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, *mode);
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);
412 OBD_FREE(mcd, sizeof(*mcd));
413 class_disconnect(exp, 0);
415 class_export_put(exp);
420 static int mds_connect_post(struct obd_export *exp, unsigned long flags)
422 struct obd_device *obd = exp->exp_obd;
423 struct mds_obd *mds = &obd->u.mds;
427 if (!(flags & OBD_OPT_MDS_CONNECTION)) {
428 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
429 atomic_inc(&mds->mds_real_clients);
430 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
431 obd->obd_name, exp->exp_client_uuid.uuid,
432 atomic_read(&mds->mds_real_clients));
433 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
435 if (mds->mds_md_name)
436 rc = mds_md_connect(obd, mds->mds_md_name);
441 static int mds_init_export(struct obd_export *exp)
443 struct mds_export_data *med = &exp->exp_mds_data;
445 INIT_LIST_HEAD(&med->med_open_head);
446 spin_lock_init(&med->med_open_lock);
450 static int mds_destroy_export(struct obd_export *export)
452 struct mds_export_data *med;
453 struct obd_device *obd = export->exp_obd;
454 struct lvfs_run_ctxt saved;
458 med = &export->exp_mds_data;
459 target_destroy_export(export);
461 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
464 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
466 /* Close any open files (which may also cause orphan unlinking). */
467 spin_lock(&med->med_open_lock);
468 while (!list_empty(&med->med_open_head)) {
469 struct list_head *tmp = med->med_open_head.next;
470 struct mds_file_data *mfd =
471 list_entry(tmp, struct mds_file_data, mfd_list);
472 BDEVNAME_DECLARE_STORAGE(btmp);
474 /* bug 1579: fix force-closing for 2.5 */
475 struct dentry *dentry = mfd->mfd_dentry;
477 list_del(&mfd->mfd_list);
478 spin_unlock(&med->med_open_lock);
480 /* If you change this message, be sure to update
481 * replay_single:test_46 */
482 CERROR("force closing client file handle for %*s (%s:%lu)\n",
483 dentry->d_name.len, dentry->d_name.name,
484 ll_bdevname(dentry->d_inode->i_sb, btmp),
485 dentry->d_inode->i_ino);
486 /* child inode->i_alloc_sem protects orphan_dec_test and
487 * is_orphan race, mds_mfd_close drops it */
488 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
489 rc = mds_mfd_close(NULL, 0, obd, mfd,
490 !(export->exp_flags & OBD_OPT_FAILOVER));
492 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
493 spin_lock(&med->med_open_lock);
495 spin_unlock(&med->med_open_lock);
496 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
500 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
504 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
506 unsigned long irqflags;
507 struct obd_device *obd;
512 LASSERT(exp != NULL);
513 obd = class_exp2obd(exp);
515 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
516 exp->exp_handle.h_cookie);
522 * suppress any inter-mds requests durring disconnecting lmv if this is
523 * detected --force mode. This is needed to avoid endless recovery.
525 if (atomic_read(&mds->mds_real_clients) > 0 &&
526 !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
527 flags |= OBD_OPT_FORCE;
529 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
530 && !atomic_read(&mds->mds_real_clients)) {
531 /* there was no client at all */
532 mds_md_disconnect(obd, flags);
535 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
536 && atomic_dec_and_test(&mds->mds_real_clients)) {
537 /* time to drop LMV connections */
538 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
539 "Disconnnect from LMV now\n",
540 obd->obd_name, exp->exp_client_uuid.uuid);
541 mds_md_disconnect(obd, flags);
544 spin_lock_irqsave(&exp->exp_lock, irqflags);
545 exp->exp_flags = flags;
546 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
548 /* disconnect early so that clients can't keep using export */
549 rc = class_disconnect(exp, flags);
550 ldlm_cancel_locks_for_export(exp);
552 /* complete all outstanding replies */
553 spin_lock_irqsave(&exp->exp_lock, irqflags);
554 while (!list_empty(&exp->exp_outstanding_replies)) {
555 struct ptlrpc_reply_state *rs =
556 list_entry(exp->exp_outstanding_replies.next,
557 struct ptlrpc_reply_state, rs_exp_list);
558 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
560 spin_lock(&svc->srv_lock);
561 list_del_init(&rs->rs_exp_list);
562 ptlrpc_schedule_difficult_reply(rs);
563 spin_unlock(&svc->srv_lock);
565 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
569 static int mds_getstatus(struct ptlrpc_request *req)
571 struct mds_obd *mds = mds_req2mds(req);
572 struct mds_body *body;
576 size = sizeof(*body);
578 rc = lustre_pack_reply(req, 1, &size, NULL);
579 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
580 CERROR("mds: out of memory for message: size=%d\n", size);
581 req->rq_status = -ENOMEM; /* superfluous? */
585 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
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 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
765 mds_pack_inode2body(obd, body, inode);
767 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
768 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
769 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
772 /* if we have LOV EA data, the OST holds size, atime, mtime. */
773 if (!(body->valid & OBD_MD_FLEASIZE) &&
774 !(body->valid & OBD_MD_FLDIREA))
775 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
776 OBD_MD_FLATIME | OBD_MD_FLMTIME);
777 } else if (S_ISLNK(inode->i_mode) &&
778 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
779 int len = req->rq_repmsg->buflens[reply_off + 1];
780 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1, 0);
782 LASSERT(symname != NULL); /* caller prepped reply */
784 if (!inode->i_op->readlink) {
787 rc = inode->i_op->readlink(dentry, symname, len);
789 CERROR("readlink failed: %d\n", rc);
790 } else if (rc != len - 1) {
791 CERROR("Unexpected readlink rc %d: expecting %d\n",
795 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
796 body->valid |= OBD_MD_LINKNAME;
797 body->eadatasize = rc + 1;
807 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
808 struct dentry *dentry,
811 int rc = 0, size[1] = {sizeof(struct mds_body)};
814 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
815 CERROR("failed MDS_GETATTR_PACK test\n");
816 req->rq_status = -ENOMEM;
820 rc = lustre_pack_reply(req, 1, size, NULL);
822 CERROR("lustre_pack_reply failed: rc %d\n", rc);
823 GOTO(out, req->rq_status = rc);
831 static int mds_getattr_pack_msg(struct ptlrpc_request *req,
835 struct mds_obd *mds = mds_req2mds(req);
836 struct mds_body *body;
837 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
840 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
841 LASSERT(body != NULL); /* checked by caller */
842 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
844 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
845 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
848 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
850 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
854 CERROR("error getting inode %lu MD: rc = %d\n",
857 } else if (rc > mds->mds_max_mdsize) {
859 CERROR("MD size %d larger than maximum possible %u\n",
860 rc, mds->mds_max_mdsize);
865 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
866 if (inode->i_size + 1 != body->eadatasize)
867 CERROR("symlink size: %Lu, reply space: %d\n",
868 inode->i_size + 1, body->eadatasize);
869 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
871 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
872 inode->i_size + 1, body->eadatasize);
875 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
876 CERROR("failed MDS_GETATTR_PACK test\n");
877 req->rq_status = -ENOMEM;
878 GOTO(out, rc = -ENOMEM);
881 rc = lustre_pack_reply(req, bufcount, size, NULL);
883 CERROR("out of memory\n");
884 GOTO(out, req->rq_status = rc);
892 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
893 char *name, int namelen)
895 struct mea *mea = NULL;
896 int mea_size, rc = 0;
899 rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
904 * dir is already splitted, check if requested filename should
905 * live at this MDS or at another one.
907 int i = mea_name2idx(mea, name, namelen - 1);
908 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
910 "inapropriate MDS(%d) for %s. should be "
911 "%lu(%d)\n", mea->mea_master, name,
912 (unsigned long)id_group(&mea->mea_ids[i]), i);
918 OBD_FREE(mea, mea_size);
922 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
923 struct lustre_handle *child_lockh, int child_part)
925 struct obd_device *obd = req->rq_export->exp_obd;
926 struct mds_obd *mds = &obd->u.mds;
927 struct ldlm_reply *rep = NULL;
928 struct lvfs_run_ctxt saved;
929 struct mds_req_sec_desc *rsd;
930 struct mds_body *body;
931 struct dentry *dparent = NULL, *dchild = NULL;
932 struct lvfs_ucred uc;
933 struct lustre_handle parent_lockh[2] = {{0}, {0}};
934 unsigned int namesize;
935 int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
939 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
940 MD_COUNTER_INCREMENT(obd, getattr_lock);
942 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
944 CERROR("Can't unpack security desc\n");
947 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
949 /* swab now, before anyone looks inside the request. */
950 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
951 lustre_swab_mds_body);
953 CERROR("Can't swab mds_body\n");
954 GOTO(cleanup, rc = -EFAULT);
957 LASSERT_REQSWAB(req, offset + 1);
958 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
960 CERROR("Can't unpack name\n");
961 GOTO(cleanup, rc = -EFAULT);
963 namesize = req->rq_reqmsg->buflens[offset + 1];
965 /* namesize less than 2 means we have empty name, probably came from
966 revalidate by cfid, so no point in having name to be set */
970 LASSERT (offset == 1 || offset == 3);
972 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
978 rc = mds_init_ucred(&uc, rsd);
980 CERROR("can't init ucred\n");
984 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
985 cleanup_phase = 1; /* kernel context */
986 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
988 LASSERT(namesize > 0);
989 if (child_lockh->cookie != 0) {
990 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
995 if (body->valid == OBD_MD_FLID) {
996 struct mds_body *mds_reply;
997 int size = sizeof(*mds_reply);
1001 dparent = mds_id2dentry(obd, &body->id1, NULL);
1002 if (IS_ERR(dparent)) {
1003 rc = PTR_ERR(dparent);
1008 * the user requested ONLY the inode number, so do a raw lookup.
1010 rc = lustre_pack_reply(req, 1, &size, NULL);
1012 CERROR("out of memory\n");
1016 dir = dparent->d_inode;
1017 LASSERT(dir->i_op->lookup_raw != NULL);
1018 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1020 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1021 sizeof(*mds_reply));
1023 id_ino(&mds_reply->id1) = inum;
1024 mds_reply->valid = OBD_MD_FLID;
1029 if (resent_req == 0) {
1030 LASSERT(id_fid(&body->id1) != 0);
1032 rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1033 parent_lockh, &dparent,
1035 MDS_INODELOCK_UPDATE,
1038 child_lockh, &dchild,
1039 LCK_PR, child_part);
1044 * let's make sure this name should leave on this mds
1047 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1051 /* we have no dentry here, drop LOOKUP bit */
1052 /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1053 child_part &= ~MDS_INODELOCK_LOOKUP;
1054 CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1055 obd->obd_name, OLID4(&body->id1));
1057 dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1058 LCK_PR, parent_lockh,
1061 MDS_INODELOCK_UPDATE);
1062 if (IS_ERR(dchild)) {
1063 CERROR("can't find inode with id "DLID4", err = %d\n",
1064 OLID4(&body->id1), (int)PTR_ERR(dchild));
1065 GOTO(cleanup, rc = PTR_ERR(dchild));
1067 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1070 struct ldlm_lock *granted_lock;
1072 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1073 granted_lock = ldlm_handle2lock(child_lockh);
1075 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1076 id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1077 child_lockh->cookie);
1079 dparent = mds_id2dentry(obd, &body->id1, NULL);
1082 dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1084 LDLM_LOCK_PUT(granted_lock);
1087 cleanup_phase = 2; /* dchild, dparent, locks */
1089 if (!DENTRY_VALID(dchild)) {
1090 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1092 * in the intent case, the policy clears this error: the
1093 * disposition is enough.
1098 intent_set_disposition(rep, DISP_LOOKUP_POS);
1101 if (req->rq_repmsg == NULL) {
1102 if (dchild->d_flags & DCACHE_CROSS_REF)
1103 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1105 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
1107 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1112 rc = mds_getattr_internal(obd, dchild, req, body, reply_offset);
1113 GOTO(cleanup, rc); /* returns the lock to the client */
1116 switch (cleanup_phase) {
1118 if (resent_req == 0) {
1119 if (rc && DENTRY_VALID(dchild))
1120 ldlm_lock_decref(child_lockh, LCK_PR);
1122 ldlm_lock_decref(parent_lockh, LCK_PR);
1124 if (parent_lockh[1].cookie != 0)
1125 ldlm_lock_decref(parent_lockh + 1, update_mode);
1132 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1133 mds_exit_ucred(&uc);
1138 static int mds_getattr(struct ptlrpc_request *req, int offset)
1140 struct obd_device *obd = req->rq_export->exp_obd;
1141 struct lvfs_run_ctxt saved;
1143 struct mds_req_sec_desc *rsd;
1144 struct mds_body *body;
1145 struct lvfs_ucred uc;
1149 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1151 CERROR("Can't unpack security desc\n");
1155 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1156 lustre_swab_mds_body);
1158 CERROR ("Can't unpack body\n");
1162 MD_COUNTER_INCREMENT(obd, getattr);
1164 rc = mds_init_ucred(&uc, rsd);
1166 CERROR("can't init ucred\n");
1170 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1171 de = mds_id2dentry(obd, &body->id1, NULL);
1173 rc = req->rq_status = PTR_ERR(de);
1177 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1179 CERROR("mds_getattr_pack_msg: %d\n", rc);
1183 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
1189 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1190 mds_exit_ucred(&uc);
1194 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1195 unsigned long max_age)
1200 spin_lock(&obd->obd_osfs_lock);
1201 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1203 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1204 spin_unlock(&obd->obd_osfs_lock);
1209 static int mds_statfs(struct ptlrpc_request *req)
1211 struct obd_device *obd = req->rq_export->exp_obd;
1212 int rc, size = sizeof(struct obd_statfs);
1215 rc = lustre_pack_reply(req, 1, &size, NULL);
1216 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1217 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1221 OBD_COUNTER_INCREMENT(obd, statfs);
1223 /* We call this so that we can cache a bit - 1 jiffie worth */
1224 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1227 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1233 req->rq_status = rc;
1237 static int mds_sync(struct ptlrpc_request *req, int offset)
1239 struct obd_device *obd = req->rq_export->exp_obd;
1240 struct mds_obd *mds = &obd->u.mds;
1241 struct mds_body *body;
1242 int rc, size = sizeof(*body);
1245 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1247 GOTO(out, rc = -EPROTO);
1249 rc = lustre_pack_reply(req, 1, &size, NULL);
1250 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1251 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1255 if (id_ino(&body->id1) == 0) {
1256 /* an id of zero is taken to mean "sync whole filesystem" */
1257 rc = fsfilt_sync(obd, mds->mds_sb);
1261 /* just any file to grab fsync method - "file" arg unused */
1262 struct file *file = mds->mds_rcvd_filp;
1265 de = mds_id2dentry(obd, &body->id1, NULL);
1267 GOTO(out, rc = PTR_ERR(de));
1269 rc = file->f_op->fsync(NULL, de, 1);
1273 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1274 mds_pack_inode2body(obd, body, de->d_inode);
1280 req->rq_status = rc;
1284 /* mds_readpage does not take a DLM lock on the inode, because the client must
1285 * already have a PR lock.
1287 * If we were to take another one here, a deadlock will result, if another
1288 * thread is already waiting for a PW lock. */
1289 static int mds_readpage(struct ptlrpc_request *req, int offset)
1291 struct obd_device *obd = req->rq_export->exp_obd;
1292 struct mds_obd *mds = &obd->u.mds;
1293 struct vfsmount *mnt;
1296 struct mds_req_sec_desc *rsd;
1297 struct mds_body *body, *repbody;
1298 struct lvfs_run_ctxt saved;
1299 int rc, size = sizeof(*repbody);
1300 struct lvfs_ucred uc;
1303 rc = lustre_pack_reply(req, 1, &size, NULL);
1304 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1305 CERROR("mds: out of memory\n");
1306 GOTO(out, rc = -ENOMEM);
1309 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1311 CERROR("Can't unpack security desc\n");
1312 GOTO (out, rc = -EFAULT);
1314 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
1316 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1317 lustre_swab_mds_body);
1319 CERROR("Can't unpack body\n");
1320 GOTO (out, rc = -EFAULT);
1323 rc = mds_init_ucred(&uc, rsd);
1325 CERROR("can't init ucred\n");
1329 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1330 de = mds_id2dentry(obd, &body->id1, &mnt);
1332 GOTO(out_pop, rc = PTR_ERR(de));
1334 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1336 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1337 /* note: in case of an error, dentry_open puts dentry */
1339 GOTO(out_pop, rc = PTR_ERR(file));
1341 /* body->size is actually the offset -eeb */
1342 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1343 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1344 body->size, de->d_inode->i_blksize);
1345 GOTO(out_file, rc = -EFAULT);
1348 /* body->nlink is actually the #bytes to read -eeb */
1349 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1350 CERROR("size %u is not multiple of blocksize %lu\n",
1351 body->nlink, de->d_inode->i_blksize);
1352 GOTO(out_file, rc = -EFAULT);
1355 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1356 repbody->size = file->f_dentry->d_inode->i_size;
1357 repbody->valid = OBD_MD_FLSIZE;
1359 /* to make this asynchronous make sure that the handling function
1360 doesn't send a reply when this function completes. Instead a
1361 callback function would send the reply */
1362 /* body->size is actually the offset -eeb */
1363 rc = mds_sendpage(req, file, body->size, body->nlink);
1367 filp_close(file, 0);
1369 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1370 mds_exit_ucred(&uc);
1372 req->rq_status = rc;
1376 /* update master MDS ID, which is stored in local inode EA. */
1377 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1378 void *data, int data_len)
1380 struct mds_obd *mds = &obd->u.mds;
1381 struct dentry *dentry;
1389 dentry = mds_id2dentry(obd, id, NULL);
1391 GOTO(out, rc = PTR_ERR(dentry));
1393 if (!dentry->d_inode) {
1394 CERROR("Can't find object "DLID4".\n",
1396 GOTO(out_dentry, rc = -EINVAL);
1399 handle = fsfilt_start(obd, dentry->d_inode,
1400 FSFILT_OP_SETATTR, NULL);
1402 GOTO(out_dentry, rc = PTR_ERR(handle));
1404 rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1405 (struct lustre_id *)data);
1407 CERROR("Can't update inode "DLID4" master id, "
1408 "error = %d.\n", OLID4(id), rc);
1409 GOTO(out_commit, rc);
1414 fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1421 EXPORT_SYMBOL(mds_update_mid);
1423 /* read master MDS ID, which is stored in local inode EA. */
1424 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1425 void *data, int data_len)
1427 struct dentry *dentry;
1434 dentry = mds_id2dentry(obd, id, NULL);
1436 GOTO(out, rc = PTR_ERR(dentry));
1438 if (!dentry->d_inode) {
1439 CERROR("Can't find object "DLID4".\n",
1441 GOTO(out_dentry, rc = -EINVAL);
1444 down(&dentry->d_inode->i_sem);
1445 rc = mds_read_inode_mid(obd, dentry->d_inode,
1446 (struct lustre_id *)data);
1447 up(&dentry->d_inode->i_sem);
1449 CERROR("Can't read inode "DLID4" master id, "
1450 "error = %d.\n", OLID4(id), rc);
1451 GOTO(out_dentry, rc);
1460 EXPORT_SYMBOL(mds_read_mid);
1462 int mds_reint(struct ptlrpc_request *req, int offset,
1463 struct lustre_handle *lockh)
1465 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1466 struct mds_update_record *rec;
1467 struct mds_req_sec_desc *rsd;
1471 OBD_ALLOC(rec, sizeof(*rec));
1475 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1477 CERROR("Can't unpack security desc\n");
1478 GOTO(out, rc = -EFAULT);
1480 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
1482 rc = mds_update_unpack(req, offset, rec);
1483 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1484 CERROR("invalid record\n");
1485 GOTO(out, req->rq_status = -EINVAL);
1488 rc = mds_init_ucred(&rec->ur_uc, rsd);
1490 CERROR("can't init ucred\n");
1494 /* rc will be used to interrupt a for loop over multiple records */
1495 rc = mds_reint_rec(rec, offset, req, lockh);
1496 mds_exit_ucred(&rec->ur_uc);
1499 OBD_FREE(rec, sizeof(*rec));
1503 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1504 struct obd_device *obd, int *process)
1506 switch (req->rq_reqmsg->opc) {
1507 case MDS_CONNECT: /* This will never get here, but for completeness. */
1508 case OST_CONNECT: /* This will never get here, but for completeness. */
1509 case MDS_DISCONNECT:
1510 case OST_DISCONNECT:
1515 case MDS_SYNC: /* used in unmounting */
1520 *process = target_queue_recovery_request(req, obd);
1524 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1526 /* XXX what should we set rq_status to here? */
1527 req->rq_status = -EAGAIN;
1528 RETURN(ptlrpc_error(req));
1532 static char *reint_names[] = {
1533 [REINT_SETATTR] "setattr",
1534 [REINT_CREATE] "create",
1535 [REINT_LINK] "link",
1536 [REINT_UNLINK] "unlink",
1537 [REINT_RENAME] "rename",
1538 [REINT_OPEN] "open",
1541 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER | \
1542 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
1543 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
1546 static void reconstruct_create(struct ptlrpc_request *req)
1548 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1549 struct mds_client_data *mcd = med->med_mcd;
1550 struct dentry *dentry;
1551 struct ost_body *body;
1552 struct lustre_id id;
1556 /* copy rc, transno and disp; steal locks */
1557 mds_req_from_mcd(req, mcd);
1558 if (req->rq_status) {
1566 id_ino(&id) = mcd->mcd_last_data;
1567 LASSERT(id_ino(&id) != 0);
1569 dentry = mds_id2dentry(req2obd(req), &id, NULL);
1570 if (IS_ERR(dentry)) {
1571 CERROR("can't find inode "LPU64"\n", id_ino(&id));
1572 req->rq_status = PTR_ERR(dentry);
1577 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1578 req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
1579 dentry->d_inode->i_generation);
1581 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1582 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
1583 body->oa.o_id = dentry->d_inode->i_ino;
1584 body->oa.o_generation = dentry->d_inode->i_generation;
1585 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1587 down(&dentry->d_inode->i_sem);
1588 rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
1589 up(&dentry->d_inode->i_sem);
1591 CERROR("Can't read inode self id, inode %lu, "
1592 "rc %d\n", dentry->d_inode->i_ino, rc);
1596 body->oa.o_fid = id_fid(&id);
1597 body->oa.o_mds = id_group(&id);
1603 static int mdt_obj_create(struct ptlrpc_request *req)
1605 struct obd_device *obd = req->rq_export->exp_obd;
1606 struct mds_obd *mds = &obd->u.mds;
1607 struct ost_body *body, *repbody;
1608 char idname[LL_ID_NAMELEN];
1609 int size = sizeof(*repbody);
1610 struct inode *parent_inode;
1611 struct lvfs_run_ctxt saved;
1612 int rc, cleanup_phase = 0;
1613 struct dentry *new = NULL;
1614 struct dentry_params dp;
1615 int mealen, flags = 0;
1616 struct lvfs_ucred uc;
1617 struct lustre_id id;
1619 void *handle = NULL;
1620 unsigned long cr_inum = 0;
1623 DEBUG_REQ(D_HA, req, "create remote object");
1624 parent_inode = mds->mds_unnamed_dir->d_inode;
1626 body = lustre_swab_reqbuf(req, 0, sizeof(*body),
1627 lustre_swab_ost_body);
1631 rc = lustre_pack_reply(req, 1, &size, NULL);
1635 MDS_CHECK_RESENT(req, reconstruct_create(req));
1638 * this only serve to inter-mds request, don't need check group database
1641 uc.luc_ghash = NULL;
1642 uc.luc_ginfo = NULL;
1643 uc.luc_uid = body->oa.o_uid;
1644 uc.luc_fsuid = body->oa.o_uid;
1645 uc.luc_fsgid = body->oa.o_gid;
1647 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1648 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1650 /* in REPLAY case inum should be given (client or other MDS fills it) */
1651 if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
1652 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
1654 * this is re-create request from MDS holding directory name.
1655 * we have to lookup given ino/gen first. if it exists (good
1656 * case) then there is nothing to do. if it does not then we
1657 * have to recreate it.
1659 id_ino(&id) = body->oa.o_id;
1660 id_gen(&id) = body->oa.o_generation;
1662 new = mds_id2dentry(obd, &id, NULL);
1663 if (!IS_ERR(new) && new->d_inode) {
1664 struct lustre_id sid;
1666 CWARN("mkdir() repairing is on its way: %lu/%lu\n",
1667 (unsigned long)id_ino(&id), (unsigned long)id_gen(&id));
1669 obdo_from_inode(&repbody->oa, new->d_inode,
1670 FILTER_VALID_FLAGS);
1672 repbody->oa.o_id = new->d_inode->i_ino;
1673 repbody->oa.o_generation = new->d_inode->i_generation;
1674 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1677 down(&new->d_inode->i_sem);
1678 rc = mds_read_inode_sid(obd, new->d_inode, &sid);
1679 up(&new->d_inode->i_sem);
1681 CERROR("Can't read inode self id "
1682 "inode %lu, rc %d.\n",
1683 new->d_inode->i_ino, rc);
1687 repbody->oa.o_fid = id_fid(&sid);
1688 repbody->oa.o_mds = id_group(&sid);
1689 LASSERT(id_fid(&sid) != 0);
1692 * here we could use fid passed in body->oa.o_fid and
1693 * thus avoid mds_read_inode_sid().
1695 cr_inum = new->d_inode->i_ino;
1696 GOTO(cleanup, rc = 0);
1700 down(&parent_inode->i_sem);
1701 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1702 if (IS_ERR(handle)) {
1703 up(&parent_inode->i_sem);
1704 CERROR("fsfilt_start() failed, rc = %d\n",
1705 (int)PTR_ERR(handle));
1706 GOTO(cleanup, rc = PTR_ERR(handle));
1708 cleanup_phase = 1; /* transaction */
1711 rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
1712 new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
1714 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1715 obd->obd_name, idname, (int) PTR_ERR(new));
1716 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
1717 up(&parent_inode->i_sem);
1718 RETURN(PTR_ERR(new));
1719 } else if (new->d_inode) {
1720 CERROR("%s: name exists. repeat\n", obd->obd_name);
1724 new->d_fsdata = (void *)&dp;
1728 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
1729 (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
1730 LASSERT(body->oa.o_id != 0);
1731 dp.p_inum = body->oa.o_id;
1732 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1733 (unsigned long)body->oa.o_id,
1734 (unsigned long)body->oa.o_generation);
1737 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1739 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
1740 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1741 new->d_inode->i_generation = body->oa.o_generation;
1742 mark_inode_dirty(new->d_inode);
1745 * avoiding asserts in cache flush case, as
1746 * @body->oa.o_id should be zero.
1748 if (body->oa.o_id) {
1749 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
1750 "BUG 3550: failed to recreate obj "
1751 LPU64" -> %lu\n", body->oa.o_id,
1752 new->d_inode->i_ino);
1754 LASSERTF(body->oa.o_generation ==
1755 new->d_inode->i_generation,
1756 "BUG 3550: failed to recreate obj/gen "
1757 LPU64"/%u -> %lu/%u\n", body->oa.o_id,
1758 body->oa.o_generation,
1759 new->d_inode->i_ino,
1760 new->d_inode->i_generation);
1764 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1765 repbody->oa.o_id = new->d_inode->i_ino;
1766 repbody->oa.o_generation = new->d_inode->i_generation;
1767 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1769 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
1770 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1771 id_group(&id) = mds->mds_num;
1773 LASSERT(body->oa.o_fid != 0);
1774 id_fid(&id) = body->oa.o_fid;
1776 LASSERT(body->oa.o_id != 0);
1777 id_ino(&id) = repbody->oa.o_id;
1778 id_gen(&id) = repbody->oa.o_generation;
1780 down(&new->d_inode->i_sem);
1781 rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
1782 up(&new->d_inode->i_sem);
1785 * make sure, that fid is up-to-date.
1787 mds_set_last_fid(obd, id_fid(&id));
1790 * allocate new sid, as object is created from scratch
1791 * and this is not replay.
1793 down(&new->d_inode->i_sem);
1794 rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
1795 up(&new->d_inode->i_sem);
1798 CERROR("Can't update lustre ID for inode %lu, "
1799 "error = %d\n", new->d_inode->i_ino, rc);
1803 /* initializing o_fid after it is allocated. */
1804 repbody->oa.o_fid = id_fid(&id);
1805 repbody->oa.o_mds = id_group(&id);
1807 rc = fsfilt_del_dir_entry(obd, new);
1808 up(&parent_inode->i_sem);
1810 CERROR("can't remove name for object: %d\n", rc);
1814 cleanup_phase = 2; /* created directory object */
1816 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
1817 (unsigned long)new->d_inode->i_ino,
1818 (unsigned long)new->d_inode->i_generation,
1819 (unsigned)new->d_inode->i_mode);
1820 cr_inum = new->d_inode->i_ino;
1822 up(&parent_inode->i_sem);
1823 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1827 if (body->oa.o_valid & OBD_MD_FLID) {
1828 /* this is new object for splitted dir. We have to prevent
1829 * recursive splitting on it -bzzz */
1830 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
1832 OBD_ALLOC(mea, mealen);
1834 GOTO(cleanup, rc = -ENOMEM);
1836 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
1837 mea->mea_master = 0;
1840 down(&new->d_inode->i_sem);
1841 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1842 up(&new->d_inode->i_sem);
1844 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
1846 OBD_FREE(mea, mealen);
1847 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
1848 obd->obd_name, new->d_inode->i_ino,
1849 new->d_inode->i_generation, flags);
1850 } else if (body->oa.o_easize) {
1851 /* we pass LCK_EX to split routine to signal that we have
1852 * exclusive access to the directory. simple because nobody
1853 * knows it already exists -bzzz */
1854 rc = mds_try_to_split_dir(obd, new, NULL,
1855 body->oa.o_easize, LCK_EX);
1857 CERROR("Can't split directory %lu, error = %d.\n",
1858 new->d_inode->i_ino, rc);
1866 switch (cleanup_phase) {
1867 case 2: /* object has been created, but we'll may want to replay it later */
1869 ptlrpc_require_repack(req);
1870 case 1: /* transaction */
1871 rc = mds_finish_transno(mds, parent_inode, handle,
1876 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1877 mds_put_group_entry(mds, uc.luc_ghash);
1881 static int mdt_get_info(struct ptlrpc_request *req)
1883 struct obd_export *exp = req->rq_export;
1888 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1890 DEBUG_REQ(D_HA, req, "no get_info key");
1893 keylen = req->rq_reqmsg->buflens[0];
1895 if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
1896 (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
1897 (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
1900 if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
1901 struct lustre_id *reply;
1902 int size = sizeof(*reply);
1904 rc = lustre_pack_reply(req, 1, &size, NULL);
1908 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
1909 rc = obd_get_info(exp, keylen, key, &size, reply);
1912 int size = sizeof(*reply);
1914 rc = lustre_pack_reply(req, 1, &size, NULL);
1918 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
1919 rc = obd_get_info(exp, keylen, key, &size, reply);
1922 req->rq_repmsg->status = 0;
1926 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1927 void *key, __u32 vallen, void *val)
1929 struct obd_device *obd;
1930 struct mds_obd *mds;
1934 obd = class_exp2obd(exp);
1936 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1937 exp->exp_handle.h_cookie);
1942 if (keylen >= strlen("mds_type") &&
1943 memcmp(key, "mds_type", keylen) == 0) {
1947 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
1949 mds->mds_obd_type = *(int*)val;
1950 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
1951 valsize = sizeof(group);
1953 /* mds number has been changed, so the corresponding obdfilter
1954 * exp need to be changed too. */
1955 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
1956 "mds_conn", valsize, &group);
1959 CDEBUG(D_IOCTL, "invalid key\n");
1963 static int mdt_set_info(struct ptlrpc_request *req)
1966 struct obd_export *exp = req->rq_export;
1967 int keylen, rc = 0, vallen;
1970 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1972 DEBUG_REQ(D_HA, req, "no set_info key");
1975 keylen = req->rq_reqmsg->buflens[0];
1977 if (keylen == strlen("mds_type") &&
1978 memcmp(key, "mds_type", keylen) == 0) {
1979 rc = lustre_pack_reply(req, 0, NULL, NULL);
1983 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
1984 vallen = req->rq_reqmsg->buflens[1];
1986 rc = obd_set_info(exp, keylen, key, vallen, val);
1987 req->rq_repmsg->status = 0;
1990 CDEBUG(D_IOCTL, "invalid key\n");
1994 static int mds_msg_check_version(struct lustre_msg *msg)
2000 case MDS_DISCONNECT:
2002 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2004 CERROR("bad opc %u version %08x, expecting %08x\n",
2005 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2010 case MDS_GETATTR_LOCK:
2014 case MDS_DONE_WRITING:
2017 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2019 CERROR("bad opc %u version %08x, expecting %08x\n",
2020 msg->opc, msg->version, LUSTRE_MDS_VERSION);
2024 case LDLM_BL_CALLBACK:
2025 case LDLM_CP_CALLBACK:
2026 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2028 CERROR("bad opc %u version %08x, expecting %08x\n",
2029 msg->opc, msg->version, LUSTRE_DLM_VERSION);
2031 case OBD_LOG_CANCEL:
2032 case LLOG_ORIGIN_HANDLE_OPEN:
2033 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2034 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2035 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2036 case LLOG_ORIGIN_HANDLE_CLOSE:
2038 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2040 CERROR("bad opc %u version %08x, expecting %08x\n",
2041 msg->opc, msg->version, LUSTRE_LOG_VERSION);
2047 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2049 CERROR("bad opc %u version %08x, expecting %08x\n",
2050 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2053 CERROR("MDS unknown opcode %d\n", msg->opc);
2061 static char str[PTL_NALFMT_SIZE];
2063 int mds_handle(struct ptlrpc_request *req)
2065 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2066 struct obd_device *obd = NULL;
2067 struct mds_obd *mds = NULL; /* quell gcc overwarning */
2071 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2073 rc = mds_msg_check_version(req->rq_reqmsg);
2075 CERROR("MDS drop mal-formed request\n");
2079 LASSERT(current->journal_info == NULL);
2080 /* XXX identical to OST */
2081 if (req->rq_reqmsg->opc != MDS_CONNECT) {
2082 struct mds_export_data *med;
2085 if (req->rq_export == NULL) {
2086 CERROR("operation %d on unconnected MDS from NID %s\n",
2087 req->rq_reqmsg->opc,
2088 ptlrpc_peernid2str(&req->rq_peer, str));
2089 req->rq_status = -ENOTCONN;
2090 GOTO(out, rc = -ENOTCONN);
2093 med = &req->rq_export->exp_mds_data;
2094 obd = req->rq_export->exp_obd;
2097 /* sanity check: if the xid matches, the request must
2098 * be marked as a resent or replayed */
2099 if (req->rq_xid == med->med_mcd->mcd_last_xid) {
2100 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2101 (MSG_RESENT | MSG_REPLAY),
2102 "rq_xid "LPU64" matches last_xid, "
2103 "expected RESENT flag\n",
2106 /* else: note the opposite is not always true; a
2107 * RESENT req after a failover will usually not match
2108 * the last_xid, since it was likely never
2109 * committed. A REPLAYed request will almost never
2110 * match the last xid, however it could for a
2111 * committed, but still retained, open. */
2113 spin_lock_bh(&obd->obd_processing_task_lock);
2114 recovering = obd->obd_recovering;
2115 spin_unlock_bh(&obd->obd_processing_task_lock);
2117 rc = mds_filter_recovery_request(req, obd,
2119 if (rc || should_process == 0) {
2121 } else if (should_process < 0) {
2122 req->rq_status = should_process;
2123 rc = ptlrpc_error(req);
2129 switch (req->rq_reqmsg->opc) {
2131 DEBUG_REQ(D_INODE, req, "connect");
2132 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2133 rc = target_handle_connect(req);
2135 /* Now that we have an export, set mds. */
2136 mds = mds_req2mds(req);
2139 case MDS_DISCONNECT:
2140 DEBUG_REQ(D_INODE, req, "disconnect");
2141 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2142 rc = target_handle_disconnect(req);
2143 req->rq_status = rc; /* superfluous? */
2147 DEBUG_REQ(D_INODE, req, "getstatus");
2148 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2149 rc = mds_getstatus(req);
2153 DEBUG_REQ(D_INODE, req, "getattr");
2154 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2155 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2158 case MDS_GETATTR_LOCK: {
2159 struct lustre_handle lockh;
2160 DEBUG_REQ(D_INODE, req, "getattr_lock");
2161 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2163 /* If this request gets a reconstructed reply, we won't be
2164 * acquiring any new locks in mds_getattr_lock, so we don't
2168 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2169 MDS_INODELOCK_UPDATE);
2170 /* this non-intent call (from an ioctl) is special */
2171 req->rq_status = rc;
2172 if (rc == 0 && lockh.cookie)
2173 ldlm_lock_decref(&lockh, LCK_PR);
2177 DEBUG_REQ(D_INODE, req, "statfs");
2178 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2179 rc = mds_statfs(req);
2183 DEBUG_REQ(D_INODE, req, "readpage");
2184 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2185 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2187 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2188 if (req->rq_reply_state) {
2189 lustre_free_reply_state (req->rq_reply_state);
2190 req->rq_reply_state = NULL;
2197 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2200 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2201 mds->mds_max_cookiesize};
2204 /* NB only peek inside req now; mds_reint() will swab it */
2206 CERROR ("Can't inspect opcode\n");
2211 if (lustre_msg_swabbed (req->rq_reqmsg))
2214 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2215 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2216 reint_names[opc] == NULL) ? reint_names[opc] :
2219 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2221 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2223 else if (opc == REINT_OPEN)
2228 rc = lustre_pack_reply(req, bufcount, size, NULL);
2232 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
2233 fail = OBD_FAIL_MDS_REINT_NET_REP;
2238 DEBUG_REQ(D_INODE, req, "close");
2239 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
2240 rc = mds_close(req, MDS_REQ_REC_OFF);
2243 case MDS_DONE_WRITING:
2244 DEBUG_REQ(D_INODE, req, "done_writing");
2245 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
2246 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2250 DEBUG_REQ(D_INODE, req, "pin");
2251 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2252 rc = mds_pin(req, MDS_REQ_REC_OFF);
2256 DEBUG_REQ(D_INODE, req, "sync");
2257 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2258 rc = mds_sync(req, MDS_REQ_REC_OFF);
2262 DEBUG_REQ(D_INODE, req, "ping");
2263 rc = target_handle_ping(req);
2266 case OBD_LOG_CANCEL:
2267 CDEBUG(D_INODE, "log cancel\n");
2268 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2269 rc = -ENOTSUPP; /* la la la */
2273 DEBUG_REQ(D_INODE, req, "enqueue");
2274 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2275 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2276 ldlm_server_blocking_ast, NULL);
2277 fail = OBD_FAIL_LDLM_REPLY;
2280 DEBUG_REQ(D_INODE, req, "convert");
2281 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2282 rc = ldlm_handle_convert(req);
2284 case LDLM_BL_CALLBACK:
2285 case LDLM_CP_CALLBACK:
2286 DEBUG_REQ(D_INODE, req, "callback");
2287 CERROR("callbacks should not happen on MDS\n");
2289 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2291 case LLOG_ORIGIN_HANDLE_OPEN:
2292 DEBUG_REQ(D_INODE, req, "llog_init");
2293 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2294 rc = llog_origin_handle_open(req);
2296 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2297 DEBUG_REQ(D_INODE, req, "llog next block");
2298 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2299 rc = llog_origin_handle_next_block(req);
2301 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2302 DEBUG_REQ(D_INODE, req, "llog prev block");
2303 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2304 rc = llog_origin_handle_prev_block(req);
2306 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2307 DEBUG_REQ(D_INODE, req, "llog read header");
2308 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2309 rc = llog_origin_handle_read_header(req);
2311 case LLOG_ORIGIN_HANDLE_CLOSE:
2312 DEBUG_REQ(D_INODE, req, "llog close");
2313 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2314 rc = llog_origin_handle_close(req);
2317 DEBUG_REQ(D_INODE, req, "ost_create");
2318 rc = mdt_obj_create(req);
2321 DEBUG_REQ(D_INODE, req, "get_info");
2322 rc = mdt_get_info(req);
2325 DEBUG_REQ(D_INODE, req, "set_info");
2326 rc = mdt_set_info(req);
2329 CDEBUG(D_INODE, "write\n");
2330 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2331 rc = ost_brw_write(req, NULL);
2332 LASSERT(current->journal_info == NULL);
2333 /* mdt_brw sends its own replies */
2337 DEBUG_REQ(D_INODE, req, "llog catinfo");
2338 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2339 rc = llog_catinfo(req);
2342 req->rq_status = -ENOTSUPP;
2343 rc = ptlrpc_error(req);
2347 LASSERT(current->journal_info == NULL);
2351 /* If we're DISCONNECTing, the mds_export_data is already freed */
2352 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
2353 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2354 struct obd_device *obd = list_entry(mds, struct obd_device,
2356 req->rq_repmsg->last_xid =
2357 le64_to_cpu(med->med_mcd->mcd_last_xid);
2359 if (!obd->obd_no_transno) {
2360 req->rq_repmsg->last_committed =
2361 obd->obd_last_committed;
2363 DEBUG_REQ(D_IOCTL, req,
2364 "not sending last_committed update");
2366 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2368 mds->mds_last_transno, obd->obd_last_committed,
2373 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
2374 if (obd && obd->obd_recovering) {
2375 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
2376 return target_queue_final_reply(req, rc);
2378 /* Lost a race with recovery; let the error path DTRT. */
2379 rc = req->rq_status = -ENOTCONN;
2382 target_send_reply(req, rc, fail);
2386 /* Update the server data on disk. This stores the new mount_count and also the
2387 * last_rcvd value to disk. If we don't have a clean shutdown, then the server
2388 * last_rcvd value may be less than that of the clients. This will alert us
2389 * that we may need to do client recovery.
2391 * Also assumes for mds_last_transno that we are not modifying it (no locking).
2393 int mds_update_server_data(struct obd_device *obd, int force_sync)
2395 struct mds_obd *mds = &obd->u.mds;
2396 struct mds_server_data *msd = mds->mds_server_data;
2397 struct file *filp = mds->mds_rcvd_filp;
2398 struct lvfs_run_ctxt saved;
2403 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2404 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2406 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2407 mds->mds_mount_count, mds->mds_last_transno);
2408 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
2410 CERROR("error writing MDS server data: rc = %d\n", rc);
2411 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2416 /* saves last allocated fid counter to file. */
2417 int mds_update_last_fid(struct obd_device *obd, void *handle,
2420 struct mds_obd *mds = &obd->u.mds;
2421 struct file *filp = mds->mds_fid_filp;
2422 struct lvfs_run_ctxt saved;
2428 spin_lock(&mds->mds_last_fid_lock);
2429 last_fid = mds->mds_last_fid;
2430 spin_unlock(&mds->mds_last_fid_lock);
2432 CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
2436 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
2437 handle, mds_commit_last_fid_cb,
2441 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2442 rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
2444 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2447 CERROR("error writing MDS last_fid #"LPU64
2448 ", err = %d\n", last_fid, rc);
2452 CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
2453 "%llu: err = %d\n", last_fid, off, rc);
2458 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
2460 struct mds_obd *mds = &obd->u.mds;
2462 spin_lock(&mds->mds_last_fid_lock);
2463 if (fid > mds->mds_last_fid)
2464 mds->mds_last_fid = fid;
2465 spin_unlock(&mds->mds_last_fid_lock);
2468 void mds_commit_last_transno_cb(struct obd_device *obd,
2469 __u64 transno, void *data,
2472 obd_transno_commit_cb(obd, transno, error);
2475 void mds_commit_last_fid_cb(struct obd_device *obd,
2476 __u64 fid, void *data,
2480 CERROR("%s: fid "LPD64" commit error: %d\n",
2481 obd->obd_name, fid, error);
2485 CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
2486 obd->obd_name, fid);
2489 __u64 mds_alloc_fid(struct obd_device *obd)
2491 struct mds_obd *mds = &obd->u.mds;
2494 spin_lock(&mds->mds_last_fid_lock);
2495 fid = ++mds->mds_last_fid;
2496 spin_unlock(&mds->mds_last_fid_lock);
2502 * allocates new lustre_id on passed @inode and saves it to inode EA.
2504 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
2505 void *handle, struct lustre_id *id)
2507 struct mds_obd *mds = &obd->u.mds;
2508 int alloc = 0, rc = 0;
2511 LASSERT(obd != NULL);
2512 LASSERT(inode != NULL);
2515 OBD_ALLOC(id, sizeof(*id));
2521 id_group(id) = mds->mds_num;
2522 id_fid(id) = mds_alloc_fid(obd);
2523 id_ino(id) = inode->i_ino;
2524 id_gen(id) = inode->i_generation;
2525 id_type(id) = (S_IFMT & inode->i_mode);
2527 rc = mds_update_inode_sid(obd, inode, handle, id);
2529 CERROR("Can't update inode FID EA, "
2534 OBD_FREE(id, sizeof(*id));
2539 * reads inode self id from inode EA. Probably later this should be replaced by
2540 * caching inode self id to avoid raeding it every time it is needed.
2542 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
2543 struct lustre_id *id)
2548 LASSERT(id != NULL);
2549 LASSERT(obd != NULL);
2550 LASSERT(inode != NULL);
2552 rc = fsfilt_get_sid(obd, inode, &id->li_fid,
2553 sizeof(id->li_fid));
2555 CERROR("fsfilt_get_sid() failed, "
2568 /* updates inode self id in EA. */
2569 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
2570 void *handle, struct lustre_id *id)
2575 LASSERT(id != NULL);
2576 LASSERT(obd != NULL);
2577 LASSERT(inode != NULL);
2579 rc = fsfilt_set_sid(obd, inode, handle, &id->li_fid,
2580 sizeof(id->li_fid));
2582 CERROR("fsfilt_set_sid() failed, rc = %d\n", rc);
2590 * reads inode id on master MDS. This is usualy done by CMOBD to update requests
2591 * to master MDS by correct store cookie, needed to find inode on master MDS
2594 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
2595 struct lustre_id *id)
2600 LASSERT(id != NULL);
2601 LASSERT(obd != NULL);
2602 LASSERT(inode != NULL);
2604 rc = fsfilt_get_mid(obd, inode, id, sizeof(*id));
2606 CERROR("fsfilt_get_mid() failed, "
2620 * updates master inode id. Usualy this is done by CMOBD after an inode is
2621 * created and relationship between cache MDS and master one should be
2624 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
2625 void *handle, struct lustre_id *id)
2630 LASSERT(id != NULL);
2631 LASSERT(obd != NULL);
2632 LASSERT(inode != NULL);
2634 rc = fsfilt_set_mid(obd, inode, handle, id, sizeof(*id));
2636 CERROR("fsfilt_set_mid() failed, rc = %d\n", rc);
2643 /* mount the file system (secretly) */
2644 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
2646 struct lustre_cfg* lcfg = buf;
2647 struct mds_obd *mds = &obd->u.mds;
2648 char *options = NULL;
2649 struct vfsmount *mnt;
2655 dev_clear_rdonly(2);
2657 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
2658 RETURN(rc = -EINVAL);
2660 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
2661 if (IS_ERR(obd->obd_fsops))
2662 RETURN(rc = PTR_ERR(obd->obd_fsops));
2664 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2666 page = __get_free_page(GFP_KERNEL);
2670 options = (char *)page;
2671 memset(options, 0, PAGE_SIZE);
2674 * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
2675 * and the rest of options are passed by mount options. Probably this
2676 * should be moved to somewhere else like startup scripts or lconf.
2678 sprintf(options, "iopen_nopriv");
2680 if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
2681 sprintf(options + strlen(options), ",%s",
2682 lcfg->lcfg_inlbuf4);
2684 /* we have to know mdsnum before touching underlying fs -bzzz */
2685 sema_init(&mds->mds_md_sem, 1);
2686 mds->mds_md_connected = 0;
2687 mds->mds_md_name = NULL;
2689 if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
2690 strncmp(lcfg->lcfg_inlbuf5, "dumb", lcfg->lcfg_inllen5)) {
2693 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
2694 obd->obd_name, lcfg->lcfg_inlbuf5);
2696 generate_random_uuid(uuid);
2697 class_uuid_unparse(uuid, &mds->mds_md_uuid);
2699 OBD_ALLOC(mds->mds_md_name, lcfg->lcfg_inllen5);
2700 if (mds->mds_md_name == NULL)
2701 RETURN(rc = -ENOMEM);
2703 memcpy(mds->mds_md_name, lcfg->lcfg_inlbuf5,
2704 lcfg->lcfg_inllen5);
2706 rc = mds_md_connect(obd, mds->mds_md_name);
2708 OBD_FREE(mds->mds_md_name, lcfg->lcfg_inllen5);
2713 mds->mds_obd_type = MDS_MASTER_OBD;
2715 if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 &&
2716 strncmp(lcfg->lcfg_inlbuf6, "dumb", lcfg->lcfg_inllen6)) {
2717 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master"))) {
2718 mds->mds_obd_type = MDS_MASTER_OBD;
2719 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache"))) {
2720 mds->mds_obd_type = MDS_CACHE_OBD;
2724 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0, lcfg->lcfg_inlbuf1, options);
2729 CERROR("do_kern_mount failed: rc = %d\n", rc);
2733 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
2735 sema_init(&mds->mds_epoch_sem, 1);
2736 atomic_set(&mds->mds_real_clients, 0);
2737 spin_lock_init(&mds->mds_transno_lock);
2738 spin_lock_init(&mds->mds_last_fid_lock);
2739 sema_init(&mds->mds_orphan_recovery_sem, 1);
2740 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2742 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
2743 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
2745 if (obd->obd_namespace == NULL) {
2746 mds_cleanup(obd, 0);
2747 GOTO(err_put, rc = -ENOMEM);
2749 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
2751 rc = mds_fs_setup(obd, mnt);
2753 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
2758 rc = llog_start_commit_thread();
2762 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2763 strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)) {
2766 generate_random_uuid(uuid);
2767 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
2769 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
2770 if (mds->mds_profile == NULL)
2771 GOTO(err_fs, rc = -ENOMEM);
2773 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
2774 lcfg->lcfg_inllen3);
2778 * setup root dir and files ID dir if lmv already connected, or there is
2781 if (mds->mds_md_exp || (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2782 strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)))
2784 rc = mds_fs_setup_rootid(obd);
2788 rc = mds_fs_setup_virtid(obd);
2793 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2794 "mds_ldlm_client", &obd->obd_ldlm_client);
2795 obd->obd_replayable = 1;
2797 rc = mds_postsetup(obd);
2804 /* No extra cleanup needed for llog_init_commit_thread() */
2805 mds_fs_cleanup(obd, 0);
2807 ldlm_namespace_free(obd->obd_namespace, 0);
2808 obd->obd_namespace = NULL;
2811 mntput(mds->mds_vfsmnt);
2815 fsfilt_put_ops(obd->obd_fsops);
2819 static int mds_postsetup(struct obd_device *obd)
2821 struct mds_obd *mds = &obd->u.mds;
2825 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
2826 obd, 0, NULL, &llog_lvfs_ops);
2830 if (mds->mds_profile) {
2831 struct llog_ctxt *lgctxt;
2832 struct lvfs_run_ctxt saved;
2833 struct lustre_profile *lprof;
2834 struct config_llog_instance cfg;
2836 cfg.cfg_instance = NULL;
2837 cfg.cfg_uuid = mds->mds_dt_uuid;
2838 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2840 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
2842 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2843 GOTO(err_llog, rc = -EINVAL);
2846 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
2847 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2852 lprof = class_get_profile(mds->mds_profile);
2853 if (lprof == NULL) {
2854 CERROR("No profile found: %s\n", mds->mds_profile);
2855 GOTO(err_cleanup, rc = -ENOENT);
2857 rc = mds_dt_connect(obd, lprof->lp_lov);
2859 GOTO(err_cleanup, rc);
2861 rc = mds_md_postsetup(obd);
2863 GOTO(err_cleanup, rc);
2870 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
2871 LLOG_CONFIG_ORIG_CTXT));
2875 int mds_postrecov(struct obd_device *obd)
2877 struct mds_obd *mds = &obd->u.mds;
2878 struct llog_ctxt *ctxt;
2879 int rc, item = 0, valsize;
2883 LASSERT(!obd->obd_recovering);
2884 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2885 LASSERT(ctxt != NULL);
2887 /* set nextid first, so we are sure it happens */
2888 rc = mds_dt_set_nextid(obd);
2890 CERROR("%s: mds_dt_set_nextid() failed\n", obd->obd_name);
2894 /* clean PENDING dir */
2895 rc = mds_cleanup_orphans(obd);
2900 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
2901 valsize = sizeof(group);
2902 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"), "mds_conn",
2907 rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
2910 CERROR("%s: failed at llog_origin_connect: %d\n",
2915 /* remove the orphaned precreated objects */
2916 rc = mds_dt_clearorphans(mds, NULL /* all OSTs */);
2921 RETURN(rc < 0 ? rc : item);
2924 /* cleanup all llogging subsystems */
2925 rc = obd_llog_finish(obd, &obd->obd_llogs,
2926 mds->mds_dt_desc.ld_tgt_count);
2928 CERROR("%s: failed to cleanup llogging subsystems\n",
2933 int mds_dt_clean(struct obd_device *obd)
2935 struct mds_obd *mds = &obd->u.mds;
2938 if (mds->mds_profile) {
2940 struct llog_ctxt *llctx;
2941 struct lvfs_run_ctxt saved;
2942 struct config_llog_instance cfg;
2943 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
2945 OBD_ALLOC(cln_prof, len);
2946 sprintf(cln_prof, "%s-clean", mds->mds_profile);
2948 cfg.cfg_instance = NULL;
2949 cfg.cfg_uuid = mds->mds_dt_uuid;
2951 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2952 llctx = llog_get_context(&obd->obd_llogs,
2953 LLOG_CONFIG_ORIG_CTXT);
2954 class_config_process_llog(llctx, cln_prof, &cfg);
2955 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2957 OBD_FREE(cln_prof, len);
2958 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2959 mds->mds_profile = NULL;
2964 int mds_md_clean(struct obd_device *obd)
2966 struct mds_obd *mds = &obd->u.mds;
2969 if (mds->mds_md_name) {
2970 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
2971 mds->mds_md_name = NULL;
2976 static int mds_precleanup(struct obd_device *obd, int flags)
2982 mds_dt_disconnect(obd, flags);
2984 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2988 static int mds_cleanup(struct obd_device *obd, int flags)
2990 struct mds_obd *mds = &obd->u.mds;
2993 if (mds->mds_sb == NULL)
2996 mds_update_server_data(obd, 1);
2997 mds_update_last_fid(obd, NULL, 1);
2999 if (mds->mds_dt_objids != NULL) {
3000 int size = mds->mds_dt_desc.ld_tgt_count *
3002 OBD_FREE(mds->mds_dt_objids, size);
3004 mds_fs_cleanup(obd, flags);
3009 * 2 seems normal on mds, (may_umount() also expects 2 fwiw), but we
3010 * only see 1 at this point in obdfilter.
3012 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
3013 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
3014 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
3016 mntput(mds->mds_vfsmnt);
3019 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3021 spin_lock_bh(&obd->obd_processing_task_lock);
3022 if (obd->obd_recovering) {
3023 target_cancel_recovery_timer(obd);
3024 obd->obd_recovering = 0;
3026 spin_unlock_bh(&obd->obd_processing_task_lock);
3029 dev_clear_rdonly(2);
3030 fsfilt_put_ops(obd->obd_fsops);
3035 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3037 struct ldlm_lock *new_lock,
3038 struct ldlm_lock **old_lock,
3039 struct lustre_handle *lockh)
3041 struct obd_export *exp = req->rq_export;
3042 struct obd_device *obd = exp->exp_obd;
3043 struct ldlm_request *dlmreq =
3044 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3045 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3046 struct list_head *iter;
3048 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3051 l_lock(&obd->obd_namespace->ns_lock);
3052 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3053 struct ldlm_lock *lock;
3054 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3055 if (lock == new_lock)
3057 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3058 lockh->cookie = lock->l_handle.h_cookie;
3059 LDLM_DEBUG(lock, "restoring lock cookie");
3060 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3063 *old_lock = LDLM_LOCK_GET(lock);
3064 l_unlock(&obd->obd_namespace->ns_lock);
3068 l_unlock(&obd->obd_namespace->ns_lock);
3070 /* If the xid matches, then we know this is a resent request,
3071 * and allow it. (It's probably an OPEN, for which we don't
3073 if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
3076 /* This remote handle isn't enqueued, so we never received or
3077 * processed this request. Clear MSG_RESENT, because it can
3078 * be handled like any normal request now. */
3080 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3082 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3086 int intent_disposition(struct ldlm_reply *rep, int flag)
3090 return (rep->lock_policy_res1 & flag);
3093 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3097 rep->lock_policy_res1 |= flag;
3100 static int mds_intent_policy(struct ldlm_namespace *ns,
3101 struct ldlm_lock **lockp, void *req_cookie,
3102 ldlm_mode_t mode, int flags, void *data)
3104 struct ptlrpc_request *req = req_cookie;
3105 struct ldlm_lock *lock = *lockp;
3106 struct ldlm_intent *it;
3107 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3108 struct ldlm_reply *rep;
3109 struct lustre_handle lockh[2] = {{0}, {0}};
3110 struct ldlm_lock *new_lock = NULL;
3111 int getattr_part = MDS_INODELOCK_UPDATE;
3112 int rc, repsize[4] = { sizeof(struct ldlm_reply),
3113 sizeof(struct mds_body),
3114 mds->mds_max_mdsize,
3115 mds->mds_max_cookiesize };
3116 int offset = MDS_REQ_INTENT_REC_OFF;
3119 LASSERT(req != NULL);
3120 MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
3122 if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3123 /* No intent was provided */
3124 int size = sizeof(struct ldlm_reply);
3125 rc = lustre_pack_reply(req, 1, &size, NULL);
3130 it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3131 lustre_swab_ldlm_intent);
3133 CERROR("Intent missing\n");
3134 RETURN(req->rq_status = -EFAULT);
3137 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3139 rc = lustre_pack_reply(req, 3, repsize, NULL);
3141 RETURN(req->rq_status = rc);
3143 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
3144 LASSERT(rep != NULL);
3146 intent_set_disposition(rep, DISP_IT_EXECD);
3148 /* execute policy */
3149 switch ((long)it->opc) {
3151 case IT_CREAT|IT_OPEN:
3152 /* XXX swab here to assert that an mds_open reint
3153 * packet is following */
3154 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3156 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
3158 /* We abort the lock if the lookup was negative and
3159 * we did not make it to the OPEN portion */
3160 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
3161 RETURN(ELDLM_LOCK_ABORTED);
3162 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
3163 !intent_disposition(rep, DISP_OPEN_OPEN))
3165 /* IT_OPEN may return lock on cross-node dentry
3166 * that we want to hold during attr retrival -bzzz */
3167 if (rc != 0 || lockh[0].cookie == 0)
3168 RETURN(ELDLM_LOCK_ABORTED);
3171 getattr_part = MDS_INODELOCK_LOOKUP;
3174 getattr_part |= MDS_INODELOCK_LOOKUP;
3176 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3177 lock, &new_lock, lockh);
3178 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
3180 /* FIXME: LDLM can set req->rq_status. MDS sets
3181 policy_res{1,2} with disposition and status.
3182 - replay: returns 0 & req->status is old status
3183 - otherwise: returns req->status */
3184 if (intent_disposition(rep, DISP_LOOKUP_NEG))
3185 rep->lock_policy_res2 = 0;
3186 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
3187 rep->lock_policy_res2)
3188 RETURN(ELDLM_LOCK_ABORTED);
3189 if (req->rq_status != 0) {
3191 rep->lock_policy_res2 = req->rq_status;
3192 RETURN(ELDLM_LOCK_ABORTED);
3196 rc = mds_lock_and_check_slave(offset, req, lockh);
3197 if ((rep->lock_policy_res2 = rc)) {
3199 rep->lock_policy_res2 = 0;
3200 RETURN(ELDLM_LOCK_ABORTED);
3204 CERROR("Unhandled intent "LPD64"\n", it->opc);
3208 /* By this point, whatever function we called above must have either
3209 * filled in 'lockh', been an intent replay, or returned an error. We
3210 * want to allow replayed RPCs to not get a lock, since we would just
3211 * drop it below anyways because lock replay is done separately by the
3212 * client afterwards. For regular RPCs we want to give the new lock to
3213 * the client instead of whatever lock it was about to get. */
3214 if (new_lock == NULL)
3215 new_lock = ldlm_handle2lock(&lockh[0]);
3216 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
3219 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
3220 it->opc, lockh[0].cookie);
3222 /* If we've already given this lock to a client once, then we should
3223 * have no readers or writers. Otherwise, we should have one reader
3224 * _or_ writer ref (which will be zeroed below) before returning the
3225 * lock to a client. */
3226 if (new_lock->l_export == req->rq_export) {
3227 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3229 LASSERT(new_lock->l_export == NULL);
3230 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3235 if (new_lock->l_export == req->rq_export) {
3236 /* Already gave this to the client, which means that we
3237 * reconstructed a reply. */
3238 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3240 RETURN(ELDLM_LOCK_REPLACED);
3243 /* Fixup the lock to be given to the client */
3244 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
3245 new_lock->l_readers = 0;
3246 new_lock->l_writers = 0;
3248 new_lock->l_export = class_export_get(req->rq_export);
3249 list_add(&new_lock->l_export_chain,
3250 &new_lock->l_export->exp_ldlm_data.led_held_locks);
3252 new_lock->l_blocking_ast = lock->l_blocking_ast;
3253 new_lock->l_completion_ast = lock->l_completion_ast;
3255 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
3256 sizeof(lock->l_remote_handle));
3258 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3260 LDLM_LOCK_PUT(new_lock);
3261 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
3263 RETURN(ELDLM_LOCK_REPLACED);
3266 int mds_attach(struct obd_device *dev, obd_count len, void *data)
3268 struct lprocfs_static_vars lvars;
3271 lprocfs_init_multi_vars(0, &lvars);
3273 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3277 return lprocfs_alloc_md_stats(dev, 0);
3280 int mds_detach(struct obd_device *dev)
3282 lprocfs_free_md_stats(dev);
3283 return lprocfs_obd_detach(dev);
3286 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
3288 struct lprocfs_static_vars lvars;
3290 lprocfs_init_multi_vars(1, &lvars);
3291 return lprocfs_obd_attach(dev, lvars.obd_vars);
3294 int mdt_detach(struct obd_device *dev)
3296 return lprocfs_obd_detach(dev);
3299 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
3301 struct mds_obd *mds = &obd->u.mds;
3306 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3307 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
3308 mds_handle, "mds", obd->obd_proc_entry);
3310 if (!mds->mds_service) {
3311 CERROR("failed to start service\n");
3315 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
3318 GOTO(err_thread, rc);
3320 mds->mds_setattr_service =
3321 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3322 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
3323 mds_handle, "mds_setattr",
3324 obd->obd_proc_entry);
3325 if (!mds->mds_setattr_service) {
3326 CERROR("failed to start getattr service\n");
3327 GOTO(err_thread, rc = -ENOMEM);
3330 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
3331 MDT_NUM_THREADS, "ll_mdt_attr");
3333 GOTO(err_thread2, rc);
3335 mds->mds_readpage_service =
3336 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3337 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
3338 mds_handle, "mds_readpage",
3339 obd->obd_proc_entry);
3340 if (!mds->mds_readpage_service) {
3341 CERROR("failed to start readpage service\n");
3342 GOTO(err_thread2, rc = -ENOMEM);
3345 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
3346 MDT_NUM_THREADS, "ll_mdt_rdpg");
3349 GOTO(err_thread3, rc);
3354 ptlrpc_unregister_service(mds->mds_readpage_service);
3356 ptlrpc_unregister_service(mds->mds_setattr_service);
3358 ptlrpc_unregister_service(mds->mds_service);
3362 static int mdt_cleanup(struct obd_device *obd, int flags)
3364 struct mds_obd *mds = &obd->u.mds;
3367 ptlrpc_stop_all_threads(mds->mds_readpage_service);
3368 ptlrpc_unregister_service(mds->mds_readpage_service);
3370 ptlrpc_stop_all_threads(mds->mds_setattr_service);
3371 ptlrpc_unregister_service(mds->mds_setattr_service);
3373 ptlrpc_stop_all_threads(mds->mds_service);
3374 ptlrpc_unregister_service(mds->mds_service);
3379 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
3380 __u64 gr, void *data)
3382 struct lustre_id id;
3383 struct obd_device *obd = data;
3387 return mds_id2dentry(obd, &id, NULL);
3390 static int mds_get_info(struct obd_export *exp, __u32 keylen,
3391 void *key, __u32 *valsize, void *val)
3393 struct obd_device *obd;
3394 struct mds_obd *mds;
3397 obd = class_exp2obd(exp);
3401 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
3402 exp->exp_handle.h_cookie);
3406 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
3407 /* get log_context handle. */
3408 unsigned long *llh_handle = val;
3409 *valsize = sizeof(unsigned long);
3410 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
3413 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
3414 /* get log_context handle. */
3415 unsigned long *sb = val;
3416 *valsize = sizeof(unsigned long);
3417 *sb = (unsigned long)obd->u.mds.mds_sb;
3421 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
3422 __u32 *mdsize = val;
3423 *valsize = sizeof(*mdsize);
3424 *mdsize = mds->mds_max_mdsize;
3428 if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
3429 __u32 *mdsnum = val;
3430 *valsize = sizeof(*mdsnum);
3431 *mdsnum = mds->mds_num;
3435 if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
3436 struct lustre_id *rootid = val;
3437 *valsize = sizeof(struct lustre_id);
3438 *rootid = mds->mds_rootid;
3442 CDEBUG(D_IOCTL, "invalid key\n");
3446 struct lvfs_callback_ops mds_lvfs_ops = {
3447 l_id2dentry: mds_lvfs_id2dentry,
3450 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
3451 int objcount, struct obd_ioobj *obj,
3452 int niocount, struct niobuf_remote *nb,
3453 struct niobuf_local *res,
3454 struct obd_trans_info *oti);
3456 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
3457 int objcount, struct obd_ioobj *obj, int niocount,
3458 struct niobuf_local *res, struct obd_trans_info *oti,
3461 /* use obd ops to offer management infrastructure */
3462 static struct obd_ops mds_obd_ops = {
3463 .o_owner = THIS_MODULE,
3464 .o_attach = mds_attach,
3465 .o_detach = mds_detach,
3466 .o_connect = mds_connect,
3467 .o_connect_post = mds_connect_post,
3468 .o_init_export = mds_init_export,
3469 .o_destroy_export = mds_destroy_export,
3470 .o_disconnect = mds_disconnect,
3471 .o_setup = mds_setup,
3472 .o_precleanup = mds_precleanup,
3473 .o_cleanup = mds_cleanup,
3474 .o_postrecov = mds_postrecov,
3475 .o_statfs = mds_obd_statfs,
3476 .o_iocontrol = mds_iocontrol,
3477 .o_create = mds_obd_create,
3478 .o_destroy = mds_obd_destroy,
3479 .o_llog_init = mds_llog_init,
3480 .o_llog_finish = mds_llog_finish,
3481 .o_notify = mds_notify,
3482 .o_get_info = mds_get_info,
3483 .o_set_info = mds_set_info,
3484 .o_preprw = mds_preprw,
3485 .o_commitrw = mds_commitrw,
3488 static struct obd_ops mdt_obd_ops = {
3489 .o_owner = THIS_MODULE,
3490 .o_attach = mdt_attach,
3491 .o_detach = mdt_detach,
3492 .o_setup = mdt_setup,
3493 .o_cleanup = mdt_cleanup,
3496 static int __init mds_init(void)
3498 struct lprocfs_static_vars lvars;
3500 mds_group_hash_init();
3502 lprocfs_init_multi_vars(0, &lvars);
3503 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
3505 lprocfs_init_multi_vars(1, &lvars);
3506 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
3512 static void /*__exit*/ mds_exit(void)
3514 mds_group_hash_cleanup();
3516 class_unregister_type(LUSTRE_MDS_NAME);
3517 class_unregister_type(LUSTRE_MDT_NAME);
3520 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3521 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
3522 MODULE_LICENSE("GPL");
3524 module_init(mds_init);
3525 module_exit(mds_exit);