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