git.oblomov.eu Git - linux-2.6/blob - fs/xfs/xfs_attr_leaf.c

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