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 "llite_internal.h"
  51
  52 #define PageChecked(page)        test_bit(PG_checked, &(page)->flags)
  53 #define SetPageChecked(page)     set_bit(PG_checked, &(page)->flags)
  54
  55 /*
  56  * (new) readdir implementation overview.
  57  *
  58  * Original lustre readdir implementation cached exact copy of raw directory
  59  * pages on the client. These pages were indexed in client page cache by
  60  * logical offset in the directory file. This design, while very simple and
  61  * intuitive had some inherent problems:
  62  *
  63  *     . it implies that byte offset to the directory entry serves as a
  64  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
  65  *     ext3/htree directory entries may move due to splits, and more
  66  *     importantly,
  67  *
  68  *     . it is incompatible with the design of split directories for cmd3,
  69  *     that assumes that names are distributed across nodes based on their
  70  *     hash, and so readdir should be done in hash order.
  71  *
  72  * New readdir implementation does readdir in hash order, and uses hash of a
  73  * file name as a telldir/seekdir cookie. This led to number of complications:
  74  *
  75  *     . hash is not unique, so it cannot be used to index cached directory
  76  *     pages on the client (note, that it requires a whole pageful of hash
  77  *     collided entries to cause two pages to have identical hashes);
  78  *
  79  *     . hash is not unique, so it cannot, strictly speaking, be used as an
  80  *     entry cookie. ext3/htree has the same problem and lustre implementation
  81  *     mimics their solution: seekdir(hash) positions directory at the first
  82  *     entry with the given hash.
  83  *
  84  * Client side.
  85  *
  86  * 0. caching
  87  *
  88  * Client caches directory pages using hash of the first entry as an index. As
  89  * noted above hash is not unique, so this solution doesn't work as is:
  90  * special processing is needed for "page hash chains" (i.e., sequences of
  91  * pages filled with entries all having the same hash value).
  92  *
  93  * First, such chains have to be detected. To this end, server returns to the
  94  * client the hash of the first entry on the page next to one returned. When
  95  * client detects that this hash is the same as hash of the first entry on the
  96  * returned page, page hash collision has to be handled. Pages in the
  97  * hash chain, except first one, are termed "overflow pages".
  98  *
  99  * Solution to index uniqueness problem is to not cache overflow
 100  * pages. Instead, when page hash collision is detected, all overflow pages
 101  * from emerging chain are immediately requested from the server and placed in
 102  * a special data structure (struct ll_dir_chain). This data structure is used
 103  * by ll_readdir() to process entries from overflow pages. When readdir
 104  * invocation finishes, overflow pages are discarded. If page hash collision
 105  * chain weren't completely processed, next call to readdir will again detect
 106  * page hash collision, again read overflow pages in, process next portion of
 107  * entries and again discard the pages. This is not as wasteful as it looks,
 108  * because, given reasonable hash, page hash collisions are extremely rare.
 109  *
 110  * 1. directory positioning
 111  *
 112  * When seekdir(hash) is called, original
 113  *
 114  *
 115  *
 116  *
 117  *
 118  *
 119  *
 120  *
 121  * Server.
 122  *
 123  * identification of and access to overflow pages
 124  *
 125  * page format
 126  *
 127  *
 128  *
 129  *
 130  *
 131  */
 132
 133 static __u32 hash_x_index(__u32 value)
 134 {
 135         return ((__u32)~0) - value;
 136 }
 137
 138 /* returns the page unlocked, but with a reference */
 139 static int ll_dir_readpage(struct file *file, struct page *page)
 140 {
 141         struct inode *inode = page->mapping->host;
 142         struct ptlrpc_request *request;
 143         struct mdt_body *body;
 144         __u32 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         rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
 153                          hash, page, &request);
 154         if (!rc) {
 155                 body = lustre_msg_buf(request->rq_repmsg, REPLY_REC_OFF,
 156                                       sizeof(*body));
 157                 LASSERT(body != NULL); /* checked by md_readpage() */
 158                 /* swabbed by md_readpage() */
 159                 LASSERT_REPSWABBED(request, REPLY_REC_OFF);
 160
 161                 if (body->valid & OBD_MD_FLSIZE)
 162                         inode->i_size = body->size;
 163                 SetPageUptodate(page);
 164         }
 165         ptlrpc_req_finished(request);
 166
 167         unlock_page(page);
 168         EXIT;
 169         return rc;
 170 }
 171
 172 struct address_space_operations ll_dir_aops = {
 173         .readpage  = ll_dir_readpage,
 174 };
 175
 176 static inline unsigned long dir_pages(struct inode *inode)
 177 {
 178         return (inode->i_size+PAGE_CACHE_SIZE-1)>>PAGE_CACHE_SHIFT;
 179 }
 180
 181 static inline unsigned ll_chunk_size(struct inode *inode)
 182 {
 183         return inode->i_sb->s_blocksize;
 184 }
 185
 186 static void ll_check_page(struct inode *dir, struct page *page)
 187 {
 188         /* XXX: check page format later */
 189         SetPageChecked(page);
 190 }
 191
 192 static inline void ll_put_page(struct page *page)
 193 {
 194         kunmap(page);
 195         page_cache_release(page);
 196 }
 197
 198 /*
 199  * Find, kmap and return page that contains given hash.
 200  */
 201 static struct page *ll_dir_page_locate(struct inode *dir, unsigned long hash,
 202                                        __u32 *start, __u32 *end)
 203 {
 204         struct address_space *mapping = dir->i_mapping;
 205         /*
 206          * Complement of hash is used as an index so that
 207          * radix_tree_gang_lookup() can be used to find a page with starting
 208          * hash _smaller_ than one we are looking for.
 209          */
 210         unsigned long offset = hash_x_index(hash);
 211         struct page *page;
 212         int found;
 213
 214         spin_lock_irq(&mapping->tree_lock);
 215         found = radix_tree_gang_lookup(&mapping->page_tree,
 216                                        (void **)&page, offset, 1);
 217         if (found > 0) {
 218                 struct lu_dirpage *dp;
 219
 220                 page_cache_get(page);
 221                 spin_unlock_irq(&mapping->tree_lock);
 222                 /*
 223                  * In contrast to find_lock_page() we are sure that directory
 224                  * page cannot be truncated (while DLM lock is held) and,
 225                  * hence, can avoid restart.
 226                  *
 227                  * In fact, page cannot be locked here at all, because
 228                  * ll_dir_readpage() does synchronous io.
 229                  */
 230                 wait_on_page(page);
 231                 if (PageUptodate(page)) {
 232                         dp = kmap(page);
 233                         *start = le32_to_cpu(dp->ldp_hash_start);
 234                         *end   = le32_to_cpu(dp->ldp_hash_end);
 235                         LASSERT(*start <= hash);
 236                         if (hash > *end || (*end != *start && hash == *end)) {
 237                                 kunmap(page);
 238                                 page_cache_release(page);
 239                                 page = NULL;
 240                         }
 241                 } else {
 242                         page_cache_release(page);
 243                         page = ERR_PTR(-EIO);
 244                 }
 245
 246         } else {
 247                 spin_unlock_irq(&mapping->tree_lock);
 248                 page = NULL;
 249         }
 250         return page;
 251 }
 252
 253 /*
 254  * Chain of hash overflow pages.
 255  */
 256 struct ll_dir_chain {
 257         /* XXX something. Later */
 258 };
 259
 260 static void ll_dir_chain_init(struct ll_dir_chain *chain)
 261 {
 262 }
 263
 264 static void ll_dir_chain_fini(struct ll_dir_chain *chain)
 265 {
 266 }
 267
 268 static struct page *ll_get_dir_page(struct inode *dir, __u32 hash, int exact,
 269                                     struct ll_dir_chain *chain)
 270 {
 271         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
 272         struct address_space *mapping = dir->i_mapping;
 273         struct lustre_handle lockh;
 274         struct page *page;
 275         struct lu_dirpage *dp;
 276         int rc;
 277         __u32 start;
 278         __u32 end;
 279
 280         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
 281                            ll_inode2fid(dir), LDLM_IBITS, &policy, LCK_CR, &lockh);
 282         if (!rc) {
 283                 struct lookup_intent it = { .it_op = IT_READDIR };
 284                 struct ptlrpc_request *request;
 285                 struct md_op_data *op_data;
 286
 287                 OBD_ALLOC_PTR(op_data);
 288                 if (op_data == NULL)
 289                         return ERR_PTR(-ENOMEM);
 290
 291                 ll_prepare_md_op_data(op_data, dir, NULL, NULL, 0, 0);
 292
 293                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, LDLM_IBITS, &it,
 294                                 LCK_CR, op_data, &lockh, NULL, 0,
 295                                 ldlm_completion_ast, ll_md_blocking_ast, dir,
 296                                 0);
 297
 298                 OBD_FREE_PTR(op_data);
 299
 300                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
 301                 if (request)
 302                         ptlrpc_req_finished(request);
 303                 if (rc < 0) {
 304                         CERROR("lock enqueue: rc: %d\n", rc);
 305                         return ERR_PTR(rc);
 306                 }
 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: %#lx != %#lx\n",
 337                               (unsigned long)hash, (unsigned long)start);
 338                         truncate_complete_page(mapping, page);
 339                         page_cache_release(page);
 340                 } else
 341                         GOTO(hash_collision, page);
 342         }
 343
 344         page = read_cache_page(mapping, hash_x_index(hash),
 345                                (filler_t*)mapping->a_ops->readpage, NULL);
 346         if (IS_ERR(page))
 347                 GOTO(out_unlock, page);
 348
 349         wait_on_page(page);
 350         (void)kmap(page);
 351         if (!PageUptodate(page))
 352                 goto fail;
 353         if (!PageChecked(page))
 354                 ll_check_page(dir, page);
 355         if (PageError(page))
 356                 goto fail;
 357 hash_collision:
 358         dp = page_address(page);
 359
 360         start = le32_to_cpu(dp->ldp_hash_start);
 361         end   = le32_to_cpu(dp->ldp_hash_end);
 362         if (end == start) {
 363                 LASSERT(start == hash);
 364                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
 365                 /*
 366                  * Fetch whole overflow chain...
 367                  *
 368                  * XXX not yet.
 369                  */
 370                 goto fail;
 371         }
 372 out_unlock:
 373         ldlm_lock_decref(&lockh, LCK_CR);
 374         return page;
 375
 376 fail:
 377         ll_put_page(page);
 378         page = ERR_PTR(-EIO);
 379         goto out_unlock;
 380 }
 381
 382 static loff_t ll_llseek(struct file *filp, loff_t off, int whence)
 383 {
 384         if (off != 0 || whence != 1 /* SEEK_CUR */) {
 385                 /*
 386                  * Except when telldir() is going on, reset readdir to the
 387                  * beginning of hash collision chain.
 388                  */
 389                 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
 390
 391                 fd->fd_dir.lfd_dup = 0;
 392         }
 393         return default_llseek(filp, off, whence);
 394 }
 395
 396 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
 397 {
 398         struct inode         *inode = filp->f_dentry->d_inode;
 399         struct ll_inode_info *info  = ll_i2info(inode);
 400         struct ll_file_data  *fd    = LUSTRE_FPRIVATE(filp);
 401         struct ll_sb_info    *sbi   = ll_i2sbi(inode);
 402         __u32                 pos   = filp->f_pos;
 403         struct page          *page;
 404         struct ll_dir_chain   chain;
 405         __u32 prevhash;
 406         int rc;
 407         int dup;
 408         int done;
 409         int shift;
 410         ENTRY;
 411
 412         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu\n",
 413                inode->i_ino, inode->i_generation, inode,
 414                (unsigned long)pos, inode->i_size);
 415
 416         if (pos == ~0)
 417                 /*
 418                  * end-of-file.
 419                  */
 420                 RETURN(0);
 421
 422         rc    = 0;
 423         dup   = 0;
 424         done  = 0;
 425         shift = 0;
 426         prevhash = ~0; /* impossible hash value */
 427         ll_dir_chain_init(&chain);
 428
 429         page = ll_get_dir_page(inode, pos, 0, &chain);
 430
 431         while (rc == 0 && !done) {
 432                 struct lu_dirpage *dp;
 433                 struct lu_dirent  *ent;
 434
 435                 if (!IS_ERR(page)) {
 436                         __u32 hash; /* no, Richard, it _is_ initialized */
 437                         __u32 next;
 438
 439                         dp = page_address(page);
 440                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
 441                              ent = lu_dirent_next(ent)) {
 442                                 char          *name;
 443                                 int            namelen;
 444                                 struct lu_fid  fid;
 445                                 ino_t          ino;
 446
 447                                 /*
 448                                  * XXX: implement correct swabbing here.
 449                                  */
 450
 451                                 hash    = le32_to_cpu(ent->lde_hash);
 452                                 namelen = le16_to_cpu(ent->lde_namelen);
 453
 454                                 if (hash < pos)
 455                                         /*
 456                                          * Skip until we find target hash
 457                                          * value.
 458                                          */
 459                                         continue;
 460
 461                                 if (namelen == 0)
 462                                         /*
 463                                          * Skip dummy record.
 464                                          */
 465                                         continue;
 466                                 /*
 467                                  * Keep track of how far we get into duplicate
 468                                  * hash segment.
 469                                  */
 470                                 if (hash == prevhash)
 471                                         dup++;
 472                                 prevhash = hash;
 473
 474                                 if (hash == fd->fd_dir.lfd_duppos &&
 475                                     fd->fd_dir.lfd_dup > 0) {
 476                                         fd->fd_dir.lfd_dup--;
 477                                         continue;
 478                                 }
 479
 480                                 fid  = ent->lde_fid;
 481                                 name = ent->lde_name;
 482                                 fid_le_to_cpu(&fid);
 483                                 ino  = ll_fid_build_ino(sbi, &fid);
 484
 485                                 done = filldir(cookie, name, namelen,
 486                                                hash, ino, DT_UNKNOWN);
 487                         }
 488                         next = le32_to_cpu(dp->ldp_hash_end);
 489                         ll_put_page(page);
 490                         if (!done) {
 491                                 pos = next;
 492                                 if (pos == ~0)
 493                                         /*
 494                                          * End of directory reached.
 495                                          */
 496                                         done = 1;
 497                                 else if (1 /* chain is exhausted*/)
 498                                         /*
 499                                          * Normal case: continue to the next
 500                                          * page.
 501                                          */
 502                                         page = ll_get_dir_page(inode, pos, 1,
 503                                                                &chain);
 504                                 else {
 505                                         /*
 506                                          * go into overflow page.
 507                                          */
 508                                 }
 509                         } else
 510                                 pos = hash;
 511                 } else {
 512                         rc = PTR_ERR(page);
 513                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
 514                                PFID(&info->lli_fid), (unsigned long)pos, rc);
 515                 }
 516         }
 517
 518         filp->f_pos = pos;
 519         filp->f_version = inode->i_version;
 520         fd->fd_dir.lfd_dup    = dup;
 521         fd->fd_dir.lfd_duppos = prevhash;
 522         touch_atime(filp->f_vfsmnt, filp->f_dentry);
 523
 524         ll_dir_chain_fini(&chain);
 525
 526         RETURN(rc);
 527 }
 528
 529 #define QCTL_COPY(out, in)              \
 530 do {                                    \
 531         Q_COPY(out, in, qc_cmd);        \
 532         Q_COPY(out, in, qc_type);       \
 533         Q_COPY(out, in, qc_id);         \
 534         Q_COPY(out, in, qc_stat);       \
 535         Q_COPY(out, in, qc_dqinfo);     \
 536         Q_COPY(out, in, qc_dqblk);      \
 537 } while (0)
 538
 539 static int ll_dir_ioctl(struct inode *inode, struct file *file,
 540                         unsigned int cmd, unsigned long arg)
 541 {
 542         struct ll_sb_info *sbi = ll_i2sbi(inode);
 543         struct obd_ioctl_data *data;
 544         ENTRY;
 545
 546         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
 547                inode->i_ino, inode->i_generation, inode, cmd);
 548
 549         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
 550         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
 551                 return -ENOTTY;
 552
 553         lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats, LPROC_LL_IOCTL);
 554         switch(cmd) {
 555         case EXT3_IOC_GETFLAGS:
 556         case EXT3_IOC_SETFLAGS:
 557                 RETURN(ll_iocontrol(inode, file, cmd, arg));
 558         case EXT3_IOC_GETVERSION_OLD:
 559         case EXT3_IOC_GETVERSION:
 560                 RETURN(put_user(inode->i_generation, (int *)arg));
 561         /* We need to special case any other ioctls we want to handle,
 562          * to send them to the MDS/OST as appropriate and to properly
 563          * network encode the arg field.
 564         case EXT3_IOC_SETVERSION_OLD:
 565         case EXT3_IOC_SETVERSION:
 566         */
 567         case IOC_MDC_LOOKUP: {
 568                 struct ptlrpc_request *request = NULL;
 569                 int namelen, rc, len = 0;
 570                 char *buf = NULL;
 571                 char *filename;
 572
 573                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 574                 if (rc)
 575                         RETURN(rc);
 576                 data = (void *)buf;
 577
 578                 filename = data->ioc_inlbuf1;
 579                 namelen = data->ioc_inllen1;
 580
 581                 if (namelen < 1) {
 582                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
 583                         GOTO(out, rc = -EINVAL);
 584                 }
 585
 586                 rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode),
 587                                      filename, namelen, OBD_MD_FLID, 0,
 588                                      &request);
 589                 if (rc < 0) {
 590                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
 591                         GOTO(out, rc);
 592                 }
 593
 594                 ptlrpc_req_finished(request);
 595
 596                 EXIT;
 597         out:
 598                 obd_ioctl_freedata(buf, len);
 599                 return rc;
 600         }
 601         case LL_IOC_LOV_SETSTRIPE: {
 602                 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
 603                 struct ptlrpc_request *request = NULL;
 604                 struct md_op_data *op_data;
 605                 struct iattr attr = { 0 };
 606                 int rc = 0;
 607
 608                 OBD_ALLOC_PTR(op_data);
 609                 if (op_data == NULL)
 610                         RETURN(-ENOMEM);
 611
 612                 ll_prepare_md_op_data(op_data, inode,
 613                                       NULL, NULL, 0, 0);
 614
 615                 LASSERT(sizeof(lum) == sizeof(*lump));
 616                 LASSERT(sizeof(lum.lmm_objects[0]) ==
 617                         sizeof(lump->lmm_objects[0]));
 618                 rc = copy_from_user(&lum, lump, sizeof(lum));
 619                 if (rc)
 620                         RETURN(-EFAULT);
 621
 622                 /*
 623                  * This is coming from userspace, so should be in
 624                  * local endian.  But the MDS would like it in little
 625                  * endian, so we swab it before we send it.
 626                  */
 627                 if (lum.lmm_magic != LOV_USER_MAGIC)
 628                         RETURN(-EINVAL);
 629
 630                 if (lum.lmm_magic != cpu_to_le32(LOV_USER_MAGIC))
 631                         lustre_swab_lov_user_md(&lum);
 632
 633                 /* swabbing is done in lov_setstripe() on server side */
 634                 rc = md_setattr(sbi->ll_md_exp, op_data,
 635                                 &attr, &lum, sizeof(lum), NULL, 0, &request);
 636                 if (rc) {
 637                         if (rc != -EPERM && rc != -EACCES)
 638                                 CERROR("md_setattr fails: rc = %d\n", rc);
 639                 }
 640                 OBD_FREE_PTR(op_data);
 641                 ptlrpc_req_finished(request);
 642                 RETURN(rc);
 643         }
 644         case LL_IOC_OBD_STATFS:
 645                 RETURN(ll_obd_statfs(inode, (void *)arg));
 646         case LL_IOC_LOV_GETSTRIPE:
 647         case LL_IOC_MDC_GETINFO:
 648         case IOC_MDC_GETFILEINFO:
 649         case IOC_MDC_GETFILESTRIPE: {
 650                 struct ptlrpc_request *request = NULL;
 651                 struct lov_user_md *lump;
 652                 struct lov_mds_md *lmm = NULL;
 653                 struct mdt_body *body;
 654                 char *filename = NULL;
 655                 int rc, lmmsize;
 656
 657                 rc = ll_get_max_mdsize(sbi, &lmmsize);
 658                 if (rc)
 659                         RETURN(rc);
 660
 661                 if (cmd == IOC_MDC_GETFILEINFO ||
 662                     cmd == IOC_MDC_GETFILESTRIPE) {
 663                         filename = getname((const char *)arg);
 664                         if (IS_ERR(filename))
 665                                 RETURN(PTR_ERR(filename));
 666
 667                         rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode),
 668                                              filename, strlen(filename) + 1,
 669                                              OBD_MD_FLEASIZE | OBD_MD_FLDIREA,
 670                                              lmmsize, &request);
 671                         if (rc < 0) {
 672                                 CDEBUG(D_INFO, "md_getattr_name failed "
 673                                        "on %s: rc %d\n", filename, rc);
 674                                 GOTO(out_name, rc);
 675                         }
 676                 } else {
 677                         rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode),
 678                                         OBD_MD_FLEASIZE | OBD_MD_FLDIREA,
 679                                         lmmsize, &request);
 680                         if (rc < 0) {
 681                                 CDEBUG(D_INFO, "md_getattr failed on inode "
 682                                        "%lu/%u: rc %d\n", inode->i_ino,
 683                                        inode->i_generation, rc);
 684                                 GOTO(out_name, rc);
 685                          }
 686                 }
 687
 688                 body = lustre_msg_buf(request->rq_repmsg, REPLY_REC_OFF,
 689                                       sizeof(*body));
 690                 LASSERT(body != NULL); /* checked by md_getattr_name */
 691                 /* swabbed by md_getattr_name */
 692                 LASSERT_REPSWABBED(request, REPLY_REC_OFF);
 693
 694                 lmmsize = body->eadatasize;
 695
 696                 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
 697                     lmmsize == 0) {
 698                         if (cmd == LL_IOC_LOV_GETSTRIPE ||
 699                             cmd == IOC_MDC_GETFILESTRIPE)
 700                                 GOTO(out_req, rc = -ENODATA);
 701                         else
 702                                 GOTO(skip_lmm, rc = 0);
 703                 }
 704
 705                 lmm = lustre_msg_buf(request->rq_repmsg, REPLY_REC_OFF + 1,
 706                                      lmmsize);
 707                 LASSERT(lmm != NULL);
 708                 LASSERT_REPSWABBED(request, REPLY_REC_OFF + 1);
 709
 710                 /*
 711                  * This is coming from the MDS, so is probably in
 712                  * little endian.  We convert it to host endian before
 713                  * passing it to userspace.
 714                  */
 715                 if (lmm->lmm_magic == __swab32(LOV_MAGIC)) {
 716                         lustre_swab_lov_user_md((struct lov_user_md *)lmm);
 717                         lustre_swab_lov_user_md_objects((struct lov_user_md *)lmm);
 718                 } else if (lmm->lmm_magic == __swab32(LOV_MAGIC_JOIN)) {
 719                         lustre_swab_lov_user_md_join((struct lov_user_md_join *)lmm);
 720                 }
 721                 if (lmm->lmm_magic == LOV_MAGIC_JOIN) {
 722                         struct lov_stripe_md *lsm;
 723                         struct lov_user_md_join *lmj;
 724                         int lmj_size, i, aindex = 0, rc;
 725
 726                         rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
 727                         if (rc < 0)
 728                                 GOTO(out_req, rc = -ENOMEM);
 729                         rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
 730                         if (rc)
 731                                 GOTO(out_free_memmd, rc);
 732
 733                         lmj_size = sizeof(struct lov_user_md_join) +
 734                                    lsm->lsm_stripe_count *
 735                                    sizeof(struct lov_user_ost_data_join);
 736                         OBD_ALLOC(lmj, lmj_size);
 737                         if (!lmj)
 738                                 GOTO(out_free_memmd, rc = -ENOMEM);
 739
 740                         memcpy(lmj, lmm, sizeof(struct lov_user_md_join));
 741                         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 742                                 struct lov_extent *lex =
 743                                         &lsm->lsm_array->lai_ext_array[aindex];
 744
 745                                 if (lex->le_loi_idx + lex->le_stripe_count <= i)
 746                                         aindex ++;
 747                                 CDEBUG(D_INFO, "aindex %d i %d l_extent_start "
 748                                        LPU64" len %d\n", aindex, i,
 749                                        lex->le_start, (int)lex->le_len);
 750                                 lmj->lmm_objects[i].l_extent_start =
 751                                         lex->le_start;
 752
 753                                 if ((int)lex->le_len == -1)
 754                                         lmj->lmm_objects[i].l_extent_end = -1;
 755                                 else
 756                                         lmj->lmm_objects[i].l_extent_end =
 757                                                 lex->le_start + lex->le_len;
 758                                 lmj->lmm_objects[i].l_object_id =
 759                                         lsm->lsm_oinfo[i].loi_id;
 760                                 lmj->lmm_objects[i].l_object_gr =
 761                                         lsm->lsm_oinfo[i].loi_gr;
 762                                 lmj->lmm_objects[i].l_ost_gen =
 763                                         lsm->lsm_oinfo[i].loi_ost_gen;
 764                                 lmj->lmm_objects[i].l_ost_idx =
 765                                         lsm->lsm_oinfo[i].loi_ost_idx;
 766                         }
 767                         lmm = (struct lov_mds_md *)lmj;
 768                         lmmsize = lmj_size;
 769                 out_free_memmd:
 770                         obd_free_memmd(sbi->ll_dt_exp, &lsm);
 771                         if (rc)
 772                                 GOTO(out_req, rc);
 773                 }
 774                 if (cmd == IOC_MDC_GETFILESTRIPE ||
 775                     cmd == LL_IOC_LOV_GETSTRIPE) {
 776                         lump = (struct lov_user_md *)arg;
 777                 } else {
 778                         struct lov_user_mds_data *lmdp;
 779                         lmdp = (struct lov_user_mds_data *)arg;
 780                         lump = &lmdp->lmd_lmm;
 781                 }
 782                 rc = copy_to_user(lump, lmm, lmmsize);
 783                 if (rc)
 784                         GOTO(out_lmm, rc = -EFAULT);
 785         skip_lmm:
 786                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
 787                         struct lov_user_mds_data *lmdp;
 788                         lstat_t st = { 0 };
 789
 790                         st.st_dev     = inode->i_sb->s_dev;
 791                         st.st_mode    = body->mode;
 792                         st.st_nlink   = body->nlink;
 793                         st.st_uid     = body->uid;
 794                         st.st_gid     = body->gid;
 795                         st.st_rdev    = body->rdev;
 796                         st.st_size    = body->size;
 797                         st.st_blksize = PAGE_SIZE;
 798                         st.st_blocks  = body->blocks;
 799                         st.st_atime   = body->atime;
 800                         st.st_mtime   = body->mtime;
 801                         st.st_ctime   = body->ctime;
 802                         st.st_ino     = inode->i_ino;
 803
 804                         lmdp = (struct lov_user_mds_data *)arg;
 805                         rc = copy_to_user(&lmdp->lmd_st, &st, sizeof(st));
 806                         if (rc)
 807                                 GOTO(out_lmm, rc = -EFAULT);
 808                 }
 809
 810                 EXIT;
 811         out_lmm:
 812                 if (lmm && lmm->lmm_magic == LOV_MAGIC_JOIN)
 813                         OBD_FREE(lmm, lmmsize);
 814         out_req:
 815                 ptlrpc_req_finished(request);
 816         out_name:
 817                 if (filename)
 818                         putname(filename);
 819                 return rc;
 820         }
 821         case IOC_LOV_GETINFO: {
 822                 struct lov_user_mds_data *lumd;
 823                 struct lov_stripe_md *lsm;
 824                 struct lov_user_md *lum;
 825                 struct lov_mds_md *lmm;
 826                 int lmmsize;
 827                 lstat_t st;
 828                 int rc;
 829
 830                 lumd = (struct lov_user_mds_data *)arg;
 831                 lum = &lumd->lmd_lmm;
 832
 833                 rc = ll_get_max_mdsize(sbi, &lmmsize);
 834                 if (rc)
 835                         RETURN(rc);
 836
 837                 OBD_ALLOC(lmm, lmmsize);
 838                 rc = copy_from_user(lmm, lum, lmmsize);
 839                 if (rc)
 840                         GOTO(free_lmm, rc = -EFAULT);
 841
 842                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
 843                 if (rc < 0)
 844                         GOTO(free_lmm, rc = -ENOMEM);
 845
 846                 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
 847                 if (rc)
 848                         GOTO(free_lsm, rc);
 849
 850                 /* Perform glimpse_size operation. */
 851                 memset(&st, 0, sizeof(st));
 852
 853                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
 854                 if (rc)
 855                         GOTO(free_lsm, rc);
 856
 857                 rc = copy_to_user(&lumd->lmd_st, &st, sizeof(st));
 858                 if (rc)
 859                         GOTO(free_lsm, rc = -EFAULT);
 860
 861                 EXIT;
 862         free_lsm:
 863                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
 864         free_lmm:
 865                 OBD_FREE(lmm, lmmsize);
 866                 return rc;
 867         }
 868         case OBD_IOC_LLOG_CATINFO: {
 869                 struct ptlrpc_request *req = NULL;
 870                 char *buf = NULL;
 871                 int rc, len = 0;
 872                 char *bufs[3] = { NULL }, *str;
 873                 int lens[3] = { sizeof(struct ptlrpc_body) };
 874                 int size[2] = { sizeof(struct ptlrpc_body) };
 875
 876                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
 877                 if (rc)
 878                         RETURN(rc);
 879                 data = (void *)buf;
 880
 881                 if (!data->ioc_inlbuf1) {
 882                         obd_ioctl_freedata(buf, len);
 883                         RETURN(-EINVAL);
 884                 }
 885
 886                 lens[REQ_REC_OFF] = data->ioc_inllen1;
 887                 bufs[REQ_REC_OFF] = data->ioc_inlbuf1;
 888                 if (data->ioc_inllen2) {
 889                         lens[REQ_REC_OFF + 1] = data->ioc_inllen2;
 890                         bufs[REQ_REC_OFF + 1] = data->ioc_inlbuf2;
 891                 } else {
 892                         lens[REQ_REC_OFF + 1] = 0;
 893                         bufs[REQ_REC_OFF + 1] = NULL;
 894                 }
 895
 896                 req = ptlrpc_prep_req(sbi2mdc(sbi)->cl_import,
 897                                       LUSTRE_LOG_VERSION, LLOG_CATINFO, 3, lens,
 898                                       bufs);
 899                 if (!req)
 900                         GOTO(out_catinfo, rc = -ENOMEM);
 901
 902                 size[REPLY_REC_OFF] = data->ioc_plen1;
 903                 ptlrpc_req_set_repsize(req, 2, size);
 904
 905                 rc = ptlrpc_queue_wait(req);
 906                 str = lustre_msg_string(req->rq_repmsg, REPLY_REC_OFF,
 907                                         data->ioc_plen1);
 908                 if (!rc)
 909                         rc = copy_to_user(data->ioc_pbuf1, str,data->ioc_plen1);
 910                 ptlrpc_req_finished(req);
 911         out_catinfo:
 912                 obd_ioctl_freedata(buf, len);
 913                 RETURN(rc);
 914         }
 915         case OBD_IOC_QUOTACHECK: {
 916                 struct obd_quotactl *oqctl;
 917                 int rc, error = 0;
 918
 919                 if (!capable(CAP_SYS_ADMIN))
 920                         RETURN(-EPERM);
 921
 922                 OBD_ALLOC_PTR(oqctl);
 923                 if (!oqctl)
 924                         RETURN(-ENOMEM);
 925                 oqctl->qc_type = arg;
 926                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
 927                 if (rc < 0) {
 928                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
 929                         error = rc;
 930                 }
 931
 932                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
 933                 if (rc < 0)
 934                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
 935
 936                 OBD_FREE_PTR(oqctl);
 937                 return error ?: rc;
 938         }
 939         case OBD_IOC_POLL_QUOTACHECK: {
 940                 struct if_quotacheck *check;
 941                 int rc;
 942
 943                 if (!capable(CAP_SYS_ADMIN))
 944                         RETURN(-EPERM);
 945
 946                 OBD_ALLOC_PTR(check);
 947                 if (!check)
 948                         RETURN(-ENOMEM);
 949
 950                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
 951                                    NULL);
 952                 if (rc) {
 953                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
 954                         if (copy_to_user((void *)arg, check, sizeof(*check)))
 955                                 rc = -EFAULT;
 956                         GOTO(out_poll, rc);
 957                 }
 958
 959                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
 960                                    NULL);
 961                 if (rc) {
 962                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
 963                         if (copy_to_user((void *)arg, check, sizeof(*check)))
 964                                 rc = -EFAULT;
 965                         GOTO(out_poll, rc);
 966                 }
 967         out_poll:
 968                 OBD_FREE_PTR(check);
 969                 RETURN(rc);
 970         }
 971 #if HAVE_QUOTA_SUPPORT
 972         case OBD_IOC_QUOTACTL: {
 973                 struct if_quotactl *qctl;
 974                 struct obd_quotactl *oqctl;
 975
 976                 int cmd, type, id, rc = 0;
 977
 978                 OBD_ALLOC_PTR(qctl);
 979                 if (!qctl)
 980                         RETURN(-ENOMEM);
 981
 982                 OBD_ALLOC_PTR(oqctl);
 983                 if (!oqctl) {
 984                         OBD_FREE_PTR(qctl);
 985                         RETURN(-ENOMEM);
 986                 }
 987                 if (copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
 988                         GOTO(out_quotactl, rc = -EFAULT);
 989
 990                 cmd = qctl->qc_cmd;
 991                 type = qctl->qc_type;
 992                 id = qctl->qc_id;
 993                 switch (cmd) {
 994                 case Q_QUOTAON:
 995                 case Q_QUOTAOFF:
 996                 case Q_SETQUOTA:
 997                 case Q_SETINFO:
 998                         if (!capable(CAP_SYS_ADMIN))
 999                                 GOTO(out_quotactl, rc = -EPERM);
1000                         break;
1001                 case Q_GETQUOTA:
1002                         if (((type == USRQUOTA && current->euid != id) ||
1003                              (type == GRPQUOTA && !in_egroup_p(id))) &&
1004                             !capable(CAP_SYS_ADMIN))
1005                                 GOTO(out_quotactl, rc = -EPERM);
1006
1007                         /* XXX: dqb_valid is borrowed as a flag to mark that
1008                          *      only mds quota is wanted */
1009                         if (qctl->qc_dqblk.dqb_valid)
1010                                 qctl->obd_uuid = sbi->ll_md_exp->exp_obd->
1011                                                         u.cli.cl_target_uuid;
1012                         break;
1013                 case Q_GETINFO:
1014                         break;
1015                 default:
1016                         CERROR("unsupported quotactl op: %#x\n", cmd);
1017                         GOTO(out_quotactl, -ENOTTY);
1018                 }
1019
1020                 QCTL_COPY(oqctl, qctl);
1021
1022                 if (qctl->obd_uuid.uuid[0]) {
1023                         struct obd_device *obd;
1024                         struct obd_uuid *uuid = &qctl->obd_uuid;
1025
1026                         obd = class_find_client_notype(uuid,
1027                                          &sbi->ll_dt_exp->exp_obd->obd_uuid);
1028                         if (!obd)
1029                                 GOTO(out_quotactl, rc = -ENOENT);
1030
1031                         if (cmd == Q_GETINFO)
1032                                 oqctl->qc_cmd = Q_GETOINFO;
1033                         else if (cmd == Q_GETQUOTA)
1034                                 oqctl->qc_cmd = Q_GETOQUOTA;
1035                         else
1036                                 GOTO(out_quotactl, rc = -EINVAL);
1037
1038                         if (sbi->ll_md_exp->exp_obd == obd) {
1039                                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1040                         } else {
1041                                 int i;
1042                                 struct obd_export *exp;
1043                                 struct lov_obd *lov = &sbi->ll_dt_exp->
1044                                                             exp_obd->u.lov;
1045
1046                                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1047                                         if (!lov->lov_tgts[i] ||
1048                                             !lov->lov_tgts[i]->ltd_active)
1049                                                 continue;
1050                                         exp = lov->lov_tgts[i]->ltd_exp;
1051                                         if (exp->exp_obd == obd) {
1052                                                 rc = obd_quotactl(exp, oqctl);
1053                                                 break;
1054                                         }
1055                                 }
1056                         }
1057
1058                         oqctl->qc_cmd = cmd;
1059                         QCTL_COPY(qctl, oqctl);
1060
1061                         if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1062                                 rc = -EFAULT;
1063
1064                         GOTO(out_quotactl, rc);
1065                 }
1066
1067                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1068                 if (rc && rc != -EBUSY && cmd == Q_QUOTAON) {
1069                         oqctl->qc_cmd = Q_QUOTAOFF;
1070                         obd_quotactl(sbi->ll_md_exp, oqctl);
1071                 }
1072
1073                 QCTL_COPY(qctl, oqctl);
1074
1075                 if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1076                         rc = -EFAULT;
1077         out_quotactl:
1078                 OBD_FREE_PTR(qctl);
1079                 OBD_FREE_PTR(oqctl);
1080                 RETURN(rc);
1081         }
1082 #endif /* HAVE_QUOTA_SUPPORT */
1083         case OBD_IOC_GETNAME: {
1084                 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1085                 if (!obd)
1086                         RETURN(-EFAULT);
1087                 if (copy_to_user((void *)arg, obd->obd_name,
1088                                 strlen(obd->obd_name) + 1))
1089                         RETURN (-EFAULT);
1090                 RETURN(0);
1091         }
1092         default:
1093                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1094         }
1095 }
1096
1097 int ll_dir_open(struct inode *inode, struct file *file)
1098 {
1099         ENTRY;
1100         RETURN(ll_file_open(inode, file));
1101 }
1102
1103 int ll_dir_release(struct inode *inode, struct file *file)
1104 {
1105         ENTRY;
1106         RETURN(ll_file_release(inode, file));
1107 }
1108
1109 struct file_operations ll_dir_operations = {
1110         .open     = ll_dir_open,
1111         .release  = ll_dir_release,
1112         .read     = generic_read_dir,
1113         .readdir  = ll_readdir,
1114         .llseek   = ll_llseek,
1115         .ioctl    = ll_dir_ioctl
1116 };
1117