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 body->valid |= OBD_MD_FID;
588 memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
591 * the last_committed and last_xid fields are filled in for all replies
592 * already - no need to do so here also.
597 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
598 void *data, int flag)
603 if (flag == LDLM_CB_CANCELING) {
604 /* Don't need to do anything here. */
608 /* XXX layering violation! -phil */
609 l_lock(&lock->l_resource->lr_namespace->ns_lock);
612 * get this: if mds_blocking_ast is racing with mds_intent_policy, such
613 * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
614 * then by the time we get the lock, we might not be the correct
615 * blocking function anymore. So check, and return early, if so.
617 if (lock->l_blocking_ast != mds_blocking_ast) {
618 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
622 lock->l_flags |= LDLM_FL_CBPENDING;
623 do_ast = (!lock->l_readers && !lock->l_writers);
624 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
627 struct lustre_handle lockh;
630 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
631 ldlm_lock2handle(lock, &lockh);
632 rc = ldlm_cli_cancel(&lockh);
634 CERROR("ldlm_cli_cancel: %d\n", rc);
636 LDLM_DEBUG(lock, "Lock still has references, will be "
642 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
651 rc = fsfilt_get_md(obd, inode, md, *size);
656 CERROR("Error %d reading eadata for ino %lu\n",
661 if (S_ISREG(inode->i_mode))
662 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
663 if (S_ISDIR(inode->i_mode))
664 rc = mds_convert_mea_ea(obd, inode, md, lmm_size);
678 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
679 * Call with lock=0 if the caller has already taken the i_sem. */
680 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
681 struct mds_body *body, struct inode *inode, int lock)
683 struct mds_obd *mds = &obd->u.mds;
689 lmm = lustre_msg_buf(msg, offset, 0);
691 /* Some problem with getting eadata when I sized the reply
693 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
697 lmm_size = msg->buflens[offset];
699 /* I don't really like this, but it is a sanity check on the client
700 * MD request. However, if the client doesn't know how much space
701 * to reserve for the MD, it shouldn't be bad to have too much space.
703 if (lmm_size > mds->mds_max_mdsize) {
704 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
705 inode->i_ino, lmm_size, mds->mds_max_mdsize);
709 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
711 if (S_ISDIR(inode->i_mode))
712 body->valid |= OBD_MD_FLDIREA;
714 body->valid |= OBD_MD_FLEASIZE;
715 body->eadatasize = lmm_size;
722 void mds_squash_root(struct mds_obd *mds, struct mds_req_sec_desc *rsd,
725 if (!mds->mds_squash_uid ||
726 (rsd->rsd_uid && rsd->rsd_fsuid))
729 if (*peernid == mds->mds_nosquash_nid)
732 CDEBUG(D_OTHER, "squash req from 0x%llx, (%d:%d/%x)=>(%d:%d/%x)\n",
733 *peernid, rsd->rsd_fsuid, rsd->rsd_fsgid, rsd->rsd_cap,
734 mds->mds_squash_uid, mds->mds_squash_gid,
735 (rsd->rsd_cap & ~CAP_FS_MASK));
737 rsd->rsd_uid = mds->mds_squash_uid;
738 rsd->rsd_fsuid = mds->mds_squash_uid;
739 rsd->rsd_fsgid = mds->mds_squash_gid;
741 /* XXX should we remove all capabilities? */
742 rsd->rsd_cap &= ~CAP_FS_MASK;
745 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
746 struct ptlrpc_request *req, struct mds_body *reqbody,
749 struct inode *inode = dentry->d_inode;
750 struct mds_body *body;
754 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
757 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
758 LASSERT(body != NULL); /* caller prepped reply */
760 if (dentry->d_flags & DCACHE_CROSS_REF) {
761 mds_pack_dentry2body(obd, body, dentry,
762 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
763 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
768 mds_pack_inode2body(obd, body, inode,
769 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
771 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
772 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
773 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
776 /* if we have LOV EA data, the OST holds size, atime, mtime. */
777 if (!(body->valid & OBD_MD_FLEASIZE) &&
778 !(body->valid & OBD_MD_FLDIREA))
779 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
780 OBD_MD_FLATIME | OBD_MD_FLMTIME);
781 } else if (S_ISLNK(inode->i_mode) &&
782 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
783 int len = req->rq_repmsg->buflens[reply_off + 1];
784 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1, 0);
786 LASSERT(symname != NULL); /* caller prepped reply */
788 if (!inode->i_op->readlink) {
791 rc = inode->i_op->readlink(dentry, symname, len);
793 CERROR("readlink failed: %d\n", rc);
794 } else if (rc != len - 1) {
795 CERROR("Unexpected readlink rc %d: expecting %d\n",
799 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
800 body->valid |= OBD_MD_LINKNAME;
801 body->eadatasize = rc + 1;
811 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
812 struct dentry *dentry,
815 int rc = 0, size[1] = {sizeof(struct mds_body)};
818 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
819 CERROR("failed MDS_GETATTR_PACK test\n");
820 req->rq_status = -ENOMEM;
824 rc = lustre_pack_reply(req, 1, size, NULL);
826 CERROR("lustre_pack_reply failed: rc %d\n", rc);
827 GOTO(out, req->rq_status = rc);
835 static int mds_getattr_pack_msg(struct ptlrpc_request *req,
839 struct mds_obd *mds = mds_req2mds(req);
840 struct mds_body *body;
841 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
844 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
845 LASSERT(body != NULL); /* checked by caller */
846 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
848 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
849 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
852 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
854 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
858 CERROR("error getting inode %lu MD: rc = %d\n",
861 } else if (rc > mds->mds_max_mdsize) {
863 CERROR("MD size %d larger than maximum possible %u\n",
864 rc, mds->mds_max_mdsize);
869 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
870 if (inode->i_size + 1 != body->eadatasize)
871 CERROR("symlink size: %Lu, reply space: %d\n",
872 inode->i_size + 1, body->eadatasize);
873 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
875 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
876 inode->i_size + 1, body->eadatasize);
879 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
880 CERROR("failed MDS_GETATTR_PACK test\n");
881 req->rq_status = -ENOMEM;
882 GOTO(out, rc = -ENOMEM);
885 rc = lustre_pack_reply(req, bufcount, size, NULL);
887 CERROR("out of memory\n");
888 GOTO(out, req->rq_status = rc);
896 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
897 char *name, int namelen)
899 struct mea *mea = NULL;
900 int mea_size, rc = 0;
903 rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
908 * dir is already splitted, check if requested filename should
909 * live at this MDS or at another one.
911 int i = mea_name2idx(mea, name, namelen - 1);
912 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
914 "inapropriate MDS(%d) for %s. should be "
915 "%lu(%d)\n", mea->mea_master, name,
916 (unsigned long)id_group(&mea->mea_ids[i]), i);
922 OBD_FREE(mea, mea_size);
926 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
927 struct lustre_handle *child_lockh, int child_part)
929 struct obd_device *obd = req->rq_export->exp_obd;
930 struct mds_obd *mds = &obd->u.mds;
931 struct ldlm_reply *rep = NULL;
932 struct lvfs_run_ctxt saved;
933 struct mds_req_sec_desc *rsd;
934 struct mds_body *body;
935 struct dentry *dparent = NULL, *dchild = NULL;
936 struct lvfs_ucred uc;
937 struct lustre_handle parent_lockh[2] = {{0}, {0}};
938 unsigned int namesize;
939 int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
943 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
944 MD_COUNTER_INCREMENT(obd, getattr_lock);
946 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
948 CERROR("Can't unpack security desc\n");
951 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
953 /* swab now, before anyone looks inside the request. */
954 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
955 lustre_swab_mds_body);
957 CERROR("Can't swab mds_body\n");
958 GOTO(cleanup, rc = -EFAULT);
961 LASSERT_REQSWAB(req, offset + 1);
962 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
964 CERROR("Can't unpack name\n");
965 GOTO(cleanup, rc = -EFAULT);
967 namesize = req->rq_reqmsg->buflens[offset + 1];
969 /* namesize less than 2 means we have empty name, probably came from
970 revalidate by cfid, so no point in having name to be set */
974 LASSERT (offset == 1 || offset == 3);
976 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
982 rc = mds_init_ucred(&uc, rsd);
984 CERROR("can't init ucred\n");
988 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
989 cleanup_phase = 1; /* kernel context */
990 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
992 LASSERT(namesize > 0);
993 if (child_lockh->cookie != 0) {
994 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
999 if (body->valid == OBD_MD_FLID) {
1000 struct mds_body *mds_reply;
1001 int size = sizeof(*mds_reply);
1005 dparent = mds_id2dentry(obd, &body->id1, NULL);
1006 if (IS_ERR(dparent)) {
1007 rc = PTR_ERR(dparent);
1012 * the user requested ONLY the inode number, so do a raw lookup.
1014 rc = lustre_pack_reply(req, 1, &size, NULL);
1016 CERROR("out of memory\n");
1020 dir = dparent->d_inode;
1021 LASSERT(dir->i_op->lookup_raw != NULL);
1022 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1024 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1025 sizeof(*mds_reply));
1027 id_ino(&mds_reply->id1) = inum;
1028 mds_reply->valid = OBD_MD_FLID;
1033 if (resent_req == 0) {
1034 LASSERT(id_fid(&body->id1) != 0);
1036 rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1037 parent_lockh, &dparent,
1039 MDS_INODELOCK_UPDATE,
1042 child_lockh, &dchild,
1043 LCK_PR, child_part);
1047 cleanup_phase = 2; /* dchild, dparent, locks */
1050 * let's make sure this name should leave on this mds
1053 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1057 /* we have no dentry here, drop LOOKUP bit */
1058 /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1059 child_part &= ~MDS_INODELOCK_LOOKUP;
1060 CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1061 obd->obd_name, OLID4(&body->id1));
1063 dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1064 LCK_PR, parent_lockh,
1067 MDS_INODELOCK_UPDATE);
1068 if (IS_ERR(dchild)) {
1069 CERROR("can't find inode with id "DLID4", err = %d\n",
1070 OLID4(&body->id1), (int)PTR_ERR(dchild));
1071 GOTO(cleanup, rc = PTR_ERR(dchild));
1073 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1076 struct ldlm_lock *granted_lock;
1078 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1079 granted_lock = ldlm_handle2lock(child_lockh);
1081 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1082 id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1083 child_lockh->cookie);
1085 dparent = mds_id2dentry(obd, &body->id1, NULL);
1088 dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1089 LASSERT(!IS_ERR(dchild));
1090 LDLM_LOCK_PUT(granted_lock);
1093 cleanup_phase = 2; /* dchild, dparent, locks */
1095 if (!DENTRY_VALID(dchild)) {
1096 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1098 * in the intent case, the policy clears this error: the
1099 * disposition is enough.
1104 intent_set_disposition(rep, DISP_LOOKUP_POS);
1107 if (req->rq_repmsg == NULL) {
1108 if (dchild->d_flags & DCACHE_CROSS_REF)
1109 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1111 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
1113 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1118 rc = mds_getattr_internal(obd, dchild, req, body, reply_offset);
1119 GOTO(cleanup, rc); /* returns the lock to the client */
1122 switch (cleanup_phase) {
1124 if (resent_req == 0) {
1125 if (rc && DENTRY_VALID(dchild))
1126 ldlm_lock_decref(child_lockh, LCK_PR);
1128 ldlm_lock_decref(parent_lockh, LCK_PR);
1130 if (parent_lockh[1].cookie != 0)
1131 ldlm_lock_decref(parent_lockh + 1, update_mode);
1138 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1139 mds_exit_ucred(&uc);
1144 static int mds_getattr(struct ptlrpc_request *req, int offset)
1146 struct obd_device *obd = req->rq_export->exp_obd;
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 MD_COUNTER_INCREMENT(obd, getattr);
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 OBD_COUNTER_INCREMENT(obd, statfs);
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;
1269 struct mds_body *rep_body;
1272 de = mds_id2dentry(obd, &body->id1, NULL);
1274 GOTO(out, rc = PTR_ERR(de));
1276 rc = file->f_op->fsync(NULL, de, 1);
1280 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1281 mds_pack_inode2body(obd, rep_body, de->d_inode,
1282 (body->valid & OBD_MD_FID) ? 1 : 0);
1288 req->rq_status = rc;
1292 /* mds_readpage does not take a DLM lock on the inode, because the client must
1293 * already have a PR lock.
1295 * If we were to take another one here, a deadlock will result, if another
1296 * thread is already waiting for a PW lock. */
1297 static int mds_readpage(struct ptlrpc_request *req, int offset)
1299 struct obd_device *obd = req->rq_export->exp_obd;
1300 struct mds_obd *mds = &obd->u.mds;
1301 struct vfsmount *mnt;
1304 struct mds_req_sec_desc *rsd;
1305 struct mds_body *body, *repbody;
1306 struct lvfs_run_ctxt saved;
1307 int rc, size = sizeof(*repbody);
1308 struct lvfs_ucred uc;
1311 rc = lustre_pack_reply(req, 1, &size, NULL);
1312 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1313 CERROR("mds: out of memory\n");
1314 GOTO(out, rc = -ENOMEM);
1317 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1319 CERROR("Can't unpack security desc\n");
1320 GOTO (out, rc = -EFAULT);
1322 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
1324 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1325 lustre_swab_mds_body);
1327 CERROR("Can't unpack body\n");
1328 GOTO (out, rc = -EFAULT);
1331 rc = mds_init_ucred(&uc, rsd);
1333 CERROR("can't init ucred\n");
1337 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1338 de = mds_id2dentry(obd, &body->id1, &mnt);
1340 GOTO(out_pop, rc = PTR_ERR(de));
1342 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1344 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1345 /* note: in case of an error, dentry_open puts dentry */
1347 GOTO(out_pop, rc = PTR_ERR(file));
1349 /* body->size is actually the offset -eeb */
1350 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1351 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1352 body->size, de->d_inode->i_blksize);
1353 GOTO(out_file, rc = -EFAULT);
1356 /* body->nlink is actually the #bytes to read -eeb */
1357 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1358 CERROR("size %u is not multiple of blocksize %lu\n",
1359 body->nlink, de->d_inode->i_blksize);
1360 GOTO(out_file, rc = -EFAULT);
1363 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1364 repbody->size = file->f_dentry->d_inode->i_size;
1365 repbody->valid = OBD_MD_FLSIZE;
1367 /* to make this asynchronous make sure that the handling function
1368 doesn't send a reply when this function completes. Instead a
1369 callback function would send the reply */
1370 /* body->size is actually the offset -eeb */
1371 rc = mds_sendpage(req, file, body->size, body->nlink);
1375 filp_close(file, 0);
1377 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1378 mds_exit_ucred(&uc);
1380 req->rq_status = rc;
1384 /* update master MDS ID, which is stored in local inode EA. */
1385 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1386 void *data, int data_len)
1388 struct mds_obd *mds = &obd->u.mds;
1389 struct dentry *dentry;
1397 dentry = mds_id2dentry(obd, id, NULL);
1399 GOTO(out, rc = PTR_ERR(dentry));
1401 if (!dentry->d_inode) {
1402 CERROR("Can't find object "DLID4".\n",
1404 GOTO(out_dentry, rc = -EINVAL);
1407 handle = fsfilt_start(obd, dentry->d_inode,
1408 FSFILT_OP_SETATTR, NULL);
1410 GOTO(out_dentry, rc = PTR_ERR(handle));
1412 rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1413 (struct lustre_id *)data);
1415 CERROR("Can't update inode "DLID4" master id, "
1416 "error = %d.\n", OLID4(id), rc);
1417 GOTO(out_commit, rc);
1422 fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1429 EXPORT_SYMBOL(mds_update_mid);
1431 /* read master MDS ID, which is stored in local inode EA. */
1432 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1433 void *data, int data_len)
1435 struct dentry *dentry;
1442 dentry = mds_id2dentry(obd, id, NULL);
1444 GOTO(out, rc = PTR_ERR(dentry));
1446 if (!dentry->d_inode) {
1447 CERROR("Can't find object "DLID4".\n",
1449 GOTO(out_dentry, rc = -EINVAL);
1452 down(&dentry->d_inode->i_sem);
1453 rc = mds_read_inode_mid(obd, dentry->d_inode,
1454 (struct lustre_id *)data);
1455 up(&dentry->d_inode->i_sem);
1457 CERROR("Can't read inode "DLID4" master id, "
1458 "error = %d.\n", OLID4(id), rc);
1459 GOTO(out_dentry, rc);
1468 EXPORT_SYMBOL(mds_read_mid);
1470 int mds_reint(struct ptlrpc_request *req, int offset,
1471 struct lustre_handle *lockh)
1473 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1474 struct mds_update_record *rec;
1475 struct mds_req_sec_desc *rsd;
1479 OBD_ALLOC(rec, sizeof(*rec));
1483 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1485 CERROR("Can't unpack security desc\n");
1486 GOTO(out, rc = -EFAULT);
1488 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
1490 rc = mds_update_unpack(req, offset, rec);
1491 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1492 CERROR("invalid record\n");
1493 GOTO(out, req->rq_status = -EINVAL);
1496 rc = mds_init_ucred(&rec->ur_uc, rsd);
1498 CERROR("can't init ucred\n");
1502 /* rc will be used to interrupt a for loop over multiple records */
1503 rc = mds_reint_rec(rec, offset, req, lockh);
1504 mds_exit_ucred(&rec->ur_uc);
1507 OBD_FREE(rec, sizeof(*rec));
1511 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1512 struct obd_device *obd, int *process)
1514 switch (req->rq_reqmsg->opc) {
1515 case MDS_CONNECT: /* This will never get here, but for completeness. */
1516 case OST_CONNECT: /* This will never get here, but for completeness. */
1517 case MDS_DISCONNECT:
1518 case OST_DISCONNECT:
1523 case MDS_SYNC: /* used in unmounting */
1528 *process = target_queue_recovery_request(req, obd);
1532 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1534 /* XXX what should we set rq_status to here? */
1535 req->rq_status = -EAGAIN;
1536 RETURN(ptlrpc_error(req));
1540 static char *reint_names[] = {
1541 [REINT_SETATTR] "setattr",
1542 [REINT_CREATE] "create",
1543 [REINT_LINK] "link",
1544 [REINT_UNLINK] "unlink",
1545 [REINT_RENAME] "rename",
1546 [REINT_OPEN] "open",
1549 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER | \
1550 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
1551 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
1554 static void reconstruct_create(struct ptlrpc_request *req)
1556 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1557 struct mds_client_data *mcd = med->med_mcd;
1558 struct dentry *dentry;
1559 struct ost_body *body;
1560 struct lustre_id id;
1564 /* copy rc, transno and disp; steal locks */
1565 mds_req_from_mcd(req, mcd);
1566 if (req->rq_status) {
1574 id_ino(&id) = mcd->mcd_last_data;
1575 LASSERT(id_ino(&id) != 0);
1577 dentry = mds_id2dentry(req2obd(req), &id, NULL);
1578 if (IS_ERR(dentry)) {
1579 CERROR("can't find inode "LPU64"\n", id_ino(&id));
1580 req->rq_status = PTR_ERR(dentry);
1585 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1586 req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
1587 dentry->d_inode->i_generation);
1589 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1590 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
1591 body->oa.o_id = dentry->d_inode->i_ino;
1592 body->oa.o_generation = dentry->d_inode->i_generation;
1593 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1595 down(&dentry->d_inode->i_sem);
1596 rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
1597 up(&dentry->d_inode->i_sem);
1599 CERROR("Can't read inode self id, inode %lu, "
1600 "rc %d\n", dentry->d_inode->i_ino, rc);
1604 body->oa.o_fid = id_fid(&id);
1605 body->oa.o_mds = id_group(&id);
1611 static int mdt_obj_create(struct ptlrpc_request *req)
1613 struct obd_device *obd = req->rq_export->exp_obd;
1614 struct mds_obd *mds = &obd->u.mds;
1615 struct ost_body *body, *repbody;
1616 char idname[LL_ID_NAMELEN];
1617 int size = sizeof(*repbody);
1618 struct inode *parent_inode;
1619 struct lvfs_run_ctxt saved;
1620 int rc, cleanup_phase = 0;
1621 struct dentry *new = NULL;
1622 struct dentry_params dp;
1623 int mealen, flags = 0;
1624 struct lvfs_ucred uc;
1625 struct lustre_id id;
1627 void *handle = NULL;
1628 unsigned long cr_inum = 0;
1631 DEBUG_REQ(D_HA, req, "create remote object");
1632 parent_inode = mds->mds_unnamed_dir->d_inode;
1634 body = lustre_swab_reqbuf(req, 0, sizeof(*body),
1635 lustre_swab_ost_body);
1639 rc = lustre_pack_reply(req, 1, &size, NULL);
1643 MDS_CHECK_RESENT(req, reconstruct_create(req));
1646 * this only serve to inter-mds request, don't need check group database
1649 uc.luc_ghash = NULL;
1650 uc.luc_ginfo = NULL;
1651 uc.luc_uid = body->oa.o_uid;
1652 uc.luc_fsuid = body->oa.o_uid;
1653 uc.luc_fsgid = body->oa.o_gid;
1655 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1656 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1658 /* in REPLAY case inum should be given (client or other MDS fills it) */
1659 if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
1660 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
1662 * this is re-create request from MDS holding directory name.
1663 * we have to lookup given ino/gen first. if it exists (good
1664 * case) then there is nothing to do. if it does not then we
1665 * have to recreate it.
1667 id_ino(&id) = body->oa.o_id;
1668 id_gen(&id) = body->oa.o_generation;
1670 new = mds_id2dentry(obd, &id, NULL);
1671 if (!IS_ERR(new) && new->d_inode) {
1672 struct lustre_id sid;
1674 CWARN("mkdir() repairing is on its way: %lu/%lu\n",
1675 (unsigned long)id_ino(&id), (unsigned long)id_gen(&id));
1677 obdo_from_inode(&repbody->oa, new->d_inode,
1678 FILTER_VALID_FLAGS);
1680 repbody->oa.o_id = new->d_inode->i_ino;
1681 repbody->oa.o_generation = new->d_inode->i_generation;
1682 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1685 down(&new->d_inode->i_sem);
1686 rc = mds_read_inode_sid(obd, new->d_inode, &sid);
1687 up(&new->d_inode->i_sem);
1689 CERROR("Can't read inode self id "
1690 "inode %lu, rc %d.\n",
1691 new->d_inode->i_ino, rc);
1695 repbody->oa.o_fid = id_fid(&sid);
1696 repbody->oa.o_mds = id_group(&sid);
1697 LASSERT(id_fid(&sid) != 0);
1700 * here we could use fid passed in body->oa.o_fid and
1701 * thus avoid mds_read_inode_sid().
1703 cr_inum = new->d_inode->i_ino;
1704 GOTO(cleanup, rc = 0);
1708 down(&parent_inode->i_sem);
1709 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1710 if (IS_ERR(handle)) {
1711 up(&parent_inode->i_sem);
1712 CERROR("fsfilt_start() failed, rc = %d\n",
1713 (int)PTR_ERR(handle));
1714 GOTO(cleanup, rc = PTR_ERR(handle));
1716 cleanup_phase = 1; /* transaction */
1719 rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
1720 new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
1722 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1723 obd->obd_name, idname, (int) PTR_ERR(new));
1724 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
1725 up(&parent_inode->i_sem);
1726 RETURN(PTR_ERR(new));
1727 } else if (new->d_inode) {
1728 CERROR("%s: name exists. repeat\n", obd->obd_name);
1732 new->d_fsdata = (void *)&dp;
1736 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
1737 (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
1738 LASSERT(body->oa.o_id != 0);
1739 dp.p_inum = body->oa.o_id;
1740 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1741 (unsigned long)body->oa.o_id,
1742 (unsigned long)body->oa.o_generation);
1745 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1747 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
1748 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1749 new->d_inode->i_generation = body->oa.o_generation;
1750 mark_inode_dirty(new->d_inode);
1753 * avoiding asserts in cache flush case, as
1754 * @body->oa.o_id should be zero.
1756 if (body->oa.o_id) {
1757 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
1758 "BUG 3550: failed to recreate obj "
1759 LPU64" -> %lu\n", body->oa.o_id,
1760 new->d_inode->i_ino);
1762 LASSERTF(body->oa.o_generation ==
1763 new->d_inode->i_generation,
1764 "BUG 3550: failed to recreate obj/gen "
1765 LPU64"/%u -> %lu/%u\n", body->oa.o_id,
1766 body->oa.o_generation,
1767 new->d_inode->i_ino,
1768 new->d_inode->i_generation);
1772 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1773 repbody->oa.o_id = new->d_inode->i_ino;
1774 repbody->oa.o_generation = new->d_inode->i_generation;
1775 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
1777 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
1778 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1779 id_group(&id) = mds->mds_num;
1781 LASSERT(body->oa.o_fid != 0);
1782 id_fid(&id) = body->oa.o_fid;
1784 LASSERT(body->oa.o_id != 0);
1785 id_ino(&id) = repbody->oa.o_id;
1786 id_gen(&id) = repbody->oa.o_generation;
1788 down(&new->d_inode->i_sem);
1789 rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
1790 up(&new->d_inode->i_sem);
1793 * make sure, that fid is up-to-date.
1795 mds_set_last_fid(obd, id_fid(&id));
1798 * allocate new sid, as object is created from scratch
1799 * and this is not replay.
1801 down(&new->d_inode->i_sem);
1802 rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
1803 up(&new->d_inode->i_sem);
1806 CERROR("Can't update lustre ID for inode %lu, "
1807 "error = %d\n", new->d_inode->i_ino, rc);
1811 /* initializing o_fid after it is allocated. */
1812 repbody->oa.o_fid = id_fid(&id);
1813 repbody->oa.o_mds = id_group(&id);
1815 rc = fsfilt_del_dir_entry(obd, new);
1816 up(&parent_inode->i_sem);
1818 CERROR("can't remove name for object: %d\n", rc);
1822 cleanup_phase = 2; /* created directory object */
1824 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
1825 (unsigned long)new->d_inode->i_ino,
1826 (unsigned long)new->d_inode->i_generation,
1827 (unsigned)new->d_inode->i_mode);
1828 cr_inum = new->d_inode->i_ino;
1830 up(&parent_inode->i_sem);
1831 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1835 if (body->oa.o_valid & OBD_MD_FLID) {
1836 /* this is new object for splitted dir. We have to prevent
1837 * recursive splitting on it -bzzz */
1838 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
1840 OBD_ALLOC(mea, mealen);
1842 GOTO(cleanup, rc = -ENOMEM);
1844 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
1845 mea->mea_master = 0;
1848 down(&new->d_inode->i_sem);
1849 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1850 up(&new->d_inode->i_sem);
1852 CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
1854 OBD_FREE(mea, mealen);
1855 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
1856 obd->obd_name, new->d_inode->i_ino,
1857 new->d_inode->i_generation, flags);
1858 } else if (body->oa.o_easize) {
1859 /* we pass LCK_EX to split routine to signal that we have
1860 * exclusive access to the directory. simple because nobody
1861 * knows it already exists -bzzz */
1862 rc = mds_try_to_split_dir(obd, new, NULL,
1863 body->oa.o_easize, LCK_EX);
1865 CERROR("Can't split directory %lu, error = %d.\n",
1866 new->d_inode->i_ino, rc);
1874 switch (cleanup_phase) {
1875 case 2: /* object has been created, but we'll may want to replay it later */
1877 ptlrpc_require_repack(req);
1878 case 1: /* transaction */
1879 rc = mds_finish_transno(mds, parent_inode, handle,
1884 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1885 mds_put_group_entry(mds, uc.luc_ghash);
1889 static int mdt_get_info(struct ptlrpc_request *req)
1891 struct obd_export *exp = req->rq_export;
1896 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1898 DEBUG_REQ(D_HA, req, "no get_info key");
1901 keylen = req->rq_reqmsg->buflens[0];
1903 if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
1904 (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
1905 (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
1908 if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
1909 struct lustre_id *reply;
1910 int size = sizeof(*reply);
1912 rc = lustre_pack_reply(req, 1, &size, NULL);
1916 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
1917 rc = obd_get_info(exp, keylen, key, &size, reply);
1920 int size = sizeof(*reply);
1922 rc = lustre_pack_reply(req, 1, &size, NULL);
1926 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
1927 rc = obd_get_info(exp, keylen, key, &size, reply);
1930 req->rq_repmsg->status = 0;
1934 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1935 void *key, __u32 vallen, void *val)
1937 struct obd_device *obd;
1938 struct mds_obd *mds;
1942 obd = class_exp2obd(exp);
1944 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1945 exp->exp_handle.h_cookie);
1950 if (keylen >= strlen("mds_type") &&
1951 memcmp(key, "mds_type", keylen) == 0) {
1955 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
1957 mds->mds_obd_type = *(int*)val;
1958 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
1959 valsize = sizeof(group);
1961 /* mds number has been changed, so the corresponding obdfilter
1962 * exp need to be changed too. */
1963 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
1964 "mds_conn", valsize, &group);
1967 CDEBUG(D_IOCTL, "invalid key\n");
1971 static int mdt_set_info(struct ptlrpc_request *req)
1974 struct obd_export *exp = req->rq_export;
1975 int keylen, rc = 0, vallen;
1978 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1980 DEBUG_REQ(D_HA, req, "no set_info key");
1983 keylen = req->rq_reqmsg->buflens[0];
1985 if (keylen == strlen("mds_type") &&
1986 memcmp(key, "mds_type", keylen) == 0) {
1987 rc = lustre_pack_reply(req, 0, NULL, NULL);
1991 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
1992 vallen = req->rq_reqmsg->buflens[1];
1994 rc = obd_set_info(exp, keylen, key, vallen, val);
1995 req->rq_repmsg->status = 0;
1998 CDEBUG(D_IOCTL, "invalid key\n");
2002 static int mds_msg_check_version(struct lustre_msg *msg)
2008 case MDS_DISCONNECT:
2010 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2012 CERROR("bad opc %u version %08x, expecting %08x\n",
2013 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2018 case MDS_GETATTR_LOCK:
2022 case MDS_DONE_WRITING:
2025 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2027 CERROR("bad opc %u version %08x, expecting %08x\n",
2028 msg->opc, msg->version, LUSTRE_MDS_VERSION);
2032 case LDLM_BL_CALLBACK:
2033 case LDLM_CP_CALLBACK:
2034 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2036 CERROR("bad opc %u version %08x, expecting %08x\n",
2037 msg->opc, msg->version, LUSTRE_DLM_VERSION);
2039 case OBD_LOG_CANCEL:
2040 case LLOG_ORIGIN_HANDLE_OPEN:
2041 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2042 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2043 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2044 case LLOG_ORIGIN_HANDLE_CLOSE:
2046 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2048 CERROR("bad opc %u version %08x, expecting %08x\n",
2049 msg->opc, msg->version, LUSTRE_LOG_VERSION);
2055 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2057 CERROR("bad opc %u version %08x, expecting %08x\n",
2058 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2061 CERROR("MDS unknown opcode %d\n", msg->opc);
2069 static char str[PTL_NALFMT_SIZE];
2071 int mds_handle(struct ptlrpc_request *req)
2073 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2074 struct obd_device *obd = NULL;
2075 struct mds_obd *mds = NULL; /* quell gcc overwarning */
2079 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2081 rc = mds_msg_check_version(req->rq_reqmsg);
2083 CERROR("MDS drop mal-formed request\n");
2087 LASSERT(current->journal_info == NULL);
2088 /* XXX identical to OST */
2089 if (req->rq_reqmsg->opc != MDS_CONNECT) {
2090 struct mds_export_data *med;
2093 if (req->rq_export == NULL) {
2094 CERROR("operation %d on unconnected MDS from NID %s\n",
2095 req->rq_reqmsg->opc,
2096 ptlrpc_peernid2str(&req->rq_peer, str));
2097 req->rq_status = -ENOTCONN;
2098 GOTO(out, rc = -ENOTCONN);
2101 med = &req->rq_export->exp_mds_data;
2102 obd = req->rq_export->exp_obd;
2105 /* sanity check: if the xid matches, the request must
2106 * be marked as a resent or replayed */
2107 if (req->rq_xid == med->med_mcd->mcd_last_xid) {
2108 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2109 (MSG_RESENT | MSG_REPLAY),
2110 "rq_xid "LPU64" matches last_xid, "
2111 "expected RESENT flag\n",
2114 /* else: note the opposite is not always true; a
2115 * RESENT req after a failover will usually not match
2116 * the last_xid, since it was likely never
2117 * committed. A REPLAYed request will almost never
2118 * match the last xid, however it could for a
2119 * committed, but still retained, open. */
2121 spin_lock_bh(&obd->obd_processing_task_lock);
2122 recovering = obd->obd_recovering;
2123 spin_unlock_bh(&obd->obd_processing_task_lock);
2125 rc = mds_filter_recovery_request(req, obd,
2127 if (rc || should_process == 0) {
2129 } else if (should_process < 0) {
2130 req->rq_status = should_process;
2131 rc = ptlrpc_error(req);
2137 switch (req->rq_reqmsg->opc) {
2139 DEBUG_REQ(D_INODE, req, "connect");
2140 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2141 rc = target_handle_connect(req);
2143 /* Now that we have an export, set mds. */
2144 mds = mds_req2mds(req);
2147 case MDS_DISCONNECT:
2148 DEBUG_REQ(D_INODE, req, "disconnect");
2149 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2150 rc = target_handle_disconnect(req);
2151 req->rq_status = rc; /* superfluous? */
2155 DEBUG_REQ(D_INODE, req, "getstatus");
2156 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2157 rc = mds_getstatus(req);
2161 DEBUG_REQ(D_INODE, req, "getattr");
2162 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2163 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2166 case MDS_GETATTR_LOCK: {
2167 struct lustre_handle lockh;
2168 DEBUG_REQ(D_INODE, req, "getattr_lock");
2169 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2171 /* If this request gets a reconstructed reply, we won't be
2172 * acquiring any new locks in mds_getattr_lock, so we don't
2176 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2177 MDS_INODELOCK_UPDATE);
2178 /* this non-intent call (from an ioctl) is special */
2179 req->rq_status = rc;
2180 if (rc == 0 && lockh.cookie)
2181 ldlm_lock_decref(&lockh, LCK_PR);
2185 DEBUG_REQ(D_INODE, req, "statfs");
2186 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2187 rc = mds_statfs(req);
2191 DEBUG_REQ(D_INODE, req, "readpage");
2192 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2193 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2195 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2196 if (req->rq_reply_state) {
2197 lustre_free_reply_state (req->rq_reply_state);
2198 req->rq_reply_state = NULL;
2205 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2208 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2209 mds->mds_max_cookiesize};
2212 /* NB only peek inside req now; mds_reint() will swab it */
2214 CERROR ("Can't inspect opcode\n");
2219 if (lustre_msg_swabbed (req->rq_reqmsg))
2222 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2223 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2224 reint_names[opc] == NULL) ? reint_names[opc] :
2227 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2229 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2231 else if (opc == REINT_OPEN)
2236 rc = lustre_pack_reply(req, bufcount, size, NULL);
2240 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
2241 fail = OBD_FAIL_MDS_REINT_NET_REP;
2246 DEBUG_REQ(D_INODE, req, "close");
2247 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
2248 rc = mds_close(req, MDS_REQ_REC_OFF);
2251 case MDS_DONE_WRITING:
2252 DEBUG_REQ(D_INODE, req, "done_writing");
2253 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
2254 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2258 DEBUG_REQ(D_INODE, req, "pin");
2259 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2260 rc = mds_pin(req, MDS_REQ_REC_OFF);
2264 DEBUG_REQ(D_INODE, req, "sync");
2265 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2266 rc = mds_sync(req, MDS_REQ_REC_OFF);
2270 DEBUG_REQ(D_INODE, req, "ping");
2271 rc = target_handle_ping(req);
2274 case OBD_LOG_CANCEL:
2275 CDEBUG(D_INODE, "log cancel\n");
2276 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2277 rc = -ENOTSUPP; /* la la la */
2281 DEBUG_REQ(D_INODE, req, "enqueue");
2282 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2283 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2284 ldlm_server_blocking_ast, NULL);
2285 fail = OBD_FAIL_LDLM_REPLY;
2288 DEBUG_REQ(D_INODE, req, "convert");
2289 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2290 rc = ldlm_handle_convert(req);
2292 case LDLM_BL_CALLBACK:
2293 case LDLM_CP_CALLBACK:
2294 DEBUG_REQ(D_INODE, req, "callback");
2295 CERROR("callbacks should not happen on MDS\n");
2297 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2299 case LLOG_ORIGIN_HANDLE_OPEN:
2300 DEBUG_REQ(D_INODE, req, "llog_init");
2301 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2302 rc = llog_origin_handle_open(req);
2304 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2305 DEBUG_REQ(D_INODE, req, "llog next block");
2306 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2307 rc = llog_origin_handle_next_block(req);
2309 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2310 DEBUG_REQ(D_INODE, req, "llog prev block");
2311 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2312 rc = llog_origin_handle_prev_block(req);
2314 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2315 DEBUG_REQ(D_INODE, req, "llog read header");
2316 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2317 rc = llog_origin_handle_read_header(req);
2319 case LLOG_ORIGIN_HANDLE_CLOSE:
2320 DEBUG_REQ(D_INODE, req, "llog close");
2321 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2322 rc = llog_origin_handle_close(req);
2325 DEBUG_REQ(D_INODE, req, "ost_create");
2326 rc = mdt_obj_create(req);
2329 DEBUG_REQ(D_INODE, req, "get_info");
2330 rc = mdt_get_info(req);
2333 DEBUG_REQ(D_INODE, req, "set_info");
2334 rc = mdt_set_info(req);
2337 CDEBUG(D_INODE, "write\n");
2338 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2339 rc = ost_brw_write(req, NULL);
2340 LASSERT(current->journal_info == NULL);
2341 /* mdt_brw sends its own replies */
2345 DEBUG_REQ(D_INODE, req, "llog catinfo");
2346 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2347 rc = llog_catinfo(req);
2350 req->rq_status = -ENOTSUPP;
2351 rc = ptlrpc_error(req);
2355 LASSERT(current->journal_info == NULL);
2359 /* If we're DISCONNECTing, the mds_export_data is already freed */
2360 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
2361 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2362 struct obd_device *obd = list_entry(mds, struct obd_device,
2364 req->rq_repmsg->last_xid =
2365 le64_to_cpu(med->med_mcd->mcd_last_xid);
2367 if (!obd->obd_no_transno) {
2368 req->rq_repmsg->last_committed =
2369 obd->obd_last_committed;
2371 DEBUG_REQ(D_IOCTL, req,
2372 "not sending last_committed update");
2374 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2376 mds->mds_last_transno, obd->obd_last_committed,
2381 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
2382 if (obd && obd->obd_recovering) {
2383 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
2384 return target_queue_final_reply(req, rc);
2386 /* Lost a race with recovery; let the error path DTRT. */
2387 rc = req->rq_status = -ENOTCONN;
2390 target_send_reply(req, rc, fail);
2394 /* Update the server data on disk. This stores the new mount_count and also the
2395 * last_rcvd value to disk. If we don't have a clean shutdown, then the server
2396 * last_rcvd value may be less than that of the clients. This will alert us
2397 * that we may need to do client recovery.
2399 * Also assumes for mds_last_transno that we are not modifying it (no locking).
2401 int mds_update_server_data(struct obd_device *obd, int force_sync)
2403 struct mds_obd *mds = &obd->u.mds;
2404 struct mds_server_data *msd = mds->mds_server_data;
2405 struct file *filp = mds->mds_rcvd_filp;
2406 struct lvfs_run_ctxt saved;
2411 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2412 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2414 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2415 mds->mds_mount_count, mds->mds_last_transno);
2416 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
2418 CERROR("error writing MDS server data: rc = %d\n", rc);
2419 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2424 /* saves last allocated fid counter to file. */
2425 int mds_update_last_fid(struct obd_device *obd, void *handle,
2428 struct mds_obd *mds = &obd->u.mds;
2429 struct file *filp = mds->mds_fid_filp;
2430 struct lvfs_run_ctxt saved;
2436 spin_lock(&mds->mds_last_fid_lock);
2437 last_fid = mds->mds_last_fid;
2438 spin_unlock(&mds->mds_last_fid_lock);
2440 CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
2444 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
2445 handle, mds_commit_last_fid_cb,
2449 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2450 rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
2452 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2455 CERROR("error writing MDS last_fid #"LPU64
2456 ", err = %d\n", last_fid, rc);
2460 CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
2461 "%llu: err = %d\n", last_fid, off, rc);
2466 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
2468 struct mds_obd *mds = &obd->u.mds;
2470 spin_lock(&mds->mds_last_fid_lock);
2471 if (fid > mds->mds_last_fid)
2472 mds->mds_last_fid = fid;
2473 spin_unlock(&mds->mds_last_fid_lock);
2476 void mds_commit_last_transno_cb(struct obd_device *obd,
2477 __u64 transno, void *data,
2480 obd_transno_commit_cb(obd, transno, error);
2483 void mds_commit_last_fid_cb(struct obd_device *obd,
2484 __u64 fid, void *data,
2488 CERROR("%s: fid "LPD64" commit error: %d\n",
2489 obd->obd_name, fid, error);
2493 CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
2494 obd->obd_name, fid);
2497 __u64 mds_alloc_fid(struct obd_device *obd)
2499 struct mds_obd *mds = &obd->u.mds;
2502 spin_lock(&mds->mds_last_fid_lock);
2503 fid = ++mds->mds_last_fid;
2504 spin_unlock(&mds->mds_last_fid_lock);
2510 * allocates new lustre_id on passed @inode and saves it to inode EA.
2512 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
2513 void *handle, struct lustre_id *id)
2515 struct mds_obd *mds = &obd->u.mds;
2516 int alloc = 0, rc = 0;
2519 LASSERT(obd != NULL);
2520 LASSERT(inode != NULL);
2523 OBD_ALLOC(id, sizeof(*id));
2529 id_group(id) = mds->mds_num;
2530 id_fid(id) = mds_alloc_fid(obd);
2531 id_ino(id) = inode->i_ino;
2532 id_gen(id) = inode->i_generation;
2533 id_type(id) = (S_IFMT & inode->i_mode);
2535 rc = mds_update_inode_sid(obd, inode, handle, id);
2537 CERROR("Can't update inode FID EA, "
2542 OBD_FREE(id, sizeof(*id));
2547 * reads inode self id from inode EA. Probably later this should be replaced by
2548 * caching inode self id to avoid raeding it every time it is needed.
2550 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
2551 struct lustre_id *id)
2556 LASSERT(id != NULL);
2557 LASSERT(obd != NULL);
2558 LASSERT(inode != NULL);
2560 rc = fsfilt_get_sid(obd, inode, &id->li_fid,
2561 sizeof(id->li_fid));
2563 CERROR("fsfilt_get_sid() failed, "
2576 /* updates inode self id in EA. */
2577 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
2578 void *handle, struct lustre_id *id)
2583 LASSERT(id != NULL);
2584 LASSERT(obd != NULL);
2585 LASSERT(inode != NULL);
2587 rc = fsfilt_set_sid(obd, inode, handle, &id->li_fid,
2588 sizeof(id->li_fid));
2590 CERROR("fsfilt_set_sid() failed, rc = %d\n", rc);
2598 * reads inode id on master MDS. This is usualy done by CMOBD to update requests
2599 * to master MDS by correct store cookie, needed to find inode on master MDS
2602 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
2603 struct lustre_id *id)
2608 LASSERT(id != NULL);
2609 LASSERT(obd != NULL);
2610 LASSERT(inode != NULL);
2612 rc = fsfilt_get_mid(obd, inode, id, sizeof(*id));
2614 CERROR("fsfilt_get_mid() failed, "
2628 * updates master inode id. Usualy this is done by CMOBD after an inode is
2629 * created and relationship between cache MDS and master one should be
2632 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
2633 void *handle, struct lustre_id *id)
2638 LASSERT(id != NULL);
2639 LASSERT(obd != NULL);
2640 LASSERT(inode != NULL);
2642 rc = fsfilt_set_mid(obd, inode, handle, id, sizeof(*id));
2644 CERROR("fsfilt_set_mid() failed, rc = %d\n", rc);
2651 /* mount the file system (secretly) */
2652 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
2654 struct lustre_cfg* lcfg = buf;
2655 struct mds_obd *mds = &obd->u.mds;
2656 char *options = NULL;
2657 struct vfsmount *mnt;
2663 dev_clear_rdonly(2);
2665 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
2666 RETURN(rc = -EINVAL);
2668 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
2669 if (IS_ERR(obd->obd_fsops))
2670 RETURN(rc = PTR_ERR(obd->obd_fsops));
2672 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2674 page = __get_free_page(GFP_KERNEL);
2678 options = (char *)page;
2679 memset(options, 0, PAGE_SIZE);
2682 * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
2683 * and the rest of options are passed by mount options. Probably this
2684 * should be moved to somewhere else like startup scripts or lconf.
2686 sprintf(options, "iopen_nopriv");
2688 if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
2689 sprintf(options + strlen(options), ",%s",
2690 lcfg->lcfg_inlbuf4);
2692 /* we have to know mdsnum before touching underlying fs -bzzz */
2693 sema_init(&mds->mds_md_sem, 1);
2694 mds->mds_md_connected = 0;
2695 mds->mds_md_name = NULL;
2697 if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
2698 strncmp(lcfg->lcfg_inlbuf5, "dumb", lcfg->lcfg_inllen5)) {
2701 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
2702 obd->obd_name, lcfg->lcfg_inlbuf5);
2704 generate_random_uuid(uuid);
2705 class_uuid_unparse(uuid, &mds->mds_md_uuid);
2707 OBD_ALLOC(mds->mds_md_name, lcfg->lcfg_inllen5);
2708 if (mds->mds_md_name == NULL)
2709 RETURN(rc = -ENOMEM);
2711 memcpy(mds->mds_md_name, lcfg->lcfg_inlbuf5,
2712 lcfg->lcfg_inllen5);
2714 rc = mds_md_connect(obd, mds->mds_md_name);
2716 OBD_FREE(mds->mds_md_name, lcfg->lcfg_inllen5);
2721 mds->mds_obd_type = MDS_MASTER_OBD;
2723 if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 &&
2724 strncmp(lcfg->lcfg_inlbuf6, "dumb", lcfg->lcfg_inllen6)) {
2725 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master"))) {
2726 mds->mds_obd_type = MDS_MASTER_OBD;
2727 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache"))) {
2728 mds->mds_obd_type = MDS_CACHE_OBD;
2732 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0, lcfg->lcfg_inlbuf1, options);
2737 CERROR("do_kern_mount failed: rc = %d\n", rc);
2741 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
2743 sema_init(&mds->mds_epoch_sem, 1);
2744 atomic_set(&mds->mds_real_clients, 0);
2745 spin_lock_init(&mds->mds_transno_lock);
2746 spin_lock_init(&mds->mds_last_fid_lock);
2747 sema_init(&mds->mds_orphan_recovery_sem, 1);
2748 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2750 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
2751 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
2753 if (obd->obd_namespace == NULL) {
2754 mds_cleanup(obd, 0);
2755 GOTO(err_put, rc = -ENOMEM);
2757 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
2759 rc = mds_fs_setup(obd, mnt);
2761 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
2766 rc = llog_start_commit_thread();
2770 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2771 strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)) {
2774 generate_random_uuid(uuid);
2775 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
2777 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
2778 if (mds->mds_profile == NULL)
2779 GOTO(err_fs, rc = -ENOMEM);
2781 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
2782 lcfg->lcfg_inllen3);
2786 * setup root dir and files ID dir if lmv already connected, or there is
2789 if (mds->mds_md_exp || (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2790 strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)))
2792 rc = mds_fs_setup_rootid(obd);
2796 rc = mds_fs_setup_virtid(obd);
2801 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2802 "mds_ldlm_client", &obd->obd_ldlm_client);
2803 obd->obd_replayable = 1;
2805 rc = mds_postsetup(obd);
2812 /* No extra cleanup needed for llog_init_commit_thread() */
2813 mds_fs_cleanup(obd, 0);
2815 ldlm_namespace_free(obd->obd_namespace, 0);
2816 obd->obd_namespace = NULL;
2819 mntput(mds->mds_vfsmnt);
2823 fsfilt_put_ops(obd->obd_fsops);
2827 static int mds_postsetup(struct obd_device *obd)
2829 struct mds_obd *mds = &obd->u.mds;
2833 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
2834 obd, 0, NULL, &llog_lvfs_ops);
2838 if (mds->mds_profile) {
2839 struct llog_ctxt *lgctxt;
2840 struct lvfs_run_ctxt saved;
2841 struct lustre_profile *lprof;
2842 struct config_llog_instance cfg;
2844 cfg.cfg_instance = NULL;
2845 cfg.cfg_uuid = mds->mds_dt_uuid;
2846 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2848 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
2850 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2851 GOTO(err_llog, rc = -EINVAL);
2854 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
2855 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2860 lprof = class_get_profile(mds->mds_profile);
2861 if (lprof == NULL) {
2862 CERROR("No profile found: %s\n", mds->mds_profile);
2863 GOTO(err_cleanup, rc = -ENOENT);
2865 rc = mds_dt_connect(obd, lprof->lp_lov);
2867 GOTO(err_cleanup, rc);
2869 rc = mds_md_postsetup(obd);
2871 GOTO(err_cleanup, rc);
2878 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
2879 LLOG_CONFIG_ORIG_CTXT));
2883 int mds_postrecov(struct obd_device *obd)
2885 struct mds_obd *mds = &obd->u.mds;
2886 struct llog_ctxt *ctxt;
2887 int rc, item = 0, valsize;
2891 LASSERT(!obd->obd_recovering);
2892 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2893 LASSERT(ctxt != NULL);
2895 /* set nextid first, so we are sure it happens */
2896 rc = mds_dt_set_nextid(obd);
2898 CERROR("%s: mds_dt_set_nextid() failed\n", obd->obd_name);
2902 /* clean PENDING dir */
2903 rc = mds_cleanup_orphans(obd);
2908 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
2909 valsize = sizeof(group);
2910 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"), "mds_conn",
2915 rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
2918 CERROR("%s: failed at llog_origin_connect: %d\n",
2923 /* remove the orphaned precreated objects */
2924 rc = mds_dt_clearorphans(mds, NULL /* all OSTs */);
2929 RETURN(rc < 0 ? rc : item);
2932 /* cleanup all llogging subsystems */
2933 rc = obd_llog_finish(obd, &obd->obd_llogs,
2934 mds->mds_dt_desc.ld_tgt_count);
2936 CERROR("%s: failed to cleanup llogging subsystems\n",
2941 int mds_dt_clean(struct obd_device *obd)
2943 struct mds_obd *mds = &obd->u.mds;
2946 if (mds->mds_profile) {
2948 struct llog_ctxt *llctx;
2949 struct lvfs_run_ctxt saved;
2950 struct config_llog_instance cfg;
2951 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
2953 OBD_ALLOC(cln_prof, len);
2954 sprintf(cln_prof, "%s-clean", mds->mds_profile);
2956 cfg.cfg_instance = NULL;
2957 cfg.cfg_uuid = mds->mds_dt_uuid;
2959 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2960 llctx = llog_get_context(&obd->obd_llogs,
2961 LLOG_CONFIG_ORIG_CTXT);
2962 class_config_process_llog(llctx, cln_prof, &cfg);
2963 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2965 OBD_FREE(cln_prof, len);
2966 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2967 mds->mds_profile = NULL;
2972 int mds_md_clean(struct obd_device *obd)
2974 struct mds_obd *mds = &obd->u.mds;
2977 if (mds->mds_md_name) {
2978 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
2979 mds->mds_md_name = NULL;
2984 static int mds_precleanup(struct obd_device *obd, int flags)
2990 mds_dt_disconnect(obd, flags);
2992 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2996 static int mds_cleanup(struct obd_device *obd, int flags)
2998 struct mds_obd *mds = &obd->u.mds;
3001 if (mds->mds_sb == NULL)
3004 mds_update_server_data(obd, 1);
3005 mds_update_last_fid(obd, NULL, 1);
3007 if (mds->mds_dt_objids != NULL) {
3008 int size = mds->mds_dt_desc.ld_tgt_count *
3010 OBD_FREE(mds->mds_dt_objids, size);
3012 mds_fs_cleanup(obd, flags);
3017 * 2 seems normal on mds, (may_umount() also expects 2 fwiw), but we
3018 * only see 1 at this point in obdfilter.
3020 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
3021 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
3022 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
3024 mntput(mds->mds_vfsmnt);
3027 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3029 spin_lock_bh(&obd->obd_processing_task_lock);
3030 if (obd->obd_recovering) {
3031 target_cancel_recovery_timer(obd);
3032 obd->obd_recovering = 0;
3034 spin_unlock_bh(&obd->obd_processing_task_lock);
3037 dev_clear_rdonly(2);
3038 fsfilt_put_ops(obd->obd_fsops);
3043 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3045 struct ldlm_lock *new_lock,
3046 struct ldlm_lock **old_lock,
3047 struct lustre_handle *lockh)
3049 struct obd_export *exp = req->rq_export;
3050 struct obd_device *obd = exp->exp_obd;
3051 struct ldlm_request *dlmreq =
3052 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3053 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3054 struct list_head *iter;
3056 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3059 l_lock(&obd->obd_namespace->ns_lock);
3060 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3061 struct ldlm_lock *lock;
3062 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3063 if (lock == new_lock)
3065 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3066 lockh->cookie = lock->l_handle.h_cookie;
3067 LDLM_DEBUG(lock, "restoring lock cookie");
3068 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3071 *old_lock = LDLM_LOCK_GET(lock);
3072 l_unlock(&obd->obd_namespace->ns_lock);
3076 l_unlock(&obd->obd_namespace->ns_lock);
3078 /* If the xid matches, then we know this is a resent request,
3079 * and allow it. (It's probably an OPEN, for which we don't
3081 if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
3084 /* This remote handle isn't enqueued, so we never received or
3085 * processed this request. Clear MSG_RESENT, because it can
3086 * be handled like any normal request now. */
3088 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3090 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3094 int intent_disposition(struct ldlm_reply *rep, int flag)
3098 return (rep->lock_policy_res1 & flag);
3101 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3105 rep->lock_policy_res1 |= flag;
3108 static int mds_intent_policy(struct ldlm_namespace *ns,
3109 struct ldlm_lock **lockp, void *req_cookie,
3110 ldlm_mode_t mode, int flags, void *data)
3112 struct ptlrpc_request *req = req_cookie;
3113 struct ldlm_lock *lock = *lockp;
3114 struct ldlm_intent *it;
3115 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3116 struct ldlm_reply *rep;
3117 struct lustre_handle lockh[2] = {{0}, {0}};
3118 struct ldlm_lock *new_lock = NULL;
3119 int getattr_part = MDS_INODELOCK_UPDATE;
3120 int rc, repsize[4] = { sizeof(struct ldlm_reply),
3121 sizeof(struct mds_body),
3122 mds->mds_max_mdsize,
3123 mds->mds_max_cookiesize };
3124 int offset = MDS_REQ_INTENT_REC_OFF;
3127 LASSERT(req != NULL);
3128 MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
3130 if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3131 /* No intent was provided */
3132 int size = sizeof(struct ldlm_reply);
3133 rc = lustre_pack_reply(req, 1, &size, NULL);
3138 it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3139 lustre_swab_ldlm_intent);
3141 CERROR("Intent missing\n");
3142 RETURN(req->rq_status = -EFAULT);
3145 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3147 rc = lustre_pack_reply(req, 3, repsize, NULL);
3149 RETURN(req->rq_status = rc);
3151 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
3152 LASSERT(rep != NULL);
3154 intent_set_disposition(rep, DISP_IT_EXECD);
3156 /* execute policy */
3157 switch ((long)it->opc) {
3159 case IT_CREAT|IT_OPEN:
3160 /* XXX swab here to assert that an mds_open reint
3161 * packet is following */
3162 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3164 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
3166 /* We abort the lock if the lookup was negative and
3167 * we did not make it to the OPEN portion */
3168 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
3169 RETURN(ELDLM_LOCK_ABORTED);
3170 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
3171 !intent_disposition(rep, DISP_OPEN_OPEN))
3173 /* IT_OPEN may return lock on cross-node dentry
3174 * that we want to hold during attr retrival -bzzz */
3175 if (rc != 0 || lockh[0].cookie == 0)
3176 RETURN(ELDLM_LOCK_ABORTED);
3179 getattr_part = MDS_INODELOCK_LOOKUP;
3182 getattr_part |= MDS_INODELOCK_LOOKUP;
3184 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3185 lock, &new_lock, lockh);
3186 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
3188 /* FIXME: LDLM can set req->rq_status. MDS sets
3189 policy_res{1,2} with disposition and status.
3190 - replay: returns 0 & req->status is old status
3191 - otherwise: returns req->status */
3192 if (intent_disposition(rep, DISP_LOOKUP_NEG))
3193 rep->lock_policy_res2 = 0;
3194 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
3195 rep->lock_policy_res2)
3196 RETURN(ELDLM_LOCK_ABORTED);
3197 if (req->rq_status != 0) {
3199 rep->lock_policy_res2 = req->rq_status;
3200 RETURN(ELDLM_LOCK_ABORTED);
3204 rc = mds_lock_and_check_slave(offset, req, lockh);
3205 if ((rep->lock_policy_res2 = rc)) {
3207 rep->lock_policy_res2 = 0;
3208 RETURN(ELDLM_LOCK_ABORTED);
3212 CERROR("Unhandled intent "LPD64"\n", it->opc);
3216 /* By this point, whatever function we called above must have either
3217 * filled in 'lockh', been an intent replay, or returned an error. We
3218 * want to allow replayed RPCs to not get a lock, since we would just
3219 * drop it below anyways because lock replay is done separately by the
3220 * client afterwards. For regular RPCs we want to give the new lock to
3221 * the client instead of whatever lock it was about to get. */
3222 if (new_lock == NULL)
3223 new_lock = ldlm_handle2lock(&lockh[0]);
3224 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
3227 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
3228 it->opc, lockh[0].cookie);
3230 /* If we've already given this lock to a client once, then we should
3231 * have no readers or writers. Otherwise, we should have one reader
3232 * _or_ writer ref (which will be zeroed below) before returning the
3233 * lock to a client. */
3234 if (new_lock->l_export == req->rq_export) {
3235 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3237 LASSERT(new_lock->l_export == NULL);
3238 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3243 if (new_lock->l_export == req->rq_export) {
3244 /* Already gave this to the client, which means that we
3245 * reconstructed a reply. */
3246 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3248 RETURN(ELDLM_LOCK_REPLACED);
3251 /* Fixup the lock to be given to the client */
3252 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
3253 new_lock->l_readers = 0;
3254 new_lock->l_writers = 0;
3256 new_lock->l_export = class_export_get(req->rq_export);
3257 list_add(&new_lock->l_export_chain,
3258 &new_lock->l_export->exp_ldlm_data.led_held_locks);
3260 new_lock->l_blocking_ast = lock->l_blocking_ast;
3261 new_lock->l_completion_ast = lock->l_completion_ast;
3263 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
3264 sizeof(lock->l_remote_handle));
3266 new_lock->l_flags &= ~LDLM_FL_LOCAL;
3268 LDLM_LOCK_PUT(new_lock);
3269 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
3271 RETURN(ELDLM_LOCK_REPLACED);
3274 int mds_attach(struct obd_device *dev, obd_count len, void *data)
3276 struct lprocfs_static_vars lvars;
3279 lprocfs_init_multi_vars(0, &lvars);
3281 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3285 return lprocfs_alloc_md_stats(dev, 0);
3288 int mds_detach(struct obd_device *dev)
3290 lprocfs_free_md_stats(dev);
3291 return lprocfs_obd_detach(dev);
3294 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
3296 struct lprocfs_static_vars lvars;
3298 lprocfs_init_multi_vars(1, &lvars);
3299 return lprocfs_obd_attach(dev, lvars.obd_vars);
3302 int mdt_detach(struct obd_device *dev)
3304 return lprocfs_obd_detach(dev);
3307 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
3309 struct mds_obd *mds = &obd->u.mds;
3314 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3315 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
3316 mds_handle, "mds", obd->obd_proc_entry);
3318 if (!mds->mds_service) {
3319 CERROR("failed to start service\n");
3323 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
3326 GOTO(err_thread, rc);
3328 mds->mds_setattr_service =
3329 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3330 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
3331 mds_handle, "mds_setattr",
3332 obd->obd_proc_entry);
3333 if (!mds->mds_setattr_service) {
3334 CERROR("failed to start getattr service\n");
3335 GOTO(err_thread, rc = -ENOMEM);
3338 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
3339 MDT_NUM_THREADS, "ll_mdt_attr");
3341 GOTO(err_thread2, rc);
3343 mds->mds_readpage_service =
3344 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3345 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
3346 mds_handle, "mds_readpage",
3347 obd->obd_proc_entry);
3348 if (!mds->mds_readpage_service) {
3349 CERROR("failed to start readpage service\n");
3350 GOTO(err_thread2, rc = -ENOMEM);
3353 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
3354 MDT_NUM_THREADS, "ll_mdt_rdpg");
3357 GOTO(err_thread3, rc);
3362 ptlrpc_unregister_service(mds->mds_readpage_service);
3364 ptlrpc_unregister_service(mds->mds_setattr_service);
3366 ptlrpc_unregister_service(mds->mds_service);
3370 static int mdt_cleanup(struct obd_device *obd, int flags)
3372 struct mds_obd *mds = &obd->u.mds;
3375 ptlrpc_stop_all_threads(mds->mds_readpage_service);
3376 ptlrpc_unregister_service(mds->mds_readpage_service);
3378 ptlrpc_stop_all_threads(mds->mds_setattr_service);
3379 ptlrpc_unregister_service(mds->mds_setattr_service);
3381 ptlrpc_stop_all_threads(mds->mds_service);
3382 ptlrpc_unregister_service(mds->mds_service);
3387 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
3388 __u64 gr, void *data)
3390 struct lustre_id id;
3391 struct obd_device *obd = data;
3395 return mds_id2dentry(obd, &id, NULL);
3398 static int mds_get_info(struct obd_export *exp, __u32 keylen,
3399 void *key, __u32 *valsize, void *val)
3401 struct obd_device *obd;
3402 struct mds_obd *mds;
3405 obd = class_exp2obd(exp);
3409 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
3410 exp->exp_handle.h_cookie);
3414 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
3415 /* get log_context handle. */
3416 unsigned long *llh_handle = val;
3417 *valsize = sizeof(unsigned long);
3418 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
3421 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
3422 /* get log_context handle. */
3423 unsigned long *sb = val;
3424 *valsize = sizeof(unsigned long);
3425 *sb = (unsigned long)obd->u.mds.mds_sb;
3429 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
3430 __u32 *mdsize = val;
3431 *valsize = sizeof(*mdsize);
3432 *mdsize = mds->mds_max_mdsize;
3436 if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
3437 __u32 *mdsnum = val;
3438 *valsize = sizeof(*mdsnum);
3439 *mdsnum = mds->mds_num;
3443 if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
3444 struct lustre_id *rootid = val;
3445 *valsize = sizeof(struct lustre_id);
3446 *rootid = mds->mds_rootid;
3450 CDEBUG(D_IOCTL, "invalid key\n");
3454 struct lvfs_callback_ops mds_lvfs_ops = {
3455 l_id2dentry: mds_lvfs_id2dentry,
3458 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
3459 int objcount, struct obd_ioobj *obj,
3460 int niocount, struct niobuf_remote *nb,
3461 struct niobuf_local *res,
3462 struct obd_trans_info *oti);
3464 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
3465 int objcount, struct obd_ioobj *obj, int niocount,
3466 struct niobuf_local *res, struct obd_trans_info *oti,
3469 /* use obd ops to offer management infrastructure */
3470 static struct obd_ops mds_obd_ops = {
3471 .o_owner = THIS_MODULE,
3472 .o_attach = mds_attach,
3473 .o_detach = mds_detach,
3474 .o_connect = mds_connect,
3475 .o_connect_post = mds_connect_post,
3476 .o_init_export = mds_init_export,
3477 .o_destroy_export = mds_destroy_export,
3478 .o_disconnect = mds_disconnect,
3479 .o_setup = mds_setup,
3480 .o_precleanup = mds_precleanup,
3481 .o_cleanup = mds_cleanup,
3482 .o_postrecov = mds_postrecov,
3483 .o_statfs = mds_obd_statfs,
3484 .o_iocontrol = mds_iocontrol,
3485 .o_create = mds_obd_create,
3486 .o_destroy = mds_obd_destroy,
3487 .o_llog_init = mds_llog_init,
3488 .o_llog_finish = mds_llog_finish,
3489 .o_notify = mds_notify,
3490 .o_get_info = mds_get_info,
3491 .o_set_info = mds_set_info,
3492 .o_preprw = mds_preprw,
3493 .o_commitrw = mds_commitrw,
3496 static struct obd_ops mdt_obd_ops = {
3497 .o_owner = THIS_MODULE,
3498 .o_attach = mdt_attach,
3499 .o_detach = mdt_detach,
3500 .o_setup = mdt_setup,
3501 .o_cleanup = mdt_cleanup,
3504 static int __init mds_init(void)
3506 struct lprocfs_static_vars lvars;
3508 mds_group_hash_init();
3510 lprocfs_init_multi_vars(0, &lvars);
3511 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
3513 lprocfs_init_multi_vars(1, &lvars);
3514 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
3520 static void /*__exit*/ mds_exit(void)
3522 mds_group_hash_cleanup();
3524 class_unregister_type(LUSTRE_MDS_NAME);
3525 class_unregister_type(LUSTRE_MDT_NAME);
3528 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3529 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
3530 MODULE_LICENSE("GPL");
3532 module_init(mds_init);
3533 module_exit(mds_exit);