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 <asm/div64.h>
46 #include "mds_internal.h"
50 * - magic in mea struct
52 int mds_md_connect(struct obd_device *obd, char *md_name)
54 struct mds_obd *mds = &obd->u.mds;
55 struct lustre_handle conn = {0};
56 int rc, valsize, value;
59 if (IS_ERR(mds->mds_md_obd))
60 RETURN(PTR_ERR(mds->mds_md_obd));
62 if (mds->mds_md_connected)
65 down(&mds->mds_md_sem);
66 if (mds->mds_md_connected) {
71 mds->mds_md_obd = class_name2obd(md_name);
72 if (!mds->mds_md_obd) {
73 CERROR("MDS cannot locate MD(LMV) %s\n",
75 mds->mds_md_obd = ERR_PTR(-ENOTCONN);
76 GOTO(err_last, rc = -ENOTCONN);
79 rc = obd_connect(&conn, mds->mds_md_obd, &obd->obd_uuid, NULL,
80 OBD_OPT_MDS_CONNECTION);
82 CERROR("MDS cannot connect to MD(LMV) %s (%d)\n",
84 mds->mds_md_obd = ERR_PTR(rc);
87 mds->mds_md_exp = class_conn2export(&conn);
88 if (mds->mds_md_exp == NULL)
89 CERROR("can't get export!\n");
91 rc = obd_register_observer(mds->mds_md_obd, obd);
93 CERROR("MDS cannot register as observer of MD(LMV) %s, "
94 "rc = %d\n", md_name, rc);
98 /* retrieve size of EA */
99 rc = obd_get_info(mds->mds_md_exp, strlen("mdsize"),
100 "mdsize", &valsize, &value);
104 if (value > mds->mds_max_mdsize)
105 mds->mds_max_mdsize = value;
107 /* find our number in LMV cluster */
108 rc = obd_get_info(mds->mds_md_exp, strlen("mdsnum"),
109 "mdsnum", &valsize, &value);
113 mds->mds_num = value;
115 rc = obd_set_info(mds->mds_md_exp, strlen("inter_mds"),
116 "inter_mds", 0, NULL);
120 if (mds->mds_mds_sec) {
121 rc = obd_set_info(mds->mds_md_exp, strlen("sec"), "sec",
122 strlen(mds->mds_mds_sec), mds->mds_mds_sec);
127 mds->mds_md_connected = 1;
128 up(&mds->mds_md_sem);
132 obd_register_observer(mds->mds_md_obd, NULL);
134 obd_disconnect(mds->mds_md_exp, 0);
135 mds->mds_md_exp = NULL;
136 mds->mds_md_obd = ERR_PTR(rc);
138 up(&mds->mds_md_sem);
142 int mds_md_postsetup(struct obd_device *obd)
144 struct mds_obd *mds = &obd->u.mds;
148 if (mds->mds_md_exp) {
149 rc = obd_init_ea_size(mds->mds_md_exp,
151 mds->mds_max_cookiesize);
157 int mds_md_disconnect(struct obd_device *obd, int flags)
159 struct mds_obd *mds = &obd->u.mds;
163 if (!mds->mds_md_connected)
166 down(&mds->mds_md_sem);
167 if (!IS_ERR(mds->mds_md_obd) && mds->mds_md_exp != NULL) {
168 LASSERT(mds->mds_md_connected);
170 obd_register_observer(mds->mds_md_obd, NULL);
172 if (flags & OBD_OPT_FORCE) {
173 struct obd_device *lmv_obd;
174 struct obd_ioctl_data ioc_data = { 0 };
176 lmv_obd = class_exp2obd(mds->mds_md_exp);
181 * making disconnecting lmv stuff do not send anything
182 * to all remote MDSs from LMV. This is needed to
183 * prevent possible hanging with endless recovery, when
184 * MDS sends disconnect to already disconnected
185 * target. Probably this is wrong, but client does the
186 * same in --force mode and I do not see why can't we do
189 lmv_obd->obd_no_recov = 1;
190 obd_iocontrol(IOC_OSC_SET_ACTIVE, mds->mds_md_exp,
191 sizeof(ioc_data), &ioc_data, NULL);
195 * if obd_disconnect() fails (probably because the export was
196 * disconnected by class_disconnect_exports()) then we just need
199 mds->mds_md_connected = 0;
200 rc = obd_disconnect(mds->mds_md_exp, flags);
202 class_export_put(mds->mds_md_exp);
205 mds->mds_md_exp = NULL;
206 mds->mds_md_obd = NULL;
208 up(&mds->mds_md_sem);
212 int mds_md_reconnect(struct obd_device *obd)
214 struct mds_obd *mds = &obd->u.mds;
215 struct obd_statfs osfs;
219 /* We don't know state of connections to another MDSes
220 * before the failure. If MDS we were connected to before
221 * the failure gets failed, then it will wait for us to
222 * reconnect and will timed recovery out. bug 4920 */
223 if (mds->mds_md_connected == 0)
225 if (mds->mds_md_obd == NULL)
228 err = obd_statfs(mds->mds_md_obd, &osfs, jiffies - HZ);
230 CERROR("can't reconnect to MDSes after recovery: %d\n", err);
234 int mds_md_get_attr(struct obd_device *obd, struct inode *inode,
235 struct mea **mea, int *mea_size)
237 struct mds_obd *mds = &obd->u.mds;
241 if (!mds->mds_md_obd)
243 if (!S_ISDIR(inode->i_mode))
246 /* first calculate mea size */
247 *mea_size = obd_alloc_diskmd(mds->mds_md_exp,
248 (struct lov_mds_md **)mea);
249 if (*mea_size < 0 || *mea == NULL)
250 return *mea_size < 0 ? *mea_size : -EINVAL;
252 rc = mds_get_md(obd, inode, *mea, mea_size, 1, 1);
255 OBD_FREE(*mea, *mea_size);
273 #define DIR_ROUND (DIR_PAD - 1)
274 #define DIR_REC_LEN(name_len) (((name_len) + 16 + DIR_ROUND) & ~DIR_ROUND)
276 /* this struct holds dir entries for particular MDS to be flushed */
278 struct list_head list;
283 struct brw_page brwc;
286 struct dirsplit_control {
287 struct obd_device *obd;
289 struct dentry *dentry;
291 struct dir_cache *cache;
294 static int dc_new_page_to_cache(struct dir_cache * dirc)
298 if (!list_empty(&dirc->list) && dirc->free > sizeof(__u16)) {
299 /* current page became full, mark the end */
300 struct dir_entry *de = dirc->cur;
304 page = alloc_page(GFP_KERNEL);
307 list_add_tail(&page->lru, &dirc->list);
308 dirc->cur = page_address(page);
309 dirc->free = PAGE_SIZE;
313 static int retrieve_generation_numbers(struct dirsplit_control *dc, void *buf)
315 struct mds_obd *mds = &dc->obd->u.mds;
316 struct dir_entry *de;
317 struct dentry *dentry;
320 end = buf + PAGE_SIZE;
321 de = (struct dir_entry *) buf;
322 while ((char *) de < end && de->namelen) {
323 /* lookup an inode */
324 LASSERT(de->namelen <= 255);
325 dentry = ll_lookup_one_len(de->name, dc->dentry,
327 if (IS_ERR(dentry)) {
328 CERROR("can't lookup %*s: %d\n", de->namelen,
329 de->name, (int) PTR_ERR(dentry));
332 if (dentry->d_inode != NULL) {
334 struct lustre_id sid;
336 down(&dentry->d_inode->i_sem);
337 rc = mds_read_inode_sid(dc->obd,
338 dentry->d_inode, &sid);
339 up(&dentry->d_inode->i_sem);
341 CERROR("Can't read inode self id, "
342 "inode %lu, rc %d\n",
343 dentry->d_inode->i_ino, rc);
347 de->fid = id_fid(&sid);
348 de->mds = mds->mds_num;
349 de->ino = dentry->d_inode->i_ino;
350 de->generation = dentry->d_inode->i_generation;
351 } else if (dentry->d_flags & DCACHE_CROSS_REF) {
352 de->fid = dentry->d_fid;
353 de->ino = dentry->d_inum;
354 de->mds = dentry->d_mdsnum;
355 de->generation = dentry->d_generation;
357 CERROR("can't lookup %*s\n", de->namelen, de->name);
363 de = (struct dir_entry *)
364 ((char *) de + DIR_REC_LEN(de->namelen));
369 static int flush_buffer_onto_mds(struct dirsplit_control *dc, int mdsnum)
371 struct mds_obd *mds = &dc->obd->u.mds;
372 struct list_head *cur, *tmp;
373 struct dir_cache *ca;
376 ca = dc->cache + mdsnum;
378 if (ca->free > sizeof(__u16)) {
379 /* current page became full, mark the end */
380 struct dir_entry *de = ca->cur;
384 list_for_each_safe(cur, tmp, &ca->list) {
387 page = list_entry(cur, struct page, lru);
388 LASSERT(page != NULL);
390 retrieve_generation_numbers(dc, page_address(page));
393 ca->brwc.disk_offset = ca->brwc.page_offset = 0;
394 ca->brwc.count = PAGE_SIZE;
396 ca->oa.o_mds = mdsnum;
397 rc = obd_brw(OBD_BRW_WRITE, mds->mds_md_exp, &ca->oa,
398 (struct lov_stripe_md *) dc->mea,
407 static int remove_entries_from_orig_dir(struct dirsplit_control *dc, int mdsnum)
409 struct list_head *cur, *tmp;
410 struct dentry *dentry;
411 struct dir_cache *ca;
412 struct dir_entry *de;
418 ca = dc->cache + mdsnum;
419 list_for_each_safe(cur, tmp, &ca->list) {
420 page = list_entry(cur, struct page, lru);
421 buf = page_address(page);
422 end = buf + PAGE_SIZE;
424 de = (struct dir_entry *) buf;
425 while ((char *) de < end && de->namelen) {
426 /* lookup an inode */
427 LASSERT(de->namelen <= 255);
429 dentry = ll_lookup_one_len(de->name, dc->dentry,
431 if (IS_ERR(dentry)) {
432 CERROR("can't lookup %*s: %d\n", de->namelen,
433 de->name, (int) PTR_ERR(dentry));
436 rc = fsfilt_del_dir_entry(dc->obd, dentry);
439 de = (struct dir_entry *)
440 ((char *) de + DIR_REC_LEN(de->namelen));
446 static int filldir(void * __buf, const char * name, int namlen,
447 loff_t offset, ino_t ino, unsigned int d_type)
449 struct dirsplit_control *dc = __buf;
450 struct mds_obd *mds = &dc->obd->u.mds;
451 struct dir_cache *ca;
452 struct dir_entry *de;
457 if (name[0] == '.' &&
458 (namlen == 1 || (namlen == 2 && name[1] == '.'))) {
459 /* skip special entries */
464 newmds = mea_name2idx(dc->mea, (char *) name, namlen);
466 if (newmds == mds->mds_num) {
467 /* this entry remains on the current MDS, skip moving */
471 OBD_ALLOC(n, namlen + 1);
472 memcpy(n, name, namlen);
473 n[namlen] = (char) 0;
475 OBD_FREE(n, namlen + 1);
477 /* check for space in buffer for new entry */
478 ca = dc->cache + newmds;
479 if (DIR_REC_LEN(namlen) > ca->free) {
480 int err = dc_new_page_to_cache(ca);
484 /* insert found entry into buffer to be flushed later */
485 /* NOTE: we'll fill generations number later, because we
486 * it's stored in inode, thus we need to lookup an entry,
487 * but directory is locked for readdir(), so we delay this */
491 de->namelen = namlen;
492 memcpy(de->name, name, namlen);
493 ca->cur += DIR_REC_LEN(namlen);
494 ca->free -= DIR_REC_LEN(namlen);
500 int scan_and_distribute(struct obd_device *obd, struct dentry *dentry,
503 struct inode *dir = dentry->d_inode;
504 struct dirsplit_control dc;
509 nlen = strlen("__iopen__/") + 10 + 1;
510 OBD_ALLOC(file_name, nlen);
514 i = sprintf(file_name, "__iopen__/0x%lx",
515 dentry->d_inode->i_ino);
517 file = filp_open(file_name, O_RDONLY, 0);
519 CERROR("can't open directory %s: %d\n",
520 file_name, (int) PTR_ERR(file));
521 OBD_FREE(file_name, nlen);
522 RETURN(PTR_ERR(file));
525 memset(&dc, 0, sizeof(dc));
530 OBD_ALLOC(dc.cache, sizeof(struct dir_cache) * mea->mea_count);
531 LASSERT(dc.cache != NULL);
532 for (i = 0; i < mea->mea_count; i++) {
533 INIT_LIST_HEAD(&dc.cache[i].list);
534 dc.cache[i].free = 0;
535 dc.cache[i].cached = 0;
538 err = vfs_readdir(file, filldir, &dc);
543 for (i = 0; i < mea->mea_count; i++) {
544 if (!dc.cache[i].cached)
546 err = flush_buffer_onto_mds(&dc, i);
551 for (i = 0; i < mea->mea_count; i++) {
552 if (!dc.cache[i].cached)
554 err = remove_entries_from_orig_dir(&dc, i);
561 for (i = 0; i < mea->mea_count; i++) {
562 struct list_head *cur, *tmp;
563 if (!dc.cache[i].cached)
565 list_for_each_safe(cur, tmp, &dc.cache[i].list) {
567 page = list_entry(cur, struct page, lru);
568 list_del(&page->lru);
572 OBD_FREE(dc.cache, sizeof(struct dir_cache) * mea->mea_count);
573 OBD_FREE(file_name, nlen);
578 #define MAX_DIR_SIZE (64 * 1024)
579 #define I_NON_SPLITTABLE (256)
581 int mds_splitting_expected(struct obd_device *obd, struct dentry *dentry)
583 struct mds_obd *mds = &obd->u.mds;
584 struct mea *mea = NULL;
588 if (!mds->mds_md_obd)
589 return MDS_NO_SPLITTABLE;
592 if (dentry->d_inode == NULL)
593 return MDS_NO_SPLITTABLE;
595 /* a dir can be splitted only */
596 if (!S_ISDIR(dentry->d_inode->i_mode))
597 return MDS_NO_SPLITTABLE;
599 /* already splittied or slave directory (part of splitted dir) */
600 if (dentry->d_inode->i_flags & I_NON_SPLITTABLE)
601 return MDS_NO_SPLITTABLE;
603 /* don't split root directory */
604 if (dentry->d_inode->i_ino == id_ino(&mds->mds_rootid))
605 return MDS_NO_SPLITTABLE;
607 /* large enough to be splitted? */
608 if (dentry->d_inode->i_size < MAX_DIR_SIZE)
609 return MDS_NO_SPLIT_EXPECTED;
611 mds_md_get_attr(obd, dentry->d_inode, &mea, &size);
613 /* already splitted or slave object: shouldn't be splitted */
614 rc = MDS_NO_SPLITTABLE;
615 /* mark to skip subsequent checks */
616 dentry->d_inode->i_flags |= I_NON_SPLITTABLE;
619 /* may be splitted */
620 rc = MDS_EXPECT_SPLIT;
627 * must not be called on already splitted directories.
629 int mds_try_to_split_dir(struct obd_device *obd, struct dentry *dentry,
630 struct mea **mea, int nstripes, int update_mode)
632 struct inode *dir = dentry->d_inode;
633 struct mds_obd *mds = &obd->u.mds;
634 struct mea *tmea = NULL;
635 struct obdo *oa = NULL;
636 int rc, mea_size = 0;
641 if (update_mode != LCK_EX)
644 /* TODO: optimization possible - we already may have mea here */
645 rc = mds_splitting_expected(obd, dentry);
646 if (rc == MDS_NO_SPLITTABLE)
648 if (rc == MDS_NO_SPLIT_EXPECTED && nstripes == 0)
650 if (nstripes && nstripes == 1)
653 LASSERT(mea == NULL || *mea == NULL);
655 CDEBUG(D_OTHER, "%s: split directory %u/%lu/%lu\n",
656 obd->obd_name, mds->mds_num, dir->i_ino,
657 (unsigned long) dir->i_generation);
661 mea_size = obd_size_diskmd(mds->mds_md_exp, NULL);
663 /* FIXME: Actually we may only want to allocate enough space for
664 * necessary amount of stripes, but on the other hand with this
665 * approach of allocating maximal possible amount of MDS slots,
666 * it would be easier to split the dir over more MDSes */
667 rc = obd_alloc_diskmd(mds->mds_md_exp, (void *)mea);
669 CERROR("obd_alloc_diskmd() failed, error %d.\n", rc);
675 (*mea)->mea_count = nstripes;
677 /* 1) create directory objects on slave MDS'es */
678 /* FIXME: should this be OBD method? */
683 oa->o_generation = dir->i_generation;
685 obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
686 OBD_MD_FLMTIME | OBD_MD_FLCTIME |
687 OBD_MD_FLUID | OBD_MD_FLGID);
689 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
690 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
691 oa->o_mode = dir->i_mode;
694 * until lmv_obd_create() properly rewritten, it is important to have
695 * here oa->o_id = dir->i_ino, as otherwise master object will have
696 * invalid store cookie (zero inode num), what will lead to -ESTALE in
697 * mds_open() or somewhere else.
699 oa->o_id = dir->i_ino;
702 rc = mds_read_inode_sid(obd, dir, &id);
705 CERROR("Can't read inode self id, inode %lu, "
706 "rc %d.\n", dir->i_ino, rc);
709 oa->o_fid = id_fid(&id);
710 oa->o_mds = mds->mds_num;
711 LASSERT(oa->o_fid != 0);
713 CDEBUG(D_OTHER, "%s: create subdirs with mode %o, uid %u, gid %u\n",
714 obd->obd_name, dir->i_mode, dir->i_uid, dir->i_gid);
716 rc = obd_create(mds->mds_md_exp, oa, NULL, 0,
717 (struct lov_stripe_md **)mea, NULL);
719 CERROR("Can't create remote inode, rc = %d\n", rc);
723 LASSERT(id_fid(&(*mea)->mea_ids[0]));
724 CDEBUG(D_OTHER, "%d dirobjects created\n", (int)(*mea)->mea_count);
726 /* 2) update dir attribute */
729 handle = fsfilt_start(obd, dir, FSFILT_OP_SETATTR, NULL);
730 if (IS_ERR(handle)) {
732 CERROR("fsfilt_start() failed: %d\n", (int) PTR_ERR(handle));
733 GOTO(err_oa, rc = PTR_ERR(handle));
736 rc = fsfilt_set_md(obd, dir, handle, *mea, mea_size, EA_MEA);
739 CERROR("fsfilt_set_md() failed, error %d.\n", rc);
743 rc = fsfilt_commit(obd, mds->mds_sb, dir, handle, 0);
746 CERROR("fsfilt_commit() failed, error %d.\n", rc);
753 /* 3) read through the dir and distribute it over objects */
754 rc = scan_and_distribute(obd, dentry, *mea);
756 obd_free_diskmd(mds->mds_md_exp, (struct lov_mds_md **)mea);
758 CERROR("scan_and_distribute() failed, error %d.\n", rc);
769 static int filter_start_page_write(struct inode *inode,
770 struct niobuf_local *lnb)
772 struct page *page = alloc_pages(GFP_HIGHUSER, 0);
774 CERROR("no memory for a temp page\n");
775 return lnb->rc = -ENOMEM;
777 POISON_PAGE(page, 0xf1);
778 page->index = lnb->offset >> PAGE_SHIFT;
783 struct dentry *filter_id2dentry(struct obd_device *obd,
784 struct dentry *dir_dentry,
785 obd_gr group, obd_id id);
787 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
788 int objcount, struct obd_ioobj *obj,
789 int niocount, struct niobuf_remote *nb,
790 struct niobuf_local *res,
791 struct obd_trans_info *oti)
793 struct niobuf_remote *rnb;
794 struct niobuf_local *lnb = NULL;
795 int rc = 0, i, tot_bytes = 0;
796 unsigned long now = jiffies;
797 struct dentry *dentry;
801 LASSERT(objcount == 1);
802 LASSERT(obj->ioo_bufcnt > 0);
804 memset(res, 0, niocount * sizeof(*res));
808 id_ino(&id) = obj->ioo_id;
809 id_gen(&id) = obj->ioo_gr;
811 dentry = mds_id2dentry(exp->exp_obd, &id, NULL);
812 if (IS_ERR(dentry)) {
813 CERROR("can't get dentry for "LPU64"/%u: %d\n",
814 id.li_stc.u.e3s.l3s_ino, id.li_stc.u.e3s.l3s_gen,
815 (int)PTR_ERR(dentry));
816 GOTO(cleanup, rc = (int) PTR_ERR(dentry));
819 if (dentry->d_inode == NULL) {
820 CERROR("trying to BRW to non-existent file "LPU64"\n",
823 GOTO(cleanup, rc = -ENOENT);
826 if (time_after(jiffies, now + 15 * HZ))
827 CERROR("slow preprw_write setup %lus\n", (jiffies - now) / HZ);
829 CDEBUG(D_INFO, "preprw_write setup: %lu jiffies\n",
832 for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt;
834 lnb->dentry = dentry;
835 lnb->offset = rnb->offset;
837 lnb->flags = rnb->flags;
839 rc = filter_start_page_write(dentry->d_inode, lnb);
841 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR, "page err %u@"
842 LPU64" %u/%u %p: rc %d\n", lnb->len, lnb->offset,
843 i, obj->ioo_bufcnt, dentry, rc);
845 __free_pages(lnb->page, 0);
849 tot_bytes += lnb->len;
852 if (time_after(jiffies, now + 15 * HZ))
853 CERROR("slow start_page_write %lus\n", (jiffies - now) / HZ);
855 CDEBUG(D_INFO, "start_page_write: %lu jiffies\n",
863 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
864 int objcount, struct obd_ioobj *obj, int niocount,
865 struct niobuf_local *res, struct obd_trans_info *oti,
868 struct obd_device *obd = exp->exp_obd;
869 struct niobuf_local *lnb;
870 struct inode *inode = NULL;
871 int rc = 0, i, cleanup_phase = 0, err, entries = 0;
874 LASSERT(objcount == 1);
875 LASSERT(current->journal_info == NULL);
878 inode = res->dentry->d_inode;
880 for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++) {
882 struct dir_entry *de;
884 buf = kmap(lnb->page);
885 LASSERT(buf != NULL);
886 end = buf + lnb->len;
887 de = (struct dir_entry *) buf;
888 while ((char *) de < end && de->namelen) {
889 err = fsfilt_add_dir_entry(obd, res->dentry, de->name,
890 de->namelen, de->ino,
891 de->generation, de->mds,
894 CERROR("can't add dir entry %*s->%u/%u/%u"
895 " to %lu/%u: %d\n", de->namelen,
896 de->name, de->mds, (unsigned)de->ino,
897 (unsigned) de->generation,
898 res->dentry->d_inode->i_ino,
899 res->dentry->d_inode->i_generation,
905 de = (struct dir_entry *)
906 ((char *) de + DIR_REC_LEN(de->namelen));
912 for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++)
913 __free_page(lnb->page);
919 int mds_choose_mdsnum(struct obd_device *obd, const char *name, int len, int flags,
920 struct ptlrpc_peer *peer, struct inode *parent)
922 struct mds_obd *mds = &obd->u.mds;
924 int i = mds->mds_num;
925 char peer_str[PTL_NALFMT_SIZE];
926 if (flags & REC_REINT_CREATE) {
928 } else if (mds->mds_md_exp != NULL && peer != NULL) {
929 LASSERT(parent != NULL);
930 /* distribute only at root level */
931 lmv = &mds->mds_md_exp->exp_obd->u.lmv;
932 if (parent->i_ino != id_ino(&mds->mds_rootid)) {
935 __u64 nid = peer->peer_id.nid;
936 __u64 count = lmv->desc.ld_tgt_count;
937 i = do_div(nid, count);
938 CWARN("client from %s creates top dir %*s on mds #%d\n",
939 ptlrpc_peernid2str(peer,peer_str), len, name, i+1);
941 } else if (mds->mds_md_exp) {
942 lmv = &mds->mds_md_exp->exp_obd->u.lmv;
943 i = raw_name2idx(MEA_MAGIC_LAST_CHAR, lmv->desc.ld_tgt_count, name, len);
948 int mds_lock_slave_objs(struct obd_device *obd, struct dentry *dentry,
949 struct lustre_handle **rlockh)
951 struct mds_obd *mds = &obd->u.mds;
952 struct mdc_op_data *op_data;
953 struct lookup_intent it;
954 struct mea *mea = NULL;
959 LASSERT(rlockh != NULL);
960 LASSERT(dentry != NULL);
961 LASSERT(dentry->d_inode != NULL);
964 if (!mds->mds_md_obd)
967 /* a dir can be splitted only */
968 if (!S_ISDIR(dentry->d_inode->i_mode))
971 rc = mds_md_get_attr(obd, dentry->d_inode,
979 if (mea->mea_count == 0)
980 /* this is slave object */
981 GOTO(cleanup, rc = 0);
983 CDEBUG(D_OTHER, "%s: lock slaves for %lu/%lu\n",
984 obd->obd_name, (unsigned long)dentry->d_inode->i_ino,
985 (unsigned long)dentry->d_inode->i_generation);
987 handle_size = sizeof(struct lustre_handle) *
990 OBD_ALLOC(*rlockh, handle_size);
992 GOTO(cleanup, rc = -ENOMEM);
994 memset(*rlockh, 0, handle_size);
995 OBD_ALLOC(op_data, sizeof(*op_data));
996 if (op_data == NULL) {
997 OBD_FREE(*rlockh, handle_size);
1000 memset(op_data, 0, sizeof(*op_data));
1002 op_data->mea1 = mea;
1003 it.it_op = IT_UNLINK;
1005 OBD_ALLOC(it.d.fs_data, sizeof(struct lustre_intent_data));
1007 rc = md_enqueue(mds->mds_md_exp, LDLM_IBITS, &it, LCK_EX,
1008 op_data, *rlockh, NULL, 0, ldlm_completion_ast,
1009 mds_blocking_ast, NULL);
1010 OBD_FREE(op_data, sizeof(*op_data));
1011 OBD_FREE(it.d.fs_data, sizeof(struct lustre_intent_data));
1014 OBD_FREE(mea, mea_size);
1018 void mds_unlock_slave_objs(struct obd_device *obd, struct dentry *dentry,
1019 struct lustre_handle *lockh)
1021 struct mds_obd *mds = &obd->u.mds;
1022 struct mea *mea = NULL;
1023 int mea_size, rc, i;
1026 if (lockh == NULL) {
1031 LASSERT(mds->mds_md_obd != NULL);
1032 LASSERT(S_ISDIR(dentry->d_inode->i_mode));
1034 rc = mds_md_get_attr(obd, dentry->d_inode, &mea, &mea_size);
1036 CERROR("locks are leaked\n");
1040 LASSERT(mea_size != 0);
1041 LASSERT(mea != NULL);
1042 LASSERT(mea->mea_count != 0);
1044 CDEBUG(D_OTHER, "%s: unlock slaves for %lu/%lu\n", obd->obd_name,
1045 (unsigned long) dentry->d_inode->i_ino,
1046 (unsigned long) dentry->d_inode->i_generation);
1048 for (i = 0; i < mea->mea_count; i++) {
1049 if (lockh[i].cookie != 0)
1050 ldlm_lock_decref(lockh + i, LCK_EX);
1053 OBD_FREE(lockh, sizeof(struct lustre_handle) * mea->mea_count);
1054 OBD_FREE(mea, mea_size);
1058 int mds_unlink_slave_objs(struct obd_device *obd, struct dentry *dentry)
1060 struct mds_obd *mds = &obd->u.mds;
1061 struct ptlrpc_request *req = NULL;
1062 struct mdc_op_data *op_data;
1063 struct mea *mea = NULL;
1067 /* clustered MD ? */
1068 if (!mds->mds_md_obd)
1071 /* a dir can be splitted only */
1072 if (!S_ISDIR(dentry->d_inode->i_mode))
1075 rc = mds_md_get_attr(obd, dentry->d_inode, &mea, &mea_size);
1082 if (mea->mea_count == 0)
1083 GOTO(cleanup, rc = 0);
1085 CDEBUG(D_OTHER, "%s: unlink slaves for %lu/%lu\n", obd->obd_name,
1086 (unsigned long)dentry->d_inode->i_ino,
1087 (unsigned long)dentry->d_inode->i_generation);
1089 OBD_ALLOC(op_data, sizeof(*op_data));
1090 if (op_data == NULL)
1093 memset(op_data, 0, sizeof(*op_data));
1094 op_data->mea1 = mea;
1095 rc = md_unlink(mds->mds_md_exp, op_data, &req);
1096 OBD_FREE(op_data, sizeof(*op_data));
1097 LASSERT(req == NULL);
1100 OBD_FREE(mea, mea_size);
1104 struct ide_tracking {
1109 int mds_ide_filldir(void *__buf, const char *name, int namelen,
1110 loff_t offset, ino_t ino, unsigned int d_type)
1112 struct ide_tracking *it = __buf;
1118 if (it->entries > 2)
1122 if (name[0] == '.' && namelen == 1)
1124 if (name[0] == '.' && name[1] == '.' && namelen == 2)
1131 /* checks if passed dentry points to empty dir. */
1132 int mds_is_dir_empty(struct obd_device *obd, struct dentry *dentry)
1134 struct ide_tracking it;
1142 nlen = strlen("__iopen__/") + 10 + 1;
1143 OBD_ALLOC(file_name, nlen);
1147 LASSERT(dentry->d_inode != NULL);
1148 i = sprintf(file_name, "__iopen__/0x%lx",
1149 dentry->d_inode->i_ino);
1151 file = filp_open(file_name, O_RDONLY, 0);
1153 CERROR("can't open directory %s: %d\n",
1154 file_name, (int) PTR_ERR(file));
1155 GOTO(cleanup, rc = PTR_ERR(file));
1158 rc = vfs_readdir(file, mds_ide_filldir, &it);
1159 filp_close(file, 0);
1161 if (it.empty && rc == 0)
1167 OBD_FREE(file_name, nlen);
1171 int mds_lock_and_check_slave(int offset, struct ptlrpc_request *req,
1172 struct lustre_handle *lockh)
1174 struct obd_device *obd = req->rq_export->exp_obd;
1175 struct dentry *dentry = NULL;
1176 struct lvfs_run_ctxt saved;
1177 int cleanup_phase = 0;
1178 struct mds_req_sec_desc *rsd;
1179 struct mds_body *body;
1180 struct lvfs_ucred uc;
1181 int rc, update_mode;
1184 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1186 CERROR("Can't unpack security desc\n");
1187 GOTO(cleanup, rc = -EFAULT);
1190 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1191 lustre_swab_mds_body);
1193 CERROR("Can't swab mds_body\n");
1194 GOTO(cleanup, rc = -EFAULT);
1196 CDEBUG(D_OTHER, "%s: check slave "DLID4"\n", obd->obd_name,
1199 dentry = mds_id2locked_dentry(obd, &body->id1, NULL, LCK_EX,
1200 lockh, &update_mode, NULL, 0,
1201 MDS_INODELOCK_UPDATE);
1202 if (IS_ERR(dentry)) {
1203 CERROR("can't find inode: %d\n", (int) PTR_ERR(dentry));
1204 GOTO(cleanup, rc = PTR_ERR(dentry));
1209 * handling the case when remote MDS checks if dir is empty
1210 * before rename. But it also does it for all entries, because
1211 * inode is stored here and remote MDS does not know if rename
1212 * point to dir or to reg file. So we check it here.
1214 if (!S_ISDIR(dentry->d_inode->i_mode))
1215 GOTO(cleanup, rc = 0);
1217 rc = mds_init_ucred(&uc, req, rsd);
1222 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1223 rc = mds_is_dir_empty(obd, dentry) ? 0 : -ENOTEMPTY;
1224 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1226 mds_exit_ucred(&uc);
1229 switch(cleanup_phase) {
1232 ldlm_lock_decref(lockh, LCK_EX);
1240 int mds_convert_mea_ea(struct obd_device *obd, struct inode *inode,
1241 struct lov_mds_md *lmm, int lmm_size)
1243 struct lov_stripe_md *lsm = NULL;
1244 struct mea_old *old;
1250 mea = (struct mea *)lmm;
1251 old = (struct mea_old *)lmm;
1253 if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
1254 mea->mea_magic == MEA_MAGIC_ALL_CHARS)
1258 * making MDS try LOV EA converting in the non-LMV configuration
1261 if (!obd->u.mds.mds_md_exp)
1264 CDEBUG(D_INODE, "converting MEA EA on %lu/%u from V0 to V1 (%u/%u)\n",
1265 inode->i_ino, inode->i_generation, old->mea_count,
1268 rc = obd_unpackmd(obd->u.mds.mds_md_exp, &lsm, lmm, lmm_size);
1272 rc = obd_packmd(obd->u.mds.mds_md_exp, &lmm, lsm);
1274 GOTO(conv_free, rc);
1277 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1278 if (IS_ERR(handle)) {
1279 rc = PTR_ERR(handle);
1280 GOTO(conv_free, rc);
1283 rc = fsfilt_set_md(obd, inode, handle, lmm, lmm_size, EA_MEA);
1284 err = fsfilt_commit(obd, obd->u.mds.mds_sb, inode, handle, 0);
1286 rc = err ? err : lmm_size;
1287 GOTO(conv_free, rc);
1289 obd_free_memmd(obd->u.mds.mds_md_exp, &lsm);