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 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/llite/dir.c
  40  *
  41  * Directory code for lustre client.
  42  */
  43
  44 #include <linux/fs.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/mm.h>
  47 #include <linux/version.h>
  48 #include <linux/smp_lock.h>
  49 #include <asm/uaccess.h>
  50 #include <linux/buffer_head.h>   // for wait_on_buffer
  51 #include <linux/pagevec.h>
  52
  53 #define DEBUG_SUBSYSTEM S_LLITE
  54
  55 #include <obd_support.h>
  56 #include <obd_class.h>
  57 #include <lustre_lib.h>
  58 #include <lustre/lustre_idl.h>
  59 #include <lustre_lite.h>
  60 #include <lustre_dlm.h>
  61 #include <lustre_fid.h>
  62 #include "llite_internal.h"
  63
  64 #ifndef HAVE_PAGE_CHECKED
  65 #ifdef HAVE_PG_FS_MISC
  66 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
  67 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
  68 #else
  69 #error PageChecked or PageFsMisc not defined in kernel
  70 #endif
  71 #endif
  72
  73 /*
  74  * (new) readdir implementation overview.
  75  *
  76  * Original lustre readdir implementation cached exact copy of raw directory
  77  * pages on the client. These pages were indexed in client page cache by
  78  * logical offset in the directory file. This design, while very simple and
  79  * intuitive had some inherent problems:
  80  *
  81  *     . it implies that byte offset to the directory entry serves as a
  82  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  83  *     ext3/htree directory entries may move due to splits, and more
  84  *     importantly,
  85  *
  86  *     . it is incompatible with the design of split directories for cmd3,
  87  *     that assumes that names are distributed across nodes based on their
  88  *     hash, and so readdir should be done in hash order.
  89  *
  90  * New readdir implementation does readdir in hash order, and uses hash of a
  91  * file name as a telldir/seekdir cookie. This led to number of complications:
  92  *
  93  *     . hash is not unique, so it cannot be used to index cached directory
  94  *     pages on the client (note, that it requires a whole pageful of hash
  95  *     collided entries to cause two pages to have identical hashes);
  96  *
  97  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  98  *     entry cookie. ext3/htree has the same problem and lustre implementation
  99  *     mimics their solution: seekdir(hash) positions directory at the first
 100  *     entry with the given hash.
 101  *
 102  * Client side.
 103  *
 104  * 0. caching
 105  *
 106  * Client caches directory pages using hash of the first entry as an index. As
 107  * noted above hash is not unique, so this solution doesn't work as is:
 108  * special processing is needed for "page hash chains" (i.e., sequences of
 109  * pages filled with entries all having the same hash value).
 110  *
 111  * First, such chains have to be detected. To this end, server returns to the
 112  * client the hash of the first entry on the page next to one returned. When
 113  * client detects that this hash is the same as hash of the first entry on the
 114  * returned page, page hash collision has to be handled. Pages in the
 115  * hash chain, except first one, are termed "overflow pages".
 116  *
 117  * Solution to index uniqueness problem is to not cache overflow
 118  * pages. Instead, when page hash collision is detected, all overflow pages
 119  * from emerging chain are immediately requested from the server and placed in
 120  * a special data structure (struct ll_dir_chain). This data structure is used
 121  * by ll_readdir() to process entries from overflow pages. When readdir
 122  * invocation finishes, overflow pages are discarded. If page hash collision
 123  * chain weren't completely processed, next call to readdir will again detect
 124  * page hash collision, again read overflow pages in, process next portion of
 125  * entries and again discard the pages. This is not as wasteful as it looks,
 126  * because, given reasonable hash, page hash collisions are extremely rare.
 127  *
 128  * 1. directory positioning
 129  *
 130  * When seekdir(hash) is called, original
 131  *
 132  *
 133  *
 134  *
 135  *
 136  *
 137  *
 138  *
 139  * Server.
 140  *
 141  * identification of and access to overflow pages
 142  *
 143  * page format
 144  *
 145  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
 146  * a header lu_dirpage which describes the start/end hash, and whether this
 147  * page is empty (contains no dir entry) or hash collide with next page.
 148  * After client receives reply, several pages will be integrated into dir page
 149  * in CFS_PAGE_SIZE (if CFS_PAGE_SIZE greater than LU_PAGE_SIZE), and the
 150  * lu_dirpage for this integrated page will be adjusted.
 151  *
 152  */
 153
 154 /* returns the page unlocked, but with a reference */
 155 static int ll_dir_readpage(struct file *file, struct page *page0)
 156 {
 157         struct inode *inode = page0->mapping->host;
 158         int hash64 = ll_i2sbi(inode)->ll_flags & LL_SBI_64BIT_HASH;
 159         struct obd_export *exp = ll_i2sbi(inode)->ll_md_exp;
 160         struct ptlrpc_request *request;
 161         struct mdt_body *body;
 162         struct md_op_data *op_data;
 163         __u64 hash;
 164         struct page **page_pool;
 165         struct page *page;
 166 #ifndef HAVE_ADD_TO_PAGE_CACHE_LRU
 167         struct pagevec lru_pvec;
 168 #endif
 169         struct lu_dirpage *dp;
 170         int max_pages = ll_i2sbi(inode)->ll_md_brw_size >> CFS_PAGE_SHIFT;
 171         int nrdpgs = 0; /* number of pages read actually */
 172         int npages;
 173         int i;
 174         int rc;
 175         ENTRY;
 176
 177         if (file) {
 178                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
 179
 180                 hash = fd->fd_dir.lfd_next;
 181         } else {
 182                 hash = ll_i2info(inode)->lli_sa_pos;
 183         }
 184         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) hash "LPU64"\n",
 185                inode->i_ino, inode->i_generation, inode, hash);
 186
 187         LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
 188
 189         OBD_ALLOC(page_pool, sizeof(page) * max_pages);
 190         if (page_pool != NULL) {
 191                 page_pool[0] = page0;
 192         } else {
 193                 page_pool = &page0;
 194                 max_pages = 1;
 195         }
 196         for (npages = 1; npages < max_pages; npages++) {
 197                 page = page_cache_alloc_cold(inode->i_mapping);
 198                 if (!page)
 199                         break;
 200                 page_pool[npages] = page;
 201         }
 202
 203         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 204                                      LUSTRE_OPC_ANY, NULL);
 205         op_data->op_npages = npages;
 206         op_data->op_offset = hash;
 207         rc = md_readpage(exp, op_data, page_pool, &request);
 208         ll_finish_md_op_data(op_data);
 209         if (rc == 0) {
 210                 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 211                 /* Checked by mdc_readpage() */
 212                 LASSERT(body != NULL);
 213
 214                 if (body->valid & OBD_MD_FLSIZE)
 215                         cl_isize_write(inode, body->size);
 216
 217                 nrdpgs = (request->rq_bulk->bd_nob_transferred+CFS_PAGE_SIZE-1)
 218                          >> CFS_PAGE_SHIFT;
 219                 SetPageUptodate(page0);
 220         }
 221         unlock_page(page0);
 222         ptlrpc_req_finished(request);
 223
 224         CDEBUG(D_VFSTRACE, "read %d/%d pages\n", nrdpgs, npages);
 225
 226         ll_pagevec_init(&lru_pvec, 0);
 227         for (i = 1; i < npages; i++) {
 228                 unsigned long offset;
 229                 int ret;
 230
 231                 page = page_pool[i];
 232
 233                 if (rc < 0 || i >= nrdpgs) {
 234                         page_cache_release(page);
 235                         continue;
 236                 }
 237
 238                 SetPageUptodate(page);
 239
 240                 dp = cfs_kmap(page);
 241                 hash = le64_to_cpu(dp->ldp_hash_start);
 242                 cfs_kunmap(page);
 243
 244                 offset = hash_x_index(hash, hash64);
 245
 246                 prefetchw(&page->flags);
 247                 ret = ll_add_to_page_cache_lru(page, inode->i_mapping, offset,
 248                                                GFP_KERNEL);
 249                 if (ret == 0) {
 250                         unlock_page(page);
 251                         page_cache_get(page);
 252                         if (ll_pagevec_add(&lru_pvec, page) == 0)
 253                                 ll_pagevec_lru_add_file(&lru_pvec);
 254                 } else {
 255                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
 256                                " %d\n", offset, ret);
 257                 }
 258                 page_cache_release(page);
 259         }
 260         ll_pagevec_lru_add_file(&lru_pvec);
 261
 262         if (page_pool != &page0)
 263                 OBD_FREE(page_pool, sizeof(struct page *) * max_pages);
 264         EXIT;
 265         return rc;
 266 }
 267
 268 #ifndef MS_HAS_NEW_AOPS
 269 struct address_space_operations ll_dir_aops = {
 270         .readpage  = ll_dir_readpage,
 271 };
 272 #else
 273 struct address_space_operations_ext ll_dir_aops = {
 274         .orig_aops.readpage  = ll_dir_readpage,
 275 };
 276 #endif
 277
 278 static void ll_check_page(struct inode *dir, struct page *page)
 279 {
 280         /* XXX: check page format later */
 281         SetPageChecked(page);
 282 }
 283
 284 void ll_release_page(struct page *page, int remove)
 285 {
 286         kunmap(page);
 287         if (remove) {
 288                 lock_page(page);
 289                 if (likely(page->mapping != NULL))
 290                         truncate_complete_page(page->mapping, page);
 291                 unlock_page(page);
 292         }
 293         page_cache_release(page);
 294 }
 295
 296 /*
 297  * Find, kmap and return page that contains given hash.
 298  */
 299 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
 300                                        __u64 *start, __u64 *end)
 301 {
 302         int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
 303         struct address_space *mapping = dir->i_mapping;
 304         /*
 305          * Complement of hash is used as an index so that
 306          * radix_tree_gang_lookup() can be used to find a page with starting
 307          * hash _smaller_ than one we are looking for.
 308          */
 309         unsigned long offset = hash_x_index(*hash, hash64);
 310         struct page *page;
 311         int found;
 312
 313         TREE_READ_LOCK_IRQ(mapping);
 314         found = radix_tree_gang_lookup(&mapping->page_tree,
 315                                        (void **)&page, offset, 1);
 316         if (found > 0) {
 317                 struct lu_dirpage *dp;
 318
 319                 page_cache_get(page);
 320                 TREE_READ_UNLOCK_IRQ(mapping);
 321                 /*
 322                  * In contrast to find_lock_page() we are sure that directory
 323                  * page cannot be truncated (while DLM lock is held) and,
 324                  * hence, can avoid restart.
 325                  *
 326                  * In fact, page cannot be locked here at all, because
 327                  * ll_dir_readpage() does synchronous io.
 328                  */
 329                 wait_on_page(page);
 330                 if (PageUptodate(page)) {
 331                         dp = cfs_kmap(page);
 332                         if (BITS_PER_LONG == 32 && hash64) {
 333                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
 334                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
 335                                 *hash  = *hash >> 32;
 336                         } else {
 337                                 *start = le64_to_cpu(dp->ldp_hash_start);
 338                                 *end   = le64_to_cpu(dp->ldp_hash_end);
 339                         }
 340                         LASSERTF(*start <= *hash, "start = "LPX64",end = "
 341                                  LPX64",hash = "LPX64"\n", *start, *end, *hash);
 342                         CDEBUG(D_VFSTRACE, "page %lu [%llu %llu], hash "LPU64"\n",
 343                                offset, *start, *end, *hash);
 344                         if (*hash > *end) {
 345                                 ll_release_page(page, 0);
 346                                 page = NULL;
 347                         } else if (*end != *start && *hash == *end) {
 348                                 /*
 349                                  * upon hash collision, remove this page,
 350                                  * otherwise put page reference, and
 351                                  * ll_get_dir_page() will issue RPC to fetch
 352                                  * the page we want.
 353                                  */
 354                                 ll_release_page(page,
 355                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
 356                                 page = NULL;
 357                         }
 358                 } else {
 359                         page_cache_release(page);
 360                         page = ERR_PTR(-EIO);
 361                 }
 362
 363         } else {
 364                 TREE_READ_UNLOCK_IRQ(mapping);
 365                 page = NULL;
 366         }
 367         return page;
 368 }
 369
 370 struct page *ll_get_dir_page(struct file *filp, struct inode *dir, __u64 hash,
 371                              struct ll_dir_chain *chain)
 372 {
 373         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 374         struct address_space *mapping = dir->i_mapping;
 375         struct lustre_handle lockh;
 376         struct lu_dirpage *dp;
 377         struct page *page;
 378         ldlm_mode_t mode;
 379         int rc;
 380         __u64 start = 0;
 381         __u64 end = 0;
 382         __u64 lhash = hash;
 383         struct ll_inode_info *lli = ll_i2info(dir);
 384         int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
 385
 386         mode = LCK_PR;
 387         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 388                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
 389         if (!rc) {
 390                 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
 391                        ll_md_blocking_ast, ldlm_completion_ast,
 392                        NULL, NULL, dir };
 393                 struct lookup_intent it = { .it_op = IT_READDIR };
 394                 struct ptlrpc_request *request;
 395                 struct md_op_data *op_data;
 396
 397                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
 398                                              LUSTRE_OPC_ANY, NULL);
 399                 if (IS_ERR(op_data))
 400                         return (void *)op_data;
 401
 402                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
 403                                 op_data, &lockh, NULL, 0, NULL, 0);
 404
 405                 ll_finish_md_op_data(op_data);
 406
 407                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 408                 if (request)
 409                         ptlrpc_req_finished(request);
 410                 if (rc < 0) {
 411                         CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
 412                                PFID(ll_inode2fid(dir)), hash, rc);
 413                         return ERR_PTR(rc);
 414                 }
 415         } else {
 416                 /* for cross-ref object, l_ast_data of the lock may not be set,
 417                  * we reset it here */
 418                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
 419                                  dir, NULL);
 420         }
 421         ldlm_lock_dump_handle(D_OTHER, &lockh);
 422
 423         cfs_down(&lli->lli_readdir_sem);
 424         page = ll_dir_page_locate(dir, &lhash, &start, &end);
 425         if (IS_ERR(page)) {
 426                 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
 427                        PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
 428                 GOTO(out_unlock, page);
 429         } else if (page != NULL) {
 430                 /*
 431                  * XXX nikita: not entirely correct handling of a corner case:
 432                  * suppose hash chain of entries with hash value HASH crosses
 433                  * border between pages P0 and P1. First both P0 and P1 are
 434                  * cached, seekdir() is called for some entry from the P0 part
 435                  * of the chain. Later P0 goes out of cache. telldir(HASH)
 436                  * happens and finds P1, as it starts with matching hash
 437                  * value. Remaining entries from P0 part of the chain are
 438                  * skipped. (Is that really a bug?)
 439                  *
 440                  * Possible solutions: 0. don't cache P1 is such case, handle
 441                  * it as an "overflow" page. 1. invalidate all pages at
 442                  * once. 2. use HASH|1 as an index for P1.
 443                  */
 444                 GOTO(hash_collision, page);
 445         }
 446
 447         page = read_cache_page(mapping, hash_x_index(hash, hash64),
 448                                (filler_t*)mapping->a_ops->readpage, filp);
 449         if (IS_ERR(page)) {
 450                 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
 451                        PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
 452                 GOTO(out_unlock, page);
 453         }
 454
 455         wait_on_page(page);
 456         (void)kmap(page);
 457         if (!PageUptodate(page)) {
 458                 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
 459                        PFID(ll_inode2fid(dir)), hash, -5);
 460                 goto fail;
 461         }
 462         if (!PageChecked(page))
 463                 ll_check_page(dir, page);
 464         if (PageError(page)) {
 465                 CERROR("page error: "DFID" at "LPU64": rc %d\n",
 466                        PFID(ll_inode2fid(dir)), hash, -5);
 467                 goto fail;
 468         }
 469 hash_collision:
 470         dp = page_address(page);
 471         if (BITS_PER_LONG == 32 && hash64) {
 472                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
 473                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
 474                 lhash = hash >> 32;
 475         } else {
 476                 start = le64_to_cpu(dp->ldp_hash_start);
 477                 end   = le64_to_cpu(dp->ldp_hash_end);
 478                 lhash = hash;
 479         }
 480         if (end == start) {
 481                 LASSERT(start == lhash);
 482                 CWARN("Page-wide hash collision: "LPU64"\n", end);
 483                 if (BITS_PER_LONG == 32 && hash64)
 484                         CWARN("Real page-wide hash collision at ["LPU64" "LPU64
 485                               "] with hash "LPU64"\n",
 486                               le64_to_cpu(dp->ldp_hash_start),
 487                               le64_to_cpu(dp->ldp_hash_end), hash);
 488                 /*
 489                  * Fetch whole overflow chain...
 490                  *
 491                  * XXX not yet.
 492                  */
 493                 goto fail;
 494         }
 495 out_unlock:
 496         cfs_up(&lli->lli_readdir_sem);
 497         ldlm_lock_decref(&lockh, mode);
 498         return page;
 499
 500 fail:
 501         ll_release_page(page, 1);
 502         page = ERR_PTR(-EIO);
 503         goto out_unlock;
 504 }
 505
 506 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 507 {
 508         struct inode         *inode      = filp->f_dentry->d_inode;
 509         struct ll_inode_info *info       = ll_i2info(inode);
 510         struct ll_sb_info    *sbi        = ll_i2sbi(inode);
 511         struct ll_file_data  *fd         = LUSTRE_FPRIVATE(filp);
 512         __u64                 pos        = fd->fd_dir.lfd_pos;
 513         int                   api32      = ll_need_32bit_api(sbi);
 514         int                   hash64     = sbi->ll_flags & LL_SBI_64BIT_HASH;
 515         struct page          *page;
 516         struct ll_dir_chain   chain;
 517         int                   done;
 518         int                   rc;
 519         ENTRY;
 520
 521         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu 32bit_api %d\n",
 522                inode->i_ino, inode->i_generation, inode,
 523                (unsigned long)pos, i_size_read(inode), api32);
 524
 525         if (pos == MDS_DIR_END_OFF)
 526                 /*
 527                  * end-of-file.
 528                  */
 529                 RETURN(0);
 530
 531         rc    = 0;
 532         done  = 0;
 533         ll_dir_chain_init(&chain);
 534
 535         fd->fd_dir.lfd_next = pos;
 536         page = ll_get_dir_page(filp, inode, pos, &chain);
 537
 538         while (rc == 0 && !done) {
 539                 struct lu_dirpage *dp;
 540                 struct lu_dirent  *ent;
 541
 542                 if (!IS_ERR(page)) {
 543                         /*
 544                          * If page is empty (end of directory is reached),
 545                          * use this value.
 546                          */
 547                         __u64 hash = MDS_DIR_END_OFF;
 548                         __u64 next;
 549
 550                         dp = page_address(page);
 551                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 552                              ent = lu_dirent_next(ent)) {
 553                                 __u16          type;
 554                                 int            namelen;
 555                                 struct lu_fid  fid;
 556                                 __u64          lhash;
 557                                 __u64          ino;
 558
 559                                 /*
 560                                  * XXX: implement correct swabbing here.
 561                                  */
 562
 563                                 hash = le64_to_cpu(ent->lde_hash);
 564                                 if (hash < pos)
 565                                         /*
 566                                          * Skip until we find target hash
 567                                          * value.
 568                                          */
 569                                         continue;
 570
 571                                 namelen = le16_to_cpu(ent->lde_namelen);
 572                                 if (namelen == 0)
 573                                         /*
 574                                          * Skip dummy record.
 575                                          */
 576                                         continue;
 577
 578                                 if (api32 && hash64)
 579                                         lhash = hash >> 32;
 580                                 else
 581                                         lhash = hash;
 582                                 fid_le_to_cpu(&fid, &ent->lde_fid);
 583                                 ino = cl_fid_build_ino(&fid, api32);
 584                                 type = ll_dirent_type_get(ent);
 585                                 /* For 'll_nfs_get_name_filldir()', it will try
 586                                  * to access the 'ent' through its 'lde_name',
 587                                  * so the parameter 'name' for 'filldir()' must
 588                                  * be part of the 'ent'. */
 589                                 done = filldir(cookie, ent->lde_name, namelen,
 590                                                lhash, ino, type);
 591                         }
 592                         next = le64_to_cpu(dp->ldp_hash_end);
 593                         if (!done) {
 594                                 pos = next;
 595                                 if (pos == MDS_DIR_END_OFF) {
 596                                         /*
 597                                          * End of directory reached.
 598                                          */
 599                                         done = 1;
 600                                         ll_release_page(page, 0);
 601                                 } else if (1 /* chain is exhausted*/) {
 602                                         /*
 603                                          * Normal case: continue to the next
 604                                          * page.
 605                                          */
 606                                         ll_release_page(page,
 607                                             le32_to_cpu(dp->ldp_flags) &
 608                                                         LDF_COLLIDE);
 609                                         fd->fd_dir.lfd_next = pos;
 610                                         page = ll_get_dir_page(filp, inode, pos,
 611                                                                &chain);
 612                                 } else {
 613                                         /*
 614                                          * go into overflow page.
 615                                          */
 616                                         LASSERT(le32_to_cpu(dp->ldp_flags) &
 617                                                 LDF_COLLIDE);
 618                                         ll_release_page(page, 1);
 619                                 }
 620                         } else {
 621                                 pos = hash;
 622                                 ll_release_page(page, 0);
 623                         }
 624                 } else {
 625                         rc = PTR_ERR(page);
 626                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 627                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 628                 }
 629         }
 630
 631         fd->fd_dir.lfd_pos = pos;
 632         if (pos == MDS_DIR_END_OFF) {
 633                 if (api32)
 634                         filp->f_pos = LL_DIR_END_OFF_32BIT;
 635                 else
 636                         filp->f_pos = LL_DIR_END_OFF;
 637         } else {
 638                 if (api32 && hash64)
 639                         filp->f_pos = pos >> 32;
 640                 else
 641                         filp->f_pos = pos;
 642         }
 643         filp->f_version = inode->i_version;
 644         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 645
 646         ll_dir_chain_fini(&chain);
 647
 648         RETURN(rc);
 649 }
 650
 651 int ll_send_mgc_param(struct obd_export *mgc, char *string)
 652 {
 653         struct mgs_send_param *msp;
 654         int rc = 0;
 655
 656         OBD_ALLOC_PTR(msp);
 657         if (!msp)
 658                 return -ENOMEM;
 659
 660         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
 661         rc = obd_set_info_async(mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
 662                                 sizeof(struct mgs_send_param), msp, NULL);
 663         if (rc)
 664                 CERROR("Failed to set parameter: %d\n", rc);
 665         OBD_FREE_PTR(msp);
 666
 667         return rc;
 668 }
 669
 670 char *ll_get_fsname(struct inode *inode)
 671 {
 672         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 673         char *ptr, *fsname;
 674         int len;
 675
 676         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
 677         len = strlen(lsi->lsi_lmd->lmd_profile);
 678         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
 679         if (ptr && (strcmp(ptr, "-client") == 0))
 680                 len -= 7;
 681         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
 682         fsname[len] = '\0';
 683
 684         return fsname;
 685 }
 686
 687 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
 688                      int set_default)
 689 {
 690         struct ll_sb_info *sbi = ll_i2sbi(inode);
 691         struct md_op_data *op_data;
 692         struct ptlrpc_request *req = NULL;
 693         int rc = 0;
 694         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 695         struct obd_device *mgc = lsi->lsi_mgc;
 696         char *fsname = NULL, *param = NULL;
 697         int lum_size;
 698
 699         if (lump != NULL) {
 700                 /*
 701                  * This is coming from userspace, so should be in
 702                  * local endian.  But the MDS would like it in little
 703                  * endian, so we swab it before we send it.
 704                  */
 705                 switch (lump->lmm_magic) {
 706                 case LOV_USER_MAGIC_V1: {
 707                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
 708                                 lustre_swab_lov_user_md_v1(lump);
 709                         lum_size = sizeof(struct lov_user_md_v1);
 710                         break;
 711                         }
 712                 case LOV_USER_MAGIC_V3: {
 713                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
 714                                 lustre_swab_lov_user_md_v3(
 715                                         (struct lov_user_md_v3 *)lump);
 716                         lum_size = sizeof(struct lov_user_md_v3);
 717                         break;
 718                         }
 719                 default: {
 720                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 721                                         " %#08x != %#08x nor %#08x\n",
 722                                         lump->lmm_magic, LOV_USER_MAGIC_V1,
 723                                         LOV_USER_MAGIC_V3);
 724                         RETURN(-EINVAL);
 725                         }
 726                }
 727         } else {
 728                 lum_size = sizeof(struct lov_user_md_v1);
 729         }
 730
 731         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 732                                      LUSTRE_OPC_ANY, NULL);
 733         if (IS_ERR(op_data))
 734                 RETURN(PTR_ERR(op_data));
 735
 736         /* swabbing is done in lov_setstripe() on server side */
 737         rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
 738                         NULL, 0, &req, NULL);
 739         ll_finish_md_op_data(op_data);
 740         ptlrpc_req_finished(req);
 741         if (rc) {
 742                 if (rc != -EPERM && rc != -EACCES)
 743                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 744         }
 745
 746         /* In the following we use the fact that LOV_USER_MAGIC_V1 and
 747          LOV_USER_MAGIC_V3 have the same initial fields so we do not
 748          need the make the distiction between the 2 versions */
 749         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
 750                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
 751
 752                 /* Get fsname and assume devname to be -MDT0000. */
 753                 fsname = ll_get_fsname(inode);
 754                 /* Set root stripesize */
 755                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
 756                         lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
 757                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 758                 if (rc)
 759                         goto end;
 760
 761                 /* Set root stripecount */
 762                 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
 763                         lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
 764                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 765                 if (rc)
 766                         goto end;
 767
 768                 /* Set root stripeoffset */
 769                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
 770                         lump ? le16_to_cpu(lump->lmm_stripe_offset) :
 771                         (typeof(lump->lmm_stripe_offset))(-1));
 772                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 773                 if (rc)
 774                         goto end;
 775 end:
 776                 if (fsname)
 777                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
 778                 if (param)
 779                         OBD_FREE(param, MGS_PARAM_MAXLEN);
 780         }
 781         return rc;
 782 }
 783
 784 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 785                      int *lmm_size, struct ptlrpc_request **request)
 786 {
 787         struct ll_sb_info *sbi = ll_i2sbi(inode);
 788         struct mdt_body   *body;
 789         struct lov_mds_md *lmm = NULL;
 790         struct ptlrpc_request *req = NULL;
 791         int rc, lmmsize;
 792         struct md_op_data *op_data;
 793
 794         rc = ll_get_max_mdsize(sbi, &lmmsize);
 795         if (rc)
 796                 RETURN(rc);
 797
 798         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
 799                                      0, lmmsize, LUSTRE_OPC_ANY,
 800                                      NULL);
 801         if (IS_ERR(op_data))
 802                 RETURN(PTR_ERR(op_data));
 803
 804         op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
 805         rc = md_getattr(sbi->ll_md_exp, op_data, &req);
 806         ll_finish_md_op_data(op_data);
 807         if (rc < 0) {
 808                 CDEBUG(D_INFO, "md_getattr failed on inode "
 809                        "%lu/%u: rc %d\n", inode->i_ino,
 810                        inode->i_generation, rc);
 811                 GOTO(out, rc);
 812         }
 813
 814         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 815         LASSERT(body != NULL);
 816
 817         lmmsize = body->eadatasize;
 818
 819         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 820             lmmsize == 0) {
 821                 GOTO(out, rc = -ENODATA);
 822         }
 823
 824         lmm = req_capsule_server_sized_get(&req->rq_pill,
 825                                            &RMF_MDT_MD, lmmsize);
 826         LASSERT(lmm != NULL);
 827
 828         /*
 829          * This is coming from the MDS, so is probably in
 830          * little endian.  We convert it to host endian before
 831          * passing it to userspace.
 832          */
 833         /* We don't swab objects for directories */
 834         switch (le32_to_cpu(lmm->lmm_magic)) {
 835         case LOV_MAGIC_V1:
 836                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
 837                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
 838                 break;
 839         case LOV_MAGIC_V3:
 840                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
 841                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
 842                 break;
 843         default:
 844                 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
 845                 rc = -EPROTO;
 846         }
 847 out:
 848         *lmmp = lmm;
 849         *lmm_size = lmmsize;
 850         *request = req;
 851         return rc;
 852 }
 853
 854 /*
 855  *  Get MDT index for the inode.
 856  */
 857 int ll_get_mdt_idx(struct inode *inode)
 858 {
 859         struct ll_sb_info *sbi = ll_i2sbi(inode);
 860         struct md_op_data *op_data;
 861         int rc, mdtidx;
 862         ENTRY;
 863
 864         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
 865                                      0, LUSTRE_OPC_ANY, NULL);
 866         if (IS_ERR(op_data))
 867                 RETURN(PTR_ERR(op_data));
 868
 869         op_data->op_valid |= OBD_MD_MDTIDX;
 870         rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
 871         mdtidx = op_data->op_mds;
 872         ll_finish_md_op_data(op_data);
 873         if (rc < 0) {
 874                 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 875                 RETURN(rc);
 876         }
 877         return mdtidx;
 878 }
 879
 880 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
 881 {
 882         void *ptr;
 883         int rc;
 884
 885         OBD_ALLOC(ptr, len);
 886         if (ptr == NULL)
 887                 return -ENOMEM;
 888         if (cfs_copy_from_user(ptr, data, len)) {
 889                 OBD_FREE(ptr, len);
 890                 return -EFAULT;
 891         }
 892         rc = obd_iocontrol(cmd, exp, len, data, NULL);
 893         OBD_FREE(ptr, len);
 894         return rc;
 895 }
 896
 897 static int quotactl_ioctl(struct ll_sb_info *sbi, struct if_quotactl *qctl)
 898 {
 899         int cmd = qctl->qc_cmd;
 900         int type = qctl->qc_type;
 901         int id = qctl->qc_id;
 902         int valid = qctl->qc_valid;
 903         int rc = 0;
 904         ENTRY;
 905
 906         switch (cmd) {
 907         case LUSTRE_Q_INVALIDATE:
 908         case LUSTRE_Q_FINVALIDATE:
 909         case Q_QUOTAON:
 910         case Q_QUOTAOFF:
 911         case Q_SETQUOTA:
 912         case Q_SETINFO:
 913                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
 914                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
 915                         RETURN(-EPERM);
 916                 break;
 917         case Q_GETQUOTA:
 918                 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
 919                      (type == GRPQUOTA && !in_egroup_p(id))) &&
 920                     (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
 921                      sbi->ll_flags & LL_SBI_RMT_CLIENT))
 922                         RETURN(-EPERM);
 923                 break;
 924         case Q_GETINFO:
 925                 break;
 926         default:
 927                 CERROR("unsupported quotactl op: %#x\n", cmd);
 928                 RETURN(-ENOTTY);
 929         }
 930
 931         if (valid != QC_GENERAL) {
 932                 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
 933                         RETURN(-EOPNOTSUPP);
 934
 935                 if (cmd == Q_GETINFO)
 936                         qctl->qc_cmd = Q_GETOINFO;
 937                 else if (cmd == Q_GETQUOTA)
 938                         qctl->qc_cmd = Q_GETOQUOTA;
 939                 else
 940                         RETURN(-EINVAL);
 941
 942                 switch (valid) {
 943                 case QC_MDTIDX:
 944                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
 945                                            sizeof(*qctl), qctl, NULL);
 946                         break;
 947                 case QC_OSTIDX:
 948                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_dt_exp,
 949                                            sizeof(*qctl), qctl, NULL);
 950                         break;
 951                 case QC_UUID:
 952                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
 953                                            sizeof(*qctl), qctl, NULL);
 954                         if (rc == -EAGAIN)
 955                                 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
 956                                                    sbi->ll_dt_exp,
 957                                                    sizeof(*qctl), qctl, NULL);
 958                         break;
 959                 default:
 960                         rc = -EINVAL;
 961                         break;
 962                 }
 963
 964                 if (rc)
 965                         RETURN(rc);
 966
 967                 qctl->qc_cmd = cmd;
 968         } else {
 969                 struct obd_quotactl *oqctl;
 970
 971                 OBD_ALLOC_PTR(oqctl);
 972                 if (oqctl == NULL)
 973                         RETURN(-ENOMEM);
 974
 975                 QCTL_COPY(oqctl, qctl);
 976                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
 977                 if (rc) {
 978                         if (rc != -EALREADY && cmd == Q_QUOTAON) {
 979                                 oqctl->qc_cmd = Q_QUOTAOFF;
 980                                 obd_quotactl(sbi->ll_md_exp, oqctl);
 981                         }
 982                         OBD_FREE_PTR(oqctl);
 983                         RETURN(rc);
 984                 }
 985                 /* If QIF_SPACE is not set, client should collect the
 986                  * space usage from OSSs by itself */
 987                 if (cmd == Q_GETQUOTA &&
 988                     !(oqctl->qc_dqblk.dqb_valid & QIF_SPACE) &&
 989                     !oqctl->qc_dqblk.dqb_curspace) {
 990                         struct obd_quotactl *oqctl_tmp;
 991
 992                         OBD_ALLOC_PTR(oqctl_tmp);
 993                         if (oqctl_tmp == NULL)
 994                                 GOTO(out, rc = -ENOMEM);
 995
 996                         oqctl_tmp->qc_cmd = Q_GETOQUOTA;
 997                         oqctl_tmp->qc_id = oqctl->qc_id;
 998                         oqctl_tmp->qc_type = oqctl->qc_type;
 999
1000                         /* collect space usage from OSTs */
1001                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1002                         rc = obd_quotactl(sbi->ll_dt_exp, oqctl_tmp);
1003                         if (!rc || rc == -EREMOTEIO) {
1004                                 oqctl->qc_dqblk.dqb_curspace =
1005                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1006                                 oqctl->qc_dqblk.dqb_valid |= QIF_SPACE;
1007                         }
1008
1009                         /* collect space & inode usage from MDTs */
1010                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1011                         oqctl_tmp->qc_dqblk.dqb_curinodes = 0;
1012                         rc = obd_quotactl(sbi->ll_md_exp, oqctl_tmp);
1013                         if (!rc || rc == -EREMOTEIO) {
1014                                 oqctl->qc_dqblk.dqb_curspace +=
1015                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1016                                 oqctl->qc_dqblk.dqb_curinodes =
1017                                         oqctl_tmp->qc_dqblk.dqb_curinodes;
1018                                 oqctl->qc_dqblk.dqb_valid |= QIF_INODES;
1019                         } else {
1020                                 oqctl->qc_dqblk.dqb_valid &= ~QIF_SPACE;
1021                         }
1022
1023                         OBD_FREE_PTR(oqctl_tmp);
1024                 }
1025 out:
1026                 QCTL_COPY(qctl, oqctl);
1027                 OBD_FREE_PTR(oqctl);
1028         }
1029
1030         RETURN(rc);
1031 }
1032
1033 #ifdef HAVE_UNLOCKED_IOCTL
1034 static long ll_dir_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1035 #else
1036 static int  ll_dir_ioctl(struct inode *unuse, struct file *file,
1037                         unsigned int cmd, unsigned long arg)
1038 #endif
1039 {
1040         struct inode *inode = file->f_dentry->d_inode;
1041         struct ll_sb_info *sbi = ll_i2sbi(inode);
1042         struct obd_ioctl_data *data;
1043         int rc = 0;
1044         ENTRY;
1045
1046         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
1047                inode->i_ino, inode->i_generation, inode, cmd);
1048
1049         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
1050         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
1051                 return -ENOTTY;
1052
1053         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
1054         switch(cmd) {
1055         case FSFILT_IOC_GETFLAGS:
1056         case FSFILT_IOC_SETFLAGS:
1057                 RETURN(ll_iocontrol(inode, file, cmd, arg));
1058         case FSFILT_IOC_GETVERSION_OLD:
1059         case FSFILT_IOC_GETVERSION:
1060                 RETURN(put_user(inode->i_generation, (int *)arg));
1061         /* We need to special case any other ioctls we want to handle,
1062          * to send them to the MDS/OST as appropriate and to properly
1063          * network encode the arg field.
1064         case FSFILT_IOC_SETVERSION_OLD:
1065         case FSFILT_IOC_SETVERSION:
1066         */
1067         case LL_IOC_GET_MDTIDX: {
1068                 int mdtidx;
1069
1070                 mdtidx = ll_get_mdt_idx(inode);
1071                 if (mdtidx < 0)
1072                         RETURN(mdtidx);
1073
1074                 if (put_user((int)mdtidx, (int*)arg))
1075                         RETURN(-EFAULT);
1076
1077                 return 0;
1078         }
1079         case IOC_MDC_LOOKUP: {
1080                 struct ptlrpc_request *request = NULL;
1081                 int namelen, len = 0;
1082                 char *buf = NULL;
1083                 char *filename;
1084                 struct md_op_data *op_data;
1085
1086                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1087                 if (rc)
1088                         RETURN(rc);
1089                 data = (void *)buf;
1090
1091                 filename = data->ioc_inlbuf1;
1092                 namelen = strlen(filename);
1093
1094                 if (namelen < 1) {
1095                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1096                         GOTO(out_free, rc = -EINVAL);
1097                 }
1098
1099                 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1100                                              0, LUSTRE_OPC_ANY, NULL);
1101                 if (IS_ERR(op_data))
1102                         GOTO(out_free, rc = PTR_ERR(op_data));
1103
1104                 op_data->op_valid = OBD_MD_FLID;
1105                 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1106                 ll_finish_md_op_data(op_data);
1107                 if (rc < 0) {
1108                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1109                         GOTO(out_free, rc);
1110                 }
1111                 ptlrpc_req_finished(request);
1112                 EXIT;
1113 out_free:
1114                 obd_ioctl_freedata(buf, len);
1115                 return rc;
1116         }
1117         case LL_IOC_LOV_SETSTRIPE: {
1118                 struct lov_user_md_v3 lumv3;
1119                 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1120                 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1121                 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1122
1123                 int set_default = 0;
1124
1125                 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1126                 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1127                         sizeof(lumv3p->lmm_objects[0]));
1128                 /* first try with v1 which is smaller than v3 */
1129                 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1130                         RETURN(-EFAULT);
1131
1132                 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1133                         if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1134                                 RETURN(-EFAULT);
1135                 }
1136
1137                 if (inode->i_sb->s_root == file->f_dentry)
1138                         set_default = 1;
1139
1140                 /* in v1 and v3 cases lumv1 points to data */
1141                 rc = ll_dir_setstripe(inode, lumv1, set_default);
1142
1143                 RETURN(rc);
1144         }
1145         case LL_IOC_OBD_STATFS:
1146                 RETURN(ll_obd_statfs(inode, (void *)arg));
1147         case LL_IOC_LOV_GETSTRIPE:
1148         case LL_IOC_MDC_GETINFO:
1149         case IOC_MDC_GETFILEINFO:
1150         case IOC_MDC_GETFILESTRIPE: {
1151                 struct ptlrpc_request *request = NULL;
1152                 struct lov_user_md *lump;
1153                 struct lov_mds_md *lmm = NULL;
1154                 struct mdt_body *body;
1155                 char *filename = NULL;
1156                 int lmmsize;
1157
1158                 if (cmd == IOC_MDC_GETFILEINFO ||
1159                     cmd == IOC_MDC_GETFILESTRIPE) {
1160                         filename = getname((const char *)arg);
1161                         if (IS_ERR(filename))
1162                                 RETURN(PTR_ERR(filename));
1163
1164                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1165                                                       &lmmsize, &request);
1166                 } else {
1167                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1168                 }
1169
1170                 if (request) {
1171                         body = req_capsule_server_get(&request->rq_pill,
1172                                                       &RMF_MDT_BODY);
1173                         LASSERT(body != NULL);
1174                 } else {
1175                         GOTO(out_req, rc);
1176                 }
1177
1178                 if (rc < 0) {
1179                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1180                                                cmd == LL_IOC_MDC_GETINFO))
1181                                 GOTO(skip_lmm, rc = 0);
1182                         else
1183                                 GOTO(out_req, rc);
1184                 }
1185
1186                 if (cmd == IOC_MDC_GETFILESTRIPE ||
1187                     cmd == LL_IOC_LOV_GETSTRIPE) {
1188                         lump = (struct lov_user_md *)arg;
1189                 } else {
1190                         struct lov_user_mds_data *lmdp;
1191                         lmdp = (struct lov_user_mds_data *)arg;
1192                         lump = &lmdp->lmd_lmm;
1193                 }
1194                 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
1195                         if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
1196                                 GOTO(out_req, rc = -EFAULT);
1197                         rc = -EOVERFLOW;
1198                 }
1199         skip_lmm:
1200                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1201                         struct lov_user_mds_data *lmdp;
1202                         lstat_t st = { 0 };
1203
1204                         st.st_dev     = inode->i_sb->s_dev;
1205                         st.st_mode    = body->mode;
1206                         st.st_nlink   = body->nlink;
1207                         st.st_uid     = body->uid;
1208                         st.st_gid     = body->gid;
1209                         st.st_rdev    = body->rdev;
1210                         st.st_size    = body->size;
1211                         st.st_blksize = CFS_PAGE_SIZE;
1212                         st.st_blocks  = body->blocks;
1213                         st.st_atime   = body->atime;
1214                         st.st_mtime   = body->mtime;
1215                         st.st_ctime   = body->ctime;
1216                         st.st_ino     = inode->i_ino;
1217
1218                         lmdp = (struct lov_user_mds_data *)arg;
1219                         if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1220                                 GOTO(out_req, rc = -EFAULT);
1221                 }
1222
1223                 EXIT;
1224         out_req:
1225                 ptlrpc_req_finished(request);
1226                 if (filename)
1227                         putname(filename);
1228                 return rc;
1229         }
1230         case IOC_LOV_GETINFO: {
1231                 struct lov_user_mds_data *lumd;
1232                 struct lov_stripe_md *lsm;
1233                 struct lov_user_md *lum;
1234                 struct lov_mds_md *lmm;
1235                 int lmmsize;
1236                 lstat_t st;
1237
1238                 lumd = (struct lov_user_mds_data *)arg;
1239                 lum = &lumd->lmd_lmm;
1240
1241                 rc = ll_get_max_mdsize(sbi, &lmmsize);
1242                 if (rc)
1243                         RETURN(rc);
1244
1245                 OBD_ALLOC_LARGE(lmm, lmmsize);
1246                 if (cfs_copy_from_user(lmm, lum, lmmsize))
1247                         GOTO(free_lmm, rc = -EFAULT);
1248
1249                 switch (lmm->lmm_magic) {
1250                 case LOV_USER_MAGIC_V1:
1251                         if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1252                                 break;
1253                         /* swab objects first so that stripes num will be sane */
1254                         lustre_swab_lov_user_md_objects(
1255                                 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1256                                 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1257                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1258                         break;
1259                 case LOV_USER_MAGIC_V3:
1260                         if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1261                                 break;
1262                         /* swab objects first so that stripes num will be sane */
1263                         lustre_swab_lov_user_md_objects(
1264                                 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1265                                 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1266                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1267                         break;
1268                 default:
1269                         GOTO(free_lmm, rc = -EINVAL);
1270                 }
1271
1272                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1273                 if (rc < 0)
1274                         GOTO(free_lmm, rc = -ENOMEM);
1275
1276                 /* Perform glimpse_size operation. */
1277                 memset(&st, 0, sizeof(st));
1278
1279                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1280                 if (rc)
1281                         GOTO(free_lsm, rc);
1282
1283                 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1284                         GOTO(free_lsm, rc = -EFAULT);
1285
1286                 EXIT;
1287         free_lsm:
1288                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1289         free_lmm:
1290                 OBD_FREE_LARGE(lmm, lmmsize);
1291                 return rc;
1292         }
1293         case OBD_IOC_LLOG_CATINFO: {
1294                 struct ptlrpc_request *req = NULL;
1295                 char                  *buf = NULL;
1296                 char                  *str;
1297                 int                    len = 0;
1298
1299                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1300                 if (rc)
1301                         RETURN(rc);
1302                 data = (void *)buf;
1303
1304                 if (!data->ioc_inlbuf1) {
1305                         obd_ioctl_freedata(buf, len);
1306                         RETURN(-EINVAL);
1307                 }
1308
1309                 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
1310                                            &RQF_LLOG_CATINFO);
1311                 if (req == NULL)
1312                         GOTO(out_catinfo, rc = -ENOMEM);
1313
1314                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
1315                                      data->ioc_inllen1);
1316                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
1317                                      data->ioc_inllen2);
1318
1319                 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
1320                 if (rc) {
1321                         ptlrpc_request_free(req);
1322                         GOTO(out_catinfo, rc);
1323                 }
1324
1325                 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
1326                 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
1327                 if (data->ioc_inllen2) {
1328                         str = req_capsule_client_get(&req->rq_pill,
1329                                                      &RMF_STRING);
1330                         memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
1331                 }
1332
1333                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
1334                                      data->ioc_plen1);
1335                 ptlrpc_request_set_replen(req);
1336
1337                 rc = ptlrpc_queue_wait(req);
1338                 if (!rc) {
1339                         str = req_capsule_server_get(&req->rq_pill,
1340                                                      &RMF_STRING);
1341                         if (cfs_copy_to_user(data->ioc_pbuf1, str,
1342                                              data->ioc_plen1))
1343                                 rc = -EFAULT;
1344                 }
1345                 ptlrpc_req_finished(req);
1346         out_catinfo:
1347                 obd_ioctl_freedata(buf, len);
1348                 RETURN(rc);
1349         }
1350         case OBD_IOC_QUOTACHECK: {
1351                 struct obd_quotactl *oqctl;
1352                 int error = 0;
1353
1354                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1355                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1356                         RETURN(-EPERM);
1357
1358                 OBD_ALLOC_PTR(oqctl);
1359                 if (!oqctl)
1360                         RETURN(-ENOMEM);
1361                 oqctl->qc_type = arg;
1362                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1363                 if (rc < 0) {
1364                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1365                         error = rc;
1366                 }
1367
1368                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1369                 if (rc < 0)
1370                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1371
1372                 OBD_FREE_PTR(oqctl);
1373                 return error ?: rc;
1374         }
1375         case OBD_IOC_POLL_QUOTACHECK: {
1376                 struct if_quotacheck *check;
1377
1378                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1379                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1380                         RETURN(-EPERM);
1381
1382                 OBD_ALLOC_PTR(check);
1383                 if (!check)
1384                         RETURN(-ENOMEM);
1385
1386                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1387                                    NULL);
1388                 if (rc) {
1389                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1390                         if (cfs_copy_to_user((void *)arg, check,
1391                                              sizeof(*check)))
1392                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1393                         GOTO(out_poll, rc);
1394                 }
1395
1396                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1397                                    NULL);
1398                 if (rc) {
1399                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1400                         if (cfs_copy_to_user((void *)arg, check,
1401                                              sizeof(*check)))
1402                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1403                         GOTO(out_poll, rc);
1404                 }
1405         out_poll:
1406                 OBD_FREE_PTR(check);
1407                 RETURN(rc);
1408         }
1409 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
1410         case LL_IOC_QUOTACTL_18: {
1411                 /* copy the old 1.x quota struct for internal use, then copy
1412                  * back into old format struct.  For 1.8 compatibility. */
1413                 struct if_quotactl_18 *qctl_18;
1414                 struct if_quotactl *qctl_20;
1415
1416                 OBD_ALLOC_PTR(qctl_18);
1417                 if (!qctl_18)
1418                         RETURN(-ENOMEM);
1419
1420                 OBD_ALLOC_PTR(qctl_20);
1421                 if (!qctl_20)
1422                         GOTO(out_quotactl_18, rc = -ENOMEM);
1423
1424                 if (cfs_copy_from_user(qctl_18, (void *)arg, sizeof(*qctl_18)))
1425                         GOTO(out_quotactl_20, rc = -ENOMEM);
1426
1427                 QCTL_COPY(qctl_20, qctl_18);
1428                 qctl_20->qc_idx = 0;
1429
1430                 /* XXX: dqb_valid was borrowed as a flag to mark that
1431                  *      only mds quota is wanted */
1432                 if (qctl_18->qc_cmd == Q_GETQUOTA &&
1433                     qctl_18->qc_dqblk.dqb_valid) {
1434                         qctl_20->qc_valid = QC_MDTIDX;
1435                         qctl_20->qc_dqblk.dqb_valid = 0;
1436                 } else if (qctl_18->obd_uuid.uuid[0] != '\0') {
1437                         qctl_20->qc_valid = QC_UUID;
1438                         qctl_20->obd_uuid = qctl_18->obd_uuid;
1439                 } else {
1440                         qctl_20->qc_valid = QC_GENERAL;
1441                 }
1442
1443                 rc = quotactl_ioctl(sbi, qctl_20);
1444
1445                 if (rc == 0) {
1446                         QCTL_COPY(qctl_18, qctl_20);
1447                         qctl_18->obd_uuid = qctl_20->obd_uuid;
1448
1449                         if (cfs_copy_to_user((void *)arg, qctl_18,
1450                                              sizeof(*qctl_18)))
1451                                 rc = -EFAULT;
1452                 }
1453
1454         out_quotactl_20:
1455                 OBD_FREE_PTR(qctl_20);
1456         out_quotactl_18:
1457                 OBD_FREE_PTR(qctl_18);
1458                 RETURN(rc);
1459         }
1460 #else
1461 #warning "remove old LL_IOC_QUOTACTL_18 compatibility code"
1462 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0) */
1463         case LL_IOC_QUOTACTL: {
1464                 struct if_quotactl *qctl;
1465
1466                 OBD_ALLOC_PTR(qctl);
1467                 if (!qctl)
1468                         RETURN(-ENOMEM);
1469
1470                 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1471                         GOTO(out_quotactl, rc = -EFAULT);
1472
1473                 rc = quotactl_ioctl(sbi, qctl);
1474
1475                 if (rc == 0 && cfs_copy_to_user((void *)arg,qctl,sizeof(*qctl)))
1476                         rc = -EFAULT;
1477
1478         out_quotactl:
1479                 OBD_FREE_PTR(qctl);
1480                 RETURN(rc);
1481         }
1482         case OBD_IOC_GETDTNAME:
1483         case OBD_IOC_GETMDNAME:
1484                 RETURN(ll_get_obd_name(inode, cmd, arg));
1485         case LL_IOC_FLUSHCTX:
1486                 RETURN(ll_flush_ctx(inode));
1487 #ifdef CONFIG_FS_POSIX_ACL
1488         case LL_IOC_RMTACL: {
1489             if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1490                 inode == inode->i_sb->s_root->d_inode) {
1491                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1492
1493                 LASSERT(fd != NULL);
1494                 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1495                 if (!rc)
1496                         fd->fd_flags |= LL_FILE_RMTACL;
1497                 RETURN(rc);
1498             } else
1499                 RETURN(0);
1500         }
1501 #endif
1502         case LL_IOC_GETOBDCOUNT: {
1503                 int count, vallen;
1504                 struct obd_export *exp;
1505
1506                 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1507                         RETURN(-EFAULT);
1508
1509                 /* get ost count when count is zero, get mdt count otherwise */
1510                 exp = count ? sbi->ll_md_exp : sbi->ll_dt_exp;
1511                 vallen = sizeof(count);
1512                 rc = obd_get_info(exp, sizeof(KEY_TGT_COUNT), KEY_TGT_COUNT,
1513                                   &vallen, &count, NULL);
1514                 if (rc) {
1515                         CERROR("get target count failed: %d\n", rc);
1516                         RETURN(rc);
1517                 }
1518
1519                 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1520                         RETURN(-EFAULT);
1521
1522                 RETURN(0);
1523         }
1524         case LL_IOC_PATH2FID:
1525                 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1526                                      sizeof(struct lu_fid)))
1527                         RETURN(-EFAULT);
1528                 RETURN(0);
1529         case LL_IOC_GET_CONNECT_FLAGS: {
1530                 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1531         }
1532         case OBD_IOC_CHANGELOG_SEND:
1533         case OBD_IOC_CHANGELOG_CLEAR:
1534                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1535                                     sizeof(struct ioc_changelog));
1536                 RETURN(rc);
1537         case OBD_IOC_FID2PATH:
1538                 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1539         case LL_IOC_HSM_CT_START:
1540                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1541                                     sizeof(struct lustre_kernelcomm));
1542                 RETURN(rc);
1543
1544         default:
1545                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1546         }
1547 }
1548
1549 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1550 {
1551         struct inode *inode = file->f_mapping->host;
1552         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1553         struct ll_sb_info *sbi = ll_i2sbi(inode);
1554         int api32 = ll_need_32bit_api(sbi);
1555         loff_t ret = -EINVAL;
1556         ENTRY;
1557
1558         cfs_mutex_lock(&inode->i_mutex);
1559         switch (origin) {
1560                 case SEEK_SET:
1561                         break;
1562                 case SEEK_CUR:
1563                         offset += file->f_pos;
1564                         break;
1565                 case SEEK_END:
1566                         if (offset > 0)
1567                                 GOTO(out, ret);
1568                         if (api32)
1569                                 offset += LL_DIR_END_OFF_32BIT;
1570                         else
1571                                 offset += LL_DIR_END_OFF;
1572                         break;
1573                 default:
1574                         GOTO(out, ret);
1575         }
1576
1577         if (offset >= 0 &&
1578             ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1579              (!api32 && offset <= LL_DIR_END_OFF))) {
1580                 if (offset != file->f_pos) {
1581                         if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1582                             (!api32 && offset == LL_DIR_END_OFF))
1583                                 fd->fd_dir.lfd_pos = MDS_DIR_END_OFF;
1584                         else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1585                                 fd->fd_dir.lfd_pos = offset << 32;
1586                         else
1587                                 fd->fd_dir.lfd_pos = offset;
1588                         file->f_pos = offset;
1589                         file->f_version = 0;
1590                 }
1591                 ret = offset;
1592         }
1593         GOTO(out, ret);
1594
1595 out:
1596         cfs_mutex_unlock(&inode->i_mutex);
1597         return ret;
1598 }
1599
1600 int ll_dir_open(struct inode *inode, struct file *file)
1601 {
1602         ENTRY;
1603         RETURN(ll_file_open(inode, file));
1604 }
1605
1606 int ll_dir_release(struct inode *inode, struct file *file)
1607 {
1608         ENTRY;
1609         RETURN(ll_file_release(inode, file));
1610 }
1611
1612 struct file_operations ll_dir_operations = {
1613         .llseek   = ll_dir_seek,
1614         .open     = ll_dir_open,
1615         .release  = ll_dir_release,
1616         .read     = generic_read_dir,
1617         .readdir  = ll_readdir,
1618 #ifdef HAVE_UNLOCKED_IOCTL
1619         .unlocked_ioctl   = ll_dir_ioctl,
1620 #else
1621         .ioctl          = ll_dir_ioctl,
1622 #endif
1623         .fsync    = ll_fsync,
1624 };