 *ap->a_vpp = NULL;
1058                 return (error);
1059         }
1060         /*
1061          * Add the new filesystem object to the directory.  This will also
1062          * bump the inode's link count.
1063          */
1064         error = hammer_ip_add_directory(&trans, dip,
1065                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1066                                         nip);
1067         if (error)
1068                 hkprintf("hammer_mkdir (add) error %d\n", error);
1069
1070         /*
1071          * Finish up.
1072          */
1073         if (error) {
1074                 hammer_rel_inode(nip, 0);
1075                 *ap->a_vpp = NULL;
1076         } else {
1077                 error = hammer_get_vnode(nip, ap->a_vpp);
1078                 hammer_rel_inode(nip, 0);
1079                 if (error == 0) {
1080                         cache_setunresolved(ap->a_nch);
1081                         cache_setvp(ap->a_nch, *ap->a_vpp);
1082                 }
1083         }
1084         hammer_done_transaction(&trans);
1085         return (error);
1086 }
1087
1088 /*
1089  * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1090  *
1091  * The operating system has already ensured that the directory entry
1092  * does not exist and done all appropriate namespace locking.
1093  */
1094 static
1095 int
1096 hammer_vop_nmknod(struct vop_nmknod_args *ap)
1097 {
1098         struct hammer_transaction trans;
1099         struct hammer_inode *dip;
1100         struct hammer_inode *nip;
1101         struct nchandle *nch;
1102         int error;
1103         int pseudofs;
1104
1105         nch = ap->a_nch;
1106         dip = VTOI(ap->a_dvp);
1107
1108         if (dip->flags & HAMMER_INODE_RO)
1109                 return (EROFS);
1110         if ((error = hammer_checkspace(dip->hmp)) != 0)
1111                 return (error);
1112
1113         /*
1114          * Create a transaction to cover the operations we perform.
1115          */
1116         hammer_start_transaction(&trans, dip->hmp);
1117
1118         /*
1119          * Create a new filesystem object of the requested type.  The
1120          * returned inode will be referenced but not locked.
1121          *
1122          * If mknod specifies a directory a pseudo-fs is created.
1123          */
1124         pseudofs = (ap->a_vap->va_type == VDIR);
1125         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1126                                     dip, pseudofs, &nip);
1127         if (error) {
1128                 hammer_done_transaction(&trans);
1129                 *ap->a_vpp = NULL;
1130                 return (error);
1131         }
1132
1133         /*
1134          * Add the new filesystem object to the directory.  This will also
1135          * bump the inode's link count.
1136          */
1137         error = hammer_ip_add_directory(&trans, dip,
1138                                         nch->ncp->nc_name, nch->ncp->nc_nlen,
1139                                         nip);
1140
1141         /*
1142          * Finish up.
1143          */
1144         if (error) {
1145                 hammer_rel_inode(nip, 0);
1146                 *ap->a_vpp = NULL;
1147         } else {
1148                 error = hammer_get_vnode(nip, ap->a_vpp);
1149                 hammer_rel_inode(nip, 0);
1150                 if (error == 0) {
1151                         cache_setunresolved(ap->a_nch);
1152                         cache_setvp(ap->a_nch, *ap->a_vpp);
1153                 }
1154         }
1155         hammer_done_transaction(&trans);
1156         return (error);
1157 }
1158
1159 /*
1160  * hammer_vop_open { vp, mode, cred, fp }
1161  */
1162 static
1163 int
1164 hammer_vop_open(struct vop_open_args *ap)
1165 {
1166         hammer_inode_t ip;
1167
1168         ip = VTOI(ap->a_vp);
1169
1170         if ((ap->a_mode & FWRITE) && (ip->flags & HAMMER_INODE_RO))
1171                 return (EROFS);
1172         return(vop_stdopen(ap));
1173 }
1174
1175 /*
1176  * hammer_vop_pathconf { vp, name, retval }
1177  */
1178 static
1179 int
1180 hammer_vop_pathconf(struct vop_pathconf_args *ap)
1181 {
1182         return EOPNOTSUPP;
1183 }
1184
1185 /*
1186  * hammer_vop_print { vp }
1187  */
1188 static
1189 int
1190 hammer_vop_print(struct vop_print_args *ap)
1191 {
1192         return EOPNOTSUPP;
1193 }
1194
1195 /*
1196  * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1197  */
1198 static
1199 int
1200 hammer_vop_readdir(struct vop_readdir_args *ap)
1201 {
1202         struct hammer_transaction trans;
1203         struct hammer_cursor cursor;
1204         struct hammer_inode *ip;
1205         struct uio *uio;
1206         hammer_base_elm_t base;
1207         int error;
1208         int cookie_index;
1209         int ncookies;
1210         off_t *cookies;
1211         off_t saveoff;
1212         int r;
1213
1214         ip = VTOI(ap->a_vp);
1215         uio = ap->a_uio;
1216         saveoff = uio->uio_offset;
1217
1218         if (ap->a_ncookies) {
1219                 ncookies = uio->uio_resid / 16 + 1;
1220                 if (ncookies > 1024)
1221                         ncookies = 1024;
1222                 cookies = kmalloc(ncookies * sizeof(off_t), M_TEMP, M_WAITOK);
1223                 cookie_index = 0;
1224         } else {
1225                 ncookies = -1;
1226                 cookies = NULL;
1227                 cookie_index = 0;
1228         }
1229
1230         hammer_simple_transaction(&trans, ip->hmp);
1231
1232         /*
1233          * Handle artificial entries
1234          */
1235         error = 0;
1236         if (saveoff == 0) {
1237                 r = vop_write_dirent(&error, uio, ip->obj_id, DT_DIR, 1, ".");
1238                 if (r)
1239                         goto done;
1240                 if (cookies)
1241                         cookies[cookie_index] = saveoff;
1242                 ++saveoff;
1243                 ++cookie_index;
1244                 if (cookie_index == ncookies)
1245                         goto done;
1246         }
1247         if (saveoff == 1) {
1248                 if (ip->ino_data.parent_obj_id) {
1249                         r = vop_write_dirent(&error, uio,
1250                                              ip->ino_data.parent_obj_id,
1251                                              DT_DIR, 2, "..");
1252                 } else {
1253                         r = vop_write_dirent(&error, uio,
1254                                              ip->obj_id, DT_DIR, 2, "..");
1255                 }
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
1266         /*
1267          * Key range (begin and end inclusive) to scan.  Directory keys
1268          * directly translate to a 64 bit 'seek' position.
1269          */
1270         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1271         cursor.key_beg.localization = ip->obj_localization +
1272                                       HAMMER_LOCALIZE_MISC;
1273         cursor.key_beg.obj_id = ip->obj_id;
1274         cursor.key_beg.create_tid = 0;
1275         cursor.key_beg.delete_tid = 0;
1276         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1277         cursor.key_beg.obj_type = 0;
1278         cursor.key_beg.key = saveoff;
1279
1280         cursor.key_end = cursor.key_beg;
1281         cursor.key_end.key = HAMMER_MAX_KEY;
1282         cursor.asof = ip->obj_asof;
1283         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1284
1285         error = hammer_ip_first(&cursor);
1286
1287         while (error == 0) {
1288                 error = hammer_ip_resolve_data(&cursor);
1289                 if (error)
1290                         break;
1291                 base = &cursor.leaf->base;
1292                 saveoff = base->key;
1293                 KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
1294
1295                 if (base->obj_id != ip->obj_id)
1296                         panic("readdir: bad record at %p", cursor.node);
1297
1298                 r = vop_write_dirent(
1299                              &error, uio, cursor.data->entry.obj_id,
1300                              hammer_get_dtype(cursor.leaf->base.obj_type),
1301                              cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF ,
1302                              (void *)cursor.data->entry.name);
1303                 if (r)
1304                         break;
1305                 ++saveoff;
1306                 if (cookies)
1307                         cookies[cookie_index] = base->key;
1308                 ++cookie_index;
1309                 if (cookie_index == ncookies)
1310                         break;
1311                 error = hammer_ip_next(&cursor);
1312         }
1313         hammer_done_cursor(&cursor);
1314
1315 done:
1316         hammer_done_transaction(&trans);
1317
1318         if (ap->a_eofflag)
1319                 *ap->a_eofflag = (error == ENOENT);
1320         uio->uio_offset = saveoff;
1321         if (error && cookie_index == 0) {
1322                 if (error == ENOENT)
1323                         error = 0;
1324                 if (cookies) {
1325                         kfree(cookies, M_TEMP);
1326                         *ap->a_ncookies = 0;
1327                         *ap->a_cookies = NULL;
1328                 }
1329         } else {
1330                 if (error == ENOENT)
1331                         error = 0;
1332                 if (cookies) {
1333                         *ap->a_ncookies = cookie_index;
1334                         *ap->a_cookies = cookies;
1335                 }
1336         }
1337         return(error);
1338 }
1339
1340 /*
1341  * hammer_vop_readlink { vp, uio, cred }
1342  */
1343 static
1344 int
1345 hammer_vop_readlink(struct vop_readlink_args *ap)
1346 {
1347         struct hammer_transaction trans;
1348         struct hammer_cursor cursor;
1349         struct hammer_inode *ip;
1350         int error;
1351
1352         ip = VTOI(ap->a_vp);
1353
1354         /*
1355          * Shortcut if the symlink data was stuffed into ino_data.
1356          */
1357         if (ip->ino_data.size <= HAMMER_INODE_BASESYMLEN) {
1358                 error = uiomove(ip->ino_data.ext.symlink,
1359                                 ip->ino_data.size, ap->a_uio);
1360                 return(error);
1361         }
1362
1363         /*
1364          * Long version
1365          */
1366         hammer_simple_transaction(&trans, ip->hmp);
1367         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
1368
1369         /*
1370          * Key range (begin and end inclusive) to scan.  Directory keys
1371          * directly translate to a 64 bit 'seek' position.
1372          */
1373         cursor.key_beg.localization = ip->obj_localization +
1374                                       HAMMER_LOCALIZE_MISC;
1375         cursor.key_beg.obj_id = ip->obj_id;
1376         cursor.key_beg.create_tid = 0;
1377         cursor.key_beg.delete_tid = 0;
1378         cursor.key_beg.rec_type = HAMMER_RECTYPE_FIX;
1379         cursor.key_beg.obj_type = 0;
1380         cursor.key_beg.key = HAMMER_FIXKEY_SYMLINK;
1381         cursor.asof = ip->obj_asof;
1382         cursor.flags |= HAMMER_CURSOR_ASOF;
1383
1384         error = hammer_ip_lookup(&cursor);
1385         if (error == 0) {
1386                 error = hammer_ip_resolve_data(&cursor);
1387                 if (error == 0) {
1388                         KKASSERT(cursor.leaf->data_len >=
1389                                  HAMMER_SYMLINK_NAME_OFF);
1390                         error = uiomove(cursor.data->symlink.name,
1391                                         cursor.leaf->data_len -
1392                                                 HAMMER_SYMLINK_NAME_OFF,
1393                                         ap->a_uio);
1394                 }
1395         }
1396         hammer_done_cursor(&cursor);
1397         hammer_done_transaction(&trans);
1398         return(error);
1399 }
1400
1401 /*
1402  * hammer_vop_nremove { nch, dvp, cred }
1403  */
1404 static
1405 int
1406 hammer_vop_nremove(struct vop_nremove_args *ap)
1407 {
1408         struct hammer_transaction trans;
1409         struct hammer_inode *dip;
1410         int error;
1411
1412         dip = VTOI(ap->a_dvp);
1413
1414         if (hammer_nohistory(dip) == 0 &&
1415             (error = hammer_checkspace(dip->hmp)) != 0) {
1416                 return (error);
1417         }
1418
1419         hammer_start_transaction(&trans, dip->hmp);
1420         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1421         hammer_done_transaction(&trans);
1422
1423         return (error);
1424 }
1425
1426 /*
1427  * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1428  */
1429 static
1430 int
1431 hammer_vop_nrename(struct vop_nrename_args *ap)
1432 {
1433         struct hammer_transaction trans;
1434         struct namecache *fncp;
1435         struct namecache *tncp;
1436         struct hammer_inode *fdip;
1437         struct hammer_inode *tdip;
1438         struct hammer_inode *ip;
1439         struct hammer_cursor cursor;
1440         int64_t namekey;
1441         int nlen, error;
1442
1443         fdip = VTOI(ap->a_fdvp);
1444         tdip = VTOI(ap->a_tdvp);
1445         fncp = ap->a_fnch->ncp;
1446         tncp = ap->a_tnch->ncp;
1447         ip = VTOI(fncp->nc_vp);
1448         KKASSERT(ip != NULL);
1449
1450         if (fdip->flags & HAMMER_INODE_RO)
1451                 return (EROFS);
1452         if (tdip->flags & HAMMER_INODE_RO)
1453                 return (EROFS);
1454         if (ip->flags & HAMMER_INODE_RO)
1455                 return (EROFS);
1456         if ((error = hammer_checkspace(fdip->hmp)) != 0)
1457                 return (error);
1458
1459         hammer_start_transaction(&trans, fdip->hmp);
1460
1461         /*
1462          * Remove tncp from the target directory and then link ip as
1463          * tncp. XXX pass trans to dounlink
1464          *
1465          * Force the inode sync-time to match the transaction so it is
1466          * in-sync with the creation of the target directory entry.
1467          */
1468         error = hammer_dounlink(&trans, ap->a_tnch, ap->a_tdvp, ap->a_cred, 0);
1469         if (error == 0 || error == ENOENT) {
1470                 error = hammer_ip_add_directory(&trans, tdip,
1471                                                 tncp->nc_name, tncp->nc_nlen,
1472                                                 ip);
1473                 if (error == 0) {
1474                         ip->ino_data.parent_obj_id = tdip->obj_id;
1475                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
1476                 }
1477         }
1478         if (error)
1479                 goto failed; /* XXX */
1480
1481         /*
1482          * Locate the record in the originating directory and remove it.
1483          *
1484          * Calculate the namekey and setup the key range for the scan.  This
1485          * works kinda like a chained hash table where the lower 32 bits
1486          * of the namekey synthesize the chain.
1487          *
1488          * The key range is inclusive of both key_beg and key_end.
1489          */
1490         namekey = hammer_directory_namekey(fncp->nc_name, fncp->nc_nlen);
1491 retry:
1492         hammer_init_cursor(&trans, &cursor, &fdip->cache[1], fdip);
1493         cursor.key_beg.localization = fdip->obj_localization +
1494                                       HAMMER_LOCALIZE_MISC;
1495         cursor.key_beg.obj_id = fdip->obj_id;
1496         cursor.key_beg.key = namekey;
1497         cursor.key_beg.create_tid = 0;
1498         cursor.key_beg.delete_tid = 0;
1499         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
1500         cursor.key_beg.obj_type = 0;
1501
1502         cursor.key_end = cursor.key_beg;
1503         cursor.key_end.key |= 0xFFFFFFFFULL;
1504         cursor.asof = fdip->obj_asof;
1505         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1506
1507         /*
1508          * Scan all matching records (the chain), locate the one matching
1509          * the requested path component.
1510          *
1511          * The hammer_ip_*() functions merge in-memory records with on-disk
1512          * records for the purposes of the search.
1513          */
1514         error = hammer_ip_first(&cursor);
1515         while (error == 0) {
1516                 if (hammer_ip_resolve_data(&cursor) != 0)
1517                         break;
1518                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
1519                 KKASSERT(nlen > 0);
1520                 if (fncp->nc_nlen == nlen &&
1521                     bcmp(fncp->nc_name, cursor.data->entry.name, nlen) == 0) {
1522                         break;
1523                 }
1524                 error = hammer_ip_next(&cursor);
1525         }
1526
1527         /*
1528          * If all is ok we have to get the inode so we can adjust nlinks.
1529          *
1530          * WARNING: hammer_ip_del_directory() may have to terminate the
1531          * cursor to avoid a recursion.  It's ok to call hammer_done_cursor()
1532          * twice.
1533          */
1534         if (error == 0)
1535                 error = hammer_ip_del_directory(&trans, &cursor, fdip, ip);
1536
1537         /*
1538          * XXX A deadlock here will break rename's atomicy for the purposes
1539          * of crash recovery.
1540          */
1541         if (error == EDEADLK) {
1542                 hammer_done_cursor(&cursor);
1543                 goto retry;
1544         }
1545
1546         /*
1547          * Cleanup and tell the kernel that the rename succeeded.
1548          */
1549         hammer_done_cursor(&cursor);
1550         if (error == 0)
1551                 cache_rename(ap->a_fnch, ap->a_tnch);
1552
1553 failed:
1554         hammer_done_transaction(&trans);
1555         return (error);
1556 }
1557
1558 /*
1559  * hammer_vop_nrmdir { nch, dvp, cred }
1560  */
1561 static
1562 int
1563 hammer_vop_nrmdir(struct vop_nrmdir_args *ap)
1564 {
1565         struct hammer_transaction trans;
1566         struct hammer_inode *dip;
1567         int error;
1568
1569         dip = VTOI(ap->a_dvp);
1570
1571         if (hammer_nohistory(dip) == 0 &&
1572             (error = hammer_checkspace(dip->hmp)) != 0) {
1573                 return (error);
1574         }
1575
1576         hammer_start_transaction(&trans, dip->hmp);
1577         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp, ap->a_cred, 0);
1578         hammer_done_transaction(&trans);
1579
1580         return (error);
1581 }
1582
1583 /*
1584  * hammer_vop_setattr { vp, vap, cred }
1585  */
1586 static
1587 int
1588 hammer_vop_setattr(struct vop_setattr_args *ap)
1589 {
1590         struct hammer_transaction trans;
1591         struct vattr *vap;
1592         struct hammer_inode *ip;
1593         int modflags;
1594         int error;
1595         int truncating;
1596         int blksize;
1597         int64_t aligned_size;
1598         u_int32_t flags;
1599
1600         vap = ap->a_vap;
1601         ip = ap->a_vp->v_data;
1602         modflags = 0;
1603
1604         if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)
1605                 return(EROFS);
1606         if (ip->flags & HAMMER_INODE_RO)
1607                 return (EROFS);
1608         if (hammer_nohistory(ip) == 0 &&
1609             (error = hammer_checkspace(ip->hmp)) != 0) {
1610                 return (error);
1611         }
1612
1613         hammer_start_transaction(&trans, ip->hmp);
1614         error = 0;
1615
1616         if (vap->va_flags != VNOVAL) {
1617                 flags = ip->ino_data.uflags;
1618                 error = vop_helper_setattr_flags(&flags, vap->va_flags,
1619                                          hammer_to_unix_xid(&ip->ino_data.uid),
1620                                          ap->a_cred);
1621                 if (error == 0) {
1622                         if (ip->ino_data.uflags != flags) {
1623                                 ip->ino_data.uflags = flags;
1624                                 modflags |= HAMMER_INODE_DDIRTY;
1625                         }
1626                         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1627                                 error = 0;
1628                                 goto done;
1629                         }
1630                 }
1631                 goto done;
1632         }
1633         if (ip->ino_data.uflags & (IMMUTABLE | APPEND)) {
1634                 error = EPERM;
1635                 goto done;
1636         }
1637         if (vap->va_uid != (uid_t)VNOVAL || vap->va_gid != (gid_t)VNOVAL) {
1638                 mode_t cur_mode = ip->ino_data.mode;
1639                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1640                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1641                 uuid_t uuid_uid;
1642                 uuid_t uuid_gid;
1643
1644                 error = vop_helper_chown(ap->a_vp, vap->va_uid, vap->va_gid,
1645                                          ap->a_cred,
1646                                          &cur_uid, &cur_gid, &cur_mode);
1647                 if (error == 0) {
1648                         hammer_guid_to_uuid(&uuid_uid, cur_uid);
1649                         hammer_guid_to_uuid(&uuid_gid, cur_gid);
1650                         if (bcmp(&uuid_uid, &ip->ino_data.uid,
1651                                  sizeof(uuid_uid)) ||
1652                             bcmp(&uuid_gid, &ip->ino_data.gid,
1653                                  sizeof(uuid_gid)) ||
1654                             ip->ino_data.mode != cur_mode
1655                         ) {
1656                                 ip->ino_data.uid = uuid_uid;
1657                                 ip->ino_data.gid = uuid_gid;
1658                                 ip->ino_data.mode = cur_mode;
1659                         }
1660                         modflags |= HAMMER_INODE_DDIRTY;
1661                 }
1662         }
1663         while (vap->va_size != VNOVAL && ip->ino_data.size != vap->va_size) {
1664                 switch(ap->a_vp->v_type) {
1665                 case VREG:
1666                         if (vap->va_size == ip->ino_data.size)
1667                                 break;
1668                         /*
1669                          * XXX break atomicy, we can deadlock the backend
1670                          * if we do not release the lock.  Probably not a
1671                          * big deal here.
1672                          */
1673                         blksize = hammer_blocksize(vap->va_size);
1674                         if (vap->va_size < ip->ino_data.size) {
1675                                 vtruncbuf(ap->a_vp, vap->va_size, blksize);
1676                                 truncating = 1;
1677                         } else {
1678                                 vnode_pager_setsize(ap->a_vp, vap->va_size);
1679                                 truncating = 0;
1680                         }
1681                         ip->ino_data.size = vap->va_size;
1682                         modflags |= HAMMER_INODE_DDIRTY;
1683
1684                         /*
1685                          * on-media truncation is cached in the inode until
1686                          * the inode is synchronized.
1687                          */
1688                         if (truncating) {
1689                                 hammer_ip_frontend_trunc(ip, vap->va_size);
1690                                 hammer_update_rsv_databufs(ip);
1691 #ifdef DEBUG_TRUNCATE
1692                                 if (HammerTruncIp == NULL)
1693                                         HammerTruncIp = ip;
1694 #endif
1695                                 if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1696                                         ip->flags |= HAMMER_INODE_TRUNCATED;
1697                                         ip->trunc_off = vap->va_size;
1698 #ifdef DEBUG_TRUNCATE
1699                                         if (ip == HammerTruncIp)
1700                                         kprintf("truncate1 %016llx\n", ip->trunc_off);
1701 #endif
1702                                 } else if (ip->trunc_off > vap->va_size) {
1703                                         ip->trunc_off = vap->va_size;
1704 #ifdef DEBUG_TRUNCATE
1705                                         if (ip == HammerTruncIp)
1706                                         kprintf("truncate2 %016llx\n", ip->trunc_off);
1707 #endif
1708                                 } else {
1709 #ifdef DEBUG_TRUNCATE
1710                                         if (ip == HammerTruncIp)
1711                                         kprintf("truncate3 %016llx (ignored)\n", vap->va_size);
1712 #endif
1713                                 }
1714                         }
1715
1716                         /*
1717                          * If truncating we have to clean out a portion of
1718                          * the last block on-disk.  We do this in the
1719                          * front-end buffer cache.
1720                          */
1721                         aligned_size = (vap->va_size + (blksize - 1)) &
1722                                        ~(int64_t)(blksize - 1);
1723                         if (truncating && vap->va_size < aligned_size) {
1724                                 struct buf *bp;
1725                                 int offset;
1726
1727                                 aligned_size -= blksize;
1728
1729                                 offset = (int)vap->va_size & (blksize - 1);
1730                                 error = bread(ap->a_vp, aligned_size,
1731                                               blksize, &bp);
1732                                 hammer_ip_frontend_trunc(ip, aligned_size);
1733                                 if (error == 0) {
1734                                         bzero(bp->b_data + offset,
1735                                               blksize - offset);
1736                                         bdwrite(bp);
1737                                 } else {
1738                                         kprintf("ERROR %d\n", error);
1739                                         brelse(bp);
1740                                 }
1741                         }
1742                         break;
1743                 case VDATABASE:
1744                         if ((ip->flags & HAMMER_INODE_TRUNCATED) == 0) {
1745                                 ip->flags |= HAMMER_INODE_TRUNCATED;
1746                                 ip->trunc_off = vap->va_size;
1747                         } else if (ip->trunc_off > vap->va_size) {
1748                                 ip->trunc_off = vap->va_size;
1749                         }
1750                         hammer_ip_frontend_trunc(ip, vap->va_size);
1751                         ip->ino_data.size = vap->va_size;
1752                         modflags |= HAMMER_INODE_DDIRTY;
1753                         break;
1754                 default:
1755                         error = EINVAL;
1756                         goto done;
1757                 }
1758                 break;
1759         }
1760         if (vap->va_atime.tv_sec != VNOVAL) {
1761                 ip->ino_data.atime =
1762                         hammer_timespec_to_time(&vap->va_atime);
1763                 modflags |= HAMMER_INODE_ATIME;
1764         }
1765         if (vap->va_mtime.tv_sec != VNOVAL) {
1766                 ip->ino_data.mtime =
1767                         hammer_timespec_to_time(&vap->va_mtime);
1768                 modflags |= HAMMER_INODE_MTIME;
1769         }
1770         if (vap->va_mode != (mode_t)VNOVAL) {
1771                 mode_t   cur_mode = ip->ino_data.mode;
1772                 uid_t cur_uid = hammer_to_unix_xid(&ip->ino_data.uid);
1773                 gid_t cur_gid = hammer_to_unix_xid(&ip->ino_data.gid);
1774
1775                 error = vop_helper_chmod(ap->a_vp, vap->va_mode, ap->a_cred,
1776                                          cur_uid, cur_gid, &cur_mode);
1777                 if (error == 0 && ip->ino_data.mode != cur_mode) {
1778                         ip->ino_data.mode = cur_mode;
1779                         modflags |= HAMMER_INODE_DDIRTY;
1780                 }
1781         }
1782 done:
1783         if (error == 0)
1784                 hammer_modify_inode(ip, modflags);
1785         hammer_done_transaction(&trans);
1786         return (error);
1787 }
1788
1789 /*
1790  * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
1791  */
1792 static
1793 int
1794 hammer_vop_nsymlink(struct vop_nsymlink_args *ap)
1795 {
1796         struct hammer_transaction trans;
1797         struct hammer_inode *dip;
1798         struct hammer_inode *nip;
1799         struct nchandle *nch;
1800         hammer_record_t record;
1801         int error;
1802         int bytes;
1803
1804         ap->a_vap->va_type = VLNK;
1805
1806         nch = ap->a_nch;
1807         dip = VTOI(ap->a_dvp);
1808
1809         if (dip->flags & HAMMER_INODE_RO)
1810                 return (EROFS);
1811         if ((error = hammer_checkspace(dip->hmp)) != 0)
1812                 return (error);
1813
1814         /*
1815          * Create a transaction to cover the operations we perform.
1816          */
1817         hammer_start_transaction(&trans, dip->hmp);
1818
1819         /*
1820          * Create a new filesystem object of the requested type.  The
1821          * returned inode will be referenced but not locked.
1822          */
1823
1824         error = hammer_create_inode(&trans, ap->a_vap, ap->a_cred,
1825                                     dip, 0, &nip);
1826         if (error) {
1827                 hammer_done_transaction(&trans);
1828                 *ap->a_vpp = NULL;
1829                 return (error);
1830         }
1831
1832         /*
1833          * Add a record representing the symlink.  symlink stores the link
1834          * as pure data, not a string, and is no \0 terminated.
1835          */
1836         if (error == 0) {
1837                 bytes = strlen(ap->a_target);
1838
1839                 if (bytes <= HAMMER_INODE_BASESYMLEN) {
1840                         bcopy(ap->a_target, nip->ino_data.ext.symlink, bytes);
1841                 } else {
1842                         record = hammer_alloc_mem_record(nip, bytes);
1843                         record->type = HAMMER_MEM_RECORD_GENERAL;
1844
1845                         record->leaf.base.localization = nip->obj_localization +
1846                                                          HAMMER_LOCALIZE_MISC;
1847                         record->leaf.base.key = HAMMER_FIXKEY_SYMLINK;
1848                         record->leaf.base.rec_type = HAMMER_RECTYPE_FIX;
1849                         record->leaf.data_len = bytes;
1850                         KKASSERT(HAMMER_SYMLINK_NAME_OFF == 0);
1851                         bcopy(ap->a_target, record->data->symlink.name, bytes);
1852                         error = hammer_ip_add_record(&trans, record);
1853                 }
1854
1855                 /*
1856                  * Set the file size to the length of the link.
1857                  */
1858                 if (error == 0) {
1859                         nip->ino_data.size = bytes;
1860                         hammer_modify_inode(nip, HAMMER_INODE_DDIRTY);
1861                 }
1862         }
1863         if (error == 0)
1864                 error = hammer_ip_add_directory(&trans, dip, nch->ncp->nc_name,
1865                                                 nch->ncp->nc_nlen, nip);
1866
1867         /*
1868          * Finish up.
1869          */
1870         if (error) {
1871                 hammer_rel_inode(nip, 0);
1872                 *ap->a_vpp = NULL;
1873         } else {
1874                 error = hammer_get_vnode(nip, ap->a_vpp);
1875                 hammer_rel_inode(nip, 0);
1876                 if (error == 0) {
1877                         cache_setunresolved(ap->a_nch);
1878                         cache_setvp(ap->a_nch, *ap->a_vpp);
1879                 }
1880         }
1881         hammer_done_transaction(&trans);
1882         return (error);
1883 }
1884
1885 /*
1886  * hammer_vop_nwhiteout { nch, dvp, cred, flags }
1887  */
1888 static
1889 int
1890 hammer_vop_nwhiteout(struct vop_nwhiteout_args *ap)
1891 {
1892         struct hammer_transaction trans;
1893         struct hammer_inode *dip;
1894         int error;
1895
1896         dip = VTOI(ap->a_dvp);
1897
1898         if (hammer_nohistory(dip) == 0 &&
1899             (error = hammer_checkspace(dip->hmp)) != 0) {
1900                 return (error);
1901         }
1902
1903         hammer_start_transaction(&trans, dip->hmp);
1904         error = hammer_dounlink(&trans, ap->a_nch, ap->a_dvp,
1905                                 ap->a_cred, ap->a_flags);
1906         hammer_done_transaction(&trans);
1907
1908         return (error);
1909 }
1910
1911 /*
1912  * hammer_vop_ioctl { vp, command, data, fflag, cred }
1913  */
1914 static
1915 int
1916 hammer_vop_ioctl(struct vop_ioctl_args *ap)
1917 {
1918         struct hammer_inode *ip = ap->a_vp->v_data;
1919
1920         return(hammer_ioctl(ip, ap->a_command, ap->a_data,
1921                             ap->a_fflag, ap->a_cred));
1922 }
1923
1924 static
1925 int
1926 hammer_vop_mountctl(struct vop_mountctl_args *ap)
1927 {
1928         struct mount *mp;
1929         int error;
1930
1931         mp = ap->a_head.a_ops->head.vv_mount;
1932
1933         switch(ap->a_op) {
1934         case MOUNTCTL_SET_EXPORT:
1935                 if (ap->a_ctllen != sizeof(struct export_args))
1936                         error = EINVAL;
1937                 error = hammer_vfs_export(mp, ap->a_op,
1938                                       (const struct export_args *)ap->a_ctl);
1939                 break;
1940         default:
1941                 error = journal_mountctl(ap);
1942                 break;
1943         }
1944         return(error);
1945 }
1946
1947 /*
1948  * hammer_vop_strategy { vp, bio }
1949  *
1950  * Strategy call, used for regular file read & write only.  Note that the
1951  * bp may represent a cluster.
1952  *
1953  * To simplify operation and allow better optimizations in the future,
1954  * this code does not make any assumptions with regards to buffer alignment
1955  * or size.
1956  */
1957 static
1958 int
1959 hammer_vop_strategy(struct vop_strategy_args *ap)
1960 {
1961         struct buf *bp;
1962         int error;
1963
1964         bp = ap->a_bio->bio_buf;
1965
1966         switch(bp->b_cmd) {
1967         case BUF_CMD_READ:
1968                 error = hammer_vop_strategy_read(ap);
1969                 break;
1970         case BUF_CMD_WRITE:
1971                 error = hammer_vop_strategy_write(ap);
1972                 break;
1973         default:
1974                 bp->b_error = error = EINVAL;
1975                 bp->b_flags |= B_ERROR;
1976                 biodone(ap->a_bio);
1977                 break;
1978         }
1979         return (error);
1980 }
1981
1982 /*
1983  * Read from a regular file.  Iterate the related records and fill in the
1984  * BIO/BUF.  Gaps are zero-filled.
1985  *
1986  * The support code in hammer_object.c should be used to deal with mixed
1987  * in-memory and on-disk records.
1988  *
1989  * NOTE: Can be called from the cluster code with an oversized buf.
1990  *
1991  * XXX atime update
1992  */
1993 static
1994 int
1995 hammer_vop_strategy_read(struct vop_strategy_args *ap)
1996 {
1997         struct hammer_transaction trans;
1998         struct hammer_inode *ip;
1999         struct hammer_cursor cursor;
2000         hammer_base_elm_t base;
2001         hammer_off_t disk_offset;
2002         struct bio *bio;
2003         struct bio *nbio;
2004         struct buf *bp;
2005         int64_t rec_offset;
2006         int64_t ran_end;
2007         int64_t tmp64;
2008         int error;
2009         int boff;
2010         int roff;
2011         int n;
2012
2013         bio = ap->a_bio;
2014         bp = bio->bio_buf;
2015         ip = ap->a_vp->v_data;
2016
2017         /*
2018          * The zone-2 disk offset may have been set by the cluster code via
2019          * a BMAP operation, or else should be NOOFFSET.
2020          *
2021          * Checking the high bits for a match against zone-2 should suffice.
2022          */
2023         nbio = push_bio(bio);
2024         if ((nbio->bio_offset & HAMMER_OFF_ZONE_MASK) ==
2025             HAMMER_ZONE_RAW_BUFFER) {
2026                 error = hammer_io_direct_read(ip->hmp, nbio);
2027                 return (error);
2028         }
2029
2030         /*
2031          * Well, that sucked.  Do it the hard way.  If all the stars are
2032          * aligned we may still be able to issue a direct-read.
2033          */
2034         hammer_simple_transaction(&trans, ip->hmp);
2035         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2036
2037         /*
2038          * Key range (begin and end inclusive) to scan.  Note that the key's
2039          * stored in the actual records represent BASE+LEN, not BASE.  The
2040          * first record containing bio_offset will have a key > bio_offset.
2041          */
2042         cursor.key_beg.localization = ip->obj_localization +
2043                                       HAMMER_LOCALIZE_MISC;
2044         cursor.key_beg.obj_id = ip->obj_id;
2045         cursor.key_beg.create_tid = 0;
2046         cursor.key_beg.delete_tid = 0;
2047         cursor.key_beg.obj_type = 0;
2048         cursor.key_beg.key = bio->bio_offset + 1;
2049         cursor.asof = ip->obj_asof;
2050         cursor.flags |= HAMMER_CURSOR_ASOF;
2051
2052         cursor.key_end = cursor.key_beg;
2053         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2054 #if 0
2055         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
2056                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
2057                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
2058                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2059         } else
2060 #endif
2061         {
2062                 ran_end = bio->bio_offset + bp->b_bufsize;
2063                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2064                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2065                 tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2066                 if (tmp64 < ran_end)
2067                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2068                 else
2069                         cursor.key_end.key = ran_end + MAXPHYS + 1;
2070         }
2071         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2072
2073         error = hammer_ip_first(&cursor);
2074         boff = 0;
2075
2076         while (error == 0) {
2077                 /*
2078                  * Get the base file offset of the record.  The key for
2079                  * data records is (base + bytes) rather then (base).
2080                  */
2081                 base = &cursor.leaf->base;
2082                 rec_offset = base->key - cursor.leaf->data_len;
2083
2084                 /*
2085                  * Calculate the gap, if any, and zero-fill it.
2086                  *
2087                  * n is the offset of the start of the record verses our
2088                  * current seek offset in the bio.
2089                  */
2090                 n = (int)(rec_offset - (bio->bio_offset + boff));
2091                 if (n > 0) {
2092                         if (n > bp->b_bufsize - boff)
2093                                 n = bp->b_bufsize - boff;
2094                         bzero((char *)bp->b_data + boff, n);
2095                         boff += n;
2096                         n = 0;
2097                 }
2098
2099                 /*
2100                  * Calculate the data offset in the record and the number
2101                  * of bytes we can copy.
2102                  *
2103                  * There are two degenerate cases.  First, boff may already
2104                  * be at bp->b_bufsize.  Secondly, the data offset within
2105                  * the record may exceed the record's size.
2106                  */
2107                 roff = -n;
2108                 rec_offset += roff;
2109                 n = cursor.leaf->data_len - roff;
2110                 if (n <= 0) {
2111                         kprintf("strategy_read: bad n=%d roff=%d\n", n, roff);
2112                         n = 0;
2113                 } else if (n > bp->b_bufsize - boff) {
2114                         n = bp->b_bufsize - boff;
2115                 }
2116
2117                 /*
2118                  * Deal with cached truncations.  This cool bit of code
2119                  * allows truncate()/ftruncate() to avoid having to sync
2120                  * the file.
2121                  *
2122                  * If the frontend is truncated then all backend records are
2123                  * subject to the frontend's truncation.
2124                  *
2125                  * If the backend is truncated then backend records on-disk
2126                  * (but not in-memory) are subject to the backend's
2127                  * truncation.  In-memory records owned by the backend
2128                  * represent data written after the truncation point on the
2129                  * backend and must not be truncated.
2130                  *
2131                  * Truncate operations deal with frontend buffer cache
2132                  * buffers and frontend-owned in-memory records synchronously.
2133                  */
2134                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2135                         if (hammer_cursor_ondisk(&cursor) ||
2136                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2137                                 if (ip->trunc_off <= rec_offset)
2138                                         n = 0;
2139                                 else if (ip->trunc_off < rec_offset + n)
2140                                         n = (int)(ip->trunc_off - rec_offset);
2141                         }
2142                 }
2143                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2144                         if (hammer_cursor_ondisk(&cursor)) {
2145                                 if (ip->sync_trunc_off <= rec_offset)
2146                                         n = 0;
2147                                 else if (ip->sync_trunc_off < rec_offset + n)
2148                                         n = (int)(ip->sync_trunc_off - rec_offset);
2149                         }
2150                 }
2151
2152                 /*
2153                  * Try to issue a direct read into our bio if possible,
2154                  * otherwise resolve the element data into a hammer_buffer
2155                  * and copy.
2156                  *
2157                  * The buffer on-disk should be zerod past any real
2158                  * truncation point, but may not be for any synthesized
2159                  * truncation point from above.
2160                  */
2161                 if (boff == 0 && n == bp->b_bufsize &&
2162                     ((cursor.leaf->data_offset + roff) & HAMMER_BUFMASK) == 0) {
2163                         disk_offset = hammer_blockmap_lookup(
2164                                                 trans.hmp,
2165                                                 cursor.leaf->data_offset + roff,
2166                                                 &error);
2167                         if (error)
2168                                 break;
2169                         nbio->bio_offset = disk_offset;
2170                         error = hammer_io_direct_read(trans.hmp, nbio);
2171                         goto done;
2172                 } else if (n) {
2173                         error = hammer_ip_resolve_data(&cursor);
2174                         if (error == 0) {
2175                                 bcopy((char *)cursor.data + roff,
2176                                       (char *)bp->b_data + boff, n);
2177                         }
2178                 }
2179                 if (error)
2180                         break;
2181
2182                 /*
2183                  * Iterate until we have filled the request.
2184                  */
2185                 boff += n;
2186                 if (boff == bp->b_bufsize)
2187                         break;
2188                 error = hammer_ip_next(&cursor);
2189         }
2190
2191         /*
2192          * There may have been a gap after the last record
2193          */
2194         if (error == ENOENT)
2195                 error = 0;
2196         if (error == 0 && boff != bp->b_bufsize) {
2197                 KKASSERT(boff < bp->b_bufsize);
2198                 bzero((char *)bp->b_data + boff, bp->b_bufsize - boff);
2199                 /* boff = bp->b_bufsize; */
2200         }
2201         bp->b_resid = 0;
2202         bp->b_error = error;
2203         if (error)
2204                 bp->b_flags |= B_ERROR;
2205         biodone(ap->a_bio);
2206
2207 done:
2208         if (cursor.node)
2209                 hammer_cache_node(&ip->cache[1], cursor.node);
2210         hammer_done_cursor(&cursor);
2211         hammer_done_transaction(&trans);
2212         return(error);
2213 }
2214
2215 /*
2216  * BMAP operation - used to support cluster_read() only.
2217  *
2218  * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2219  *
2220  * This routine may return EOPNOTSUPP if the opration is not supported for
2221  * the specified offset.  The contents of the pointer arguments do not
2222  * need to be initialized in that case.
2223  *
2224  * If a disk address is available and properly aligned return 0 with
2225  * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2226  * to the run-length relative to that offset.  Callers may assume that
2227  * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2228  * large, so return EOPNOTSUPP if it is not sufficiently large.
2229  */
2230 static
2231 int
2232 hammer_vop_bmap(struct vop_bmap_args *ap)
2233 {
2234         struct hammer_transaction trans;
2235         struct hammer_inode *ip;
2236         struct hammer_cursor cursor;
2237         hammer_base_elm_t base;
2238         int64_t rec_offset;
2239         int64_t ran_end;
2240         int64_t tmp64;
2241         int64_t base_offset;
2242         int64_t base_disk_offset;
2243         int64_t last_offset;
2244         hammer_off_t last_disk_offset;
2245         hammer_off_t disk_offset;
2246         int     rec_len;
2247         int     error;
2248         int     blksize;
2249
2250         ip = ap->a_vp->v_data;
2251
2252         /*
2253          * We can only BMAP regular files.  We can't BMAP database files,
2254          * directories, etc.
2255          */
2256         if (ip->ino_data.obj_type != HAMMER_OBJTYPE_REGFILE)
2257                 return(EOPNOTSUPP);
2258
2259         /*
2260          * bmap is typically called with runp/runb both NULL when used
2261          * for writing.  We do not support BMAP for writing atm.
2262          */
2263         if (ap->a_cmd != BUF_CMD_READ)
2264                 return(EOPNOTSUPP);
2265
2266         /*
2267          * Scan the B-Tree to acquire blockmap addresses, then translate
2268          * to raw addresses.
2269          */
2270         hammer_simple_transaction(&trans, ip->hmp);
2271 #if 0
2272         kprintf("bmap_beg %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2273 #endif
2274         hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
2275
2276         /*
2277          * Key range (begin and end inclusive) to scan.  Note that the key's
2278          * stored in the actual records represent BASE+LEN, not BASE.  The
2279          * first record containing bio_offset will have a key > bio_offset.
2280          */
2281         cursor.key_beg.localization = ip->obj_localization +
2282                                       HAMMER_LOCALIZE_MISC;
2283         cursor.key_beg.obj_id = ip->obj_id;
2284         cursor.key_beg.create_tid = 0;
2285         cursor.key_beg.delete_tid = 0;
2286         cursor.key_beg.obj_type = 0;
2287         if (ap->a_runb)
2288                 cursor.key_beg.key = ap->a_loffset - MAXPHYS + 1;
2289         else
2290                 cursor.key_beg.key = ap->a_loffset + 1;
2291         if (cursor.key_beg.key < 0)
2292                 cursor.key_beg.key = 0;
2293         cursor.asof = ip->obj_asof;
2294         cursor.flags |= HAMMER_CURSOR_ASOF;
2295
2296         cursor.key_end = cursor.key_beg;
2297         KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
2298
2299         ran_end = ap->a_loffset + MAXPHYS;
2300         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
2301         cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
2302         tmp64 = ran_end + MAXPHYS + 1;  /* work-around GCC-4 bug */
2303         if (tmp64 < ran_end)
2304                 cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
2305         else
2306                 cursor.key_end.key = ran_end + MAXPHYS + 1;
2307
2308         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
2309
2310         error = hammer_ip_first(&cursor);
2311         base_offset = last_offset = 0;
2312         base_disk_offset = last_disk_offset = 0;
2313
2314         while (error == 0) {
2315                 /*
2316                  * Get the base file offset of the record.  The key for
2317                  * data records is (base + bytes) rather then (base).
2318                  *
2319                  * NOTE: rec_offset + rec_len may exceed the end-of-file.
2320                  * The extra bytes should be zero on-disk and the BMAP op
2321                  * should still be ok.
2322                  */
2323                 base = &cursor.leaf->base;
2324                 rec_offset = base->key - cursor.leaf->data_len;
2325                 rec_len    = cursor.leaf->data_len;
2326
2327                 /*
2328                  * Incorporate any cached truncation.
2329                  *
2330                  * NOTE: Modifications to rec_len based on synthesized
2331                  * truncation points remove the guarantee that any extended
2332                  * data on disk is zero (since the truncations may not have
2333                  * taken place on-media yet).
2334                  */
2335                 if (ip->flags & HAMMER_INODE_TRUNCATED) {
2336                         if (hammer_cursor_ondisk(&cursor) ||
2337                             cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
2338                                 if (ip->trunc_off <= rec_offset)
2339                                         rec_len = 0;
2340                                 else if (ip->trunc_off < rec_offset + rec_len)
2341                                         rec_len = (int)(ip->trunc_off - rec_offset);
2342                         }
2343                 }
2344                 if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
2345                         if (hammer_cursor_ondisk(&cursor)) {
2346                                 if (ip->sync_trunc_off <= rec_offset)
2347                                         rec_len = 0;
2348                                 else if (ip->sync_trunc_off < rec_offset + rec_len)
2349                                         rec_len = (int)(ip->sync_trunc_off - rec_offset);
2350                         }
2351                 }
2352
2353                 /*
2354                  * Accumulate information.  If we have hit a discontiguous
2355                  * block reset base_offset unless we are already beyond the
2356                  * requested offset.  If we are, that's it, we stop.
2357                  */
2358                 disk_offset = hammer_blockmap_lookup(trans.hmp,
2359                                                      cursor.leaf->data_offset,
2360                                                      &error);
2361                 if (error)
2362                         break;
2363                 if (rec_offset != last_offset ||
2364                     disk_offset != last_disk_offset) {
2365                         if (rec_offset > ap->a_loffset)
2366                                 break;
2367                         base_offset = rec_offset;
2368                         base_disk_offset = disk_offset;
2369                 }
2370                 last_offset = rec_offset + rec_len;
2371                 last_disk_offset = disk_offset + rec_len;
2372
2373                 error = hammer_ip_next(&cursor);
2374         }
2375
2376 #if 0
2377         kprintf("BMAP %016llx:  %016llx - %016llx\n",
2378                 ap->a_loffset, base_offset, last_offset);
2379         kprintf("BMAP %16s:  %016llx - %016llx\n",
2380                 "", base_disk_offset, last_disk_offset);
2381 #endif
2382
2383         if (cursor.node) {
2384                 hammer_cache_node(&ip->cache[1], cursor.node);
2385 #if 0
2386                 kprintf("bmap_end2 %016llx ip->cache %p\n", ap->a_loffset, ip->cache[1]);
2387 #endif
2388         }
2389         hammer_done_cursor(&cursor);
2390         hammer_done_transaction(&trans);
2391
2392         /*
2393          * If we couldn't find any records or the records we did find were
2394          * all behind the requested offset, return failure.  A forward
2395          * truncation can leave a hole w/ no on-disk records.
2396          */
2397         if (last_offset == 0 || last_offset < ap->a_loffset)
2398                 return (EOPNOTSUPP);
2399
2400         /*
2401          * Figure out the block size at the requested offset and adjust
2402          * our limits so the cluster_read() does not create inappropriately
2403          * sized buffer cache buffers.
2404          */
2405         blksize = hammer_blocksize(ap->a_loffset);
2406         if (hammer_blocksize(base_offset) != blksize) {
2407                 base_offset = hammer_blockdemarc(base_offset, ap->a_loffset);
2408         }
2409         if (last_offset != ap->a_loffset &&
2410             hammer_blocksize(last_offset - 1) != blksize) {
2411                 last_offset = hammer_blockdemarc(ap->a_loffset,
2412                                                  last_offset - 1);
2413         }
2414
2415         /*
2416          * Returning EOPNOTSUPP simply prevents the direct-IO optimization
2417          * from occuring.
2418          */
2419         disk_offset = base_disk_offset + (ap->a_loffset - base_offset);
2420
2421         /*
2422          * If doffsetp is not aligned or the forward run size does
2423          * not cover a whole buffer, disallow the direct I/O.
2424          */
2425         if ((disk_offset & HAMMER_BUFMASK) ||
2426             (last_offset - ap->a_loffset) < blksize) {
2427                 error = EOPNOTSUPP;
2428         } else {
2429                 *ap->a_doffsetp = disk_offset;
2430                 if (ap->a_runb) {
2431                         *ap->a_runb = ap->a_loffset - base_offset;
2432                         KKASSERT(*ap->a_runb >= 0);
2433                 }
2434                 if (ap->a_runp) {
2435                         *ap->a_runp = last_offset - ap->a_loffset;
2436                         KKASSERT(*ap->a_runp >= 0);
2437                 }
2438                 error = 0;
2439         }
2440         return(error);
2441 }
2442
2443 /*
2444  * Write to a regular file.   Because this is a strategy call the OS is
2445  * trying to actually get data onto the media.
2446  */
2447 static
2448 int
2449 hammer_vop_strategy_write(struct vop_strategy_args *ap)
2450 {
2451         hammer_record_t record;
2452         hammer_mount_t hmp;
2453         hammer_inode_t ip;
2454         struct bio *bio;
2455         struct buf *bp;
2456         int bytes;
2457         int error;
2458
2459         bio = ap->a_bio;
2460         bp = bio->bio_buf;
2461         ip = ap->a_vp->v_data;
2462         hmp = ip->hmp;
2463
2464         KKASSERT(bp->b_bufsize == hammer_blocksize(bio->bio_offset));
2465
2466         if (ip->flags & HAMMER_INODE_RO) {
2467                 bp->b_error = EROFS;
2468                 bp->b_flags |= B_ERROR;
2469                 biodone(ap->a_bio);
2470                 hammer_cleanup_write_io(ip);
2471                 return(EROFS);
2472         }
2473
2474         /*
2475          * Interlock with inode destruction (no in-kernel or directory
2476          * topology visibility).  If we queue new IO while trying to
2477          * destroy the inode we can deadlock the vtrunc call in
2478          * hammer_inode_unloadable_check().
2479          */
2480         if (ip->flags & (HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) {
2481                 bp->b_resid = 0;
2482                 biodone(ap->a_bio);
2483                 hammer_cleanup_write_io(ip);
2484                 return(0);
2485         }
2486
2487         /*
2488          * Reserve space and issue a direct-write from the front-end.
2489          * NOTE: The direct_io code will hammer_bread/bcopy smaller
2490          * allocations.
2491          *
2492          * An in-memory record will be installed to reference the storage
2493          * until the flusher can get to it.
2494          *
2495          * Since we own the high level bio the front-end will not try to
2496          * do a direct-read until the write completes.
2497          *
2498          * NOTE: The only time we do not reserve a full-sized buffers
2499          * worth of data is if the file is small.  We do not try to
2500          * allocate a fragment (from the small-data zone) at the end of
2501          * an otherwise large file as this can lead to wildly separated
2502          * data.
2503          */
2504         KKASSERT((bio->bio_offset & HAMMER_BUFMASK) == 0);
2505         KKASSERT(bio->bio_offset < ip->ino_data.size);
2506         if (bio->bio_offset || ip->ino_data.size > HAMMER_BUFSIZE / 2)
2507                 bytes = bp->b_bufsize;
2508         else
2509                 bytes = ((int)ip->ino_data.size + 15) & ~15;
2510
2511         record = hammer_ip_add_bulk(ip, bio->bio_offset, bp->b_data,
2512                                     bytes, &error);
2513         if (record) {
2514                 hammer_io_direct_write(hmp, &record->leaf, bio);
2515                 hammer_rel_mem_record(record);
2516                 if (ip->rsv_recs > 1 && hmp->rsv_recs > hammer_limit_recs)
2517                         hammer_flush_inode(ip, 0);
2518         } else {
2519                 bp->b_bio2.bio_offset = NOOFFSET;
2520                 bp->b_error = error;
2521                 bp->b_flags |= B_ERROR;
2522                 biodone(ap->a_bio);
2523         }
2524         hammer_cleanup_write_io(ip);
2525         return(error);
2526 }
2527
2528 /*
2529  * Clean-up after disposing of a dirty frontend buffer's data.
2530  * This is somewhat heuristical so try to be robust.
2531  */
2532 static void
2533 hammer_cleanup_write_io(hammer_inode_t ip)
2534 {
2535         if (ip->rsv_databufs) {
2536                 --ip->rsv_databufs;
2537                 --ip->hmp->rsv_databufs;
2538         }
2539 }
2540
2541 /*
2542  * We can lose track of dirty buffer cache buffers if we truncate, this
2543  * routine will resynchronize the count.
2544  */
2545 static
2546 void
2547 hammer_update_rsv_databufs(hammer_inode_t ip)
2548 {
2549         struct buf *bp;
2550         int delta;
2551         int n;
2552
2553         if (ip->vp) {
2554                 n = 0;
2555                 RB_FOREACH(bp, buf_rb_tree, &ip->vp->v_rbdirty_tree) {
2556                         ++n;
2557                 }
2558         } else {
2559                 n = 0;
2560         }
2561         delta = n - ip->rsv_databufs;
2562         ip->rsv_databufs += delta;
2563         ip->hmp->rsv_databufs += delta;
2564 }
2565
2566 /*
2567  * dounlink - disconnect a directory entry
2568  *
2569  * XXX whiteout support not really in yet
2570  */
2571 static int
2572 hammer_dounlink(hammer_transaction_t trans, struct nchandle *nch,
2573                 struct vnode *dvp, struct ucred *cred, int flags)
2574 {
2575         struct namecache *ncp;
2576         hammer_inode_t dip;
2577         hammer_inode_t ip;
2578         struct hammer_cursor cursor;
2579         int64_t namekey;
2580         int nlen, error;
2581
2582         /*
2583          * Calculate the namekey and setup the key range for the scan.  This
2584          * works kinda like a chained hash table where the lower 32 bits
2585          * of the namekey synthesize the chain.
2586          *
2587          * The key range is inclusive of both key_beg and key_end.
2588          */
2589         dip = VTOI(dvp);
2590         ncp = nch->ncp;
2591
2592         if (dip->flags & HAMMER_INODE_RO)
2593                 return (EROFS);
2594
2595         namekey = hammer_directory_namekey(ncp->nc_name, ncp->nc_nlen);
2596 retry:
2597         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
2598         cursor.key_beg.localization = dip->obj_localization +
2599                                       HAMMER_LOCALIZE_MISC;
2600         cursor.key_beg.obj_id = dip->obj_id;
2601         cursor.key_beg.key = namekey;
2602         cursor.key_beg.create_tid = 0;
2603         cursor.key_beg.delete_tid = 0;
2604         cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
2605         cursor.key_beg.obj_type = 0;
2606
2607         cursor.key_end = cursor.key_beg;
2608         cursor.key_end.key |= 0xFFFFFFFFULL;
2609         cursor.asof = dip->obj_asof;
2610         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2611
2612         /*
2613          * Scan all matching records (the chain), locate the one matching
2614          * the requested path component.  info->last_error contains the
2615          * error code on search termination and could be 0, ENOENT, or
2616          * something else.
2617          *
2618          * The hammer_ip_*() functions merge in-memory records with on-disk
2619          * records for the purposes of the search.
2620          */
2621         error = hammer_ip_first(&cursor);
2622
2623         while (error == 0) {
2624                 error = hammer_ip_resolve_data(&cursor);
2625                 if (error)
2626                         break;
2627                 nlen = cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF;
2628                 KKASSERT(nlen > 0);
2629                 if (ncp->nc_nlen == nlen &&
2630                     bcmp(ncp->nc_name, cursor.data->entry.name, nlen) == 0) {
2631                         break;
2632                 }
2633                 error = hammer_ip_next(&cursor);
2634         }
2635
2636         /*
2637          * If all is ok we have to get the inode so we can adjust nlinks.
2638          * To avoid a deadlock with the flusher we must release the inode
2639          * lock on the directory when acquiring the inode for the entry.
2640          *
2641          * If the target is a directory, it must be empty.
2642          */
2643         if (error == 0) {
2644                 hammer_unlock(&cursor.ip->lock);
2645                 ip = hammer_get_inode(trans, dip, cursor.data->entry.obj_id,
2646                                       dip->hmp->asof,
2647                                       cursor.data->entry.localization,
2648                                       0, &error);
2649                 hammer_lock_sh(&cursor.ip->lock);
2650                 if (error == ENOENT) {
2651                         kprintf("obj_id %016llx\n", cursor.data->entry.obj_id);
2652                         Debugger("ENOENT unlinking object that should exist");
2653                 }
2654
2655                 /*
2656                  * If we are trying to remove a directory the directory must
2657                  * be empty.
2658                  *
2659                  * WARNING: hammer_ip_check_directory_empty() may have to
2660                  * terminate the cursor to avoid a deadlock.  It is ok to
2661                  * call hammer_done_cursor() twice.
2662                  */
2663                 if (error == 0 && ip->ino_data.obj_type ==
2664                                   HAMMER_OBJTYPE_DIRECTORY) {
2665                         error = hammer_ip_check_directory_empty(trans, ip);
2666                 }
2667
2668                 /*
2669                  * Delete the directory entry.
2670                  *
2671                  * WARNING: hammer_ip_del_directory() may have to terminate
2672                  * the cursor to avoid a deadlock.  It is ok to call
2673                  * hammer_done_cursor() twice.
2674                  */
2675                 if (error == 0) {
2676                         error = hammer_ip_del_directory(trans, &cursor,
2677                                                         dip, ip);
2678                 }
2679                 hammer_done_cursor(&cursor);
2680                 if (error == 0) {
2681                         cache_setunresolved(nch);
2682                         cache_setvp(nch, NULL);
2683                         /* XXX locking */
2684                         if (ip->vp)
2685                                 cache_inval_vp(ip->vp, CINV_DESTROY);
2686                 }
2687                 if (ip)
2688                         hammer_rel_inode(ip, 0);
2689         } else {
2690                 hammer_done_cursor(&cursor);
2691         }
2692         if (error == EDEADLK)
2693                 goto retry;
2694
2695         return (error);
2696 }
2697
2698 /************************************************************************
2699  *                          FIFO AND SPECFS OPS                         *
2700  ************************************************************************
2701  *
2702  */
2703
2704 static int
2705 hammer_vop_fifoclose (struct vop_close_args *ap)
2706 {
2707         /* XXX update itimes */
2708         return (VOCALL(&fifo_vnode_vops, &ap->a_head));
2709 }
2710
2711 static int
2712 hammer_vop_fiforead (struct vop_read_args *ap)
2713 {
2714         int error;
2715
2716         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2717         /* XXX update access time */
2718         return (error);
2719 }
2720
2721 static int
2722 hammer_vop_fifowrite (struct vop_write_args *ap)
2723 {
2724         int error;
2725
2726         error = VOCALL(&fifo_vnode_vops, &ap->a_head);
2727         /* XXX update access time */
2728         return (error);
2729 }
2730
2731 static int
2732 hammer_vop_specclose (struct vop_close_args *ap)
2733 {
2734         /* XXX update itimes */
2735         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2736 }
2737
2738 static int
2739 hammer_vop_specread (struct vop_read_args *ap)
2740 {
2741         /* XXX update access time */
2742         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2743 }
2744
2745 static int
2746 hammer_vop_specwrite (struct vop_write_args *ap)
2747 {
2748         /* XXX update last change time */
2749         return (VOCALL(&spec_vnode_vops, &ap->a_head));
2750 }
2751