Whamcloud - gitweb
LU-1576 llite: correct page usage count
[fs/lustre-release.git] / lustre / llite / dir.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/llite/dir.c
37  *
38  * Directory code for lustre client.
39  */
40
41 #include <linux/fs.h>
42 #include <linux/pagemap.h>
43 #include <linux/mm.h>
44 #include <linux/version.h>
45 #include <linux/smp_lock.h>
46 #include <asm/uaccess.h>
47 #include <linux/buffer_head.h>   // for wait_on_buffer
48 #include <linux/pagevec.h>
49
50 #define DEBUG_SUBSYSTEM S_LLITE
51
52 #include <obd_support.h>
53 #include <obd_class.h>
54 #include <lustre_lib.h>
55 #include <lustre/lustre_idl.h>
56 #include <lustre_lite.h>
57 #include <lustre_dlm.h>
58 #include <lustre_fid.h>
59 #include "llite_internal.h"
60
61 #ifndef HAVE_PAGE_CHECKED
62 #ifdef HAVE_PG_FS_MISC
63 #define PageChecked(page)        test_bit(PG_fs_misc, &(page)->flags)
64 #define SetPageChecked(page)     set_bit(PG_fs_misc, &(page)->flags)
65 #else
66 #error PageChecked or PageFsMisc not defined in kernel
67 #endif
68 #endif
69
70 /*
71  * (new) readdir implementation overview.
72  *
73  * Original lustre readdir implementation cached exact copy of raw directory
74  * pages on the client. These pages were indexed in client page cache by
75  * logical offset in the directory file. This design, while very simple and
76  * intuitive had some inherent problems:
77  *
78  *     . it implies that byte offset to the directory entry serves as a
79  *     telldir(3)/seekdir(3) cookie, but that offset is not stable: in
80  *     ext3/htree directory entries may move due to splits, and more
81  *     importantly,
82  *
83  *     . it is incompatible with the design of split directories for cmd3,
84  *     that assumes that names are distributed across nodes based on their
85  *     hash, and so readdir should be done in hash order.
86  *
87  * New readdir implementation does readdir in hash order, and uses hash of a
88  * file name as a telldir/seekdir cookie. This led to number of complications:
89  *
90  *     . hash is not unique, so it cannot be used to index cached directory
91  *     pages on the client (note, that it requires a whole pageful of hash
92  *     collided entries to cause two pages to have identical hashes);
93  *
94  *     . hash is not unique, so it cannot, strictly speaking, be used as an
95  *     entry cookie. ext3/htree has the same problem and lustre implementation
96  *     mimics their solution: seekdir(hash) positions directory at the first
97  *     entry with the given hash.
98  *
99  * Client side.
100  *
101  * 0. caching
102  *
103  * Client caches directory pages using hash of the first entry as an index. As
104  * noted above hash is not unique, so this solution doesn't work as is:
105  * special processing is needed for "page hash chains" (i.e., sequences of
106  * pages filled with entries all having the same hash value).
107  *
108  * First, such chains have to be detected. To this end, server returns to the
109  * client the hash of the first entry on the page next to one returned. When
110  * client detects that this hash is the same as hash of the first entry on the
111  * returned page, page hash collision has to be handled. Pages in the
112  * hash chain, except first one, are termed "overflow pages".
113  *
114  * Solution to index uniqueness problem is to not cache overflow
115  * pages. Instead, when page hash collision is detected, all overflow pages
116  * from emerging chain are immediately requested from the server and placed in
117  * a special data structure (struct ll_dir_chain). This data structure is used
118  * by ll_readdir() to process entries from overflow pages. When readdir
119  * invocation finishes, overflow pages are discarded. If page hash collision
120  * chain weren't completely processed, next call to readdir will again detect
121  * page hash collision, again read overflow pages in, process next portion of
122  * entries and again discard the pages. This is not as wasteful as it looks,
123  * because, given reasonable hash, page hash collisions are extremely rare.
124  *
125  * 1. directory positioning
126  *
127  * When seekdir(hash) is called, original
128  *
129  *
130  *
131  *
132  *
133  *
134  *
135  *
136  * Server.
137  *
138  * identification of and access to overflow pages
139  *
140  * page format
141  *
142  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
143  * a header lu_dirpage which describes the start/end hash, and whether this
144  * page is empty (contains no dir entry) or hash collide with next page.
145  * After client receives reply, several pages will be integrated into dir page
146  * in CFS_PAGE_SIZE (if CFS_PAGE_SIZE greater than LU_PAGE_SIZE), and the
147  * lu_dirpage for this integrated page will be adjusted.
148  *
149  */
150
151 /* returns the page unlocked, but with a reference */
152 static int ll_dir_readpage(struct file *file, struct page *page0)
153 {
154         struct inode *inode = page0->mapping->host;
155         int hash64 = ll_i2sbi(inode)->ll_flags & LL_SBI_64BIT_HASH;
156         struct obd_export *exp = ll_i2sbi(inode)->ll_md_exp;
157         struct ptlrpc_request *request;
158         struct mdt_body *body;
159         struct md_op_data *op_data;
160         __u64 hash;
161         struct page **page_pool;
162         struct page *page;
163 #ifndef HAVE_ADD_TO_PAGE_CACHE_LRU
164         struct pagevec lru_pvec;
165 #endif
166         struct lu_dirpage *dp;
167         int max_pages = ll_i2sbi(inode)->ll_md_brw_size >> CFS_PAGE_SHIFT;
168         int nrdpgs = 0; /* number of pages read actually */
169         int npages;
170         int i;
171         int rc;
172         ENTRY;
173
174         if (file) {
175                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
176
177                 hash = fd->fd_dir.lfd_next;
178         } else {
179                 hash = ll_i2info(inode)->lli_sa_pos;
180         }
181         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) hash "LPU64"\n",
182                inode->i_ino, inode->i_generation, inode, hash);
183
184         LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
185
186         OBD_ALLOC(page_pool, sizeof(page) * max_pages);
187         if (page_pool != NULL) {
188                 page_pool[0] = page0;
189         } else {
190                 page_pool = &page0;
191                 max_pages = 1;
192         }
193         for (npages = 1; npages < max_pages; npages++) {
194                 page = page_cache_alloc_cold(inode->i_mapping);
195                 if (!page)
196                         break;
197                 page_pool[npages] = page;
198         }
199
200         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
201                                      LUSTRE_OPC_ANY, NULL);
202         op_data->op_npages = npages;
203         op_data->op_offset = hash;
204         rc = md_readpage(exp, op_data, page_pool, &request);
205         ll_finish_md_op_data(op_data);
206         if (rc == 0) {
207                 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
208                 /* Checked by mdc_readpage() */
209                 LASSERT(body != NULL);
210
211                 if (body->valid & OBD_MD_FLSIZE)
212                         cl_isize_write(inode, body->size);
213
214                 nrdpgs = (request->rq_bulk->bd_nob_transferred+CFS_PAGE_SIZE-1)
215                          >> CFS_PAGE_SHIFT;
216                 SetPageUptodate(page0);
217         }
218         unlock_page(page0);
219         ptlrpc_req_finished(request);
220
221         CDEBUG(D_VFSTRACE, "read %d/%d pages\n", nrdpgs, npages);
222
223         ll_pagevec_init(&lru_pvec, 0);
224         for (i = 1; i < npages; i++) {
225                 unsigned long offset;
226                 int ret;
227
228                 page = page_pool[i];
229
230                 if (rc < 0 || i >= nrdpgs) {
231                         page_cache_release(page);
232                         continue;
233                 }
234
235                 SetPageUptodate(page);
236
237                 dp = cfs_kmap(page);
238                 hash = le64_to_cpu(dp->ldp_hash_start);
239                 cfs_kunmap(page);
240
241                 offset = hash_x_index(hash, hash64);
242
243                 prefetchw(&page->flags);
244                 ret = ll_add_to_page_cache_lru(page, inode->i_mapping, offset,
245                                                GFP_KERNEL);
246                 if (ret == 0) {
247                         unlock_page(page);
248                         if (ll_pagevec_add(&lru_pvec, page) == 0)
249                                 ll_pagevec_lru_add_file(&lru_pvec);
250                 } else {
251                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
252                                " %d\n", offset, ret);
253                 }
254                 page_cache_release(page);
255         }
256         ll_pagevec_lru_add_file(&lru_pvec);
257
258         if (page_pool != &page0)
259                 OBD_FREE(page_pool, sizeof(struct page *) * max_pages);
260         EXIT;
261         return rc;
262 }
263
264 #ifndef MS_HAS_NEW_AOPS
265 struct address_space_operations ll_dir_aops = {
266         .readpage  = ll_dir_readpage,
267 };
268 #else
269 struct address_space_operations_ext ll_dir_aops = {
270         .orig_aops.readpage  = ll_dir_readpage,
271 };
272 #endif
273
274 static void ll_check_page(struct inode *dir, struct page *page)
275 {
276         /* XXX: check page format later */
277         SetPageChecked(page);
278 }
279
280 void ll_release_page(struct page *page, int remove)
281 {
282         kunmap(page);
283         if (remove) {
284                 lock_page(page);
285                 if (likely(page->mapping != NULL))
286                         truncate_complete_page(page->mapping, page);
287                 unlock_page(page);
288         }
289         page_cache_release(page);
290 }
291
292 /*
293  * Find, kmap and return page that contains given hash.
294  */
295 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
296                                        __u64 *start, __u64 *end)
297 {
298         int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
299         struct address_space *mapping = dir->i_mapping;
300         /*
301          * Complement of hash is used as an index so that
302          * radix_tree_gang_lookup() can be used to find a page with starting
303          * hash _smaller_ than one we are looking for.
304          */
305         unsigned long offset = hash_x_index(*hash, hash64);
306         struct page *page;
307         int found;
308
309         TREE_READ_LOCK_IRQ(mapping);
310         found = radix_tree_gang_lookup(&mapping->page_tree,
311                                        (void **)&page, offset, 1);
312         if (found > 0) {
313                 struct lu_dirpage *dp;
314
315                 page_cache_get(page);
316                 TREE_READ_UNLOCK_IRQ(mapping);
317                 /*
318                  * In contrast to find_lock_page() we are sure that directory
319                  * page cannot be truncated (while DLM lock is held) and,
320                  * hence, can avoid restart.
321                  *
322                  * In fact, page cannot be locked here at all, because
323                  * ll_dir_readpage() does synchronous io.
324                  */
325                 wait_on_page(page);
326                 if (PageUptodate(page)) {
327                         dp = cfs_kmap(page);
328                         if (BITS_PER_LONG == 32 && hash64) {
329                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
330                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
331                                 *hash  = *hash >> 32;
332                         } else {
333                                 *start = le64_to_cpu(dp->ldp_hash_start);
334                                 *end   = le64_to_cpu(dp->ldp_hash_end);
335                         }
336                         LASSERTF(*start <= *hash, "start = "LPX64",end = "
337                                  LPX64",hash = "LPX64"\n", *start, *end, *hash);
338                         CDEBUG(D_VFSTRACE, "page %lu [%llu %llu], hash "LPU64"\n",
339                                offset, *start, *end, *hash);
340                         if (*hash > *end) {
341                                 ll_release_page(page, 0);
342                                 page = NULL;
343                         } else if (*end != *start && *hash == *end) {
344                                 /*
345                                  * upon hash collision, remove this page,
346                                  * otherwise put page reference, and
347                                  * ll_get_dir_page() will issue RPC to fetch
348                                  * the page we want.
349                                  */
350                                 ll_release_page(page,
351                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
352                                 page = NULL;
353                         }
354                 } else {
355                         page_cache_release(page);
356                         page = ERR_PTR(-EIO);
357                 }
358
359         } else {
360                 TREE_READ_UNLOCK_IRQ(mapping);
361                 page = NULL;
362         }
363         return page;
364 }
365
366 struct page *ll_get_dir_page(struct file *filp, struct inode *dir, __u64 hash,
367                              struct ll_dir_chain *chain)
368 {
369         ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
370         struct address_space *mapping = dir->i_mapping;
371         struct lustre_handle lockh;
372         struct lu_dirpage *dp;
373         struct page *page;
374         ldlm_mode_t mode;
375         int rc;
376         __u64 start = 0;
377         __u64 end = 0;
378         __u64 lhash = hash;
379         struct ll_inode_info *lli = ll_i2info(dir);
380         int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
381
382         mode = LCK_PR;
383         rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
384                            ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
385         if (!rc) {
386                 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
387                        ll_md_blocking_ast, ldlm_completion_ast,
388                        NULL, NULL, dir };
389                 struct lookup_intent it = { .it_op = IT_READDIR };
390                 struct ptlrpc_request *request;
391                 struct md_op_data *op_data;
392
393                 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
394                                              LUSTRE_OPC_ANY, NULL);
395                 if (IS_ERR(op_data))
396                         return (void *)op_data;
397
398                 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
399                                 op_data, &lockh, NULL, 0, NULL, 0);
400
401                 ll_finish_md_op_data(op_data);
402
403                 request = (struct ptlrpc_request *)it.d.lustre.it_data;
404                 if (request)
405                         ptlrpc_req_finished(request);
406                 if (rc < 0) {
407                         CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
408                                PFID(ll_inode2fid(dir)), hash, rc);
409                         return ERR_PTR(rc);
410                 }
411         } else {
412                 /* for cross-ref object, l_ast_data of the lock may not be set,
413                  * we reset it here */
414                 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
415                                  dir, NULL);
416         }
417         ldlm_lock_dump_handle(D_OTHER, &lockh);
418
419         cfs_mutex_lock(&lli->lli_readdir_mutex);
420         page = ll_dir_page_locate(dir, &lhash, &start, &end);
421         if (IS_ERR(page)) {
422                 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
423                        PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
424                 GOTO(out_unlock, page);
425         } else if (page != NULL) {
426                 /*
427                  * XXX nikita: not entirely correct handling of a corner case:
428                  * suppose hash chain of entries with hash value HASH crosses
429                  * border between pages P0 and P1. First both P0 and P1 are
430                  * cached, seekdir() is called for some entry from the P0 part
431                  * of the chain. Later P0 goes out of cache. telldir(HASH)
432                  * happens and finds P1, as it starts with matching hash
433                  * value. Remaining entries from P0 part of the chain are
434                  * skipped. (Is that really a bug?)
435                  *
436                  * Possible solutions: 0. don't cache P1 is such case, handle
437                  * it as an "overflow" page. 1. invalidate all pages at
438                  * once. 2. use HASH|1 as an index for P1.
439                  */
440                 GOTO(hash_collision, page);
441         }
442
443         page = read_cache_page(mapping, hash_x_index(hash, hash64),
444                                (filler_t*)mapping->a_ops->readpage, filp);
445         if (IS_ERR(page)) {
446                 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
447                        PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
448                 GOTO(out_unlock, page);
449         }
450
451         wait_on_page(page);
452         (void)kmap(page);
453         if (!PageUptodate(page)) {
454                 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
455                        PFID(ll_inode2fid(dir)), hash, -5);
456                 goto fail;
457         }
458         if (!PageChecked(page))
459                 ll_check_page(dir, page);
460         if (PageError(page)) {
461                 CERROR("page error: "DFID" at "LPU64": rc %d\n",
462                        PFID(ll_inode2fid(dir)), hash, -5);
463                 goto fail;
464         }
465 hash_collision:
466         dp = page_address(page);
467         if (BITS_PER_LONG == 32 && hash64) {
468                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
469                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
470                 lhash = hash >> 32;
471         } else {
472                 start = le64_to_cpu(dp->ldp_hash_start);
473                 end   = le64_to_cpu(dp->ldp_hash_end);
474                 lhash = hash;
475         }
476         if (end == start) {
477                 LASSERT(start == lhash);
478                 CWARN("Page-wide hash collision: "LPU64"\n", end);
479                 if (BITS_PER_LONG == 32 && hash64)
480                         CWARN("Real page-wide hash collision at ["LPU64" "LPU64
481                               "] with hash "LPU64"\n",
482                               le64_to_cpu(dp->ldp_hash_start),
483                               le64_to_cpu(dp->ldp_hash_end), hash);
484                 /*
485                  * Fetch whole overflow chain...
486                  *
487                  * XXX not yet.
488                  */
489                 goto fail;
490         }
491 out_unlock:
492         cfs_mutex_unlock(&lli->lli_readdir_mutex);
493         ldlm_lock_decref(&lockh, mode);
494         return page;
495
496 fail:
497         ll_release_page(page, 1);
498         page = ERR_PTR(-EIO);
499         goto out_unlock;
500 }
501
502 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
503 {
504         struct inode         *inode      = filp->f_dentry->d_inode;
505         struct ll_inode_info *info       = ll_i2info(inode);
506         struct ll_sb_info    *sbi        = ll_i2sbi(inode);
507         struct ll_file_data  *fd         = LUSTRE_FPRIVATE(filp);
508         __u64                 pos        = fd->fd_dir.lfd_pos;
509         int                   api32      = ll_need_32bit_api(sbi);
510         int                   hash64     = sbi->ll_flags & LL_SBI_64BIT_HASH;
511         struct page          *page;
512         struct ll_dir_chain   chain;
513         int                   done;
514         int                   rc;
515         ENTRY;
516
517         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu 32bit_api %d\n",
518                inode->i_ino, inode->i_generation, inode,
519                (unsigned long)pos, i_size_read(inode), api32);
520
521         if (pos == MDS_DIR_END_OFF)
522                 /*
523                  * end-of-file.
524                  */
525                 GOTO(out, rc = 0);
526
527         rc    = 0;
528         done  = 0;
529         ll_dir_chain_init(&chain);
530
531         fd->fd_dir.lfd_next = pos;
532         page = ll_get_dir_page(filp, inode, pos, &chain);
533
534         while (rc == 0 && !done) {
535                 struct lu_dirpage *dp;
536                 struct lu_dirent  *ent;
537
538                 if (!IS_ERR(page)) {
539                         /*
540                          * If page is empty (end of directory is reached),
541                          * use this value.
542                          */
543                         __u64 hash = MDS_DIR_END_OFF;
544                         __u64 next;
545
546                         dp = page_address(page);
547                         for (ent = lu_dirent_start(dp); ent != NULL && !done;
548                              ent = lu_dirent_next(ent)) {
549                                 __u16          type;
550                                 int            namelen;
551                                 struct lu_fid  fid;
552                                 __u64          lhash;
553                                 __u64          ino;
554
555                                 /*
556                                  * XXX: implement correct swabbing here.
557                                  */
558
559                                 hash = le64_to_cpu(ent->lde_hash);
560                                 if (hash < pos)
561                                         /*
562                                          * Skip until we find target hash
563                                          * value.
564                                          */
565                                         continue;
566
567                                 namelen = le16_to_cpu(ent->lde_namelen);
568                                 if (namelen == 0)
569                                         /*
570                                          * Skip dummy record.
571                                          */
572                                         continue;
573
574                                 if (api32 && hash64)
575                                         lhash = hash >> 32;
576                                 else
577                                         lhash = hash;
578                                 fid_le_to_cpu(&fid, &ent->lde_fid);
579                                 ino = cl_fid_build_ino(&fid, api32);
580                                 type = ll_dirent_type_get(ent);
581                                 /* For 'll_nfs_get_name_filldir()', it will try
582                                  * to access the 'ent' through its 'lde_name',
583                                  * so the parameter 'name' for 'filldir()' must
584                                  * be part of the 'ent'. */
585                                 done = filldir(cookie, ent->lde_name, namelen,
586                                                lhash, ino, type);
587                         }
588                         next = le64_to_cpu(dp->ldp_hash_end);
589                         if (!done) {
590                                 pos = next;
591                                 if (pos == MDS_DIR_END_OFF) {
592                                         /*
593                                          * End of directory reached.
594                                          */
595                                         done = 1;
596                                         ll_release_page(page, 0);
597                                 } else if (1 /* chain is exhausted*/) {
598                                         /*
599                                          * Normal case: continue to the next
600                                          * page.
601                                          */
602                                         ll_release_page(page,
603                                             le32_to_cpu(dp->ldp_flags) &
604                                                         LDF_COLLIDE);
605                                         fd->fd_dir.lfd_next = pos;
606                                         page = ll_get_dir_page(filp, inode, pos,
607                                                                &chain);
608                                 } else {
609                                         /*
610                                          * go into overflow page.
611                                          */
612                                         LASSERT(le32_to_cpu(dp->ldp_flags) &
613                                                 LDF_COLLIDE);
614                                         ll_release_page(page, 1);
615                                 }
616                         } else {
617                                 pos = hash;
618                                 ll_release_page(page, 0);
619                         }
620                 } else {
621                         rc = PTR_ERR(page);
622                         CERROR("error reading dir "DFID" at %lu: rc %d\n",
623                                PFID(&info->lli_fid), (unsigned long)pos, rc);
624                 }
625         }
626
627         fd->fd_dir.lfd_pos = pos;
628         if (pos == MDS_DIR_END_OFF) {
629                 if (api32)
630                         filp->f_pos = LL_DIR_END_OFF_32BIT;
631                 else
632                         filp->f_pos = LL_DIR_END_OFF;
633         } else {
634                 if (api32 && hash64)
635                         filp->f_pos = pos >> 32;
636                 else
637                         filp->f_pos = pos;
638         }
639         filp->f_version = inode->i_version;
640         touch_atime(filp->f_vfsmnt, filp->f_dentry);
641
642         ll_dir_chain_fini(&chain);
643
644 out:
645         if (!rc)
646                 ll_stats_ops_tally(sbi, LPROC_LL_READDIR, 1);
647
648         RETURN(rc);
649 }
650
651 int ll_send_mgc_param(struct obd_export *mgc, char *string)
652 {
653         struct mgs_send_param *msp;
654         int rc = 0;
655
656         OBD_ALLOC_PTR(msp);
657         if (!msp)
658                 return -ENOMEM;
659
660         strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
661         rc = obd_set_info_async(NULL, mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
662                                 sizeof(struct mgs_send_param), msp, NULL);
663         if (rc)
664                 CERROR("Failed to set parameter: %d\n", rc);
665         OBD_FREE_PTR(msp);
666
667         return rc;
668 }
669
670 char *ll_get_fsname(struct inode *inode)
671 {
672         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
673         char *ptr, *fsname;
674         int len;
675
676         OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
677         len = strlen(lsi->lsi_lmd->lmd_profile);
678         ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
679         if (ptr && (strcmp(ptr, "-client") == 0))
680                 len -= 7;
681         strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
682         fsname[len] = '\0';
683
684         return fsname;
685 }
686
687 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
688                      int set_default)
689 {
690         struct ll_sb_info *sbi = ll_i2sbi(inode);
691         struct md_op_data *op_data;
692         struct ptlrpc_request *req = NULL;
693         int rc = 0;
694         struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
695         struct obd_device *mgc = lsi->lsi_mgc;
696         char *fsname = NULL, *param = NULL;
697         int lum_size;
698
699         if (lump != NULL) {
700                 /*
701                  * This is coming from userspace, so should be in
702                  * local endian.  But the MDS would like it in little
703                  * endian, so we swab it before we send it.
704                  */
705                 switch (lump->lmm_magic) {
706                 case LOV_USER_MAGIC_V1: {
707                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
708                                 lustre_swab_lov_user_md_v1(lump);
709                         lum_size = sizeof(struct lov_user_md_v1);
710                         break;
711                         }
712                 case LOV_USER_MAGIC_V3: {
713                         if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
714                                 lustre_swab_lov_user_md_v3(
715                                         (struct lov_user_md_v3 *)lump);
716                         lum_size = sizeof(struct lov_user_md_v3);
717                         break;
718                         }
719                 default: {
720                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
721                                         " %#08x != %#08x nor %#08x\n",
722                                         lump->lmm_magic, LOV_USER_MAGIC_V1,
723                                         LOV_USER_MAGIC_V3);
724                         RETURN(-EINVAL);
725                         }
726                }
727         } else {
728                 lum_size = sizeof(struct lov_user_md_v1);
729         }
730
731         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
732                                      LUSTRE_OPC_ANY, NULL);
733         if (IS_ERR(op_data))
734                 RETURN(PTR_ERR(op_data));
735
736         /* swabbing is done in lov_setstripe() on server side */
737         rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
738                         NULL, 0, &req, NULL);
739         ll_finish_md_op_data(op_data);
740         ptlrpc_req_finished(req);
741         if (rc) {
742                 if (rc != -EPERM && rc != -EACCES)
743                         CERROR("mdc_setattr fails: rc = %d\n", rc);
744         }
745
746         /* In the following we use the fact that LOV_USER_MAGIC_V1 and
747          LOV_USER_MAGIC_V3 have the same initial fields so we do not
748          need the make the distiction between the 2 versions */
749         if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
750                 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
751
752                 /* Get fsname and assume devname to be -MDT0000. */
753                 fsname = ll_get_fsname(inode);
754                 /* Set root stripesize */
755                 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
756                         lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
757                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
758                 if (rc)
759                         goto end;
760
761                 /* Set root stripecount */
762                 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
763                         lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
764                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
765                 if (rc)
766                         goto end;
767
768                 /* Set root stripeoffset */
769                 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
770                         lump ? le16_to_cpu(lump->lmm_stripe_offset) :
771                         (typeof(lump->lmm_stripe_offset))(-1));
772                 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
773                 if (rc)
774                         goto end;
775 end:
776                 if (fsname)
777                         OBD_FREE(fsname, MGS_PARAM_MAXLEN);
778                 if (param)
779                         OBD_FREE(param, MGS_PARAM_MAXLEN);
780         }
781         return rc;
782 }
783
784 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
785                      int *lmm_size, struct ptlrpc_request **request)
786 {
787         struct ll_sb_info *sbi = ll_i2sbi(inode);
788         struct mdt_body   *body;
789         struct lov_mds_md *lmm = NULL;
790         struct ptlrpc_request *req = NULL;
791         int rc, lmmsize;
792         struct md_op_data *op_data;
793
794         rc = ll_get_max_mdsize(sbi, &lmmsize);
795         if (rc)
796                 RETURN(rc);
797
798         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
799                                      0, lmmsize, LUSTRE_OPC_ANY,
800                                      NULL);
801         if (IS_ERR(op_data))
802                 RETURN(PTR_ERR(op_data));
803
804         op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
805         rc = md_getattr(sbi->ll_md_exp, op_data, &req);
806         ll_finish_md_op_data(op_data);
807         if (rc < 0) {
808                 CDEBUG(D_INFO, "md_getattr failed on inode "
809                        "%lu/%u: rc %d\n", inode->i_ino,
810                        inode->i_generation, rc);
811                 GOTO(out, rc);
812         }
813
814         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
815         LASSERT(body != NULL);
816
817         lmmsize = body->eadatasize;
818
819         if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
820             lmmsize == 0) {
821                 GOTO(out, rc = -ENODATA);
822         }
823
824         lmm = req_capsule_server_sized_get(&req->rq_pill,
825                                            &RMF_MDT_MD, lmmsize);
826         LASSERT(lmm != NULL);
827
828         /*
829          * This is coming from the MDS, so is probably in
830          * little endian.  We convert it to host endian before
831          * passing it to userspace.
832          */
833         /* We don't swab objects for directories */
834         switch (le32_to_cpu(lmm->lmm_magic)) {
835         case LOV_MAGIC_V1:
836                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
837                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
838                 break;
839         case LOV_MAGIC_V3:
840                 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
841                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
842                 break;
843         default:
844                 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
845                 rc = -EPROTO;
846         }
847 out:
848         *lmmp = lmm;
849         *lmm_size = lmmsize;
850         *request = req;
851         return rc;
852 }
853
854 /*
855  *  Get MDT index for the inode.
856  */
857 int ll_get_mdt_idx(struct inode *inode)
858 {
859         struct ll_sb_info *sbi = ll_i2sbi(inode);
860         struct md_op_data *op_data;
861         int rc, mdtidx;
862         ENTRY;
863
864         op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
865                                      0, LUSTRE_OPC_ANY, NULL);
866         if (IS_ERR(op_data))
867                 RETURN(PTR_ERR(op_data));
868
869         op_data->op_valid |= OBD_MD_MDTIDX;
870         rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
871         mdtidx = op_data->op_mds;
872         ll_finish_md_op_data(op_data);
873         if (rc < 0) {
874                 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
875                 RETURN(rc);
876         }
877         return mdtidx;
878 }
879
880 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
881 {
882         void *ptr;
883         int rc;
884
885         OBD_ALLOC(ptr, len);
886         if (ptr == NULL)
887                 return -ENOMEM;
888         if (cfs_copy_from_user(ptr, data, len)) {
889                 OBD_FREE(ptr, len);
890                 return -EFAULT;
891         }
892         rc = obd_iocontrol(cmd, exp, len, data, NULL);
893         OBD_FREE(ptr, len);
894         return rc;
895 }
896
897 static int quotactl_ioctl(struct ll_sb_info *sbi, struct if_quotactl *qctl)
898 {
899         int cmd = qctl->qc_cmd;
900         int type = qctl->qc_type;
901         int id = qctl->qc_id;
902         int valid = qctl->qc_valid;
903         int rc = 0;
904         ENTRY;
905
906         switch (cmd) {
907         case LUSTRE_Q_INVALIDATE:
908         case LUSTRE_Q_FINVALIDATE:
909         case Q_QUOTAON:
910         case Q_QUOTAOFF:
911         case Q_SETQUOTA:
912         case Q_SETINFO:
913                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
914                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
915                         RETURN(-EPERM);
916                 break;
917         case Q_GETQUOTA:
918                 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
919                      (type == GRPQUOTA && !in_egroup_p(id))) &&
920                     (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
921                      sbi->ll_flags & LL_SBI_RMT_CLIENT))
922                         RETURN(-EPERM);
923                 break;
924         case Q_GETINFO:
925                 break;
926         default:
927                 CERROR("unsupported quotactl op: %#x\n", cmd);
928                 RETURN(-ENOTTY);
929         }
930
931         if (valid != QC_GENERAL) {
932                 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
933                         RETURN(-EOPNOTSUPP);
934
935                 if (cmd == Q_GETINFO)
936                         qctl->qc_cmd = Q_GETOINFO;
937                 else if (cmd == Q_GETQUOTA)
938                         qctl->qc_cmd = Q_GETOQUOTA;
939                 else
940                         RETURN(-EINVAL);
941
942                 switch (valid) {
943                 case QC_MDTIDX:
944                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
945                                            sizeof(*qctl), qctl, NULL);
946                         break;
947                 case QC_OSTIDX:
948                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_dt_exp,
949                                            sizeof(*qctl), qctl, NULL);
950                         break;
951                 case QC_UUID:
952                         rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
953                                            sizeof(*qctl), qctl, NULL);
954                         if (rc == -EAGAIN)
955                                 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
956                                                    sbi->ll_dt_exp,
957                                                    sizeof(*qctl), qctl, NULL);
958                         break;
959                 default:
960                         rc = -EINVAL;
961                         break;
962                 }
963
964                 if (rc)
965                         RETURN(rc);
966
967                 qctl->qc_cmd = cmd;
968         } else {
969                 struct obd_quotactl *oqctl;
970
971                 OBD_ALLOC_PTR(oqctl);
972                 if (oqctl == NULL)
973                         RETURN(-ENOMEM);
974
975                 QCTL_COPY(oqctl, qctl);
976                 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
977                 if (rc) {
978                         if (rc != -EALREADY && cmd == Q_QUOTAON) {
979                                 oqctl->qc_cmd = Q_QUOTAOFF;
980                                 obd_quotactl(sbi->ll_md_exp, oqctl);
981                         }
982                         OBD_FREE_PTR(oqctl);
983                         RETURN(rc);
984                 }
985                 /* If QIF_SPACE is not set, client should collect the
986                  * space usage from OSSs by itself */
987                 if (cmd == Q_GETQUOTA &&
988                     !(oqctl->qc_dqblk.dqb_valid & QIF_SPACE) &&
989                     !oqctl->qc_dqblk.dqb_curspace) {
990                         struct obd_quotactl *oqctl_tmp;
991
992                         OBD_ALLOC_PTR(oqctl_tmp);
993                         if (oqctl_tmp == NULL)
994                                 GOTO(out, rc = -ENOMEM);
995
996                         oqctl_tmp->qc_cmd = Q_GETOQUOTA;
997                         oqctl_tmp->qc_id = oqctl->qc_id;
998                         oqctl_tmp->qc_type = oqctl->qc_type;
999
1000                         /* collect space usage from OSTs */
1001                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1002                         rc = obd_quotactl(sbi->ll_dt_exp, oqctl_tmp);
1003                         if (!rc || rc == -EREMOTEIO) {
1004                                 oqctl->qc_dqblk.dqb_curspace =
1005                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1006                                 oqctl->qc_dqblk.dqb_valid |= QIF_SPACE;
1007                         }
1008
1009                         /* collect space & inode usage from MDTs */
1010                         oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1011                         oqctl_tmp->qc_dqblk.dqb_curinodes = 0;
1012                         rc = obd_quotactl(sbi->ll_md_exp, oqctl_tmp);
1013                         if (!rc || rc == -EREMOTEIO) {
1014                                 oqctl->qc_dqblk.dqb_curspace +=
1015                                         oqctl_tmp->qc_dqblk.dqb_curspace;
1016                                 oqctl->qc_dqblk.dqb_curinodes =
1017                                         oqctl_tmp->qc_dqblk.dqb_curinodes;
1018                                 oqctl->qc_dqblk.dqb_valid |= QIF_INODES;
1019                         } else {
1020                                 oqctl->qc_dqblk.dqb_valid &= ~QIF_SPACE;
1021                         }
1022
1023                         OBD_FREE_PTR(oqctl_tmp);
1024                 }
1025 out:
1026                 QCTL_COPY(qctl, oqctl);
1027                 OBD_FREE_PTR(oqctl);
1028         }
1029
1030         RETURN(rc);
1031 }
1032
1033 static long ll_dir_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1034 {
1035         struct inode *inode = file->f_dentry->d_inode;
1036         struct ll_sb_info *sbi = ll_i2sbi(inode);
1037         struct obd_ioctl_data *data;
1038         int rc = 0;
1039         ENTRY;
1040
1041         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
1042                inode->i_ino, inode->i_generation, inode, cmd);
1043
1044         /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
1045         if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
1046                 return -ENOTTY;
1047
1048         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
1049         switch(cmd) {
1050         case FSFILT_IOC_GETFLAGS:
1051         case FSFILT_IOC_SETFLAGS:
1052                 RETURN(ll_iocontrol(inode, file, cmd, arg));
1053         case FSFILT_IOC_GETVERSION_OLD:
1054         case FSFILT_IOC_GETVERSION:
1055                 RETURN(put_user(inode->i_generation, (int *)arg));
1056         /* We need to special case any other ioctls we want to handle,
1057          * to send them to the MDS/OST as appropriate and to properly
1058          * network encode the arg field.
1059         case FSFILT_IOC_SETVERSION_OLD:
1060         case FSFILT_IOC_SETVERSION:
1061         */
1062         case LL_IOC_GET_MDTIDX: {
1063                 int mdtidx;
1064
1065                 mdtidx = ll_get_mdt_idx(inode);
1066                 if (mdtidx < 0)
1067                         RETURN(mdtidx);
1068
1069                 if (put_user((int)mdtidx, (int*)arg))
1070                         RETURN(-EFAULT);
1071
1072                 return 0;
1073         }
1074         case IOC_MDC_LOOKUP: {
1075                 struct ptlrpc_request *request = NULL;
1076                 int namelen, len = 0;
1077                 char *buf = NULL;
1078                 char *filename;
1079                 struct md_op_data *op_data;
1080
1081                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1082                 if (rc)
1083                         RETURN(rc);
1084                 data = (void *)buf;
1085
1086                 filename = data->ioc_inlbuf1;
1087                 namelen = strlen(filename);
1088
1089                 if (namelen < 1) {
1090                         CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1091                         GOTO(out_free, rc = -EINVAL);
1092                 }
1093
1094                 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1095                                              0, LUSTRE_OPC_ANY, NULL);
1096                 if (IS_ERR(op_data))
1097                         GOTO(out_free, rc = PTR_ERR(op_data));
1098
1099                 op_data->op_valid = OBD_MD_FLID;
1100                 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1101                 ll_finish_md_op_data(op_data);
1102                 if (rc < 0) {
1103                         CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1104                         GOTO(out_free, rc);
1105                 }
1106                 ptlrpc_req_finished(request);
1107                 EXIT;
1108 out_free:
1109                 obd_ioctl_freedata(buf, len);
1110                 return rc;
1111         }
1112         case LL_IOC_LOV_SETSTRIPE: {
1113                 struct lov_user_md_v3 lumv3;
1114                 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1115                 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1116                 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1117
1118                 int set_default = 0;
1119
1120                 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1121                 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1122                         sizeof(lumv3p->lmm_objects[0]));
1123                 /* first try with v1 which is smaller than v3 */
1124                 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1125                         RETURN(-EFAULT);
1126
1127                 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1128                         if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1129                                 RETURN(-EFAULT);
1130                 }
1131
1132                 if (inode->i_sb->s_root == file->f_dentry)
1133                         set_default = 1;
1134
1135                 /* in v1 and v3 cases lumv1 points to data */
1136                 rc = ll_dir_setstripe(inode, lumv1, set_default);
1137
1138                 RETURN(rc);
1139         }
1140         case LL_IOC_OBD_STATFS:
1141                 RETURN(ll_obd_statfs(inode, (void *)arg));
1142         case LL_IOC_LOV_GETSTRIPE:
1143         case LL_IOC_MDC_GETINFO:
1144         case IOC_MDC_GETFILEINFO:
1145         case IOC_MDC_GETFILESTRIPE: {
1146                 struct ptlrpc_request *request = NULL;
1147                 struct lov_user_md *lump;
1148                 struct lov_mds_md *lmm = NULL;
1149                 struct mdt_body *body;
1150                 char *filename = NULL;
1151                 int lmmsize;
1152
1153                 if (cmd == IOC_MDC_GETFILEINFO ||
1154                     cmd == IOC_MDC_GETFILESTRIPE) {
1155                         filename = getname((const char *)arg);
1156                         if (IS_ERR(filename))
1157                                 RETURN(PTR_ERR(filename));
1158
1159                         rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1160                                                       &lmmsize, &request);
1161                 } else {
1162                         rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1163                 }
1164
1165                 if (request) {
1166                         body = req_capsule_server_get(&request->rq_pill,
1167                                                       &RMF_MDT_BODY);
1168                         LASSERT(body != NULL);
1169                 } else {
1170                         GOTO(out_req, rc);
1171                 }
1172
1173                 if (rc < 0) {
1174                         if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1175                                                cmd == LL_IOC_MDC_GETINFO))
1176                                 GOTO(skip_lmm, rc = 0);
1177                         else
1178                                 GOTO(out_req, rc);
1179                 }
1180
1181                 if (cmd == IOC_MDC_GETFILESTRIPE ||
1182                     cmd == LL_IOC_LOV_GETSTRIPE) {
1183                         lump = (struct lov_user_md *)arg;
1184                 } else {
1185                         struct lov_user_mds_data *lmdp;
1186                         lmdp = (struct lov_user_mds_data *)arg;
1187                         lump = &lmdp->lmd_lmm;
1188                 }
1189                 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
1190                         if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
1191                                 GOTO(out_req, rc = -EFAULT);
1192                         rc = -EOVERFLOW;
1193                 }
1194         skip_lmm:
1195                 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1196                         struct lov_user_mds_data *lmdp;
1197                         lstat_t st = { 0 };
1198
1199                         st.st_dev     = inode->i_sb->s_dev;
1200                         st.st_mode    = body->mode;
1201                         st.st_nlink   = body->nlink;
1202                         st.st_uid     = body->uid;
1203                         st.st_gid     = body->gid;
1204                         st.st_rdev    = body->rdev;
1205                         st.st_size    = body->size;
1206                         st.st_blksize = CFS_PAGE_SIZE;
1207                         st.st_blocks  = body->blocks;
1208                         st.st_atime   = body->atime;
1209                         st.st_mtime   = body->mtime;
1210                         st.st_ctime   = body->ctime;
1211                         st.st_ino     = inode->i_ino;
1212
1213                         lmdp = (struct lov_user_mds_data *)arg;
1214                         if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1215                                 GOTO(out_req, rc = -EFAULT);
1216                 }
1217
1218                 EXIT;
1219         out_req:
1220                 ptlrpc_req_finished(request);
1221                 if (filename)
1222                         putname(filename);
1223                 return rc;
1224         }
1225         case IOC_LOV_GETINFO: {
1226                 struct lov_user_mds_data *lumd;
1227                 struct lov_stripe_md *lsm;
1228                 struct lov_user_md *lum;
1229                 struct lov_mds_md *lmm;
1230                 int lmmsize;
1231                 lstat_t st;
1232
1233                 lumd = (struct lov_user_mds_data *)arg;
1234                 lum = &lumd->lmd_lmm;
1235
1236                 rc = ll_get_max_mdsize(sbi, &lmmsize);
1237                 if (rc)
1238                         RETURN(rc);
1239
1240                 OBD_ALLOC_LARGE(lmm, lmmsize);
1241                 if (cfs_copy_from_user(lmm, lum, lmmsize))
1242                         GOTO(free_lmm, rc = -EFAULT);
1243
1244                 switch (lmm->lmm_magic) {
1245                 case LOV_USER_MAGIC_V1:
1246                         if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1247                                 break;
1248                         /* swab objects first so that stripes num will be sane */
1249                         lustre_swab_lov_user_md_objects(
1250                                 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1251                                 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1252                         lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1253                         break;
1254                 case LOV_USER_MAGIC_V3:
1255                         if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1256                                 break;
1257                         /* swab objects first so that stripes num will be sane */
1258                         lustre_swab_lov_user_md_objects(
1259                                 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1260                                 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1261                         lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1262                         break;
1263                 default:
1264                         GOTO(free_lmm, rc = -EINVAL);
1265                 }
1266
1267                 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1268                 if (rc < 0)
1269                         GOTO(free_lmm, rc = -ENOMEM);
1270
1271                 /* Perform glimpse_size operation. */
1272                 memset(&st, 0, sizeof(st));
1273
1274                 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1275                 if (rc)
1276                         GOTO(free_lsm, rc);
1277
1278                 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1279                         GOTO(free_lsm, rc = -EFAULT);
1280
1281                 EXIT;
1282         free_lsm:
1283                 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1284         free_lmm:
1285                 OBD_FREE_LARGE(lmm, lmmsize);
1286                 return rc;
1287         }
1288         case OBD_IOC_LLOG_CATINFO: {
1289                 struct ptlrpc_request *req = NULL;
1290                 char                  *buf = NULL;
1291                 char                  *str;
1292                 int                    len = 0;
1293
1294                 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1295                 if (rc)
1296                         RETURN(rc);
1297                 data = (void *)buf;
1298
1299                 if (!data->ioc_inlbuf1) {
1300                         obd_ioctl_freedata(buf, len);
1301                         RETURN(-EINVAL);
1302                 }
1303
1304                 req = ptlrpc_request_alloc(sbi2mdc(sbi)->cl_import,
1305                                            &RQF_LLOG_CATINFO);
1306                 if (req == NULL)
1307                         GOTO(out_catinfo, rc = -ENOMEM);
1308
1309                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
1310                                      data->ioc_inllen1);
1311                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_CLIENT,
1312                                      data->ioc_inllen2);
1313
1314                 rc = ptlrpc_request_pack(req, LUSTRE_LOG_VERSION, LLOG_CATINFO);
1315                 if (rc) {
1316                         ptlrpc_request_free(req);
1317                         GOTO(out_catinfo, rc);
1318                 }
1319
1320                 str = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
1321                 memcpy(str, data->ioc_inlbuf1, data->ioc_inllen1);
1322                 if (data->ioc_inllen2) {
1323                         str = req_capsule_client_get(&req->rq_pill,
1324                                                      &RMF_STRING);
1325                         memcpy(str, data->ioc_inlbuf2, data->ioc_inllen2);
1326                 }
1327
1328                 req_capsule_set_size(&req->rq_pill, &RMF_STRING, RCL_SERVER,
1329                                      data->ioc_plen1);
1330                 ptlrpc_request_set_replen(req);
1331
1332                 rc = ptlrpc_queue_wait(req);
1333                 if (!rc) {
1334                         str = req_capsule_server_get(&req->rq_pill,
1335                                                      &RMF_STRING);
1336                         if (cfs_copy_to_user(data->ioc_pbuf1, str,
1337                                              data->ioc_plen1))
1338                                 rc = -EFAULT;
1339                 }
1340                 ptlrpc_req_finished(req);
1341         out_catinfo:
1342                 obd_ioctl_freedata(buf, len);
1343                 RETURN(rc);
1344         }
1345         case OBD_IOC_QUOTACHECK: {
1346                 struct obd_quotactl *oqctl;
1347                 int error = 0;
1348
1349                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1350                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1351                         RETURN(-EPERM);
1352
1353                 OBD_ALLOC_PTR(oqctl);
1354                 if (!oqctl)
1355                         RETURN(-ENOMEM);
1356                 oqctl->qc_type = arg;
1357                 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1358                 if (rc < 0) {
1359                         CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1360                         error = rc;
1361                 }
1362
1363                 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1364                 if (rc < 0)
1365                         CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1366
1367                 OBD_FREE_PTR(oqctl);
1368                 return error ?: rc;
1369         }
1370         case OBD_IOC_POLL_QUOTACHECK: {
1371                 struct if_quotacheck *check;
1372
1373                 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1374                     sbi->ll_flags & LL_SBI_RMT_CLIENT)
1375                         RETURN(-EPERM);
1376
1377                 OBD_ALLOC_PTR(check);
1378                 if (!check)
1379                         RETURN(-ENOMEM);
1380
1381                 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1382                                    NULL);
1383                 if (rc) {
1384                         CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1385                         if (cfs_copy_to_user((void *)arg, check,
1386                                              sizeof(*check)))
1387                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1388                         GOTO(out_poll, rc);
1389                 }
1390
1391                 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1392                                    NULL);
1393                 if (rc) {
1394                         CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1395                         if (cfs_copy_to_user((void *)arg, check,
1396                                              sizeof(*check)))
1397                                 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1398                         GOTO(out_poll, rc);
1399                 }
1400         out_poll:
1401                 OBD_FREE_PTR(check);
1402                 RETURN(rc);
1403         }
1404 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0)
1405         case LL_IOC_QUOTACTL_18: {
1406                 /* copy the old 1.x quota struct for internal use, then copy
1407                  * back into old format struct.  For 1.8 compatibility. */
1408                 struct if_quotactl_18 *qctl_18;
1409                 struct if_quotactl *qctl_20;
1410
1411                 OBD_ALLOC_PTR(qctl_18);
1412                 if (!qctl_18)
1413                         RETURN(-ENOMEM);
1414
1415                 OBD_ALLOC_PTR(qctl_20);
1416                 if (!qctl_20)
1417                         GOTO(out_quotactl_18, rc = -ENOMEM);
1418
1419                 if (cfs_copy_from_user(qctl_18, (void *)arg, sizeof(*qctl_18)))
1420                         GOTO(out_quotactl_20, rc = -ENOMEM);
1421
1422                 QCTL_COPY(qctl_20, qctl_18);
1423                 qctl_20->qc_idx = 0;
1424
1425                 /* XXX: dqb_valid was borrowed as a flag to mark that
1426                  *      only mds quota is wanted */
1427                 if (qctl_18->qc_cmd == Q_GETQUOTA &&
1428                     qctl_18->qc_dqblk.dqb_valid) {
1429                         qctl_20->qc_valid = QC_MDTIDX;
1430                         qctl_20->qc_dqblk.dqb_valid = 0;
1431                 } else if (qctl_18->obd_uuid.uuid[0] != '\0') {
1432                         qctl_20->qc_valid = QC_UUID;
1433                         qctl_20->obd_uuid = qctl_18->obd_uuid;
1434                 } else {
1435                         qctl_20->qc_valid = QC_GENERAL;
1436                 }
1437
1438                 rc = quotactl_ioctl(sbi, qctl_20);
1439
1440                 if (rc == 0) {
1441                         QCTL_COPY(qctl_18, qctl_20);
1442                         qctl_18->obd_uuid = qctl_20->obd_uuid;
1443
1444                         if (cfs_copy_to_user((void *)arg, qctl_18,
1445                                              sizeof(*qctl_18)))
1446                                 rc = -EFAULT;
1447                 }
1448
1449         out_quotactl_20:
1450                 OBD_FREE_PTR(qctl_20);
1451         out_quotactl_18:
1452                 OBD_FREE_PTR(qctl_18);
1453                 RETURN(rc);
1454         }
1455 #else
1456 #warning "remove old LL_IOC_QUOTACTL_18 compatibility code"
1457 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2,7,50,0) */
1458         case LL_IOC_QUOTACTL: {
1459                 struct if_quotactl *qctl;
1460
1461                 OBD_ALLOC_PTR(qctl);
1462                 if (!qctl)
1463                         RETURN(-ENOMEM);
1464
1465                 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1466                         GOTO(out_quotactl, rc = -EFAULT);
1467
1468                 rc = quotactl_ioctl(sbi, qctl);
1469
1470                 if (rc == 0 && cfs_copy_to_user((void *)arg,qctl,sizeof(*qctl)))
1471                         rc = -EFAULT;
1472
1473         out_quotactl:
1474                 OBD_FREE_PTR(qctl);
1475                 RETURN(rc);
1476         }
1477         case OBD_IOC_GETDTNAME:
1478         case OBD_IOC_GETMDNAME:
1479                 RETURN(ll_get_obd_name(inode, cmd, arg));
1480         case LL_IOC_FLUSHCTX:
1481                 RETURN(ll_flush_ctx(inode));
1482 #ifdef CONFIG_FS_POSIX_ACL
1483         case LL_IOC_RMTACL: {
1484             if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1485                 inode == inode->i_sb->s_root->d_inode) {
1486                 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1487
1488                 LASSERT(fd != NULL);
1489                 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1490                 if (!rc)
1491                         fd->fd_flags |= LL_FILE_RMTACL;
1492                 RETURN(rc);
1493             } else
1494                 RETURN(0);
1495         }
1496 #endif
1497         case LL_IOC_GETOBDCOUNT: {
1498                 int count, vallen;
1499                 struct obd_export *exp;
1500
1501                 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1502                         RETURN(-EFAULT);
1503
1504                 /* get ost count when count is zero, get mdt count otherwise */
1505                 exp = count ? sbi->ll_md_exp : sbi->ll_dt_exp;
1506                 vallen = sizeof(count);
1507                 rc = obd_get_info(NULL, exp, sizeof(KEY_TGT_COUNT),
1508                                   KEY_TGT_COUNT, &vallen, &count, NULL);
1509                 if (rc) {
1510                         CERROR("get target count failed: %d\n", rc);
1511                         RETURN(rc);
1512                 }
1513
1514                 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1515                         RETURN(-EFAULT);
1516
1517                 RETURN(0);
1518         }
1519         case LL_IOC_PATH2FID:
1520                 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1521                                      sizeof(struct lu_fid)))
1522                         RETURN(-EFAULT);
1523                 RETURN(0);
1524         case LL_IOC_GET_CONNECT_FLAGS: {
1525                 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1526         }
1527         case OBD_IOC_CHANGELOG_SEND:
1528         case OBD_IOC_CHANGELOG_CLEAR:
1529                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1530                                     sizeof(struct ioc_changelog));
1531                 RETURN(rc);
1532         case OBD_IOC_FID2PATH:
1533                 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1534         case LL_IOC_HSM_CT_START:
1535                 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1536                                     sizeof(struct lustre_kernelcomm));
1537                 RETURN(rc);
1538
1539         default:
1540                 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1541         }
1542 }
1543
1544 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1545 {
1546         struct inode *inode = file->f_mapping->host;
1547         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1548         struct ll_sb_info *sbi = ll_i2sbi(inode);
1549         int api32 = ll_need_32bit_api(sbi);
1550         loff_t ret = -EINVAL;
1551         ENTRY;
1552
1553         cfs_mutex_lock(&inode->i_mutex);
1554         switch (origin) {
1555                 case SEEK_SET:
1556                         break;
1557                 case SEEK_CUR:
1558                         offset += file->f_pos;
1559                         break;
1560                 case SEEK_END:
1561                         if (offset > 0)
1562                                 GOTO(out, ret);
1563                         if (api32)
1564                                 offset += LL_DIR_END_OFF_32BIT;
1565                         else
1566                                 offset += LL_DIR_END_OFF;
1567                         break;
1568                 default:
1569                         GOTO(out, ret);
1570         }
1571
1572         if (offset >= 0 &&
1573             ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1574              (!api32 && offset <= LL_DIR_END_OFF))) {
1575                 if (offset != file->f_pos) {
1576                         if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1577                             (!api32 && offset == LL_DIR_END_OFF))
1578                                 fd->fd_dir.lfd_pos = MDS_DIR_END_OFF;
1579                         else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1580                                 fd->fd_dir.lfd_pos = offset << 32;
1581                         else
1582                                 fd->fd_dir.lfd_pos = offset;
1583                         file->f_pos = offset;
1584                         file->f_version = 0;
1585                 }
1586                 ret = offset;
1587         }
1588         GOTO(out, ret);
1589
1590 out:
1591         cfs_mutex_unlock(&inode->i_mutex);
1592         return ret;
1593 }
1594
1595 int ll_dir_open(struct inode *inode, struct file *file)
1596 {
1597         ENTRY;
1598         RETURN(ll_file_open(inode, file));
1599 }
1600
1601 int ll_dir_release(struct inode *inode, struct file *file)
1602 {
1603         ENTRY;
1604         RETURN(ll_file_release(inode, file));
1605 }
1606
1607 struct file_operations ll_dir_operations = {
1608         .llseek   = ll_dir_seek,
1609         .open     = ll_dir_open,
1610         .release  = ll_dir_release,
1611         .read     = generic_read_dir,
1612         .readdir  = ll_readdir,
1613         .unlocked_ioctl   = ll_dir_ioctl,
1614         .fsync    = ll_fsync,
1615 };