1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
5 * Lustre Metadata Server (mds) handling of striped file data
7 * Copyright (C) 2001-2003 Cluster File Systems, Inc.
9 * This file is part of Lustre, http://www.lustre.org.
11 * Lustre is free software; you can redistribute it and/or
12 * modify it under the terms of version 2 of the GNU General Public
13 * License as published by the Free Software Foundation.
15 * Lustre is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with Lustre; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
26 # define EXPORT_SYMTAB
28 #define DEBUG_SUBSYSTEM S_MDS
30 #include <linux/module.h>
31 #include <linux/dcache.h>
32 #include <linux/namei.h>
33 #include <linux/obd_support.h>
34 #include <linux/obd_class.h>
35 #include <linux/obd.h>
36 #include <linux/lustre_lib.h>
37 #include <linux/lustre_mds.h>
38 #include <linux/lustre_idl.h>
39 #include <linux/obd_class.h>
40 #include <linux/obd_lov.h>
41 #include <linux/lustre_lib.h>
42 #include <linux/lustre_fsfilt.h>
43 #include <linux/lustre_lite.h>
44 #include <linux/lustre_sec.h>
45 #include <asm/div64.h>
47 #include "mds_internal.h"
51 * - magic in mea struct
53 int mds_md_connect(struct obd_device *obd, char *md_name)
55 struct mds_obd *mds = &obd->u.mds;
56 struct lustre_handle conn = {0};
57 unsigned long sec_flags = PTLRPC_SEC_FL_MDS;
62 if (IS_ERR(mds->mds_md_obd))
63 RETURN(PTR_ERR(mds->mds_md_obd));
65 if (mds->mds_md_connected)
68 down(&mds->mds_md_sem);
69 if (mds->mds_md_connected) {
74 mds->mds_md_obd = class_name2obd(md_name);
75 if (!mds->mds_md_obd) {
76 CERROR("MDS cannot locate MD(LMV) %s\n",
78 mds->mds_md_obd = ERR_PTR(-ENOTCONN);
79 GOTO(err_last, rc = -ENOTCONN);
82 rc = obd_connect(&conn, mds->mds_md_obd, &obd->obd_uuid, NULL,
83 OBD_OPT_MDS_CONNECTION);
85 CERROR("MDS cannot connect to MD(LMV) %s (%d)\n",
87 mds->mds_md_obd = ERR_PTR(rc);
90 mds->mds_md_exp = class_conn2export(&conn);
91 if (mds->mds_md_exp == NULL)
92 CERROR("can't get export!\n");
94 rc = obd_register_observer(mds->mds_md_obd, obd);
96 CERROR("MDS cannot register as observer of MD(LMV) %s, "
97 "rc = %d\n", md_name, rc);
101 /* retrieve size of EA */
102 rc = obd_get_info(mds->mds_md_exp, strlen("mdsize"),
103 "mdsize", &valsize, &value);
107 if (value > mds->mds_max_mdsize)
108 mds->mds_max_mdsize = value;
110 /* find our number in LMV cluster */
111 rc = obd_get_info(mds->mds_md_exp, strlen("mdsnum"),
112 "mdsnum", &valsize, &value);
116 mds->mds_num = value;
118 rc = obd_set_info(mds->mds_md_exp, strlen("inter_mds"),
119 "inter_mds", 0, NULL);
123 if (mds->mds_mds_sec) {
124 rc = obd_set_info(mds->mds_md_exp, strlen("sec"), "sec",
125 strlen(mds->mds_mds_sec), mds->mds_mds_sec);
130 rc = obd_set_info(mds->mds_md_exp, strlen("sec_flags"), "sec_flags",
131 sizeof(sec_flags), &sec_flags);
135 mds->mds_md_connected = 1;
136 up(&mds->mds_md_sem);
140 obd_register_observer(mds->mds_md_obd, NULL);
142 obd_disconnect(mds->mds_md_exp, 0);
143 mds->mds_md_exp = NULL;
144 mds->mds_md_obd = ERR_PTR(rc);
146 up(&mds->mds_md_sem);
150 int mds_md_postsetup(struct obd_device *obd)
152 struct mds_obd *mds = &obd->u.mds;
156 if (mds->mds_md_exp) {
157 rc = obd_init_ea_size(mds->mds_md_exp,
159 mds->mds_max_cookiesize);
165 int mds_md_disconnect(struct obd_device *obd, int flags)
167 struct mds_obd *mds = &obd->u.mds;
171 if (!mds->mds_md_connected)
174 down(&mds->mds_md_sem);
175 if (!IS_ERR(mds->mds_md_obd) && mds->mds_md_exp != NULL) {
176 LASSERT(mds->mds_md_connected);
178 obd_register_observer(mds->mds_md_obd, NULL);
180 if (flags & OBD_OPT_FORCE) {
181 struct obd_device *lmv_obd;
182 struct obd_ioctl_data ioc_data = { 0 };
184 lmv_obd = class_exp2obd(mds->mds_md_exp);
189 * making disconnecting lmv stuff do not send anything
190 * to all remote MDSs from LMV. This is needed to
191 * prevent possible hanging with endless recovery, when
192 * MDS sends disconnect to already disconnected
193 * target. Probably this is wrong, but client does the
194 * same in --force mode and I do not see why can't we do
197 lmv_obd->obd_no_recov = 1;
198 obd_iocontrol(IOC_OSC_SET_ACTIVE, mds->mds_md_exp,
199 sizeof(ioc_data), &ioc_data, NULL);
203 * if obd_disconnect() fails (probably because the export was
204 * disconnected by class_disconnect_exports()) then we just need
207 mds->mds_md_connected = 0;
208 rc = obd_disconnect(mds->mds_md_exp, flags);
210 class_export_put(mds->mds_md_exp);
213 mds->mds_md_exp = NULL;
214 mds->mds_md_obd = NULL;
216 up(&mds->mds_md_sem);
220 int mds_md_reconnect(struct obd_device *obd)
222 struct mds_obd *mds = &obd->u.mds;
223 struct obd_statfs osfs;
227 /* We don't know state of connections to another MDSes
228 * before the failure. If MDS we were connected to before
229 * the failure gets failed, then it will wait for us to
230 * reconnect and will timed recovery out. bug 4920 */
231 if (mds->mds_md_connected == 0)
233 if (mds->mds_md_obd == NULL)
236 err = obd_statfs(mds->mds_md_obd, &osfs, jiffies - HZ);
238 CERROR("can't reconnect to MDSes after recovery: %d\n", err);
242 int mds_md_get_attr(struct obd_device *obd, struct inode *inode,
243 struct mea **mea, int *mea_size)
245 struct mds_obd *mds = &obd->u.mds;
249 if (!mds->mds_md_obd)
251 if (!S_ISDIR(inode->i_mode))
254 /* first calculate mea size */
255 *mea_size = obd_alloc_diskmd(mds->mds_md_exp,
256 (struct lov_mds_md **)mea);
257 if (*mea_size < 0 || *mea == NULL)
258 return *mea_size < 0 ? *mea_size : -EINVAL;
260 rc = mds_get_md(obd, inode, *mea, mea_size, 1, 1);
263 OBD_FREE(*mea, *mea_size);
281 #define DIR_ROUND (DIR_PAD - 1)
282 #define DIR_REC_LEN(name_len) (((name_len) + 16 + DIR_ROUND) & ~DIR_ROUND)
284 /* this struct holds dir entries for particular MDS to be flushed */
286 struct list_head list;
291 struct brw_page brwc;
294 struct dirsplit_control {
295 struct obd_device *obd;
297 struct dentry *dentry;
299 struct dir_cache *cache;
302 static int dc_new_page_to_cache(struct dir_cache * dirc)
306 if (!list_empty(&dirc->list) && dirc->free > sizeof(__u16)) {
307 /* current page became full, mark the end */
308 struct dir_entry *de = dirc->cur;
312 page = alloc_page(GFP_KERNEL);
315 list_add_tail(&page->lru, &dirc->list);
316 dirc->cur = page_address(page);
317 dirc->free = PAGE_SIZE;
321 static int retrieve_generation_numbers(struct dirsplit_control *dc, void *buf)
323 struct mds_obd *mds = &dc->obd->u.mds;
324 struct dir_entry *de;
325 struct dentry *dentry;
328 end = buf + PAGE_SIZE;
329 de = (struct dir_entry *) buf;
330 while ((char *) de < end && de->namelen) {
331 /* lookup an inode */
332 LASSERT(de->namelen <= 255);
333 dentry = ll_lookup_one_len(de->name, dc->dentry,
335 if (IS_ERR(dentry)) {
336 CERROR("can't lookup %*s: %d\n", de->namelen,
337 de->name, (int) PTR_ERR(dentry));
340 if (dentry->d_inode != NULL) {
342 struct lustre_id sid;
344 down(&dentry->d_inode->i_sem);
345 rc = mds_read_inode_sid(dc->obd,
346 dentry->d_inode, &sid);
347 up(&dentry->d_inode->i_sem);
349 CERROR("Can't read inode self id, "
350 "inode %lu, rc %d\n",
351 dentry->d_inode->i_ino, rc);
355 de->fid = id_fid(&sid);
356 de->mds = mds->mds_num;
357 de->ino = dentry->d_inode->i_ino;
358 de->generation = dentry->d_inode->i_generation;
359 } else if (dentry->d_flags & DCACHE_CROSS_REF) {
360 de->fid = dentry->d_fid;
361 de->ino = dentry->d_inum;
362 de->mds = dentry->d_mdsnum;
363 de->generation = dentry->d_generation;
365 CERROR("can't lookup %*s\n", de->namelen, de->name);
371 de = (struct dir_entry *)
372 ((char *) de + DIR_REC_LEN(de->namelen));
377 static int flush_buffer_onto_mds(struct dirsplit_control *dc, int mdsnum)
379 struct mds_obd *mds = &dc->obd->u.mds;
380 struct list_head *cur, *tmp;
381 struct dir_cache *ca;
384 ca = dc->cache + mdsnum;
386 if (ca->free > sizeof(__u16)) {
387 /* current page became full, mark the end */
388 struct dir_entry *de = ca->cur;
392 list_for_each_safe(cur, tmp, &ca->list) {
395 page = list_entry(cur, struct page, lru);
396 LASSERT(page != NULL);
398 retrieve_generation_numbers(dc, page_address(page));
401 ca->brwc.disk_offset = ca->brwc.page_offset = 0;
402 ca->brwc.count = PAGE_SIZE;
404 ca->oa.o_mds = mdsnum;
405 rc = obd_brw(OBD_BRW_WRITE, mds->mds_md_exp, &ca->oa,
406 (struct lov_stripe_md *) dc->mea,
415 static int remove_entries_from_orig_dir(struct dirsplit_control *dc, int mdsnum)
417 struct list_head *cur, *tmp;
418 struct dentry *dentry;
419 struct dir_cache *ca;
420 struct dir_entry *de;
426 ca = dc->cache + mdsnum;
427 list_for_each_safe(cur, tmp, &ca->list) {
428 page = list_entry(cur, struct page, lru);
429 buf = page_address(page);
430 end = buf + PAGE_SIZE;
432 de = (struct dir_entry *) buf;
433 while ((char *) de < end && de->namelen) {
434 /* lookup an inode */
435 LASSERT(de->namelen <= 255);
437 dentry = ll_lookup_one_len(de->name, dc->dentry,
439 if (IS_ERR(dentry)) {
440 CERROR("can't lookup %*s: %d\n", de->namelen,
441 de->name, (int) PTR_ERR(dentry));
444 rc = fsfilt_del_dir_entry(dc->obd, dentry);
447 de = (struct dir_entry *)
448 ((char *) de + DIR_REC_LEN(de->namelen));
454 static int filldir(void * __buf, const char * name, int namlen,
455 loff_t offset, ino_t ino, unsigned int d_type)
457 struct dirsplit_control *dc = __buf;
458 struct mds_obd *mds = &dc->obd->u.mds;
459 struct dir_cache *ca;
460 struct dir_entry *de;
465 if (name[0] == '.' &&
466 (namlen == 1 || (namlen == 2 && name[1] == '.'))) {
467 /* skip special entries */
472 newmds = mea_name2idx(dc->mea, (char *) name, namlen);
474 if (newmds == mds->mds_num) {
475 /* this entry remains on the current MDS, skip moving */
479 OBD_ALLOC(n, namlen + 1);
481 /* XXX: should be memory allocation checked here in some more smart
485 memcpy(n, name, namlen);
486 n[namlen] = (char) 0;
488 OBD_FREE(n, namlen + 1);
490 /* check for space in buffer for new entry */
491 ca = dc->cache + newmds;
492 if (DIR_REC_LEN(namlen) > ca->free) {
493 int err = dc_new_page_to_cache(ca);
497 /* insert found entry into buffer to be flushed later */
498 /* NOTE: we'll fill generations number later, because we
499 * it's stored in inode, thus we need to lookup an entry,
500 * but directory is locked for readdir(), so we delay this */
504 de->namelen = namlen;
505 memcpy(de->name, name, namlen);
506 ca->cur += DIR_REC_LEN(namlen);
507 ca->free -= DIR_REC_LEN(namlen);
513 int scan_and_distribute(struct obd_device *obd, struct dentry *dentry,
516 struct inode *dir = dentry->d_inode;
517 struct dirsplit_control dc;
522 nlen = strlen("__iopen__/") + 10 + 1;
523 OBD_ALLOC(file_name, nlen);
527 i = sprintf(file_name, "__iopen__/0x%lx",
528 dentry->d_inode->i_ino);
530 file = filp_open(file_name, O_RDONLY, 0);
532 CERROR("can't open directory %s: %d\n",
533 file_name, (int) PTR_ERR(file));
534 OBD_FREE(file_name, nlen);
535 RETURN(PTR_ERR(file));
538 memset(&dc, 0, sizeof(dc));
543 OBD_ALLOC(dc.cache, sizeof(struct dir_cache) * mea->mea_count);
544 LASSERT(dc.cache != NULL);
545 for (i = 0; i < mea->mea_count; i++) {
546 INIT_LIST_HEAD(&dc.cache[i].list);
547 dc.cache[i].free = 0;
548 dc.cache[i].cached = 0;
551 err = vfs_readdir(file, filldir, &dc);
556 for (i = 0; i < mea->mea_count; i++) {
557 if (!dc.cache[i].cached)
559 err = flush_buffer_onto_mds(&dc, i);
564 for (i = 0; i < mea->mea_count; i++) {
565 if (!dc.cache[i].cached)
567 err = remove_entries_from_orig_dir(&dc, i);
574 for (i = 0; i < mea->mea_count; i++) {
575 struct list_head *cur, *tmp;
576 if (!dc.cache[i].cached)
578 list_for_each_safe(cur, tmp, &dc.cache[i].list) {
580 page = list_entry(cur, struct page, lru);
581 list_del(&page->lru);
585 OBD_FREE(dc.cache, sizeof(struct dir_cache) * mea->mea_count);
586 OBD_FREE(file_name, nlen);
591 #define MAX_DIR_SIZE (64 * 1024)
592 #define I_NON_SPLITTABLE (256)
594 int mds_splitting_expected(struct obd_device *obd, struct dentry *dentry)
596 struct mds_obd *mds = &obd->u.mds;
597 struct mea *mea = NULL;
601 if (!mds->mds_md_obd)
602 return MDS_NO_SPLITTABLE;
605 if (dentry->d_inode == NULL)
606 return MDS_NO_SPLITTABLE;
608 /* a dir can be splitted only */
609 if (!S_ISDIR(dentry->d_inode->i_mode))
610 return MDS_NO_SPLITTABLE;
612 /* already splittied or slave directory (part of splitted dir) */
613 if (dentry->d_inode->i_flags & I_NON_SPLITTABLE)
614 return MDS_NO_SPLITTABLE;
616 /* don't split root directory */
617 if (dentry->d_inode->i_ino == id_ino(&mds->mds_rootid))
618 return MDS_NO_SPLITTABLE;
620 /* large enough to be splitted? */
621 if (dentry->d_inode->i_size < MAX_DIR_SIZE)
622 return MDS_NO_SPLIT_EXPECTED;
624 mds_md_get_attr(obd, dentry->d_inode, &mea, &size);
626 /* already splitted or slave object: shouldn't be splitted */
627 rc = MDS_NO_SPLITTABLE;
628 /* mark to skip subsequent checks */
629 dentry->d_inode->i_flags |= I_NON_SPLITTABLE;
632 /* may be splitted */
633 rc = MDS_EXPECT_SPLIT;
640 * must not be called on already splitted directories.
642 int mds_try_to_split_dir(struct obd_device *obd, struct dentry *dentry,
643 struct mea **mea, int nstripes, int update_mode)
645 struct inode *dir = dentry->d_inode;
646 struct mds_obd *mds = &obd->u.mds;
647 struct mea *tmea = NULL;
648 struct obdo *oa = NULL;
649 int rc, mea_size = 0;
654 if (update_mode != LCK_EX)
657 /* TODO: optimization possible - we already may have mea here */
658 rc = mds_splitting_expected(obd, dentry);
659 if (rc == MDS_NO_SPLITTABLE)
661 if (rc == MDS_NO_SPLIT_EXPECTED && nstripes == 0)
663 if (nstripes && nstripes == 1)
666 LASSERT(mea == NULL || *mea == NULL);
668 CDEBUG(D_OTHER, "%s: split directory %u/%lu/%lu\n",
669 obd->obd_name, mds->mds_num, dir->i_ino,
670 (unsigned long) dir->i_generation);
674 mea_size = obd_size_diskmd(mds->mds_md_exp, NULL);
676 /* FIXME: Actually we may only want to allocate enough space for
677 * necessary amount of stripes, but on the other hand with this
678 * approach of allocating maximal possible amount of MDS slots,
679 * it would be easier to split the dir over more MDSes */
680 rc = obd_alloc_diskmd(mds->mds_md_exp, (void *)mea);
682 CERROR("obd_alloc_diskmd() failed, error %d.\n", rc);
688 (*mea)->mea_count = nstripes;
690 /* 1) create directory objects on slave MDS'es */
691 /* FIXME: should this be OBD method? */
696 oa->o_generation = dir->i_generation;
698 obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
699 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
700 OBD_MD_FLUID | OBD_MD_FLGID);
702 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
703 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
704 oa->o_mode = dir->i_mode;
707 * until lmv_obd_create() properly rewritten, it is important to have
708 * here oa->o_id = dir->i_ino, as otherwise master object will have
709 * invalid store cookie (zero inode num), what will lead to -ESTALE in
710 * mds_open() or somewhere else.
712 oa->o_id = dir->i_ino;
714 /* get parent id: ldlm lock on the parent protects ea */
715 rc = mds_read_inode_sid(obd, dir, &id);
717 CERROR("Can't read inode self id, inode %lu, "
718 "rc %d.\n", dir->i_ino, rc);
721 oa->o_fid = id_fid(&id);
722 oa->o_mds = mds->mds_num;
723 LASSERT(oa->o_fid != 0);
725 CDEBUG(D_OTHER, "%s: create subdirs with mode %o, uid %u, gid %u\n",
726 obd->obd_name, dir->i_mode, dir->i_uid, dir->i_gid);
728 rc = obd_create(mds->mds_md_exp, oa, NULL, 0,
729 (struct lov_stripe_md **)mea, NULL);
731 CERROR("Can't create remote inode, rc = %d\n", rc);
735 LASSERT(id_fid(&(*mea)->mea_ids[0]));
736 CDEBUG(D_OTHER, "%d dirobjects created\n", (int)(*mea)->mea_count);
738 /* 2) update dir attribute */
741 handle = fsfilt_start(obd, dir, FSFILT_OP_SETATTR, NULL);
742 if (IS_ERR(handle)) {
744 CERROR("fsfilt_start() failed: %d\n", (int) PTR_ERR(handle));
745 GOTO(err_oa, rc = PTR_ERR(handle));
748 rc = fsfilt_set_md(obd, dir, handle, *mea, mea_size, EA_MEA);
751 CERROR("fsfilt_set_md() failed, error %d.\n", rc);
755 rc = fsfilt_commit(obd, mds->mds_sb, dir, handle, 0);
758 CERROR("fsfilt_commit() failed, error %d.\n", rc);
765 /* 3) read through the dir and distribute it over objects */
766 rc = scan_and_distribute(obd, dentry, *mea);
768 obd_free_diskmd(mds->mds_md_exp, (struct lov_mds_md **)mea);
770 CERROR("scan_and_distribute() failed, error %d.\n", rc);
781 static int filter_start_page_write(struct inode *inode,
782 struct niobuf_local *lnb)
784 struct page *page = alloc_pages(GFP_HIGHUSER, 0);
786 CERROR("no memory for a temp page\n");
787 return lnb->rc = -ENOMEM;
789 POISON_PAGE(page, 0xf1);
790 page->index = lnb->offset >> PAGE_SHIFT;
795 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
796 int objcount, struct obd_ioobj *obj,
797 int niocount, struct niobuf_remote *nb,
798 struct niobuf_local *res,
799 struct obd_trans_info *oti, struct lustre_capa *capa)
801 struct niobuf_remote *rnb;
802 struct niobuf_local *lnb = NULL;
803 int rc = 0, i, tot_bytes = 0;
804 unsigned long now = jiffies;
805 struct dentry *dentry;
809 LASSERT(objcount == 1);
810 LASSERT(obj->ioo_bufcnt > 0);
812 memset(res, 0, niocount * sizeof(*res));
816 id_ino(&id) = obj->ioo_id;
817 id_gen(&id) = obj->ioo_gr;
819 dentry = mds_id2dentry(exp->exp_obd, &id, NULL);
820 if (IS_ERR(dentry)) {
821 CERROR("can't get dentry for "LPU64"/%u: %d\n",
822 id.li_stc.u.e3s.l3s_ino, id.li_stc.u.e3s.l3s_gen,
823 (int)PTR_ERR(dentry));
824 GOTO(cleanup, rc = (int) PTR_ERR(dentry));
827 if (dentry->d_inode == NULL) {
828 CERROR("trying to BRW to non-existent file "LPU64"\n",
831 GOTO(cleanup, rc = -ENOENT);
834 if (time_after(jiffies, now + 15 * HZ))
835 CERROR("slow preprw_write setup %lus\n", (jiffies - now) / HZ);
837 CDEBUG(D_INFO, "preprw_write setup: %lu jiffies\n",
840 for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
842 lnb->dentry = dentry;
843 lnb->offset = rnb->offset;
845 lnb->flags = rnb->flags;
847 rc = filter_start_page_write(dentry->d_inode, lnb);
849 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR, "page err %u@"
850 LPU64" %u/%u %p: rc %d\n", lnb->len, lnb->offset,
851 i, obj->ioo_bufcnt, dentry, rc);
853 __free_pages(lnb->page, 0);
857 tot_bytes += lnb->len;
860 if (time_after(jiffies, now + 15 * HZ))
861 CERROR("slow start_page_write %lus\n", (jiffies - now) / HZ);
863 CDEBUG(D_INFO, "start_page_write: %lu jiffies\n",
871 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
872 int objcount, struct obd_ioobj *obj, int niocount,
873 struct niobuf_local *res, struct obd_trans_info *oti,
876 struct obd_device *obd = exp->exp_obd;
877 struct niobuf_local *lnb;
878 struct inode *inode = NULL;
879 int rc = 0, i, cleanup_phase = 0, err, entries = 0;
882 LASSERT(objcount == 1);
883 LASSERT(current->journal_info == NULL);
886 inode = res->dentry->d_inode;
888 for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
890 struct dir_entry *de;
892 buf = kmap(lnb->page);
893 LASSERT(buf != NULL);
894 end = buf + lnb->len;
895 de = (struct dir_entry *) buf;
896 while ((char *) de < end && de->namelen) {
897 err = fsfilt_add_dir_entry(obd, res->dentry, de->name,
898 de->namelen, de->ino,
899 de->generation, de->mds,
902 CERROR("can't add dir entry %*s->%u/%u/%u"
903 " to %lu/%u: %d\n", de->namelen,
904 de->name, de->mds, (unsigned)de->ino,
905 (unsigned) de->generation,
906 res->dentry->d_inode->i_ino,
907 res->dentry->d_inode->i_generation,
913 de = (struct dir_entry *)
914 ((char *) de + DIR_REC_LEN(de->namelen));
920 for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++)
921 __free_page(lnb->page);
927 int mds_choose_mdsnum(struct obd_device *obd, const char *name, int len, int flags,
928 struct ptlrpc_peer *peer, struct inode *parent, int local)
930 struct mds_obd *mds = &obd->u.mds;
931 char peer_str[PTL_NALFMT_SIZE];
932 int i = mds->mds_num;
937 RETURN(mds->mds_num);
939 if (flags & REC_REINT_CREATE) {
941 } else if (mds->mds_md_exp != NULL && peer != NULL) {
942 LASSERT(parent != NULL);
943 /* distribute only at root level */
944 lmv = &mds->mds_md_exp->exp_obd->u.lmv;
945 if (parent->i_ino != id_ino(&mds->mds_rootid)) {
948 __u64 nid = peer->peer_id.nid;
949 __u64 count = lmv->desc.ld_tgt_count;
950 i = do_div(nid, count);
951 CDEBUG(D_OTHER, "client from %s creates top dir %*s "
953 ptlrpc_peernid2str(peer,peer_str), len, name, i+1);
955 } else if (mds->mds_md_exp) {
956 lmv = &mds->mds_md_exp->exp_obd->u.lmv;
957 i = raw_name2idx(MEA_MAGIC_LAST_CHAR, lmv->desc.ld_tgt_count, name, len);
962 int mds_lock_slave_objs(struct obd_device *obd, struct dentry *dentry,
963 struct lustre_handle **rlockh)
965 struct mds_obd *mds = &obd->u.mds;
966 struct mdc_op_data *op_data;
967 struct lookup_intent it;
968 struct mea *mea = NULL;
973 LASSERT(rlockh != NULL);
974 LASSERT(dentry != NULL);
975 LASSERT(dentry->d_inode != NULL);
978 if (!mds->mds_md_obd)
981 /* a dir can be splitted only */
982 if (!S_ISDIR(dentry->d_inode->i_mode))
985 rc = mds_md_get_attr(obd, dentry->d_inode,
993 if (mea->mea_count == 0)
994 /* this is slave object */
995 GOTO(cleanup, rc = 0);
997 CDEBUG(D_OTHER, "%s: lock slaves for %lu/%lu\n",
998 obd->obd_name, (unsigned long)dentry->d_inode->i_ino,
999 (unsigned long)dentry->d_inode->i_generation);
1001 handle_size = sizeof(struct lustre_handle) *
1004 OBD_ALLOC(*rlockh, handle_size);
1005 if (*rlockh == NULL)
1006 GOTO(cleanup, rc = -ENOMEM);
1008 memset(*rlockh, 0, handle_size);
1009 OBD_ALLOC(op_data, sizeof(*op_data));
1010 if (op_data == NULL) {
1011 OBD_FREE(*rlockh, handle_size);
1014 memset(op_data, 0, sizeof(*op_data));
1016 op_data->mea1 = mea;
1017 it.it_op = IT_UNLINK;
1019 OBD_ALLOC(it.d.fs_data, sizeof(struct lustre_intent_data));
1020 if (!it.d.fs_data) {
1021 OBD_FREE(*rlockh, handle_size);
1022 OBD_FREE(op_data, sizeof(*op_data));
1026 rc = md_enqueue(mds->mds_md_exp, LDLM_IBITS, &it, LCK_EX,
1027 op_data, *rlockh, NULL, 0, ldlm_completion_ast,
1028 mds_blocking_ast, NULL);
1030 OBD_FREE(op_data, sizeof(*op_data));
1031 OBD_FREE(it.d.fs_data, sizeof(struct lustre_intent_data));
1034 OBD_FREE(mea, mea_size);
1038 void mds_unlock_slave_objs(struct obd_device *obd, struct dentry *dentry,
1039 struct lustre_handle *lockh)
1041 struct mds_obd *mds = &obd->u.mds;
1042 struct mea *mea = NULL;
1043 int mea_size, rc, i;
1046 if (lockh == NULL) {
1051 LASSERT(mds->mds_md_obd != NULL);
1052 LASSERT(S_ISDIR(dentry->d_inode->i_mode));
1054 rc = mds_md_get_attr(obd, dentry->d_inode, &mea, &mea_size);
1056 CERROR("locks are leaked\n");
1060 LASSERT(mea_size != 0);
1061 LASSERT(mea != NULL);
1062 LASSERT(mea->mea_count != 0);
1064 CDEBUG(D_OTHER, "%s: unlock slaves for %lu/%lu\n", obd->obd_name,
1065 (unsigned long) dentry->d_inode->i_ino,
1066 (unsigned long) dentry->d_inode->i_generation);
1068 for (i = 0; i < mea->mea_count; i++) {
1069 if (lockh[i].cookie != 0)
1070 ldlm_lock_decref(lockh + i, LCK_EX);
1073 OBD_FREE(lockh, sizeof(struct lustre_handle) * mea->mea_count);
1074 OBD_FREE(mea, mea_size);
1078 int mds_unlink_slave_objs(struct obd_device *obd, struct dentry *dentry)
1080 struct mds_obd *mds = &obd->u.mds;
1081 struct ptlrpc_request *req = NULL;
1082 struct mdc_op_data *op_data;
1083 struct mea *mea = NULL;
1087 /* clustered MD ? */
1088 if (!mds->mds_md_obd)
1091 /* a dir can be splitted only */
1092 if (!S_ISDIR(dentry->d_inode->i_mode))
1095 rc = mds_md_get_attr(obd, dentry->d_inode, &mea, &mea_size);
1102 if (mea->mea_count == 0)
1103 GOTO(cleanup, rc = 0);
1105 CDEBUG(D_OTHER, "%s: unlink slaves for %lu/%lu\n", obd->obd_name,
1106 (unsigned long)dentry->d_inode->i_ino,
1107 (unsigned long)dentry->d_inode->i_generation);
1109 OBD_ALLOC(op_data, sizeof(*op_data));
1110 if (op_data == NULL)
1113 memset(op_data, 0, sizeof(*op_data));
1114 op_data->mea1 = mea;
1115 rc = md_unlink(mds->mds_md_exp, op_data, &req);
1116 OBD_FREE(op_data, sizeof(*op_data));
1117 LASSERT(req == NULL);
1120 OBD_FREE(mea, mea_size);
1124 struct ide_tracking {
1129 int mds_ide_filldir(void *__buf, const char *name, int namelen,
1130 loff_t offset, ino_t ino, unsigned int d_type)
1132 struct ide_tracking *it = __buf;
1138 if (it->entries > 2)
1142 if (name[0] == '.' && namelen == 1)
1144 if (name[0] == '.' && name[1] == '.' && namelen == 2)
1151 /* checks if passed dentry points to empty dir. */
1152 int mds_is_dir_empty(struct obd_device *obd, struct dentry *dentry)
1154 struct ide_tracking it;
1162 nlen = strlen("__iopen__/") + 10 + 1;
1163 OBD_ALLOC(file_name, nlen);
1167 LASSERT(dentry->d_inode != NULL);
1168 i = sprintf(file_name, "__iopen__/0x%lx",
1169 dentry->d_inode->i_ino);
1171 file = filp_open(file_name, O_RDONLY, 0);
1173 CERROR("can't open directory %s: %d\n",
1174 file_name, (int) PTR_ERR(file));
1175 GOTO(cleanup, rc = PTR_ERR(file));
1178 rc = vfs_readdir(file, mds_ide_filldir, &it);
1179 filp_close(file, 0);
1181 if (it.empty && rc == 0)
1187 OBD_FREE(file_name, nlen);
1191 int mds_lock_and_check_slave(int offset, struct ptlrpc_request *req,
1192 struct lustre_handle *lockh)
1194 struct obd_device *obd = req->rq_export->exp_obd;
1195 struct dentry *dentry = NULL;
1196 struct lvfs_run_ctxt saved;
1197 int cleanup_phase = 0;
1198 struct mds_req_sec_desc *rsd;
1199 struct mds_body *body;
1200 struct lvfs_ucred uc;
1201 int rc, update_mode;
1204 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1206 CERROR("Can't unpack security desc\n");
1207 GOTO(cleanup, rc = -EFAULT);
1210 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1211 lustre_swab_mds_body);
1213 CERROR("Can't swab mds_body\n");
1214 GOTO(cleanup, rc = -EFAULT);
1216 CDEBUG(D_OTHER, "%s: check slave "DLID4"\n", obd->obd_name,
1219 dentry = mds_id2locked_dentry(obd, &body->id1, NULL, LCK_EX,
1220 lockh, &update_mode, NULL, 0,
1221 MDS_INODELOCK_UPDATE);
1222 if (IS_ERR(dentry)) {
1223 CERROR("can't find inode: %d\n", (int) PTR_ERR(dentry));
1224 GOTO(cleanup, rc = PTR_ERR(dentry));
1229 * handling the case when remote MDS checks if dir is empty
1230 * before rename. But it also does it for all entries, because
1231 * inode is stored here and remote MDS does not know if rename
1232 * point to dir or to reg file. So we check it here.
1234 if (!S_ISDIR(dentry->d_inode->i_mode))
1235 GOTO(cleanup, rc = 0);
1237 rc = mds_init_ucred(&uc, req, rsd);
1242 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1243 rc = mds_is_dir_empty(obd, dentry) ? 0 : -ENOTEMPTY;
1244 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1246 mds_exit_ucred(&uc);
1249 switch(cleanup_phase) {
1252 ldlm_lock_decref(lockh, LCK_EX);
1260 int mds_convert_mea_ea(struct obd_device *obd, struct inode *inode,
1261 struct lov_mds_md *lmm, int lmm_size)
1263 struct lov_stripe_md *lsm = NULL;
1264 struct mea_old *old;
1270 mea = (struct mea *)lmm;
1271 old = (struct mea_old *)lmm;
1273 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
1274 mea->mea_magic == MEA_MAGIC_ALL_CHARS)
1278 * making MDS try LOV EA converting in the non-LMV configuration
1281 if (!obd->u.mds.mds_md_exp)
1284 CDEBUG(D_INODE, "converting MEA EA on %lu/%u from V0 to V1 (%u/%u)\n",
1285 inode->i_ino, inode->i_generation, old->mea_count,
1288 rc = obd_unpackmd(obd->u.mds.mds_md_exp, &lsm, lmm, lmm_size);
1292 rc = obd_packmd(obd->u.mds.mds_md_exp, &lmm, lsm);
1294 GOTO(conv_free, rc);
1297 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1298 if (IS_ERR(handle)) {
1299 rc = PTR_ERR(handle);
1300 GOTO(conv_free, rc);
1303 rc = fsfilt_set_md(obd, inode, handle, lmm, lmm_size, EA_MEA);
1304 err = fsfilt_commit(obd, obd->u.mds.mds_sb, inode, handle, 0);
1306 rc = err ? err : lmm_size;
1307 GOTO(conv_free, rc);
1309 obd_free_memmd(obd->u.mds.mds_md_exp, &lsm);