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                 RETURN(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         RETURN(rc);
 648 }
 649
 650 int ll_send_mgc_param(struct obd_export *mgc, char *string)
 651 {
 652         struct mgs_send_param *msp;
 653         int rc = 0;
 654
 655         OBD_ALLOC_PTR(msp);
 656         if (!msp)
 657                 return -ENOMEM;
 658
 659         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
 660         rc = obd_set_info_async(mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
 661                                 sizeof(struct mgs_send_param), msp, NULL);
 662         if (rc)
 663                 CERROR("Failed to set parameter: %d\n", rc);
 664         OBD_FREE_PTR(msp);
 665
 666         return rc;
 667 }
 668
 669 char *ll_get_fsname(struct inode *inode)
 670 {
 671         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 672         char *ptr, *fsname;
 673         int len;
 674
 675         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
 676         len = strlen(lsi->lsi_lmd->lmd_profile);
 677         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
 678         if (ptr && (strcmp(ptr, "-client") == 0))
 679                 len -= 7;
 680         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
 681         fsname[len] = '\0';
 682
 683         return fsname;
 684 }
 685
 686 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
 687                      int set_default)
 688 {
 689         struct ll_sb_info *sbi = ll_i2sbi(inode);
 690         struct md_op_data *op_data;
 691         struct ptlrpc_request *req = NULL;
 692         int rc = 0;
 693         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 694         struct obd_device *mgc = lsi->lsi_mgc;
 695         char *fsname = NULL, *param = NULL;
 696         int lum_size;
 697
 698         if (lump != NULL) {
 699                 /*
 700                  * This is coming from userspace, so should be in
 701                  * local endian.  But the MDS would like it in little
 702                  * endian, so we swab it before we send it.
 703                  */
 704                 switch (lump->lmm_magic) {
 705                 case LOV_USER_MAGIC_V1: {
 706                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
 707                                 lustre_swab_lov_user_md_v1(lump);
 708                         lum_size = sizeof(struct lov_user_md_v1);
 709                         break;
 710                         }
 711                 case LOV_USER_MAGIC_V3: {
 712                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
 713                                 lustre_swab_lov_user_md_v3(
 714                                         (struct lov_user_md_v3 *)lump);
 715                         lum_size = sizeof(struct lov_user_md_v3);
 716                         break;
 717                         }
 718                 default: {
 719                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 720                                         " %#08x != %#08x nor %#08x\n",
 721                                         lump->lmm_magic, LOV_USER_MAGIC_V1,
 722                                         LOV_USER_MAGIC_V3);
 723                         RETURN(-EINVAL);
 724                         }
 725                }
 726         } else {
 727                 lum_size = sizeof(struct lov_user_md_v1);
 728         }
 729
 730         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 731                                      LUSTRE_OPC_ANY, NULL);
 732         if (IS_ERR(op_data))
 733                 RETURN(PTR_ERR(op_data));
 734
 735         /* swabbing is done in lov_setstripe() on server side */
 736         rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
 737                         NULL, 0, &req, NULL);
 738         ll_finish_md_op_data(op_data);
 739         ptlrpc_req_finished(req);
 740         if (rc) {
 741                 if (rc != -EPERM && rc != -EACCES)
 742                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 743         }
 744
 745         /* In the following we use the fact that LOV_USER_MAGIC_V1 and
 746          LOV_USER_MAGIC_V3 have the same initial fields so we do not
 747          need the make the distiction between the 2 versions */
 748         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
 749                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
 750
 751                 /* Get fsname and assume devname to be -MDT0000. */
 752                 fsname = ll_get_fsname(inode);
 753                 /* Set root stripesize */
 754                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
 755                         lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
 756                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 757                 if (rc)
 758                         goto end;
 759
 760                 /* Set root stripecount */
 761                 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
 762                         lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
 763                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 764                 if (rc)
 765                         goto end;
 766
 767                 /* Set root stripeoffset */
 768                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
 769                         lump ? le16_to_cpu(lump->lmm_stripe_offset) :
 770                         (typeof(lump->lmm_stripe_offset))(-1));
 771                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 772                 if (rc)
 773                         goto end;
 774 end:
 775                 if (fsname)
 776                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
 777                 if (param)
 778                         OBD_FREE(param, MGS_PARAM_MAXLEN);
 779         }
 780         return rc;
 781 }
 782
 783 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 784                      int *lmm_size, struct ptlrpc_request **request)
 785 {
 786         struct ll_sb_info *sbi = ll_i2sbi(inode);
 787         struct mdt_body   *body;
 788         struct lov_mds_md *lmm = NULL;
 789         struct ptlrpc_request *req = NULL;
 790         int rc, lmmsize;
 791         struct md_op_data *op_data;
 792
 793         rc = ll_get_max_mdsize(sbi, &lmmsize);
 794         if (rc)
 795                 RETURN(rc);
 796
 797         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
 798                                      0, lmmsize, LUSTRE_OPC_ANY,
 799                                      NULL);
 800         if (IS_ERR(op_data))
 801                 RETURN(PTR_ERR(op_data));
 802
 803         op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
 804         rc = md_getattr(sbi->ll_md_exp, op_data, &req);
 805         ll_finish_md_op_data(op_data);
 806         if (rc < 0) {
 807                 CDEBUG(D_INFO, "md_getattr failed on inode "
 808                        "%lu/%u: rc %d\n", inode->i_ino,
 809                        inode->i_generation, rc);
 810                 GOTO(out, rc);
 811         }
 812
 813         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 814         LASSERT(body != NULL);
 815
 816         lmmsize = body->eadatasize;
 817
 818         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 819             lmmsize == 0) {
 820                 GOTO(out, rc = -ENODATA);
 821         }
 822
 823         lmm = req_capsule_server_sized_get(&req->rq_pill,
 824                                            &RMF_MDT_MD, lmmsize);
 825         LASSERT(lmm != NULL);
 826
 827         /*
 828          * This is coming from the MDS, so is probably in
 829          * little endian.  We convert it to host endian before
 830          * passing it to userspace.
 831          */
 832         /* We don't swab objects for directories */
 833         switch (le32_to_cpu(lmm->lmm_magic)) {
 834         case LOV_MAGIC_V1:
 835                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
 836                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
 837                 break;
 838         case LOV_MAGIC_V3:
 839                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
 840                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
 841                 break;
 842         default:
 843                 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
 844                 rc = -EPROTO;
 845         }
 846 out:
 847         *lmmp = lmm;
 848         *lmm_size = lmmsize;
 849         *request = req;
 850         return rc;
 851 }
 852
 853 /*
 854  *  Get MDT index for the inode.
 855  */
 856 int ll_get_mdt_idx(struct inode *inode)
 857 {
 858         struct ll_sb_info *sbi = ll_i2sbi(inode);
 859         struct md_op_data *op_data;
 860         int rc, mdtidx;
 861         ENTRY;
 862
 863         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
 864                                      0, LUSTRE_OPC_ANY, NULL);
 865         if (IS_ERR(op_data))
 866                 RETURN(PTR_ERR(op_data));
 867
 868         op_data->op_valid |= OBD_MD_MDTIDX;
 869         rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
 870         mdtidx = op_data->op_mds;
 871         ll_finish_md_op_data(op_data);
 872         if (rc < 0) {
 873                 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 874                 RETURN(rc);
 875         }
 876         return mdtidx;
 877 }
 878
 879 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
 880 {
 881         void *ptr;
 882         int rc;
 883
 884         OBD_ALLOC(ptr, len);
 885         if (ptr == NULL)
 886                 return -ENOMEM;
 887         if (cfs_copy_from_user(ptr, data, len)) {
 888                 OBD_FREE(ptr, len);
 889                 return -EFAULT;
 890         }
 891         rc = obd_iocontrol(cmd, exp, len, data, NULL);
 892         OBD_FREE(ptr, len);
 893         return rc;
 894 }
 895
 896 static int quotactl_ioctl(struct ll_sb_info *sbi, struct if_quotactl *qctl)
 897 {
 898         int cmd = qctl->qc_cmd;
 899         int type = qctl->qc_type;
 900         int id = qctl->qc_id;
 901         int valid = qctl->qc_valid;
 902         int rc = 0;
 903         ENTRY;
 904
 905         switch (cmd) {
 906         case LUSTRE_Q_INVALIDATE:
 907         case LUSTRE_Q_FINVALIDATE:
 908         case Q_QUOTAON:
 909         case Q_QUOTAOFF:
 910         case Q_SETQUOTA:
 911         case Q_SETINFO:
 912                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
 913                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
 914                         RETURN(-EPERM);
 915                 break;
 916         case Q_GETQUOTA:
 917                 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
 918                      (type == GRPQUOTA && !in_egroup_p(id))) &&
 919                     (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
 920                      sbi->ll_flags & LL_SBI_RMT_CLIENT))
 921                         RETURN(-EPERM);
 922                 break;
 923         case Q_GETINFO:
 924                 break;
 925         default:
 926                 CERROR("unsupported quotactl op: %#x\n", cmd);
 927                 RETURN(-ENOTTY);
 928         }
 929
 930         if (valid != QC_GENERAL) {
 931                 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
 932                         RETURN(-EOPNOTSUPP);
 933
 934                 if (cmd == Q_GETINFO)
 935                         qctl->qc_cmd = Q_GETOINFO;
 936                 else if (cmd == Q_GETQUOTA)
 937                         qctl->qc_cmd = Q_GETOQUOTA;
 938                 else
 939                         RETURN(-EINVAL);
 940
 941                 switch (valid) {
 942                 case QC_MDTIDX:
 943                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
 944                                            sizeof(*qctl), qctl, NULL);
 945                         break;
 946                 case QC_OSTIDX:
 947                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_dt_exp,
 948                                            sizeof(*qctl), qctl, NULL);
 949                         break;
 950                 case QC_UUID:
 951                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
 952                                            sizeof(*qctl), qctl, NULL);
 953                         if (rc == -EAGAIN)
 954                                 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
 955                                                    sbi->ll_dt_exp,
 956                                                    sizeof(*qctl), qctl, NULL);
 957                         break;
 958                 default:
 959                         rc = -EINVAL;
 960                         break;
 961                 }
 962
 963                 if (rc)
 964                         RETURN(rc);
 965
 966                 qctl->qc_cmd = cmd;
 967         } else {
 968                 struct obd_quotactl *oqctl;
 969
 970                 OBD_ALLOC_PTR(oqctl);
 971                 if (oqctl == NULL)
 972                         RETURN(-ENOMEM);
 973
 974                 QCTL_COPY(oqctl, qctl);
 975                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
 976                 if (rc) {
 977                         if (rc != -EALREADY && cmd == Q_QUOTAON) {
 978                                 oqctl->qc_cmd = Q_QUOTAOFF;
 979                                 obd_quotactl(sbi->ll_md_exp, oqctl);
 980                         }
 981                         OBD_FREE_PTR(oqctl);
 982                         RETURN(rc);
 983                 }
 984                 /* If QIF_SPACE is not set, client should collect the
 985                  * space usage from OSSs by itself */
 986                 if (cmd == Q_GETQUOTA &&
 987                     !(oqctl->qc_dqblk.dqb_valid & QIF_SPACE) &&
 988                     !oqctl->qc_dqblk.dqb_curspace) {
 989                         struct obd_quotactl *oqctl_tmp;
 990
 991                         OBD_ALLOC_PTR(oqctl_tmp);
 992                         if (oqctl_tmp == NULL)
 993                                 GOTO(out, rc = -ENOMEM);
 994
 995                         oqctl_tmp->qc_cmd = Q_GETOQUOTA;
 996                         oqctl_tmp->qc_id = oqctl->qc_id;
 997                         oqctl_tmp->qc_type = oqctl->qc_type;
 998
 999                         /* collect space usage from OSTs */
1000                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1001                         rc = obd_quotactl(sbi->ll_dt_exp, oqctl_tmp);
1002                         if (!rc || rc == -EREMOTEIO) {
1003                                 oqctl->qc_dqblk.dqb_curspace =
1004                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1005                                 oqctl->qc_dqblk.dqb_valid |= QIF_SPACE;
1006                         }
1007
1008                         /* collect space & inode usage from MDTs */
1009                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1010                         oqctl_tmp->qc_dqblk.dqb_curinodes = 0;
1011                         rc = obd_quotactl(sbi->ll_md_exp, oqctl_tmp);
1012                         if (!rc || rc == -EREMOTEIO) {
1013                                 oqctl->qc_dqblk.dqb_curspace +=
1014                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1015                                 oqctl->qc_dqblk.dqb_curinodes =
1016                                         oqctl_tmp->qc_dqblk.dqb_curinodes;
1017                                 oqctl->qc_dqblk.dqb_valid |= QIF_INODES;
1018                         } else {
1019                                 oqctl->qc_dqblk.dqb_valid &= ~QIF_SPACE;
1020                         }
1021
1022                         OBD_FREE_PTR(oqctl_tmp);
1023                 }
1024 out:
1025                 QCTL_COPY(qctl, oqctl);
1026                 OBD_FREE_PTR(oqctl);
1027         }
1028
1029         RETURN(rc);
1030 }
1031
1032 #ifdef HAVE_UNLOCKED_IOCTL
1033 static long ll_dir_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1034 #else
1035 static int  ll_dir_ioctl(struct inode *unuse, struct file *file,
1036                         unsigned int cmd, unsigned long arg)
1037 #endif
1038 {
1039         struct inode *inode = file->f_dentry->d_inode;
1040         struct ll_sb_info *sbi = ll_i2sbi(inode);
1041         struct obd_ioctl_data *data;
1042         int rc = 0;
1043         ENTRY;
1044
1045         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
1046                inode->i_ino, inode->i_generation, inode, cmd);
1047
1048         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
1049         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
1050                 return -ENOTTY;
1051
1052         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
1053         switch(cmd) {
1054         case FSFILT_IOC_GETFLAGS:
1055         case FSFILT_IOC_SETFLAGS:
1056                 RETURN(ll_iocontrol(inode, file, cmd, arg));
1057         case FSFILT_IOC_GETVERSION_OLD:
1058         case FSFILT_IOC_GETVERSION:
1059                 RETURN(put_user(inode->i_generation, (int *)arg));
1060         /* We need to special case any other ioctls we want to handle,
1061          * to send them to the MDS/OST as appropriate and to properly
1062          * network encode the arg field.
1063         case FSFILT_IOC_SETVERSION_OLD:
1064         case FSFILT_IOC_SETVERSION:
1065         */
1066         case LL_IOC_GET_MDTIDX: {
1067                 int mdtidx;
1068
1069                 mdtidx = ll_get_mdt_idx(inode);
1070                 if (mdtidx < 0)
1071                         RETURN(mdtidx);
1072
1073                 if (put_user((int)mdtidx, (int*)arg))
1074                         RETURN(-EFAULT);
1075
1076                 return 0;
1077         }
1078         case IOC_MDC_LOOKUP: {
1079                 struct ptlrpc_request *request = NULL;
1080                 int namelen, len = 0;
1081                 char *buf = NULL;
1082                 char *filename;
1083                 struct md_op_data *op_data;
1084
1085                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1086                 if (rc)
1087                         RETURN(rc);
1088                 data = (void *)buf;
1089
1090                 filename = data->ioc_inlbuf1;
1091                 namelen = strlen(filename);
1092
1093                 if (namelen < 1) {
1094                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1095                         GOTO(out_free, rc = -EINVAL);
1096                 }
1097
1098                 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1099                                              0, LUSTRE_OPC_ANY, NULL);
1100                 if (IS_ERR(op_data))
1101                         GOTO(out_free, rc = PTR_ERR(op_data));
1102
1103                 op_data->op_valid = OBD_MD_FLID;
1104                 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1105                 ll_finish_md_op_data(op_data);
1106                 if (rc < 0) {
1107                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1108                         GOTO(out_free, rc);
1109                 }
1110                 ptlrpc_req_finished(request);
1111                 EXIT;
1112 out_free:
1113                 obd_ioctl_freedata(buf, len);
1114                 return rc;
1115         }
1116         case LL_IOC_LOV_SETSTRIPE: {
1117                 struct lov_user_md_v3 lumv3;
1118                 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1119                 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1120                 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1121
1122                 int set_default = 0;
1123
1124                 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1125                 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1126                         sizeof(lumv3p->lmm_objects[0]));
1127                 /* first try with v1 which is smaller than v3 */
1128                 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1129                         RETURN(-EFAULT);
1130
1131                 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1132                         if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1133                                 RETURN(-EFAULT);
1134                 }
1135
1136                 if (inode->i_sb->s_root == file->f_dentry)
1137                         set_default = 1;
1138
1139                 /* in v1 and v3 cases lumv1 points to data */
1140                 rc = ll_dir_setstripe(inode, lumv1, set_default);
1141
1142                 RETURN(rc);
1143         }
1144         case LL_IOC_OBD_STATFS:
1145                 RETURN(ll_obd_statfs(inode, (void *)arg));
1146         case LL_IOC_LOV_GETSTRIPE:
1147         case LL_IOC_MDC_GETINFO:
1148         case IOC_MDC_GETFILEINFO:
1149         case IOC_MDC_GETFILESTRIPE: {
1150                 struct ptlrpc_request *request = NULL;
1151                 struct lov_user_md *lump;
1152                 struct lov_mds_md *lmm = NULL;
1153                 struct mdt_body *body;
1154                 char *filename = NULL;
1155                 int lmmsize;
1156
1157                 if (cmd == IOC_MDC_GETFILEINFO ||
1158                     cmd == IOC_MDC_GETFILESTRIPE) {
1159                         filename = getname((const char *)arg);
1160                         if (IS_ERR(filename))
1161                                 RETURN(PTR_ERR(filename));
1162
1163                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1164                                                       &lmmsize, &request);
1165                 } else {
1166                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1167                 }
1168
1169                 if (request) {
1170                         body = req_capsule_server_get(&request->rq_pill,
1171                                                       &RMF_MDT_BODY);
1172                         LASSERT(body != NULL);
1173                 } else {
1174                         GOTO(out_req, rc);
1175                 }
1176
1177                 if (rc < 0) {
1178                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1179                                                cmd == LL_IOC_MDC_GETINFO))
1180                                 GOTO(skip_lmm, rc = 0);
1181                         else
1182                                 GOTO(out_req, rc);
1183                 }
1184
1185                 if (cmd == IOC_MDC_GETFILESTRIPE ||
1186                     cmd == LL_IOC_LOV_GETSTRIPE) {
1187                         lump = (struct lov_user_md *)arg;
1188                 } else {
1189                         struct lov_user_mds_data *lmdp;
1190                         lmdp = (struct lov_user_mds_data *)arg;
1191                         lump = &lmdp->lmd_lmm;
1192                 }
1193                 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
1194                         if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
1195                                 GOTO(out_req, rc = -EFAULT);
1196                         rc = -EOVERFLOW;
1197                 }
1198         skip_lmm:
1199                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1200                         struct lov_user_mds_data *lmdp;
1201                         lstat_t st = { 0 };
1202
1203                         st.st_dev     = inode->i_sb->s_dev;
1204                         st.st_mode    = body->mode;
1205                         st.st_nlink   = body->nlink;
1206                         st.st_uid     = body->uid;
1207                         st.st_gid     = body->gid;
1208                         st.st_rdev    = body->rdev;
1209                         st.st_size    = body->size;
1210                         st.st_blksize = CFS_PAGE_SIZE;
1211                         st.st_blocks  = body->blocks;
1212                         st.st_atime   = body->atime;
1213                         st.st_mtime   = body->mtime;
1214                         st.st_ctime   = body->ctime;
1215                         st.st_ino     = inode->i_ino;
1216
1217                         lmdp = (struct lov_user_mds_data *)arg;
1218                         if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1219                                 GOTO(out_req, rc = -EFAULT);
1220                 }
1221
1222                 EXIT;
1223         out_req:
1224                 ptlrpc_req_finished(request);
1225                 if (filename)
1226                         putname(filename);
1227                 return rc;
1228         }
1229         case IOC_LOV_GETINFO: {
1230                 struct lov_user_mds_data *lumd;
1231                 struct lov_stripe_md *lsm;
1232                 struct lov_user_md *lum;
1233                 struct lov_mds_md *lmm;
1234                 int lmmsize;
1235                 lstat_t st;
1236
1237                 lumd = (struct lov_user_mds_data *)arg;
1238                 lum = &lumd->lmd_lmm;
1239
1240                 rc = ll_get_max_mdsize(sbi, &lmmsize);
1241                 if (rc)
1242                         RETURN(rc);
1243
1244                 OBD_ALLOC_LARGE(lmm, lmmsize);
1245                 if (cfs_copy_from_user(lmm, lum, lmmsize))
1246                         GOTO(free_lmm, rc = -EFAULT);
1247
1248                 switch (lmm->lmm_magic) {
1249                 case LOV_USER_MAGIC_V1:
1250                         if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1251                                 break;
1252                         /* swab objects first so that stripes num will be sane */
1253                         lustre_swab_lov_user_md_objects(
1254                                 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1255                                 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1256                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1257                         break;
1258                 case LOV_USER_MAGIC_V3:
1259                         if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1260                                 break;
1261                         /* swab objects first so that stripes num will be sane */
1262                         lustre_swab_lov_user_md_objects(
1263                                 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1264                                 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1265                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1266                         break;
1267                 default:
1268                         GOTO(free_lmm, rc = -EINVAL);
1269                 }
1270
1271                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1272                 if (rc < 0)
1273                         GOTO(free_lmm, rc = -ENOMEM);
1274
1275                 /* Perform glimpse_size operation. */
1276                 memset(&st, 0, sizeof(st));
1277
1278                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1279                 if (rc)
1280                         GOTO(free_lsm, rc);
1281
1282                 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1283                         GOTO(free_lsm, rc = -EFAULT);
1284
1285                 EXIT;
1286         free_lsm:
1287                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1288         free_lmm:
1289                 OBD_FREE_LARGE(lmm, lmmsize);
1290                 return rc;
1291         }
1292         case OBD_IOC_LLOG_CATINFO: {
1293                 struct ptlrpc_request *req = NULL;
1294                 char                  *buf = NULL;
1295                 char                  *str;
1296                 int                    len = 0;
1297
1298                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1299                 if (rc)
1300                         RETURN(rc);
1301                 data = (void *)buf;
1302
1303                 if (!data->ioc_inlbuf1) {
1304                         obd_ioctl_freedata(buf, len);
1305                         RETURN(-EINVAL);
1306                 }
1307
1308                 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
1309                                            &RQF_LLOG_CATINFO);
1310                 if (req == NULL)
1311                         GOTO(out_catinfo, rc = -ENOMEM);
1312
1313                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
1314                                      data->ioc_inllen1);
1315                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
1316                                      data->ioc_inllen2);
1317
1318                 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
1319                 if (rc) {
1320                         ptlrpc_request_free(req);
1321                         GOTO(out_catinfo, rc);
1322                 }
1323
1324                 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
1325                 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
1326                 if (data->ioc_inllen2) {
1327                         str = req_capsule_client_get(&req->rq_pill,
1328                                                      &RMF_STRING);
1329                         memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
1330                 }
1331
1332                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
1333                                      data->ioc_plen1);
1334                 ptlrpc_request_set_replen(req);
1335
1336                 rc = ptlrpc_queue_wait(req);
1337                 if (!rc) {
1338                         str = req_capsule_server_get(&req->rq_pill,
1339                                                      &RMF_STRING);
1340                         if (cfs_copy_to_user(data->ioc_pbuf1, str,
1341                                              data->ioc_plen1))
1342                                 rc = -EFAULT;
1343                 }
1344                 ptlrpc_req_finished(req);
1345         out_catinfo:
1346                 obd_ioctl_freedata(buf, len);
1347                 RETURN(rc);
1348         }
1349         case OBD_IOC_QUOTACHECK: {
1350                 struct obd_quotactl *oqctl;
1351                 int error = 0;
1352
1353                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1354                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1355                         RETURN(-EPERM);
1356
1357                 OBD_ALLOC_PTR(oqctl);
1358                 if (!oqctl)
1359                         RETURN(-ENOMEM);
1360                 oqctl->qc_type = arg;
1361                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1362                 if (rc < 0) {
1363                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1364                         error = rc;
1365                 }
1366
1367                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1368                 if (rc < 0)
1369                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1370
1371                 OBD_FREE_PTR(oqctl);
1372                 return error ?: rc;
1373         }
1374         case OBD_IOC_POLL_QUOTACHECK: {
1375                 struct if_quotacheck *check;
1376
1377                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1378                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1379                         RETURN(-EPERM);
1380
1381                 OBD_ALLOC_PTR(check);
1382                 if (!check)
1383                         RETURN(-ENOMEM);
1384
1385                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1386                                    NULL);
1387                 if (rc) {
1388                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1389                         if (cfs_copy_to_user((void *)arg, check,
1390                                              sizeof(*check)))
1391                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1392                         GOTO(out_poll, rc);
1393                 }
1394
1395                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1396                                    NULL);
1397                 if (rc) {
1398                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1399                         if (cfs_copy_to_user((void *)arg, check,
1400                                              sizeof(*check)))
1401                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1402                         GOTO(out_poll, rc);
1403                 }
1404         out_poll:
1405                 OBD_FREE_PTR(check);
1406                 RETURN(rc);
1407         }
1408 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
1409         case LL_IOC_QUOTACTL_18: {
1410                 /* copy the old 1.x quota struct for internal use, then copy
1411                  * back into old format struct.  For 1.8 compatibility. */
1412                 struct if_quotactl_18 *qctl_18;
1413                 struct if_quotactl *qctl_20;
1414
1415                 OBD_ALLOC_PTR(qctl_18);
1416                 if (!qctl_18)
1417                         RETURN(-ENOMEM);
1418
1419                 OBD_ALLOC_PTR(qctl_20);
1420                 if (!qctl_20)
1421                         GOTO(out_quotactl_18, rc = -ENOMEM);
1422
1423                 if (cfs_copy_from_user(qctl_18, (void *)arg, sizeof(*qctl_18)))
1424                         GOTO(out_quotactl_20, rc = -ENOMEM);
1425
1426                 QCTL_COPY(qctl_20, qctl_18);
1427                 qctl_20->qc_idx = 0;
1428
1429                 /* XXX: dqb_valid was borrowed as a flag to mark that
1430                  *      only mds quota is wanted */
1431                 if (qctl_18->qc_cmd == Q_GETQUOTA &&
1432                     qctl_18->qc_dqblk.dqb_valid) {
1433                         qctl_20->qc_valid = QC_MDTIDX;
1434                         qctl_20->qc_dqblk.dqb_valid = 0;
1435                 } else if (qctl_18->obd_uuid.uuid[0] != '\0') {
1436                         qctl_20->qc_valid = QC_UUID;
1437                         qctl_20->obd_uuid = qctl_18->obd_uuid;
1438                 } else {
1439                         qctl_20->qc_valid = QC_GENERAL;
1440                 }
1441
1442                 rc = quotactl_ioctl(sbi, qctl_20);
1443
1444                 if (rc == 0) {
1445                         QCTL_COPY(qctl_18, qctl_20);
1446                         qctl_18->obd_uuid = qctl_20->obd_uuid;
1447
1448                         if (cfs_copy_to_user((void *)arg, qctl_18,
1449                                              sizeof(*qctl_18)))
1450                                 rc = -EFAULT;
1451                 }
1452
1453         out_quotactl_20:
1454                 OBD_FREE_PTR(qctl_20);
1455         out_quotactl_18:
1456                 OBD_FREE_PTR(qctl_18);
1457                 RETURN(rc);
1458         }
1459 #else
1460 #warning "remove old LL_IOC_QUOTACTL_18 compatibility code"
1461 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0) */
1462         case LL_IOC_QUOTACTL: {
1463                 struct if_quotactl *qctl;
1464
1465                 OBD_ALLOC_PTR(qctl);
1466                 if (!qctl)
1467                         RETURN(-ENOMEM);
1468
1469                 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1470                         GOTO(out_quotactl, rc = -EFAULT);
1471
1472                 rc = quotactl_ioctl(sbi, qctl);
1473
1474                 if (rc == 0 && cfs_copy_to_user((void *)arg,qctl,sizeof(*qctl)))
1475                         rc = -EFAULT;
1476
1477         out_quotactl:
1478                 OBD_FREE_PTR(qctl);
1479                 RETURN(rc);
1480         }
1481         case OBD_IOC_GETDTNAME:
1482         case OBD_IOC_GETMDNAME:
1483                 RETURN(ll_get_obd_name(inode, cmd, arg));
1484         case LL_IOC_FLUSHCTX:
1485                 RETURN(ll_flush_ctx(inode));
1486 #ifdef CONFIG_FS_POSIX_ACL
1487         case LL_IOC_RMTACL: {
1488             if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1489                 inode == inode->i_sb->s_root->d_inode) {
1490                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1491
1492                 LASSERT(fd != NULL);
1493                 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1494                 if (!rc)
1495                         fd->fd_flags |= LL_FILE_RMTACL;
1496                 RETURN(rc);
1497             } else
1498                 RETURN(0);
1499         }
1500 #endif
1501         case LL_IOC_GETOBDCOUNT: {
1502                 int count, vallen;
1503                 struct obd_export *exp;
1504
1505                 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1506                         RETURN(-EFAULT);
1507
1508                 /* get ost count when count is zero, get mdt count otherwise */
1509                 exp = count ? sbi->ll_md_exp : sbi->ll_dt_exp;
1510                 vallen = sizeof(count);
1511                 rc = obd_get_info(exp, sizeof(KEY_TGT_COUNT), KEY_TGT_COUNT,
1512                                   &vallen, &count, NULL);
1513                 if (rc) {
1514                         CERROR("get target count failed: %d\n", rc);
1515                         RETURN(rc);
1516                 }
1517
1518                 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1519                         RETURN(-EFAULT);
1520
1521                 RETURN(0);
1522         }
1523         case LL_IOC_PATH2FID:
1524                 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1525                                      sizeof(struct lu_fid)))
1526                         RETURN(-EFAULT);
1527                 RETURN(0);
1528         case LL_IOC_GET_CONNECT_FLAGS: {
1529                 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1530         }
1531         case OBD_IOC_CHANGELOG_SEND:
1532         case OBD_IOC_CHANGELOG_CLEAR:
1533                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1534                                     sizeof(struct ioc_changelog));
1535                 RETURN(rc);
1536         case OBD_IOC_FID2PATH:
1537                 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1538         case LL_IOC_HSM_CT_START:
1539                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1540                                     sizeof(struct lustre_kernelcomm));
1541                 RETURN(rc);
1542
1543         default:
1544                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1545         }
1546 }
1547
1548 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1549 {
1550         struct inode *inode = file->f_mapping->host;
1551         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1552         struct ll_sb_info *sbi = ll_i2sbi(inode);
1553         int api32 = ll_need_32bit_api(sbi);
1554         loff_t ret = -EINVAL;
1555         ENTRY;
1556
1557         cfs_mutex_lock(&inode->i_mutex);
1558         switch (origin) {
1559                 case SEEK_SET:
1560                         break;
1561                 case SEEK_CUR:
1562                         offset += file->f_pos;
1563                         break;
1564                 case SEEK_END:
1565                         if (offset > 0)
1566                                 GOTO(out, ret);
1567                         if (api32)
1568                                 offset += LL_DIR_END_OFF_32BIT;
1569                         else
1570                                 offset += LL_DIR_END_OFF;
1571                         break;
1572                 default:
1573                         GOTO(out, ret);
1574         }
1575
1576         if (offset >= 0 &&
1577             ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1578              (!api32 && offset <= LL_DIR_END_OFF))) {
1579                 if (offset != file->f_pos) {
1580                         if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1581                             (!api32 && offset == LL_DIR_END_OFF))
1582                                 fd->fd_dir.lfd_pos = MDS_DIR_END_OFF;
1583                         else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1584                                 fd->fd_dir.lfd_pos = offset << 32;
1585                         else
1586                                 fd->fd_dir.lfd_pos = offset;
1587                         file->f_pos = offset;
1588                         file->f_version = 0;
1589                 }
1590                 ret = offset;
1591         }
1592         GOTO(out, ret);
1593
1594 out:
1595         cfs_mutex_unlock(&inode->i_mutex);
1596         return ret;
1597 }
1598
1599 int ll_dir_open(struct inode *inode, struct file *file)
1600 {
1601         ENTRY;
1602         RETURN(ll_file_open(inode, file));
1603 }
1604
1605 int ll_dir_release(struct inode *inode, struct file *file)
1606 {
1607         ENTRY;
1608         RETURN(ll_file_release(inode, file));
1609 }
1610
1611 struct file_operations ll_dir_operations = {
1612         .llseek   = ll_dir_seek,
1613         .open     = ll_dir_open,
1614         .release  = ll_dir_release,
1615         .read     = generic_read_dir,
1616         .readdir  = ll_readdir,
1617 #ifdef HAVE_UNLOCKED_IOCTL
1618         .unlocked_ioctl   = ll_dir_ioctl,
1619 #else
1620         .ioctl          = ll_dir_ioctl,
1621 #endif
1622         .fsync    = ll_fsync,
1623 };