lustre/llite/dir.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 1992, 1993, 1994, 1995
   5  * Remy Card (card@masi.ibp.fr)
   6  * Laboratoire MASI - Institut Blaise Pascal
   7  * Universite Pierre et Marie Curie (Paris VI)
   8  *
   9  *  from
  10  *
  11  *  linux/fs/minix/dir.c
  12  *  linux/fs/ext2/dir.c
  13  *
  14  *  Copyright (C) 1991, 1992  Linus Torvalds
  15  *
  16  *  ext2 directory handling functions
  17  *
  18  *  Big-endian to little-endian byte-swapping/bitmaps by
  19  *        David S. Miller (davem@caip.rutgers.edu), 1995
  20  *
  21  *  All code that works with directory layout had been switched to pagecache
  22  *  and moved here. AV
  23  *
  24  *  Adapted for Lustre Light
  25  *  Copyright (C) 2002-2003, Cluster File Systems, Inc.
  26  *
  27  */
  28
  29 #include <linux/fs.h>
  30 #include <linux/ext2_fs.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/mm.h>
  33 #include <linux/version.h>
  34 #include <linux/smp_lock.h>
  35 #include <asm/uaccess.h>
  36 #include <linux/buffer_head.h>   // for wait_on_buffer
  37
  38 #define DEBUG_SUBSYSTEM S_LLITE
  39
  40 #include <obd_support.h>
  41 #include <obd_class.h>
  42 #include <lustre_lib.h>
  43 #include <lustre/lustre_idl.h>
  44 #include <lustre_lite.h>
  45 #include <lustre_dlm.h>
  46 #include <lustre_fid.h>
  47 #include "llite_internal.h"
  48
  49 #ifdef HAVE_PG_FS_MISC
  50 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
  51 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
  52 #endif
  53
  54 /*
  55  * (new) readdir implementation overview.
  56  *
  57  * Original lustre readdir implementation cached exact copy of raw directory
  58  * pages on the client. These pages were indexed in client page cache by
  59  * logical offset in the directory file. This design, while very simple and
  60  * intuitive had some inherent problems:
  61  *
  62  *     . it implies that byte offset to the directory entry serves as a
  63  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  64  *     ext3/htree directory entries may move due to splits, and more
  65  *     importantly,
  66  *
  67  *     . it is incompatible with the design of split directories for cmd3,
  68  *     that assumes that names are distributed across nodes based on their
  69  *     hash, and so readdir should be done in hash order.
  70  *
  71  * New readdir implementation does readdir in hash order, and uses hash of a
  72  * file name as a telldir/seekdir cookie. This led to number of complications:
  73  *
  74  *     . hash is not unique, so it cannot be used to index cached directory
  75  *     pages on the client (note, that it requires a whole pageful of hash
  76  *     collided entries to cause two pages to have identical hashes);
  77  *
  78  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  79  *     entry cookie. ext3/htree has the same problem and lustre implementation
  80  *     mimics their solution: seekdir(hash) positions directory at the first
  81  *     entry with the given hash.
  82  *
  83  * Client side.
  84  *
  85  * 0. caching
  86  *
  87  * Client caches directory pages using hash of the first entry as an index. As
  88  * noted above hash is not unique, so this solution doesn't work as is:
  89  * special processing is needed for "page hash chains" (i.e., sequences of
  90  * pages filled with entries all having the same hash value).
  91  *
  92  * First, such chains have to be detected. To this end, server returns to the
  93  * client the hash of the first entry on the page next to one returned. When
  94  * client detects that this hash is the same as hash of the first entry on the
  95  * returned page, page hash collision has to be handled. Pages in the
  96  * hash chain, except first one, are termed "overflow pages".
  97  *
  98  * Solution to index uniqueness problem is to not cache overflow
  99  * pages. Instead, when page hash collision is detected, all overflow pages
 100  * from emerging chain are immediately requested from the server and placed in
 101  * a special data structure (struct ll_dir_chain). This data structure is used
 102  * by ll_readdir() to process entries from overflow pages. When readdir
 103  * invocation finishes, overflow pages are discarded. If page hash collision
 104  * chain weren't completely processed, next call to readdir will again detect
 105  * page hash collision, again read overflow pages in, process next portion of
 106  * entries and again discard the pages. This is not as wasteful as it looks,
 107  * because, given reasonable hash, page hash collisions are extremely rare.
 108  *
 109  * 1. directory positioning
 110  *
 111  * When seekdir(hash) is called, original
 112  *
 113  *
 114  *
 115  *
 116  *
 117  *
 118  *
 119  *
 120  * Server.
 121  *
 122  * identification of and access to overflow pages
 123  *
 124  * page format
 125  *
 126  *
 127  *
 128  *
 129  *
 130  */
 131
 132 #ifdef HAVE_PG_FS_MISC
 133 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
 134 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
 135 #endif
 136
 137 /* returns the page unlocked, but with a reference */
 138 static int ll_dir_readpage(struct file *file, struct page *page)
 139 {
 140         struct inode *inode = page->mapping->host;
 141         struct ptlrpc_request *request;
 142         struct mdt_body *body;
 143         struct obd_capa *oc;
 144         __u64 hash;
 145         int rc;
 146         ENTRY;
 147
 148         hash = hash_x_index(page->index);
 149         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) off %lu\n",
 150                inode->i_ino, inode->i_generation, inode, (unsigned long)hash);
 151
 152         oc = ll_mdscapa_get(inode);
 153         rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
 154                          oc, hash, page, &request);
 155         capa_put(oc);
 156         if (!rc) {
 157                 body = lustre_msg_buf(request->rq_repmsg, REPLY_REC_OFF,
 158                                       sizeof(*body));
 159                 /* Checked by mdc_readpage() */
 160                 LASSERT(body != NULL);
 161
 162                 /* Swabbed by mdc_readpage() */
 163                 LASSERT(lustre_rep_swabbed(request, REPLY_REC_OFF));
 164
 165                 if (body->valid & OBD_MD_FLSIZE) {
 166                         ll_inode_size_lock(inode, 0);
 167                         i_size_write(inode, body->size);
 168                         ll_inode_size_unlock(inode, 0);
 169                 }
 170                 SetPageUptodate(page);
 171         }
 172         ptlrpc_req_finished(request);
 173
 174         unlock_page(page);
 175         EXIT;
 176         return rc;
 177 }
 178
 179 struct address_space_operations ll_dir_aops = {
 180         .readpage  = ll_dir_readpage,
 181 };
 182
 183 static inline unsigned long dir_pages(struct inode *inode)
 184 {
 185         return (i_size_read(inode) + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 186 }
 187
 188 static inline unsigned ll_chunk_size(struct inode *inode)
 189 {
 190         return inode->i_sb->s_blocksize;
 191 }
 192
 193 static void ll_check_page(struct inode *dir, struct page *page)
 194 {
 195         /* XXX: check page format later */
 196         SetPageChecked(page);
 197 }
 198
 199 static inline void ll_put_page(struct page *page)
 200 {
 201         kunmap(page);
 202         page_cache_release(page);
 203 }
 204
 205 /*
 206  * Find, kmap and return page that contains given hash.
 207  */
 208 static struct page *ll_dir_page_locate(struct inode *dir, unsigned long hash,
 209                                        __u32 *start, __u32 *end)
 210 {
 211         struct address_space *mapping = dir->i_mapping;
 212         /*
 213          * Complement of hash is used as an index so that
 214          * radix_tree_gang_lookup() can be used to find a page with starting
 215          * hash _smaller_ than one we are looking for.
 216          */
 217         unsigned long offset = hash_x_index(hash);
 218         struct page *page;
 219         int found;
 220
 221         TREE_READ_LOCK_IRQ(mapping);
 222         found = radix_tree_gang_lookup(&mapping->page_tree,
 223                                        (void **)&page, offset, 1);
 224         if (found > 0) {
 225                 struct lu_dirpage *dp;
 226
 227                 page_cache_get(page);
 228                 TREE_READ_UNLOCK_IRQ(mapping);
 229                 /*
 230                  * In contrast to find_lock_page() we are sure that directory
 231                  * page cannot be truncated (while DLM lock is held) and,
 232                  * hence, can avoid restart.
 233                  *
 234                  * In fact, page cannot be locked here at all, because
 235                  * ll_dir_readpage() does synchronous io.
 236                  */
 237                 wait_on_page(page);
 238                 if (PageUptodate(page)) {
 239                         dp = kmap(page);
 240                         *start = le32_to_cpu(dp->ldp_hash_start);
 241                         *end   = le32_to_cpu(dp->ldp_hash_end);
 242                         LASSERT(*start <= hash);
 243                         if (hash > *end || (*end != *start && hash == *end)) {
 244                                 kunmap(page);
 245                                 lock_page(page);
 246                                 ll_truncate_complete_page(page);
 247                                 unlock_page(page);
 248                                 page_cache_release(page);
 249                                 page = NULL;
 250                         }
 251                 } else {
 252                         page_cache_release(page);
 253                         page = ERR_PTR(-EIO);
 254                 }
 255
 256         } else {
 257                 TREE_READ_UNLOCK_IRQ(mapping);
 258                 page = NULL;
 259         }
 260         return page;
 261 }
 262
 263 static struct page *ll_get_dir_page(struct inode *dir, __u32 hash, int exact,
 264                                     struct ll_dir_chain *chain)
 265 {
 266         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 267         struct address_space *mapping = dir->i_mapping;
 268         struct lustre_handle lockh;
 269         struct lu_dirpage *dp;
 270         struct page *page;
 271         ldlm_mode_t mode;
 272         int rc;
 273         __u32 start;
 274         __u32 end;
 275
 276         mode = LCK_PR;
 277         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 278                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
 279         if (!rc) {
 280                 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
 281                        ll_md_blocking_ast, ldlm_completion_ast, NULL, dir };
 282                 struct lookup_intent it = { .it_op = IT_READDIR };
 283                 struct ptlrpc_request *request;
 284                 struct md_op_data *op_data;
 285
 286                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
 287                                              LUSTRE_OPC_ANY, NULL);
 288                 if (IS_ERR(op_data))
 289                         return (void *)op_data;
 290
 291                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
 292                                 op_data, &lockh, NULL, 0, 0);
 293
 294                 ll_finish_md_op_data(op_data);
 295
 296                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 297                 if (request)
 298                         ptlrpc_req_finished(request);
 299                 if (rc < 0) {
 300                         CERROR("lock enqueue: rc: %d\n", rc);
 301                         return ERR_PTR(rc);
 302                 }
 303         } else {
 304                 /* for cross-ref object, l_ast_data of the lock may not be set,
 305                  * we reset it here */
 306                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie, dir);
 307         }
 308         ldlm_lock_dump_handle(D_OTHER, &lockh);
 309
 310         page = ll_dir_page_locate(dir, hash, &start, &end);
 311         if (IS_ERR(page))
 312                 GOTO(out_unlock, page);
 313
 314         if (page != NULL) {
 315                 /*
 316                  * XXX nikita: not entirely correct handling of a corner case:
 317                  * suppose hash chain of entries with hash value HASH crosses
 318                  * border between pages P0 and P1. First both P0 and P1 are
 319                  * cached, seekdir() is called for some entry from the P0 part
 320                  * of the chain. Later P0 goes out of cache. telldir(HASH)
 321                  * happens and finds P1, as it starts with matching hash
 322                  * value. Remaining entries from P0 part of the chain are
 323                  * skipped. (Is that really a bug?)
 324                  *
 325                  * Possible solutions: 0. don't cache P1 is such case, handle
 326                  * it as an "overflow" page. 1. invalidate all pages at
 327                  * once. 2. use HASH|1 as an index for P1.
 328                  */
 329                 if (exact && hash != start) {
 330                         /*
 331                          * readdir asked for a page starting _exactly_ from
 332                          * given hash, but cache contains stale page, with
 333                          * entries with smaller hash values. Stale page should
 334                          * be invalidated, and new one fetched.
 335                          */
 336                         CWARN("Stale readpage page %p: %#lx != %#lx\n", page,
 337                               (unsigned long)hash, (unsigned long)start);
 338                         lock_page(page);
 339                         ll_truncate_complete_page(page);
 340                         unlock_page(page);
 341                         page_cache_release(page);
 342                 } else
 343                         GOTO(hash_collision, page);
 344         }
 345
 346         page = read_cache_page(mapping, hash_x_index(hash),
 347                                (filler_t*)mapping->a_ops->readpage, NULL);
 348         if (IS_ERR(page))
 349                 GOTO(out_unlock, page);
 350
 351         wait_on_page(page);
 352         (void)kmap(page);
 353         if (!PageUptodate(page))
 354                 goto fail;
 355         if (!PageChecked(page))
 356                 ll_check_page(dir, page);
 357         if (PageError(page))
 358                 goto fail;
 359 hash_collision:
 360         dp = page_address(page);
 361
 362         start = le32_to_cpu(dp->ldp_hash_start);
 363         end   = le32_to_cpu(dp->ldp_hash_end);
 364         if (end == start) {
 365                 LASSERT(start == hash);
 366                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
 367                 /*
 368                  * Fetch whole overflow chain...
 369                  *
 370                  * XXX not yet.
 371                  */
 372                 goto fail;
 373         }
 374 out_unlock:
 375         ldlm_lock_decref(&lockh, mode);
 376         return page;
 377
 378 fail:
 379         ll_put_page(page);
 380         page = ERR_PTR(-EIO);
 381         goto out_unlock;
 382 }
 383
 384 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 385 {
 386         struct inode         *inode = filp->f_dentry->d_inode;
 387         struct ll_inode_info *info  = ll_i2info(inode);
 388         struct ll_sb_info    *sbi   = ll_i2sbi(inode);
 389         __u32                 pos   = filp->f_pos;
 390         struct page          *page;
 391         struct ll_dir_chain   chain;
 392         int rc;
 393         int done;
 394         int shift;
 395         ENTRY;
 396
 397         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu\n",
 398                inode->i_ino, inode->i_generation, inode,
 399                (unsigned long)pos, i_size_read(inode));
 400
 401         if (pos == DIR_END_OFF)
 402                 /*
 403                  * end-of-file.
 404                  */
 405                 RETURN(0);
 406
 407         rc    = 0;
 408         done  = 0;
 409         shift = 0;
 410         ll_dir_chain_init(&chain);
 411
 412         page = ll_get_dir_page(inode, pos, 0, &chain);
 413
 414         while (rc == 0 && !done) {
 415                 struct lu_dirpage *dp;
 416                 struct lu_dirent  *ent;
 417
 418                 if (!IS_ERR(page)) {
 419                         /*
 420                          * If page is empty (end of directoryis reached),
 421                          * use this value.
 422                          */
 423                         __u32 hash = DIR_END_OFF;
 424                         __u32 next;
 425
 426                         dp = page_address(page);
 427                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 428                              ent = lu_dirent_next(ent)) {
 429                                 char          *name;
 430                                 int            namelen;
 431                                 struct lu_fid  fid;
 432                                 ino_t          ino;
 433
 434                                 /*
 435                                  * XXX: implement correct swabbing here.
 436                                  */
 437
 438                                 hash    = le32_to_cpu(ent->lde_hash);
 439                                 namelen = le16_to_cpu(ent->lde_namelen);
 440
 441                                 if (hash < pos)
 442                                         /*
 443                                          * Skip until we find target hash
 444                                          * value.
 445                                          */
 446                                         continue;
 447
 448                                 if (namelen == 0)
 449                                         /*
 450                                          * Skip dummy record.
 451                                          */
 452                                         continue;
 453
 454                                 fid  = ent->lde_fid;
 455                                 name = ent->lde_name;
 456                                 fid_le_to_cpu(&fid, &fid);
 457                                 ino  = ll_fid_build_ino(sbi, &fid);
 458
 459                                 done = filldir(cookie, name, namelen,
 460                                                (loff_t)hash, ino, DT_UNKNOWN);
 461                         }
 462                         next = le32_to_cpu(dp->ldp_hash_end);
 463                         ll_put_page(page);
 464                         if (!done) {
 465                                 pos = next;
 466                                 if (pos == DIR_END_OFF)
 467                                         /*
 468                                          * End of directory reached.
 469                                          */
 470                                         done = 1;
 471                                 else if (1 /* chain is exhausted*/)
 472                                         /*
 473                                          * Normal case: continue to the next
 474                                          * page.
 475                                          */
 476                                         page = ll_get_dir_page(inode, pos, 1,
 477                                                                &chain);
 478                                 else {
 479                                         /*
 480                                          * go into overflow page.
 481                                          */
 482                                 }
 483                         } else
 484                                 pos = hash;
 485                 } else {
 486                         rc = PTR_ERR(page);
 487                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 488                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 489                 }
 490         }
 491
 492         filp->f_pos = (loff_t)(__s32)pos;
 493         filp->f_version = inode->i_version;
 494         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 495
 496         ll_dir_chain_fini(&chain);
 497
 498         RETURN(rc);
 499 }
 500
 501 #define QCTL_COPY(out, in)              \
 502 do {                                    \
 503         Q_COPY(out, in, qc_cmd);        \
 504         Q_COPY(out, in, qc_type);       \
 505         Q_COPY(out, in, qc_id);         \
 506         Q_COPY(out, in, qc_stat);       \
 507         Q_COPY(out, in, qc_dqinfo);     \
 508         Q_COPY(out, in, qc_dqblk);      \
 509 } while (0)
 510
 511 int ll_send_mgc_param(struct obd_export *mgc, char *string)
 512 {
 513         struct mgs_send_param *msp;
 514         int rc = 0;
 515
 516         OBD_ALLOC_PTR(msp);
 517         if (!msp)
 518                 return -ENOMEM;
 519
 520         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
 521         rc = obd_set_info_async(mgc, strlen(KEY_SET_INFO), KEY_SET_INFO,
 522                                 sizeof(struct mgs_send_param), msp, NULL);
 523         if (rc)
 524                 CERROR("Failed to set parameter: %d\n", rc);
 525         OBD_FREE_PTR(msp);
 526
 527         return rc;
 528 }
 529
 530 char *ll_get_fsname(struct inode *inode)
 531 {
 532         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 533         char *ptr, *fsname;
 534         int len;
 535
 536         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
 537         len = strlen(lsi->lsi_lmd->lmd_profile);
 538         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
 539         if (ptr && (strcmp(ptr, "-client") == 0))
 540                 len -= 7;
 541         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
 542         fsname[len] = '\0';
 543
 544         return fsname;
 545 }
 546
 547 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
 548                      int set_default)
 549 {
 550         struct ll_sb_info *sbi = ll_i2sbi(inode);
 551         struct md_op_data *op_data;
 552         struct ptlrpc_request *req = NULL;
 553         int rc = 0;
 554         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 555         struct obd_device *mgc = lsi->lsi_mgc;
 556         char *fsname = NULL, *param = NULL;
 557
 558         /*
 559          * This is coming from userspace, so should be in
 560          * local endian.  But the MDS would like it in little
 561          * endian, so we swab it before we send it.
 562          */
 563         if (lump->lmm_magic != LOV_USER_MAGIC)
 564                 RETURN(-EINVAL);
 565
 566         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC))
 567                 lustre_swab_lov_user_md(lump);
 568
 569         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 570                                      LUSTRE_OPC_ANY, NULL);
 571         if (IS_ERR(op_data))
 572                 RETURN(PTR_ERR(op_data));
 573
 574         /* swabbing is done in lov_setstripe() on server side */
 575         rc = md_setattr(sbi->ll_md_exp, op_data, lump, sizeof(*lump),
 576                         NULL, 0, &req, NULL);
 577         ll_finish_md_op_data(op_data);
 578         ptlrpc_req_finished(req);
 579         if (rc) {
 580                 if (rc != -EPERM && rc != -EACCES)
 581                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 582         }
 583
 584         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
 585                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
 586
 587                 /* Get fsname and assume devname to be -MDT0000. */
 588                 fsname = ll_get_fsname(inode);
 589                 /* Set root stripesize */
 590                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
 591                         lump->lmm_stripe_size);
 592                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 593                 if (rc)
 594                         goto end;
 595
 596                 /* Set root stripecount */
 597                 sprintf(param, "%s-MDT0000.lov.stripecount=%u", fsname,
 598                         lump->lmm_stripe_count);
 599                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 600                 if (rc)
 601                         goto end;
 602
 603                 /* Set root stripeoffset */
 604                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%u", fsname,
 605                         lump->lmm_stripe_offset);
 606                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 607                 if (rc)
 608                         goto end;
 609 end:
 610                 if (fsname)
 611                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
 612                 if (param)
 613                         OBD_FREE(param, MGS_PARAM_MAXLEN);
 614         }
 615         return rc;
 616 }
 617
 618 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 619                      int *lmm_size, struct ptlrpc_request **request)
 620 {
 621         struct ll_sb_info *sbi = ll_i2sbi(inode);
 622         struct mdt_body   *body;
 623         struct lov_mds_md *lmm = NULL;
 624         struct ptlrpc_request *req = NULL;
 625         int rc, lmmsize;
 626         struct obd_capa *oc;
 627
 628         rc = ll_get_max_mdsize(sbi, &lmmsize);
 629         if (rc)
 630                 RETURN(rc);
 631
 632         oc = ll_mdscapa_get(inode);
 633         rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode),
 634                         oc, OBD_MD_FLEASIZE | OBD_MD_FLDIREA,
 635                         lmmsize, &req);
 636         capa_put(oc);
 637         if (rc < 0) {
 638                 CDEBUG(D_INFO, "md_getattr failed on inode "
 639                        "%lu/%u: rc %d\n", inode->i_ino,
 640                        inode->i_generation, rc);
 641                 GOTO(out, rc);
 642         }
 643
 644         body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
 645         LASSERT(body != NULL); /* checked by md_getattr_name */
 646         /* swabbed by mdc_getattr_name */
 647         LASSERT(lustre_rep_swabbed(req, REPLY_REC_OFF));
 648
 649         lmmsize = body->eadatasize;
 650
 651         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 652             lmmsize == 0) {
 653                 GOTO(out, rc = -ENODATA);
 654         }
 655
 656         lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, lmmsize);
 657         LASSERT(lmm != NULL);
 658         LASSERT(lustre_rep_swabbed(req, REPLY_REC_OFF + 1));
 659
 660         /*
 661          * This is coming from the MDS, so is probably in
 662          * little endian.  We convert it to host endian before
 663          * passing it to userspace.
 664          */
 665         if (lmm->lmm_magic == __swab32(LOV_MAGIC)) {
 666                 lustre_swab_lov_user_md((struct lov_user_md *)lmm);
 667                 lustre_swab_lov_user_md_objects((struct lov_user_md *)lmm);
 668         }
 669 out:
 670         *lmmp = lmm;
 671         *lmm_size = lmmsize;
 672         *request = req;
 673         return rc;
 674 }
 675
 676 static int ll_dir_ioctl(struct inode *inode, struct file *file,
 677                         unsigned int cmd, unsigned long arg)
 678 {
 679         struct ll_sb_info *sbi = ll_i2sbi(inode);
 680         struct obd_ioctl_data *data;
 681         ENTRY;
 682
 683         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
 684                inode->i_ino, inode->i_generation, inode, cmd);
 685
 686         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
 687         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
 688                 return -ENOTTY;
 689
 690         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
 691         switch(cmd) {
 692         case EXT3_IOC_GETFLAGS:
 693         case EXT3_IOC_SETFLAGS:
 694                 RETURN(ll_iocontrol(inode, file, cmd, arg));
 695         case EXT3_IOC_GETVERSION_OLD:
 696         case EXT3_IOC_GETVERSION:
 697                 RETURN(put_user(inode->i_generation, (int *)arg));
 698         /* We need to special case any other ioctls we want to handle,
 699          * to send them to the MDS/OST as appropriate and to properly
 700          * network encode the arg field.
 701         case EXT3_IOC_SETVERSION_OLD:
 702         case EXT3_IOC_SETVERSION:
 703         */
 704         case IOC_MDC_LOOKUP: {
 705                 struct ptlrpc_request *request = NULL;
 706                 int namelen, rc, len = 0;
 707                 char *buf = NULL;
 708                 char *filename;
 709                 struct obd_capa *oc;
 710
 711                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 712                 if (rc)
 713                         RETURN(rc);
 714                 data = (void *)buf;
 715
 716                 filename = data->ioc_inlbuf1;
 717                 namelen = data->ioc_inllen1;
 718
 719                 if (namelen < 1) {
 720                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
 721                         GOTO(out, rc = -EINVAL);
 722                 }
 723
 724                 oc = ll_mdscapa_get(inode);
 725                 rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode), oc,
 726                                      filename, namelen, OBD_MD_FLID, 0,
 727                                      &request);
 728                 capa_put(oc);
 729                 if (rc < 0) {
 730                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 731                         GOTO(out, rc);
 732                 }
 733
 734                 ptlrpc_req_finished(request);
 735
 736                 EXIT;
 737         out:
 738                 obd_ioctl_freedata(buf, len);
 739                 return rc;
 740         }
 741         case LL_IOC_LOV_SETSTRIPE: {
 742                 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
 743                 int rc = 0;
 744                 int set_default = 0;
 745
 746                 LASSERT(sizeof(lum) == sizeof(*lump));
 747                 LASSERT(sizeof(lum.lmm_objects[0]) ==
 748                         sizeof(lump->lmm_objects[0]));
 749                 rc = copy_from_user(&lum, lump, sizeof(lum));
 750                 if (rc)
 751                         RETURN(-EFAULT);
 752
 753                 if (inode->i_sb->s_root == file->f_dentry)
 754                         set_default = 1;
 755
 756                 rc = ll_dir_setstripe(inode, &lum, set_default);
 757
 758                 RETURN(rc);
 759         }
 760         case LL_IOC_OBD_STATFS:
 761                 RETURN(ll_obd_statfs(inode, (void *)arg));
 762         case LL_IOC_LOV_GETSTRIPE:
 763         case LL_IOC_MDC_GETINFO:
 764         case IOC_MDC_GETFILEINFO:
 765         case IOC_MDC_GETFILESTRIPE: {
 766                 struct ptlrpc_request *request = NULL;
 767                 struct lov_user_md *lump;
 768                 struct lov_mds_md *lmm = NULL;
 769                 struct mdt_body *body;
 770                 char *filename = NULL;
 771                 int rc, lmmsize;
 772
 773                 if (cmd == IOC_MDC_GETFILEINFO ||
 774                     cmd == IOC_MDC_GETFILESTRIPE) {
 775                         filename = getname((const char *)arg);
 776                         if (IS_ERR(filename))
 777                                 RETURN(PTR_ERR(filename));
 778
 779                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
 780                                                       &lmmsize, &request);
 781                 } else {
 782                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
 783                 }
 784
 785                 if (request) {
 786                         body = lustre_msg_buf(request->rq_repmsg,
 787                                               REPLY_REC_OFF, sizeof(*body));
 788                         LASSERT(body != NULL); /* checked by md_getattr_name */
 789                         /* swabbed by md_getattr_name */
 790                         LASSERT(lustre_rep_swabbed(request, REPLY_REC_OFF));
 791                 } else {
 792                         GOTO(out_req, rc);
 793                 }
 794
 795                 if (rc < 0) {
 796                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
 797                                                cmd == LL_IOC_MDC_GETINFO))
 798                                 GOTO(skip_lmm, rc = 0);
 799                         else
 800                                 GOTO(out_req, rc);
 801                 }
 802
 803                 if (cmd == IOC_MDC_GETFILESTRIPE ||
 804                     cmd == LL_IOC_LOV_GETSTRIPE) {
 805                         lump = (struct lov_user_md *)arg;
 806                 } else {
 807                         struct lov_user_mds_data *lmdp;
 808                         lmdp = (struct lov_user_mds_data *)arg;
 809                         lump = &lmdp->lmd_lmm;
 810                 }
 811                 rc = copy_to_user(lump, lmm, lmmsize);
 812                 if (rc)
 813                         GOTO(out_lmm, rc = -EFAULT);
 814         skip_lmm:
 815                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
 816                         struct lov_user_mds_data *lmdp;
 817                         lstat_t st = { 0 };
 818
 819                         st.st_dev     = inode->i_sb->s_dev;
 820                         st.st_mode    = body->mode;
 821                         st.st_nlink   = body->nlink;
 822                         st.st_uid     = body->uid;
 823                         st.st_gid     = body->gid;
 824                         st.st_rdev    = body->rdev;
 825                         st.st_size    = body->size;
 826                         st.st_blksize = CFS_PAGE_SIZE;
 827                         st.st_blocks  = body->blocks;
 828                         st.st_atime   = body->atime;
 829                         st.st_mtime   = body->mtime;
 830                         st.st_ctime   = body->ctime;
 831                         st.st_ino     = inode->i_ino;
 832
 833                         lmdp = (struct lov_user_mds_data *)arg;
 834                         rc = copy_to_user(&lmdp->lmd_st, &st, sizeof(st));
 835                         if (rc)
 836                                 GOTO(out_lmm, rc = -EFAULT);
 837                 }
 838
 839                 EXIT;
 840         out_lmm:
 841                 if (lmm && lmm->lmm_magic == LOV_MAGIC_JOIN)
 842                         OBD_FREE(lmm, lmmsize);
 843         out_req:
 844                 ptlrpc_req_finished(request);
 845                 if (filename)
 846                         putname(filename);
 847                 return rc;
 848         }
 849         case IOC_LOV_GETINFO: {
 850                 struct lov_user_mds_data *lumd;
 851                 struct lov_stripe_md *lsm;
 852                 struct lov_user_md *lum;
 853                 struct lov_mds_md *lmm;
 854                 int lmmsize;
 855                 lstat_t st;
 856                 int rc;
 857
 858                 lumd = (struct lov_user_mds_data *)arg;
 859                 lum = &lumd->lmd_lmm;
 860
 861                 rc = ll_get_max_mdsize(sbi, &lmmsize);
 862                 if (rc)
 863                         RETURN(rc);
 864
 865                 OBD_ALLOC(lmm, lmmsize);
 866                 rc = copy_from_user(lmm, lum, lmmsize);
 867                 if (rc)
 868                         GOTO(free_lmm, rc = -EFAULT);
 869
 870                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
 871                 if (rc < 0)
 872                         GOTO(free_lmm, rc = -ENOMEM);
 873
 874                 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
 875                 if (rc)
 876                         GOTO(free_lsm, rc);
 877
 878                 /* Perform glimpse_size operation. */
 879                 memset(&st, 0, sizeof(st));
 880
 881                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
 882                 if (rc)
 883                         GOTO(free_lsm, rc);
 884
 885                 rc = copy_to_user(&lumd->lmd_st, &st, sizeof(st));
 886                 if (rc)
 887                         GOTO(free_lsm, rc = -EFAULT);
 888
 889                 EXIT;
 890         free_lsm:
 891                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
 892         free_lmm:
 893                 OBD_FREE(lmm, lmmsize);
 894                 return rc;
 895         }
 896         case OBD_IOC_LLOG_CATINFO: {
 897                 struct ptlrpc_request *req = NULL;
 898                 char *buf = NULL;
 899                 int rc, len = 0;
 900                 char *bufs[3] = { NULL }, *str;
 901                 int lens[3] = { sizeof(struct ptlrpc_body) };
 902                 int size[2] = { sizeof(struct ptlrpc_body) };
 903
 904                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 905                 if (rc)
 906                         RETURN(rc);
 907                 data = (void *)buf;
 908
 909                 if (!data->ioc_inlbuf1) {
 910                         obd_ioctl_freedata(buf, len);
 911                         RETURN(-EINVAL);
 912                 }
 913
 914                 lens[REQ_REC_OFF] = data->ioc_inllen1;
 915                 bufs[REQ_REC_OFF] = data->ioc_inlbuf1;
 916                 if (data->ioc_inllen2) {
 917                         lens[REQ_REC_OFF + 1] = data->ioc_inllen2;
 918                         bufs[REQ_REC_OFF + 1] = data->ioc_inlbuf2;
 919                 } else {
 920                         lens[REQ_REC_OFF + 1] = 0;
 921                         bufs[REQ_REC_OFF + 1] = NULL;
 922                 }
 923
 924                 req = ptlrpc_prep_req(sbi2mdc(sbi)->cl_import,
 925                                       LUSTRE_LOG_VERSION, LLOG_CATINFO, 3, lens,
 926                                       bufs);
 927                 if (!req)
 928                         GOTO(out_catinfo, rc = -ENOMEM);
 929
 930                 size[REPLY_REC_OFF] = data->ioc_plen1;
 931                 ptlrpc_req_set_repsize(req, 2, size);
 932
 933                 rc = ptlrpc_queue_wait(req);
 934                 if (!rc) {
 935                         str = lustre_msg_string(req->rq_repmsg, REPLY_REC_OFF,
 936                                                 data->ioc_plen1);
 937                         rc = copy_to_user(data->ioc_pbuf1, str, data->ioc_plen1);
 938                 }
 939                 ptlrpc_req_finished(req);
 940         out_catinfo:
 941                 obd_ioctl_freedata(buf, len);
 942                 RETURN(rc);
 943         }
 944         case OBD_IOC_QUOTACHECK: {
 945                 struct obd_quotactl *oqctl;
 946                 int rc, error = 0;
 947
 948                 if (!capable(CAP_SYS_ADMIN))
 949                         RETURN(-EPERM);
 950
 951                 OBD_ALLOC_PTR(oqctl);
 952                 if (!oqctl)
 953                         RETURN(-ENOMEM);
 954                 oqctl->qc_type = arg;
 955                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
 956                 if (rc < 0) {
 957                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
 958                         error = rc;
 959                 }
 960
 961                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
 962                 if (rc < 0)
 963                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
 964
 965                 OBD_FREE_PTR(oqctl);
 966                 return error ?: rc;
 967         }
 968         case OBD_IOC_POLL_QUOTACHECK: {
 969                 struct if_quotacheck *check;
 970                 int rc;
 971
 972                 if (!capable(CAP_SYS_ADMIN))
 973                         RETURN(-EPERM);
 974
 975                 OBD_ALLOC_PTR(check);
 976                 if (!check)
 977                         RETURN(-ENOMEM);
 978
 979                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
 980                                    NULL);
 981                 if (rc) {
 982                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
 983                         if (copy_to_user((void *)arg, check, sizeof(*check)))
 984                                 rc = -EFAULT;
 985                         GOTO(out_poll, rc);
 986                 }
 987
 988                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
 989                                    NULL);
 990                 if (rc) {
 991                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
 992                         if (copy_to_user((void *)arg, check, sizeof(*check)))
 993                                 rc = -EFAULT;
 994                         GOTO(out_poll, rc);
 995                 }
 996         out_poll:
 997                 OBD_FREE_PTR(check);
 998                 RETURN(rc);
 999         }
1000 #ifdef HAVE_QUOTA_SUPPORT
1001         case OBD_IOC_QUOTACTL: {
1002                 struct if_quotactl *qctl;
1003                 struct obd_quotactl *oqctl;
1004
1005                 int cmd, type, id, rc = 0;
1006
1007                 OBD_ALLOC_PTR(qctl);
1008                 if (!qctl)
1009                         RETURN(-ENOMEM);
1010
1011                 OBD_ALLOC_PTR(oqctl);
1012                 if (!oqctl) {
1013                         OBD_FREE_PTR(qctl);
1014                         RETURN(-ENOMEM);
1015                 }
1016                 if (copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1017                         GOTO(out_quotactl, rc = -EFAULT);
1018
1019                 cmd = qctl->qc_cmd;
1020                 type = qctl->qc_type;
1021                 id = qctl->qc_id;
1022                 switch (cmd) {
1023                 case Q_QUOTAON:
1024                 case Q_QUOTAOFF:
1025                 case Q_SETQUOTA:
1026                 case Q_SETINFO:
1027                         if (!capable(CAP_SYS_ADMIN))
1028                                 GOTO(out_quotactl, rc = -EPERM);
1029                         break;
1030                 case Q_GETQUOTA:
1031                         if (((type == USRQUOTA && current->euid != id) ||
1032                              (type == GRPQUOTA && !in_egroup_p(id))) &&
1033                             !capable(CAP_SYS_ADMIN))
1034                                 GOTO(out_quotactl, rc = -EPERM);
1035
1036                         /* XXX: dqb_valid is borrowed as a flag to mark that
1037                          *      only mds quota is wanted */
1038                         if (qctl->qc_dqblk.dqb_valid)
1039                                 qctl->obd_uuid = sbi->ll_md_exp->exp_obd->
1040                                                         u.cli.cl_target_uuid;
1041                         break;
1042                 case Q_GETINFO:
1043                         break;
1044                 default:
1045                         CERROR("unsupported quotactl op: %#x\n", cmd);
1046                         GOTO(out_quotactl, rc = -ENOTTY);
1047                 }
1048
1049                 QCTL_COPY(oqctl, qctl);
1050
1051                 if (qctl->obd_uuid.uuid[0]) {
1052                         struct obd_device *obd;
1053                         struct obd_uuid *uuid = &qctl->obd_uuid;
1054
1055                         obd = class_find_client_notype(uuid,
1056                                          &sbi->ll_dt_exp->exp_obd->obd_uuid);
1057                         if (!obd)
1058                                 GOTO(out_quotactl, rc = -ENOENT);
1059
1060                         if (cmd == Q_GETINFO)
1061                                 oqctl->qc_cmd = Q_GETOINFO;
1062                         else if (cmd == Q_GETQUOTA)
1063                                 oqctl->qc_cmd = Q_GETOQUOTA;
1064                         else
1065                                 GOTO(out_quotactl, rc = -EINVAL);
1066
1067                         if (sbi->ll_md_exp->exp_obd == obd) {
1068                                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1069                         } else {
1070                                 int i;
1071                                 struct obd_export *exp;
1072                                 struct lov_obd *lov = &sbi->ll_dt_exp->
1073                                                             exp_obd->u.lov;
1074
1075                                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1076                                         if (!lov->lov_tgts[i] ||
1077                                             !lov->lov_tgts[i]->ltd_active)
1078                                                 continue;
1079                                         exp = lov->lov_tgts[i]->ltd_exp;
1080                                         if (exp->exp_obd == obd) {
1081                                                 rc = obd_quotactl(exp, oqctl);
1082                                                 break;
1083                                         }
1084                                 }
1085                         }
1086
1087                         oqctl->qc_cmd = cmd;
1088                         QCTL_COPY(qctl, oqctl);
1089
1090                         if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1091                                 rc = -EFAULT;
1092
1093                         GOTO(out_quotactl, rc);
1094                 }
1095
1096                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1097                 if (rc && rc != -EBUSY && cmd == Q_QUOTAON) {
1098                         oqctl->qc_cmd = Q_QUOTAOFF;
1099                         obd_quotactl(sbi->ll_md_exp, oqctl);
1100                 }
1101
1102                 QCTL_COPY(qctl, oqctl);
1103
1104                 if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1105                         rc = -EFAULT;
1106         out_quotactl:
1107                 OBD_FREE_PTR(qctl);
1108                 OBD_FREE_PTR(oqctl);
1109                 RETURN(rc);
1110         }
1111 #endif /* HAVE_QUOTA_SUPPORT */
1112         case OBD_IOC_GETNAME: {
1113                 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1114                 if (!obd)
1115                         RETURN(-EFAULT);
1116                 if (copy_to_user((void *)arg, obd->obd_name,
1117                                 strlen(obd->obd_name) + 1))
1118                         RETURN (-EFAULT);
1119                 RETURN(0);
1120         }
1121         case LL_IOC_FLUSHCTX:
1122                 RETURN(ll_flush_ctx(inode));
1123         case LL_IOC_GETFACL: {
1124                 struct rmtacl_ioctl_data ioc;
1125
1126                 if (copy_from_user(&ioc, (void *)arg, sizeof(ioc)))
1127                         RETURN(-EFAULT);
1128
1129                 RETURN(ll_ioctl_getfacl(inode, &ioc));
1130         }
1131         case LL_IOC_SETFACL: {
1132                 struct rmtacl_ioctl_data ioc;
1133
1134                 if (copy_from_user(&ioc, (void *)arg, sizeof(ioc)))
1135                         RETURN(-EFAULT);
1136
1137                 RETURN(ll_ioctl_setfacl(inode, &ioc));
1138         }
1139         default:
1140                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1141         }
1142 }
1143
1144 int ll_dir_open(struct inode *inode, struct file *file)
1145 {
1146         ENTRY;
1147         RETURN(ll_file_open(inode, file));
1148 }
1149
1150 int ll_dir_release(struct inode *inode, struct file *file)
1151 {
1152         ENTRY;
1153         RETURN(ll_file_release(inode, file));
1154 }
1155
1156 struct file_operations ll_dir_operations = {
1157         .open     = ll_dir_open,
1158         .release  = ll_dir_release,
1159         .read     = generic_read_dir,
1160         .readdir  = ll_readdir,
1161         .ioctl    = ll_dir_ioctl
1162 };
1163