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_close_xid = 0;
610 mcd->mcd_last_result = 0;
611 mcd->mcd_last_data = 0;
614 if (!(flags & OBD_OPT_MDS_CONNECTION)) {
615 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
616 atomic_inc(&mds->mds_real_clients);
617 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
618 obd->obd_name, exp->exp_client_uuid.uuid,
619 atomic_read(&mds->mds_real_clients));
620 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
622 if (mds->mds_md_name)
623 rc = mds_md_connect(obd, mds->mds_md_name);
628 static int mds_init_export(struct obd_export *exp)
630 struct mds_export_data *med = &exp->exp_mds_data;
632 INIT_LIST_HEAD(&med->med_open_head);
633 spin_lock_init(&med->med_open_lock);
637 static int mds_destroy_export(struct obd_export *export)
639 struct obd_device *obd = export->exp_obd;
640 struct mds_export_data *med = &export->exp_mds_data;
641 struct lvfs_run_ctxt saved;
645 mds_free_export_data(med);
646 target_destroy_export(export);
648 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
651 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
653 /* Close any open files (which may also cause orphan unlinking). */
654 spin_lock(&med->med_open_lock);
655 while (!list_empty(&med->med_open_head)) {
656 struct list_head *tmp = med->med_open_head.next;
657 struct mds_file_data *mfd =
658 list_entry(tmp, struct mds_file_data, mfd_list);
659 struct lustre_id sid;
661 BDEVNAME_DECLARE_STORAGE(btmp);
663 /* bug 1579: fix force-closing for 2.5 */
664 struct dentry *dentry = mfd->mfd_dentry;
666 list_del(&mfd->mfd_list);
667 spin_unlock(&med->med_open_lock);
669 down(&dentry->d_inode->i_sem);
670 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
671 up(&dentry->d_inode->i_sem);
673 CERROR("Can't read inode self id, inode %lu, "
674 "rc %d\n", dentry->d_inode->i_ino, rc);
675 memset(&sid, 0, sizeof(sid));
678 /* If you change this message, be sure to update
679 * replay_single:test_46 */
680 CERROR("force closing client file handle for %.*s (%s:"
681 DLID4")\n", dentry->d_name.len, dentry->d_name.name,
682 ll_bdevname(dentry->d_inode->i_sb, btmp),
685 /* child inode->i_alloc_sem protects orphan_dec_test and
686 * is_orphan race, mds_mfd_close drops it */
687 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
688 rc = mds_mfd_close(NULL, 0, obd, mfd,
689 !(export->exp_flags & OBD_OPT_FAILOVER));
691 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
692 spin_lock(&med->med_open_lock);
694 spin_unlock(&med->med_open_lock);
695 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
699 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
703 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
705 unsigned long irqflags;
706 struct obd_device *obd;
711 LASSERT(exp != NULL);
712 obd = class_exp2obd(exp);
714 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
715 exp->exp_handle.h_cookie);
721 * suppress any inter-mds requests durring disconnecting lmv if this is
722 * detected --force mode. This is needed to avoid endless recovery.
724 if (atomic_read(&mds->mds_real_clients) > 0 &&
725 !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
726 flags |= OBD_OPT_FORCE;
728 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
729 && !atomic_read(&mds->mds_real_clients)) {
730 /* there was no client at all */
731 mds_md_disconnect(obd, flags);
734 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
735 && atomic_dec_and_test(&mds->mds_real_clients)) {
736 /* time to drop LMV connections */
737 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
738 "Disconnnect from LMV now\n",
739 obd->obd_name, exp->exp_client_uuid.uuid);
740 mds_md_disconnect(obd, flags);
743 spin_lock_irqsave(&exp->exp_lock, irqflags);
744 exp->exp_flags = flags;
745 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
747 /* disconnect early so that clients can't keep using export */
748 rc = class_disconnect(exp, flags);
749 ldlm_cancel_locks_for_export(exp);
751 /* complete all outstanding replies */
752 spin_lock_irqsave(&exp->exp_lock, irqflags);
753 while (!list_empty(&exp->exp_outstanding_replies)) {
754 struct ptlrpc_reply_state *rs =
755 list_entry(exp->exp_outstanding_replies.next,
756 struct ptlrpc_reply_state, rs_exp_list);
757 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
759 spin_lock(&svc->srv_lock);
760 list_del_init(&rs->rs_exp_list);
761 ptlrpc_schedule_difficult_reply(rs);
762 spin_unlock(&svc->srv_lock);
764 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
768 static int mds_getstatus(struct ptlrpc_request *req)
770 struct mds_obd *mds = mds_req2mds(req);
771 struct mds_body *body;
775 size = sizeof(*body);
777 rc = lustre_pack_reply(req, 1, &size, NULL);
778 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
779 CERROR("mds: out of memory for message: size=%d\n", size);
780 req->rq_status = -ENOMEM; /* superfluous? */
784 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
785 body->valid |= OBD_MD_FID;
787 memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
790 * the last_committed and last_xid fields are filled in for all replies
791 * already - no need to do so here also.
796 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
797 void *data, int flag)
802 if (flag == LDLM_CB_CANCELING) {
803 /* Don't need to do anything here. */
807 /* XXX layering violation! -phil */
808 lock_res_and_lock(lock);
811 * get this: if mds_blocking_ast is racing with mds_intent_policy, such
812 * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
813 * then by the time we get the lock, we might not be the correct
814 * blocking function anymore. So check, and return early, if so.
816 if (lock->l_blocking_ast != mds_blocking_ast) {
817 unlock_res_and_lock(lock);
821 lock->l_flags |= LDLM_FL_CBPENDING;
822 do_ast = (!lock->l_readers && !lock->l_writers);
823 unlock_res_and_lock(lock);
826 struct lustre_handle lockh;
829 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
830 ldlm_lock2handle(lock, &lockh);
831 rc = ldlm_cli_cancel(&lockh);
833 CERROR("ldlm_cli_cancel: %d\n", rc);
835 LDLM_DEBUG(lock, "Lock still has references, will be "
841 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
842 void *md, int size, int mea)
846 if (S_ISREG(inode->i_mode)) {
847 rc = mds_convert_lov_ea(obd, inode, md, size);
848 } else if (S_ISDIR(inode->i_mode)) {
850 rc = mds_convert_mea_ea(obd, inode, md, size);
852 rc = mds_convert_lov_ea(obd, inode, md, size);
855 CERROR("Invalid EA format (nor LOV or MEA) "
856 "is detected. Inode %lu/%u\n",
857 inode->i_ino, inode->i_generation);
863 int mds_get_md(struct obd_device *obd, struct inode *inode,
864 void *md, int *size, int lock, int mea)
873 rc = fsfilt_get_md(obd, inode, md, *size,
874 (mea ? EA_MEA : EA_LOV));
876 CERROR("Error %d reading eadata for ino %lu\n",
880 rc = mds_convert_md(obd, inode, md,
895 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
896 * Call with lock=0 if the caller has already taken the i_sem. */
897 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
898 struct mds_body *body, struct inode *inode, int lock, int mea)
900 struct mds_obd *mds = &obd->u.mds;
905 lmm = lustre_msg_buf(msg, offset, 0);
907 /* Some problem with getting eadata when I sized the reply
909 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
913 lmm_size = msg->buflens[offset];
915 /* I don't really like this, but it is a sanity check on the client
916 * MD request. However, if the client doesn't know how much space
917 * to reserve for the MD, it shouldn't be bad to have too much space.
919 if (lmm_size > mds->mds_max_mdsize) {
920 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
921 inode->i_ino, lmm_size, mds->mds_max_mdsize);
925 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
927 body->valid |= S_ISDIR(inode->i_mode) ?
928 OBD_MD_FLDIREA : OBD_MD_FLEASIZE;
931 body->valid |= OBD_MD_MEA;
933 body->eadatasize = lmm_size;
940 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
941 struct mds_body *repbody, int reply_off)
943 struct inode *inode = dentry->d_inode;
948 symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
949 LASSERT(symname != NULL);
950 len = req->rq_repmsg->buflens[reply_off + 1];
952 rc = inode->i_op->readlink(dentry, symname, len);
954 CERROR("readlink failed: %d\n", rc);
955 } else if (rc != len - 1) {
956 CERROR ("Unexpected readlink rc %d: expecting %d\n",
960 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
961 repbody->valid |= OBD_MD_LINKNAME;
962 repbody->eadatasize = rc + 1;
963 symname[rc] = 0; /* NULL terminate */
970 int mds_pack_xattr(struct dentry *dentry, struct ptlrpc_request *req,
971 struct mds_body *repbody, int req_off, int reply_off)
973 struct inode *inode = dentry->d_inode;
979 ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
980 len = req->rq_repmsg->buflens[reply_off + 1];
982 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
986 if (!strcmp(ea_name, XATTR_NAME_LUSTRE_ACL)) {
987 struct rmtacl_upcall_desc desc;
989 if (len != LUSTRE_ACL_SIZE_MAX || !value) {
990 CERROR("no reply buffer prepared\n");
994 memset(&desc, 0, sizeof(desc));
996 desc.cmd = lustre_msg_string(req->rq_reqmsg, req_off + 2, 0);
997 desc.cmdlen = req->rq_reqmsg->buflens[req_off + 2];
998 desc.res = (char *) value;
999 desc.reslen = LUSTRE_ACL_SIZE_MAX;
1001 mds_do_remote_acl_upcall(&desc);
1003 if (desc.upcall_status)
1004 RETURN(desc.upcall_status);
1006 if (desc.reslen > LUSTRE_ACL_SIZE_MAX) {
1007 CERROR("downcall claim reslen %u\n", desc.reslen);
1010 /* like remote setfacl, steal "flags" in mds_body as the
1011 * exececution status
1013 repbody->flags = desc.status;
1014 repbody->valid |= OBD_MD_FLXATTR;
1015 repbody->eadatasize = desc.reslen;
1020 if (inode->i_op && inode->i_op->getxattr)
1021 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
1024 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1025 CERROR("getxattr failed: %d", rc);
1027 repbody->valid |= OBD_MD_FLXATTR;
1028 repbody->eadatasize = rc;
1035 int mds_pack_xattr_list(struct dentry *dentry, struct ptlrpc_request *req,
1036 struct mds_body *repbody, int reply_off)
1038 struct inode *inode = dentry->d_inode;
1043 len = req->rq_repmsg->buflens[reply_off + 1];
1045 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
1048 if (inode->i_op && inode->i_op->getxattr)
1049 rc = inode->i_op->listxattr(dentry, value, len);
1052 CERROR("listxattr failed: %d", rc);
1054 repbody->valid |= OBD_MD_FLXATTRLIST;
1055 repbody->eadatasize = rc;
1062 int mds_pack_posix_acl(struct lustre_msg *repmsg, int offset,
1063 struct mds_body *body, struct inode *inode)
1065 struct dentry de = { .d_inode = inode };
1066 __u32 buflen, *sizep;
1071 if (!inode->i_op->getxattr)
1074 buflen = repmsg->buflens[offset + 1];
1075 buf = lustre_msg_buf(repmsg, offset + 1, buflen);
1077 size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1078 if (size == -ENODATA || size == -EOPNOTSUPP)
1084 sizep = lustre_msg_buf(repmsg, offset, 4);
1086 CERROR("can't locate returned acl size buf\n");
1090 *sizep = cpu_to_le32(size);
1091 body->valid |= OBD_MD_FLACL;
1097 int mds_pack_remote_perm(struct ptlrpc_request *req, int reply_off,
1098 struct mds_body *body, struct inode *inode)
1100 struct lustre_sec_desc *lsd;
1101 struct mds_remote_perm *perm;
1104 LASSERT(inode->i_op);
1105 LASSERT(inode->i_op->permission);
1106 LASSERT(req->rq_export->exp_mds_data.med_remote);
1108 perm = (struct mds_remote_perm *)
1109 lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(perm));
1113 memset(perm, 0, sizeof(*perm));
1115 /* obtain authenticated uid/gid and LSD permissions, which
1116 * might be different from current process context, from LSD
1118 lsd = mds_get_lsd(current->uid);
1120 CWARN("can't LSD of uid %u\n", current->uid);
1124 perm->mrp_auth_uid = lsd->lsd_uid;
1125 perm->mrp_auth_gid = lsd->lsd_gid;
1127 lsd_perms = mds_lsd_get_perms(lsd, 1, 0, req->rq_peer.peer_id.nid);
1128 if (lsd_perms & LSD_PERM_SETUID)
1129 perm->mrp_allow_setuid = 1;
1130 if (lsd_perms & LSD_PERM_SETGID)
1131 perm->mrp_allow_setgid = 1;
1135 /* permission bits of current user
1136 * XXX this is low efficient, could we do it in one blow?
1138 if (inode->i_op->permission(inode, MAY_EXEC, NULL) == 0)
1139 perm->mrp_perm |= MAY_EXEC;
1140 if (inode->i_op->permission(inode, MAY_WRITE, NULL) == 0)
1141 perm->mrp_perm |= MAY_WRITE;
1142 if (inode->i_op->permission(inode, MAY_READ, NULL) == 0)
1143 perm->mrp_perm |= MAY_READ;
1145 body->valid |= (OBD_MD_FLACL | OBD_MD_FLRMTACL);
1150 int mds_pack_acl(struct ptlrpc_request *req, int reply_off,
1151 struct mds_body *body, struct inode *inode)
1155 if (!req->rq_export->exp_mds_data.med_remote)
1156 rc = mds_pack_posix_acl(req->rq_repmsg, reply_off, body, inode);
1158 rc = mds_pack_remote_perm(req, reply_off, body, inode);
1163 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1164 struct ptlrpc_request *req, int req_off,
1165 struct mds_body *reqbody, int reply_off,
1166 struct mds_req_sec_desc *rsd)
1168 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1169 struct inode *inode = dentry->d_inode;
1170 struct mds_body *body;
1174 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1177 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1178 LASSERT(body != NULL); /* caller prepped reply */
1180 if (dentry->d_flags & DCACHE_CROSS_REF) {
1181 mds_pack_dentry2body(obd, body, dentry,
1182 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1183 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1188 mds_pack_inode2body(obd, body, inode,
1189 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1191 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1192 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1194 /* guessing what kind og attribute do we need. */
1195 int is_mea = (S_ISDIR(inode->i_mode) &&
1196 (reqbody->valid & OBD_MD_MEA) != 0);
1198 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
1199 body, inode, 1, is_mea);
1201 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1202 if (!(body->valid & OBD_MD_FLEASIZE) &&
1203 !(body->valid & OBD_MD_FLDIREA))
1204 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1205 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1206 } else if (S_ISLNK(inode->i_mode) &&
1207 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1208 rc = mds_pack_link(dentry, req, body, reply_off);
1209 } else if (reqbody->valid & OBD_MD_FLXATTR) {
1210 rc = mds_pack_xattr(dentry, req, body, req_off, reply_off);
1211 } else if (reqbody->valid & OBD_MD_FLXATTRLIST) {
1212 rc = mds_pack_xattr_list(dentry, req, body, reply_off);
1215 if (reqbody->valid & OBD_MD_FLACL) {
1216 int inc = (reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1;
1217 rc = mds_pack_acl(req, reply_off + inc, body, inode);
1221 mds_body_do_reverse_map(med, body);
1226 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1227 struct dentry *dentry,
1230 int rc = 0, size[1] = {sizeof(struct mds_body)};
1233 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1234 CERROR("failed MDS_GETATTR_PACK test\n");
1235 req->rq_status = -ENOMEM;
1239 rc = lustre_pack_reply(req, 1, size, NULL);
1241 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1242 GOTO(out, req->rq_status = rc);
1250 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1253 struct inode *inode = de->d_inode;
1254 struct mds_obd *mds = mds_req2mds(req);
1255 struct mds_body *body;
1256 int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1259 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1260 LASSERT(body != NULL); /* checked by caller */
1261 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
1263 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1264 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1267 down(&inode->i_sem);
1268 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1269 ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1272 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1273 CERROR("error getting inode %lu MD: rc = %d\n",
1276 } else if (rc > mds->mds_max_mdsize) {
1278 CERROR("MD size %d larger than maximum possible %u\n",
1279 rc, mds->mds_max_mdsize);
1281 size[bufcount] = rc;
1284 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1285 if (inode->i_size + 1 != body->eadatasize)
1286 CERROR("symlink size: %Lu, reply space: %d\n",
1287 inode->i_size + 1, body->eadatasize);
1288 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1290 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1291 inode->i_size + 1, body->eadatasize);
1292 } else if ((body->valid & OBD_MD_FLXATTR)) {
1293 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1297 if (!strcmp(ea_name, XATTR_NAME_LUSTRE_ACL)) {
1298 size[bufcount] = LUSTRE_ACL_SIZE_MAX;
1300 if (inode->i_op && inode->i_op->getxattr)
1301 rc = inode->i_op->getxattr(de, ea_name,
1305 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1306 CERROR("error get inode %lu EA: %d\n",
1310 size[bufcount] = min_t(int,
1311 body->eadatasize, rc);
1315 } else if (body->valid & OBD_MD_FLXATTRLIST) {
1317 if (inode->i_op && inode->i_op->getxattr)
1318 rc = inode->i_op->listxattr(de, NULL, 0);
1321 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1322 CERROR("error getting inode %lu EA: rc = %d\n",
1326 size[bufcount] = min_t(int, body->eadatasize, rc);
1331 /* may co-exist with OBD_MD_FLEASIZE */
1332 if (body->valid & OBD_MD_FLACL) {
1333 if (req->rq_export->exp_mds_data.med_remote) {
1334 size[bufcount++] = sizeof(struct mds_remote_perm);
1336 size[bufcount++] = 4;
1337 size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1341 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1342 CERROR("failed MDS_GETATTR_PACK test\n");
1343 req->rq_status = -ENOMEM;
1344 GOTO(out, rc = -ENOMEM);
1347 rc = lustre_pack_reply(req, bufcount, size, NULL);
1349 CERROR("out of memory\n");
1350 GOTO(out, req->rq_status = rc);
1358 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1359 char *name, int namelen)
1361 struct mea *mea = NULL;
1362 int mea_size, rc = 0;
1365 rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1370 * dir is already splitted, check if requested filename should
1371 * live at this MDS or at another one.
1373 int i = mea_name2idx(mea, name, namelen - 1);
1374 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1376 "inapropriate MDS(%d) for %s. should be "
1377 "%lu(%d)\n", mea->mea_master, name,
1378 (unsigned long)id_group(&mea->mea_ids[i]), i);
1384 OBD_FREE(mea, mea_size);
1388 int mds_getattr_size(struct obd_device *obd, struct dentry *dentry,
1389 struct ptlrpc_request *req, struct mds_body *body)
1391 struct inode *inode = dentry->d_inode;
1394 LASSERT(body != NULL);
1396 if (dentry->d_inode == NULL || !S_ISREG(inode->i_mode))
1399 if (obd->obd_recovering) {
1400 CDEBUG(D_ERROR, "size for "DLID4" is unknown yet (recovering)\n",
1405 if (atomic_read(&inode->i_writecount)) {
1406 /* some one has opened the file for write.
1407 * mds doesn't know actual size */
1408 CDEBUG(D_OTHER, "MDS doesn't know actual size for "DLID4"\n",
1412 CDEBUG(D_OTHER, "MDS returns "LPD64"/"LPD64" for"DLID4"\n",
1413 body->size, body->blocks, OLID4(&body->id1));
1414 body->valid |= OBD_MD_FLSIZE;
1418 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1419 struct lustre_handle *child_lockh, int child_part)
1421 struct obd_device *obd = req->rq_export->exp_obd;
1422 struct mds_obd *mds = &obd->u.mds;
1423 struct ldlm_reply *rep = NULL;
1424 struct lvfs_run_ctxt saved;
1425 struct mds_req_sec_desc *rsd;
1426 struct mds_body *body;
1427 struct dentry *dparent = NULL, *dchild = NULL;
1428 struct lvfs_ucred uc = {NULL, NULL,};
1429 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1430 unsigned int namesize;
1431 int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1435 LASSERT(!strcmp(obd->obd_type->typ_name, OBD_MDS_DEVICENAME));
1436 MD_COUNTER_INCREMENT(obd, getattr_lock);
1438 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1440 CERROR("Can't unpack security desc\n");
1444 /* swab now, before anyone looks inside the request. */
1445 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1446 lustre_swab_mds_body);
1448 CERROR("Can't swab mds_body\n");
1449 GOTO(cleanup, rc = -EFAULT);
1452 LASSERT_REQSWAB(req, offset + 1);
1453 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1455 CERROR("Can't unpack name\n");
1456 GOTO(cleanup, rc = -EFAULT);
1458 namesize = req->rq_reqmsg->buflens[offset + 1];
1460 /* namesize less than 2 means we have empty name, probably came from
1461 revalidate by cfid, so no point in having name to be set */
1465 LASSERT (offset == 1 || offset == 3);
1467 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1473 rc = mds_init_ucred(&uc, req, rsd);
1478 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1479 cleanup_phase = 1; /* kernel context */
1480 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1482 LASSERT(namesize > 0);
1483 if (child_lockh->cookie != 0) {
1484 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1488 if (body->valid == OBD_MD_FLID) {
1489 struct mds_body *mds_reply;
1490 int size = sizeof(*mds_reply);
1494 dparent = mds_id2dentry(obd, &body->id1, NULL);
1495 if (IS_ERR(dparent)) {
1496 rc = PTR_ERR(dparent);
1500 * the user requested ONLY the inode number, so do a raw lookup.
1502 rc = lustre_pack_reply(req, 1, &size, NULL);
1504 CERROR("out of memory\n");
1508 dir = dparent->d_inode;
1509 LASSERT(dir->i_op->lookup_raw != NULL);
1510 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1512 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1513 sizeof(*mds_reply));
1515 id_ino(&mds_reply->id1) = inum;
1516 mds_reply->valid = OBD_MD_FLID;
1520 if (resent_req == 0) {
1521 LASSERT(id_fid(&body->id1) != 0);
1523 rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1524 parent_lockh, &dparent,
1526 MDS_INODELOCK_UPDATE,
1529 child_lockh, &dchild,
1530 LCK_PR, child_part);
1534 cleanup_phase = 2; /* dchild, dparent, locks */
1537 * let's make sure this name should leave on this mds
1540 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1544 /* we have no dentry here, drop LOOKUP bit */
1545 /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1546 child_part &= ~MDS_INODELOCK_LOOKUP;
1547 CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1548 obd->obd_name, OLID4(&body->id1));
1550 dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1551 LCK_PR, parent_lockh,
1554 MDS_INODELOCK_UPDATE);
1555 if (IS_ERR(dchild)) {
1556 CERROR("can't find inode with id "DLID4", err = %d\n",
1557 OLID4(&body->id1), (int)PTR_ERR(dchild));
1558 GOTO(cleanup, rc = PTR_ERR(dchild));
1560 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1563 struct ldlm_lock *granted_lock;
1565 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1566 granted_lock = ldlm_handle2lock(child_lockh);
1568 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1569 id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1570 child_lockh->cookie);
1573 /* usual named request */
1574 dparent = mds_id2dentry(obd, &body->id1, NULL);
1575 LASSERT(!IS_ERR(dparent));
1576 dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1577 LASSERT(!IS_ERR(dchild));
1579 /* client wants to get attr. by id */
1580 dchild = mds_id2dentry(obd, &body->id1, NULL);
1581 LASSERT(!IS_ERR(dchild));
1583 LDLM_LOCK_PUT(granted_lock);
1586 cleanup_phase = 2; /* dchild, dparent, locks */
1588 if (!DENTRY_VALID(dchild)) {
1589 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1591 * in the intent case, the policy clears this error: the
1592 * disposition is enough.
1597 intent_set_disposition(rep, DISP_LOOKUP_POS);
1600 if (req->rq_repmsg == NULL) {
1601 if (dchild->d_flags & DCACHE_CROSS_REF)
1602 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1604 rc = mds_getattr_pack_msg(req, dchild, offset);
1606 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1611 rc = mds_getattr_internal(obd, dchild, req, offset, body,
1614 GOTO(cleanup, rc); /* returns the lock to the client */
1616 /* probably MDS knows actual size? */
1617 body = lustre_msg_buf(req->rq_repmsg, reply_offset, sizeof(*body));
1618 LASSERT(body != NULL);
1619 mds_getattr_size(obd, dchild, req, body);
1624 switch (cleanup_phase) {
1626 if (resent_req == 0) {
1627 if (rc && DENTRY_VALID(dchild))
1628 ldlm_lock_decref(child_lockh, LCK_PR);
1630 ldlm_lock_decref(parent_lockh, LCK_PR);
1632 if (parent_lockh[1].cookie != 0)
1633 ldlm_lock_decref(parent_lockh + 1, update_mode);
1640 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1642 mds_exit_ucred(&uc);
1647 static int mds_getattr(struct ptlrpc_request *req, int offset)
1649 struct obd_device *obd = req->rq_export->exp_obd;
1650 struct lvfs_run_ctxt saved;
1652 struct mds_req_sec_desc *rsd;
1653 struct mds_body *body;
1654 struct lvfs_ucred uc = {NULL, NULL,};
1658 MD_COUNTER_INCREMENT(obd, getattr);
1660 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1662 CERROR("Can't unpack security desc\n");
1666 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1667 lustre_swab_mds_body);
1669 CERROR ("Can't unpack body\n");
1673 rc = mds_init_ucred(&uc, req, rsd);
1675 mds_exit_ucred(&uc);
1679 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1680 de = mds_id2dentry(obd, &body->id1, NULL);
1682 rc = req->rq_status = PTR_ERR(de);
1686 rc = mds_getattr_pack_msg(req, de, offset);
1688 CERROR("mds_getattr_pack_msg: %d\n", rc);
1692 req->rq_status = mds_getattr_internal(obd, de, req, offset, body,
1698 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1699 mds_exit_ucred(&uc);
1703 static int mds_access_check(struct ptlrpc_request *req, int offset)
1705 struct obd_device *obd = req->rq_export->exp_obd;
1706 struct lvfs_run_ctxt saved;
1708 struct mds_req_sec_desc *rsd;
1709 struct mds_body *body;
1710 struct lvfs_ucred uc;
1711 int rep_size[2] = {sizeof(*body),
1712 sizeof(struct mds_remote_perm)};
1716 if (!req->rq_export->exp_mds_data.med_remote) {
1717 CERROR("from local client "LPU64"\n", req->rq_peer.peer_id.nid);
1721 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1723 CERROR("Can't unpack security desc\n");
1727 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1728 lustre_swab_mds_body);
1730 CERROR ("Can't unpack body\n");
1734 MD_COUNTER_INCREMENT(obd, access_check);
1736 rc = mds_init_ucred(&uc, req, rsd);
1738 CERROR("init ucred error: %d\n", rc);
1741 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1743 de = mds_id2dentry(obd, &body->id1, NULL);
1745 CERROR("grab ino "LPU64": err %ld\n",
1746 body->id1.li_stc.u.e3s.l3s_ino, PTR_ERR(de));
1747 GOTO(out_pop, rc = PTR_ERR(de));
1750 rc = lustre_pack_reply(req, 2, rep_size, NULL);
1752 CERROR("pack reply error: %d\n", rc);
1753 GOTO(out_dput, rc = -EINVAL);
1756 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1759 rc = mds_pack_remote_perm(req, 1, body, de->d_inode);
1765 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1766 mds_exit_ucred(&uc);
1770 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1771 unsigned long max_age)
1776 spin_lock(&obd->obd_osfs_lock);
1777 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1779 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1780 spin_unlock(&obd->obd_osfs_lock);
1785 static int mds_statfs(struct ptlrpc_request *req)
1787 struct obd_device *obd = req->rq_export->exp_obd;
1788 int rc, size = sizeof(struct obd_statfs);
1791 /* This will trigger a watchdog timeout */
1792 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1793 (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1795 rc = lustre_pack_reply(req, 1, &size, NULL);
1796 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1797 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1801 OBD_COUNTER_INCREMENT(obd, statfs);
1803 /* We call this so that we can cache a bit - 1 jiffie worth */
1804 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1807 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1813 req->rq_status = rc;
1817 static int mds_sync(struct ptlrpc_request *req, int offset)
1819 struct obd_device *obd = req->rq_export->exp_obd;
1820 struct mds_obd *mds = &obd->u.mds;
1821 struct mds_body *body;
1822 int rc, size = sizeof(*body);
1825 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1826 lustre_swab_mds_body);
1828 GOTO(out, rc = -EPROTO);
1830 rc = lustre_pack_reply(req, 1, &size, NULL);
1831 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1832 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1836 if (id_ino(&body->id1) == 0) {
1837 /* an id of zero is taken to mean "sync whole filesystem" */
1838 rc = fsfilt_sync(obd, mds->mds_sb);
1842 /* just any file to grab fsync method - "file" arg unused */
1843 struct file *file = mds->mds_rcvd_filp;
1844 struct mds_body *rep_body;
1847 de = mds_id2dentry(obd, &body->id1, NULL);
1849 GOTO(out, rc = PTR_ERR(de));
1851 rc = file->f_op->fsync(NULL, de, 1);
1855 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1856 mds_pack_inode2body(obd, rep_body, de->d_inode,
1857 (body->valid & OBD_MD_FID) ? 1 : 0);
1863 req->rq_status = rc;
1867 /* mds_readpage does not take a DLM lock on the inode, because the client must
1868 * already have a PR lock.
1870 * If we were to take another one here, a deadlock will result, if another
1871 * thread is already waiting for a PW lock. */
1872 static int mds_readpage(struct ptlrpc_request *req, int offset)
1874 struct obd_device *obd = req->rq_export->exp_obd;
1875 struct vfsmount *mnt;
1878 struct mds_req_sec_desc *rsd;
1879 struct mds_body *body, *repbody;
1880 struct lvfs_run_ctxt saved;
1881 int rc, size = sizeof(*repbody);
1882 struct lvfs_ucred uc = {NULL, NULL,};
1885 rc = lustre_pack_reply(req, 1, &size, NULL);
1886 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1887 CERROR("mds: out of memory\n");
1888 GOTO(out, rc = -ENOMEM);
1891 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1893 CERROR("Can't unpack security desc\n");
1894 GOTO (out, rc = -EFAULT);
1897 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1898 lustre_swab_mds_body);
1900 CERROR("Can't unpack body\n");
1901 GOTO (out, rc = -EFAULT);
1904 rc = mds_init_ucred(&uc, req, rsd);
1909 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1910 de = mds_id2dentry(obd, &body->id1, &mnt);
1912 GOTO(out_pop, rc = PTR_ERR(de));
1914 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1916 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1917 /* note: in case of an error, dentry_open puts dentry */
1919 GOTO(out_pop, rc = PTR_ERR(file));
1921 /* body->size is actually the offset -eeb */
1922 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1923 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1924 body->size, de->d_inode->i_blksize);
1925 GOTO(out_file, rc = -EFAULT);
1928 /* body->nlink is actually the #bytes to read -eeb */
1929 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1930 CERROR("size %u is not multiple of blocksize %lu\n",
1931 body->nlink, de->d_inode->i_blksize);
1932 GOTO(out_file, rc = -EFAULT);
1935 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1936 repbody->size = file->f_dentry->d_inode->i_size;
1937 repbody->valid = OBD_MD_FLSIZE;
1939 /* to make this asynchronous make sure that the handling function
1940 doesn't send a reply when this function completes. Instead a
1941 callback function would send the reply */
1942 /* body->size is actually the offset -eeb */
1943 rc = mds_sendpage(req, file, body->size, body->nlink);
1947 filp_close(file, 0);
1949 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1951 mds_exit_ucred(&uc);
1952 req->rq_status = rc;
1956 /* update master MDS ID, which is stored in local inode EA. */
1957 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1958 void *data, int data_len)
1960 struct mds_obd *mds = &obd->u.mds;
1961 struct dentry *dentry;
1969 dentry = mds_id2dentry(obd, id, NULL);
1971 GOTO(out, rc = PTR_ERR(dentry));
1973 if (!dentry->d_inode) {
1974 CERROR("Can't find object "DLID4".\n",
1976 GOTO(out_dentry, rc = -EINVAL);
1979 handle = fsfilt_start(obd, dentry->d_inode,
1980 FSFILT_OP_SETATTR, NULL);
1982 GOTO(out_dentry, rc = PTR_ERR(handle));
1984 rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1985 (struct lustre_id *)data);
1987 CERROR("Can't update inode "DLID4" master id, "
1988 "error = %d.\n", OLID4(id), rc);
1989 GOTO(out_commit, rc);
1994 fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
2001 EXPORT_SYMBOL(mds_update_mid);
2003 /* read master MDS ID, which is stored in local inode EA. */
2004 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
2005 void *data, int data_len)
2007 struct dentry *dentry;
2014 dentry = mds_id2dentry(obd, id, NULL);
2016 GOTO(out, rc = PTR_ERR(dentry));
2018 if (!dentry->d_inode) {
2019 CERROR("Can't find object "DLID4".\n",
2021 GOTO(out_dentry, rc = -EINVAL);
2024 down(&dentry->d_inode->i_sem);
2025 rc = mds_read_inode_mid(obd, dentry->d_inode,
2026 (struct lustre_id *)data);
2027 up(&dentry->d_inode->i_sem);
2029 CERROR("Can't read inode "DLID4" master id, "
2030 "error = %d.\n", OLID4(id), rc);
2031 GOTO(out_dentry, rc);
2040 EXPORT_SYMBOL(mds_read_mid);
2042 int mds_read_md(struct obd_device *obd, struct lustre_id *id,
2043 char **data, int *datalen)
2045 struct dentry *dentry;
2046 struct mds_obd *mds = &obd->u.mds;
2047 int rc = 0, mea = 0;
2054 dentry = mds_id2dentry(obd, id, NULL);
2056 GOTO(out, rc = PTR_ERR(dentry));
2058 if (!dentry->d_inode) {
2059 CERROR("Can't find object "DLID4".\n",
2061 GOTO(out_dentry, rc = -EINVAL);
2063 if (S_ISDIR(dentry->d_inode->i_mode)) {
2064 *datalen = obd_packmd(mds->mds_md_exp, NULL, NULL);
2067 *datalen = obd_packmd(mds->mds_dt_exp, NULL, NULL);
2070 OBD_ALLOC(ea, *datalen);
2073 GOTO(out_dentry, rc = PTR_ERR(dentry));
2076 down(&dentry->d_inode->i_sem);
2077 rc = fsfilt_get_md(obd, dentry->d_inode, *data, *datalen,
2078 (mea ? EA_MEA : EA_LOV));
2079 up(&dentry->d_inode->i_sem);
2082 CERROR("Error %d reading eadata for ino %lu\n",
2083 rc, dentry->d_inode->i_ino);
2089 EXPORT_SYMBOL(mds_read_md);
2091 int mds_reint(struct ptlrpc_request *req, int offset,
2092 struct lustre_handle *lockh)
2094 struct mds_update_record *rec;
2095 struct mds_req_sec_desc *rsd;
2099 OBD_ALLOC(rec, sizeof(*rec));
2103 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
2105 CERROR("Can't unpack security desc\n");
2106 GOTO(out, rc = -EFAULT);
2109 rc = mds_update_unpack(req, offset, rec);
2110 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
2111 CERROR("invalid record\n");
2112 GOTO(out, req->rq_status = -EINVAL);
2115 rc = mds_init_ucred(&rec->ur_uc, req, rsd);
2120 /* rc will be used to interrupt a for loop over multiple records */
2121 rc = mds_reint_rec(rec, offset, req, lockh);
2124 mds_exit_ucred(&rec->ur_uc);
2125 OBD_FREE(rec, sizeof(*rec));
2129 static int mds_filter_recovery_request(struct ptlrpc_request *req,
2130 struct obd_device *obd, int *process)
2132 switch (req->rq_reqmsg->opc) {
2133 case MDS_CONNECT: /* This will never get here, but for completeness. */
2134 case OST_CONNECT: /* This will never get here, but for completeness. */
2135 case MDS_DISCONNECT:
2136 case OST_DISCONNECT:
2141 case MDS_SYNC: /* used in unmounting */
2146 *process = target_queue_recovery_request(req, obd);
2150 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
2152 /* XXX what should we set rq_status to here? */
2153 req->rq_status = -EAGAIN;
2154 RETURN(ptlrpc_error(req));
2158 static char *reint_names[] = {
2159 [REINT_SETATTR] "setattr",
2160 [REINT_CREATE] "create",
2161 [REINT_LINK] "link",
2162 [REINT_UNLINK] "unlink",
2163 [REINT_RENAME] "rename",
2164 [REINT_OPEN] "open",
2167 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER | \
2168 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
2169 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
2172 static void reconstruct_create(struct ptlrpc_request *req)
2174 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2175 struct mds_client_data *mcd = med->med_mcd;
2176 struct dentry *dentry;
2177 struct ost_body *body;
2178 struct lustre_id id;
2182 /* copy rc, transno and disp; steal locks */
2183 mds_req_from_mcd(req, mcd);
2184 if (req->rq_status) {
2192 id_ino(&id) = mcd->mcd_last_data;
2193 LASSERT(id_ino(&id) != 0);
2195 dentry = mds_id2dentry(req2obd(req), &id, NULL);
2196 if (IS_ERR(dentry)) {
2197 CERROR("can't find inode "LPU64"\n", id_ino(&id));
2198 req->rq_status = PTR_ERR(dentry);
2203 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
2204 req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
2205 dentry->d_inode->i_generation);
2207 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2208 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
2209 body->oa.o_id = dentry->d_inode->i_ino;
2210 body->oa.o_generation = dentry->d_inode->i_generation;
2211 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2213 down(&dentry->d_inode->i_sem);
2214 rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
2215 up(&dentry->d_inode->i_sem);
2217 CERROR("Can't read inode self id, inode %lu, "
2218 "rc %d\n", dentry->d_inode->i_ino, rc);
2222 body->oa.o_fid = id_fid(&id);
2223 body->oa.o_mds = id_group(&id);
2229 static int mds_inode_init_acl(struct obd_device *obd, void *handle,
2230 struct dentry *de, void *xattr, int xattr_size)
2232 struct inode *inode = de->d_inode;
2233 struct posix_acl *acl;
2240 LASSERT(xattr_size > 0);
2242 if (!inode->i_op->getxattr || !inode->i_op->setxattr) {
2243 CERROR("backend fs dosen't support xattr\n");
2247 /* set default acl */
2248 if (S_ISDIR(inode->i_mode)) {
2249 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_DEFAULT,
2250 xattr, xattr_size, 0);
2252 CERROR("set default acl err: %d\n", rc);
2257 /* set access acl */
2258 acl = posix_acl_from_xattr(xattr, xattr_size);
2259 if (acl == NULL || IS_ERR(acl)) {
2260 CERROR("insane attr data\n");
2261 return PTR_ERR(acl);
2264 if (posix_acl_valid(acl)) {
2265 CERROR("default acl not valid: %d\n", rc);
2270 mode = inode->i_mode;
2271 rc = posix_acl_create_masq(acl, &mode);
2273 CERROR("create masq err %d\n", rc);
2277 if (inode->i_mode != mode) {
2278 struct iattr iattr = { .ia_valid = ATTR_MODE,
2282 rc2 = fsfilt_setattr(obd, de, handle, &iattr, 0);
2284 CERROR("setattr mode err: %d\n", rc2);
2291 /* we didn't change acl except mode bits of some
2292 * entries, so should be fit into original size.
2294 rc = posix_acl_to_xattr(acl, xattr, xattr_size);
2297 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_ACCESS,
2298 xattr, xattr_size, 0);
2300 CERROR("set access acl err: %d\n", rc);
2303 posix_acl_release(acl);
2307 static int mdt_obj_create(struct ptlrpc_request *req)
2309 struct obd_device *obd = req->rq_export->exp_obd;
2310 struct mds_obd *mds = &obd->u.mds;
2311 struct ost_body *body, *repbody;
2314 char idname[LL_ID_NAMELEN];
2315 int size = sizeof(*repbody);
2316 struct inode *parent_inode;
2317 struct lvfs_run_ctxt saved;
2318 int rc, cleanup_phase = 0;
2319 struct dentry *new = NULL;
2320 struct dentry_params dp;
2321 int mealen, flags = 0;
2322 struct lvfs_ucred uc;
2323 struct lustre_id id;
2325 void *handle = NULL;
2326 unsigned long cr_inum = 0;
2330 DEBUG_REQ(D_HA, req, "create remote object");
2331 parent_inode = mds->mds_unnamed_dir->d_inode;
2333 body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2334 lustre_swab_ost_body);
2338 /* acl data is packed transparently, no swab here */
2339 LASSERT(req->rq_reqmsg->bufcount >= 2);
2340 acl_size = req->rq_reqmsg->buflens[1];
2342 acl = lustre_msg_buf(req->rq_reqmsg, 1, acl_size);
2344 CERROR("No default acl buf?\n");
2349 rc = lustre_pack_reply(req, 1, &size, NULL);
2353 MDS_CHECK_RESENT(req, reconstruct_create(req));
2356 uc.luc_ginfo = NULL;
2357 uc.luc_uid = body->oa.o_uid;
2358 uc.luc_gid = body->oa.o_gid;
2359 uc.luc_fsuid = body->oa.o_uid;
2360 uc.luc_fsgid = body->oa.o_gid;
2362 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2363 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2365 /* in REPLAY case inum should be given (client or other MDS fills it) */
2366 if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2367 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2369 * this is re-create request from MDS holding directory name.
2370 * we have to lookup given ino/gen first. if it exists (good
2371 * case) then there is nothing to do. if it does not then we
2372 * have to recreate it.
2374 id_ino(&id) = body->oa.o_id;
2375 id_gen(&id) = body->oa.o_generation;
2377 new = mds_id2dentry(obd, &id, NULL);
2378 if (!IS_ERR(new) && new->d_inode) {
2379 struct lustre_id sid;
2381 CDEBUG(D_OTHER, "mkdir repairing %lu/%lu\n",
2382 (unsigned long)id_ino(&id),
2383 (unsigned long)id_gen(&id));
2385 obdo_from_inode(&repbody->oa, new->d_inode,
2386 FILTER_VALID_FLAGS);
2388 repbody->oa.o_id = new->d_inode->i_ino;
2389 repbody->oa.o_generation = new->d_inode->i_generation;
2390 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2393 down(&new->d_inode->i_sem);
2394 rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2395 up(&new->d_inode->i_sem);
2397 CERROR("Can't read inode self id "
2398 "inode %lu, rc %d.\n",
2399 new->d_inode->i_ino, rc);
2403 repbody->oa.o_fid = id_fid(&sid);
2404 repbody->oa.o_mds = id_group(&sid);
2405 LASSERT(id_fid(&sid) != 0);
2408 * here we could use fid passed in body->oa.o_fid and
2409 * thus avoid mds_read_inode_sid().
2411 cr_inum = new->d_inode->i_ino;
2412 GOTO(cleanup, rc = 0);
2416 down(&parent_inode->i_sem);
2417 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2418 if (IS_ERR(handle)) {
2419 up(&parent_inode->i_sem);
2420 CERROR("fsfilt_start() failed, rc = %d\n",
2421 (int)PTR_ERR(handle));
2422 GOTO(cleanup, rc = PTR_ERR(handle));
2424 cleanup_phase = 1; /* transaction */
2427 rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2428 new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2430 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2431 obd->obd_name, idname, (int) PTR_ERR(new));
2432 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2433 up(&parent_inode->i_sem);
2434 RETURN(PTR_ERR(new));
2435 } else if (new->d_inode) {
2436 CERROR("%s: name exists. repeat\n", obd->obd_name);
2439 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2440 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2441 fid = body->oa.o_fid;
2443 fid = mds_alloc_fid(obd);
2445 new->d_fsdata = (void *)&dp;
2449 dp.p_group = mds->mds_num;
2451 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2452 (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2453 LASSERT(body->oa.o_id != 0);
2454 dp.p_inum = body->oa.o_id;
2455 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2456 (unsigned long)body->oa.o_id,
2457 (unsigned long)body->oa.o_generation);
2460 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2463 rc = mds_inode_init_acl(obd, handle, new,
2466 up(&parent_inode->i_sem);
2470 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2471 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2472 new->d_inode->i_generation = body->oa.o_generation;
2473 mark_inode_dirty(new->d_inode);
2476 * avoiding asserts in cache flush case, as
2477 * @body->oa.o_id should be zero.
2479 if (body->oa.o_id) {
2480 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2481 "BUG 3550: failed to recreate obj "
2482 LPU64" -> %lu\n", body->oa.o_id,
2483 new->d_inode->i_ino);
2485 LASSERTF(body->oa.o_generation ==
2486 new->d_inode->i_generation,
2487 "BUG 3550: failed to recreate obj/gen "
2488 LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2489 body->oa.o_generation,
2490 new->d_inode->i_ino,
2491 new->d_inode->i_generation);
2495 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2496 repbody->oa.o_id = new->d_inode->i_ino;
2497 repbody->oa.o_generation = new->d_inode->i_generation;
2498 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2500 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2501 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2502 id_group(&id) = mds->mds_num;
2504 LASSERT(body->oa.o_fid != 0);
2505 id_fid(&id) = body->oa.o_fid;
2507 LASSERT(body->oa.o_id != 0);
2508 id_ino(&id) = repbody->oa.o_id;
2509 id_gen(&id) = repbody->oa.o_generation;
2511 down(&new->d_inode->i_sem);
2512 rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
2513 up(&new->d_inode->i_sem);
2516 * make sure, that fid is up-to-date.
2518 mds_set_last_fid(obd, id_fid(&id));
2521 * allocate new sid, as object is created from scratch
2522 * and this is not replay.
2524 down(&new->d_inode->i_sem);
2525 rc = mds_set_inode_sid(obd, new->d_inode, handle, &id, fid);
2526 up(&new->d_inode->i_sem);
2529 CERROR("Can't update lustre ID for inode %lu, "
2530 "error = %d\n", new->d_inode->i_ino, rc);
2534 /* initializing o_fid after it is allocated. */
2535 repbody->oa.o_fid = id_fid(&id);
2536 repbody->oa.o_mds = id_group(&id);
2538 rc = fsfilt_del_dir_entry(obd, new);
2539 up(&parent_inode->i_sem);
2541 CERROR("can't remove name for object: %d\n", rc);
2545 cleanup_phase = 2; /* created directory object */
2547 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2548 (unsigned long)new->d_inode->i_ino,
2549 (unsigned long)new->d_inode->i_generation,
2550 (unsigned)new->d_inode->i_mode);
2551 cr_inum = new->d_inode->i_ino;
2553 up(&parent_inode->i_sem);
2554 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2558 if (body->oa.o_valid & OBD_MD_FLID) {
2559 /* this is new object for splitted dir. We have to prevent
2560 * recursive splitting on it -bzzz */
2561 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2563 OBD_ALLOC(mea, mealen);
2565 GOTO(cleanup, rc = -ENOMEM);
2567 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2568 mea->mea_master = 0;
2571 down(&new->d_inode->i_sem);
2572 rc = fsfilt_set_md(obd, new->d_inode, handle,
2573 mea, mealen, EA_MEA);
2574 up(&new->d_inode->i_sem);
2576 CERROR("fsfilt_set_md() failed, "
2579 OBD_FREE(mea, mealen);
2581 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2582 obd->obd_name, new->d_inode->i_ino,
2583 new->d_inode->i_generation, flags);
2584 } else if (body->oa.o_easize) {
2585 /* we pass LCK_EX to split routine to signal that we have
2586 * exclusive access to the directory. simple because nobody
2587 * knows it already exists -bzzz */
2588 rc = mds_try_to_split_dir(obd, new, NULL,
2589 body->oa.o_easize, LCK_EX);
2591 CERROR("Can't split directory %lu, error = %d.\n",
2592 new->d_inode->i_ino, rc);
2600 switch (cleanup_phase) {
2601 case 2: /* object has been created, but we'll may want to replay it later */
2603 ptlrpc_require_repack(req);
2604 case 1: /* transaction */
2605 rc = mds_finish_transno(mds, parent_inode, handle,
2610 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2614 static int mdt_get_info(struct ptlrpc_request *req)
2616 struct obd_export *exp = req->rq_export;
2621 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2623 DEBUG_REQ(D_HA, req, "no get_info key");
2626 keylen = req->rq_reqmsg->buflens[0];
2628 if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2629 (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2630 (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2633 if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2634 struct lustre_id *reply;
2635 int size = sizeof(*reply);
2637 rc = lustre_pack_reply(req, 1, &size, NULL);
2641 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2642 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2645 int size = sizeof(*reply);
2647 rc = lustre_pack_reply(req, 1, &size, NULL);
2651 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2652 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2655 req->rq_repmsg->status = 0;
2659 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2660 void *key, __u32 vallen, void *val)
2662 struct obd_device *obd;
2663 struct mds_obd *mds;
2667 obd = class_exp2obd(exp);
2669 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2670 exp->exp_handle.h_cookie);
2675 if (keylen >= strlen("mds_type") &&
2676 memcmp(key, "mds_type", keylen) == 0) {
2680 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
2682 mds->mds_obd_type = *(int*)val;
2683 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
2684 valsize = sizeof(group);
2686 /* mds number has been changed, so the corresponding obdfilter
2687 * exp need to be changed too. */
2688 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
2689 "mds_conn", valsize, &group);
2692 CDEBUG(D_IOCTL, "invalid key\n");
2696 static int mdt_set_info(struct ptlrpc_request *req)
2699 struct obd_export *exp = req->rq_export;
2700 int keylen, rc = 0, vallen;
2703 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2705 DEBUG_REQ(D_HA, req, "no set_info key");
2708 keylen = req->rq_reqmsg->buflens[0];
2710 if (keylen == strlen("mds_type") &&
2711 memcmp(key, "mds_type", keylen) == 0) {
2712 rc = lustre_pack_reply(req, 0, NULL, NULL);
2716 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2717 vallen = req->rq_reqmsg->buflens[1];
2719 rc = obd_set_info(exp, keylen, key, vallen, val);
2720 req->rq_repmsg->status = 0;
2723 CDEBUG(D_IOCTL, "invalid key\n");
2727 static void mds_revoke_export_locks(struct obd_export *exp)
2729 struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2730 struct list_head work;
2731 struct ldlm_lock *lock, *next;
2732 struct ldlm_lock_desc desc;
2734 if (!exp->u.eu_mds_data.med_remote)
2738 CERROR("implement right locking here! -bzzz\n");
2739 INIT_LIST_HEAD(&work);
2740 spin_lock(&exp->exp_ldlm_data.led_lock);
2741 list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2743 lock_res_and_lock(lock);
2744 if (lock->l_req_mode != lock->l_granted_mode) {
2745 unlock_res_and_lock(lock);
2749 LASSERT(lock->l_resource);
2750 if (lock->l_resource->lr_type != LDLM_IBITS &&
2751 lock->l_resource->lr_type != LDLM_PLAIN) {
2752 unlock_res_and_lock(lock);
2756 if (lock->l_flags & LDLM_FL_AST_SENT) {
2757 unlock_res_and_lock(lock);
2761 lock->l_flags |= LDLM_FL_AST_SENT;
2762 unlock_res_and_lock(lock);
2764 /* the desc just pretend to exclusive */
2765 ldlm_lock2desc(lock, &desc);
2766 desc.l_req_mode = LCK_EX;
2767 desc.l_granted_mode = 0;
2769 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2771 spin_unlock(&exp->exp_ldlm_data.led_lock);
2776 static int mds_msg_check_version(struct lustre_msg *msg)
2782 case MDS_DISCONNECT:
2784 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2786 CERROR("bad opc %u version %08x, expecting %08x\n",
2787 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2792 case MDS_GETATTR_LOCK:
2793 case MDS_ACCESS_CHECK:
2797 case MDS_DONE_WRITING:
2800 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2802 CERROR("bad opc %u version %08x, expecting %08x\n",
2803 msg->opc, msg->version, LUSTRE_MDS_VERSION);
2807 case LDLM_BL_CALLBACK:
2808 case LDLM_CP_CALLBACK:
2809 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2811 CERROR("bad opc %u version %08x, expecting %08x\n",
2812 msg->opc, msg->version, LUSTRE_DLM_VERSION);
2814 case OBD_LOG_CANCEL:
2815 case LLOG_ORIGIN_HANDLE_OPEN:
2816 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2817 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2818 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2819 case LLOG_ORIGIN_HANDLE_CLOSE:
2821 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2823 CERROR("bad opc %u version %08x, expecting %08x\n",
2824 msg->opc, msg->version, LUSTRE_LOG_VERSION);
2830 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2832 CERROR("bad opc %u version %08x, expecting %08x\n",
2833 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2836 case SEC_INIT_CONTINUE:
2841 CERROR("MDS unknown opcode %d\n", msg->opc);
2849 int mds_handle(struct ptlrpc_request *req)
2851 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2852 struct obd_device *obd = NULL;
2853 struct mds_obd *mds = NULL; /* quell gcc overwarning */
2857 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2859 rc = mds_msg_check_version(req->rq_reqmsg);
2861 CERROR("MDS drop mal-formed request\n");
2865 /* Security opc should NOT trigger any recovery events */
2866 if (req->rq_reqmsg->opc == SEC_INIT ||
2867 req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
2868 if (req->rq_export) {
2869 mds_req_add_idmapping(req,
2870 &req->rq_export->exp_mds_data);
2871 mds_revoke_export_locks(req->rq_export);
2874 } else if (req->rq_reqmsg->opc == SEC_FINI) {
2875 if (req->rq_export) {
2876 mds_req_del_idmapping(req,
2877 &req->rq_export->exp_mds_data);
2878 mds_revoke_export_locks(req->rq_export);
2883 LASSERT(current->journal_info == NULL);
2884 /* XXX identical to OST */
2885 if (req->rq_reqmsg->opc != MDS_CONNECT) {
2886 struct mds_export_data *med;
2889 if (req->rq_export == NULL) {
2890 CERROR("operation %d on unconnected MDS from %s\n",
2891 req->rq_reqmsg->opc,
2893 req->rq_status = -ENOTCONN;
2894 GOTO(out, rc = -ENOTCONN);
2897 med = &req->rq_export->exp_mds_data;
2898 obd = req->rq_export->exp_obd;
2901 /* sanity check: if the xid matches, the request must
2902 * be marked as a resent or replayed */
2903 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
2904 req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
2905 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2906 (MSG_RESENT | MSG_REPLAY),
2907 "rq_xid "LPU64" matches last_xid, "
2908 "expected RESENT flag\n",
2911 /* else: note the opposite is not always true; a
2912 * RESENT req after a failover will usually not match
2913 * the last_xid, since it was likely never
2914 * committed. A REPLAYed request will almost never
2915 * match the last xid, however it could for a
2916 * committed, but still retained, open. */
2918 spin_lock_bh(&obd->obd_processing_task_lock);
2919 recovering = obd->obd_recovering;
2920 spin_unlock_bh(&obd->obd_processing_task_lock);
2922 rc = mds_filter_recovery_request(req, obd,
2924 if (rc || should_process == 0) {
2926 } else if (should_process < 0) {
2927 req->rq_status = should_process;
2928 rc = ptlrpc_error(req);
2934 switch (req->rq_reqmsg->opc) {
2936 DEBUG_REQ(D_INODE, req, "connect");
2937 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2938 rc = target_handle_connect(req);
2940 struct mds_export_data *med;
2942 LASSERT(req->rq_export);
2943 med = &req->rq_export->u.eu_mds_data;
2944 mds_init_export_data(req, med);
2945 mds_req_add_idmapping(req, med);
2947 /* Now that we have an export, set mds. */
2948 obd = req->rq_export->exp_obd;
2949 mds = mds_req2mds(req);
2953 case MDS_DISCONNECT:
2954 DEBUG_REQ(D_INODE, req, "disconnect");
2955 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2956 rc = target_handle_disconnect(req);
2957 req->rq_status = rc; /* superfluous? */
2961 DEBUG_REQ(D_INODE, req, "getstatus");
2962 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2963 rc = mds_getstatus(req);
2967 DEBUG_REQ(D_INODE, req, "getattr");
2968 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2969 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2972 case MDS_ACCESS_CHECK:
2973 DEBUG_REQ(D_INODE, req, "access_check");
2974 OBD_FAIL_RETURN(OBD_FAIL_MDS_ACCESS_CHECK_NET, 0);
2975 rc = mds_access_check(req, MDS_REQ_REC_OFF);
2978 case MDS_GETATTR_LOCK: {
2979 struct lustre_handle lockh;
2980 DEBUG_REQ(D_INODE, req, "getattr_lock");
2981 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2983 /* If this request gets a reconstructed reply, we won't be
2984 * acquiring any new locks in mds_getattr_lock, so we don't
2988 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2989 MDS_INODELOCK_UPDATE);
2990 /* this non-intent call (from an ioctl) is special */
2991 req->rq_status = rc;
2992 if (rc == 0 && lockh.cookie)
2993 ldlm_lock_decref(&lockh, LCK_PR);
2997 DEBUG_REQ(D_INODE, req, "statfs");
2998 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2999 rc = mds_statfs(req);
3003 DEBUG_REQ(D_INODE, req, "readpage");
3004 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
3005 rc = mds_readpage(req, MDS_REQ_REC_OFF);
3007 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
3008 if (req->rq_reply_state) {
3009 lustre_free_reply_state (req->rq_reply_state);
3010 req->rq_reply_state = NULL;
3017 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
3020 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
3021 mds->mds_max_cookiesize};
3024 /* NB only peek inside req now; mds_reint() will swab it */
3026 CERROR ("Can't inspect opcode\n");
3031 if (lustre_msg_swabbed (req->rq_reqmsg))
3034 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
3035 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
3036 reint_names[opc] == NULL) ? reint_names[opc] :
3039 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
3041 if (opc == REINT_UNLINK || opc == REINT_RENAME)
3043 else if (opc == REINT_OPEN)
3048 /* for SETATTR: I have different reply setting for
3049 * remote setfacl, so delay the reply buffer allocation.
3051 if (opc != REINT_SETATTR) {
3052 rc = lustre_pack_reply(req, bufcount, size, NULL);
3057 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
3058 fail = OBD_FAIL_MDS_REINT_NET_REP;
3063 DEBUG_REQ(D_INODE, req, "close");
3064 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
3065 rc = mds_close(req, MDS_REQ_REC_OFF);
3068 case MDS_DONE_WRITING:
3069 DEBUG_REQ(D_INODE, req, "done_writing");
3070 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
3071 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
3075 DEBUG_REQ(D_INODE, req, "pin");
3076 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
3077 rc = mds_pin(req, MDS_REQ_REC_OFF);
3081 DEBUG_REQ(D_INODE, req, "sync");
3082 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
3083 rc = mds_sync(req, MDS_REQ_REC_OFF);
3087 DEBUG_REQ(D_INODE, req, "ping");
3088 rc = target_handle_ping(req);
3091 case OBD_LOG_CANCEL:
3092 CDEBUG(D_INODE, "log cancel\n");
3093 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
3094 rc = -ENOTSUPP; /* la la la */
3098 DEBUG_REQ(D_INODE, req, "enqueue");
3099 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
3100 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
3101 ldlm_server_blocking_ast, NULL);
3102 fail = OBD_FAIL_LDLM_REPLY;
3105 DEBUG_REQ(D_INODE, req, "convert");
3106 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
3107 rc = ldlm_handle_convert(req);
3109 case LDLM_BL_CALLBACK:
3110 case LDLM_CP_CALLBACK:
3111 DEBUG_REQ(D_INODE, req, "callback");
3112 CERROR("callbacks should not happen on MDS\n");
3114 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
3116 case LLOG_ORIGIN_HANDLE_OPEN:
3117 DEBUG_REQ(D_INODE, req, "llog_init");
3118 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3119 rc = llog_origin_handle_open(req);
3121 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
3122 DEBUG_REQ(D_INODE, req, "llog next block");
3123 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3124 rc = llog_origin_handle_next_block(req);
3126 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
3127 DEBUG_REQ(D_INODE, req, "llog prev block");
3128 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3129 rc = llog_origin_handle_prev_block(req);
3131 case LLOG_ORIGIN_HANDLE_READ_HEADER:
3132 DEBUG_REQ(D_INODE, req, "llog read header");
3133 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3134 rc = llog_origin_handle_read_header(req);
3136 case LLOG_ORIGIN_HANDLE_CLOSE:
3137 DEBUG_REQ(D_INODE, req, "llog close");
3138 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3139 rc = llog_origin_handle_close(req);
3142 DEBUG_REQ(D_INODE, req, "ost_create");
3143 rc = mdt_obj_create(req);
3146 DEBUG_REQ(D_INODE, req, "get_info");
3147 rc = mdt_get_info(req);
3150 DEBUG_REQ(D_INODE, req, "set_info");
3151 rc = mdt_set_info(req);
3154 CDEBUG(D_INODE, "write\n");
3155 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
3156 rc = ost_brw_write(req, NULL);
3157 LASSERT(current->journal_info == NULL);
3158 /* mdt_brw sends its own replies */
3162 DEBUG_REQ(D_INODE, req, "llog catinfo");
3163 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3164 rc = llog_catinfo(req);
3167 req->rq_status = -ENOTSUPP;
3168 rc = ptlrpc_error(req);
3172 LASSERT(current->journal_info == NULL);
3176 /* If we're DISCONNECTing, the mds_export_data is already freed */
3177 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
3178 struct mds_export_data *med = &req->rq_export->exp_mds_data;
3179 struct obd_device *obd = list_entry(mds, struct obd_device,
3181 req->rq_repmsg->last_xid =
3182 le64_to_cpu(med->med_mcd->mcd_last_xid);
3184 if (!obd->obd_no_transno) {
3185 req->rq_repmsg->last_committed =
3186 obd->obd_last_committed;
3188 DEBUG_REQ(D_IOCTL, req,
3189 "not sending last_committed update");
3191 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
3193 mds->mds_last_transno, obd->obd_last_committed,
3199 target_send_reply(req, rc, fail);
3203 /* Update the server data on disk. This stores the new mount_count and also the
3204 * last_rcvd value to disk. If we don't have a clean shutdown, then the server
3205 * last_rcvd value may be less than that of the clients. This will alert us
3206 * that we may need to do client recovery.
3208 * Also assumes for mds_last_transno that we are not modifying it (no locking).
3210 int mds_update_server_data(struct obd_device *obd, int force_sync)
3212 struct mds_obd *mds = &obd->u.mds;
3213 struct mds_server_data *msd = mds->mds_server_data;
3214 struct file *filp = mds->mds_rcvd_filp;
3215 struct lvfs_run_ctxt saved;
3220 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3221 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
3223 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
3224 mds->mds_mount_count, mds->mds_last_transno);
3225 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
3227 CERROR("error writing MDS server data: rc = %d\n", rc);
3228 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3233 /* saves last allocated fid counter to file. */
3234 int mds_update_last_fid(struct obd_device *obd, void *handle,
3237 struct mds_obd *mds = &obd->u.mds;
3238 struct file *filp = mds->mds_fid_filp;
3239 struct lvfs_run_ctxt saved;
3245 spin_lock(&mds->mds_last_fid_lock);
3246 last_fid = mds->mds_last_fid;
3247 spin_unlock(&mds->mds_last_fid_lock);
3249 CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
3253 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
3254 handle, mds_commit_last_fid_cb,
3258 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3259 rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
3261 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3264 CERROR("error writing MDS last_fid #"LPU64
3265 ", err = %d\n", last_fid, rc);
3269 CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
3270 "%llu: err = %d\n", last_fid, off, rc);
3275 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
3277 struct mds_obd *mds = &obd->u.mds;
3279 spin_lock(&mds->mds_last_fid_lock);
3280 if (fid > mds->mds_last_fid)
3281 mds->mds_last_fid = fid;
3282 spin_unlock(&mds->mds_last_fid_lock);
3285 void mds_commit_last_transno_cb(struct obd_device *obd,
3286 __u64 transno, void *data,
3289 obd_transno_commit_cb(obd, transno, error);
3292 void mds_commit_last_fid_cb(struct obd_device *obd,
3293 __u64 fid, void *data,
3297 CERROR("%s: fid "LPD64" commit error: %d\n",
3298 obd->obd_name, fid, error);
3302 CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
3303 obd->obd_name, fid);
3306 __u64 mds_alloc_fid(struct obd_device *obd)
3308 struct mds_obd *mds = &obd->u.mds;
3311 spin_lock(&mds->mds_last_fid_lock);
3312 fid = ++mds->mds_last_fid;
3313 spin_unlock(&mds->mds_last_fid_lock);
3318 * update new lustre_id on passed @inode and saves it to inode EA.
3320 int mds_set_inode_sid(struct obd_device *obd, struct inode *inode,
3321 void *handle, struct lustre_id *id, __u64 fid)
3323 struct mds_obd *mds = &obd->u.mds;
3324 int alloc = 0, rc = 0;
3327 LASSERT(obd != NULL);
3328 LASSERT(inode != NULL);
3331 OBD_ALLOC(id, sizeof(*id));
3336 id_group(id) = mds->mds_num;
3338 id_ino(id) = inode->i_ino;
3339 id_gen(id) = inode->i_generation;
3340 id_type(id) = (S_IFMT & inode->i_mode);
3342 rc = mds_update_inode_sid(obd, inode, handle, id);
3344 CERROR("Can't update inode FID EA, "
3349 OBD_FREE(id, sizeof(*id));
3354 * reads inode self id from inode EA. Probably later this should be replaced by
3355 * caching inode self id to avoid raeding it every time it is needed.
3357 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
3358 struct lustre_id *id)
3363 LASSERT(id != NULL);
3364 LASSERT(obd != NULL);
3365 LASSERT(inode != NULL);
3367 rc = fsfilt_get_md(obd, inode, &id->li_fid,
3368 sizeof(id->li_fid), EA_SID);
3370 CERROR("fsfilt_get_md() failed, "
3383 /* updates inode self id in EA. */
3384 int mds_update_inode_sid(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->li_fid,
3395 sizeof(id->li_fid), EA_SID);
3397 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
3405 * reads inode id on master MDS. This is usualy done by CMOBD to update requests
3406 * to master MDS by correct store cookie, needed to find inode on master MDS
3409 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
3410 struct lustre_id *id)
3415 LASSERT(id != NULL);
3416 LASSERT(obd != NULL);
3417 LASSERT(inode != NULL);
3419 rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_MID);
3421 CERROR("fsfilt_get_md() failed, rc = %d\n", rc);
3434 * updates master inode id. Usualy this is done by CMOBD after an inode is
3435 * created and relationship between cache MDS and master one should be
3438 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
3439 void *handle, struct lustre_id *id)
3444 LASSERT(id != NULL);
3445 LASSERT(obd != NULL);
3446 LASSERT(inode != NULL);
3448 rc = fsfilt_set_md(obd, inode, handle, id,
3449 sizeof(*id), EA_MID);
3451 CERROR("fsfilt_set_md() failed, "
3459 /* mount the file system (secretly) */
3460 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3462 struct lustre_cfg* lcfg = buf;
3463 struct mds_obd *mds = &obd->u.mds;
3464 struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3465 char *options = NULL;
3466 struct vfsmount *mnt;
3472 if (lcfg->lcfg_bufcount < 3)
3473 RETURN(rc = -EINVAL);
3475 if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
3476 RETURN(rc = -EINVAL);
3478 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
3479 if (IS_ERR(obd->obd_fsops))
3480 RETURN(rc = PTR_ERR(obd->obd_fsops));
3482 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3484 page = get_zeroed_page(GFP_KERNEL);
3488 options = (char *)page;
3491 * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3492 * and the rest of options are passed by mount options. Probably this
3493 * should be moved to somewhere else like startup scripts or lconf. */
3494 sprintf(options, "iopen_nopriv");
3496 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0 && lustre_cfg_buf(lcfg, 4))
3497 sprintf(options + strlen(options), ",%s",
3498 lustre_cfg_string(lcfg, 4));
3500 /* we have to know mdsnum before touching underlying fs -bzzz */
3501 atomic_set(&mds->mds_open_count, 0);
3502 sema_init(&mds->mds_md_sem, 1);
3503 sema_init(&mds->mds_create_sem, 1);
3504 mds->mds_md_connected = 0;
3505 mds->mds_md_name = NULL;
3507 if (LUSTRE_CFG_BUFLEN(lcfg, 5) > 0 && lustre_cfg_buf(lcfg, 5) &&
3508 strncmp(lustre_cfg_string(lcfg, 5), "dumb", LUSTRE_CFG_BUFLEN(lcfg, 5))) {
3511 generate_random_uuid(uuid);
3512 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3514 OBD_ALLOC(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3515 if (mds->mds_md_name == NULL)
3516 RETURN(rc = -ENOMEM);
3518 memcpy(mds->mds_md_name, lustre_cfg_buf(lcfg, 5),
3519 LUSTRE_CFG_BUFLEN(lcfg, 5));
3521 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3522 obd->obd_name, mds->mds_md_name);
3524 rc = mds_md_connect(obd, mds->mds_md_name);
3526 OBD_FREE(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3531 mds->mds_obd_type = MDS_MASTER_OBD;
3533 if (LUSTRE_CFG_BUFLEN(lcfg, 6) > 0 && lustre_cfg_buf(lcfg, 6) &&
3534 strncmp(lustre_cfg_string(lcfg, 6), "dumb",
3535 LUSTRE_CFG_BUFLEN(lcfg, 6))) {
3536 if (!memcmp(lustre_cfg_string(lcfg, 6), "master",
3537 strlen("master"))) {
3538 mds->mds_obd_type = MDS_MASTER_OBD;
3539 } else if (!memcmp(lustre_cfg_string(lcfg, 6), "cache",
3541 mds->mds_obd_type = MDS_CACHE_OBD;
3545 rc = lvfs_mount_fs(lustre_cfg_string(lcfg, 1),
3546 lustre_cfg_string(lcfg, 2),
3547 options, 0, &lvfs_ctxt);
3551 if (rc || !lvfs_ctxt) {
3552 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3556 mnt = lvfs_ctxt->loc_mnt;
3557 mds->mds_lvfs_ctxt = lvfs_ctxt;
3558 ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3560 CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
3562 sema_init(&mds->mds_epoch_sem, 1);
3563 atomic_set(&mds->mds_real_clients, 0);
3564 spin_lock_init(&mds->mds_transno_lock);
3565 spin_lock_init(&mds->mds_last_fid_lock);
3566 sema_init(&mds->mds_orphan_recovery_sem, 1);
3567 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3569 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3570 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3572 if (obd->obd_namespace == NULL) {
3573 mds_cleanup(obd, 0);
3574 GOTO(err_put, rc = -ENOMEM);
3576 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3578 rc = mds_fs_setup(obd, mnt);
3580 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3585 rc = llog_start_commit_thread();
3591 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 && lustre_cfg_buf(lcfg, 3) &&
3592 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3593 LUSTRE_CFG_BUFLEN(lcfg, 3))) {
3596 generate_random_uuid(uuid);
3597 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3599 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
3600 if (mds->mds_profile == NULL)
3601 GOTO(err_fs, rc = -ENOMEM);
3603 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
3604 LUSTRE_CFG_BUFLEN(lcfg, 3));
3608 * setup root dir and files ID dir if lmv already connected, or there is
3611 if (mds->mds_md_exp || (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 &&
3612 lustre_cfg_buf(lcfg, 3) &&
3613 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3614 LUSTRE_CFG_BUFLEN(lcfg, 3)))) {
3615 rc = mds_fs_setup_rootid(obd);
3619 rc = mds_fs_setup_virtid(obd);
3624 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3625 "mds_ldlm_client", &obd->obd_ldlm_client);
3626 obd->obd_replayable = 1;
3628 rc = mds_postsetup(obd);
3635 /* No extra cleanup needed for llog_init_commit_thread() */
3636 mds_fs_cleanup(obd, 0);
3638 ldlm_namespace_free(obd->obd_namespace, 0);
3639 obd->obd_namespace = NULL;
3642 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3646 fsfilt_put_ops(obd->obd_fsops);
3650 static int mds_fs_post_setup(struct obd_device *obd)
3652 struct mds_obd *mds = &obd->u.mds;
3653 struct dentry *dentry;
3657 dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
3658 if (IS_ERR(dentry)) {
3659 CERROR("Can't find ROOT, err = %d\n",
3660 (int)PTR_ERR(dentry));
3661 RETURN(PTR_ERR(dentry));
3664 rc = fsfilt_post_setup(obd, dentry);
3670 static int mds_postsetup(struct obd_device *obd)
3672 struct mds_obd *mds = &obd->u.mds;
3676 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3677 obd, 0, NULL, &llog_lvfs_ops);
3681 if (mds->mds_profile) {
3682 struct llog_ctxt *lgctxt;
3683 struct lvfs_run_ctxt saved;
3684 struct lustre_profile *lprof;
3685 struct config_llog_instance cfg;
3687 cfg.cfg_instance = NULL;
3688 cfg.cfg_uuid = mds->mds_dt_uuid;
3689 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3691 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3693 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3694 GOTO(err_llog, rc = -EINVAL);
3697 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3698 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3703 lprof = class_get_profile(mds->mds_profile);
3704 if (lprof == NULL) {
3705 CERROR("No profile found: %s\n", mds->mds_profile);
3706 GOTO(err_cleanup, rc = -ENOENT);
3708 rc = mds_dt_connect(obd, lprof->lp_lov);
3710 GOTO(err_cleanup, rc);
3712 rc = mds_md_postsetup(obd);
3714 GOTO(err_cleanup, rc);
3716 rc = mds_fs_post_setup(obd);
3718 CERROR("can not post setup fsfilt\n");
3723 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3724 LLOG_CONFIG_ORIG_CTXT));
3728 int mds_postrecov_common(struct obd_device *obd)
3730 struct mds_obd *mds = &obd->u.mds;
3731 struct llog_ctxt *ctxt;
3732 int rc, item = 0, valsize;
3736 LASSERT(!obd->obd_recovering);
3737 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3738 LASSERT(ctxt != NULL);
3740 /* clean PENDING dir */
3741 rc = mds_cleanup_orphans(obd);
3746 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3747 valsize = sizeof(group);
3748 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
3749 "mds_conn", valsize, &group);
3753 rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3756 CERROR("%s: failed at llog_origin_connect: %d\n",
3761 /* remove the orphaned precreated objects */
3762 rc = mds_dt_clear_orphans(mds, NULL /* all OSTs */);
3767 RETURN(rc < 0 ? rc : item);
3770 /* cleanup all llogging subsystems */
3771 rc = obd_llog_finish(obd, &obd->obd_llogs,
3772 mds->mds_dt_desc.ld_tgt_count);
3774 CERROR("%s: failed to cleanup llogging subsystems\n",
3779 int mds_postrecov(struct obd_device *obd)
3783 rc = mds_postrecov_common(obd);
3785 rc = mds_md_reconnect(obd);
3789 int mds_dt_clean(struct obd_device *obd)
3791 struct mds_obd *mds = &obd->u.mds;
3794 if (mds->mds_profile) {
3796 struct llog_ctxt *llctx;
3797 struct lvfs_run_ctxt saved;
3798 struct config_llog_instance cfg;
3799 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3801 OBD_ALLOC(cln_prof, len);
3802 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3804 cfg.cfg_instance = NULL;
3805 cfg.cfg_uuid = mds->mds_dt_uuid;
3807 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3808 llctx = llog_get_context(&obd->obd_llogs,
3809 LLOG_CONFIG_ORIG_CTXT);
3810 class_config_process_llog(llctx, cln_prof, &cfg);
3811 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3813 OBD_FREE(cln_prof, len);
3814 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3815 mds->mds_profile = NULL;
3820 int mds_md_clean(struct obd_device *obd)
3822 struct mds_obd *mds = &obd->u.mds;
3825 if (mds->mds_md_name) {
3826 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3827 mds->mds_md_name = NULL;
3832 static int mds_precleanup(struct obd_device *obd, int flags)
3838 mds_dt_disconnect(obd, flags);
3840 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3844 extern void lgss_svc_cache_purge_all(void);
3845 static int mds_cleanup(struct obd_device *obd, int flags)
3847 struct mds_obd *mds = &obd->u.mds;
3850 if (mds->mds_sb == NULL)
3853 mds_update_server_data(obd, 1);
3854 mds_update_last_fid(obd, NULL, 1);
3856 if (mds->mds_dt_objids != NULL) {
3857 int size = mds->mds_dt_desc.ld_tgt_count *
3859 OBD_FREE(mds->mds_dt_objids, size);
3861 mds_fs_cleanup(obd, flags);
3865 /* 2 seems normal on mds, (may_umount() also expects 2
3866 fwiw), but we only see 1 at this point in obdfilter. */
3867 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3871 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3873 spin_lock_bh(&obd->obd_processing_task_lock);
3874 if (obd->obd_recovering) {
3875 target_cancel_recovery_timer(obd);
3876 obd->obd_recovering = 0;
3878 spin_unlock_bh(&obd->obd_processing_task_lock);
3881 fsfilt_put_ops(obd->obd_fsops);
3885 lgss_svc_cache_purge_all();
3888 spin_lock(&mds->mds_denylist_lock);
3889 while (!list_empty( &mds->mds_denylist ) ) {
3890 deny_sec_t *p_deny_sec = list_entry(mds->mds_denylist.next,
3892 list_del(&p_deny_sec->list);
3893 OBD_FREE(p_deny_sec, sizeof(*p_deny_sec));
3895 spin_unlock(&mds->mds_denylist_lock);
3900 static int set_security(const char *value, char **sec)
3902 if (!strcmp(value, "null"))
3904 else if (!strcmp(value, "krb5i"))
3906 else if (!strcmp(value, "krb5p"))
3909 CERROR("Unrecognized security flavor %s\n", value);
3916 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
3918 struct lustre_cfg *lcfg = buf;
3919 struct mds_obd *mds = &obd->u.mds;
3923 switch(lcfg->lcfg_command) {
3924 case LCFG_SET_SECURITY: {
3925 if ((LUSTRE_CFG_BUFLEN(lcfg, 1) == 0) ||
3926 (LUSTRE_CFG_BUFLEN(lcfg, 2) == 0))
3927 GOTO(out, rc = -EINVAL);
3929 if (!strcmp(lustre_cfg_string(lcfg, 1), "mds_sec"))
3930 rc = set_security(lustre_cfg_string(lcfg, 2),
3932 else if (!strcmp(lustre_cfg_string(lcfg, 1), "oss_sec"))
3933 rc = set_security(lustre_cfg_string(lcfg, 2),
3935 else if (!strcmp(lustre_cfg_string(lcfg, 1), "deny_sec")){
3936 spin_lock(&mds->mds_denylist_lock);
3937 rc = add_deny_security(lustre_cfg_string(lcfg, 2),
3938 &mds->mds_denylist);
3939 spin_unlock(&mds->mds_denylist_lock);
3941 CERROR("Unrecognized key\n");
3947 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
3948 GOTO(out, rc = -EINVAL);
3954 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3956 struct ldlm_lock *new_lock,
3957 struct ldlm_lock **old_lock,
3958 struct lustre_handle *lockh)
3960 struct obd_export *exp = req->rq_export;
3961 struct obd_device *obd = exp->exp_obd;
3962 struct ldlm_request *dlmreq =
3963 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3964 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3965 struct list_head *iter;
3967 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3970 spin_lock(&obd->obd_namespace->ns_hash_lock);
3971 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3972 struct ldlm_lock *lock;
3973 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3974 if (lock == new_lock)
3976 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3977 lockh->cookie = lock->l_handle.h_cookie;
3978 LDLM_DEBUG(lock, "restoring lock cookie");
3979 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3982 *old_lock = LDLM_LOCK_GET(lock);
3983 spin_unlock(&obd->obd_namespace->ns_hash_lock);
3987 spin_unlock(&obd->obd_namespace->ns_hash_lock);
3989 /* If the xid matches, then we know this is a resent request,
3990 * and allow it. (It's probably an OPEN, for which we don't
3993 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
3997 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
4000 /* This remote handle isn't enqueued, so we never received or
4001 * processed this request. Clear MSG_RESENT, because it can
4002 * be handled like any normal request now. */
4004 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
4006 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
4010 int intent_disposition(struct ldlm_reply *rep, int flag)
4014 return (rep->lock_policy_res1 & flag);
4017 void intent_set_disposition(struct ldlm_reply *rep, int flag)
4021 rep->lock_policy_res1 |= flag;
4024 static int mds_intent_policy(struct ldlm_namespace *ns,
4025 struct ldlm_lock **lockp, void *req_cookie,
4026 ldlm_mode_t mode, int flags, void *data)
4028 struct ptlrpc_request *req = req_cookie;
4029 struct ldlm_lock *lock = *lockp;
4030 struct ldlm_intent *it;
4031 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
4032 struct ldlm_reply *rep;
4033 struct lustre_handle lockh[2] = {{0}, {0}};
4034 struct ldlm_lock *new_lock = NULL;
4035 int getattr_part = MDS_INODELOCK_UPDATE;
4036 int rc, reply_buffers;
4037 int repsize[5] = {sizeof(struct ldlm_reply),
4038 sizeof(struct mds_body),
4039 mds->mds_max_mdsize};
4041 int offset = MDS_REQ_INTENT_REC_OFF;
4044 LASSERT(req != NULL);
4045 MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
4047 if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
4048 /* No intent was provided */
4049 int size = sizeof(struct ldlm_reply);
4050 rc = lustre_pack_reply(req, 1, &size, NULL);
4055 it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
4056 lustre_swab_ldlm_intent);
4058 CERROR("Intent missing\n");
4059 RETURN(req->rq_status = -EFAULT);
4062 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
4065 if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
4066 if (req->rq_export->exp_mds_data.med_remote) {
4068 repsize[3] = sizeof(struct mds_remote_perm);
4072 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
4076 rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
4078 RETURN(req->rq_status = rc);
4080 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
4081 LASSERT(rep != NULL);
4083 intent_set_disposition(rep, DISP_IT_EXECD);
4085 /* execute policy */
4086 switch ((long)it->opc) {
4088 case IT_CREAT|IT_OPEN:
4089 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4091 /* XXX swab here to assert that an mds_open reint
4092 * packet is following */
4093 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4095 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
4097 if (rep->lock_policy_res2) {
4099 * mds_open() returns ENOLCK where it should return
4100 * zero, but it has no lock to return.
4102 if (rep->lock_policy_res2 == ENOLCK)
4103 rep->lock_policy_res2 = 0;
4105 RETURN(ELDLM_LOCK_ABORTED);
4109 * IT_OPEN may return lock on cross-node dentry that we want to
4110 * hold during attr retrival -bzzz
4112 if (lockh[0].cookie == 0)
4113 RETURN(ELDLM_LOCK_ABORTED);
4117 getattr_part = MDS_INODELOCK_LOOKUP;
4120 getattr_part |= MDS_INODELOCK_LOOKUP;
4122 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4123 lock, &new_lock, lockh);
4124 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
4126 /* FIXME: LDLM can set req->rq_status. MDS sets
4127 policy_res{1,2} with disposition and status.
4128 - replay: returns 0 & req->status is old status
4129 - otherwise: returns req->status */
4130 if (intent_disposition(rep, DISP_LOOKUP_NEG))
4131 rep->lock_policy_res2 = 0;
4132 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
4133 rep->lock_policy_res2)
4134 RETURN(ELDLM_LOCK_ABORTED);
4135 if (req->rq_status != 0) {
4137 rep->lock_policy_res2 = req->rq_status;
4138 RETURN(ELDLM_LOCK_ABORTED);
4142 rc = mds_lock_and_check_slave(offset, req, lockh);
4143 if ((rep->lock_policy_res2 = rc)) {
4145 rep->lock_policy_res2 = 0;
4146 RETURN(ELDLM_LOCK_ABORTED);
4150 CERROR("Unhandled intent "LPD64"\n", it->opc);
4154 /* By this point, whatever function we called above must have either
4155 * filled in 'lockh', been an intent replay, or returned an error. We
4156 * want to allow replayed RPCs to not get a lock, since we would just
4157 * drop it below anyways because lock replay is done separately by the
4158 * client afterwards. For regular RPCs we want to give the new lock to
4159 * the client instead of whatever lock it was about to get. */
4160 if (new_lock == NULL)
4161 new_lock = ldlm_handle2lock(&lockh[0]);
4162 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
4165 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
4166 it->opc, lockh[0].cookie);
4168 /* If we've already given this lock to a client once, then we should
4169 * have no readers or writers. Otherwise, we should have one reader
4170 * _or_ writer ref (which will be zeroed below) before returning the
4171 * lock to a client. */
4172 if (new_lock->l_export == req->rq_export) {
4173 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
4175 LASSERT(new_lock->l_export == NULL);
4176 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
4181 if (new_lock->l_export == req->rq_export) {
4182 /* Already gave this to the client, which means that we
4183 * reconstructed a reply. */
4184 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
4186 RETURN(ELDLM_LOCK_REPLACED);
4189 /* Fixup the lock to be given to the client */
4190 lock_res_and_lock(new_lock);
4191 new_lock->l_readers = 0;
4192 new_lock->l_writers = 0;
4194 new_lock->l_export = class_export_get(req->rq_export);
4196 spin_lock(&new_lock->l_export->exp_ldlm_data.led_lock);
4197 list_add(&new_lock->l_export_chain,
4198 &new_lock->l_export->exp_ldlm_data.led_held_locks);
4199 spin_unlock(&new_lock->l_export->exp_ldlm_data.led_lock);
4201 new_lock->l_blocking_ast = lock->l_blocking_ast;
4202 new_lock->l_completion_ast = lock->l_completion_ast;
4204 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
4205 sizeof(lock->l_remote_handle));
4207 new_lock->l_flags &= ~LDLM_FL_LOCAL;
4209 unlock_res_and_lock(new_lock);
4210 LDLM_LOCK_PUT(new_lock);
4212 RETURN(ELDLM_LOCK_REPLACED);
4215 int mds_attach(struct obd_device *dev, obd_count len, void *data)
4217 struct lprocfs_static_vars lvars;
4219 struct mds_obd *mds = &dev->u.mds;
4221 spin_lock_init(&mds->mds_denylist_lock);
4222 INIT_LIST_HEAD(&mds->mds_denylist);
4224 lprocfs_init_multi_vars(0, &lvars);
4226 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
4230 return lprocfs_alloc_md_stats(dev, 0);
4233 int mds_detach(struct obd_device *dev)
4235 lprocfs_free_md_stats(dev);
4236 return lprocfs_obd_detach(dev);
4239 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
4241 struct lprocfs_static_vars lvars;
4243 lprocfs_init_multi_vars(1, &lvars);
4244 return lprocfs_obd_attach(dev, lvars.obd_vars);
4247 int mdt_detach(struct obd_device *dev)
4249 return lprocfs_obd_detach(dev);
4252 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
4254 struct mds_obd *mds = &obd->u.mds;
4259 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4260 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
4261 MDS_SERVICE_WATCHDOG_TIMEOUT,
4262 mds_handle, "mds", obd->obd_proc_entry);
4264 if (!mds->mds_service) {
4265 CERROR("failed to start service\n");
4269 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
4272 GOTO(err_thread, rc);
4274 mds->mds_setattr_service =
4275 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4276 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
4277 MDS_SERVICE_WATCHDOG_TIMEOUT,
4278 mds_handle, "mds_setattr",
4279 obd->obd_proc_entry);
4280 if (!mds->mds_setattr_service) {
4281 CERROR("failed to start getattr service\n");
4282 GOTO(err_thread, rc = -ENOMEM);
4285 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
4286 MDT_NUM_THREADS, "ll_mdt_attr");
4288 GOTO(err_thread2, rc);
4290 mds->mds_readpage_service =
4291 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4292 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
4293 MDS_SERVICE_WATCHDOG_TIMEOUT,
4294 mds_handle, "mds_readpage",
4295 obd->obd_proc_entry);
4296 if (!mds->mds_readpage_service) {
4297 CERROR("failed to start readpage service\n");
4298 GOTO(err_thread2, rc = -ENOMEM);
4301 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
4302 MDT_NUM_THREADS, "ll_mdt_rdpg");
4305 GOTO(err_thread3, rc);
4310 ptlrpc_unregister_service(mds->mds_readpage_service);
4312 ptlrpc_unregister_service(mds->mds_setattr_service);
4314 ptlrpc_unregister_service(mds->mds_service);
4318 static int mdt_cleanup(struct obd_device *obd, int flags)
4320 struct mds_obd *mds = &obd->u.mds;
4323 ptlrpc_stop_all_threads(mds->mds_readpage_service);
4324 ptlrpc_unregister_service(mds->mds_readpage_service);
4326 ptlrpc_stop_all_threads(mds->mds_setattr_service);
4327 ptlrpc_unregister_service(mds->mds_setattr_service);
4329 ptlrpc_stop_all_threads(mds->mds_service);
4330 ptlrpc_unregister_service(mds->mds_service);
4335 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
4336 __u64 gr, void *data)
4338 struct lustre_id id;
4339 struct obd_device *obd = data;
4343 return mds_id2dentry(obd, &id, NULL);
4346 static int mds_get_info(struct obd_export *exp, __u32 keylen,
4347 void *key, __u32 *valsize, void *val)
4349 struct obd_device *obd;
4350 struct mds_obd *mds;
4353 obd = class_exp2obd(exp);
4357 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
4358 exp->exp_handle.h_cookie);
4362 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
4363 /* get log_context handle. */
4364 unsigned long *llh_handle = val;
4365 *valsize = sizeof(unsigned long);
4366 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
4369 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
4370 /* get log_context handle. */
4371 unsigned long *sb = val;
4372 *valsize = sizeof(unsigned long);
4373 *sb = (unsigned long)obd->u.mds.mds_sb;
4377 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
4378 __u32 *mdsize = val;
4379 *valsize = sizeof(*mdsize);
4380 *mdsize = mds->mds_max_mdsize;
4384 if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
4385 __u32 *mdsnum = val;
4386 *valsize = sizeof(*mdsnum);
4387 *mdsnum = mds->mds_num;
4391 if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
4392 struct lustre_id *rootid = val;
4393 *valsize = sizeof(*rootid);
4394 *rootid = mds->mds_rootid;
4398 if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
4399 struct lov_desc *desc = val;
4400 *valsize = sizeof(*desc);
4401 *desc = mds->mds_dt_desc;
4405 CDEBUG(D_IOCTL, "invalid key\n");
4409 struct lvfs_callback_ops mds_lvfs_ops = {
4410 l_id2dentry: mds_lvfs_id2dentry,
4413 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
4414 int objcount, struct obd_ioobj *obj,
4415 int niocount, struct niobuf_remote *nb,
4416 struct niobuf_local *res,
4417 struct obd_trans_info *oti);
4419 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
4420 int objcount, struct obd_ioobj *obj, int niocount,
4421 struct niobuf_local *res, struct obd_trans_info *oti,
4424 /* use obd ops to offer management infrastructure */
4425 static struct obd_ops mds_obd_ops = {
4426 .o_owner = THIS_MODULE,
4427 .o_attach = mds_attach,
4428 .o_detach = mds_detach,
4429 .o_connect = mds_connect,
4430 .o_connect_post = mds_connect_post,
4431 .o_init_export = mds_init_export,
4432 .o_destroy_export = mds_destroy_export,
4433 .o_disconnect = mds_disconnect,
4434 .o_setup = mds_setup,
4435 .o_precleanup = mds_precleanup,
4436 .o_cleanup = mds_cleanup,
4437 .o_process_config = mds_process_config,
4438 .o_postrecov = mds_postrecov,
4439 .o_statfs = mds_obd_statfs,
4440 .o_iocontrol = mds_iocontrol,
4441 .o_create = mds_obd_create,
4442 .o_destroy = mds_obd_destroy,
4443 .o_llog_init = mds_llog_init,
4444 .o_llog_finish = mds_llog_finish,
4445 .o_notify = mds_notify,
4446 .o_get_info = mds_get_info,
4447 .o_set_info = mds_set_info,
4448 .o_preprw = mds_preprw,
4449 .o_commitrw = mds_commitrw,
4452 static struct obd_ops mdt_obd_ops = {
4453 .o_owner = THIS_MODULE,
4454 .o_attach = mdt_attach,
4455 .o_detach = mdt_detach,
4456 .o_setup = mdt_setup,
4457 .o_cleanup = mdt_cleanup,
4460 static int __init mds_init(void)
4462 struct lprocfs_static_vars lvars;
4464 mds_init_lsd_cache();
4465 mds_init_rmtacl_upcall_cache();
4467 lprocfs_init_multi_vars(0, &lvars);
4468 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4469 OBD_MDS_DEVICENAME);
4470 lprocfs_init_multi_vars(1, &lvars);
4471 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4472 OBD_MDT_DEVICENAME);
4477 static void /*__exit*/ mds_exit(void)
4479 mds_cleanup_rmtacl_upcall_cache();
4480 mds_cleanup_lsd_cache();
4482 class_unregister_type(OBD_MDS_DEVICENAME);
4483 class_unregister_type(OBD_MDT_DEVICENAME);
4486 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4487 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4488 MODULE_LICENSE("GPL");
4490 module_init(mds_init);
4491 module_exit(mds_exit);