1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 1992, 1993, 1994, 1995
5 * Remy Card (card@masi.ibp.fr)
6 * Laboratoire MASI - Institut Blaise Pascal
7 * Universite Pierre et Marie Curie (Paris VI)
11 * linux/fs/minix/dir.c
14 * Copyright (C) 1991, 1992 Linus Torvalds
16 * ext2 directory handling functions
18 * Big-endian to little-endian byte-swapping/bitmaps by
19 * David S. Miller (davem@caip.rutgers.edu), 1995
21 * All code that works with directory layout had been switched to pagecache
24 * Adapted for Lustre Light
25 * Copyright (C) 2002-2003, Cluster File Systems, Inc.
30 #include <linux/ext2_fs.h>
31 #include <linux/pagemap.h>
33 #include <linux/version.h>
34 #include <linux/smp_lock.h>
35 #include <asm/uaccess.h>
36 #include <linux/buffer_head.h> // for wait_on_buffer
38 #define DEBUG_SUBSYSTEM S_LLITE
40 #include <obd_support.h>
41 #include <obd_class.h>
42 #include <lustre_lib.h>
43 #include <lustre/lustre_idl.h>
44 #include <lustre_lite.h>
45 #include <lustre_dlm.h>
46 #include <lustre_fid.h>
47 #include "llite_internal.h"
49 #ifndef HAVE_PAGE_CHECKED
50 #ifdef HAVE_PG_FS_MISC
51 #define PageChecked(page) test_bit(PG_fs_misc, &(page)->flags)
52 #define SetPageChecked(page) set_bit(PG_fs_misc, &(page)->flags)
54 #error PageChecked or PageFsMisc not defined in kernel
59 * (new) readdir implementation overview.
61 * Original lustre readdir implementation cached exact copy of raw directory
62 * pages on the client. These pages were indexed in client page cache by
63 * logical offset in the directory file. This design, while very simple and
64 * intuitive had some inherent problems:
66 * . it implies that byte offset to the directory entry serves as a
67 * telldir(3)/seekdir(3) cookie, but that offset is not stable: in
68 * ext3/htree directory entries may move due to splits, and more
71 * . it is incompatible with the design of split directories for cmd3,
72 * that assumes that names are distributed across nodes based on their
73 * hash, and so readdir should be done in hash order.
75 * New readdir implementation does readdir in hash order, and uses hash of a
76 * file name as a telldir/seekdir cookie. This led to number of complications:
78 * . hash is not unique, so it cannot be used to index cached directory
79 * pages on the client (note, that it requires a whole pageful of hash
80 * collided entries to cause two pages to have identical hashes);
82 * . hash is not unique, so it cannot, strictly speaking, be used as an
83 * entry cookie. ext3/htree has the same problem and lustre implementation
84 * mimics their solution: seekdir(hash) positions directory at the first
85 * entry with the given hash.
91 * Client caches directory pages using hash of the first entry as an index. As
92 * noted above hash is not unique, so this solution doesn't work as is:
93 * special processing is needed for "page hash chains" (i.e., sequences of
94 * pages filled with entries all having the same hash value).
96 * First, such chains have to be detected. To this end, server returns to the
97 * client the hash of the first entry on the page next to one returned. When
98 * client detects that this hash is the same as hash of the first entry on the
99 * returned page, page hash collision has to be handled. Pages in the
100 * hash chain, except first one, are termed "overflow pages".
102 * Solution to index uniqueness problem is to not cache overflow
103 * pages. Instead, when page hash collision is detected, all overflow pages
104 * from emerging chain are immediately requested from the server and placed in
105 * a special data structure (struct ll_dir_chain). This data structure is used
106 * by ll_readdir() to process entries from overflow pages. When readdir
107 * invocation finishes, overflow pages are discarded. If page hash collision
108 * chain weren't completely processed, next call to readdir will again detect
109 * page hash collision, again read overflow pages in, process next portion of
110 * entries and again discard the pages. This is not as wasteful as it looks,
111 * because, given reasonable hash, page hash collisions are extremely rare.
113 * 1. directory positioning
115 * When seekdir(hash) is called, original
126 * identification of and access to overflow pages
136 /* returns the page unlocked, but with a reference */
137 static int ll_dir_readpage(struct file *file, struct page *page)
139 struct inode *inode = page->mapping->host;
140 struct ptlrpc_request *request;
141 struct mdt_body *body;
147 hash = hash_x_index(page->index);
148 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) off %lu\n",
149 inode->i_ino, inode->i_generation, inode, (unsigned long)hash);
151 oc = ll_mdscapa_get(inode);
152 rc = md_readpage(ll_i2sbi(inode)->ll_md_exp, ll_inode2fid(inode),
153 oc, hash, page, &request);
156 body = lustre_msg_buf(request->rq_repmsg, REPLY_REC_OFF,
158 /* Checked by mdc_readpage() */
159 LASSERT(body != NULL);
161 /* Swabbed by mdc_readpage() */
162 LASSERT(lustre_rep_swabbed(request, REPLY_REC_OFF));
164 if (body->valid & OBD_MD_FLSIZE) {
165 ll_inode_size_lock(inode, 0);
166 i_size_write(inode, body->size);
167 ll_inode_size_unlock(inode, 0);
169 SetPageUptodate(page);
171 ptlrpc_req_finished(request);
178 struct address_space_operations ll_dir_aops = {
179 .readpage = ll_dir_readpage,
182 static inline unsigned long dir_pages(struct inode *inode)
184 return (i_size_read(inode) + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
187 static inline unsigned ll_chunk_size(struct inode *inode)
189 return inode->i_sb->s_blocksize;
192 static void ll_check_page(struct inode *dir, struct page *page)
194 /* XXX: check page format later */
195 SetPageChecked(page);
198 static inline void ll_put_page(struct page *page)
201 page_cache_release(page);
205 * Find, kmap and return page that contains given hash.
207 static struct page *ll_dir_page_locate(struct inode *dir, unsigned long hash,
208 __u32 *start, __u32 *end)
210 struct address_space *mapping = dir->i_mapping;
212 * Complement of hash is used as an index so that
213 * radix_tree_gang_lookup() can be used to find a page with starting
214 * hash _smaller_ than one we are looking for.
216 unsigned long offset = hash_x_index(hash);
220 TREE_READ_LOCK_IRQ(mapping);
221 found = radix_tree_gang_lookup(&mapping->page_tree,
222 (void **)&page, offset, 1);
224 struct lu_dirpage *dp;
226 page_cache_get(page);
227 TREE_READ_UNLOCK_IRQ(mapping);
229 * In contrast to find_lock_page() we are sure that directory
230 * page cannot be truncated (while DLM lock is held) and,
231 * hence, can avoid restart.
233 * In fact, page cannot be locked here at all, because
234 * ll_dir_readpage() does synchronous io.
237 if (PageUptodate(page)) {
239 *start = le32_to_cpu(dp->ldp_hash_start);
240 *end = le32_to_cpu(dp->ldp_hash_end);
241 LASSERT(*start <= hash);
242 if (hash > *end || (*end != *start && hash == *end)) {
245 ll_truncate_complete_page(page);
247 page_cache_release(page);
251 page_cache_release(page);
252 page = ERR_PTR(-EIO);
256 TREE_READ_UNLOCK_IRQ(mapping);
262 static struct page *ll_get_dir_page(struct inode *dir, __u32 hash, int exact,
263 struct ll_dir_chain *chain)
265 ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_UPDATE} };
266 struct address_space *mapping = dir->i_mapping;
267 struct lustre_handle lockh;
268 struct lu_dirpage *dp;
276 rc = md_lock_match(ll_i2sbi(dir)->ll_md_exp, LDLM_FL_BLOCK_GRANTED,
277 ll_inode2fid(dir), LDLM_IBITS, &policy, mode, &lockh);
279 struct ldlm_enqueue_info einfo = { LDLM_IBITS, mode,
280 ll_md_blocking_ast, ldlm_completion_ast, NULL, dir };
281 struct lookup_intent it = { .it_op = IT_READDIR };
282 struct ptlrpc_request *request;
283 struct md_op_data *op_data;
285 op_data = ll_prep_md_op_data(NULL, dir, NULL, NULL, 0, 0,
286 LUSTRE_OPC_ANY, NULL);
288 return (void *)op_data;
290 rc = md_enqueue(ll_i2sbi(dir)->ll_md_exp, &einfo, &it,
291 op_data, &lockh, NULL, 0, 0);
293 ll_finish_md_op_data(op_data);
295 request = (struct ptlrpc_request *)it.d.lustre.it_data;
297 ptlrpc_req_finished(request);
299 CERROR("lock enqueue: rc: %d\n", rc);
303 /* for cross-ref object, l_ast_data of the lock may not be set,
304 * we reset it here */
305 md_set_lock_data(ll_i2sbi(dir)->ll_md_exp, &lockh.cookie, dir);
307 ldlm_lock_dump_handle(D_OTHER, &lockh);
309 page = ll_dir_page_locate(dir, hash, &start, &end);
311 GOTO(out_unlock, page);
315 * XXX nikita: not entirely correct handling of a corner case:
316 * suppose hash chain of entries with hash value HASH crosses
317 * border between pages P0 and P1. First both P0 and P1 are
318 * cached, seekdir() is called for some entry from the P0 part
319 * of the chain. Later P0 goes out of cache. telldir(HASH)
320 * happens and finds P1, as it starts with matching hash
321 * value. Remaining entries from P0 part of the chain are
322 * skipped. (Is that really a bug?)
324 * Possible solutions: 0. don't cache P1 is such case, handle
325 * it as an "overflow" page. 1. invalidate all pages at
326 * once. 2. use HASH|1 as an index for P1.
328 if (exact && hash != start) {
330 * readdir asked for a page starting _exactly_ from
331 * given hash, but cache contains stale page, with
332 * entries with smaller hash values. Stale page should
333 * be invalidated, and new one fetched.
335 CWARN("Stale readpage page %p: %#lx != %#lx\n", page,
336 (unsigned long)hash, (unsigned long)start);
338 ll_truncate_complete_page(page);
340 page_cache_release(page);
342 GOTO(hash_collision, page);
345 page = read_cache_page(mapping, hash_x_index(hash),
346 (filler_t*)mapping->a_ops->readpage, NULL);
348 GOTO(out_unlock, page);
352 if (!PageUptodate(page))
354 if (!PageChecked(page))
355 ll_check_page(dir, page);
359 dp = page_address(page);
361 start = le32_to_cpu(dp->ldp_hash_start);
362 end = le32_to_cpu(dp->ldp_hash_end);
364 LASSERT(start == hash);
365 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
367 * Fetch whole overflow chain...
374 ldlm_lock_decref(&lockh, mode);
379 page = ERR_PTR(-EIO);
383 int ll_readdir(struct file *filp, void *cookie, filldir_t filldir)
385 struct inode *inode = filp->f_dentry->d_inode;
386 struct ll_inode_info *info = ll_i2info(inode);
387 struct ll_sb_info *sbi = ll_i2sbi(inode);
388 __u32 pos = filp->f_pos;
390 struct ll_dir_chain chain;
396 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) pos %lu/%llu\n",
397 inode->i_ino, inode->i_generation, inode,
398 (unsigned long)pos, i_size_read(inode));
400 if (pos == DIR_END_OFF)
409 ll_dir_chain_init(&chain);
411 page = ll_get_dir_page(inode, pos, 0, &chain);
413 while (rc == 0 && !done) {
414 struct lu_dirpage *dp;
415 struct lu_dirent *ent;
419 * If page is empty (end of directoryis reached),
422 __u32 hash = DIR_END_OFF;
425 dp = page_address(page);
426 for (ent = lu_dirent_start(dp); ent != NULL && !done;
427 ent = lu_dirent_next(ent)) {
434 * XXX: implement correct swabbing here.
437 hash = le32_to_cpu(ent->lde_hash);
438 namelen = le16_to_cpu(ent->lde_namelen);
442 * Skip until we find target hash
454 name = ent->lde_name;
455 fid_le_to_cpu(&fid, &fid);
456 ino = ll_fid_build_ino(sbi, &fid);
458 done = filldir(cookie, name, namelen,
459 (loff_t)hash, ino, DT_UNKNOWN);
461 next = le32_to_cpu(dp->ldp_hash_end);
465 if (pos == DIR_END_OFF)
467 * End of directory reached.
470 else if (1 /* chain is exhausted*/)
472 * Normal case: continue to the next
475 page = ll_get_dir_page(inode, pos, 1,
479 * go into overflow page.
486 CERROR("error reading dir "DFID" at %lu: rc %d\n",
487 PFID(&info->lli_fid), (unsigned long)pos, rc);
491 filp->f_pos = (loff_t)(__s32)pos;
492 filp->f_version = inode->i_version;
493 touch_atime(filp->f_vfsmnt, filp->f_dentry);
495 ll_dir_chain_fini(&chain);
500 #define QCTL_COPY(out, in) \
502 Q_COPY(out, in, qc_cmd); \
503 Q_COPY(out, in, qc_type); \
504 Q_COPY(out, in, qc_id); \
505 Q_COPY(out, in, qc_stat); \
506 Q_COPY(out, in, qc_dqinfo); \
507 Q_COPY(out, in, qc_dqblk); \
510 int ll_send_mgc_param(struct obd_export *mgc, char *string)
512 struct mgs_send_param *msp;
519 strncpy(msp->mgs_param, string, MGS_PARAM_MAXLEN);
520 rc = obd_set_info_async(mgc, strlen(KEY_SET_INFO), KEY_SET_INFO,
521 sizeof(struct mgs_send_param), msp, NULL);
523 CERROR("Failed to set parameter: %d\n", rc);
529 char *ll_get_fsname(struct inode *inode)
531 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
535 OBD_ALLOC(fsname, MGS_PARAM_MAXLEN);
536 len = strlen(lsi->lsi_lmd->lmd_profile);
537 ptr = strrchr(lsi->lsi_lmd->lmd_profile, '-');
538 if (ptr && (strcmp(ptr, "-client") == 0))
540 strncpy(fsname, lsi->lsi_lmd->lmd_profile, len);
546 int ll_dir_setstripe(struct inode *inode, struct lov_user_md *lump,
549 struct ll_sb_info *sbi = ll_i2sbi(inode);
550 struct md_op_data *op_data;
551 struct ptlrpc_request *req = NULL;
553 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
554 struct obd_device *mgc = lsi->lsi_mgc;
555 char *fsname = NULL, *param = NULL;
558 * This is coming from userspace, so should be in
559 * local endian. But the MDS would like it in little
560 * endian, so we swab it before we send it.
562 if (lump->lmm_magic != LOV_USER_MAGIC)
565 if (lump->lmm_magic != cpu_to_le32(LOV_USER_MAGIC))
566 lustre_swab_lov_user_md(lump);
568 op_data = ll_prep_md_op_data(NULL, inode, NULL, NULL, 0, 0,
569 LUSTRE_OPC_ANY, NULL);
571 RETURN(PTR_ERR(op_data));
573 /* swabbing is done in lov_setstripe() on server side */
574 rc = md_setattr(sbi->ll_md_exp, op_data, lump, sizeof(*lump),
575 NULL, 0, &req, NULL);
576 ll_finish_md_op_data(op_data);
577 ptlrpc_req_finished(req);
579 if (rc != -EPERM && rc != -EACCES)
580 CERROR("mdc_setattr fails: rc = %d\n", rc);
583 if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
584 OBD_ALLOC(param, MGS_PARAM_MAXLEN);
586 /* Get fsname and assume devname to be -MDT0000. */
587 fsname = ll_get_fsname(inode);
588 /* Set root stripesize */
589 sprintf(param, "%s-MDT0000.lov.stripesize=%u", fsname,
590 lump->lmm_stripe_size);
591 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
595 /* Set root stripecount */
596 sprintf(param, "%s-MDT0000.lov.stripecount=%u", fsname,
597 lump->lmm_stripe_count);
598 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
602 /* Set root stripeoffset */
603 sprintf(param, "%s-MDT0000.lov.stripeoffset=%u", fsname,
604 lump->lmm_stripe_offset);
605 rc = ll_send_mgc_param(mgc->u.cli.cl_mgc_mgsexp, param);
610 OBD_FREE(fsname, MGS_PARAM_MAXLEN);
612 OBD_FREE(param, MGS_PARAM_MAXLEN);
617 int ll_dir_getstripe(struct inode *inode, struct lov_mds_md **lmmp,
618 int *lmm_size, struct ptlrpc_request **request)
620 struct ll_sb_info *sbi = ll_i2sbi(inode);
621 struct mdt_body *body;
622 struct lov_mds_md *lmm = NULL;
623 struct ptlrpc_request *req = NULL;
627 rc = ll_get_max_mdsize(sbi, &lmmsize);
631 oc = ll_mdscapa_get(inode);
632 rc = md_getattr(sbi->ll_md_exp, ll_inode2fid(inode),
633 oc, OBD_MD_FLEASIZE | OBD_MD_FLDIREA,
637 CDEBUG(D_INFO, "md_getattr failed on inode "
638 "%lu/%u: rc %d\n", inode->i_ino,
639 inode->i_generation, rc);
643 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
644 LASSERT(body != NULL); /* checked by md_getattr_name */
645 /* swabbed by mdc_getattr_name */
646 LASSERT(lustre_rep_swabbed(req, REPLY_REC_OFF));
648 lmmsize = body->eadatasize;
650 if (!(body->valid & (OBD_MD_FLEASIZE | OBD_MD_FLDIREA)) ||
652 GOTO(out, rc = -ENODATA);
655 lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, lmmsize);
656 LASSERT(lmm != NULL);
657 LASSERT(lustre_rep_swabbed(req, REPLY_REC_OFF + 1));
660 * This is coming from the MDS, so is probably in
661 * little endian. We convert it to host endian before
662 * passing it to userspace.
664 if (lmm->lmm_magic == __swab32(LOV_MAGIC)) {
665 lustre_swab_lov_user_md((struct lov_user_md *)lmm);
666 lustre_swab_lov_user_md_objects((struct lov_user_md *)lmm);
675 static int ll_dir_ioctl(struct inode *inode, struct file *file,
676 unsigned int cmd, unsigned long arg)
678 struct ll_sb_info *sbi = ll_i2sbi(inode);
679 struct obd_ioctl_data *data;
682 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), cmd=%#x\n",
683 inode->i_ino, inode->i_generation, inode, cmd);
685 /* asm-ppc{,64} declares TCGETS, et. al. as type 't' not 'T' */
686 if (_IOC_TYPE(cmd) == 'T' || _IOC_TYPE(cmd) == 't') /* tty ioctls */
689 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_IOCTL, 1);
691 case EXT3_IOC_GETFLAGS:
692 case EXT3_IOC_SETFLAGS:
693 RETURN(ll_iocontrol(inode, file, cmd, arg));
694 case EXT3_IOC_GETVERSION_OLD:
695 case EXT3_IOC_GETVERSION:
696 RETURN(put_user(inode->i_generation, (int *)arg));
697 /* We need to special case any other ioctls we want to handle,
698 * to send them to the MDS/OST as appropriate and to properly
699 * network encode the arg field.
700 case EXT3_IOC_SETVERSION_OLD:
701 case EXT3_IOC_SETVERSION:
703 case IOC_MDC_LOOKUP: {
704 struct ptlrpc_request *request = NULL;
705 int namelen, rc, len = 0;
710 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
715 filename = data->ioc_inlbuf1;
716 namelen = data->ioc_inllen1;
719 CDEBUG(D_INFO, "IOC_MDC_LOOKUP missing filename\n");
720 GOTO(out, rc = -EINVAL);
723 oc = ll_mdscapa_get(inode);
724 rc = md_getattr_name(sbi->ll_md_exp, ll_inode2fid(inode), oc,
725 filename, namelen, OBD_MD_FLID, 0,
726 ll_i2suppgid(inode), &request);
729 CDEBUG(D_INFO, "md_getattr_name: %d\n", rc);
733 ptlrpc_req_finished(request);
737 obd_ioctl_freedata(buf, len);
740 case LL_IOC_LOV_SETSTRIPE: {
741 struct lov_user_md lum, *lump = (struct lov_user_md *)arg;
745 LASSERT(sizeof(lum) == sizeof(*lump));
746 LASSERT(sizeof(lum.lmm_objects[0]) ==
747 sizeof(lump->lmm_objects[0]));
748 rc = copy_from_user(&lum, lump, sizeof(lum));
752 if (inode->i_sb->s_root == file->f_dentry)
755 rc = ll_dir_setstripe(inode, &lum, set_default);
759 case LL_IOC_OBD_STATFS:
760 RETURN(ll_obd_statfs(inode, (void *)arg));
761 case LL_IOC_LOV_GETSTRIPE:
762 case LL_IOC_MDC_GETINFO:
763 case IOC_MDC_GETFILEINFO:
764 case IOC_MDC_GETFILESTRIPE: {
765 struct ptlrpc_request *request = NULL;
766 struct lov_user_md *lump;
767 struct lov_mds_md *lmm = NULL;
768 struct mdt_body *body;
769 char *filename = NULL;
772 if (cmd == IOC_MDC_GETFILEINFO ||
773 cmd == IOC_MDC_GETFILESTRIPE) {
774 filename = getname((const char *)arg);
775 if (IS_ERR(filename))
776 RETURN(PTR_ERR(filename));
778 rc = ll_lov_getstripe_ea_info(inode, filename, &lmm,
781 rc = ll_dir_getstripe(inode, &lmm, &lmmsize, &request);
785 body = lustre_msg_buf(request->rq_repmsg,
786 REPLY_REC_OFF, sizeof(*body));
787 LASSERT(body != NULL); /* checked by md_getattr_name */
788 /* swabbed by md_getattr_name */
789 LASSERT(lustre_rep_swabbed(request, REPLY_REC_OFF));
795 if (rc == -ENODATA && (cmd == IOC_MDC_GETFILEINFO ||
796 cmd == LL_IOC_MDC_GETINFO))
797 GOTO(skip_lmm, rc = 0);
802 if (cmd == IOC_MDC_GETFILESTRIPE ||
803 cmd == LL_IOC_LOV_GETSTRIPE) {
804 lump = (struct lov_user_md *)arg;
806 struct lov_user_mds_data *lmdp;
807 lmdp = (struct lov_user_mds_data *)arg;
808 lump = &lmdp->lmd_lmm;
810 rc = copy_to_user(lump, lmm, lmmsize);
812 GOTO(out_lmm, rc = -EFAULT);
814 if (cmd == IOC_MDC_GETFILEINFO || cmd == LL_IOC_MDC_GETINFO) {
815 struct lov_user_mds_data *lmdp;
818 st.st_dev = inode->i_sb->s_dev;
819 st.st_mode = body->mode;
820 st.st_nlink = body->nlink;
821 st.st_uid = body->uid;
822 st.st_gid = body->gid;
823 st.st_rdev = body->rdev;
824 st.st_size = body->size;
825 st.st_blksize = CFS_PAGE_SIZE;
826 st.st_blocks = body->blocks;
827 st.st_atime = body->atime;
828 st.st_mtime = body->mtime;
829 st.st_ctime = body->ctime;
830 st.st_ino = inode->i_ino;
832 lmdp = (struct lov_user_mds_data *)arg;
833 rc = copy_to_user(&lmdp->lmd_st, &st, sizeof(st));
835 GOTO(out_lmm, rc = -EFAULT);
840 if (lmm && lmm->lmm_magic == LOV_MAGIC_JOIN)
841 OBD_FREE(lmm, lmmsize);
843 ptlrpc_req_finished(request);
848 case IOC_LOV_GETINFO: {
849 struct lov_user_mds_data *lumd;
850 struct lov_stripe_md *lsm;
851 struct lov_user_md *lum;
852 struct lov_mds_md *lmm;
857 lumd = (struct lov_user_mds_data *)arg;
858 lum = &lumd->lmd_lmm;
860 rc = ll_get_max_mdsize(sbi, &lmmsize);
864 OBD_ALLOC(lmm, lmmsize);
865 rc = copy_from_user(lmm, lum, lmmsize);
867 GOTO(free_lmm, rc = -EFAULT);
869 rc = obd_unpackmd(sbi->ll_dt_exp, &lsm, lmm, lmmsize);
871 GOTO(free_lmm, rc = -ENOMEM);
873 rc = obd_checkmd(sbi->ll_dt_exp, sbi->ll_md_exp, lsm);
877 /* Perform glimpse_size operation. */
878 memset(&st, 0, sizeof(st));
880 rc = ll_glimpse_ioctl(sbi, lsm, &st);
884 rc = copy_to_user(&lumd->lmd_st, &st, sizeof(st));
886 GOTO(free_lsm, rc = -EFAULT);
890 obd_free_memmd(sbi->ll_dt_exp, &lsm);
892 OBD_FREE(lmm, lmmsize);
895 case OBD_IOC_LLOG_CATINFO: {
896 struct ptlrpc_request *req = NULL;
899 char *bufs[3] = { NULL }, *str;
900 int lens[3] = { sizeof(struct ptlrpc_body) };
901 int size[2] = { sizeof(struct ptlrpc_body) };
903 rc = obd_ioctl_getdata(&buf, &len, (void *)arg);
908 if (!data->ioc_inlbuf1) {
909 obd_ioctl_freedata(buf, len);
913 lens[REQ_REC_OFF] = data->ioc_inllen1;
914 bufs[REQ_REC_OFF] = data->ioc_inlbuf1;
915 if (data->ioc_inllen2) {
916 lens[REQ_REC_OFF + 1] = data->ioc_inllen2;
917 bufs[REQ_REC_OFF + 1] = data->ioc_inlbuf2;
919 lens[REQ_REC_OFF + 1] = 0;
920 bufs[REQ_REC_OFF + 1] = NULL;
923 req = ptlrpc_prep_req(sbi2mdc(sbi)->cl_import,
924 LUSTRE_LOG_VERSION, LLOG_CATINFO, 3, lens,
927 GOTO(out_catinfo, rc = -ENOMEM);
929 size[REPLY_REC_OFF] = data->ioc_plen1;
930 ptlrpc_req_set_repsize(req, 2, size);
932 rc = ptlrpc_queue_wait(req);
934 str = lustre_msg_string(req->rq_repmsg, REPLY_REC_OFF,
936 rc = copy_to_user(data->ioc_pbuf1, str, data->ioc_plen1);
938 ptlrpc_req_finished(req);
940 obd_ioctl_freedata(buf, len);
943 case OBD_IOC_QUOTACHECK: {
944 struct obd_quotactl *oqctl;
947 if (!capable(CAP_SYS_ADMIN))
950 OBD_ALLOC_PTR(oqctl);
953 oqctl->qc_type = arg;
954 rc = obd_quotacheck(sbi->ll_md_exp, oqctl);
956 CDEBUG(D_INFO, "md_quotacheck failed: rc %d\n", rc);
960 rc = obd_quotacheck(sbi->ll_dt_exp, oqctl);
962 CDEBUG(D_INFO, "obd_quotacheck failed: rc %d\n", rc);
967 case OBD_IOC_POLL_QUOTACHECK: {
968 struct if_quotacheck *check;
971 if (!capable(CAP_SYS_ADMIN))
974 OBD_ALLOC_PTR(check);
978 rc = obd_iocontrol(cmd, sbi->ll_md_exp, 0, (void *)check,
981 CDEBUG(D_QUOTA, "mdc ioctl %d failed: %d\n", cmd, rc);
982 if (copy_to_user((void *)arg, check, sizeof(*check)))
987 rc = obd_iocontrol(cmd, sbi->ll_dt_exp, 0, (void *)check,
990 CDEBUG(D_QUOTA, "osc ioctl %d failed: %d\n", cmd, rc);
991 if (copy_to_user((void *)arg, check, sizeof(*check)))
999 #ifdef HAVE_QUOTA_SUPPORT
1000 case OBD_IOC_QUOTACTL: {
1001 struct if_quotactl *qctl;
1002 struct obd_quotactl *oqctl;
1004 int cmd, type, id, rc = 0;
1006 OBD_ALLOC_PTR(qctl);
1010 OBD_ALLOC_PTR(oqctl);
1015 if (copy_from_user(qctl, (void *)arg, sizeof(*qctl)))
1016 GOTO(out_quotactl, rc = -EFAULT);
1019 type = qctl->qc_type;
1026 if (!capable(CAP_SYS_ADMIN))
1027 GOTO(out_quotactl, rc = -EPERM);
1030 if (((type == USRQUOTA && current->euid != id) ||
1031 (type == GRPQUOTA && !in_egroup_p(id))) &&
1032 !capable(CAP_SYS_ADMIN))
1033 GOTO(out_quotactl, rc = -EPERM);
1035 /* XXX: dqb_valid is borrowed as a flag to mark that
1036 * only mds quota is wanted */
1037 if (qctl->qc_dqblk.dqb_valid)
1038 qctl->obd_uuid = sbi->ll_md_exp->exp_obd->
1039 u.cli.cl_target_uuid;
1044 CERROR("unsupported quotactl op: %#x\n", cmd);
1045 GOTO(out_quotactl, rc = -ENOTTY);
1048 QCTL_COPY(oqctl, qctl);
1050 if (qctl->obd_uuid.uuid[0]) {
1051 struct obd_device *obd;
1052 struct obd_uuid *uuid = &qctl->obd_uuid;
1054 obd = class_find_client_notype(uuid,
1055 &sbi->ll_dt_exp->exp_obd->obd_uuid);
1057 GOTO(out_quotactl, rc = -ENOENT);
1059 if (cmd == Q_GETINFO)
1060 oqctl->qc_cmd = Q_GETOINFO;
1061 else if (cmd == Q_GETQUOTA)
1062 oqctl->qc_cmd = Q_GETOQUOTA;
1064 GOTO(out_quotactl, rc = -EINVAL);
1066 if (sbi->ll_md_exp->exp_obd == obd) {
1067 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1070 struct obd_export *exp;
1071 struct lov_obd *lov = &sbi->ll_dt_exp->
1074 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1075 if (!lov->lov_tgts[i] ||
1076 !lov->lov_tgts[i]->ltd_active)
1078 exp = lov->lov_tgts[i]->ltd_exp;
1079 if (exp->exp_obd == obd) {
1080 rc = obd_quotactl(exp, oqctl);
1086 oqctl->qc_cmd = cmd;
1087 QCTL_COPY(qctl, oqctl);
1089 if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1092 GOTO(out_quotactl, rc);
1095 rc = obd_quotactl(sbi->ll_md_exp, oqctl);
1096 if (rc && rc != -EBUSY && cmd == Q_QUOTAON) {
1097 oqctl->qc_cmd = Q_QUOTAOFF;
1098 obd_quotactl(sbi->ll_md_exp, oqctl);
1101 QCTL_COPY(qctl, oqctl);
1103 if (copy_to_user((void *)arg, qctl, sizeof(*qctl)))
1107 OBD_FREE_PTR(oqctl);
1110 #endif /* HAVE_QUOTA_SUPPORT */
1111 case OBD_IOC_GETNAME: {
1112 struct obd_device *obd = class_exp2obd(sbi->ll_dt_exp);
1115 if (copy_to_user((void *)arg, obd->obd_name,
1116 strlen(obd->obd_name) + 1))
1120 case LL_IOC_FLUSHCTX:
1121 RETURN(ll_flush_ctx(inode));
1122 #ifdef CONFIG_FS_POSIX_ACL
1123 case LL_IOC_RMTACL: {
1124 if (sbi->ll_flags & LL_SBI_RMT_CLIENT &&
1125 inode == inode->i_sb->s_root->d_inode) {
1126 struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1129 LASSERT(fd != NULL);
1130 rc = rct_add(&sbi->ll_rct, cfs_curproc_pid(), arg);
1132 fd->fd_flags |= LL_FILE_RMTACL;
1139 RETURN(obd_iocontrol(cmd, sbi->ll_dt_exp,0,NULL,(void *)arg));
1143 int ll_dir_open(struct inode *inode, struct file *file)
1146 RETURN(ll_file_open(inode, file));
1149 int ll_dir_release(struct inode *inode, struct file *file)
1152 RETURN(ll_file_release(inode, file));
1155 struct file_operations ll_dir_operations = {
1156 .open = ll_dir_open,
1157 .release = ll_dir_release,
1158 .read = generic_read_dir,
1159 .readdir = ll_readdir,
1160 .ioctl = ll_dir_ioctl