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  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/llite/dir.c
  37  *
  38  * Directory code for lustre client.
  39  */
  40
  41 #include <linux/fs.h>
  42 #include <linux/pagemap.h>
  43 #include <linux/mm.h>
  44 #include <linux/version.h>
  45 #include <linux/smp_lock.h>
  46 #include <asm/uaccess.h>
  47 #include <linux/buffer_head.h>   // for wait_on_buffer
  48
  49 #define DEBUG_SUBSYSTEM S_LLITE
  50
  51 #include <obd_support.h>
  52 #include <obd_class.h>
  53 #include <lustre_lib.h>
  54 #include <lustre/lustre_idl.h>
  55 #include <lustre_lite.h>
  56 #include <lustre_dlm.h>
  57 #include <lustre_fid.h>
  58 #include "llite_internal.h"
  59
  60 #ifndef HAVE_PAGE_CHECKED
  61 #ifdef HAVE_PG_FS_MISC
  62 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
  63 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
  64 #else
  65 #error PageChecked or PageFsMisc not defined in kernel
  66 #endif
  67 #endif
  68
  69 /*
  70  * (new) readdir implementation overview.
  71  *
  72  * Original lustre readdir implementation cached exact copy of raw directory
  73  * pages on the client. These pages were indexed in client page cache by
  74  * logical offset in the directory file. This design, while very simple and
  75  * intuitive had some inherent problems:
  76  *
  77  *     . it implies that byte offset to the directory entry serves as a
  78  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  79  *     ext3/htree directory entries may move due to splits, and more
  80  *     importantly,
  81  *
  82  *     . it is incompatible with the design of split directories for cmd3,
  83  *     that assumes that names are distributed across nodes based on their
  84  *     hash, and so readdir should be done in hash order.
  85  *
  86  * New readdir implementation does readdir in hash order, and uses hash of a
  87  * file name as a telldir/seekdir cookie. This led to number of complications:
  88  *
  89  *     . hash is not unique, so it cannot be used to index cached directory
  90  *     pages on the client (note, that it requires a whole pageful of hash
  91  *     collided entries to cause two pages to have identical hashes);
  92  *
  93  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  94  *     entry cookie. ext3/htree has the same problem and lustre implementation
  95  *     mimics their solution: seekdir(hash) positions directory at the first
  96  *     entry with the given hash.
  97  *
  98  * Client side.
  99  *
 100  * 0. caching
 101  *
 102  * Client caches directory pages using hash of the first entry as an index. As
 103  * noted above hash is not unique, so this solution doesn't work as is:
 104  * special processing is needed for "page hash chains" (i.e., sequences of
 105  * pages filled with entries all having the same hash value).
 106  *
 107  * First, such chains have to be detected. To this end, server returns to the
 108  * client the hash of the first entry on the page next to one returned. When
 109  * client detects that this hash is the same as hash of the first entry on the
 110  * returned page, page hash collision has to be handled. Pages in the
 111  * hash chain, except first one, are termed "overflow pages".
 112  *
 113  * Solution to index uniqueness problem is to not cache overflow
 114  * pages. Instead, when page hash collision is detected, all overflow pages
 115  * from emerging chain are immediately requested from the server and placed in
 116  * a special data structure (struct ll_dir_chain). This data structure is used
 117  * by ll_readdir() to process entries from overflow pages. When readdir
 118  * invocation finishes, overflow pages are discarded. If page hash collision
 119  * chain weren't completely processed, next call to readdir will again detect
 120  * page hash collision, again read overflow pages in, process next portion of
 121  * entries and again discard the pages. This is not as wasteful as it looks,
 122  * because, given reasonable hash, page hash collisions are extremely rare.
 123  *
 124  * 1. directory positioning
 125  *
 126  * When seekdir(hash) is called, original
 127  *
 128  *
 129  *
 130  *
 131  *
 132  *
 133  *
 134  *
 135  * Server.
 136  *
 137  * identification of and access to overflow pages
 138  *
 139  * page format
 140  *
 141  *
 142  *
 143  *
 144  *
 145  */
 146
 147 /* returns the page unlocked, but with a reference */
 148 static int ll_dir_readpage(struct file *file, struct page *page)
 149 {
 150         struct inode *inode = page->mapping->host;
 151         struct ptlrpc_request *request;
 152         struct mdt_body *body;
 153         struct obd_capa *oc;
 154         __u64 hash;
 155         int rc;
 156         ENTRY;
 157
 158         hash = (__u64)hash_x_index(page->index);
 159         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) off %lu\n",
 160                inode->i_ino, inode->i_generation, inode, (unsigned long)hash);
 161
 162         oc = ll_mdscapa_get(inode);
 163         rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
 164                          oc, hash, page, &request);
 165         capa_put(oc);
 166         if (!rc) {
 167                 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
 168                 /* Checked by mdc_readpage() */
 169                 LASSERT(body != NULL);
 170
 171                 if (body->valid & OBD_MD_FLSIZE) {
 172                         ll_inode_size_lock(inode, 0);
 173                         i_size_write(inode, body->size);
 174                         ll_inode_size_unlock(inode, 0);
 175                 }
 176                 SetPageUptodate(page);
 177         }
 178         ptlrpc_req_finished(request);
 179
 180         unlock_page(page);
 181         EXIT;
 182         return rc;
 183 }
 184
 185 struct address_space_operations ll_dir_aops = {
 186         .readpage  = ll_dir_readpage,
 187 };
 188
 189 static void ll_check_page(struct inode *dir, struct page *page)
 190 {
 191         /* XXX: check page format later */
 192         SetPageChecked(page);
 193 }
 194
 195 static void ll_release_page(struct page *page, __u64 hash,
 196                             __u64 start, __u64 end)
 197 {
 198         kunmap(page);
 199         lock_page(page);
 200         if (likely(page->mapping != NULL)) {
 201                 ll_truncate_complete_page(page);
 202                 unlock_page(page);
 203         } else {
 204                 unlock_page(page);
 205                 CWARN("NULL mapping page %p, truncated by others: "
 206                       "hash("LPX64") | start("LPX64") | end("LPX64")\n",
 207                       page, hash, start, end);
 208         }
 209         page_cache_release(page);
 210 }
 211
 212 /*
 213  * Find, kmap and return page that contains given hash.
 214  */
 215 static struct page *ll_dir_page_locate(struct inode *dir, __u64 hash,
 216                                        __u64 *start, __u64 *end)
 217 {
 218         struct address_space *mapping = dir->i_mapping;
 219         /*
 220          * Complement of hash is used as an index so that
 221          * radix_tree_gang_lookup() can be used to find a page with starting
 222          * hash _smaller_ than one we are looking for.
 223          */
 224         unsigned long offset = hash_x_index((__u32)hash);
 225         struct page *page;
 226         int found;
 227
 228         TREE_READ_LOCK_IRQ(mapping);
 229         found = radix_tree_gang_lookup(&mapping->page_tree,
 230                                        (void **)&page, offset, 1);
 231         if (found > 0) {
 232                 struct lu_dirpage *dp;
 233
 234                 page_cache_get(page);
 235                 TREE_READ_UNLOCK_IRQ(mapping);
 236                 /*
 237                  * In contrast to find_lock_page() we are sure that directory
 238                  * page cannot be truncated (while DLM lock is held) and,
 239                  * hence, can avoid restart.
 240                  *
 241                  * In fact, page cannot be locked here at all, because
 242                  * ll_dir_readpage() does synchronous io.
 243                  */
 244                 wait_on_page(page);
 245                 if (PageUptodate(page)) {
 246                         dp = kmap(page);
 247                         *start = le64_to_cpu(dp->ldp_hash_start);
 248                         *end   = le64_to_cpu(dp->ldp_hash_end);
 249                         LASSERT(*start <= hash);
 250                         if (hash > *end || (*end != *start && hash == *end)) {
 251                                 ll_release_page(page, hash, *start, *end);
 252                                 page = NULL;
 253                         }
 254                 } else {
 255                         page_cache_release(page);
 256                         page = ERR_PTR(-EIO);
 257                 }
 258
 259         } else {
 260                 TREE_READ_UNLOCK_IRQ(mapping);
 261                 page = NULL;
 262         }
 263         return page;
 264 }
 265
 266 struct page *ll_get_dir_page(struct inode *dir, __u64 hash, int exact,
 267                              struct ll_dir_chain *chain)
 268 {
 269         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 270         struct address_space *mapping = dir->i_mapping;
 271         struct lustre_handle lockh;
 272         struct lu_dirpage *dp;
 273         struct page *page;
 274         ldlm_mode_t mode;
 275         int rc;
 276         __u64 start = 0;
 277         __u64 end = 0;
 278
 279         mode = LCK_PR;
 280         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 281                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
 282         if (!rc) {
 283                 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
 284                        ll_md_blocking_ast, ldlm_completion_ast, NULL, dir };
 285                 struct lookup_intent it = { .it_op = IT_READDIR };
 286                 struct ptlrpc_request *request;
 287                 struct md_op_data *op_data;
 288
 289                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
 290                                              LUSTRE_OPC_ANY, NULL);
 291                 if (IS_ERR(op_data))
 292                         return (void *)op_data;
 293
 294                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
 295                                 op_data, &lockh, NULL, 0, NULL, 0);
 296
 297                 ll_finish_md_op_data(op_data);
 298
 299                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 300                 if (request)
 301                         ptlrpc_req_finished(request);
 302                 if (rc < 0) {
 303                         CERROR("lock enqueue: rc: %d\n", rc);
 304                         return ERR_PTR(rc);
 305                 }
 306         } else {
 307                 /* for cross-ref object, l_ast_data of the lock may not be set,
 308                  * we reset it here */
 309                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie, dir);
 310         }
 311         ldlm_lock_dump_handle(D_OTHER, &lockh);
 312
 313         page = ll_dir_page_locate(dir, hash, &start, &end);
 314         if (IS_ERR(page))
 315                 GOTO(out_unlock, page);
 316
 317         if (page != NULL) {
 318                 /*
 319                  * XXX nikita: not entirely correct handling of a corner case:
 320                  * suppose hash chain of entries with hash value HASH crosses
 321                  * border between pages P0 and P1. First both P0 and P1 are
 322                  * cached, seekdir() is called for some entry from the P0 part
 323                  * of the chain. Later P0 goes out of cache. telldir(HASH)
 324                  * happens and finds P1, as it starts with matching hash
 325                  * value. Remaining entries from P0 part of the chain are
 326                  * skipped. (Is that really a bug?)
 327                  *
 328                  * Possible solutions: 0. don't cache P1 is such case, handle
 329                  * it as an "overflow" page. 1. invalidate all pages at
 330                  * once. 2. use HASH|1 as an index for P1.
 331                  */
 332                 if (exact && hash != start) {
 333                         /*
 334                          * readdir asked for a page starting _exactly_ from
 335                          * given hash, but cache contains stale page, with
 336                          * entries with smaller hash values. Stale page should
 337                          * be invalidated, and new one fetched.
 338                          */
 339                         CWARN("Stale readpage page %p: "LPX64" != "LPX64"\n",
 340                               page, hash, start);
 341                         ll_release_page(page, hash, start, end);
 342                 } else {
 343                         GOTO(hash_collision, page);
 344                 }
 345         }
 346
 347         page = read_cache_page(mapping, hash_x_index((__u32)hash),
 348                                (filler_t*)mapping->a_ops->readpage, NULL);
 349         if (IS_ERR(page))
 350                 GOTO(out_unlock, page);
 351
 352         wait_on_page(page);
 353         (void)kmap(page);
 354         if (!PageUptodate(page))
 355                 goto fail;
 356         if (!PageChecked(page))
 357                 ll_check_page(dir, page);
 358         if (PageError(page))
 359                 goto fail;
 360 hash_collision:
 361         dp = page_address(page);
 362
 363         start = le64_to_cpu(dp->ldp_hash_start);
 364         end   = le64_to_cpu(dp->ldp_hash_end);
 365         if (end == start) {
 366                 LASSERT(start == hash);
 367                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
 368                 /*
 369                  * Fetch whole overflow chain...
 370                  *
 371                  * XXX not yet.
 372                  */
 373                 goto fail;
 374         }
 375 out_unlock:
 376         ldlm_lock_decref(&lockh, mode);
 377         return page;
 378
 379 fail:
 380         ll_put_page(page);
 381         page = ERR_PTR(-EIO);
 382         goto out_unlock;
 383 }
 384
 385 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 386 {
 387         struct inode         *inode = filp->f_dentry->d_inode;
 388         struct ll_inode_info *info  = ll_i2info(inode);
 389         struct ll_sb_info    *sbi   = ll_i2sbi(inode);
 390         __u64                 pos   = filp->f_pos;
 391         struct page          *page;
 392         struct ll_dir_chain   chain;
 393         int rc;
 394         int done;
 395         int shift;
 396         ENTRY;
 397
 398         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu\n",
 399                inode->i_ino, inode->i_generation, inode,
 400                (unsigned long)pos, i_size_read(inode));
 401
 402         if (pos == DIR_END_OFF)
 403                 /*
 404                  * end-of-file.
 405                  */
 406                 RETURN(0);
 407
 408         rc    = 0;
 409         done  = 0;
 410         shift = 0;
 411         ll_dir_chain_init(&chain);
 412
 413         page = ll_get_dir_page(inode, pos, 0, &chain);
 414
 415         while (rc == 0 && !done) {
 416                 struct lu_dirpage *dp;
 417                 struct lu_dirent  *ent;
 418
 419                 if (!IS_ERR(page)) {
 420                         /*
 421                          * If page is empty (end of directoryis reached),
 422                          * use this value.
 423                          */
 424                         __u64 hash = DIR_END_OFF;
 425                         __u64 next;
 426
 427                         dp = page_address(page);
 428                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 429                              ent = lu_dirent_next(ent)) {
 430                                 char          *name;
 431                                 int            namelen;
 432                                 struct lu_fid  fid;
 433                                 ino_t          ino;
 434
 435                                 /*
 436                                  * XXX: implement correct swabbing here.
 437                                  */
 438
 439                                 hash    = le64_to_cpu(ent->lde_hash);
 440                                 namelen = le16_to_cpu(ent->lde_namelen);
 441
 442                                 if (hash < pos)
 443                                         /*
 444                                          * Skip until we find target hash
 445                                          * value.
 446                                          */
 447                                         continue;
 448
 449                                 if (namelen == 0)
 450                                         /*
 451                                          * Skip dummy record.
 452                                          */
 453                                         continue;
 454
 455                                 fid  = ent->lde_fid;
 456                                 name = ent->lde_name;
 457                                 fid_le_to_cpu(&fid, &fid);
 458                                 ino  = ll_fid_build_ino(sbi, &fid);
 459
 460                                 done = filldir(cookie, name, namelen,
 461                                                (loff_t)hash, ino, DT_UNKNOWN);
 462                         }
 463                         next = le64_to_cpu(dp->ldp_hash_end);
 464                         ll_put_page(page);
 465                         if (!done) {
 466                                 pos = next;
 467                                 if (pos == DIR_END_OFF)
 468                                         /*
 469                                          * End of directory reached.
 470                                          */
 471                                         done = 1;
 472                                 else if (1 /* chain is exhausted*/)
 473                                         /*
 474                                          * Normal case: continue to the next
 475                                          * page.
 476                                          */
 477                                         page = ll_get_dir_page(inode, pos, 1,
 478                                                                &chain);
 479                                 else {
 480                                         /*
 481                                          * go into overflow page.
 482                                          */
 483                                 }
 484                         } else
 485                                 pos = hash;
 486                 } else {
 487                         rc = PTR_ERR(page);
 488                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 489                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 490                 }
 491         }
 492
 493         filp->f_pos = (loff_t)(__s32)pos;
 494         filp->f_version = inode->i_version;
 495         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 496
 497         ll_dir_chain_fini(&chain);
 498
 499         RETURN(rc);
 500 }
 501
 502 #define QCTL_COPY(out, in)              \
 503 do {                                    \
 504         Q_COPY(out, in, qc_cmd);        \
 505         Q_COPY(out, in, qc_type);       \
 506         Q_COPY(out, in, qc_id);         \
 507         Q_COPY(out, in, qc_stat);       \
 508         Q_COPY(out, in, qc_dqinfo);     \
 509         Q_COPY(out, in, qc_dqblk);      \
 510 } while (0)
 511
 512 int ll_send_mgc_param(struct obd_export *mgc, char *string)
 513 {
 514         struct mgs_send_param *msp;
 515         int rc = 0;
 516
 517         OBD_ALLOC_PTR(msp);
 518         if (!msp)
 519                 return -ENOMEM;
 520
 521         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
 522         rc = obd_set_info_async(mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
 523                                 sizeof(struct mgs_send_param), msp, NULL);
 524         if (rc)
 525                 CERROR("Failed to set parameter: %d\n", rc);
 526         OBD_FREE_PTR(msp);
 527
 528         return rc;
 529 }
 530
 531 char *ll_get_fsname(struct inode *inode)
 532 {
 533         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 534         char *ptr, *fsname;
 535         int len;
 536
 537         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
 538         len = strlen(lsi->lsi_lmd->lmd_profile);
 539         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
 540         if (ptr && (strcmp(ptr, "-client") == 0))
 541                 len -= 7;
 542         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
 543         fsname[len] = '\0';
 544
 545         return fsname;
 546 }
 547
 548 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
 549                      int set_default)
 550 {
 551         struct ll_sb_info *sbi = ll_i2sbi(inode);
 552         struct md_op_data *op_data;
 553         struct ptlrpc_request *req = NULL;
 554         int rc = 0;
 555         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 556         struct obd_device *mgc = lsi->lsi_mgc;
 557         char *fsname = NULL, *param = NULL;
 558
 559         /*
 560          * This is coming from userspace, so should be in
 561          * local endian.  But the MDS would like it in little
 562          * endian, so we swab it before we send it.
 563          */
 564         if (lump->lmm_magic != LOV_USER_MAGIC)
 565                 RETURN(-EINVAL);
 566
 567         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC))
 568                 lustre_swab_lov_user_md(lump);
 569
 570         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 571                                      LUSTRE_OPC_ANY, NULL);
 572         if (IS_ERR(op_data))
 573                 RETURN(PTR_ERR(op_data));
 574
 575         /* swabbing is done in lov_setstripe() on server side */
 576         rc = md_setattr(sbi->ll_md_exp, op_data, lump, sizeof(*lump),
 577                         NULL, 0, &req, NULL);
 578         ll_finish_md_op_data(op_data);
 579         ptlrpc_req_finished(req);
 580         if (rc) {
 581                 if (rc != -EPERM && rc != -EACCES)
 582                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 583         }
 584
 585         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
 586                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
 587
 588                 /* Get fsname and assume devname to be -MDT0000. */
 589                 fsname = ll_get_fsname(inode);
 590                 /* Set root stripesize */
 591                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
 592                         lump->lmm_stripe_size);
 593                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 594                 if (rc)
 595                         goto end;
 596
 597                 /* Set root stripecount */
 598                 sprintf(param, "%s-MDT0000.lov.stripecount=%u", fsname,
 599                         lump->lmm_stripe_count);
 600                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 601                 if (rc)
 602                         goto end;
 603
 604                 /* Set root stripeoffset */
 605                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%u", fsname,
 606                         lump->lmm_stripe_offset);
 607                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 608                 if (rc)
 609                         goto end;
 610 end:
 611                 if (fsname)
 612                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
 613                 if (param)
 614                         OBD_FREE(param, MGS_PARAM_MAXLEN);
 615         }
 616         return rc;
 617 }
 618
 619 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 620                      int *lmm_size, struct ptlrpc_request **request)
 621 {
 622         struct ll_sb_info *sbi = ll_i2sbi(inode);
 623         struct mdt_body   *body;
 624         struct lov_mds_md *lmm = NULL;
 625         struct ptlrpc_request *req = NULL;
 626         int rc, lmmsize;
 627         struct obd_capa *oc;
 628
 629         rc = ll_get_max_mdsize(sbi, &lmmsize);
 630         if (rc)
 631                 RETURN(rc);
 632
 633         oc = ll_mdscapa_get(inode);
 634         rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode),
 635                         oc, OBD_MD_FLEASIZE | OBD_MD_FLDIREA,
 636                         lmmsize, &req);
 637         capa_put(oc);
 638         if (rc < 0) {
 639                 CDEBUG(D_INFO, "md_getattr failed on inode "
 640                        "%lu/%u: rc %d\n", inode->i_ino,
 641                        inode->i_generation, rc);
 642                 GOTO(out, rc);
 643         }
 644
 645         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 646         LASSERT(body != NULL);
 647
 648         lmmsize = body->eadatasize;
 649
 650         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 651             lmmsize == 0) {
 652                 GOTO(out, rc = -ENODATA);
 653         }
 654
 655         lmm = req_capsule_server_sized_get(&req->rq_pill,
 656                                            &RMF_MDT_MD, lmmsize);
 657         LASSERT(lmm != NULL);
 658
 659         /*
 660          * This is coming from the MDS, so is probably in
 661          * little endian.  We convert it to host endian before
 662          * passing it to userspace.
 663          */
 664         if (lmm->lmm_magic == __swab32(LOV_MAGIC)) {
 665                 lustre_swab_lov_user_md((struct lov_user_md *)lmm);
 666         }
 667 out:
 668         *lmmp = lmm;
 669         *lmm_size = lmmsize;
 670         *request = req;
 671         return rc;
 672 }
 673
 674 static int ll_dir_ioctl(struct inode *inode, struct file *file,
 675                         unsigned int cmd, unsigned long arg)
 676 {
 677         struct ll_sb_info *sbi = ll_i2sbi(inode);
 678         struct obd_ioctl_data *data;
 679         ENTRY;
 680
 681         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
 682                inode->i_ino, inode->i_generation, inode, cmd);
 683
 684         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
 685         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
 686                 return -ENOTTY;
 687
 688         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
 689         switch(cmd) {
 690         case EXT3_IOC_GETFLAGS:
 691         case EXT3_IOC_SETFLAGS:
 692                 RETURN(ll_iocontrol(inode, file, cmd, arg));
 693         case EXT3_IOC_GETVERSION_OLD:
 694         case EXT3_IOC_GETVERSION:
 695                 RETURN(put_user(inode->i_generation, (int *)arg));
 696         /* We need to special case any other ioctls we want to handle,
 697          * to send them to the MDS/OST as appropriate and to properly
 698          * network encode the arg field.
 699         case EXT3_IOC_SETVERSION_OLD:
 700         case EXT3_IOC_SETVERSION:
 701         */
 702         case IOC_MDC_LOOKUP: {
 703                 struct ptlrpc_request *request = NULL;
 704                 int namelen, rc, len = 0;
 705                 char *buf = NULL;
 706                 char *filename;
 707                 struct obd_capa *oc;
 708
 709                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 710                 if (rc)
 711                         RETURN(rc);
 712                 data = (void *)buf;
 713
 714                 filename = data->ioc_inlbuf1;
 715                 namelen = data->ioc_inllen1;
 716
 717                 if (namelen < 1) {
 718                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
 719                         GOTO(out, rc = -EINVAL);
 720                 }
 721
 722                 oc = ll_mdscapa_get(inode);
 723                 rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode), oc,
 724                                      filename, namelen, OBD_MD_FLID, 0,
 725                                      ll_i2suppgid(inode), &request);
 726                 capa_put(oc);
 727                 if (rc < 0) {
 728                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 729                         GOTO(out, rc);
 730                 }
 731
 732                 ptlrpc_req_finished(request);
 733
 734                 EXIT;
 735         out:
 736                 obd_ioctl_freedata(buf, len);
 737                 return rc;
 738         }
 739         case LL_IOC_LOV_SETSTRIPE: {
 740                 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
 741                 int rc = 0;
 742                 int set_default = 0;
 743
 744                 LASSERT(sizeof(lum) == sizeof(*lump));
 745                 LASSERT(sizeof(lum.lmm_objects[0]) ==
 746                         sizeof(lump->lmm_objects[0]));
 747                 rc = copy_from_user(&lum, lump, sizeof(lum));
 748                 if (rc)
 749                         RETURN(-EFAULT);
 750
 751                 if (inode->i_sb->s_root == file->f_dentry)
 752                         set_default = 1;
 753
 754                 rc = ll_dir_setstripe(inode, &lum, set_default);
 755
 756                 RETURN(rc);
 757         }
 758         case LL_IOC_OBD_STATFS:
 759                 RETURN(ll_obd_statfs(inode, (void *)arg));
 760         case LL_IOC_LOV_GETSTRIPE:
 761         case LL_IOC_MDC_GETINFO:
 762         case IOC_MDC_GETFILEINFO:
 763         case IOC_MDC_GETFILESTRIPE: {
 764                 struct ptlrpc_request *request = NULL;
 765                 struct lov_user_md *lump;
 766                 struct lov_mds_md *lmm = NULL;
 767                 struct mdt_body *body;
 768                 char *filename = NULL;
 769                 int rc, lmmsize;
 770
 771                 if (cmd == IOC_MDC_GETFILEINFO ||
 772                     cmd == IOC_MDC_GETFILESTRIPE) {
 773                         filename = getname((const char *)arg);
 774                         if (IS_ERR(filename))
 775                                 RETURN(PTR_ERR(filename));
 776
 777                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
 778                                                       &lmmsize, &request);
 779                 } else {
 780                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
 781                 }
 782
 783                 if (request) {
 784                         body = req_capsule_server_get(&request->rq_pill,
 785                                                       &RMF_MDT_BODY);
 786                         LASSERT(body != NULL);
 787                 } else {
 788                         GOTO(out_req, rc);
 789                 }
 790
 791                 if (rc < 0) {
 792                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
 793                                                cmd == LL_IOC_MDC_GETINFO))
 794                                 GOTO(skip_lmm, rc = 0);
 795                         else
 796                                 GOTO(out_req, rc);
 797                 }
 798
 799                 if (cmd == IOC_MDC_GETFILESTRIPE ||
 800                     cmd == LL_IOC_LOV_GETSTRIPE) {
 801                         lump = (struct lov_user_md *)arg;
 802                 } else {
 803                         struct lov_user_mds_data *lmdp;
 804                         lmdp = (struct lov_user_mds_data *)arg;
 805                         lump = &lmdp->lmd_lmm;
 806                 }
 807                 rc = copy_to_user(lump, lmm, lmmsize);
 808                 if (rc)
 809                         GOTO(out_lmm, rc = -EFAULT);
 810         skip_lmm:
 811                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
 812                         struct lov_user_mds_data *lmdp;
 813                         lstat_t st = { 0 };
 814
 815                         st.st_dev     = inode->i_sb->s_dev;
 816                         st.st_mode    = body->mode;
 817                         st.st_nlink   = body->nlink;
 818                         st.st_uid     = body->uid;
 819                         st.st_gid     = body->gid;
 820                         st.st_rdev    = body->rdev;
 821                         st.st_size    = body->size;
 822                         st.st_blksize = CFS_PAGE_SIZE;
 823                         st.st_blocks  = body->blocks;
 824                         st.st_atime   = body->atime;
 825                         st.st_mtime   = body->mtime;
 826                         st.st_ctime   = body->ctime;
 827                         st.st_ino     = inode->i_ino;
 828
 829                         lmdp = (struct lov_user_mds_data *)arg;
 830                         rc = copy_to_user(&lmdp->lmd_st, &st, sizeof(st));
 831                         if (rc)
 832                                 GOTO(out_lmm, rc = -EFAULT);
 833                 }
 834
 835                 EXIT;
 836         out_lmm:
 837                 if (lmm && lmm->lmm_magic == LOV_MAGIC_JOIN)
 838                         OBD_FREE(lmm, lmmsize);
 839         out_req:
 840                 ptlrpc_req_finished(request);
 841                 if (filename)
 842                         putname(filename);
 843                 return rc;
 844         }
 845         case IOC_LOV_GETINFO: {
 846                 struct lov_user_mds_data *lumd;
 847                 struct lov_stripe_md *lsm;
 848                 struct lov_user_md *lum;
 849                 struct lov_mds_md *lmm;
 850                 int lmmsize;
 851                 lstat_t st;
 852                 int rc;
 853
 854                 lumd = (struct lov_user_mds_data *)arg;
 855                 lum = &lumd->lmd_lmm;
 856
 857                 rc = ll_get_max_mdsize(sbi, &lmmsize);
 858                 if (rc)
 859                         RETURN(rc);
 860
 861                 OBD_ALLOC(lmm, lmmsize);
 862                 rc = copy_from_user(lmm, lum, lmmsize);
 863                 if (rc)
 864                         GOTO(free_lmm, rc = -EFAULT);
 865
 866                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
 867                 if (rc < 0)
 868                         GOTO(free_lmm, rc = -ENOMEM);
 869
 870                 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
 871                 if (rc)
 872                         GOTO(free_lsm, rc);
 873
 874                 /* Perform glimpse_size operation. */
 875                 memset(&st, 0, sizeof(st));
 876
 877                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
 878                 if (rc)
 879                         GOTO(free_lsm, rc);
 880
 881                 rc = copy_to_user(&lumd->lmd_st, &st, sizeof(st));
 882                 if (rc)
 883                         GOTO(free_lsm, rc = -EFAULT);
 884
 885                 EXIT;
 886         free_lsm:
 887                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
 888         free_lmm:
 889                 OBD_FREE(lmm, lmmsize);
 890                 return rc;
 891         }
 892         case OBD_IOC_LLOG_CATINFO: {
 893                 struct ptlrpc_request *req = NULL;
 894                 char                  *buf = NULL;
 895                 char                  *str;
 896                 int                    len = 0;
 897                 int                    rc;
 898
 899                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 900                 if (rc)
 901                         RETURN(rc);
 902                 data = (void *)buf;
 903
 904                 if (!data->ioc_inlbuf1) {
 905                         obd_ioctl_freedata(buf, len);
 906                         RETURN(-EINVAL);
 907                 }
 908
 909                 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
 910                                            &RQF_LLOG_CATINFO);
 911                 if (req == NULL)
 912                         GOTO(out_catinfo, rc = -ENOMEM);
 913
 914                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 915                                      data->ioc_inllen1);
 916                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
 917                                      data->ioc_inllen2);
 918
 919                 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
 920                 if (rc) {
 921                         ptlrpc_request_free(req);
 922                         GOTO(out_catinfo, rc);
 923                 }
 924
 925                 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 926                 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
 927                 if (data->ioc_inllen2) {
 928                         str = req_capsule_client_get(&req->rq_pill,
 929                                                      &RMF_STRING);
 930                         memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
 931                 }
 932
 933                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
 934                                      data->ioc_plen1);
 935                 ptlrpc_request_set_replen(req);
 936
 937                 rc = ptlrpc_queue_wait(req);
 938                 if (!rc) {
 939                         str = req_capsule_server_get(&req->rq_pill,
 940                                                      &RMF_STRING);
 941                         rc = copy_to_user(data->ioc_pbuf1, str, data->ioc_plen1);
 942                 }
 943                 ptlrpc_req_finished(req);
 944         out_catinfo:
 945                 obd_ioctl_freedata(buf, len);
 946                 RETURN(rc);
 947         }
 948         case OBD_IOC_QUOTACHECK: {
 949                 struct obd_quotactl *oqctl;
 950                 int rc, error = 0;
 951
 952                 if (!cfs_capable(CFS_CAP_SYS_ADMIN))
 953                         RETURN(-EPERM);
 954
 955                 OBD_ALLOC_PTR(oqctl);
 956                 if (!oqctl)
 957                         RETURN(-ENOMEM);
 958                 oqctl->qc_type = arg;
 959                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
 960                 if (rc < 0) {
 961                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
 962                         error = rc;
 963                 }
 964
 965                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
 966                 if (rc < 0)
 967                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
 968
 969                 OBD_FREE_PTR(oqctl);
 970                 return error ?: rc;
 971         }
 972         case OBD_IOC_POLL_QUOTACHECK: {
 973                 struct if_quotacheck *check;
 974                 int rc;
 975
 976                 if (!cfs_capable(CFS_CAP_SYS_ADMIN))
 977                         RETURN(-EPERM);
 978
 979                 OBD_ALLOC_PTR(check);
 980                 if (!check)
 981                         RETURN(-ENOMEM);
 982
 983                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
 984                                    NULL);
 985                 if (rc) {
 986                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
 987                         if (copy_to_user((void *)arg, check, sizeof(*check)))
 988                                 rc = -EFAULT;
 989                         GOTO(out_poll, rc);
 990                 }
 991
 992                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
 993                                    NULL);
 994                 if (rc) {
 995                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
 996                         if (copy_to_user((void *)arg, check, sizeof(*check)))
 997                                 rc = -EFAULT;
 998                         GOTO(out_poll, rc);
 999                 }
1000         out_poll:
1001                 OBD_FREE_PTR(check);
1002                 RETURN(rc);
1003         }
1004 #ifdef HAVE_QUOTA_SUPPORT
1005         case OBD_IOC_QUOTACTL: {
1006                 struct if_quotactl *qctl;
1007                 struct obd_quotactl *oqctl;
1008
1009                 int cmd, type, id, rc = 0;
1010
1011                 OBD_ALLOC_PTR(qctl);
1012                 if (!qctl)
1013                         RETURN(-ENOMEM);
1014
1015                 OBD_ALLOC_PTR(oqctl);
1016                 if (!oqctl) {
1017                         OBD_FREE_PTR(qctl);
1018                         RETURN(-ENOMEM);
1019                 }
1020                 if (copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1021                         GOTO(out_quotactl, rc = -EFAULT);
1022
1023                 cmd = qctl->qc_cmd;
1024                 type = qctl->qc_type;
1025                 id = qctl->qc_id;
1026                 switch (cmd) {
1027                 case Q_QUOTAON:
1028                 case Q_QUOTAOFF:
1029                 case Q_SETQUOTA:
1030                 case Q_SETINFO:
1031                         if (!cfs_capable(CFS_CAP_SYS_ADMIN))
1032                                 GOTO(out_quotactl, rc = -EPERM);
1033                         break;
1034                 case Q_GETQUOTA:
1035                         if (((type == USRQUOTA && current->euid != id) ||
1036                              (type == GRPQUOTA && !in_egroup_p(id))) &&
1037                             !cfs_capable(CFS_CAP_SYS_ADMIN))
1038                                 GOTO(out_quotactl, rc = -EPERM);
1039
1040                         /* XXX: dqb_valid is borrowed as a flag to mark that
1041                          *      only mds quota is wanted */
1042                         if (qctl->qc_dqblk.dqb_valid)
1043                                 qctl->obd_uuid = sbi->ll_md_exp->exp_obd->
1044                                                         u.cli.cl_target_uuid;
1045                         break;
1046                 case Q_GETINFO:
1047                         break;
1048                 default:
1049                         CERROR("unsupported quotactl op: %#x\n", cmd);
1050                         GOTO(out_quotactl, rc = -ENOTTY);
1051                 }
1052
1053                 QCTL_COPY(oqctl, qctl);
1054
1055                 if (qctl->obd_uuid.uuid[0]) {
1056                         struct obd_device *obd;
1057                         struct obd_uuid *uuid = &qctl->obd_uuid;
1058
1059                         obd = class_find_client_notype(uuid,
1060                                          &sbi->ll_dt_exp->exp_obd->obd_uuid);
1061                         if (!obd)
1062                                 GOTO(out_quotactl, rc = -ENOENT);
1063
1064                         if (cmd == Q_GETINFO)
1065                                 oqctl->qc_cmd = Q_GETOINFO;
1066                         else if (cmd == Q_GETQUOTA)
1067                                 oqctl->qc_cmd = Q_GETOQUOTA;
1068                         else
1069                                 GOTO(out_quotactl, rc = -EINVAL);
1070
1071                         if (sbi->ll_md_exp->exp_obd == obd) {
1072                                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1073                         } else {
1074                                 int i;
1075                                 struct obd_export *exp;
1076                                 struct lov_obd *lov = &sbi->ll_dt_exp->
1077                                                             exp_obd->u.lov;
1078
1079                                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1080                                         if (!lov->lov_tgts[i] ||
1081                                             !lov->lov_tgts[i]->ltd_active)
1082                                                 continue;
1083                                         exp = lov->lov_tgts[i]->ltd_exp;
1084                                         if (exp->exp_obd == obd) {
1085                                                 rc = obd_quotactl(exp, oqctl);
1086                                                 break;
1087                                         }
1088                                 }
1089                         }
1090
1091                         oqctl->qc_cmd = cmd;
1092                         QCTL_COPY(qctl, oqctl);
1093
1094                         if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1095                                 rc = -EFAULT;
1096
1097                         GOTO(out_quotactl, rc);
1098                 }
1099
1100                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1101                 if (rc && rc != -EBUSY && cmd == Q_QUOTAON) {
1102                         oqctl->qc_cmd = Q_QUOTAOFF;
1103                         obd_quotactl(sbi->ll_md_exp, oqctl);
1104                 }
1105
1106                 QCTL_COPY(qctl, oqctl);
1107
1108                 if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1109                         rc = -EFAULT;
1110         out_quotactl:
1111                 OBD_FREE_PTR(qctl);
1112                 OBD_FREE_PTR(oqctl);
1113                 RETURN(rc);
1114         }
1115 #endif /* HAVE_QUOTA_SUPPORT */
1116         case OBD_IOC_GETNAME: {
1117                 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1118                 if (!obd)
1119                         RETURN(-EFAULT);
1120                 if (copy_to_user((void *)arg, obd->obd_name,
1121                                 strlen(obd->obd_name) + 1))
1122                         RETURN (-EFAULT);
1123                 RETURN(0);
1124         }
1125         case LL_IOC_FLUSHCTX:
1126                 RETURN(ll_flush_ctx(inode));
1127 #ifdef CONFIG_FS_POSIX_ACL
1128         case LL_IOC_RMTACL: {
1129             if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1130                 inode == inode->i_sb->s_root->d_inode) {
1131                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1132                 int rc;
1133
1134                 LASSERT(fd != NULL);
1135                 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1136                 if (!rc)
1137                         fd->fd_flags |= LL_FILE_RMTACL;
1138                 RETURN(rc);
1139             } else
1140                 RETURN(0);
1141         }
1142 #endif
1143         default:
1144                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1145         }
1146 }
1147
1148 int ll_dir_open(struct inode *inode, struct file *file)
1149 {
1150         ENTRY;
1151         RETURN(ll_file_open(inode, file));
1152 }
1153
1154 int ll_dir_release(struct inode *inode, struct file *file)
1155 {
1156         ENTRY;
1157         RETURN(ll_file_release(inode, file));
1158 }
1159
1160 struct file_operations ll_dir_operations = {
1161         .open     = ll_dir_open,
1162         .release  = ll_dir_release,
1163         .read     = generic_read_dir,
1164         .readdir  = ll_readdir,
1165         .ioctl    = ll_dir_ioctl
1166 };