lustre/llite/dir.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/llite/dir.c
  37  *
  38  * Directory code for lustre client.
  39  */
  40
  41 #include <linux/fs.h>
  42 #include <linux/pagemap.h>
  43 #include <linux/mm.h>
  44 #include <linux/version.h>
  45 #include <asm/uaccess.h>
  46 #include <linux/buffer_head.h>   // for wait_on_buffer
  47 #include <linux/pagevec.h>
  48
  49 #define DEBUG_SUBSYSTEM S_LLITE
  50
  51 #include <lustre/lustre_idl.h>
  52 #include <obd_support.h>
  53 #include <obd_class.h>
  54 #include <lustre_lib.h>
  55 #include <lustre/lustre_idl.h>
  56 #include <lustre_lite.h>
  57 #include <lustre_dlm.h>
  58 #include <lustre_fid.h>
  59 #include "llite_internal.h"
  60
  61 /*
  62  * (new) readdir implementation overview.
  63  *
  64  * Original lustre readdir implementation cached exact copy of raw directory
  65  * pages on the client. These pages were indexed in client page cache by
  66  * logical offset in the directory file. This design, while very simple and
  67  * intuitive had some inherent problems:
  68  *
  69  *     . it implies that byte offset to the directory entry serves as a
  70  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  71  *     ext3/htree directory entries may move due to splits, and more
  72  *     importantly,
  73  *
  74  *     . it is incompatible with the design of split directories for cmd3,
  75  *     that assumes that names are distributed across nodes based on their
  76  *     hash, and so readdir should be done in hash order.
  77  *
  78  * New readdir implementation does readdir in hash order, and uses hash of a
  79  * file name as a telldir/seekdir cookie. This led to number of complications:
  80  *
  81  *     . hash is not unique, so it cannot be used to index cached directory
  82  *     pages on the client (note, that it requires a whole pageful of hash
  83  *     collided entries to cause two pages to have identical hashes);
  84  *
  85  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  86  *     entry cookie. ext3/htree has the same problem and lustre implementation
  87  *     mimics their solution: seekdir(hash) positions directory at the first
  88  *     entry with the given hash.
  89  *
  90  * Client side.
  91  *
  92  * 0. caching
  93  *
  94  * Client caches directory pages using hash of the first entry as an index. As
  95  * noted above hash is not unique, so this solution doesn't work as is:
  96  * special processing is needed for "page hash chains" (i.e., sequences of
  97  * pages filled with entries all having the same hash value).
  98  *
  99  * First, such chains have to be detected. To this end, server returns to the
 100  * client the hash of the first entry on the page next to one returned. When
 101  * client detects that this hash is the same as hash of the first entry on the
 102  * returned page, page hash collision has to be handled. Pages in the
 103  * hash chain, except first one, are termed "overflow pages".
 104  *
 105  * Solution to index uniqueness problem is to not cache overflow
 106  * pages. Instead, when page hash collision is detected, all overflow pages
 107  * from emerging chain are immediately requested from the server and placed in
 108  * a special data structure (struct ll_dir_chain). This data structure is used
 109  * by ll_readdir() to process entries from overflow pages. When readdir
 110  * invocation finishes, overflow pages are discarded. If page hash collision
 111  * chain weren't completely processed, next call to readdir will again detect
 112  * page hash collision, again read overflow pages in, process next portion of
 113  * entries and again discard the pages. This is not as wasteful as it looks,
 114  * because, given reasonable hash, page hash collisions are extremely rare.
 115  *
 116  * 1. directory positioning
 117  *
 118  * When seekdir(hash) is called, original
 119  *
 120  *
 121  *
 122  *
 123  *
 124  *
 125  *
 126  *
 127  * Server.
 128  *
 129  * identification of and access to overflow pages
 130  *
 131  * page format
 132  *
 133  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
 134  * a header lu_dirpage which describes the start/end hash, and whether this
 135  * page is empty (contains no dir entry) or hash collide with next page.
 136  * After client receives reply, several pages will be integrated into dir page
 137  * in CFS_PAGE_SIZE (if CFS_PAGE_SIZE greater than LU_PAGE_SIZE), and the
 138  * lu_dirpage for this integrated page will be adjusted.
 139  *
 140  */
 141
 142 /* returns the page unlocked, but with a reference */
 143 static int ll_dir_filler(void *_hash, struct page *page0)
 144 {
 145         struct inode *inode = page0->mapping->host;
 146         int hash64 = ll_i2sbi(inode)->ll_flags & LL_SBI_64BIT_HASH;
 147         struct obd_export *exp = ll_i2sbi(inode)->ll_md_exp;
 148         struct ptlrpc_request *request;
 149         struct mdt_body *body;
 150         struct md_op_data *op_data;
 151         __u64 hash = *((__u64 *)_hash);
 152         struct page **page_pool;
 153         struct page *page;
 154 #ifndef HAVE_ADD_TO_PAGE_CACHE_LRU
 155         struct pagevec lru_pvec;
 156 #endif
 157         struct lu_dirpage *dp;
 158         int max_pages = ll_i2sbi(inode)->ll_md_brw_size >> CFS_PAGE_SHIFT;
 159         int nrdpgs = 0; /* number of pages read actually */
 160         int npages;
 161         int i;
 162         int rc;
 163         ENTRY;
 164
 165         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) hash "LPU64"\n",
 166                inode->i_ino, inode->i_generation, inode, hash);
 167
 168         LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
 169
 170         OBD_ALLOC(page_pool, sizeof(page) * max_pages);
 171         if (page_pool != NULL) {
 172                 page_pool[0] = page0;
 173         } else {
 174                 page_pool = &page0;
 175                 max_pages = 1;
 176         }
 177         for (npages = 1; npages < max_pages; npages++) {
 178                 page = page_cache_alloc_cold(inode->i_mapping);
 179                 if (!page)
 180                         break;
 181                 page_pool[npages] = page;
 182         }
 183
 184         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 185                                      LUSTRE_OPC_ANY, NULL);
 186         op_data->op_npages = npages;
 187         op_data->op_offset = hash;
 188         rc = md_readpage(exp, op_data, page_pool, &request);
 189         ll_finish_md_op_data(op_data);
 190         if (rc == 0) {
 191                 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 192                 /* Checked by mdc_readpage() */
 193                 LASSERT(body != NULL);
 194
 195                 if (body->valid & OBD_MD_FLSIZE)
 196                         cl_isize_write(inode, body->size);
 197
 198                 nrdpgs = (request->rq_bulk->bd_nob_transferred+CFS_PAGE_SIZE-1)
 199                          >> CFS_PAGE_SHIFT;
 200                 SetPageUptodate(page0);
 201         }
 202         unlock_page(page0);
 203         ptlrpc_req_finished(request);
 204
 205         CDEBUG(D_VFSTRACE, "read %d/%d pages\n", nrdpgs, npages);
 206
 207         ll_pagevec_init(&lru_pvec, 0);
 208         for (i = 1; i < npages; i++) {
 209                 unsigned long offset;
 210                 int ret;
 211
 212                 page = page_pool[i];
 213
 214                 if (rc < 0 || i >= nrdpgs) {
 215                         page_cache_release(page);
 216                         continue;
 217                 }
 218
 219                 SetPageUptodate(page);
 220
 221                 dp = cfs_kmap(page);
 222                 hash = le64_to_cpu(dp->ldp_hash_start);
 223                 cfs_kunmap(page);
 224
 225                 offset = hash_x_index(hash, hash64);
 226
 227                 prefetchw(&page->flags);
 228                 ret = ll_add_to_page_cache_lru(page, inode->i_mapping, offset,
 229                                                GFP_KERNEL);
 230                 if (ret == 0) {
 231                         unlock_page(page);
 232                         if (ll_pagevec_add(&lru_pvec, page) == 0)
 233                                 ll_pagevec_lru_add_file(&lru_pvec);
 234                 } else {
 235                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
 236                                " %d\n", offset, ret);
 237                 }
 238                 page_cache_release(page);
 239         }
 240         ll_pagevec_lru_add_file(&lru_pvec);
 241
 242         if (page_pool != &page0)
 243                 OBD_FREE(page_pool, sizeof(struct page *) * max_pages);
 244         EXIT;
 245         return rc;
 246 }
 247
 248 static void ll_check_page(struct inode *dir, struct page *page)
 249 {
 250         /* XXX: check page format later */
 251         SetPageChecked(page);
 252 }
 253
 254 void ll_release_page(struct page *page, int remove)
 255 {
 256         kunmap(page);
 257         if (remove) {
 258                 lock_page(page);
 259                 if (likely(page->mapping != NULL))
 260                         truncate_complete_page(page->mapping, page);
 261                 unlock_page(page);
 262         }
 263         page_cache_release(page);
 264 }
 265
 266 /*
 267  * Find, kmap and return page that contains given hash.
 268  */
 269 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
 270                                        __u64 *start, __u64 *end)
 271 {
 272         int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
 273         struct address_space *mapping = dir->i_mapping;
 274         /*
 275          * Complement of hash is used as an index so that
 276          * radix_tree_gang_lookup() can be used to find a page with starting
 277          * hash _smaller_ than one we are looking for.
 278          */
 279         unsigned long offset = hash_x_index(*hash, hash64);
 280         struct page *page;
 281         int found;
 282
 283         TREE_READ_LOCK_IRQ(mapping);
 284         found = radix_tree_gang_lookup(&mapping->page_tree,
 285                                        (void **)&page, offset, 1);
 286         if (found > 0) {
 287                 struct lu_dirpage *dp;
 288
 289                 page_cache_get(page);
 290                 TREE_READ_UNLOCK_IRQ(mapping);
 291                 /*
 292                  * In contrast to find_lock_page() we are sure that directory
 293                  * page cannot be truncated (while DLM lock is held) and,
 294                  * hence, can avoid restart.
 295                  *
 296                  * In fact, page cannot be locked here at all, because
 297                  * ll_dir_filler() does synchronous io.
 298                  */
 299                 wait_on_page(page);
 300                 if (PageUptodate(page)) {
 301                         dp = cfs_kmap(page);
 302                         if (BITS_PER_LONG == 32 && hash64) {
 303                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
 304                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
 305                                 *hash  = *hash >> 32;
 306                         } else {
 307                                 *start = le64_to_cpu(dp->ldp_hash_start);
 308                                 *end   = le64_to_cpu(dp->ldp_hash_end);
 309                         }
 310                         LASSERTF(*start <= *hash, "start = "LPX64",end = "
 311                                  LPX64",hash = "LPX64"\n", *start, *end, *hash);
 312                         CDEBUG(D_VFSTRACE, "page %lu [%llu %llu], hash "LPU64"\n",
 313                                offset, *start, *end, *hash);
 314                         if (*hash > *end) {
 315                                 ll_release_page(page, 0);
 316                                 page = NULL;
 317                         } else if (*end != *start && *hash == *end) {
 318                                 /*
 319                                  * upon hash collision, remove this page,
 320                                  * otherwise put page reference, and
 321                                  * ll_get_dir_page() will issue RPC to fetch
 322                                  * the page we want.
 323                                  */
 324                                 ll_release_page(page,
 325                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
 326                                 page = NULL;
 327                         }
 328                 } else {
 329                         page_cache_release(page);
 330                         page = ERR_PTR(-EIO);
 331                 }
 332
 333         } else {
 334                 TREE_READ_UNLOCK_IRQ(mapping);
 335                 page = NULL;
 336         }
 337         return page;
 338 }
 339
 340 struct page *ll_get_dir_page(struct inode *dir, __u64 hash,
 341                              struct ll_dir_chain *chain)
 342 {
 343         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 344         struct address_space *mapping = dir->i_mapping;
 345         struct lustre_handle lockh;
 346         struct lu_dirpage *dp;
 347         struct page *page;
 348         ldlm_mode_t mode;
 349         int rc;
 350         __u64 start = 0;
 351         __u64 end = 0;
 352         __u64 lhash = hash;
 353         struct ll_inode_info *lli = ll_i2info(dir);
 354         int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
 355
 356         mode = LCK_PR;
 357         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 358                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
 359         if (!rc) {
 360                 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
 361                        ll_md_blocking_ast, ldlm_completion_ast,
 362                        NULL, NULL, dir };
 363                 struct lookup_intent it = { .it_op = IT_READDIR };
 364                 struct ptlrpc_request *request;
 365                 struct md_op_data *op_data;
 366
 367                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
 368                                              LUSTRE_OPC_ANY, NULL);
 369                 if (IS_ERR(op_data))
 370                         return (void *)op_data;
 371
 372                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
 373                                 op_data, &lockh, NULL, 0, NULL, 0);
 374
 375                 ll_finish_md_op_data(op_data);
 376
 377                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 378                 if (request)
 379                         ptlrpc_req_finished(request);
 380                 if (rc < 0) {
 381                         CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
 382                                PFID(ll_inode2fid(dir)), hash, rc);
 383                         return ERR_PTR(rc);
 384                 }
 385         } else {
 386                 /* for cross-ref object, l_ast_data of the lock may not be set,
 387                  * we reset it here */
 388                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
 389                                  dir, NULL);
 390         }
 391         ldlm_lock_dump_handle(D_OTHER, &lockh);
 392
 393         cfs_mutex_lock(&lli->lli_readdir_mutex);
 394         page = ll_dir_page_locate(dir, &lhash, &start, &end);
 395         if (IS_ERR(page)) {
 396                 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
 397                        PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
 398                 GOTO(out_unlock, page);
 399         } else if (page != NULL) {
 400                 /*
 401                  * XXX nikita: not entirely correct handling of a corner case:
 402                  * suppose hash chain of entries with hash value HASH crosses
 403                  * border between pages P0 and P1. First both P0 and P1 are
 404                  * cached, seekdir() is called for some entry from the P0 part
 405                  * of the chain. Later P0 goes out of cache. telldir(HASH)
 406                  * happens and finds P1, as it starts with matching hash
 407                  * value. Remaining entries from P0 part of the chain are
 408                  * skipped. (Is that really a bug?)
 409                  *
 410                  * Possible solutions: 0. don't cache P1 is such case, handle
 411                  * it as an "overflow" page. 1. invalidate all pages at
 412                  * once. 2. use HASH|1 as an index for P1.
 413                  */
 414                 GOTO(hash_collision, page);
 415         }
 416
 417         page = read_cache_page(mapping, hash_x_index(hash, hash64),
 418                                ll_dir_filler, &lhash);
 419         if (IS_ERR(page)) {
 420                 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
 421                        PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
 422                 GOTO(out_unlock, page);
 423         }
 424
 425         wait_on_page(page);
 426         (void)kmap(page);
 427         if (!PageUptodate(page)) {
 428                 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
 429                        PFID(ll_inode2fid(dir)), hash, -5);
 430                 goto fail;
 431         }
 432         if (!PageChecked(page))
 433                 ll_check_page(dir, page);
 434         if (PageError(page)) {
 435                 CERROR("page error: "DFID" at "LPU64": rc %d\n",
 436                        PFID(ll_inode2fid(dir)), hash, -5);
 437                 goto fail;
 438         }
 439 hash_collision:
 440         dp = page_address(page);
 441         if (BITS_PER_LONG == 32 && hash64) {
 442                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
 443                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
 444                 lhash = hash >> 32;
 445         } else {
 446                 start = le64_to_cpu(dp->ldp_hash_start);
 447                 end   = le64_to_cpu(dp->ldp_hash_end);
 448                 lhash = hash;
 449         }
 450         if (end == start) {
 451                 LASSERT(start == lhash);
 452                 CWARN("Page-wide hash collision: "LPU64"\n", end);
 453                 if (BITS_PER_LONG == 32 && hash64)
 454                         CWARN("Real page-wide hash collision at ["LPU64" "LPU64
 455                               "] with hash "LPU64"\n",
 456                               le64_to_cpu(dp->ldp_hash_start),
 457                               le64_to_cpu(dp->ldp_hash_end), hash);
 458                 /*
 459                  * Fetch whole overflow chain...
 460                  *
 461                  * XXX not yet.
 462                  */
 463                 goto fail;
 464         }
 465 out_unlock:
 466         cfs_mutex_unlock(&lli->lli_readdir_mutex);
 467         ldlm_lock_decref(&lockh, mode);
 468         return page;
 469
 470 fail:
 471         ll_release_page(page, 1);
 472         page = ERR_PTR(-EIO);
 473         goto out_unlock;
 474 }
 475
 476 int ll_dir_read(struct inode *inode, __u64 *_pos, void *cookie,
 477                 filldir_t filldir)
 478 {
 479         struct ll_inode_info *info       = ll_i2info(inode);
 480         struct ll_sb_info    *sbi        = ll_i2sbi(inode);
 481         __u64                 pos        = *_pos;
 482         int                   api32      = ll_need_32bit_api(sbi);
 483         int                   hash64     = sbi->ll_flags & LL_SBI_64BIT_HASH;
 484         struct page          *page;
 485         struct ll_dir_chain   chain;
 486         int                   done = 0;
 487         int                   rc = 0;
 488         ENTRY;
 489
 490         ll_dir_chain_init(&chain);
 491
 492         page = ll_get_dir_page(inode, pos, &chain);
 493
 494         while (rc == 0 && !done) {
 495                 struct lu_dirpage *dp;
 496                 struct lu_dirent  *ent;
 497
 498                 if (!IS_ERR(page)) {
 499                         /*
 500                          * If page is empty (end of directory is reached),
 501                          * use this value.
 502                          */
 503                         __u64 hash = MDS_DIR_END_OFF;
 504                         __u64 next;
 505
 506                         dp = page_address(page);
 507                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 508                              ent = lu_dirent_next(ent)) {
 509                                 __u16          type;
 510                                 int            namelen;
 511                                 struct lu_fid  fid;
 512                                 __u64          lhash;
 513                                 __u64          ino;
 514
 515                                 /*
 516                                  * XXX: implement correct swabbing here.
 517                                  */
 518
 519                                 hash = le64_to_cpu(ent->lde_hash);
 520                                 if (hash < pos)
 521                                         /*
 522                                          * Skip until we find target hash
 523                                          * value.
 524                                          */
 525                                         continue;
 526
 527                                 namelen = le16_to_cpu(ent->lde_namelen);
 528                                 if (namelen == 0)
 529                                         /*
 530                                          * Skip dummy record.
 531                                          */
 532                                         continue;
 533
 534                                 if (api32 && hash64)
 535                                         lhash = hash >> 32;
 536                                 else
 537                                         lhash = hash;
 538                                 fid_le_to_cpu(&fid, &ent->lde_fid);
 539                                 ino = cl_fid_build_ino(&fid, api32);
 540                                 type = ll_dirent_type_get(ent);
 541                                 /* For 'll_nfs_get_name_filldir()', it will try
 542                                  * to access the 'ent' through its 'lde_name',
 543                                  * so the parameter 'name' for 'filldir()' must
 544                                  * be part of the 'ent'. */
 545                                 done = filldir(cookie, ent->lde_name, namelen,
 546                                                lhash, ino, type);
 547                         }
 548                         next = le64_to_cpu(dp->ldp_hash_end);
 549                         if (!done) {
 550                                 pos = next;
 551                                 if (pos == MDS_DIR_END_OFF) {
 552                                         /*
 553                                          * End of directory reached.
 554                                          */
 555                                         done = 1;
 556                                         ll_release_page(page, 0);
 557                                 } else if (1 /* chain is exhausted*/) {
 558                                         /*
 559                                          * Normal case: continue to the next
 560                                          * page.
 561                                          */
 562                                         ll_release_page(page,
 563                                             le32_to_cpu(dp->ldp_flags) &
 564                                                         LDF_COLLIDE);
 565                                         next = pos;
 566                                         page = ll_get_dir_page(inode, pos,
 567                                                                &chain);
 568                                 } else {
 569                                         /*
 570                                          * go into overflow page.
 571                                          */
 572                                         LASSERT(le32_to_cpu(dp->ldp_flags) &
 573                                                 LDF_COLLIDE);
 574                                         ll_release_page(page, 1);
 575                                 }
 576                         } else {
 577                                 pos = hash;
 578                                 ll_release_page(page, 0);
 579                         }
 580                 } else {
 581                         rc = PTR_ERR(page);
 582                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 583                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 584                 }
 585         }
 586
 587         *_pos = pos;
 588         ll_dir_chain_fini(&chain);
 589         RETURN(rc);
 590 }
 591
 592 static int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 593 {
 594         struct inode            *inode  = filp->f_dentry->d_inode;
 595         struct ll_file_data     *lfd    = LUSTRE_FPRIVATE(filp);
 596         struct ll_sb_info       *sbi    = ll_i2sbi(inode);
 597         __u64                   pos     = lfd->lfd_pos;
 598         int                     hash64  = sbi->ll_flags & LL_SBI_64BIT_HASH;
 599         int                     api32   = ll_need_32bit_api(sbi);
 600         int                     rc;
 601         ENTRY;
 602
 603         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu "
 604                " 32bit_api %d\n", inode->i_ino, inode->i_generation,
 605                inode, (unsigned long)pos, i_size_read(inode), api32);
 606
 607         if (pos == MDS_DIR_END_OFF)
 608                 /*
 609                  * end-of-file.
 610                  */
 611                 GOTO(out, rc = 0);
 612
 613         rc = ll_dir_read(inode, &pos, cookie, filldir);
 614         lfd->lfd_pos = pos;
 615         if (pos == MDS_DIR_END_OFF) {
 616                 if (api32)
 617                         filp->f_pos = LL_DIR_END_OFF_32BIT;
 618                 else
 619                         filp->f_pos = LL_DIR_END_OFF;
 620         } else {
 621                 if (api32 && hash64)
 622                         filp->f_pos = pos >> 32;
 623                 else
 624                         filp->f_pos = pos;
 625         }
 626         filp->f_version = inode->i_version;
 627         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 628
 629 out:
 630         if (!rc)
 631                 ll_stats_ops_tally(sbi, LPROC_LL_READDIR, 1);
 632
 633         RETURN(rc);
 634 }
 635
 636 int ll_send_mgc_param(struct obd_export *mgc, char *string)
 637 {
 638         struct mgs_send_param *msp;
 639         int rc = 0;
 640
 641         OBD_ALLOC_PTR(msp);
 642         if (!msp)
 643                 return -ENOMEM;
 644
 645         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
 646         rc = obd_set_info_async(NULL, mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
 647                                 sizeof(struct mgs_send_param), msp, NULL);
 648         if (rc)
 649                 CERROR("Failed to set parameter: %d\n", rc);
 650         OBD_FREE_PTR(msp);
 651
 652         return rc;
 653 }
 654
 655 char *ll_get_fsname(struct inode *inode)
 656 {
 657         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 658         char *ptr, *fsname;
 659         int len;
 660
 661         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
 662         len = strlen(lsi->lsi_lmd->lmd_profile);
 663         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
 664         if (ptr && (strcmp(ptr, "-client") == 0))
 665                 len -= 7;
 666         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
 667         fsname[len] = '\0';
 668
 669         return fsname;
 670 }
 671
 672 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
 673                      int set_default)
 674 {
 675         struct ll_sb_info *sbi = ll_i2sbi(inode);
 676         struct md_op_data *op_data;
 677         struct ptlrpc_request *req = NULL;
 678         int rc = 0;
 679         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 680         struct obd_device *mgc = lsi->lsi_mgc;
 681         char *fsname = NULL, *param = NULL;
 682         int lum_size;
 683
 684         if (lump != NULL) {
 685                 /*
 686                  * This is coming from userspace, so should be in
 687                  * local endian.  But the MDS would like it in little
 688                  * endian, so we swab it before we send it.
 689                  */
 690                 switch (lump->lmm_magic) {
 691                 case LOV_USER_MAGIC_V1: {
 692                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
 693                                 lustre_swab_lov_user_md_v1(lump);
 694                         lum_size = sizeof(struct lov_user_md_v1);
 695                         break;
 696                         }
 697                 case LOV_USER_MAGIC_V3: {
 698                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
 699                                 lustre_swab_lov_user_md_v3(
 700                                         (struct lov_user_md_v3 *)lump);
 701                         lum_size = sizeof(struct lov_user_md_v3);
 702                         break;
 703                         }
 704                 default: {
 705                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 706                                         " %#08x != %#08x nor %#08x\n",
 707                                         lump->lmm_magic, LOV_USER_MAGIC_V1,
 708                                         LOV_USER_MAGIC_V3);
 709                         RETURN(-EINVAL);
 710                         }
 711                }
 712         } else {
 713                 lum_size = sizeof(struct lov_user_md_v1);
 714         }
 715
 716         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 717                                      LUSTRE_OPC_ANY, NULL);
 718         if (IS_ERR(op_data))
 719                 RETURN(PTR_ERR(op_data));
 720
 721         /* swabbing is done in lov_setstripe() on server side */
 722         rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
 723                         NULL, 0, &req, NULL);
 724         ll_finish_md_op_data(op_data);
 725         ptlrpc_req_finished(req);
 726         if (rc) {
 727                 if (rc != -EPERM && rc != -EACCES)
 728                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 729         }
 730
 731         /* In the following we use the fact that LOV_USER_MAGIC_V1 and
 732          LOV_USER_MAGIC_V3 have the same initial fields so we do not
 733          need the make the distiction between the 2 versions */
 734         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
 735                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
 736
 737                 /* Get fsname and assume devname to be -MDT0000. */
 738                 fsname = ll_get_fsname(inode);
 739                 /* Set root stripesize */
 740                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
 741                         lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
 742                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 743                 if (rc)
 744                         goto end;
 745
 746                 /* Set root stripecount */
 747                 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
 748                         lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
 749                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 750                 if (rc)
 751                         goto end;
 752
 753                 /* Set root stripeoffset */
 754                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
 755                         lump ? le16_to_cpu(lump->lmm_stripe_offset) :
 756                         (typeof(lump->lmm_stripe_offset))(-1));
 757                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 758                 if (rc)
 759                         goto end;
 760 end:
 761                 if (fsname)
 762                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
 763                 if (param)
 764                         OBD_FREE(param, MGS_PARAM_MAXLEN);
 765         }
 766         return rc;
 767 }
 768
 769 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 770                      int *lmm_size, struct ptlrpc_request **request)
 771 {
 772         struct ll_sb_info *sbi = ll_i2sbi(inode);
 773         struct mdt_body   *body;
 774         struct lov_mds_md *lmm = NULL;
 775         struct ptlrpc_request *req = NULL;
 776         int rc, lmmsize;
 777         struct md_op_data *op_data;
 778
 779         rc = ll_get_max_mdsize(sbi, &lmmsize);
 780         if (rc)
 781                 RETURN(rc);
 782
 783         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
 784                                      0, lmmsize, LUSTRE_OPC_ANY,
 785                                      NULL);
 786         if (IS_ERR(op_data))
 787                 RETURN(PTR_ERR(op_data));
 788
 789         op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
 790         rc = md_getattr(sbi->ll_md_exp, op_data, &req);
 791         ll_finish_md_op_data(op_data);
 792         if (rc < 0) {
 793                 CDEBUG(D_INFO, "md_getattr failed on inode "
 794                        "%lu/%u: rc %d\n", inode->i_ino,
 795                        inode->i_generation, rc);
 796                 GOTO(out, rc);
 797         }
 798
 799         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 800         LASSERT(body != NULL);
 801
 802         lmmsize = body->eadatasize;
 803
 804         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 805             lmmsize == 0) {
 806                 GOTO(out, rc = -ENODATA);
 807         }
 808
 809         lmm = req_capsule_server_sized_get(&req->rq_pill,
 810                                            &RMF_MDT_MD, lmmsize);
 811         LASSERT(lmm != NULL);
 812
 813         /*
 814          * This is coming from the MDS, so is probably in
 815          * little endian.  We convert it to host endian before
 816          * passing it to userspace.
 817          */
 818         /* We don't swab objects for directories */
 819         switch (le32_to_cpu(lmm->lmm_magic)) {
 820         case LOV_MAGIC_V1:
 821                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
 822                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
 823                 break;
 824         case LOV_MAGIC_V3:
 825                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
 826                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
 827                 break;
 828         default:
 829                 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
 830                 rc = -EPROTO;
 831         }
 832 out:
 833         *lmmp = lmm;
 834         *lmm_size = lmmsize;
 835         *request = req;
 836         return rc;
 837 }
 838
 839 /*
 840  *  Get MDT index for the inode.
 841  */
 842 int ll_get_mdt_idx(struct inode *inode)
 843 {
 844         struct ll_sb_info *sbi = ll_i2sbi(inode);
 845         struct md_op_data *op_data;
 846         int rc, mdtidx;
 847         ENTRY;
 848
 849         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
 850                                      0, LUSTRE_OPC_ANY, NULL);
 851         if (IS_ERR(op_data))
 852                 RETURN(PTR_ERR(op_data));
 853
 854         op_data->op_valid |= OBD_MD_MDTIDX;
 855         rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
 856         mdtidx = op_data->op_mds;
 857         ll_finish_md_op_data(op_data);
 858         if (rc < 0) {
 859                 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 860                 RETURN(rc);
 861         }
 862         return mdtidx;
 863 }
 864
 865 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
 866 {
 867         void *ptr;
 868         int rc;
 869
 870         OBD_ALLOC(ptr, len);
 871         if (ptr == NULL)
 872                 return -ENOMEM;
 873         if (cfs_copy_from_user(ptr, data, len)) {
 874                 OBD_FREE(ptr, len);
 875                 return -EFAULT;
 876         }
 877         rc = obd_iocontrol(cmd, exp, len, data, NULL);
 878         OBD_FREE(ptr, len);
 879         return rc;
 880 }
 881
 882 static int quotactl_ioctl(struct ll_sb_info *sbi, struct if_quotactl *qctl)
 883 {
 884         int cmd = qctl->qc_cmd;
 885         int type = qctl->qc_type;
 886         int id = qctl->qc_id;
 887         int valid = qctl->qc_valid;
 888         int rc = 0;
 889         ENTRY;
 890
 891         switch (cmd) {
 892         case LUSTRE_Q_INVALIDATE:
 893         case LUSTRE_Q_FINVALIDATE:
 894         case Q_QUOTAON:
 895         case Q_QUOTAOFF:
 896         case Q_SETQUOTA:
 897         case Q_SETINFO:
 898                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
 899                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
 900                         RETURN(-EPERM);
 901                 break;
 902         case Q_GETQUOTA:
 903                 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
 904                      (type == GRPQUOTA && !in_egroup_p(id))) &&
 905                     (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
 906                      sbi->ll_flags & LL_SBI_RMT_CLIENT))
 907                         RETURN(-EPERM);
 908                 break;
 909         case Q_GETINFO:
 910                 break;
 911         default:
 912                 CERROR("unsupported quotactl op: %#x\n", cmd);
 913                 RETURN(-ENOTTY);
 914         }
 915
 916         if (valid != QC_GENERAL) {
 917                 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
 918                         RETURN(-EOPNOTSUPP);
 919
 920                 if (cmd == Q_GETINFO)
 921                         qctl->qc_cmd = Q_GETOINFO;
 922                 else if (cmd == Q_GETQUOTA)
 923                         qctl->qc_cmd = Q_GETOQUOTA;
 924                 else
 925                         RETURN(-EINVAL);
 926
 927                 switch (valid) {
 928                 case QC_MDTIDX:
 929                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
 930                                            sizeof(*qctl), qctl, NULL);
 931                         break;
 932                 case QC_OSTIDX:
 933                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_dt_exp,
 934                                            sizeof(*qctl), qctl, NULL);
 935                         break;
 936                 case QC_UUID:
 937                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
 938                                            sizeof(*qctl), qctl, NULL);
 939                         if (rc == -EAGAIN)
 940                                 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
 941                                                    sbi->ll_dt_exp,
 942                                                    sizeof(*qctl), qctl, NULL);
 943                         break;
 944                 default:
 945                         rc = -EINVAL;
 946                         break;
 947                 }
 948
 949                 if (rc)
 950                         RETURN(rc);
 951
 952                 qctl->qc_cmd = cmd;
 953         } else {
 954                 struct obd_quotactl *oqctl;
 955
 956                 OBD_ALLOC_PTR(oqctl);
 957                 if (oqctl == NULL)
 958                         RETURN(-ENOMEM);
 959
 960                 QCTL_COPY(oqctl, qctl);
 961                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
 962                 if (rc) {
 963                         if (rc != -EALREADY && cmd == Q_QUOTAON) {
 964                                 oqctl->qc_cmd = Q_QUOTAOFF;
 965                                 obd_quotactl(sbi->ll_md_exp, oqctl);
 966                         }
 967                         OBD_FREE_PTR(oqctl);
 968                         RETURN(rc);
 969                 }
 970                 /* If QIF_SPACE is not set, client should collect the
 971                  * space usage from OSSs by itself */
 972                 if (cmd == Q_GETQUOTA &&
 973                     !(oqctl->qc_dqblk.dqb_valid & QIF_SPACE) &&
 974                     !oqctl->qc_dqblk.dqb_curspace) {
 975                         struct obd_quotactl *oqctl_tmp;
 976
 977                         OBD_ALLOC_PTR(oqctl_tmp);
 978                         if (oqctl_tmp == NULL)
 979                                 GOTO(out, rc = -ENOMEM);
 980
 981                         oqctl_tmp->qc_cmd = Q_GETOQUOTA;
 982                         oqctl_tmp->qc_id = oqctl->qc_id;
 983                         oqctl_tmp->qc_type = oqctl->qc_type;
 984
 985                         /* collect space usage from OSTs */
 986                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
 987                         rc = obd_quotactl(sbi->ll_dt_exp, oqctl_tmp);
 988                         if (!rc || rc == -EREMOTEIO) {
 989                                 oqctl->qc_dqblk.dqb_curspace =
 990                                         oqctl_tmp->qc_dqblk.dqb_curspace;
 991                                 oqctl->qc_dqblk.dqb_valid |= QIF_SPACE;
 992                         }
 993
 994                         /* collect space & inode usage from MDTs */
 995                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
 996                         oqctl_tmp->qc_dqblk.dqb_curinodes = 0;
 997                         rc = obd_quotactl(sbi->ll_md_exp, oqctl_tmp);
 998                         if (!rc || rc == -EREMOTEIO) {
 999                                 oqctl->qc_dqblk.dqb_curspace +=
1000                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1001                                 oqctl->qc_dqblk.dqb_curinodes =
1002                                         oqctl_tmp->qc_dqblk.dqb_curinodes;
1003                                 oqctl->qc_dqblk.dqb_valid |= QIF_INODES;
1004                         } else {
1005                                 oqctl->qc_dqblk.dqb_valid &= ~QIF_SPACE;
1006                         }
1007
1008                         OBD_FREE_PTR(oqctl_tmp);
1009                 }
1010 out:
1011                 QCTL_COPY(qctl, oqctl);
1012                 OBD_FREE_PTR(oqctl);
1013         }
1014
1015         RETURN(rc);
1016 }
1017
1018 static long ll_dir_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1019 {
1020         struct inode *inode = file->f_dentry->d_inode;
1021         struct ll_sb_info *sbi = ll_i2sbi(inode);
1022         struct obd_ioctl_data *data;
1023         int rc = 0;
1024         ENTRY;
1025
1026         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
1027                inode->i_ino, inode->i_generation, inode, cmd);
1028
1029         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
1030         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
1031                 return -ENOTTY;
1032
1033         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
1034         switch(cmd) {
1035         case FSFILT_IOC_GETFLAGS:
1036         case FSFILT_IOC_SETFLAGS:
1037                 RETURN(ll_iocontrol(inode, file, cmd, arg));
1038         case FSFILT_IOC_GETVERSION_OLD:
1039         case FSFILT_IOC_GETVERSION:
1040                 RETURN(put_user(inode->i_generation, (int *)arg));
1041         /* We need to special case any other ioctls we want to handle,
1042          * to send them to the MDS/OST as appropriate and to properly
1043          * network encode the arg field.
1044         case FSFILT_IOC_SETVERSION_OLD:
1045         case FSFILT_IOC_SETVERSION:
1046         */
1047         case LL_IOC_GET_MDTIDX: {
1048                 int mdtidx;
1049
1050                 mdtidx = ll_get_mdt_idx(inode);
1051                 if (mdtidx < 0)
1052                         RETURN(mdtidx);
1053
1054                 if (put_user((int)mdtidx, (int*)arg))
1055                         RETURN(-EFAULT);
1056
1057                 return 0;
1058         }
1059         case IOC_MDC_LOOKUP: {
1060                 struct ptlrpc_request *request = NULL;
1061                 int namelen, len = 0;
1062                 char *buf = NULL;
1063                 char *filename;
1064                 struct md_op_data *op_data;
1065
1066                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1067                 if (rc)
1068                         RETURN(rc);
1069                 data = (void *)buf;
1070
1071                 filename = data->ioc_inlbuf1;
1072                 namelen = strlen(filename);
1073
1074                 if (namelen < 1) {
1075                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1076                         GOTO(out_free, rc = -EINVAL);
1077                 }
1078
1079                 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1080                                              0, LUSTRE_OPC_ANY, NULL);
1081                 if (IS_ERR(op_data))
1082                         GOTO(out_free, rc = PTR_ERR(op_data));
1083
1084                 op_data->op_valid = OBD_MD_FLID;
1085                 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1086                 ll_finish_md_op_data(op_data);
1087                 if (rc < 0) {
1088                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1089                         GOTO(out_free, rc);
1090                 }
1091                 ptlrpc_req_finished(request);
1092                 EXIT;
1093 out_free:
1094                 obd_ioctl_freedata(buf, len);
1095                 return rc;
1096         }
1097         case LL_IOC_LOV_SETSTRIPE: {
1098                 struct lov_user_md_v3 lumv3;
1099                 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1100                 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1101                 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1102
1103                 int set_default = 0;
1104
1105                 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1106                 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1107                         sizeof(lumv3p->lmm_objects[0]));
1108                 /* first try with v1 which is smaller than v3 */
1109                 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1110                         RETURN(-EFAULT);
1111
1112                 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1113                         if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1114                                 RETURN(-EFAULT);
1115                 }
1116
1117                 if (inode->i_sb->s_root == file->f_dentry)
1118                         set_default = 1;
1119
1120                 /* in v1 and v3 cases lumv1 points to data */
1121                 rc = ll_dir_setstripe(inode, lumv1, set_default);
1122
1123                 RETURN(rc);
1124         }
1125         case LL_IOC_OBD_STATFS:
1126                 RETURN(ll_obd_statfs(inode, (void *)arg));
1127         case LL_IOC_LOV_GETSTRIPE:
1128         case LL_IOC_MDC_GETINFO:
1129         case IOC_MDC_GETFILEINFO:
1130         case IOC_MDC_GETFILESTRIPE: {
1131                 struct ptlrpc_request *request = NULL;
1132                 struct lov_user_md *lump;
1133                 struct lov_mds_md *lmm = NULL;
1134                 struct mdt_body *body;
1135                 char *filename = NULL;
1136                 int lmmsize;
1137
1138                 if (cmd == IOC_MDC_GETFILEINFO ||
1139                     cmd == IOC_MDC_GETFILESTRIPE) {
1140                         filename = getname((const char *)arg);
1141                         if (IS_ERR(filename))
1142                                 RETURN(PTR_ERR(filename));
1143
1144                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1145                                                       &lmmsize, &request);
1146                 } else {
1147                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1148                 }
1149
1150                 if (request) {
1151                         body = req_capsule_server_get(&request->rq_pill,
1152                                                       &RMF_MDT_BODY);
1153                         LASSERT(body != NULL);
1154                 } else {
1155                         GOTO(out_req, rc);
1156                 }
1157
1158                 if (rc < 0) {
1159                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1160                                                cmd == LL_IOC_MDC_GETINFO))
1161                                 GOTO(skip_lmm, rc = 0);
1162                         else
1163                                 GOTO(out_req, rc);
1164                 }
1165
1166                 if (cmd == IOC_MDC_GETFILESTRIPE ||
1167                     cmd == LL_IOC_LOV_GETSTRIPE) {
1168                         lump = (struct lov_user_md *)arg;
1169                 } else {
1170                         struct lov_user_mds_data *lmdp;
1171                         lmdp = (struct lov_user_mds_data *)arg;
1172                         lump = &lmdp->lmd_lmm;
1173                 }
1174                 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
1175                         if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
1176                                 GOTO(out_req, rc = -EFAULT);
1177                         rc = -EOVERFLOW;
1178                 }
1179         skip_lmm:
1180                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1181                         struct lov_user_mds_data *lmdp;
1182                         lstat_t st = { 0 };
1183
1184                         st.st_dev     = inode->i_sb->s_dev;
1185                         st.st_mode    = body->mode;
1186                         st.st_nlink   = body->nlink;
1187                         st.st_uid     = body->uid;
1188                         st.st_gid     = body->gid;
1189                         st.st_rdev    = body->rdev;
1190                         st.st_size    = body->size;
1191                         st.st_blksize = CFS_PAGE_SIZE;
1192                         st.st_blocks  = body->blocks;
1193                         st.st_atime   = body->atime;
1194                         st.st_mtime   = body->mtime;
1195                         st.st_ctime   = body->ctime;
1196                         st.st_ino     = inode->i_ino;
1197
1198                         lmdp = (struct lov_user_mds_data *)arg;
1199                         if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1200                                 GOTO(out_req, rc = -EFAULT);
1201                 }
1202
1203                 EXIT;
1204         out_req:
1205                 ptlrpc_req_finished(request);
1206                 if (filename)
1207                         putname(filename);
1208                 return rc;
1209         }
1210         case IOC_LOV_GETINFO: {
1211                 struct lov_user_mds_data *lumd;
1212                 struct lov_stripe_md *lsm;
1213                 struct lov_user_md *lum;
1214                 struct lov_mds_md *lmm;
1215                 int lmmsize;
1216                 lstat_t st;
1217
1218                 lumd = (struct lov_user_mds_data *)arg;
1219                 lum = &lumd->lmd_lmm;
1220
1221                 rc = ll_get_max_mdsize(sbi, &lmmsize);
1222                 if (rc)
1223                         RETURN(rc);
1224
1225                 OBD_ALLOC_LARGE(lmm, lmmsize);
1226                 if (cfs_copy_from_user(lmm, lum, lmmsize))
1227                         GOTO(free_lmm, rc = -EFAULT);
1228
1229                 switch (lmm->lmm_magic) {
1230                 case LOV_USER_MAGIC_V1:
1231                         if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1232                                 break;
1233                         /* swab objects first so that stripes num will be sane */
1234                         lustre_swab_lov_user_md_objects(
1235                                 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1236                                 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1237                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1238                         break;
1239                 case LOV_USER_MAGIC_V3:
1240                         if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1241                                 break;
1242                         /* swab objects first so that stripes num will be sane */
1243                         lustre_swab_lov_user_md_objects(
1244                                 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1245                                 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1246                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1247                         break;
1248                 default:
1249                         GOTO(free_lmm, rc = -EINVAL);
1250                 }
1251
1252                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1253                 if (rc < 0)
1254                         GOTO(free_lmm, rc = -ENOMEM);
1255
1256                 /* Perform glimpse_size operation. */
1257                 memset(&st, 0, sizeof(st));
1258
1259                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1260                 if (rc)
1261                         GOTO(free_lsm, rc);
1262
1263                 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1264                         GOTO(free_lsm, rc = -EFAULT);
1265
1266                 EXIT;
1267         free_lsm:
1268                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1269         free_lmm:
1270                 OBD_FREE_LARGE(lmm, lmmsize);
1271                 return rc;
1272         }
1273         case OBD_IOC_LLOG_CATINFO: {
1274                 RETURN(-EOPNOTSUPP);
1275         }
1276         case OBD_IOC_QUOTACHECK: {
1277                 struct obd_quotactl *oqctl;
1278                 int error = 0;
1279
1280                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1281                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1282                         RETURN(-EPERM);
1283
1284                 OBD_ALLOC_PTR(oqctl);
1285                 if (!oqctl)
1286                         RETURN(-ENOMEM);
1287                 oqctl->qc_type = arg;
1288                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1289                 if (rc < 0) {
1290                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1291                         error = rc;
1292                 }
1293
1294                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1295                 if (rc < 0)
1296                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1297
1298                 OBD_FREE_PTR(oqctl);
1299                 return error ?: rc;
1300         }
1301         case OBD_IOC_POLL_QUOTACHECK: {
1302                 struct if_quotacheck *check;
1303
1304                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1305                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1306                         RETURN(-EPERM);
1307
1308                 OBD_ALLOC_PTR(check);
1309                 if (!check)
1310                         RETURN(-ENOMEM);
1311
1312                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1313                                    NULL);
1314                 if (rc) {
1315                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1316                         if (cfs_copy_to_user((void *)arg, check,
1317                                              sizeof(*check)))
1318                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1319                         GOTO(out_poll, rc);
1320                 }
1321
1322                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1323                                    NULL);
1324                 if (rc) {
1325                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1326                         if (cfs_copy_to_user((void *)arg, check,
1327                                              sizeof(*check)))
1328                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1329                         GOTO(out_poll, rc);
1330                 }
1331         out_poll:
1332                 OBD_FREE_PTR(check);
1333                 RETURN(rc);
1334         }
1335 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
1336         case LL_IOC_QUOTACTL_18: {
1337                 /* copy the old 1.x quota struct for internal use, then copy
1338                  * back into old format struct.  For 1.8 compatibility. */
1339                 struct if_quotactl_18 *qctl_18;
1340                 struct if_quotactl *qctl_20;
1341
1342                 OBD_ALLOC_PTR(qctl_18);
1343                 if (!qctl_18)
1344                         RETURN(-ENOMEM);
1345
1346                 OBD_ALLOC_PTR(qctl_20);
1347                 if (!qctl_20)
1348                         GOTO(out_quotactl_18, rc = -ENOMEM);
1349
1350                 if (cfs_copy_from_user(qctl_18, (void *)arg, sizeof(*qctl_18)))
1351                         GOTO(out_quotactl_20, rc = -ENOMEM);
1352
1353                 QCTL_COPY(qctl_20, qctl_18);
1354                 qctl_20->qc_idx = 0;
1355
1356                 /* XXX: dqb_valid was borrowed as a flag to mark that
1357                  *      only mds quota is wanted */
1358                 if (qctl_18->qc_cmd == Q_GETQUOTA &&
1359                     qctl_18->qc_dqblk.dqb_valid) {
1360                         qctl_20->qc_valid = QC_MDTIDX;
1361                         qctl_20->qc_dqblk.dqb_valid = 0;
1362                 } else if (qctl_18->obd_uuid.uuid[0] != '\0') {
1363                         qctl_20->qc_valid = QC_UUID;
1364                         qctl_20->obd_uuid = qctl_18->obd_uuid;
1365                 } else {
1366                         qctl_20->qc_valid = QC_GENERAL;
1367                 }
1368
1369                 rc = quotactl_ioctl(sbi, qctl_20);
1370
1371                 if (rc == 0) {
1372                         QCTL_COPY(qctl_18, qctl_20);
1373                         qctl_18->obd_uuid = qctl_20->obd_uuid;
1374
1375                         if (cfs_copy_to_user((void *)arg, qctl_18,
1376                                              sizeof(*qctl_18)))
1377                                 rc = -EFAULT;
1378                 }
1379
1380         out_quotactl_20:
1381                 OBD_FREE_PTR(qctl_20);
1382         out_quotactl_18:
1383                 OBD_FREE_PTR(qctl_18);
1384                 RETURN(rc);
1385         }
1386 #else
1387 #warning "remove old LL_IOC_QUOTACTL_18 compatibility code"
1388 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0) */
1389         case LL_IOC_QUOTACTL: {
1390                 struct if_quotactl *qctl;
1391
1392                 OBD_ALLOC_PTR(qctl);
1393                 if (!qctl)
1394                         RETURN(-ENOMEM);
1395
1396                 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1397                         GOTO(out_quotactl, rc = -EFAULT);
1398
1399                 rc = quotactl_ioctl(sbi, qctl);
1400
1401                 if (rc == 0 && cfs_copy_to_user((void *)arg,qctl,sizeof(*qctl)))
1402                         rc = -EFAULT;
1403
1404         out_quotactl:
1405                 OBD_FREE_PTR(qctl);
1406                 RETURN(rc);
1407         }
1408         case OBD_IOC_GETDTNAME:
1409         case OBD_IOC_GETMDNAME:
1410                 RETURN(ll_get_obd_name(inode, cmd, arg));
1411         case LL_IOC_FLUSHCTX:
1412                 RETURN(ll_flush_ctx(inode));
1413 #ifdef CONFIG_FS_POSIX_ACL
1414         case LL_IOC_RMTACL: {
1415             if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1416                 inode == inode->i_sb->s_root->d_inode) {
1417                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1418
1419                 LASSERT(fd != NULL);
1420                 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1421                 if (!rc)
1422                         fd->fd_flags |= LL_FILE_RMTACL;
1423                 RETURN(rc);
1424             } else
1425                 RETURN(0);
1426         }
1427 #endif
1428         case LL_IOC_GETOBDCOUNT: {
1429                 int count, vallen;
1430                 struct obd_export *exp;
1431
1432                 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1433                         RETURN(-EFAULT);
1434
1435                 /* get ost count when count is zero, get mdt count otherwise */
1436                 exp = count ? sbi->ll_md_exp : sbi->ll_dt_exp;
1437                 vallen = sizeof(count);
1438                 rc = obd_get_info(NULL, exp, sizeof(KEY_TGT_COUNT),
1439                                   KEY_TGT_COUNT, &vallen, &count, NULL);
1440                 if (rc) {
1441                         CERROR("get target count failed: %d\n", rc);
1442                         RETURN(rc);
1443                 }
1444
1445                 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1446                         RETURN(-EFAULT);
1447
1448                 RETURN(0);
1449         }
1450         case LL_IOC_PATH2FID:
1451                 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1452                                      sizeof(struct lu_fid)))
1453                         RETURN(-EFAULT);
1454                 RETURN(0);
1455         case LL_IOC_GET_CONNECT_FLAGS: {
1456                 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1457         }
1458         case OBD_IOC_CHANGELOG_SEND:
1459         case OBD_IOC_CHANGELOG_CLEAR:
1460                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1461                                     sizeof(struct ioc_changelog));
1462                 RETURN(rc);
1463         case OBD_IOC_FID2PATH:
1464                 RETURN(ll_fid2path(inode, (void *)arg));
1465         case LL_IOC_HSM_CT_START:
1466                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1467                                     sizeof(struct lustre_kernelcomm));
1468                 RETURN(rc);
1469
1470         default:
1471                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1472         }
1473 }
1474
1475 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1476 {
1477         struct inode *inode = file->f_mapping->host;
1478         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1479         struct ll_sb_info *sbi = ll_i2sbi(inode);
1480         int api32 = ll_need_32bit_api(sbi);
1481         loff_t ret = -EINVAL;
1482         ENTRY;
1483
1484         cfs_mutex_lock(&inode->i_mutex);
1485         switch (origin) {
1486                 case SEEK_SET:
1487                         break;
1488                 case SEEK_CUR:
1489                         offset += file->f_pos;
1490                         break;
1491                 case SEEK_END:
1492                         if (offset > 0)
1493                                 GOTO(out, ret);
1494                         if (api32)
1495                                 offset += LL_DIR_END_OFF_32BIT;
1496                         else
1497                                 offset += LL_DIR_END_OFF;
1498                         break;
1499                 default:
1500                         GOTO(out, ret);
1501         }
1502
1503         if (offset >= 0 &&
1504             ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1505              (!api32 && offset <= LL_DIR_END_OFF))) {
1506                 if (offset != file->f_pos) {
1507                         if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1508                             (!api32 && offset == LL_DIR_END_OFF))
1509                                 fd->lfd_pos = MDS_DIR_END_OFF;
1510                         else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1511                                 fd->lfd_pos = offset << 32;
1512                         else
1513                                 fd->lfd_pos = offset;
1514                         file->f_pos = offset;
1515                         file->f_version = 0;
1516                 }
1517                 ret = offset;
1518         }
1519         GOTO(out, ret);
1520
1521 out:
1522         cfs_mutex_unlock(&inode->i_mutex);
1523         return ret;
1524 }
1525
1526 int ll_dir_open(struct inode *inode, struct file *file)
1527 {
1528         ENTRY;
1529         RETURN(ll_file_open(inode, file));
1530 }
1531
1532 int ll_dir_release(struct inode *inode, struct file *file)
1533 {
1534         ENTRY;
1535         RETURN(ll_file_release(inode, file));
1536 }
1537
1538 struct file_operations ll_dir_operations = {
1539         .llseek   = ll_dir_seek,
1540         .open     = ll_dir_open,
1541         .release  = ll_dir_release,
1542         .read     = generic_read_dir,
1543         .readdir  = ll_readdir,
1544         .unlocked_ioctl   = ll_dir_ioctl,
1545         .fsync    = ll_fsync,
1546 };