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 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  37 # include <linux/locks.h>   // for wait_on_buffer
  38 #else
  39 # include <linux/buffer_head.h>   // for wait_on_buffer
  40 #endif
  41
  42 #define DEBUG_SUBSYSTEM S_LLITE
  43
  44 #include <obd_support.h>
  45 #include <obd_class.h>
  46 #include <lustre_lib.h>
  47 #include <lustre/lustre_idl.h>
  48 #include <lustre_lite.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_fid.h>
  51 #include "llite_internal.h"
  52
  53 #ifdef HAVE_PG_FS_MISC
  54 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
  55 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
  56 #endif
  57
  58 /*
  59  * (new) readdir implementation overview.
  60  *
  61  * Original lustre readdir implementation cached exact copy of raw directory
  62  * pages on the client. These pages were indexed in client page cache by
  63  * logical offset in the directory file. This design, while very simple and
  64  * intuitive had some inherent problems:
  65  *
  66  *     . it implies that byte offset to the directory entry serves as a
  67  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  68  *     ext3/htree directory entries may move due to splits, and more
  69  *     importantly,
  70  *
  71  *     . it is incompatible with the design of split directories for cmd3,
  72  *     that assumes that names are distributed across nodes based on their
  73  *     hash, and so readdir should be done in hash order.
  74  *
  75  * New readdir implementation does readdir in hash order, and uses hash of a
  76  * file name as a telldir/seekdir cookie. This led to number of complications:
  77  *
  78  *     . hash is not unique, so it cannot be used to index cached directory
  79  *     pages on the client (note, that it requires a whole pageful of hash
  80  *     collided entries to cause two pages to have identical hashes);
  81  *
  82  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  83  *     entry cookie. ext3/htree has the same problem and lustre implementation
  84  *     mimics their solution: seekdir(hash) positions directory at the first
  85  *     entry with the given hash.
  86  *
  87  * Client side.
  88  *
  89  * 0. caching
  90  *
  91  * Client caches directory pages using hash of the first entry as an index. As
  92  * noted above hash is not unique, so this solution doesn't work as is:
  93  * special processing is needed for "page hash chains" (i.e., sequences of
  94  * pages filled with entries all having the same hash value).
  95  *
  96  * First, such chains have to be detected. To this end, server returns to the
  97  * client the hash of the first entry on the page next to one returned. When
  98  * client detects that this hash is the same as hash of the first entry on the
  99  * returned page, page hash collision has to be handled. Pages in the
 100  * hash chain, except first one, are termed "overflow pages".
 101  *
 102  * Solution to index uniqueness problem is to not cache overflow
 103  * pages. Instead, when page hash collision is detected, all overflow pages
 104  * from emerging chain are immediately requested from the server and placed in
 105  * a special data structure (struct ll_dir_chain). This data structure is used
 106  * by ll_readdir() to process entries from overflow pages. When readdir
 107  * invocation finishes, overflow pages are discarded. If page hash collision
 108  * chain weren't completely processed, next call to readdir will again detect
 109  * page hash collision, again read overflow pages in, process next portion of
 110  * entries and again discard the pages. This is not as wasteful as it looks,
 111  * because, given reasonable hash, page hash collisions are extremely rare.
 112  *
 113  * 1. directory positioning
 114  *
 115  * When seekdir(hash) is called, original
 116  *
 117  *
 118  *
 119  *
 120  *
 121  *
 122  *
 123  *
 124  * Server.
 125  *
 126  * identification of and access to overflow pages
 127  *
 128  * page format
 129  *
 130  *
 131  *
 132  *
 133  *
 134  */
 135
 136 static __u32 hash_x_index(__u32 value)
 137 {
 138         return ((__u32)~0) - value;
 139 }
 140 #ifdef HAVE_PG_FS_MISC
 141 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
 142 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
 143 #endif
 144
 145 /* returns the page unlocked, but with a reference */
 146 static int ll_dir_readpage(struct file *file, struct page *page)
 147 {
 148         struct inode *inode = page->mapping->host;
 149         struct ptlrpc_request *request;
 150         struct mdt_body *body;
 151         struct obd_capa *oc;
 152         __u64 hash;
 153         int rc;
 154         ENTRY;
 155
 156         hash = hash_x_index(page->index);
 157         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) off %lu\n",
 158                inode->i_ino, inode->i_generation, inode, (unsigned long)hash);
 159
 160         oc = ll_mdscapa_get(inode);
 161         rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
 162                          oc, hash, page, &request);
 163         capa_put(oc);
 164         if (!rc) {
 165                 body = lustre_msg_buf(request->rq_repmsg, REPLY_REC_OFF,
 166                                       sizeof(*body));
 167                 /* Checked by mdc_readpage() */
 168                 LASSERT(body != NULL);
 169
 170                 /* Swabbed by mdc_readpage() */
 171                 LASSERT_REPSWABBED(request, REPLY_REC_OFF);
 172
 173                 if (body->valid & OBD_MD_FLSIZE)
 174                         i_size_write(inode, body->size);
 175                 SetPageUptodate(page);
 176         }
 177         ptlrpc_req_finished(request);
 178
 179         unlock_page(page);
 180         EXIT;
 181         return rc;
 182 }
 183
 184 struct address_space_operations ll_dir_aops = {
 185         .readpage  = ll_dir_readpage,
 186 };
 187
 188 static inline unsigned long dir_pages(struct inode *inode)
 189 {
 190         return (i_size_read(inode) + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 191 }
 192
 193 static inline unsigned ll_chunk_size(struct inode *inode)
 194 {
 195         return inode->i_sb->s_blocksize;
 196 }
 197
 198 static void ll_check_page(struct inode *dir, struct page *page)
 199 {
 200         /* XXX: check page format later */
 201         SetPageChecked(page);
 202 }
 203
 204 static inline void ll_put_page(struct page *page)
 205 {
 206         kunmap(page);
 207         page_cache_release(page);
 208 }
 209
 210 /*
 211  * Find, kmap and return page that contains given hash.
 212  */
 213 static struct page *ll_dir_page_locate(struct inode *dir, unsigned long hash,
 214                                        __u32 *start, __u32 *end)
 215 {
 216         struct address_space *mapping = dir->i_mapping;
 217         /*
 218          * Complement of hash is used as an index so that
 219          * radix_tree_gang_lookup() can be used to find a page with starting
 220          * hash _smaller_ than one we are looking for.
 221          */
 222         unsigned long offset = hash_x_index(hash);
 223         struct page *page;
 224         int found;
 225
 226         spin_lock_irq(&mapping->tree_lock);
 227         found = radix_tree_gang_lookup(&mapping->page_tree,
 228                                        (void **)&page, offset, 1);
 229         if (found > 0) {
 230                 struct lu_dirpage *dp;
 231
 232                 page_cache_get(page);
 233                 spin_unlock_irq(&mapping->tree_lock);
 234                 /*
 235                  * In contrast to find_lock_page() we are sure that directory
 236                  * page cannot be truncated (while DLM lock is held) and,
 237                  * hence, can avoid restart.
 238                  *
 239                  * In fact, page cannot be locked here at all, because
 240                  * ll_dir_readpage() does synchronous io.
 241                  */
 242                 wait_on_page(page);
 243                 if (PageUptodate(page)) {
 244                         dp = kmap(page);
 245                         *start = le32_to_cpu(dp->ldp_hash_start);
 246                         *end   = le32_to_cpu(dp->ldp_hash_end);
 247                         LASSERT(*start <= hash);
 248                         if (hash > *end || (*end != *start && hash == *end)) {
 249                                 kunmap(page);
 250                                 lock_page(page);
 251                                 ll_truncate_complete_page(page);
 252                                 unlock_page(page);
 253                                 page_cache_release(page);
 254                                 page = NULL;
 255                         }
 256                 } else {
 257                         page_cache_release(page);
 258                         page = ERR_PTR(-EIO);
 259                 }
 260
 261         } else {
 262                 spin_unlock_irq(&mapping->tree_lock);
 263                 page = NULL;
 264         }
 265         return page;
 266 }
 267
 268 /*
 269  * Chain of hash overflow pages.
 270  */
 271 struct ll_dir_chain {
 272         /* XXX something. Later */
 273 };
 274
 275 static void ll_dir_chain_init(struct ll_dir_chain *chain)
 276 {
 277 }
 278
 279 static void ll_dir_chain_fini(struct ll_dir_chain *chain)
 280 {
 281 }
 282
 283 static struct page *ll_get_dir_page(struct inode *dir, __u32 hash, int exact,
 284                                     struct ll_dir_chain *chain)
 285 {
 286         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 287         struct address_space *mapping = dir->i_mapping;
 288         struct lustre_handle lockh;
 289         struct lu_dirpage *dp;
 290         struct page *page;
 291         ldlm_mode_t mode;
 292         int rc;
 293         __u32 start;
 294         __u32 end;
 295
 296         mode = LCK_PR;
 297         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 298                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
 299         if (!rc) {
 300                 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
 301                        ll_md_blocking_ast, ldlm_completion_ast, NULL, dir };
 302                 struct lookup_intent it = { .it_op = IT_READDIR };
 303                 struct ptlrpc_request *request;
 304                 struct md_op_data *op_data;
 305
 306                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
 307                                              LUSTRE_OPC_ANY, NULL);
 308                 if (IS_ERR(op_data))
 309                         return (void *)op_data;
 310
 311                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
 312                                 op_data, &lockh, NULL, 0, 0);
 313
 314                 ll_finish_md_op_data(op_data);
 315
 316                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 317                 if (request)
 318                         ptlrpc_req_finished(request);
 319                 if (rc < 0) {
 320                         CERROR("lock enqueue: rc: %d\n", rc);
 321                         return ERR_PTR(rc);
 322                 }
 323         } else {
 324                 /* for cross-ref object, l_ast_data of the lock may not be set,
 325                  * we reset it here */
 326                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie, dir);
 327         }
 328         ldlm_lock_dump_handle(D_OTHER, &lockh);
 329
 330         page = ll_dir_page_locate(dir, hash, &start, &end);
 331         if (IS_ERR(page))
 332                 GOTO(out_unlock, page);
 333
 334         if (page != NULL) {
 335                 /*
 336                  * XXX nikita: not entirely correct handling of a corner case:
 337                  * suppose hash chain of entries with hash value HASH crosses
 338                  * border between pages P0 and P1. First both P0 and P1 are
 339                  * cached, seekdir() is called for some entry from the P0 part
 340                  * of the chain. Later P0 goes out of cache. telldir(HASH)
 341                  * happens and finds P1, as it starts with matching hash
 342                  * value. Remaining entries from P0 part of the chain are
 343                  * skipped. (Is that really a bug?)
 344                  *
 345                  * Possible solutions: 0. don't cache P1 is such case, handle
 346                  * it as an "overflow" page. 1. invalidate all pages at
 347                  * once. 2. use HASH|1 as an index for P1.
 348                  */
 349                 if (exact && hash != start) {
 350                         /*
 351                          * readdir asked for a page starting _exactly_ from
 352                          * given hash, but cache contains stale page, with
 353                          * entries with smaller hash values. Stale page should
 354                          * be invalidated, and new one fetched.
 355                          */
 356                         CWARN("Stale readpage page %p: %#lx != %#lx\n", page,
 357                               (unsigned long)hash, (unsigned long)start);
 358                         lock_page(page);
 359                         ll_truncate_complete_page(page);
 360                         unlock_page(page);
 361                         page_cache_release(page);
 362                 } else
 363                         GOTO(hash_collision, page);
 364         }
 365
 366         page = read_cache_page(mapping, hash_x_index(hash),
 367                                (filler_t*)mapping->a_ops->readpage, NULL);
 368         if (IS_ERR(page))
 369                 GOTO(out_unlock, page);
 370
 371         wait_on_page(page);
 372         (void)kmap(page);
 373         if (!PageUptodate(page))
 374                 goto fail;
 375         if (!PageChecked(page))
 376                 ll_check_page(dir, page);
 377         if (PageError(page))
 378                 goto fail;
 379 hash_collision:
 380         dp = page_address(page);
 381
 382         start = le32_to_cpu(dp->ldp_hash_start);
 383         end   = le32_to_cpu(dp->ldp_hash_end);
 384         if (end == start) {
 385                 LASSERT(start == hash);
 386                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
 387                 /*
 388                  * Fetch whole overflow chain...
 389                  *
 390                  * XXX not yet.
 391                  */
 392                 goto fail;
 393         }
 394 out_unlock:
 395         ldlm_lock_decref(&lockh, mode);
 396         return page;
 397
 398 fail:
 399         ll_put_page(page);
 400         page = ERR_PTR(-EIO);
 401         goto out_unlock;
 402 }
 403
 404 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 405 {
 406         struct inode         *inode = filp->f_dentry->d_inode;
 407         struct ll_inode_info *info  = ll_i2info(inode);
 408         struct ll_sb_info    *sbi   = ll_i2sbi(inode);
 409         __u32                 pos   = filp->f_pos;
 410         struct page          *page;
 411         struct ll_dir_chain   chain;
 412         int rc;
 413         int done;
 414         int shift;
 415         ENTRY;
 416
 417         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu\n",
 418                inode->i_ino, inode->i_generation, inode,
 419                (unsigned long)pos, i_size_read(inode));
 420
 421         if (pos == DIR_END_OFF)
 422                 /*
 423                  * end-of-file.
 424                  */
 425                 RETURN(0);
 426
 427         rc    = 0;
 428         done  = 0;
 429         shift = 0;
 430         ll_dir_chain_init(&chain);
 431
 432         page = ll_get_dir_page(inode, pos, 0, &chain);
 433
 434         while (rc == 0 && !done) {
 435                 struct lu_dirpage *dp;
 436                 struct lu_dirent  *ent;
 437
 438                 if (!IS_ERR(page)) {
 439                         /*
 440                          * If page is empty (end of directoryis reached),
 441                          * use this value.
 442                          */
 443                         __u32 hash = DIR_END_OFF;
 444                         __u32 next;
 445
 446                         dp = page_address(page);
 447                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 448                              ent = lu_dirent_next(ent)) {
 449                                 char          *name;
 450                                 int            namelen;
 451                                 struct lu_fid  fid;
 452                                 ino_t          ino;
 453
 454                                 /*
 455                                  * XXX: implement correct swabbing here.
 456                                  */
 457
 458                                 hash    = le32_to_cpu(ent->lde_hash);
 459                                 namelen = le16_to_cpu(ent->lde_namelen);
 460
 461                                 if (hash < pos)
 462                                         /*
 463                                          * Skip until we find target hash
 464                                          * value.
 465                                          */
 466                                         continue;
 467
 468                                 if (namelen == 0)
 469                                         /*
 470                                          * Skip dummy record.
 471                                          */
 472                                         continue;
 473
 474                                 fid  = ent->lde_fid;
 475                                 name = ent->lde_name;
 476                                 fid_le_to_cpu(&fid, &fid);
 477                                 ino  = ll_fid_build_ino(sbi, &fid);
 478
 479                                 done = filldir(cookie, name, namelen,
 480                                                (loff_t)hash, ino, DT_UNKNOWN);
 481                         }
 482                         next = le32_to_cpu(dp->ldp_hash_end);
 483                         ll_put_page(page);
 484                         if (!done) {
 485                                 pos = next;
 486                                 if (pos == DIR_END_OFF)
 487                                         /*
 488                                          * End of directory reached.
 489                                          */
 490                                         done = 1;
 491                                 else if (1 /* chain is exhausted*/)
 492                                         /*
 493                                          * Normal case: continue to the next
 494                                          * page.
 495                                          */
 496                                         page = ll_get_dir_page(inode, pos, 1,
 497                                                                &chain);
 498                                 else {
 499                                         /*
 500                                          * go into overflow page.
 501                                          */
 502                                 }
 503                         } else
 504                                 pos = hash;
 505                 } else {
 506                         rc = PTR_ERR(page);
 507                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 508                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 509                 }
 510         }
 511
 512         filp->f_pos = (loff_t)(__s32)pos;
 513         filp->f_version = inode->i_version;
 514         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 515
 516         ll_dir_chain_fini(&chain);
 517
 518         RETURN(rc);
 519 }
 520
 521 #define QCTL_COPY(out, in)              \
 522 do {                                    \
 523         Q_COPY(out, in, qc_cmd);        \
 524         Q_COPY(out, in, qc_type);       \
 525         Q_COPY(out, in, qc_id);         \
 526         Q_COPY(out, in, qc_stat);       \
 527         Q_COPY(out, in, qc_dqinfo);     \
 528         Q_COPY(out, in, qc_dqblk);      \
 529 } while (0)
 530
 531 int ll_send_mgc_param(struct obd_export *mgc, char *string)
 532 {
 533         struct mgs_send_param *msp;
 534         int rc = 0;
 535
 536         OBD_ALLOC_PTR(msp);
 537         if (!msp)
 538                 return -ENOMEM;
 539
 540         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
 541         rc = obd_set_info_async(mgc, strlen(KEY_SET_INFO), KEY_SET_INFO,
 542                                 sizeof(struct mgs_send_param), msp, NULL);
 543         if (rc)
 544                 CERROR("Failed to set parameter: %d\n", rc);
 545         OBD_FREE_PTR(msp);
 546
 547         return rc;
 548 }
 549
 550 char *ll_get_fsname(struct inode *inode)
 551 {
 552         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 553         char *ptr, *fsname;
 554         int len;
 555
 556         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
 557         len = strlen(lsi->lsi_lmd->lmd_profile);
 558         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
 559         if (ptr && (strcmp(ptr, "-client") == 0))
 560                 len -= 7;
 561         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
 562         fsname[len] = '\0';
 563
 564         return fsname;
 565 }
 566
 567 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
 568                      int set_default)
 569 {
 570         struct ll_sb_info *sbi = ll_i2sbi(inode);
 571         struct md_op_data *op_data;
 572         struct ptlrpc_request *req = NULL;
 573         int rc = 0;
 574         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
 575         struct obd_device *mgc = lsi->lsi_mgc;
 576         char *fsname = NULL, *param = NULL;
 577
 578         /*
 579          * This is coming from userspace, so should be in
 580          * local endian.  But the MDS would like it in little
 581          * endian, so we swab it before we send it.
 582          */
 583         if (lump->lmm_magic != LOV_USER_MAGIC)
 584                 RETURN(-EINVAL);
 585
 586         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC))
 587                 lustre_swab_lov_user_md(lump);
 588
 589         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
 590                                      LUSTRE_OPC_ANY, NULL);
 591         if (IS_ERR(op_data))
 592                 RETURN(PTR_ERR(op_data));
 593
 594         /* swabbing is done in lov_setstripe() on server side */
 595         rc = md_setattr(sbi->ll_md_exp, op_data, lump, sizeof(*lump),
 596                         NULL, 0, &req, NULL);
 597         ll_finish_md_op_data(op_data);
 598         ptlrpc_req_finished(req);
 599         if (rc) {
 600                 if (rc != -EPERM && rc != -EACCES)
 601                         CERROR("mdc_setattr fails: rc = %d\n", rc);
 602         }
 603
 604         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
 605                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
 606
 607                 /* Get fsname and assume devname to be -MDT0000. */
 608                 fsname = ll_get_fsname(inode);
 609                 /* Set root stripesize */
 610                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
 611                         lump->lmm_stripe_size);
 612                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 613                 if (rc)
 614                         goto end;
 615
 616                 /* Set root stripecount */
 617                 sprintf(param, "%s-MDT0000.lov.stripecount=%u", fsname,
 618                         lump->lmm_stripe_count);
 619                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 620                 if (rc)
 621                         goto end;
 622
 623                 /* Set root stripeoffset */
 624                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%u", fsname,
 625                         lump->lmm_stripe_offset);
 626                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
 627                 if (rc)
 628                         goto end;
 629 end:
 630                 if (fsname)
 631                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
 632                 if (param)
 633                         OBD_FREE(param, MGS_PARAM_MAXLEN);
 634         }
 635         return rc;
 636 }
 637
 638 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
 639                      int *lmm_size, struct ptlrpc_request **request)
 640 {
 641         struct ll_sb_info *sbi = ll_i2sbi(inode);
 642         struct mdt_body   *body;
 643         struct lov_mds_md *lmm = NULL;
 644         struct ptlrpc_request *req = NULL;
 645         int rc, lmmsize;
 646         struct obd_capa *oc;
 647
 648         rc = ll_get_max_mdsize(sbi, &lmmsize);
 649         if (rc)
 650                 RETURN(rc);
 651
 652         oc = ll_mdscapa_get(inode);
 653         rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode),
 654                         oc, OBD_MD_FLEASIZE | OBD_MD_FLDIREA,
 655                         lmmsize, &req);
 656         capa_put(oc);
 657         if (rc < 0) {
 658                 CDEBUG(D_INFO, "md_getattr failed on inode "
 659                        "%lu/%u: rc %d\n", inode->i_ino,
 660                        inode->i_generation, rc);
 661                 GOTO(out, rc);
 662         }
 663
 664         body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
 665         LASSERT(body != NULL); /* checked by md_getattr_name */
 666         /* swabbed by mdc_getattr_name */
 667         LASSERT_REPSWABBED(req, REPLY_REC_OFF);
 668
 669         lmmsize = body->eadatasize;
 670
 671         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 672             lmmsize == 0) {
 673                 GOTO(out, rc = -ENODATA);
 674         }
 675
 676         lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, lmmsize);
 677         LASSERT(lmm != NULL);
 678         LASSERT_REPSWABBED(req, REPLY_REC_OFF + 1);
 679
 680         /*
 681          * This is coming from the MDS, so is probably in
 682          * little endian.  We convert it to host endian before
 683          * passing it to userspace.
 684          */
 685         if (lmm->lmm_magic == __swab32(LOV_MAGIC)) {
 686                 lustre_swab_lov_user_md((struct lov_user_md *)lmm);
 687                 lustre_swab_lov_user_md_objects((struct lov_user_md *)lmm);
 688         }
 689 out:
 690         *lmmp = lmm;
 691         *lmm_size = lmmsize;
 692         *request = req;
 693         return rc;
 694 }
 695
 696 static int ll_dir_ioctl(struct inode *inode, struct file *file,
 697                         unsigned int cmd, unsigned long arg)
 698 {
 699         struct ll_sb_info *sbi = ll_i2sbi(inode);
 700         struct obd_ioctl_data *data;
 701         ENTRY;
 702
 703         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
 704                inode->i_ino, inode->i_generation, inode, cmd);
 705
 706         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
 707         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
 708                 return -ENOTTY;
 709
 710         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
 711         switch(cmd) {
 712         case EXT3_IOC_GETFLAGS:
 713         case EXT3_IOC_SETFLAGS:
 714                 RETURN(ll_iocontrol(inode, file, cmd, arg));
 715         case EXT3_IOC_GETVERSION_OLD:
 716         case EXT3_IOC_GETVERSION:
 717                 RETURN(put_user(inode->i_generation, (int *)arg));
 718         /* We need to special case any other ioctls we want to handle,
 719          * to send them to the MDS/OST as appropriate and to properly
 720          * network encode the arg field.
 721         case EXT3_IOC_SETVERSION_OLD:
 722         case EXT3_IOC_SETVERSION:
 723         */
 724         case IOC_MDC_LOOKUP: {
 725                 struct ptlrpc_request *request = NULL;
 726                 int namelen, rc, len = 0;
 727                 char *buf = NULL;
 728                 char *filename;
 729                 struct obd_capa *oc;
 730
 731                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 732                 if (rc)
 733                         RETURN(rc);
 734                 data = (void *)buf;
 735
 736                 filename = data->ioc_inlbuf1;
 737                 namelen = data->ioc_inllen1;
 738
 739                 if (namelen < 1) {
 740                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
 741                         GOTO(out, rc = -EINVAL);
 742                 }
 743
 744                 oc = ll_mdscapa_get(inode);
 745                 rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode), oc,
 746                                      filename, namelen, OBD_MD_FLID, 0,
 747                                      &request);
 748                 capa_put(oc);
 749                 if (rc < 0) {
 750                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 751                         GOTO(out, rc);
 752                 }
 753
 754                 ptlrpc_req_finished(request);
 755
 756                 EXIT;
 757         out:
 758                 obd_ioctl_freedata(buf, len);
 759                 return rc;
 760         }
 761         case LL_IOC_LOV_SETSTRIPE: {
 762                 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
 763                 int rc = 0;
 764                 int set_default = 0;
 765
 766                 LASSERT(sizeof(lum) == sizeof(*lump));
 767                 LASSERT(sizeof(lum.lmm_objects[0]) ==
 768                         sizeof(lump->lmm_objects[0]));
 769                 rc = copy_from_user(&lum, lump, sizeof(lum));
 770                 if (rc)
 771                         RETURN(-EFAULT);
 772
 773                 if (inode->i_sb->s_root == file->f_dentry)
 774                         set_default = 1;
 775
 776                 rc = ll_dir_setstripe(inode, &lum, set_default);
 777
 778                 RETURN(rc);
 779         }
 780         case LL_IOC_OBD_STATFS:
 781                 RETURN(ll_obd_statfs(inode, (void *)arg));
 782         case LL_IOC_LOV_GETSTRIPE:
 783         case LL_IOC_MDC_GETINFO:
 784         case IOC_MDC_GETFILEINFO:
 785         case IOC_MDC_GETFILESTRIPE: {
 786                 struct ptlrpc_request *request = NULL;
 787                 struct lov_user_md *lump;
 788                 struct lov_mds_md *lmm = NULL;
 789                 struct mdt_body *body;
 790                 char *filename = NULL;
 791                 int rc, lmmsize;
 792
 793                 if (cmd == IOC_MDC_GETFILEINFO ||
 794                     cmd == IOC_MDC_GETFILESTRIPE) {
 795                         filename = getname((const char *)arg);
 796                         if (IS_ERR(filename))
 797                                 RETURN(PTR_ERR(filename));
 798
 799                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
 800                                                       &lmmsize, &request);
 801                 } else {
 802                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
 803                 }
 804
 805                 if (request) {
 806                         body = lustre_msg_buf(request->rq_repmsg,
 807                                               REPLY_REC_OFF, sizeof(*body));
 808                         LASSERT(body != NULL); /* checked by md_getattr_name */
 809                         /* swabbed by md_getattr_name */
 810                         LASSERT_REPSWABBED(request, REPLY_REC_OFF);
 811                 } else {
 812                         GOTO(out_req, rc);
 813                 }
 814
 815                 if (rc < 0) {
 816                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
 817                                                cmd == LL_IOC_MDC_GETINFO))
 818                                 GOTO(skip_lmm, rc = 0);
 819                         else
 820                                 GOTO(out_req, rc);
 821                 }
 822
 823                 if (cmd == IOC_MDC_GETFILESTRIPE ||
 824                     cmd == LL_IOC_LOV_GETSTRIPE) {
 825                         lump = (struct lov_user_md *)arg;
 826                 } else {
 827                         struct lov_user_mds_data *lmdp;
 828                         lmdp = (struct lov_user_mds_data *)arg;
 829                         lump = &lmdp->lmd_lmm;
 830                 }
 831                 rc = copy_to_user(lump, lmm, lmmsize);
 832                 if (rc)
 833                         GOTO(out_lmm, rc = -EFAULT);
 834         skip_lmm:
 835                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
 836                         struct lov_user_mds_data *lmdp;
 837                         lstat_t st = { 0 };
 838
 839                         st.st_dev     = inode->i_sb->s_dev;
 840                         st.st_mode    = body->mode;
 841                         st.st_nlink   = body->nlink;
 842                         st.st_uid     = body->uid;
 843                         st.st_gid     = body->gid;
 844                         st.st_rdev    = body->rdev;
 845                         st.st_size    = body->size;
 846                         st.st_blksize = CFS_PAGE_SIZE;
 847                         st.st_blocks  = body->blocks;
 848                         st.st_atime   = body->atime;
 849                         st.st_mtime   = body->mtime;
 850                         st.st_ctime   = body->ctime;
 851                         st.st_ino     = inode->i_ino;
 852
 853                         lmdp = (struct lov_user_mds_data *)arg;
 854                         rc = copy_to_user(&lmdp->lmd_st, &st, sizeof(st));
 855                         if (rc)
 856                                 GOTO(out_lmm, rc = -EFAULT);
 857                 }
 858
 859                 EXIT;
 860         out_lmm:
 861                 if (lmm && lmm->lmm_magic == LOV_MAGIC_JOIN)
 862                         OBD_FREE(lmm, lmmsize);
 863         out_req:
 864                 ptlrpc_req_finished(request);
 865                 if (filename)
 866                         putname(filename);
 867                 return rc;
 868         }
 869         case IOC_LOV_GETINFO: {
 870                 struct lov_user_mds_data *lumd;
 871                 struct lov_stripe_md *lsm;
 872                 struct lov_user_md *lum;
 873                 struct lov_mds_md *lmm;
 874                 int lmmsize;
 875                 lstat_t st;
 876                 int rc;
 877
 878                 lumd = (struct lov_user_mds_data *)arg;
 879                 lum = &lumd->lmd_lmm;
 880
 881                 rc = ll_get_max_mdsize(sbi, &lmmsize);
 882                 if (rc)
 883                         RETURN(rc);
 884
 885                 OBD_ALLOC(lmm, lmmsize);
 886                 rc = copy_from_user(lmm, lum, lmmsize);
 887                 if (rc)
 888                         GOTO(free_lmm, rc = -EFAULT);
 889
 890                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
 891                 if (rc < 0)
 892                         GOTO(free_lmm, rc = -ENOMEM);
 893
 894                 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
 895                 if (rc)
 896                         GOTO(free_lsm, rc);
 897
 898                 /* Perform glimpse_size operation. */
 899                 memset(&st, 0, sizeof(st));
 900
 901                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
 902                 if (rc)
 903                         GOTO(free_lsm, rc);
 904
 905                 rc = copy_to_user(&lumd->lmd_st, &st, sizeof(st));
 906                 if (rc)
 907                         GOTO(free_lsm, rc = -EFAULT);
 908
 909                 EXIT;
 910         free_lsm:
 911                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
 912         free_lmm:
 913                 OBD_FREE(lmm, lmmsize);
 914                 return rc;
 915         }
 916         case OBD_IOC_LLOG_CATINFO: {
 917                 struct ptlrpc_request *req = NULL;
 918                 char *buf = NULL;
 919                 int rc, len = 0;
 920                 char *bufs[3] = { NULL }, *str;
 921                 int lens[3] = { sizeof(struct ptlrpc_body) };
 922                 int size[2] = { sizeof(struct ptlrpc_body) };
 923
 924                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 925                 if (rc)
 926                         RETURN(rc);
 927                 data = (void *)buf;
 928
 929                 if (!data->ioc_inlbuf1) {
 930                         obd_ioctl_freedata(buf, len);
 931                         RETURN(-EINVAL);
 932                 }
 933
 934                 lens[REQ_REC_OFF] = data->ioc_inllen1;
 935                 bufs[REQ_REC_OFF] = data->ioc_inlbuf1;
 936                 if (data->ioc_inllen2) {
 937                         lens[REQ_REC_OFF + 1] = data->ioc_inllen2;
 938                         bufs[REQ_REC_OFF + 1] = data->ioc_inlbuf2;
 939                 } else {
 940                         lens[REQ_REC_OFF + 1] = 0;
 941                         bufs[REQ_REC_OFF + 1] = NULL;
 942                 }
 943
 944                 req = ptlrpc_prep_req(sbi2mdc(sbi)->cl_import,
 945                                       LUSTRE_LOG_VERSION, LLOG_CATINFO, 3, lens,
 946                                       bufs);
 947                 if (!req)
 948                         GOTO(out_catinfo, rc = -ENOMEM);
 949
 950                 size[REPLY_REC_OFF] = data->ioc_plen1;
 951                 ptlrpc_req_set_repsize(req, 2, size);
 952
 953                 rc = ptlrpc_queue_wait(req);
 954                 if (!rc) {
 955                         str = lustre_msg_string(req->rq_repmsg, REPLY_REC_OFF,
 956                                                 data->ioc_plen1);
 957                         rc = copy_to_user(data->ioc_pbuf1, str, data->ioc_plen1);
 958                 }
 959                 ptlrpc_req_finished(req);
 960         out_catinfo:
 961                 obd_ioctl_freedata(buf, len);
 962                 RETURN(rc);
 963         }
 964         case OBD_IOC_QUOTACHECK: {
 965                 struct obd_quotactl *oqctl;
 966                 int rc, error = 0;
 967
 968                 if (!capable(CAP_SYS_ADMIN))
 969                         RETURN(-EPERM);
 970
 971                 OBD_ALLOC_PTR(oqctl);
 972                 if (!oqctl)
 973                         RETURN(-ENOMEM);
 974                 oqctl->qc_type = arg;
 975                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
 976                 if (rc < 0) {
 977                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
 978                         error = rc;
 979                 }
 980
 981                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
 982                 if (rc < 0)
 983                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
 984
 985                 OBD_FREE_PTR(oqctl);
 986                 return error ?: rc;
 987         }
 988         case OBD_IOC_POLL_QUOTACHECK: {
 989                 struct if_quotacheck *check;
 990                 int rc;
 991
 992                 if (!capable(CAP_SYS_ADMIN))
 993                         RETURN(-EPERM);
 994
 995                 OBD_ALLOC_PTR(check);
 996                 if (!check)
 997                         RETURN(-ENOMEM);
 998
 999                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1000                                    NULL);
1001                 if (rc) {
1002                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1003                         if (copy_to_user((void *)arg, check, sizeof(*check)))
1004                                 rc = -EFAULT;
1005                         GOTO(out_poll, rc);
1006                 }
1007
1008                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1009                                    NULL);
1010                 if (rc) {
1011                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1012                         if (copy_to_user((void *)arg, check, sizeof(*check)))
1013                                 rc = -EFAULT;
1014                         GOTO(out_poll, rc);
1015                 }
1016         out_poll:
1017                 OBD_FREE_PTR(check);
1018                 RETURN(rc);
1019         }
1020 #ifdef HAVE_QUOTA_SUPPORT
1021         case OBD_IOC_QUOTACTL: {
1022                 struct if_quotactl *qctl;
1023                 struct obd_quotactl *oqctl;
1024
1025                 int cmd, type, id, rc = 0;
1026
1027                 OBD_ALLOC_PTR(qctl);
1028                 if (!qctl)
1029                         RETURN(-ENOMEM);
1030
1031                 OBD_ALLOC_PTR(oqctl);
1032                 if (!oqctl) {
1033                         OBD_FREE_PTR(qctl);
1034                         RETURN(-ENOMEM);
1035                 }
1036                 if (copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1037                         GOTO(out_quotactl, rc = -EFAULT);
1038
1039                 cmd = qctl->qc_cmd;
1040                 type = qctl->qc_type;
1041                 id = qctl->qc_id;
1042                 switch (cmd) {
1043                 case Q_QUOTAON:
1044                 case Q_QUOTAOFF:
1045                 case Q_SETQUOTA:
1046                 case Q_SETINFO:
1047                         if (!capable(CAP_SYS_ADMIN))
1048                                 GOTO(out_quotactl, rc = -EPERM);
1049                         break;
1050                 case Q_GETQUOTA:
1051                         if (((type == USRQUOTA && current->euid != id) ||
1052                              (type == GRPQUOTA && !in_egroup_p(id))) &&
1053                             !capable(CAP_SYS_ADMIN))
1054                                 GOTO(out_quotactl, rc = -EPERM);
1055
1056                         /* XXX: dqb_valid is borrowed as a flag to mark that
1057                          *      only mds quota is wanted */
1058                         if (qctl->qc_dqblk.dqb_valid)
1059                                 qctl->obd_uuid = sbi->ll_md_exp->exp_obd->
1060                                                         u.cli.cl_target_uuid;
1061                         break;
1062                 case Q_GETINFO:
1063                         break;
1064                 default:
1065                         CERROR("unsupported quotactl op: %#x\n", cmd);
1066                         GOTO(out_quotactl, rc = -ENOTTY);
1067                 }
1068
1069                 QCTL_COPY(oqctl, qctl);
1070
1071                 if (qctl->obd_uuid.uuid[0]) {
1072                         struct obd_device *obd;
1073                         struct obd_uuid *uuid = &qctl->obd_uuid;
1074
1075                         obd = class_find_client_notype(uuid,
1076                                          &sbi->ll_dt_exp->exp_obd->obd_uuid);
1077                         if (!obd)
1078                                 GOTO(out_quotactl, rc = -ENOENT);
1079
1080                         if (cmd == Q_GETINFO)
1081                                 oqctl->qc_cmd = Q_GETOINFO;
1082                         else if (cmd == Q_GETQUOTA)
1083                                 oqctl->qc_cmd = Q_GETOQUOTA;
1084                         else
1085                                 GOTO(out_quotactl, rc = -EINVAL);
1086
1087                         if (sbi->ll_md_exp->exp_obd == obd) {
1088                                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1089                         } else {
1090                                 int i;
1091                                 struct obd_export *exp;
1092                                 struct lov_obd *lov = &sbi->ll_dt_exp->
1093                                                             exp_obd->u.lov;
1094
1095                                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1096                                         if (!lov->lov_tgts[i] ||
1097                                             !lov->lov_tgts[i]->ltd_active)
1098                                                 continue;
1099                                         exp = lov->lov_tgts[i]->ltd_exp;
1100                                         if (exp->exp_obd == obd) {
1101                                                 rc = obd_quotactl(exp, oqctl);
1102                                                 break;
1103                                         }
1104                                 }
1105                         }
1106
1107                         oqctl->qc_cmd = cmd;
1108                         QCTL_COPY(qctl, oqctl);
1109
1110                         if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1111                                 rc = -EFAULT;
1112
1113                         GOTO(out_quotactl, rc);
1114                 }
1115
1116                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1117                 if (rc && rc != -EBUSY && cmd == Q_QUOTAON) {
1118                         oqctl->qc_cmd = Q_QUOTAOFF;
1119                         obd_quotactl(sbi->ll_md_exp, oqctl);
1120                 }
1121
1122                 QCTL_COPY(qctl, oqctl);
1123
1124                 if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1125                         rc = -EFAULT;
1126         out_quotactl:
1127                 OBD_FREE_PTR(qctl);
1128                 OBD_FREE_PTR(oqctl);
1129                 RETURN(rc);
1130         }
1131 #endif /* HAVE_QUOTA_SUPPORT */
1132         case OBD_IOC_GETNAME: {
1133                 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1134                 if (!obd)
1135                         RETURN(-EFAULT);
1136                 if (copy_to_user((void *)arg, obd->obd_name,
1137                                 strlen(obd->obd_name) + 1))
1138                         RETURN (-EFAULT);
1139                 RETURN(0);
1140         }
1141         case LL_IOC_FLUSHCTX:
1142                 RETURN(ll_flush_ctx(inode));
1143         case LL_IOC_GETFACL: {
1144                 struct rmtacl_ioctl_data ioc;
1145
1146                 if (copy_from_user(&ioc, (void *)arg, sizeof(ioc)))
1147                         RETURN(-EFAULT);
1148
1149                 RETURN(ll_ioctl_getfacl(inode, &ioc));
1150         }
1151         case LL_IOC_SETFACL: {
1152                 struct rmtacl_ioctl_data ioc;
1153
1154                 if (copy_from_user(&ioc, (void *)arg, sizeof(ioc)))
1155                         RETURN(-EFAULT);
1156
1157                 RETURN(ll_ioctl_setfacl(inode, &ioc));
1158         }
1159         default:
1160                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1161         }
1162 }
1163
1164 int ll_dir_open(struct inode *inode, struct file *file)
1165 {
1166         ENTRY;
1167         RETURN(ll_file_open(inode, file));
1168 }
1169
1170 int ll_dir_release(struct inode *inode, struct file *file)
1171 {
1172         ENTRY;
1173         RETURN(ll_file_release(inode, file));
1174 }
1175
1176 struct file_operations ll_dir_operations = {
1177         .open     = ll_dir_open,
1178         .release  = ll_dir_release,
1179         .read     = generic_read_dir,
1180         .readdir  = ll_readdir,
1181         .ioctl    = ll_dir_ioctl
1182 };
1183