lustre/llite/dir.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/llite/dir.c
  39  *
  40  * Directory code for lustre client.
  41  */
  42
  43 #include <linux/fs.h>
  44 #include <linux/pagemap.h>
  45 #include <linux/mm.h>
  46 #include <linux/version.h>
  47 #include <linux/smp_lock.h>
  48 #include <asm/uaccess.h>
  49 #include <linux/buffer_head.h>   // for wait_on_buffer
  50 #include <linux/pagevec.h>
  51
  52 #define DEBUG_SUBSYSTEM S_LLITE
  53
  54 #include <obd_support.h>
  55 #include <obd_class.h>
  56 #include <lustre_lib.h>
  57 #include <lustre/lustre_idl.h>
  58 #include <lustre_lite.h>
  59 #include <lustre_dlm.h>
  60 #include <lustre_fid.h>
  61 #include "llite_internal.h"
  62
  63 #ifndef HAVE_PAGE_CHECKED
  64 #ifdef HAVE_PG_FS_MISC
  65 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
  66 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
  67 #else
  68 #error PageChecked or PageFsMisc not defined in kernel
  69 #endif
  70 #endif
  71
  72 /*
  73  * (new) readdir implementation overview.
  74  *
  75  * Original lustre readdir implementation cached exact copy of raw directory
  76  * pages on the client. These pages were indexed in client page cache by
  77  * logical offset in the directory file. This design, while very simple and
  78  * intuitive had some inherent problems:
  79  *
  80  *     . it implies that byte offset to the directory entry serves as a
  81  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  82  *     ext3/htree directory entries may move due to splits, and more
  83  *     importantly,
  84  *
  85  *     . it is incompatible with the design of split directories for cmd3,
  86  *     that assumes that names are distributed across nodes based on their
  87  *     hash, and so readdir should be done in hash order.
  88  *
  89  * New readdir implementation does readdir in hash order, and uses hash of a
  90  * file name as a telldir/seekdir cookie. This led to number of complications:
  91  *
  92  *     . hash is not unique, so it cannot be used to index cached directory
  93  *     pages on the client (note, that it requires a whole pageful of hash
  94  *     collided entries to cause two pages to have identical hashes);
  95  *
  96  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  97  *     entry cookie. ext3/htree has the same problem and lustre implementation
  98  *     mimics their solution: seekdir(hash) positions directory at the first
  99  *     entry with the given hash.
 100  *
 101  * Client side.
 102  *
 103  * 0. caching
 104  *
 105  * Client caches directory pages using hash of the first entry as an index. As
 106  * noted above hash is not unique, so this solution doesn't work as is:
 107  * special processing is needed for "page hash chains" (i.e., sequences of
 108  * pages filled with entries all having the same hash value).
 109  *
 110  * First, such chains have to be detected. To this end, server returns to the
 111  * client the hash of the first entry on the page next to one returned. When
 112  * client detects that this hash is the same as hash of the first entry on the
 113  * returned page, page hash collision has to be handled. Pages in the
 114  * hash chain, except first one, are termed "overflow pages".
 115  *
 116  * Solution to index uniqueness problem is to not cache overflow
 117  * pages. Instead, when page hash collision is detected, all overflow pages
 118  * from emerging chain are immediately requested from the server and placed in
 119  * a special data structure (struct ll_dir_chain). This data structure is used
 120  * by ll_readdir() to process entries from overflow pages. When readdir
 121  * invocation finishes, overflow pages are discarded. If page hash collision
 122  * chain weren't completely processed, next call to readdir will again detect
 123  * page hash collision, again read overflow pages in, process next portion of
 124  * entries and again discard the pages. This is not as wasteful as it looks,
 125  * because, given reasonable hash, page hash collisions are extremely rare.
 126  *
 127  * 1. directory positioning
 128  *
 129  * When seekdir(hash) is called, original
 130  *
 131  *
 132  *
 133  *
 134  *
 135  *
 136  *
 137  *
 138  * Server.
 139  *
 140  * identification of and access to overflow pages
 141  *
 142  * page format
 143  *
 144  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
 145  * a header lu_dirpage which describes the start/end hash, and whether this
 146  * page is empty (contains no dir entry) or hash collide with next page.
 147  * After client receives reply, several pages will be integrated into dir page
 148  * in CFS_PAGE_SIZE (if CFS_PAGE_SIZE greater than LU_PAGE_SIZE), and the
 149  * lu_dirpage for this integrated page will be adjusted.
 150  *
 151  */
 152
 153 /* returns the page unlocked, but with a reference */
 154 static int ll_dir_readpage(struct file *file, struct page *page0)
 155 {
 156         struct inode *inode = page0->mapping->host;
 157         int hash64 = ll_i2sbi(inode)->ll_flags & LL_SBI_64BIT_HASH;
 158         struct obd_export *exp = ll_i2sbi(inode)->ll_md_exp;
 159         struct ptlrpc_request *request;
 160         struct mdt_body *body;
 161         struct md_op_data *op_data;
 162         __u64 hash;
 163         struct page **page_pool;
 164         struct page *page;
 165 #ifndef HAVE_ADD_TO_PAGE_CACHE_LRU
 166         struct pagevec lru_pvec;
 167 #endif
 168         struct lu_dirpage *dp;
 169         int max_pages = ll_i2sbi(inode)->ll_md_brw_size >> CFS_PAGE_SHIFT;
 170         int nrdpgs = 0; /* number of pages read actually */
 171         int npages;
 172         int i;
 173         int rc;
 174         ENTRY;
 175
 176         if (file) {
 177                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
 178
 179                 hash = fd->fd_dir.lfd_next;
 180         } else {
 181                 hash = ll_i2info(inode)->lli_sa_pos;
 182         }
 183         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) hash "LPU64"\n",
 184                inode->i_ino, inode->i_generation, inode, hash);
 185
 186         LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
 187
 188         OBD_ALLOC(page_pool, sizeof(page) * max_pages);
 189         if (page_pool != NULL) {
 190                 page_pool[0] = page0;
 191         } else {
 192                 page_pool = &page0;
 193                 max_pages = 1;
 194         }
 195         for (npages = 1; npages < max_pages; npages++) {
 196                 page = page_cache_alloc_cold(inode->i_mapping);
 197                 if (!page)
 198                         break;
 199                 page_pool[npages] = page;
 200         }
 201
 202         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 203                                      LUSTRE_OPC_ANY, NULL);
 204         op_data->op_npages = npages;
 205         op_data->op_offset = hash;
 206         rc = md_readpage(exp, op_data, page_pool, &request);
 207         ll_finish_md_op_data(op_data);
 208         if (rc == 0) {
 209                 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 210                 /* Checked by mdc_readpage() */
 211                 LASSERT(body != NULL);
 212
 213                 if (body->valid & OBD_MD_FLSIZE)
 214                         cl_isize_write(inode, body->size);
 215
 216                 nrdpgs = (request->rq_bulk->bd_nob_transferred+CFS_PAGE_SIZE-1)
 217                          >> CFS_PAGE_SHIFT;
 218                 SetPageUptodate(page0);
 219         }
 220         unlock_page(page0);
 221         ptlrpc_req_finished(request);
 222
 223         CDEBUG(D_VFSTRACE, "read %d/%d pages\n", nrdpgs, npages);
 224
 225         ll_pagevec_init(&lru_pvec, 0);
 226         for (i = 1; i < npages; i++) {
 227                 unsigned long offset;
 228                 int ret;
 229
 230                 page = page_pool[i];
 231
 232                 if (rc < 0 || i >= nrdpgs) {
 233                         page_cache_release(page);
 234                         continue;
 235                 }
 236
 237                 SetPageUptodate(page);
 238
 239                 dp = cfs_kmap(page);
 240                 hash = le64_to_cpu(dp->ldp_hash_start);
 241                 cfs_kunmap(page);
 242
 243                 offset = hash_x_index(hash, hash64);
 244
 245                 prefetchw(&page->flags);
 246                 ret = ll_add_to_page_cache_lru(page, inode->i_mapping, offset,
 247                                                GFP_KERNEL);
 248                 if (ret == 0) {
 249                         unlock_page(page);
 250                         page_cache_get(page);
 251                         if (ll_pagevec_add(&lru_pvec, page) == 0)
 252                                 ll_pagevec_lru_add_file(&lru_pvec);
 253                 } else {
 254                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
 255                                " %d\n", offset, ret);
 256                 }
 257                 page_cache_release(page);
 258         }
 259         ll_pagevec_lru_add_file(&lru_pvec);
 260
 261         if (page_pool != &page0)
 262                 OBD_FREE(page_pool, sizeof(struct page *) * max_pages);
 263         EXIT;
 264         return rc;
 265 }
 266
 267 #ifndef MS_HAS_NEW_AOPS
 268 struct address_space_operations ll_dir_aops = {
 269         .readpage  = ll_dir_readpage,
 270 };
 271 #else
 272 struct address_space_operations_ext ll_dir_aops = {
 273         .orig_aops.readpage  = ll_dir_readpage,
 274 };
 275 #endif
 276
 277 static void ll_check_page(struct inode *dir, struct page *page)
 278 {
 279         /* XXX: check page format later */
 280         SetPageChecked(page);
 281 }
 282
 283 void ll_release_page(struct page *page, int remove)
 284 {
 285         kunmap(page);
 286         if (remove) {
 287                 lock_page(page);
 288                 if (likely(page->mapping != NULL))
 289                         truncate_complete_page(page->mapping, page);
 290                 unlock_page(page);
 291         }
 292         page_cache_release(page);
 293 }
 294
 295 /*
 296  * Find, kmap and return page that contains given hash.
 297  */
 298 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
 299                                        __u64 *start, __u64 *end)
 300 {
 301         int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
 302         struct address_space *mapping = dir->i_mapping;
 303         /*
 304          * Complement of hash is used as an index so that
 305          * radix_tree_gang_lookup() can be used to find a page with starting
 306          * hash _smaller_ than one we are looking for.
 307          */
 308         unsigned long offset = hash_x_index(*hash, hash64);
 309         struct page *page;
 310         int found;
 311
 312         TREE_READ_LOCK_IRQ(mapping);
 313         found = radix_tree_gang_lookup(&mapping->page_tree,
 314                                        (void **)&page, offset, 1);
 315         if (found > 0) {
 316                 struct lu_dirpage *dp;
 317
 318                 page_cache_get(page);
 319                 TREE_READ_UNLOCK_IRQ(mapping);
 320                 /*
 321                  * In contrast to find_lock_page() we are sure that directory
 322                  * page cannot be truncated (while DLM lock is held) and,
 323                  * hence, can avoid restart.
 324                  *
 325                  * In fact, page cannot be locked here at all, because
 326                  * ll_dir_readpage() does synchronous io.
 327                  */
 328                 wait_on_page(page);
 329                 if (PageUptodate(page)) {
 330                         dp = cfs_kmap(page);
 331                         if (BITS_PER_LONG == 32 && hash64) {
 332                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
 333                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
 334                                 *hash  = *hash >> 32;
 335                         } else {
 336                                 *start = le64_to_cpu(dp->ldp_hash_start);
 337                                 *end   = le64_to_cpu(dp->ldp_hash_end);
 338                         }
 339                         LASSERTF(*start <= *hash, "start = "LPX64",end = "
 340                                  LPX64",hash = "LPX64"\n", *start, *end, *hash);
 341                         CDEBUG(D_VFSTRACE, "page %lu [%llu %llu], hash "LPU64"\n",
 342                                offset, *start, *end, *hash);
 343                         if (*hash > *end) {
 344                                 ll_release_page(page, 0);
 345                                 page = NULL;
 346                         } else if (*end != *start && *hash == *end) {
 347                                 /*
 348                                  * upon hash collision, remove this page,
 349                                  * otherwise put page reference, and
 350                                  * ll_get_dir_page() will issue RPC to fetch
 351                                  * the page we want.
 352                                  */
 353                                 ll_release_page(page,
 354                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
 355                                 page = NULL;
 356                         }
 357                 } else {
 358                         page_cache_release(page);
 359                         page = ERR_PTR(-EIO);
 360                 }
 361
 362         } else {
 363                 TREE_READ_UNLOCK_IRQ(mapping);
 364                 page = NULL;
 365         }
 366         return page;
 367 }
 368
 369 struct page *ll_get_dir_page(struct file *filp, struct inode *dir, __u64 hash,
 370                              struct ll_dir_chain *chain)
 371 {
 372         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 373         struct address_space *mapping = dir->i_mapping;
 374         struct lustre_handle lockh;
 375         struct lu_dirpage *dp;
 376         struct page *page;
 377         ldlm_mode_t mode;
 378         int rc;
 379         __u64 start = 0;
 380         __u64 end = 0;
 381         __u64 lhash = hash;
 382         struct ll_inode_info *lli = ll_i2info(dir);
 383         int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
 384
 385         mode = LCK_PR;
 386         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 387                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
 388         if (!rc) {
 389                 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
 390                        ll_md_blocking_ast, ldlm_completion_ast,
 391                        NULL, NULL, dir };
 392                 struct lookup_intent it = { .it_op = IT_READDIR };
 393                 struct ptlrpc_request *request;
 394                 struct md_op_data *op_data;
 395
 396                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
 397                                              LUSTRE_OPC_ANY, NULL);
 398                 if (IS_ERR(op_data))
 399                         return (void *)op_data;
 400
 401                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
 402                                 op_data, &lockh, NULL, 0, NULL, 0);
 403
 404                 ll_finish_md_op_data(op_data);
 405
 406                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 407                 if (request)
 408                         ptlrpc_req_finished(request);
 409                 if (rc < 0) {
 410                         CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
 411                                PFID(ll_inode2fid(dir)), hash, rc);
 412                         return ERR_PTR(rc);
 413                 }
 414         } else {
 415                 /* for cross-ref object, l_ast_data of the lock may not be set,
 416                  * we reset it here */
 417                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
 418                                  dir, NULL);
 419         }
 420         ldlm_lock_dump_handle(D_OTHER, &lockh);
 421
 422         cfs_mutex_lock(&lli->lli_readdir_mutex);
 423         page = ll_dir_page_locate(dir, &lhash, &start, &end);
 424         if (IS_ERR(page)) {
 425                 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
 426                        PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
 427                 GOTO(out_unlock, page);
 428         } else if (page != NULL) {
 429                 /*
 430                  * XXX nikita: not entirely correct handling of a corner case:
 431                  * suppose hash chain of entries with hash value HASH crosses
 432                  * border between pages P0 and P1. First both P0 and P1 are
 433                  * cached, seekdir() is called for some entry from the P0 part
 434                  * of the chain. Later P0 goes out of cache. telldir(HASH)
 435                  * happens and finds P1, as it starts with matching hash
 436                  * value. Remaining entries from P0 part of the chain are
 437                  * skipped. (Is that really a bug?)
 438                  *
 439                  * Possible solutions: 0. don't cache P1 is such case, handle
 440                  * it as an "overflow" page. 1. invalidate all pages at
 441                  * once. 2. use HASH|1 as an index for P1.
 442                  */
 443                 GOTO(hash_collision, page);
 444         }
 445
 446         page = read_cache_page(mapping, hash_x_index(hash, hash64),
 447                                (filler_t*)mapping->a_ops->readpage, filp);
 448         if (IS_ERR(page)) {
 449                 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
 450                        PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
 451                 GOTO(out_unlock, page);
 452         }
 453
 454         wait_on_page(page);
 455         (void)kmap(page);
 456         if (!PageUptodate(page)) {
 457                 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
 458                        PFID(ll_inode2fid(dir)), hash, -5);
 459                 goto fail;
 460         }
 461         if (!PageChecked(page))
 462                 ll_check_page(dir, page);
 463         if (PageError(page)) {
 464                 CERROR("page error: "DFID" at "LPU64": rc %d\n",
 465                        PFID(ll_inode2fid(dir)), hash, -5);
 466                 goto fail;
 467         }
 468 hash_collision:
 469         dp = page_address(page);
 470         if (BITS_PER_LONG == 32 && hash64) {
 471                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
 472                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
 473                 lhash = hash >> 32;
 474         } else {
 475                 start = le64_to_cpu(dp->ldp_hash_start);
 476                 end   = le64_to_cpu(dp->ldp_hash_end);
 477                 lhash = hash;
 478         }
 479         if (end == start) {
 480                 LASSERT(start == lhash);
 481                 CWARN("Page-wide hash collision: "LPU64"\n", end);
 482                 if (BITS_PER_LONG == 32 && hash64)
 483                         CWARN("Real page-wide hash collision at ["LPU64" "LPU64
 484                               "] with hash "LPU64"\n",
 485                               le64_to_cpu(dp->ldp_hash_start),
 486                               le64_to_cpu(dp->ldp_hash_end), hash);
 487                 /*
 488                  * Fetch whole overflow chain...
 489                  *
 490                  * XXX not yet.
 491                  */
 492                 goto fail;
 493         }
 494 out_unlock:
 495         cfs_mutex_unlock(&lli->lli_readdir_mutex);
 496         ldlm_lock_decref(&lockh, mode);
 497         return page;
 498
 499 fail:
 500         ll_release_page(page, 1);
 501         page = ERR_PTR(-EIO);
 502         goto out_unlock;
 503 }
 504
 505 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 506 {
 507         struct inode         *inode      = filp->f_dentry->d_inode;
 508         struct ll_inode_info *info       = ll_i2info(inode);
 509         struct ll_sb_info    *sbi        = ll_i2sbi(inode);
 510         struct ll_file_data  *fd         = LUSTRE_FPRIVATE(filp);
 511         __u64                 pos        = fd->fd_dir.lfd_pos;
 512         int                   api32      = ll_need_32bit_api(sbi);
 513         int                   hash64     = sbi->ll_flags & LL_SBI_64BIT_HASH;
 514         struct page          *page;
 515         struct ll_dir_chain   chain;
 516         int                   done;
 517         int                   rc;
 518         ENTRY;
 519
 520         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu 32bit_api %d\n",
 521                inode->i_ino, inode->i_generation, inode,
 522                (unsigned long)pos, i_size_read(inode), api32);
 523
 524         if (pos == MDS_DIR_END_OFF)
 525                 /*
 526                  * end-of-file.
 527                  */
 528                 GOTO(out, rc = 0);
 529
 530         rc    = 0;
 531         done  = 0;
 532         ll_dir_chain_init(&chain);
 533
 534         fd->fd_dir.lfd_next = pos;
 535         page = ll_get_dir_page(filp, inode, pos, &chain);
 536
 537         while (rc == 0 && !done) {
 538                 struct lu_dirpage *dp;
 539                 struct lu_dirent  *ent;
 540
 541                 if (!IS_ERR(page)) {
 542                         /*
 543                          * If page is empty (end of directory is reached),
 544                          * use this value.
 545                          */
 546                         __u64 hash = MDS_DIR_END_OFF;
 547                         __u64 next;
 548
 549                         dp = page_address(page);
 550                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 551                              ent = lu_dirent_next(ent)) {
 552                                 __u16          type;
 553                                 int            namelen;
 554                                 struct lu_fid  fid;
 555                                 __u64          lhash;
 556                                 __u64          ino;
 557
 558                                 /*
 559                                  * XXX: implement correct swabbing here.
 560                                  */
 561
 562                                 hash = le64_to_cpu(ent->lde_hash);
 563                                 if (hash < pos)
 564                                         /*
 565                                          * Skip until we find target hash
 566                                          * value.
 567                                          */
 568                                         continue;
 569
 570                                 namelen = le16_to_cpu(ent->lde_namelen);
 571                                 if (namelen == 0)
 572                                         /*
 573                                          * Skip dummy record.
 574                                          */
 575                                         continue;
 576
 577                                 if (api32 && hash64)
 578                                         lhash = hash >> 32;
 579                                 else
 580                                         lhash = hash;
 581                                 fid_le_to_cpu(&fid, &ent->lde_fid);
 582                                 ino = cl_fid_build_ino(&fid, api32);
 583                                 type = ll_dirent_type_get(ent);
 584                                 /* For 'll_nfs_get_name_filldir()', it will try
 585                                  * to access the 'ent' through its 'lde_name',
 586                                  * so the parameter 'name' for 'filldir()' must
 587                                  * be part of the 'ent'. */
 588                                 done = filldir(cookie, ent->lde_name, namelen,
 589                                                lhash, ino, type);
 590                         }
 591                         next = le64_to_cpu(dp->ldp_hash_end);
 592                         if (!done) {
 593                                 pos = next;
 594                                 if (pos == MDS_DIR_END_OFF) {
 595                                         /*
 596                                          * End of directory reached.
 597                                          */
 598                                         done = 1;
 599                                         ll_release_page(page, 0);
 600                                 } else if (1 /* chain is exhausted*/) {
 601                                         /*
 602                                          * Normal case: continue to the next
 603                                          * page.
 604                                          */
 605                                         ll_release_page(page,
 606                                             le32_to_cpu(dp->ldp_flags) &
 607                                                         LDF_COLLIDE);
 608                                         fd->fd_dir.lfd_next = pos;
 609                                         page = ll_get_dir_page(filp, inode, pos,
 610                                                                &chain);
 611                                 } else {
 612                                         /*
 613                                          * go into overflow page.
 614                                          */
 615                                         LASSERT(le32_to_cpu(dp->ldp_flags) &
 616                                                 LDF_COLLIDE);
 617                                         ll_release_page(page, 1);
 618                                 }
 619                         } else {
 620                                 pos = hash;
 621                                 ll_release_page(page, 0);
 622                         }
 623                 } else {
 624                         rc = PTR_ERR(page);
 625                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 626                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 627                 }
 628         }
 629
 630         fd->fd_dir.lfd_pos = pos;
 631         if (pos == MDS_DIR_END_OFF) {
 632                 if (api32)
 633                         filp->f_pos = LL_DIR_END_OFF_32BIT;
 634                 else
 635                         filp->f_pos = LL_DIR_END_OFF;
 636         } else {
 637                 if (api32 && hash64)
 638                         filp->f_pos = pos >> 32;
 639                 else
 640                         filp->f_pos = pos;
 641         }
 642         filp->f_version = inode->i_version;
 643         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 644
 645         ll_dir_chain_fini(&chain);
 646
 647 out:
 648         if (!rc)
 649                 ll_stats_ops_tally(sbi, LPROC_LL_READDIR, 1);
 650
 651         RETURN(rc);
 652 }
 653
 654 int ll_send_mgc_param(struct obd_export *mgc, char *string)
 655 {
 656         struct mgs_send_param *msp;
 657         int rc = 0;
 658
 659         OBD_ALLOC_PTR(msp);
 660         if (!msp)
 661                 return -ENOMEM;
 662
 663         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
 664         rc = obd_set_info_async(mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
 665                                 sizeof(struct mgs_send_param), msp, NULL);
 666         if (rc)
 667                 CERROR("Failed to set parameter: %d\n", rc);
 668         OBD_FREE_PTR(msp);
 669
 670         return rc;
 671 }
 672
 673 char *ll_get_fsname(struct inode *inode)
 674 {
 675         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 676         char *ptr, *fsname;
 677         int len;
 678
 679         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
 680         len = strlen(lsi->lsi_lmd->lmd_profile);
 681         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
 682         if (ptr && (strcmp(ptr, "-client") == 0))
 683                 len -= 7;
 684         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
 685         fsname[len] = '\0';
 686
 687         return fsname;
 688 }
 689
 690 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
 691                      int set_default)
 692 {
 693         struct ll_sb_info *sbi = ll_i2sbi(inode);
 694         struct md_op_data *op_data;
 695         struct ptlrpc_request *req = NULL;
 696         int rc = 0;
 697         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 698         struct obd_device *mgc = lsi->lsi_mgc;
 699         char *fsname = NULL, *param = NULL;
 700         int lum_size;
 701
 702         if (lump != NULL) {
 703                 /*
 704                  * This is coming from userspace, so should be in
 705                  * local endian.  But the MDS would like it in little
 706                  * endian, so we swab it before we send it.
 707                  */
 708                 switch (lump->lmm_magic) {
 709                 case LOV_USER_MAGIC_V1: {
 710                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
 711                                 lustre_swab_lov_user_md_v1(lump);
 712                         lum_size = sizeof(struct lov_user_md_v1);
 713                         break;
 714                         }
 715                 case LOV_USER_MAGIC_V3: {
 716                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
 717                                 lustre_swab_lov_user_md_v3(
 718                                         (struct lov_user_md_v3 *)lump);
 719                         lum_size = sizeof(struct lov_user_md_v3);
 720                         break;
 721                         }
 722                 default: {
 723                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 724                                         " %#08x != %#08x nor %#08x\n",
 725                                         lump->lmm_magic, LOV_USER_MAGIC_V1,
 726                                         LOV_USER_MAGIC_V3);
 727                         RETURN(-EINVAL);
 728                         }
 729                }
 730         } else {
 731                 lum_size = sizeof(struct lov_user_md_v1);
 732         }
 733
 734         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 735                                      LUSTRE_OPC_ANY, NULL);
 736         if (IS_ERR(op_data))
 737                 RETURN(PTR_ERR(op_data));
 738
 739         /* swabbing is done in lov_setstripe() on server side */
 740         rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
 741                         NULL, 0, &req, NULL);
 742         ll_finish_md_op_data(op_data);
 743         ptlrpc_req_finished(req);
 744         if (rc) {
 745                 if (rc != -EPERM && rc != -EACCES)
 746                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 747         }
 748
 749         /* In the following we use the fact that LOV_USER_MAGIC_V1 and
 750          LOV_USER_MAGIC_V3 have the same initial fields so we do not
 751          need the make the distiction between the 2 versions */
 752         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
 753                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
 754
 755                 /* Get fsname and assume devname to be -MDT0000. */
 756                 fsname = ll_get_fsname(inode);
 757                 /* Set root stripesize */
 758                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
 759                         lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
 760                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 761                 if (rc)
 762                         goto end;
 763
 764                 /* Set root stripecount */
 765                 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
 766                         lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
 767                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 768                 if (rc)
 769                         goto end;
 770
 771                 /* Set root stripeoffset */
 772                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
 773                         lump ? le16_to_cpu(lump->lmm_stripe_offset) :
 774                         (typeof(lump->lmm_stripe_offset))(-1));
 775                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 776                 if (rc)
 777                         goto end;
 778 end:
 779                 if (fsname)
 780                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
 781                 if (param)
 782                         OBD_FREE(param, MGS_PARAM_MAXLEN);
 783         }
 784         return rc;
 785 }
 786
 787 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 788                      int *lmm_size, struct ptlrpc_request **request)
 789 {
 790         struct ll_sb_info *sbi = ll_i2sbi(inode);
 791         struct mdt_body   *body;
 792         struct lov_mds_md *lmm = NULL;
 793         struct ptlrpc_request *req = NULL;
 794         int rc, lmmsize;
 795         struct md_op_data *op_data;
 796
 797         rc = ll_get_max_mdsize(sbi, &lmmsize);
 798         if (rc)
 799                 RETURN(rc);
 800
 801         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
 802                                      0, lmmsize, LUSTRE_OPC_ANY,
 803                                      NULL);
 804         if (IS_ERR(op_data))
 805                 RETURN(PTR_ERR(op_data));
 806
 807         op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
 808         rc = md_getattr(sbi->ll_md_exp, op_data, &req);
 809         ll_finish_md_op_data(op_data);
 810         if (rc < 0) {
 811                 CDEBUG(D_INFO, "md_getattr failed on inode "
 812                        "%lu/%u: rc %d\n", inode->i_ino,
 813                        inode->i_generation, rc);
 814                 GOTO(out, rc);
 815         }
 816
 817         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 818         LASSERT(body != NULL);
 819
 820         lmmsize = body->eadatasize;
 821
 822         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 823             lmmsize == 0) {
 824                 GOTO(out, rc = -ENODATA);
 825         }
 826
 827         lmm = req_capsule_server_sized_get(&req->rq_pill,
 828                                            &RMF_MDT_MD, lmmsize);
 829         LASSERT(lmm != NULL);
 830
 831         /*
 832          * This is coming from the MDS, so is probably in
 833          * little endian.  We convert it to host endian before
 834          * passing it to userspace.
 835          */
 836         /* We don't swab objects for directories */
 837         switch (le32_to_cpu(lmm->lmm_magic)) {
 838         case LOV_MAGIC_V1:
 839                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
 840                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
 841                 break;
 842         case LOV_MAGIC_V3:
 843                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
 844                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
 845                 break;
 846         default:
 847                 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
 848                 rc = -EPROTO;
 849         }
 850 out:
 851         *lmmp = lmm;
 852         *lmm_size = lmmsize;
 853         *request = req;
 854         return rc;
 855 }
 856
 857 /*
 858  *  Get MDT index for the inode.
 859  */
 860 int ll_get_mdt_idx(struct inode *inode)
 861 {
 862         struct ll_sb_info *sbi = ll_i2sbi(inode);
 863         struct md_op_data *op_data;
 864         int rc, mdtidx;
 865         ENTRY;
 866
 867         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
 868                                      0, LUSTRE_OPC_ANY, NULL);
 869         if (IS_ERR(op_data))
 870                 RETURN(PTR_ERR(op_data));
 871
 872         op_data->op_valid |= OBD_MD_MDTIDX;
 873         rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
 874         mdtidx = op_data->op_mds;
 875         ll_finish_md_op_data(op_data);
 876         if (rc < 0) {
 877                 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 878                 RETURN(rc);
 879         }
 880         return mdtidx;
 881 }
 882
 883 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
 884 {
 885         void *ptr;
 886         int rc;
 887
 888         OBD_ALLOC(ptr, len);
 889         if (ptr == NULL)
 890                 return -ENOMEM;
 891         if (cfs_copy_from_user(ptr, data, len)) {
 892                 OBD_FREE(ptr, len);
 893                 return -EFAULT;
 894         }
 895         rc = obd_iocontrol(cmd, exp, len, data, NULL);
 896         OBD_FREE(ptr, len);
 897         return rc;
 898 }
 899
 900 static int quotactl_ioctl(struct ll_sb_info *sbi, struct if_quotactl *qctl)
 901 {
 902         int cmd = qctl->qc_cmd;
 903         int type = qctl->qc_type;
 904         int id = qctl->qc_id;
 905         int valid = qctl->qc_valid;
 906         int rc = 0;
 907         ENTRY;
 908
 909         switch (cmd) {
 910         case LUSTRE_Q_INVALIDATE:
 911         case LUSTRE_Q_FINVALIDATE:
 912         case Q_QUOTAON:
 913         case Q_QUOTAOFF:
 914         case Q_SETQUOTA:
 915         case Q_SETINFO:
 916                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
 917                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
 918                         RETURN(-EPERM);
 919                 break;
 920         case Q_GETQUOTA:
 921                 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
 922                      (type == GRPQUOTA && !in_egroup_p(id))) &&
 923                     (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
 924                      sbi->ll_flags & LL_SBI_RMT_CLIENT))
 925                         RETURN(-EPERM);
 926                 break;
 927         case Q_GETINFO:
 928                 break;
 929         default:
 930                 CERROR("unsupported quotactl op: %#x\n", cmd);
 931                 RETURN(-ENOTTY);
 932         }
 933
 934         if (valid != QC_GENERAL) {
 935                 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
 936                         RETURN(-EOPNOTSUPP);
 937
 938                 if (cmd == Q_GETINFO)
 939                         qctl->qc_cmd = Q_GETOINFO;
 940                 else if (cmd == Q_GETQUOTA)
 941                         qctl->qc_cmd = Q_GETOQUOTA;
 942                 else
 943                         RETURN(-EINVAL);
 944
 945                 switch (valid) {
 946                 case QC_MDTIDX:
 947                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
 948                                            sizeof(*qctl), qctl, NULL);
 949                         break;
 950                 case QC_OSTIDX:
 951                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_dt_exp,
 952                                            sizeof(*qctl), qctl, NULL);
 953                         break;
 954                 case QC_UUID:
 955                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
 956                                            sizeof(*qctl), qctl, NULL);
 957                         if (rc == -EAGAIN)
 958                                 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
 959                                                    sbi->ll_dt_exp,
 960                                                    sizeof(*qctl), qctl, NULL);
 961                         break;
 962                 default:
 963                         rc = -EINVAL;
 964                         break;
 965                 }
 966
 967                 if (rc)
 968                         RETURN(rc);
 969
 970                 qctl->qc_cmd = cmd;
 971         } else {
 972                 struct obd_quotactl *oqctl;
 973
 974                 OBD_ALLOC_PTR(oqctl);
 975                 if (oqctl == NULL)
 976                         RETURN(-ENOMEM);
 977
 978                 QCTL_COPY(oqctl, qctl);
 979                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
 980                 if (rc) {
 981                         if (rc != -EALREADY && cmd == Q_QUOTAON) {
 982                                 oqctl->qc_cmd = Q_QUOTAOFF;
 983                                 obd_quotactl(sbi->ll_md_exp, oqctl);
 984                         }
 985                         OBD_FREE_PTR(oqctl);
 986                         RETURN(rc);
 987                 }
 988                 /* If QIF_SPACE is not set, client should collect the
 989                  * space usage from OSSs by itself */
 990                 if (cmd == Q_GETQUOTA &&
 991                     !(oqctl->qc_dqblk.dqb_valid & QIF_SPACE) &&
 992                     !oqctl->qc_dqblk.dqb_curspace) {
 993                         struct obd_quotactl *oqctl_tmp;
 994
 995                         OBD_ALLOC_PTR(oqctl_tmp);
 996                         if (oqctl_tmp == NULL)
 997                                 GOTO(out, rc = -ENOMEM);
 998
 999                         oqctl_tmp->qc_cmd = Q_GETOQUOTA;
1000                         oqctl_tmp->qc_id = oqctl->qc_id;
1001                         oqctl_tmp->qc_type = oqctl->qc_type;
1002
1003                         /* collect space usage from OSTs */
1004                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1005                         rc = obd_quotactl(sbi->ll_dt_exp, oqctl_tmp);
1006                         if (!rc || rc == -EREMOTEIO) {
1007                                 oqctl->qc_dqblk.dqb_curspace =
1008                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1009                                 oqctl->qc_dqblk.dqb_valid |= QIF_SPACE;
1010                         }
1011
1012                         /* collect space & inode usage from MDTs */
1013                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1014                         oqctl_tmp->qc_dqblk.dqb_curinodes = 0;
1015                         rc = obd_quotactl(sbi->ll_md_exp, oqctl_tmp);
1016                         if (!rc || rc == -EREMOTEIO) {
1017                                 oqctl->qc_dqblk.dqb_curspace +=
1018                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1019                                 oqctl->qc_dqblk.dqb_curinodes =
1020                                         oqctl_tmp->qc_dqblk.dqb_curinodes;
1021                                 oqctl->qc_dqblk.dqb_valid |= QIF_INODES;
1022                         } else {
1023                                 oqctl->qc_dqblk.dqb_valid &= ~QIF_SPACE;
1024                         }
1025
1026                         OBD_FREE_PTR(oqctl_tmp);
1027                 }
1028 out:
1029                 QCTL_COPY(qctl, oqctl);
1030                 OBD_FREE_PTR(oqctl);
1031         }
1032
1033         RETURN(rc);
1034 }
1035
1036 #ifdef HAVE_UNLOCKED_IOCTL
1037 static long ll_dir_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1038 #else
1039 static int  ll_dir_ioctl(struct inode *unuse, struct file *file,
1040                         unsigned int cmd, unsigned long arg)
1041 #endif
1042 {
1043         struct inode *inode = file->f_dentry->d_inode;
1044         struct ll_sb_info *sbi = ll_i2sbi(inode);
1045         struct obd_ioctl_data *data;
1046         int rc = 0;
1047         ENTRY;
1048
1049         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
1050                inode->i_ino, inode->i_generation, inode, cmd);
1051
1052         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
1053         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
1054                 return -ENOTTY;
1055
1056         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
1057         switch(cmd) {
1058         case FSFILT_IOC_GETFLAGS:
1059         case FSFILT_IOC_SETFLAGS:
1060                 RETURN(ll_iocontrol(inode, file, cmd, arg));
1061         case FSFILT_IOC_GETVERSION_OLD:
1062         case FSFILT_IOC_GETVERSION:
1063                 RETURN(put_user(inode->i_generation, (int *)arg));
1064         /* We need to special case any other ioctls we want to handle,
1065          * to send them to the MDS/OST as appropriate and to properly
1066          * network encode the arg field.
1067         case FSFILT_IOC_SETVERSION_OLD:
1068         case FSFILT_IOC_SETVERSION:
1069         */
1070         case LL_IOC_GET_MDTIDX: {
1071                 int mdtidx;
1072
1073                 mdtidx = ll_get_mdt_idx(inode);
1074                 if (mdtidx < 0)
1075                         RETURN(mdtidx);
1076
1077                 if (put_user((int)mdtidx, (int*)arg))
1078                         RETURN(-EFAULT);
1079
1080                 return 0;
1081         }
1082         case IOC_MDC_LOOKUP: {
1083                 struct ptlrpc_request *request = NULL;
1084                 int namelen, len = 0;
1085                 char *buf = NULL;
1086                 char *filename;
1087                 struct md_op_data *op_data;
1088
1089                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1090                 if (rc)
1091                         RETURN(rc);
1092                 data = (void *)buf;
1093
1094                 filename = data->ioc_inlbuf1;
1095                 namelen = strlen(filename);
1096
1097                 if (namelen < 1) {
1098                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1099                         GOTO(out_free, rc = -EINVAL);
1100                 }
1101
1102                 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1103                                              0, LUSTRE_OPC_ANY, NULL);
1104                 if (IS_ERR(op_data))
1105                         GOTO(out_free, rc = PTR_ERR(op_data));
1106
1107                 op_data->op_valid = OBD_MD_FLID;
1108                 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1109                 ll_finish_md_op_data(op_data);
1110                 if (rc < 0) {
1111                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1112                         GOTO(out_free, rc);
1113                 }
1114                 ptlrpc_req_finished(request);
1115                 EXIT;
1116 out_free:
1117                 obd_ioctl_freedata(buf, len);
1118                 return rc;
1119         }
1120         case LL_IOC_LOV_SETSTRIPE: {
1121                 struct lov_user_md_v3 lumv3;
1122                 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1123                 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1124                 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1125
1126                 int set_default = 0;
1127
1128                 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1129                 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1130                         sizeof(lumv3p->lmm_objects[0]));
1131                 /* first try with v1 which is smaller than v3 */
1132                 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1133                         RETURN(-EFAULT);
1134
1135                 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1136                         if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1137                                 RETURN(-EFAULT);
1138                 }
1139
1140                 if (inode->i_sb->s_root == file->f_dentry)
1141                         set_default = 1;
1142
1143                 /* in v1 and v3 cases lumv1 points to data */
1144                 rc = ll_dir_setstripe(inode, lumv1, set_default);
1145
1146                 RETURN(rc);
1147         }
1148         case LL_IOC_OBD_STATFS:
1149                 RETURN(ll_obd_statfs(inode, (void *)arg));
1150         case LL_IOC_LOV_GETSTRIPE:
1151         case LL_IOC_MDC_GETINFO:
1152         case IOC_MDC_GETFILEINFO:
1153         case IOC_MDC_GETFILESTRIPE: {
1154                 struct ptlrpc_request *request = NULL;
1155                 struct lov_user_md *lump;
1156                 struct lov_mds_md *lmm = NULL;
1157                 struct mdt_body *body;
1158                 char *filename = NULL;
1159                 int lmmsize;
1160
1161                 if (cmd == IOC_MDC_GETFILEINFO ||
1162                     cmd == IOC_MDC_GETFILESTRIPE) {
1163                         filename = getname((const char *)arg);
1164                         if (IS_ERR(filename))
1165                                 RETURN(PTR_ERR(filename));
1166
1167                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1168                                                       &lmmsize, &request);
1169                 } else {
1170                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1171                 }
1172
1173                 if (request) {
1174                         body = req_capsule_server_get(&request->rq_pill,
1175                                                       &RMF_MDT_BODY);
1176                         LASSERT(body != NULL);
1177                 } else {
1178                         GOTO(out_req, rc);
1179                 }
1180
1181                 if (rc < 0) {
1182                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1183                                                cmd == LL_IOC_MDC_GETINFO))
1184                                 GOTO(skip_lmm, rc = 0);
1185                         else
1186                                 GOTO(out_req, rc);
1187                 }
1188
1189                 if (cmd == IOC_MDC_GETFILESTRIPE ||
1190                     cmd == LL_IOC_LOV_GETSTRIPE) {
1191                         lump = (struct lov_user_md *)arg;
1192                 } else {
1193                         struct lov_user_mds_data *lmdp;
1194                         lmdp = (struct lov_user_mds_data *)arg;
1195                         lump = &lmdp->lmd_lmm;
1196                 }
1197                 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
1198                         if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
1199                                 GOTO(out_req, rc = -EFAULT);
1200                         rc = -EOVERFLOW;
1201                 }
1202         skip_lmm:
1203                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1204                         struct lov_user_mds_data *lmdp;
1205                         lstat_t st = { 0 };
1206
1207                         st.st_dev     = inode->i_sb->s_dev;
1208                         st.st_mode    = body->mode;
1209                         st.st_nlink   = body->nlink;
1210                         st.st_uid     = body->uid;
1211                         st.st_gid     = body->gid;
1212                         st.st_rdev    = body->rdev;
1213                         st.st_size    = body->size;
1214                         st.st_blksize = CFS_PAGE_SIZE;
1215                         st.st_blocks  = body->blocks;
1216                         st.st_atime   = body->atime;
1217                         st.st_mtime   = body->mtime;
1218                         st.st_ctime   = body->ctime;
1219                         st.st_ino     = inode->i_ino;
1220
1221                         lmdp = (struct lov_user_mds_data *)arg;
1222                         if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1223                                 GOTO(out_req, rc = -EFAULT);
1224                 }
1225
1226                 EXIT;
1227         out_req:
1228                 ptlrpc_req_finished(request);
1229                 if (filename)
1230                         putname(filename);
1231                 return rc;
1232         }
1233         case IOC_LOV_GETINFO: {
1234                 struct lov_user_mds_data *lumd;
1235                 struct lov_stripe_md *lsm;
1236                 struct lov_user_md *lum;
1237                 struct lov_mds_md *lmm;
1238                 int lmmsize;
1239                 lstat_t st;
1240
1241                 lumd = (struct lov_user_mds_data *)arg;
1242                 lum = &lumd->lmd_lmm;
1243
1244                 rc = ll_get_max_mdsize(sbi, &lmmsize);
1245                 if (rc)
1246                         RETURN(rc);
1247
1248                 OBD_ALLOC_LARGE(lmm, lmmsize);
1249                 if (cfs_copy_from_user(lmm, lum, lmmsize))
1250                         GOTO(free_lmm, rc = -EFAULT);
1251
1252                 switch (lmm->lmm_magic) {
1253                 case LOV_USER_MAGIC_V1:
1254                         if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
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_v1 *)lmm)->lmm_objects,
1259                                 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1260                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1261                         break;
1262                 case LOV_USER_MAGIC_V3:
1263                         if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1264                                 break;
1265                         /* swab objects first so that stripes num will be sane */
1266                         lustre_swab_lov_user_md_objects(
1267                                 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1268                                 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1269                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1270                         break;
1271                 default:
1272                         GOTO(free_lmm, rc = -EINVAL);
1273                 }
1274
1275                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1276                 if (rc < 0)
1277                         GOTO(free_lmm, rc = -ENOMEM);
1278
1279                 /* Perform glimpse_size operation. */
1280                 memset(&st, 0, sizeof(st));
1281
1282                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1283                 if (rc)
1284                         GOTO(free_lsm, rc);
1285
1286                 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1287                         GOTO(free_lsm, rc = -EFAULT);
1288
1289                 EXIT;
1290         free_lsm:
1291                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1292         free_lmm:
1293                 OBD_FREE_LARGE(lmm, lmmsize);
1294                 return rc;
1295         }
1296         case OBD_IOC_LLOG_CATINFO: {
1297                 struct ptlrpc_request *req = NULL;
1298                 char                  *buf = NULL;
1299                 char                  *str;
1300                 int                    len = 0;
1301
1302                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1303                 if (rc)
1304                         RETURN(rc);
1305                 data = (void *)buf;
1306
1307                 if (!data->ioc_inlbuf1) {
1308                         obd_ioctl_freedata(buf, len);
1309                         RETURN(-EINVAL);
1310                 }
1311
1312                 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
1313                                            &RQF_LLOG_CATINFO);
1314                 if (req == NULL)
1315                         GOTO(out_catinfo, rc = -ENOMEM);
1316
1317                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
1318                                      data->ioc_inllen1);
1319                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
1320                                      data->ioc_inllen2);
1321
1322                 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
1323                 if (rc) {
1324                         ptlrpc_request_free(req);
1325                         GOTO(out_catinfo, rc);
1326                 }
1327
1328                 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
1329                 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
1330                 if (data->ioc_inllen2) {
1331                         str = req_capsule_client_get(&req->rq_pill,
1332                                                      &RMF_STRING);
1333                         memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
1334                 }
1335
1336                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
1337                                      data->ioc_plen1);
1338                 ptlrpc_request_set_replen(req);
1339
1340                 rc = ptlrpc_queue_wait(req);
1341                 if (!rc) {
1342                         str = req_capsule_server_get(&req->rq_pill,
1343                                                      &RMF_STRING);
1344                         if (cfs_copy_to_user(data->ioc_pbuf1, str,
1345                                              data->ioc_plen1))
1346                                 rc = -EFAULT;
1347                 }
1348                 ptlrpc_req_finished(req);
1349         out_catinfo:
1350                 obd_ioctl_freedata(buf, len);
1351                 RETURN(rc);
1352         }
1353         case OBD_IOC_QUOTACHECK: {
1354                 struct obd_quotactl *oqctl;
1355                 int error = 0;
1356
1357                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1358                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1359                         RETURN(-EPERM);
1360
1361                 OBD_ALLOC_PTR(oqctl);
1362                 if (!oqctl)
1363                         RETURN(-ENOMEM);
1364                 oqctl->qc_type = arg;
1365                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1366                 if (rc < 0) {
1367                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1368                         error = rc;
1369                 }
1370
1371                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1372                 if (rc < 0)
1373                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1374
1375                 OBD_FREE_PTR(oqctl);
1376                 return error ?: rc;
1377         }
1378         case OBD_IOC_POLL_QUOTACHECK: {
1379                 struct if_quotacheck *check;
1380
1381                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1382                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1383                         RETURN(-EPERM);
1384
1385                 OBD_ALLOC_PTR(check);
1386                 if (!check)
1387                         RETURN(-ENOMEM);
1388
1389                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1390                                    NULL);
1391                 if (rc) {
1392                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1393                         if (cfs_copy_to_user((void *)arg, check,
1394                                              sizeof(*check)))
1395                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1396                         GOTO(out_poll, rc);
1397                 }
1398
1399                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1400                                    NULL);
1401                 if (rc) {
1402                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1403                         if (cfs_copy_to_user((void *)arg, check,
1404                                              sizeof(*check)))
1405                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1406                         GOTO(out_poll, rc);
1407                 }
1408         out_poll:
1409                 OBD_FREE_PTR(check);
1410                 RETURN(rc);
1411         }
1412 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
1413         case LL_IOC_QUOTACTL_18: {
1414                 /* copy the old 1.x quota struct for internal use, then copy
1415                  * back into old format struct.  For 1.8 compatibility. */
1416                 struct if_quotactl_18 *qctl_18;
1417                 struct if_quotactl *qctl_20;
1418
1419                 OBD_ALLOC_PTR(qctl_18);
1420                 if (!qctl_18)
1421                         RETURN(-ENOMEM);
1422
1423                 OBD_ALLOC_PTR(qctl_20);
1424                 if (!qctl_20)
1425                         GOTO(out_quotactl_18, rc = -ENOMEM);
1426
1427                 if (cfs_copy_from_user(qctl_18, (void *)arg, sizeof(*qctl_18)))
1428                         GOTO(out_quotactl_20, rc = -ENOMEM);
1429
1430                 QCTL_COPY(qctl_20, qctl_18);
1431                 qctl_20->qc_idx = 0;
1432
1433                 /* XXX: dqb_valid was borrowed as a flag to mark that
1434                  *      only mds quota is wanted */
1435                 if (qctl_18->qc_cmd == Q_GETQUOTA &&
1436                     qctl_18->qc_dqblk.dqb_valid) {
1437                         qctl_20->qc_valid = QC_MDTIDX;
1438                         qctl_20->qc_dqblk.dqb_valid = 0;
1439                 } else if (qctl_18->obd_uuid.uuid[0] != '\0') {
1440                         qctl_20->qc_valid = QC_UUID;
1441                         qctl_20->obd_uuid = qctl_18->obd_uuid;
1442                 } else {
1443                         qctl_20->qc_valid = QC_GENERAL;
1444                 }
1445
1446                 rc = quotactl_ioctl(sbi, qctl_20);
1447
1448                 if (rc == 0) {
1449                         QCTL_COPY(qctl_18, qctl_20);
1450                         qctl_18->obd_uuid = qctl_20->obd_uuid;
1451
1452                         if (cfs_copy_to_user((void *)arg, qctl_18,
1453                                              sizeof(*qctl_18)))
1454                                 rc = -EFAULT;
1455                 }
1456
1457         out_quotactl_20:
1458                 OBD_FREE_PTR(qctl_20);
1459         out_quotactl_18:
1460                 OBD_FREE_PTR(qctl_18);
1461                 RETURN(rc);
1462         }
1463 #else
1464 #warning "remove old LL_IOC_QUOTACTL_18 compatibility code"
1465 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0) */
1466         case LL_IOC_QUOTACTL: {
1467                 struct if_quotactl *qctl;
1468
1469                 OBD_ALLOC_PTR(qctl);
1470                 if (!qctl)
1471                         RETURN(-ENOMEM);
1472
1473                 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1474                         GOTO(out_quotactl, rc = -EFAULT);
1475
1476                 rc = quotactl_ioctl(sbi, qctl);
1477
1478                 if (rc == 0 && cfs_copy_to_user((void *)arg,qctl,sizeof(*qctl)))
1479                         rc = -EFAULT;
1480
1481         out_quotactl:
1482                 OBD_FREE_PTR(qctl);
1483                 RETURN(rc);
1484         }
1485         case OBD_IOC_GETDTNAME:
1486         case OBD_IOC_GETMDNAME:
1487                 RETURN(ll_get_obd_name(inode, cmd, arg));
1488         case LL_IOC_FLUSHCTX:
1489                 RETURN(ll_flush_ctx(inode));
1490 #ifdef CONFIG_FS_POSIX_ACL
1491         case LL_IOC_RMTACL: {
1492             if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1493                 inode == inode->i_sb->s_root->d_inode) {
1494                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1495
1496                 LASSERT(fd != NULL);
1497                 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1498                 if (!rc)
1499                         fd->fd_flags |= LL_FILE_RMTACL;
1500                 RETURN(rc);
1501             } else
1502                 RETURN(0);
1503         }
1504 #endif
1505         case LL_IOC_GETOBDCOUNT: {
1506                 int count, vallen;
1507                 struct obd_export *exp;
1508
1509                 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1510                         RETURN(-EFAULT);
1511
1512                 /* get ost count when count is zero, get mdt count otherwise */
1513                 exp = count ? sbi->ll_md_exp : sbi->ll_dt_exp;
1514                 vallen = sizeof(count);
1515                 rc = obd_get_info(exp, sizeof(KEY_TGT_COUNT), KEY_TGT_COUNT,
1516                                   &vallen, &count, NULL);
1517                 if (rc) {
1518                         CERROR("get target count failed: %d\n", rc);
1519                         RETURN(rc);
1520                 }
1521
1522                 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1523                         RETURN(-EFAULT);
1524
1525                 RETURN(0);
1526         }
1527         case LL_IOC_PATH2FID:
1528                 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1529                                      sizeof(struct lu_fid)))
1530                         RETURN(-EFAULT);
1531                 RETURN(0);
1532         case LL_IOC_GET_CONNECT_FLAGS: {
1533                 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1534         }
1535         case OBD_IOC_CHANGELOG_SEND:
1536         case OBD_IOC_CHANGELOG_CLEAR:
1537                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1538                                     sizeof(struct ioc_changelog));
1539                 RETURN(rc);
1540         case OBD_IOC_FID2PATH:
1541                 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1542         case LL_IOC_HSM_CT_START:
1543                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1544                                     sizeof(struct lustre_kernelcomm));
1545                 RETURN(rc);
1546
1547         default:
1548                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1549         }
1550 }
1551
1552 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1553 {
1554         struct inode *inode = file->f_mapping->host;
1555         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1556         struct ll_sb_info *sbi = ll_i2sbi(inode);
1557         int api32 = ll_need_32bit_api(sbi);
1558         loff_t ret = -EINVAL;
1559         ENTRY;
1560
1561         cfs_mutex_lock(&inode->i_mutex);
1562         switch (origin) {
1563                 case SEEK_SET:
1564                         break;
1565                 case SEEK_CUR:
1566                         offset += file->f_pos;
1567                         break;
1568                 case SEEK_END:
1569                         if (offset > 0)
1570                                 GOTO(out, ret);
1571                         if (api32)
1572                                 offset += LL_DIR_END_OFF_32BIT;
1573                         else
1574                                 offset += LL_DIR_END_OFF;
1575                         break;
1576                 default:
1577                         GOTO(out, ret);
1578         }
1579
1580         if (offset >= 0 &&
1581             ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1582              (!api32 && offset <= LL_DIR_END_OFF))) {
1583                 if (offset != file->f_pos) {
1584                         if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1585                             (!api32 && offset == LL_DIR_END_OFF))
1586                                 fd->fd_dir.lfd_pos = MDS_DIR_END_OFF;
1587                         else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1588                                 fd->fd_dir.lfd_pos = offset << 32;
1589                         else
1590                                 fd->fd_dir.lfd_pos = offset;
1591                         file->f_pos = offset;
1592                         file->f_version = 0;
1593                 }
1594                 ret = offset;
1595         }
1596         GOTO(out, ret);
1597
1598 out:
1599         cfs_mutex_unlock(&inode->i_mutex);
1600         return ret;
1601 }
1602
1603 int ll_dir_open(struct inode *inode, struct file *file)
1604 {
1605         ENTRY;
1606         RETURN(ll_file_open(inode, file));
1607 }
1608
1609 int ll_dir_release(struct inode *inode, struct file *file)
1610 {
1611         ENTRY;
1612         RETURN(ll_file_release(inode, file));
1613 }
1614
1615 struct file_operations ll_dir_operations = {
1616         .llseek   = ll_dir_seek,
1617         .open     = ll_dir_open,
1618         .release  = ll_dir_release,
1619         .read     = generic_read_dir,
1620         .readdir  = ll_readdir,
1621 #ifdef HAVE_UNLOCKED_IOCTL
1622         .unlocked_ioctl   = ll_dir_ioctl,
1623 #else
1624         .ioctl          = ll_dir_ioctl,
1625 #endif
1626         .fsync    = ll_fsync,
1627 };