1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
5 * Lustre Metadata Server (mds) request handler
7 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
8 * Author: Peter Braam <braam@clusterfs.com>
9 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
11 * Author: Mike Shaver <shaver@clusterfs.com>
13 * This file is part of Lustre, http://www.lustre.org.
15 * Lustre is free software; you can redistribute it and/or
16 * modify it under the terms of version 2 of the GNU General Public
17 * License as published by the Free Software Foundation.
19 * Lustre is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * GNU General Public License for more details.
24 * You should have received a copy of the GNU General Public License
25 * along with Lustre; if not, write to the Free Software
26 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
30 # define EXPORT_SYMTAB
32 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/module.h>
35 #include <linux/lustre_mds.h>
36 #include <linux/lustre_dlm.h>
37 #include <linux/init.h>
38 #include <linux/obd_class.h>
39 #include <linux/random.h>
41 #include <linux/jbd.h>
42 #include <linux/namei.h>
43 #include <linux/ext3_fs.h>
44 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
45 # include <linux/smp_lock.h>
46 # include <linux/buffer_head.h>
47 # include <linux/workqueue.h>
48 # include <linux/mount.h>
50 # include <linux/locks.h>
52 #include <linux/obd_lov.h>
53 #include <linux/obd_ost.h>
54 #include <linux/lustre_mds.h>
55 #include <linux/lustre_fsfilt.h>
56 #include <linux/lprocfs_status.h>
57 #include <linux/lustre_commit_confd.h>
58 #include <linux/lustre_acl.h>
59 #include "mds_internal.h"
60 #include <linux/lustre_sec.h>
62 static int mds_intent_policy(struct ldlm_namespace *ns,
63 struct ldlm_lock **lockp, void *req_cookie,
64 ldlm_mode_t mode, int flags, void *data);
65 static int mds_postsetup(struct obd_device *obd);
66 static int mds_cleanup(struct obd_device *obd, int flags);
69 /* Assumes caller has already pushed into the kernel filesystem context */
70 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
71 loff_t offset, int count)
73 struct ptlrpc_bulk_desc *desc;
74 struct l_wait_info lwi;
76 int rc = 0, npages, i, tmpcount, tmpsize = 0;
79 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
81 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
82 OBD_ALLOC(pages, sizeof(*pages) * npages);
84 GOTO(out, rc = -ENOMEM);
86 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
89 GOTO(out_free, rc = -ENOMEM);
91 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
92 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
94 pages[i] = alloc_pages(GFP_KERNEL, 0);
96 GOTO(cleanup_buf, rc = -ENOMEM);
98 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
101 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
102 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
103 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
104 tmpsize, offset, file->f_dentry->d_inode->i_ino,
105 file->f_dentry->d_inode->i_size);
107 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
108 kmap(pages[i]), tmpsize, &offset);
112 GOTO(cleanup_buf, rc = -EIO);
115 LASSERT(desc->bd_nob == count);
117 rc = ptlrpc_start_bulk_transfer(desc);
119 GOTO(cleanup_buf, rc);
121 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
122 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
123 OBD_FAIL_MDS_SENDPAGE, rc = -EIO);
124 GOTO(abort_bulk, rc);
127 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
128 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
129 LASSERT (rc == 0 || rc == -ETIMEDOUT);
132 if (desc->bd_success &&
133 desc->bd_nob_transferred == count)
134 GOTO(cleanup_buf, rc);
136 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
139 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
140 (rc == -ETIMEDOUT) ? "timeout" : "network error",
141 desc->bd_nob_transferred, count,
142 req->rq_export->exp_client_uuid.uuid,
143 req->rq_export->exp_connection->c_remote_uuid.uuid);
145 ptlrpc_fail_export(req->rq_export);
149 ptlrpc_abort_bulk (desc);
151 for (i = 0; i < npages; i++)
153 __free_pages(pages[i], 0);
155 ptlrpc_free_bulk(desc);
157 OBD_FREE(pages, sizeof(*pages) * npages);
162 extern char *ldlm_lockname[];
164 int mds_lock_mode_for_dir(struct obd_device *obd,
165 struct dentry *dentry, int mode)
167 int ret_mode = 0, split;
169 /* any dir access needs couple locks:
170 * 1) on part of dir we gonna lookup/modify in
171 * 2) on a whole dir to protect it from concurrent splitting
172 * and to flush client's cache for readdir()
173 * so, for a given mode and dentry this routine decides what
174 * lock mode to use for lock #2:
175 * 1) if caller's gonna lookup in dir then we need to protect
176 * dir from being splitted only - LCK_CR
177 * 2) if caller's gonna modify dir then we need to protect
178 * dir from being splitted and to flush cache - LCK_CW
179 * 3) if caller's gonna modify dir and that dir seems ready
180 * for splitting then we need to protect it from any
181 * type of access (lookup/modify/split) - LCK_EX -bzzz */
183 split = mds_splitting_expected(obd, dentry);
186 * it is important to check here only for MDS_NO_SPLITTABLE. The reason
187 * is that MDS_NO_SPLITTABLE means dir is not splittable in principle
188 * and another thread will not split it on the quiet. But if we have
189 * MDS_NO_SPLIT_EXPECTED, this means, that dir may be splitted anytime,
190 * but not now (for current thread) and we should consider that it can
191 * happen soon and go that branch which can yield LCK_EX to protect from
192 * possible splitting.
194 if (split == MDS_NO_SPLITTABLE) {
196 * this inode won't be splitted. so we need not to protect from
197 * just flush client's cache on modification.
204 if (mode == LCK_EX) {
206 } else if (mode == LCK_PR) {
208 } else if (mode == LCK_PW) {
210 * caller gonna modify directory. We use concurrent
211 * write lock here to retract client's cache for
214 if (split == MDS_EXPECT_SPLIT) {
216 * splitting possible. serialize any access the
217 * idea is that first one seen dir is splittable
218 * is given exclusive lock and split
219 * directory. caller passes lock mode to
220 * mds_try_to_split_dir() and splitting would be
221 * done with exclusive lock only -bzzz.
223 CDEBUG(D_OTHER, "%s: gonna split %lu/%lu\n",
225 (unsigned long)dentry->d_inode->i_ino,
226 (unsigned long)dentry->d_inode->i_generation);
237 /* only valid locked dentries or errors should be returned */
238 struct dentry *mds_id2locked_dentry(struct obd_device *obd, struct lustre_id *id,
239 struct vfsmount **mnt, int lock_mode,
240 struct lustre_handle *lockh, int *mode,
241 char *name, int namelen, __u64 lockpart)
243 struct dentry *de = mds_id2dentry(obd, id, mnt), *retval = de;
244 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
245 struct ldlm_res_id res_id = { .name = {0} };
253 res_id.name[0] = id_fid(id);
254 res_id.name[1] = id_group(id);
257 if (name && IS_PDIROPS(de->d_inode)) {
258 ldlm_policy_data_t cpolicy =
259 { .l_inodebits = { MDS_INODELOCK_UPDATE } };
260 LASSERT(mode != NULL);
261 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
263 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
265 &cpolicy, *mode, &flags,
267 ldlm_completion_ast, NULL, NULL,
268 NULL, 0, NULL, lockh + 1);
269 if (rc != ELDLM_OK) {
271 RETURN(ERR_PTR(-ENOLCK));
276 res_id.name[2] = full_name_hash((unsigned char *)name, namelen);
278 CDEBUG(D_INFO, "take lock on "DLID4":"LPX64"\n",
279 OLID4(id), res_id.name[2]);
282 #warning "No PDIROPS support in the kernel"
284 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
285 LDLM_IBITS, &policy, lock_mode, &flags,
286 mds_blocking_ast, ldlm_completion_ast,
287 NULL, NULL, NULL, 0, NULL, lockh);
288 if (rc != ELDLM_OK) {
290 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
293 ldlm_lock_decref(lockh + 1, *mode);
300 #ifndef DCACHE_DISCONNECTED
301 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
304 /* Look up an entry by inode number. This function ONLY returns valid dget'd
305 * dentries with an initialized inode or errors */
306 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
307 struct vfsmount **mnt)
309 unsigned long ino = (unsigned long)id_ino(id);
310 __u32 generation = (__u32)id_gen(id);
311 struct mds_obd *mds = &obd->u.mds;
312 struct dentry *result;
317 RETURN(ERR_PTR(-ESTALE));
319 snprintf(idname, sizeof(idname), "0x%lx", ino);
321 CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
322 ino, generation, mds->mds_sb);
324 /* under ext3 this is neither supposed to return bad inodes nor NULL
326 result = ll_lookup_one_len(idname, mds->mds_id_de,
331 inode = result->d_inode;
333 RETURN(ERR_PTR(-ENOENT));
335 if (is_bad_inode(inode)) {
336 CERROR("bad inode returned %lu/%u\n",
337 inode->i_ino, inode->i_generation);
339 RETURN(ERR_PTR(-ENOENT));
342 /* here we disabled generation check, as root inode i_generation
343 * of cache mds and real mds are different. */
344 if (inode->i_ino != id_ino(&mds->mds_rootid) && generation &&
345 inode->i_generation != generation) {
346 /* we didn't find the right inode.. */
347 if (id_group(id) != mds->mds_num) {
348 CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
349 "%u != %u, mds %u != %u, request to wrong MDS?\n",
350 inode->i_ino, (unsigned long)inode->i_nlink,
351 atomic_read(&inode->i_count), inode->i_generation,
352 generation, mds->mds_num, (unsigned)id_group(id));
354 CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
355 "%u != %u, inode is recreated while request handled?\n",
356 inode->i_ino, (unsigned long)inode->i_nlink,
357 atomic_read(&inode->i_count), inode->i_generation,
361 RETURN(ERR_PTR(-ENOENT));
365 *mnt = mds->mds_vfsmnt;
373 int mds_req_add_idmapping(struct ptlrpc_request *req,
374 struct mds_export_data *med)
376 struct mds_req_sec_desc *rsd;
377 struct lustre_sec_desc *lsd;
380 if (!med->med_remote)
383 /* maybe we should do it more completely: invalidate the gss ctxt? */
384 if (req->rq_mapped_uid == MDS_IDMAP_NOTFOUND) {
385 CWARN("didn't find mapped uid\n");
389 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
391 CERROR("Can't unpack security desc\n");
395 lsd = mds_get_lsd(req->rq_mapped_uid);
397 CERROR("can't get LSD(%u), no mapping added\n",
402 rc = mds_idmap_add(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
403 rsd->rsd_gid, lsd->lsd_gid);
409 int mds_req_del_idmapping(struct ptlrpc_request *req,
410 struct mds_export_data *med)
412 struct mds_req_sec_desc *rsd;
413 struct lustre_sec_desc *lsd;
416 if (!med->med_remote)
419 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
421 CERROR("Can't unpack security desc\n");
425 LASSERT(req->rq_mapped_uid != -1);
426 lsd = mds_get_lsd(req->rq_mapped_uid);
428 CERROR("can't get LSD(%u), no idmapping deleted\n",
433 rc = mds_idmap_del(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
434 rsd->rsd_gid, lsd->lsd_gid);
439 static int mds_init_export_data(struct ptlrpc_request *req,
440 struct mds_export_data *med)
442 struct obd_connect_data *data, *reply;
443 int ask_remote, ask_local;
446 data = lustre_msg_buf(req->rq_reqmsg, 5, sizeof(*data));
447 reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*data));
448 LASSERT(data && reply);
450 if (med->med_initialized) {
451 CDEBUG(D_SEC, "med already initialized, reconnect?\n");
455 ask_remote = data->ocd_connect_flags & OBD_CONNECT_REMOTE;
456 ask_local = data->ocd_connect_flags & OBD_CONNECT_LOCAL;
458 /* currently the policy is simple: satisfy client as possible
461 if (req->rq_auth_uid == -1) {
463 CWARN("null sec is used, force to be local\n");
467 if (!req->rq_remote_realm)
468 CWARN("local realm asked to be remote\n");
470 } else if (ask_local) {
471 if (req->rq_remote_realm)
472 CWARN("remote realm asked to be local\n");
475 med->med_remote = (req->rq_remote_realm != 0);
478 med->med_nllu = data->ocd_nllu[0];
479 med->med_nllg = data->ocd_nllu[1];
481 med->med_initialized = 1;
483 reply->ocd_connect_flags &= ~(OBD_CONNECT_REMOTE | OBD_CONNECT_LOCAL);
484 if (med->med_remote) {
486 med->med_idmap = mds_idmap_alloc();
489 CERROR("Failed to alloc idmap, following request from "
490 "this client will be refused\n");
492 reply->ocd_connect_flags |= OBD_CONNECT_REMOTE;
493 CDEBUG(D_SEC, "set client as remote\n");
495 reply->ocd_connect_flags |= OBD_CONNECT_LOCAL;
496 CDEBUG(D_SEC, "set client as local\n");
502 static void mds_free_export_data(struct mds_export_data *med)
507 LASSERT(med->med_remote);
508 mds_idmap_free(med->med_idmap);
509 med->med_idmap = NULL;
512 /* Establish a connection to the MDS.
514 * This will set up an export structure for the client to hold state data about
515 * that client, like open files, the last operation number it did on the server,
518 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
519 struct obd_uuid *cluuid, struct obd_connect_data *data,
522 struct mds_export_data *med;
523 struct mds_client_data *mcd;
524 struct obd_export *exp;
528 if (!conn || !obd || !cluuid)
531 /* XXX There is a small race between checking the list and adding a new
532 * connection for the same UUID, but the real threat (list corruption
533 * when multiple different clients connect) is solved.
535 * There is a second race between adding the export to the list, and
536 * filling in the client data below. Hence skipping the case of NULL
537 * mcd above. We should already be controlling multiple connects at the
538 * client, and we can't hold the spinlock over memory allocations
539 * without risk of deadlocking.
541 rc = class_connect(conn, obd, cluuid);
544 exp = class_conn2export(conn);
546 LASSERT(exp != NULL);
547 med = &exp->exp_mds_data;
549 OBD_ALLOC(mcd, sizeof(*mcd));
551 CERROR("%s: out of memory for client data.\n",
553 GOTO(out, rc = -ENOMEM);
556 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
559 rc = mds_client_add(obd, &obd->u.mds, med, -1);
567 OBD_FREE(mcd, sizeof(*mcd));
568 class_disconnect(exp, 0);
570 class_export_put(exp);
575 static int mds_connect_post(struct obd_export *exp, unsigned initial,
578 struct obd_device *obd = exp->exp_obd;
579 struct mds_obd *mds = &obd->u.mds;
580 struct mds_export_data *med;
581 struct mds_client_data *mcd;
585 med = &exp->exp_mds_data;
589 /* some one reconnect initially, we have to reset
590 * data existing export can have. bug 6102 */
591 if (mcd->mcd_last_xid != 0)
592 CDEBUG(D_HA, "initial reconnect to existing export\n");
593 mcd->mcd_last_transno = 0;
594 mcd->mcd_last_xid = 0;
595 mcd->mcd_last_result = 0;
596 mcd->mcd_last_data = 0;
599 if (!(flags & OBD_OPT_MDS_CONNECTION)) {
600 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
601 atomic_inc(&mds->mds_real_clients);
602 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
603 obd->obd_name, exp->exp_client_uuid.uuid,
604 atomic_read(&mds->mds_real_clients));
605 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
607 if (mds->mds_md_name)
608 rc = mds_md_connect(obd, mds->mds_md_name);
613 static int mds_init_export(struct obd_export *exp)
615 struct mds_export_data *med = &exp->exp_mds_data;
617 INIT_LIST_HEAD(&med->med_open_head);
618 spin_lock_init(&med->med_open_lock);
622 static int mds_destroy_export(struct obd_export *export)
624 struct obd_device *obd = export->exp_obd;
625 struct mds_export_data *med = &export->exp_mds_data;
626 struct lvfs_run_ctxt saved;
630 mds_free_export_data(med);
631 target_destroy_export(export);
633 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
636 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
638 /* Close any open files (which may also cause orphan unlinking). */
639 spin_lock(&med->med_open_lock);
640 while (!list_empty(&med->med_open_head)) {
641 struct list_head *tmp = med->med_open_head.next;
642 struct mds_file_data *mfd =
643 list_entry(tmp, struct mds_file_data, mfd_list);
644 struct lustre_id sid;
646 BDEVNAME_DECLARE_STORAGE(btmp);
648 /* bug 1579: fix force-closing for 2.5 */
649 struct dentry *dentry = mfd->mfd_dentry;
651 list_del(&mfd->mfd_list);
652 spin_unlock(&med->med_open_lock);
654 down(&dentry->d_inode->i_sem);
655 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
656 up(&dentry->d_inode->i_sem);
658 CERROR("Can't read inode self id, inode %lu, "
659 "rc %d\n", dentry->d_inode->i_ino, rc);
660 memset(&sid, 0, sizeof(sid));
663 /* If you change this message, be sure to update
664 * replay_single:test_46 */
665 CERROR("force closing client file handle for %.*s (%s:"
666 DLID4")\n", dentry->d_name.len, dentry->d_name.name,
667 ll_bdevname(dentry->d_inode->i_sb, btmp),
670 /* child inode->i_alloc_sem protects orphan_dec_test and
671 * is_orphan race, mds_mfd_close drops it */
672 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
673 rc = mds_mfd_close(NULL, 0, obd, mfd,
674 !(export->exp_flags & OBD_OPT_FAILOVER));
676 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
677 spin_lock(&med->med_open_lock);
679 spin_unlock(&med->med_open_lock);
680 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
684 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
688 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
690 unsigned long irqflags;
691 struct obd_device *obd;
696 LASSERT(exp != NULL);
697 obd = class_exp2obd(exp);
699 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
700 exp->exp_handle.h_cookie);
706 * suppress any inter-mds requests durring disconnecting lmv if this is
707 * detected --force mode. This is needed to avoid endless recovery.
709 if (atomic_read(&mds->mds_real_clients) > 0 &&
710 !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
711 flags |= OBD_OPT_FORCE;
713 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
714 && !atomic_read(&mds->mds_real_clients)) {
715 /* there was no client at all */
716 mds_md_disconnect(obd, flags);
719 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
720 && atomic_dec_and_test(&mds->mds_real_clients)) {
721 /* time to drop LMV connections */
722 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
723 "Disconnnect from LMV now\n",
724 obd->obd_name, exp->exp_client_uuid.uuid);
725 mds_md_disconnect(obd, flags);
728 spin_lock_irqsave(&exp->exp_lock, irqflags);
729 exp->exp_flags = flags;
730 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
732 /* disconnect early so that clients can't keep using export */
733 rc = class_disconnect(exp, flags);
734 ldlm_cancel_locks_for_export(exp);
736 /* complete all outstanding replies */
737 spin_lock_irqsave(&exp->exp_lock, irqflags);
738 while (!list_empty(&exp->exp_outstanding_replies)) {
739 struct ptlrpc_reply_state *rs =
740 list_entry(exp->exp_outstanding_replies.next,
741 struct ptlrpc_reply_state, rs_exp_list);
742 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
744 spin_lock(&svc->srv_lock);
745 list_del_init(&rs->rs_exp_list);
746 ptlrpc_schedule_difficult_reply(rs);
747 spin_unlock(&svc->srv_lock);
749 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
753 static int mds_getstatus(struct ptlrpc_request *req)
755 struct mds_obd *mds = mds_req2mds(req);
756 struct mds_body *body;
760 size = sizeof(*body);
762 rc = lustre_pack_reply(req, 1, &size, NULL);
763 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
764 CERROR("mds: out of memory for message: size=%d\n", size);
765 req->rq_status = -ENOMEM; /* superfluous? */
769 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
770 body->valid |= OBD_MD_FID;
772 memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
775 * the last_committed and last_xid fields are filled in for all replies
776 * already - no need to do so here also.
781 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
782 void *data, int flag)
787 if (flag == LDLM_CB_CANCELING) {
788 /* Don't need to do anything here. */
792 /* XXX layering violation! -phil */
793 l_lock(&lock->l_resource->lr_namespace->ns_lock);
796 * get this: if mds_blocking_ast is racing with mds_intent_policy, such
797 * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
798 * then by the time we get the lock, we might not be the correct
799 * blocking function anymore. So check, and return early, if so.
801 if (lock->l_blocking_ast != mds_blocking_ast) {
802 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
806 lock->l_flags |= LDLM_FL_CBPENDING;
807 do_ast = (!lock->l_readers && !lock->l_writers);
808 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
811 struct lustre_handle lockh;
814 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
815 ldlm_lock2handle(lock, &lockh);
816 rc = ldlm_cli_cancel(&lockh);
818 CERROR("ldlm_cli_cancel: %d\n", rc);
820 LDLM_DEBUG(lock, "Lock still has references, will be "
826 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
827 void *md, int size, int mea)
831 if (S_ISREG(inode->i_mode)) {
832 rc = mds_convert_lov_ea(obd, inode, md, size);
833 } else if (S_ISDIR(inode->i_mode)) {
835 rc = mds_convert_mea_ea(obd, inode, md, size);
837 rc = mds_convert_lov_ea(obd, inode, md, size);
840 CERROR("Invalid EA format (nor LOV or MEA) "
841 "is detected. Inode %lu/%u\n",
842 inode->i_ino, inode->i_generation);
848 int mds_get_md(struct obd_device *obd, struct inode *inode,
849 void *md, int *size, int lock, int mea)
858 rc = fsfilt_get_md(obd, inode, md, *size,
859 (mea ? EA_MEA : EA_LOV));
861 CERROR("Error %d reading eadata for ino %lu\n",
865 rc = mds_convert_md(obd, inode, md,
880 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
881 * Call with lock=0 if the caller has already taken the i_sem. */
882 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
883 struct mds_body *body, struct inode *inode, int lock, int mea)
885 struct mds_obd *mds = &obd->u.mds;
890 lmm = lustre_msg_buf(msg, offset, 0);
892 /* Some problem with getting eadata when I sized the reply
894 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
898 lmm_size = msg->buflens[offset];
900 /* I don't really like this, but it is a sanity check on the client
901 * MD request. However, if the client doesn't know how much space
902 * to reserve for the MD, it shouldn't be bad to have too much space.
904 if (lmm_size > mds->mds_max_mdsize) {
905 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
906 inode->i_ino, lmm_size, mds->mds_max_mdsize);
910 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
912 body->valid |= S_ISDIR(inode->i_mode) ?
913 OBD_MD_FLDIREA : OBD_MD_FLEASIZE;
916 body->valid |= OBD_MD_MEA;
918 body->eadatasize = lmm_size;
925 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
926 struct mds_body *repbody, int reply_off)
928 struct inode *inode = dentry->d_inode;
933 symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
934 LASSERT(symname != NULL);
935 len = req->rq_repmsg->buflens[reply_off + 1];
937 rc = inode->i_op->readlink(dentry, symname, len);
939 CERROR("readlink failed: %d\n", rc);
940 } else if (rc != len - 1) {
941 CERROR ("Unexpected readlink rc %d: expecting %d\n",
945 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
946 repbody->valid |= OBD_MD_LINKNAME;
947 repbody->eadatasize = rc + 1;
948 symname[rc] = 0; /* NULL terminate */
955 int mds_pack_ea(struct dentry *dentry, struct ptlrpc_request *req,
956 struct mds_body *repbody, int req_off, int reply_off)
958 struct inode *inode = dentry->d_inode;
964 ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
965 len = req->rq_repmsg->buflens[reply_off + 1];
967 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
970 if (inode->i_op && inode->i_op->getxattr)
971 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
974 if (rc != -ENODATA && rc != -EOPNOTSUPP)
975 CERROR("getxattr failed: %d", rc);
977 repbody->valid |= OBD_MD_FLEA;
978 repbody->eadatasize = rc;
985 int mds_pack_ealist(struct dentry *dentry, struct ptlrpc_request *req,
986 struct mds_body *repbody, int reply_off)
988 struct inode *inode = dentry->d_inode;
993 len = req->rq_repmsg->buflens[reply_off + 1];
995 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
998 if (inode->i_op && inode->i_op->getxattr)
999 rc = inode->i_op->listxattr(dentry, value, len);
1002 CERROR("listxattr failed: %d", rc);
1004 repbody->valid |= OBD_MD_FLEALIST;
1005 repbody->eadatasize = rc;
1011 int mds_pack_acl(struct obd_device *obd, struct lustre_msg *repmsg, int offset,
1012 struct mds_body *body, struct inode *inode)
1014 struct dentry de = { .d_inode = inode };
1015 __u32 buflen, *sizep;
1020 if (!inode->i_op->getxattr)
1023 buflen = repmsg->buflens[offset + 1];
1024 buf = lustre_msg_buf(repmsg, offset + 1, buflen);
1026 size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1027 if (size == -ENODATA || size == -EOPNOTSUPP)
1033 sizep = lustre_msg_buf(repmsg, offset, 4);
1035 CERROR("can't locate returned acl size buf\n");
1039 *sizep = cpu_to_le32(size);
1040 body->valid |= OBD_MD_FLACL_ACCESS;
1046 * here we take simple rule: once uid/fsuid is root, we also squash
1047 * the gid/fsgid, don't care setuid/setgid attributes.
1049 int mds_squash_root(struct mds_obd *mds, struct mds_req_sec_desc *rsd,
1052 if (!mds->mds_squash_uid || *peernid == mds->mds_nosquash_nid)
1055 if (rsd->rsd_uid && rsd->rsd_fsuid)
1058 CDEBUG(D_SEC, "squash req from "LPX64":"
1059 "(%u:%u-%u:%u/%x)=>(%u:%u-%u:%u/%x)\n", *peernid,
1060 rsd->rsd_uid, rsd->rsd_gid,
1061 rsd->rsd_fsuid, rsd->rsd_fsgid, rsd->rsd_cap,
1062 rsd->rsd_uid ? rsd->rsd_uid : mds->mds_squash_uid,
1063 rsd->rsd_uid ? rsd->rsd_gid : mds->mds_squash_gid,
1064 rsd->rsd_fsuid ? rsd->rsd_fsuid : mds->mds_squash_uid,
1065 rsd->rsd_fsuid ? rsd->rsd_fsgid : mds->mds_squash_gid,
1066 rsd->rsd_cap & ~CAP_FS_MASK);
1068 if (rsd->rsd_uid == 0) {
1069 rsd->rsd_uid = mds->mds_squash_uid;
1070 rsd->rsd_gid = mds->mds_squash_gid;
1072 if (rsd->rsd_fsuid == 0) {
1073 rsd->rsd_fsuid = mds->mds_squash_uid;
1074 rsd->rsd_fsgid = mds->mds_squash_gid;
1076 rsd->rsd_cap &= ~CAP_FS_MASK;
1081 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1082 struct ptlrpc_request *req, int req_off,
1083 struct mds_body *reqbody, int reply_off)
1085 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1086 struct inode *inode = dentry->d_inode;
1087 struct mds_body *body;
1091 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1094 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1095 LASSERT(body != NULL); /* caller prepped reply */
1097 if (dentry->d_flags & DCACHE_CROSS_REF) {
1098 mds_pack_dentry2body(obd, body, dentry,
1099 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1100 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1105 mds_pack_inode2body(obd, body, inode,
1106 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1108 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1109 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1111 /* guessing what kind og attribute do we need. */
1112 int is_mea = (S_ISDIR(inode->i_mode) &&
1113 (reqbody->valid & OBD_MD_MEA) != 0);
1115 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
1116 body, inode, 1, is_mea);
1118 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1119 if (!(body->valid & OBD_MD_FLEASIZE) &&
1120 !(body->valid & OBD_MD_FLDIREA))
1121 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1122 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1123 } else if (S_ISLNK(inode->i_mode) &&
1124 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1125 rc = mds_pack_link(dentry, req, body, reply_off);
1126 } else if (reqbody->valid & OBD_MD_FLEA) {
1127 rc = mds_pack_ea(dentry, req, body, req_off, reply_off);
1128 } else if (reqbody->valid & OBD_MD_FLEALIST) {
1129 rc = mds_pack_ealist(dentry, req, body, reply_off);
1132 if (reqbody->valid & OBD_MD_FLACL_ACCESS) {
1133 int inc = (reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1;
1134 rc = mds_pack_acl(obd, req->rq_repmsg, reply_off + inc,
1139 mds_body_do_reverse_map(med, body);
1144 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1145 struct dentry *dentry,
1148 int rc = 0, size[1] = {sizeof(struct mds_body)};
1151 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1152 CERROR("failed MDS_GETATTR_PACK test\n");
1153 req->rq_status = -ENOMEM;
1157 rc = lustre_pack_reply(req, 1, size, NULL);
1159 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1160 GOTO(out, req->rq_status = rc);
1168 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1171 struct inode *inode = de->d_inode;
1172 struct mds_obd *mds = mds_req2mds(req);
1173 struct mds_body *body;
1174 int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1177 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1178 LASSERT(body != NULL); /* checked by caller */
1179 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
1181 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1182 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1185 down(&inode->i_sem);
1186 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1187 ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1190 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1191 CERROR("error getting inode %lu MD: rc = %d\n",
1194 } else if (rc > mds->mds_max_mdsize) {
1196 CERROR("MD size %d larger than maximum possible %u\n",
1197 rc, mds->mds_max_mdsize);
1199 size[bufcount] = rc;
1202 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1203 if (inode->i_size + 1 != body->eadatasize)
1204 CERROR("symlink size: %Lu, reply space: %d\n",
1205 inode->i_size + 1, body->eadatasize);
1206 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1208 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1209 inode->i_size + 1, body->eadatasize);
1210 } else if ((body->valid & OBD_MD_FLEA)) {
1211 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1214 if (inode->i_op && inode->i_op->getxattr)
1215 rc = inode->i_op->getxattr(de, ea_name, NULL, 0);
1218 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1219 CERROR("error getting inode %lu EA: rc = %d\n",
1223 size[bufcount] = min_t(int, body->eadatasize, rc);
1226 } else if (body->valid & OBD_MD_FLEALIST) {
1228 if (inode->i_op && inode->i_op->getxattr)
1229 rc = inode->i_op->listxattr(de, NULL, 0);
1232 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1233 CERROR("error getting inode %lu EA: rc = %d\n",
1237 size[bufcount] = min_t(int, body->eadatasize, rc);
1242 /* may co-exist with OBD_MD_FLEASIZE */
1243 if (body->valid & OBD_MD_FLACL_ACCESS) {
1244 size[bufcount++] = 4;
1245 size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1248 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1249 CERROR("failed MDS_GETATTR_PACK test\n");
1250 req->rq_status = -ENOMEM;
1251 GOTO(out, rc = -ENOMEM);
1254 rc = lustre_pack_reply(req, bufcount, size, NULL);
1256 CERROR("out of memory\n");
1257 GOTO(out, req->rq_status = rc);
1265 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1266 char *name, int namelen)
1268 struct mea *mea = NULL;
1269 int mea_size, rc = 0;
1272 rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1277 * dir is already splitted, check if requested filename should
1278 * live at this MDS or at another one.
1280 int i = mea_name2idx(mea, name, namelen - 1);
1281 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1283 "inapropriate MDS(%d) for %s. should be "
1284 "%lu(%d)\n", mea->mea_master, name,
1285 (unsigned long)id_group(&mea->mea_ids[i]), i);
1291 OBD_FREE(mea, mea_size);
1295 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1296 struct lustre_handle *child_lockh, int child_part)
1298 struct obd_device *obd = req->rq_export->exp_obd;
1299 struct mds_obd *mds = &obd->u.mds;
1300 struct ldlm_reply *rep = NULL;
1301 struct lvfs_run_ctxt saved;
1302 struct mds_req_sec_desc *rsd;
1303 struct mds_body *body;
1304 struct dentry *dparent = NULL, *dchild = NULL;
1305 struct lvfs_ucred uc = {NULL, NULL,};
1306 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1307 unsigned int namesize;
1308 int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1312 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
1313 MD_COUNTER_INCREMENT(obd, getattr_lock);
1315 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1317 CERROR("Can't unpack security desc\n");
1321 /* swab now, before anyone looks inside the request. */
1322 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1323 lustre_swab_mds_body);
1325 CERROR("Can't swab mds_body\n");
1326 GOTO(cleanup, rc = -EFAULT);
1329 LASSERT_REQSWAB(req, offset + 1);
1330 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1332 CERROR("Can't unpack name\n");
1333 GOTO(cleanup, rc = -EFAULT);
1335 namesize = req->rq_reqmsg->buflens[offset + 1];
1337 /* namesize less than 2 means we have empty name, probably came from
1338 revalidate by cfid, so no point in having name to be set */
1342 LASSERT (offset == 1 || offset == 3);
1344 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1350 rc = mds_init_ucred(&uc, req, rsd);
1355 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1356 cleanup_phase = 1; /* kernel context */
1357 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1359 LASSERT(namesize > 0);
1360 if (child_lockh->cookie != 0) {
1361 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1365 if (body->valid == OBD_MD_FLID) {
1366 struct mds_body *mds_reply;
1367 int size = sizeof(*mds_reply);
1371 dparent = mds_id2dentry(obd, &body->id1, NULL);
1372 if (IS_ERR(dparent)) {
1373 rc = PTR_ERR(dparent);
1377 * the user requested ONLY the inode number, so do a raw lookup.
1379 rc = lustre_pack_reply(req, 1, &size, NULL);
1381 CERROR("out of memory\n");
1385 dir = dparent->d_inode;
1386 LASSERT(dir->i_op->lookup_raw != NULL);
1387 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1389 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1390 sizeof(*mds_reply));
1392 id_ino(&mds_reply->id1) = inum;
1393 mds_reply->valid = OBD_MD_FLID;
1397 if (resent_req == 0) {
1398 LASSERT(id_fid(&body->id1) != 0);
1400 rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1401 parent_lockh, &dparent,
1403 MDS_INODELOCK_UPDATE,
1406 child_lockh, &dchild,
1407 LCK_PR, child_part);
1411 cleanup_phase = 2; /* dchild, dparent, locks */
1414 * let's make sure this name should leave on this mds
1417 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1421 /* we have no dentry here, drop LOOKUP bit */
1422 /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1423 child_part &= ~MDS_INODELOCK_LOOKUP;
1424 CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1425 obd->obd_name, OLID4(&body->id1));
1427 dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1428 LCK_PR, parent_lockh,
1431 MDS_INODELOCK_UPDATE);
1432 if (IS_ERR(dchild)) {
1433 CERROR("can't find inode with id "DLID4", err = %d\n",
1434 OLID4(&body->id1), (int)PTR_ERR(dchild));
1435 GOTO(cleanup, rc = PTR_ERR(dchild));
1437 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1440 struct ldlm_lock *granted_lock;
1442 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1443 granted_lock = ldlm_handle2lock(child_lockh);
1445 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1446 id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1447 child_lockh->cookie);
1450 /* usual named request */
1451 dparent = mds_id2dentry(obd, &body->id1, NULL);
1452 LASSERT(!IS_ERR(dparent));
1453 dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1454 LASSERT(!IS_ERR(dchild));
1456 /* client wants to get attr. by id */
1457 dchild = mds_id2dentry(obd, &body->id1, NULL);
1458 LASSERT(!IS_ERR(dchild));
1460 LDLM_LOCK_PUT(granted_lock);
1463 cleanup_phase = 2; /* dchild, dparent, locks */
1465 if (!DENTRY_VALID(dchild)) {
1466 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1468 * in the intent case, the policy clears this error: the
1469 * disposition is enough.
1474 intent_set_disposition(rep, DISP_LOOKUP_POS);
1477 if (req->rq_repmsg == NULL) {
1478 if (dchild->d_flags & DCACHE_CROSS_REF)
1479 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1481 rc = mds_getattr_pack_msg(req, dchild, offset);
1483 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1488 rc = mds_getattr_internal(obd, dchild, req, offset, body, reply_offset);
1489 GOTO(cleanup, rc); /* returns the lock to the client */
1492 switch (cleanup_phase) {
1494 if (resent_req == 0) {
1495 if (rc && DENTRY_VALID(dchild))
1496 ldlm_lock_decref(child_lockh, LCK_PR);
1498 ldlm_lock_decref(parent_lockh, LCK_PR);
1500 if (parent_lockh[1].cookie != 0)
1501 ldlm_lock_decref(parent_lockh + 1, update_mode);
1508 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1510 mds_exit_ucred(&uc);
1515 static int mds_getattr(struct ptlrpc_request *req, int offset)
1517 struct obd_device *obd = req->rq_export->exp_obd;
1518 struct lvfs_run_ctxt saved;
1520 struct mds_req_sec_desc *rsd;
1521 struct mds_body *body;
1522 struct lvfs_ucred uc = {NULL, NULL,};
1526 MD_COUNTER_INCREMENT(obd, getattr);
1528 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1530 CERROR("Can't unpack security desc\n");
1534 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1535 lustre_swab_mds_body);
1537 CERROR ("Can't unpack body\n");
1541 rc = mds_init_ucred(&uc, req, rsd);
1543 mds_exit_ucred(&uc);
1547 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1548 de = mds_id2dentry(obd, &body->id1, NULL);
1550 rc = req->rq_status = PTR_ERR(de);
1554 rc = mds_getattr_pack_msg(req, de, offset);
1556 CERROR("mds_getattr_pack_msg: %d\n", rc);
1560 req->rq_status = mds_getattr_internal(obd, de, req, offset, body, 0);
1565 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1566 mds_exit_ucred(&uc);
1570 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1571 unsigned long max_age)
1576 spin_lock(&obd->obd_osfs_lock);
1577 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1579 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1580 spin_unlock(&obd->obd_osfs_lock);
1585 static int mds_statfs(struct ptlrpc_request *req)
1587 struct obd_device *obd = req->rq_export->exp_obd;
1588 int rc, size = sizeof(struct obd_statfs);
1591 /* This will trigger a watchdog timeout */
1592 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1593 (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1595 rc = lustre_pack_reply(req, 1, &size, NULL);
1596 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1597 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1601 OBD_COUNTER_INCREMENT(obd, statfs);
1603 /* We call this so that we can cache a bit - 1 jiffie worth */
1604 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1607 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1613 req->rq_status = rc;
1617 static int mds_sync(struct ptlrpc_request *req, int offset)
1619 struct obd_device *obd = req->rq_export->exp_obd;
1620 struct mds_obd *mds = &obd->u.mds;
1621 struct mds_body *body;
1622 int rc, size = sizeof(*body);
1625 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1626 lustre_swab_mds_body);
1628 GOTO(out, rc = -EPROTO);
1630 rc = lustre_pack_reply(req, 1, &size, NULL);
1631 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1632 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1636 if (id_ino(&body->id1) == 0) {
1637 /* an id of zero is taken to mean "sync whole filesystem" */
1638 rc = fsfilt_sync(obd, mds->mds_sb);
1642 /* just any file to grab fsync method - "file" arg unused */
1643 struct file *file = mds->mds_rcvd_filp;
1644 struct mds_body *rep_body;
1647 de = mds_id2dentry(obd, &body->id1, NULL);
1649 GOTO(out, rc = PTR_ERR(de));
1651 rc = file->f_op->fsync(NULL, de, 1);
1655 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1656 mds_pack_inode2body(obd, rep_body, de->d_inode,
1657 (body->valid & OBD_MD_FID) ? 1 : 0);
1663 req->rq_status = rc;
1667 /* mds_readpage does not take a DLM lock on the inode, because the client must
1668 * already have a PR lock.
1670 * If we were to take another one here, a deadlock will result, if another
1671 * thread is already waiting for a PW lock. */
1672 static int mds_readpage(struct ptlrpc_request *req, int offset)
1674 struct obd_device *obd = req->rq_export->exp_obd;
1675 struct vfsmount *mnt;
1678 struct mds_req_sec_desc *rsd;
1679 struct mds_body *body, *repbody;
1680 struct lvfs_run_ctxt saved;
1681 int rc, size = sizeof(*repbody);
1682 struct lvfs_ucred uc = {NULL, NULL,};
1685 rc = lustre_pack_reply(req, 1, &size, NULL);
1686 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1687 CERROR("mds: out of memory\n");
1688 GOTO(out, rc = -ENOMEM);
1691 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1693 CERROR("Can't unpack security desc\n");
1694 GOTO (out, rc = -EFAULT);
1697 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1698 lustre_swab_mds_body);
1700 CERROR("Can't unpack body\n");
1701 GOTO (out, rc = -EFAULT);
1704 rc = mds_init_ucred(&uc, req, rsd);
1709 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1710 de = mds_id2dentry(obd, &body->id1, &mnt);
1712 GOTO(out_pop, rc = PTR_ERR(de));
1714 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1716 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1717 /* note: in case of an error, dentry_open puts dentry */
1719 GOTO(out_pop, rc = PTR_ERR(file));
1721 /* body->size is actually the offset -eeb */
1722 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1723 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1724 body->size, de->d_inode->i_blksize);
1725 GOTO(out_file, rc = -EFAULT);
1728 /* body->nlink is actually the #bytes to read -eeb */
1729 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1730 CERROR("size %u is not multiple of blocksize %lu\n",
1731 body->nlink, de->d_inode->i_blksize);
1732 GOTO(out_file, rc = -EFAULT);
1735 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1736 repbody->size = file->f_dentry->d_inode->i_size;
1737 repbody->valid = OBD_MD_FLSIZE;
1739 /* to make this asynchronous make sure that the handling function
1740 doesn't send a reply when this function completes. Instead a
1741 callback function would send the reply */
1742 /* body->size is actually the offset -eeb */
1743 rc = mds_sendpage(req, file, body->size, body->nlink);
1747 filp_close(file, 0);
1749 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1751 mds_exit_ucred(&uc);
1752 req->rq_status = rc;
1756 /* update master MDS ID, which is stored in local inode EA. */
1757 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1758 void *data, int data_len)
1760 struct mds_obd *mds = &obd->u.mds;
1761 struct dentry *dentry;
1769 dentry = mds_id2dentry(obd, id, NULL);
1771 GOTO(out, rc = PTR_ERR(dentry));
1773 if (!dentry->d_inode) {
1774 CERROR("Can't find object "DLID4".\n",
1776 GOTO(out_dentry, rc = -EINVAL);
1779 handle = fsfilt_start(obd, dentry->d_inode,
1780 FSFILT_OP_SETATTR, NULL);
1782 GOTO(out_dentry, rc = PTR_ERR(handle));
1784 rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1785 (struct lustre_id *)data);
1787 CERROR("Can't update inode "DLID4" master id, "
1788 "error = %d.\n", OLID4(id), rc);
1789 GOTO(out_commit, rc);
1794 fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1801 EXPORT_SYMBOL(mds_update_mid);
1803 /* read master MDS ID, which is stored in local inode EA. */
1804 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1805 void *data, int data_len)
1807 struct dentry *dentry;
1814 dentry = mds_id2dentry(obd, id, NULL);
1816 GOTO(out, rc = PTR_ERR(dentry));
1818 if (!dentry->d_inode) {
1819 CERROR("Can't find object "DLID4".\n",
1821 GOTO(out_dentry, rc = -EINVAL);
1824 down(&dentry->d_inode->i_sem);
1825 rc = mds_read_inode_mid(obd, dentry->d_inode,
1826 (struct lustre_id *)data);
1827 up(&dentry->d_inode->i_sem);
1829 CERROR("Can't read inode "DLID4" master id, "
1830 "error = %d.\n", OLID4(id), rc);
1831 GOTO(out_dentry, rc);
1840 EXPORT_SYMBOL(mds_read_mid);
1842 int mds_read_md(struct obd_device *obd, struct lustre_id *id,
1843 char **data, int *datalen)
1845 struct dentry *dentry;
1846 struct mds_obd *mds = &obd->u.mds;
1847 int rc = 0, mea = 0;
1854 dentry = mds_id2dentry(obd, id, NULL);
1856 GOTO(out, rc = PTR_ERR(dentry));
1858 if (!dentry->d_inode) {
1859 CERROR("Can't find object "DLID4".\n",
1861 GOTO(out_dentry, rc = -EINVAL);
1863 if (S_ISDIR(dentry->d_inode->i_mode)) {
1864 *datalen = obd_packmd(mds->mds_md_exp, NULL, NULL);
1867 *datalen = obd_packmd(mds->mds_dt_exp, NULL, NULL);
1870 OBD_ALLOC(ea, *datalen);
1873 GOTO(out_dentry, rc = PTR_ERR(dentry));
1876 down(&dentry->d_inode->i_sem);
1877 rc = fsfilt_get_md(obd, dentry->d_inode, *data, *datalen,
1878 (mea ? EA_MEA : EA_LOV));
1879 up(&dentry->d_inode->i_sem);
1882 CERROR("Error %d reading eadata for ino %lu\n",
1883 rc, dentry->d_inode->i_ino);
1889 EXPORT_SYMBOL(mds_read_md);
1891 int mds_reint(struct ptlrpc_request *req, int offset,
1892 struct lustre_handle *lockh)
1894 struct mds_update_record *rec;
1895 struct mds_req_sec_desc *rsd;
1899 OBD_ALLOC(rec, sizeof(*rec));
1903 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1905 CERROR("Can't unpack security desc\n");
1906 GOTO(out, rc = -EFAULT);
1909 rc = mds_update_unpack(req, offset, rec);
1910 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1911 CERROR("invalid record\n");
1912 GOTO(out, req->rq_status = -EINVAL);
1915 rc = mds_init_ucred(&rec->ur_uc, req, rsd);
1920 /* rc will be used to interrupt a for loop over multiple records */
1921 rc = mds_reint_rec(rec, offset, req, lockh);
1924 mds_exit_ucred(&rec->ur_uc);
1925 OBD_FREE(rec, sizeof(*rec));
1929 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1930 struct obd_device *obd, int *process)
1932 switch (req->rq_reqmsg->opc) {
1933 case MDS_CONNECT: /* This will never get here, but for completeness. */
1934 case OST_CONNECT: /* This will never get here, but for completeness. */
1935 case MDS_DISCONNECT:
1936 case OST_DISCONNECT:
1941 case MDS_SYNC: /* used in unmounting */
1946 *process = target_queue_recovery_request(req, obd);
1950 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1952 /* XXX what should we set rq_status to here? */
1953 req->rq_status = -EAGAIN;
1954 RETURN(ptlrpc_error(req));
1958 static char *reint_names[] = {
1959 [REINT_SETATTR] "setattr",
1960 [REINT_CREATE] "create",
1961 [REINT_LINK] "link",
1962 [REINT_UNLINK] "unlink",
1963 [REINT_RENAME] "rename",
1964 [REINT_OPEN] "open",
1967 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER | \
1968 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
1969 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
1972 static void reconstruct_create(struct ptlrpc_request *req)
1974 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1975 struct mds_client_data *mcd = med->med_mcd;
1976 struct dentry *dentry;
1977 struct ost_body *body;
1978 struct lustre_id id;
1982 /* copy rc, transno and disp; steal locks */
1983 mds_req_from_mcd(req, mcd);
1984 if (req->rq_status) {
1992 id_ino(&id) = mcd->mcd_last_data;
1993 LASSERT(id_ino(&id) != 0);
1995 dentry = mds_id2dentry(req2obd(req), &id, NULL);
1996 if (IS_ERR(dentry)) {
1997 CERROR("can't find inode "LPU64"\n", id_ino(&id));
1998 req->rq_status = PTR_ERR(dentry);
2003 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
2004 req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
2005 dentry->d_inode->i_generation);
2007 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2008 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
2009 body->oa.o_id = dentry->d_inode->i_ino;
2010 body->oa.o_generation = dentry->d_inode->i_generation;
2011 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2013 down(&dentry->d_inode->i_sem);
2014 rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
2015 up(&dentry->d_inode->i_sem);
2017 CERROR("Can't read inode self id, inode %lu, "
2018 "rc %d\n", dentry->d_inode->i_ino, rc);
2022 body->oa.o_fid = id_fid(&id);
2023 body->oa.o_mds = id_group(&id);
2029 static int mds_inode_init_acl(struct obd_device *obd, void *handle,
2030 struct dentry *de, void *xattr, int xattr_size)
2032 struct inode *inode = de->d_inode;
2033 struct posix_acl *acl;
2040 LASSERT(xattr_size > 0);
2042 if (!inode->i_op->getxattr || !inode->i_op->setxattr) {
2043 CERROR("backend fs dosen't support xattr\n");
2047 /* set default acl */
2048 if (S_ISDIR(inode->i_mode)) {
2049 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_DEFAULT,
2050 xattr, xattr_size, 0);
2052 CERROR("set default acl err: %d\n", rc);
2057 /* set access acl */
2058 acl = posix_acl_from_xattr(xattr, xattr_size);
2059 if (acl == NULL || IS_ERR(acl)) {
2060 CERROR("insane attr data\n");
2061 return PTR_ERR(acl);
2064 if (posix_acl_valid(acl)) {
2065 CERROR("default acl not valid: %d\n", rc);
2070 mode = inode->i_mode;
2071 rc = posix_acl_create_masq(acl, &mode);
2073 CERROR("create masq err %d\n", rc);
2077 if (inode->i_mode != mode) {
2078 struct iattr iattr = { .ia_valid = ATTR_MODE,
2082 rc2 = fsfilt_setattr(obd, de, handle, &iattr, 0);
2084 CERROR("setattr mode err: %d\n", rc2);
2091 /* we didn't change acl except mode bits of some
2092 * entries, so should be fit into original size.
2094 rc = posix_acl_to_xattr(acl, xattr, xattr_size);
2097 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_ACCESS,
2098 xattr, xattr_size, 0);
2100 CERROR("set access acl err: %d\n", rc);
2103 posix_acl_release(acl);
2107 static int mdt_obj_create(struct ptlrpc_request *req)
2109 struct obd_device *obd = req->rq_export->exp_obd;
2110 struct mds_obd *mds = &obd->u.mds;
2111 struct ost_body *body, *repbody;
2114 char idname[LL_ID_NAMELEN];
2115 int size = sizeof(*repbody);
2116 struct inode *parent_inode;
2117 struct lvfs_run_ctxt saved;
2118 int rc, cleanup_phase = 0;
2119 struct dentry *new = NULL;
2120 struct dentry_params dp;
2121 int mealen, flags = 0;
2122 struct lvfs_ucred uc;
2123 struct lustre_id id;
2125 void *handle = NULL;
2126 unsigned long cr_inum = 0;
2129 DEBUG_REQ(D_HA, req, "create remote object");
2130 parent_inode = mds->mds_unnamed_dir->d_inode;
2132 body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2133 lustre_swab_ost_body);
2137 /* acl data is packed transparently, no swab here */
2138 LASSERT(req->rq_reqmsg->bufcount >= 2);
2139 acl_size = req->rq_reqmsg->buflens[1];
2141 acl = lustre_msg_buf(req->rq_reqmsg, 1, acl_size);
2143 CERROR("No default acl buf?\n");
2148 rc = lustre_pack_reply(req, 1, &size, NULL);
2152 MDS_CHECK_RESENT(req, reconstruct_create(req));
2155 uc.luc_ginfo = NULL;
2156 uc.luc_uid = body->oa.o_uid;
2157 uc.luc_gid = body->oa.o_gid;
2158 uc.luc_fsuid = body->oa.o_uid;
2159 uc.luc_fsgid = body->oa.o_gid;
2161 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2162 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2164 /* in REPLAY case inum should be given (client or other MDS fills it) */
2165 if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2166 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2168 * this is re-create request from MDS holding directory name.
2169 * we have to lookup given ino/gen first. if it exists (good
2170 * case) then there is nothing to do. if it does not then we
2171 * have to recreate it.
2173 id_ino(&id) = body->oa.o_id;
2174 id_gen(&id) = body->oa.o_generation;
2176 new = mds_id2dentry(obd, &id, NULL);
2177 if (!IS_ERR(new) && new->d_inode) {
2178 struct lustre_id sid;
2180 CWARN("mkdir() repairing is on its way: %lu/%lu\n",
2181 (unsigned long)id_ino(&id), (unsigned long)id_gen(&id));
2183 obdo_from_inode(&repbody->oa, new->d_inode,
2184 FILTER_VALID_FLAGS);
2186 repbody->oa.o_id = new->d_inode->i_ino;
2187 repbody->oa.o_generation = new->d_inode->i_generation;
2188 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2191 down(&new->d_inode->i_sem);
2192 rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2193 up(&new->d_inode->i_sem);
2195 CERROR("Can't read inode self id "
2196 "inode %lu, rc %d.\n",
2197 new->d_inode->i_ino, rc);
2201 repbody->oa.o_fid = id_fid(&sid);
2202 repbody->oa.o_mds = id_group(&sid);
2203 LASSERT(id_fid(&sid) != 0);
2206 * here we could use fid passed in body->oa.o_fid and
2207 * thus avoid mds_read_inode_sid().
2209 cr_inum = new->d_inode->i_ino;
2210 GOTO(cleanup, rc = 0);
2214 down(&parent_inode->i_sem);
2215 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2216 if (IS_ERR(handle)) {
2217 up(&parent_inode->i_sem);
2218 CERROR("fsfilt_start() failed, rc = %d\n",
2219 (int)PTR_ERR(handle));
2220 GOTO(cleanup, rc = PTR_ERR(handle));
2222 cleanup_phase = 1; /* transaction */
2225 rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2226 new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2228 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2229 obd->obd_name, idname, (int) PTR_ERR(new));
2230 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2231 up(&parent_inode->i_sem);
2232 RETURN(PTR_ERR(new));
2233 } else if (new->d_inode) {
2234 CERROR("%s: name exists. repeat\n", obd->obd_name);
2238 new->d_fsdata = (void *)&dp;
2242 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2243 (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2244 LASSERT(body->oa.o_id != 0);
2245 dp.p_inum = body->oa.o_id;
2246 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2247 (unsigned long)body->oa.o_id,
2248 (unsigned long)body->oa.o_generation);
2251 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2254 rc = mds_inode_init_acl(obd, handle, new,
2257 up(&parent_inode->i_sem);
2261 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2262 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2263 new->d_inode->i_generation = body->oa.o_generation;
2264 mark_inode_dirty(new->d_inode);
2267 * avoiding asserts in cache flush case, as
2268 * @body->oa.o_id should be zero.
2270 if (body->oa.o_id) {
2271 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2272 "BUG 3550: failed to recreate obj "
2273 LPU64" -> %lu\n", body->oa.o_id,
2274 new->d_inode->i_ino);
2276 LASSERTF(body->oa.o_generation ==
2277 new->d_inode->i_generation,
2278 "BUG 3550: failed to recreate obj/gen "
2279 LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2280 body->oa.o_generation,
2281 new->d_inode->i_ino,
2282 new->d_inode->i_generation);
2286 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2287 repbody->oa.o_id = new->d_inode->i_ino;
2288 repbody->oa.o_generation = new->d_inode->i_generation;
2289 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2291 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2292 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2293 id_group(&id) = mds->mds_num;
2295 LASSERT(body->oa.o_fid != 0);
2296 id_fid(&id) = body->oa.o_fid;
2298 LASSERT(body->oa.o_id != 0);
2299 id_ino(&id) = repbody->oa.o_id;
2300 id_gen(&id) = repbody->oa.o_generation;
2302 down(&new->d_inode->i_sem);
2303 rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
2304 up(&new->d_inode->i_sem);
2307 * make sure, that fid is up-to-date.
2309 mds_set_last_fid(obd, id_fid(&id));
2312 * allocate new sid, as object is created from scratch
2313 * and this is not replay.
2315 down(&new->d_inode->i_sem);
2316 rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
2317 up(&new->d_inode->i_sem);
2320 CERROR("Can't update lustre ID for inode %lu, "
2321 "error = %d\n", new->d_inode->i_ino, rc);
2325 /* initializing o_fid after it is allocated. */
2326 repbody->oa.o_fid = id_fid(&id);
2327 repbody->oa.o_mds = id_group(&id);
2329 rc = fsfilt_del_dir_entry(obd, new);
2330 up(&parent_inode->i_sem);
2332 CERROR("can't remove name for object: %d\n", rc);
2336 cleanup_phase = 2; /* created directory object */
2338 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2339 (unsigned long)new->d_inode->i_ino,
2340 (unsigned long)new->d_inode->i_generation,
2341 (unsigned)new->d_inode->i_mode);
2342 cr_inum = new->d_inode->i_ino;
2344 up(&parent_inode->i_sem);
2345 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2349 if (body->oa.o_valid & OBD_MD_FLID) {
2350 /* this is new object for splitted dir. We have to prevent
2351 * recursive splitting on it -bzzz */
2352 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2354 OBD_ALLOC(mea, mealen);
2356 GOTO(cleanup, rc = -ENOMEM);
2358 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2359 mea->mea_master = 0;
2362 down(&new->d_inode->i_sem);
2363 rc = fsfilt_set_md(obd, new->d_inode, handle,
2364 mea, mealen, EA_MEA);
2365 up(&new->d_inode->i_sem);
2367 CERROR("fsfilt_set_md() failed, "
2370 OBD_FREE(mea, mealen);
2372 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2373 obd->obd_name, new->d_inode->i_ino,
2374 new->d_inode->i_generation, flags);
2375 } else if (body->oa.o_easize) {
2376 /* we pass LCK_EX to split routine to signal that we have
2377 * exclusive access to the directory. simple because nobody
2378 * knows it already exists -bzzz */
2379 rc = mds_try_to_split_dir(obd, new, NULL,
2380 body->oa.o_easize, LCK_EX);
2382 CERROR("Can't split directory %lu, error = %d.\n",
2383 new->d_inode->i_ino, rc);
2391 switch (cleanup_phase) {
2392 case 2: /* object has been created, but we'll may want to replay it later */
2394 ptlrpc_require_repack(req);
2395 case 1: /* transaction */
2396 rc = mds_finish_transno(mds, parent_inode, handle,
2401 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2405 static int mdt_get_info(struct ptlrpc_request *req)
2407 struct obd_export *exp = req->rq_export;
2412 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2414 DEBUG_REQ(D_HA, req, "no get_info key");
2417 keylen = req->rq_reqmsg->buflens[0];
2419 if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2420 (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2421 (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2424 if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2425 struct lustre_id *reply;
2426 int size = sizeof(*reply);
2428 rc = lustre_pack_reply(req, 1, &size, NULL);
2432 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2433 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2436 int size = sizeof(*reply);
2438 rc = lustre_pack_reply(req, 1, &size, NULL);
2442 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2443 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2446 req->rq_repmsg->status = 0;
2450 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2451 void *key, __u32 vallen, void *val)
2453 struct obd_device *obd;
2454 struct mds_obd *mds;
2458 obd = class_exp2obd(exp);
2460 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2461 exp->exp_handle.h_cookie);
2466 if (keylen >= strlen("mds_type") &&
2467 memcmp(key, "mds_type", keylen) == 0) {
2471 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
2473 mds->mds_obd_type = *(int*)val;
2474 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
2475 valsize = sizeof(group);
2477 /* mds number has been changed, so the corresponding obdfilter
2478 * exp need to be changed too. */
2479 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
2480 "mds_conn", valsize, &group);
2483 CDEBUG(D_IOCTL, "invalid key\n");
2487 static int mdt_set_info(struct ptlrpc_request *req)
2490 struct obd_export *exp = req->rq_export;
2491 int keylen, rc = 0, vallen;
2494 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2496 DEBUG_REQ(D_HA, req, "no set_info key");
2499 keylen = req->rq_reqmsg->buflens[0];
2501 if (keylen == strlen("mds_type") &&
2502 memcmp(key, "mds_type", keylen) == 0) {
2503 rc = lustre_pack_reply(req, 0, NULL, NULL);
2507 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2508 vallen = req->rq_reqmsg->buflens[1];
2510 rc = obd_set_info(exp, keylen, key, vallen, val);
2511 req->rq_repmsg->status = 0;
2514 CDEBUG(D_IOCTL, "invalid key\n");
2518 static void mds_revoke_export_locks(struct obd_export *exp)
2520 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
2521 struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2522 struct ldlm_lock *lock, *next;
2523 struct ldlm_lock_desc desc;
2525 if (!exp->u.eu_mds_data.med_remote)
2529 l_lock(&ns->ns_lock);
2530 list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2531 if (lock->l_req_mode != lock->l_granted_mode)
2534 LASSERT(lock->l_resource);
2535 if (lock->l_resource->lr_type != LDLM_IBITS &&
2536 lock->l_resource->lr_type != LDLM_PLAIN)
2539 if (lock->l_flags & LDLM_FL_AST_SENT)
2542 lock->l_flags |= LDLM_FL_AST_SENT;
2544 /* the desc just pretend to exclusive */
2545 ldlm_lock2desc(lock, &desc);
2546 desc.l_req_mode = LCK_EX;
2547 desc.l_granted_mode = 0;
2549 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2551 l_unlock(&ns->ns_lock);
2555 static int mds_msg_check_version(struct lustre_msg *msg)
2561 case MDS_DISCONNECT:
2563 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2565 CERROR("bad opc %u version %08x, expecting %08x\n",
2566 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2571 case MDS_GETATTR_LOCK:
2575 case MDS_DONE_WRITING:
2578 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2580 CERROR("bad opc %u version %08x, expecting %08x\n",
2581 msg->opc, msg->version, LUSTRE_MDS_VERSION);
2585 case LDLM_BL_CALLBACK:
2586 case LDLM_CP_CALLBACK:
2587 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2589 CERROR("bad opc %u version %08x, expecting %08x\n",
2590 msg->opc, msg->version, LUSTRE_DLM_VERSION);
2592 case OBD_LOG_CANCEL:
2593 case LLOG_ORIGIN_HANDLE_OPEN:
2594 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2595 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2596 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2597 case LLOG_ORIGIN_HANDLE_CLOSE:
2599 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2601 CERROR("bad opc %u version %08x, expecting %08x\n",
2602 msg->opc, msg->version, LUSTRE_LOG_VERSION);
2608 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2610 CERROR("bad opc %u version %08x, expecting %08x\n",
2611 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2614 case SEC_INIT_CONTINUE:
2619 CERROR("MDS unknown opcode %d\n", msg->opc);
2627 int mds_handle(struct ptlrpc_request *req)
2629 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2630 struct obd_device *obd = NULL;
2631 struct mds_obd *mds = NULL; /* quell gcc overwarning */
2635 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2637 rc = mds_msg_check_version(req->rq_reqmsg);
2639 CERROR("MDS drop mal-formed request\n");
2643 /* Security opc should NOT trigger any recovery events */
2644 if (req->rq_reqmsg->opc == SEC_INIT ||
2645 req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
2646 if (req->rq_export) {
2647 mds_req_add_idmapping(req,
2648 &req->rq_export->exp_mds_data);
2649 mds_revoke_export_locks(req->rq_export);
2652 } else if (req->rq_reqmsg->opc == SEC_FINI) {
2653 if (req->rq_export) {
2654 mds_req_del_idmapping(req,
2655 &req->rq_export->exp_mds_data);
2656 mds_revoke_export_locks(req->rq_export);
2661 LASSERT(current->journal_info == NULL);
2662 /* XXX identical to OST */
2663 if (req->rq_reqmsg->opc != MDS_CONNECT) {
2664 struct mds_export_data *med;
2667 if (req->rq_export == NULL) {
2668 CERROR("operation %d on unconnected MDS from %s\n",
2669 req->rq_reqmsg->opc,
2671 req->rq_status = -ENOTCONN;
2672 GOTO(out, rc = -ENOTCONN);
2675 med = &req->rq_export->exp_mds_data;
2676 obd = req->rq_export->exp_obd;
2679 /* sanity check: if the xid matches, the request must
2680 * be marked as a resent or replayed */
2681 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
2682 req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
2683 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2684 (MSG_RESENT | MSG_REPLAY),
2685 "rq_xid "LPU64" matches last_xid, "
2686 "expected RESENT flag\n",
2689 /* else: note the opposite is not always true; a
2690 * RESENT req after a failover will usually not match
2691 * the last_xid, since it was likely never
2692 * committed. A REPLAYed request will almost never
2693 * match the last xid, however it could for a
2694 * committed, but still retained, open. */
2696 spin_lock_bh(&obd->obd_processing_task_lock);
2697 recovering = obd->obd_recovering;
2698 spin_unlock_bh(&obd->obd_processing_task_lock);
2700 rc = mds_filter_recovery_request(req, obd,
2702 if (rc || should_process == 0) {
2704 } else if (should_process < 0) {
2705 req->rq_status = should_process;
2706 rc = ptlrpc_error(req);
2712 switch (req->rq_reqmsg->opc) {
2714 DEBUG_REQ(D_INODE, req, "connect");
2715 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2716 rc = target_handle_connect(req);
2718 struct mds_export_data *med;
2720 LASSERT(req->rq_export);
2721 med = &req->rq_export->u.eu_mds_data;
2722 mds_init_export_data(req, med);
2723 mds_req_add_idmapping(req, med);
2725 /* Now that we have an export, set mds. */
2726 obd = req->rq_export->exp_obd;
2727 mds = mds_req2mds(req);
2731 case MDS_DISCONNECT:
2732 DEBUG_REQ(D_INODE, req, "disconnect");
2733 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2734 rc = target_handle_disconnect(req);
2735 req->rq_status = rc; /* superfluous? */
2739 DEBUG_REQ(D_INODE, req, "getstatus");
2740 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2741 rc = mds_getstatus(req);
2745 DEBUG_REQ(D_INODE, req, "getattr");
2746 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2747 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2750 case MDS_GETATTR_LOCK: {
2751 struct lustre_handle lockh;
2752 DEBUG_REQ(D_INODE, req, "getattr_lock");
2753 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2755 /* If this request gets a reconstructed reply, we won't be
2756 * acquiring any new locks in mds_getattr_lock, so we don't
2760 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2761 MDS_INODELOCK_UPDATE);
2762 /* this non-intent call (from an ioctl) is special */
2763 req->rq_status = rc;
2764 if (rc == 0 && lockh.cookie)
2765 ldlm_lock_decref(&lockh, LCK_PR);
2769 DEBUG_REQ(D_INODE, req, "statfs");
2770 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2771 rc = mds_statfs(req);
2775 DEBUG_REQ(D_INODE, req, "readpage");
2776 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2777 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2779 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2780 if (req->rq_reply_state) {
2781 lustre_free_reply_state (req->rq_reply_state);
2782 req->rq_reply_state = NULL;
2789 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2792 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2793 mds->mds_max_cookiesize};
2796 /* NB only peek inside req now; mds_reint() will swab it */
2798 CERROR ("Can't inspect opcode\n");
2803 if (lustre_msg_swabbed (req->rq_reqmsg))
2806 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2807 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2808 reint_names[opc] == NULL) ? reint_names[opc] :
2811 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2813 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2815 else if (opc == REINT_OPEN)
2820 rc = lustre_pack_reply(req, bufcount, size, NULL);
2824 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
2825 fail = OBD_FAIL_MDS_REINT_NET_REP;
2830 DEBUG_REQ(D_INODE, req, "close");
2831 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
2832 rc = mds_close(req, MDS_REQ_REC_OFF);
2835 case MDS_DONE_WRITING:
2836 DEBUG_REQ(D_INODE, req, "done_writing");
2837 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
2838 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2842 DEBUG_REQ(D_INODE, req, "pin");
2843 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2844 rc = mds_pin(req, MDS_REQ_REC_OFF);
2848 DEBUG_REQ(D_INODE, req, "sync");
2849 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2850 rc = mds_sync(req, MDS_REQ_REC_OFF);
2854 DEBUG_REQ(D_INODE, req, "ping");
2855 rc = target_handle_ping(req);
2858 case OBD_LOG_CANCEL:
2859 CDEBUG(D_INODE, "log cancel\n");
2860 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2861 rc = -ENOTSUPP; /* la la la */
2865 DEBUG_REQ(D_INODE, req, "enqueue");
2866 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2867 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2868 ldlm_server_blocking_ast, NULL);
2869 fail = OBD_FAIL_LDLM_REPLY;
2872 DEBUG_REQ(D_INODE, req, "convert");
2873 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2874 rc = ldlm_handle_convert(req);
2876 case LDLM_BL_CALLBACK:
2877 case LDLM_CP_CALLBACK:
2878 DEBUG_REQ(D_INODE, req, "callback");
2879 CERROR("callbacks should not happen on MDS\n");
2881 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2883 case LLOG_ORIGIN_HANDLE_OPEN:
2884 DEBUG_REQ(D_INODE, req, "llog_init");
2885 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2886 rc = llog_origin_handle_open(req);
2888 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2889 DEBUG_REQ(D_INODE, req, "llog next block");
2890 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2891 rc = llog_origin_handle_next_block(req);
2893 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2894 DEBUG_REQ(D_INODE, req, "llog prev block");
2895 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2896 rc = llog_origin_handle_prev_block(req);
2898 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2899 DEBUG_REQ(D_INODE, req, "llog read header");
2900 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2901 rc = llog_origin_handle_read_header(req);
2903 case LLOG_ORIGIN_HANDLE_CLOSE:
2904 DEBUG_REQ(D_INODE, req, "llog close");
2905 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2906 rc = llog_origin_handle_close(req);
2909 DEBUG_REQ(D_INODE, req, "ost_create");
2910 rc = mdt_obj_create(req);
2913 DEBUG_REQ(D_INODE, req, "get_info");
2914 rc = mdt_get_info(req);
2917 DEBUG_REQ(D_INODE, req, "set_info");
2918 rc = mdt_set_info(req);
2921 CDEBUG(D_INODE, "write\n");
2922 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2923 rc = ost_brw_write(req, NULL);
2924 LASSERT(current->journal_info == NULL);
2925 /* mdt_brw sends its own replies */
2929 DEBUG_REQ(D_INODE, req, "llog catinfo");
2930 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2931 rc = llog_catinfo(req);
2934 req->rq_status = -ENOTSUPP;
2935 rc = ptlrpc_error(req);
2939 LASSERT(current->journal_info == NULL);
2943 /* If we're DISCONNECTing, the mds_export_data is already freed */
2944 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
2945 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2946 struct obd_device *obd = list_entry(mds, struct obd_device,
2948 req->rq_repmsg->last_xid =
2949 le64_to_cpu(med->med_mcd->mcd_last_xid);
2951 if (!obd->obd_no_transno) {
2952 req->rq_repmsg->last_committed =
2953 obd->obd_last_committed;
2955 DEBUG_REQ(D_IOCTL, req,
2956 "not sending last_committed update");
2958 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2960 mds->mds_last_transno, obd->obd_last_committed,
2966 target_send_reply(req, rc, fail);
2970 /* Update the server data on disk. This stores the new mount_count and also the
2971 * last_rcvd value to disk. If we don't have a clean shutdown, then the server
2972 * last_rcvd value may be less than that of the clients. This will alert us
2973 * that we may need to do client recovery.
2975 * Also assumes for mds_last_transno that we are not modifying it (no locking).
2977 int mds_update_server_data(struct obd_device *obd, int force_sync)
2979 struct mds_obd *mds = &obd->u.mds;
2980 struct mds_server_data *msd = mds->mds_server_data;
2981 struct file *filp = mds->mds_rcvd_filp;
2982 struct lvfs_run_ctxt saved;
2987 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2988 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2990 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2991 mds->mds_mount_count, mds->mds_last_transno);
2992 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
2994 CERROR("error writing MDS server data: rc = %d\n", rc);
2995 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3000 /* saves last allocated fid counter to file. */
3001 int mds_update_last_fid(struct obd_device *obd, void *handle,
3004 struct mds_obd *mds = &obd->u.mds;
3005 struct file *filp = mds->mds_fid_filp;
3006 struct lvfs_run_ctxt saved;
3012 spin_lock(&mds->mds_last_fid_lock);
3013 last_fid = mds->mds_last_fid;
3014 spin_unlock(&mds->mds_last_fid_lock);
3016 CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
3020 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
3021 handle, mds_commit_last_fid_cb,
3025 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3026 rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
3028 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3031 CERROR("error writing MDS last_fid #"LPU64
3032 ", err = %d\n", last_fid, rc);
3036 CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
3037 "%llu: err = %d\n", last_fid, off, rc);
3042 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
3044 struct mds_obd *mds = &obd->u.mds;
3046 spin_lock(&mds->mds_last_fid_lock);
3047 if (fid > mds->mds_last_fid)
3048 mds->mds_last_fid = fid;
3049 spin_unlock(&mds->mds_last_fid_lock);
3052 void mds_commit_last_transno_cb(struct obd_device *obd,
3053 __u64 transno, void *data,
3056 obd_transno_commit_cb(obd, transno, error);
3059 void mds_commit_last_fid_cb(struct obd_device *obd,
3060 __u64 fid, void *data,
3064 CERROR("%s: fid "LPD64" commit error: %d\n",
3065 obd->obd_name, fid, error);
3069 CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
3070 obd->obd_name, fid);
3073 __u64 mds_alloc_fid(struct obd_device *obd)
3075 struct mds_obd *mds = &obd->u.mds;
3078 spin_lock(&mds->mds_last_fid_lock);
3079 fid = ++mds->mds_last_fid;
3080 spin_unlock(&mds->mds_last_fid_lock);
3086 * allocates new lustre_id on passed @inode and saves it to inode EA.
3088 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
3089 void *handle, struct lustre_id *id)
3091 struct mds_obd *mds = &obd->u.mds;
3092 int alloc = 0, rc = 0;
3095 LASSERT(obd != NULL);
3096 LASSERT(inode != NULL);
3099 OBD_ALLOC(id, sizeof(*id));
3105 id_group(id) = mds->mds_num;
3106 id_fid(id) = mds_alloc_fid(obd);
3107 id_ino(id) = inode->i_ino;
3108 id_gen(id) = inode->i_generation;
3109 id_type(id) = (S_IFMT & inode->i_mode);
3111 rc = mds_update_inode_sid(obd, inode, handle, id);
3113 CERROR("Can't update inode FID EA, "
3118 OBD_FREE(id, sizeof(*id));
3123 * reads inode self id from inode EA. Probably later this should be replaced by
3124 * caching inode self id to avoid raeding it every time it is needed.
3126 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
3127 struct lustre_id *id)
3132 LASSERT(id != NULL);
3133 LASSERT(obd != NULL);
3134 LASSERT(inode != NULL);
3136 rc = fsfilt_get_md(obd, inode, &id->li_fid,
3137 sizeof(id->li_fid), EA_SID);
3139 CERROR("fsfilt_get_md() failed, "
3152 /* updates inode self id in EA. */
3153 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
3154 void *handle, struct lustre_id *id)
3159 LASSERT(id != NULL);
3160 LASSERT(obd != NULL);
3161 LASSERT(inode != NULL);
3163 rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3164 sizeof(id->li_fid), EA_SID);
3166 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
3174 * reads inode id on master MDS. This is usualy done by CMOBD to update requests
3175 * to master MDS by correct store cookie, needed to find inode on master MDS
3178 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
3179 struct lustre_id *id)
3184 LASSERT(id != NULL);
3185 LASSERT(obd != NULL);
3186 LASSERT(inode != NULL);
3188 rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_MID);
3190 CERROR("fsfilt_get_md() failed, rc = %d\n", rc);
3203 * updates master inode id. Usualy this is done by CMOBD after an inode is
3204 * created and relationship between cache MDS and master one should be
3207 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
3208 void *handle, struct lustre_id *id)
3213 LASSERT(id != NULL);
3214 LASSERT(obd != NULL);
3215 LASSERT(inode != NULL);
3217 rc = fsfilt_set_md(obd, inode, handle, id,
3218 sizeof(*id), EA_MID);
3220 CERROR("fsfilt_set_md() failed, "
3228 /* mount the file system (secretly) */
3229 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3231 struct lustre_cfg* lcfg = buf;
3232 struct mds_obd *mds = &obd->u.mds;
3233 struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3234 char *options = NULL;
3235 struct vfsmount *mnt;
3241 if (lcfg->lcfg_bufcount < 3)
3242 RETURN(rc = -EINVAL);
3244 if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
3245 RETURN(rc = -EINVAL);
3247 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
3248 if (IS_ERR(obd->obd_fsops))
3249 RETURN(rc = PTR_ERR(obd->obd_fsops));
3251 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3253 page = get_zeroed_page(GFP_KERNEL);
3257 options = (char *)page;
3260 * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3261 * and the rest of options are passed by mount options. Probably this
3262 * should be moved to somewhere else like startup scripts or lconf. */
3263 sprintf(options, "iopen_nopriv");
3265 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0 && lustre_cfg_buf(lcfg, 4))
3266 sprintf(options + strlen(options), ",%s",
3267 lustre_cfg_string(lcfg, 4));
3269 /* we have to know mdsnum before touching underlying fs -bzzz */
3270 atomic_set(&mds->mds_open_count, 0);
3271 sema_init(&mds->mds_md_sem, 1);
3272 sema_init(&mds->mds_create_sem, 1);
3273 mds->mds_md_connected = 0;
3274 mds->mds_md_name = NULL;
3276 if (LUSTRE_CFG_BUFLEN(lcfg, 5) > 0 && lustre_cfg_buf(lcfg, 5) &&
3277 strncmp(lustre_cfg_string(lcfg, 5), "dumb", LUSTRE_CFG_BUFLEN(lcfg, 5))) {
3280 generate_random_uuid(uuid);
3281 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3283 OBD_ALLOC(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3284 if (mds->mds_md_name == NULL)
3285 RETURN(rc = -ENOMEM);
3287 memcpy(mds->mds_md_name, lustre_cfg_buf(lcfg, 5),
3288 LUSTRE_CFG_BUFLEN(lcfg, 5));
3290 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3291 obd->obd_name, mds->mds_md_name);
3293 rc = mds_md_connect(obd, mds->mds_md_name);
3295 OBD_FREE(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3300 mds->mds_obd_type = MDS_MASTER_OBD;
3302 if (LUSTRE_CFG_BUFLEN(lcfg, 6) > 0 && lustre_cfg_buf(lcfg, 6) &&
3303 strncmp(lustre_cfg_string(lcfg, 6), "dumb",
3304 LUSTRE_CFG_BUFLEN(lcfg, 6))) {
3305 if (!memcmp(lustre_cfg_string(lcfg, 6), "master",
3306 strlen("master"))) {
3307 mds->mds_obd_type = MDS_MASTER_OBD;
3308 } else if (!memcmp(lustre_cfg_string(lcfg, 6), "cache",
3310 mds->mds_obd_type = MDS_CACHE_OBD;
3314 rc = lvfs_mount_fs(lustre_cfg_string(lcfg, 1),
3315 lustre_cfg_string(lcfg, 2),
3316 options, 0, &lvfs_ctxt);
3320 if (rc || !lvfs_ctxt) {
3321 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3325 mnt = lvfs_ctxt->loc_mnt;
3326 mds->mds_lvfs_ctxt = lvfs_ctxt;
3327 ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3329 CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
3331 sema_init(&mds->mds_epoch_sem, 1);
3332 atomic_set(&mds->mds_real_clients, 0);
3333 spin_lock_init(&mds->mds_transno_lock);
3334 spin_lock_init(&mds->mds_last_fid_lock);
3335 sema_init(&mds->mds_orphan_recovery_sem, 1);
3336 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3338 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3339 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3341 if (obd->obd_namespace == NULL) {
3342 mds_cleanup(obd, 0);
3343 GOTO(err_put, rc = -ENOMEM);
3345 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3347 rc = mds_fs_setup(obd, mnt);
3349 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3354 rc = llog_start_commit_thread();
3360 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 && lustre_cfg_buf(lcfg, 3) &&
3361 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3362 LUSTRE_CFG_BUFLEN(lcfg, 3))) {
3365 generate_random_uuid(uuid);
3366 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3368 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
3369 if (mds->mds_profile == NULL)
3370 GOTO(err_fs, rc = -ENOMEM);
3372 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
3373 LUSTRE_CFG_BUFLEN(lcfg, 3));
3377 * setup root dir and files ID dir if lmv already connected, or there is
3380 if (mds->mds_md_exp || (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 &&
3381 lustre_cfg_buf(lcfg, 3) &&
3382 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3383 LUSTRE_CFG_BUFLEN(lcfg, 3)))) {
3384 rc = mds_fs_setup_rootid(obd);
3388 rc = mds_fs_setup_virtid(obd);
3393 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3394 "mds_ldlm_client", &obd->obd_ldlm_client);
3395 obd->obd_replayable = 1;
3397 rc = mds_postsetup(obd);
3404 /* No extra cleanup needed for llog_init_commit_thread() */
3405 mds_fs_cleanup(obd, 0);
3407 ldlm_namespace_free(obd->obd_namespace, 0);
3408 obd->obd_namespace = NULL;
3411 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3415 fsfilt_put_ops(obd->obd_fsops);
3419 static int mds_postsetup(struct obd_device *obd)
3421 struct mds_obd *mds = &obd->u.mds;
3425 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3426 obd, 0, NULL, &llog_lvfs_ops);
3430 if (mds->mds_profile) {
3431 struct llog_ctxt *lgctxt;
3432 struct lvfs_run_ctxt saved;
3433 struct lustre_profile *lprof;
3434 struct config_llog_instance cfg;
3436 cfg.cfg_instance = NULL;
3437 cfg.cfg_uuid = mds->mds_dt_uuid;
3438 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3440 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3442 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3443 GOTO(err_llog, rc = -EINVAL);
3446 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3447 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3452 lprof = class_get_profile(mds->mds_profile);
3453 if (lprof == NULL) {
3454 CERROR("No profile found: %s\n", mds->mds_profile);
3455 GOTO(err_cleanup, rc = -ENOENT);
3457 rc = mds_dt_connect(obd, lprof->lp_lov);
3459 GOTO(err_cleanup, rc);
3461 rc = mds_md_postsetup(obd);
3463 GOTO(err_cleanup, rc);
3470 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3471 LLOG_CONFIG_ORIG_CTXT));
3475 int mds_postrecov_common(struct obd_device *obd)
3477 struct mds_obd *mds = &obd->u.mds;
3478 struct llog_ctxt *ctxt;
3479 int rc, item = 0, valsize;
3483 LASSERT(!obd->obd_recovering);
3484 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3485 LASSERT(ctxt != NULL);
3487 /* clean PENDING dir */
3488 rc = mds_cleanup_orphans(obd);
3493 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3494 valsize = sizeof(group);
3495 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
3496 "mds_conn", valsize, &group);
3500 rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3503 CERROR("%s: failed at llog_origin_connect: %d\n",
3508 /* remove the orphaned precreated objects */
3509 rc = mds_dt_clear_orphans(mds, NULL /* all OSTs */);
3514 RETURN(rc < 0 ? rc : item);
3517 /* cleanup all llogging subsystems */
3518 rc = obd_llog_finish(obd, &obd->obd_llogs,
3519 mds->mds_dt_desc.ld_tgt_count);
3521 CERROR("%s: failed to cleanup llogging subsystems\n",
3526 int mds_postrecov(struct obd_device *obd)
3530 rc = mds_postrecov_common(obd);
3532 rc = mds_md_reconnect(obd);
3536 int mds_dt_clean(struct obd_device *obd)
3538 struct mds_obd *mds = &obd->u.mds;
3541 if (mds->mds_profile) {
3543 struct llog_ctxt *llctx;
3544 struct lvfs_run_ctxt saved;
3545 struct config_llog_instance cfg;
3546 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3548 OBD_ALLOC(cln_prof, len);
3549 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3551 cfg.cfg_instance = NULL;
3552 cfg.cfg_uuid = mds->mds_dt_uuid;
3554 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3555 llctx = llog_get_context(&obd->obd_llogs,
3556 LLOG_CONFIG_ORIG_CTXT);
3557 class_config_process_llog(llctx, cln_prof, &cfg);
3558 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3560 OBD_FREE(cln_prof, len);
3561 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3562 mds->mds_profile = NULL;
3567 int mds_md_clean(struct obd_device *obd)
3569 struct mds_obd *mds = &obd->u.mds;
3572 if (mds->mds_md_name) {
3573 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3574 mds->mds_md_name = NULL;
3579 static int mds_precleanup(struct obd_device *obd, int flags)
3585 mds_dt_disconnect(obd, flags);
3587 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3591 extern void lgss_svc_cache_purge_all(void);
3592 static int mds_cleanup(struct obd_device *obd, int flags)
3594 struct mds_obd *mds = &obd->u.mds;
3597 if (mds->mds_sb == NULL)
3600 mds_update_server_data(obd, 1);
3601 mds_update_last_fid(obd, NULL, 1);
3603 if (mds->mds_dt_objids != NULL) {
3604 int size = mds->mds_dt_desc.ld_tgt_count *
3606 OBD_FREE(mds->mds_dt_objids, size);
3608 mds_fs_cleanup(obd, flags);
3612 /* 2 seems normal on mds, (may_umount() also expects 2
3613 fwiw), but we only see 1 at this point in obdfilter. */
3614 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3618 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3620 spin_lock_bh(&obd->obd_processing_task_lock);
3621 if (obd->obd_recovering) {
3622 target_cancel_recovery_timer(obd);
3623 obd->obd_recovering = 0;
3625 spin_unlock_bh(&obd->obd_processing_task_lock);
3628 fsfilt_put_ops(obd->obd_fsops);
3632 lgss_svc_cache_purge_all();
3635 spin_lock(&mds->mds_denylist_lock);
3636 while (!list_empty( &mds->mds_denylist ) ) {
3637 deny_sec_t *p_deny_sec = list_entry(mds->mds_denylist.next,
3639 list_del(&p_deny_sec->list);
3640 OBD_FREE(p_deny_sec, sizeof(*p_deny_sec));
3642 spin_unlock(&mds->mds_denylist_lock);
3643 if(mds->mds_mds_sec)
3644 OBD_FREE(mds->mds_mds_sec, strlen(mds->mds_mds_sec) + 1);
3645 if(mds->mds_ost_sec)
3646 OBD_FREE(mds->mds_ost_sec, strlen(mds->mds_ost_sec) + 1);
3651 static int set_security(const char *value, char **sec)
3655 if (!strcmp(value, "null") ||
3656 !strcmp(value, "krb5i") ||
3657 !strcmp(value, "krb5p")) {
3658 OBD_ALLOC(*sec, strlen(value) + 1);
3661 memcpy(*sec, value, strlen(value) + 1);
3663 CERROR("Unrecognized value, force use NULL\n");
3670 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
3672 struct lustre_cfg *lcfg = buf;
3673 struct mds_obd *mds = &obd->u.mds;
3677 switch(lcfg->lcfg_command) {
3678 case LCFG_SET_SECURITY: {
3679 if ((LUSTRE_CFG_BUFLEN(lcfg, 1) == 0) ||
3680 (LUSTRE_CFG_BUFLEN(lcfg, 2) == 0))
3681 GOTO(out, rc = -EINVAL);
3683 if (!strcmp(lustre_cfg_string(lcfg, 1), "mds_sec"))
3684 rc = set_security(lustre_cfg_string(lcfg, 2),
3686 else if (!strcmp(lustre_cfg_string(lcfg, 1), "oss_sec"))
3687 rc = set_security(lustre_cfg_string(lcfg, 2),
3689 else if (!strcmp(lustre_cfg_string(lcfg, 1), "deny_sec")){
3690 spin_lock(&mds->mds_denylist_lock);
3691 rc = add_deny_security(lustre_cfg_string(lcfg, 2),
3692 &mds->mds_denylist);
3693 spin_unlock(&mds->mds_denylist_lock);
3695 CERROR("Unrecognized key\n");
3701 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
3702 GOTO(out, rc = -EINVAL);
3710 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3712 struct ldlm_lock *new_lock,
3713 struct ldlm_lock **old_lock,
3714 struct lustre_handle *lockh)
3716 struct obd_export *exp = req->rq_export;
3717 struct obd_device *obd = exp->exp_obd;
3718 struct ldlm_request *dlmreq =
3719 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3720 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3721 struct list_head *iter;
3723 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3726 l_lock(&obd->obd_namespace->ns_lock);
3727 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3728 struct ldlm_lock *lock;
3729 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3730 if (lock == new_lock)
3732 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3733 lockh->cookie = lock->l_handle.h_cookie;
3734 LDLM_DEBUG(lock, "restoring lock cookie");
3735 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3738 *old_lock = LDLM_LOCK_GET(lock);
3739 l_unlock(&obd->obd_namespace->ns_lock);
3743 l_unlock(&obd->obd_namespace->ns_lock);
3745 /* If the xid matches, then we know this is a resent request,
3746 * and allow it. (It's probably an OPEN, for which we don't
3749 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
3753 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
3756 /* This remote handle isn't enqueued, so we never received or
3757 * processed this request. Clear MSG_RESENT, because it can
3758 * be handled like any normal request now. */
3760 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3762 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3766 int intent_disposition(struct ldlm_reply *rep, int flag)
3770 return (rep->lock_policy_res1 & flag);
3773 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3777 rep->lock_policy_res1 |= flag;
3780 static int mds_intent_policy(struct ldlm_namespace *ns,
3781 struct ldlm_lock **lockp, void *req_cookie,
3782 ldlm_mode_t mode, int flags, void *data)
3784 struct ptlrpc_request *req = req_cookie;
3785 struct ldlm_lock *lock = *lockp;
3786 struct ldlm_intent *it;
3787 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3788 struct ldlm_reply *rep;
3789 struct lustre_handle lockh[2] = {{0}, {0}};
3790 struct ldlm_lock *new_lock = NULL;
3791 int getattr_part = MDS_INODELOCK_UPDATE;
3792 int rc, reply_buffers;
3793 int repsize[5] = {sizeof(struct ldlm_reply),
3794 sizeof(struct mds_body),
3795 mds->mds_max_mdsize};
3797 int offset = MDS_REQ_INTENT_REC_OFF;
3800 LASSERT(req != NULL);
3801 MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
3803 if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3804 /* No intent was provided */
3805 int size = sizeof(struct ldlm_reply);
3806 rc = lustre_pack_reply(req, 1, &size, NULL);
3811 it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3812 lustre_swab_ldlm_intent);
3814 CERROR("Intent missing\n");
3815 RETURN(req->rq_status = -EFAULT);
3818 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3821 if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
3824 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
3827 rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
3829 RETURN(req->rq_status = rc);
3831 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
3832 LASSERT(rep != NULL);
3834 intent_set_disposition(rep, DISP_IT_EXECD);
3836 /* execute policy */
3837 switch ((long)it->opc) {
3839 case IT_CREAT|IT_OPEN:
3840 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3842 /* XXX swab here to assert that an mds_open reint
3843 * packet is following */
3844 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3846 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
3848 if (rep->lock_policy_res2) {
3850 * mds_open() returns ENOLCK where it should return
3851 * zero, but it has no lock to return.
3853 if (rep->lock_policy_res2 == ENOLCK)
3854 rep->lock_policy_res2 = 0;
3856 RETURN(ELDLM_LOCK_ABORTED);
3860 * IT_OPEN may return lock on cross-node dentry that we want to
3861 * hold during attr retrival -bzzz
3863 if (lockh[0].cookie == 0)
3864 RETURN(ELDLM_LOCK_ABORTED);
3868 getattr_part = MDS_INODELOCK_LOOKUP;
3871 getattr_part |= MDS_INODELOCK_LOOKUP;
3873 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3874 lock, &new_lock, lockh);
3875 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
3877 /* FIXME: LDLM can set req->rq_status. MDS sets
3878 policy_res{1,2} with disposition and status.
3879 - replay: returns 0 & req->status is old status
3880 - otherwise: returns req->status */
3881 if (intent_disposition(rep, DISP_LOOKUP_NEG))
3882 rep->lock_policy_res2 = 0;
3883 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
3884 rep->lock_policy_res2)
3885 RETURN(ELDLM_LOCK_ABORTED);
3886 if (req->rq_status != 0) {
3888 rep->lock_policy_res2 = req->rq_status;
3889 RETURN(ELDLM_LOCK_ABORTED);
3893 rc = mds_lock_and_check_slave(offset, req, lockh);
3894 if ((rep->lock_policy_res2 = rc)) {
3896 rep->lock_policy_res2 = 0;
3897 RETURN(ELDLM_LOCK_ABORTED);
3901 CERROR("Unhandled intent "LPD64"\n", it->opc);
3905 /* By this point, whatever function we called above must have either
3906 * filled in 'lockh', been an intent replay, or returned an error. We
3907 * want to allow replayed RPCs to not get a lock, since we would just
3908 * drop it below anyways because lock replay is done separately by the
3909 * client afterwards. For regular RPCs we want to give the new lock to
3910 * the client instead of whatever lock it was about to get. */
3911 if (new_lock == NULL)
3912 new_lock = ldlm_handle2lock(&lockh[0]);
3913 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
3916 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
3917 it->opc, lockh[0].cookie);
3919 /* If we've already given this lock to a client once, then we should
3920 * have no readers or writers. Otherwise, we should have one reader
3921 * _or_ writer ref (which will be zeroed below) before returning the
3922 * lock to a client. */
3923 if (new_lock->l_export == req->rq_export) {
3924 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3926 LASSERT(new_lock->l_export == NULL);
3927 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3932 if (new_lock->l_export == req->rq_export) {
3933 /* Already gave this to the client, which means that we
3934 * reconstructed a reply. */
3935 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3937 RETURN(ELDLM_LOCK_REPLACED);
3940 /* Fixup the lock to be given to the client */
3941 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
3942 new_lock->l_readers = 0;
3943 new_lock->l_writers = 0;
3945 new_lock->l_export = class_export_get(req->rq_export);
3946 list_add(&new_lock->l_export_chain,
3947 &new_lock->l_export->exp_ldlm_data.led_held_locks);
3949 new_lock->l_blocking_ast = lock->l_blocking_ast;
3950 new_lock->l_completion_ast = lock->l_completion_ast;
3952 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
3953 sizeof(lock->l_remote_handle));
3955 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3957 LDLM_LOCK_PUT(new_lock);
3958 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
3960 RETURN(ELDLM_LOCK_REPLACED);
3963 int mds_attach(struct obd_device *dev, obd_count len, void *data)
3965 struct lprocfs_static_vars lvars;
3967 struct mds_obd *mds = &dev->u.mds;
3969 spin_lock_init(&mds->mds_denylist_lock);
3970 INIT_LIST_HEAD(&mds->mds_denylist);
3972 lprocfs_init_multi_vars(0, &lvars);
3974 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3978 return lprocfs_alloc_md_stats(dev, 0);
3981 int mds_detach(struct obd_device *dev)
3983 lprocfs_free_md_stats(dev);
3984 return lprocfs_obd_detach(dev);
3987 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
3989 struct lprocfs_static_vars lvars;
3991 lprocfs_init_multi_vars(1, &lvars);
3992 return lprocfs_obd_attach(dev, lvars.obd_vars);
3995 int mdt_detach(struct obd_device *dev)
3997 return lprocfs_obd_detach(dev);
4000 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
4002 struct mds_obd *mds = &obd->u.mds;
4007 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4008 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
4009 MDS_SERVICE_WATCHDOG_TIMEOUT,
4010 mds_handle, "mds", obd->obd_proc_entry);
4012 if (!mds->mds_service) {
4013 CERROR("failed to start service\n");
4017 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
4020 GOTO(err_thread, rc);
4022 mds->mds_setattr_service =
4023 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4024 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
4025 MDS_SERVICE_WATCHDOG_TIMEOUT,
4026 mds_handle, "mds_setattr",
4027 obd->obd_proc_entry);
4028 if (!mds->mds_setattr_service) {
4029 CERROR("failed to start getattr service\n");
4030 GOTO(err_thread, rc = -ENOMEM);
4033 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
4034 MDT_NUM_THREADS, "ll_mdt_attr");
4036 GOTO(err_thread2, rc);
4038 mds->mds_readpage_service =
4039 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4040 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
4041 MDS_SERVICE_WATCHDOG_TIMEOUT,
4042 mds_handle, "mds_readpage",
4043 obd->obd_proc_entry);
4044 if (!mds->mds_readpage_service) {
4045 CERROR("failed to start readpage service\n");
4046 GOTO(err_thread2, rc = -ENOMEM);
4049 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
4050 MDT_NUM_THREADS, "ll_mdt_rdpg");
4053 GOTO(err_thread3, rc);
4058 ptlrpc_unregister_service(mds->mds_readpage_service);
4060 ptlrpc_unregister_service(mds->mds_setattr_service);
4062 ptlrpc_unregister_service(mds->mds_service);
4066 static int mdt_cleanup(struct obd_device *obd, int flags)
4068 struct mds_obd *mds = &obd->u.mds;
4071 ptlrpc_stop_all_threads(mds->mds_readpage_service);
4072 ptlrpc_unregister_service(mds->mds_readpage_service);
4074 ptlrpc_stop_all_threads(mds->mds_setattr_service);
4075 ptlrpc_unregister_service(mds->mds_setattr_service);
4077 ptlrpc_stop_all_threads(mds->mds_service);
4078 ptlrpc_unregister_service(mds->mds_service);
4083 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
4084 __u64 gr, void *data)
4086 struct lustre_id id;
4087 struct obd_device *obd = data;
4091 return mds_id2dentry(obd, &id, NULL);
4094 static int mds_get_info(struct obd_export *exp, __u32 keylen,
4095 void *key, __u32 *valsize, void *val)
4097 struct obd_device *obd;
4098 struct mds_obd *mds;
4101 obd = class_exp2obd(exp);
4105 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
4106 exp->exp_handle.h_cookie);
4110 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
4111 /* get log_context handle. */
4112 unsigned long *llh_handle = val;
4113 *valsize = sizeof(unsigned long);
4114 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
4117 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
4118 /* get log_context handle. */
4119 unsigned long *sb = val;
4120 *valsize = sizeof(unsigned long);
4121 *sb = (unsigned long)obd->u.mds.mds_sb;
4125 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
4126 __u32 *mdsize = val;
4127 *valsize = sizeof(*mdsize);
4128 *mdsize = mds->mds_max_mdsize;
4132 if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
4133 __u32 *mdsnum = val;
4134 *valsize = sizeof(*mdsnum);
4135 *mdsnum = mds->mds_num;
4139 if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
4140 struct lustre_id *rootid = val;
4141 *valsize = sizeof(struct lustre_id);
4142 *rootid = mds->mds_rootid;
4146 CDEBUG(D_IOCTL, "invalid key\n");
4150 struct lvfs_callback_ops mds_lvfs_ops = {
4151 l_id2dentry: mds_lvfs_id2dentry,
4154 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
4155 int objcount, struct obd_ioobj *obj,
4156 int niocount, struct niobuf_remote *nb,
4157 struct niobuf_local *res,
4158 struct obd_trans_info *oti);
4160 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
4161 int objcount, struct obd_ioobj *obj, int niocount,
4162 struct niobuf_local *res, struct obd_trans_info *oti,
4165 /* use obd ops to offer management infrastructure */
4166 static struct obd_ops mds_obd_ops = {
4167 .o_owner = THIS_MODULE,
4168 .o_attach = mds_attach,
4169 .o_detach = mds_detach,
4170 .o_connect = mds_connect,
4171 .o_connect_post = mds_connect_post,
4172 .o_init_export = mds_init_export,
4173 .o_destroy_export = mds_destroy_export,
4174 .o_disconnect = mds_disconnect,
4175 .o_setup = mds_setup,
4176 .o_precleanup = mds_precleanup,
4177 .o_cleanup = mds_cleanup,
4178 .o_process_config = mds_process_config,
4179 .o_postrecov = mds_postrecov,
4180 .o_statfs = mds_obd_statfs,
4181 .o_iocontrol = mds_iocontrol,
4182 .o_create = mds_obd_create,
4183 .o_destroy = mds_obd_destroy,
4184 .o_llog_init = mds_llog_init,
4185 .o_llog_finish = mds_llog_finish,
4186 .o_notify = mds_notify,
4187 .o_get_info = mds_get_info,
4188 .o_set_info = mds_set_info,
4189 .o_preprw = mds_preprw,
4190 .o_commitrw = mds_commitrw,
4193 static struct obd_ops mdt_obd_ops = {
4194 .o_owner = THIS_MODULE,
4195 .o_attach = mdt_attach,
4196 .o_detach = mdt_detach,
4197 .o_setup = mdt_setup,
4198 .o_cleanup = mdt_cleanup,
4201 static int __init mds_init(void)
4203 struct lprocfs_static_vars lvars;
4205 mds_init_lsd_cache();
4207 lprocfs_init_multi_vars(0, &lvars);
4208 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4210 lprocfs_init_multi_vars(1, &lvars);
4211 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4217 static void /*__exit*/ mds_exit(void)
4219 mds_cleanup_lsd_cache();
4221 class_unregister_type(LUSTRE_MDS_NAME);
4222 class_unregister_type(LUSTRE_MDT_NAME);
4225 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4226 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4227 MODULE_LICENSE("GPL");
4229 module_init(mds_init);
4230 module_exit(mds_exit);