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 #include <linux/buffer_head.h>   // for wait_on_buffer
  37
  38 #define DEBUG_SUBSYSTEM S_LLITE
  39
  40 #include <obd_support.h>
  41 #include <obd_class.h>
  42 #include <lustre_lib.h>
  43 #include <lustre/lustre_idl.h>
  44 #include <lustre_lite.h>
  45 #include <lustre_dlm.h>
  46 #include <lustre_fid.h>
  47 #include "llite_internal.h"
  48
  49 #ifdef HAVE_PG_FS_MISC
  50 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
  51 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
  52 #endif
  53
  54 /*
  55  * (new) readdir implementation overview.
  56  *
  57  * Original lustre readdir implementation cached exact copy of raw directory
  58  * pages on the client. These pages were indexed in client page cache by
  59  * logical offset in the directory file. This design, while very simple and
  60  * intuitive had some inherent problems:
  61  *
  62  *     . it implies that byte offset to the directory entry serves as a
  63  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  64  *     ext3/htree directory entries may move due to splits, and more
  65  *     importantly,
  66  *
  67  *     . it is incompatible with the design of split directories for cmd3,
  68  *     that assumes that names are distributed across nodes based on their
  69  *     hash, and so readdir should be done in hash order.
  70  *
  71  * New readdir implementation does readdir in hash order, and uses hash of a
  72  * file name as a telldir/seekdir cookie. This led to number of complications:
  73  *
  74  *     . hash is not unique, so it cannot be used to index cached directory
  75  *     pages on the client (note, that it requires a whole pageful of hash
  76  *     collided entries to cause two pages to have identical hashes);
  77  *
  78  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  79  *     entry cookie. ext3/htree has the same problem and lustre implementation
  80  *     mimics their solution: seekdir(hash) positions directory at the first
  81  *     entry with the given hash.
  82  *
  83  * Client side.
  84  *
  85  * 0. caching
  86  *
  87  * Client caches directory pages using hash of the first entry as an index. As
  88  * noted above hash is not unique, so this solution doesn't work as is:
  89  * special processing is needed for "page hash chains" (i.e., sequences of
  90  * pages filled with entries all having the same hash value).
  91  *
  92  * First, such chains have to be detected. To this end, server returns to the
  93  * client the hash of the first entry on the page next to one returned. When
  94  * client detects that this hash is the same as hash of the first entry on the
  95  * returned page, page hash collision has to be handled. Pages in the
  96  * hash chain, except first one, are termed "overflow pages".
  97  *
  98  * Solution to index uniqueness problem is to not cache overflow
  99  * pages. Instead, when page hash collision is detected, all overflow pages
 100  * from emerging chain are immediately requested from the server and placed in
 101  * a special data structure (struct ll_dir_chain). This data structure is used
 102  * by ll_readdir() to process entries from overflow pages. When readdir
 103  * invocation finishes, overflow pages are discarded. If page hash collision
 104  * chain weren't completely processed, next call to readdir will again detect
 105  * page hash collision, again read overflow pages in, process next portion of
 106  * entries and again discard the pages. This is not as wasteful as it looks,
 107  * because, given reasonable hash, page hash collisions are extremely rare.
 108  *
 109  * 1. directory positioning
 110  *
 111  * When seekdir(hash) is called, original
 112  *
 113  *
 114  *
 115  *
 116  *
 117  *
 118  *
 119  *
 120  * Server.
 121  *
 122  * identification of and access to overflow pages
 123  *
 124  * page format
 125  *
 126  *
 127  *
 128  *
 129  *
 130  */
 131
 132 static __u32 hash_x_index(__u32 value)
 133 {
 134         return ((__u32)~0) - value;
 135 }
 136 #ifdef HAVE_PG_FS_MISC
 137 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
 138 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
 139 #endif
 140
 141 /* returns the page unlocked, but with a reference */
 142 static int ll_dir_readpage(struct file *file, struct page *page)
 143 {
 144         struct inode *inode = page->mapping->host;
 145         struct ptlrpc_request *request;
 146         struct mdt_body *body;
 147         struct obd_capa *oc;
 148         __u64 hash;
 149         int rc;
 150         ENTRY;
 151
 152         hash = hash_x_index(page->index);
 153         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) off %lu\n",
 154                inode->i_ino, inode->i_generation, inode, (unsigned long)hash);
 155
 156         oc = ll_mdscapa_get(inode);
 157         rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
 158                          oc, hash, page, &request);
 159         capa_put(oc);
 160         if (!rc) {
 161                 body = lustre_msg_buf(request->rq_repmsg, REPLY_REC_OFF,
 162                                       sizeof(*body));
 163                 /* Checked by mdc_readpage() */
 164                 LASSERT(body != NULL);
 165
 166                 /* Swabbed by mdc_readpage() */
 167                 LASSERT(lustre_rep_swabbed(request, REPLY_REC_OFF));
 168
 169                 if (body->valid & OBD_MD_FLSIZE)
 170                         i_size_write(inode, body->size);
 171                 SetPageUptodate(page);
 172         }
 173         ptlrpc_req_finished(request);
 174
 175         unlock_page(page);
 176         EXIT;
 177         return rc;
 178 }
 179
 180 struct address_space_operations ll_dir_aops = {
 181         .readpage  = ll_dir_readpage,
 182 };
 183
 184 static inline unsigned long dir_pages(struct inode *inode)
 185 {
 186         return (i_size_read(inode) + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 187 }
 188
 189 static inline unsigned ll_chunk_size(struct inode *inode)
 190 {
 191         return inode->i_sb->s_blocksize;
 192 }
 193
 194 static void ll_check_page(struct inode *dir, struct page *page)
 195 {
 196         /* XXX: check page format later */
 197         SetPageChecked(page);
 198 }
 199
 200 static inline void ll_put_page(struct page *page)
 201 {
 202         kunmap(page);
 203         page_cache_release(page);
 204 }
 205
 206 /*
 207  * Find, kmap and return page that contains given hash.
 208  */
 209 static struct page *ll_dir_page_locate(struct inode *dir, unsigned long hash,
 210                                        __u32 *start, __u32 *end)
 211 {
 212         struct address_space *mapping = dir->i_mapping;
 213         /*
 214          * Complement of hash is used as an index so that
 215          * radix_tree_gang_lookup() can be used to find a page with starting
 216          * hash _smaller_ than one we are looking for.
 217          */
 218         unsigned long offset = hash_x_index(hash);
 219         struct page *page;
 220         int found;
 221
 222         TREE_READ_LOCK_IRQ(mapping);
 223         found = radix_tree_gang_lookup(&mapping->page_tree,
 224                                        (void **)&page, offset, 1);
 225         if (found > 0) {
 226                 struct lu_dirpage *dp;
 227
 228                 page_cache_get(page);
 229                 TREE_READ_UNLOCK_IRQ(mapping);
 230                 /*
 231                  * In contrast to find_lock_page() we are sure that directory
 232                  * page cannot be truncated (while DLM lock is held) and,
 233                  * hence, can avoid restart.
 234                  *
 235                  * In fact, page cannot be locked here at all, because
 236                  * ll_dir_readpage() does synchronous io.
 237                  */
 238                 wait_on_page(page);
 239                 if (PageUptodate(page)) {
 240                         dp = kmap(page);
 241                         *start = le32_to_cpu(dp->ldp_hash_start);
 242                         *end   = le32_to_cpu(dp->ldp_hash_end);
 243                         LASSERT(*start <= hash);
 244                         if (hash > *end || (*end != *start && hash == *end)) {
 245                                 kunmap(page);
 246                                 lock_page(page);
 247                                 ll_truncate_complete_page(page);
 248                                 unlock_page(page);
 249                                 page_cache_release(page);
 250                                 page = NULL;
 251                         }
 252                 } else {
 253                         page_cache_release(page);
 254                         page = ERR_PTR(-EIO);
 255                 }
 256
 257         } else {
 258                 TREE_READ_UNLOCK_IRQ(mapping);
 259                 page = NULL;
 260         }
 261         return page;
 262 }
 263
 264 /*
 265  * Chain of hash overflow pages.
 266  */
 267 struct ll_dir_chain {
 268         /* XXX something. Later */
 269 };
 270
 271 static void ll_dir_chain_init(struct ll_dir_chain *chain)
 272 {
 273 }
 274
 275 static void ll_dir_chain_fini(struct ll_dir_chain *chain)
 276 {
 277 }
 278
 279 static struct page *ll_get_dir_page(struct inode *dir, __u32 hash, int exact,
 280                                     struct ll_dir_chain *chain)
 281 {
 282         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 283         struct address_space *mapping = dir->i_mapping;
 284         struct lustre_handle lockh;
 285         struct lu_dirpage *dp;
 286         struct page *page;
 287         ldlm_mode_t mode;
 288         int rc;
 289         __u32 start;
 290         __u32 end;
 291
 292         mode = LCK_PR;
 293         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 294                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
 295         if (!rc) {
 296                 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
 297                        ll_md_blocking_ast, ldlm_completion_ast, NULL, dir };
 298                 struct lookup_intent it = { .it_op = IT_READDIR };
 299                 struct ptlrpc_request *request;
 300                 struct md_op_data *op_data;
 301
 302                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
 303                                              LUSTRE_OPC_ANY, NULL);
 304                 if (IS_ERR(op_data))
 305                         return (void *)op_data;
 306
 307                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
 308                                 op_data, &lockh, NULL, 0, 0);
 309
 310                 ll_finish_md_op_data(op_data);
 311
 312                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 313                 if (request)
 314                         ptlrpc_req_finished(request);
 315                 if (rc < 0) {
 316                         CERROR("lock enqueue: rc: %d\n", rc);
 317                         return ERR_PTR(rc);
 318                 }
 319         } else {
 320                 /* for cross-ref object, l_ast_data of the lock may not be set,
 321                  * we reset it here */
 322                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie, dir);
 323         }
 324         ldlm_lock_dump_handle(D_OTHER, &lockh);
 325
 326         page = ll_dir_page_locate(dir, hash, &start, &end);
 327         if (IS_ERR(page))
 328                 GOTO(out_unlock, page);
 329
 330         if (page != NULL) {
 331                 /*
 332                  * XXX nikita: not entirely correct handling of a corner case:
 333                  * suppose hash chain of entries with hash value HASH crosses
 334                  * border between pages P0 and P1. First both P0 and P1 are
 335                  * cached, seekdir() is called for some entry from the P0 part
 336                  * of the chain. Later P0 goes out of cache. telldir(HASH)
 337                  * happens and finds P1, as it starts with matching hash
 338                  * value. Remaining entries from P0 part of the chain are
 339                  * skipped. (Is that really a bug?)
 340                  *
 341                  * Possible solutions: 0. don't cache P1 is such case, handle
 342                  * it as an "overflow" page. 1. invalidate all pages at
 343                  * once. 2. use HASH|1 as an index for P1.
 344                  */
 345                 if (exact && hash != start) {
 346                         /*
 347                          * readdir asked for a page starting _exactly_ from
 348                          * given hash, but cache contains stale page, with
 349                          * entries with smaller hash values. Stale page should
 350                          * be invalidated, and new one fetched.
 351                          */
 352                         CWARN("Stale readpage page %p: %#lx != %#lx\n", page,
 353                               (unsigned long)hash, (unsigned long)start);
 354                         lock_page(page);
 355                         ll_truncate_complete_page(page);
 356                         unlock_page(page);
 357                         page_cache_release(page);
 358                 } else
 359                         GOTO(hash_collision, page);
 360         }
 361
 362         page = read_cache_page(mapping, hash_x_index(hash),
 363                                (filler_t*)mapping->a_ops->readpage, NULL);
 364         if (IS_ERR(page))
 365                 GOTO(out_unlock, page);
 366
 367         wait_on_page(page);
 368         (void)kmap(page);
 369         if (!PageUptodate(page))
 370                 goto fail;
 371         if (!PageChecked(page))
 372                 ll_check_page(dir, page);
 373         if (PageError(page))
 374                 goto fail;
 375 hash_collision:
 376         dp = page_address(page);
 377
 378         start = le32_to_cpu(dp->ldp_hash_start);
 379         end   = le32_to_cpu(dp->ldp_hash_end);
 380         if (end == start) {
 381                 LASSERT(start == hash);
 382                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
 383                 /*
 384                  * Fetch whole overflow chain...
 385                  *
 386                  * XXX not yet.
 387                  */
 388                 goto fail;
 389         }
 390 out_unlock:
 391         ldlm_lock_decref(&lockh, mode);
 392         return page;
 393
 394 fail:
 395         ll_put_page(page);
 396         page = ERR_PTR(-EIO);
 397         goto out_unlock;
 398 }
 399
 400 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 401 {
 402         struct inode         *inode = filp->f_dentry->d_inode;
 403         struct ll_inode_info *info  = ll_i2info(inode);
 404         struct ll_sb_info    *sbi   = ll_i2sbi(inode);
 405         __u32                 pos   = filp->f_pos;
 406         struct page          *page;
 407         struct ll_dir_chain   chain;
 408         int rc;
 409         int done;
 410         int shift;
 411         ENTRY;
 412
 413         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu\n",
 414                inode->i_ino, inode->i_generation, inode,
 415                (unsigned long)pos, i_size_read(inode));
 416
 417         if (pos == DIR_END_OFF)
 418                 /*
 419                  * end-of-file.
 420                  */
 421                 RETURN(0);
 422
 423         rc    = 0;
 424         done  = 0;
 425         shift = 0;
 426         ll_dir_chain_init(&chain);
 427
 428         page = ll_get_dir_page(inode, pos, 0, &chain);
 429
 430         while (rc == 0 && !done) {
 431                 struct lu_dirpage *dp;
 432                 struct lu_dirent  *ent;
 433
 434                 if (!IS_ERR(page)) {
 435                         /*
 436                          * If page is empty (end of directoryis reached),
 437                          * use this value.
 438                          */
 439                         __u32 hash = DIR_END_OFF;
 440                         __u32 next;
 441
 442                         dp = page_address(page);
 443                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 444                              ent = lu_dirent_next(ent)) {
 445                                 char          *name;
 446                                 int            namelen;
 447                                 struct lu_fid  fid;
 448                                 ino_t          ino;
 449
 450                                 /*
 451                                  * XXX: implement correct swabbing here.
 452                                  */
 453
 454                                 hash    = le32_to_cpu(ent->lde_hash);
 455                                 namelen = le16_to_cpu(ent->lde_namelen);
 456
 457                                 if (hash < pos)
 458                                         /*
 459                                          * Skip until we find target hash
 460                                          * value.
 461                                          */
 462                                         continue;
 463
 464                                 if (namelen == 0)
 465                                         /*
 466                                          * Skip dummy record.
 467                                          */
 468                                         continue;
 469
 470                                 fid  = ent->lde_fid;
 471                                 name = ent->lde_name;
 472                                 fid_le_to_cpu(&fid, &fid);
 473                                 ino  = ll_fid_build_ino(sbi, &fid);
 474
 475                                 done = filldir(cookie, name, namelen,
 476                                                (loff_t)hash, ino, DT_UNKNOWN);
 477                         }
 478                         next = le32_to_cpu(dp->ldp_hash_end);
 479                         ll_put_page(page);
 480                         if (!done) {
 481                                 pos = next;
 482                                 if (pos == DIR_END_OFF)
 483                                         /*
 484                                          * End of directory reached.
 485                                          */
 486                                         done = 1;
 487                                 else if (1 /* chain is exhausted*/)
 488                                         /*
 489                                          * Normal case: continue to the next
 490                                          * page.
 491                                          */
 492                                         page = ll_get_dir_page(inode, pos, 1,
 493                                                                &chain);
 494                                 else {
 495                                         /*
 496                                          * go into overflow page.
 497                                          */
 498                                 }
 499                         } else
 500                                 pos = hash;
 501                 } else {
 502                         rc = PTR_ERR(page);
 503                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 504                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 505                 }
 506         }
 507
 508         filp->f_pos = (loff_t)(__s32)pos;
 509         filp->f_version = inode->i_version;
 510         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 511
 512         ll_dir_chain_fini(&chain);
 513
 514         RETURN(rc);
 515 }
 516
 517 #define QCTL_COPY(out, in)              \
 518 do {                                    \
 519         Q_COPY(out, in, qc_cmd);        \
 520         Q_COPY(out, in, qc_type);       \
 521         Q_COPY(out, in, qc_id);         \
 522         Q_COPY(out, in, qc_stat);       \
 523         Q_COPY(out, in, qc_dqinfo);     \
 524         Q_COPY(out, in, qc_dqblk);      \
 525 } while (0)
 526
 527 int ll_send_mgc_param(struct obd_export *mgc, char *string)
 528 {
 529         struct mgs_send_param *msp;
 530         int rc = 0;
 531
 532         OBD_ALLOC_PTR(msp);
 533         if (!msp)
 534                 return -ENOMEM;
 535
 536         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
 537         rc = obd_set_info_async(mgc, strlen(KEY_SET_INFO), KEY_SET_INFO,
 538                                 sizeof(struct mgs_send_param), msp, NULL);
 539         if (rc)
 540                 CERROR("Failed to set parameter: %d\n", rc);
 541         OBD_FREE_PTR(msp);
 542
 543         return rc;
 544 }
 545
 546 char *ll_get_fsname(struct inode *inode)
 547 {
 548         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 549         char *ptr, *fsname;
 550         int len;
 551
 552         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
 553         len = strlen(lsi->lsi_lmd->lmd_profile);
 554         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
 555         if (ptr && (strcmp(ptr, "-client") == 0))
 556                 len -= 7;
 557         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
 558         fsname[len] = '\0';
 559
 560         return fsname;
 561 }
 562
 563 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
 564                      int set_default)
 565 {
 566         struct ll_sb_info *sbi = ll_i2sbi(inode);
 567         struct md_op_data *op_data;
 568         struct ptlrpc_request *req = NULL;
 569         int rc = 0;
 570         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 571         struct obd_device *mgc = lsi->lsi_mgc;
 572         char *fsname = NULL, *param = NULL;
 573
 574         /*
 575          * This is coming from userspace, so should be in
 576          * local endian.  But the MDS would like it in little
 577          * endian, so we swab it before we send it.
 578          */
 579         if (lump->lmm_magic != LOV_USER_MAGIC)
 580                 RETURN(-EINVAL);
 581
 582         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC))
 583                 lustre_swab_lov_user_md(lump);
 584
 585         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 586                                      LUSTRE_OPC_ANY, NULL);
 587         if (IS_ERR(op_data))
 588                 RETURN(PTR_ERR(op_data));
 589
 590         /* swabbing is done in lov_setstripe() on server side */
 591         rc = md_setattr(sbi->ll_md_exp, op_data, lump, sizeof(*lump),
 592                         NULL, 0, &req, NULL);
 593         ll_finish_md_op_data(op_data);
 594         ptlrpc_req_finished(req);
 595         if (rc) {
 596                 if (rc != -EPERM && rc != -EACCES)
 597                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 598         }
 599
 600         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
 601                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
 602
 603                 /* Get fsname and assume devname to be -MDT0000. */
 604                 fsname = ll_get_fsname(inode);
 605                 /* Set root stripesize */
 606                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
 607                         lump->lmm_stripe_size);
 608                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 609                 if (rc)
 610                         goto end;
 611
 612                 /* Set root stripecount */
 613                 sprintf(param, "%s-MDT0000.lov.stripecount=%u", fsname,
 614                         lump->lmm_stripe_count);
 615                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 616                 if (rc)
 617                         goto end;
 618
 619                 /* Set root stripeoffset */
 620                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%u", fsname,
 621                         lump->lmm_stripe_offset);
 622                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 623                 if (rc)
 624                         goto end;
 625 end:
 626                 if (fsname)
 627                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
 628                 if (param)
 629                         OBD_FREE(param, MGS_PARAM_MAXLEN);
 630         }
 631         return rc;
 632 }
 633
 634 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 635                      int *lmm_size, struct ptlrpc_request **request)
 636 {
 637         struct ll_sb_info *sbi = ll_i2sbi(inode);
 638         struct mdt_body   *body;
 639         struct lov_mds_md *lmm = NULL;
 640         struct ptlrpc_request *req = NULL;
 641         int rc, lmmsize;
 642         struct obd_capa *oc;
 643
 644         rc = ll_get_max_mdsize(sbi, &lmmsize);
 645         if (rc)
 646                 RETURN(rc);
 647
 648         oc = ll_mdscapa_get(inode);
 649         rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode),
 650                         oc, OBD_MD_FLEASIZE | OBD_MD_FLDIREA,
 651                         lmmsize, &req);
 652         capa_put(oc);
 653         if (rc < 0) {
 654                 CDEBUG(D_INFO, "md_getattr failed on inode "
 655                        "%lu/%u: rc %d\n", inode->i_ino,
 656                        inode->i_generation, rc);
 657                 GOTO(out, rc);
 658         }
 659
 660         body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
 661         LASSERT(body != NULL); /* checked by md_getattr_name */
 662         /* swabbed by mdc_getattr_name */
 663         LASSERT(lustre_rep_swabbed(req, REPLY_REC_OFF));
 664
 665         lmmsize = body->eadatasize;
 666
 667         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 668             lmmsize == 0) {
 669                 GOTO(out, rc = -ENODATA);
 670         }
 671
 672         lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, lmmsize);
 673         LASSERT(lmm != NULL);
 674         LASSERT(lustre_rep_swabbed(req, REPLY_REC_OFF + 1));
 675
 676         /*
 677          * This is coming from the MDS, so is probably in
 678          * little endian.  We convert it to host endian before
 679          * passing it to userspace.
 680          */
 681         if (lmm->lmm_magic == __swab32(LOV_MAGIC)) {
 682                 lustre_swab_lov_user_md((struct lov_user_md *)lmm);
 683                 lustre_swab_lov_user_md_objects((struct lov_user_md *)lmm);
 684         }
 685 out:
 686         *lmmp = lmm;
 687         *lmm_size = lmmsize;
 688         *request = req;
 689         return rc;
 690 }
 691
 692 static int ll_dir_ioctl(struct inode *inode, struct file *file,
 693                         unsigned int cmd, unsigned long arg)
 694 {
 695         struct ll_sb_info *sbi = ll_i2sbi(inode);
 696         struct obd_ioctl_data *data;
 697         ENTRY;
 698
 699         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
 700                inode->i_ino, inode->i_generation, inode, cmd);
 701
 702         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
 703         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
 704                 return -ENOTTY;
 705
 706         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
 707         switch(cmd) {
 708         case EXT3_IOC_GETFLAGS:
 709         case EXT3_IOC_SETFLAGS:
 710                 RETURN(ll_iocontrol(inode, file, cmd, arg));
 711         case EXT3_IOC_GETVERSION_OLD:
 712         case EXT3_IOC_GETVERSION:
 713                 RETURN(put_user(inode->i_generation, (int *)arg));
 714         /* We need to special case any other ioctls we want to handle,
 715          * to send them to the MDS/OST as appropriate and to properly
 716          * network encode the arg field.
 717         case EXT3_IOC_SETVERSION_OLD:
 718         case EXT3_IOC_SETVERSION:
 719         */
 720         case IOC_MDC_LOOKUP: {
 721                 struct ptlrpc_request *request = NULL;
 722                 int namelen, rc, len = 0;
 723                 char *buf = NULL;
 724                 char *filename;
 725                 struct obd_capa *oc;
 726
 727                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 728                 if (rc)
 729                         RETURN(rc);
 730                 data = (void *)buf;
 731
 732                 filename = data->ioc_inlbuf1;
 733                 namelen = data->ioc_inllen1;
 734
 735                 if (namelen < 1) {
 736                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
 737                         GOTO(out, rc = -EINVAL);
 738                 }
 739
 740                 oc = ll_mdscapa_get(inode);
 741                 rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode), oc,
 742                                      filename, namelen, OBD_MD_FLID, 0,
 743                                      &request);
 744                 capa_put(oc);
 745                 if (rc < 0) {
 746                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 747                         GOTO(out, rc);
 748                 }
 749
 750                 ptlrpc_req_finished(request);
 751
 752                 EXIT;
 753         out:
 754                 obd_ioctl_freedata(buf, len);
 755                 return rc;
 756         }
 757         case LL_IOC_LOV_SETSTRIPE: {
 758                 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
 759                 int rc = 0;
 760                 int set_default = 0;
 761
 762                 LASSERT(sizeof(lum) == sizeof(*lump));
 763                 LASSERT(sizeof(lum.lmm_objects[0]) ==
 764                         sizeof(lump->lmm_objects[0]));
 765                 rc = copy_from_user(&lum, lump, sizeof(lum));
 766                 if (rc)
 767                         RETURN(-EFAULT);
 768
 769                 if (inode->i_sb->s_root == file->f_dentry)
 770                         set_default = 1;
 771
 772                 rc = ll_dir_setstripe(inode, &lum, set_default);
 773
 774                 RETURN(rc);
 775         }
 776         case LL_IOC_OBD_STATFS:
 777                 RETURN(ll_obd_statfs(inode, (void *)arg));
 778         case LL_IOC_LOV_GETSTRIPE:
 779         case LL_IOC_MDC_GETINFO:
 780         case IOC_MDC_GETFILEINFO:
 781         case IOC_MDC_GETFILESTRIPE: {
 782                 struct ptlrpc_request *request = NULL;
 783                 struct lov_user_md *lump;
 784                 struct lov_mds_md *lmm = NULL;
 785                 struct mdt_body *body;
 786                 char *filename = NULL;
 787                 int rc, lmmsize;
 788
 789                 if (cmd == IOC_MDC_GETFILEINFO ||
 790                     cmd == IOC_MDC_GETFILESTRIPE) {
 791                         filename = getname((const char *)arg);
 792                         if (IS_ERR(filename))
 793                                 RETURN(PTR_ERR(filename));
 794
 795                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
 796                                                       &lmmsize, &request);
 797                 } else {
 798                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
 799                 }
 800
 801                 if (request) {
 802                         body = lustre_msg_buf(request->rq_repmsg,
 803                                               REPLY_REC_OFF, sizeof(*body));
 804                         LASSERT(body != NULL); /* checked by md_getattr_name */
 805                         /* swabbed by md_getattr_name */
 806                         LASSERT(lustre_rep_swabbed(request, REPLY_REC_OFF));
 807                 } else {
 808                         GOTO(out_req, rc);
 809                 }
 810
 811                 if (rc < 0) {
 812                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
 813                                                cmd == LL_IOC_MDC_GETINFO))
 814                                 GOTO(skip_lmm, rc = 0);
 815                         else
 816                                 GOTO(out_req, rc);
 817                 }
 818
 819                 if (cmd == IOC_MDC_GETFILESTRIPE ||
 820                     cmd == LL_IOC_LOV_GETSTRIPE) {
 821                         lump = (struct lov_user_md *)arg;
 822                 } else {
 823                         struct lov_user_mds_data *lmdp;
 824                         lmdp = (struct lov_user_mds_data *)arg;
 825                         lump = &lmdp->lmd_lmm;
 826                 }
 827                 rc = copy_to_user(lump, lmm, lmmsize);
 828                 if (rc)
 829                         GOTO(out_lmm, rc = -EFAULT);
 830         skip_lmm:
 831                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
 832                         struct lov_user_mds_data *lmdp;
 833                         lstat_t st = { 0 };
 834
 835                         st.st_dev     = inode->i_sb->s_dev;
 836                         st.st_mode    = body->mode;
 837                         st.st_nlink   = body->nlink;
 838                         st.st_uid     = body->uid;
 839                         st.st_gid     = body->gid;
 840                         st.st_rdev    = body->rdev;
 841                         st.st_size    = body->size;
 842                         st.st_blksize = CFS_PAGE_SIZE;
 843                         st.st_blocks  = body->blocks;
 844                         st.st_atime   = body->atime;
 845                         st.st_mtime   = body->mtime;
 846                         st.st_ctime   = body->ctime;
 847                         st.st_ino     = inode->i_ino;
 848
 849                         lmdp = (struct lov_user_mds_data *)arg;
 850                         rc = copy_to_user(&lmdp->lmd_st, &st, sizeof(st));
 851                         if (rc)
 852                                 GOTO(out_lmm, rc = -EFAULT);
 853                 }
 854
 855                 EXIT;
 856         out_lmm:
 857                 if (lmm && lmm->lmm_magic == LOV_MAGIC_JOIN)
 858                         OBD_FREE(lmm, lmmsize);
 859         out_req:
 860                 ptlrpc_req_finished(request);
 861                 if (filename)
 862                         putname(filename);
 863                 return rc;
 864         }
 865         case IOC_LOV_GETINFO: {
 866                 struct lov_user_mds_data *lumd;
 867                 struct lov_stripe_md *lsm;
 868                 struct lov_user_md *lum;
 869                 struct lov_mds_md *lmm;
 870                 int lmmsize;
 871                 lstat_t st;
 872                 int rc;
 873
 874                 lumd = (struct lov_user_mds_data *)arg;
 875                 lum = &lumd->lmd_lmm;
 876
 877                 rc = ll_get_max_mdsize(sbi, &lmmsize);
 878                 if (rc)
 879                         RETURN(rc);
 880
 881                 OBD_ALLOC(lmm, lmmsize);
 882                 rc = copy_from_user(lmm, lum, lmmsize);
 883                 if (rc)
 884                         GOTO(free_lmm, rc = -EFAULT);
 885
 886                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
 887                 if (rc < 0)
 888                         GOTO(free_lmm, rc = -ENOMEM);
 889
 890                 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
 891                 if (rc)
 892                         GOTO(free_lsm, rc);
 893
 894                 /* Perform glimpse_size operation. */
 895                 memset(&st, 0, sizeof(st));
 896
 897                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
 898                 if (rc)
 899                         GOTO(free_lsm, rc);
 900
 901                 rc = copy_to_user(&lumd->lmd_st, &st, sizeof(st));
 902                 if (rc)
 903                         GOTO(free_lsm, rc = -EFAULT);
 904
 905                 EXIT;
 906         free_lsm:
 907                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
 908         free_lmm:
 909                 OBD_FREE(lmm, lmmsize);
 910                 return rc;
 911         }
 912         case OBD_IOC_LLOG_CATINFO: {
 913                 struct ptlrpc_request *req = NULL;
 914                 char *buf = NULL;
 915                 int rc, len = 0;
 916                 char *bufs[3] = { NULL }, *str;
 917                 int lens[3] = { sizeof(struct ptlrpc_body) };
 918                 int size[2] = { sizeof(struct ptlrpc_body) };
 919
 920                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 921                 if (rc)
 922                         RETURN(rc);
 923                 data = (void *)buf;
 924
 925                 if (!data->ioc_inlbuf1) {
 926                         obd_ioctl_freedata(buf, len);
 927                         RETURN(-EINVAL);
 928                 }
 929
 930                 lens[REQ_REC_OFF] = data->ioc_inllen1;
 931                 bufs[REQ_REC_OFF] = data->ioc_inlbuf1;
 932                 if (data->ioc_inllen2) {
 933                         lens[REQ_REC_OFF + 1] = data->ioc_inllen2;
 934                         bufs[REQ_REC_OFF + 1] = data->ioc_inlbuf2;
 935                 } else {
 936                         lens[REQ_REC_OFF + 1] = 0;
 937                         bufs[REQ_REC_OFF + 1] = NULL;
 938                 }
 939
 940                 req = ptlrpc_prep_req(sbi2mdc(sbi)->cl_import,
 941                                       LUSTRE_LOG_VERSION, LLOG_CATINFO, 3, lens,
 942                                       bufs);
 943                 if (!req)
 944                         GOTO(out_catinfo, rc = -ENOMEM);
 945
 946                 size[REPLY_REC_OFF] = data->ioc_plen1;
 947                 ptlrpc_req_set_repsize(req, 2, size);
 948
 949                 rc = ptlrpc_queue_wait(req);
 950                 if (!rc) {
 951                         str = lustre_msg_string(req->rq_repmsg, REPLY_REC_OFF,
 952                                                 data->ioc_plen1);
 953                         rc = copy_to_user(data->ioc_pbuf1, str, data->ioc_plen1);
 954                 }
 955                 ptlrpc_req_finished(req);
 956         out_catinfo:
 957                 obd_ioctl_freedata(buf, len);
 958                 RETURN(rc);
 959         }
 960         case OBD_IOC_QUOTACHECK: {
 961                 struct obd_quotactl *oqctl;
 962                 int rc, error = 0;
 963
 964                 if (!capable(CAP_SYS_ADMIN))
 965                         RETURN(-EPERM);
 966
 967                 OBD_ALLOC_PTR(oqctl);
 968                 if (!oqctl)
 969                         RETURN(-ENOMEM);
 970                 oqctl->qc_type = arg;
 971                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
 972                 if (rc < 0) {
 973                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
 974                         error = rc;
 975                 }
 976
 977                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
 978                 if (rc < 0)
 979                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
 980
 981                 OBD_FREE_PTR(oqctl);
 982                 return error ?: rc;
 983         }
 984         case OBD_IOC_POLL_QUOTACHECK: {
 985                 struct if_quotacheck *check;
 986                 int rc;
 987
 988                 if (!capable(CAP_SYS_ADMIN))
 989                         RETURN(-EPERM);
 990
 991                 OBD_ALLOC_PTR(check);
 992                 if (!check)
 993                         RETURN(-ENOMEM);
 994
 995                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
 996                                    NULL);
 997                 if (rc) {
 998                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
 999                         if (copy_to_user((void *)arg, check, sizeof(*check)))
1000                                 rc = -EFAULT;
1001                         GOTO(out_poll, rc);
1002                 }
1003
1004                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1005                                    NULL);
1006                 if (rc) {
1007                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1008                         if (copy_to_user((void *)arg, check, sizeof(*check)))
1009                                 rc = -EFAULT;
1010                         GOTO(out_poll, rc);
1011                 }
1012         out_poll:
1013                 OBD_FREE_PTR(check);
1014                 RETURN(rc);
1015         }
1016 #ifdef HAVE_QUOTA_SUPPORT
1017         case OBD_IOC_QUOTACTL: {
1018                 struct if_quotactl *qctl;
1019                 struct obd_quotactl *oqctl;
1020
1021                 int cmd, type, id, rc = 0;
1022
1023                 OBD_ALLOC_PTR(qctl);
1024                 if (!qctl)
1025                         RETURN(-ENOMEM);
1026
1027                 OBD_ALLOC_PTR(oqctl);
1028                 if (!oqctl) {
1029                         OBD_FREE_PTR(qctl);
1030                         RETURN(-ENOMEM);
1031                 }
1032                 if (copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1033                         GOTO(out_quotactl, rc = -EFAULT);
1034
1035                 cmd = qctl->qc_cmd;
1036                 type = qctl->qc_type;
1037                 id = qctl->qc_id;
1038                 switch (cmd) {
1039                 case Q_QUOTAON:
1040                 case Q_QUOTAOFF:
1041                 case Q_SETQUOTA:
1042                 case Q_SETINFO:
1043                         if (!capable(CAP_SYS_ADMIN))
1044                                 GOTO(out_quotactl, rc = -EPERM);
1045                         break;
1046                 case Q_GETQUOTA:
1047                         if (((type == USRQUOTA && current->euid != id) ||
1048                              (type == GRPQUOTA && !in_egroup_p(id))) &&
1049                             !capable(CAP_SYS_ADMIN))
1050                                 GOTO(out_quotactl, rc = -EPERM);
1051
1052                         /* XXX: dqb_valid is borrowed as a flag to mark that
1053                          *      only mds quota is wanted */
1054                         if (qctl->qc_dqblk.dqb_valid)
1055                                 qctl->obd_uuid = sbi->ll_md_exp->exp_obd->
1056                                                         u.cli.cl_target_uuid;
1057                         break;
1058                 case Q_GETINFO:
1059                         break;
1060                 default:
1061                         CERROR("unsupported quotactl op: %#x\n", cmd);
1062                         GOTO(out_quotactl, rc = -ENOTTY);
1063                 }
1064
1065                 QCTL_COPY(oqctl, qctl);
1066
1067                 if (qctl->obd_uuid.uuid[0]) {
1068                         struct obd_device *obd;
1069                         struct obd_uuid *uuid = &qctl->obd_uuid;
1070
1071                         obd = class_find_client_notype(uuid,
1072                                          &sbi->ll_dt_exp->exp_obd->obd_uuid);
1073                         if (!obd)
1074                                 GOTO(out_quotactl, rc = -ENOENT);
1075
1076                         if (cmd == Q_GETINFO)
1077                                 oqctl->qc_cmd = Q_GETOINFO;
1078                         else if (cmd == Q_GETQUOTA)
1079                                 oqctl->qc_cmd = Q_GETOQUOTA;
1080                         else
1081                                 GOTO(out_quotactl, rc = -EINVAL);
1082
1083                         if (sbi->ll_md_exp->exp_obd == obd) {
1084                                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1085                         } else {
1086                                 int i;
1087                                 struct obd_export *exp;
1088                                 struct lov_obd *lov = &sbi->ll_dt_exp->
1089                                                             exp_obd->u.lov;
1090
1091                                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1092                                         if (!lov->lov_tgts[i] ||
1093                                             !lov->lov_tgts[i]->ltd_active)
1094                                                 continue;
1095                                         exp = lov->lov_tgts[i]->ltd_exp;
1096                                         if (exp->exp_obd == obd) {
1097                                                 rc = obd_quotactl(exp, oqctl);
1098                                                 break;
1099                                         }
1100                                 }
1101                         }
1102
1103                         oqctl->qc_cmd = cmd;
1104                         QCTL_COPY(qctl, oqctl);
1105
1106                         if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1107                                 rc = -EFAULT;
1108
1109                         GOTO(out_quotactl, rc);
1110                 }
1111
1112                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1113                 if (rc && rc != -EBUSY && cmd == Q_QUOTAON) {
1114                         oqctl->qc_cmd = Q_QUOTAOFF;
1115                         obd_quotactl(sbi->ll_md_exp, oqctl);
1116                 }
1117
1118                 QCTL_COPY(qctl, oqctl);
1119
1120                 if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1121                         rc = -EFAULT;
1122         out_quotactl:
1123                 OBD_FREE_PTR(qctl);
1124                 OBD_FREE_PTR(oqctl);
1125                 RETURN(rc);
1126         }
1127 #endif /* HAVE_QUOTA_SUPPORT */
1128         case OBD_IOC_GETNAME: {
1129                 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1130                 if (!obd)
1131                         RETURN(-EFAULT);
1132                 if (copy_to_user((void *)arg, obd->obd_name,
1133                                 strlen(obd->obd_name) + 1))
1134                         RETURN (-EFAULT);
1135                 RETURN(0);
1136         }
1137         case LL_IOC_FLUSHCTX:
1138                 RETURN(ll_flush_ctx(inode));
1139         case LL_IOC_GETFACL: {
1140                 struct rmtacl_ioctl_data ioc;
1141
1142                 if (copy_from_user(&ioc, (void *)arg, sizeof(ioc)))
1143                         RETURN(-EFAULT);
1144
1145                 RETURN(ll_ioctl_getfacl(inode, &ioc));
1146         }
1147         case LL_IOC_SETFACL: {
1148                 struct rmtacl_ioctl_data ioc;
1149
1150                 if (copy_from_user(&ioc, (void *)arg, sizeof(ioc)))
1151                         RETURN(-EFAULT);
1152
1153                 RETURN(ll_ioctl_setfacl(inode, &ioc));
1154         }
1155         default:
1156                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1157         }
1158 }
1159
1160 int ll_dir_open(struct inode *inode, struct file *file)
1161 {
1162         ENTRY;
1163         RETURN(ll_file_open(inode, file));
1164 }
1165
1166 int ll_dir_release(struct inode *inode, struct file *file)
1167 {
1168         ENTRY;
1169         RETURN(ll_file_release(inode, file));
1170 }
1171
1172 struct file_operations ll_dir_operations = {
1173         .open     = ll_dir_open,
1174         .release  = ll_dir_release,
1175         .read     = generic_read_dir,
1176         .readdir  = ll_readdir,
1177         .ioctl    = ll_dir_ioctl
1178 };
1179