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/lustre_mds.h>
32 #include <linux/lustre_idl.h>
33 #include <linux/obd_class.h>
34 #include <linux/obd_lov.h>
35 #include <linux/lustre_lib.h>
36 #include <linux/lustre_fsfilt.h>
38 #include "mds_internal.h"
43 * - magic in mea struct
45 int mds_md_connect(struct obd_device *obd, char *md_name)
47 struct mds_obd *mds = &obd->u.mds;
48 struct lustre_handle conn = {0};
49 int rc, valsize, value;
52 if (IS_ERR(mds->mds_md_obd))
53 RETURN(PTR_ERR(mds->mds_md_obd));
55 if (mds->mds_md_connected)
58 down(&mds->mds_md_sem);
59 if (mds->mds_md_connected) {
64 mds->mds_md_obd = class_name2obd(md_name);
65 if (!mds->mds_md_obd) {
66 CERROR("MDS cannot locate MD(LMV) %s\n",
68 mds->mds_md_obd = ERR_PTR(-ENOTCONN);
69 GOTO(err_last, rc = -ENOTCONN);
72 rc = obd_connect(&conn, mds->mds_md_obd,
73 &obd->obd_uuid, OBD_OPT_MDS_CONNECTION);
75 CERROR("MDS cannot connect to MD(LMV) %s (%d)\n",
77 mds->mds_md_obd = ERR_PTR(rc);
80 mds->mds_md_exp = class_conn2export(&conn);
81 if (mds->mds_md_exp == NULL)
82 CERROR("can't get export!\n");
84 rc = obd_register_observer(mds->mds_md_obd, obd);
86 CERROR("MDS cannot register as observer of MD(LMV) %s, "
87 "rc = %d\n", md_name, rc);
91 /* retrieve size of EA */
92 rc = obd_get_info(mds->mds_md_exp, strlen("mdsize"),
93 "mdsize", &valsize, &value);
97 if (value > mds->mds_max_mdsize)
98 mds->mds_max_mdsize = value;
100 /* find our number in LMV cluster */
101 rc = obd_get_info(mds->mds_md_exp, strlen("mdsnum"),
102 "mdsnum", &valsize, &value);
106 mds->mds_num = value;
108 rc = obd_set_info(mds->mds_md_exp, strlen("inter_mds"),
109 "inter_mds", 0, NULL);
113 mds->mds_md_connected = 1;
114 up(&mds->mds_md_sem);
118 obd_register_observer(mds->mds_md_obd, NULL);
120 obd_disconnect(mds->mds_md_exp, 0);
121 mds->mds_md_exp = NULL;
122 mds->mds_md_obd = ERR_PTR(rc);
124 up(&mds->mds_md_sem);
128 int mds_md_postsetup(struct obd_device *obd)
130 struct mds_obd *mds = &obd->u.mds;
134 if (mds->mds_md_exp) {
135 rc = obd_init_ea_size(mds->mds_md_exp,
137 mds->mds_max_cookiesize);
143 int mds_md_disconnect(struct obd_device *obd, int flags)
145 struct mds_obd *mds = &obd->u.mds;
149 if (!mds->mds_md_connected)
152 down(&mds->mds_md_sem);
153 if (!IS_ERR(mds->mds_md_obd) && mds->mds_md_exp != NULL) {
154 LASSERT(mds->mds_md_connected);
156 obd_register_observer(mds->mds_md_obd, NULL);
158 if (flags & OBD_OPT_FORCE) {
159 struct obd_device *lmv_obd;
160 struct obd_ioctl_data ioc_data = { 0 };
162 lmv_obd = class_exp2obd(mds->mds_md_exp);
167 * making disconnecting lmv stuff do not send anything
168 * to all remote MDSs from LMV. This is needed to
169 * prevent possible hanging with endless recovery, when
170 * MDS sends disconnect to already disconnected
171 * target. Probably this is wrong, but client does the
172 * same in --force mode and I do not see why can't we do
175 lmv_obd->obd_no_recov = 1;
176 obd_iocontrol(IOC_OSC_SET_ACTIVE, mds->mds_md_exp,
177 sizeof(ioc_data), &ioc_data, NULL);
181 * if obd_disconnect() fails (probably because the export was
182 * disconnected by class_disconnect_exports()) then we just need
185 mds->mds_md_connected = 0;
186 rc = obd_disconnect(mds->mds_md_exp, flags);
188 class_export_put(mds->mds_md_exp);
191 mds->mds_md_exp = NULL;
192 mds->mds_md_obd = NULL;
194 up(&mds->mds_md_sem);
198 int mds_md_get_attr(struct obd_device *obd, struct inode *inode,
199 struct mea **mea, int *mea_size)
201 struct mds_obd *mds = &obd->u.mds;
205 if (!mds->mds_md_obd)
207 if (!S_ISDIR(inode->i_mode))
210 /* first calculate mea size */
211 *mea_size = obd_alloc_diskmd(mds->mds_md_exp,
212 (struct lov_mds_md **)mea);
213 if (*mea_size < 0 || *mea == NULL)
214 return *mea_size < 0 ? *mea_size : -EINVAL;
216 rc = mds_get_md(obd, inode, *mea, mea_size, 1);
219 OBD_FREE(*mea, *mea_size);
237 #define DIR_ROUND (DIR_PAD - 1)
238 #define DIR_REC_LEN(name_len) (((name_len) + 16 + DIR_ROUND) & ~DIR_ROUND)
240 /* this struct holds dir entries for particular MDS to be flushed */
242 struct list_head list;
247 struct brw_page brwc;
250 struct dirsplit_control {
251 struct obd_device *obd;
253 struct dentry *dentry;
255 struct dir_cache *cache;
258 static int dc_new_page_to_cache(struct dir_cache * dirc)
262 if (!list_empty(&dirc->list) && dirc->free > sizeof(__u16)) {
263 /* current page became full, mark the end */
264 struct dir_entry *de = dirc->cur;
268 page = alloc_page(GFP_KERNEL);
271 list_add_tail(&page->lru, &dirc->list);
272 dirc->cur = page_address(page);
273 dirc->free = PAGE_SIZE;
277 static int retrieve_generation_numbers(struct dirsplit_control *dc, void *buf)
279 struct mds_obd *mds = &dc->obd->u.mds;
280 struct dir_entry *de;
281 struct dentry *dentry;
284 end = buf + PAGE_SIZE;
285 de = (struct dir_entry *) buf;
286 while ((char *) de < end && de->namelen) {
287 /* lookup an inode */
288 LASSERT(de->namelen <= 255);
289 dentry = ll_lookup_one_len(de->name, dc->dentry,
291 if (IS_ERR(dentry)) {
292 CERROR("can't lookup %*s: %d\n", de->namelen,
293 de->name, (int) PTR_ERR(dentry));
296 if (dentry->d_inode != NULL) {
298 struct lustre_id sid;
300 down(&dentry->d_inode->i_sem);
301 rc = mds_read_inode_sid(dc->obd,
302 dentry->d_inode, &sid);
303 up(&dentry->d_inode->i_sem);
305 CERROR("Can't read inode self id, "
306 "inode %lu, rc %d\n",
307 dentry->d_inode->i_ino, rc);
311 de->fid = id_fid(&sid);
312 de->mds = mds->mds_num;
313 de->ino = dentry->d_inode->i_ino;
314 de->generation = dentry->d_inode->i_generation;
315 } else if (dentry->d_flags & DCACHE_CROSS_REF) {
316 de->fid = dentry->d_fid;
317 de->ino = dentry->d_inum;
318 de->mds = dentry->d_mdsnum;
319 de->generation = dentry->d_generation;
321 CERROR("can't lookup %*s\n", de->namelen, de->name);
327 de = (struct dir_entry *)
328 ((char *) de + DIR_REC_LEN(de->namelen));
333 static int flush_buffer_onto_mds(struct dirsplit_control *dc, int mdsnum)
335 struct mds_obd *mds = &dc->obd->u.mds;
336 struct list_head *cur, *tmp;
337 struct dir_cache *ca;
340 ca = dc->cache + mdsnum;
342 if (ca->free > sizeof(__u16)) {
343 /* current page became full, mark the end */
344 struct dir_entry *de = ca->cur;
348 list_for_each_safe(cur, tmp, &ca->list) {
351 page = list_entry(cur, struct page, lru);
352 LASSERT(page != NULL);
354 retrieve_generation_numbers(dc, page_address(page));
357 ca->brwc.disk_offset = ca->brwc.page_offset = 0;
358 ca->brwc.count = PAGE_SIZE;
360 ca->oa.o_mds = mdsnum;
361 rc = obd_brw(OBD_BRW_WRITE, mds->mds_md_exp, &ca->oa,
362 (struct lov_stripe_md *) dc->mea,
371 static int remove_entries_from_orig_dir(struct dirsplit_control *dc, int mdsnum)
373 struct list_head *cur, *tmp;
374 struct dentry *dentry;
375 struct dir_cache *ca;
376 struct dir_entry *de;
382 ca = dc->cache + mdsnum;
383 list_for_each_safe(cur, tmp, &ca->list) {
384 page = list_entry(cur, struct page, lru);
385 buf = page_address(page);
386 end = buf + PAGE_SIZE;
388 de = (struct dir_entry *) buf;
389 while ((char *) de < end && de->namelen) {
390 /* lookup an inode */
391 LASSERT(de->namelen <= 255);
393 dentry = ll_lookup_one_len(de->name, dc->dentry,
395 if (IS_ERR(dentry)) {
396 CERROR("can't lookup %*s: %d\n", de->namelen,
397 de->name, (int) PTR_ERR(dentry));
400 rc = fsfilt_del_dir_entry(dc->obd, dentry);
403 de = (struct dir_entry *)
404 ((char *) de + DIR_REC_LEN(de->namelen));
410 static int filldir(void * __buf, const char * name, int namlen,
411 loff_t offset, ino_t ino, unsigned int d_type)
413 struct dirsplit_control *dc = __buf;
414 struct mds_obd *mds = &dc->obd->u.mds;
415 struct dir_cache *ca;
416 struct dir_entry *de;
421 if (name[0] == '.' &&
422 (namlen == 1 || (namlen == 2 && name[1] == '.'))) {
423 /* skip special entries */
428 newmds = mea_name2idx(dc->mea, (char *) name, namlen);
430 if (newmds == mds->mds_num) {
431 /* this entry remains on the current MDS, skip moving */
435 OBD_ALLOC(n, namlen + 1);
436 memcpy(n, name, namlen);
437 n[namlen] = (char) 0;
439 OBD_FREE(n, namlen + 1);
441 /* check for space in buffer for new entry */
442 ca = dc->cache + newmds;
443 if (DIR_REC_LEN(namlen) > ca->free) {
444 int err = dc_new_page_to_cache(ca);
448 /* insert found entry into buffer to be flushed later */
449 /* NOTE: we'll fill generations number later, because we
450 * it's stored in inode, thus we need to lookup an entry,
451 * but directory is locked for readdir(), so we delay this */
455 de->namelen = namlen;
456 memcpy(de->name, name, namlen);
457 ca->cur += DIR_REC_LEN(namlen);
458 ca->free -= DIR_REC_LEN(namlen);
464 int scan_and_distribute(struct obd_device *obd, struct dentry *dentry,
467 struct inode *dir = dentry->d_inode;
468 struct dirsplit_control dc;
473 nlen = strlen("__iopen__/") + 10 + 1;
474 OBD_ALLOC(file_name, nlen);
478 i = sprintf(file_name, "__iopen__/0x%lx",
479 dentry->d_inode->i_ino);
481 file = filp_open(file_name, O_RDONLY, 0);
483 CERROR("can't open directory %s: %d\n",
484 file_name, (int) PTR_ERR(file));
485 OBD_FREE(file_name, nlen);
486 RETURN(PTR_ERR(file));
489 memset(&dc, 0, sizeof(dc));
494 OBD_ALLOC(dc.cache, sizeof(struct dir_cache) * mea->mea_count);
495 LASSERT(dc.cache != NULL);
496 for (i = 0; i < mea->mea_count; i++) {
497 INIT_LIST_HEAD(&dc.cache[i].list);
498 dc.cache[i].free = 0;
499 dc.cache[i].cached = 0;
502 err = vfs_readdir(file, filldir, &dc);
507 for (i = 0; i < mea->mea_count; i++) {
508 if (!dc.cache[i].cached)
510 err = flush_buffer_onto_mds(&dc, i);
515 for (i = 0; i < mea->mea_count; i++) {
516 if (!dc.cache[i].cached)
518 err = remove_entries_from_orig_dir(&dc, i);
525 for (i = 0; i < mea->mea_count; i++) {
526 struct list_head *cur, *tmp;
527 if (!dc.cache[i].cached)
529 list_for_each_safe(cur, tmp, &dc.cache[i].list) {
531 page = list_entry(cur, struct page, lru);
532 list_del(&page->lru);
536 OBD_FREE(dc.cache, sizeof(struct dir_cache) * mea->mea_count);
537 OBD_FREE(file_name, nlen);
542 #define MAX_DIR_SIZE (64 * 1024)
543 #define I_NON_SPLITTABLE (256)
545 int mds_splitting_expected(struct obd_device *obd, struct dentry *dentry)
547 struct mds_obd *mds = &obd->u.mds;
548 struct mea *mea = NULL;
552 if (!mds->mds_md_obd)
553 return MDS_NO_SPLITTABLE;
556 if (dentry->d_inode == NULL)
557 return MDS_NO_SPLITTABLE;
559 /* a dir can be splitted only */
560 if (!S_ISDIR(dentry->d_inode->i_mode))
561 return MDS_NO_SPLITTABLE;
563 /* already splittied or slave directory (part of splitted dir) */
564 if (dentry->d_inode->i_flags & I_NON_SPLITTABLE)
565 return MDS_NO_SPLITTABLE;
567 /* don't split root directory */
568 if (dentry->d_inode->i_ino == id_ino(&mds->mds_rootid))
569 return MDS_NO_SPLITTABLE;
571 /* large enough to be splitted? */
572 if (dentry->d_inode->i_size < MAX_DIR_SIZE)
573 return MDS_NO_SPLIT_EXPECTED;
575 mds_md_get_attr(obd, dentry->d_inode, &mea, &size);
577 /* already splitted or slave object: shouldn't be splitted */
578 rc = MDS_NO_SPLITTABLE;
579 /* mark to skip subsequent checks */
580 dentry->d_inode->i_flags |= I_NON_SPLITTABLE;
583 /* may be splitted */
584 rc = MDS_EXPECT_SPLIT;
591 * must not be called on already splitted directories.
593 int mds_try_to_split_dir(struct obd_device *obd, struct dentry *dentry,
594 struct mea **mea, int nstripes, int update_mode)
596 struct inode *dir = dentry->d_inode;
597 struct mds_obd *mds = &obd->u.mds;
598 struct mea *tmea = NULL;
599 struct obdo *oa = NULL;
600 int rc, mea_size = 0;
605 if (update_mode != LCK_EX)
608 /* TODO: optimization possible - we already may have mea here */
609 rc = mds_splitting_expected(obd, dentry);
610 if (rc == MDS_NO_SPLITTABLE)
612 if (rc == MDS_NO_SPLIT_EXPECTED && nstripes == 0)
614 if (nstripes && nstripes == 1)
617 LASSERT(mea == NULL || *mea == NULL);
619 CDEBUG(D_OTHER, "%s: split directory %u/%lu/%lu\n",
620 obd->obd_name, mds->mds_num, dir->i_ino,
621 (unsigned long) dir->i_generation);
625 mea_size = obd_size_diskmd(mds->mds_md_exp, NULL);
627 /* FIXME: Actually we may only want to allocate enough space for
628 * necessary amount of stripes, but on the other hand with this
629 * approach of allocating maximal possible amount of MDS slots,
630 * it would be easier to split the dir over more MDSes */
631 rc = obd_alloc_diskmd(mds->mds_md_exp, (void *)mea);
633 CERROR("obd_alloc_diskmd() failed, error %d.\n", rc);
639 (*mea)->mea_count = nstripes;
641 /* 1) create directory objects on slave MDS'es */
642 /* FIXME: should this be OBD method? */
647 oa->o_generation = dir->i_generation;
649 obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
650 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
651 OBD_MD_FLUID | OBD_MD_FLGID);
653 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
654 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
655 oa->o_mode = dir->i_mode;
658 * until lmv_obd_create() properly rewritten, it is important to have
659 * here oa->o_id = dir->i_ino, as otherwise master object will have
660 * invalid store cookie (zero inode num), what will lead to -ESTALE in
661 * mds_open() or somewhere else.
663 oa->o_id = dir->i_ino;
666 rc = mds_read_inode_sid(obd, dir, &id);
669 CERROR("Can't read inode self id, inode %lu, "
670 "rc %d.\n", dir->i_ino, rc);
673 oa->o_fid = id_fid(&id);
674 oa->o_mds = mds->mds_num;
675 LASSERT(oa->o_fid != 0);
677 CDEBUG(D_OTHER, "%s: create subdirs with mode %o, uid %u, gid %u\n",
678 obd->obd_name, dir->i_mode, dir->i_uid, dir->i_gid);
680 rc = obd_create(mds->mds_md_exp, oa,
681 (struct lov_stripe_md **)mea, NULL);
683 CERROR("Can't create remote inode, rc = %d\n", rc);
687 LASSERT(id_fid(&(*mea)->mea_ids[0]));
688 CDEBUG(D_OTHER, "%d dirobjects created\n", (int)(*mea)->mea_count);
690 /* 2) update dir attribute */
693 handle = fsfilt_start(obd, dir, FSFILT_OP_SETATTR, NULL);
694 if (IS_ERR(handle)) {
696 CERROR("fsfilt_start() failed: %d\n", (int) PTR_ERR(handle));
697 GOTO(err_oa, rc = PTR_ERR(handle));
700 rc = fsfilt_set_md(obd, dir, handle, *mea, mea_size);
703 CERROR("fsfilt_set_md() failed, error %d.\n", rc);
707 rc = fsfilt_commit(obd, mds->mds_sb, dir, handle, 0);
710 CERROR("fsfilt_commit() failed, error %d.\n", rc);
717 /* 3) read through the dir and distribute it over objects */
718 rc = scan_and_distribute(obd, dentry, *mea);
720 obd_free_diskmd(mds->mds_md_exp, (struct lov_mds_md **)mea);
722 CERROR("scan_and_distribute() failed, error %d.\n", rc);
733 static int filter_start_page_write(struct inode *inode,
734 struct niobuf_local *lnb)
736 struct page *page = alloc_pages(GFP_HIGHUSER, 0);
738 CERROR("no memory for a temp page\n");
739 return lnb->rc = -ENOMEM;
741 POISON_PAGE(page, 0xf1);
742 page->index = lnb->offset >> PAGE_SHIFT;
747 struct dentry *filter_id2dentry(struct obd_device *obd,
748 struct dentry *dir_dentry,
749 obd_gr group, obd_id id);
751 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
752 int objcount, struct obd_ioobj *obj,
753 int niocount, struct niobuf_remote *nb,
754 struct niobuf_local *res,
755 struct obd_trans_info *oti)
757 struct niobuf_remote *rnb;
758 struct niobuf_local *lnb = NULL;
759 int rc = 0, i, tot_bytes = 0;
760 unsigned long now = jiffies;
761 struct dentry *dentry;
765 LASSERT(objcount == 1);
766 LASSERT(obj->ioo_bufcnt > 0);
768 memset(res, 0, niocount * sizeof(*res));
772 id_ino(&id) = obj->ioo_id;
773 id_gen(&id) = obj->ioo_gr;
775 dentry = mds_id2dentry(exp->exp_obd, &id, NULL);
776 if (IS_ERR(dentry)) {
777 CERROR("can't get dentry for "LPU64"/%u: %d\n",
778 id.li_stc.u.e3s.l3s_ino, id.li_stc.u.e3s.l3s_gen,
779 (int)PTR_ERR(dentry));
780 GOTO(cleanup, rc = (int) PTR_ERR(dentry));
783 if (dentry->d_inode == NULL) {
784 CERROR("trying to BRW to non-existent file "LPU64"\n",
787 GOTO(cleanup, rc = -ENOENT);
790 if (time_after(jiffies, now + 15 * HZ))
791 CERROR("slow preprw_write setup %lus\n", (jiffies - now) / HZ);
793 CDEBUG(D_INFO, "preprw_write setup: %lu jiffies\n",
796 for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
798 lnb->dentry = dentry;
799 lnb->offset = rnb->offset;
801 lnb->flags = rnb->flags;
803 rc = filter_start_page_write(dentry->d_inode, lnb);
805 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR, "page err %u@"
806 LPU64" %u/%u %p: rc %d\n", lnb->len, lnb->offset,
807 i, obj->ioo_bufcnt, dentry, rc);
809 __free_pages(lnb->page, 0);
813 tot_bytes += lnb->len;
816 if (time_after(jiffies, now + 15 * HZ))
817 CERROR("slow start_page_write %lus\n", (jiffies - now) / HZ);
819 CDEBUG(D_INFO, "start_page_write: %lu jiffies\n",
827 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
828 int objcount, struct obd_ioobj *obj, int niocount,
829 struct niobuf_local *res, struct obd_trans_info *oti,
832 struct obd_device *obd = exp->exp_obd;
833 struct niobuf_local *lnb;
834 struct inode *inode = NULL;
835 int rc = 0, i, cleanup_phase = 0, err, entries = 0;
838 LASSERT(objcount == 1);
839 LASSERT(current->journal_info == NULL);
842 inode = res->dentry->d_inode;
844 for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
846 struct dir_entry *de;
848 buf = kmap(lnb->page);
849 LASSERT(buf != NULL);
850 end = buf + lnb->len;
851 de = (struct dir_entry *) buf;
852 while ((char *) de < end && de->namelen) {
853 err = fsfilt_add_dir_entry(obd, res->dentry, de->name,
854 de->namelen, de->ino,
855 de->generation, de->mds,
858 CERROR("can't add dir entry %*s->%u/%u/%u"
859 " to %lu/%u: %d\n", de->namelen,
860 de->name, de->mds, (unsigned)de->ino,
861 (unsigned) de->generation,
862 res->dentry->d_inode->i_ino,
863 res->dentry->d_inode->i_generation,
869 de = (struct dir_entry *)
870 ((char *) de + DIR_REC_LEN(de->namelen));
876 for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++)
877 __free_page(lnb->page);
883 int mds_choose_mdsnum(struct obd_device *obd, const char *name, int len, int flags)
886 struct mds_obd *mds = &obd->u.mds;
887 int i = mds->mds_num;
889 if (flags & REC_REINT_CREATE) {
891 } else if (mds->mds_md_exp) {
892 lmv = &mds->mds_md_exp->exp_obd->u.lmv;
893 i = raw_name2idx(MEA_MAGIC_LAST_CHAR, lmv->desc.ld_tgt_count, name, len);
898 int mds_lock_slave_objs(struct obd_device *obd, struct dentry *dentry,
899 struct lustre_handle **rlockh)
901 struct mds_obd *mds = &obd->u.mds;
902 struct mdc_op_data *op_data;
903 struct lookup_intent it;
904 struct mea *mea = NULL;
909 LASSERT(rlockh != NULL);
910 LASSERT(dentry != NULL);
911 LASSERT(dentry->d_inode != NULL);
914 if (!mds->mds_md_obd)
917 /* a dir can be splitted only */
918 if (!S_ISDIR(dentry->d_inode->i_mode))
921 rc = mds_md_get_attr(obd, dentry->d_inode,
929 if (mea->mea_count == 0)
930 /* this is slave object */
931 GOTO(cleanup, rc = 0);
933 CDEBUG(D_OTHER, "%s: lock slaves for %lu/%lu\n",
934 obd->obd_name, (unsigned long)dentry->d_inode->i_ino,
935 (unsigned long)dentry->d_inode->i_generation);
937 handle_size = sizeof(struct lustre_handle) *
940 OBD_ALLOC(*rlockh, handle_size);
942 GOTO(cleanup, rc = -ENOMEM);
944 memset(*rlockh, 0, handle_size);
945 OBD_ALLOC(op_data, sizeof(*op_data));
946 if (op_data == NULL) {
947 OBD_FREE(*rlockh, handle_size);
950 memset(op_data, 0, sizeof(*op_data));
953 it.it_op = IT_UNLINK;
955 rc = md_enqueue(mds->mds_md_exp, LDLM_IBITS, &it, LCK_EX,
956 op_data, *rlockh, NULL, 0, ldlm_completion_ast,
957 mds_blocking_ast, NULL);
958 OBD_FREE(op_data, sizeof(*op_data));
961 OBD_FREE(mea, mea_size);
965 void mds_unlock_slave_objs(struct obd_device *obd, struct dentry *dentry,
966 struct lustre_handle *lockh)
968 struct mds_obd *mds = &obd->u.mds;
969 struct mea *mea = NULL;
978 LASSERT(mds->mds_md_obd != NULL);
979 LASSERT(S_ISDIR(dentry->d_inode->i_mode));
981 rc = mds_md_get_attr(obd, dentry->d_inode, &mea, &mea_size);
983 CERROR("locks are leaked\n");
987 LASSERT(mea_size != 0);
988 LASSERT(mea != NULL);
989 LASSERT(mea->mea_count != 0);
991 CDEBUG(D_OTHER, "%s: unlock slaves for %lu/%lu\n", obd->obd_name,
992 (unsigned long) dentry->d_inode->i_ino,
993 (unsigned long) dentry->d_inode->i_generation);
995 for (i = 0; i < mea->mea_count; i++) {
996 if (lockh[i].cookie != 0)
997 ldlm_lock_decref(lockh + i, LCK_EX);
1000 OBD_FREE(lockh, sizeof(struct lustre_handle) * mea->mea_count);
1001 OBD_FREE(mea, mea_size);
1005 int mds_unlink_slave_objs(struct obd_device *obd, struct dentry *dentry)
1007 struct mds_obd *mds = &obd->u.mds;
1008 struct ptlrpc_request *req = NULL;
1009 struct mdc_op_data *op_data;
1010 struct mea *mea = NULL;
1014 /* clustered MD ? */
1015 if (!mds->mds_md_obd)
1018 /* a dir can be splitted only */
1019 if (!S_ISDIR(dentry->d_inode->i_mode))
1022 rc = mds_md_get_attr(obd, dentry->d_inode, &mea, &mea_size);
1029 if (mea->mea_count == 0)
1030 GOTO(cleanup, rc = 0);
1032 CDEBUG(D_OTHER, "%s: unlink slaves for %lu/%lu\n", obd->obd_name,
1033 (unsigned long)dentry->d_inode->i_ino,
1034 (unsigned long)dentry->d_inode->i_generation);
1036 OBD_ALLOC(op_data, sizeof(*op_data));
1037 if (op_data == NULL)
1040 memset(op_data, 0, sizeof(*op_data));
1041 op_data->mea1 = mea;
1042 rc = md_unlink(mds->mds_md_exp, op_data, &req);
1043 OBD_FREE(op_data, sizeof(*op_data));
1044 LASSERT(req == NULL);
1047 OBD_FREE(mea, mea_size);
1051 struct ide_tracking {
1056 int mds_ide_filldir(void *__buf, const char *name, int namelen,
1057 loff_t offset, ino_t ino, unsigned int d_type)
1059 struct ide_tracking *it = __buf;
1065 if (it->entries > 2)
1069 if (name[0] == '.' && namelen == 1)
1071 if (name[0] == '.' && name[1] == '.' && namelen == 2)
1078 /* checks if passed dentry points to empty dir. */
1079 int mds_is_dir_empty(struct obd_device *obd, struct dentry *dentry)
1081 struct ide_tracking it;
1089 nlen = strlen("__iopen__/") + 10 + 1;
1090 OBD_ALLOC(file_name, nlen);
1094 LASSERT(dentry->d_inode != NULL);
1095 i = sprintf(file_name, "__iopen__/0x%lx",
1096 dentry->d_inode->i_ino);
1098 file = filp_open(file_name, O_RDONLY, 0);
1100 CERROR("can't open directory %s: %d\n",
1101 file_name, (int) PTR_ERR(file));
1102 GOTO(cleanup, rc = PTR_ERR(file));
1105 rc = vfs_readdir(file, mds_ide_filldir, &it);
1106 filp_close(file, 0);
1108 if (it.empty && rc == 0)
1114 OBD_FREE(file_name, nlen);
1118 int mds_lock_and_check_slave(int offset, struct ptlrpc_request *req,
1119 struct lustre_handle *lockh)
1121 struct obd_device *obd = req->rq_export->exp_obd;
1122 struct dentry *dentry = NULL;
1123 struct lvfs_run_ctxt saved;
1124 int cleanup_phase = 0;
1125 struct mds_req_sec_desc *rsd;
1126 struct mds_body *body;
1127 struct lvfs_ucred uc;
1128 int rc, update_mode;
1131 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1133 CERROR("Can't unpack security desc\n");
1134 GOTO(cleanup, rc = -EFAULT);
1136 mds_squash_root(&obd->u.mds, rsd, &req->rq_peer.peer_id.nid);
1138 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1139 lustre_swab_mds_body);
1141 CERROR("Can't swab mds_body\n");
1142 GOTO(cleanup, rc = -EFAULT);
1144 CDEBUG(D_OTHER, "%s: check slave "DLID4"\n", obd->obd_name,
1147 dentry = mds_id2locked_dentry(obd, &body->id1, NULL, LCK_EX,
1148 lockh, &update_mode, NULL, 0,
1149 MDS_INODELOCK_UPDATE);
1150 if (IS_ERR(dentry)) {
1151 CERROR("can't find inode: %d\n", (int) PTR_ERR(dentry));
1152 GOTO(cleanup, rc = PTR_ERR(dentry));
1156 LASSERT(S_ISDIR(dentry->d_inode->i_mode));
1158 rc = mds_init_ucred(&uc, rsd);
1160 CERROR("can't init ucred\n");
1164 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1167 if (!mds_is_dir_empty(obd, dentry))
1172 switch(cleanup_phase) {
1175 ldlm_lock_decref(lockh, LCK_EX);
1177 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1178 mds_exit_ucred(&uc);
1185 int mds_convert_mea_ea(struct obd_device *obd, struct inode *inode,
1186 struct lov_mds_md *lmm, int lmmsize)
1188 int i, rc, err, size;
1189 struct mea_old *old;
1195 mea = (struct mea *) lmm;
1196 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
1197 mea->mea_magic == MEA_MAGIC_ALL_CHARS)
1200 old = (struct mea_old *) lmm;
1202 rc = sizeof(struct lustre_id) * old->mea_count +
1203 sizeof(struct mea_old);
1205 if (old->mea_count > 256 || old->mea_master > 256 || lmmsize < rc
1206 || old->mea_master > old->mea_count) {
1207 CWARN("unknown MEA format, dont convert it\n");
1208 CWARN(" count %u, master %u, size %u\n",
1209 old->mea_count, old->mea_master, rc);
1213 CWARN("converting MEA EA on %lu/%u from V0 to V1 (%u/%u)\n",
1214 inode->i_ino, inode->i_generation, old->mea_count,
1217 size = sizeof(struct lustre_id) * old->mea_count +
1220 OBD_ALLOC(new, size);
1224 new->mea_magic = MEA_MAGIC_LAST_CHAR;
1225 new->mea_count = old->mea_count;
1226 new->mea_master = old->mea_master;
1227 for (i = 0; i < new->mea_count; i++)
1228 new->mea_ids[i] = old->mea_ids[i];
1230 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1231 if (IS_ERR(handle)) {
1232 rc = PTR_ERR(handle);
1233 GOTO(conv_free, rc);
1236 rc = fsfilt_set_md(obd, inode, handle, (struct lov_mds_md *) new, size);
1239 memcpy(lmm, new, size);
1241 err = fsfilt_commit(obd, obd->u.mds.mds_sb, inode, handle, 0);
1243 rc = err ? err : size;
1246 OBD_FREE(new, size);