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 <obd_support.h>
  52 #include <obd_class.h>
  53 #include <lustre_lib.h>
  54 #include <lustre/lustre_idl.h>
  55 #include <lustre_lite.h>
  56 #include <lustre_dlm.h>
  57 #include <lustre_fid.h>
  58 #include "llite_internal.h"
  59
  60 #ifndef HAVE_PAGE_CHECKED
  61 #ifdef HAVE_PG_FS_MISC
  62 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
  63 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
  64 #else
  65 #error PageChecked or PageFsMisc not defined in kernel
  66 #endif
  67 #endif
  68
  69 /*
  70  * (new) readdir implementation overview.
  71  *
  72  * Original lustre readdir implementation cached exact copy of raw directory
  73  * pages on the client. These pages were indexed in client page cache by
  74  * logical offset in the directory file. This design, while very simple and
  75  * intuitive had some inherent problems:
  76  *
  77  *     . it implies that byte offset to the directory entry serves as a
  78  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  79  *     ext3/htree directory entries may move due to splits, and more
  80  *     importantly,
  81  *
  82  *     . it is incompatible with the design of split directories for cmd3,
  83  *     that assumes that names are distributed across nodes based on their
  84  *     hash, and so readdir should be done in hash order.
  85  *
  86  * New readdir implementation does readdir in hash order, and uses hash of a
  87  * file name as a telldir/seekdir cookie. This led to number of complications:
  88  *
  89  *     . hash is not unique, so it cannot be used to index cached directory
  90  *     pages on the client (note, that it requires a whole pageful of hash
  91  *     collided entries to cause two pages to have identical hashes);
  92  *
  93  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  94  *     entry cookie. ext3/htree has the same problem and lustre implementation
  95  *     mimics their solution: seekdir(hash) positions directory at the first
  96  *     entry with the given hash.
  97  *
  98  * Client side.
  99  *
 100  * 0. caching
 101  *
 102  * Client caches directory pages using hash of the first entry as an index. As
 103  * noted above hash is not unique, so this solution doesn't work as is:
 104  * special processing is needed for "page hash chains" (i.e., sequences of
 105  * pages filled with entries all having the same hash value).
 106  *
 107  * First, such chains have to be detected. To this end, server returns to the
 108  * client the hash of the first entry on the page next to one returned. When
 109  * client detects that this hash is the same as hash of the first entry on the
 110  * returned page, page hash collision has to be handled. Pages in the
 111  * hash chain, except first one, are termed "overflow pages".
 112  *
 113  * Solution to index uniqueness problem is to not cache overflow
 114  * pages. Instead, when page hash collision is detected, all overflow pages
 115  * from emerging chain are immediately requested from the server and placed in
 116  * a special data structure (struct ll_dir_chain). This data structure is used
 117  * by ll_readdir() to process entries from overflow pages. When readdir
 118  * invocation finishes, overflow pages are discarded. If page hash collision
 119  * chain weren't completely processed, next call to readdir will again detect
 120  * page hash collision, again read overflow pages in, process next portion of
 121  * entries and again discard the pages. This is not as wasteful as it looks,
 122  * because, given reasonable hash, page hash collisions are extremely rare.
 123  *
 124  * 1. directory positioning
 125  *
 126  * When seekdir(hash) is called, original
 127  *
 128  *
 129  *
 130  *
 131  *
 132  *
 133  *
 134  *
 135  * Server.
 136  *
 137  * identification of and access to overflow pages
 138  *
 139  * page format
 140  *
 141  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
 142  * a header lu_dirpage which describes the start/end hash, and whether this
 143  * page is empty (contains no dir entry) or hash collide with next page.
 144  * After client receives reply, several pages will be integrated into dir page
 145  * in CFS_PAGE_SIZE (if CFS_PAGE_SIZE greater than LU_PAGE_SIZE), and the
 146  * lu_dirpage for this integrated page will be adjusted.
 147  *
 148  */
 149
 150 /* returns the page unlocked, but with a reference */
 151 static int ll_dir_readpage(struct file *file, struct page *page0)
 152 {
 153         struct inode *inode = page0->mapping->host;
 154         int hash64 = ll_i2sbi(inode)->ll_flags & LL_SBI_64BIT_HASH;
 155         struct obd_export *exp = ll_i2sbi(inode)->ll_md_exp;
 156         struct ptlrpc_request *request;
 157         struct mdt_body *body;
 158         struct md_op_data *op_data;
 159         __u64 hash;
 160         struct page **page_pool;
 161         struct page *page;
 162 #ifndef HAVE_ADD_TO_PAGE_CACHE_LRU
 163         struct pagevec lru_pvec;
 164 #endif
 165         struct lu_dirpage *dp;
 166         int max_pages = ll_i2sbi(inode)->ll_md_brw_size >> CFS_PAGE_SHIFT;
 167         int nrdpgs = 0; /* number of pages read actually */
 168         int npages;
 169         int i;
 170         int rc;
 171         ENTRY;
 172
 173         if (file) {
 174                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
 175
 176                 hash = fd->fd_dir.lfd_next;
 177         } else {
 178                 hash = ll_i2info(inode)->lli_sa_pos;
 179         }
 180         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) hash "LPU64"\n",
 181                inode->i_ino, inode->i_generation, inode, hash);
 182
 183         LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
 184
 185         OBD_ALLOC(page_pool, sizeof(page) * max_pages);
 186         if (page_pool != NULL) {
 187                 page_pool[0] = page0;
 188         } else {
 189                 page_pool = &page0;
 190                 max_pages = 1;
 191         }
 192         for (npages = 1; npages < max_pages; npages++) {
 193                 page = page_cache_alloc_cold(inode->i_mapping);
 194                 if (!page)
 195                         break;
 196                 page_pool[npages] = page;
 197         }
 198
 199         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 200                                      LUSTRE_OPC_ANY, NULL);
 201         op_data->op_npages = npages;
 202         op_data->op_offset = hash;
 203         rc = md_readpage(exp, op_data, page_pool, &request);
 204         ll_finish_md_op_data(op_data);
 205         if (rc == 0) {
 206                 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 207                 /* Checked by mdc_readpage() */
 208                 LASSERT(body != NULL);
 209
 210                 if (body->valid & OBD_MD_FLSIZE)
 211                         cl_isize_write(inode, body->size);
 212
 213                 nrdpgs = (request->rq_bulk->bd_nob_transferred+CFS_PAGE_SIZE-1)
 214                          >> CFS_PAGE_SHIFT;
 215                 SetPageUptodate(page0);
 216         }
 217         unlock_page(page0);
 218         ptlrpc_req_finished(request);
 219
 220         CDEBUG(D_VFSTRACE, "read %d/%d pages\n", nrdpgs, npages);
 221
 222         ll_pagevec_init(&lru_pvec, 0);
 223         for (i = 1; i < npages; i++) {
 224                 unsigned long offset;
 225                 int ret;
 226
 227                 page = page_pool[i];
 228
 229                 if (rc < 0 || i >= nrdpgs) {
 230                         page_cache_release(page);
 231                         continue;
 232                 }
 233
 234                 SetPageUptodate(page);
 235
 236                 dp = cfs_kmap(page);
 237                 hash = le64_to_cpu(dp->ldp_hash_start);
 238                 cfs_kunmap(page);
 239
 240                 offset = hash_x_index(hash, hash64);
 241
 242                 prefetchw(&page->flags);
 243                 ret = ll_add_to_page_cache_lru(page, inode->i_mapping, offset,
 244                                                GFP_KERNEL);
 245                 if (ret == 0) {
 246                         unlock_page(page);
 247                         if (ll_pagevec_add(&lru_pvec, page) == 0)
 248                                 ll_pagevec_lru_add_file(&lru_pvec);
 249                 } else {
 250                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
 251                                " %d\n", offset, ret);
 252                 }
 253                 page_cache_release(page);
 254         }
 255         ll_pagevec_lru_add_file(&lru_pvec);
 256
 257         if (page_pool != &page0)
 258                 OBD_FREE(page_pool, sizeof(struct page *) * max_pages);
 259         EXIT;
 260         return rc;
 261 }
 262
 263 #ifndef MS_HAS_NEW_AOPS
 264 struct address_space_operations ll_dir_aops = {
 265         .readpage  = ll_dir_readpage,
 266 };
 267 #else
 268 struct address_space_operations_ext ll_dir_aops = {
 269         .orig_aops.readpage  = ll_dir_readpage,
 270 };
 271 #endif
 272
 273 static void ll_check_page(struct inode *dir, struct page *page)
 274 {
 275         /* XXX: check page format later */
 276         SetPageChecked(page);
 277 }
 278
 279 void ll_release_page(struct page *page, int remove)
 280 {
 281         kunmap(page);
 282         if (remove) {
 283                 lock_page(page);
 284                 if (likely(page->mapping != NULL))
 285                         truncate_complete_page(page->mapping, page);
 286                 unlock_page(page);
 287         }
 288         page_cache_release(page);
 289 }
 290
 291 /*
 292  * Find, kmap and return page that contains given hash.
 293  */
 294 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
 295                                        __u64 *start, __u64 *end)
 296 {
 297         int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
 298         struct address_space *mapping = dir->i_mapping;
 299         /*
 300          * Complement of hash is used as an index so that
 301          * radix_tree_gang_lookup() can be used to find a page with starting
 302          * hash _smaller_ than one we are looking for.
 303          */
 304         unsigned long offset = hash_x_index(*hash, hash64);
 305         struct page *page;
 306         int found;
 307
 308         TREE_READ_LOCK_IRQ(mapping);
 309         found = radix_tree_gang_lookup(&mapping->page_tree,
 310                                        (void **)&page, offset, 1);
 311         if (found > 0) {
 312                 struct lu_dirpage *dp;
 313
 314                 page_cache_get(page);
 315                 TREE_READ_UNLOCK_IRQ(mapping);
 316                 /*
 317                  * In contrast to find_lock_page() we are sure that directory
 318                  * page cannot be truncated (while DLM lock is held) and,
 319                  * hence, can avoid restart.
 320                  *
 321                  * In fact, page cannot be locked here at all, because
 322                  * ll_dir_readpage() does synchronous io.
 323                  */
 324                 wait_on_page(page);
 325                 if (PageUptodate(page)) {
 326                         dp = cfs_kmap(page);
 327                         if (BITS_PER_LONG == 32 && hash64) {
 328                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
 329                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
 330                                 *hash  = *hash >> 32;
 331                         } else {
 332                                 *start = le64_to_cpu(dp->ldp_hash_start);
 333                                 *end   = le64_to_cpu(dp->ldp_hash_end);
 334                         }
 335                         LASSERTF(*start <= *hash, "start = "LPX64",end = "
 336                                  LPX64",hash = "LPX64"\n", *start, *end, *hash);
 337                         CDEBUG(D_VFSTRACE, "page %lu [%llu %llu], hash "LPU64"\n",
 338                                offset, *start, *end, *hash);
 339                         if (*hash > *end) {
 340                                 ll_release_page(page, 0);
 341                                 page = NULL;
 342                         } else if (*end != *start && *hash == *end) {
 343                                 /*
 344                                  * upon hash collision, remove this page,
 345                                  * otherwise put page reference, and
 346                                  * ll_get_dir_page() will issue RPC to fetch
 347                                  * the page we want.
 348                                  */
 349                                 ll_release_page(page,
 350                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
 351                                 page = NULL;
 352                         }
 353                 } else {
 354                         page_cache_release(page);
 355                         page = ERR_PTR(-EIO);
 356                 }
 357
 358         } else {
 359                 TREE_READ_UNLOCK_IRQ(mapping);
 360                 page = NULL;
 361         }
 362         return page;
 363 }
 364
 365 struct page *ll_get_dir_page(struct file *filp, struct inode *dir, __u64 hash,
 366                              struct ll_dir_chain *chain)
 367 {
 368         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 369         struct address_space *mapping = dir->i_mapping;
 370         struct lustre_handle lockh;
 371         struct lu_dirpage *dp;
 372         struct page *page;
 373         ldlm_mode_t mode;
 374         int rc;
 375         __u64 start = 0;
 376         __u64 end = 0;
 377         __u64 lhash = hash;
 378         struct ll_inode_info *lli = ll_i2info(dir);
 379         int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
 380
 381         mode = LCK_PR;
 382         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 383                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
 384         if (!rc) {
 385                 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
 386                        ll_md_blocking_ast, ldlm_completion_ast,
 387                        NULL, NULL, dir };
 388                 struct lookup_intent it = { .it_op = IT_READDIR };
 389                 struct ptlrpc_request *request;
 390                 struct md_op_data *op_data;
 391
 392                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
 393                                              LUSTRE_OPC_ANY, NULL);
 394                 if (IS_ERR(op_data))
 395                         return (void *)op_data;
 396
 397                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
 398                                 op_data, &lockh, NULL, 0, NULL, 0);
 399
 400                 ll_finish_md_op_data(op_data);
 401
 402                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 403                 if (request)
 404                         ptlrpc_req_finished(request);
 405                 if (rc < 0) {
 406                         CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
 407                                PFID(ll_inode2fid(dir)), hash, rc);
 408                         return ERR_PTR(rc);
 409                 }
 410         } else {
 411                 /* for cross-ref object, l_ast_data of the lock may not be set,
 412                  * we reset it here */
 413                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
 414                                  dir, NULL);
 415         }
 416         ldlm_lock_dump_handle(D_OTHER, &lockh);
 417
 418         cfs_mutex_lock(&lli->lli_readdir_mutex);
 419         page = ll_dir_page_locate(dir, &lhash, &start, &end);
 420         if (IS_ERR(page)) {
 421                 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
 422                        PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
 423                 GOTO(out_unlock, page);
 424         } else if (page != NULL) {
 425                 /*
 426                  * XXX nikita: not entirely correct handling of a corner case:
 427                  * suppose hash chain of entries with hash value HASH crosses
 428                  * border between pages P0 and P1. First both P0 and P1 are
 429                  * cached, seekdir() is called for some entry from the P0 part
 430                  * of the chain. Later P0 goes out of cache. telldir(HASH)
 431                  * happens and finds P1, as it starts with matching hash
 432                  * value. Remaining entries from P0 part of the chain are
 433                  * skipped. (Is that really a bug?)
 434                  *
 435                  * Possible solutions: 0. don't cache P1 is such case, handle
 436                  * it as an "overflow" page. 1. invalidate all pages at
 437                  * once. 2. use HASH|1 as an index for P1.
 438                  */
 439                 GOTO(hash_collision, page);
 440         }
 441
 442         page = read_cache_page(mapping, hash_x_index(hash, hash64),
 443                                (filler_t*)mapping->a_ops->readpage, filp);
 444         if (IS_ERR(page)) {
 445                 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
 446                        PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
 447                 GOTO(out_unlock, page);
 448         }
 449
 450         wait_on_page(page);
 451         (void)kmap(page);
 452         if (!PageUptodate(page)) {
 453                 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
 454                        PFID(ll_inode2fid(dir)), hash, -5);
 455                 goto fail;
 456         }
 457         if (!PageChecked(page))
 458                 ll_check_page(dir, page);
 459         if (PageError(page)) {
 460                 CERROR("page error: "DFID" at "LPU64": rc %d\n",
 461                        PFID(ll_inode2fid(dir)), hash, -5);
 462                 goto fail;
 463         }
 464 hash_collision:
 465         dp = page_address(page);
 466         if (BITS_PER_LONG == 32 && hash64) {
 467                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
 468                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
 469                 lhash = hash >> 32;
 470         } else {
 471                 start = le64_to_cpu(dp->ldp_hash_start);
 472                 end   = le64_to_cpu(dp->ldp_hash_end);
 473                 lhash = hash;
 474         }
 475         if (end == start) {
 476                 LASSERT(start == lhash);
 477                 CWARN("Page-wide hash collision: "LPU64"\n", end);
 478                 if (BITS_PER_LONG == 32 && hash64)
 479                         CWARN("Real page-wide hash collision at ["LPU64" "LPU64
 480                               "] with hash "LPU64"\n",
 481                               le64_to_cpu(dp->ldp_hash_start),
 482                               le64_to_cpu(dp->ldp_hash_end), hash);
 483                 /*
 484                  * Fetch whole overflow chain...
 485                  *
 486                  * XXX not yet.
 487                  */
 488                 goto fail;
 489         }
 490 out_unlock:
 491         cfs_mutex_unlock(&lli->lli_readdir_mutex);
 492         ldlm_lock_decref(&lockh, mode);
 493         return page;
 494
 495 fail:
 496         ll_release_page(page, 1);
 497         page = ERR_PTR(-EIO);
 498         goto out_unlock;
 499 }
 500
 501 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 502 {
 503         struct inode         *inode      = filp->f_dentry->d_inode;
 504         struct ll_inode_info *info       = ll_i2info(inode);
 505         struct ll_sb_info    *sbi        = ll_i2sbi(inode);
 506         struct ll_file_data  *fd         = LUSTRE_FPRIVATE(filp);
 507         __u64                 pos        = fd->fd_dir.lfd_pos;
 508         int                   api32      = ll_need_32bit_api(sbi);
 509         int                   hash64     = sbi->ll_flags & LL_SBI_64BIT_HASH;
 510         struct page          *page;
 511         struct ll_dir_chain   chain;
 512         int                   done;
 513         int                   rc;
 514         ENTRY;
 515
 516         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu 32bit_api %d\n",
 517                inode->i_ino, inode->i_generation, inode,
 518                (unsigned long)pos, i_size_read(inode), api32);
 519
 520         if (pos == MDS_DIR_END_OFF)
 521                 /*
 522                  * end-of-file.
 523                  */
 524                 GOTO(out, rc = 0);
 525
 526         rc    = 0;
 527         done  = 0;
 528         ll_dir_chain_init(&chain);
 529
 530         fd->fd_dir.lfd_next = pos;
 531         page = ll_get_dir_page(filp, inode, pos, &chain);
 532
 533         while (rc == 0 && !done) {
 534                 struct lu_dirpage *dp;
 535                 struct lu_dirent  *ent;
 536
 537                 if (!IS_ERR(page)) {
 538                         /*
 539                          * If page is empty (end of directory is reached),
 540                          * use this value.
 541                          */
 542                         __u64 hash = MDS_DIR_END_OFF;
 543                         __u64 next;
 544
 545                         dp = page_address(page);
 546                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 547                              ent = lu_dirent_next(ent)) {
 548                                 __u16          type;
 549                                 int            namelen;
 550                                 struct lu_fid  fid;
 551                                 __u64          lhash;
 552                                 __u64          ino;
 553
 554                                 /*
 555                                  * XXX: implement correct swabbing here.
 556                                  */
 557
 558                                 hash = le64_to_cpu(ent->lde_hash);
 559                                 if (hash < pos)
 560                                         /*
 561                                          * Skip until we find target hash
 562                                          * value.
 563                                          */
 564                                         continue;
 565
 566                                 namelen = le16_to_cpu(ent->lde_namelen);
 567                                 if (namelen == 0)
 568                                         /*
 569                                          * Skip dummy record.
 570                                          */
 571                                         continue;
 572
 573                                 if (api32 && hash64)
 574                                         lhash = hash >> 32;
 575                                 else
 576                                         lhash = hash;
 577                                 fid_le_to_cpu(&fid, &ent->lde_fid);
 578                                 ino = cl_fid_build_ino(&fid, api32);
 579                                 type = ll_dirent_type_get(ent);
 580                                 /* For 'll_nfs_get_name_filldir()', it will try
 581                                  * to access the 'ent' through its 'lde_name',
 582                                  * so the parameter 'name' for 'filldir()' must
 583                                  * be part of the 'ent'. */
 584                                 done = filldir(cookie, ent->lde_name, namelen,
 585                                                lhash, ino, type);
 586                         }
 587                         next = le64_to_cpu(dp->ldp_hash_end);
 588                         if (!done) {
 589                                 pos = next;
 590                                 if (pos == MDS_DIR_END_OFF) {
 591                                         /*
 592                                          * End of directory reached.
 593                                          */
 594                                         done = 1;
 595                                         ll_release_page(page, 0);
 596                                 } else if (1 /* chain is exhausted*/) {
 597                                         /*
 598                                          * Normal case: continue to the next
 599                                          * page.
 600                                          */
 601                                         ll_release_page(page,
 602                                             le32_to_cpu(dp->ldp_flags) &
 603                                                         LDF_COLLIDE);
 604                                         fd->fd_dir.lfd_next = pos;
 605                                         page = ll_get_dir_page(filp, inode, pos,
 606                                                                &chain);
 607                                 } else {
 608                                         /*
 609                                          * go into overflow page.
 610                                          */
 611                                         LASSERT(le32_to_cpu(dp->ldp_flags) &
 612                                                 LDF_COLLIDE);
 613                                         ll_release_page(page, 1);
 614                                 }
 615                         } else {
 616                                 pos = hash;
 617                                 ll_release_page(page, 0);
 618                         }
 619                 } else {
 620                         rc = PTR_ERR(page);
 621                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 622                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 623                 }
 624         }
 625
 626         fd->fd_dir.lfd_pos = pos;
 627         if (pos == MDS_DIR_END_OFF) {
 628                 if (api32)
 629                         filp->f_pos = LL_DIR_END_OFF_32BIT;
 630                 else
 631                         filp->f_pos = LL_DIR_END_OFF;
 632         } else {
 633                 if (api32 && hash64)
 634                         filp->f_pos = pos >> 32;
 635                 else
 636                         filp->f_pos = pos;
 637         }
 638         filp->f_version = inode->i_version;
 639         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 640
 641         ll_dir_chain_fini(&chain);
 642
 643 out:
 644         if (!rc)
 645                 ll_stats_ops_tally(sbi, LPROC_LL_READDIR, 1);
 646
 647         RETURN(rc);
 648 }
 649
 650 int ll_send_mgc_param(struct obd_export *mgc, char *string)
 651 {
 652         struct mgs_send_param *msp;
 653         int rc = 0;
 654
 655         OBD_ALLOC_PTR(msp);
 656         if (!msp)
 657                 return -ENOMEM;
 658
 659         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
 660         rc = obd_set_info_async(NULL, mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
 661                                 sizeof(struct mgs_send_param), msp, NULL);
 662         if (rc)
 663                 CERROR("Failed to set parameter: %d\n", rc);
 664         OBD_FREE_PTR(msp);
 665
 666         return rc;
 667 }
 668
 669 char *ll_get_fsname(struct inode *inode)
 670 {
 671         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 672         char *ptr, *fsname;
 673         int len;
 674
 675         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
 676         len = strlen(lsi->lsi_lmd->lmd_profile);
 677         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
 678         if (ptr && (strcmp(ptr, "-client") == 0))
 679                 len -= 7;
 680         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
 681         fsname[len] = '\0';
 682
 683         return fsname;
 684 }
 685
 686 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
 687                      int set_default)
 688 {
 689         struct ll_sb_info *sbi = ll_i2sbi(inode);
 690         struct md_op_data *op_data;
 691         struct ptlrpc_request *req = NULL;
 692         int rc = 0;
 693         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 694         struct obd_device *mgc = lsi->lsi_mgc;
 695         char *fsname = NULL, *param = NULL;
 696         int lum_size;
 697
 698         if (lump != NULL) {
 699                 /*
 700                  * This is coming from userspace, so should be in
 701                  * local endian.  But the MDS would like it in little
 702                  * endian, so we swab it before we send it.
 703                  */
 704                 switch (lump->lmm_magic) {
 705                 case LOV_USER_MAGIC_V1: {
 706                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
 707                                 lustre_swab_lov_user_md_v1(lump);
 708                         lum_size = sizeof(struct lov_user_md_v1);
 709                         break;
 710                         }
 711                 case LOV_USER_MAGIC_V3: {
 712                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
 713                                 lustre_swab_lov_user_md_v3(
 714                                         (struct lov_user_md_v3 *)lump);
 715                         lum_size = sizeof(struct lov_user_md_v3);
 716                         break;
 717                         }
 718                 default: {
 719                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 720                                         " %#08x != %#08x nor %#08x\n",
 721                                         lump->lmm_magic, LOV_USER_MAGIC_V1,
 722                                         LOV_USER_MAGIC_V3);
 723                         RETURN(-EINVAL);
 724                         }
 725                }
 726         } else {
 727                 lum_size = sizeof(struct lov_user_md_v1);
 728         }
 729
 730         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 731                                      LUSTRE_OPC_ANY, NULL);
 732         if (IS_ERR(op_data))
 733                 RETURN(PTR_ERR(op_data));
 734
 735         /* swabbing is done in lov_setstripe() on server side */
 736         rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
 737                         NULL, 0, &req, NULL);
 738         ll_finish_md_op_data(op_data);
 739         ptlrpc_req_finished(req);
 740         if (rc) {
 741                 if (rc != -EPERM && rc != -EACCES)
 742                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 743         }
 744
 745         /* In the following we use the fact that LOV_USER_MAGIC_V1 and
 746          LOV_USER_MAGIC_V3 have the same initial fields so we do not
 747          need the make the distiction between the 2 versions */
 748         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
 749                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
 750
 751                 /* Get fsname and assume devname to be -MDT0000. */
 752                 fsname = ll_get_fsname(inode);
 753                 /* Set root stripesize */
 754                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
 755                         lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
 756                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 757                 if (rc)
 758                         goto end;
 759
 760                 /* Set root stripecount */
 761                 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
 762                         lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
 763                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 764                 if (rc)
 765                         goto end;
 766
 767                 /* Set root stripeoffset */
 768                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
 769                         lump ? le16_to_cpu(lump->lmm_stripe_offset) :
 770                         (typeof(lump->lmm_stripe_offset))(-1));
 771                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 772                 if (rc)
 773                         goto end;
 774 end:
 775                 if (fsname)
 776                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
 777                 if (param)
 778                         OBD_FREE(param, MGS_PARAM_MAXLEN);
 779         }
 780         return rc;
 781 }
 782
 783 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 784                      int *lmm_size, struct ptlrpc_request **request)
 785 {
 786         struct ll_sb_info *sbi = ll_i2sbi(inode);
 787         struct mdt_body   *body;
 788         struct lov_mds_md *lmm = NULL;
 789         struct ptlrpc_request *req = NULL;
 790         int rc, lmmsize;
 791         struct md_op_data *op_data;
 792
 793         rc = ll_get_max_mdsize(sbi, &lmmsize);
 794         if (rc)
 795                 RETURN(rc);
 796
 797         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
 798                                      0, lmmsize, LUSTRE_OPC_ANY,
 799                                      NULL);
 800         if (IS_ERR(op_data))
 801                 RETURN(PTR_ERR(op_data));
 802
 803         op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
 804         rc = md_getattr(sbi->ll_md_exp, op_data, &req);
 805         ll_finish_md_op_data(op_data);
 806         if (rc < 0) {
 807                 CDEBUG(D_INFO, "md_getattr failed on inode "
 808                        "%lu/%u: rc %d\n", inode->i_ino,
 809                        inode->i_generation, rc);
 810                 GOTO(out, rc);
 811         }
 812
 813         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 814         LASSERT(body != NULL);
 815
 816         lmmsize = body->eadatasize;
 817
 818         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 819             lmmsize == 0) {
 820                 GOTO(out, rc = -ENODATA);
 821         }
 822
 823         lmm = req_capsule_server_sized_get(&req->rq_pill,
 824                                            &RMF_MDT_MD, lmmsize);
 825         LASSERT(lmm != NULL);
 826
 827         /*
 828          * This is coming from the MDS, so is probably in
 829          * little endian.  We convert it to host endian before
 830          * passing it to userspace.
 831          */
 832         /* We don't swab objects for directories */
 833         switch (le32_to_cpu(lmm->lmm_magic)) {
 834         case LOV_MAGIC_V1:
 835                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
 836                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
 837                 break;
 838         case LOV_MAGIC_V3:
 839                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
 840                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
 841                 break;
 842         default:
 843                 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
 844                 rc = -EPROTO;
 845         }
 846 out:
 847         *lmmp = lmm;
 848         *lmm_size = lmmsize;
 849         *request = req;
 850         return rc;
 851 }
 852
 853 /*
 854  *  Get MDT index for the inode.
 855  */
 856 int ll_get_mdt_idx(struct inode *inode)
 857 {
 858         struct ll_sb_info *sbi = ll_i2sbi(inode);
 859         struct md_op_data *op_data;
 860         int rc, mdtidx;
 861         ENTRY;
 862
 863         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
 864                                      0, LUSTRE_OPC_ANY, NULL);
 865         if (IS_ERR(op_data))
 866                 RETURN(PTR_ERR(op_data));
 867
 868         op_data->op_valid |= OBD_MD_MDTIDX;
 869         rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
 870         mdtidx = op_data->op_mds;
 871         ll_finish_md_op_data(op_data);
 872         if (rc < 0) {
 873                 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 874                 RETURN(rc);
 875         }
 876         return mdtidx;
 877 }
 878
 879 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
 880 {
 881         void *ptr;
 882         int rc;
 883
 884         OBD_ALLOC(ptr, len);
 885         if (ptr == NULL)
 886                 return -ENOMEM;
 887         if (cfs_copy_from_user(ptr, data, len)) {
 888                 OBD_FREE(ptr, len);
 889                 return -EFAULT;
 890         }
 891         rc = obd_iocontrol(cmd, exp, len, data, NULL);
 892         OBD_FREE(ptr, len);
 893         return rc;
 894 }
 895
 896 static int quotactl_ioctl(struct ll_sb_info *sbi, struct if_quotactl *qctl)
 897 {
 898         int cmd = qctl->qc_cmd;
 899         int type = qctl->qc_type;
 900         int id = qctl->qc_id;
 901         int valid = qctl->qc_valid;
 902         int rc = 0;
 903         ENTRY;
 904
 905         switch (cmd) {
 906         case LUSTRE_Q_INVALIDATE:
 907         case LUSTRE_Q_FINVALIDATE:
 908         case Q_QUOTAON:
 909         case Q_QUOTAOFF:
 910         case Q_SETQUOTA:
 911         case Q_SETINFO:
 912                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
 913                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
 914                         RETURN(-EPERM);
 915                 break;
 916         case Q_GETQUOTA:
 917                 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
 918                      (type == GRPQUOTA && !in_egroup_p(id))) &&
 919                     (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
 920                      sbi->ll_flags & LL_SBI_RMT_CLIENT))
 921                         RETURN(-EPERM);
 922                 break;
 923         case Q_GETINFO:
 924                 break;
 925         default:
 926                 CERROR("unsupported quotactl op: %#x\n", cmd);
 927                 RETURN(-ENOTTY);
 928         }
 929
 930         if (valid != QC_GENERAL) {
 931                 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
 932                         RETURN(-EOPNOTSUPP);
 933
 934                 if (cmd == Q_GETINFO)
 935                         qctl->qc_cmd = Q_GETOINFO;
 936                 else if (cmd == Q_GETQUOTA)
 937                         qctl->qc_cmd = Q_GETOQUOTA;
 938                 else
 939                         RETURN(-EINVAL);
 940
 941                 switch (valid) {
 942                 case QC_MDTIDX:
 943                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
 944                                            sizeof(*qctl), qctl, NULL);
 945                         break;
 946                 case QC_OSTIDX:
 947                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_dt_exp,
 948                                            sizeof(*qctl), qctl, NULL);
 949                         break;
 950                 case QC_UUID:
 951                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
 952                                            sizeof(*qctl), qctl, NULL);
 953                         if (rc == -EAGAIN)
 954                                 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
 955                                                    sbi->ll_dt_exp,
 956                                                    sizeof(*qctl), qctl, NULL);
 957                         break;
 958                 default:
 959                         rc = -EINVAL;
 960                         break;
 961                 }
 962
 963                 if (rc)
 964                         RETURN(rc);
 965
 966                 qctl->qc_cmd = cmd;
 967         } else {
 968                 struct obd_quotactl *oqctl;
 969
 970                 OBD_ALLOC_PTR(oqctl);
 971                 if (oqctl == NULL)
 972                         RETURN(-ENOMEM);
 973
 974                 QCTL_COPY(oqctl, qctl);
 975                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
 976                 if (rc) {
 977                         if (rc != -EALREADY && cmd == Q_QUOTAON) {
 978                                 oqctl->qc_cmd = Q_QUOTAOFF;
 979                                 obd_quotactl(sbi->ll_md_exp, oqctl);
 980                         }
 981                         OBD_FREE_PTR(oqctl);
 982                         RETURN(rc);
 983                 }
 984                 /* If QIF_SPACE is not set, client should collect the
 985                  * space usage from OSSs by itself */
 986                 if (cmd == Q_GETQUOTA &&
 987                     !(oqctl->qc_dqblk.dqb_valid & QIF_SPACE) &&
 988                     !oqctl->qc_dqblk.dqb_curspace) {
 989                         struct obd_quotactl *oqctl_tmp;
 990
 991                         OBD_ALLOC_PTR(oqctl_tmp);
 992                         if (oqctl_tmp == NULL)
 993                                 GOTO(out, rc = -ENOMEM);
 994
 995                         oqctl_tmp->qc_cmd = Q_GETOQUOTA;
 996                         oqctl_tmp->qc_id = oqctl->qc_id;
 997                         oqctl_tmp->qc_type = oqctl->qc_type;
 998
 999                         /* collect space usage from OSTs */
1000                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1001                         rc = obd_quotactl(sbi->ll_dt_exp, oqctl_tmp);
1002                         if (!rc || rc == -EREMOTEIO) {
1003                                 oqctl->qc_dqblk.dqb_curspace =
1004                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1005                                 oqctl->qc_dqblk.dqb_valid |= QIF_SPACE;
1006                         }
1007
1008                         /* collect space & inode usage from MDTs */
1009                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1010                         oqctl_tmp->qc_dqblk.dqb_curinodes = 0;
1011                         rc = obd_quotactl(sbi->ll_md_exp, oqctl_tmp);
1012                         if (!rc || rc == -EREMOTEIO) {
1013                                 oqctl->qc_dqblk.dqb_curspace +=
1014                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1015                                 oqctl->qc_dqblk.dqb_curinodes =
1016                                         oqctl_tmp->qc_dqblk.dqb_curinodes;
1017                                 oqctl->qc_dqblk.dqb_valid |= QIF_INODES;
1018                         } else {
1019                                 oqctl->qc_dqblk.dqb_valid &= ~QIF_SPACE;
1020                         }
1021
1022                         OBD_FREE_PTR(oqctl_tmp);
1023                 }
1024 out:
1025                 QCTL_COPY(qctl, oqctl);
1026                 OBD_FREE_PTR(oqctl);
1027         }
1028
1029         RETURN(rc);
1030 }
1031
1032 static long ll_dir_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1033 {
1034         struct inode *inode = file->f_dentry->d_inode;
1035         struct ll_sb_info *sbi = ll_i2sbi(inode);
1036         struct obd_ioctl_data *data;
1037         int rc = 0;
1038         ENTRY;
1039
1040         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
1041                inode->i_ino, inode->i_generation, inode, cmd);
1042
1043         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
1044         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
1045                 return -ENOTTY;
1046
1047         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
1048         switch(cmd) {
1049         case FSFILT_IOC_GETFLAGS:
1050         case FSFILT_IOC_SETFLAGS:
1051                 RETURN(ll_iocontrol(inode, file, cmd, arg));
1052         case FSFILT_IOC_GETVERSION_OLD:
1053         case FSFILT_IOC_GETVERSION:
1054                 RETURN(put_user(inode->i_generation, (int *)arg));
1055         /* We need to special case any other ioctls we want to handle,
1056          * to send them to the MDS/OST as appropriate and to properly
1057          * network encode the arg field.
1058         case FSFILT_IOC_SETVERSION_OLD:
1059         case FSFILT_IOC_SETVERSION:
1060         */
1061         case LL_IOC_GET_MDTIDX: {
1062                 int mdtidx;
1063
1064                 mdtidx = ll_get_mdt_idx(inode);
1065                 if (mdtidx < 0)
1066                         RETURN(mdtidx);
1067
1068                 if (put_user((int)mdtidx, (int*)arg))
1069                         RETURN(-EFAULT);
1070
1071                 return 0;
1072         }
1073         case IOC_MDC_LOOKUP: {
1074                 struct ptlrpc_request *request = NULL;
1075                 int namelen, len = 0;
1076                 char *buf = NULL;
1077                 char *filename;
1078                 struct md_op_data *op_data;
1079
1080                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1081                 if (rc)
1082                         RETURN(rc);
1083                 data = (void *)buf;
1084
1085                 filename = data->ioc_inlbuf1;
1086                 namelen = strlen(filename);
1087
1088                 if (namelen < 1) {
1089                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1090                         GOTO(out_free, rc = -EINVAL);
1091                 }
1092
1093                 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1094                                              0, LUSTRE_OPC_ANY, NULL);
1095                 if (IS_ERR(op_data))
1096                         GOTO(out_free, rc = PTR_ERR(op_data));
1097
1098                 op_data->op_valid = OBD_MD_FLID;
1099                 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1100                 ll_finish_md_op_data(op_data);
1101                 if (rc < 0) {
1102                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1103                         GOTO(out_free, rc);
1104                 }
1105                 ptlrpc_req_finished(request);
1106                 EXIT;
1107 out_free:
1108                 obd_ioctl_freedata(buf, len);
1109                 return rc;
1110         }
1111         case LL_IOC_LOV_SETSTRIPE: {
1112                 struct lov_user_md_v3 lumv3;
1113                 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1114                 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1115                 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1116
1117                 int set_default = 0;
1118
1119                 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1120                 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1121                         sizeof(lumv3p->lmm_objects[0]));
1122                 /* first try with v1 which is smaller than v3 */
1123                 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1124                         RETURN(-EFAULT);
1125
1126                 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1127                         if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1128                                 RETURN(-EFAULT);
1129                 }
1130
1131                 if (inode->i_sb->s_root == file->f_dentry)
1132                         set_default = 1;
1133
1134                 /* in v1 and v3 cases lumv1 points to data */
1135                 rc = ll_dir_setstripe(inode, lumv1, set_default);
1136
1137                 RETURN(rc);
1138         }
1139         case LL_IOC_OBD_STATFS:
1140                 RETURN(ll_obd_statfs(inode, (void *)arg));
1141         case LL_IOC_LOV_GETSTRIPE:
1142         case LL_IOC_MDC_GETINFO:
1143         case IOC_MDC_GETFILEINFO:
1144         case IOC_MDC_GETFILESTRIPE: {
1145                 struct ptlrpc_request *request = NULL;
1146                 struct lov_user_md *lump;
1147                 struct lov_mds_md *lmm = NULL;
1148                 struct mdt_body *body;
1149                 char *filename = NULL;
1150                 int lmmsize;
1151
1152                 if (cmd == IOC_MDC_GETFILEINFO ||
1153                     cmd == IOC_MDC_GETFILESTRIPE) {
1154                         filename = getname((const char *)arg);
1155                         if (IS_ERR(filename))
1156                                 RETURN(PTR_ERR(filename));
1157
1158                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1159                                                       &lmmsize, &request);
1160                 } else {
1161                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1162                 }
1163
1164                 if (request) {
1165                         body = req_capsule_server_get(&request->rq_pill,
1166                                                       &RMF_MDT_BODY);
1167                         LASSERT(body != NULL);
1168                 } else {
1169                         GOTO(out_req, rc);
1170                 }
1171
1172                 if (rc < 0) {
1173                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1174                                                cmd == LL_IOC_MDC_GETINFO))
1175                                 GOTO(skip_lmm, rc = 0);
1176                         else
1177                                 GOTO(out_req, rc);
1178                 }
1179
1180                 if (cmd == IOC_MDC_GETFILESTRIPE ||
1181                     cmd == LL_IOC_LOV_GETSTRIPE) {
1182                         lump = (struct lov_user_md *)arg;
1183                 } else {
1184                         struct lov_user_mds_data *lmdp;
1185                         lmdp = (struct lov_user_mds_data *)arg;
1186                         lump = &lmdp->lmd_lmm;
1187                 }
1188                 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
1189                         if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
1190                                 GOTO(out_req, rc = -EFAULT);
1191                         rc = -EOVERFLOW;
1192                 }
1193         skip_lmm:
1194                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1195                         struct lov_user_mds_data *lmdp;
1196                         lstat_t st = { 0 };
1197
1198                         st.st_dev     = inode->i_sb->s_dev;
1199                         st.st_mode    = body->mode;
1200                         st.st_nlink   = body->nlink;
1201                         st.st_uid     = body->uid;
1202                         st.st_gid     = body->gid;
1203                         st.st_rdev    = body->rdev;
1204                         st.st_size    = body->size;
1205                         st.st_blksize = CFS_PAGE_SIZE;
1206                         st.st_blocks  = body->blocks;
1207                         st.st_atime   = body->atime;
1208                         st.st_mtime   = body->mtime;
1209                         st.st_ctime   = body->ctime;
1210                         st.st_ino     = inode->i_ino;
1211
1212                         lmdp = (struct lov_user_mds_data *)arg;
1213                         if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1214                                 GOTO(out_req, rc = -EFAULT);
1215                 }
1216
1217                 EXIT;
1218         out_req:
1219                 ptlrpc_req_finished(request);
1220                 if (filename)
1221                         putname(filename);
1222                 return rc;
1223         }
1224         case IOC_LOV_GETINFO: {
1225                 struct lov_user_mds_data *lumd;
1226                 struct lov_stripe_md *lsm;
1227                 struct lov_user_md *lum;
1228                 struct lov_mds_md *lmm;
1229                 int lmmsize;
1230                 lstat_t st;
1231
1232                 lumd = (struct lov_user_mds_data *)arg;
1233                 lum = &lumd->lmd_lmm;
1234
1235                 rc = ll_get_max_mdsize(sbi, &lmmsize);
1236                 if (rc)
1237                         RETURN(rc);
1238
1239                 OBD_ALLOC_LARGE(lmm, lmmsize);
1240                 if (cfs_copy_from_user(lmm, lum, lmmsize))
1241                         GOTO(free_lmm, rc = -EFAULT);
1242
1243                 switch (lmm->lmm_magic) {
1244                 case LOV_USER_MAGIC_V1:
1245                         if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1246                                 break;
1247                         /* swab objects first so that stripes num will be sane */
1248                         lustre_swab_lov_user_md_objects(
1249                                 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1250                                 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1251                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1252                         break;
1253                 case LOV_USER_MAGIC_V3:
1254                         if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1255                                 break;
1256                         /* swab objects first so that stripes num will be sane */
1257                         lustre_swab_lov_user_md_objects(
1258                                 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1259                                 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1260                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1261                         break;
1262                 default:
1263                         GOTO(free_lmm, rc = -EINVAL);
1264                 }
1265
1266                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1267                 if (rc < 0)
1268                         GOTO(free_lmm, rc = -ENOMEM);
1269
1270                 /* Perform glimpse_size operation. */
1271                 memset(&st, 0, sizeof(st));
1272
1273                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1274                 if (rc)
1275                         GOTO(free_lsm, rc);
1276
1277                 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1278                         GOTO(free_lsm, rc = -EFAULT);
1279
1280                 EXIT;
1281         free_lsm:
1282                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1283         free_lmm:
1284                 OBD_FREE_LARGE(lmm, lmmsize);
1285                 return rc;
1286         }
1287         case OBD_IOC_LLOG_CATINFO: {
1288                 RETURN(-EOPNOTSUPP);
1289         }
1290         case OBD_IOC_QUOTACHECK: {
1291                 struct obd_quotactl *oqctl;
1292                 int error = 0;
1293
1294                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1295                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1296                         RETURN(-EPERM);
1297
1298                 OBD_ALLOC_PTR(oqctl);
1299                 if (!oqctl)
1300                         RETURN(-ENOMEM);
1301                 oqctl->qc_type = arg;
1302                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1303                 if (rc < 0) {
1304                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1305                         error = rc;
1306                 }
1307
1308                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1309                 if (rc < 0)
1310                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1311
1312                 OBD_FREE_PTR(oqctl);
1313                 return error ?: rc;
1314         }
1315         case OBD_IOC_POLL_QUOTACHECK: {
1316                 struct if_quotacheck *check;
1317
1318                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1319                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1320                         RETURN(-EPERM);
1321
1322                 OBD_ALLOC_PTR(check);
1323                 if (!check)
1324                         RETURN(-ENOMEM);
1325
1326                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1327                                    NULL);
1328                 if (rc) {
1329                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1330                         if (cfs_copy_to_user((void *)arg, check,
1331                                              sizeof(*check)))
1332                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1333                         GOTO(out_poll, rc);
1334                 }
1335
1336                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1337                                    NULL);
1338                 if (rc) {
1339                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1340                         if (cfs_copy_to_user((void *)arg, check,
1341                                              sizeof(*check)))
1342                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1343                         GOTO(out_poll, rc);
1344                 }
1345         out_poll:
1346                 OBD_FREE_PTR(check);
1347                 RETURN(rc);
1348         }
1349 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
1350         case LL_IOC_QUOTACTL_18: {
1351                 /* copy the old 1.x quota struct for internal use, then copy
1352                  * back into old format struct.  For 1.8 compatibility. */
1353                 struct if_quotactl_18 *qctl_18;
1354                 struct if_quotactl *qctl_20;
1355
1356                 OBD_ALLOC_PTR(qctl_18);
1357                 if (!qctl_18)
1358                         RETURN(-ENOMEM);
1359
1360                 OBD_ALLOC_PTR(qctl_20);
1361                 if (!qctl_20)
1362                         GOTO(out_quotactl_18, rc = -ENOMEM);
1363
1364                 if (cfs_copy_from_user(qctl_18, (void *)arg, sizeof(*qctl_18)))
1365                         GOTO(out_quotactl_20, rc = -ENOMEM);
1366
1367                 QCTL_COPY(qctl_20, qctl_18);
1368                 qctl_20->qc_idx = 0;
1369
1370                 /* XXX: dqb_valid was borrowed as a flag to mark that
1371                  *      only mds quota is wanted */
1372                 if (qctl_18->qc_cmd == Q_GETQUOTA &&
1373                     qctl_18->qc_dqblk.dqb_valid) {
1374                         qctl_20->qc_valid = QC_MDTIDX;
1375                         qctl_20->qc_dqblk.dqb_valid = 0;
1376                 } else if (qctl_18->obd_uuid.uuid[0] != '\0') {
1377                         qctl_20->qc_valid = QC_UUID;
1378                         qctl_20->obd_uuid = qctl_18->obd_uuid;
1379                 } else {
1380                         qctl_20->qc_valid = QC_GENERAL;
1381                 }
1382
1383                 rc = quotactl_ioctl(sbi, qctl_20);
1384
1385                 if (rc == 0) {
1386                         QCTL_COPY(qctl_18, qctl_20);
1387                         qctl_18->obd_uuid = qctl_20->obd_uuid;
1388
1389                         if (cfs_copy_to_user((void *)arg, qctl_18,
1390                                              sizeof(*qctl_18)))
1391                                 rc = -EFAULT;
1392                 }
1393
1394         out_quotactl_20:
1395                 OBD_FREE_PTR(qctl_20);
1396         out_quotactl_18:
1397                 OBD_FREE_PTR(qctl_18);
1398                 RETURN(rc);
1399         }
1400 #else
1401 #warning "remove old LL_IOC_QUOTACTL_18 compatibility code"
1402 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0) */
1403         case LL_IOC_QUOTACTL: {
1404                 struct if_quotactl *qctl;
1405
1406                 OBD_ALLOC_PTR(qctl);
1407                 if (!qctl)
1408                         RETURN(-ENOMEM);
1409
1410                 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1411                         GOTO(out_quotactl, rc = -EFAULT);
1412
1413                 rc = quotactl_ioctl(sbi, qctl);
1414
1415                 if (rc == 0 && cfs_copy_to_user((void *)arg,qctl,sizeof(*qctl)))
1416                         rc = -EFAULT;
1417
1418         out_quotactl:
1419                 OBD_FREE_PTR(qctl);
1420                 RETURN(rc);
1421         }
1422         case OBD_IOC_GETDTNAME:
1423         case OBD_IOC_GETMDNAME:
1424                 RETURN(ll_get_obd_name(inode, cmd, arg));
1425         case LL_IOC_FLUSHCTX:
1426                 RETURN(ll_flush_ctx(inode));
1427 #ifdef CONFIG_FS_POSIX_ACL
1428         case LL_IOC_RMTACL: {
1429             if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1430                 inode == inode->i_sb->s_root->d_inode) {
1431                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1432
1433                 LASSERT(fd != NULL);
1434                 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1435                 if (!rc)
1436                         fd->fd_flags |= LL_FILE_RMTACL;
1437                 RETURN(rc);
1438             } else
1439                 RETURN(0);
1440         }
1441 #endif
1442         case LL_IOC_GETOBDCOUNT: {
1443                 int count, vallen;
1444                 struct obd_export *exp;
1445
1446                 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1447                         RETURN(-EFAULT);
1448
1449                 /* get ost count when count is zero, get mdt count otherwise */
1450                 exp = count ? sbi->ll_md_exp : sbi->ll_dt_exp;
1451                 vallen = sizeof(count);
1452                 rc = obd_get_info(NULL, exp, sizeof(KEY_TGT_COUNT),
1453                                   KEY_TGT_COUNT, &vallen, &count, NULL);
1454                 if (rc) {
1455                         CERROR("get target count failed: %d\n", rc);
1456                         RETURN(rc);
1457                 }
1458
1459                 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1460                         RETURN(-EFAULT);
1461
1462                 RETURN(0);
1463         }
1464         case LL_IOC_PATH2FID:
1465                 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1466                                      sizeof(struct lu_fid)))
1467                         RETURN(-EFAULT);
1468                 RETURN(0);
1469         case LL_IOC_GET_CONNECT_FLAGS: {
1470                 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1471         }
1472         case OBD_IOC_CHANGELOG_SEND:
1473         case OBD_IOC_CHANGELOG_CLEAR:
1474                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1475                                     sizeof(struct ioc_changelog));
1476                 RETURN(rc);
1477         case OBD_IOC_FID2PATH:
1478                 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1479         case LL_IOC_HSM_CT_START:
1480                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1481                                     sizeof(struct lustre_kernelcomm));
1482                 RETURN(rc);
1483
1484         default:
1485                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1486         }
1487 }
1488
1489 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1490 {
1491         struct inode *inode = file->f_mapping->host;
1492         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1493         struct ll_sb_info *sbi = ll_i2sbi(inode);
1494         int api32 = ll_need_32bit_api(sbi);
1495         loff_t ret = -EINVAL;
1496         ENTRY;
1497
1498         cfs_mutex_lock(&inode->i_mutex);
1499         switch (origin) {
1500                 case SEEK_SET:
1501                         break;
1502                 case SEEK_CUR:
1503                         offset += file->f_pos;
1504                         break;
1505                 case SEEK_END:
1506                         if (offset > 0)
1507                                 GOTO(out, ret);
1508                         if (api32)
1509                                 offset += LL_DIR_END_OFF_32BIT;
1510                         else
1511                                 offset += LL_DIR_END_OFF;
1512                         break;
1513                 default:
1514                         GOTO(out, ret);
1515         }
1516
1517         if (offset >= 0 &&
1518             ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1519              (!api32 && offset <= LL_DIR_END_OFF))) {
1520                 if (offset != file->f_pos) {
1521                         if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1522                             (!api32 && offset == LL_DIR_END_OFF))
1523                                 fd->fd_dir.lfd_pos = MDS_DIR_END_OFF;
1524                         else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1525                                 fd->fd_dir.lfd_pos = offset << 32;
1526                         else
1527                                 fd->fd_dir.lfd_pos = offset;
1528                         file->f_pos = offset;
1529                         file->f_version = 0;
1530                 }
1531                 ret = offset;
1532         }
1533         GOTO(out, ret);
1534
1535 out:
1536         cfs_mutex_unlock(&inode->i_mutex);
1537         return ret;
1538 }
1539
1540 int ll_dir_open(struct inode *inode, struct file *file)
1541 {
1542         ENTRY;
1543         RETURN(ll_file_open(inode, file));
1544 }
1545
1546 int ll_dir_release(struct inode *inode, struct file *file)
1547 {
1548         ENTRY;
1549         RETURN(ll_file_release(inode, file));
1550 }
1551
1552 struct file_operations ll_dir_operations = {
1553         .llseek   = ll_dir_seek,
1554         .open     = ll_dir_open,
1555         .release  = ll_dir_release,
1556         .read     = generic_read_dir,
1557         .readdir  = ll_readdir,
1558         .unlocked_ioctl   = ll_dir_ioctl,
1559         .fsync    = ll_fsync,
1560 };