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/ext3_fs.h>
43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
44 # include <linux/smp_lock.h>
45 # include <linux/buffer_head.h>
46 # include <linux/workqueue.h>
47 # include <linux/mount.h>
49 # include <linux/locks.h>
51 #include <linux/obd_lov.h>
52 #include <linux/obd_ost.h>
53 #include <linux/lustre_mds.h>
54 #include <linux/lustre_fsfilt.h>
55 #include <linux/lprocfs_status.h>
56 #include <linux/lustre_commit_confd.h>
58 #include "mds_internal.h"
60 static int mds_intent_policy(struct ldlm_namespace *ns,
61 struct ldlm_lock **lockp, void *req_cookie,
62 ldlm_mode_t mode, int flags, void *data);
63 static int mds_postsetup(struct obd_device *obd);
64 static int mds_cleanup(struct obd_device *obd, int flags);
67 /* Assumes caller has already pushed into the kernel filesystem context */
68 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
69 loff_t offset, int count)
71 struct ptlrpc_bulk_desc *desc;
72 struct l_wait_info lwi;
74 int rc = 0, npages, i, tmpcount, tmpsize = 0;
77 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
79 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
80 OBD_ALLOC(pages, sizeof(*pages) * npages);
82 GOTO(out, rc = -ENOMEM);
84 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
87 GOTO(out_free, rc = -ENOMEM);
89 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
90 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
92 pages[i] = alloc_pages(GFP_KERNEL, 0);
94 GOTO(cleanup_buf, rc = -ENOMEM);
96 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
99 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
100 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
101 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
102 tmpsize, offset, file->f_dentry->d_inode->i_ino,
103 file->f_dentry->d_inode->i_size);
105 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
106 kmap(pages[i]), tmpsize, &offset);
110 GOTO(cleanup_buf, rc = -EIO);
113 LASSERT(desc->bd_nob == count);
115 rc = ptlrpc_start_bulk_transfer(desc);
117 GOTO(cleanup_buf, rc);
119 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
120 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
121 OBD_FAIL_MDS_SENDPAGE, rc = -EIO);
122 GOTO(abort_bulk, rc);
125 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
126 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
127 LASSERT (rc == 0 || rc == -ETIMEDOUT);
130 if (desc->bd_success &&
131 desc->bd_nob_transferred == count)
132 GOTO(cleanup_buf, rc);
134 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
137 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
138 (rc == -ETIMEDOUT) ? "timeout" : "network error",
139 desc->bd_nob_transferred, count,
140 req->rq_export->exp_client_uuid.uuid,
141 req->rq_export->exp_connection->c_remote_uuid.uuid);
143 ptlrpc_fail_export(req->rq_export);
147 ptlrpc_abort_bulk (desc);
149 for (i = 0; i < npages; i++)
151 __free_pages(pages[i], 0);
153 ptlrpc_free_bulk(desc);
155 OBD_FREE(pages, sizeof(*pages) * npages);
160 int mds_lock_mode_for_dir(struct obd_device *obd,
161 struct dentry *dentry, int mode)
165 /* any dir access needs couple locks:
166 * 1) on part of dir we gonna lookup/modify in
167 * 2) on a whole dir to protect it from concurrent splitting
168 * and to flush client's cache for readdir()
169 * so, for a given mode and dentry this routine decides what
170 * lock mode to use for lock #2:
171 * 1) if caller's gonna lookup in dir then we need to protect
172 * dir from being splitted only - LCK_CR
173 * 2) if caller's gonna modify dir then we need to protect
174 * dir from being splitted and to flush cache - LCK_CW
175 * 3) if caller's gonna modify dir and that dir seems ready
176 * for splitting then we need to protect it from any
177 * type of access (lookup/modify/split) - LCK_EX -bzzz */
179 split = mds_splitting_expected(obd, dentry);
182 * it is important to check here only for MDS_NO_SPLITTABLE. The reason
183 * is that MDS_NO_SPLITTABLE means dir is not splittable in principle
184 * and another thread will not split it on the quiet. But if we have
185 * MDS_NO_SPLIT_EXPECTED, this means, that dir may be splitted anytime,
186 * but not now (for current thread) and we should consider that it can
187 * happen soon and go that branch which can yield LCK_EX to protect from
188 * possible splitting.
190 if (split == MDS_NO_SPLITTABLE) {
192 * this inode won't be splitted. so we need not to protect from
193 * just flush client's cache on modification.
202 if (mode == LCK_PR) {
204 } else if (mode == LCK_PW) {
206 * caller gonna modify directory.we use concurrent write
207 * lock here to retract client's cache for readdir.
210 if (split == MDS_EXPECT_SPLIT) {
212 * splitting possible. serialize any access the
213 * idea is that first one seen dir is splittable
214 * is given exclusive lock and split
215 * directory. caller passes lock mode to
216 * mds_try_to_split_dir() and splitting would be
217 * done with exclusive lock only -bzzz.
219 CDEBUG(D_OTHER, "%s: gonna split %lu/%lu\n",
221 (unsigned long)dentry->d_inode->i_ino,
222 (unsigned long)dentry->d_inode->i_generation);
230 /* only valid locked dentries or errors should be returned */
231 struct dentry *mds_id2locked_dentry(struct obd_device *obd, struct lustre_id *id,
232 struct vfsmount **mnt, int lock_mode,
233 struct lustre_handle *lockh, int *mode,
234 char *name, int namelen, __u64 lockpart)
236 struct dentry *de = mds_id2dentry(obd, id, mnt), *retval = de;
237 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
238 struct ldlm_res_id res_id = { .name = {0} };
246 res_id.name[0] = id_fid(id);
247 res_id.name[1] = id_group(id);
250 if (name && IS_PDIROPS(de->d_inode)) {
251 ldlm_policy_data_t cpolicy =
252 { .l_inodebits = { MDS_INODELOCK_UPDATE } };
253 LASSERT(mode != NULL);
254 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
256 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
258 &cpolicy, *mode, &flags,
260 ldlm_completion_ast, NULL, NULL,
261 NULL, 0, NULL, lockh + 1);
262 if (rc != ELDLM_OK) {
264 RETURN(ERR_PTR(-ENOLCK));
269 res_id.name[2] = full_name_hash(name, namelen);
271 CDEBUG(D_INFO, "take lock on "DLID4":"LPX64"\n",
272 OLID4(id), res_id.name[2]);
275 #warning "No PDIROPS support in the kernel"
277 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
278 LDLM_IBITS, &policy, lock_mode, &flags,
279 mds_blocking_ast, ldlm_completion_ast,
280 NULL, NULL, NULL, 0, NULL, lockh);
281 if (rc != ELDLM_OK) {
283 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
286 ldlm_lock_decref(lockh + 1, LCK_CW);
293 #ifndef DCACHE_DISCONNECTED
294 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
298 /* Look up an entry by inode number. This function ONLY returns valid dget'd
299 * dentries with an initialized inode or errors */
300 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
301 struct vfsmount **mnt)
303 unsigned long ino = (unsigned long)id_ino(id);
304 __u32 generation = (__u32)id_gen(id);
305 struct mds_obd *mds = &obd->u.mds;
306 struct dentry *result;
311 RETURN(ERR_PTR(-ESTALE));
313 snprintf(idname, sizeof(idname), "0x%lx", ino);
315 CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
316 ino, generation, mds->mds_sb);
318 /* under ext3 this is neither supposed to return bad inodes nor NULL
320 result = ll_lookup_one_len(idname, mds->mds_id_de,
325 inode = result->d_inode;
327 RETURN(ERR_PTR(-ENOENT));
329 if (is_bad_inode(inode)) {
330 CERROR("bad inode returned %lu/%u\n",
331 inode->i_ino, inode->i_generation);
333 RETURN(ERR_PTR(-ENOENT));
336 /* here we disabled generation check, as root inode i_generation
337 * of cache mds and real mds are different. */
338 if (inode->i_ino != id_ino(&mds->mds_rootid) && generation &&
339 inode->i_generation != generation) {
340 /* we didn't find the right inode.. */
341 CERROR("bad inode %lu, link: %lu, ct: %d, generation %u/%u\n",
342 inode->i_ino, (unsigned long)inode->i_nlink,
343 atomic_read(&inode->i_count), inode->i_generation,
346 RETURN(ERR_PTR(-ENOENT));
350 *mnt = mds->mds_vfsmnt;
358 /* Establish a connection to the MDS.
360 * This will set up an export structure for the client to hold state data about
361 * that client, like open files, the last operation number it did on the server,
364 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
365 struct obd_uuid *cluuid, unsigned long flags)
367 struct mds_export_data *med;
368 struct mds_client_data *mcd;
369 struct obd_export *exp;
373 if (!conn || !obd || !cluuid)
376 /* XXX There is a small race between checking the list and adding a new
377 * connection for the same UUID, but the real threat (list corruption
378 * when multiple different clients connect) is solved.
380 * There is a second race between adding the export to the list, and
381 * filling in the client data below. Hence skipping the case of NULL
382 * mcd above. We should already be controlling multiple connects at the
383 * client, and we can't hold the spinlock over memory allocations
384 * without risk of deadlocking.
386 rc = class_connect(conn, obd, cluuid);
389 exp = class_conn2export(conn);
391 LASSERT(exp != NULL);
392 med = &exp->exp_mds_data;
394 OBD_ALLOC(mcd, sizeof(*mcd));
396 CERROR("%s: out of memory for client data.\n",
398 GOTO(out, rc = -ENOMEM);
401 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
404 rc = mds_client_add(obd, &obd->u.mds, med, -1);
411 OBD_FREE(mcd, sizeof(*mcd));
412 class_disconnect(exp, 0);
414 class_export_put(exp);
418 static int mds_connect_post(struct obd_export *exp, unsigned long connect_flags)
420 struct obd_device *obd = exp->exp_obd;
421 struct mds_obd *mds = &obd->u.mds;
425 if (!(connect_flags & OBD_OPT_MDS_CONNECTION)) {
426 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
427 atomic_inc(&mds->mds_real_clients);
428 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
429 obd->obd_name, exp->exp_client_uuid.uuid,
430 atomic_read(&mds->mds_real_clients));
431 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
433 if (mds->mds_lmv_name)
434 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
439 static int mds_init_export(struct obd_export *exp)
441 struct mds_export_data *med = &exp->exp_mds_data;
443 INIT_LIST_HEAD(&med->med_open_head);
444 spin_lock_init(&med->med_open_lock);
448 static int mds_destroy_export(struct obd_export *export)
450 struct mds_export_data *med;
451 struct obd_device *obd = export->exp_obd;
452 struct lvfs_run_ctxt saved;
456 med = &export->exp_mds_data;
457 target_destroy_export(export);
459 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
462 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
464 /* Close any open files (which may also cause orphan unlinking). */
465 spin_lock(&med->med_open_lock);
466 while (!list_empty(&med->med_open_head)) {
467 struct list_head *tmp = med->med_open_head.next;
468 struct mds_file_data *mfd =
469 list_entry(tmp, struct mds_file_data, mfd_list);
470 BDEVNAME_DECLARE_STORAGE(btmp);
472 /* bug 1579: fix force-closing for 2.5 */
473 struct dentry *dentry = mfd->mfd_dentry;
475 list_del(&mfd->mfd_list);
476 spin_unlock(&med->med_open_lock);
478 /* If you change this message, be sure to update
479 * replay_single:test_46 */
480 CERROR("force closing client file handle for %*s (%s:%lu)\n",
481 dentry->d_name.len, dentry->d_name.name,
482 ll_bdevname(dentry->d_inode->i_sb, btmp),
483 dentry->d_inode->i_ino);
484 /* child inode->i_alloc_sem protects orphan_dec_test and
485 * is_orphan race, mds_mfd_close drops it */
486 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
487 rc = mds_mfd_close(NULL, 0, obd, mfd,
488 !(export->exp_flags & OBD_OPT_FAILOVER));
490 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
491 spin_lock(&med->med_open_lock);
493 spin_unlock(&med->med_open_lock);
494 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
498 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
502 static int mds_disconnect(struct obd_export *exp, int flags)
504 unsigned long irqflags;
505 struct obd_device *obd;
510 LASSERT(exp != NULL);
511 obd = class_exp2obd(exp);
513 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
514 exp->exp_handle.h_cookie);
520 * suppress any inter-mds requests durring disconnecting lmv if this is
521 * detected --force mode. This is needed to avoid endless recovery.
523 if (atomic_read(&mds->mds_real_clients) > 0 &&
524 !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
525 flags |= OBD_OPT_FORCE;
527 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
528 && !atomic_read(&mds->mds_real_clients)) {
529 /* there was no client at all */
530 mds_lmv_disconnect(obd, flags);
533 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
534 && atomic_dec_and_test(&mds->mds_real_clients)) {
535 /* time to drop LMV connections */
536 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
537 "Disconnnect from LMV now\n",
538 obd->obd_name, exp->exp_client_uuid.uuid);
539 mds_lmv_disconnect(obd, flags);
542 spin_lock_irqsave(&exp->exp_lock, irqflags);
543 exp->exp_flags = flags;
544 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
546 /* disconnect early so that clients can't keep using export */
547 rc = class_disconnect(exp, flags);
548 ldlm_cancel_locks_for_export(exp);
550 /* complete all outstanding replies */
551 spin_lock_irqsave(&exp->exp_lock, irqflags);
552 while (!list_empty(&exp->exp_outstanding_replies)) {
553 struct ptlrpc_reply_state *rs =
554 list_entry(exp->exp_outstanding_replies.next,
555 struct ptlrpc_reply_state, rs_exp_list);
556 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
558 spin_lock(&svc->srv_lock);
559 list_del_init(&rs->rs_exp_list);
560 ptlrpc_schedule_difficult_reply(rs);
561 spin_unlock(&svc->srv_lock);
563 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
567 static int mds_getstatus(struct ptlrpc_request *req)
569 struct mds_obd *mds = mds_req2mds(req);
570 struct mds_body *body;
574 size = sizeof(*body);
576 rc = lustre_pack_reply(req, 1, &size, NULL);
577 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
578 CERROR("mds: out of memory for message: size=%d\n", size);
579 req->rq_status = -ENOMEM; /* superfluous? */
583 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
585 body->valid |= OBD_MD_FID;
586 memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
589 * the last_committed and last_xid fields are filled in for all replies
590 * already - no need to do so here also.
595 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
596 void *data, int flag)
601 if (flag == LDLM_CB_CANCELING) {
602 /* Don't need to do anything here. */
606 /* XXX layering violation! -phil */
607 l_lock(&lock->l_resource->lr_namespace->ns_lock);
610 * get this: if mds_blocking_ast is racing with mds_intent_policy, such
611 * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
612 * then by the time we get the lock, we might not be the correct
613 * blocking function anymore. So check, and return early, if so.
615 if (lock->l_blocking_ast != mds_blocking_ast) {
616 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
620 lock->l_flags |= LDLM_FL_CBPENDING;
621 do_ast = (!lock->l_readers && !lock->l_writers);
622 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
625 struct lustre_handle lockh;
628 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
629 ldlm_lock2handle(lock, &lockh);
630 rc = ldlm_cli_cancel(&lockh);
632 CERROR("ldlm_cli_cancel: %d\n", rc);
634 LDLM_DEBUG(lock, "Lock still has references, will be "
640 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
649 rc = fsfilt_get_md(obd, inode, md, *size);
654 CERROR("Error %d reading eadata for ino %lu\n",
659 if (S_ISREG(inode->i_mode))
660 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
661 if (S_ISDIR(inode->i_mode))
662 rc = mds_convert_mea_ea(obd, inode, md, lmm_size);
676 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
677 * Call with lock=0 if the caller has already taken the i_sem. */
678 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
679 struct mds_body *body, struct inode *inode, int lock)
681 struct mds_obd *mds = &obd->u.mds;
687 lmm = lustre_msg_buf(msg, offset, 0);
689 /* Some problem with getting eadata when I sized the reply
691 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
695 lmm_size = msg->buflens[offset];
697 /* I don't really like this, but it is a sanity check on the client
698 * MD request. However, if the client doesn't know how much space
699 * to reserve for the MD, it shouldn't be bad to have too much space.
701 if (lmm_size > mds->mds_max_mdsize) {
702 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
703 inode->i_ino, lmm_size, mds->mds_max_mdsize);
707 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
709 if (S_ISDIR(inode->i_mode))
710 body->valid |= OBD_MD_FLDIREA;
712 body->valid |= OBD_MD_FLEASIZE;
713 body->eadatasize = lmm_size;
720 void mds_squash_root(struct mds_obd *mds, struct mds_req_sec_desc *rsd,
723 if (!mds->mds_squash_uid ||
724 (rsd->rsd_uid && rsd->rsd_fsuid))
727 if (*peernid == mds->mds_nosquash_nid)
730 CDEBUG(D_OTHER, "squash req from 0x%llx, (%d:%d/%x)=>(%d:%d/%x)\n",
731 *peernid, rsd->rsd_fsuid, rsd->rsd_fsgid, rsd->rsd_cap,
732 mds->mds_squash_uid, mds->mds_squash_gid,
733 (rsd->rsd_cap & ~CAP_FS_MASK));
735 rsd->rsd_uid = mds->mds_squash_uid;
736 rsd->rsd_fsuid = mds->mds_squash_uid;
737 rsd->rsd_fsgid = mds->mds_squash_gid;
739 /* XXX should we remove all capabilities? */
740 rsd->rsd_cap &= ~CAP_FS_MASK;
743 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
744 struct ptlrpc_request *req, struct mds_body *reqbody,
747 struct inode *inode = dentry->d_inode;
748 struct mds_body *body;
752 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
755 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
756 LASSERT(body != NULL); /* caller prepped reply */
758 if (dentry->d_flags & DCACHE_CROSS_REF) {
759 mds_pack_dentry2body(obd, body, dentry,
760 (reqbody->valid & OBD_MD_FID));
761 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
766 mds_pack_inode2body(obd, body, inode, (reqbody->valid & OBD_MD_FID));
768 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
769 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
770 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
773 /* if we have LOV EA data, the OST holds size, atime, mtime. */
774 if (!(body->valid & OBD_MD_FLEASIZE) &&
775 !(body->valid & OBD_MD_FLDIREA))
776 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
777 OBD_MD_FLATIME | OBD_MD_FLMTIME);
778 } else if (S_ISLNK(inode->i_mode) &&
779 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
780 int len = req->rq_repmsg->buflens[reply_off + 1];
781 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1, 0);
783 LASSERT(symname != NULL); /* caller prepped reply */
785 if (!inode->i_op->readlink) {
788 rc = inode->i_op->readlink(dentry, symname, len);
790 CERROR("readlink failed: %d\n", rc);
791 } else if (rc != len - 1) {
792 CERROR("Unexpected readlink rc %d: expecting %d\n",
796 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
797 body->valid |= OBD_MD_LINKNAME;
798 body->eadatasize = rc + 1;
808 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
809 struct dentry *dentry,
812 int rc = 0, size[1] = {sizeof(struct mds_body)};
815 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
816 CERROR("failed MDS_GETATTR_PACK test\n");
817 req->rq_status = -ENOMEM;
821 rc = lustre_pack_reply(req, 1, size, NULL);
824 CERROR("lustre_pack_reply failed: rc %d\n", rc);
825 GOTO(out, req->rq_status = rc);
833 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
836 struct mds_obd *mds = mds_req2mds(req);
837 struct mds_body *body;
838 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
841 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
842 LASSERT(body != NULL); /* checked by caller */
843 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
845 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
846 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
849 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
851 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
855 CERROR("error getting inode %lu MD: rc = %d\n",
858 } else if (rc > mds->mds_max_mdsize) {
860 CERROR("MD size %d larger than maximum possible %u\n",
861 rc, mds->mds_max_mdsize);
866 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
867 if (inode->i_size + 1 != body->eadatasize)
868 CERROR("symlink size: %Lu, reply space: %d\n",
869 inode->i_size + 1, body->eadatasize);
870 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
872 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
873 inode->i_size + 1, body->eadatasize);
876 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
877 CERROR("failed MDS_GETATTR_PACK test\n");
878 req->rq_status = -ENOMEM;
879 GOTO(out, rc = -ENOMEM);
882 rc = lustre_pack_reply(req, bufcount, size, NULL);
884 CERROR("out of memory\n");
885 GOTO(out, req->rq_status = rc);
893 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
894 char *name, int namelen)
896 struct mea *mea = NULL;
897 int mea_size, rc = 0;
900 rc = mds_get_lmv_attr(obd, inode, &mea, &mea_size);
905 * dir is already splitted, check if requested filename should
906 * live at this MDS or at another one.
908 int i = mea_name2idx(mea, name, namelen - 1);
909 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
911 "inapropriate MDS(%d) for %s. should be "
912 "%lu(%d)\n", mea->mea_master, name,
913 (unsigned long)id_group(&mea->mea_ids[i]), i);
919 OBD_FREE(mea, mea_size);
923 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
924 struct lustre_handle *child_lockh, int child_part)
926 struct obd_device *obd = req->rq_export->exp_obd;
927 struct mds_obd *mds = &obd->u.mds;
928 struct ldlm_reply *rep = NULL;
929 struct lvfs_run_ctxt saved;
930 struct mds_req_sec_desc *rsd;
931 struct mds_body *body;
932 struct dentry *dparent = NULL, *dchild = NULL;
933 struct lvfs_ucred uc;
934 struct lustre_handle parent_lockh[2] = {{0}, {0}};
935 unsigned int namesize;
936 int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
940 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
941 MDS_UPDATE_COUNTER((&obd->u.mds), MDS_GETATTR_NAME_COUNT);
943 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
945 CERROR("Can't unpack security desc\n");
948 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
950 /* swab now, before anyone looks inside the request. */
951 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
952 lustre_swab_mds_body);
954 CERROR("Can't swab mds_body\n");
955 GOTO(cleanup, rc = -EFAULT);
958 LASSERT_REQSWAB(req, offset + 1);
959 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
961 CERROR("Can't unpack name\n");
962 GOTO(cleanup, rc = -EFAULT);
964 namesize = req->rq_reqmsg->buflens[offset + 1];
966 /* namesize less than 2 means we have empty name, probably came from
967 revalidate by cfid, so no point in having name to be set */
971 LASSERT (offset == 1 || offset == 3);
973 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
979 rc = mds_init_ucred(&uc, rsd);
981 CERROR("can't init ucred\n");
985 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
986 cleanup_phase = 1; /* kernel context */
987 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
989 LASSERT(namesize > 0);
990 if (child_lockh->cookie != 0) {
991 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
996 if (body->valid == OBD_MD_FLID) {
997 struct mds_body *mds_reply;
998 int size = sizeof(*mds_reply);
1002 dparent = mds_id2dentry(obd, &body->id1, NULL);
1003 if (IS_ERR(dparent)) {
1004 rc = PTR_ERR(dparent);
1009 * the user requested ONLY the inode number, so do a raw lookup.
1011 rc = lustre_pack_reply(req, 1, &size, NULL);
1013 CERROR("out of memory\n");
1017 dir = dparent->d_inode;
1018 LASSERT(dir->i_op->lookup_raw != NULL);
1019 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1021 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1022 sizeof(*mds_reply));
1024 id_ino(&mds_reply->id1) = inum;
1025 mds_reply->valid = OBD_MD_FLID;
1030 if (resent_req == 0) {
1031 LASSERT(id_fid(&body->id1) != 0);
1033 rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1034 parent_lockh, &dparent,
1036 MDS_INODELOCK_UPDATE,
1039 child_lockh, &dchild,
1040 LCK_PR, child_part);
1045 * let's make sure this name should leave on this mds
1048 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1052 /* we have no dentry here, drop LOOKUP bit */
1053 /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1054 child_part &= ~MDS_INODELOCK_LOOKUP;
1055 CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1056 obd->obd_name, OLID4(&body->id1));
1058 dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1059 LCK_PR, parent_lockh,
1062 MDS_INODELOCK_UPDATE);
1063 if (IS_ERR(dchild)) {
1064 CERROR("can't find inode with id "DLID4", err = %d\n",
1065 OLID4(&body->id1), (int)PTR_ERR(dchild));
1066 GOTO(cleanup, rc = PTR_ERR(dchild));
1068 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1070 if (parent_lockh[1].cookie)
1071 ldlm_lock_decref(parent_lockh + 1, update_mode);
1075 struct ldlm_lock *granted_lock;
1077 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1078 granted_lock = ldlm_handle2lock(child_lockh);
1080 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1081 id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1082 child_lockh->cookie);
1084 dparent = mds_id2dentry(obd, &body->id1, NULL);
1087 dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1089 LDLM_LOCK_PUT(granted_lock);
1092 cleanup_phase = 2; /* dchild, dparent, locks */
1094 if (!DENTRY_VALID(dchild)) {
1095 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1097 * in the intent case, the policy clears this error: the
1098 * disposition is enough.
1103 intent_set_disposition(rep, DISP_LOOKUP_POS);
1106 if (req->rq_repmsg == NULL) {
1107 if (dchild->d_flags & DCACHE_CROSS_REF)
1108 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1110 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
1112 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1117 rc = mds_getattr_internal(obd, dchild, req, body, reply_offset);
1118 GOTO(cleanup, rc); /* returns the lock to the client */
1121 switch (cleanup_phase) {
1123 if (resent_req == 0) {
1124 if (rc && DENTRY_VALID(dchild))
1125 ldlm_lock_decref(child_lockh, LCK_PR);
1127 ldlm_lock_decref(parent_lockh, LCK_PR);
1129 if (parent_lockh[1].cookie != 0)
1130 ldlm_lock_decref(parent_lockh + 1, update_mode);
1137 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1138 mds_exit_ucred(&uc);
1143 static int mds_getattr(struct ptlrpc_request *req, int offset)
1145 struct obd_device *obd = req->rq_export->exp_obd;
1146 struct mds_obd *mds = &obd->u.mds;
1147 struct lvfs_run_ctxt saved;
1149 struct mds_req_sec_desc *rsd;
1150 struct mds_body *body;
1151 struct lvfs_ucred uc;
1155 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1157 CERROR("Can't unpack security desc\n");
1161 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1162 lustre_swab_mds_body);
1164 CERROR ("Can't unpack body\n");
1168 MDS_UPDATE_COUNTER(mds, MDS_GETATTR_COUNT);
1170 rc = mds_init_ucred(&uc, rsd);
1172 CERROR("can't init ucred\n");
1176 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1177 de = mds_id2dentry(obd, &body->id1, NULL);
1179 rc = req->rq_status = PTR_ERR(de);
1183 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1185 CERROR("mds_getattr_pack_msg: %d\n", rc);
1189 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
1195 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1196 mds_exit_ucred(&uc);
1200 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1201 unsigned long max_age)
1206 spin_lock(&obd->obd_osfs_lock);
1207 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1209 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1210 spin_unlock(&obd->obd_osfs_lock);
1215 static int mds_statfs(struct ptlrpc_request *req)
1217 struct obd_device *obd = req->rq_export->exp_obd;
1218 int rc, size = sizeof(struct obd_statfs);
1221 rc = lustre_pack_reply(req, 1, &size, NULL);
1222 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1223 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1227 MDS_UPDATE_COUNTER((&obd->u.mds), MDS_STATFS_COUNT);
1229 /* We call this so that we can cache a bit - 1 jiffie worth */
1230 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1233 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1239 req->rq_status = rc;
1243 static int mds_sync(struct ptlrpc_request *req, int offset)
1245 struct obd_device *obd = req->rq_export->exp_obd;
1246 struct mds_obd *mds = &obd->u.mds;
1247 struct mds_body *body;
1248 int rc, size = sizeof(*body);
1251 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1253 GOTO(out, rc = -EPROTO);
1255 rc = lustre_pack_reply(req, 1, &size, NULL);
1256 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1257 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1261 if (id_ino(&body->id1) == 0) {
1262 /* an id of zero is taken to mean "sync whole filesystem" */
1263 rc = fsfilt_sync(obd, mds->mds_sb);
1267 /* just any file to grab fsync method - "file" arg unused */
1268 struct file *file = mds->mds_rcvd_filp;
1271 de = mds_id2dentry(obd, &body->id1, NULL);
1273 GOTO(out, rc = PTR_ERR(de));
1275 rc = file->f_op->fsync(NULL, de, 1);
1279 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1280 mds_pack_inode2body(obd, body, de->d_inode, 0);
1286 req->rq_status = rc;
1290 /* mds_readpage does not take a DLM lock on the inode, because the client must
1291 * already have a PR lock.
1293 * If we were to take another one here, a deadlock will result, if another
1294 * thread is already waiting for a PW lock. */
1295 static int mds_readpage(struct ptlrpc_request *req, int offset)
1297 struct obd_device *obd = req->rq_export->exp_obd;
1298 struct mds_obd *mds = &obd->u.mds;
1299 struct vfsmount *mnt;
1302 struct mds_req_sec_desc *rsd;
1303 struct mds_body *body, *repbody;
1304 struct lvfs_run_ctxt saved;
1305 int rc, size = sizeof(*repbody);
1306 struct lvfs_ucred uc;
1309 rc = lustre_pack_reply(req, 1, &size, NULL);
1310 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1311 CERROR("mds: out of memory\n");
1312 GOTO(out, rc = -ENOMEM);
1315 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1317 CERROR("Can't unpack security desc\n");
1318 GOTO (out, rc = -EFAULT);
1320 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
1322 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1323 lustre_swab_mds_body);
1325 CERROR("Can't unpack body\n");
1326 GOTO (out, rc = -EFAULT);
1329 rc = mds_init_ucred(&uc, rsd);
1331 CERROR("can't init ucred\n");
1335 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1336 de = mds_id2dentry(obd, &body->id1, &mnt);
1338 GOTO(out_pop, rc = PTR_ERR(de));
1340 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1342 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1343 /* note: in case of an error, dentry_open puts dentry */
1345 GOTO(out_pop, rc = PTR_ERR(file));
1347 /* body->size is actually the offset -eeb */
1348 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1349 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1350 body->size, de->d_inode->i_blksize);
1351 GOTO(out_file, rc = -EFAULT);
1354 /* body->nlink is actually the #bytes to read -eeb */
1355 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1356 CERROR("size %u is not multiple of blocksize %lu\n",
1357 body->nlink, de->d_inode->i_blksize);
1358 GOTO(out_file, rc = -EFAULT);
1361 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1362 repbody->size = file->f_dentry->d_inode->i_size;
1363 repbody->valid = OBD_MD_FLSIZE;
1365 /* to make this asynchronous make sure that the handling function
1366 doesn't send a reply when this function completes. Instead a
1367 callback function would send the reply */
1368 /* body->size is actually the offset -eeb */
1369 rc = mds_sendpage(req, file, body->size, body->nlink);
1373 filp_close(file, 0);
1375 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1376 mds_exit_ucred(&uc);
1378 req->rq_status = rc;
1382 /* update master MDS ID, which is stored in local inode EA. */
1383 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1384 void *data, int data_len)
1386 struct mds_obd *mds = &obd->u.mds;
1387 struct dentry *dentry;
1395 dentry = mds_id2dentry(obd, id, NULL);
1397 GOTO(out, rc = PTR_ERR(dentry));
1399 if (!dentry->d_inode) {
1400 CERROR("Can't find object "DLID4".\n",
1402 GOTO(out_dentry, rc = -EINVAL);
1405 handle = fsfilt_start(obd, dentry->d_inode,
1406 FSFILT_OP_SETATTR, NULL);
1408 GOTO(out_dentry, rc = PTR_ERR(handle));
1410 rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1411 (struct lustre_id *)data);
1413 CERROR("Can't update inode "DLID4" master id, "
1414 "error = %d.\n", OLID4(id), rc);
1415 GOTO(out_commit, rc);
1420 fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1427 EXPORT_SYMBOL(mds_update_mid);
1429 /* read master MDS ID, which is stored in local inode EA. */
1430 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1431 void *data, int data_len)
1433 struct dentry *dentry;
1440 dentry = mds_id2dentry(obd, id, NULL);
1442 GOTO(out, rc = PTR_ERR(dentry));
1444 if (!dentry->d_inode) {
1445 CERROR("Can't find object "DLID4".\n",
1447 GOTO(out_dentry, rc = -EINVAL);
1450 down(&dentry->d_inode->i_sem);
1451 rc = mds_read_inode_mid(obd, dentry->d_inode,
1452 (struct lustre_id *)data);
1453 up(&dentry->d_inode->i_sem);
1455 CERROR("Can't read inode "DLID4" master id, "
1456 "error = %d.\n", OLID4(id), rc);
1457 GOTO(out_dentry, rc);
1466 EXPORT_SYMBOL(mds_read_mid);
1468 int mds_reint(struct ptlrpc_request *req, int offset,
1469 struct lustre_handle *lockh)
1471 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1472 struct mds_update_record *rec;
1473 struct mds_req_sec_desc *rsd;
1477 OBD_ALLOC(rec, sizeof(*rec));
1481 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1483 CERROR("Can't unpack security desc\n");
1484 GOTO(out, rc = -EFAULT);
1486 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
1488 rc = mds_update_unpack(req, offset, rec);
1489 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1490 CERROR("invalid record\n");
1491 GOTO(out, req->rq_status = -EINVAL);
1494 rc = mds_init_ucred(&rec->ur_uc, rsd);
1496 CERROR("can't init ucred\n");
1500 /* rc will be used to interrupt a for loop over multiple records */
1501 rc = mds_reint_rec(rec, offset, req, lockh);
1502 mds_exit_ucred(&rec->ur_uc);
1505 OBD_FREE(rec, sizeof(*rec));
1509 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1510 struct obd_device *obd, int *process)
1512 switch (req->rq_reqmsg->opc) {
1513 case MDS_CONNECT: /* This will never get here, but for completeness. */
1514 case OST_CONNECT: /* This will never get here, but for completeness. */
1515 case MDS_DISCONNECT:
1516 case OST_DISCONNECT:
1521 case MDS_SYNC: /* used in unmounting */
1526 *process = target_queue_recovery_request(req, obd);
1530 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1532 /* XXX what should we set rq_status to here? */
1533 req->rq_status = -EAGAIN;
1534 RETURN(ptlrpc_error(req));
1538 static char *reint_names[] = {
1539 [REINT_SETATTR] "setattr",
1540 [REINT_CREATE] "create",
1541 [REINT_LINK] "link",
1542 [REINT_UNLINK] "unlink",
1543 [REINT_RENAME] "rename",
1544 [REINT_OPEN] "open",
1547 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER | \
1548 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
1549 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
1552 static void reconstruct_create(struct ptlrpc_request *req)
1554 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1555 struct mds_client_data *mcd = med->med_mcd;
1556 struct dentry *dentry;
1557 struct ost_body *body;
1558 struct lustre_id id;
1562 /* copy rc, transno and disp; steal locks */
1563 mds_req_from_mcd(req, mcd);
1564 if (req->rq_status) {
1572 id_ino(&id) = mcd->mcd_last_data;
1573 LASSERT(id_ino(&id) != 0);
1575 dentry = mds_id2dentry(req2obd(req), &id, NULL);
1576 if (IS_ERR(dentry)) {
1577 CERROR("can't find inode "LPU64"\n", id_ino(&id));
1578 req->rq_status = PTR_ERR(dentry);
1583 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1584 req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
1585 dentry->d_inode->i_generation);
1587 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1588 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
1589 body->oa.o_id = dentry->d_inode->i_ino;
1590 body->oa.o_generation = dentry->d_inode->i_generation;
1591 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1593 down(&dentry->d_inode->i_sem);
1594 rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
1595 up(&dentry->d_inode->i_sem);
1597 CERROR("Can't read inode self id, inode %lu, "
1598 "rc %d\n", dentry->d_inode->i_ino, rc);
1602 body->oa.o_fid = id_fid(&id);
1603 body->oa.o_mds = id_group(&id);
1609 static int mdt_obj_create(struct ptlrpc_request *req)
1611 struct obd_device *obd = req->rq_export->exp_obd;
1612 struct mds_obd *mds = &obd->u.mds;
1613 struct ost_body *body, *repbody;
1614 char idname[LL_ID_NAMELEN];
1615 int size = sizeof(*repbody);
1616 struct inode *parent_inode;
1617 struct lvfs_run_ctxt saved;
1618 int rc, cleanup_phase = 0;
1619 struct dentry *new = NULL;
1620 struct dentry_params dp;
1621 int mealen, flags = 0;
1622 struct lvfs_ucred uc;
1623 struct lustre_id id;
1625 void *handle = NULL;
1626 unsigned long cr_inum = 0;
1629 DEBUG_REQ(D_HA, req, "create remote object");
1631 parent_inode = mds->mds_unnamed_dir->d_inode;
1633 body = lustre_swab_reqbuf(req, 0, sizeof(*body),
1634 lustre_swab_ost_body);
1638 rc = lustre_pack_reply(req, 1, &size, NULL);
1642 MDS_CHECK_RESENT(req, reconstruct_create(req));
1645 * this only serve to inter-mds request, don't need check group database
1648 uc.luc_ghash = NULL;
1649 uc.luc_ginfo = NULL;
1650 uc.luc_uid = body->oa.o_uid;
1651 uc.luc_fsuid = body->oa.o_uid;
1652 uc.luc_fsgid = body->oa.o_gid;
1653 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1655 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1657 /* in REPLAY case inum should be given (client or other MDS fills it) */
1658 if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
1659 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
1660 /* this is re-create request from MDS holding directory name.
1661 * we have to lookup given ino/gen first. if it exists
1662 * (good case) then there is nothing to do. if it does not then
1663 * we have to recreate it */
1664 id_ino(&id) = body->oa.o_id;
1665 id_gen(&id) = body->oa.o_generation;
1667 new = mds_id2dentry(obd, &id, NULL);
1668 if (!IS_ERR(new) && new->d_inode) {
1669 struct lustre_id sid;
1671 CWARN("mkdir() repairing is on its way: %lu/%lu\n",
1672 (unsigned long)id_ino(&id), (unsigned long)id_gen(&id));
1674 obdo_from_inode(&repbody->oa, new->d_inode,
1675 FILTER_VALID_FLAGS);
1677 repbody->oa.o_id = new->d_inode->i_ino;
1678 repbody->oa.o_generation = new->d_inode->i_generation;
1679 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1682 down(&new->d_inode->i_sem);
1683 rc = mds_read_inode_sid(obd, new->d_inode, &sid);
1684 up(&new->d_inode->i_sem);
1686 CERROR("Can't read inode self id "
1687 "inode %lu, rc %d.\n",
1688 new->d_inode->i_ino, rc);
1692 repbody->oa.o_fid = id_fid(&sid);
1693 repbody->oa.o_mds = id_group(&sid);
1694 cr_inum = new->d_inode->i_ino;
1695 GOTO(cleanup, rc = 0);
1699 down(&parent_inode->i_sem);
1700 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1701 if (IS_ERR(handle)) {
1702 up(&parent_inode->i_sem);
1703 CERROR("fsfilt_start() failed, rc = %d\n",
1704 (int)PTR_ERR(handle));
1705 GOTO(cleanup, rc = PTR_ERR(handle));
1707 cleanup_phase = 1; /* transaction */
1710 rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
1711 new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
1713 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1714 obd->obd_name, idname, (int) PTR_ERR(new));
1715 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
1716 up(&parent_inode->i_sem);
1717 RETURN(PTR_ERR(new));
1718 } else if (new->d_inode) {
1719 CERROR("%s: name exists. repeat\n", obd->obd_name);
1723 new->d_fsdata = (void *)&dp;
1727 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
1728 (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
1729 LASSERT(body->oa.o_id != 0);
1730 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1731 (unsigned long)body->oa.o_id,
1732 (unsigned long)body->oa.o_generation);
1733 dp.p_inum = body->oa.o_id;
1736 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1738 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1739 new->d_inode->i_generation = body->oa.o_generation;
1740 mark_inode_dirty(new->d_inode);
1743 * avoiding asserts in cache flush case, as
1744 * @body->oa.o_id should be zero.
1746 if (body->oa.o_id) {
1747 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
1748 "BUG 3550: failed to recreate obj "
1749 LPU64" -> %lu\n", body->oa.o_id,
1750 new->d_inode->i_ino);
1752 LASSERTF(body->oa.o_generation ==
1753 new->d_inode->i_generation,
1754 "BUG 3550: failed to recreate obj/gen "
1755 LPU64"/%u -> %lu/%u\n", body->oa.o_id,
1756 body->oa.o_generation,
1757 new->d_inode->i_ino,
1758 new->d_inode->i_generation);
1762 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1763 repbody->oa.o_id = new->d_inode->i_ino;
1764 repbody->oa.o_generation = new->d_inode->i_generation;
1765 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
1767 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1768 /* building lustre_id on passed @oa. */
1769 id_group(&id) = mds->mds_num;
1771 LASSERT(body->oa.o_fid != 0);
1772 id_fid(&id) = body->oa.o_fid;
1774 id_ino(&id) = repbody->oa.o_id;
1775 id_gen(&id) = repbody->oa.o_generation;
1777 down(&new->d_inode->i_sem);
1778 rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
1779 up(&new->d_inode->i_sem);
1782 * make sure, that fid is up-to-date.
1784 mds_set_last_fid(obd, id_fid(&id));
1786 down(&new->d_inode->i_sem);
1787 rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
1788 up(&new->d_inode->i_sem);
1791 CERROR("Can't update lustre ID for inode %lu, "
1792 "error = %d\n", new->d_inode->i_ino, rc);
1796 /* initializing o_fid after it is allocated. */
1797 repbody->oa.o_fid = id_fid(&id);
1798 repbody->oa.o_mds = id_group(&id);
1800 rc = fsfilt_del_dir_entry(obd, new);
1801 up(&parent_inode->i_sem);
1803 CERROR("can't remove name for object: %d\n", rc);
1807 cleanup_phase = 2; /* created directory object */
1809 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
1810 (unsigned long)new->d_inode->i_ino,
1811 (unsigned long)new->d_inode->i_generation,
1812 (unsigned)new->d_inode->i_mode);
1813 cr_inum = new->d_inode->i_ino;
1815 up(&parent_inode->i_sem);
1816 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1820 if (body->oa.o_valid & OBD_MD_FLID) {
1821 /* this is new object for splitted dir. We have to prevent
1822 * recursive splitting on it -bzzz */
1823 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1825 OBD_ALLOC(mea, mealen);
1827 GOTO(cleanup, rc = -ENOMEM);
1829 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
1830 mea->mea_master = 0;
1833 down(&new->d_inode->i_sem);
1834 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1835 up(&new->d_inode->i_sem);
1837 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
1839 OBD_FREE(mea, mealen);
1840 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
1841 obd->obd_name, new->d_inode->i_ino,
1842 new->d_inode->i_generation, flags);
1843 } else if (body->oa.o_easize) {
1844 /* we pass LCK_EX to split routine to signal that we have
1845 * exclusive access to the directory. simple because nobody
1846 * knows it already exists -bzzz */
1847 rc = mds_try_to_split_dir(obd, new, NULL,
1848 body->oa.o_easize, LCK_EX);
1850 CERROR("Can't split directory %lu, error = %d.\n",
1851 new->d_inode->i_ino, rc);
1859 switch (cleanup_phase) {
1860 case 2: /* object has been created, but we'll may want to replay it later */
1862 ptlrpc_require_repack(req);
1863 case 1: /* transaction */
1864 rc = mds_finish_transno(mds, parent_inode, handle,
1869 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1870 mds_put_group_entry(mds, uc.luc_ghash);
1874 static int mdt_get_info(struct ptlrpc_request *req)
1876 struct obd_export *exp = req->rq_export;
1881 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1883 DEBUG_REQ(D_HA, req, "no get_info key");
1886 keylen = req->rq_reqmsg->buflens[0];
1888 if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
1889 (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
1890 (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
1893 if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
1894 struct lustre_id *reply;
1895 int size = sizeof(*reply);
1897 rc = lustre_pack_reply(req, 1, &size, NULL);
1901 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
1902 rc = obd_get_info(exp, keylen, key, &size, reply);
1905 int size = sizeof(*reply);
1907 rc = lustre_pack_reply(req, 1, &size, NULL);
1911 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
1912 rc = obd_get_info(exp, keylen, key, &size, reply);
1915 req->rq_repmsg->status = 0;
1919 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1920 void *key, __u32 vallen, void *val)
1922 struct obd_device *obd;
1923 struct mds_obd *mds;
1927 obd = class_exp2obd(exp);
1929 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1930 exp->exp_handle.h_cookie);
1935 if (keylen >= strlen("mds_type") &&
1936 memcmp(key, "mds_type", keylen) == 0) {
1940 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
1942 mds->mds_obd_type = *(int*)val;
1943 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
1944 valsize = sizeof(group);
1946 /* mds number has been changed, so the corresponding obdfilter
1947 * exp need to be changed too. */
1948 rc = obd_set_info(mds->mds_lov_exp, strlen("mds_conn"),
1949 "mds_conn", valsize, &group);
1952 CDEBUG(D_IOCTL, "invalid key\n");
1956 static int mdt_set_info(struct ptlrpc_request *req)
1959 struct obd_export *exp = req->rq_export;
1960 int keylen, rc = 0, vallen;
1963 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1965 DEBUG_REQ(D_HA, req, "no set_info key");
1968 keylen = req->rq_reqmsg->buflens[0];
1970 if (keylen == strlen("mds_type") &&
1971 memcmp(key, "mds_type", keylen) == 0) {
1972 rc = lustre_pack_reply(req, 0, NULL, NULL);
1976 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
1977 vallen = req->rq_reqmsg->buflens[1];
1979 rc = obd_set_info(exp, keylen, key, vallen, val);
1980 req->rq_repmsg->status = 0;
1983 CDEBUG(D_IOCTL, "invalid key\n");
1987 static int mds_msg_check_version(struct lustre_msg *msg)
1993 case MDS_DISCONNECT:
1995 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1997 CERROR("bad opc %u version %08x, expecting %08x\n",
1998 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2003 case MDS_GETATTR_LOCK:
2007 case MDS_DONE_WRITING:
2010 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2012 CERROR("bad opc %u version %08x, expecting %08x\n",
2013 msg->opc, msg->version, LUSTRE_MDS_VERSION);
2017 case LDLM_BL_CALLBACK:
2018 case LDLM_CP_CALLBACK:
2019 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2021 CERROR("bad opc %u version %08x, expecting %08x\n",
2022 msg->opc, msg->version, LUSTRE_DLM_VERSION);
2024 case OBD_LOG_CANCEL:
2025 case LLOG_ORIGIN_HANDLE_OPEN:
2026 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2027 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2028 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2029 case LLOG_ORIGIN_HANDLE_CLOSE:
2031 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2033 CERROR("bad opc %u version %08x, expecting %08x\n",
2034 msg->opc, msg->version, LUSTRE_LOG_VERSION);
2040 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2042 CERROR("bad opc %u version %08x, expecting %08x\n",
2043 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2046 CERROR("MDS unknown opcode %d\n", msg->opc);
2054 static char str[PTL_NALFMT_SIZE];
2055 int mds_handle(struct ptlrpc_request *req)
2057 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2059 struct mds_obd *mds = NULL; /* quell gcc overwarning */
2060 struct obd_device *obd = NULL;
2063 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2065 rc = mds_msg_check_version(req->rq_reqmsg);
2067 CERROR("MDS drop mal-formed request\n");
2071 LASSERT(current->journal_info == NULL);
2072 /* XXX identical to OST */
2073 if (req->rq_reqmsg->opc != MDS_CONNECT) {
2074 struct mds_export_data *med;
2077 if (req->rq_export == NULL) {
2078 CERROR("operation %d on unconnected MDS from NID %s\n",
2079 req->rq_reqmsg->opc,
2080 ptlrpc_peernid2str(&req->rq_peer, str));
2081 req->rq_status = -ENOTCONN;
2082 GOTO(out, rc = -ENOTCONN);
2085 med = &req->rq_export->exp_mds_data;
2086 obd = req->rq_export->exp_obd;
2089 /* sanity check: if the xid matches, the request must
2090 * be marked as a resent or replayed */
2091 if (req->rq_xid == med->med_mcd->mcd_last_xid) {
2092 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2093 (MSG_RESENT | MSG_REPLAY),
2094 "rq_xid "LPU64" matches last_xid, "
2095 "expected RESENT flag\n",
2098 /* else: note the opposite is not always true; a
2099 * RESENT req after a failover will usually not match
2100 * the last_xid, since it was likely never
2101 * committed. A REPLAYed request will almost never
2102 * match the last xid, however it could for a
2103 * committed, but still retained, open. */
2105 spin_lock_bh(&obd->obd_processing_task_lock);
2106 recovering = obd->obd_recovering;
2107 spin_unlock_bh(&obd->obd_processing_task_lock);
2109 rc = mds_filter_recovery_request(req, obd,
2111 if (rc || should_process == 0) {
2113 } else if (should_process < 0) {
2114 req->rq_status = should_process;
2115 rc = ptlrpc_error(req);
2121 switch (req->rq_reqmsg->opc) {
2123 DEBUG_REQ(D_INODE, req, "connect");
2124 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2125 rc = target_handle_connect(req);
2127 /* Now that we have an export, set mds. */
2128 mds = mds_req2mds(req);
2131 case MDS_DISCONNECT:
2132 DEBUG_REQ(D_INODE, req, "disconnect");
2133 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2134 rc = target_handle_disconnect(req);
2135 req->rq_status = rc; /* superfluous? */
2139 DEBUG_REQ(D_INODE, req, "getstatus");
2140 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2141 rc = mds_getstatus(req);
2145 DEBUG_REQ(D_INODE, req, "getattr");
2146 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2147 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2150 case MDS_GETATTR_LOCK: {
2151 struct lustre_handle lockh;
2152 DEBUG_REQ(D_INODE, req, "getattr_lock");
2153 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2155 /* If this request gets a reconstructed reply, we won't be
2156 * acquiring any new locks in mds_getattr_lock, so we don't
2160 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2161 MDS_INODELOCK_UPDATE);
2162 /* this non-intent call (from an ioctl) is special */
2163 req->rq_status = rc;
2164 if (rc == 0 && lockh.cookie)
2165 ldlm_lock_decref(&lockh, LCK_PR);
2169 DEBUG_REQ(D_INODE, req, "statfs");
2170 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2171 rc = mds_statfs(req);
2175 DEBUG_REQ(D_INODE, req, "readpage");
2176 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2177 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2179 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2180 if (req->rq_reply_state) {
2181 lustre_free_reply_state (req->rq_reply_state);
2182 req->rq_reply_state = NULL;
2189 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2192 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2193 mds->mds_max_cookiesize};
2196 /* NB only peek inside req now; mds_reint() will swab it */
2198 CERROR ("Can't inspect opcode\n");
2203 if (lustre_msg_swabbed (req->rq_reqmsg))
2206 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2207 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2208 reint_names[opc] == NULL) ? reint_names[opc] :
2211 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2213 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2215 else if (opc == REINT_OPEN)
2220 rc = lustre_pack_reply(req, bufcount, size, NULL);
2224 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
2225 fail = OBD_FAIL_MDS_REINT_NET_REP;
2230 DEBUG_REQ(D_INODE, req, "close");
2231 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
2232 rc = mds_close(req, MDS_REQ_REC_OFF);
2235 case MDS_DONE_WRITING:
2236 DEBUG_REQ(D_INODE, req, "done_writing");
2237 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
2238 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2242 DEBUG_REQ(D_INODE, req, "pin");
2243 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2244 rc = mds_pin(req, MDS_REQ_REC_OFF);
2248 DEBUG_REQ(D_INODE, req, "sync");
2249 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2250 rc = mds_sync(req, MDS_REQ_REC_OFF);
2254 DEBUG_REQ(D_INODE, req, "ping");
2255 rc = target_handle_ping(req);
2258 case OBD_LOG_CANCEL:
2259 CDEBUG(D_INODE, "log cancel\n");
2260 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2261 rc = -ENOTSUPP; /* la la la */
2265 DEBUG_REQ(D_INODE, req, "enqueue");
2266 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2267 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2268 ldlm_server_blocking_ast, NULL);
2269 fail = OBD_FAIL_LDLM_REPLY;
2272 DEBUG_REQ(D_INODE, req, "convert");
2273 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2274 rc = ldlm_handle_convert(req);
2276 case LDLM_BL_CALLBACK:
2277 case LDLM_CP_CALLBACK:
2278 DEBUG_REQ(D_INODE, req, "callback");
2279 CERROR("callbacks should not happen on MDS\n");
2281 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2283 case LLOG_ORIGIN_HANDLE_OPEN:
2284 DEBUG_REQ(D_INODE, req, "llog_init");
2285 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2286 rc = llog_origin_handle_open(req);
2288 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2289 DEBUG_REQ(D_INODE, req, "llog next block");
2290 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2291 rc = llog_origin_handle_next_block(req);
2293 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2294 DEBUG_REQ(D_INODE, req, "llog prev block");
2295 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2296 rc = llog_origin_handle_prev_block(req);
2298 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2299 DEBUG_REQ(D_INODE, req, "llog read header");
2300 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2301 rc = llog_origin_handle_read_header(req);
2303 case LLOG_ORIGIN_HANDLE_CLOSE:
2304 DEBUG_REQ(D_INODE, req, "llog close");
2305 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2306 rc = llog_origin_handle_close(req);
2309 DEBUG_REQ(D_INODE, req, "ost_create");
2310 rc = mdt_obj_create(req);
2313 DEBUG_REQ(D_INODE, req, "get_info");
2314 rc = mdt_get_info(req);
2317 DEBUG_REQ(D_INODE, req, "set_info");
2318 rc = mdt_set_info(req);
2321 CDEBUG(D_INODE, "write\n");
2322 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2323 rc = ost_brw_write(req, NULL);
2324 LASSERT(current->journal_info == NULL);
2325 /* mdt_brw sends its own replies */
2329 DEBUG_REQ(D_INODE, req, "llog catinfo");
2330 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2331 rc = llog_catinfo(req);
2334 req->rq_status = -ENOTSUPP;
2335 rc = ptlrpc_error(req);
2339 LASSERT(current->journal_info == NULL);
2343 /* If we're DISCONNECTing, the mds_export_data is already freed */
2344 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
2345 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2346 struct obd_device *obd = list_entry(mds, struct obd_device,
2348 req->rq_repmsg->last_xid =
2349 le64_to_cpu(med->med_mcd->mcd_last_xid);
2351 if (!obd->obd_no_transno) {
2352 req->rq_repmsg->last_committed =
2353 obd->obd_last_committed;
2355 DEBUG_REQ(D_IOCTL, req,
2356 "not sending last_committed update");
2358 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2360 mds->mds_last_transno, obd->obd_last_committed,
2365 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
2366 if (obd && obd->obd_recovering) {
2367 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
2368 return target_queue_final_reply(req, rc);
2370 /* Lost a race with recovery; let the error path DTRT. */
2371 rc = req->rq_status = -ENOTCONN;
2374 target_send_reply(req, rc, fail);
2378 /* Update the server data on disk. This stores the new mount_count and also the
2379 * last_rcvd value to disk. If we don't have a clean shutdown, then the server
2380 * last_rcvd value may be less than that of the clients. This will alert us
2381 * that we may need to do client recovery.
2383 * Also assumes for mds_last_transno that we are not modifying it (no locking).
2385 int mds_update_server_data(struct obd_device *obd, int force_sync)
2387 struct mds_obd *mds = &obd->u.mds;
2388 struct mds_server_data *msd = mds->mds_server_data;
2389 struct file *filp = mds->mds_rcvd_filp;
2390 struct lvfs_run_ctxt saved;
2395 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2396 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2398 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2399 mds->mds_mount_count, mds->mds_last_transno);
2400 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
2402 CERROR("error writing MDS server data: rc = %d\n", rc);
2403 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2408 /* saves last allocated fid counter to file. */
2409 int mds_update_last_fid(struct obd_device *obd, void *handle,
2412 struct mds_obd *mds = &obd->u.mds;
2413 struct file *filp = mds->mds_fid_filp;
2414 struct lvfs_run_ctxt saved;
2419 down(&mds->mds_last_fid_sem);
2420 if (mds->mds_last_fid_changed) {
2421 CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
2425 fsfilt_add_journal_cb(obd, mds->mds_sb,
2426 mds->mds_last_fid, handle,
2427 mds_commit_last_fid_cb, NULL);
2430 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2431 rc = fsfilt_write_record(obd, filp, &mds->mds_last_fid,
2432 sizeof(mds->mds_last_fid),
2435 CERROR("error writing MDS last_fid #"LPU64
2436 ", err = %d\n", mds->mds_last_fid, rc);
2438 mds->mds_last_fid_changed = 0;
2441 CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
2442 "%llu: err = %d\n", mds->mds_last_fid,
2444 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2446 up(&mds->mds_last_fid_sem);
2451 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
2453 struct mds_obd *mds = &obd->u.mds;
2455 down(&mds->mds_last_fid_sem);
2456 if (fid > mds->mds_last_fid) {
2457 mds->mds_last_fid = fid;
2458 mds->mds_last_fid_changed = 1;
2460 up(&mds->mds_last_fid_sem);
2463 void mds_commit_last_transno_cb(struct obd_device *obd,
2464 __u64 transno, void *data,
2467 obd_transno_commit_cb(obd, transno, error);
2470 void mds_commit_last_fid_cb(struct obd_device *obd,
2471 __u64 fid, void *data,
2475 CERROR("%s: fid "LPD64" commit error: %d\n",
2476 obd->obd_name, fid, error);
2480 CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
2481 obd->obd_name, fid);
2485 * allocates new lustre_id on passed @inode and saves it to inode EA.
2487 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
2488 void *handle, struct lustre_id *id)
2490 struct mds_obd *mds = &obd->u.mds;
2494 LASSERT(id != NULL);
2495 LASSERT(obd != NULL);
2497 id_group(id) = mds->mds_num;
2499 down(&mds->mds_last_fid_sem);
2500 mds->mds_last_fid_changed = 1;
2501 id_fid(id) = ++mds->mds_last_fid;
2502 up(&mds->mds_last_fid_sem);
2504 id_ino(id) = inode->i_ino;
2505 id_gen(id) = inode->i_generation;
2506 id_type(id) = (S_IFMT & inode->i_mode);
2508 rc = mds_update_inode_sid(obd, inode, handle, id);
2510 CERROR("Can't update inode FID EA, "
2518 * reads inode self id from inode EA. Probably later this should be replaced by
2519 * caching inode self id to avoid raeding it every time it is needed.
2521 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
2522 struct lustre_id *id)
2527 LASSERT(id != NULL);
2528 LASSERT(obd != NULL);
2529 LASSERT(inode != NULL);
2531 rc = fsfilt_get_sid(obd, inode, &id->li_fid,
2532 sizeof(id->li_fid));
2534 CERROR("fsfilt_get_sid() failed, "
2547 /* updates inode self id in EA. */
2548 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
2549 void *handle, struct lustre_id *id)
2554 LASSERT(id != NULL);
2555 LASSERT(obd != NULL);
2556 LASSERT(inode != NULL);
2558 rc = fsfilt_set_sid(obd, inode, handle, &id->li_fid,
2559 sizeof(id->li_fid));
2561 CERROR("fsfilt_set_sid() failed, rc = %d\n", rc);
2569 * reads inode id on master MDS. This is usualy done by CMOBD to update requests
2570 * to master MDS by correct store cookie, needed to find inode on master MDS
2573 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
2574 struct lustre_id *id)
2579 LASSERT(id != NULL);
2580 LASSERT(obd != NULL);
2581 LASSERT(inode != NULL);
2583 rc = fsfilt_get_mid(obd, inode, id, sizeof(*id));
2585 CERROR("fsfilt_get_mid() failed, "
2599 * updates master inode id. Usualy this is done by CMOBD after an inode is
2600 * created and relationship between cache MDS and master one should be
2603 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
2604 void *handle, struct lustre_id *id)
2609 LASSERT(id != NULL);
2610 LASSERT(obd != NULL);
2611 LASSERT(inode != NULL);
2613 rc = fsfilt_set_mid(obd, inode, handle, id, sizeof(*id));
2615 CERROR("fsfilt_set_mid() failed, rc = %d\n", rc);
2622 /* mount the file system (secretly) */
2623 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
2625 struct lustre_cfg* lcfg = buf;
2626 struct mds_obd *mds = &obd->u.mds;
2627 char *options = NULL;
2628 struct vfsmount *mnt;
2634 dev_clear_rdonly(2);
2636 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
2637 RETURN(rc = -EINVAL);
2639 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
2640 if (IS_ERR(obd->obd_fsops))
2641 RETURN(rc = PTR_ERR(obd->obd_fsops));
2643 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2645 page = __get_free_page(GFP_KERNEL);
2649 options = (char *)page;
2650 memset(options, 0, PAGE_SIZE);
2653 * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
2654 * and the rest of options are passed by mount options. Probably this
2655 * should be moved to somewhere else like startup scripts or lconf.
2657 sprintf(options, "iopen_nopriv");
2659 if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
2660 sprintf(options + strlen(options), ",%s",
2661 lcfg->lcfg_inlbuf4);
2663 /* we have to know mdsnum before touching underlying fs -bzzz */
2664 sema_init(&mds->mds_lmv_sem, 1);
2665 mds->mds_lmv_connected = 0;
2666 if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
2667 strcmp(lcfg->lcfg_inlbuf5, "dumb")) {
2670 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
2671 obd->obd_name, lcfg->lcfg_inlbuf5);
2673 generate_random_uuid(uuid);
2674 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
2676 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2677 if (mds->mds_lmv_name == NULL)
2678 RETURN(rc = -ENOMEM);
2680 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf5,
2681 lcfg->lcfg_inllen5);
2683 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
2685 OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2690 mds->mds_obd_type = MDS_MASTER_OBD;
2692 if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 &&
2693 strcmp(lcfg->lcfg_inlbuf6, "dumb")) {
2694 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master"))) {
2695 mds->mds_obd_type = MDS_MASTER_OBD;
2696 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache"))) {
2697 mds->mds_obd_type = MDS_CACHE_OBD;
2701 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0, lcfg->lcfg_inlbuf1, options);
2706 CERROR("do_kern_mount failed: rc = %d\n", rc);
2710 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
2712 mds->mds_last_fid_changed = 0;
2713 sema_init(&mds->mds_epoch_sem, 1);
2714 sema_init(&mds->mds_last_fid_sem, 1);
2715 atomic_set(&mds->mds_real_clients, 0);
2716 spin_lock_init(&mds->mds_transno_lock);
2717 sema_init(&mds->mds_orphan_recovery_sem, 1);
2718 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2720 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
2721 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
2723 if (obd->obd_namespace == NULL) {
2724 mds_cleanup(obd, 0);
2725 GOTO(err_put, rc = -ENOMEM);
2727 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
2729 rc = mds_fs_setup(obd, mnt);
2731 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
2736 rc = llog_start_commit_thread();
2741 * this check for @dumb string is needed to handle mounting MDS with
2742 * smfs. Read lconf:MDSDEV.write_conf() for more details.
2744 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2745 strcmp(lcfg->lcfg_inlbuf3, "dumb")) {
2748 generate_random_uuid(uuid);
2749 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
2751 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
2752 if (mds->mds_profile == NULL)
2753 GOTO(err_fs, rc = -ENOMEM);
2755 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
2756 lcfg->lcfg_inllen3);
2759 * setup root id in the case this is not clients write
2760 * setup. This is important, as in the case of LMV we need
2761 * mds->mds_num to be already assigned to form correct root fid.
2763 rc = mds_fs_setup_rootid(obd);
2767 /* setup lustre id for ID directory. */
2768 rc = mds_fs_setup_virtid(obd);
2773 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2774 "mds_ldlm_client", &obd->obd_ldlm_client);
2775 obd->obd_replayable = 1;
2777 mds->mds_counters = lprocfs_alloc_mds_counters();
2779 rc = mds_postsetup(obd);
2786 /* No extra cleanup needed for llog_init_commit_thread() */
2787 mds_fs_cleanup(obd, 0);
2789 ldlm_namespace_free(obd->obd_namespace, 0);
2790 obd->obd_namespace = NULL;
2793 mntput(mds->mds_vfsmnt);
2797 fsfilt_put_ops(obd->obd_fsops);
2801 static int mds_postsetup(struct obd_device *obd)
2803 struct mds_obd *mds = &obd->u.mds;
2807 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
2808 obd, 0, NULL, &llog_lvfs_ops);
2812 if (mds->mds_profile) {
2813 struct llog_ctxt *lgctxt;
2814 struct lvfs_run_ctxt saved;
2815 struct lustre_profile *lprof;
2816 struct config_llog_instance cfg;
2818 cfg.cfg_instance = NULL;
2819 cfg.cfg_uuid = mds->mds_lov_uuid;
2820 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2822 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
2824 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2825 GOTO(err_llog, rc = -EINVAL);
2828 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
2829 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2834 lprof = class_get_profile(mds->mds_profile);
2835 if (lprof == NULL) {
2836 CERROR("No profile found: %s\n", mds->mds_profile);
2837 GOTO(err_cleanup, rc = -ENOENT);
2839 rc = mds_lov_connect(obd, lprof->lp_lov);
2841 GOTO(err_cleanup, rc);
2843 rc = mds_lmv_postsetup(obd);
2845 GOTO(err_cleanup, rc);
2852 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
2853 LLOG_CONFIG_ORIG_CTXT));
2857 int mds_postrecov(struct obd_device *obd)
2859 struct mds_obd *mds = &obd->u.mds;
2860 struct llog_ctxt *ctxt;
2861 int rc, item = 0, valsize;
2865 LASSERT(!obd->obd_recovering);
2866 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2867 LASSERT(ctxt != NULL);
2869 /* set nextid first, so we are sure it happens */
2870 rc = mds_lov_set_nextid(obd);
2872 CERROR("%s: mds_lov_set_nextid failed\n", obd->obd_name);
2876 /* clean PENDING dir */
2877 rc = mds_cleanup_orphans(obd);
2882 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
2883 valsize = sizeof(group);
2884 rc = obd_set_info(mds->mds_lov_exp, strlen("mds_conn"), "mds_conn",
2889 rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
2892 CERROR("%s: failed at llog_origin_connect: %d\n",
2897 /* remove the orphaned precreated objects */
2898 rc = mds_lov_clearorphans(mds, NULL /* all OSTs */);
2903 RETURN(rc < 0 ? rc : item);
2906 /* cleanup all llogging subsystems */
2907 rc = obd_llog_finish(obd, &obd->obd_llogs,
2908 mds->mds_lov_desc.ld_tgt_count);
2910 CERROR("%s: failed to cleanup llogging subsystems\n",
2915 int mds_lov_clean(struct obd_device *obd)
2917 struct mds_obd *mds = &obd->u.mds;
2920 if (mds->mds_profile) {
2922 struct llog_ctxt *llctx;
2923 struct lvfs_run_ctxt saved;
2924 struct config_llog_instance cfg;
2925 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
2927 OBD_ALLOC(cln_prof, len);
2928 sprintf(cln_prof, "%s-clean", mds->mds_profile);
2930 cfg.cfg_instance = NULL;
2931 cfg.cfg_uuid = mds->mds_lov_uuid;
2933 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2934 llctx = llog_get_context(&obd->obd_llogs,
2935 LLOG_CONFIG_ORIG_CTXT);
2936 class_config_process_llog(llctx, cln_prof, &cfg);
2937 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2939 OBD_FREE(cln_prof, len);
2940 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2941 mds->mds_profile = NULL;
2946 int mds_lmv_clean(struct obd_device *obd)
2948 struct mds_obd *mds = &obd->u.mds;
2951 if (mds->mds_lmv_name) {
2952 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
2953 mds->mds_lmv_name = NULL;
2958 static int mds_precleanup(struct obd_device *obd, int flags)
2964 mds_lov_disconnect(obd, flags);
2966 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2970 static int mds_cleanup(struct obd_device *obd, int flags)
2972 struct mds_obd *mds = &obd->u.mds;
2975 if (mds->mds_sb == NULL)
2978 mds_update_server_data(obd, 1);
2979 mds_update_last_fid(obd, NULL, 1);
2981 if (mds->mds_lov_objids != NULL) {
2982 int size = mds->mds_lov_desc.ld_tgt_count *
2984 OBD_FREE(mds->mds_lov_objids, size);
2986 mds_fs_cleanup(obd, flags);
2991 * 2 seems normal on mds, (may_umount() also expects 2 fwiw), but we
2992 * only see 1 at this point in obdfilter.
2994 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
2995 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
2996 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
2998 mntput(mds->mds_vfsmnt);
3001 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3003 if (mds->mds_counters) {
3004 lprocfs_free_mds_counters(mds->mds_counters);
3007 spin_lock_bh(&obd->obd_processing_task_lock);
3008 if (obd->obd_recovering) {
3009 target_cancel_recovery_timer(obd);
3010 obd->obd_recovering = 0;
3012 spin_unlock_bh(&obd->obd_processing_task_lock);
3015 dev_clear_rdonly(2);
3016 fsfilt_put_ops(obd->obd_fsops);
3021 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3023 struct ldlm_lock *new_lock,
3024 struct ldlm_lock **old_lock,
3025 struct lustre_handle *lockh)
3027 struct obd_export *exp = req->rq_export;
3028 struct obd_device *obd = exp->exp_obd;
3029 struct ldlm_request *dlmreq =
3030 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3031 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3032 struct list_head *iter;
3034 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3037 l_lock(&obd->obd_namespace->ns_lock);
3038 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3039 struct ldlm_lock *lock;
3040 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3041 if (lock == new_lock)
3043 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3044 lockh->cookie = lock->l_handle.h_cookie;
3045 LDLM_DEBUG(lock, "restoring lock cookie");
3046 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3049 *old_lock = LDLM_LOCK_GET(lock);
3050 l_unlock(&obd->obd_namespace->ns_lock);
3054 l_unlock(&obd->obd_namespace->ns_lock);
3056 /* If the xid matches, then we know this is a resent request,
3057 * and allow it. (It's probably an OPEN, for which we don't
3059 if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
3062 /* This remote handle isn't enqueued, so we never received or
3063 * processed this request. Clear MSG_RESENT, because it can
3064 * be handled like any normal request now. */
3066 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3068 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3072 int intent_disposition(struct ldlm_reply *rep, int flag)
3076 return (rep->lock_policy_res1 & flag);
3079 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3083 rep->lock_policy_res1 |= flag;
3086 static int mds_intent_policy(struct ldlm_namespace *ns,
3087 struct ldlm_lock **lockp, void *req_cookie,
3088 ldlm_mode_t mode, int flags, void *data)
3090 struct ptlrpc_request *req = req_cookie;
3091 struct ldlm_lock *lock = *lockp;
3092 struct ldlm_intent *it;
3093 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3094 struct ldlm_reply *rep;
3095 struct lustre_handle lockh[2] = {{0}, {0}};
3096 struct ldlm_lock *new_lock = NULL;
3097 int getattr_part = MDS_INODELOCK_UPDATE;
3098 int rc, repsize[4] = { sizeof(struct ldlm_reply),
3099 sizeof(struct mds_body),
3100 mds->mds_max_mdsize,
3101 mds->mds_max_cookiesize };
3102 int offset = MDS_REQ_INTENT_REC_OFF;
3105 LASSERT(req != NULL);
3107 if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3108 /* No intent was provided */
3109 int size = sizeof(struct ldlm_reply);
3110 rc = lustre_pack_reply(req, 1, &size, NULL);
3115 it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3116 lustre_swab_ldlm_intent);
3118 CERROR("Intent missing\n");
3119 RETURN(req->rq_status = -EFAULT);
3122 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3124 rc = lustre_pack_reply(req, 3, repsize, NULL);
3126 RETURN(req->rq_status = rc);
3128 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
3129 LASSERT(rep != NULL);
3131 intent_set_disposition(rep, DISP_IT_EXECD);
3133 /* execute policy */
3134 switch ((long)it->opc) {
3136 case IT_CREAT|IT_OPEN:
3137 /* XXX swab here to assert that an mds_open reint
3138 * packet is following */
3139 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3141 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
3143 /* We abort the lock if the lookup was negative and
3144 * we did not make it to the OPEN portion */
3145 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
3146 RETURN(ELDLM_LOCK_ABORTED);
3147 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
3148 !intent_disposition(rep, DISP_OPEN_OPEN))
3150 /* IT_OPEN may return lock on cross-node dentry
3151 * that we want to hold during attr retrival -bzzz */
3152 if (rc != 0 || lockh[0].cookie == 0)
3153 RETURN(ELDLM_LOCK_ABORTED);
3156 getattr_part = MDS_INODELOCK_LOOKUP;
3159 getattr_part |= MDS_INODELOCK_LOOKUP;
3161 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3162 lock, &new_lock, lockh);
3163 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
3165 /* FIXME: LDLM can set req->rq_status. MDS sets
3166 policy_res{1,2} with disposition and status.
3167 - replay: returns 0 & req->status is old status
3168 - otherwise: returns req->status */
3169 if (intent_disposition(rep, DISP_LOOKUP_NEG))
3170 rep->lock_policy_res2 = 0;
3171 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
3172 rep->lock_policy_res2)
3173 RETURN(ELDLM_LOCK_ABORTED);
3174 if (req->rq_status != 0) {
3176 rep->lock_policy_res2 = req->rq_status;
3177 RETURN(ELDLM_LOCK_ABORTED);
3181 rc = mds_lock_and_check_slave(offset, req, lockh);
3182 if ((rep->lock_policy_res2 = rc)) {
3184 rep->lock_policy_res2 = 0;
3185 RETURN(ELDLM_LOCK_ABORTED);
3189 CERROR("Unhandled intent "LPD64"\n", it->opc);
3193 /* By this point, whatever function we called above must have either
3194 * filled in 'lockh', been an intent replay, or returned an error. We
3195 * want to allow replayed RPCs to not get a lock, since we would just
3196 * drop it below anyways because lock replay is done separately by the
3197 * client afterwards. For regular RPCs we want to give the new lock to
3198 * the client instead of whatever lock it was about to get. */
3199 if (new_lock == NULL)
3200 new_lock = ldlm_handle2lock(&lockh[0]);
3201 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
3204 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
3205 it->opc, lockh[0].cookie);
3207 /* If we've already given this lock to a client once, then we should
3208 * have no readers or writers. Otherwise, we should have one reader
3209 * _or_ writer ref (which will be zeroed below) before returning the
3210 * lock to a client. */
3211 if (new_lock->l_export == req->rq_export) {
3212 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3214 LASSERT(new_lock->l_export == NULL);
3215 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3220 if (new_lock->l_export == req->rq_export) {
3221 /* Already gave this to the client, which means that we
3222 * reconstructed a reply. */
3223 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3225 RETURN(ELDLM_LOCK_REPLACED);
3228 /* Fixup the lock to be given to the client */
3229 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
3230 new_lock->l_readers = 0;
3231 new_lock->l_writers = 0;
3233 new_lock->l_export = class_export_get(req->rq_export);
3234 list_add(&new_lock->l_export_chain,
3235 &new_lock->l_export->exp_ldlm_data.led_held_locks);
3237 new_lock->l_blocking_ast = lock->l_blocking_ast;
3238 new_lock->l_completion_ast = lock->l_completion_ast;
3240 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
3241 sizeof(lock->l_remote_handle));
3243 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3245 LDLM_LOCK_PUT(new_lock);
3246 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
3248 RETURN(ELDLM_LOCK_REPLACED);
3251 int mds_attach(struct obd_device *dev, obd_count len, void *data)
3253 struct lprocfs_static_vars lvars;
3255 lprocfs_init_multi_vars(0, &lvars);
3256 return lprocfs_obd_attach(dev, lvars.obd_vars);
3259 int mds_detach(struct obd_device *dev)
3261 return lprocfs_obd_detach(dev);
3264 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
3266 struct lprocfs_static_vars lvars;
3268 lprocfs_init_multi_vars(1, &lvars);
3269 return lprocfs_obd_attach(dev, lvars.obd_vars);
3272 int mdt_detach(struct obd_device *dev)
3274 return lprocfs_obd_detach(dev);
3277 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
3279 struct mds_obd *mds = &obd->u.mds;
3284 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3285 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
3286 mds_handle, "mds", obd->obd_proc_entry);
3288 if (!mds->mds_service) {
3289 CERROR("failed to start service\n");
3293 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
3296 GOTO(err_thread, rc);
3298 mds->mds_setattr_service =
3299 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3300 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
3301 mds_handle, "mds_setattr",
3302 obd->obd_proc_entry);
3303 if (!mds->mds_setattr_service) {
3304 CERROR("failed to start getattr service\n");
3305 GOTO(err_thread, rc = -ENOMEM);
3308 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
3309 MDT_NUM_THREADS, "ll_mdt_attr");
3311 GOTO(err_thread2, rc);
3313 mds->mds_readpage_service =
3314 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3315 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
3316 mds_handle, "mds_readpage",
3317 obd->obd_proc_entry);
3318 if (!mds->mds_readpage_service) {
3319 CERROR("failed to start readpage service\n");
3320 GOTO(err_thread2, rc = -ENOMEM);
3323 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
3324 MDT_NUM_THREADS, "ll_mdt_rdpg");
3327 GOTO(err_thread3, rc);
3332 ptlrpc_unregister_service(mds->mds_readpage_service);
3334 ptlrpc_unregister_service(mds->mds_setattr_service);
3336 ptlrpc_unregister_service(mds->mds_service);
3340 static int mdt_cleanup(struct obd_device *obd, int flags)
3342 struct mds_obd *mds = &obd->u.mds;
3345 ptlrpc_stop_all_threads(mds->mds_readpage_service);
3346 ptlrpc_unregister_service(mds->mds_readpage_service);
3348 ptlrpc_stop_all_threads(mds->mds_setattr_service);
3349 ptlrpc_unregister_service(mds->mds_setattr_service);
3351 ptlrpc_stop_all_threads(mds->mds_service);
3352 ptlrpc_unregister_service(mds->mds_service);
3357 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
3358 __u64 gr, void *data)
3360 struct lustre_id id;
3361 struct obd_device *obd = data;
3365 return mds_id2dentry(obd, &id, NULL);
3368 static int mds_get_info(struct obd_export *exp, __u32 keylen,
3369 void *key, __u32 *valsize, void *val)
3371 struct obd_device *obd;
3372 struct mds_obd *mds;
3375 obd = class_exp2obd(exp);
3379 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
3380 exp->exp_handle.h_cookie);
3384 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
3385 /* get log_context handle. */
3386 unsigned long *llh_handle = val;
3387 *valsize = sizeof(unsigned long);
3388 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
3391 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
3392 /* get log_context handle. */
3393 unsigned long *sb = val;
3394 *valsize = sizeof(unsigned long);
3395 *sb = (unsigned long)obd->u.mds.mds_sb;
3399 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
3400 __u32 *mdsize = val;
3401 *valsize = sizeof(*mdsize);
3402 *mdsize = mds->mds_max_mdsize;
3406 if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
3407 __u32 *mdsnum = val;
3408 *valsize = sizeof(*mdsnum);
3409 *mdsnum = mds->mds_num;
3413 if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
3414 struct lustre_id *rootid = val;
3415 *valsize = sizeof(struct lustre_id);
3416 *rootid = mds->mds_rootid;
3420 CDEBUG(D_IOCTL, "invalid key\n");
3424 struct lvfs_callback_ops mds_lvfs_ops = {
3425 l_id2dentry: mds_lvfs_id2dentry,
3428 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
3429 int objcount, struct obd_ioobj *obj,
3430 int niocount, struct niobuf_remote *nb,
3431 struct niobuf_local *res,
3432 struct obd_trans_info *oti);
3434 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
3435 int objcount, struct obd_ioobj *obj, int niocount,
3436 struct niobuf_local *res, struct obd_trans_info *oti,
3439 /* use obd ops to offer management infrastructure */
3440 static struct obd_ops mds_obd_ops = {
3441 .o_owner = THIS_MODULE,
3442 .o_attach = mds_attach,
3443 .o_detach = mds_detach,
3444 .o_connect = mds_connect,
3445 .o_connect_post = mds_connect_post,
3446 .o_init_export = mds_init_export,
3447 .o_destroy_export = mds_destroy_export,
3448 .o_disconnect = mds_disconnect,
3449 .o_setup = mds_setup,
3450 .o_precleanup = mds_precleanup,
3451 .o_cleanup = mds_cleanup,
3452 .o_postrecov = mds_postrecov,
3453 .o_statfs = mds_obd_statfs,
3454 .o_iocontrol = mds_iocontrol,
3455 .o_create = mds_obd_create,
3456 .o_destroy = mds_obd_destroy,
3457 .o_llog_init = mds_llog_init,
3458 .o_llog_finish = mds_llog_finish,
3459 .o_notify = mds_notify,
3460 .o_get_info = mds_get_info,
3461 .o_set_info = mds_set_info,
3462 .o_preprw = mds_preprw,
3463 .o_commitrw = mds_commitrw,
3466 static struct obd_ops mdt_obd_ops = {
3467 .o_owner = THIS_MODULE,
3468 .o_attach = mdt_attach,
3469 .o_detach = mdt_detach,
3470 .o_setup = mdt_setup,
3471 .o_cleanup = mdt_cleanup,
3474 static int __init mds_init(void)
3476 struct lprocfs_static_vars lvars;
3478 mds_group_hash_init();
3480 lprocfs_init_multi_vars(0, &lvars);
3481 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
3483 lprocfs_init_multi_vars(1, &lvars);
3484 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
3490 static void /*__exit*/ mds_exit(void)
3492 mds_group_hash_cleanup();
3494 class_unregister_type(LUSTRE_MDS_NAME);
3495 class_unregister_type(LUSTRE_MDT_NAME);
3498 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3499 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
3500 MODULE_LICENSE("GPL");
3502 module_init(mds_init);
3503 module_exit(mds_exit);