fs/xfs/xfs_attr_leaf.c

   1 /*
   2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.  All Rights Reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of version 2 of the GNU General Public License as
   6  * published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it would be useful, but
   9  * WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  11  *
  12  * Further, this software is distributed without any warranty that it is
  13  * free of the rightful claim of any third person regarding infringement
  14  * or the like.  Any license provided herein, whether implied or
  15  * otherwise, applies only to this software file.  Patent licenses, if
  16  * any, provided herein do not apply to combinations of this program with
  17  * other software, or any other product whatsoever.
  18  *
  19  * You should have received a copy of the GNU General Public License along
  20  * with this program; if not, write the Free Software Foundation, Inc., 59
  21  * Temple Place - Suite 330, Boston MA 02111-1307, USA.
  22  *
  23  * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
  24  * Mountain View, CA  94043, or:
  25  *
  26  * http://www.sgi.com
  27  *
  28  * For further information regarding this notice, see:
  29  *
  30  * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
  31  */
  32 /*
  33  * xfs_attr_leaf.c
  34  *
  35  * GROT: figure out how to recover gracefully when bmap returns ENOSPC.
  36  */
  37
  38 #include "xfs.h"
  39
  40 #include "xfs_macros.h"
  41 #include "xfs_types.h"
  42 #include "xfs_inum.h"
  43 #include "xfs_log.h"
  44 #include "xfs_trans.h"
  45 #include "xfs_sb.h"
  46 #include "xfs_ag.h"
  47 #include "xfs_dir.h"
  48 #include "xfs_dir2.h"
  49 #include "xfs_dmapi.h"
  50 #include "xfs_mount.h"
  51 #include "xfs_alloc_btree.h"
  52 #include "xfs_bmap_btree.h"
  53 #include "xfs_ialloc_btree.h"
  54 #include "xfs_alloc.h"
  55 #include "xfs_btree.h"
  56 #include "xfs_attr_sf.h"
  57 #include "xfs_dir_sf.h"
  58 #include "xfs_dir2_sf.h"
  59 #include "xfs_dinode.h"
  60 #include "xfs_inode_item.h"
  61 #include "xfs_inode.h"
  62 #include "xfs_bmap.h"
  63 #include "xfs_da_btree.h"
  64 #include "xfs_attr.h"
  65 #include "xfs_attr_leaf.h"
  66 #include "xfs_error.h"
  67 #include "xfs_bit.h"
  68
  69 /*
  70  * xfs_attr_leaf.c
  71  *
  72  * Routines to implement leaf blocks of attributes as Btrees of hashed names.
  73  */
  74
  75 /*========================================================================
  76  * Function prototypes for the kernel.
  77  *========================================================================*/
  78
  79 /*
  80  * Routines used for growing the Btree.
  81  */
  82 STATIC int xfs_attr_leaf_create(xfs_da_args_t *args, xfs_dablk_t which_block,
  83                                     xfs_dabuf_t **bpp);
  84 STATIC int xfs_attr_leaf_add_work(xfs_dabuf_t *leaf_buffer, xfs_da_args_t *args,
  85                                               int freemap_index);
  86 STATIC void xfs_attr_leaf_compact(xfs_trans_t *trans, xfs_dabuf_t *leaf_buffer);
  87 STATIC void xfs_attr_leaf_rebalance(xfs_da_state_t *state,
  88                                                    xfs_da_state_blk_t *blk1,
  89                                                    xfs_da_state_blk_t *blk2);
  90 STATIC int xfs_attr_leaf_figure_balance(xfs_da_state_t *state,
  91                                            xfs_da_state_blk_t *leaf_blk_1,
  92                                            xfs_da_state_blk_t *leaf_blk_2,
  93                                            int *number_entries_in_blk1,
  94                                            int *number_usedbytes_in_blk1);
  95
  96 /*
  97  * Routines used for shrinking the Btree.
  98  */
  99 STATIC int xfs_attr_node_inactive(xfs_trans_t **trans, xfs_inode_t *dp,
 100                                   xfs_dabuf_t *bp, int level);
 101 STATIC int xfs_attr_leaf_inactive(xfs_trans_t **trans, xfs_inode_t *dp,
 102                                   xfs_dabuf_t *bp);
 103 STATIC int xfs_attr_leaf_freextent(xfs_trans_t **trans, xfs_inode_t *dp,
 104                                    xfs_dablk_t blkno, int blkcnt);
 105
 106 /*
 107  * Utility routines.
 108  */
 109 STATIC void xfs_attr_leaf_moveents(xfs_attr_leafblock_t *src_leaf,
 110                                          int src_start,
 111                                          xfs_attr_leafblock_t *dst_leaf,
 112                                          int dst_start, int move_count,
 113                                          xfs_mount_t *mp);
 114 STATIC int xfs_attr_leaf_entsize(xfs_attr_leafblock_t *leaf, int index);
 115 STATIC int xfs_attr_put_listent(xfs_attr_list_context_t *context,
 116                              attrnames_t *, char *name, int namelen,
 117                              int valuelen);
 118
 119
 120 /*========================================================================
 121  * External routines when dirsize < XFS_LITINO(mp).
 122  *========================================================================*/
 123
 124 /*
 125  * Create the initial contents of a shortform attribute list.
 126  */
 127 int
 128 xfs_attr_shortform_create(xfs_da_args_t *args)
 129 {
 130         xfs_attr_sf_hdr_t *hdr;
 131         xfs_inode_t *dp;
 132         xfs_ifork_t *ifp;
 133
 134         dp = args->dp;
 135         ASSERT(dp != NULL);
 136         ifp = dp->i_afp;
 137         ASSERT(ifp != NULL);
 138         ASSERT(ifp->if_bytes == 0);
 139         if (dp->i_d.di_aformat == XFS_DINODE_FMT_EXTENTS) {
 140                 ifp->if_flags &= ~XFS_IFEXTENTS;        /* just in case */
 141                 dp->i_d.di_aformat = XFS_DINODE_FMT_LOCAL;
 142                 ifp->if_flags |= XFS_IFINLINE;
 143         } else {
 144                 ASSERT(ifp->if_flags & XFS_IFINLINE);
 145         }
 146         xfs_idata_realloc(dp, sizeof(*hdr), XFS_ATTR_FORK);
 147         hdr = (xfs_attr_sf_hdr_t *)ifp->if_u1.if_data;
 148         hdr->count = 0;
 149         INT_SET(hdr->totsize, ARCH_CONVERT, sizeof(*hdr));
 150         xfs_trans_log_inode(args->trans, dp, XFS_ILOG_CORE | XFS_ILOG_ADATA);
 151         return(0);
 152 }
 153
 154 /*
 155  * Add a name/value pair to the shortform attribute list.
 156  * Overflow from the inode has already been checked for.
 157  */
 158 int
 159 xfs_attr_shortform_add(xfs_da_args_t *args)
 160 {
 161         xfs_attr_shortform_t *sf;
 162         xfs_attr_sf_entry_t *sfe;
 163         int i, offset, size;
 164         xfs_inode_t *dp;
 165         xfs_ifork_t *ifp;
 166
 167         dp = args->dp;
 168         ifp = dp->i_afp;
 169         ASSERT(ifp->if_flags & XFS_IFINLINE);
 170         sf = (xfs_attr_shortform_t *)ifp->if_u1.if_data;
 171         sfe = &sf->list[0];
 172         for (i = 0; i < INT_GET(sf->hdr.count, ARCH_CONVERT);
 173                                 sfe = XFS_ATTR_SF_NEXTENTRY(sfe), i++) {
 174                 if (sfe->namelen != args->namelen)
 175                         continue;
 176                 if (memcmp(args->name, sfe->nameval, args->namelen) != 0)
 177                         continue;
 178                 if (((args->flags & ATTR_SECURE) != 0) !=
 179                     ((sfe->flags & XFS_ATTR_SECURE) != 0))
 180                         continue;
 181                 if (((args->flags & ATTR_ROOT) != 0) !=
 182                     ((sfe->flags & XFS_ATTR_ROOT) != 0))
 183                         continue;
 184                 return(XFS_ERROR(EEXIST));
 185         }
 186
 187         offset = (char *)sfe - (char *)sf;
 188         size = XFS_ATTR_SF_ENTSIZE_BYNAME(args->namelen, args->valuelen);
 189         xfs_idata_realloc(dp, size, XFS_ATTR_FORK);
 190         sf = (xfs_attr_shortform_t *)ifp->if_u1.if_data;
 191         sfe = (xfs_attr_sf_entry_t *)((char *)sf + offset);
 192
 193         sfe->namelen = args->namelen;
 194         INT_SET(sfe->valuelen, ARCH_CONVERT, args->valuelen);
 195         sfe->flags = (args->flags & ATTR_SECURE) ? XFS_ATTR_SECURE :
 196                         ((args->flags & ATTR_ROOT) ? XFS_ATTR_ROOT : 0);
 197         memcpy(sfe->nameval, args->name, args->namelen);
 198         memcpy(&sfe->nameval[args->namelen], args->value, args->valuelen);
 199         INT_MOD(sf->hdr.count, ARCH_CONVERT, 1);
 200         INT_MOD(sf->hdr.totsize, ARCH_CONVERT, size);
 201         xfs_trans_log_inode(args->trans, dp, XFS_ILOG_CORE | XFS_ILOG_ADATA);
 202
 203         return(0);
 204 }
 205
 206 /*
 207  * Remove a name from the shortform attribute list structure.
 208  */
 209 int
 210 xfs_attr_shortform_remove(xfs_da_args_t *args)
 211 {
 212         xfs_attr_shortform_t *sf;
 213         xfs_attr_sf_entry_t *sfe;
 214         int base, size=0, end, totsize, i;
 215         xfs_inode_t *dp;
 216
 217         /*
 218          * Remove the attribute.
 219          */
 220         dp = args->dp;
 221         base = sizeof(xfs_attr_sf_hdr_t);
 222         sf = (xfs_attr_shortform_t *)dp->i_afp->if_u1.if_data;
 223         sfe = &sf->list[0];
 224         for (i = 0; i < INT_GET(sf->hdr.count, ARCH_CONVERT);
 225                                 sfe = XFS_ATTR_SF_NEXTENTRY(sfe),
 226                                         base += size, i++) {
 227                 size = XFS_ATTR_SF_ENTSIZE(sfe);
 228                 if (sfe->namelen != args->namelen)
 229                         continue;
 230                 if (memcmp(sfe->nameval, args->name, args->namelen) != 0)
 231                         continue;
 232                 if (((args->flags & ATTR_SECURE) != 0) !=
 233                     ((sfe->flags & XFS_ATTR_SECURE) != 0))
 234                         continue;
 235                 if (((args->flags & ATTR_ROOT) != 0) !=
 236                     ((sfe->flags & XFS_ATTR_ROOT) != 0))
 237                         continue;
 238                 break;
 239         }
 240         if (i == INT_GET(sf->hdr.count, ARCH_CONVERT))
 241                 return(XFS_ERROR(ENOATTR));
 242
 243         end = base + size;
 244         totsize = INT_GET(sf->hdr.totsize, ARCH_CONVERT);
 245         if (end != totsize) {
 246                 memmove(&((char *)sf)[base], &((char *)sf)[end],
 247                                                         totsize - end);
 248         }
 249         INT_MOD(sf->hdr.count, ARCH_CONVERT, -1);
 250         INT_MOD(sf->hdr.totsize, ARCH_CONVERT, -size);
 251         xfs_idata_realloc(dp, -size, XFS_ATTR_FORK);
 252         xfs_trans_log_inode(args->trans, dp, XFS_ILOG_CORE | XFS_ILOG_ADATA);
 253
 254         return(0);
 255 }
 256
 257 /*
 258  * Look up a name in a shortform attribute list structure.
 259  */
 260 /*ARGSUSED*/
 261 int
 262 xfs_attr_shortform_lookup(xfs_da_args_t *args)
 263 {
 264         xfs_attr_shortform_t *sf;
 265         xfs_attr_sf_entry_t *sfe;
 266         int i;
 267         xfs_ifork_t *ifp;
 268
 269         ifp = args->dp->i_afp;
 270         ASSERT(ifp->if_flags & XFS_IFINLINE);
 271         sf = (xfs_attr_shortform_t *)ifp->if_u1.if_data;
 272         sfe = &sf->list[0];
 273         for (i = 0; i < INT_GET(sf->hdr.count, ARCH_CONVERT);
 274                                 sfe = XFS_ATTR_SF_NEXTENTRY(sfe), i++) {
 275                 if (sfe->namelen != args->namelen)
 276                         continue;
 277                 if (memcmp(args->name, sfe->nameval, args->namelen) != 0)
 278                         continue;
 279                 if (((args->flags & ATTR_SECURE) != 0) !=
 280                     ((sfe->flags & XFS_ATTR_SECURE) != 0))
 281                         continue;
 282                 if (((args->flags & ATTR_ROOT) != 0) !=
 283                     ((sfe->flags & XFS_ATTR_ROOT) != 0))
 284                         continue;
 285                 return(XFS_ERROR(EEXIST));
 286         }
 287         return(XFS_ERROR(ENOATTR));
 288 }
 289
 290 /*
 291  * Look up a name in a shortform attribute list structure.
 292  */
 293 /*ARGSUSED*/
 294 int
 295 xfs_attr_shortform_getvalue(xfs_da_args_t *args)
 296 {
 297         xfs_attr_shortform_t *sf;
 298         xfs_attr_sf_entry_t *sfe;
 299         int i;
 300
 301         ASSERT(args->dp->i_d.di_aformat == XFS_IFINLINE);
 302         sf = (xfs_attr_shortform_t *)args->dp->i_afp->if_u1.if_data;
 303         sfe = &sf->list[0];
 304         for (i = 0; i < INT_GET(sf->hdr.count, ARCH_CONVERT);
 305                                 sfe = XFS_ATTR_SF_NEXTENTRY(sfe), i++) {
 306                 if (sfe->namelen != args->namelen)
 307                         continue;
 308                 if (memcmp(args->name, sfe->nameval, args->namelen) != 0)
 309                         continue;
 310                 if (((args->flags & ATTR_SECURE) != 0) !=
 311                     ((sfe->flags & XFS_ATTR_SECURE) != 0))
 312                         continue;
 313                 if (((args->flags & ATTR_ROOT) != 0) !=
 314                     ((sfe->flags & XFS_ATTR_ROOT) != 0))
 315                         continue;
 316                 if (args->flags & ATTR_KERNOVAL) {
 317                         args->valuelen = INT_GET(sfe->valuelen, ARCH_CONVERT);
 318                         return(XFS_ERROR(EEXIST));
 319                 }
 320                 if (args->valuelen < INT_GET(sfe->valuelen, ARCH_CONVERT)) {
 321                         args->valuelen = INT_GET(sfe->valuelen, ARCH_CONVERT);
 322                         return(XFS_ERROR(ERANGE));
 323                 }
 324                 args->valuelen = INT_GET(sfe->valuelen, ARCH_CONVERT);
 325                 memcpy(args->value, &sfe->nameval[args->namelen],
 326                                                     args->valuelen);
 327                 return(XFS_ERROR(EEXIST));
 328         }
 329         return(XFS_ERROR(ENOATTR));
 330 }
 331
 332 /*
 333  * Convert from using the shortform to the leaf.
 334  */
 335 int
 336 xfs_attr_shortform_to_leaf(xfs_da_args_t *args)
 337 {
 338         xfs_inode_t *dp;
 339         xfs_attr_shortform_t *sf;
 340         xfs_attr_sf_entry_t *sfe;
 341         xfs_da_args_t nargs;
 342         char *tmpbuffer;
 343         int error, i, size;
 344         xfs_dablk_t blkno;
 345         xfs_dabuf_t *bp;
 346         xfs_ifork_t *ifp;
 347
 348         dp = args->dp;
 349         ifp = dp->i_afp;
 350         sf = (xfs_attr_shortform_t *)ifp->if_u1.if_data;
 351         size = INT_GET(sf->hdr.totsize, ARCH_CONVERT);
 352         tmpbuffer = kmem_alloc(size, KM_SLEEP);
 353         ASSERT(tmpbuffer != NULL);
 354         memcpy(tmpbuffer, ifp->if_u1.if_data, size);
 355         sf = (xfs_attr_shortform_t *)tmpbuffer;
 356
 357         xfs_idata_realloc(dp, -size, XFS_ATTR_FORK);
 358         bp = NULL;
 359         error = xfs_da_grow_inode(args, &blkno);
 360         if (error) {
 361                 /*
 362                  * If we hit an IO error middle of the transaction inside
 363                  * grow_inode(), we may have inconsistent data. Bail out.
 364                  */
 365                 if (error == EIO)
 366                         goto out;
 367                 xfs_idata_realloc(dp, size, XFS_ATTR_FORK);     /* try to put */
 368                 memcpy(ifp->if_u1.if_data, tmpbuffer, size);    /* it back */
 369                 goto out;
 370         }
 371
 372         ASSERT(blkno == 0);
 373         error = xfs_attr_leaf_create(args, blkno, &bp);
 374         if (error) {
 375                 error = xfs_da_shrink_inode(args, 0, bp);
 376                 bp = NULL;
 377                 if (error)
 378                         goto out;
 379                 xfs_idata_realloc(dp, size, XFS_ATTR_FORK);     /* try to put */
 380                 memcpy(ifp->if_u1.if_data, tmpbuffer, size);    /* it back */
 381                 goto out;
 382         }
 383
 384         memset((char *)&nargs, 0, sizeof(nargs));
 385         nargs.dp = dp;
 386         nargs.firstblock = args->firstblock;
 387         nargs.flist = args->flist;
 388         nargs.total = args->total;
 389         nargs.whichfork = XFS_ATTR_FORK;
 390         nargs.trans = args->trans;
 391         nargs.oknoent = 1;
 392
 393         sfe = &sf->list[0];
 394         for (i = 0; i < INT_GET(sf->hdr.count, ARCH_CONVERT); i++) {
 395                 nargs.name = (char *)sfe->nameval;
 396                 nargs.namelen = sfe->namelen;
 397                 nargs.value = (char *)&sfe->nameval[nargs.namelen];
 398                 nargs.valuelen = INT_GET(sfe->valuelen, ARCH_CONVERT);
 399                 nargs.hashval = xfs_da_hashname((char *)sfe->nameval,
 400                                                 sfe->namelen);
 401                 nargs.flags = (sfe->flags & XFS_ATTR_SECURE) ? ATTR_SECURE :
 402                                 ((sfe->flags & XFS_ATTR_ROOT) ? ATTR_ROOT : 0);
 403                 error = xfs_attr_leaf_lookup_int(bp, &nargs); /* set a->index */
 404                 ASSERT(error == ENOATTR);
 405                 error = xfs_attr_leaf_add(bp, &nargs);
 406                 ASSERT(error != ENOSPC);
 407                 if (error)
 408                         goto out;
 409                 sfe = XFS_ATTR_SF_NEXTENTRY(sfe);
 410         }
 411         error = 0;
 412
 413 out:
 414         if(bp)
 415                 xfs_da_buf_done(bp);
 416         kmem_free(tmpbuffer, size);
 417         return(error);
 418 }
 419
 420 STATIC int
 421 xfs_attr_shortform_compare(const void *a, const void *b)
 422 {
 423         xfs_attr_sf_sort_t *sa, *sb;
 424
 425         sa = (xfs_attr_sf_sort_t *)a;
 426         sb = (xfs_attr_sf_sort_t *)b;
 427         if (INT_GET(sa->hash, ARCH_CONVERT)
 428                                 < INT_GET(sb->hash, ARCH_CONVERT)) {
 429                 return(-1);
 430         } else if (INT_GET(sa->hash, ARCH_CONVERT)
 431                                 > INT_GET(sb->hash, ARCH_CONVERT)) {
 432                 return(1);
 433         } else {
 434                 return(sa->entno - sb->entno);
 435         }
 436 }
 437
 438 /*
 439  * Copy out entries of shortform attribute lists for attr_list().
 440  * Shortform atrtribute lists are not stored in hashval sorted order.
 441  * If the output buffer is not large enough to hold them all, then we
 442  * we have to calculate each entries' hashvalue and sort them before
 443  * we can begin returning them to the user.
 444  */
 445 /*ARGSUSED*/
 446 int
 447 xfs_attr_shortform_list(xfs_attr_list_context_t *context)
 448 {
 449         attrlist_cursor_kern_t *cursor;
 450         xfs_attr_sf_sort_t *sbuf, *sbp;
 451         xfs_attr_shortform_t *sf;
 452         xfs_attr_sf_entry_t *sfe;
 453         xfs_inode_t *dp;
 454         int sbsize, nsbuf, count, i;
 455
 456         ASSERT(context != NULL);
 457         dp = context->dp;
 458         ASSERT(dp != NULL);
 459         ASSERT(dp->i_afp != NULL);
 460         sf = (xfs_attr_shortform_t *)dp->i_afp->if_u1.if_data;
 461         ASSERT(sf != NULL);
 462         if (!sf->hdr.count)
 463                 return(0);
 464         cursor = context->cursor;
 465         ASSERT(cursor != NULL);
 466
 467         xfs_attr_trace_l_c("sf start", context);
 468
 469         /*
 470          * If the buffer is large enough, do not bother with sorting.
 471          * Note the generous fudge factor of 16 overhead bytes per entry.
 472          */
 473         if ((dp->i_afp->if_bytes + INT_GET(sf->hdr.count, ARCH_CONVERT) * 16)
 474                                                         < context->bufsize) {
 475                 for (i = 0, sfe = &sf->list[0];
 476                                 i < INT_GET(sf->hdr.count, ARCH_CONVERT); i++) {
 477                         attrnames_t     *namesp;
 478
 479                         if (((context->flags & ATTR_SECURE) != 0) !=
 480                             ((sfe->flags & XFS_ATTR_SECURE) != 0) &&
 481                             !(context->flags & ATTR_KERNORMALS)) {
 482                                 sfe = XFS_ATTR_SF_NEXTENTRY(sfe);
 483                                 continue;
 484                         }
 485                         if (((context->flags & ATTR_ROOT) != 0) !=
 486                             ((sfe->flags & XFS_ATTR_ROOT) != 0) &&
 487                             !(context->flags & ATTR_KERNROOTLS)) {
 488                                 sfe = XFS_ATTR_SF_NEXTENTRY(sfe);
 489                                 continue;
 490                         }
 491                         namesp = (sfe->flags & XFS_ATTR_SECURE) ? &attr_secure:
 492                                 ((sfe->flags & XFS_ATTR_ROOT) ? &attr_trusted :
 493                                   &attr_user);
 494                         if (context->flags & ATTR_KERNOVAL) {
 495                                 ASSERT(context->flags & ATTR_KERNAMELS);
 496                                 context->count += namesp->attr_namelen +
 497                                         INT_GET(sfe->namelen, ARCH_CONVERT) + 1;
 498                         }
 499                         else {
 500                                 if (xfs_attr_put_listent(context, namesp,
 501                                                    (char *)sfe->nameval,
 502                                                    (int)sfe->namelen,
 503                                                    (int)INT_GET(sfe->valuelen,
 504                                                                 ARCH_CONVERT)))
 505                                         break;
 506                         }
 507                         sfe = XFS_ATTR_SF_NEXTENTRY(sfe);
 508                 }
 509                 xfs_attr_trace_l_c("sf big-gulp", context);
 510                 return(0);
 511         }
 512
 513         /*
 514          * It didn't all fit, so we have to sort everything on hashval.
 515          */
 516         sbsize = INT_GET(sf->hdr.count, ARCH_CONVERT) * sizeof(*sbuf);
 517         sbp = sbuf = kmem_alloc(sbsize, KM_SLEEP);
 518
 519         /*
 520          * Scan the attribute list for the rest of the entries, storing
 521          * the relevant info from only those that match into a buffer.
 522          */
 523         nsbuf = 0;
 524         for (i = 0, sfe = &sf->list[0];
 525                         i < INT_GET(sf->hdr.count, ARCH_CONVERT); i++) {
 526                 if (unlikely(
 527                     ((char *)sfe < (char *)sf) ||
 528                     ((char *)sfe >= ((char *)sf + dp->i_afp->if_bytes)))) {
 529                         XFS_CORRUPTION_ERROR("xfs_attr_shortform_list",
 530                                              XFS_ERRLEVEL_LOW,
 531                                              context->dp->i_mount, sfe);
 532                         xfs_attr_trace_l_c("sf corrupted", context);
 533                         kmem_free(sbuf, sbsize);
 534                         return XFS_ERROR(EFSCORRUPTED);
 535                 }
 536                 if (((context->flags & ATTR_SECURE) != 0) !=
 537                     ((sfe->flags & XFS_ATTR_SECURE) != 0) &&
 538                     !(context->flags & ATTR_KERNORMALS)) {
 539                         sfe = XFS_ATTR_SF_NEXTENTRY(sfe);
 540                         continue;
 541                 }
 542                 if (((context->flags & ATTR_ROOT) != 0) !=
 543                     ((sfe->flags & XFS_ATTR_ROOT) != 0) &&
 544                     !(context->flags & ATTR_KERNROOTLS)) {
 545                         sfe = XFS_ATTR_SF_NEXTENTRY(sfe);
 546                         continue;
 547                 }
 548                 sbp->entno = i;
 549                 INT_SET(sbp->hash, ARCH_CONVERT,
 550                         xfs_da_hashname((char *)sfe->nameval, sfe->namelen));
 551                 sbp->name = (char *)sfe->nameval;
 552                 sbp->namelen = sfe->namelen;
 553                 /* These are bytes, and both on-disk, don't endian-flip */
 554                 sbp->valuelen = sfe->valuelen;
 555                 sbp->flags = sfe->flags;
 556                 sfe = XFS_ATTR_SF_NEXTENTRY(sfe);
 557                 sbp++;
 558                 nsbuf++;
 559         }
 560
 561         /*
 562          * Sort the entries on hash then entno.
 563          */
 564         qsort(sbuf, nsbuf, sizeof(*sbuf), xfs_attr_shortform_compare);
 565
 566         /*
 567          * Re-find our place IN THE SORTED LIST.
 568          */
 569         count = 0;
 570         cursor->initted = 1;
 571         cursor->blkno = 0;
 572         for (sbp = sbuf, i = 0; i < nsbuf; i++, sbp++) {
 573                 if (INT_GET(sbp->hash, ARCH_CONVERT) == cursor->hashval) {
 574                         if (cursor->offset == count) {
 575                                 break;
 576                         }
 577                         count++;
 578                 } else if (INT_GET(sbp->hash, ARCH_CONVERT) > cursor->hashval) {
 579                         break;
 580                 }
 581         }
 582         if (i == nsbuf) {
 583                 kmem_free(sbuf, sbsize);
 584                 xfs_attr_trace_l_c("blk end", context);
 585                 return(0);
 586         }
 587
 588         /*
 589          * Loop putting entries into the user buffer.
 590          */
 591         for ( ; i < nsbuf; i++, sbp++) {
 592                 attrnames_t     *namesp;
 593
 594                 namesp = (sbp->flags & XFS_ATTR_SECURE) ? &attr_secure :
 595                         ((sbp->flags & XFS_ATTR_ROOT) ? &attr_trusted :
 596                           &attr_user);
 597
 598                 if (cursor->hashval != INT_GET(sbp->hash, ARCH_CONVERT)) {
 599                         cursor->hashval = INT_GET(sbp->hash, ARCH_CONVERT);
 600                         cursor->offset = 0;
 601                 }
 602                 if (context->flags & ATTR_KERNOVAL) {
 603                         ASSERT(context->flags & ATTR_KERNAMELS);
 604                         context->count += namesp->attr_namelen +
 605                                                 sbp->namelen + 1;
 606                 } else {
 607                         if (xfs_attr_put_listent(context, namesp,
 608                                         sbp->name, sbp->namelen,
 609                                         INT_GET(sbp->valuelen, ARCH_CONVERT)))
 610                                 break;
 611                 }
 612                 cursor->offset++;
 613         }
 614
 615         kmem_free(sbuf, sbsize);
 616         xfs_attr_trace_l_c("sf E-O-F", context);
 617         return(0);
 618 }
 619
 620 /*
 621  * Check a leaf attribute block to see if all the entries would fit into
 622  * a shortform attribute list.
 623  */
 624 int
 625 xfs_attr_shortform_allfit(xfs_dabuf_t *bp, xfs_inode_t *dp)
 626 {
 627         xfs_attr_leafblock_t *leaf;
 628         xfs_attr_leaf_entry_t *entry;
 629         xfs_attr_leaf_name_local_t *name_loc;
 630         int bytes, i;
 631
 632         leaf = bp->data;
 633         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
 634                                                 == XFS_ATTR_LEAF_MAGIC);
 635
 636         entry = &leaf->entries[0];
 637         bytes = sizeof(struct xfs_attr_sf_hdr);
 638         for (i = 0; i < INT_GET(leaf->hdr.count, ARCH_CONVERT); entry++, i++) {
 639                 if (entry->flags & XFS_ATTR_INCOMPLETE)
 640                         continue;               /* don't copy partial entries */
 641                 if (!(entry->flags & XFS_ATTR_LOCAL))
 642                         return(0);
 643                 name_loc = XFS_ATTR_LEAF_NAME_LOCAL(leaf, i);
 644                 if (name_loc->namelen >= XFS_ATTR_SF_ENTSIZE_MAX)
 645                         return(0);
 646                 if (INT_GET(name_loc->valuelen, ARCH_CONVERT) >= XFS_ATTR_SF_ENTSIZE_MAX)
 647                         return(0);
 648                 bytes += sizeof(struct xfs_attr_sf_entry)-1
 649                                 + name_loc->namelen
 650                                 + INT_GET(name_loc->valuelen, ARCH_CONVERT);
 651         }
 652         return( bytes < XFS_IFORK_ASIZE(dp) );
 653 }
 654
 655 /*
 656  * Convert a leaf attribute list to shortform attribute list
 657  */
 658 int
 659 xfs_attr_leaf_to_shortform(xfs_dabuf_t *bp, xfs_da_args_t *args)
 660 {
 661         xfs_attr_leafblock_t *leaf;
 662         xfs_attr_leaf_entry_t *entry;
 663         xfs_attr_leaf_name_local_t *name_loc;
 664         xfs_da_args_t nargs;
 665         xfs_inode_t *dp;
 666         char *tmpbuffer;
 667         int error, i;
 668
 669         dp = args->dp;
 670         tmpbuffer = kmem_alloc(XFS_LBSIZE(dp->i_mount), KM_SLEEP);
 671         ASSERT(tmpbuffer != NULL);
 672
 673         ASSERT(bp != NULL);
 674         memcpy(tmpbuffer, bp->data, XFS_LBSIZE(dp->i_mount));
 675         leaf = (xfs_attr_leafblock_t *)tmpbuffer;
 676         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
 677                                                 == XFS_ATTR_LEAF_MAGIC);
 678         memset(bp->data, 0, XFS_LBSIZE(dp->i_mount));
 679
 680         /*
 681          * Clean out the prior contents of the attribute list.
 682          */
 683         error = xfs_da_shrink_inode(args, 0, bp);
 684         if (error)
 685                 goto out;
 686         error = xfs_attr_shortform_create(args);
 687         if (error)
 688                 goto out;
 689
 690         /*
 691          * Copy the attributes
 692          */
 693         memset((char *)&nargs, 0, sizeof(nargs));
 694         nargs.dp = dp;
 695         nargs.firstblock = args->firstblock;
 696         nargs.flist = args->flist;
 697         nargs.total = args->total;
 698         nargs.whichfork = XFS_ATTR_FORK;
 699         nargs.trans = args->trans;
 700         nargs.oknoent = 1;
 701         entry = &leaf->entries[0];
 702         for (i = 0; i < INT_GET(leaf->hdr.count, ARCH_CONVERT); entry++, i++) {
 703                 if (entry->flags & XFS_ATTR_INCOMPLETE)
 704                         continue;       /* don't copy partial entries */
 705                 if (!entry->nameidx)
 706                         continue;
 707                 ASSERT(entry->flags & XFS_ATTR_LOCAL);
 708                 name_loc = XFS_ATTR_LEAF_NAME_LOCAL(leaf, i);
 709                 nargs.name = (char *)name_loc->nameval;
 710                 nargs.namelen = name_loc->namelen;
 711                 nargs.value = (char *)&name_loc->nameval[nargs.namelen];
 712                 nargs.valuelen = INT_GET(name_loc->valuelen, ARCH_CONVERT);
 713                 nargs.hashval = INT_GET(entry->hashval, ARCH_CONVERT);
 714                 nargs.flags = (entry->flags & XFS_ATTR_SECURE) ? ATTR_SECURE :
 715                               ((entry->flags & XFS_ATTR_ROOT) ? ATTR_ROOT : 0);
 716                 xfs_attr_shortform_add(&nargs);
 717         }
 718         error = 0;
 719
 720 out:
 721         kmem_free(tmpbuffer, XFS_LBSIZE(dp->i_mount));
 722         return(error);
 723 }
 724
 725 /*
 726  * Convert from using a single leaf to a root node and a leaf.
 727  */
 728 int
 729 xfs_attr_leaf_to_node(xfs_da_args_t *args)
 730 {
 731         xfs_attr_leafblock_t *leaf;
 732         xfs_da_intnode_t *node;
 733         xfs_inode_t *dp;
 734         xfs_dabuf_t *bp1, *bp2;
 735         xfs_dablk_t blkno;
 736         int error;
 737
 738         dp = args->dp;
 739         bp1 = bp2 = NULL;
 740         error = xfs_da_grow_inode(args, &blkno);
 741         if (error)
 742                 goto out;
 743         error = xfs_da_read_buf(args->trans, args->dp, 0, -1, &bp1,
 744                                              XFS_ATTR_FORK);
 745         if (error)
 746                 goto out;
 747         ASSERT(bp1 != NULL);
 748         bp2 = NULL;
 749         error = xfs_da_get_buf(args->trans, args->dp, blkno, -1, &bp2,
 750                                             XFS_ATTR_FORK);
 751         if (error)
 752                 goto out;
 753         ASSERT(bp2 != NULL);
 754         memcpy(bp2->data, bp1->data, XFS_LBSIZE(dp->i_mount));
 755         xfs_da_buf_done(bp1);
 756         bp1 = NULL;
 757         xfs_da_log_buf(args->trans, bp2, 0, XFS_LBSIZE(dp->i_mount) - 1);
 758
 759         /*
 760          * Set up the new root node.
 761          */
 762         error = xfs_da_node_create(args, 0, 1, &bp1, XFS_ATTR_FORK);
 763         if (error)
 764                 goto out;
 765         node = bp1->data;
 766         leaf = bp2->data;
 767         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
 768                                                 == XFS_ATTR_LEAF_MAGIC);
 769         /* both on-disk, don't endian-flip twice */
 770         node->btree[0].hashval =
 771                 leaf->entries[INT_GET(leaf->hdr.count, ARCH_CONVERT)-1 ].hashval;
 772         INT_SET(node->btree[0].before, ARCH_CONVERT, blkno);
 773         INT_SET(node->hdr.count, ARCH_CONVERT, 1);
 774         xfs_da_log_buf(args->trans, bp1, 0, XFS_LBSIZE(dp->i_mount) - 1);
 775         error = 0;
 776 out:
 777         if (bp1)
 778                 xfs_da_buf_done(bp1);
 779         if (bp2)
 780                 xfs_da_buf_done(bp2);
 781         return(error);
 782 }
 783
 784
 785 /*========================================================================
 786  * Routines used for growing the Btree.
 787  *========================================================================*/
 788
 789 /*
 790  * Create the initial contents of a leaf attribute list
 791  * or a leaf in a node attribute list.
 792  */
 793 STATIC int
 794 xfs_attr_leaf_create(xfs_da_args_t *args, xfs_dablk_t blkno, xfs_dabuf_t **bpp)
 795 {
 796         xfs_attr_leafblock_t *leaf;
 797         xfs_attr_leaf_hdr_t *hdr;
 798         xfs_inode_t *dp;
 799         xfs_dabuf_t *bp;
 800         int error;
 801
 802         dp = args->dp;
 803         ASSERT(dp != NULL);
 804         error = xfs_da_get_buf(args->trans, args->dp, blkno, -1, &bp,
 805                                             XFS_ATTR_FORK);
 806         if (error)
 807                 return(error);
 808         ASSERT(bp != NULL);
 809         leaf = bp->data;
 810         memset((char *)leaf, 0, XFS_LBSIZE(dp->i_mount));
 811         hdr = &leaf->hdr;
 812         INT_SET(hdr->info.magic, ARCH_CONVERT, XFS_ATTR_LEAF_MAGIC);
 813         INT_SET(hdr->firstused, ARCH_CONVERT, XFS_LBSIZE(dp->i_mount));
 814         if (!hdr->firstused) {
 815                 INT_SET(hdr->firstused, ARCH_CONVERT,
 816                         XFS_LBSIZE(dp->i_mount) - XFS_ATTR_LEAF_NAME_ALIGN);
 817         }
 818
 819         INT_SET(hdr->freemap[0].base, ARCH_CONVERT,
 820                                                 sizeof(xfs_attr_leaf_hdr_t));
 821         INT_SET(hdr->freemap[0].size, ARCH_CONVERT,
 822                                           INT_GET(hdr->firstused, ARCH_CONVERT)
 823                                         - INT_GET(hdr->freemap[0].base,
 824                                                                 ARCH_CONVERT));
 825
 826         xfs_da_log_buf(args->trans, bp, 0, XFS_LBSIZE(dp->i_mount) - 1);
 827
 828         *bpp = bp;
 829         return(0);
 830 }
 831
 832 /*
 833  * Split the leaf node, rebalance, then add the new entry.
 834  */
 835 int
 836 xfs_attr_leaf_split(xfs_da_state_t *state, xfs_da_state_blk_t *oldblk,
 837                                    xfs_da_state_blk_t *newblk)
 838 {
 839         xfs_dablk_t blkno;
 840         int error;
 841
 842         /*
 843          * Allocate space for a new leaf node.
 844          */
 845         ASSERT(oldblk->magic == XFS_ATTR_LEAF_MAGIC);
 846         error = xfs_da_grow_inode(state->args, &blkno);
 847         if (error)
 848                 return(error);
 849         error = xfs_attr_leaf_create(state->args, blkno, &newblk->bp);
 850         if (error)
 851                 return(error);
 852         newblk->blkno = blkno;
 853         newblk->magic = XFS_ATTR_LEAF_MAGIC;
 854
 855         /*
 856          * Rebalance the entries across the two leaves.
 857          * NOTE: rebalance() currently depends on the 2nd block being empty.
 858          */
 859         xfs_attr_leaf_rebalance(state, oldblk, newblk);
 860         error = xfs_da_blk_link(state, oldblk, newblk);
 861         if (error)
 862                 return(error);
 863
 864         /*
 865          * Save info on "old" attribute for "atomic rename" ops, leaf_add()
 866          * modifies the index/blkno/rmtblk/rmtblkcnt fields to show the
 867          * "new" attrs info.  Will need the "old" info to remove it later.
 868          *
 869          * Insert the "new" entry in the correct block.
 870          */
 871         if (state->inleaf)
 872                 error = xfs_attr_leaf_add(oldblk->bp, state->args);
 873         else
 874                 error = xfs_attr_leaf_add(newblk->bp, state->args);
 875
 876         /*
 877          * Update last hashval in each block since we added the name.
 878          */
 879         oldblk->hashval = xfs_attr_leaf_lasthash(oldblk->bp, NULL);
 880         newblk->hashval = xfs_attr_leaf_lasthash(newblk->bp, NULL);
 881         return(error);
 882 }
 883
 884 /*
 885  * Add a name to the leaf attribute list structure.
 886  */
 887 int
 888 xfs_attr_leaf_add(xfs_dabuf_t *bp, xfs_da_args_t *args)
 889 {
 890         xfs_attr_leafblock_t *leaf;
 891         xfs_attr_leaf_hdr_t *hdr;
 892         xfs_attr_leaf_map_t *map;
 893         int tablesize, entsize, sum, tmp, i;
 894
 895         leaf = bp->data;
 896         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
 897                                                 == XFS_ATTR_LEAF_MAGIC);
 898         ASSERT((args->index >= 0)
 899                 && (args->index <= INT_GET(leaf->hdr.count, ARCH_CONVERT)));
 900         hdr = &leaf->hdr;
 901         entsize = xfs_attr_leaf_newentsize(args->namelen, args->valuelen,
 902                            args->trans->t_mountp->m_sb.sb_blocksize, NULL);
 903
 904         /*
 905          * Search through freemap for first-fit on new name length.
 906          * (may need to figure in size of entry struct too)
 907          */
 908         tablesize = (INT_GET(hdr->count, ARCH_CONVERT) + 1)
 909                                         * sizeof(xfs_attr_leaf_entry_t)
 910                                         + sizeof(xfs_attr_leaf_hdr_t);
 911         map = &hdr->freemap[XFS_ATTR_LEAF_MAPSIZE-1];
 912         for (sum = 0, i = XFS_ATTR_LEAF_MAPSIZE-1; i >= 0; map--, i--) {
 913                 if (tablesize > INT_GET(hdr->firstused, ARCH_CONVERT)) {
 914                         sum += INT_GET(map->size, ARCH_CONVERT);
 915                         continue;
 916                 }
 917                 if (!map->size)
 918                         continue;       /* no space in this map */
 919                 tmp = entsize;
 920                 if (INT_GET(map->base, ARCH_CONVERT)
 921                                 < INT_GET(hdr->firstused, ARCH_CONVERT))
 922                         tmp += sizeof(xfs_attr_leaf_entry_t);
 923                 if (INT_GET(map->size, ARCH_CONVERT) >= tmp) {
 924                         tmp = xfs_attr_leaf_add_work(bp, args, i);
 925                         return(tmp);
 926                 }
 927                 sum += INT_GET(map->size, ARCH_CONVERT);
 928         }
 929
 930         /*
 931          * If there are no holes in the address space of the block,
 932          * and we don't have enough freespace, then compaction will do us
 933          * no good and we should just give up.
 934          */
 935         if (!hdr->holes && (sum < entsize))
 936                 return(XFS_ERROR(ENOSPC));
 937
 938         /*
 939          * Compact the entries to coalesce free space.
 940          * This may change the hdr->count via dropping INCOMPLETE entries.
 941          */
 942         xfs_attr_leaf_compact(args->trans, bp);
 943
 944         /*
 945          * After compaction, the block is guaranteed to have only one
 946          * free region, in freemap[0].  If it is not big enough, give up.
 947          */
 948         if (INT_GET(hdr->freemap[0].size, ARCH_CONVERT)
 949                                 < (entsize + sizeof(xfs_attr_leaf_entry_t)))
 950                 return(XFS_ERROR(ENOSPC));
 951
 952         return(xfs_attr_leaf_add_work(bp, args, 0));
 953 }
 954
 955 /*
 956  * Add a name to a leaf attribute list structure.
 957  */
 958 STATIC int
 959 xfs_attr_leaf_add_work(xfs_dabuf_t *bp, xfs_da_args_t *args, int mapindex)
 960 {
 961         xfs_attr_leafblock_t *leaf;
 962         xfs_attr_leaf_hdr_t *hdr;
 963         xfs_attr_leaf_entry_t *entry;
 964         xfs_attr_leaf_name_local_t *name_loc;
 965         xfs_attr_leaf_name_remote_t *name_rmt;
 966         xfs_attr_leaf_map_t *map;
 967         xfs_mount_t *mp;
 968         int tmp, i;
 969
 970         leaf = bp->data;
 971         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
 972                                                 == XFS_ATTR_LEAF_MAGIC);
 973         hdr = &leaf->hdr;
 974         ASSERT((mapindex >= 0) && (mapindex < XFS_ATTR_LEAF_MAPSIZE));
 975         ASSERT((args->index >= 0)
 976                 && (args->index <= INT_GET(hdr->count, ARCH_CONVERT)));
 977
 978         /*
 979          * Force open some space in the entry array and fill it in.
 980          */
 981         entry = &leaf->entries[args->index];
 982         if (args->index < INT_GET(hdr->count, ARCH_CONVERT)) {
 983                 tmp  = INT_GET(hdr->count, ARCH_CONVERT) - args->index;
 984                 tmp *= sizeof(xfs_attr_leaf_entry_t);
 985                 memmove((char *)(entry+1), (char *)entry, tmp);
 986                 xfs_da_log_buf(args->trans, bp,
 987                     XFS_DA_LOGRANGE(leaf, entry, tmp + sizeof(*entry)));
 988         }
 989         INT_MOD(hdr->count, ARCH_CONVERT, 1);
 990
 991         /*
 992          * Allocate space for the new string (at the end of the run).
 993          */
 994         map = &hdr->freemap[mapindex];
 995         mp = args->trans->t_mountp;
 996         ASSERT(INT_GET(map->base, ARCH_CONVERT) < XFS_LBSIZE(mp));
 997         ASSERT((INT_GET(map->base, ARCH_CONVERT) & 0x3) == 0);
 998         ASSERT(INT_GET(map->size, ARCH_CONVERT) >=
 999                 xfs_attr_leaf_newentsize(args->namelen, args->valuelen,
1000                                          mp->m_sb.sb_blocksize, NULL));
1001         ASSERT(INT_GET(map->size, ARCH_CONVERT) < XFS_LBSIZE(mp));
1002         ASSERT((INT_GET(map->size, ARCH_CONVERT) & 0x3) == 0);
1003         INT_MOD(map->size, ARCH_CONVERT,
1004                 -xfs_attr_leaf_newentsize(args->namelen, args->valuelen,
1005                                           mp->m_sb.sb_blocksize, &tmp));
1006         INT_SET(entry->nameidx, ARCH_CONVERT,
1007                                         INT_GET(map->base, ARCH_CONVERT)
1008                                       + INT_GET(map->size, ARCH_CONVERT));
1009         INT_SET(entry->hashval, ARCH_CONVERT, args->hashval);
1010         entry->flags = tmp ? XFS_ATTR_LOCAL : 0;
1011         entry->flags |= (args->flags & ATTR_SECURE) ? XFS_ATTR_SECURE :
1012                         ((args->flags & ATTR_ROOT) ? XFS_ATTR_ROOT : 0);
1013         if (args->rename) {
1014                 entry->flags |= XFS_ATTR_INCOMPLETE;
1015                 if ((args->blkno2 == args->blkno) &&
1016                     (args->index2 <= args->index)) {
1017                         args->index2++;
1018                 }
1019         }
1020         xfs_da_log_buf(args->trans, bp,
1021                           XFS_DA_LOGRANGE(leaf, entry, sizeof(*entry)));
1022         ASSERT((args->index == 0) || (INT_GET(entry->hashval, ARCH_CONVERT)
1023                                                 >= INT_GET((entry-1)->hashval,
1024                                                             ARCH_CONVERT)));
1025         ASSERT((args->index == INT_GET(hdr->count, ARCH_CONVERT)-1) ||
1026                (INT_GET(entry->hashval, ARCH_CONVERT)
1027                             <= (INT_GET((entry+1)->hashval, ARCH_CONVERT))));
1028
1029         /*
1030          * Copy the attribute name and value into the new space.
1031          *
1032          * For "remote" attribute values, simply note that we need to
1033          * allocate space for the "remote" value.  We can't actually
1034          * allocate the extents in this transaction, and we can't decide
1035          * which blocks they should be as we might allocate more blocks
1036          * as part of this transaction (a split operation for example).
1037          */
1038         if (entry->flags & XFS_ATTR_LOCAL) {
1039                 name_loc = XFS_ATTR_LEAF_NAME_LOCAL(leaf, args->index);
1040                 name_loc->namelen = args->namelen;
1041                 INT_SET(name_loc->valuelen, ARCH_CONVERT, args->valuelen);
1042                 memcpy((char *)name_loc->nameval, args->name, args->namelen);
1043                 memcpy((char *)&name_loc->nameval[args->namelen], args->value,
1044                                    INT_GET(name_loc->valuelen, ARCH_CONVERT));
1045         } else {
1046                 name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf, args->index);
1047                 name_rmt->namelen = args->namelen;
1048                 memcpy((char *)name_rmt->name, args->name, args->namelen);
1049                 entry->flags |= XFS_ATTR_INCOMPLETE;
1050                 /* just in case */
1051                 name_rmt->valuelen = 0;
1052                 name_rmt->valueblk = 0;
1053                 args->rmtblkno = 1;
1054                 args->rmtblkcnt = XFS_B_TO_FSB(mp, args->valuelen);
1055         }
1056         xfs_da_log_buf(args->trans, bp,
1057              XFS_DA_LOGRANGE(leaf, XFS_ATTR_LEAF_NAME(leaf, args->index),
1058                                    xfs_attr_leaf_entsize(leaf, args->index)));
1059
1060         /*
1061          * Update the control info for this leaf node
1062          */
1063         if (INT_GET(entry->nameidx, ARCH_CONVERT)
1064                                 < INT_GET(hdr->firstused, ARCH_CONVERT)) {
1065                 /* both on-disk, don't endian-flip twice */
1066                 hdr->firstused = entry->nameidx;
1067         }
1068         ASSERT(INT_GET(hdr->firstused, ARCH_CONVERT)
1069                                 >= ((INT_GET(hdr->count, ARCH_CONVERT)
1070                                         * sizeof(*entry))+sizeof(*hdr)));
1071         tmp = (INT_GET(hdr->count, ARCH_CONVERT)-1)
1072                                         * sizeof(xfs_attr_leaf_entry_t)
1073                                         + sizeof(xfs_attr_leaf_hdr_t);
1074         map = &hdr->freemap[0];
1075         for (i = 0; i < XFS_ATTR_LEAF_MAPSIZE; map++, i++) {
1076                 if (INT_GET(map->base, ARCH_CONVERT) == tmp) {
1077                         INT_MOD(map->base, ARCH_CONVERT,
1078                                         sizeof(xfs_attr_leaf_entry_t));
1079                         INT_MOD(map->size, ARCH_CONVERT,
1080                                         -sizeof(xfs_attr_leaf_entry_t));
1081                 }
1082         }
1083         INT_MOD(hdr->usedbytes, ARCH_CONVERT,
1084                                 xfs_attr_leaf_entsize(leaf, args->index));
1085         xfs_da_log_buf(args->trans, bp,
1086                 XFS_DA_LOGRANGE(leaf, hdr, sizeof(*hdr)));
1087         return(0);
1088 }
1089
1090 /*
1091  * Garbage collect a leaf attribute list block by copying it to a new buffer.
1092  */
1093 STATIC void
1094 xfs_attr_leaf_compact(xfs_trans_t *trans, xfs_dabuf_t *bp)
1095 {
1096         xfs_attr_leafblock_t *leaf_s, *leaf_d;
1097         xfs_attr_leaf_hdr_t *hdr_s, *hdr_d;
1098         xfs_mount_t *mp;
1099         char *tmpbuffer;
1100
1101         mp = trans->t_mountp;
1102         tmpbuffer = kmem_alloc(XFS_LBSIZE(mp), KM_SLEEP);
1103         ASSERT(tmpbuffer != NULL);
1104         memcpy(tmpbuffer, bp->data, XFS_LBSIZE(mp));
1105         memset(bp->data, 0, XFS_LBSIZE(mp));
1106
1107         /*
1108          * Copy basic information
1109          */
1110         leaf_s = (xfs_attr_leafblock_t *)tmpbuffer;
1111         leaf_d = bp->data;
1112         hdr_s = &leaf_s->hdr;
1113         hdr_d = &leaf_d->hdr;
1114         hdr_d->info = hdr_s->info;      /* struct copy */
1115         INT_SET(hdr_d->firstused, ARCH_CONVERT, XFS_LBSIZE(mp));
1116         /* handle truncation gracefully */
1117         if (!hdr_d->firstused) {
1118                 INT_SET(hdr_d->firstused, ARCH_CONVERT,
1119                                 XFS_LBSIZE(mp) - XFS_ATTR_LEAF_NAME_ALIGN);
1120         }
1121         hdr_d->usedbytes = 0;
1122         hdr_d->count = 0;
1123         hdr_d->holes = 0;
1124         INT_SET(hdr_d->freemap[0].base, ARCH_CONVERT,
1125                                         sizeof(xfs_attr_leaf_hdr_t));
1126         INT_SET(hdr_d->freemap[0].size, ARCH_CONVERT,
1127                                 INT_GET(hdr_d->firstused, ARCH_CONVERT)
1128                               - INT_GET(hdr_d->freemap[0].base, ARCH_CONVERT));
1129
1130         /*
1131          * Copy all entry's in the same (sorted) order,
1132          * but allocate name/value pairs packed and in sequence.
1133          */
1134         xfs_attr_leaf_moveents(leaf_s, 0, leaf_d, 0,
1135                                 (int)INT_GET(hdr_s->count, ARCH_CONVERT), mp);
1136
1137         xfs_da_log_buf(trans, bp, 0, XFS_LBSIZE(mp) - 1);
1138
1139         kmem_free(tmpbuffer, XFS_LBSIZE(mp));
1140 }
1141
1142 /*
1143  * Redistribute the attribute list entries between two leaf nodes,
1144  * taking into account the size of the new entry.
1145  *
1146  * NOTE: if new block is empty, then it will get the upper half of the
1147  * old block.  At present, all (one) callers pass in an empty second block.
1148  *
1149  * This code adjusts the args->index/blkno and args->index2/blkno2 fields
1150  * to match what it is doing in splitting the attribute leaf block.  Those
1151  * values are used in "atomic rename" operations on attributes.  Note that
1152  * the "new" and "old" values can end up in different blocks.
1153  */
1154 STATIC void
1155 xfs_attr_leaf_rebalance(xfs_da_state_t *state, xfs_da_state_blk_t *blk1,
1156                                        xfs_da_state_blk_t *blk2)
1157 {
1158         xfs_da_args_t *args;
1159         xfs_da_state_blk_t *tmp_blk;
1160         xfs_attr_leafblock_t *leaf1, *leaf2;
1161         xfs_attr_leaf_hdr_t *hdr1, *hdr2;
1162         int count, totallen, max, space, swap;
1163
1164         /*
1165          * Set up environment.
1166          */
1167         ASSERT(blk1->magic == XFS_ATTR_LEAF_MAGIC);
1168         ASSERT(blk2->magic == XFS_ATTR_LEAF_MAGIC);
1169         leaf1 = blk1->bp->data;
1170         leaf2 = blk2->bp->data;
1171         ASSERT(INT_GET(leaf1->hdr.info.magic, ARCH_CONVERT)
1172                                                 == XFS_ATTR_LEAF_MAGIC);
1173         ASSERT(INT_GET(leaf2->hdr.info.magic, ARCH_CONVERT)
1174                                                 == XFS_ATTR_LEAF_MAGIC);
1175         args = state->args;
1176
1177         /*
1178          * Check ordering of blocks, reverse if it makes things simpler.
1179          *
1180          * NOTE: Given that all (current) callers pass in an empty
1181          * second block, this code should never set "swap".
1182          */
1183         swap = 0;
1184         if (xfs_attr_leaf_order(blk1->bp, blk2->bp)) {
1185                 tmp_blk = blk1;
1186                 blk1 = blk2;
1187                 blk2 = tmp_blk;
1188                 leaf1 = blk1->bp->data;
1189                 leaf2 = blk2->bp->data;
1190                 swap = 1;
1191         }
1192         hdr1 = &leaf1->hdr;
1193         hdr2 = &leaf2->hdr;
1194
1195         /*
1196          * Examine entries until we reduce the absolute difference in
1197          * byte usage between the two blocks to a minimum.  Then get
1198          * the direction to copy and the number of elements to move.
1199          *
1200          * "inleaf" is true if the new entry should be inserted into blk1.
1201          * If "swap" is also true, then reverse the sense of "inleaf".
1202          */
1203         state->inleaf = xfs_attr_leaf_figure_balance(state, blk1, blk2,
1204                                                             &count, &totallen);
1205         if (swap)
1206                 state->inleaf = !state->inleaf;
1207
1208         /*
1209          * Move any entries required from leaf to leaf:
1210          */
1211         if (count < INT_GET(hdr1->count, ARCH_CONVERT)) {
1212                 /*
1213                  * Figure the total bytes to be added to the destination leaf.
1214                  */
1215                 /* number entries being moved */
1216                 count = INT_GET(hdr1->count, ARCH_CONVERT) - count;
1217                 space  = INT_GET(hdr1->usedbytes, ARCH_CONVERT) - totallen;
1218                 space += count * sizeof(xfs_attr_leaf_entry_t);
1219
1220                 /*
1221                  * leaf2 is the destination, compact it if it looks tight.
1222                  */
1223                 max  = INT_GET(hdr2->firstused, ARCH_CONVERT)
1224                                                 - sizeof(xfs_attr_leaf_hdr_t);
1225                 max -= INT_GET(hdr2->count, ARCH_CONVERT)
1226                                         * sizeof(xfs_attr_leaf_entry_t);
1227                 if (space > max) {
1228                         xfs_attr_leaf_compact(args->trans, blk2->bp);
1229                 }
1230
1231                 /*
1232                  * Move high entries from leaf1 to low end of leaf2.
1233                  */
1234                 xfs_attr_leaf_moveents(leaf1,
1235                                 INT_GET(hdr1->count, ARCH_CONVERT)-count,
1236                                 leaf2, 0, count, state->mp);
1237
1238                 xfs_da_log_buf(args->trans, blk1->bp, 0, state->blocksize-1);
1239                 xfs_da_log_buf(args->trans, blk2->bp, 0, state->blocksize-1);
1240         } else if (count > INT_GET(hdr1->count, ARCH_CONVERT)) {
1241                 /*
1242                  * I assert that since all callers pass in an empty
1243                  * second buffer, this code should never execute.
1244                  */
1245
1246                 /*
1247                  * Figure the total bytes to be added to the destination leaf.
1248                  */
1249                 /* number entries being moved */
1250                 count -= INT_GET(hdr1->count, ARCH_CONVERT);
1251                 space  = totallen - INT_GET(hdr1->usedbytes, ARCH_CONVERT);
1252                 space += count * sizeof(xfs_attr_leaf_entry_t);
1253
1254                 /*
1255                  * leaf1 is the destination, compact it if it looks tight.
1256                  */
1257                 max  = INT_GET(hdr1->firstused, ARCH_CONVERT)
1258                                                 - sizeof(xfs_attr_leaf_hdr_t);
1259                 max -= INT_GET(hdr1->count, ARCH_CONVERT)
1260                                         * sizeof(xfs_attr_leaf_entry_t);
1261                 if (space > max) {
1262                         xfs_attr_leaf_compact(args->trans, blk1->bp);
1263                 }
1264
1265                 /*
1266                  * Move low entries from leaf2 to high end of leaf1.
1267                  */
1268                 xfs_attr_leaf_moveents(leaf2, 0, leaf1,
1269                                 (int)INT_GET(hdr1->count, ARCH_CONVERT), count,
1270                                 state->mp);
1271
1272                 xfs_da_log_buf(args->trans, blk1->bp, 0, state->blocksize-1);
1273                 xfs_da_log_buf(args->trans, blk2->bp, 0, state->blocksize-1);
1274         }
1275
1276         /*
1277          * Copy out last hashval in each block for B-tree code.
1278          */
1279         blk1->hashval =
1280             INT_GET(leaf1->entries[INT_GET(leaf1->hdr.count,
1281                                     ARCH_CONVERT)-1].hashval, ARCH_CONVERT);
1282         blk2->hashval =
1283             INT_GET(leaf2->entries[INT_GET(leaf2->hdr.count,
1284                                     ARCH_CONVERT)-1].hashval, ARCH_CONVERT);
1285
1286         /*
1287          * Adjust the expected index for insertion.
1288          * NOTE: this code depends on the (current) situation that the
1289          * second block was originally empty.
1290          *
1291          * If the insertion point moved to the 2nd block, we must adjust
1292          * the index.  We must also track the entry just following the
1293          * new entry for use in an "atomic rename" operation, that entry
1294          * is always the "old" entry and the "new" entry is what we are
1295          * inserting.  The index/blkno fields refer to the "old" entry,
1296          * while the index2/blkno2 fields refer to the "new" entry.
1297          */
1298         if (blk1->index > INT_GET(leaf1->hdr.count, ARCH_CONVERT)) {
1299                 ASSERT(state->inleaf == 0);
1300                 blk2->index = blk1->index
1301                                 - INT_GET(leaf1->hdr.count, ARCH_CONVERT);
1302                 args->index = args->index2 = blk2->index;
1303                 args->blkno = args->blkno2 = blk2->blkno;
1304         } else if (blk1->index == INT_GET(leaf1->hdr.count, ARCH_CONVERT)) {
1305                 if (state->inleaf) {
1306                         args->index = blk1->index;
1307                         args->blkno = blk1->blkno;
1308                         args->index2 = 0;
1309                         args->blkno2 = blk2->blkno;
1310                 } else {
1311                         blk2->index = blk1->index
1312                                     - INT_GET(leaf1->hdr.count, ARCH_CONVERT);
1313                         args->index = args->index2 = blk2->index;
1314                         args->blkno = args->blkno2 = blk2->blkno;
1315                 }
1316         } else {
1317                 ASSERT(state->inleaf == 1);
1318                 args->index = args->index2 = blk1->index;
1319                 args->blkno = args->blkno2 = blk1->blkno;
1320         }
1321 }
1322
1323 /*
1324  * Examine entries until we reduce the absolute difference in
1325  * byte usage between the two blocks to a minimum.
1326  * GROT: Is this really necessary?  With other than a 512 byte blocksize,
1327  * GROT: there will always be enough room in either block for a new entry.
1328  * GROT: Do a double-split for this case?
1329  */
1330 STATIC int
1331 xfs_attr_leaf_figure_balance(xfs_da_state_t *state,
1332                                     xfs_da_state_blk_t *blk1,
1333                                     xfs_da_state_blk_t *blk2,
1334                                     int *countarg, int *usedbytesarg)
1335 {
1336         xfs_attr_leafblock_t *leaf1, *leaf2;
1337         xfs_attr_leaf_hdr_t *hdr1, *hdr2;
1338         xfs_attr_leaf_entry_t *entry;
1339         int count, max, index, totallen, half;
1340         int lastdelta, foundit, tmp;
1341
1342         /*
1343          * Set up environment.
1344          */
1345         leaf1 = blk1->bp->data;
1346         leaf2 = blk2->bp->data;
1347         hdr1 = &leaf1->hdr;
1348         hdr2 = &leaf2->hdr;
1349         foundit = 0;
1350         totallen = 0;
1351
1352         /*
1353          * Examine entries until we reduce the absolute difference in
1354          * byte usage between the two blocks to a minimum.
1355          */
1356         max = INT_GET(hdr1->count, ARCH_CONVERT)
1357                         + INT_GET(hdr2->count, ARCH_CONVERT);
1358         half  = (max+1) * sizeof(*entry);
1359         half += INT_GET(hdr1->usedbytes, ARCH_CONVERT)
1360                                 + INT_GET(hdr2->usedbytes, ARCH_CONVERT)
1361                                 + xfs_attr_leaf_newentsize(
1362                                                 state->args->namelen,
1363                                                 state->args->valuelen,
1364                                                 state->blocksize, NULL);
1365         half /= 2;
1366         lastdelta = state->blocksize;
1367         entry = &leaf1->entries[0];
1368         for (count = index = 0; count < max; entry++, index++, count++) {
1369
1370 #define XFS_ATTR_ABS(A) (((A) < 0) ? -(A) : (A))
1371                 /*
1372                  * The new entry is in the first block, account for it.
1373                  */
1374                 if (count == blk1->index) {
1375                         tmp = totallen + sizeof(*entry) +
1376                                 xfs_attr_leaf_newentsize(
1377                                                 state->args->namelen,
1378                                                 state->args->valuelen,
1379                                                 state->blocksize, NULL);
1380                         if (XFS_ATTR_ABS(half - tmp) > lastdelta)
1381                                 break;
1382                         lastdelta = XFS_ATTR_ABS(half - tmp);
1383                         totallen = tmp;
1384                         foundit = 1;
1385                 }
1386
1387                 /*
1388                  * Wrap around into the second block if necessary.
1389                  */
1390                 if (count == INT_GET(hdr1->count, ARCH_CONVERT)) {
1391                         leaf1 = leaf2;
1392                         entry = &leaf1->entries[0];
1393                         index = 0;
1394                 }
1395
1396                 /*
1397                  * Figure out if next leaf entry would be too much.
1398                  */
1399                 tmp = totallen + sizeof(*entry) + xfs_attr_leaf_entsize(leaf1,
1400                                                                         index);
1401                 if (XFS_ATTR_ABS(half - tmp) > lastdelta)
1402                         break;
1403                 lastdelta = XFS_ATTR_ABS(half - tmp);
1404                 totallen = tmp;
1405 #undef XFS_ATTR_ABS
1406         }
1407
1408         /*
1409          * Calculate the number of usedbytes that will end up in lower block.
1410          * If new entry not in lower block, fix up the count.
1411          */
1412         totallen -= count * sizeof(*entry);
1413         if (foundit) {
1414                 totallen -= sizeof(*entry) +
1415                                 xfs_attr_leaf_newentsize(
1416                                                 state->args->namelen,
1417                                                 state->args->valuelen,
1418                                                 state->blocksize, NULL);
1419         }
1420
1421         *countarg = count;
1422         *usedbytesarg = totallen;
1423         return(foundit);
1424 }
1425
1426 /*========================================================================
1427  * Routines used for shrinking the Btree.
1428  *========================================================================*/
1429
1430 /*
1431  * Check a leaf block and its neighbors to see if the block should be
1432  * collapsed into one or the other neighbor.  Always keep the block
1433  * with the smaller block number.
1434  * If the current block is over 50% full, don't try to join it, return 0.
1435  * If the block is empty, fill in the state structure and return 2.
1436  * If it can be collapsed, fill in the state structure and return 1.
1437  * If nothing can be done, return 0.
1438  *
1439  * GROT: allow for INCOMPLETE entries in calculation.
1440  */
1441 int
1442 xfs_attr_leaf_toosmall(xfs_da_state_t *state, int *action)
1443 {
1444         xfs_attr_leafblock_t *leaf;
1445         xfs_da_state_blk_t *blk;
1446         xfs_da_blkinfo_t *info;
1447         int count, bytes, forward, error, retval, i;
1448         xfs_dablk_t blkno;
1449         xfs_dabuf_t *bp;
1450
1451         /*
1452          * Check for the degenerate case of the block being over 50% full.
1453          * If so, it's not worth even looking to see if we might be able
1454          * to coalesce with a sibling.
1455          */
1456         blk = &state->path.blk[ state->path.active-1 ];
1457         info = blk->bp->data;
1458         ASSERT(INT_GET(info->magic, ARCH_CONVERT) == XFS_ATTR_LEAF_MAGIC);
1459         leaf = (xfs_attr_leafblock_t *)info;
1460         count = INT_GET(leaf->hdr.count, ARCH_CONVERT);
1461         bytes = sizeof(xfs_attr_leaf_hdr_t) +
1462                 count * sizeof(xfs_attr_leaf_entry_t) +
1463                 INT_GET(leaf->hdr.usedbytes, ARCH_CONVERT);
1464         if (bytes > (state->blocksize >> 1)) {
1465                 *action = 0;    /* blk over 50%, don't try to join */
1466                 return(0);
1467         }
1468
1469         /*
1470          * Check for the degenerate case of the block being empty.
1471          * If the block is empty, we'll simply delete it, no need to
1472          * coalesce it with a sibling block.  We choose (aribtrarily)
1473          * to merge with the forward block unless it is NULL.
1474          */
1475         if (count == 0) {
1476                 /*
1477                  * Make altpath point to the block we want to keep and
1478                  * path point to the block we want to drop (this one).
1479                  */
1480                 forward = info->forw;
1481                 memcpy(&state->altpath, &state->path, sizeof(state->path));
1482                 error = xfs_da_path_shift(state, &state->altpath, forward,
1483                                                  0, &retval);
1484                 if (error)
1485                         return(error);
1486                 if (retval) {
1487                         *action = 0;
1488                 } else {
1489                         *action = 2;
1490                 }
1491                 return(0);
1492         }
1493
1494         /*
1495          * Examine each sibling block to see if we can coalesce with
1496          * at least 25% free space to spare.  We need to figure out
1497          * whether to merge with the forward or the backward block.
1498          * We prefer coalescing with the lower numbered sibling so as
1499          * to shrink an attribute list over time.
1500          */
1501         /* start with smaller blk num */
1502         forward = (INT_GET(info->forw, ARCH_CONVERT)
1503                                         < INT_GET(info->back, ARCH_CONVERT));
1504         for (i = 0; i < 2; forward = !forward, i++) {
1505                 if (forward)
1506                         blkno = INT_GET(info->forw, ARCH_CONVERT);
1507                 else
1508                         blkno = INT_GET(info->back, ARCH_CONVERT);
1509                 if (blkno == 0)
1510                         continue;
1511                 error = xfs_da_read_buf(state->args->trans, state->args->dp,
1512                                         blkno, -1, &bp, XFS_ATTR_FORK);
1513                 if (error)
1514                         return(error);
1515                 ASSERT(bp != NULL);
1516
1517                 leaf = (xfs_attr_leafblock_t *)info;
1518                 count  = INT_GET(leaf->hdr.count, ARCH_CONVERT);
1519                 bytes  = state->blocksize - (state->blocksize>>2);
1520                 bytes -= INT_GET(leaf->hdr.usedbytes, ARCH_CONVERT);
1521                 leaf = bp->data;
1522                 ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
1523                                                 == XFS_ATTR_LEAF_MAGIC);
1524                 count += INT_GET(leaf->hdr.count, ARCH_CONVERT);
1525                 bytes -= INT_GET(leaf->hdr.usedbytes, ARCH_CONVERT);
1526                 bytes -= count * sizeof(xfs_attr_leaf_entry_t);
1527                 bytes -= sizeof(xfs_attr_leaf_hdr_t);
1528                 xfs_da_brelse(state->args->trans, bp);
1529                 if (bytes >= 0)
1530                         break;  /* fits with at least 25% to spare */
1531         }
1532         if (i >= 2) {
1533                 *action = 0;
1534                 return(0);
1535         }
1536
1537         /*
1538          * Make altpath point to the block we want to keep (the lower
1539          * numbered block) and path point to the block we want to drop.
1540          */
1541         memcpy(&state->altpath, &state->path, sizeof(state->path));
1542         if (blkno < blk->blkno) {
1543                 error = xfs_da_path_shift(state, &state->altpath, forward,
1544                                                  0, &retval);
1545         } else {
1546                 error = xfs_da_path_shift(state, &state->path, forward,
1547                                                  0, &retval);
1548         }
1549         if (error)
1550                 return(error);
1551         if (retval) {
1552                 *action = 0;
1553         } else {
1554                 *action = 1;
1555         }
1556         return(0);
1557 }
1558
1559 /*
1560  * Remove a name from the leaf attribute list structure.
1561  *
1562  * Return 1 if leaf is less than 37% full, 0 if >= 37% full.
1563  * If two leaves are 37% full, when combined they will leave 25% free.
1564  */
1565 int
1566 xfs_attr_leaf_remove(xfs_dabuf_t *bp, xfs_da_args_t *args)
1567 {
1568         xfs_attr_leafblock_t *leaf;
1569         xfs_attr_leaf_hdr_t *hdr;
1570         xfs_attr_leaf_map_t *map;
1571         xfs_attr_leaf_entry_t *entry;
1572         int before, after, smallest, entsize;
1573         int tablesize, tmp, i;
1574         xfs_mount_t *mp;
1575
1576         leaf = bp->data;
1577         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
1578                                                 == XFS_ATTR_LEAF_MAGIC);
1579         hdr = &leaf->hdr;
1580         mp = args->trans->t_mountp;
1581         ASSERT((INT_GET(hdr->count, ARCH_CONVERT) > 0)
1582                 && (INT_GET(hdr->count, ARCH_CONVERT) < (XFS_LBSIZE(mp)/8)));
1583         ASSERT((args->index >= 0)
1584                 && (args->index < INT_GET(hdr->count, ARCH_CONVERT)));
1585         ASSERT(INT_GET(hdr->firstused, ARCH_CONVERT)
1586                                 >= ((INT_GET(hdr->count, ARCH_CONVERT)
1587                                         * sizeof(*entry))+sizeof(*hdr)));
1588         entry = &leaf->entries[args->index];
1589         ASSERT(INT_GET(entry->nameidx, ARCH_CONVERT)
1590                                 >= INT_GET(hdr->firstused, ARCH_CONVERT));
1591         ASSERT(INT_GET(entry->nameidx, ARCH_CONVERT) < XFS_LBSIZE(mp));
1592
1593         /*
1594          * Scan through free region table:
1595          *    check for adjacency of free'd entry with an existing one,
1596          *    find smallest free region in case we need to replace it,
1597          *    adjust any map that borders the entry table,
1598          */
1599         tablesize = INT_GET(hdr->count, ARCH_CONVERT)
1600                                         * sizeof(xfs_attr_leaf_entry_t)
1601                                         + sizeof(xfs_attr_leaf_hdr_t);
1602         map = &hdr->freemap[0];
1603         tmp = INT_GET(map->size, ARCH_CONVERT);
1604         before = after = -1;
1605         smallest = XFS_ATTR_LEAF_MAPSIZE - 1;
1606         entsize = xfs_attr_leaf_entsize(leaf, args->index);
1607         for (i = 0; i < XFS_ATTR_LEAF_MAPSIZE; map++, i++) {
1608                 ASSERT(INT_GET(map->base, ARCH_CONVERT) < XFS_LBSIZE(mp));
1609                 ASSERT(INT_GET(map->size, ARCH_CONVERT) < XFS_LBSIZE(mp));
1610                 if (INT_GET(map->base, ARCH_CONVERT) == tablesize) {
1611                         INT_MOD(map->base, ARCH_CONVERT,
1612                                         -sizeof(xfs_attr_leaf_entry_t));
1613                         INT_MOD(map->size, ARCH_CONVERT,
1614                                         sizeof(xfs_attr_leaf_entry_t));
1615                 }
1616
1617                 if ((INT_GET(map->base, ARCH_CONVERT)
1618                                         + INT_GET(map->size, ARCH_CONVERT))
1619                                 == INT_GET(entry->nameidx, ARCH_CONVERT)) {
1620                         before = i;
1621                 } else if (INT_GET(map->base, ARCH_CONVERT)
1622                         == (INT_GET(entry->nameidx, ARCH_CONVERT) + entsize)) {
1623                         after = i;
1624                 } else if (INT_GET(map->size, ARCH_CONVERT) < tmp) {
1625                         tmp = INT_GET(map->size, ARCH_CONVERT);
1626                         smallest = i;
1627                 }
1628         }
1629
1630         /*
1631          * Coalesce adjacent freemap regions,
1632          * or replace the smallest region.
1633          */
1634         if ((before >= 0) || (after >= 0)) {
1635                 if ((before >= 0) && (after >= 0)) {
1636                         map = &hdr->freemap[before];
1637                         INT_MOD(map->size, ARCH_CONVERT, entsize);
1638                         INT_MOD(map->size, ARCH_CONVERT,
1639                                 INT_GET(hdr->freemap[after].size,
1640                                                         ARCH_CONVERT));
1641                         hdr->freemap[after].base = 0;
1642                         hdr->freemap[after].size = 0;
1643                 } else if (before >= 0) {
1644                         map = &hdr->freemap[before];
1645                         INT_MOD(map->size, ARCH_CONVERT, entsize);
1646                 } else {
1647                         map = &hdr->freemap[after];
1648                         /* both on-disk, don't endian flip twice */
1649                         map->base = entry->nameidx;
1650                         INT_MOD(map->size, ARCH_CONVERT, entsize);
1651                 }
1652         } else {
1653                 /*
1654                  * Replace smallest region (if it is smaller than free'd entry)
1655                  */
1656                 map = &hdr->freemap[smallest];
1657                 if (INT_GET(map->size, ARCH_CONVERT) < entsize) {
1658                         INT_SET(map->base, ARCH_CONVERT,
1659                                         INT_GET(entry->nameidx, ARCH_CONVERT));
1660                         INT_SET(map->size, ARCH_CONVERT, entsize);
1661                 }
1662         }
1663
1664         /*
1665          * Did we remove the first entry?
1666          */
1667         if (INT_GET(entry->nameidx, ARCH_CONVERT)
1668                                 == INT_GET(hdr->firstused, ARCH_CONVERT))
1669                 smallest = 1;
1670         else
1671                 smallest = 0;
1672
1673         /*
1674          * Compress the remaining entries and zero out the removed stuff.
1675          */
1676         memset(XFS_ATTR_LEAF_NAME(leaf, args->index), 0, entsize);
1677         INT_MOD(hdr->usedbytes, ARCH_CONVERT, -entsize);
1678         xfs_da_log_buf(args->trans, bp,
1679              XFS_DA_LOGRANGE(leaf, XFS_ATTR_LEAF_NAME(leaf, args->index),
1680                                    entsize));
1681
1682         tmp = (INT_GET(hdr->count, ARCH_CONVERT) - args->index)
1683                                         * sizeof(xfs_attr_leaf_entry_t);
1684         memmove((char *)entry, (char *)(entry+1), tmp);
1685         INT_MOD(hdr->count, ARCH_CONVERT, -1);
1686         xfs_da_log_buf(args->trans, bp,
1687             XFS_DA_LOGRANGE(leaf, entry, tmp + sizeof(*entry)));
1688         entry = &leaf->entries[INT_GET(hdr->count, ARCH_CONVERT)];
1689         memset((char *)entry, 0, sizeof(xfs_attr_leaf_entry_t));
1690
1691         /*
1692          * If we removed the first entry, re-find the first used byte
1693          * in the name area.  Note that if the entry was the "firstused",
1694          * then we don't have a "hole" in our block resulting from
1695          * removing the name.
1696          */
1697         if (smallest) {
1698                 tmp = XFS_LBSIZE(mp);
1699                 entry = &leaf->entries[0];
1700                 for (i = INT_GET(hdr->count, ARCH_CONVERT)-1;
1701                                                 i >= 0; entry++, i--) {
1702                         ASSERT(INT_GET(entry->nameidx, ARCH_CONVERT)
1703                                 >= INT_GET(hdr->firstused, ARCH_CONVERT));
1704                         ASSERT(INT_GET(entry->nameidx, ARCH_CONVERT)
1705                                                         < XFS_LBSIZE(mp));
1706                         if (INT_GET(entry->nameidx, ARCH_CONVERT) < tmp)
1707                                 tmp = INT_GET(entry->nameidx, ARCH_CONVERT);
1708                 }
1709                 INT_SET(hdr->firstused, ARCH_CONVERT, tmp);
1710                 if (!hdr->firstused) {
1711                         INT_SET(hdr->firstused, ARCH_CONVERT,
1712                                         tmp - XFS_ATTR_LEAF_NAME_ALIGN);
1713                 }
1714         } else {
1715                 hdr->holes = 1;         /* mark as needing compaction */
1716         }
1717         xfs_da_log_buf(args->trans, bp,
1718                           XFS_DA_LOGRANGE(leaf, hdr, sizeof(*hdr)));
1719
1720         /*
1721          * Check if leaf is less than 50% full, caller may want to
1722          * "join" the leaf with a sibling if so.
1723          */
1724         tmp  = sizeof(xfs_attr_leaf_hdr_t);
1725         tmp += INT_GET(leaf->hdr.count, ARCH_CONVERT)
1726                                         * sizeof(xfs_attr_leaf_entry_t);
1727         tmp += INT_GET(leaf->hdr.usedbytes, ARCH_CONVERT);
1728         return(tmp < mp->m_attr_magicpct); /* leaf is < 37% full */
1729 }
1730
1731 /*
1732  * Move all the attribute list entries from drop_leaf into save_leaf.
1733  */
1734 void
1735 xfs_attr_leaf_unbalance(xfs_da_state_t *state, xfs_da_state_blk_t *drop_blk,
1736                                        xfs_da_state_blk_t *save_blk)
1737 {
1738         xfs_attr_leafblock_t *drop_leaf, *save_leaf, *tmp_leaf;
1739         xfs_attr_leaf_hdr_t *drop_hdr, *save_hdr, *tmp_hdr;
1740         xfs_mount_t *mp;
1741         char *tmpbuffer;
1742
1743         /*
1744          * Set up environment.
1745          */
1746         mp = state->mp;
1747         ASSERT(drop_blk->magic == XFS_ATTR_LEAF_MAGIC);
1748         ASSERT(save_blk->magic == XFS_ATTR_LEAF_MAGIC);
1749         drop_leaf = drop_blk->bp->data;
1750         save_leaf = save_blk->bp->data;
1751         ASSERT(INT_GET(drop_leaf->hdr.info.magic, ARCH_CONVERT)
1752                                                 == XFS_ATTR_LEAF_MAGIC);
1753         ASSERT(INT_GET(save_leaf->hdr.info.magic, ARCH_CONVERT)
1754                                                 == XFS_ATTR_LEAF_MAGIC);
1755         drop_hdr = &drop_leaf->hdr;
1756         save_hdr = &save_leaf->hdr;
1757
1758         /*
1759          * Save last hashval from dying block for later Btree fixup.
1760          */
1761         drop_blk->hashval =
1762                 INT_GET(drop_leaf->entries[INT_GET(drop_leaf->hdr.count,
1763                                                 ARCH_CONVERT)-1].hashval,
1764                                                                 ARCH_CONVERT);
1765
1766         /*
1767          * Check if we need a temp buffer, or can we do it in place.
1768          * Note that we don't check "leaf" for holes because we will
1769          * always be dropping it, toosmall() decided that for us already.
1770          */
1771         if (save_hdr->holes == 0) {
1772                 /*
1773                  * dest leaf has no holes, so we add there.  May need
1774                  * to make some room in the entry array.
1775                  */
1776                 if (xfs_attr_leaf_order(save_blk->bp, drop_blk->bp)) {
1777                         xfs_attr_leaf_moveents(drop_leaf, 0, save_leaf, 0,
1778                              (int)INT_GET(drop_hdr->count, ARCH_CONVERT), mp);
1779                 } else {
1780                         xfs_attr_leaf_moveents(drop_leaf, 0, save_leaf,
1781                                   INT_GET(save_hdr->count, ARCH_CONVERT),
1782                                   (int)INT_GET(drop_hdr->count, ARCH_CONVERT),
1783                                   mp);
1784                 }
1785         } else {
1786                 /*
1787                  * Destination has holes, so we make a temporary copy
1788                  * of the leaf and add them both to that.
1789                  */
1790                 tmpbuffer = kmem_alloc(state->blocksize, KM_SLEEP);
1791                 ASSERT(tmpbuffer != NULL);
1792                 memset(tmpbuffer, 0, state->blocksize);
1793                 tmp_leaf = (xfs_attr_leafblock_t *)tmpbuffer;
1794                 tmp_hdr = &tmp_leaf->hdr;
1795                 tmp_hdr->info = save_hdr->info; /* struct copy */
1796                 tmp_hdr->count = 0;
1797                 INT_SET(tmp_hdr->firstused, ARCH_CONVERT, state->blocksize);
1798                 if (!tmp_hdr->firstused) {
1799                         INT_SET(tmp_hdr->firstused, ARCH_CONVERT,
1800                                 state->blocksize - XFS_ATTR_LEAF_NAME_ALIGN);
1801                 }
1802                 tmp_hdr->usedbytes = 0;
1803                 if (xfs_attr_leaf_order(save_blk->bp, drop_blk->bp)) {
1804                         xfs_attr_leaf_moveents(drop_leaf, 0, tmp_leaf, 0,
1805                                 (int)INT_GET(drop_hdr->count, ARCH_CONVERT),
1806                                 mp);
1807                         xfs_attr_leaf_moveents(save_leaf, 0, tmp_leaf,
1808                                   INT_GET(tmp_leaf->hdr.count, ARCH_CONVERT),
1809                                  (int)INT_GET(save_hdr->count, ARCH_CONVERT),
1810                                  mp);
1811                 } else {
1812                         xfs_attr_leaf_moveents(save_leaf, 0, tmp_leaf, 0,
1813                                 (int)INT_GET(save_hdr->count, ARCH_CONVERT),
1814                                 mp);
1815                         xfs_attr_leaf_moveents(drop_leaf, 0, tmp_leaf,
1816                                 INT_GET(tmp_leaf->hdr.count, ARCH_CONVERT),
1817                                 (int)INT_GET(drop_hdr->count, ARCH_CONVERT),
1818                                 mp);
1819                 }
1820                 memcpy((char *)save_leaf, (char *)tmp_leaf, state->blocksize);
1821                 kmem_free(tmpbuffer, state->blocksize);
1822         }
1823
1824         xfs_da_log_buf(state->args->trans, save_blk->bp, 0,
1825                                            state->blocksize - 1);
1826
1827         /*
1828          * Copy out last hashval in each block for B-tree code.
1829          */
1830         save_blk->hashval =
1831                 INT_GET(save_leaf->entries[INT_GET(save_leaf->hdr.count,
1832                                                 ARCH_CONVERT)-1].hashval,
1833                                                                 ARCH_CONVERT);
1834 }
1835
1836 /*========================================================================
1837  * Routines used for finding things in the Btree.
1838  *========================================================================*/
1839
1840 /*
1841  * Look up a name in a leaf attribute list structure.
1842  * This is the internal routine, it uses the caller's buffer.
1843  *
1844  * Note that duplicate keys are allowed, but only check within the
1845  * current leaf node.  The Btree code must check in adjacent leaf nodes.
1846  *
1847  * Return in args->index the index into the entry[] array of either
1848  * the found entry, or where the entry should have been (insert before
1849  * that entry).
1850  *
1851  * Don't change the args->value unless we find the attribute.
1852  */
1853 int
1854 xfs_attr_leaf_lookup_int(xfs_dabuf_t *bp, xfs_da_args_t *args)
1855 {
1856         xfs_attr_leafblock_t *leaf;
1857         xfs_attr_leaf_entry_t *entry;
1858         xfs_attr_leaf_name_local_t *name_loc;
1859         xfs_attr_leaf_name_remote_t *name_rmt;
1860         int probe, span;
1861         xfs_dahash_t hashval;
1862
1863         leaf = bp->data;
1864         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
1865                                                 == XFS_ATTR_LEAF_MAGIC);
1866         ASSERT(INT_GET(leaf->hdr.count, ARCH_CONVERT)
1867                                         < (XFS_LBSIZE(args->dp->i_mount)/8));
1868
1869         /*
1870          * Binary search.  (note: small blocks will skip this loop)
1871          */
1872         hashval = args->hashval;
1873         probe = span = INT_GET(leaf->hdr.count, ARCH_CONVERT) / 2;
1874         for (entry = &leaf->entries[probe]; span > 4;
1875                    entry = &leaf->entries[probe]) {
1876                 span /= 2;
1877                 if (INT_GET(entry->hashval, ARCH_CONVERT) < hashval)
1878                         probe += span;
1879                 else if (INT_GET(entry->hashval, ARCH_CONVERT) > hashval)
1880                         probe -= span;
1881                 else
1882                         break;
1883         }
1884         ASSERT((probe >= 0) &&
1885                (!leaf->hdr.count
1886                || (probe < INT_GET(leaf->hdr.count, ARCH_CONVERT))));
1887         ASSERT((span <= 4) || (INT_GET(entry->hashval, ARCH_CONVERT)
1888                                                         == hashval));
1889
1890         /*
1891          * Since we may have duplicate hashval's, find the first matching
1892          * hashval in the leaf.
1893          */
1894         while ((probe > 0) && (INT_GET(entry->hashval, ARCH_CONVERT)
1895                                                         >= hashval)) {
1896                 entry--;
1897                 probe--;
1898         }
1899         while ((probe < INT_GET(leaf->hdr.count, ARCH_CONVERT))
1900                 && (INT_GET(entry->hashval, ARCH_CONVERT) < hashval)) {
1901                 entry++;
1902                 probe++;
1903         }
1904         if ((probe == INT_GET(leaf->hdr.count, ARCH_CONVERT))
1905                     || (INT_GET(entry->hashval, ARCH_CONVERT) != hashval)) {
1906                 args->index = probe;
1907                 return(XFS_ERROR(ENOATTR));
1908         }
1909
1910         /*
1911          * Duplicate keys may be present, so search all of them for a match.
1912          */
1913         for (  ; (probe < INT_GET(leaf->hdr.count, ARCH_CONVERT))
1914                         && (INT_GET(entry->hashval, ARCH_CONVERT) == hashval);
1915                         entry++, probe++) {
1916 /*
1917  * GROT: Add code to remove incomplete entries.
1918  */
1919                 /*
1920                  * If we are looking for INCOMPLETE entries, show only those.
1921                  * If we are looking for complete entries, show only those.
1922                  */
1923                 if ((args->flags & XFS_ATTR_INCOMPLETE) !=
1924                     (entry->flags & XFS_ATTR_INCOMPLETE)) {
1925                         continue;
1926                 }
1927                 if (entry->flags & XFS_ATTR_LOCAL) {
1928                         name_loc = XFS_ATTR_LEAF_NAME_LOCAL(leaf, probe);
1929                         if (name_loc->namelen != args->namelen)
1930                                 continue;
1931                         if (memcmp(args->name, (char *)name_loc->nameval,
1932                                              args->namelen) != 0)
1933                                 continue;
1934                         if (((args->flags & ATTR_SECURE) != 0) !=
1935                             ((entry->flags & XFS_ATTR_SECURE) != 0))
1936                                 continue;
1937                         if (((args->flags & ATTR_ROOT) != 0) !=
1938                             ((entry->flags & XFS_ATTR_ROOT) != 0))
1939                                 continue;
1940                         args->index = probe;
1941                         return(XFS_ERROR(EEXIST));
1942                 } else {
1943                         name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf, probe);
1944                         if (name_rmt->namelen != args->namelen)
1945                                 continue;
1946                         if (memcmp(args->name, (char *)name_rmt->name,
1947                                              args->namelen) != 0)
1948                                 continue;
1949                         if (((args->flags & ATTR_SECURE) != 0) !=
1950                             ((entry->flags & XFS_ATTR_SECURE) != 0))
1951                                 continue;
1952                         if (((args->flags & ATTR_ROOT) != 0) !=
1953                             ((entry->flags & XFS_ATTR_ROOT) != 0))
1954                                 continue;
1955                         args->index = probe;
1956                         args->rmtblkno
1957                                   = INT_GET(name_rmt->valueblk, ARCH_CONVERT);
1958                         args->rmtblkcnt = XFS_B_TO_FSB(args->dp->i_mount,
1959                                                    INT_GET(name_rmt->valuelen,
1960                                                                 ARCH_CONVERT));
1961                         return(XFS_ERROR(EEXIST));
1962                 }
1963         }
1964         args->index = probe;
1965         return(XFS_ERROR(ENOATTR));
1966 }
1967
1968 /*
1969  * Get the value associated with an attribute name from a leaf attribute
1970  * list structure.
1971  */
1972 int
1973 xfs_attr_leaf_getvalue(xfs_dabuf_t *bp, xfs_da_args_t *args)
1974 {
1975         int valuelen;
1976         xfs_attr_leafblock_t *leaf;
1977         xfs_attr_leaf_entry_t *entry;
1978         xfs_attr_leaf_name_local_t *name_loc;
1979         xfs_attr_leaf_name_remote_t *name_rmt;
1980
1981         leaf = bp->data;
1982         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
1983                                                 == XFS_ATTR_LEAF_MAGIC);
1984         ASSERT(INT_GET(leaf->hdr.count, ARCH_CONVERT)
1985                                         < (XFS_LBSIZE(args->dp->i_mount)/8));
1986         ASSERT(args->index < ((int)INT_GET(leaf->hdr.count, ARCH_CONVERT)));
1987
1988         entry = &leaf->entries[args->index];
1989         if (entry->flags & XFS_ATTR_LOCAL) {
1990                 name_loc = XFS_ATTR_LEAF_NAME_LOCAL(leaf, args->index);
1991                 ASSERT(name_loc->namelen == args->namelen);
1992                 ASSERT(memcmp(args->name, name_loc->nameval, args->namelen) == 0);
1993                 valuelen = INT_GET(name_loc->valuelen, ARCH_CONVERT);
1994                 if (args->flags & ATTR_KERNOVAL) {
1995                         args->valuelen = valuelen;
1996                         return(0);
1997                 }
1998                 if (args->valuelen < valuelen) {
1999                         args->valuelen = valuelen;
2000                         return(XFS_ERROR(ERANGE));
2001                 }
2002                 args->valuelen = valuelen;
2003                 memcpy(args->value, &name_loc->nameval[args->namelen], valuelen);
2004         } else {
2005                 name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf, args->index);
2006                 ASSERT(name_rmt->namelen == args->namelen);
2007                 ASSERT(memcmp(args->name, name_rmt->name, args->namelen) == 0);
2008                 valuelen = INT_GET(name_rmt->valuelen, ARCH_CONVERT);
2009                 args->rmtblkno = INT_GET(name_rmt->valueblk, ARCH_CONVERT);
2010                 args->rmtblkcnt = XFS_B_TO_FSB(args->dp->i_mount, valuelen);
2011                 if (args->flags & ATTR_KERNOVAL) {
2012                         args->valuelen = valuelen;
2013                         return(0);
2014                 }
2015                 if (args->valuelen < valuelen) {
2016                         args->valuelen = valuelen;
2017                         return(XFS_ERROR(ERANGE));
2018                 }
2019                 args->valuelen = valuelen;
2020         }
2021         return(0);
2022 }
2023
2024 /*========================================================================
2025  * Utility routines.
2026  *========================================================================*/
2027
2028 /*
2029  * Move the indicated entries from one leaf to another.
2030  * NOTE: this routine modifies both source and destination leaves.
2031  */
2032 /*ARGSUSED*/
2033 STATIC void
2034 xfs_attr_leaf_moveents(xfs_attr_leafblock_t *leaf_s, int start_s,
2035                         xfs_attr_leafblock_t *leaf_d, int start_d,
2036                         int count, xfs_mount_t *mp)
2037 {
2038         xfs_attr_leaf_hdr_t *hdr_s, *hdr_d;
2039         xfs_attr_leaf_entry_t *entry_s, *entry_d;
2040         int desti, tmp, i;
2041
2042         /*
2043          * Check for nothing to do.
2044          */
2045         if (count == 0)
2046                 return;
2047
2048         /*
2049          * Set up environment.
2050          */
2051         ASSERT(INT_GET(leaf_s->hdr.info.magic, ARCH_CONVERT)
2052                                                 == XFS_ATTR_LEAF_MAGIC);
2053         ASSERT(INT_GET(leaf_d->hdr.info.magic, ARCH_CONVERT)
2054                                                 == XFS_ATTR_LEAF_MAGIC);
2055         hdr_s = &leaf_s->hdr;
2056         hdr_d = &leaf_d->hdr;
2057         ASSERT((INT_GET(hdr_s->count, ARCH_CONVERT) > 0)
2058                                 && (INT_GET(hdr_s->count, ARCH_CONVERT)
2059                                                 < (XFS_LBSIZE(mp)/8)));
2060         ASSERT(INT_GET(hdr_s->firstused, ARCH_CONVERT) >=
2061                 ((INT_GET(hdr_s->count, ARCH_CONVERT)
2062                                         * sizeof(*entry_s))+sizeof(*hdr_s)));
2063         ASSERT(INT_GET(hdr_d->count, ARCH_CONVERT) < (XFS_LBSIZE(mp)/8));
2064         ASSERT(INT_GET(hdr_d->firstused, ARCH_CONVERT) >=
2065                 ((INT_GET(hdr_d->count, ARCH_CONVERT)
2066                                         * sizeof(*entry_d))+sizeof(*hdr_d)));
2067
2068         ASSERT(start_s < INT_GET(hdr_s->count, ARCH_CONVERT));
2069         ASSERT(start_d <= INT_GET(hdr_d->count, ARCH_CONVERT));
2070         ASSERT(count <= INT_GET(hdr_s->count, ARCH_CONVERT));
2071
2072         /*
2073          * Move the entries in the destination leaf up to make a hole?
2074          */
2075         if (start_d < INT_GET(hdr_d->count, ARCH_CONVERT)) {
2076                 tmp  = INT_GET(hdr_d->count, ARCH_CONVERT) - start_d;
2077                 tmp *= sizeof(xfs_attr_leaf_entry_t);
2078                 entry_s = &leaf_d->entries[start_d];
2079                 entry_d = &leaf_d->entries[start_d + count];
2080                 memmove((char *)entry_d, (char *)entry_s, tmp);
2081         }
2082
2083         /*
2084          * Copy all entry's in the same (sorted) order,
2085          * but allocate attribute info packed and in sequence.
2086          */
2087         entry_s = &leaf_s->entries[start_s];
2088         entry_d = &leaf_d->entries[start_d];
2089         desti = start_d;
2090         for (i = 0; i < count; entry_s++, entry_d++, desti++, i++) {
2091                 ASSERT(INT_GET(entry_s->nameidx, ARCH_CONVERT)
2092                                 >= INT_GET(hdr_s->firstused, ARCH_CONVERT));
2093                 tmp = xfs_attr_leaf_entsize(leaf_s, start_s + i);
2094 #ifdef GROT
2095                 /*
2096                  * Code to drop INCOMPLETE entries.  Difficult to use as we
2097                  * may also need to change the insertion index.  Code turned
2098                  * off for 6.2, should be revisited later.
2099                  */
2100                 if (entry_s->flags & XFS_ATTR_INCOMPLETE) { /* skip partials? */
2101                         memset(XFS_ATTR_LEAF_NAME(leaf_s, start_s + i), 0, tmp);
2102                         INT_MOD(hdr_s->usedbytes, ARCH_CONVERT, -tmp);
2103                         INT_MOD(hdr_s->count, ARCH_CONVERT, -1);
2104                         entry_d--;      /* to compensate for ++ in loop hdr */
2105                         desti--;
2106                         if ((start_s + i) < offset)
2107                                 result++;       /* insertion index adjustment */
2108                 } else {
2109 #endif /* GROT */
2110                         INT_MOD(hdr_d->firstused, ARCH_CONVERT, -tmp);
2111                         /* both on-disk, don't endian flip twice */
2112                         entry_d->hashval = entry_s->hashval;
2113                         /* both on-disk, don't endian flip twice */
2114                         entry_d->nameidx = hdr_d->firstused;
2115                         entry_d->flags = entry_s->flags;
2116                         ASSERT(INT_GET(entry_d->nameidx, ARCH_CONVERT) + tmp
2117                                                         <= XFS_LBSIZE(mp));
2118                         memmove(XFS_ATTR_LEAF_NAME(leaf_d, desti),
2119                                 XFS_ATTR_LEAF_NAME(leaf_s, start_s + i), tmp);
2120                         ASSERT(INT_GET(entry_s->nameidx, ARCH_CONVERT) + tmp
2121                                                         <= XFS_LBSIZE(mp));
2122                         memset(XFS_ATTR_LEAF_NAME(leaf_s, start_s + i), 0, tmp);
2123                         INT_MOD(hdr_s->usedbytes, ARCH_CONVERT, -tmp);
2124                         INT_MOD(hdr_d->usedbytes, ARCH_CONVERT, tmp);
2125                         INT_MOD(hdr_s->count, ARCH_CONVERT, -1);
2126                         INT_MOD(hdr_d->count, ARCH_CONVERT, 1);
2127                         tmp = INT_GET(hdr_d->count, ARCH_CONVERT)
2128                                                 * sizeof(xfs_attr_leaf_entry_t)
2129                                                 + sizeof(xfs_attr_leaf_hdr_t);
2130                         ASSERT(INT_GET(hdr_d->firstused, ARCH_CONVERT) >= tmp);
2131 #ifdef GROT
2132                 }
2133 #endif /* GROT */
2134         }
2135
2136         /*
2137          * Zero out the entries we just copied.
2138          */
2139         if (start_s == INT_GET(hdr_s->count, ARCH_CONVERT)) {
2140                 tmp = count * sizeof(xfs_attr_leaf_entry_t);
2141                 entry_s = &leaf_s->entries[start_s];
2142                 ASSERT(((char *)entry_s + tmp) <=
2143                        ((char *)leaf_s + XFS_LBSIZE(mp)));
2144                 memset((char *)entry_s, 0, tmp);
2145         } else {
2146                 /*
2147                  * Move the remaining entries down to fill the hole,
2148                  * then zero the entries at the top.
2149                  */
2150                 tmp  = INT_GET(hdr_s->count, ARCH_CONVERT) - count;
2151                 tmp *= sizeof(xfs_attr_leaf_entry_t);
2152                 entry_s = &leaf_s->entries[start_s + count];
2153                 entry_d = &leaf_s->entries[start_s];
2154                 memmove((char *)entry_d, (char *)entry_s, tmp);
2155
2156                 tmp = count * sizeof(xfs_attr_leaf_entry_t);
2157                 entry_s = &leaf_s->entries[INT_GET(hdr_s->count,
2158                                                         ARCH_CONVERT)];
2159                 ASSERT(((char *)entry_s + tmp) <=
2160                        ((char *)leaf_s + XFS_LBSIZE(mp)));
2161                 memset((char *)entry_s, 0, tmp);
2162         }
2163
2164         /*
2165          * Fill in the freemap information
2166          */
2167         INT_SET(hdr_d->freemap[0].base, ARCH_CONVERT,
2168                                         sizeof(xfs_attr_leaf_hdr_t));
2169         INT_MOD(hdr_d->freemap[0].base, ARCH_CONVERT,
2170                                 INT_GET(hdr_d->count, ARCH_CONVERT)
2171                                         * sizeof(xfs_attr_leaf_entry_t));
2172         INT_SET(hdr_d->freemap[0].size, ARCH_CONVERT,
2173                                 INT_GET(hdr_d->firstused, ARCH_CONVERT)
2174                               - INT_GET(hdr_d->freemap[0].base, ARCH_CONVERT));
2175         hdr_d->freemap[1].base = 0;
2176         hdr_d->freemap[2].base = 0;
2177         hdr_d->freemap[1].size = 0;
2178         hdr_d->freemap[2].size = 0;
2179         hdr_s->holes = 1;       /* leaf may not be compact */
2180 }
2181
2182 /*
2183  * Compare two leaf blocks "order".
2184  * Return 0 unless leaf2 should go before leaf1.
2185  */
2186 int
2187 xfs_attr_leaf_order(xfs_dabuf_t *leaf1_bp, xfs_dabuf_t *leaf2_bp)
2188 {
2189         xfs_attr_leafblock_t *leaf1, *leaf2;
2190
2191         leaf1 = leaf1_bp->data;
2192         leaf2 = leaf2_bp->data;
2193         ASSERT((INT_GET(leaf1->hdr.info.magic, ARCH_CONVERT)
2194                                                 == XFS_ATTR_LEAF_MAGIC) &&
2195                (INT_GET(leaf2->hdr.info.magic, ARCH_CONVERT)
2196                                                 == XFS_ATTR_LEAF_MAGIC));
2197         if (   (INT_GET(leaf1->hdr.count, ARCH_CONVERT) > 0)
2198             && (INT_GET(leaf2->hdr.count, ARCH_CONVERT) > 0)
2199             && (   (INT_GET(leaf2->entries[ 0 ].hashval, ARCH_CONVERT) <
2200                       INT_GET(leaf1->entries[ 0 ].hashval, ARCH_CONVERT))
2201                 || (INT_GET(leaf2->entries[INT_GET(leaf2->hdr.count,
2202                                 ARCH_CONVERT)-1].hashval, ARCH_CONVERT) <
2203                       INT_GET(leaf1->entries[INT_GET(leaf1->hdr.count,
2204                                 ARCH_CONVERT)-1].hashval, ARCH_CONVERT))) ) {
2205                 return(1);
2206         }
2207         return(0);
2208 }
2209
2210 /*
2211  * Pick up the last hashvalue from a leaf block.
2212  */
2213 xfs_dahash_t
2214 xfs_attr_leaf_lasthash(xfs_dabuf_t *bp, int *count)
2215 {
2216         xfs_attr_leafblock_t *leaf;
2217
2218         leaf = bp->data;
2219         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
2220                                                 == XFS_ATTR_LEAF_MAGIC);
2221         if (count)
2222                 *count = INT_GET(leaf->hdr.count, ARCH_CONVERT);
2223         if (!leaf->hdr.count)
2224                 return(0);
2225         return(INT_GET(leaf->entries[INT_GET(leaf->hdr.count,
2226                                 ARCH_CONVERT)-1].hashval, ARCH_CONVERT));
2227 }
2228
2229 /*
2230  * Calculate the number of bytes used to store the indicated attribute
2231  * (whether local or remote only calculate bytes in this block).
2232  */
2233 STATIC int
2234 xfs_attr_leaf_entsize(xfs_attr_leafblock_t *leaf, int index)
2235 {
2236         xfs_attr_leaf_name_local_t *name_loc;
2237         xfs_attr_leaf_name_remote_t *name_rmt;
2238         int size;
2239
2240         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
2241                                                 == XFS_ATTR_LEAF_MAGIC);
2242         if (leaf->entries[index].flags & XFS_ATTR_LOCAL) {
2243                 name_loc = XFS_ATTR_LEAF_NAME_LOCAL(leaf, index);
2244                 size = XFS_ATTR_LEAF_ENTSIZE_LOCAL(name_loc->namelen,
2245                                                    INT_GET(name_loc->valuelen,
2246                                                                 ARCH_CONVERT));
2247         } else {
2248                 name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf, index);
2249                 size = XFS_ATTR_LEAF_ENTSIZE_REMOTE(name_rmt->namelen);
2250         }
2251         return(size);
2252 }
2253
2254 /*
2255  * Calculate the number of bytes that would be required to store the new
2256  * attribute (whether local or remote only calculate bytes in this block).
2257  * This routine decides as a side effect whether the attribute will be
2258  * a "local" or a "remote" attribute.
2259  */
2260 int
2261 xfs_attr_leaf_newentsize(int namelen, int valuelen, int blocksize, int *local)
2262 {
2263         int size;
2264
2265         size = XFS_ATTR_LEAF_ENTSIZE_LOCAL(namelen, valuelen);
2266         if (size < XFS_ATTR_LEAF_ENTSIZE_LOCAL_MAX(blocksize)) {
2267                 if (local) {
2268                         *local = 1;
2269                 }
2270         } else {
2271                 size = XFS_ATTR_LEAF_ENTSIZE_REMOTE(namelen);
2272                 if (local) {
2273                         *local = 0;
2274                 }
2275         }
2276         return(size);
2277 }
2278
2279 /*
2280  * Copy out attribute list entries for attr_list(), for leaf attribute lists.
2281  */
2282 int
2283 xfs_attr_leaf_list_int(xfs_dabuf_t *bp, xfs_attr_list_context_t *context)
2284 {
2285         attrlist_cursor_kern_t *cursor;
2286         xfs_attr_leafblock_t *leaf;
2287         xfs_attr_leaf_entry_t *entry;
2288         xfs_attr_leaf_name_local_t *name_loc;
2289         xfs_attr_leaf_name_remote_t *name_rmt;
2290         int retval, i;
2291
2292         ASSERT(bp != NULL);
2293         leaf = bp->data;
2294         cursor = context->cursor;
2295         cursor->initted = 1;
2296
2297         xfs_attr_trace_l_cl("blk start", context, leaf);
2298
2299         /*
2300          * Re-find our place in the leaf block if this is a new syscall.
2301          */
2302         if (context->resynch) {
2303                 entry = &leaf->entries[0];
2304                 for (i = 0; i < INT_GET(leaf->hdr.count, ARCH_CONVERT);
2305                                                         entry++, i++) {
2306                         if (INT_GET(entry->hashval, ARCH_CONVERT)
2307                                                         == cursor->hashval) {
2308                                 if (cursor->offset == context->dupcnt) {
2309                                         context->dupcnt = 0;
2310                                         break;
2311                                 }
2312                                 context->dupcnt++;
2313                         } else if (INT_GET(entry->hashval, ARCH_CONVERT)
2314                                                         > cursor->hashval) {
2315                                 context->dupcnt = 0;
2316                                 break;
2317                         }
2318                 }
2319                 if (i == INT_GET(leaf->hdr.count, ARCH_CONVERT)) {
2320                         xfs_attr_trace_l_c("not found", context);
2321                         return(0);
2322                 }
2323         } else {
2324                 entry = &leaf->entries[0];
2325                 i = 0;
2326         }
2327         context->resynch = 0;
2328
2329         /*
2330          * We have found our place, start copying out the new attributes.
2331          */
2332         retval = 0;
2333         for (  ; (i < INT_GET(leaf->hdr.count, ARCH_CONVERT))
2334              && (retval == 0); entry++, i++) {
2335                 attrnames_t     *namesp;
2336
2337                 if (INT_GET(entry->hashval, ARCH_CONVERT) != cursor->hashval) {
2338                         cursor->hashval = INT_GET(entry->hashval, ARCH_CONVERT);
2339                         cursor->offset = 0;
2340                 }
2341
2342                 if (entry->flags & XFS_ATTR_INCOMPLETE)
2343                         continue;               /* skip incomplete entries */
2344                 if (((context->flags & ATTR_SECURE) != 0) !=
2345                     ((entry->flags & XFS_ATTR_SECURE) != 0) &&
2346                     !(context->flags & ATTR_KERNORMALS))
2347                         continue;               /* skip non-matching entries */
2348                 if (((context->flags & ATTR_ROOT) != 0) !=
2349                     ((entry->flags & XFS_ATTR_ROOT) != 0) &&
2350                     !(context->flags & ATTR_KERNROOTLS))
2351                         continue;               /* skip non-matching entries */
2352
2353                 namesp = (entry->flags & XFS_ATTR_SECURE) ? &attr_secure :
2354                         ((entry->flags & XFS_ATTR_ROOT) ? &attr_trusted :
2355                           &attr_user);
2356
2357                 if (entry->flags & XFS_ATTR_LOCAL) {
2358                         name_loc = XFS_ATTR_LEAF_NAME_LOCAL(leaf, i);
2359                         if (context->flags & ATTR_KERNOVAL) {
2360                                 ASSERT(context->flags & ATTR_KERNAMELS);
2361                                 context->count += namesp->attr_namelen +
2362                                                 (int)name_loc->namelen + 1;
2363                         } else {
2364                                 retval = xfs_attr_put_listent(context, namesp,
2365                                         (char *)name_loc->nameval,
2366                                         (int)name_loc->namelen,
2367                                         (int)INT_GET(name_loc->valuelen,
2368                                                                 ARCH_CONVERT));
2369                         }
2370                 } else {
2371                         name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf, i);
2372                         if (context->flags & ATTR_KERNOVAL) {
2373                                 ASSERT(context->flags & ATTR_KERNAMELS);
2374                                 context->count += namesp->attr_namelen +
2375                                                 (int)name_rmt->namelen + 1;
2376                         } else {
2377                                 retval = xfs_attr_put_listent(context, namesp,
2378                                         (char *)name_rmt->name,
2379                                         (int)name_rmt->namelen,
2380                                         (int)INT_GET(name_rmt->valuelen,
2381                                                                 ARCH_CONVERT));
2382                         }
2383                 }
2384                 if (retval == 0) {
2385                         cursor->offset++;
2386                 }
2387         }
2388         xfs_attr_trace_l_cl("blk end", context, leaf);
2389         return(retval);
2390 }
2391
2392 #define ATTR_ENTBASESIZE                /* minimum bytes used by an attr */ \
2393         (((struct attrlist_ent *) 0)->a_name - (char *) 0)
2394 #define ATTR_ENTSIZE(namelen)           /* actual bytes used by an attr */ \
2395         ((ATTR_ENTBASESIZE + (namelen) + 1 + sizeof(u_int32_t)-1) \
2396          & ~(sizeof(u_int32_t)-1))
2397
2398 /*
2399  * Format an attribute and copy it out to the user's buffer.
2400  * Take care to check values and protect against them changing later,
2401  * we may be reading them directly out of a user buffer.
2402  */
2403 /*ARGSUSED*/
2404 STATIC int
2405 xfs_attr_put_listent(xfs_attr_list_context_t *context,
2406                      attrnames_t *namesp, char *name, int namelen, int valuelen)
2407 {
2408         attrlist_ent_t *aep;
2409         int arraytop;
2410
2411         ASSERT(!(context->flags & ATTR_KERNOVAL));
2412         if (context->flags & ATTR_KERNAMELS) {
2413                 char *offset;
2414
2415                 ASSERT(context->count >= 0);
2416
2417                 arraytop = context->count + namesp->attr_namelen + namelen + 1;
2418                 if (arraytop > context->firstu) {
2419                         context->count = -1;    /* insufficient space */
2420                         return(1);
2421                 }
2422                 offset = (char *)context->alist + context->count;
2423                 strncpy(offset, namesp->attr_name, namesp->attr_namelen);
2424                 offset += namesp->attr_namelen;
2425                 strncpy(offset, name, namelen);                 /* real name */
2426                 offset += namelen;
2427                 *offset = '\0';
2428                 context->count += namesp->attr_namelen + namelen + 1;
2429                 return(0);
2430         }
2431
2432         ASSERT(context->count >= 0);
2433         ASSERT(context->count < (ATTR_MAX_VALUELEN/8));
2434         ASSERT(context->firstu >= sizeof(*context->alist));
2435         ASSERT(context->firstu <= context->bufsize);
2436
2437         arraytop = sizeof(*context->alist) +
2438                         context->count * sizeof(context->alist->al_offset[0]);
2439         context->firstu -= ATTR_ENTSIZE(namelen);
2440         if (context->firstu < arraytop) {
2441                 xfs_attr_trace_l_c("buffer full", context);
2442                 context->alist->al_more = 1;
2443                 return(1);
2444         }
2445
2446         aep = (attrlist_ent_t *)&(((char *)context->alist)[ context->firstu ]);
2447         aep->a_valuelen = valuelen;
2448         memcpy(aep->a_name, name, namelen);
2449         aep->a_name[ namelen ] = 0;
2450         context->alist->al_offset[ context->count++ ] = context->firstu;
2451         context->alist->al_count = context->count;
2452         xfs_attr_trace_l_c("add", context);
2453         return(0);
2454 }
2455
2456 /*========================================================================
2457  * Manage the INCOMPLETE flag in a leaf entry
2458  *========================================================================*/
2459
2460 /*
2461  * Clear the INCOMPLETE flag on an entry in a leaf block.
2462  */
2463 int
2464 xfs_attr_leaf_clearflag(xfs_da_args_t *args)
2465 {
2466         xfs_attr_leafblock_t *leaf;
2467         xfs_attr_leaf_entry_t *entry;
2468         xfs_attr_leaf_name_remote_t *name_rmt;
2469         xfs_dabuf_t *bp;
2470         int error;
2471 #ifdef DEBUG
2472         xfs_attr_leaf_name_local_t *name_loc;
2473         int namelen;
2474         char *name;
2475 #endif /* DEBUG */
2476
2477         /*
2478          * Set up the operation.
2479          */
2480         error = xfs_da_read_buf(args->trans, args->dp, args->blkno, -1, &bp,
2481                                              XFS_ATTR_FORK);
2482         if (error) {
2483                 return(error);
2484         }
2485         ASSERT(bp != NULL);
2486
2487         leaf = bp->data;
2488         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
2489                                                 == XFS_ATTR_LEAF_MAGIC);
2490         ASSERT(args->index < INT_GET(leaf->hdr.count, ARCH_CONVERT));
2491         ASSERT(args->index >= 0);
2492         entry = &leaf->entries[ args->index ];
2493         ASSERT(entry->flags & XFS_ATTR_INCOMPLETE);
2494
2495 #ifdef DEBUG
2496         if (entry->flags & XFS_ATTR_LOCAL) {
2497                 name_loc = XFS_ATTR_LEAF_NAME_LOCAL(leaf, args->index);
2498                 namelen = name_loc->namelen;
2499                 name = (char *)name_loc->nameval;
2500         } else {
2501                 name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf, args->index);
2502                 namelen = name_rmt->namelen;
2503                 name = (char *)name_rmt->name;
2504         }
2505         ASSERT(INT_GET(entry->hashval, ARCH_CONVERT) == args->hashval);
2506         ASSERT(namelen == args->namelen);
2507         ASSERT(memcmp(name, args->name, namelen) == 0);
2508 #endif /* DEBUG */
2509
2510         entry->flags &= ~XFS_ATTR_INCOMPLETE;
2511         xfs_da_log_buf(args->trans, bp,
2512                          XFS_DA_LOGRANGE(leaf, entry, sizeof(*entry)));
2513
2514         if (args->rmtblkno) {
2515                 ASSERT((entry->flags & XFS_ATTR_LOCAL) == 0);
2516                 name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf, args->index);
2517                 INT_SET(name_rmt->valueblk, ARCH_CONVERT, args->rmtblkno);
2518                 INT_SET(name_rmt->valuelen, ARCH_CONVERT, args->valuelen);
2519                 xfs_da_log_buf(args->trans, bp,
2520                          XFS_DA_LOGRANGE(leaf, name_rmt, sizeof(*name_rmt)));
2521         }
2522         xfs_da_buf_done(bp);
2523
2524         /*
2525          * Commit the flag value change and start the next trans in series.
2526          */
2527         error = xfs_attr_rolltrans(&args->trans, args->dp);
2528
2529         return(error);
2530 }
2531
2532 /*
2533  * Set the INCOMPLETE flag on an entry in a leaf block.
2534  */
2535 int
2536 xfs_attr_leaf_setflag(xfs_da_args_t *args)
2537 {
2538         xfs_attr_leafblock_t *leaf;
2539         xfs_attr_leaf_entry_t *entry;
2540         xfs_attr_leaf_name_remote_t *name_rmt;
2541         xfs_dabuf_t *bp;
2542         int error;
2543
2544         /*
2545          * Set up the operation.
2546          */
2547         error = xfs_da_read_buf(args->trans, args->dp, args->blkno, -1, &bp,
2548                                              XFS_ATTR_FORK);
2549         if (error) {
2550                 return(error);
2551         }
2552         ASSERT(bp != NULL);
2553
2554         leaf = bp->data;
2555         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
2556                                                 == XFS_ATTR_LEAF_MAGIC);
2557         ASSERT(args->index < INT_GET(leaf->hdr.count, ARCH_CONVERT));
2558         ASSERT(args->index >= 0);
2559         entry = &leaf->entries[ args->index ];
2560
2561         ASSERT((entry->flags & XFS_ATTR_INCOMPLETE) == 0);
2562         entry->flags |= XFS_ATTR_INCOMPLETE;
2563         xfs_da_log_buf(args->trans, bp,
2564                         XFS_DA_LOGRANGE(leaf, entry, sizeof(*entry)));
2565         if ((entry->flags & XFS_ATTR_LOCAL) == 0) {
2566                 name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf, args->index);
2567                 name_rmt->valueblk = 0;
2568                 name_rmt->valuelen = 0;
2569                 xfs_da_log_buf(args->trans, bp,
2570                          XFS_DA_LOGRANGE(leaf, name_rmt, sizeof(*name_rmt)));
2571         }
2572         xfs_da_buf_done(bp);
2573
2574         /*
2575          * Commit the flag value change and start the next trans in series.
2576          */
2577         error = xfs_attr_rolltrans(&args->trans, args->dp);
2578
2579         return(error);
2580 }
2581
2582 /*
2583  * In a single transaction, clear the INCOMPLETE flag on the leaf entry
2584  * given by args->blkno/index and set the INCOMPLETE flag on the leaf
2585  * entry given by args->blkno2/index2.
2586  *
2587  * Note that they could be in different blocks, or in the same block.
2588  */
2589 int
2590 xfs_attr_leaf_flipflags(xfs_da_args_t *args)
2591 {
2592         xfs_attr_leafblock_t *leaf1, *leaf2;
2593         xfs_attr_leaf_entry_t *entry1, *entry2;
2594         xfs_attr_leaf_name_remote_t *name_rmt;
2595         xfs_dabuf_t *bp1, *bp2;
2596         int error;
2597 #ifdef DEBUG
2598         xfs_attr_leaf_name_local_t *name_loc;
2599         int namelen1, namelen2;
2600         char *name1, *name2;
2601 #endif /* DEBUG */
2602
2603         /*
2604          * Read the block containing the "old" attr
2605          */
2606         error = xfs_da_read_buf(args->trans, args->dp, args->blkno, -1, &bp1,
2607                                              XFS_ATTR_FORK);
2608         if (error) {
2609                 return(error);
2610         }
2611         ASSERT(bp1 != NULL);
2612
2613         /*
2614          * Read the block containing the "new" attr, if it is different
2615          */
2616         if (args->blkno2 != args->blkno) {
2617                 error = xfs_da_read_buf(args->trans, args->dp, args->blkno2,
2618                                         -1, &bp2, XFS_ATTR_FORK);
2619                 if (error) {
2620                         return(error);
2621                 }
2622                 ASSERT(bp2 != NULL);
2623         } else {
2624                 bp2 = bp1;
2625         }
2626
2627         leaf1 = bp1->data;
2628         ASSERT(INT_GET(leaf1->hdr.info.magic, ARCH_CONVERT)
2629                                                 == XFS_ATTR_LEAF_MAGIC);
2630         ASSERT(args->index < INT_GET(leaf1->hdr.count, ARCH_CONVERT));
2631         ASSERT(args->index >= 0);
2632         entry1 = &leaf1->entries[ args->index ];
2633
2634         leaf2 = bp2->data;
2635         ASSERT(INT_GET(leaf2->hdr.info.magic, ARCH_CONVERT)
2636                                                 == XFS_ATTR_LEAF_MAGIC);
2637         ASSERT(args->index2 < INT_GET(leaf2->hdr.count, ARCH_CONVERT));
2638         ASSERT(args->index2 >= 0);
2639         entry2 = &leaf2->entries[ args->index2 ];
2640
2641 #ifdef DEBUG
2642         if (entry1->flags & XFS_ATTR_LOCAL) {
2643                 name_loc = XFS_ATTR_LEAF_NAME_LOCAL(leaf1, args->index);
2644                 namelen1 = name_loc->namelen;
2645                 name1 = (char *)name_loc->nameval;
2646         } else {
2647                 name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf1, args->index);
2648                 namelen1 = name_rmt->namelen;
2649                 name1 = (char *)name_rmt->name;
2650         }
2651         if (entry2->flags & XFS_ATTR_LOCAL) {
2652                 name_loc = XFS_ATTR_LEAF_NAME_LOCAL(leaf2, args->index2);
2653                 namelen2 = name_loc->namelen;
2654                 name2 = (char *)name_loc->nameval;
2655         } else {
2656                 name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf2, args->index2);
2657                 namelen2 = name_rmt->namelen;
2658                 name2 = (char *)name_rmt->name;
2659         }
2660         ASSERT(INT_GET(entry1->hashval, ARCH_CONVERT) == INT_GET(entry2->hashval, ARCH_CONVERT));
2661         ASSERT(namelen1 == namelen2);
2662         ASSERT(memcmp(name1, name2, namelen1) == 0);
2663 #endif /* DEBUG */
2664
2665         ASSERT(entry1->flags & XFS_ATTR_INCOMPLETE);
2666         ASSERT((entry2->flags & XFS_ATTR_INCOMPLETE) == 0);
2667
2668         entry1->flags &= ~XFS_ATTR_INCOMPLETE;
2669         xfs_da_log_buf(args->trans, bp1,
2670                           XFS_DA_LOGRANGE(leaf1, entry1, sizeof(*entry1)));
2671         if (args->rmtblkno) {
2672                 ASSERT((entry1->flags & XFS_ATTR_LOCAL) == 0);
2673                 name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf1, args->index);
2674                 INT_SET(name_rmt->valueblk, ARCH_CONVERT, args->rmtblkno);
2675                 INT_SET(name_rmt->valuelen, ARCH_CONVERT, args->valuelen);
2676                 xfs_da_log_buf(args->trans, bp1,
2677                          XFS_DA_LOGRANGE(leaf1, name_rmt, sizeof(*name_rmt)));
2678         }
2679
2680         entry2->flags |= XFS_ATTR_INCOMPLETE;
2681         xfs_da_log_buf(args->trans, bp2,
2682                           XFS_DA_LOGRANGE(leaf2, entry2, sizeof(*entry2)));
2683         if ((entry2->flags & XFS_ATTR_LOCAL) == 0) {
2684                 name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf2, args->index2);
2685                 name_rmt->valueblk = 0;
2686                 name_rmt->valuelen = 0;
2687                 xfs_da_log_buf(args->trans, bp2,
2688                          XFS_DA_LOGRANGE(leaf2, name_rmt, sizeof(*name_rmt)));
2689         }
2690         xfs_da_buf_done(bp1);
2691         if (bp1 != bp2)
2692                 xfs_da_buf_done(bp2);
2693
2694         /*
2695          * Commit the flag value change and start the next trans in series.
2696          */
2697         error = xfs_attr_rolltrans(&args->trans, args->dp);
2698
2699         return(error);
2700 }
2701
2702 /*========================================================================
2703  * Indiscriminately delete the entire attribute fork
2704  *========================================================================*/
2705
2706 /*
2707  * Recurse (gasp!) through the attribute nodes until we find leaves.
2708  * We're doing a depth-first traversal in order to invalidate everything.
2709  */
2710 int
2711 xfs_attr_root_inactive(xfs_trans_t **trans, xfs_inode_t *dp)
2712 {
2713         xfs_da_blkinfo_t *info;
2714         xfs_daddr_t blkno;
2715         xfs_dabuf_t *bp;
2716         int error;
2717
2718         /*
2719          * Read block 0 to see what we have to work with.
2720          * We only get here if we have extents, since we remove
2721          * the extents in reverse order the extent containing
2722          * block 0 must still be there.
2723          */
2724         error = xfs_da_read_buf(*trans, dp, 0, -1, &bp, XFS_ATTR_FORK);
2725         if (error)
2726                 return(error);
2727         blkno = xfs_da_blkno(bp);
2728
2729         /*
2730          * Invalidate the tree, even if the "tree" is only a single leaf block.
2731          * This is a depth-first traversal!
2732          */
2733         info = bp->data;
2734         if (INT_GET(info->magic, ARCH_CONVERT) == XFS_DA_NODE_MAGIC) {
2735                 error = xfs_attr_node_inactive(trans, dp, bp, 1);
2736         } else if (INT_GET(info->magic, ARCH_CONVERT) == XFS_ATTR_LEAF_MAGIC) {
2737                 error = xfs_attr_leaf_inactive(trans, dp, bp);
2738         } else {
2739                 error = XFS_ERROR(EIO);
2740                 xfs_da_brelse(*trans, bp);
2741         }
2742         if (error)
2743                 return(error);
2744
2745         /*
2746          * Invalidate the incore copy of the root block.
2747          */
2748         error = xfs_da_get_buf(*trans, dp, 0, blkno, &bp, XFS_ATTR_FORK);
2749         if (error)
2750                 return(error);
2751         xfs_da_binval(*trans, bp);      /* remove from cache */
2752         /*
2753          * Commit the invalidate and start the next transaction.
2754          */
2755         error = xfs_attr_rolltrans(trans, dp);
2756
2757         return (error);
2758 }
2759
2760 /*
2761  * Recurse (gasp!) through the attribute nodes until we find leaves.
2762  * We're doing a depth-first traversal in order to invalidate everything.
2763  */
2764 STATIC int
2765 xfs_attr_node_inactive(xfs_trans_t **trans, xfs_inode_t *dp, xfs_dabuf_t *bp,
2766                                    int level)
2767 {
2768         xfs_da_blkinfo_t *info;
2769         xfs_da_intnode_t *node;
2770         xfs_dablk_t child_fsb;
2771         xfs_daddr_t parent_blkno, child_blkno;
2772         int error, count, i;
2773         xfs_dabuf_t *child_bp;
2774
2775         /*
2776          * Since this code is recursive (gasp!) we must protect ourselves.
2777          */
2778         if (level > XFS_DA_NODE_MAXDEPTH) {
2779                 xfs_da_brelse(*trans, bp);      /* no locks for later trans */
2780                 return(XFS_ERROR(EIO));
2781         }
2782
2783         node = bp->data;
2784         ASSERT(INT_GET(node->hdr.info.magic, ARCH_CONVERT)
2785                                                 == XFS_DA_NODE_MAGIC);
2786         parent_blkno = xfs_da_blkno(bp);        /* save for re-read later */
2787         count = INT_GET(node->hdr.count, ARCH_CONVERT);
2788         if (!count) {
2789                 xfs_da_brelse(*trans, bp);
2790                 return(0);
2791         }
2792         child_fsb = INT_GET(node->btree[0].before, ARCH_CONVERT);
2793         xfs_da_brelse(*trans, bp);      /* no locks for later trans */
2794
2795         /*
2796          * If this is the node level just above the leaves, simply loop
2797          * over the leaves removing all of them.  If this is higher up
2798          * in the tree, recurse downward.
2799          */
2800         for (i = 0; i < count; i++) {
2801                 /*
2802                  * Read the subsidiary block to see what we have to work with.
2803                  * Don't do this in a transaction.  This is a depth-first
2804                  * traversal of the tree so we may deal with many blocks
2805                  * before we come back to this one.
2806                  */
2807                 error = xfs_da_read_buf(*trans, dp, child_fsb, -2, &child_bp,
2808                                                 XFS_ATTR_FORK);
2809                 if (error)
2810                         return(error);
2811                 if (child_bp) {
2812                                                 /* save for re-read later */
2813                         child_blkno = xfs_da_blkno(child_bp);
2814
2815                         /*
2816                          * Invalidate the subtree, however we have to.
2817                          */
2818                         info = child_bp->data;
2819                         if (INT_GET(info->magic, ARCH_CONVERT)
2820                                                         == XFS_DA_NODE_MAGIC) {
2821                                 error = xfs_attr_node_inactive(trans, dp,
2822                                                 child_bp, level+1);
2823                         } else if (INT_GET(info->magic, ARCH_CONVERT)
2824                                                 == XFS_ATTR_LEAF_MAGIC) {
2825                                 error = xfs_attr_leaf_inactive(trans, dp,
2826                                                 child_bp);
2827                         } else {
2828                                 error = XFS_ERROR(EIO);
2829                                 xfs_da_brelse(*trans, child_bp);
2830                         }
2831                         if (error)
2832                                 return(error);
2833
2834                         /*
2835                          * Remove the subsidiary block from the cache
2836                          * and from the log.
2837                          */
2838                         error = xfs_da_get_buf(*trans, dp, 0, child_blkno,
2839                                 &child_bp, XFS_ATTR_FORK);
2840                         if (error)
2841                                 return(error);
2842                         xfs_da_binval(*trans, child_bp);
2843                 }
2844
2845                 /*
2846                  * If we're not done, re-read the parent to get the next
2847                  * child block number.
2848                  */
2849                 if ((i+1) < count) {
2850                         error = xfs_da_read_buf(*trans, dp, 0, parent_blkno,
2851                                 &bp, XFS_ATTR_FORK);
2852                         if (error)
2853                                 return(error);
2854                         child_fsb = INT_GET(node->btree[i+1].before, ARCH_CONVERT);
2855                         xfs_da_brelse(*trans, bp);
2856                 }
2857                 /*
2858                  * Atomically commit the whole invalidate stuff.
2859                  */
2860                 if ((error = xfs_attr_rolltrans(trans, dp)))
2861                         return (error);
2862         }
2863
2864         return(0);
2865 }
2866
2867 /*
2868  * Invalidate all of the "remote" value regions pointed to by a particular
2869  * leaf block.
2870  * Note that we must release the lock on the buffer so that we are not
2871  * caught holding something that the logging code wants to flush to disk.
2872  */
2873 STATIC int
2874 xfs_attr_leaf_inactive(xfs_trans_t **trans, xfs_inode_t *dp, xfs_dabuf_t *bp)
2875 {
2876         xfs_attr_leafblock_t *leaf;
2877         xfs_attr_leaf_entry_t *entry;
2878         xfs_attr_leaf_name_remote_t *name_rmt;
2879         xfs_attr_inactive_list_t *list, *lp;
2880         int error, count, size, tmp, i;
2881
2882         leaf = bp->data;
2883         ASSERT(INT_GET(leaf->hdr.info.magic, ARCH_CONVERT)
2884                                                 == XFS_ATTR_LEAF_MAGIC);
2885
2886         /*
2887          * Count the number of "remote" value extents.
2888          */
2889         count = 0;
2890         entry = &leaf->entries[0];
2891         for (i = 0; i < INT_GET(leaf->hdr.count, ARCH_CONVERT); entry++, i++) {
2892                 if (   INT_GET(entry->nameidx, ARCH_CONVERT)
2893                     && ((entry->flags & XFS_ATTR_LOCAL) == 0)) {
2894                         name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf, i);
2895                         if (name_rmt->valueblk)
2896                                 count++;
2897                 }
2898         }
2899
2900         /*
2901          * If there are no "remote" values, we're done.
2902          */
2903         if (count == 0) {
2904                 xfs_da_brelse(*trans, bp);
2905                 return(0);
2906         }
2907
2908         /*
2909          * Allocate storage for a list of all the "remote" value extents.
2910          */
2911         size = count * sizeof(xfs_attr_inactive_list_t);
2912         list = (xfs_attr_inactive_list_t *)kmem_alloc(size, KM_SLEEP);
2913
2914         /*
2915          * Identify each of the "remote" value extents.
2916          */
2917         lp = list;
2918         entry = &leaf->entries[0];
2919         for (i = 0; i < INT_GET(leaf->hdr.count, ARCH_CONVERT); entry++, i++) {
2920                 if (   INT_GET(entry->nameidx, ARCH_CONVERT)
2921                     && ((entry->flags & XFS_ATTR_LOCAL) == 0)) {
2922                         name_rmt = XFS_ATTR_LEAF_NAME_REMOTE(leaf, i);
2923                         if (name_rmt->valueblk) {
2924                                 /* both on-disk, don't endian flip twice */
2925                                 lp->valueblk = name_rmt->valueblk;
2926                                 INT_SET(lp->valuelen, ARCH_CONVERT,
2927                                                 XFS_B_TO_FSB(dp->i_mount,
2928                                                     INT_GET(name_rmt->valuelen,
2929                                                               ARCH_CONVERT)));
2930                                 lp++;
2931                         }
2932                 }
2933         }
2934         xfs_da_brelse(*trans, bp);      /* unlock for trans. in freextent() */
2935
2936         /*
2937          * Invalidate each of the "remote" value extents.
2938          */
2939         error = 0;
2940         for (lp = list, i = 0; i < count; i++, lp++) {
2941                 tmp = xfs_attr_leaf_freextent(trans, dp,
2942                                                      INT_GET(lp->valueblk,
2943                                                                 ARCH_CONVERT),
2944                                                      INT_GET(lp->valuelen,
2945                                                                 ARCH_CONVERT));
2946                 if (error == 0)
2947                         error = tmp;    /* save only the 1st errno */
2948         }
2949
2950         kmem_free((xfs_caddr_t)list, size);
2951         return(error);
2952 }
2953
2954 /*
2955  * Look at all the extents for this logical region,
2956  * invalidate any buffers that are incore/in transactions.
2957  */
2958 STATIC int
2959 xfs_attr_leaf_freextent(xfs_trans_t **trans, xfs_inode_t *dp,
2960                                     xfs_dablk_t blkno, int blkcnt)
2961 {
2962         xfs_bmbt_irec_t map;
2963         xfs_dablk_t tblkno;
2964         int tblkcnt, dblkcnt, nmap, error;
2965         xfs_daddr_t dblkno;
2966         xfs_buf_t *bp;
2967
2968         /*
2969          * Roll through the "value", invalidating the attribute value's
2970          * blocks.
2971          */
2972         tblkno = blkno;
2973         tblkcnt = blkcnt;
2974         while (tblkcnt > 0) {
2975                 /*
2976                  * Try to remember where we decided to put the value.
2977                  */
2978                 nmap = 1;
2979                 error = xfs_bmapi(*trans, dp, (xfs_fileoff_t)tblkno, tblkcnt,
2980                                         XFS_BMAPI_ATTRFORK | XFS_BMAPI_METADATA,
2981                                         NULL, 0, &map, &nmap, NULL);
2982                 if (error) {
2983                         return(error);
2984                 }
2985                 ASSERT(nmap == 1);
2986                 ASSERT(map.br_startblock != DELAYSTARTBLOCK);
2987
2988                 /*
2989                  * If it's a hole, these are already unmapped
2990                  * so there's nothing to invalidate.
2991                  */
2992                 if (map.br_startblock != HOLESTARTBLOCK) {
2993
2994                         dblkno = XFS_FSB_TO_DADDR(dp->i_mount,
2995                                                   map.br_startblock);
2996                         dblkcnt = XFS_FSB_TO_BB(dp->i_mount,
2997                                                 map.br_blockcount);
2998                         bp = xfs_trans_get_buf(*trans,
2999                                         dp->i_mount->m_ddev_targp,
3000                                         dblkno, dblkcnt, XFS_BUF_LOCK);
3001                         xfs_trans_binval(*trans, bp);
3002                         /*
3003                          * Roll to next transaction.
3004                          */
3005                         if ((error = xfs_attr_rolltrans(trans, dp)))
3006                                 return (error);
3007                 }
3008
3009                 tblkno += map.br_blockcount;
3010                 tblkcnt -= map.br_blockcount;
3011         }
3012
3013         return(0);
3014 }
3015
3016
3017 /*
3018  * Roll from one trans in the sequence of PERMANENT transactions to the next.
3019  */
3020 int
3021 xfs_attr_rolltrans(xfs_trans_t **transp, xfs_inode_t *dp)
3022 {
3023         xfs_trans_t *trans;
3024         unsigned int logres, count;
3025         int     error;
3026
3027         /*
3028          * Ensure that the inode is always logged.
3029          */
3030         trans = *transp;
3031         xfs_trans_log_inode(trans, dp, XFS_ILOG_CORE);
3032
3033         /*
3034          * Copy the critical parameters from one trans to the next.
3035          */
3036         logres = trans->t_log_res;
3037         count = trans->t_log_count;
3038         *transp = xfs_trans_dup(trans);
3039
3040         /*
3041          * Commit the current transaction.
3042          * If this commit failed, then it'd just unlock those items that
3043          * are not marked ihold. That also means that a filesystem shutdown
3044          * is in progress. The caller takes the responsibility to cancel
3045          * the duplicate transaction that gets returned.
3046          */
3047         if ((error = xfs_trans_commit(trans, 0, NULL)))
3048                 return (error);
3049
3050         trans = *transp;
3051
3052         /*
3053          * Reserve space in the log for th next transaction.
3054          * This also pushes items in the "AIL", the list of logged items,
3055          * out to disk if they are taking up space at the tail of the log
3056          * that we want to use.  This requires that either nothing be locked
3057          * across this call, or that anything that is locked be logged in
3058          * the prior and the next transactions.
3059          */
3060         error = xfs_trans_reserve(trans, 0, logres, 0,
3061                                   XFS_TRANS_PERM_LOG_RES, count);
3062         /*
3063          *  Ensure that the inode is in the new transaction and locked.
3064          */
3065         if (!error) {
3066                 xfs_trans_ijoin(trans, dp, XFS_ILOCK_EXCL);
3067                 xfs_trans_ihold(trans, dp);
3068         }
3069         return (error);
3070
3071 }