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} };
246 int flags = LDLM_FL_ATOMIC_CB, rc;
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));
274 flags = LDLM_FL_ATOMIC_CB;
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);
295 } else if (de->d_inode && de->d_inode->i_nlink == 0) {
296 /* as sometimes we lookup inode by ino/generation through
297 iopen mechanism, it's possible to find already unlinked
298 inode with nlink == 0. let's interpretate the case as
300 CWARN("found already unlinked inode %lu/%u\n",
301 de->d_inode->i_ino, de->d_inode->i_generation);
303 retval = ERR_PTR(-ENOENT);
304 ldlm_lock_decref(lockh, lock_mode);
307 ldlm_lock_decref(lockh + 1, *mode);
314 #ifndef DCACHE_DISCONNECTED
315 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
318 /* Look up an entry by inode number. This function ONLY returns valid dget'd
319 * dentries with an initialized inode or errors */
320 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
321 struct vfsmount **mnt)
323 unsigned long ino = (unsigned long)id_ino(id);
324 __u32 generation = (__u32)id_gen(id);
325 struct mds_obd *mds = &obd->u.mds;
326 struct dentry *result;
331 RETURN(ERR_PTR(-ESTALE));
333 snprintf(idname, sizeof(idname), "0x%lx", ino);
335 CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
336 ino, generation, mds->mds_sb);
338 /* under ext3 this is neither supposed to return bad inodes nor NULL
340 result = ll_lookup_one_len(idname, mds->mds_id_de,
345 inode = result->d_inode;
347 RETURN(ERR_PTR(-ENOENT));
349 if (is_bad_inode(inode)) {
350 CERROR("bad inode returned %lu/%u\n",
351 inode->i_ino, inode->i_generation);
353 RETURN(ERR_PTR(-ENOENT));
356 /* here we disabled generation check, as root inode i_generation
357 * of cache mds and real mds are different. */
358 if (inode->i_ino != id_ino(&mds->mds_rootid) && generation &&
359 inode->i_generation != generation) {
360 /* we didn't find the right inode.. */
361 if (id_group(id) != mds->mds_num) {
362 CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
363 "%u != %u, mds %u != %u, request to wrong MDS?\n",
364 inode->i_ino, (unsigned long)inode->i_nlink,
365 atomic_read(&inode->i_count), inode->i_generation,
366 generation, mds->mds_num, (unsigned)id_group(id));
368 CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
369 "%u != %u, inode is recreated while request handled?\n",
370 inode->i_ino, (unsigned long)inode->i_nlink,
371 atomic_read(&inode->i_count), inode->i_generation,
375 RETURN(ERR_PTR(-ENOENT));
379 *mnt = mds->mds_vfsmnt;
387 int mds_req_add_idmapping(struct ptlrpc_request *req,
388 struct mds_export_data *med)
390 struct mds_req_sec_desc *rsd;
391 struct lustre_sec_desc *lsd;
394 if (!med->med_remote)
397 /* maybe we should do it more completely: invalidate the gss ctxt? */
398 if (req->rq_mapped_uid == MDS_IDMAP_NOTFOUND) {
399 CWARN("didn't find mapped uid\n");
403 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
405 CERROR("Can't unpack security desc\n");
409 lsd = mds_get_lsd(req->rq_mapped_uid);
411 CERROR("can't get LSD(%u), no mapping added\n",
416 rc = mds_idmap_add(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
417 rsd->rsd_gid, lsd->lsd_gid);
423 int mds_req_del_idmapping(struct ptlrpc_request *req,
424 struct mds_export_data *med)
426 struct mds_req_sec_desc *rsd;
427 struct lustre_sec_desc *lsd;
430 if (!med->med_remote)
433 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
435 CERROR("Can't unpack security desc\n");
439 LASSERT(req->rq_mapped_uid != -1);
440 lsd = mds_get_lsd(req->rq_mapped_uid);
442 CERROR("can't get LSD(%u), no idmapping deleted\n",
447 rc = mds_idmap_del(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
448 rsd->rsd_gid, lsd->lsd_gid);
453 static int mds_init_export_data(struct ptlrpc_request *req,
454 struct mds_export_data *med)
456 struct obd_connect_data *data, *reply;
457 int ask_remote, ask_local;
460 data = lustre_msg_buf(req->rq_reqmsg, 5, sizeof(*data));
461 reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*data));
462 LASSERT(data && reply);
464 if (med->med_initialized) {
465 CDEBUG(D_SEC, "med already initialized, reconnect?\n");
469 ask_remote = data->ocd_connect_flags & OBD_CONNECT_REMOTE;
470 ask_local = data->ocd_connect_flags & OBD_CONNECT_LOCAL;
472 /* currently the policy is simple: satisfy client as possible
475 if (req->rq_auth_uid == -1) {
477 CWARN("null sec is used, force to be local\n");
481 if (!req->rq_remote_realm)
482 CWARN("local realm asked to be remote\n");
484 } else if (ask_local) {
485 if (req->rq_remote_realm)
486 CWARN("remote realm asked to be local\n");
489 med->med_remote = (req->rq_remote_realm != 0);
492 med->med_nllu = data->ocd_nllu[0];
493 med->med_nllg = data->ocd_nllu[1];
495 med->med_initialized = 1;
497 reply->ocd_connect_flags &= ~(OBD_CONNECT_REMOTE | OBD_CONNECT_LOCAL);
498 if (med->med_remote) {
500 med->med_idmap = mds_idmap_alloc();
503 CERROR("Failed to alloc idmap, following request from "
504 "this client will be refused\n");
506 reply->ocd_connect_flags |= OBD_CONNECT_REMOTE;
507 CDEBUG(D_SEC, "set client as remote\n");
509 reply->ocd_connect_flags |= OBD_CONNECT_LOCAL;
510 CDEBUG(D_SEC, "set client as local\n");
516 static void mds_free_export_data(struct mds_export_data *med)
521 LASSERT(med->med_remote);
522 mds_idmap_free(med->med_idmap);
523 med->med_idmap = NULL;
526 /* Establish a connection to the MDS.
528 * This will set up an export structure for the client to hold state data about
529 * that client, like open files, the last operation number it did on the server,
532 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
533 struct obd_uuid *cluuid, struct obd_connect_data *data,
536 struct mds_export_data *med;
537 struct mds_client_data *mcd;
538 struct obd_export *exp;
542 if (!conn || !obd || !cluuid)
545 /* XXX There is a small race between checking the list and adding a new
546 * connection for the same UUID, but the real threat (list corruption
547 * when multiple different clients connect) is solved.
549 * There is a second race between adding the export to the list, and
550 * filling in the client data below. Hence skipping the case of NULL
551 * mcd above. We should already be controlling multiple connects at the
552 * client, and we can't hold the spinlock over memory allocations
553 * without risk of deadlocking.
555 rc = class_connect(conn, obd, cluuid);
558 exp = class_conn2export(conn);
560 LASSERT(exp != NULL);
561 med = &exp->exp_mds_data;
563 OBD_ALLOC(mcd, sizeof(*mcd));
565 CERROR("%s: out of memory for client data.\n",
567 GOTO(out, rc = -ENOMEM);
570 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
573 rc = mds_client_add(obd, &obd->u.mds, med, -1);
581 OBD_FREE(mcd, sizeof(*mcd));
582 class_disconnect(exp, 0);
584 class_export_put(exp);
589 static int mds_connect_post(struct obd_export *exp, unsigned initial,
592 struct obd_device *obd = exp->exp_obd;
593 struct mds_obd *mds = &obd->u.mds;
594 struct mds_export_data *med;
595 struct mds_client_data *mcd;
599 med = &exp->exp_mds_data;
603 /* some one reconnect initially, we have to reset
604 * data existing export can have. bug 6102 */
605 if (mcd->mcd_last_xid != 0)
606 CDEBUG(D_HA, "initial reconnect to existing export\n");
607 mcd->mcd_last_transno = 0;
608 mcd->mcd_last_xid = 0;
609 mcd->mcd_last_result = 0;
610 mcd->mcd_last_data = 0;
613 if (!(flags & OBD_OPT_MDS_CONNECTION)) {
614 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
615 atomic_inc(&mds->mds_real_clients);
616 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
617 obd->obd_name, exp->exp_client_uuid.uuid,
618 atomic_read(&mds->mds_real_clients));
619 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
621 if (mds->mds_md_name)
622 rc = mds_md_connect(obd, mds->mds_md_name);
627 static int mds_init_export(struct obd_export *exp)
629 struct mds_export_data *med = &exp->exp_mds_data;
631 INIT_LIST_HEAD(&med->med_open_head);
632 spin_lock_init(&med->med_open_lock);
636 static int mds_destroy_export(struct obd_export *export)
638 struct obd_device *obd = export->exp_obd;
639 struct mds_export_data *med = &export->exp_mds_data;
640 struct lvfs_run_ctxt saved;
644 mds_free_export_data(med);
645 target_destroy_export(export);
647 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
650 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
652 /* Close any open files (which may also cause orphan unlinking). */
653 spin_lock(&med->med_open_lock);
654 while (!list_empty(&med->med_open_head)) {
655 struct list_head *tmp = med->med_open_head.next;
656 struct mds_file_data *mfd =
657 list_entry(tmp, struct mds_file_data, mfd_list);
658 struct lustre_id sid;
660 BDEVNAME_DECLARE_STORAGE(btmp);
662 /* bug 1579: fix force-closing for 2.5 */
663 struct dentry *dentry = mfd->mfd_dentry;
665 list_del(&mfd->mfd_list);
666 spin_unlock(&med->med_open_lock);
668 down(&dentry->d_inode->i_sem);
669 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
670 up(&dentry->d_inode->i_sem);
672 CERROR("Can't read inode self id, inode %lu, "
673 "rc %d\n", dentry->d_inode->i_ino, rc);
674 memset(&sid, 0, sizeof(sid));
677 /* If you change this message, be sure to update
678 * replay_single:test_46 */
679 CERROR("force closing client file handle for %.*s (%s:"
680 DLID4")\n", dentry->d_name.len, dentry->d_name.name,
681 ll_bdevname(dentry->d_inode->i_sb, btmp),
684 /* child inode->i_alloc_sem protects orphan_dec_test and
685 * is_orphan race, mds_mfd_close drops it */
686 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
687 rc = mds_mfd_close(NULL, 0, obd, mfd,
688 !(export->exp_flags & OBD_OPT_FAILOVER));
690 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
691 spin_lock(&med->med_open_lock);
693 spin_unlock(&med->med_open_lock);
694 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
698 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
702 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
704 unsigned long irqflags;
705 struct obd_device *obd;
710 LASSERT(exp != NULL);
711 obd = class_exp2obd(exp);
713 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
714 exp->exp_handle.h_cookie);
720 * suppress any inter-mds requests durring disconnecting lmv if this is
721 * detected --force mode. This is needed to avoid endless recovery.
723 if (atomic_read(&mds->mds_real_clients) > 0 &&
724 !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
725 flags |= OBD_OPT_FORCE;
727 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
728 && !atomic_read(&mds->mds_real_clients)) {
729 /* there was no client at all */
730 mds_md_disconnect(obd, flags);
733 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
734 && atomic_dec_and_test(&mds->mds_real_clients)) {
735 /* time to drop LMV connections */
736 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
737 "Disconnnect from LMV now\n",
738 obd->obd_name, exp->exp_client_uuid.uuid);
739 mds_md_disconnect(obd, flags);
742 spin_lock_irqsave(&exp->exp_lock, irqflags);
743 exp->exp_flags = flags;
744 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
746 /* disconnect early so that clients can't keep using export */
747 rc = class_disconnect(exp, flags);
748 ldlm_cancel_locks_for_export(exp);
750 /* complete all outstanding replies */
751 spin_lock_irqsave(&exp->exp_lock, irqflags);
752 while (!list_empty(&exp->exp_outstanding_replies)) {
753 struct ptlrpc_reply_state *rs =
754 list_entry(exp->exp_outstanding_replies.next,
755 struct ptlrpc_reply_state, rs_exp_list);
756 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
758 spin_lock(&svc->srv_lock);
759 list_del_init(&rs->rs_exp_list);
760 ptlrpc_schedule_difficult_reply(rs);
761 spin_unlock(&svc->srv_lock);
763 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
767 static int mds_getstatus(struct ptlrpc_request *req)
769 struct mds_obd *mds = mds_req2mds(req);
770 struct mds_body *body;
774 size = sizeof(*body);
776 rc = lustre_pack_reply(req, 1, &size, NULL);
777 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
778 CERROR("mds: out of memory for message: size=%d\n", size);
779 req->rq_status = -ENOMEM; /* superfluous? */
783 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
784 body->valid |= OBD_MD_FID;
786 memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
789 * the last_committed and last_xid fields are filled in for all replies
790 * already - no need to do so here also.
795 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
796 void *data, int flag)
801 if (flag == LDLM_CB_CANCELING) {
802 /* Don't need to do anything here. */
806 /* XXX layering violation! -phil */
807 lock_res(lock->l_resource);
810 * get this: if mds_blocking_ast is racing with mds_intent_policy, such
811 * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
812 * then by the time we get the lock, we might not be the correct
813 * blocking function anymore. So check, and return early, if so.
815 if (lock->l_blocking_ast != mds_blocking_ast) {
816 unlock_res(lock->l_resource);
820 lock->l_flags |= LDLM_FL_CBPENDING;
821 do_ast = (!lock->l_readers && !lock->l_writers);
822 unlock_res(lock->l_resource);
825 struct lustre_handle lockh;
828 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
829 ldlm_lock2handle(lock, &lockh);
830 rc = ldlm_cli_cancel(&lockh);
832 CERROR("ldlm_cli_cancel: %d\n", rc);
834 LDLM_DEBUG(lock, "Lock still has references, will be "
840 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
841 void *md, int size, int mea)
845 if (S_ISREG(inode->i_mode)) {
846 rc = mds_convert_lov_ea(obd, inode, md, size);
847 } else if (S_ISDIR(inode->i_mode)) {
849 rc = mds_convert_mea_ea(obd, inode, md, size);
851 rc = mds_convert_lov_ea(obd, inode, md, size);
854 CERROR("Invalid EA format (nor LOV or MEA) "
855 "is detected. Inode %lu/%u\n",
856 inode->i_ino, inode->i_generation);
862 int mds_get_md(struct obd_device *obd, struct inode *inode,
863 void *md, int *size, int lock, int mea)
872 rc = fsfilt_get_md(obd, inode, md, *size,
873 (mea ? EA_MEA : EA_LOV));
875 CERROR("Error %d reading eadata for ino %lu\n",
879 rc = mds_convert_md(obd, inode, md,
894 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
895 * Call with lock=0 if the caller has already taken the i_sem. */
896 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
897 struct mds_body *body, struct inode *inode, int lock, int mea)
899 struct mds_obd *mds = &obd->u.mds;
904 lmm = lustre_msg_buf(msg, offset, 0);
906 /* Some problem with getting eadata when I sized the reply
908 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
912 lmm_size = msg->buflens[offset];
914 /* I don't really like this, but it is a sanity check on the client
915 * MD request. However, if the client doesn't know how much space
916 * to reserve for the MD, it shouldn't be bad to have too much space.
918 if (lmm_size > mds->mds_max_mdsize) {
919 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
920 inode->i_ino, lmm_size, mds->mds_max_mdsize);
924 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
926 body->valid |= S_ISDIR(inode->i_mode) ?
927 OBD_MD_FLDIREA : OBD_MD_FLEASIZE;
930 body->valid |= OBD_MD_MEA;
932 body->eadatasize = lmm_size;
939 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
940 struct mds_body *repbody, int reply_off)
942 struct inode *inode = dentry->d_inode;
947 symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
948 LASSERT(symname != NULL);
949 len = req->rq_repmsg->buflens[reply_off + 1];
951 rc = inode->i_op->readlink(dentry, symname, len);
953 CERROR("readlink failed: %d\n", rc);
954 } else if (rc != len - 1) {
955 CERROR ("Unexpected readlink rc %d: expecting %d\n",
959 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
960 repbody->valid |= OBD_MD_LINKNAME;
961 repbody->eadatasize = rc + 1;
962 symname[rc] = 0; /* NULL terminate */
969 int mds_pack_ea(struct dentry *dentry, struct ptlrpc_request *req,
970 struct mds_body *repbody, int req_off, int reply_off)
972 struct inode *inode = dentry->d_inode;
978 ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
979 len = req->rq_repmsg->buflens[reply_off + 1];
981 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
984 if (inode->i_op && inode->i_op->getxattr)
985 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
988 if (rc != -ENODATA && rc != -EOPNOTSUPP)
989 CERROR("getxattr failed: %d", rc);
991 repbody->valid |= OBD_MD_FLEA;
992 repbody->eadatasize = rc;
999 int mds_pack_ealist(struct dentry *dentry, struct ptlrpc_request *req,
1000 struct mds_body *repbody, int reply_off)
1002 struct inode *inode = dentry->d_inode;
1007 len = req->rq_repmsg->buflens[reply_off + 1];
1009 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
1012 if (inode->i_op && inode->i_op->getxattr)
1013 rc = inode->i_op->listxattr(dentry, value, len);
1016 CERROR("listxattr failed: %d", rc);
1018 repbody->valid |= OBD_MD_FLEALIST;
1019 repbody->eadatasize = rc;
1026 int mds_pack_posix_acl(struct lustre_msg *repmsg, int offset,
1027 struct mds_body *body, struct inode *inode)
1029 struct dentry de = { .d_inode = inode };
1030 __u32 buflen, *sizep;
1035 if (!inode->i_op->getxattr)
1038 buflen = repmsg->buflens[offset + 1];
1039 buf = lustre_msg_buf(repmsg, offset + 1, buflen);
1041 size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1042 if (size == -ENODATA || size == -EOPNOTSUPP)
1048 sizep = lustre_msg_buf(repmsg, offset, 4);
1050 CERROR("can't locate returned acl size buf\n");
1054 *sizep = cpu_to_le32(size);
1055 body->valid |= OBD_MD_FLACL_ACCESS;
1061 int mds_pack_remote_perm(struct ptlrpc_request *req, int reply_off,
1062 struct mds_body *body, struct inode *inode)
1064 struct lustre_sec_desc *lsd;
1065 struct mds_remote_perm *perm;
1068 LASSERT(inode->i_op);
1069 LASSERT(inode->i_op->permission);
1070 LASSERT(req->rq_export->exp_mds_data.med_remote);
1072 perm = (struct mds_remote_perm *)
1073 lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(perm));
1077 memset(perm, 0, sizeof(*perm));
1079 /* obtain authenticated uid/gid and LSD permissions, which
1080 * might be different from current process context, from LSD
1082 lsd = mds_get_lsd(current->uid);
1084 CWARN("can't LSD of uid %u\n", current->uid);
1088 perm->mrp_auth_uid = lsd->lsd_uid;
1089 perm->mrp_auth_gid = lsd->lsd_gid;
1091 lsd_perms = mds_lsd_get_perms(lsd, 1, 0, req->rq_peer.peer_id.nid);
1092 if (lsd_perms & LSD_PERM_SETUID)
1093 perm->mrp_allow_setuid = 1;
1094 if (lsd_perms & LSD_PERM_SETGID)
1095 perm->mrp_allow_setgid = 1;
1099 /* permission bits of current user
1100 * XXX this is low efficient, could we do it in one blow?
1102 if (inode->i_op->permission(inode, MAY_EXEC, NULL) == 0)
1103 perm->mrp_perm |= MAY_EXEC;
1104 if (inode->i_op->permission(inode, MAY_WRITE, NULL) == 0)
1105 perm->mrp_perm |= MAY_WRITE;
1106 if (inode->i_op->permission(inode, MAY_READ, NULL) == 0)
1107 perm->mrp_perm |= MAY_READ;
1109 body->valid |= (OBD_MD_FLACL_ACCESS | OBD_MD_RACL);
1114 int mds_pack_acl(struct ptlrpc_request *req, int reply_off,
1115 struct mds_body *body, struct inode *inode)
1119 if (!req->rq_export->exp_mds_data.med_remote)
1120 rc = mds_pack_posix_acl(req->rq_repmsg, reply_off, body, inode);
1122 rc = mds_pack_remote_perm(req, reply_off, body, inode);
1127 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1128 struct ptlrpc_request *req, int req_off,
1129 struct mds_body *reqbody, int reply_off,
1130 struct mds_req_sec_desc *rsd)
1132 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1133 struct inode *inode = dentry->d_inode;
1134 struct mds_body *body;
1138 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1141 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1142 LASSERT(body != NULL); /* caller prepped reply */
1144 if (dentry->d_flags & DCACHE_CROSS_REF) {
1145 mds_pack_dentry2body(obd, body, dentry,
1146 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1147 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1152 mds_pack_inode2body(obd, body, inode,
1153 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1155 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1156 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1158 /* guessing what kind og attribute do we need. */
1159 int is_mea = (S_ISDIR(inode->i_mode) &&
1160 (reqbody->valid & OBD_MD_MEA) != 0);
1162 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
1163 body, inode, 1, is_mea);
1165 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1166 if (!(body->valid & OBD_MD_FLEASIZE) &&
1167 !(body->valid & OBD_MD_FLDIREA))
1168 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1169 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1170 } else if (S_ISLNK(inode->i_mode) &&
1171 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1172 rc = mds_pack_link(dentry, req, body, reply_off);
1173 } else if (reqbody->valid & OBD_MD_FLEA) {
1174 rc = mds_pack_ea(dentry, req, body, req_off, reply_off);
1175 } else if (reqbody->valid & OBD_MD_FLEALIST) {
1176 rc = mds_pack_ealist(dentry, req, body, reply_off);
1179 if (reqbody->valid & OBD_MD_FLACL_ACCESS) {
1180 int inc = (reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1;
1181 rc = mds_pack_acl(req, reply_off + inc, body, inode);
1185 mds_body_do_reverse_map(med, body);
1190 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1191 struct dentry *dentry,
1194 int rc = 0, size[1] = {sizeof(struct mds_body)};
1197 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1198 CERROR("failed MDS_GETATTR_PACK test\n");
1199 req->rq_status = -ENOMEM;
1203 rc = lustre_pack_reply(req, 1, size, NULL);
1205 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1206 GOTO(out, req->rq_status = rc);
1214 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1217 struct inode *inode = de->d_inode;
1218 struct mds_obd *mds = mds_req2mds(req);
1219 struct mds_body *body;
1220 int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1223 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1224 LASSERT(body != NULL); /* checked by caller */
1225 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
1227 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1228 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1231 down(&inode->i_sem);
1232 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1233 ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1236 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1237 CERROR("error getting inode %lu MD: rc = %d\n",
1240 } else if (rc > mds->mds_max_mdsize) {
1242 CERROR("MD size %d larger than maximum possible %u\n",
1243 rc, mds->mds_max_mdsize);
1245 size[bufcount] = rc;
1248 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1249 if (inode->i_size + 1 != body->eadatasize)
1250 CERROR("symlink size: %Lu, reply space: %d\n",
1251 inode->i_size + 1, body->eadatasize);
1252 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1254 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1255 inode->i_size + 1, body->eadatasize);
1256 } else if ((body->valid & OBD_MD_FLEA)) {
1257 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1260 if (inode->i_op && inode->i_op->getxattr)
1261 rc = inode->i_op->getxattr(de, ea_name, NULL, 0);
1264 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1265 CERROR("error getting inode %lu EA: rc = %d\n",
1269 size[bufcount] = min_t(int, body->eadatasize, rc);
1272 } else if (body->valid & OBD_MD_FLEALIST) {
1274 if (inode->i_op && inode->i_op->getxattr)
1275 rc = inode->i_op->listxattr(de, NULL, 0);
1278 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1279 CERROR("error getting inode %lu EA: rc = %d\n",
1283 size[bufcount] = min_t(int, body->eadatasize, rc);
1288 /* may co-exist with OBD_MD_FLEASIZE */
1289 if (body->valid & OBD_MD_FLACL_ACCESS) {
1290 if (req->rq_export->exp_mds_data.med_remote) {
1291 size[bufcount++] = sizeof(struct mds_remote_perm);
1293 size[bufcount++] = 4;
1294 size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1298 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1299 CERROR("failed MDS_GETATTR_PACK test\n");
1300 req->rq_status = -ENOMEM;
1301 GOTO(out, rc = -ENOMEM);
1304 rc = lustre_pack_reply(req, bufcount, size, NULL);
1306 CERROR("out of memory\n");
1307 GOTO(out, req->rq_status = rc);
1315 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1316 char *name, int namelen)
1318 struct mea *mea = NULL;
1319 int mea_size, rc = 0;
1322 rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1327 * dir is already splitted, check if requested filename should
1328 * live at this MDS or at another one.
1330 int i = mea_name2idx(mea, name, namelen - 1);
1331 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1333 "inapropriate MDS(%d) for %s. should be "
1334 "%lu(%d)\n", mea->mea_master, name,
1335 (unsigned long)id_group(&mea->mea_ids[i]), i);
1341 OBD_FREE(mea, mea_size);
1345 int mds_getattr_size(struct obd_device *obd, struct dentry *dentry,
1346 struct ptlrpc_request *req, struct mds_body *body)
1348 struct inode *inode = dentry->d_inode;
1351 LASSERT(body != NULL);
1353 if (dentry->d_inode == NULL || !S_ISREG(inode->i_mode))
1356 if (obd->obd_recovering) {
1357 CDEBUG(D_ERROR, "size for "DLID4" is unknown yet (recovering)\n",
1362 if (atomic_read(&inode->i_writecount)) {
1363 /* some one has opened the file for write.
1364 * mds doesn't know actual size */
1365 CDEBUG(D_OTHER, "MDS doesn't know actual size for "DLID4"\n",
1369 CDEBUG(D_OTHER, "MDS returns "LPD64"/"LPD64" for"DLID4"\n",
1370 body->size, body->blocks, OLID4(&body->id1));
1371 body->valid |= OBD_MD_FLSIZE;
1375 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1376 struct lustre_handle *child_lockh, int child_part)
1378 struct obd_device *obd = req->rq_export->exp_obd;
1379 struct mds_obd *mds = &obd->u.mds;
1380 struct ldlm_reply *rep = NULL;
1381 struct lvfs_run_ctxt saved;
1382 struct mds_req_sec_desc *rsd;
1383 struct mds_body *body;
1384 struct dentry *dparent = NULL, *dchild = NULL;
1385 struct lvfs_ucred uc = {NULL, NULL,};
1386 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1387 unsigned int namesize;
1388 int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1392 LASSERT(!strcmp(obd->obd_type->typ_name, OBD_MDS_DEVICENAME));
1393 MD_COUNTER_INCREMENT(obd, getattr_lock);
1395 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1397 CERROR("Can't unpack security desc\n");
1401 /* swab now, before anyone looks inside the request. */
1402 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1403 lustre_swab_mds_body);
1405 CERROR("Can't swab mds_body\n");
1406 GOTO(cleanup, rc = -EFAULT);
1409 LASSERT_REQSWAB(req, offset + 1);
1410 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1412 CERROR("Can't unpack name\n");
1413 GOTO(cleanup, rc = -EFAULT);
1415 namesize = req->rq_reqmsg->buflens[offset + 1];
1417 /* namesize less than 2 means we have empty name, probably came from
1418 revalidate by cfid, so no point in having name to be set */
1422 LASSERT (offset == 1 || offset == 3);
1424 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1430 rc = mds_init_ucred(&uc, req, rsd);
1435 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1436 cleanup_phase = 1; /* kernel context */
1437 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1439 LASSERT(namesize > 0);
1440 if (child_lockh->cookie != 0) {
1441 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1445 if (body->valid == OBD_MD_FLID) {
1446 struct mds_body *mds_reply;
1447 int size = sizeof(*mds_reply);
1451 dparent = mds_id2dentry(obd, &body->id1, NULL);
1452 if (IS_ERR(dparent)) {
1453 rc = PTR_ERR(dparent);
1457 * the user requested ONLY the inode number, so do a raw lookup.
1459 rc = lustre_pack_reply(req, 1, &size, NULL);
1461 CERROR("out of memory\n");
1465 dir = dparent->d_inode;
1466 LASSERT(dir->i_op->lookup_raw != NULL);
1467 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1469 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1470 sizeof(*mds_reply));
1472 id_ino(&mds_reply->id1) = inum;
1473 mds_reply->valid = OBD_MD_FLID;
1477 if (resent_req == 0) {
1478 LASSERT(id_fid(&body->id1) != 0);
1480 rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1481 parent_lockh, &dparent,
1483 MDS_INODELOCK_UPDATE,
1486 child_lockh, &dchild,
1487 LCK_PR, child_part);
1491 cleanup_phase = 2; /* dchild, dparent, locks */
1494 * let's make sure this name should leave on this mds
1497 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1501 /* we have no dentry here, drop LOOKUP bit */
1502 /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1503 child_part &= ~MDS_INODELOCK_LOOKUP;
1504 CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1505 obd->obd_name, OLID4(&body->id1));
1507 dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1508 LCK_PR, parent_lockh,
1511 MDS_INODELOCK_UPDATE);
1512 if (IS_ERR(dchild)) {
1513 CERROR("can't find inode with id "DLID4", err = %d\n",
1514 OLID4(&body->id1), (int)PTR_ERR(dchild));
1515 GOTO(cleanup, rc = PTR_ERR(dchild));
1517 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1520 struct ldlm_lock *granted_lock;
1522 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1523 granted_lock = ldlm_handle2lock(child_lockh);
1525 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1526 id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1527 child_lockh->cookie);
1530 /* usual named request */
1531 dparent = mds_id2dentry(obd, &body->id1, NULL);
1532 LASSERT(!IS_ERR(dparent));
1533 dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1534 LASSERT(!IS_ERR(dchild));
1536 /* client wants to get attr. by id */
1537 dchild = mds_id2dentry(obd, &body->id1, NULL);
1538 LASSERT(!IS_ERR(dchild));
1540 LDLM_LOCK_PUT(granted_lock);
1543 cleanup_phase = 2; /* dchild, dparent, locks */
1545 if (!DENTRY_VALID(dchild)) {
1546 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1548 * in the intent case, the policy clears this error: the
1549 * disposition is enough.
1554 intent_set_disposition(rep, DISP_LOOKUP_POS);
1557 if (req->rq_repmsg == NULL) {
1558 if (dchild->d_flags & DCACHE_CROSS_REF)
1559 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1561 rc = mds_getattr_pack_msg(req, dchild, offset);
1563 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1568 rc = mds_getattr_internal(obd, dchild, req, offset, body,
1571 GOTO(cleanup, rc); /* returns the lock to the client */
1573 /* probably MDS knows actual size? */
1574 body = lustre_msg_buf(req->rq_repmsg, reply_offset, sizeof(*body));
1575 LASSERT(body != NULL);
1576 mds_getattr_size(obd, dchild, req, body);
1581 switch (cleanup_phase) {
1583 if (resent_req == 0) {
1584 if (rc && DENTRY_VALID(dchild))
1585 ldlm_lock_decref(child_lockh, LCK_PR);
1587 ldlm_lock_decref(parent_lockh, LCK_PR);
1589 if (parent_lockh[1].cookie != 0)
1590 ldlm_lock_decref(parent_lockh + 1, update_mode);
1597 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1599 mds_exit_ucred(&uc);
1604 static int mds_getattr(struct ptlrpc_request *req, int offset)
1606 struct obd_device *obd = req->rq_export->exp_obd;
1607 struct lvfs_run_ctxt saved;
1609 struct mds_req_sec_desc *rsd;
1610 struct mds_body *body;
1611 struct lvfs_ucred uc = {NULL, NULL,};
1615 MD_COUNTER_INCREMENT(obd, getattr);
1617 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1619 CERROR("Can't unpack security desc\n");
1623 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1624 lustre_swab_mds_body);
1626 CERROR ("Can't unpack body\n");
1630 rc = mds_init_ucred(&uc, req, rsd);
1632 mds_exit_ucred(&uc);
1636 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1637 de = mds_id2dentry(obd, &body->id1, NULL);
1639 rc = req->rq_status = PTR_ERR(de);
1643 rc = mds_getattr_pack_msg(req, de, offset);
1645 CERROR("mds_getattr_pack_msg: %d\n", rc);
1649 req->rq_status = mds_getattr_internal(obd, de, req, offset, body,
1655 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1656 mds_exit_ucred(&uc);
1660 static int mds_access_check(struct ptlrpc_request *req, int offset)
1662 struct obd_device *obd = req->rq_export->exp_obd;
1663 struct lvfs_run_ctxt saved;
1665 struct mds_req_sec_desc *rsd;
1666 struct mds_body *body;
1667 struct lvfs_ucred uc;
1668 int rep_size[2] = {sizeof(*body),
1669 sizeof(struct mds_remote_perm)};
1673 if (!req->rq_export->exp_mds_data.med_remote) {
1674 CERROR("from local client "LPU64"\n", req->rq_peer.peer_id.nid);
1678 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1680 CERROR("Can't unpack security desc\n");
1684 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1685 lustre_swab_mds_body);
1687 CERROR ("Can't unpack body\n");
1691 MD_COUNTER_INCREMENT(obd, access_check);
1693 rc = mds_init_ucred(&uc, req, rsd);
1695 CERROR("init ucred error: %d\n", rc);
1698 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1700 de = mds_id2dentry(obd, &body->id1, NULL);
1702 CERROR("grab ino "LPU64": err %ld\n",
1703 body->id1.li_stc.u.e3s.l3s_ino, PTR_ERR(de));
1704 GOTO(out_pop, rc = PTR_ERR(de));
1707 rc = lustre_pack_reply(req, 2, rep_size, NULL);
1709 CERROR("pack reply error: %d\n", rc);
1710 GOTO(out_dput, rc = -EINVAL);
1713 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1716 rc = mds_pack_remote_perm(req, 1, body, de->d_inode);
1722 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1723 mds_exit_ucred(&uc);
1727 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1728 unsigned long max_age)
1733 spin_lock(&obd->obd_osfs_lock);
1734 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1736 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1737 spin_unlock(&obd->obd_osfs_lock);
1742 static int mds_statfs(struct ptlrpc_request *req)
1744 struct obd_device *obd = req->rq_export->exp_obd;
1745 int rc, size = sizeof(struct obd_statfs);
1748 /* This will trigger a watchdog timeout */
1749 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1750 (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1752 rc = lustre_pack_reply(req, 1, &size, NULL);
1753 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1754 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1758 OBD_COUNTER_INCREMENT(obd, statfs);
1760 /* We call this so that we can cache a bit - 1 jiffie worth */
1761 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1764 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1770 req->rq_status = rc;
1774 static int mds_sync(struct ptlrpc_request *req, int offset)
1776 struct obd_device *obd = req->rq_export->exp_obd;
1777 struct mds_obd *mds = &obd->u.mds;
1778 struct mds_body *body;
1779 int rc, size = sizeof(*body);
1782 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1783 lustre_swab_mds_body);
1785 GOTO(out, rc = -EPROTO);
1787 rc = lustre_pack_reply(req, 1, &size, NULL);
1788 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1789 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1793 if (id_ino(&body->id1) == 0) {
1794 /* an id of zero is taken to mean "sync whole filesystem" */
1795 rc = fsfilt_sync(obd, mds->mds_sb);
1799 /* just any file to grab fsync method - "file" arg unused */
1800 struct file *file = mds->mds_rcvd_filp;
1801 struct mds_body *rep_body;
1804 de = mds_id2dentry(obd, &body->id1, NULL);
1806 GOTO(out, rc = PTR_ERR(de));
1808 rc = file->f_op->fsync(NULL, de, 1);
1812 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1813 mds_pack_inode2body(obd, rep_body, de->d_inode,
1814 (body->valid & OBD_MD_FID) ? 1 : 0);
1820 req->rq_status = rc;
1824 /* mds_readpage does not take a DLM lock on the inode, because the client must
1825 * already have a PR lock.
1827 * If we were to take another one here, a deadlock will result, if another
1828 * thread is already waiting for a PW lock. */
1829 static int mds_readpage(struct ptlrpc_request *req, int offset)
1831 struct obd_device *obd = req->rq_export->exp_obd;
1832 struct vfsmount *mnt;
1835 struct mds_req_sec_desc *rsd;
1836 struct mds_body *body, *repbody;
1837 struct lvfs_run_ctxt saved;
1838 int rc, size = sizeof(*repbody);
1839 struct lvfs_ucred uc = {NULL, NULL,};
1842 rc = lustre_pack_reply(req, 1, &size, NULL);
1843 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1844 CERROR("mds: out of memory\n");
1845 GOTO(out, rc = -ENOMEM);
1848 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1850 CERROR("Can't unpack security desc\n");
1851 GOTO (out, rc = -EFAULT);
1854 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1855 lustre_swab_mds_body);
1857 CERROR("Can't unpack body\n");
1858 GOTO (out, rc = -EFAULT);
1861 rc = mds_init_ucred(&uc, req, rsd);
1866 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1867 de = mds_id2dentry(obd, &body->id1, &mnt);
1869 GOTO(out_pop, rc = PTR_ERR(de));
1871 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1873 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1874 /* note: in case of an error, dentry_open puts dentry */
1876 GOTO(out_pop, rc = PTR_ERR(file));
1878 /* body->size is actually the offset -eeb */
1879 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1880 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1881 body->size, de->d_inode->i_blksize);
1882 GOTO(out_file, rc = -EFAULT);
1885 /* body->nlink is actually the #bytes to read -eeb */
1886 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1887 CERROR("size %u is not multiple of blocksize %lu\n",
1888 body->nlink, de->d_inode->i_blksize);
1889 GOTO(out_file, rc = -EFAULT);
1892 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1893 repbody->size = file->f_dentry->d_inode->i_size;
1894 repbody->valid = OBD_MD_FLSIZE;
1896 /* to make this asynchronous make sure that the handling function
1897 doesn't send a reply when this function completes. Instead a
1898 callback function would send the reply */
1899 /* body->size is actually the offset -eeb */
1900 rc = mds_sendpage(req, file, body->size, body->nlink);
1904 filp_close(file, 0);
1906 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1908 mds_exit_ucred(&uc);
1909 req->rq_status = rc;
1913 /* update master MDS ID, which is stored in local inode EA. */
1914 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1915 void *data, int data_len)
1917 struct mds_obd *mds = &obd->u.mds;
1918 struct dentry *dentry;
1926 dentry = mds_id2dentry(obd, id, NULL);
1928 GOTO(out, rc = PTR_ERR(dentry));
1930 if (!dentry->d_inode) {
1931 CERROR("Can't find object "DLID4".\n",
1933 GOTO(out_dentry, rc = -EINVAL);
1936 handle = fsfilt_start(obd, dentry->d_inode,
1937 FSFILT_OP_SETATTR, NULL);
1939 GOTO(out_dentry, rc = PTR_ERR(handle));
1941 rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1942 (struct lustre_id *)data);
1944 CERROR("Can't update inode "DLID4" master id, "
1945 "error = %d.\n", OLID4(id), rc);
1946 GOTO(out_commit, rc);
1951 fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1958 EXPORT_SYMBOL(mds_update_mid);
1960 /* read master MDS ID, which is stored in local inode EA. */
1961 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1962 void *data, int data_len)
1964 struct dentry *dentry;
1971 dentry = mds_id2dentry(obd, id, NULL);
1973 GOTO(out, rc = PTR_ERR(dentry));
1975 if (!dentry->d_inode) {
1976 CERROR("Can't find object "DLID4".\n",
1978 GOTO(out_dentry, rc = -EINVAL);
1981 down(&dentry->d_inode->i_sem);
1982 rc = mds_read_inode_mid(obd, dentry->d_inode,
1983 (struct lustre_id *)data);
1984 up(&dentry->d_inode->i_sem);
1986 CERROR("Can't read inode "DLID4" master id, "
1987 "error = %d.\n", OLID4(id), rc);
1988 GOTO(out_dentry, rc);
1997 EXPORT_SYMBOL(mds_read_mid);
1999 int mds_read_md(struct obd_device *obd, struct lustre_id *id,
2000 char **data, int *datalen)
2002 struct dentry *dentry;
2003 struct mds_obd *mds = &obd->u.mds;
2004 int rc = 0, mea = 0;
2011 dentry = mds_id2dentry(obd, id, NULL);
2013 GOTO(out, rc = PTR_ERR(dentry));
2015 if (!dentry->d_inode) {
2016 CERROR("Can't find object "DLID4".\n",
2018 GOTO(out_dentry, rc = -EINVAL);
2020 if (S_ISDIR(dentry->d_inode->i_mode)) {
2021 *datalen = obd_packmd(mds->mds_md_exp, NULL, NULL);
2024 *datalen = obd_packmd(mds->mds_dt_exp, NULL, NULL);
2027 OBD_ALLOC(ea, *datalen);
2030 GOTO(out_dentry, rc = PTR_ERR(dentry));
2033 down(&dentry->d_inode->i_sem);
2034 rc = fsfilt_get_md(obd, dentry->d_inode, *data, *datalen,
2035 (mea ? EA_MEA : EA_LOV));
2036 up(&dentry->d_inode->i_sem);
2039 CERROR("Error %d reading eadata for ino %lu\n",
2040 rc, dentry->d_inode->i_ino);
2046 EXPORT_SYMBOL(mds_read_md);
2048 int mds_reint(struct ptlrpc_request *req, int offset,
2049 struct lustre_handle *lockh)
2051 struct mds_update_record *rec;
2052 struct mds_req_sec_desc *rsd;
2056 OBD_ALLOC(rec, sizeof(*rec));
2060 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
2062 CERROR("Can't unpack security desc\n");
2063 GOTO(out, rc = -EFAULT);
2066 rc = mds_update_unpack(req, offset, rec);
2067 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
2068 CERROR("invalid record\n");
2069 GOTO(out, req->rq_status = -EINVAL);
2072 rc = mds_init_ucred(&rec->ur_uc, req, rsd);
2077 /* rc will be used to interrupt a for loop over multiple records */
2078 rc = mds_reint_rec(rec, offset, req, lockh);
2081 mds_exit_ucred(&rec->ur_uc);
2082 OBD_FREE(rec, sizeof(*rec));
2086 static int mds_filter_recovery_request(struct ptlrpc_request *req,
2087 struct obd_device *obd, int *process)
2089 switch (req->rq_reqmsg->opc) {
2090 case MDS_CONNECT: /* This will never get here, but for completeness. */
2091 case OST_CONNECT: /* This will never get here, but for completeness. */
2092 case MDS_DISCONNECT:
2093 case OST_DISCONNECT:
2098 case MDS_SYNC: /* used in unmounting */
2103 *process = target_queue_recovery_request(req, obd);
2107 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
2109 /* XXX what should we set rq_status to here? */
2110 req->rq_status = -EAGAIN;
2111 RETURN(ptlrpc_error(req));
2115 static char *reint_names[] = {
2116 [REINT_SETATTR] "setattr",
2117 [REINT_CREATE] "create",
2118 [REINT_LINK] "link",
2119 [REINT_UNLINK] "unlink",
2120 [REINT_RENAME] "rename",
2121 [REINT_OPEN] "open",
2124 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER | \
2125 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
2126 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
2129 static void reconstruct_create(struct ptlrpc_request *req)
2131 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2132 struct mds_client_data *mcd = med->med_mcd;
2133 struct dentry *dentry;
2134 struct ost_body *body;
2135 struct lustre_id id;
2139 /* copy rc, transno and disp; steal locks */
2140 mds_req_from_mcd(req, mcd);
2141 if (req->rq_status) {
2149 id_ino(&id) = mcd->mcd_last_data;
2150 LASSERT(id_ino(&id) != 0);
2152 dentry = mds_id2dentry(req2obd(req), &id, NULL);
2153 if (IS_ERR(dentry)) {
2154 CERROR("can't find inode "LPU64"\n", id_ino(&id));
2155 req->rq_status = PTR_ERR(dentry);
2160 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
2161 req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
2162 dentry->d_inode->i_generation);
2164 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2165 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
2166 body->oa.o_id = dentry->d_inode->i_ino;
2167 body->oa.o_generation = dentry->d_inode->i_generation;
2168 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2170 down(&dentry->d_inode->i_sem);
2171 rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
2172 up(&dentry->d_inode->i_sem);
2174 CERROR("Can't read inode self id, inode %lu, "
2175 "rc %d\n", dentry->d_inode->i_ino, rc);
2179 body->oa.o_fid = id_fid(&id);
2180 body->oa.o_mds = id_group(&id);
2186 static int mds_inode_init_acl(struct obd_device *obd, void *handle,
2187 struct dentry *de, void *xattr, int xattr_size)
2189 struct inode *inode = de->d_inode;
2190 struct posix_acl *acl;
2197 LASSERT(xattr_size > 0);
2199 if (!inode->i_op->getxattr || !inode->i_op->setxattr) {
2200 CERROR("backend fs dosen't support xattr\n");
2204 /* set default acl */
2205 if (S_ISDIR(inode->i_mode)) {
2206 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_DEFAULT,
2207 xattr, xattr_size, 0);
2209 CERROR("set default acl err: %d\n", rc);
2214 /* set access acl */
2215 acl = posix_acl_from_xattr(xattr, xattr_size);
2216 if (acl == NULL || IS_ERR(acl)) {
2217 CERROR("insane attr data\n");
2218 return PTR_ERR(acl);
2221 if (posix_acl_valid(acl)) {
2222 CERROR("default acl not valid: %d\n", rc);
2227 mode = inode->i_mode;
2228 rc = posix_acl_create_masq(acl, &mode);
2230 CERROR("create masq err %d\n", rc);
2234 if (inode->i_mode != mode) {
2235 struct iattr iattr = { .ia_valid = ATTR_MODE,
2239 rc2 = fsfilt_setattr(obd, de, handle, &iattr, 0);
2241 CERROR("setattr mode err: %d\n", rc2);
2248 /* we didn't change acl except mode bits of some
2249 * entries, so should be fit into original size.
2251 rc = posix_acl_to_xattr(acl, xattr, xattr_size);
2254 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_ACCESS,
2255 xattr, xattr_size, 0);
2257 CERROR("set access acl err: %d\n", rc);
2260 posix_acl_release(acl);
2264 static int mdt_obj_create(struct ptlrpc_request *req)
2266 struct obd_device *obd = req->rq_export->exp_obd;
2267 struct mds_obd *mds = &obd->u.mds;
2268 struct ost_body *body, *repbody;
2271 char idname[LL_ID_NAMELEN];
2272 int size = sizeof(*repbody);
2273 struct inode *parent_inode;
2274 struct lvfs_run_ctxt saved;
2275 int rc, cleanup_phase = 0;
2276 struct dentry *new = NULL;
2277 struct dentry_params dp;
2278 int mealen, flags = 0;
2279 struct lvfs_ucred uc;
2280 struct lustre_id id;
2282 void *handle = NULL;
2283 unsigned long cr_inum = 0;
2286 DEBUG_REQ(D_HA, req, "create remote object");
2287 parent_inode = mds->mds_unnamed_dir->d_inode;
2289 body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2290 lustre_swab_ost_body);
2294 /* acl data is packed transparently, no swab here */
2295 LASSERT(req->rq_reqmsg->bufcount >= 2);
2296 acl_size = req->rq_reqmsg->buflens[1];
2298 acl = lustre_msg_buf(req->rq_reqmsg, 1, acl_size);
2300 CERROR("No default acl buf?\n");
2305 rc = lustre_pack_reply(req, 1, &size, NULL);
2309 MDS_CHECK_RESENT(req, reconstruct_create(req));
2312 uc.luc_ginfo = NULL;
2313 uc.luc_uid = body->oa.o_uid;
2314 uc.luc_gid = body->oa.o_gid;
2315 uc.luc_fsuid = body->oa.o_uid;
2316 uc.luc_fsgid = body->oa.o_gid;
2318 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2319 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2321 /* in REPLAY case inum should be given (client or other MDS fills it) */
2322 if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2323 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2325 * this is re-create request from MDS holding directory name.
2326 * we have to lookup given ino/gen first. if it exists (good
2327 * case) then there is nothing to do. if it does not then we
2328 * have to recreate it.
2330 id_ino(&id) = body->oa.o_id;
2331 id_gen(&id) = body->oa.o_generation;
2333 new = mds_id2dentry(obd, &id, NULL);
2334 if (!IS_ERR(new) && new->d_inode) {
2335 struct lustre_id sid;
2337 CDEBUG(D_OTHER, "mkdir repairing %lu/%lu\n",
2338 (unsigned long)id_ino(&id),
2339 (unsigned long)id_gen(&id));
2341 obdo_from_inode(&repbody->oa, new->d_inode,
2342 FILTER_VALID_FLAGS);
2344 repbody->oa.o_id = new->d_inode->i_ino;
2345 repbody->oa.o_generation = new->d_inode->i_generation;
2346 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2349 down(&new->d_inode->i_sem);
2350 rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2351 up(&new->d_inode->i_sem);
2353 CERROR("Can't read inode self id "
2354 "inode %lu, rc %d.\n",
2355 new->d_inode->i_ino, rc);
2359 repbody->oa.o_fid = id_fid(&sid);
2360 repbody->oa.o_mds = id_group(&sid);
2361 LASSERT(id_fid(&sid) != 0);
2364 * here we could use fid passed in body->oa.o_fid and
2365 * thus avoid mds_read_inode_sid().
2367 cr_inum = new->d_inode->i_ino;
2368 GOTO(cleanup, rc = 0);
2372 down(&parent_inode->i_sem);
2373 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2374 if (IS_ERR(handle)) {
2375 up(&parent_inode->i_sem);
2376 CERROR("fsfilt_start() failed, rc = %d\n",
2377 (int)PTR_ERR(handle));
2378 GOTO(cleanup, rc = PTR_ERR(handle));
2380 cleanup_phase = 1; /* transaction */
2383 rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2384 new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2386 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2387 obd->obd_name, idname, (int) PTR_ERR(new));
2388 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2389 up(&parent_inode->i_sem);
2390 RETURN(PTR_ERR(new));
2391 } else if (new->d_inode) {
2392 CERROR("%s: name exists. repeat\n", obd->obd_name);
2396 new->d_fsdata = (void *)&dp;
2400 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2401 (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2402 LASSERT(body->oa.o_id != 0);
2403 dp.p_inum = body->oa.o_id;
2404 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2405 (unsigned long)body->oa.o_id,
2406 (unsigned long)body->oa.o_generation);
2409 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2412 rc = mds_inode_init_acl(obd, handle, new,
2415 up(&parent_inode->i_sem);
2419 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2420 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2421 new->d_inode->i_generation = body->oa.o_generation;
2422 mark_inode_dirty(new->d_inode);
2425 * avoiding asserts in cache flush case, as
2426 * @body->oa.o_id should be zero.
2428 if (body->oa.o_id) {
2429 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2430 "BUG 3550: failed to recreate obj "
2431 LPU64" -> %lu\n", body->oa.o_id,
2432 new->d_inode->i_ino);
2434 LASSERTF(body->oa.o_generation ==
2435 new->d_inode->i_generation,
2436 "BUG 3550: failed to recreate obj/gen "
2437 LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2438 body->oa.o_generation,
2439 new->d_inode->i_ino,
2440 new->d_inode->i_generation);
2444 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2445 repbody->oa.o_id = new->d_inode->i_ino;
2446 repbody->oa.o_generation = new->d_inode->i_generation;
2447 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2449 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2450 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2451 id_group(&id) = mds->mds_num;
2453 LASSERT(body->oa.o_fid != 0);
2454 id_fid(&id) = body->oa.o_fid;
2456 LASSERT(body->oa.o_id != 0);
2457 id_ino(&id) = repbody->oa.o_id;
2458 id_gen(&id) = repbody->oa.o_generation;
2460 down(&new->d_inode->i_sem);
2461 rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
2462 up(&new->d_inode->i_sem);
2465 * make sure, that fid is up-to-date.
2467 mds_set_last_fid(obd, id_fid(&id));
2470 * allocate new sid, as object is created from scratch
2471 * and this is not replay.
2473 down(&new->d_inode->i_sem);
2474 rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
2475 up(&new->d_inode->i_sem);
2478 CERROR("Can't update lustre ID for inode %lu, "
2479 "error = %d\n", new->d_inode->i_ino, rc);
2483 /* initializing o_fid after it is allocated. */
2484 repbody->oa.o_fid = id_fid(&id);
2485 repbody->oa.o_mds = id_group(&id);
2487 rc = fsfilt_del_dir_entry(obd, new);
2488 up(&parent_inode->i_sem);
2490 CERROR("can't remove name for object: %d\n", rc);
2494 cleanup_phase = 2; /* created directory object */
2496 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2497 (unsigned long)new->d_inode->i_ino,
2498 (unsigned long)new->d_inode->i_generation,
2499 (unsigned)new->d_inode->i_mode);
2500 cr_inum = new->d_inode->i_ino;
2502 up(&parent_inode->i_sem);
2503 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2507 if (body->oa.o_valid & OBD_MD_FLID) {
2508 /* this is new object for splitted dir. We have to prevent
2509 * recursive splitting on it -bzzz */
2510 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2512 OBD_ALLOC(mea, mealen);
2514 GOTO(cleanup, rc = -ENOMEM);
2516 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2517 mea->mea_master = 0;
2520 down(&new->d_inode->i_sem);
2521 rc = fsfilt_set_md(obd, new->d_inode, handle,
2522 mea, mealen, EA_MEA);
2523 up(&new->d_inode->i_sem);
2525 CERROR("fsfilt_set_md() failed, "
2528 OBD_FREE(mea, mealen);
2530 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2531 obd->obd_name, new->d_inode->i_ino,
2532 new->d_inode->i_generation, flags);
2533 } else if (body->oa.o_easize) {
2534 /* we pass LCK_EX to split routine to signal that we have
2535 * exclusive access to the directory. simple because nobody
2536 * knows it already exists -bzzz */
2537 rc = mds_try_to_split_dir(obd, new, NULL,
2538 body->oa.o_easize, LCK_EX);
2540 CERROR("Can't split directory %lu, error = %d.\n",
2541 new->d_inode->i_ino, rc);
2549 switch (cleanup_phase) {
2550 case 2: /* object has been created, but we'll may want to replay it later */
2552 ptlrpc_require_repack(req);
2553 case 1: /* transaction */
2554 rc = mds_finish_transno(mds, parent_inode, handle,
2559 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2563 static int mdt_get_info(struct ptlrpc_request *req)
2565 struct obd_export *exp = req->rq_export;
2570 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2572 DEBUG_REQ(D_HA, req, "no get_info key");
2575 keylen = req->rq_reqmsg->buflens[0];
2577 if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2578 (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2579 (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2582 if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2583 struct lustre_id *reply;
2584 int size = sizeof(*reply);
2586 rc = lustre_pack_reply(req, 1, &size, NULL);
2590 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2591 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2594 int size = sizeof(*reply);
2596 rc = lustre_pack_reply(req, 1, &size, NULL);
2600 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2601 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2604 req->rq_repmsg->status = 0;
2608 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2609 void *key, __u32 vallen, void *val)
2611 struct obd_device *obd;
2612 struct mds_obd *mds;
2616 obd = class_exp2obd(exp);
2618 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2619 exp->exp_handle.h_cookie);
2624 if (keylen >= strlen("mds_type") &&
2625 memcmp(key, "mds_type", keylen) == 0) {
2629 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
2631 mds->mds_obd_type = *(int*)val;
2632 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
2633 valsize = sizeof(group);
2635 /* mds number has been changed, so the corresponding obdfilter
2636 * exp need to be changed too. */
2637 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
2638 "mds_conn", valsize, &group);
2641 CDEBUG(D_IOCTL, "invalid key\n");
2645 static int mdt_set_info(struct ptlrpc_request *req)
2648 struct obd_export *exp = req->rq_export;
2649 int keylen, rc = 0, vallen;
2652 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2654 DEBUG_REQ(D_HA, req, "no set_info key");
2657 keylen = req->rq_reqmsg->buflens[0];
2659 if (keylen == strlen("mds_type") &&
2660 memcmp(key, "mds_type", keylen) == 0) {
2661 rc = lustre_pack_reply(req, 0, NULL, NULL);
2665 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2666 vallen = req->rq_reqmsg->buflens[1];
2668 rc = obd_set_info(exp, keylen, key, vallen, val);
2669 req->rq_repmsg->status = 0;
2672 CDEBUG(D_IOCTL, "invalid key\n");
2676 static void mds_revoke_export_locks(struct obd_export *exp)
2678 struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2679 struct list_head work;
2680 struct ldlm_lock *lock, *next;
2681 struct ldlm_lock_desc desc;
2683 if (!exp->u.eu_mds_data.med_remote)
2687 CERROR("implement right locking here! -bzzz\n");
2688 INIT_LIST_HEAD(&work);
2689 spin_lock(&exp->exp_ldlm_data.led_lock);
2690 list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2692 lock_res(lock->l_resource);
2693 if (lock->l_req_mode != lock->l_granted_mode) {
2694 unlock_res(lock->l_resource);
2698 LASSERT(lock->l_resource);
2699 if (lock->l_resource->lr_type != LDLM_IBITS &&
2700 lock->l_resource->lr_type != LDLM_PLAIN) {
2701 unlock_res(lock->l_resource);
2705 if (lock->l_flags & LDLM_FL_AST_SENT) {
2706 unlock_res(lock->l_resource);
2710 lock->l_flags |= LDLM_FL_AST_SENT;
2711 unlock_res(lock->l_resource);
2713 /* the desc just pretend to exclusive */
2714 ldlm_lock2desc(lock, &desc);
2715 desc.l_req_mode = LCK_EX;
2716 desc.l_granted_mode = 0;
2718 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2720 spin_unlock(&exp->exp_ldlm_data.led_lock);
2725 static int mds_msg_check_version(struct lustre_msg *msg)
2731 case MDS_DISCONNECT:
2733 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2735 CERROR("bad opc %u version %08x, expecting %08x\n",
2736 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2741 case MDS_GETATTR_LOCK:
2742 case MDS_ACCESS_CHECK:
2746 case MDS_DONE_WRITING:
2749 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2751 CERROR("bad opc %u version %08x, expecting %08x\n",
2752 msg->opc, msg->version, LUSTRE_MDS_VERSION);
2756 case LDLM_BL_CALLBACK:
2757 case LDLM_CP_CALLBACK:
2758 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2760 CERROR("bad opc %u version %08x, expecting %08x\n",
2761 msg->opc, msg->version, LUSTRE_DLM_VERSION);
2763 case OBD_LOG_CANCEL:
2764 case LLOG_ORIGIN_HANDLE_OPEN:
2765 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2766 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2767 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2768 case LLOG_ORIGIN_HANDLE_CLOSE:
2770 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2772 CERROR("bad opc %u version %08x, expecting %08x\n",
2773 msg->opc, msg->version, LUSTRE_LOG_VERSION);
2779 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2781 CERROR("bad opc %u version %08x, expecting %08x\n",
2782 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2785 case SEC_INIT_CONTINUE:
2790 CERROR("MDS unknown opcode %d\n", msg->opc);
2798 int mds_handle(struct ptlrpc_request *req)
2800 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2801 struct obd_device *obd = NULL;
2802 struct mds_obd *mds = NULL; /* quell gcc overwarning */
2806 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2808 rc = mds_msg_check_version(req->rq_reqmsg);
2810 CERROR("MDS drop mal-formed request\n");
2814 /* Security opc should NOT trigger any recovery events */
2815 if (req->rq_reqmsg->opc == SEC_INIT ||
2816 req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
2817 if (req->rq_export) {
2818 mds_req_add_idmapping(req,
2819 &req->rq_export->exp_mds_data);
2820 mds_revoke_export_locks(req->rq_export);
2823 } else if (req->rq_reqmsg->opc == SEC_FINI) {
2824 if (req->rq_export) {
2825 mds_req_del_idmapping(req,
2826 &req->rq_export->exp_mds_data);
2827 mds_revoke_export_locks(req->rq_export);
2832 LASSERT(current->journal_info == NULL);
2833 /* XXX identical to OST */
2834 if (req->rq_reqmsg->opc != MDS_CONNECT) {
2835 struct mds_export_data *med;
2838 if (req->rq_export == NULL) {
2839 CERROR("operation %d on unconnected MDS from %s\n",
2840 req->rq_reqmsg->opc,
2842 req->rq_status = -ENOTCONN;
2843 GOTO(out, rc = -ENOTCONN);
2846 med = &req->rq_export->exp_mds_data;
2847 obd = req->rq_export->exp_obd;
2850 /* sanity check: if the xid matches, the request must
2851 * be marked as a resent or replayed */
2852 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
2853 req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
2854 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2855 (MSG_RESENT | MSG_REPLAY),
2856 "rq_xid "LPU64" matches last_xid, "
2857 "expected RESENT flag\n",
2860 /* else: note the opposite is not always true; a
2861 * RESENT req after a failover will usually not match
2862 * the last_xid, since it was likely never
2863 * committed. A REPLAYed request will almost never
2864 * match the last xid, however it could for a
2865 * committed, but still retained, open. */
2867 spin_lock_bh(&obd->obd_processing_task_lock);
2868 recovering = obd->obd_recovering;
2869 spin_unlock_bh(&obd->obd_processing_task_lock);
2871 rc = mds_filter_recovery_request(req, obd,
2873 if (rc || should_process == 0) {
2875 } else if (should_process < 0) {
2876 req->rq_status = should_process;
2877 rc = ptlrpc_error(req);
2883 switch (req->rq_reqmsg->opc) {
2885 DEBUG_REQ(D_INODE, req, "connect");
2886 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2887 rc = target_handle_connect(req);
2889 struct mds_export_data *med;
2891 LASSERT(req->rq_export);
2892 med = &req->rq_export->u.eu_mds_data;
2893 mds_init_export_data(req, med);
2894 mds_req_add_idmapping(req, med);
2896 /* Now that we have an export, set mds. */
2897 obd = req->rq_export->exp_obd;
2898 mds = mds_req2mds(req);
2902 case MDS_DISCONNECT:
2903 DEBUG_REQ(D_INODE, req, "disconnect");
2904 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2905 rc = target_handle_disconnect(req);
2906 req->rq_status = rc; /* superfluous? */
2910 DEBUG_REQ(D_INODE, req, "getstatus");
2911 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2912 rc = mds_getstatus(req);
2916 DEBUG_REQ(D_INODE, req, "getattr");
2917 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2918 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2921 case MDS_ACCESS_CHECK:
2922 DEBUG_REQ(D_INODE, req, "access_check");
2923 OBD_FAIL_RETURN(OBD_FAIL_MDS_ACCESS_CHECK_NET, 0);
2924 rc = mds_access_check(req, MDS_REQ_REC_OFF);
2927 case MDS_GETATTR_LOCK: {
2928 struct lustre_handle lockh;
2929 DEBUG_REQ(D_INODE, req, "getattr_lock");
2930 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2932 /* If this request gets a reconstructed reply, we won't be
2933 * acquiring any new locks in mds_getattr_lock, so we don't
2937 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2938 MDS_INODELOCK_UPDATE);
2939 /* this non-intent call (from an ioctl) is special */
2940 req->rq_status = rc;
2941 if (rc == 0 && lockh.cookie)
2942 ldlm_lock_decref(&lockh, LCK_PR);
2946 DEBUG_REQ(D_INODE, req, "statfs");
2947 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2948 rc = mds_statfs(req);
2952 DEBUG_REQ(D_INODE, req, "readpage");
2953 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2954 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2956 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2957 if (req->rq_reply_state) {
2958 lustre_free_reply_state (req->rq_reply_state);
2959 req->rq_reply_state = NULL;
2966 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2969 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2970 mds->mds_max_cookiesize};
2973 /* NB only peek inside req now; mds_reint() will swab it */
2975 CERROR ("Can't inspect opcode\n");
2980 if (lustre_msg_swabbed (req->rq_reqmsg))
2983 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2984 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2985 reint_names[opc] == NULL) ? reint_names[opc] :
2988 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2990 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2992 else if (opc == REINT_OPEN)
2997 rc = lustre_pack_reply(req, bufcount, size, NULL);
3001 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
3002 fail = OBD_FAIL_MDS_REINT_NET_REP;
3007 DEBUG_REQ(D_INODE, req, "close");
3008 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
3009 rc = mds_close(req, MDS_REQ_REC_OFF);
3012 case MDS_DONE_WRITING:
3013 DEBUG_REQ(D_INODE, req, "done_writing");
3014 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
3015 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
3019 DEBUG_REQ(D_INODE, req, "pin");
3020 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
3021 rc = mds_pin(req, MDS_REQ_REC_OFF);
3025 DEBUG_REQ(D_INODE, req, "sync");
3026 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
3027 rc = mds_sync(req, MDS_REQ_REC_OFF);
3031 DEBUG_REQ(D_INODE, req, "ping");
3032 rc = target_handle_ping(req);
3035 case OBD_LOG_CANCEL:
3036 CDEBUG(D_INODE, "log cancel\n");
3037 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
3038 rc = -ENOTSUPP; /* la la la */
3042 DEBUG_REQ(D_INODE, req, "enqueue");
3043 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
3044 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
3045 ldlm_server_blocking_ast, NULL);
3046 fail = OBD_FAIL_LDLM_REPLY;
3049 DEBUG_REQ(D_INODE, req, "convert");
3050 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
3051 rc = ldlm_handle_convert(req);
3053 case LDLM_BL_CALLBACK:
3054 case LDLM_CP_CALLBACK:
3055 DEBUG_REQ(D_INODE, req, "callback");
3056 CERROR("callbacks should not happen on MDS\n");
3058 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
3060 case LLOG_ORIGIN_HANDLE_OPEN:
3061 DEBUG_REQ(D_INODE, req, "llog_init");
3062 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3063 rc = llog_origin_handle_open(req);
3065 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
3066 DEBUG_REQ(D_INODE, req, "llog next block");
3067 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3068 rc = llog_origin_handle_next_block(req);
3070 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
3071 DEBUG_REQ(D_INODE, req, "llog prev block");
3072 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3073 rc = llog_origin_handle_prev_block(req);
3075 case LLOG_ORIGIN_HANDLE_READ_HEADER:
3076 DEBUG_REQ(D_INODE, req, "llog read header");
3077 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3078 rc = llog_origin_handle_read_header(req);
3080 case LLOG_ORIGIN_HANDLE_CLOSE:
3081 DEBUG_REQ(D_INODE, req, "llog close");
3082 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3083 rc = llog_origin_handle_close(req);
3086 DEBUG_REQ(D_INODE, req, "ost_create");
3087 rc = mdt_obj_create(req);
3090 DEBUG_REQ(D_INODE, req, "get_info");
3091 rc = mdt_get_info(req);
3094 DEBUG_REQ(D_INODE, req, "set_info");
3095 rc = mdt_set_info(req);
3098 CDEBUG(D_INODE, "write\n");
3099 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
3100 rc = ost_brw_write(req, NULL);
3101 LASSERT(current->journal_info == NULL);
3102 /* mdt_brw sends its own replies */
3106 DEBUG_REQ(D_INODE, req, "llog catinfo");
3107 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3108 rc = llog_catinfo(req);
3111 req->rq_status = -ENOTSUPP;
3112 rc = ptlrpc_error(req);
3116 LASSERT(current->journal_info == NULL);
3120 /* If we're DISCONNECTing, the mds_export_data is already freed */
3121 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
3122 struct mds_export_data *med = &req->rq_export->exp_mds_data;
3123 struct obd_device *obd = list_entry(mds, struct obd_device,
3125 req->rq_repmsg->last_xid =
3126 le64_to_cpu(med->med_mcd->mcd_last_xid);
3128 if (!obd->obd_no_transno) {
3129 req->rq_repmsg->last_committed =
3130 obd->obd_last_committed;
3132 DEBUG_REQ(D_IOCTL, req,
3133 "not sending last_committed update");
3135 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
3137 mds->mds_last_transno, obd->obd_last_committed,
3143 target_send_reply(req, rc, fail);
3147 /* Update the server data on disk. This stores the new mount_count and also the
3148 * last_rcvd value to disk. If we don't have a clean shutdown, then the server
3149 * last_rcvd value may be less than that of the clients. This will alert us
3150 * that we may need to do client recovery.
3152 * Also assumes for mds_last_transno that we are not modifying it (no locking).
3154 int mds_update_server_data(struct obd_device *obd, int force_sync)
3156 struct mds_obd *mds = &obd->u.mds;
3157 struct mds_server_data *msd = mds->mds_server_data;
3158 struct file *filp = mds->mds_rcvd_filp;
3159 struct lvfs_run_ctxt saved;
3164 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3165 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
3167 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
3168 mds->mds_mount_count, mds->mds_last_transno);
3169 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
3171 CERROR("error writing MDS server data: rc = %d\n", rc);
3172 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3177 /* saves last allocated fid counter to file. */
3178 int mds_update_last_fid(struct obd_device *obd, void *handle,
3181 struct mds_obd *mds = &obd->u.mds;
3182 struct file *filp = mds->mds_fid_filp;
3183 struct lvfs_run_ctxt saved;
3189 spin_lock(&mds->mds_last_fid_lock);
3190 last_fid = mds->mds_last_fid;
3191 spin_unlock(&mds->mds_last_fid_lock);
3193 CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
3197 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
3198 handle, mds_commit_last_fid_cb,
3202 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3203 rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
3205 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3208 CERROR("error writing MDS last_fid #"LPU64
3209 ", err = %d\n", last_fid, rc);
3213 CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
3214 "%llu: err = %d\n", last_fid, off, rc);
3219 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
3221 struct mds_obd *mds = &obd->u.mds;
3223 spin_lock(&mds->mds_last_fid_lock);
3224 if (fid > mds->mds_last_fid)
3225 mds->mds_last_fid = fid;
3226 spin_unlock(&mds->mds_last_fid_lock);
3229 void mds_commit_last_transno_cb(struct obd_device *obd,
3230 __u64 transno, void *data,
3233 obd_transno_commit_cb(obd, transno, error);
3236 void mds_commit_last_fid_cb(struct obd_device *obd,
3237 __u64 fid, void *data,
3241 CERROR("%s: fid "LPD64" commit error: %d\n",
3242 obd->obd_name, fid, error);
3246 CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
3247 obd->obd_name, fid);
3250 __u64 mds_alloc_fid(struct obd_device *obd)
3252 struct mds_obd *mds = &obd->u.mds;
3255 spin_lock(&mds->mds_last_fid_lock);
3256 fid = ++mds->mds_last_fid;
3257 spin_unlock(&mds->mds_last_fid_lock);
3263 * allocates new lustre_id on passed @inode and saves it to inode EA.
3265 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
3266 void *handle, struct lustre_id *id)
3268 struct mds_obd *mds = &obd->u.mds;
3269 int alloc = 0, rc = 0;
3272 LASSERT(obd != NULL);
3273 LASSERT(inode != NULL);
3276 OBD_ALLOC(id, sizeof(*id));
3282 id_group(id) = mds->mds_num;
3283 id_fid(id) = mds_alloc_fid(obd);
3284 id_ino(id) = inode->i_ino;
3285 id_gen(id) = inode->i_generation;
3286 id_type(id) = (S_IFMT & inode->i_mode);
3288 rc = mds_update_inode_sid(obd, inode, handle, id);
3290 CERROR("Can't update inode FID EA, "
3295 OBD_FREE(id, sizeof(*id));
3300 * reads inode self id from inode EA. Probably later this should be replaced by
3301 * caching inode self id to avoid raeding it every time it is needed.
3303 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
3304 struct lustre_id *id)
3309 LASSERT(id != NULL);
3310 LASSERT(obd != NULL);
3311 LASSERT(inode != NULL);
3313 rc = fsfilt_get_md(obd, inode, &id->li_fid,
3314 sizeof(id->li_fid), EA_SID);
3316 CERROR("fsfilt_get_md() failed, "
3329 /* updates inode self id in EA. */
3330 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
3331 void *handle, struct lustre_id *id)
3336 LASSERT(id != NULL);
3337 LASSERT(obd != NULL);
3338 LASSERT(inode != NULL);
3340 rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3341 sizeof(id->li_fid), EA_SID);
3343 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
3351 * reads inode id on master MDS. This is usualy done by CMOBD to update requests
3352 * to master MDS by correct store cookie, needed to find inode on master MDS
3355 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
3356 struct lustre_id *id)
3361 LASSERT(id != NULL);
3362 LASSERT(obd != NULL);
3363 LASSERT(inode != NULL);
3365 rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_MID);
3367 CERROR("fsfilt_get_md() failed, rc = %d\n", rc);
3380 * updates master inode id. Usualy this is done by CMOBD after an inode is
3381 * created and relationship between cache MDS and master one should be
3384 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
3385 void *handle, struct lustre_id *id)
3390 LASSERT(id != NULL);
3391 LASSERT(obd != NULL);
3392 LASSERT(inode != NULL);
3394 rc = fsfilt_set_md(obd, inode, handle, id,
3395 sizeof(*id), EA_MID);
3397 CERROR("fsfilt_set_md() failed, "
3405 /* mount the file system (secretly) */
3406 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3408 struct lustre_cfg* lcfg = buf;
3409 struct mds_obd *mds = &obd->u.mds;
3410 struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3411 char *options = NULL;
3412 struct vfsmount *mnt;
3418 if (lcfg->lcfg_bufcount < 3)
3419 RETURN(rc = -EINVAL);
3421 if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
3422 RETURN(rc = -EINVAL);
3424 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
3425 if (IS_ERR(obd->obd_fsops))
3426 RETURN(rc = PTR_ERR(obd->obd_fsops));
3428 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3430 page = get_zeroed_page(GFP_KERNEL);
3434 options = (char *)page;
3437 * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3438 * and the rest of options are passed by mount options. Probably this
3439 * should be moved to somewhere else like startup scripts or lconf. */
3440 sprintf(options, "iopen_nopriv");
3442 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0 && lustre_cfg_buf(lcfg, 4))
3443 sprintf(options + strlen(options), ",%s",
3444 lustre_cfg_string(lcfg, 4));
3446 /* we have to know mdsnum before touching underlying fs -bzzz */
3447 atomic_set(&mds->mds_open_count, 0);
3448 sema_init(&mds->mds_md_sem, 1);
3449 sema_init(&mds->mds_create_sem, 1);
3450 mds->mds_md_connected = 0;
3451 mds->mds_md_name = NULL;
3453 if (LUSTRE_CFG_BUFLEN(lcfg, 5) > 0 && lustre_cfg_buf(lcfg, 5) &&
3454 strncmp(lustre_cfg_string(lcfg, 5), "dumb", LUSTRE_CFG_BUFLEN(lcfg, 5))) {
3457 generate_random_uuid(uuid);
3458 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3460 OBD_ALLOC(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3461 if (mds->mds_md_name == NULL)
3462 RETURN(rc = -ENOMEM);
3464 memcpy(mds->mds_md_name, lustre_cfg_buf(lcfg, 5),
3465 LUSTRE_CFG_BUFLEN(lcfg, 5));
3467 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3468 obd->obd_name, mds->mds_md_name);
3470 rc = mds_md_connect(obd, mds->mds_md_name);
3472 OBD_FREE(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3477 mds->mds_obd_type = MDS_MASTER_OBD;
3479 if (LUSTRE_CFG_BUFLEN(lcfg, 6) > 0 && lustre_cfg_buf(lcfg, 6) &&
3480 strncmp(lustre_cfg_string(lcfg, 6), "dumb",
3481 LUSTRE_CFG_BUFLEN(lcfg, 6))) {
3482 if (!memcmp(lustre_cfg_string(lcfg, 6), "master",
3483 strlen("master"))) {
3484 mds->mds_obd_type = MDS_MASTER_OBD;
3485 } else if (!memcmp(lustre_cfg_string(lcfg, 6), "cache",
3487 mds->mds_obd_type = MDS_CACHE_OBD;
3491 rc = lvfs_mount_fs(lustre_cfg_string(lcfg, 1),
3492 lustre_cfg_string(lcfg, 2),
3493 options, 0, &lvfs_ctxt);
3497 if (rc || !lvfs_ctxt) {
3498 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3502 mnt = lvfs_ctxt->loc_mnt;
3503 mds->mds_lvfs_ctxt = lvfs_ctxt;
3504 ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3506 CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
3508 sema_init(&mds->mds_epoch_sem, 1);
3509 atomic_set(&mds->mds_real_clients, 0);
3510 spin_lock_init(&mds->mds_transno_lock);
3511 spin_lock_init(&mds->mds_last_fid_lock);
3512 sema_init(&mds->mds_orphan_recovery_sem, 1);
3513 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3515 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3516 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3518 if (obd->obd_namespace == NULL) {
3519 mds_cleanup(obd, 0);
3520 GOTO(err_put, rc = -ENOMEM);
3522 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3524 rc = mds_fs_setup(obd, mnt);
3526 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3531 rc = llog_start_commit_thread();
3537 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 && lustre_cfg_buf(lcfg, 3) &&
3538 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3539 LUSTRE_CFG_BUFLEN(lcfg, 3))) {
3542 generate_random_uuid(uuid);
3543 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3545 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
3546 if (mds->mds_profile == NULL)
3547 GOTO(err_fs, rc = -ENOMEM);
3549 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
3550 LUSTRE_CFG_BUFLEN(lcfg, 3));
3554 * setup root dir and files ID dir if lmv already connected, or there is
3557 if (mds->mds_md_exp || (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 &&
3558 lustre_cfg_buf(lcfg, 3) &&
3559 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3560 LUSTRE_CFG_BUFLEN(lcfg, 3)))) {
3561 rc = mds_fs_setup_rootid(obd);
3565 rc = mds_fs_setup_virtid(obd);
3570 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3571 "mds_ldlm_client", &obd->obd_ldlm_client);
3572 obd->obd_replayable = 1;
3574 rc = mds_postsetup(obd);
3581 /* No extra cleanup needed for llog_init_commit_thread() */
3582 mds_fs_cleanup(obd, 0);
3584 ldlm_namespace_free(obd->obd_namespace, 0);
3585 obd->obd_namespace = NULL;
3588 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3592 fsfilt_put_ops(obd->obd_fsops);
3596 static int mds_postsetup(struct obd_device *obd)
3598 struct mds_obd *mds = &obd->u.mds;
3602 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3603 obd, 0, NULL, &llog_lvfs_ops);
3607 if (mds->mds_profile) {
3608 struct llog_ctxt *lgctxt;
3609 struct lvfs_run_ctxt saved;
3610 struct lustre_profile *lprof;
3611 struct config_llog_instance cfg;
3613 cfg.cfg_instance = NULL;
3614 cfg.cfg_uuid = mds->mds_dt_uuid;
3615 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3617 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3619 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3620 GOTO(err_llog, rc = -EINVAL);
3623 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3624 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3629 lprof = class_get_profile(mds->mds_profile);
3630 if (lprof == NULL) {
3631 CERROR("No profile found: %s\n", mds->mds_profile);
3632 GOTO(err_cleanup, rc = -ENOENT);
3634 rc = mds_dt_connect(obd, lprof->lp_lov);
3636 GOTO(err_cleanup, rc);
3638 rc = mds_md_postsetup(obd);
3640 GOTO(err_cleanup, rc);
3647 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3648 LLOG_CONFIG_ORIG_CTXT));
3652 int mds_postrecov_common(struct obd_device *obd)
3654 struct mds_obd *mds = &obd->u.mds;
3655 struct llog_ctxt *ctxt;
3656 int rc, item = 0, valsize;
3660 LASSERT(!obd->obd_recovering);
3661 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3662 LASSERT(ctxt != NULL);
3664 /* clean PENDING dir */
3665 rc = mds_cleanup_orphans(obd);
3670 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3671 valsize = sizeof(group);
3672 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
3673 "mds_conn", valsize, &group);
3677 rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3680 CERROR("%s: failed at llog_origin_connect: %d\n",
3685 /* remove the orphaned precreated objects */
3686 rc = mds_dt_clear_orphans(mds, NULL /* all OSTs */);
3691 RETURN(rc < 0 ? rc : item);
3694 /* cleanup all llogging subsystems */
3695 rc = obd_llog_finish(obd, &obd->obd_llogs,
3696 mds->mds_dt_desc.ld_tgt_count);
3698 CERROR("%s: failed to cleanup llogging subsystems\n",
3703 int mds_postrecov(struct obd_device *obd)
3707 rc = mds_postrecov_common(obd);
3709 rc = mds_md_reconnect(obd);
3713 int mds_dt_clean(struct obd_device *obd)
3715 struct mds_obd *mds = &obd->u.mds;
3718 if (mds->mds_profile) {
3720 struct llog_ctxt *llctx;
3721 struct lvfs_run_ctxt saved;
3722 struct config_llog_instance cfg;
3723 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3725 OBD_ALLOC(cln_prof, len);
3726 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3728 cfg.cfg_instance = NULL;
3729 cfg.cfg_uuid = mds->mds_dt_uuid;
3731 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3732 llctx = llog_get_context(&obd->obd_llogs,
3733 LLOG_CONFIG_ORIG_CTXT);
3734 class_config_process_llog(llctx, cln_prof, &cfg);
3735 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3737 OBD_FREE(cln_prof, len);
3738 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3739 mds->mds_profile = NULL;
3744 int mds_md_clean(struct obd_device *obd)
3746 struct mds_obd *mds = &obd->u.mds;
3749 if (mds->mds_md_name) {
3750 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3751 mds->mds_md_name = NULL;
3756 static int mds_precleanup(struct obd_device *obd, int flags)
3762 mds_dt_disconnect(obd, flags);
3764 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3768 extern void lgss_svc_cache_purge_all(void);
3769 static int mds_cleanup(struct obd_device *obd, int flags)
3771 struct mds_obd *mds = &obd->u.mds;
3774 if (mds->mds_sb == NULL)
3777 mds_update_server_data(obd, 1);
3778 mds_update_last_fid(obd, NULL, 1);
3780 if (mds->mds_dt_objids != NULL) {
3781 int size = mds->mds_dt_desc.ld_tgt_count *
3783 OBD_FREE(mds->mds_dt_objids, size);
3785 mds_fs_cleanup(obd, flags);
3789 /* 2 seems normal on mds, (may_umount() also expects 2
3790 fwiw), but we only see 1 at this point in obdfilter. */
3791 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3795 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3797 spin_lock_bh(&obd->obd_processing_task_lock);
3798 if (obd->obd_recovering) {
3799 target_cancel_recovery_timer(obd);
3800 obd->obd_recovering = 0;
3802 spin_unlock_bh(&obd->obd_processing_task_lock);
3805 fsfilt_put_ops(obd->obd_fsops);
3809 lgss_svc_cache_purge_all();
3812 spin_lock(&mds->mds_denylist_lock);
3813 while (!list_empty( &mds->mds_denylist ) ) {
3814 deny_sec_t *p_deny_sec = list_entry(mds->mds_denylist.next,
3816 list_del(&p_deny_sec->list);
3817 OBD_FREE(p_deny_sec, sizeof(*p_deny_sec));
3819 spin_unlock(&mds->mds_denylist_lock);
3824 static int set_security(const char *value, char **sec)
3826 if (!strcmp(value, "null"))
3828 else if (!strcmp(value, "krb5i"))
3830 else if (!strcmp(value, "krb5p"))
3833 CERROR("Unrecognized security flavor %s\n", value);
3840 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
3842 struct lustre_cfg *lcfg = buf;
3843 struct mds_obd *mds = &obd->u.mds;
3847 switch(lcfg->lcfg_command) {
3848 case LCFG_SET_SECURITY: {
3849 if ((LUSTRE_CFG_BUFLEN(lcfg, 1) == 0) ||
3850 (LUSTRE_CFG_BUFLEN(lcfg, 2) == 0))
3851 GOTO(out, rc = -EINVAL);
3853 if (!strcmp(lustre_cfg_string(lcfg, 1), "mds_sec"))
3854 rc = set_security(lustre_cfg_string(lcfg, 2),
3856 else if (!strcmp(lustre_cfg_string(lcfg, 1), "oss_sec"))
3857 rc = set_security(lustre_cfg_string(lcfg, 2),
3859 else if (!strcmp(lustre_cfg_string(lcfg, 1), "deny_sec")){
3860 spin_lock(&mds->mds_denylist_lock);
3861 rc = add_deny_security(lustre_cfg_string(lcfg, 2),
3862 &mds->mds_denylist);
3863 spin_unlock(&mds->mds_denylist_lock);
3865 CERROR("Unrecognized key\n");
3871 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
3872 GOTO(out, rc = -EINVAL);
3878 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3880 struct ldlm_lock *new_lock,
3881 struct ldlm_lock **old_lock,
3882 struct lustre_handle *lockh)
3884 struct obd_export *exp = req->rq_export;
3885 struct obd_device *obd = exp->exp_obd;
3886 struct ldlm_request *dlmreq =
3887 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3888 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3889 struct list_head *iter;
3891 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3894 spin_lock(&obd->obd_namespace->ns_hash_lock);
3895 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3896 struct ldlm_lock *lock;
3897 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3898 if (lock == new_lock)
3900 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3901 lockh->cookie = lock->l_handle.h_cookie;
3902 LDLM_DEBUG(lock, "restoring lock cookie");
3903 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3906 *old_lock = LDLM_LOCK_GET(lock);
3907 spin_unlock(&obd->obd_namespace->ns_hash_lock);
3911 spin_unlock(&obd->obd_namespace->ns_hash_lock);
3913 /* If the xid matches, then we know this is a resent request,
3914 * and allow it. (It's probably an OPEN, for which we don't
3917 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
3921 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
3924 /* This remote handle isn't enqueued, so we never received or
3925 * processed this request. Clear MSG_RESENT, because it can
3926 * be handled like any normal request now. */
3928 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3930 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3934 int intent_disposition(struct ldlm_reply *rep, int flag)
3938 return (rep->lock_policy_res1 & flag);
3941 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3945 rep->lock_policy_res1 |= flag;
3948 static int mds_intent_policy(struct ldlm_namespace *ns,
3949 struct ldlm_lock **lockp, void *req_cookie,
3950 ldlm_mode_t mode, int flags, void *data)
3952 struct ptlrpc_request *req = req_cookie;
3953 struct ldlm_lock *lock = *lockp;
3954 struct ldlm_intent *it;
3955 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3956 struct ldlm_reply *rep;
3957 struct lustre_handle lockh[2] = {{0}, {0}};
3958 struct ldlm_lock *new_lock = NULL;
3959 int getattr_part = MDS_INODELOCK_UPDATE;
3960 int rc, reply_buffers;
3961 int repsize[5] = {sizeof(struct ldlm_reply),
3962 sizeof(struct mds_body),
3963 mds->mds_max_mdsize};
3965 int offset = MDS_REQ_INTENT_REC_OFF;
3968 LASSERT(req != NULL);
3969 MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
3971 if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3972 /* No intent was provided */
3973 int size = sizeof(struct ldlm_reply);
3974 rc = lustre_pack_reply(req, 1, &size, NULL);
3979 it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3980 lustre_swab_ldlm_intent);
3982 CERROR("Intent missing\n");
3983 RETURN(req->rq_status = -EFAULT);
3986 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3989 if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
3990 if (req->rq_export->exp_mds_data.med_remote) {
3992 repsize[3] = sizeof(struct mds_remote_perm);
3996 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
4000 rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
4002 RETURN(req->rq_status = rc);
4004 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
4005 LASSERT(rep != NULL);
4007 intent_set_disposition(rep, DISP_IT_EXECD);
4009 /* execute policy */
4010 switch ((long)it->opc) {
4012 case IT_CREAT|IT_OPEN:
4013 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4015 /* XXX swab here to assert that an mds_open reint
4016 * packet is following */
4017 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4019 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
4021 if (rep->lock_policy_res2) {
4023 * mds_open() returns ENOLCK where it should return
4024 * zero, but it has no lock to return.
4026 if (rep->lock_policy_res2 == ENOLCK)
4027 rep->lock_policy_res2 = 0;
4029 RETURN(ELDLM_LOCK_ABORTED);
4033 * IT_OPEN may return lock on cross-node dentry that we want to
4034 * hold during attr retrival -bzzz
4036 if (lockh[0].cookie == 0)
4037 RETURN(ELDLM_LOCK_ABORTED);
4041 getattr_part = MDS_INODELOCK_LOOKUP;
4044 getattr_part |= MDS_INODELOCK_LOOKUP;
4046 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4047 lock, &new_lock, lockh);
4048 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
4050 /* FIXME: LDLM can set req->rq_status. MDS sets
4051 policy_res{1,2} with disposition and status.
4052 - replay: returns 0 & req->status is old status
4053 - otherwise: returns req->status */
4054 if (intent_disposition(rep, DISP_LOOKUP_NEG))
4055 rep->lock_policy_res2 = 0;
4056 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
4057 rep->lock_policy_res2)
4058 RETURN(ELDLM_LOCK_ABORTED);
4059 if (req->rq_status != 0) {
4061 rep->lock_policy_res2 = req->rq_status;
4062 RETURN(ELDLM_LOCK_ABORTED);
4066 rc = mds_lock_and_check_slave(offset, req, lockh);
4067 if ((rep->lock_policy_res2 = rc)) {
4069 rep->lock_policy_res2 = 0;
4070 RETURN(ELDLM_LOCK_ABORTED);
4074 CERROR("Unhandled intent "LPD64"\n", it->opc);
4078 /* By this point, whatever function we called above must have either
4079 * filled in 'lockh', been an intent replay, or returned an error. We
4080 * want to allow replayed RPCs to not get a lock, since we would just
4081 * drop it below anyways because lock replay is done separately by the
4082 * client afterwards. For regular RPCs we want to give the new lock to
4083 * the client instead of whatever lock it was about to get. */
4084 if (new_lock == NULL)
4085 new_lock = ldlm_handle2lock(&lockh[0]);
4086 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
4089 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
4090 it->opc, lockh[0].cookie);
4092 /* If we've already given this lock to a client once, then we should
4093 * have no readers or writers. Otherwise, we should have one reader
4094 * _or_ writer ref (which will be zeroed below) before returning the
4095 * lock to a client. */
4096 if (new_lock->l_export == req->rq_export) {
4097 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
4099 LASSERT(new_lock->l_export == NULL);
4100 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
4105 if (new_lock->l_export == req->rq_export) {
4106 /* Already gave this to the client, which means that we
4107 * reconstructed a reply. */
4108 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
4110 RETURN(ELDLM_LOCK_REPLACED);
4113 /* Fixup the lock to be given to the client */
4114 lock_res(new_lock->l_resource);
4115 new_lock->l_readers = 0;
4116 new_lock->l_writers = 0;
4118 new_lock->l_export = class_export_get(req->rq_export);
4120 spin_lock(&new_lock->l_export->exp_ldlm_data.led_lock);
4121 list_add(&new_lock->l_export_chain,
4122 &new_lock->l_export->exp_ldlm_data.led_held_locks);
4123 spin_unlock(&new_lock->l_export->exp_ldlm_data.led_lock);
4125 new_lock->l_blocking_ast = lock->l_blocking_ast;
4126 new_lock->l_completion_ast = lock->l_completion_ast;
4128 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
4129 sizeof(lock->l_remote_handle));
4131 new_lock->l_flags &= ~LDLM_FL_LOCAL;
4133 unlock_res(new_lock->l_resource);
4134 LDLM_LOCK_PUT(new_lock);
4136 RETURN(ELDLM_LOCK_REPLACED);
4139 int mds_attach(struct obd_device *dev, obd_count len, void *data)
4141 struct lprocfs_static_vars lvars;
4143 struct mds_obd *mds = &dev->u.mds;
4145 spin_lock_init(&mds->mds_denylist_lock);
4146 INIT_LIST_HEAD(&mds->mds_denylist);
4148 lprocfs_init_multi_vars(0, &lvars);
4150 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
4154 return lprocfs_alloc_md_stats(dev, 0);
4157 int mds_detach(struct obd_device *dev)
4159 lprocfs_free_md_stats(dev);
4160 return lprocfs_obd_detach(dev);
4163 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
4165 struct lprocfs_static_vars lvars;
4167 lprocfs_init_multi_vars(1, &lvars);
4168 return lprocfs_obd_attach(dev, lvars.obd_vars);
4171 int mdt_detach(struct obd_device *dev)
4173 return lprocfs_obd_detach(dev);
4176 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
4178 struct mds_obd *mds = &obd->u.mds;
4183 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4184 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
4185 MDS_SERVICE_WATCHDOG_TIMEOUT,
4186 mds_handle, "mds", obd->obd_proc_entry);
4188 if (!mds->mds_service) {
4189 CERROR("failed to start service\n");
4193 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
4196 GOTO(err_thread, rc);
4198 mds->mds_setattr_service =
4199 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4200 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
4201 MDS_SERVICE_WATCHDOG_TIMEOUT,
4202 mds_handle, "mds_setattr",
4203 obd->obd_proc_entry);
4204 if (!mds->mds_setattr_service) {
4205 CERROR("failed to start getattr service\n");
4206 GOTO(err_thread, rc = -ENOMEM);
4209 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
4210 MDT_NUM_THREADS, "ll_mdt_attr");
4212 GOTO(err_thread2, rc);
4214 mds->mds_readpage_service =
4215 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4216 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
4217 MDS_SERVICE_WATCHDOG_TIMEOUT,
4218 mds_handle, "mds_readpage",
4219 obd->obd_proc_entry);
4220 if (!mds->mds_readpage_service) {
4221 CERROR("failed to start readpage service\n");
4222 GOTO(err_thread2, rc = -ENOMEM);
4225 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
4226 MDT_NUM_THREADS, "ll_mdt_rdpg");
4229 GOTO(err_thread3, rc);
4234 ptlrpc_unregister_service(mds->mds_readpage_service);
4236 ptlrpc_unregister_service(mds->mds_setattr_service);
4238 ptlrpc_unregister_service(mds->mds_service);
4242 static int mdt_cleanup(struct obd_device *obd, int flags)
4244 struct mds_obd *mds = &obd->u.mds;
4247 ptlrpc_stop_all_threads(mds->mds_readpage_service);
4248 ptlrpc_unregister_service(mds->mds_readpage_service);
4250 ptlrpc_stop_all_threads(mds->mds_setattr_service);
4251 ptlrpc_unregister_service(mds->mds_setattr_service);
4253 ptlrpc_stop_all_threads(mds->mds_service);
4254 ptlrpc_unregister_service(mds->mds_service);
4259 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
4260 __u64 gr, void *data)
4262 struct lustre_id id;
4263 struct obd_device *obd = data;
4267 return mds_id2dentry(obd, &id, NULL);
4270 static int mds_get_info(struct obd_export *exp, __u32 keylen,
4271 void *key, __u32 *valsize, void *val)
4273 struct obd_device *obd;
4274 struct mds_obd *mds;
4277 obd = class_exp2obd(exp);
4281 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
4282 exp->exp_handle.h_cookie);
4286 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
4287 /* get log_context handle. */
4288 unsigned long *llh_handle = val;
4289 *valsize = sizeof(unsigned long);
4290 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
4293 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
4294 /* get log_context handle. */
4295 unsigned long *sb = val;
4296 *valsize = sizeof(unsigned long);
4297 *sb = (unsigned long)obd->u.mds.mds_sb;
4301 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
4302 __u32 *mdsize = val;
4303 *valsize = sizeof(*mdsize);
4304 *mdsize = mds->mds_max_mdsize;
4308 if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
4309 __u32 *mdsnum = val;
4310 *valsize = sizeof(*mdsnum);
4311 *mdsnum = mds->mds_num;
4315 if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
4316 struct lustre_id *rootid = val;
4317 *valsize = sizeof(*rootid);
4318 *rootid = mds->mds_rootid;
4322 if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
4323 struct lov_desc *desc = val;
4324 *valsize = sizeof(*desc);
4325 *desc = mds->mds_dt_desc;
4329 CDEBUG(D_IOCTL, "invalid key\n");
4333 struct lvfs_callback_ops mds_lvfs_ops = {
4334 l_id2dentry: mds_lvfs_id2dentry,
4337 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
4338 int objcount, struct obd_ioobj *obj,
4339 int niocount, struct niobuf_remote *nb,
4340 struct niobuf_local *res,
4341 struct obd_trans_info *oti);
4343 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
4344 int objcount, struct obd_ioobj *obj, int niocount,
4345 struct niobuf_local *res, struct obd_trans_info *oti,
4348 /* use obd ops to offer management infrastructure */
4349 static struct obd_ops mds_obd_ops = {
4350 .o_owner = THIS_MODULE,
4351 .o_attach = mds_attach,
4352 .o_detach = mds_detach,
4353 .o_connect = mds_connect,
4354 .o_connect_post = mds_connect_post,
4355 .o_init_export = mds_init_export,
4356 .o_destroy_export = mds_destroy_export,
4357 .o_disconnect = mds_disconnect,
4358 .o_setup = mds_setup,
4359 .o_precleanup = mds_precleanup,
4360 .o_cleanup = mds_cleanup,
4361 .o_process_config = mds_process_config,
4362 .o_postrecov = mds_postrecov,
4363 .o_statfs = mds_obd_statfs,
4364 .o_iocontrol = mds_iocontrol,
4365 .o_create = mds_obd_create,
4366 .o_destroy = mds_obd_destroy,
4367 .o_llog_init = mds_llog_init,
4368 .o_llog_finish = mds_llog_finish,
4369 .o_notify = mds_notify,
4370 .o_get_info = mds_get_info,
4371 .o_set_info = mds_set_info,
4372 .o_preprw = mds_preprw,
4373 .o_commitrw = mds_commitrw,
4376 static struct obd_ops mdt_obd_ops = {
4377 .o_owner = THIS_MODULE,
4378 .o_attach = mdt_attach,
4379 .o_detach = mdt_detach,
4380 .o_setup = mdt_setup,
4381 .o_cleanup = mdt_cleanup,
4384 static int __init mds_init(void)
4386 struct lprocfs_static_vars lvars;
4388 mds_init_lsd_cache();
4390 lprocfs_init_multi_vars(0, &lvars);
4391 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4392 OBD_MDS_DEVICENAME);
4393 lprocfs_init_multi_vars(1, &lvars);
4394 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4395 OBD_MDT_DEVICENAME);
4400 static void /*__exit*/ mds_exit(void)
4402 mds_cleanup_lsd_cache();
4404 class_unregister_type(OBD_MDS_DEVICENAME);
4405 class_unregister_type(OBD_MDT_DEVICENAME);
4408 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4409 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4410 MODULE_LICENSE("GPL");
4412 module_init(mds_init);
4413 module_exit(mds_exit);