4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Directory code for lustre client.
42 #include <linux/pagemap.h>
44 #include <linux/version.h>
45 #include <asm/uaccess.h>
46 #include <linux/buffer_head.h> // for wait_on_buffer
47 #include <linux/pagevec.h>
49 #define DEBUG_SUBSYSTEM S_LLITE
51 #include <obd_support.h>
52 #include <obd_class.h>
53 #include <lustre_lib.h>
54 #include <lustre/lustre_idl.h>
55 #include <lustre_lite.h>
56 #include <lustre_dlm.h>
57 #include <lustre_fid.h>
58 #include "llite_internal.h"
61 * (new) readdir implementation overview.
63 * Original lustre readdir implementation cached exact copy of raw directory
64 * pages on the client. These pages were indexed in client page cache by
65 * logical offset in the directory file. This design, while very simple and
66 * intuitive had some inherent problems:
68 * . it implies that byte offset to the directory entry serves as a
69 * telldir(3)/seekdir(3) cookie, but that offset is not stable: in
70 * ext3/htree directory entries may move due to splits, and more
73 * . it is incompatible with the design of split directories for cmd3,
74 * that assumes that names are distributed across nodes based on their
75 * hash, and so readdir should be done in hash order.
77 * New readdir implementation does readdir in hash order, and uses hash of a
78 * file name as a telldir/seekdir cookie. This led to number of complications:
80 * . hash is not unique, so it cannot be used to index cached directory
81 * pages on the client (note, that it requires a whole pageful of hash
82 * collided entries to cause two pages to have identical hashes);
84 * . hash is not unique, so it cannot, strictly speaking, be used as an
85 * entry cookie. ext3/htree has the same problem and lustre implementation
86 * mimics their solution: seekdir(hash) positions directory at the first
87 * entry with the given hash.
93 * Client caches directory pages using hash of the first entry as an index. As
94 * noted above hash is not unique, so this solution doesn't work as is:
95 * special processing is needed for "page hash chains" (i.e., sequences of
96 * pages filled with entries all having the same hash value).
98 * First, such chains have to be detected. To this end, server returns to the
99 * client the hash of the first entry on the page next to one returned. When
100 * client detects that this hash is the same as hash of the first entry on the
101 * returned page, page hash collision has to be handled. Pages in the
102 * hash chain, except first one, are termed "overflow pages".
104 * Solution to index uniqueness problem is to not cache overflow
105 * pages. Instead, when page hash collision is detected, all overflow pages
106 * from emerging chain are immediately requested from the server and placed in
107 * a special data structure (struct ll_dir_chain). This data structure is used
108 * by ll_readdir() to process entries from overflow pages. When readdir
109 * invocation finishes, overflow pages are discarded. If page hash collision
110 * chain weren't completely processed, next call to readdir will again detect
111 * page hash collision, again read overflow pages in, process next portion of
112 * entries and again discard the pages. This is not as wasteful as it looks,
113 * because, given reasonable hash, page hash collisions are extremely rare.
115 * 1. directory positioning
117 * When seekdir(hash) is called, original
128 * identification of and access to overflow pages
132 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
133 * a header lu_dirpage which describes the start/end hash, and whether this
134 * page is empty (contains no dir entry) or hash collide with next page.
135 * After client receives reply, several pages will be integrated into dir page
136 * in CFS_PAGE_SIZE (if CFS_PAGE_SIZE greater than LU_PAGE_SIZE), and the
137 * lu_dirpage for this integrated page will be adjusted.
141 /* returns the page unlocked, but with a reference */
142 static int ll_dir_readpage(struct file *file, struct page *page0)
144 struct inode *inode = page0->mapping->host;
145 int hash64 = ll_i2sbi(inode)->ll_flags & LL_SBI_64BIT_HASH;
146 struct obd_export *exp = ll_i2sbi(inode)->ll_md_exp;
147 struct ptlrpc_request *request;
148 struct mdt_body *body;
149 struct md_op_data *op_data;
151 struct page **page_pool;
153 #ifndef HAVE_ADD_TO_PAGE_CACHE_LRU
154 struct pagevec lru_pvec;
156 struct lu_dirpage *dp;
157 int max_pages = ll_i2sbi(inode)->ll_md_brw_size >> CFS_PAGE_SHIFT;
158 int nrdpgs = 0; /* number of pages read actually */
165 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
167 hash = fd->fd_dir.lfd_next;
169 hash = ll_i2info(inode)->lli_sa_pos;
171 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) hash "LPU64"\n",
172 inode->i_ino, inode->i_generation, inode, hash);
174 LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
176 OBD_ALLOC(page_pool, sizeof(page) * max_pages);
177 if (page_pool != NULL) {
178 page_pool[0] = page0;
183 for (npages = 1; npages < max_pages; npages++) {
184 page = page_cache_alloc_cold(inode->i_mapping);
187 page_pool[npages] = page;
190 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
191 LUSTRE_OPC_ANY, NULL);
192 op_data->op_npages = npages;
193 op_data->op_offset = hash;
194 rc = md_readpage(exp, op_data, page_pool, &request);
195 ll_finish_md_op_data(op_data);
197 body = req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY);
198 /* Checked by mdc_readpage() */
199 LASSERT(body != NULL);
201 if (body->valid & OBD_MD_FLSIZE)
202 cl_isize_write(inode, body->size);
204 nrdpgs = (request->rq_bulk->bd_nob_transferred+CFS_PAGE_SIZE-1)
206 SetPageUptodate(page0);
209 ptlrpc_req_finished(request);
211 CDEBUG(D_VFSTRACE, "read %d/%d pages\n", nrdpgs, npages);
213 ll_pagevec_init(&lru_pvec, 0);
214 for (i = 1; i < npages; i++) {
215 unsigned long offset;
220 if (rc < 0 || i >= nrdpgs) {
221 page_cache_release(page);
225 SetPageUptodate(page);
228 hash = le64_to_cpu(dp->ldp_hash_start);
231 offset = hash_x_index(hash, hash64);
233 prefetchw(&page->flags);
234 ret = ll_add_to_page_cache_lru(page, inode->i_mapping, offset,
238 if (ll_pagevec_add(&lru_pvec, page) == 0)
239 ll_pagevec_lru_add_file(&lru_pvec);
241 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
242 " %d\n", offset, ret);
244 page_cache_release(page);
246 ll_pagevec_lru_add_file(&lru_pvec);
248 if (page_pool != &page0)
249 OBD_FREE(page_pool, sizeof(struct page *) * max_pages);
254 #ifndef MS_HAS_NEW_AOPS
255 struct address_space_operations ll_dir_aops = {
256 .readpage = ll_dir_readpage,
259 struct address_space_operations_ext ll_dir_aops = {
260 .orig_aops.readpage = ll_dir_readpage,
264 static void ll_check_page(struct inode *dir, struct page *page)
266 /* XXX: check page format later */
267 SetPageChecked(page);
270 void ll_release_page(struct page *page, int remove)
275 if (likely(page->mapping != NULL))
276 truncate_complete_page(page->mapping, page);
279 page_cache_release(page);
283 * Find, kmap and return page that contains given hash.
285 static struct page *ll_dir_page_locate(struct inode *dir, __u64 *hash,
286 __u64 *start, __u64 *end)
288 int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
289 struct address_space *mapping = dir->i_mapping;
291 * Complement of hash is used as an index so that
292 * radix_tree_gang_lookup() can be used to find a page with starting
293 * hash _smaller_ than one we are looking for.
295 unsigned long offset = hash_x_index(*hash, hash64);
299 TREE_READ_LOCK_IRQ(mapping);
300 found = radix_tree_gang_lookup(&mapping->page_tree,
301 (void **)&page, offset, 1);
303 struct lu_dirpage *dp;
305 page_cache_get(page);
306 TREE_READ_UNLOCK_IRQ(mapping);
308 * In contrast to find_lock_page() we are sure that directory
309 * page cannot be truncated (while DLM lock is held) and,
310 * hence, can avoid restart.
312 * In fact, page cannot be locked here at all, because
313 * ll_dir_readpage() does synchronous io.
316 if (PageUptodate(page)) {
318 if (BITS_PER_LONG == 32 && hash64) {
319 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
320 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
323 *start = le64_to_cpu(dp->ldp_hash_start);
324 *end = le64_to_cpu(dp->ldp_hash_end);
326 LASSERTF(*start <= *hash, "start = "LPX64",end = "
327 LPX64",hash = "LPX64"\n", *start, *end, *hash);
328 CDEBUG(D_VFSTRACE, "page %lu [%llu %llu], hash "LPU64"\n",
329 offset, *start, *end, *hash);
331 ll_release_page(page, 0);
333 } else if (*end != *start && *hash == *end) {
335 * upon hash collision, remove this page,
336 * otherwise put page reference, and
337 * ll_get_dir_page() will issue RPC to fetch
340 ll_release_page(page,
341 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
345 page_cache_release(page);
346 page = ERR_PTR(-EIO);
350 TREE_READ_UNLOCK_IRQ(mapping);
356 struct page *ll_get_dir_page(struct file *filp, struct inode *dir, __u64 hash,
357 struct ll_dir_chain *chain)
359 ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
360 struct address_space *mapping = dir->i_mapping;
361 struct lustre_handle lockh;
362 struct lu_dirpage *dp;
369 struct ll_inode_info *lli = ll_i2info(dir);
370 int hash64 = ll_i2sbi(dir)->ll_flags & LL_SBI_64BIT_HASH;
373 rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
374 ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
376 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
377 ll_md_blocking_ast, ldlm_completion_ast,
379 struct lookup_intent it = { .it_op = IT_READDIR };
380 struct ptlrpc_request *request;
381 struct md_op_data *op_data;
383 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
384 LUSTRE_OPC_ANY, NULL);
386 return (void *)op_data;
388 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
389 op_data, &lockh, NULL, 0, NULL, 0);
391 ll_finish_md_op_data(op_data);
393 request = (struct ptlrpc_request *)it.d.lustre.it_data;
395 ptlrpc_req_finished(request);
397 CERROR("lock enqueue: "DFID" at "LPU64": rc %d\n",
398 PFID(ll_inode2fid(dir)), hash, rc);
402 /* for cross-ref object, l_ast_data of the lock may not be set,
403 * we reset it here */
404 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie,
407 ldlm_lock_dump_handle(D_OTHER, &lockh);
409 cfs_mutex_lock(&lli->lli_readdir_mutex);
410 page = ll_dir_page_locate(dir, &lhash, &start, &end);
412 CERROR("dir page locate: "DFID" at "LPU64": rc %ld\n",
413 PFID(ll_inode2fid(dir)), lhash, PTR_ERR(page));
414 GOTO(out_unlock, page);
415 } else if (page != NULL) {
417 * XXX nikita: not entirely correct handling of a corner case:
418 * suppose hash chain of entries with hash value HASH crosses
419 * border between pages P0 and P1. First both P0 and P1 are
420 * cached, seekdir() is called for some entry from the P0 part
421 * of the chain. Later P0 goes out of cache. telldir(HASH)
422 * happens and finds P1, as it starts with matching hash
423 * value. Remaining entries from P0 part of the chain are
424 * skipped. (Is that really a bug?)
426 * Possible solutions: 0. don't cache P1 is such case, handle
427 * it as an "overflow" page. 1. invalidate all pages at
428 * once. 2. use HASH|1 as an index for P1.
430 GOTO(hash_collision, page);
433 page = read_cache_page(mapping, hash_x_index(hash, hash64),
434 (filler_t*)mapping->a_ops->readpage, filp);
436 CERROR("read cache page: "DFID" at "LPU64": rc %ld\n",
437 PFID(ll_inode2fid(dir)), hash, PTR_ERR(page));
438 GOTO(out_unlock, page);
443 if (!PageUptodate(page)) {
444 CERROR("page not updated: "DFID" at "LPU64": rc %d\n",
445 PFID(ll_inode2fid(dir)), hash, -5);
448 if (!PageChecked(page))
449 ll_check_page(dir, page);
450 if (PageError(page)) {
451 CERROR("page error: "DFID" at "LPU64": rc %d\n",
452 PFID(ll_inode2fid(dir)), hash, -5);
456 dp = page_address(page);
457 if (BITS_PER_LONG == 32 && hash64) {
458 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
459 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
462 start = le64_to_cpu(dp->ldp_hash_start);
463 end = le64_to_cpu(dp->ldp_hash_end);
467 LASSERT(start == lhash);
468 CWARN("Page-wide hash collision: "LPU64"\n", end);
469 if (BITS_PER_LONG == 32 && hash64)
470 CWARN("Real page-wide hash collision at ["LPU64" "LPU64
471 "] with hash "LPU64"\n",
472 le64_to_cpu(dp->ldp_hash_start),
473 le64_to_cpu(dp->ldp_hash_end), hash);
475 * Fetch whole overflow chain...
482 cfs_mutex_unlock(&lli->lli_readdir_mutex);
483 ldlm_lock_decref(&lockh, mode);
487 ll_release_page(page, 1);
488 page = ERR_PTR(-EIO);
492 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
494 struct inode *inode = filp->f_dentry->d_inode;
495 struct ll_inode_info *info = ll_i2info(inode);
496 struct ll_sb_info *sbi = ll_i2sbi(inode);
497 struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
498 __u64 pos = fd->fd_dir.lfd_pos;
499 int api32 = ll_need_32bit_api(sbi);
500 int hash64 = sbi->ll_flags & LL_SBI_64BIT_HASH;
502 struct ll_dir_chain chain;
507 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu 32bit_api %d\n",
508 inode->i_ino, inode->i_generation, inode,
509 (unsigned long)pos, i_size_read(inode), api32);
511 if (pos == MDS_DIR_END_OFF)
519 ll_dir_chain_init(&chain);
521 fd->fd_dir.lfd_next = pos;
522 page = ll_get_dir_page(filp, inode, pos, &chain);
524 while (rc == 0 && !done) {
525 struct lu_dirpage *dp;
526 struct lu_dirent *ent;
530 * If page is empty (end of directory is reached),
533 __u64 hash = MDS_DIR_END_OFF;
536 dp = page_address(page);
537 for (ent = lu_dirent_start(dp); ent != NULL && !done;
538 ent = lu_dirent_next(ent)) {
546 * XXX: implement correct swabbing here.
549 hash = le64_to_cpu(ent->lde_hash);
552 * Skip until we find target hash
557 namelen = le16_to_cpu(ent->lde_namelen);
568 fid_le_to_cpu(&fid, &ent->lde_fid);
569 ino = cl_fid_build_ino(&fid, api32);
570 type = ll_dirent_type_get(ent);
571 /* For 'll_nfs_get_name_filldir()', it will try
572 * to access the 'ent' through its 'lde_name',
573 * so the parameter 'name' for 'filldir()' must
574 * be part of the 'ent'. */
575 done = filldir(cookie, ent->lde_name, namelen,
578 next = le64_to_cpu(dp->ldp_hash_end);
581 if (pos == MDS_DIR_END_OFF) {
583 * End of directory reached.
586 ll_release_page(page, 0);
587 } else if (1 /* chain is exhausted*/) {
589 * Normal case: continue to the next
592 ll_release_page(page,
593 le32_to_cpu(dp->ldp_flags) &
595 fd->fd_dir.lfd_next = pos;
596 page = ll_get_dir_page(filp, inode, pos,
600 * go into overflow page.
602 LASSERT(le32_to_cpu(dp->ldp_flags) &
604 ll_release_page(page, 1);
608 ll_release_page(page, 0);
612 CERROR("error reading dir "DFID" at %lu: rc %d\n",
613 PFID(&info->lli_fid), (unsigned long)pos, rc);
617 fd->fd_dir.lfd_pos = pos;
618 if (pos == MDS_DIR_END_OFF) {
620 filp->f_pos = LL_DIR_END_OFF_32BIT;
622 filp->f_pos = LL_DIR_END_OFF;
625 filp->f_pos = pos >> 32;
629 filp->f_version = inode->i_version;
630 touch_atime(filp->f_vfsmnt, filp->f_dentry);
632 ll_dir_chain_fini(&chain);
636 ll_stats_ops_tally(sbi, LPROC_LL_READDIR, 1);
641 int ll_send_mgc_param(struct obd_export *mgc, char *string)
643 struct mgs_send_param *msp;
650 strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
651 rc = obd_set_info_async(NULL, mgc, sizeof(KEY_SET_INFO), KEY_SET_INFO,
652 sizeof(struct mgs_send_param), msp, NULL);
654 CERROR("Failed to set parameter: %d\n", rc);
660 char *ll_get_fsname(struct inode *inode)
662 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
666 OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
667 len = strlen(lsi->lsi_lmd->lmd_profile);
668 ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
669 if (ptr && (strcmp(ptr, "-client") == 0))
671 strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
677 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
680 struct ll_sb_info *sbi = ll_i2sbi(inode);
681 struct md_op_data *op_data;
682 struct ptlrpc_request *req = NULL;
684 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
685 struct obd_device *mgc = lsi->lsi_mgc;
686 char *fsname = NULL, *param = NULL;
691 * This is coming from userspace, so should be in
692 * local endian. But the MDS would like it in little
693 * endian, so we swab it before we send it.
695 switch (lump->lmm_magic) {
696 case LOV_USER_MAGIC_V1: {
697 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V1))
698 lustre_swab_lov_user_md_v1(lump);
699 lum_size = sizeof(struct lov_user_md_v1);
702 case LOV_USER_MAGIC_V3: {
703 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC_V3))
704 lustre_swab_lov_user_md_v3(
705 (struct lov_user_md_v3 *)lump);
706 lum_size = sizeof(struct lov_user_md_v3);
710 CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
711 " %#08x != %#08x nor %#08x\n",
712 lump->lmm_magic, LOV_USER_MAGIC_V1,
718 lum_size = sizeof(struct lov_user_md_v1);
721 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
722 LUSTRE_OPC_ANY, NULL);
724 RETURN(PTR_ERR(op_data));
726 /* swabbing is done in lov_setstripe() on server side */
727 rc = md_setattr(sbi->ll_md_exp, op_data, lump, lum_size,
728 NULL, 0, &req, NULL);
729 ll_finish_md_op_data(op_data);
730 ptlrpc_req_finished(req);
732 if (rc != -EPERM && rc != -EACCES)
733 CERROR("mdc_setattr fails: rc = %d\n", rc);
736 /* In the following we use the fact that LOV_USER_MAGIC_V1 and
737 LOV_USER_MAGIC_V3 have the same initial fields so we do not
738 need the make the distiction between the 2 versions */
739 if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
740 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
742 /* Get fsname and assume devname to be -MDT0000. */
743 fsname = ll_get_fsname(inode);
744 /* Set root stripesize */
745 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
746 lump ? le32_to_cpu(lump->lmm_stripe_size) : 0);
747 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
751 /* Set root stripecount */
752 sprintf(param, "%s-MDT0000.lov.stripecount=%hd", fsname,
753 lump ? le16_to_cpu(lump->lmm_stripe_count) : 0);
754 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
758 /* Set root stripeoffset */
759 sprintf(param, "%s-MDT0000.lov.stripeoffset=%hd", fsname,
760 lump ? le16_to_cpu(lump->lmm_stripe_offset) :
761 (typeof(lump->lmm_stripe_offset))(-1));
762 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
767 OBD_FREE(fsname, MGS_PARAM_MAXLEN);
769 OBD_FREE(param, MGS_PARAM_MAXLEN);
774 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
775 int *lmm_size, struct ptlrpc_request **request)
777 struct ll_sb_info *sbi = ll_i2sbi(inode);
778 struct mdt_body *body;
779 struct lov_mds_md *lmm = NULL;
780 struct ptlrpc_request *req = NULL;
782 struct md_op_data *op_data;
784 rc = ll_get_max_mdsize(sbi, &lmmsize);
788 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL,
789 0, lmmsize, LUSTRE_OPC_ANY,
792 RETURN(PTR_ERR(op_data));
794 op_data->op_valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
795 rc = md_getattr(sbi->ll_md_exp, op_data, &req);
796 ll_finish_md_op_data(op_data);
798 CDEBUG(D_INFO, "md_getattr failed on inode "
799 "%lu/%u: rc %d\n", inode->i_ino,
800 inode->i_generation, rc);
804 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
805 LASSERT(body != NULL);
807 lmmsize = body->eadatasize;
809 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
811 GOTO(out, rc = -ENODATA);
814 lmm = req_capsule_server_sized_get(&req->rq_pill,
815 &RMF_MDT_MD, lmmsize);
816 LASSERT(lmm != NULL);
819 * This is coming from the MDS, so is probably in
820 * little endian. We convert it to host endian before
821 * passing it to userspace.
823 /* We don't swab objects for directories */
824 switch (le32_to_cpu(lmm->lmm_magic)) {
826 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
827 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
830 if (LOV_MAGIC != cpu_to_le32(LOV_MAGIC))
831 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
834 CERROR("unknown magic: %lX\n", (unsigned long)lmm->lmm_magic);
845 * Get MDT index for the inode.
847 int ll_get_mdt_idx(struct inode *inode)
849 struct ll_sb_info *sbi = ll_i2sbi(inode);
850 struct md_op_data *op_data;
854 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0,
855 0, LUSTRE_OPC_ANY, NULL);
857 RETURN(PTR_ERR(op_data));
859 op_data->op_valid |= OBD_MD_MDTIDX;
860 rc = md_getattr(sbi->ll_md_exp, op_data, NULL);
861 mdtidx = op_data->op_mds;
862 ll_finish_md_op_data(op_data);
864 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
870 static int copy_and_ioctl(int cmd, struct obd_export *exp, void *data, int len)
878 if (cfs_copy_from_user(ptr, data, len)) {
882 rc = obd_iocontrol(cmd, exp, len, data, NULL);
887 static int quotactl_ioctl(struct ll_sb_info *sbi, struct if_quotactl *qctl)
889 int cmd = qctl->qc_cmd;
890 int type = qctl->qc_type;
891 int id = qctl->qc_id;
892 int valid = qctl->qc_valid;
897 case LUSTRE_Q_INVALIDATE:
898 case LUSTRE_Q_FINVALIDATE:
903 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
904 sbi->ll_flags & LL_SBI_RMT_CLIENT)
908 if (((type == USRQUOTA && cfs_curproc_euid() != id) ||
909 (type == GRPQUOTA && !in_egroup_p(id))) &&
910 (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
911 sbi->ll_flags & LL_SBI_RMT_CLIENT))
917 CERROR("unsupported quotactl op: %#x\n", cmd);
921 if (valid != QC_GENERAL) {
922 if (sbi->ll_flags & LL_SBI_RMT_CLIENT)
925 if (cmd == Q_GETINFO)
926 qctl->qc_cmd = Q_GETOINFO;
927 else if (cmd == Q_GETQUOTA)
928 qctl->qc_cmd = Q_GETOQUOTA;
934 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
935 sizeof(*qctl), qctl, NULL);
938 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_dt_exp,
939 sizeof(*qctl), qctl, NULL);
942 rc = obd_iocontrol(OBD_IOC_QUOTACTL, sbi->ll_md_exp,
943 sizeof(*qctl), qctl, NULL);
945 rc = obd_iocontrol(OBD_IOC_QUOTACTL,
947 sizeof(*qctl), qctl, NULL);
959 struct obd_quotactl *oqctl;
961 OBD_ALLOC_PTR(oqctl);
965 QCTL_COPY(oqctl, qctl);
966 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
968 if (rc != -EALREADY && cmd == Q_QUOTAON) {
969 oqctl->qc_cmd = Q_QUOTAOFF;
970 obd_quotactl(sbi->ll_md_exp, oqctl);
975 /* If QIF_SPACE is not set, client should collect the
976 * space usage from OSSs by itself */
977 if (cmd == Q_GETQUOTA &&
978 !(oqctl->qc_dqblk.dqb_valid & QIF_SPACE) &&
979 !oqctl->qc_dqblk.dqb_curspace) {
980 struct obd_quotactl *oqctl_tmp;
982 OBD_ALLOC_PTR(oqctl_tmp);
983 if (oqctl_tmp == NULL)
984 GOTO(out, rc = -ENOMEM);
986 oqctl_tmp->qc_cmd = Q_GETOQUOTA;
987 oqctl_tmp->qc_id = oqctl->qc_id;
988 oqctl_tmp->qc_type = oqctl->qc_type;
990 /* collect space usage from OSTs */
991 oqctl_tmp->qc_dqblk.dqb_curspace = 0;
992 rc = obd_quotactl(sbi->ll_dt_exp, oqctl_tmp);
993 if (!rc || rc == -EREMOTEIO) {
994 oqctl->qc_dqblk.dqb_curspace =
995 oqctl_tmp->qc_dqblk.dqb_curspace;
996 oqctl->qc_dqblk.dqb_valid |= QIF_SPACE;
999 /* collect space & inode usage from MDTs */
1000 oqctl_tmp->qc_dqblk.dqb_curspace = 0;
1001 oqctl_tmp->qc_dqblk.dqb_curinodes = 0;
1002 rc = obd_quotactl(sbi->ll_md_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_curinodes =
1007 oqctl_tmp->qc_dqblk.dqb_curinodes;
1008 oqctl->qc_dqblk.dqb_valid |= QIF_INODES;
1010 oqctl->qc_dqblk.dqb_valid &= ~QIF_SPACE;
1013 OBD_FREE_PTR(oqctl_tmp);
1016 QCTL_COPY(qctl, oqctl);
1017 OBD_FREE_PTR(oqctl);
1023 static long ll_dir_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1025 struct inode *inode = file->f_dentry->d_inode;
1026 struct ll_sb_info *sbi = ll_i2sbi(inode);
1027 struct obd_ioctl_data *data;
1031 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
1032 inode->i_ino, inode->i_generation, inode, cmd);
1034 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
1035 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
1038 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
1040 case FSFILT_IOC_GETFLAGS:
1041 case FSFILT_IOC_SETFLAGS:
1042 RETURN(ll_iocontrol(inode, file, cmd, arg));
1043 case FSFILT_IOC_GETVERSION_OLD:
1044 case FSFILT_IOC_GETVERSION:
1045 RETURN(put_user(inode->i_generation, (int *)arg));
1046 /* We need to special case any other ioctls we want to handle,
1047 * to send them to the MDS/OST as appropriate and to properly
1048 * network encode the arg field.
1049 case FSFILT_IOC_SETVERSION_OLD:
1050 case FSFILT_IOC_SETVERSION:
1052 case LL_IOC_GET_MDTIDX: {
1055 mdtidx = ll_get_mdt_idx(inode);
1059 if (put_user((int)mdtidx, (int*)arg))
1064 case IOC_MDC_LOOKUP: {
1065 struct ptlrpc_request *request = NULL;
1066 int namelen, len = 0;
1069 struct md_op_data *op_data;
1071 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
1076 filename = data->ioc_inlbuf1;
1077 namelen = strlen(filename);
1080 CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
1081 GOTO(out_free, rc = -EINVAL);
1084 op_data = ll_prep_md_op_data(NULL, inode, NULL, filename, namelen,
1085 0, LUSTRE_OPC_ANY, NULL);
1086 if (IS_ERR(op_data))
1087 GOTO(out_free, rc = PTR_ERR(op_data));
1089 op_data->op_valid = OBD_MD_FLID;
1090 rc = md_getattr_name(sbi->ll_md_exp, op_data, &request);
1091 ll_finish_md_op_data(op_data);
1093 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
1096 ptlrpc_req_finished(request);
1099 obd_ioctl_freedata(buf, len);
1102 case LL_IOC_LOV_SETSTRIPE: {
1103 struct lov_user_md_v3 lumv3;
1104 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
1105 struct lov_user_md_v1 *lumv1p = (struct lov_user_md_v1 *)arg;
1106 struct lov_user_md_v3 *lumv3p = (struct lov_user_md_v3 *)arg;
1108 int set_default = 0;
1110 LASSERT(sizeof(lumv3) == sizeof(*lumv3p));
1111 LASSERT(sizeof(lumv3.lmm_objects[0]) ==
1112 sizeof(lumv3p->lmm_objects[0]));
1113 /* first try with v1 which is smaller than v3 */
1114 if (cfs_copy_from_user(lumv1, lumv1p, sizeof(*lumv1)))
1117 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
1118 if (cfs_copy_from_user(&lumv3, lumv3p, sizeof(lumv3)))
1122 if (inode->i_sb->s_root == file->f_dentry)
1125 /* in v1 and v3 cases lumv1 points to data */
1126 rc = ll_dir_setstripe(inode, lumv1, set_default);
1130 case LL_IOC_OBD_STATFS:
1131 RETURN(ll_obd_statfs(inode, (void *)arg));
1132 case LL_IOC_LOV_GETSTRIPE:
1133 case LL_IOC_MDC_GETINFO:
1134 case IOC_MDC_GETFILEINFO:
1135 case IOC_MDC_GETFILESTRIPE: {
1136 struct ptlrpc_request *request = NULL;
1137 struct lov_user_md *lump;
1138 struct lov_mds_md *lmm = NULL;
1139 struct mdt_body *body;
1140 char *filename = NULL;
1143 if (cmd == IOC_MDC_GETFILEINFO ||
1144 cmd == IOC_MDC_GETFILESTRIPE) {
1145 filename = getname((const char *)arg);
1146 if (IS_ERR(filename))
1147 RETURN(PTR_ERR(filename));
1149 rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
1150 &lmmsize, &request);
1152 rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
1156 body = req_capsule_server_get(&request->rq_pill,
1158 LASSERT(body != NULL);
1164 if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
1165 cmd == LL_IOC_MDC_GETINFO))
1166 GOTO(skip_lmm, rc = 0);
1171 if (cmd == IOC_MDC_GETFILESTRIPE ||
1172 cmd == LL_IOC_LOV_GETSTRIPE) {
1173 lump = (struct lov_user_md *)arg;
1175 struct lov_user_mds_data *lmdp;
1176 lmdp = (struct lov_user_mds_data *)arg;
1177 lump = &lmdp->lmd_lmm;
1179 if (cfs_copy_to_user(lump, lmm, lmmsize)) {
1180 if (cfs_copy_to_user(lump, lmm, sizeof(*lump)))
1181 GOTO(out_req, rc = -EFAULT);
1185 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
1186 struct lov_user_mds_data *lmdp;
1189 st.st_dev = inode->i_sb->s_dev;
1190 st.st_mode = body->mode;
1191 st.st_nlink = body->nlink;
1192 st.st_uid = body->uid;
1193 st.st_gid = body->gid;
1194 st.st_rdev = body->rdev;
1195 st.st_size = body->size;
1196 st.st_blksize = CFS_PAGE_SIZE;
1197 st.st_blocks = body->blocks;
1198 st.st_atime = body->atime;
1199 st.st_mtime = body->mtime;
1200 st.st_ctime = body->ctime;
1201 st.st_ino = inode->i_ino;
1203 lmdp = (struct lov_user_mds_data *)arg;
1204 if (cfs_copy_to_user(&lmdp->lmd_st, &st, sizeof(st)))
1205 GOTO(out_req, rc = -EFAULT);
1210 ptlrpc_req_finished(request);
1215 case IOC_LOV_GETINFO: {
1216 struct lov_user_mds_data *lumd;
1217 struct lov_stripe_md *lsm;
1218 struct lov_user_md *lum;
1219 struct lov_mds_md *lmm;
1223 lumd = (struct lov_user_mds_data *)arg;
1224 lum = &lumd->lmd_lmm;
1226 rc = ll_get_max_mdsize(sbi, &lmmsize);
1230 OBD_ALLOC_LARGE(lmm, lmmsize);
1231 if (cfs_copy_from_user(lmm, lum, lmmsize))
1232 GOTO(free_lmm, rc = -EFAULT);
1234 switch (lmm->lmm_magic) {
1235 case LOV_USER_MAGIC_V1:
1236 if (LOV_USER_MAGIC_V1 == cpu_to_le32(LOV_USER_MAGIC_V1))
1238 /* swab objects first so that stripes num will be sane */
1239 lustre_swab_lov_user_md_objects(
1240 ((struct lov_user_md_v1 *)lmm)->lmm_objects,
1241 ((struct lov_user_md_v1 *)lmm)->lmm_stripe_count);
1242 lustre_swab_lov_user_md_v1((struct lov_user_md_v1 *)lmm);
1244 case LOV_USER_MAGIC_V3:
1245 if (LOV_USER_MAGIC_V3 == cpu_to_le32(LOV_USER_MAGIC_V3))
1247 /* swab objects first so that stripes num will be sane */
1248 lustre_swab_lov_user_md_objects(
1249 ((struct lov_user_md_v3 *)lmm)->lmm_objects,
1250 ((struct lov_user_md_v3 *)lmm)->lmm_stripe_count);
1251 lustre_swab_lov_user_md_v3((struct lov_user_md_v3 *)lmm);
1254 GOTO(free_lmm, rc = -EINVAL);
1257 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
1259 GOTO(free_lmm, rc = -ENOMEM);
1261 /* Perform glimpse_size operation. */
1262 memset(&st, 0, sizeof(st));
1264 rc = ll_glimpse_ioctl(sbi, lsm, &st);
1268 if (cfs_copy_to_user(&lumd->lmd_st, &st, sizeof(st)))
1269 GOTO(free_lsm, rc = -EFAULT);
1273 obd_free_memmd(sbi->ll_dt_exp, &lsm);
1275 OBD_FREE_LARGE(lmm, lmmsize);
1278 case OBD_IOC_LLOG_CATINFO: {
1279 RETURN(-EOPNOTSUPP);
1281 case OBD_IOC_QUOTACHECK: {
1282 struct obd_quotactl *oqctl;
1285 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1286 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1289 OBD_ALLOC_PTR(oqctl);
1292 oqctl->qc_type = arg;
1293 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
1295 CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
1299 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
1301 CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
1303 OBD_FREE_PTR(oqctl);
1306 case OBD_IOC_POLL_QUOTACHECK: {
1307 struct if_quotacheck *check;
1309 if (!cfs_capable(CFS_CAP_SYS_ADMIN) ||
1310 sbi->ll_flags & LL_SBI_RMT_CLIENT)
1313 OBD_ALLOC_PTR(check);
1317 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
1320 CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
1321 if (cfs_copy_to_user((void *)arg, check,
1323 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1327 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
1330 CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
1331 if (cfs_copy_to_user((void *)arg, check,
1333 CDEBUG(D_QUOTA, "cfs_copy_to_user failed\n");
1337 OBD_FREE_PTR(check);
1340 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0)
1341 case LL_IOC_QUOTACTL_18: {
1342 /* copy the old 1.x quota struct for internal use, then copy
1343 * back into old format struct. For 1.8 compatibility. */
1344 struct if_quotactl_18 *qctl_18;
1345 struct if_quotactl *qctl_20;
1347 OBD_ALLOC_PTR(qctl_18);
1351 OBD_ALLOC_PTR(qctl_20);
1353 GOTO(out_quotactl_18, rc = -ENOMEM);
1355 if (cfs_copy_from_user(qctl_18, (void *)arg, sizeof(*qctl_18)))
1356 GOTO(out_quotactl_20, rc = -ENOMEM);
1358 QCTL_COPY(qctl_20, qctl_18);
1359 qctl_20->qc_idx = 0;
1361 /* XXX: dqb_valid was borrowed as a flag to mark that
1362 * only mds quota is wanted */
1363 if (qctl_18->qc_cmd == Q_GETQUOTA &&
1364 qctl_18->qc_dqblk.dqb_valid) {
1365 qctl_20->qc_valid = QC_MDTIDX;
1366 qctl_20->qc_dqblk.dqb_valid = 0;
1367 } else if (qctl_18->obd_uuid.uuid[0] != '\0') {
1368 qctl_20->qc_valid = QC_UUID;
1369 qctl_20->obd_uuid = qctl_18->obd_uuid;
1371 qctl_20->qc_valid = QC_GENERAL;
1374 rc = quotactl_ioctl(sbi, qctl_20);
1377 QCTL_COPY(qctl_18, qctl_20);
1378 qctl_18->obd_uuid = qctl_20->obd_uuid;
1380 if (cfs_copy_to_user((void *)arg, qctl_18,
1386 OBD_FREE_PTR(qctl_20);
1388 OBD_FREE_PTR(qctl_18);
1392 #warning "remove old LL_IOC_QUOTACTL_18 compatibility code"
1393 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 7, 50, 0) */
1394 case LL_IOC_QUOTACTL: {
1395 struct if_quotactl *qctl;
1397 OBD_ALLOC_PTR(qctl);
1401 if (cfs_copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1402 GOTO(out_quotactl, rc = -EFAULT);
1404 rc = quotactl_ioctl(sbi, qctl);
1406 if (rc == 0 && cfs_copy_to_user((void *)arg,qctl,sizeof(*qctl)))
1413 case OBD_IOC_GETDTNAME:
1414 case OBD_IOC_GETMDNAME:
1415 RETURN(ll_get_obd_name(inode, cmd, arg));
1416 case LL_IOC_FLUSHCTX:
1417 RETURN(ll_flush_ctx(inode));
1418 #ifdef CONFIG_FS_POSIX_ACL
1419 case LL_IOC_RMTACL: {
1420 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1421 inode == inode->i_sb->s_root->d_inode) {
1422 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1424 LASSERT(fd != NULL);
1425 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1427 fd->fd_flags |= LL_FILE_RMTACL;
1433 case LL_IOC_GETOBDCOUNT: {
1435 struct obd_export *exp;
1437 if (cfs_copy_from_user(&count, (int *)arg, sizeof(int)))
1440 /* get ost count when count is zero, get mdt count otherwise */
1441 exp = count ? sbi->ll_md_exp : sbi->ll_dt_exp;
1442 vallen = sizeof(count);
1443 rc = obd_get_info(NULL, exp, sizeof(KEY_TGT_COUNT),
1444 KEY_TGT_COUNT, &vallen, &count, NULL);
1446 CERROR("get target count failed: %d\n", rc);
1450 if (cfs_copy_to_user((int *)arg, &count, sizeof(int)))
1455 case LL_IOC_PATH2FID:
1456 if (cfs_copy_to_user((void *)arg, ll_inode2fid(inode),
1457 sizeof(struct lu_fid)))
1460 case LL_IOC_GET_CONNECT_FLAGS: {
1461 RETURN(obd_iocontrol(cmd, sbi->ll_md_exp, 0, NULL, (void*)arg));
1463 case OBD_IOC_CHANGELOG_SEND:
1464 case OBD_IOC_CHANGELOG_CLEAR:
1465 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1466 sizeof(struct ioc_changelog));
1468 case OBD_IOC_FID2PATH:
1469 RETURN(ll_fid2path(ll_i2mdexp(inode), (void *)arg));
1470 case LL_IOC_HSM_CT_START:
1471 rc = copy_and_ioctl(cmd, sbi->ll_md_exp, (void *)arg,
1472 sizeof(struct lustre_kernelcomm));
1476 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1480 static loff_t ll_dir_seek(struct file *file, loff_t offset, int origin)
1482 struct inode *inode = file->f_mapping->host;
1483 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1484 struct ll_sb_info *sbi = ll_i2sbi(inode);
1485 int api32 = ll_need_32bit_api(sbi);
1486 loff_t ret = -EINVAL;
1489 cfs_mutex_lock(&inode->i_mutex);
1494 offset += file->f_pos;
1500 offset += LL_DIR_END_OFF_32BIT;
1502 offset += LL_DIR_END_OFF;
1509 ((api32 && offset <= LL_DIR_END_OFF_32BIT) ||
1510 (!api32 && offset <= LL_DIR_END_OFF))) {
1511 if (offset != file->f_pos) {
1512 if ((api32 && offset == LL_DIR_END_OFF_32BIT) ||
1513 (!api32 && offset == LL_DIR_END_OFF))
1514 fd->fd_dir.lfd_pos = MDS_DIR_END_OFF;
1515 else if (api32 && sbi->ll_flags & LL_SBI_64BIT_HASH)
1516 fd->fd_dir.lfd_pos = offset << 32;
1518 fd->fd_dir.lfd_pos = offset;
1519 file->f_pos = offset;
1520 file->f_version = 0;
1527 cfs_mutex_unlock(&inode->i_mutex);
1531 int ll_dir_open(struct inode *inode, struct file *file)
1534 RETURN(ll_file_open(inode, file));
1537 int ll_dir_release(struct inode *inode, struct file *file)
1540 RETURN(ll_file_release(inode, file));
1543 struct file_operations ll_dir_operations = {
1544 .llseek = ll_dir_seek,
1545 .open = ll_dir_open,
1546 .release = ll_dir_release,
1547 .read = generic_read_dir,
1548 .readdir = ll_readdir,
1549 .unlocked_ioctl = ll_dir_ioctl,