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/pagemap.h>
  31 #include <linux/mm.h>
  32 #include <linux/version.h>
  33 #include <linux/smp_lock.h>
  34 #include <asm/uaccess.h>
  35 #include <linux/buffer_head.h>   // for wait_on_buffer
  36
  37 #define DEBUG_SUBSYSTEM S_LLITE
  38
  39 #include <obd_support.h>
  40 #include <obd_class.h>
  41 #include <lustre_lib.h>
  42 #include <lustre/lustre_idl.h>
  43 #include <lustre_lite.h>
  44 #include <lustre_dlm.h>
  45 #include <lustre_fid.h>
  46 #include "llite_internal.h"
  47
  48 #ifndef HAVE_PAGE_CHECKED
  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 #else
  53 #error PageChecked or PageFsMisc not defined in kernel
  54 #endif
  55 #endif
  56
  57 /*
  58  * (new) readdir implementation overview.
  59  *
  60  * Original lustre readdir implementation cached exact copy of raw directory
  61  * pages on the client. These pages were indexed in client page cache by
  62  * logical offset in the directory file. This design, while very simple and
  63  * intuitive had some inherent problems:
  64  *
  65  *     . it implies that byte offset to the directory entry serves as a
  66  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  67  *     ext3/htree directory entries may move due to splits, and more
  68  *     importantly,
  69  *
  70  *     . it is incompatible with the design of split directories for cmd3,
  71  *     that assumes that names are distributed across nodes based on their
  72  *     hash, and so readdir should be done in hash order.
  73  *
  74  * New readdir implementation does readdir in hash order, and uses hash of a
  75  * file name as a telldir/seekdir cookie. This led to number of complications:
  76  *
  77  *     . hash is not unique, so it cannot be used to index cached directory
  78  *     pages on the client (note, that it requires a whole pageful of hash
  79  *     collided entries to cause two pages to have identical hashes);
  80  *
  81  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  82  *     entry cookie. ext3/htree has the same problem and lustre implementation
  83  *     mimics their solution: seekdir(hash) positions directory at the first
  84  *     entry with the given hash.
  85  *
  86  * Client side.
  87  *
  88  * 0. caching
  89  *
  90  * Client caches directory pages using hash of the first entry as an index. As
  91  * noted above hash is not unique, so this solution doesn't work as is:
  92  * special processing is needed for "page hash chains" (i.e., sequences of
  93  * pages filled with entries all having the same hash value).
  94  *
  95  * First, such chains have to be detected. To this end, server returns to the
  96  * client the hash of the first entry on the page next to one returned. When
  97  * client detects that this hash is the same as hash of the first entry on the
  98  * returned page, page hash collision has to be handled. Pages in the
  99  * hash chain, except first one, are termed "overflow pages".
 100  *
 101  * Solution to index uniqueness problem is to not cache overflow
 102  * pages. Instead, when page hash collision is detected, all overflow pages
 103  * from emerging chain are immediately requested from the server and placed in
 104  * a special data structure (struct ll_dir_chain). This data structure is used
 105  * by ll_readdir() to process entries from overflow pages. When readdir
 106  * invocation finishes, overflow pages are discarded. If page hash collision
 107  * chain weren't completely processed, next call to readdir will again detect
 108  * page hash collision, again read overflow pages in, process next portion of
 109  * entries and again discard the pages. This is not as wasteful as it looks,
 110  * because, given reasonable hash, page hash collisions are extremely rare.
 111  *
 112  * 1. directory positioning
 113  *
 114  * When seekdir(hash) is called, original
 115  *
 116  *
 117  *
 118  *
 119  *
 120  *
 121  *
 122  *
 123  * Server.
 124  *
 125  * identification of and access to overflow pages
 126  *
 127  * page format
 128  *
 129  *
 130  *
 131  *
 132  *
 133  */
 134
 135 /* returns the page unlocked, but with a reference */
 136 static int ll_dir_readpage(struct file *file, struct page *page)
 137 {
 138         struct inode *inode = page->mapping->host;
 139         struct ptlrpc_request *request;
 140         struct mdt_body *body;
 141         struct obd_capa *oc;
 142         __u64 hash;
 143         int rc;
 144         ENTRY;
 145
 146         hash = (__u64)hash_x_index(page->index);
 147         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) off %lu\n",
 148                inode->i_ino, inode->i_generation, inode, (unsigned long)hash);
 149
 150         oc = ll_mdscapa_get(inode);
 151         rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
 152                          oc, hash, page, &request);
 153         capa_put(oc);
 154         if (!rc) {
 155                 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 156                 /* Checked by mdc_readpage() */
 157                 LASSERT(body != NULL);
 158
 159                 if (body->valid & OBD_MD_FLSIZE) {
 160                         ll_inode_size_lock(inode, 0);
 161                         i_size_write(inode, body->size);
 162                         ll_inode_size_unlock(inode, 0);
 163                 }
 164                 SetPageUptodate(page);
 165         }
 166         ptlrpc_req_finished(request);
 167
 168         unlock_page(page);
 169         EXIT;
 170         return rc;
 171 }
 172
 173 struct address_space_operations ll_dir_aops = {
 174         .readpage  = ll_dir_readpage,
 175 };
 176
 177 static void ll_check_page(struct inode *dir, struct page *page)
 178 {
 179         /* XXX: check page format later */
 180         SetPageChecked(page);
 181 }
 182
 183 static void ll_release_page(struct page *page, __u64 hash,
 184                             __u64 start, __u64 end)
 185 {
 186         kunmap(page);
 187         lock_page(page);
 188         if (likely(page->mapping != NULL)) {
 189                 ll_truncate_complete_page(page);
 190                 unlock_page(page);
 191         } else {
 192                 unlock_page(page);
 193                 CWARN("NULL mapping page %p, truncated by others: "
 194                       "hash(%#llx) | start(%#llx) | end(%#llx)\n",
 195                       page, hash, start, end);
 196         }
 197         page_cache_release(page);
 198 }
 199
 200 /*
 201  * Find, kmap and return page that contains given hash.
 202  */
 203 static struct page *ll_dir_page_locate(struct inode *dir, __u64 hash,
 204                                        __u64 *start, __u64 *end)
 205 {
 206         struct address_space *mapping = dir->i_mapping;
 207         /*
 208          * Complement of hash is used as an index so that
 209          * radix_tree_gang_lookup() can be used to find a page with starting
 210          * hash _smaller_ than one we are looking for.
 211          */
 212         unsigned long offset = hash_x_index((__u32)hash);
 213         struct page *page;
 214         int found;
 215
 216         TREE_READ_LOCK_IRQ(mapping);
 217         found = radix_tree_gang_lookup(&mapping->page_tree,
 218                                        (void **)&page, offset, 1);
 219         if (found > 0) {
 220                 struct lu_dirpage *dp;
 221
 222                 page_cache_get(page);
 223                 TREE_READ_UNLOCK_IRQ(mapping);
 224                 /*
 225                  * In contrast to find_lock_page() we are sure that directory
 226                  * page cannot be truncated (while DLM lock is held) and,
 227                  * hence, can avoid restart.
 228                  *
 229                  * In fact, page cannot be locked here at all, because
 230                  * ll_dir_readpage() does synchronous io.
 231                  */
 232                 wait_on_page(page);
 233                 if (PageUptodate(page)) {
 234                         dp = kmap(page);
 235                         *start = le64_to_cpu(dp->ldp_hash_start);
 236                         *end   = le64_to_cpu(dp->ldp_hash_end);
 237                         LASSERT(*start <= hash);
 238                         if (hash > *end || (*end != *start && hash == *end)) {
 239                                 ll_release_page(page, hash, *start, *end);
 240                                 page = NULL;
 241                         }
 242                 } else {
 243                         page_cache_release(page);
 244                         page = ERR_PTR(-EIO);
 245                 }
 246
 247         } else {
 248                 TREE_READ_UNLOCK_IRQ(mapping);
 249                 page = NULL;
 250         }
 251         return page;
 252 }
 253
 254 struct page *ll_get_dir_page(struct inode *dir, __u64 hash, int exact,
 255                              struct ll_dir_chain *chain)
 256 {
 257         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 258         struct address_space *mapping = dir->i_mapping;
 259         struct lustre_handle lockh;
 260         struct lu_dirpage *dp;
 261         struct page *page;
 262         ldlm_mode_t mode;
 263         int rc;
 264         __u64 start;
 265         __u64 end;
 266
 267         mode = LCK_PR;
 268         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 269                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
 270         if (!rc) {
 271                 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
 272                        ll_md_blocking_ast, ldlm_completion_ast, NULL, dir };
 273                 struct lookup_intent it = { .it_op = IT_READDIR };
 274                 struct ptlrpc_request *request;
 275                 struct md_op_data *op_data;
 276
 277                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
 278                                              LUSTRE_OPC_ANY, NULL);
 279                 if (IS_ERR(op_data))
 280                         return (void *)op_data;
 281
 282                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
 283                                 op_data, &lockh, NULL, 0, 0);
 284
 285                 ll_finish_md_op_data(op_data);
 286
 287                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 288                 if (request)
 289                         ptlrpc_req_finished(request);
 290                 if (rc < 0) {
 291                         CERROR("lock enqueue: rc: %d\n", rc);
 292                         return ERR_PTR(rc);
 293                 }
 294         } else {
 295                 /* for cross-ref object, l_ast_data of the lock may not be set,
 296                  * we reset it here */
 297                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie, dir);
 298         }
 299         ldlm_lock_dump_handle(D_OTHER, &lockh);
 300
 301         page = ll_dir_page_locate(dir, hash, &start, &end);
 302         if (IS_ERR(page))
 303                 GOTO(out_unlock, page);
 304
 305         if (page != NULL) {
 306                 /*
 307                  * XXX nikita: not entirely correct handling of a corner case:
 308                  * suppose hash chain of entries with hash value HASH crosses
 309                  * border between pages P0 and P1. First both P0 and P1 are
 310                  * cached, seekdir() is called for some entry from the P0 part
 311                  * of the chain. Later P0 goes out of cache. telldir(HASH)
 312                  * happens and finds P1, as it starts with matching hash
 313                  * value. Remaining entries from P0 part of the chain are
 314                  * skipped. (Is that really a bug?)
 315                  *
 316                  * Possible solutions: 0. don't cache P1 is such case, handle
 317                  * it as an "overflow" page. 1. invalidate all pages at
 318                  * once. 2. use HASH|1 as an index for P1.
 319                  */
 320                 if (exact && hash != start) {
 321                         /*
 322                          * readdir asked for a page starting _exactly_ from
 323                          * given hash, but cache contains stale page, with
 324                          * entries with smaller hash values. Stale page should
 325                          * be invalidated, and new one fetched.
 326                          */
 327                         CWARN("Stale readpage page %p: %#llx != %#llx\n",
 328                               page, hash, start);
 329                         ll_release_page(page, hash, start, end);
 330                 } else {
 331                         GOTO(hash_collision, page);
 332                 }
 333         }
 334
 335         page = read_cache_page(mapping, hash_x_index((__u32)hash),
 336                                (filler_t*)mapping->a_ops->readpage, NULL);
 337         if (IS_ERR(page))
 338                 GOTO(out_unlock, page);
 339
 340         wait_on_page(page);
 341         (void)kmap(page);
 342         if (!PageUptodate(page))
 343                 goto fail;
 344         if (!PageChecked(page))
 345                 ll_check_page(dir, page);
 346         if (PageError(page))
 347                 goto fail;
 348 hash_collision:
 349         dp = page_address(page);
 350
 351         start = le64_to_cpu(dp->ldp_hash_start);
 352         end   = le64_to_cpu(dp->ldp_hash_end);
 353         if (end == start) {
 354                 LASSERT(start == hash);
 355                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
 356                 /*
 357                  * Fetch whole overflow chain...
 358                  *
 359                  * XXX not yet.
 360                  */
 361                 goto fail;
 362         }
 363 out_unlock:
 364         ldlm_lock_decref(&lockh, mode);
 365         return page;
 366
 367 fail:
 368         ll_put_page(page);
 369         page = ERR_PTR(-EIO);
 370         goto out_unlock;
 371 }
 372
 373 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 374 {
 375         struct inode         *inode = filp->f_dentry->d_inode;
 376         struct ll_inode_info *info  = ll_i2info(inode);
 377         struct ll_sb_info    *sbi   = ll_i2sbi(inode);
 378         __u64                 pos   = filp->f_pos;
 379         struct page          *page;
 380         struct ll_dir_chain   chain;
 381         int rc;
 382         int done;
 383         int shift;
 384         ENTRY;
 385
 386         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu\n",
 387                inode->i_ino, inode->i_generation, inode,
 388                (unsigned long)pos, i_size_read(inode));
 389
 390         if (pos == DIR_END_OFF)
 391                 /*
 392                  * end-of-file.
 393                  */
 394                 RETURN(0);
 395
 396         rc    = 0;
 397         done  = 0;
 398         shift = 0;
 399         ll_dir_chain_init(&chain);
 400
 401         page = ll_get_dir_page(inode, pos, 0, &chain);
 402
 403         while (rc == 0 && !done) {
 404                 struct lu_dirpage *dp;
 405                 struct lu_dirent  *ent;
 406
 407                 if (!IS_ERR(page)) {
 408                         /*
 409                          * If page is empty (end of directoryis reached),
 410                          * use this value.
 411                          */
 412                         __u64 hash = DIR_END_OFF;
 413                         __u64 next;
 414
 415                         dp = page_address(page);
 416                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 417                              ent = lu_dirent_next(ent)) {
 418                                 char          *name;
 419                                 int            namelen;
 420                                 struct lu_fid  fid;
 421                                 ino_t          ino;
 422
 423                                 /*
 424                                  * XXX: implement correct swabbing here.
 425                                  */
 426
 427                                 hash    = le64_to_cpu(ent->lde_hash);
 428                                 namelen = le16_to_cpu(ent->lde_namelen);
 429
 430                                 if (hash < pos)
 431                                         /*
 432                                          * Skip until we find target hash
 433                                          * value.
 434                                          */
 435                                         continue;
 436
 437                                 if (namelen == 0)
 438                                         /*
 439                                          * Skip dummy record.
 440                                          */
 441                                         continue;
 442
 443                                 fid  = ent->lde_fid;
 444                                 name = ent->lde_name;
 445                                 fid_le_to_cpu(&fid, &fid);
 446                                 ino  = ll_fid_build_ino(sbi, &fid);
 447
 448                                 done = filldir(cookie, name, namelen,
 449                                                (loff_t)hash, ino, DT_UNKNOWN);
 450                         }
 451                         next = le64_to_cpu(dp->ldp_hash_end);
 452                         ll_put_page(page);
 453                         if (!done) {
 454                                 pos = next;
 455                                 if (pos == DIR_END_OFF)
 456                                         /*
 457                                          * End of directory reached.
 458                                          */
 459                                         done = 1;
 460                                 else if (1 /* chain is exhausted*/)
 461                                         /*
 462                                          * Normal case: continue to the next
 463                                          * page.
 464                                          */
 465                                         page = ll_get_dir_page(inode, pos, 1,
 466                                                                &chain);
 467                                 else {
 468                                         /*
 469                                          * go into overflow page.
 470                                          */
 471                                 }
 472                         } else
 473                                 pos = hash;
 474                 } else {
 475                         rc = PTR_ERR(page);
 476                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 477                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 478                 }
 479         }
 480
 481         filp->f_pos = (loff_t)(__s32)pos;
 482         filp->f_version = inode->i_version;
 483         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 484
 485         ll_dir_chain_fini(&chain);
 486
 487         RETURN(rc);
 488 }
 489
 490 #define QCTL_COPY(out, in)              \
 491 do {                                    \
 492         Q_COPY(out, in, qc_cmd);        \
 493         Q_COPY(out, in, qc_type);       \
 494         Q_COPY(out, in, qc_id);         \
 495         Q_COPY(out, in, qc_stat);       \
 496         Q_COPY(out, in, qc_dqinfo);     \
 497         Q_COPY(out, in, qc_dqblk);      \
 498 } while (0)
 499
 500 int ll_send_mgc_param(struct obd_export *mgc, char *string)
 501 {
 502         struct mgs_send_param *msp;
 503         int rc = 0;
 504
 505         OBD_ALLOC_PTR(msp);
 506         if (!msp)
 507                 return -ENOMEM;
 508
 509         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
 510         rc = obd_set_info_async(mgc, strlen(KEY_SET_INFO), KEY_SET_INFO,
 511                                 sizeof(struct mgs_send_param), msp, NULL);
 512         if (rc)
 513                 CERROR("Failed to set parameter: %d\n", rc);
 514         OBD_FREE_PTR(msp);
 515
 516         return rc;
 517 }
 518
 519 char *ll_get_fsname(struct inode *inode)
 520 {
 521         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 522         char *ptr, *fsname;
 523         int len;
 524
 525         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
 526         len = strlen(lsi->lsi_lmd->lmd_profile);
 527         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
 528         if (ptr && (strcmp(ptr, "-client") == 0))
 529                 len -= 7;
 530         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
 531         fsname[len] = '\0';
 532
 533         return fsname;
 534 }
 535
 536 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
 537                      int set_default)
 538 {
 539         struct ll_sb_info *sbi = ll_i2sbi(inode);
 540         struct md_op_data *op_data;
 541         struct ptlrpc_request *req = NULL;
 542         int rc = 0;
 543         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 544         struct obd_device *mgc = lsi->lsi_mgc;
 545         char *fsname = NULL, *param = NULL;
 546
 547         /*
 548          * This is coming from userspace, so should be in
 549          * local endian.  But the MDS would like it in little
 550          * endian, so we swab it before we send it.
 551          */
 552         if (lump->lmm_magic != LOV_USER_MAGIC)
 553                 RETURN(-EINVAL);
 554
 555         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC))
 556                 lustre_swab_lov_user_md(lump);
 557
 558         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 559                                      LUSTRE_OPC_ANY, NULL);
 560         if (IS_ERR(op_data))
 561                 RETURN(PTR_ERR(op_data));
 562
 563         /* swabbing is done in lov_setstripe() on server side */
 564         rc = md_setattr(sbi->ll_md_exp, op_data, lump, sizeof(*lump),
 565                         NULL, 0, &req, NULL);
 566         ll_finish_md_op_data(op_data);
 567         ptlrpc_req_finished(req);
 568         if (rc) {
 569                 if (rc != -EPERM && rc != -EACCES)
 570                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 571         }
 572
 573         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
 574                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
 575
 576                 /* Get fsname and assume devname to be -MDT0000. */
 577                 fsname = ll_get_fsname(inode);
 578                 /* Set root stripesize */
 579                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
 580                         lump->lmm_stripe_size);
 581                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 582                 if (rc)
 583                         goto end;
 584
 585                 /* Set root stripecount */
 586                 sprintf(param, "%s-MDT0000.lov.stripecount=%u", fsname,
 587                         lump->lmm_stripe_count);
 588                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 589                 if (rc)
 590                         goto end;
 591
 592                 /* Set root stripeoffset */
 593                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%u", fsname,
 594                         lump->lmm_stripe_offset);
 595                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 596                 if (rc)
 597                         goto end;
 598 end:
 599                 if (fsname)
 600                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
 601                 if (param)
 602                         OBD_FREE(param, MGS_PARAM_MAXLEN);
 603         }
 604         return rc;
 605 }
 606
 607 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 608                      int *lmm_size, struct ptlrpc_request **request)
 609 {
 610         struct ll_sb_info *sbi = ll_i2sbi(inode);
 611         struct mdt_body   *body;
 612         struct lov_mds_md *lmm = NULL;
 613         struct ptlrpc_request *req = NULL;
 614         int rc, lmmsize;
 615         struct obd_capa *oc;
 616
 617         rc = ll_get_max_mdsize(sbi, &lmmsize);
 618         if (rc)
 619                 RETURN(rc);
 620
 621         oc = ll_mdscapa_get(inode);
 622         rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode),
 623                         oc, OBD_MD_FLEASIZE | OBD_MD_FLDIREA,
 624                         lmmsize, &req);
 625         capa_put(oc);
 626         if (rc < 0) {
 627                 CDEBUG(D_INFO, "md_getattr failed on inode "
 628                        "%lu/%u: rc %d\n", inode->i_ino,
 629                        inode->i_generation, rc);
 630                 GOTO(out, rc);
 631         }
 632
 633         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 634         LASSERT(body != NULL);
 635
 636         lmmsize = body->eadatasize;
 637
 638         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 639             lmmsize == 0) {
 640                 GOTO(out, rc = -ENODATA);
 641         }
 642
 643         lmm = req_capsule_server_sized_get(&req->rq_pill,
 644                                            &RMF_MDT_MD, lmmsize);
 645         LASSERT(lmm != NULL);
 646
 647         /*
 648          * This is coming from the MDS, so is probably in
 649          * little endian.  We convert it to host endian before
 650          * passing it to userspace.
 651          */
 652         if (lmm->lmm_magic == __swab32(LOV_MAGIC)) {
 653                 lustre_swab_lov_user_md((struct lov_user_md *)lmm);
 654         }
 655 out:
 656         *lmmp = lmm;
 657         *lmm_size = lmmsize;
 658         *request = req;
 659         return rc;
 660 }
 661
 662 static int ll_dir_ioctl(struct inode *inode, struct file *file,
 663                         unsigned int cmd, unsigned long arg)
 664 {
 665         struct ll_sb_info *sbi = ll_i2sbi(inode);
 666         struct obd_ioctl_data *data;
 667         ENTRY;
 668
 669         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
 670                inode->i_ino, inode->i_generation, inode, cmd);
 671
 672         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
 673         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
 674                 return -ENOTTY;
 675
 676         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
 677         switch(cmd) {
 678         case EXT3_IOC_GETFLAGS:
 679         case EXT3_IOC_SETFLAGS:
 680                 RETURN(ll_iocontrol(inode, file, cmd, arg));
 681         case EXT3_IOC_GETVERSION_OLD:
 682         case EXT3_IOC_GETVERSION:
 683                 RETURN(put_user(inode->i_generation, (int *)arg));
 684         /* We need to special case any other ioctls we want to handle,
 685          * to send them to the MDS/OST as appropriate and to properly
 686          * network encode the arg field.
 687         case EXT3_IOC_SETVERSION_OLD:
 688         case EXT3_IOC_SETVERSION:
 689         */
 690         case IOC_MDC_LOOKUP: {
 691                 struct ptlrpc_request *request = NULL;
 692                 int namelen, rc, len = 0;
 693                 char *buf = NULL;
 694                 char *filename;
 695                 struct obd_capa *oc;
 696
 697                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 698                 if (rc)
 699                         RETURN(rc);
 700                 data = (void *)buf;
 701
 702                 filename = data->ioc_inlbuf1;
 703                 namelen = data->ioc_inllen1;
 704
 705                 if (namelen < 1) {
 706                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
 707                         GOTO(out, rc = -EINVAL);
 708                 }
 709
 710                 oc = ll_mdscapa_get(inode);
 711                 rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode), oc,
 712                                      filename, namelen, OBD_MD_FLID, 0,
 713                                      ll_i2suppgid(inode), &request);
 714                 capa_put(oc);
 715                 if (rc < 0) {
 716                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 717                         GOTO(out, rc);
 718                 }
 719
 720                 ptlrpc_req_finished(request);
 721
 722                 EXIT;
 723         out:
 724                 obd_ioctl_freedata(buf, len);
 725                 return rc;
 726         }
 727         case LL_IOC_LOV_SETSTRIPE: {
 728                 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
 729                 int rc = 0;
 730                 int set_default = 0;
 731
 732                 LASSERT(sizeof(lum) == sizeof(*lump));
 733                 LASSERT(sizeof(lum.lmm_objects[0]) ==
 734                         sizeof(lump->lmm_objects[0]));
 735                 rc = copy_from_user(&lum, lump, sizeof(lum));
 736                 if (rc)
 737                         RETURN(-EFAULT);
 738
 739                 if (inode->i_sb->s_root == file->f_dentry)
 740                         set_default = 1;
 741
 742                 rc = ll_dir_setstripe(inode, &lum, set_default);
 743
 744                 RETURN(rc);
 745         }
 746         case LL_IOC_OBD_STATFS:
 747                 RETURN(ll_obd_statfs(inode, (void *)arg));
 748         case LL_IOC_LOV_GETSTRIPE:
 749         case LL_IOC_MDC_GETINFO:
 750         case IOC_MDC_GETFILEINFO:
 751         case IOC_MDC_GETFILESTRIPE: {
 752                 struct ptlrpc_request *request = NULL;
 753                 struct lov_user_md *lump;
 754                 struct lov_mds_md *lmm = NULL;
 755                 struct mdt_body *body;
 756                 char *filename = NULL;
 757                 int rc, lmmsize;
 758
 759                 if (cmd == IOC_MDC_GETFILEINFO ||
 760                     cmd == IOC_MDC_GETFILESTRIPE) {
 761                         filename = getname((const char *)arg);
 762                         if (IS_ERR(filename))
 763                                 RETURN(PTR_ERR(filename));
 764
 765                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
 766                                                       &lmmsize, &request);
 767                 } else {
 768                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
 769                 }
 770
 771                 if (request) {
 772                         body = req_capsule_server_get(&request->rq_pill,
 773                                                       &RMF_MDT_BODY);
 774                         LASSERT(body != NULL);
 775                 } else {
 776                         GOTO(out_req, rc);
 777                 }
 778
 779                 if (rc < 0) {
 780                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
 781                                                cmd == LL_IOC_MDC_GETINFO))
 782                                 GOTO(skip_lmm, rc = 0);
 783                         else
 784                                 GOTO(out_req, rc);
 785                 }
 786
 787                 if (cmd == IOC_MDC_GETFILESTRIPE ||
 788                     cmd == LL_IOC_LOV_GETSTRIPE) {
 789                         lump = (struct lov_user_md *)arg;
 790                 } else {
 791                         struct lov_user_mds_data *lmdp;
 792                         lmdp = (struct lov_user_mds_data *)arg;
 793                         lump = &lmdp->lmd_lmm;
 794                 }
 795                 rc = copy_to_user(lump, lmm, lmmsize);
 796                 if (rc)
 797                         GOTO(out_lmm, rc = -EFAULT);
 798         skip_lmm:
 799                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
 800                         struct lov_user_mds_data *lmdp;
 801                         lstat_t st = { 0 };
 802
 803                         st.st_dev     = inode->i_sb->s_dev;
 804                         st.st_mode    = body->mode;
 805                         st.st_nlink   = body->nlink;
 806                         st.st_uid     = body->uid;
 807                         st.st_gid     = body->gid;
 808                         st.st_rdev    = body->rdev;
 809                         st.st_size    = body->size;
 810                         st.st_blksize = CFS_PAGE_SIZE;
 811                         st.st_blocks  = body->blocks;
 812                         st.st_atime   = body->atime;
 813                         st.st_mtime   = body->mtime;
 814                         st.st_ctime   = body->ctime;
 815                         st.st_ino     = inode->i_ino;
 816
 817                         lmdp = (struct lov_user_mds_data *)arg;
 818                         rc = copy_to_user(&lmdp->lmd_st, &st, sizeof(st));
 819                         if (rc)
 820                                 GOTO(out_lmm, rc = -EFAULT);
 821                 }
 822
 823                 EXIT;
 824         out_lmm:
 825                 if (lmm && lmm->lmm_magic == LOV_MAGIC_JOIN)
 826                         OBD_FREE(lmm, lmmsize);
 827         out_req:
 828                 ptlrpc_req_finished(request);
 829                 if (filename)
 830                         putname(filename);
 831                 return rc;
 832         }
 833         case IOC_LOV_GETINFO: {
 834                 struct lov_user_mds_data *lumd;
 835                 struct lov_stripe_md *lsm;
 836                 struct lov_user_md *lum;
 837                 struct lov_mds_md *lmm;
 838                 int lmmsize;
 839                 lstat_t st;
 840                 int rc;
 841
 842                 lumd = (struct lov_user_mds_data *)arg;
 843                 lum = &lumd->lmd_lmm;
 844
 845                 rc = ll_get_max_mdsize(sbi, &lmmsize);
 846                 if (rc)
 847                         RETURN(rc);
 848
 849                 OBD_ALLOC(lmm, lmmsize);
 850                 rc = copy_from_user(lmm, lum, lmmsize);
 851                 if (rc)
 852                         GOTO(free_lmm, rc = -EFAULT);
 853
 854                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
 855                 if (rc < 0)
 856                         GOTO(free_lmm, rc = -ENOMEM);
 857
 858                 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
 859                 if (rc)
 860                         GOTO(free_lsm, rc);
 861
 862                 /* Perform glimpse_size operation. */
 863                 memset(&st, 0, sizeof(st));
 864
 865                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
 866                 if (rc)
 867                         GOTO(free_lsm, rc);
 868
 869                 rc = copy_to_user(&lumd->lmd_st, &st, sizeof(st));
 870                 if (rc)
 871                         GOTO(free_lsm, rc = -EFAULT);
 872
 873                 EXIT;
 874         free_lsm:
 875                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
 876         free_lmm:
 877                 OBD_FREE(lmm, lmmsize);
 878                 return rc;
 879         }
 880         case OBD_IOC_LLOG_CATINFO: {
 881                 struct ptlrpc_request *req = NULL;
 882                 char                  *buf = NULL;
 883                 char                  *str;
 884                 int                    len = 0;
 885                 int                    rc;
 886
 887                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 888                 if (rc)
 889                         RETURN(rc);
 890                 data = (void *)buf;
 891
 892                 if (!data->ioc_inlbuf1) {
 893                         obd_ioctl_freedata(buf, len);
 894                         RETURN(-EINVAL);
 895                 }
 896
 897                 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
 898                                            &RQF_LLOG_CATINFO);
 899                 if (req == NULL)
 900                         GOTO(out_catinfo, rc = -ENOMEM);
 901
 902                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 903                                      data->ioc_inllen1);
 904                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
 905                                      data->ioc_inllen2);
 906
 907                 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
 908                 if (rc) {
 909                         ptlrpc_request_free(req);
 910                         GOTO(out_catinfo, rc);
 911                 }
 912
 913                 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 914                 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
 915                 if (data->ioc_inllen2) {
 916                         str = req_capsule_client_get(&req->rq_pill,
 917                                                      &RMF_STRING);
 918                         memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
 919                 }
 920
 921                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
 922                                      data->ioc_plen1);
 923                 ptlrpc_request_set_replen(req);
 924
 925                 rc = ptlrpc_queue_wait(req);
 926                 if (!rc) {
 927                         str = req_capsule_server_get(&req->rq_pill,
 928                                                      &RMF_STRING);
 929                         rc = copy_to_user(data->ioc_pbuf1, str, data->ioc_plen1);
 930                 }
 931                 ptlrpc_req_finished(req);
 932         out_catinfo:
 933                 obd_ioctl_freedata(buf, len);
 934                 RETURN(rc);
 935         }
 936         case OBD_IOC_QUOTACHECK: {
 937                 struct obd_quotactl *oqctl;
 938                 int rc, error = 0;
 939
 940                 if (!capable(CAP_SYS_ADMIN))
 941                         RETURN(-EPERM);
 942
 943                 OBD_ALLOC_PTR(oqctl);
 944                 if (!oqctl)
 945                         RETURN(-ENOMEM);
 946                 oqctl->qc_type = arg;
 947                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
 948                 if (rc < 0) {
 949                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
 950                         error = rc;
 951                 }
 952
 953                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
 954                 if (rc < 0)
 955                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
 956
 957                 OBD_FREE_PTR(oqctl);
 958                 return error ?: rc;
 959         }
 960         case OBD_IOC_POLL_QUOTACHECK: {
 961                 struct if_quotacheck *check;
 962                 int rc;
 963
 964                 if (!capable(CAP_SYS_ADMIN))
 965                         RETURN(-EPERM);
 966
 967                 OBD_ALLOC_PTR(check);
 968                 if (!check)
 969                         RETURN(-ENOMEM);
 970
 971                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
 972                                    NULL);
 973                 if (rc) {
 974                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
 975                         if (copy_to_user((void *)arg, check, sizeof(*check)))
 976                                 rc = -EFAULT;
 977                         GOTO(out_poll, rc);
 978                 }
 979
 980                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
 981                                    NULL);
 982                 if (rc) {
 983                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
 984                         if (copy_to_user((void *)arg, check, sizeof(*check)))
 985                                 rc = -EFAULT;
 986                         GOTO(out_poll, rc);
 987                 }
 988         out_poll:
 989                 OBD_FREE_PTR(check);
 990                 RETURN(rc);
 991         }
 992 #ifdef HAVE_QUOTA_SUPPORT
 993         case OBD_IOC_QUOTACTL: {
 994                 struct if_quotactl *qctl;
 995                 struct obd_quotactl *oqctl;
 996
 997                 int cmd, type, id, rc = 0;
 998
 999                 OBD_ALLOC_PTR(qctl);
1000                 if (!qctl)
1001                         RETURN(-ENOMEM);
1002
1003                 OBD_ALLOC_PTR(oqctl);
1004                 if (!oqctl) {
1005                         OBD_FREE_PTR(qctl);
1006                         RETURN(-ENOMEM);
1007                 }
1008                 if (copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1009                         GOTO(out_quotactl, rc = -EFAULT);
1010
1011                 cmd = qctl->qc_cmd;
1012                 type = qctl->qc_type;
1013                 id = qctl->qc_id;
1014                 switch (cmd) {
1015                 case Q_QUOTAON:
1016                 case Q_QUOTAOFF:
1017                 case Q_SETQUOTA:
1018                 case Q_SETINFO:
1019                         if (!capable(CAP_SYS_ADMIN))
1020                                 GOTO(out_quotactl, rc = -EPERM);
1021                         break;
1022                 case Q_GETQUOTA:
1023                         if (((type == USRQUOTA && current->euid != id) ||
1024                              (type == GRPQUOTA && !in_egroup_p(id))) &&
1025                             !capable(CAP_SYS_ADMIN))
1026                                 GOTO(out_quotactl, rc = -EPERM);
1027
1028                         /* XXX: dqb_valid is borrowed as a flag to mark that
1029                          *      only mds quota is wanted */
1030                         if (qctl->qc_dqblk.dqb_valid)
1031                                 qctl->obd_uuid = sbi->ll_md_exp->exp_obd->
1032                                                         u.cli.cl_target_uuid;
1033                         break;
1034                 case Q_GETINFO:
1035                         break;
1036                 default:
1037                         CERROR("unsupported quotactl op: %#x\n", cmd);
1038                         GOTO(out_quotactl, rc = -ENOTTY);
1039                 }
1040
1041                 QCTL_COPY(oqctl, qctl);
1042
1043                 if (qctl->obd_uuid.uuid[0]) {
1044                         struct obd_device *obd;
1045                         struct obd_uuid *uuid = &qctl->obd_uuid;
1046
1047                         obd = class_find_client_notype(uuid,
1048                                          &sbi->ll_dt_exp->exp_obd->obd_uuid);
1049                         if (!obd)
1050                                 GOTO(out_quotactl, rc = -ENOENT);
1051
1052                         if (cmd == Q_GETINFO)
1053                                 oqctl->qc_cmd = Q_GETOINFO;
1054                         else if (cmd == Q_GETQUOTA)
1055                                 oqctl->qc_cmd = Q_GETOQUOTA;
1056                         else
1057                                 GOTO(out_quotactl, rc = -EINVAL);
1058
1059                         if (sbi->ll_md_exp->exp_obd == obd) {
1060                                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1061                         } else {
1062                                 int i;
1063                                 struct obd_export *exp;
1064                                 struct lov_obd *lov = &sbi->ll_dt_exp->
1065                                                             exp_obd->u.lov;
1066
1067                                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1068                                         if (!lov->lov_tgts[i] ||
1069                                             !lov->lov_tgts[i]->ltd_active)
1070                                                 continue;
1071                                         exp = lov->lov_tgts[i]->ltd_exp;
1072                                         if (exp->exp_obd == obd) {
1073                                                 rc = obd_quotactl(exp, oqctl);
1074                                                 break;
1075                                         }
1076                                 }
1077                         }
1078
1079                         oqctl->qc_cmd = cmd;
1080                         QCTL_COPY(qctl, oqctl);
1081
1082                         if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1083                                 rc = -EFAULT;
1084
1085                         GOTO(out_quotactl, rc);
1086                 }
1087
1088                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1089                 if (rc && rc != -EBUSY && cmd == Q_QUOTAON) {
1090                         oqctl->qc_cmd = Q_QUOTAOFF;
1091                         obd_quotactl(sbi->ll_md_exp, oqctl);
1092                 }
1093
1094                 QCTL_COPY(qctl, oqctl);
1095
1096                 if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1097                         rc = -EFAULT;
1098         out_quotactl:
1099                 OBD_FREE_PTR(qctl);
1100                 OBD_FREE_PTR(oqctl);
1101                 RETURN(rc);
1102         }
1103 #endif /* HAVE_QUOTA_SUPPORT */
1104         case OBD_IOC_GETNAME: {
1105                 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1106                 if (!obd)
1107                         RETURN(-EFAULT);
1108                 if (copy_to_user((void *)arg, obd->obd_name,
1109                                 strlen(obd->obd_name) + 1))
1110                         RETURN (-EFAULT);
1111                 RETURN(0);
1112         }
1113         case LL_IOC_FLUSHCTX:
1114                 RETURN(ll_flush_ctx(inode));
1115 #ifdef CONFIG_FS_POSIX_ACL
1116         case LL_IOC_RMTACL: {
1117             if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1118                 inode == inode->i_sb->s_root->d_inode) {
1119                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1120                 int rc;
1121
1122                 LASSERT(fd != NULL);
1123                 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1124                 if (!rc)
1125                         fd->fd_flags |= LL_FILE_RMTACL;
1126                 RETURN(rc);
1127             } else
1128                 RETURN(0);
1129         }
1130 #endif
1131         default:
1132                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1133         }
1134 }
1135
1136 int ll_dir_open(struct inode *inode, struct file *file)
1137 {
1138         ENTRY;
1139         RETURN(ll_file_open(inode, file));
1140 }
1141
1142 int ll_dir_release(struct inode *inode, struct file *file)
1143 {
1144         ENTRY;
1145         RETURN(ll_file_release(inode, file));
1146 }
1147
1148 struct file_operations ll_dir_operations = {
1149         .open     = ll_dir_open,
1150         .release  = ll_dir_release,
1151         .read     = generic_read_dir,
1152         .readdir  = ll_readdir,
1153         .ioctl    = ll_dir_ioctl
1154 };
1155