lustre/llite/dir.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 1992, 1993, 1994, 1995
   5  * Remy Card (card@masi.ibp.fr)
   6  * Laboratoire MASI - Institut Blaise Pascal
   7  * Universite Pierre et Marie Curie (Paris VI)
   8  *
   9  *  from
  10  *
  11  *  linux/fs/minix/dir.c
  12  *  linux/fs/ext2/dir.c
  13  *
  14  *  Copyright (C) 1991, 1992  Linus Torvalds
  15  *
  16  *  ext2 directory handling functions
  17  *
  18  *  Big-endian to little-endian byte-swapping/bitmaps by
  19  *        David S. Miller (davem@caip.rutgers.edu), 1995
  20  *
  21  *  All code that works with directory layout had been switched to pagecache
  22  *  and moved here. AV
  23  *
  24  *  Adapted for Lustre Light
  25  *  Copyright (C) 2002-2003, Cluster File Systems, Inc.
  26  *
  27  */
  28
  29 #include <linux/fs.h>
  30 #include <linux/ext2_fs.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/mm.h>
  33 #include <linux/version.h>
  34 #include <linux/smp_lock.h>
  35 #include <asm/uaccess.h>
  36 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  37 # include <linux/locks.h>   // for wait_on_buffer
  38 #else
  39 # include <linux/buffer_head.h>   // for wait_on_buffer
  40 #endif
  41
  42 #define DEBUG_SUBSYSTEM S_LLITE
  43
  44 #include <obd_support.h>
  45 #include <obd_class.h>
  46 #include <lustre_lib.h>
  47 #include <lustre/lustre_idl.h>
  48 #include <lustre_lite.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_fid.h>
  51 #include "llite_internal.h"
  52
  53 #ifdef HAVE_PG_FS_MISC
  54 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
  55 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
  56 #endif
  57
  58 /*
  59  * (new) readdir implementation overview.
  60  *
  61  * Original lustre readdir implementation cached exact copy of raw directory
  62  * pages on the client. These pages were indexed in client page cache by
  63  * logical offset in the directory file. This design, while very simple and
  64  * intuitive had some inherent problems:
  65  *
  66  *     . it implies that byte offset to the directory entry serves as a
  67  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  68  *     ext3/htree directory entries may move due to splits, and more
  69  *     importantly,
  70  *
  71  *     . it is incompatible with the design of split directories for cmd3,
  72  *     that assumes that names are distributed across nodes based on their
  73  *     hash, and so readdir should be done in hash order.
  74  *
  75  * New readdir implementation does readdir in hash order, and uses hash of a
  76  * file name as a telldir/seekdir cookie. This led to number of complications:
  77  *
  78  *     . hash is not unique, so it cannot be used to index cached directory
  79  *     pages on the client (note, that it requires a whole pageful of hash
  80  *     collided entries to cause two pages to have identical hashes);
  81  *
  82  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  83  *     entry cookie. ext3/htree has the same problem and lustre implementation
  84  *     mimics their solution: seekdir(hash) positions directory at the first
  85  *     entry with the given hash.
  86  *
  87  * Client side.
  88  *
  89  * 0. caching
  90  *
  91  * Client caches directory pages using hash of the first entry as an index. As
  92  * noted above hash is not unique, so this solution doesn't work as is:
  93  * special processing is needed for "page hash chains" (i.e., sequences of
  94  * pages filled with entries all having the same hash value).
  95  *
  96  * First, such chains have to be detected. To this end, server returns to the
  97  * client the hash of the first entry on the page next to one returned. When
  98  * client detects that this hash is the same as hash of the first entry on the
  99  * returned page, page hash collision has to be handled. Pages in the
 100  * hash chain, except first one, are termed "overflow pages".
 101  *
 102  * Solution to index uniqueness problem is to not cache overflow
 103  * pages. Instead, when page hash collision is detected, all overflow pages
 104  * from emerging chain are immediately requested from the server and placed in
 105  * a special data structure (struct ll_dir_chain). This data structure is used
 106  * by ll_readdir() to process entries from overflow pages. When readdir
 107  * invocation finishes, overflow pages are discarded. If page hash collision
 108  * chain weren't completely processed, next call to readdir will again detect
 109  * page hash collision, again read overflow pages in, process next portion of
 110  * entries and again discard the pages. This is not as wasteful as it looks,
 111  * because, given reasonable hash, page hash collisions are extremely rare.
 112  *
 113  * 1. directory positioning
 114  *
 115  * When seekdir(hash) is called, original
 116  *
 117  *
 118  *
 119  *
 120  *
 121  *
 122  *
 123  *
 124  * Server.
 125  *
 126  * identification of and access to overflow pages
 127  *
 128  * page format
 129  *
 130  *
 131  *
 132  *
 133  *
 134  */
 135
 136 static __u32 hash_x_index(__u32 value)
 137 {
 138         return ((__u32)~0) - value;
 139 }
 140 #ifdef HAVE_PG_FS_MISC
 141 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
 142 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
 143 #endif
 144
 145 /* returns the page unlocked, but with a reference */
 146 static int ll_dir_readpage(struct file *file, struct page *page)
 147 {
 148         struct inode *inode = page->mapping->host;
 149         struct ptlrpc_request *request;
 150         struct mdt_body *body;
 151         struct obd_capa *oc;
 152         __u64 hash;
 153         int rc;
 154         ENTRY;
 155
 156         hash = hash_x_index(page->index);
 157         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) off %lu\n",
 158                inode->i_ino, inode->i_generation, inode, (unsigned long)hash);
 159
 160         oc = ll_mdscapa_get(inode);
 161         rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
 162                          oc, hash, page, &request);
 163         capa_put(oc);
 164         if (!rc) {
 165                 body = lustre_msg_buf(request->rq_repmsg, REPLY_REC_OFF,
 166                                       sizeof(*body));
 167                 /* Checked by mdc_readpage() */
 168                 LASSERT(body != NULL);
 169
 170                 /* Swabbed by mdc_readpage() */
 171                 LASSERT_REPSWABBED(request, REPLY_REC_OFF);
 172
 173                 if (body->valid & OBD_MD_FLSIZE)
 174                         inode->i_size = body->size;
 175                 SetPageUptodate(page);
 176         }
 177         ptlrpc_req_finished(request);
 178
 179         unlock_page(page);
 180         EXIT;
 181         return rc;
 182 }
 183
 184 struct address_space_operations ll_dir_aops = {
 185         .readpage  = ll_dir_readpage,
 186 };
 187
 188 static inline unsigned long dir_pages(struct inode *inode)
 189 {
 190         return (inode->i_size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 191 }
 192
 193 static inline unsigned ll_chunk_size(struct inode *inode)
 194 {
 195         return inode->i_sb->s_blocksize;
 196 }
 197
 198 static void ll_check_page(struct inode *dir, struct page *page)
 199 {
 200         /* XXX: check page format later */
 201         SetPageChecked(page);
 202 }
 203
 204 static inline void ll_put_page(struct page *page)
 205 {
 206         kunmap(page);
 207         page_cache_release(page);
 208 }
 209
 210 /*
 211  * Find, kmap and return page that contains given hash.
 212  */
 213 static struct page *ll_dir_page_locate(struct inode *dir, unsigned long hash,
 214                                        __u32 *start, __u32 *end)
 215 {
 216         struct address_space *mapping = dir->i_mapping;
 217         /*
 218          * Complement of hash is used as an index so that
 219          * radix_tree_gang_lookup() can be used to find a page with starting
 220          * hash _smaller_ than one we are looking for.
 221          */
 222         unsigned long offset = hash_x_index(hash);
 223         struct page *page;
 224         int found;
 225
 226         spin_lock_irq(&mapping->tree_lock);
 227         found = radix_tree_gang_lookup(&mapping->page_tree,
 228                                        (void **)&page, offset, 1);
 229         if (found > 0) {
 230                 struct lu_dirpage *dp;
 231
 232                 page_cache_get(page);
 233                 spin_unlock_irq(&mapping->tree_lock);
 234                 /*
 235                  * In contrast to find_lock_page() we are sure that directory
 236                  * page cannot be truncated (while DLM lock is held) and,
 237                  * hence, can avoid restart.
 238                  *
 239                  * In fact, page cannot be locked here at all, because
 240                  * ll_dir_readpage() does synchronous io.
 241                  */
 242                 wait_on_page(page);
 243                 if (PageUptodate(page)) {
 244                         dp = kmap(page);
 245                         *start = le32_to_cpu(dp->ldp_hash_start);
 246                         *end   = le32_to_cpu(dp->ldp_hash_end);
 247                         LASSERT(*start <= hash);
 248                         if (hash > *end || (*end != *start && hash == *end)) {
 249                                 kunmap(page);
 250                                 lock_page(page);
 251                                 ll_truncate_complete_page(page);
 252                                 unlock_page(page);
 253                                 page_cache_release(page);
 254                                 page = NULL;
 255                         }
 256                 } else {
 257                         page_cache_release(page);
 258                         page = ERR_PTR(-EIO);
 259                 }
 260
 261         } else {
 262                 spin_unlock_irq(&mapping->tree_lock);
 263                 page = NULL;
 264         }
 265         return page;
 266 }
 267
 268 /*
 269  * Chain of hash overflow pages.
 270  */
 271 struct ll_dir_chain {
 272         /* XXX something. Later */
 273 };
 274
 275 static void ll_dir_chain_init(struct ll_dir_chain *chain)
 276 {
 277 }
 278
 279 static void ll_dir_chain_fini(struct ll_dir_chain *chain)
 280 {
 281 }
 282
 283 static struct page *ll_get_dir_page(struct inode *dir, __u32 hash, int exact,
 284                                     struct ll_dir_chain *chain)
 285 {
 286         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 287         struct address_space *mapping = dir->i_mapping;
 288         struct lustre_handle lockh;
 289         struct lu_dirpage *dp;
 290         struct page *page;
 291         ldlm_mode_t mode;
 292         int rc;
 293         __u32 start;
 294         __u32 end;
 295
 296         mode = LCK_PR;
 297         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 298                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
 299         if (!rc) {
 300                 struct lookup_intent it = { .it_op = IT_READDIR };
 301                 struct ptlrpc_request *request;
 302                 struct md_op_data *op_data;
 303
 304                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
 305                                              LUSTRE_OPC_ANY);
 306                 if (IS_ERR(op_data))
 307                         return (void *)op_data;
 308
 309                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, LDLM_IBITS, &it,
 310                                 mode, op_data, &lockh, NULL, 0,
 311                                 ldlm_completion_ast, ll_md_blocking_ast, dir,
 312                                 0);
 313
 314                 ll_finish_md_op_data(op_data);
 315
 316                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 317                 if (request)
 318                         ptlrpc_req_finished(request);
 319                 if (rc < 0) {
 320                         CERROR("lock enqueue: rc: %d\n", rc);
 321                         return ERR_PTR(rc);
 322                 }
 323         } else {
 324                 /* for cross-ref object, l_ast_data of the lock may not be set,
 325                  * we reset it here */
 326                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie, dir);
 327         }
 328         ldlm_lock_dump_handle(D_OTHER, &lockh);
 329
 330         page = ll_dir_page_locate(dir, hash, &start, &end);
 331         if (IS_ERR(page))
 332                 GOTO(out_unlock, page);
 333
 334         if (page != NULL) {
 335                 /*
 336                  * XXX nikita: not entirely correct handling of a corner case:
 337                  * suppose hash chain of entries with hash value HASH crosses
 338                  * border between pages P0 and P1. First both P0 and P1 are
 339                  * cached, seekdir() is called for some entry from the P0 part
 340                  * of the chain. Later P0 goes out of cache. telldir(HASH)
 341                  * happens and finds P1, as it starts with matching hash
 342                  * value. Remaining entries from P0 part of the chain are
 343                  * skipped. (Is that really a bug?)
 344                  *
 345                  * Possible solutions: 0. don't cache P1 is such case, handle
 346                  * it as an "overflow" page. 1. invalidate all pages at
 347                  * once. 2. use HASH|1 as an index for P1.
 348                  */
 349                 if (exact && hash != start) {
 350                         /*
 351                          * readdir asked for a page starting _exactly_ from
 352                          * given hash, but cache contains stale page, with
 353                          * entries with smaller hash values. Stale page should
 354                          * be invalidated, and new one fetched.
 355                          */
 356                         CWARN("Stale readpage page %p: %#lx != %#lx\n", page,
 357                               (unsigned long)hash, (unsigned long)start);
 358                         lock_page(page);
 359                         ll_truncate_complete_page(page);
 360                         unlock_page(page);
 361                         page_cache_release(page);
 362                 } else
 363                         GOTO(hash_collision, page);
 364         }
 365
 366         page = read_cache_page(mapping, hash_x_index(hash),
 367                                (filler_t*)mapping->a_ops->readpage, NULL);
 368         if (IS_ERR(page))
 369                 GOTO(out_unlock, page);
 370
 371         wait_on_page(page);
 372         (void)kmap(page);
 373         if (!PageUptodate(page))
 374                 goto fail;
 375         if (!PageChecked(page))
 376                 ll_check_page(dir, page);
 377         if (PageError(page))
 378                 goto fail;
 379 hash_collision:
 380         dp = page_address(page);
 381
 382         start = le32_to_cpu(dp->ldp_hash_start);
 383         end   = le32_to_cpu(dp->ldp_hash_end);
 384         if (end == start) {
 385                 LASSERT(start == hash);
 386                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
 387                 /*
 388                  * Fetch whole overflow chain...
 389                  *
 390                  * XXX not yet.
 391                  */
 392                 goto fail;
 393         }
 394 out_unlock:
 395         ldlm_lock_decref(&lockh, mode);
 396         return page;
 397
 398 fail:
 399         ll_put_page(page);
 400         page = ERR_PTR(-EIO);
 401         goto out_unlock;
 402 }
 403
 404 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 405 {
 406         struct inode         *inode = filp->f_dentry->d_inode;
 407         struct ll_inode_info *info  = ll_i2info(inode);
 408         struct ll_sb_info    *sbi   = ll_i2sbi(inode);
 409         __u32                 pos   = filp->f_pos;
 410         struct page          *page;
 411         struct ll_dir_chain   chain;
 412         int rc;
 413         int done;
 414         int shift;
 415         ENTRY;
 416
 417         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu\n",
 418                inode->i_ino, inode->i_generation, inode,
 419                (unsigned long)pos, inode->i_size);
 420
 421         if (pos == DIR_END_OFF)
 422                 /*
 423                  * end-of-file.
 424                  */
 425                 RETURN(0);
 426
 427         rc    = 0;
 428         done  = 0;
 429         shift = 0;
 430         ll_dir_chain_init(&chain);
 431
 432         page = ll_get_dir_page(inode, pos, 0, &chain);
 433
 434         while (rc == 0 && !done) {
 435                 struct lu_dirpage *dp;
 436                 struct lu_dirent  *ent;
 437
 438                 if (!IS_ERR(page)) {
 439                         /*
 440                          * If page is empty (end of directoryis reached),
 441                          * use this value.
 442                          */
 443                         __u32 hash = DIR_END_OFF;
 444                         __u32 next;
 445
 446                         dp = page_address(page);
 447                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 448                              ent = lu_dirent_next(ent)) {
 449                                 char          *name;
 450                                 int            namelen;
 451                                 struct lu_fid  fid;
 452                                 ino_t          ino;
 453
 454                                 /*
 455                                  * XXX: implement correct swabbing here.
 456                                  */
 457
 458                                 hash    = le32_to_cpu(ent->lde_hash);
 459                                 namelen = le16_to_cpu(ent->lde_namelen);
 460
 461                                 if (hash < pos)
 462                                         /*
 463                                          * Skip until we find target hash
 464                                          * value.
 465                                          */
 466                                         continue;
 467
 468                                 if (namelen == 0)
 469                                         /*
 470                                          * Skip dummy record.
 471                                          */
 472                                         continue;
 473
 474                                 fid  = ent->lde_fid;
 475                                 name = ent->lde_name;
 476                                 fid_le_to_cpu(&fid, &fid);
 477                                 ino  = ll_fid_build_ino(sbi, &fid);
 478
 479                                 done = filldir(cookie, name, namelen,
 480                                                (loff_t)hash, ino, DT_UNKNOWN);
 481                         }
 482                         next = le32_to_cpu(dp->ldp_hash_end);
 483                         ll_put_page(page);
 484                         if (!done) {
 485                                 pos = next;
 486                                 if (pos == DIR_END_OFF)
 487                                         /*
 488                                          * End of directory reached.
 489                                          */
 490                                         done = 1;
 491                                 else if (1 /* chain is exhausted*/)
 492                                         /*
 493                                          * Normal case: continue to the next
 494                                          * page.
 495                                          */
 496                                         page = ll_get_dir_page(inode, pos, 1,
 497                                                                &chain);
 498                                 else {
 499                                         /*
 500                                          * go into overflow page.
 501                                          */
 502                                 }
 503                         } else
 504                                 pos = hash;
 505                 } else {
 506                         rc = PTR_ERR(page);
 507                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 508                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 509                 }
 510         }
 511
 512         filp->f_pos = (loff_t)(__s32)pos;
 513         filp->f_version = inode->i_version;
 514         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 515
 516         ll_dir_chain_fini(&chain);
 517
 518         RETURN(rc);
 519 }
 520
 521 #define QCTL_COPY(out, in)              \
 522 do {                                    \
 523         Q_COPY(out, in, qc_cmd);        \
 524         Q_COPY(out, in, qc_type);       \
 525         Q_COPY(out, in, qc_id);         \
 526         Q_COPY(out, in, qc_stat);       \
 527         Q_COPY(out, in, qc_dqinfo);     \
 528         Q_COPY(out, in, qc_dqblk);      \
 529 } while (0)
 530
 531 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump)
 532 {
 533         struct ll_sb_info *sbi = ll_i2sbi(inode);
 534         struct md_op_data *op_data;
 535         struct ptlrpc_request *req = NULL;
 536         int rc = 0;
 537
 538         /*
 539          * This is coming from userspace, so should be in
 540          * local endian.  But the MDS would like it in little
 541          * endian, so we swab it before we send it.
 542          */
 543         if (lump->lmm_magic != LOV_USER_MAGIC)
 544                 RETURN(-EINVAL);
 545
 546         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC))
 547                 lustre_swab_lov_user_md(lump);
 548
 549         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 550                                      LUSTRE_OPC_ANY);
 551         if (IS_ERR(op_data))
 552                 RETURN(PTR_ERR(op_data));
 553
 554         /* swabbing is done in lov_setstripe() on server side */
 555         rc = md_setattr(sbi->ll_md_exp, op_data, lump, sizeof(*lump),
 556                         NULL, 0, &req);
 557         ll_finish_md_op_data(op_data);
 558         ptlrpc_req_finished(req);
 559         if (rc) {
 560                 if (rc != -EPERM && rc != -EACCES)
 561                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 562         }
 563         return rc;
 564
 565 }
 566
 567 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 568                      int *lmm_size, struct ptlrpc_request **request)
 569 {
 570         struct ll_sb_info *sbi = ll_i2sbi(inode);
 571         struct mdt_body   *body;
 572         struct lov_mds_md *lmm = NULL;
 573         struct ptlrpc_request *req = NULL;
 574         int rc, lmmsize;
 575         struct obd_capa *oc;
 576
 577         rc = ll_get_max_mdsize(sbi, &lmmsize);
 578         if (rc)
 579                 RETURN(rc);
 580
 581         oc = ll_mdscapa_get(inode);
 582         rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode),
 583                         oc, OBD_MD_FLEASIZE | OBD_MD_FLDIREA,
 584                         lmmsize, &req);
 585         capa_put(oc);
 586         if (rc < 0) {
 587                 CDEBUG(D_INFO, "md_getattr failed on inode "
 588                        "%lu/%u: rc %d\n", inode->i_ino,
 589                        inode->i_generation, rc);
 590                 GOTO(out, rc);
 591         }
 592
 593         body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
 594         LASSERT(body != NULL); /* checked by md_getattr_name */
 595         /* swabbed by mdc_getattr_name */
 596         LASSERT_REPSWABBED(req, REPLY_REC_OFF);
 597
 598         lmmsize = body->eadatasize;
 599
 600         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 601             lmmsize == 0) {
 602                 GOTO(out, rc = -ENODATA);
 603         }
 604
 605         lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, lmmsize);
 606         LASSERT(lmm != NULL);
 607         LASSERT_REPSWABBED(req, REPLY_REC_OFF + 1);
 608
 609         /*
 610          * This is coming from the MDS, so is probably in
 611          * little endian.  We convert it to host endian before
 612          * passing it to userspace.
 613          */
 614         if (lmm->lmm_magic == __swab32(LOV_MAGIC)) {
 615                 lustre_swab_lov_user_md((struct lov_user_md *)lmm);
 616                 lustre_swab_lov_user_md_objects((struct lov_user_md *)lmm);
 617         }
 618 out:
 619         *lmmp = lmm;
 620         *lmm_size = lmmsize;
 621         *request = req;
 622         return rc;
 623 }
 624
 625 static int ll_dir_ioctl(struct inode *inode, struct file *file,
 626                         unsigned int cmd, unsigned long arg)
 627 {
 628         struct ll_sb_info *sbi = ll_i2sbi(inode);
 629         struct obd_ioctl_data *data;
 630         ENTRY;
 631
 632         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
 633                inode->i_ino, inode->i_generation, inode, cmd);
 634
 635         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
 636         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
 637                 return -ENOTTY;
 638
 639         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
 640         switch(cmd) {
 641         case EXT3_IOC_GETFLAGS:
 642         case EXT3_IOC_SETFLAGS:
 643                 RETURN(ll_iocontrol(inode, file, cmd, arg));
 644         case EXT3_IOC_GETVERSION_OLD:
 645         case EXT3_IOC_GETVERSION:
 646                 RETURN(put_user(inode->i_generation, (int *)arg));
 647         /* We need to special case any other ioctls we want to handle,
 648          * to send them to the MDS/OST as appropriate and to properly
 649          * network encode the arg field.
 650         case EXT3_IOC_SETVERSION_OLD:
 651         case EXT3_IOC_SETVERSION:
 652         */
 653         case IOC_MDC_LOOKUP: {
 654                 struct ptlrpc_request *request = NULL;
 655                 int namelen, rc, len = 0;
 656                 char *buf = NULL;
 657                 char *filename;
 658                 struct obd_capa *oc;
 659
 660                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 661                 if (rc)
 662                         RETURN(rc);
 663                 data = (void *)buf;
 664
 665                 filename = data->ioc_inlbuf1;
 666                 namelen = data->ioc_inllen1;
 667
 668                 if (namelen < 1) {
 669                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
 670                         GOTO(out, rc = -EINVAL);
 671                 }
 672
 673                 oc = ll_mdscapa_get(inode);
 674                 rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode), oc,
 675                                      filename, namelen, OBD_MD_FLID, 0,
 676                                      &request);
 677                 capa_put(oc);
 678                 if (rc < 0) {
 679                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 680                         GOTO(out, rc);
 681                 }
 682
 683                 ptlrpc_req_finished(request);
 684
 685                 EXIT;
 686         out:
 687                 obd_ioctl_freedata(buf, len);
 688                 return rc;
 689         }
 690         case LL_IOC_LOV_SETSTRIPE: {
 691                 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
 692                 int rc = 0;
 693
 694                 LASSERT(sizeof(lum) == sizeof(*lump));
 695                 LASSERT(sizeof(lum.lmm_objects[0]) ==
 696                         sizeof(lump->lmm_objects[0]));
 697                 rc = copy_from_user(&lum, lump, sizeof(lum));
 698                 if (rc)
 699                         RETURN(-EFAULT);
 700
 701                 rc = ll_dir_setstripe(inode, &lum);
 702
 703                 RETURN(rc);
 704         }
 705         case LL_IOC_OBD_STATFS:
 706                 RETURN(ll_obd_statfs(inode, (void *)arg));
 707         case LL_IOC_LOV_GETSTRIPE:
 708         case LL_IOC_MDC_GETINFO:
 709         case IOC_MDC_GETFILEINFO:
 710         case IOC_MDC_GETFILESTRIPE: {
 711                 struct ptlrpc_request *request = NULL;
 712                 struct lov_user_md *lump;
 713                 struct lov_mds_md *lmm = NULL;
 714                 struct mdt_body *body;
 715                 char *filename = NULL;
 716                 int rc, lmmsize;
 717
 718                 if (cmd == IOC_MDC_GETFILEINFO ||
 719                     cmd == IOC_MDC_GETFILESTRIPE) {
 720                         filename = getname((const char *)arg);
 721                         if (IS_ERR(filename))
 722                                 RETURN(PTR_ERR(filename));
 723
 724                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
 725                                                       &lmmsize, &request);
 726                 } else {
 727                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
 728                 }
 729
 730                 if (request) {
 731                         body = lustre_msg_buf(request->rq_repmsg,
 732                                               REPLY_REC_OFF, sizeof(*body));
 733                         LASSERT(body != NULL); /* checked by md_getattr_name */
 734                         /* swabbed by md_getattr_name */
 735                         LASSERT_REPSWABBED(request, REPLY_REC_OFF);
 736                 } else {
 737                         GOTO(out_req, rc);
 738                 }
 739
 740                 if (rc < 0) {
 741                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
 742                                                cmd == LL_IOC_MDC_GETINFO))
 743                                 GOTO(skip_lmm, rc = 0);
 744                         else
 745                                 GOTO(out_req, rc);
 746                 }
 747
 748                 if (cmd == IOC_MDC_GETFILESTRIPE ||
 749                     cmd == LL_IOC_LOV_GETSTRIPE) {
 750                         lump = (struct lov_user_md *)arg;
 751                 } else {
 752                         struct lov_user_mds_data *lmdp;
 753                         lmdp = (struct lov_user_mds_data *)arg;
 754                         lump = &lmdp->lmd_lmm;
 755                 }
 756                 rc = copy_to_user(lump, lmm, lmmsize);
 757                 if (rc)
 758                         GOTO(out_lmm, rc = -EFAULT);
 759         skip_lmm:
 760                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
 761                         struct lov_user_mds_data *lmdp;
 762                         lstat_t st = { 0 };
 763
 764                         st.st_dev     = inode->i_sb->s_dev;
 765                         st.st_mode    = body->mode;
 766                         st.st_nlink   = body->nlink;
 767                         st.st_uid     = body->uid;
 768                         st.st_gid     = body->gid;
 769                         st.st_rdev    = body->rdev;
 770                         st.st_size    = body->size;
 771                         st.st_blksize = CFS_PAGE_SIZE;
 772                         st.st_blocks  = body->blocks;
 773                         st.st_atime   = body->atime;
 774                         st.st_mtime   = body->mtime;
 775                         st.st_ctime   = body->ctime;
 776                         st.st_ino     = inode->i_ino;
 777
 778                         lmdp = (struct lov_user_mds_data *)arg;
 779                         rc = copy_to_user(&lmdp->lmd_st, &st, sizeof(st));
 780                         if (rc)
 781                                 GOTO(out_lmm, rc = -EFAULT);
 782                 }
 783
 784                 EXIT;
 785         out_lmm:
 786                 if (lmm && lmm->lmm_magic == LOV_MAGIC_JOIN)
 787                         OBD_FREE(lmm, lmmsize);
 788         out_req:
 789                 ptlrpc_req_finished(request);
 790                 if (filename)
 791                         putname(filename);
 792                 return rc;
 793         }
 794         case IOC_LOV_GETINFO: {
 795                 struct lov_user_mds_data *lumd;
 796                 struct lov_stripe_md *lsm;
 797                 struct lov_user_md *lum;
 798                 struct lov_mds_md *lmm;
 799                 int lmmsize;
 800                 lstat_t st;
 801                 int rc;
 802
 803                 lumd = (struct lov_user_mds_data *)arg;
 804                 lum = &lumd->lmd_lmm;
 805
 806                 rc = ll_get_max_mdsize(sbi, &lmmsize);
 807                 if (rc)
 808                         RETURN(rc);
 809
 810                 OBD_ALLOC(lmm, lmmsize);
 811                 rc = copy_from_user(lmm, lum, lmmsize);
 812                 if (rc)
 813                         GOTO(free_lmm, rc = -EFAULT);
 814
 815                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
 816                 if (rc < 0)
 817                         GOTO(free_lmm, rc = -ENOMEM);
 818
 819                 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
 820                 if (rc)
 821                         GOTO(free_lsm, rc);
 822
 823                 /* Perform glimpse_size operation. */
 824                 memset(&st, 0, sizeof(st));
 825
 826                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
 827                 if (rc)
 828                         GOTO(free_lsm, rc);
 829
 830                 rc = copy_to_user(&lumd->lmd_st, &st, sizeof(st));
 831                 if (rc)
 832                         GOTO(free_lsm, rc = -EFAULT);
 833
 834                 EXIT;
 835         free_lsm:
 836                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
 837         free_lmm:
 838                 OBD_FREE(lmm, lmmsize);
 839                 return rc;
 840         }
 841         case OBD_IOC_LLOG_CATINFO: {
 842                 struct ptlrpc_request *req = NULL;
 843                 char *buf = NULL;
 844                 int rc, len = 0;
 845                 char *bufs[3] = { NULL }, *str;
 846                 int lens[3] = { sizeof(struct ptlrpc_body) };
 847                 int size[2] = { sizeof(struct ptlrpc_body) };
 848
 849                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 850                 if (rc)
 851                         RETURN(rc);
 852                 data = (void *)buf;
 853
 854                 if (!data->ioc_inlbuf1) {
 855                         obd_ioctl_freedata(buf, len);
 856                         RETURN(-EINVAL);
 857                 }
 858
 859                 lens[REQ_REC_OFF] = data->ioc_inllen1;
 860                 bufs[REQ_REC_OFF] = data->ioc_inlbuf1;
 861                 if (data->ioc_inllen2) {
 862                         lens[REQ_REC_OFF + 1] = data->ioc_inllen2;
 863                         bufs[REQ_REC_OFF + 1] = data->ioc_inlbuf2;
 864                 } else {
 865                         lens[REQ_REC_OFF + 1] = 0;
 866                         bufs[REQ_REC_OFF + 1] = NULL;
 867                 }
 868
 869                 req = ptlrpc_prep_req(sbi2mdc(sbi)->cl_import,
 870                                       LUSTRE_LOG_VERSION, LLOG_CATINFO, 3, lens,
 871                                       bufs);
 872                 if (!req)
 873                         GOTO(out_catinfo, rc = -ENOMEM);
 874
 875                 size[REPLY_REC_OFF] = data->ioc_plen1;
 876                 ptlrpc_req_set_repsize(req, 2, size);
 877
 878                 rc = ptlrpc_queue_wait(req);
 879                 if (!rc) {
 880                         str = lustre_msg_string(req->rq_repmsg, REPLY_REC_OFF,
 881                                                 data->ioc_plen1);
 882                         rc = copy_to_user(data->ioc_pbuf1, str, data->ioc_plen1);
 883                 }
 884                 ptlrpc_req_finished(req);
 885         out_catinfo:
 886                 obd_ioctl_freedata(buf, len);
 887                 RETURN(rc);
 888         }
 889         case OBD_IOC_QUOTACHECK: {
 890                 struct obd_quotactl *oqctl;
 891                 int rc, error = 0;
 892
 893                 if (!capable(CAP_SYS_ADMIN))
 894                         RETURN(-EPERM);
 895
 896                 OBD_ALLOC_PTR(oqctl);
 897                 if (!oqctl)
 898                         RETURN(-ENOMEM);
 899                 oqctl->qc_type = arg;
 900                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
 901                 if (rc < 0) {
 902                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
 903                         error = rc;
 904                 }
 905
 906                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
 907                 if (rc < 0)
 908                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
 909
 910                 OBD_FREE_PTR(oqctl);
 911                 return error ?: rc;
 912         }
 913         case OBD_IOC_POLL_QUOTACHECK: {
 914                 struct if_quotacheck *check;
 915                 int rc;
 916
 917                 if (!capable(CAP_SYS_ADMIN))
 918                         RETURN(-EPERM);
 919
 920                 OBD_ALLOC_PTR(check);
 921                 if (!check)
 922                         RETURN(-ENOMEM);
 923
 924                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
 925                                    NULL);
 926                 if (rc) {
 927                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
 928                         if (copy_to_user((void *)arg, check, sizeof(*check)))
 929                                 rc = -EFAULT;
 930                         GOTO(out_poll, rc);
 931                 }
 932
 933                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
 934                                    NULL);
 935                 if (rc) {
 936                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
 937                         if (copy_to_user((void *)arg, check, sizeof(*check)))
 938                                 rc = -EFAULT;
 939                         GOTO(out_poll, rc);
 940                 }
 941         out_poll:
 942                 OBD_FREE_PTR(check);
 943                 RETURN(rc);
 944         }
 945 #ifdef HAVE_QUOTA_SUPPORT
 946         case OBD_IOC_QUOTACTL: {
 947                 struct if_quotactl *qctl;
 948                 struct obd_quotactl *oqctl;
 949
 950                 int cmd, type, id, rc = 0;
 951
 952                 OBD_ALLOC_PTR(qctl);
 953                 if (!qctl)
 954                         RETURN(-ENOMEM);
 955
 956                 OBD_ALLOC_PTR(oqctl);
 957                 if (!oqctl) {
 958                         OBD_FREE_PTR(qctl);
 959                         RETURN(-ENOMEM);
 960                 }
 961                 if (copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
 962                         GOTO(out_quotactl, rc = -EFAULT);
 963
 964                 cmd = qctl->qc_cmd;
 965                 type = qctl->qc_type;
 966                 id = qctl->qc_id;
 967                 switch (cmd) {
 968                 case Q_QUOTAON:
 969                 case Q_QUOTAOFF:
 970                 case Q_SETQUOTA:
 971                 case Q_SETINFO:
 972                         if (!capable(CAP_SYS_ADMIN))
 973                                 GOTO(out_quotactl, rc = -EPERM);
 974                         break;
 975                 case Q_GETQUOTA:
 976                         if (((type == USRQUOTA && current->euid != id) ||
 977                              (type == GRPQUOTA && !in_egroup_p(id))) &&
 978                             !capable(CAP_SYS_ADMIN))
 979                                 GOTO(out_quotactl, rc = -EPERM);
 980
 981                         /* XXX: dqb_valid is borrowed as a flag to mark that
 982                          *      only mds quota is wanted */
 983                         if (qctl->qc_dqblk.dqb_valid)
 984                                 qctl->obd_uuid = sbi->ll_md_exp->exp_obd->
 985                                                         u.cli.cl_target_uuid;
 986                         break;
 987                 case Q_GETINFO:
 988                         break;
 989                 default:
 990                         CERROR("unsupported quotactl op: %#x\n", cmd);
 991                         GOTO(out_quotactl, rc = -ENOTTY);
 992                 }
 993
 994                 QCTL_COPY(oqctl, qctl);
 995
 996                 if (qctl->obd_uuid.uuid[0]) {
 997                         struct obd_device *obd;
 998                         struct obd_uuid *uuid = &qctl->obd_uuid;
 999
1000                         obd = class_find_client_notype(uuid,
1001                                          &sbi->ll_dt_exp->exp_obd->obd_uuid);
1002                         if (!obd)
1003                                 GOTO(out_quotactl, rc = -ENOENT);
1004
1005                         if (cmd == Q_GETINFO)
1006                                 oqctl->qc_cmd = Q_GETOINFO;
1007                         else if (cmd == Q_GETQUOTA)
1008                                 oqctl->qc_cmd = Q_GETOQUOTA;
1009                         else
1010                                 GOTO(out_quotactl, rc = -EINVAL);
1011
1012                         if (sbi->ll_md_exp->exp_obd == obd) {
1013                                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1014                         } else {
1015                                 int i;
1016                                 struct obd_export *exp;
1017                                 struct lov_obd *lov = &sbi->ll_dt_exp->
1018                                                             exp_obd->u.lov;
1019
1020                                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1021                                         if (!lov->lov_tgts[i] ||
1022                                             !lov->lov_tgts[i]->ltd_active)
1023                                                 continue;
1024                                         exp = lov->lov_tgts[i]->ltd_exp;
1025                                         if (exp->exp_obd == obd) {
1026                                                 rc = obd_quotactl(exp, oqctl);
1027                                                 break;
1028                                         }
1029                                 }
1030                         }
1031
1032                         oqctl->qc_cmd = cmd;
1033                         QCTL_COPY(qctl, oqctl);
1034
1035                         if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1036                                 rc = -EFAULT;
1037
1038                         GOTO(out_quotactl, rc);
1039                 }
1040
1041                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1042                 if (rc && rc != -EBUSY && cmd == Q_QUOTAON) {
1043                         oqctl->qc_cmd = Q_QUOTAOFF;
1044                         obd_quotactl(sbi->ll_md_exp, oqctl);
1045                 }
1046
1047                 QCTL_COPY(qctl, oqctl);
1048
1049                 if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1050                         rc = -EFAULT;
1051         out_quotactl:
1052                 OBD_FREE_PTR(qctl);
1053                 OBD_FREE_PTR(oqctl);
1054                 RETURN(rc);
1055         }
1056 #endif /* HAVE_QUOTA_SUPPORT */
1057         case OBD_IOC_GETNAME: {
1058                 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1059                 if (!obd)
1060                         RETURN(-EFAULT);
1061                 if (copy_to_user((void *)arg, obd->obd_name,
1062                                 strlen(obd->obd_name) + 1))
1063                         RETURN (-EFAULT);
1064                 RETURN(0);
1065         }
1066         case LL_IOC_FLUSHCTX:
1067                 RETURN(ll_flush_ctx(inode));
1068         case LL_IOC_GETFACL: {
1069                 struct rmtacl_ioctl_data ioc;
1070
1071                 if (copy_from_user(&ioc, (void *)arg, sizeof(ioc)))
1072                         RETURN(-EFAULT);
1073
1074                 RETURN(ll_ioctl_getfacl(inode, &ioc));
1075         }
1076         case LL_IOC_SETFACL: {
1077                 struct rmtacl_ioctl_data ioc;
1078
1079                 if (copy_from_user(&ioc, (void *)arg, sizeof(ioc)))
1080                         RETURN(-EFAULT);
1081
1082                 RETURN(ll_ioctl_setfacl(inode, &ioc));
1083         }
1084         default:
1085                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1086         }
1087 }
1088
1089 int ll_dir_open(struct inode *inode, struct file *file)
1090 {
1091         ENTRY;
1092         RETURN(ll_file_open(inode, file));
1093 }
1094
1095 int ll_dir_release(struct inode *inode, struct file *file)
1096 {
1097         ENTRY;
1098         RETURN(ll_file_release(inode, file));
1099 }
1100
1101 struct file_operations ll_dir_operations = {
1102         .open     = ll_dir_open,
1103         .release  = ll_dir_release,
1104         .read     = generic_read_dir,
1105         .readdir  = ll_readdir,
1106         .ioctl    = ll_dir_ioctl
1107 };
1108