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/lustre_mds.h>
53 #include <linux/lustre_fsfilt.h>
54 #include <linux/lprocfs_status.h>
55 #include <linux/lustre_commit_confd.h>
57 #include "mds_internal.h"
59 static int mds_intent_policy(struct ldlm_namespace *ns,
60 struct ldlm_lock **lockp, void *req_cookie,
61 ldlm_mode_t mode, int flags, void *data);
62 static int mds_postsetup(struct obd_device *obd);
63 static int mds_cleanup(struct obd_device *obd, int flags);
65 /* Assumes caller has already pushed into the kernel filesystem context */
66 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
67 loff_t offset, int count)
69 struct ptlrpc_bulk_desc *desc;
70 struct l_wait_info lwi;
72 int rc = 0, npages, i, tmpcount, tmpsize = 0;
75 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
77 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
78 OBD_ALLOC(pages, sizeof(*pages) * npages);
80 GOTO(out, rc = -ENOMEM);
82 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
85 GOTO(out_free, rc = -ENOMEM);
87 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
88 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
90 pages[i] = alloc_pages(GFP_KERNEL, 0);
92 GOTO(cleanup_buf, rc = -ENOMEM);
94 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
97 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
98 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
99 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
100 tmpsize, offset, file->f_dentry->d_inode->i_ino,
101 file->f_dentry->d_inode->i_size);
103 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
104 page_address(pages[i]), tmpsize, &offset);
107 GOTO(cleanup_buf, rc = -EIO);
110 LASSERT(desc->bd_nob == count);
112 rc = ptlrpc_start_bulk_transfer(desc);
114 GOTO(cleanup_buf, rc);
116 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
117 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
118 OBD_FAIL_MDS_SENDPAGE, rc);
119 GOTO(abort_bulk, rc);
122 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
123 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
124 LASSERT (rc == 0 || rc == -ETIMEDOUT);
127 if (desc->bd_success &&
128 desc->bd_nob_transferred == count)
129 GOTO(cleanup_buf, rc);
131 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
134 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
135 (rc == -ETIMEDOUT) ? "timeout" : "network error",
136 desc->bd_nob_transferred, count,
137 req->rq_export->exp_client_uuid.uuid,
138 req->rq_export->exp_connection->c_remote_uuid.uuid);
140 ptlrpc_fail_export(req->rq_export);
144 ptlrpc_abort_bulk (desc);
146 for (i = 0; i < npages; i++)
148 __free_pages(pages[i], 0);
150 ptlrpc_free_bulk(desc);
152 OBD_FREE(pages, sizeof(*pages) * npages);
157 /* only valid locked dentries or errors should be returned */
158 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
159 struct vfsmount **mnt, int lock_mode,
160 struct lustre_handle *lockh,
161 char *name, int namelen, __u64 lockpart)
163 struct mds_obd *mds = &obd->u.mds;
164 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
165 struct ldlm_res_id res_id = { .name = {0} };
167 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
174 res_id.name[0] = de->d_inode->i_ino;
175 res_id.name[1] = de->d_inode->i_generation;
176 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
177 LDLM_IBITS, &policy, lock_mode, &flags,
178 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
179 NULL, 0, NULL, lockh);
180 if (rc != ELDLM_OK) {
182 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
188 #ifndef DCACHE_DISCONNECTED
189 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
193 /* Look up an entry by inode number. */
194 /* this function ONLY returns valid dget'd dentries with an initialized inode
196 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
197 struct vfsmount **mnt)
200 unsigned long ino = fid->id;
201 __u32 generation = fid->generation;
203 struct dentry *result;
206 RETURN(ERR_PTR(-ESTALE));
208 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
210 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
211 ino, generation, mds->mds_sb);
213 /* under ext3 this is neither supposed to return bad inodes
215 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
219 inode = result->d_inode;
221 RETURN(ERR_PTR(-ENOENT));
223 if (generation && inode->i_generation != generation) {
224 /* we didn't find the right inode.. */
225 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
226 inode->i_ino, (unsigned long)inode->i_nlink,
227 atomic_read(&inode->i_count), inode->i_generation,
230 RETURN(ERR_PTR(-ENOENT));
234 *mnt = mds->mds_vfsmnt;
242 /* Establish a connection to the MDS.
244 * This will set up an export structure for the client to hold state data
245 * about that client, like open files, the last operation number it did
246 * on the server, etc.
248 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
249 struct obd_uuid *cluuid)
251 struct obd_export *exp;
252 struct mds_export_data *med; /* */
253 struct mds_client_data *mcd;
254 int rc, abort_recovery;
257 if (!conn || !obd || !cluuid)
260 /* Check for aborted recovery. */
261 spin_lock_bh(&obd->obd_processing_task_lock);
262 abort_recovery = obd->obd_abort_recovery;
263 spin_unlock_bh(&obd->obd_processing_task_lock);
265 target_abort_recovery(obd);
267 /* XXX There is a small race between checking the list and adding a
268 * new connection for the same UUID, but the real threat (list
269 * corruption when multiple different clients connect) is solved.
271 * There is a second race between adding the export to the list,
272 * and filling in the client data below. Hence skipping the case
273 * of NULL mcd above. We should already be controlling multiple
274 * connects at the client, and we can't hold the spinlock over
275 * memory allocations without risk of deadlocking.
277 rc = class_connect(conn, obd, cluuid);
280 exp = class_conn2export(conn);
282 med = &exp->exp_mds_data;
284 OBD_ALLOC(mcd, sizeof(*mcd));
286 CERROR("mds: out of memory for client data\n");
287 GOTO(out, rc = -ENOMEM);
290 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
293 rc = mds_client_add(obd, &obd->u.mds, med, -1);
298 OBD_FREE(mcd, sizeof(*mcd));
299 class_disconnect(exp, 0);
301 class_export_put(exp);
306 static int mds_init_export(struct obd_export *exp)
308 struct mds_export_data *med = &exp->exp_mds_data;
310 INIT_LIST_HEAD(&med->med_open_head);
311 spin_lock_init(&med->med_open_lock);
315 static int mds_destroy_export(struct obd_export *export)
317 struct mds_export_data *med;
318 struct obd_device *obd = export->exp_obd;
319 struct lvfs_run_ctxt saved;
323 med = &export->exp_mds_data;
324 target_destroy_export(export);
326 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
329 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
331 /* Close any open files (which may also cause orphan unlinking). */
332 spin_lock(&med->med_open_lock);
333 while (!list_empty(&med->med_open_head)) {
334 struct list_head *tmp = med->med_open_head.next;
335 struct mds_file_data *mfd =
336 list_entry(tmp, struct mds_file_data, mfd_list);
337 BDEVNAME_DECLARE_STORAGE(btmp);
339 /* bug 1579: fix force-closing for 2.5 */
340 struct dentry *dentry = mfd->mfd_dentry;
342 list_del(&mfd->mfd_list);
343 spin_unlock(&med->med_open_lock);
345 CERROR("force closing client file handle for %*s (%s:%lu)\n",
346 dentry->d_name.len, dentry->d_name.name,
347 ll_bdevname(dentry->d_inode->i_sb, btmp),
348 dentry->d_inode->i_ino);
349 rc = mds_mfd_close(NULL, obd, mfd,
350 !(export->exp_flags & OBD_OPT_FAILOVER));
353 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
354 spin_lock(&med->med_open_lock);
356 spin_unlock(&med->med_open_lock);
357 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
360 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
365 static int mds_disconnect(struct obd_export *exp, int flags)
367 unsigned long irqflags;
372 class_export_get(exp);
374 spin_lock_irqsave(&exp->exp_lock, irqflags);
375 exp->exp_flags = flags;
376 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
378 /* Disconnect early so that clients can't keep using export */
379 rc = class_disconnect(exp, flags);
380 ldlm_cancel_locks_for_export(exp);
382 /* complete all outstanding replies */
383 spin_lock_irqsave(&exp->exp_lock, irqflags);
384 while (!list_empty(&exp->exp_outstanding_replies)) {
385 struct ptlrpc_reply_state *rs =
386 list_entry(exp->exp_outstanding_replies.next,
387 struct ptlrpc_reply_state, rs_exp_list);
388 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
390 spin_lock(&svc->srv_lock);
391 list_del_init(&rs->rs_exp_list);
392 ptlrpc_schedule_difficult_reply(rs);
393 spin_unlock(&svc->srv_lock);
395 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
397 class_export_put(exp);
401 static int mds_getstatus(struct ptlrpc_request *req)
403 struct mds_obd *mds = mds_req2mds(req);
404 struct mds_body *body;
405 int rc, size = sizeof(*body);
408 rc = lustre_pack_reply(req, 1, &size, NULL);
409 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
410 CERROR("mds: out of memory for message: size=%d\n", size);
411 req->rq_status = -ENOMEM; /* superfluous? */
415 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
416 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
418 /* the last_committed and last_xid fields are filled in for all
419 * replies already - no need to do so here also.
424 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
425 void *data, int flag)
430 if (flag == LDLM_CB_CANCELING) {
431 /* Don't need to do anything here. */
435 /* XXX layering violation! -phil */
436 l_lock(&lock->l_resource->lr_namespace->ns_lock);
437 /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
438 * such that mds_blocking_ast is called just before l_i_p takes the
439 * ns_lock, then by the time we get the lock, we might not be the
440 * correct blocking function anymore. So check, and return early, if
442 if (lock->l_blocking_ast != mds_blocking_ast) {
443 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
447 lock->l_flags |= LDLM_FL_CBPENDING;
448 do_ast = (!lock->l_readers && !lock->l_writers);
449 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
452 struct lustre_handle lockh;
455 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
456 ldlm_lock2handle(lock, &lockh);
457 rc = ldlm_cli_cancel(&lockh);
459 CERROR("ldlm_cli_cancel: %d\n", rc);
461 LDLM_DEBUG(lock, "Lock still has references, will be "
467 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
475 rc = fsfilt_get_md(obd, inode, md, *size);
480 CERROR("Error %d reading eadata for ino %lu\n",
484 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
498 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
499 * Call with lock=0 if the caller has already taken the i_sem. */
500 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
501 struct mds_body *body, struct inode *inode, int lock)
503 struct mds_obd *mds = &obd->u.mds;
509 lmm = lustre_msg_buf(msg, offset, 0);
511 /* Some problem with getting eadata when I sized the reply
513 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
517 lmm_size = msg->buflens[offset];
519 /* I don't really like this, but it is a sanity check on the client
520 * MD request. However, if the client doesn't know how much space
521 * to reserve for the MD, it shouldn't be bad to have too much space.
523 if (lmm_size > mds->mds_max_mdsize) {
524 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
525 inode->i_ino, lmm_size, mds->mds_max_mdsize);
529 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
531 if (S_ISDIR(inode->i_mode))
532 body->valid |= OBD_MD_FLDIREA;
534 body->valid |= OBD_MD_FLEASIZE;
535 body->eadatasize = lmm_size;
542 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
543 struct ptlrpc_request *req,
544 struct mds_body *reqbody, int reply_off)
546 struct mds_body *body;
547 struct inode *inode = dentry->d_inode;
554 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
555 LASSERT(body != NULL); /* caller prepped reply */
557 mds_pack_inode2fid(&body->fid1, inode);
558 mds_pack_inode2body(body, inode);
560 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
561 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
562 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
565 /* If we have LOV EA data, the OST holds size, atime, mtime */
566 if (!(body->valid & OBD_MD_FLEASIZE) &&
567 !(body->valid & OBD_MD_FLDIREA))
568 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
569 OBD_MD_FLATIME | OBD_MD_FLMTIME);
570 } else if (S_ISLNK(inode->i_mode) &&
571 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
572 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
575 LASSERT (symname != NULL); /* caller prepped reply */
576 len = req->rq_repmsg->buflens[reply_off + 1];
578 rc = inode->i_op->readlink(dentry, symname, len);
580 CERROR("readlink failed: %d\n", rc);
581 } else if (rc != len - 1) {
582 CERROR ("Unexpected readlink rc %d: expecting %d\n",
586 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
587 body->valid |= OBD_MD_LINKNAME;
588 body->eadatasize = rc + 1;
589 symname[rc] = 0; /* NULL terminate */
597 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
600 struct mds_obd *mds = mds_req2mds(req);
601 struct mds_body *body;
602 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
605 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
606 LASSERT(body != NULL); /* checked by caller */
607 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
609 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
610 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
613 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
615 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
619 CERROR("error getting inode %lu MD: rc = %d\n",
622 } else if (rc > mds->mds_max_mdsize) {
624 CERROR("MD size %d larger than maximum possible %u\n",
625 rc, mds->mds_max_mdsize);
630 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
631 if (inode->i_size + 1 != body->eadatasize)
632 CERROR("symlink size: %Lu, reply space: %d\n",
633 inode->i_size + 1, body->eadatasize);
634 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
636 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
637 inode->i_size + 1, body->eadatasize);
640 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
641 CERROR("failed MDS_GETATTR_PACK test\n");
642 req->rq_status = -ENOMEM;
643 GOTO(out, rc = -ENOMEM);
646 rc = lustre_pack_reply(req, bufcount, size, NULL);
648 CERROR("out of memory\n");
649 GOTO(out, req->rq_status = rc);
657 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
658 struct lustre_handle *child_lockh, int child_part)
660 struct obd_device *obd = req->rq_export->exp_obd;
661 struct ldlm_reply *rep = NULL;
662 struct lvfs_run_ctxt saved;
663 struct mds_body *body;
664 struct dentry *dparent = NULL, *dchild = NULL;
665 struct lvfs_ucred uc;
666 struct lustre_handle parent_lockh;
668 int rc = 0, cleanup_phase = 0, resent_req = 0;
672 LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
674 /* Swab now, before anyone looks inside the request */
676 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
677 lustre_swab_mds_body);
679 CERROR("Can't swab mds_body\n");
680 GOTO(cleanup, rc = -EFAULT);
683 LASSERT_REQSWAB(req, offset + 1);
684 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
686 CERROR("Can't unpack name\n");
687 GOTO(cleanup, rc = -EFAULT);
689 namesize = req->rq_reqmsg->buflens[offset + 1];
691 LASSERT (offset == 0 || offset == 2);
692 /* if requests were at offset 2, the getattr reply goes back at 1 */
694 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
698 uc.luc_fsuid = body->fsuid;
699 uc.luc_fsgid = body->fsgid;
700 uc.luc_cap = body->capability;
701 uc.luc_suppgid1 = body->suppgid;
702 uc.luc_suppgid2 = -1;
703 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
704 cleanup_phase = 1; /* kernel context */
705 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
707 /* FIXME: handle raw lookup */
709 if (body->valid == OBD_MD_FLID) {
710 struct mds_body *mds_reply;
711 int size = sizeof(*mds_reply);
713 // The user requested ONLY the inode number, so do a raw lookup
714 rc = lustre_pack_reply(req, 1, &size, NULL);
716 CERROR("out of memory\n");
720 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
722 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
724 mds_reply->fid1.id = inum;
725 mds_reply->valid = OBD_MD_FLID;
730 if (child_lockh->cookie != 0) {
731 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
735 if (resent_req == 0) {
736 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
737 &parent_lockh, &dparent,
738 LCK_PR, MDS_INODELOCK_UPDATE,
740 child_lockh, &dchild, LCK_PR,
745 struct ldlm_lock *granted_lock;
746 struct ll_fid child_fid;
747 struct ldlm_resource *res;
748 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
749 granted_lock = ldlm_handle2lock(child_lockh);
750 LASSERT(granted_lock);
752 res = granted_lock->l_resource;
753 child_fid.id = res->lr_name.name[0];
754 child_fid.generation = res->lr_name.name[1];
755 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
757 LDLM_LOCK_PUT(granted_lock);
760 cleanup_phase = 2; /* dchild, dparent, locks */
762 if (dchild->d_inode == NULL) {
763 intent_set_disposition(rep, DISP_LOOKUP_NEG);
764 /* in the intent case, the policy clears this error:
765 the disposition is enough */
766 GOTO(cleanup, rc = -ENOENT);
768 intent_set_disposition(rep, DISP_LOOKUP_POS);
771 if (req->rq_repmsg == NULL) {
772 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
774 CERROR ("mds_getattr_pack_msg: %d\n", rc);
779 rc = mds_getattr_internal(obd, dchild, req, body, offset);
780 GOTO(cleanup, rc); /* returns the lock to the client */
783 switch (cleanup_phase) {
785 if (resent_req == 0) {
786 if (rc && dchild->d_inode)
787 ldlm_lock_decref(child_lockh, LCK_PR);
788 ldlm_lock_decref(&parent_lockh, LCK_PR);
793 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
799 static int mds_getattr(int offset, struct ptlrpc_request *req)
801 struct mds_obd *mds = mds_req2mds(req);
802 struct obd_device *obd = req->rq_export->exp_obd;
803 struct lvfs_run_ctxt saved;
805 struct mds_body *body;
806 struct lvfs_ucred uc;
810 body = lustre_swab_reqbuf (req, offset, sizeof (*body),
811 lustre_swab_mds_body);
813 CERROR ("Can't unpack body\n");
817 uc.luc_fsuid = body->fsuid;
818 uc.luc_fsgid = body->fsgid;
819 uc.luc_cap = body->capability;
820 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
821 de = mds_fid2dentry(mds, &body->fid1, NULL);
823 rc = req->rq_status = -ENOENT;
824 GOTO(out_pop, PTR_ERR(de));
827 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
829 CERROR ("mds_getattr_pack_msg: %d\n", rc);
833 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
838 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
843 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
844 unsigned long max_age)
848 spin_lock(&obd->obd_osfs_lock);
849 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
851 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
852 spin_unlock(&obd->obd_osfs_lock);
857 static int mds_statfs(struct ptlrpc_request *req)
859 struct obd_device *obd = req->rq_export->exp_obd;
860 int rc, size = sizeof(struct obd_statfs);
863 rc = lustre_pack_reply(req, 1, &size, NULL);
864 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
865 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
869 /* We call this so that we can cache a bit - 1 jiffie worth */
870 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
873 CERROR("mds_obd_statfs failed: rc %d\n", rc);
883 static int mds_sync(struct ptlrpc_request *req)
885 struct obd_device *obd = req->rq_export->exp_obd;
886 struct mds_obd *mds = &obd->u.mds;
887 struct mds_body *body;
888 int rc, size = sizeof(*body);
891 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
893 GOTO(out, rc = -EPROTO);
895 rc = lustre_pack_reply(req, 1, &size, NULL);
896 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
897 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
901 if (body->fid1.id == 0) {
902 /* a fid of zero is taken to mean "sync whole filesystem" */
903 rc = fsfilt_sync(obd, mds->mds_sb);
907 /* just any file to grab fsync method - "file" arg unused */
908 struct file *file = mds->mds_rcvd_filp;
911 de = mds_fid2dentry(mds, &body->fid1, NULL);
913 GOTO(out, rc = PTR_ERR(de));
915 rc = file->f_op->fsync(NULL, de, 1);
920 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
921 mds_pack_inode2fid(&body->fid1, de->d_inode);
922 mds_pack_inode2body(body, de->d_inode);
929 /* mds_readpage does not take a DLM lock on the inode, because the client must
930 * already have a PR lock.
932 * If we were to take another one here, a deadlock will result, if another
933 * thread is already waiting for a PW lock. */
934 static int mds_readpage(struct ptlrpc_request *req)
936 struct obd_device *obd = req->rq_export->exp_obd;
937 struct vfsmount *mnt;
940 struct mds_body *body, *repbody;
941 struct lvfs_run_ctxt saved;
942 int rc, size = sizeof(*repbody);
943 struct lvfs_ucred uc;
946 rc = lustre_pack_reply(req, 1, &size, NULL);
947 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
948 CERROR("mds: out of memory\n");
949 GOTO(out, rc = -ENOMEM);
952 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
954 GOTO (out, rc = -EFAULT);
956 uc.luc_fsuid = body->fsuid;
957 uc.luc_fsgid = body->fsgid;
958 uc.luc_cap = body->capability;
959 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
960 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
962 GOTO(out_pop, rc = PTR_ERR(de));
964 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
966 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
967 /* note: in case of an error, dentry_open puts dentry */
969 GOTO(out_pop, rc = PTR_ERR(file));
971 /* body->size is actually the offset -eeb */
972 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
973 CERROR("offset "LPU64" not on a block boundary of %lu\n",
974 body->size, de->d_inode->i_blksize);
975 GOTO(out_file, rc = -EFAULT);
978 /* body->nlink is actually the #bytes to read -eeb */
979 if (body->nlink & (de->d_inode->i_blksize - 1)) {
980 CERROR("size %u is not multiple of blocksize %lu\n",
981 body->nlink, de->d_inode->i_blksize);
982 GOTO(out_file, rc = -EFAULT);
985 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
986 repbody->size = file->f_dentry->d_inode->i_size;
987 repbody->valid = OBD_MD_FLSIZE;
989 /* to make this asynchronous make sure that the handling function
990 doesn't send a reply when this function completes. Instead a
991 callback function would send the reply */
992 /* body->size is actually the offset -eeb */
993 rc = mds_sendpage(req, file, body->size, body->nlink);
998 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1000 req->rq_status = rc;
1004 int mds_reint(struct ptlrpc_request *req, int offset,
1005 struct lustre_handle *lockh)
1007 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
1010 OBD_ALLOC(rec, sizeof(*rec));
1014 rc = mds_update_unpack(req, offset, rec);
1015 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1016 CERROR("invalid record\n");
1017 GOTO(out, req->rq_status = -EINVAL);
1019 /* rc will be used to interrupt a for loop over multiple records */
1020 rc = mds_reint_rec(rec, offset, req, lockh);
1022 OBD_FREE(rec, sizeof(*rec));
1026 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1027 struct obd_device *obd, int *process)
1029 switch (req->rq_reqmsg->opc) {
1030 case MDS_CONNECT: /* This will never get here, but for completeness. */
1031 case OST_CONNECT: /* This will never get here, but for completeness. */
1032 case MDS_DISCONNECT:
1033 case OST_DISCONNECT:
1038 case MDS_SYNC: /* used in unmounting */
1042 *process = target_queue_recovery_request(req, obd);
1046 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1048 /* XXX what should we set rq_status to here? */
1049 req->rq_status = -EAGAIN;
1050 RETURN(ptlrpc_error(req));
1054 static char *reint_names[] = {
1055 [REINT_SETATTR] "setattr",
1056 [REINT_CREATE] "create",
1057 [REINT_LINK] "link",
1058 [REINT_UNLINK] "unlink",
1059 [REINT_RENAME] "rename",
1060 [REINT_OPEN] "open",
1063 int mds_handle(struct ptlrpc_request *req)
1065 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1067 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1068 struct obd_device *obd = NULL;
1071 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1073 LASSERT(current->journal_info == NULL);
1074 /* XXX identical to OST */
1075 if (req->rq_reqmsg->opc != MDS_CONNECT) {
1076 struct mds_export_data *med;
1077 int recovering, abort_recovery;
1079 if (req->rq_export == NULL) {
1080 CERROR("lustre_mds: operation %d on unconnected MDS\n",
1081 req->rq_reqmsg->opc);
1082 req->rq_status = -ENOTCONN;
1083 GOTO(out, rc = -ENOTCONN);
1086 med = &req->rq_export->exp_mds_data;
1087 obd = req->rq_export->exp_obd;
1090 /* Check for aborted recovery. */
1091 spin_lock_bh(&obd->obd_processing_task_lock);
1092 abort_recovery = obd->obd_abort_recovery;
1093 recovering = obd->obd_recovering;
1094 spin_unlock_bh(&obd->obd_processing_task_lock);
1095 if (abort_recovery) {
1096 target_abort_recovery(obd);
1097 } else if (recovering) {
1098 rc = mds_filter_recovery_request(req, obd,
1100 if (rc || !should_process)
1105 switch (req->rq_reqmsg->opc) {
1107 DEBUG_REQ(D_INODE, req, "connect");
1108 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1109 rc = target_handle_connect(req, mds_handle);
1111 /* Now that we have an export, set mds. */
1112 mds = mds_req2mds(req);
1115 case MDS_DISCONNECT:
1116 DEBUG_REQ(D_INODE, req, "disconnect");
1117 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1118 rc = target_handle_disconnect(req);
1119 req->rq_status = rc; /* superfluous? */
1123 DEBUG_REQ(D_INODE, req, "getstatus");
1124 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1125 rc = mds_getstatus(req);
1129 DEBUG_REQ(D_INODE, req, "getattr");
1130 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1131 rc = mds_getattr(0, req);
1134 case MDS_GETATTR_NAME: {
1135 struct lustre_handle lockh;
1136 DEBUG_REQ(D_INODE, req, "getattr_name");
1137 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1139 /* If this request gets a reconstructed reply, we won't be
1140 * acquiring any new locks in mds_getattr_name, so we don't
1144 rc = mds_getattr_name(0, req, &lockh, MDS_INODELOCK_UPDATE);
1145 /* this non-intent call (from an ioctl) is special */
1146 req->rq_status = rc;
1147 if (rc == 0 && lockh.cookie)
1148 ldlm_lock_decref(&lockh, LCK_PR);
1152 DEBUG_REQ(D_INODE, req, "statfs");
1153 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1154 rc = mds_statfs(req);
1158 DEBUG_REQ(D_INODE, req, "readpage");
1159 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1160 rc = mds_readpage(req);
1162 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1163 if (req->rq_reply_state) {
1164 lustre_free_reply_state (req->rq_reply_state);
1165 req->rq_reply_state = NULL;
1173 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1175 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1176 mds->mds_max_cookiesize};
1179 /* NB only peek inside req now; mds_reint() will swab it */
1181 CERROR ("Can't inspect opcode\n");
1186 if (lustre_msg_swabbed (req->rq_reqmsg))
1189 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1190 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1191 reint_names[opc] == NULL) ? reint_names[opc] :
1194 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1196 if (opc == REINT_UNLINK)
1198 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1203 rc = lustre_pack_reply(req, bufcount, size, NULL);
1207 rc = mds_reint(req, 0, NULL);
1208 fail = OBD_FAIL_MDS_REINT_NET_REP;
1213 DEBUG_REQ(D_INODE, req, "close");
1214 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1215 rc = mds_close(req);
1218 case MDS_DONE_WRITING:
1219 DEBUG_REQ(D_INODE, req, "done_writing");
1220 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1221 rc = mds_done_writing(req);
1225 DEBUG_REQ(D_INODE, req, "pin");
1226 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1231 DEBUG_REQ(D_INODE, req, "sync");
1232 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1237 DEBUG_REQ(D_INODE, req, "ping");
1238 rc = target_handle_ping(req);
1241 case OBD_LOG_CANCEL:
1242 CDEBUG(D_INODE, "log cancel\n");
1243 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1244 rc = -ENOTSUPP; /* la la la */
1248 DEBUG_REQ(D_INODE, req, "enqueue");
1249 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1250 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1251 ldlm_server_blocking_ast, NULL);
1254 DEBUG_REQ(D_INODE, req, "convert");
1255 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1256 rc = ldlm_handle_convert(req);
1258 case LDLM_BL_CALLBACK:
1259 case LDLM_CP_CALLBACK:
1260 DEBUG_REQ(D_INODE, req, "callback");
1261 CERROR("callbacks should not happen on MDS\n");
1263 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1265 case LLOG_ORIGIN_HANDLE_CREATE:
1266 DEBUG_REQ(D_INODE, req, "llog_init");
1267 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1268 rc = llog_origin_handle_create(req);
1270 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1271 DEBUG_REQ(D_INODE, req, "llog next block");
1272 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1273 rc = llog_origin_handle_next_block(req);
1275 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1276 DEBUG_REQ(D_INODE, req, "llog prev block");
1277 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1278 rc = llog_origin_handle_prev_block(req);
1280 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1281 DEBUG_REQ(D_INODE, req, "llog read header");
1282 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1283 rc = llog_origin_handle_read_header(req);
1285 case LLOG_ORIGIN_HANDLE_CLOSE:
1286 DEBUG_REQ(D_INODE, req, "llog close");
1287 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1288 rc = llog_origin_handle_close(req);
1291 DEBUG_REQ(D_INODE, req, "llog catinfo");
1292 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1293 rc = llog_catinfo(req);
1296 req->rq_status = -ENOTSUPP;
1297 rc = ptlrpc_error(req);
1301 LASSERT(current->journal_info == NULL);
1305 /* If we're DISCONNECTing, the mds_export_data is already freed */
1306 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1307 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1308 struct obd_device *obd = list_entry(mds, struct obd_device,
1310 req->rq_repmsg->last_xid =
1311 le64_to_cpu(med->med_mcd->mcd_last_xid);
1313 if (!obd->obd_no_transno) {
1314 req->rq_repmsg->last_committed =
1315 obd->obd_last_committed;
1317 DEBUG_REQ(D_IOCTL, req,
1318 "not sending last_committed update");
1320 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1322 mds->mds_last_transno, obd->obd_last_committed,
1327 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1328 if (obd && obd->obd_recovering) {
1329 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1330 return target_queue_final_reply(req, rc);
1332 /* Lost a race with recovery; let the error path DTRT. */
1333 rc = req->rq_status = -ENOTCONN;
1336 target_send_reply(req, rc, fail);
1340 /* Update the server data on disk. This stores the new mount_count and
1341 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1342 * then the server last_rcvd value may be less than that of the clients.
1343 * This will alert us that we may need to do client recovery.
1345 * Also assumes for mds_last_transno that we are not modifying it (no locking).
1347 int mds_update_server_data(struct obd_device *obd, int force_sync)
1349 struct mds_obd *mds = &obd->u.mds;
1350 struct mds_server_data *msd = mds->mds_server_data;
1351 struct file *filp = mds->mds_rcvd_filp;
1352 struct lvfs_run_ctxt saved;
1357 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1358 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1360 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1361 mds->mds_mount_count, mds->mds_last_transno);
1362 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1364 CERROR("error writing MDS server data: rc = %d\n", rc);
1365 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1371 /* mount the file system (secretly) */
1372 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1374 struct lprocfs_static_vars lvars;
1375 struct lustre_cfg* lcfg = buf;
1376 struct mds_obd *mds = &obd->u.mds;
1377 char *options = NULL;
1378 struct vfsmount *mnt;
1383 dev_clear_rdonly(2);
1385 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1386 RETURN(rc = -EINVAL);
1388 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1389 if (IS_ERR(obd->obd_fsops))
1390 RETURN(rc = PTR_ERR(obd->obd_fsops));
1392 page = __get_free_page(GFP_KERNEL);
1396 options = (char *)page;
1397 memset(options, 0, PAGE_SIZE);
1399 /* here we use "iopen_nopriv" hardcoded, because it affects MDS utility
1400 * and the rest of options are passed by mount options. Probably this
1401 * should be moved to somewhere else like startup scripts or lconf. */
1402 sprintf(options, "iopen_nopriv");
1404 if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
1405 sprintf(options + strlen(options), ",%s",
1406 lcfg->lcfg_inlbuf4);
1408 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1409 lcfg->lcfg_inlbuf1, options);
1415 CERROR("do_kern_mount failed: rc = %d\n", rc);
1419 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1421 sema_init(&mds->mds_orphan_recovery_sem, 1);
1422 sema_init(&mds->mds_epoch_sem, 1);
1423 spin_lock_init(&mds->mds_transno_lock);
1424 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1425 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1426 atomic_set(&mds->mds_open_count, 0);
1428 obd->obd_namespace = ldlm_namespace_new("mds_server",
1429 LDLM_NAMESPACE_SERVER);
1430 if (obd->obd_namespace == NULL) {
1431 mds_cleanup(obd, 0);
1432 GOTO(err_put, rc = -ENOMEM);
1434 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1436 rc = mds_fs_setup(obd, mnt);
1438 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1442 rc = llog_start_commit_thread();
1446 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1449 generate_random_uuid(uuid);
1450 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1452 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1453 if (mds->mds_profile == NULL)
1454 GOTO(err_fs, rc = -ENOMEM);
1456 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1457 lcfg->lcfg_inllen3);
1461 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1462 "mds_ldlm_client", &obd->obd_ldlm_client);
1463 obd->obd_replayable = 1;
1465 rc = mds_postsetup(obd);
1469 lprocfs_init_vars(mds, &lvars);
1470 lprocfs_obd_setup(obd, lvars.obd_vars);
1475 /* No extra cleanup needed for llog_init_commit_thread() */
1476 mds_fs_cleanup(obd, 0);
1478 ldlm_namespace_free(obd->obd_namespace, 0);
1479 obd->obd_namespace = NULL;
1482 mntput(mds->mds_vfsmnt);
1486 fsfilt_put_ops(obd->obd_fsops);
1490 static int mds_postsetup(struct obd_device *obd)
1492 struct mds_obd *mds = &obd->u.mds;
1496 rc = obd_llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0,
1497 NULL, &llog_lvfs_ops);
1501 /* This check for @dumb string is needed to handle
1502 mounting MDS with smfs. Read lconf:MDSDEV.write_conf() for
1503 more detail explanation. */
1504 if (mds->mds_profile && strcmp(mds->mds_profile, "dumb")) {
1505 struct lvfs_run_ctxt saved;
1506 struct lustre_profile *lprof;
1507 struct config_llog_instance cfg;
1509 cfg.cfg_instance = NULL;
1510 cfg.cfg_uuid = mds->mds_lov_uuid;
1511 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1512 rc = class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1513 mds->mds_profile, &cfg);
1514 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1518 lprof = class_get_profile(mds->mds_profile);
1519 if (lprof == NULL) {
1520 CERROR("No profile found: %s\n", mds->mds_profile);
1521 GOTO(err_cleanup, rc = -ENOENT);
1523 rc = mds_lov_connect(obd, lprof->lp_osc);
1525 GOTO(err_cleanup, rc);
1533 obd_llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1537 static int mds_postrecov(struct obd_device *obd)
1542 LASSERT(!obd->obd_recovering);
1543 LASSERT(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT) != NULL);
1545 rc = llog_connect(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT),
1546 obd->u.mds.mds_lov_desc.ld_tgt_count,
1549 CERROR("faild at llog_origin_connect: %d\n", rc);
1552 rc = mds_cleanup_orphans(obd);
1554 rc2 = mds_lov_set_nextid(obd);
1560 int mds_lov_clean(struct obd_device *obd)
1562 struct mds_obd *mds = &obd->u.mds;
1564 if (mds->mds_profile) {
1566 struct config_llog_instance cfg;
1567 struct lvfs_run_ctxt saved;
1568 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1570 OBD_ALLOC(cln_prof, len);
1571 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1573 cfg.cfg_instance = NULL;
1574 cfg.cfg_uuid = mds->mds_lov_uuid;
1576 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1577 class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1579 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1581 OBD_FREE(cln_prof, len);
1582 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1583 mds->mds_profile = NULL;
1588 static int mds_precleanup(struct obd_device *obd, int flags)
1593 mds_lov_disconnect(obd, flags);
1595 obd_llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1599 static int mds_cleanup(struct obd_device *obd, int flags)
1601 struct mds_obd *mds = &obd->u.mds;
1604 if (mds->mds_sb == NULL)
1607 lprocfs_obd_cleanup(obd);
1609 mds_update_server_data(obd, 1);
1610 if (mds->mds_lov_objids != NULL) {
1611 OBD_FREE(mds->mds_lov_objids,
1612 mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1614 mds_fs_cleanup(obd, flags);
1618 /* 2 seems normal on mds, (may_umount() also expects 2
1619 fwiw), but we only see 1 at this point in obdfilter. */
1620 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1621 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1622 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1624 mntput(mds->mds_vfsmnt);
1628 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1630 spin_lock_bh(&obd->obd_processing_task_lock);
1631 if (obd->obd_recovering) {
1632 target_cancel_recovery_timer(obd);
1633 obd->obd_recovering = 0;
1635 spin_unlock_bh(&obd->obd_processing_task_lock);
1638 dev_clear_rdonly(2);
1639 fsfilt_put_ops(obd->obd_fsops);
1644 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
1645 struct ldlm_lock *new_lock,
1646 struct lustre_handle *lockh)
1648 struct obd_export *exp = req->rq_export;
1649 struct obd_device *obd = exp->exp_obd;
1650 struct ldlm_request *dlmreq =
1651 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
1652 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
1653 struct list_head *iter;
1655 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
1658 l_lock(&obd->obd_namespace->ns_lock);
1659 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
1660 struct ldlm_lock *lock;
1661 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1662 if (lock == new_lock)
1664 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
1665 lockh->cookie = lock->l_handle.h_cookie;
1666 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
1668 l_unlock(&obd->obd_namespace->ns_lock);
1672 l_unlock(&obd->obd_namespace->ns_lock);
1674 /* This remote handle isn't enqueued, so we never received or
1675 * processed this request. Clear MSG_RESENT, because it can
1676 * be handled like any normal request now. */
1678 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1680 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
1684 int intent_disposition(struct ldlm_reply *rep, int flag)
1688 return (rep->lock_policy_res1 & flag);
1691 void intent_set_disposition(struct ldlm_reply *rep, int flag)
1695 rep->lock_policy_res1 |= flag;
1698 static int mds_intent_policy(struct ldlm_namespace *ns,
1699 struct ldlm_lock **lockp, void *req_cookie,
1700 ldlm_mode_t mode, int flags, void *data)
1702 struct ptlrpc_request *req = req_cookie;
1703 struct ldlm_lock *lock = *lockp;
1704 struct ldlm_intent *it;
1705 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1706 struct ldlm_reply *rep;
1707 struct lustre_handle lockh = { 0 };
1708 struct ldlm_lock *new_lock;
1709 int getattr_part = MDS_INODELOCK_UPDATE;
1710 int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
1711 sizeof(struct mds_body),
1712 mds->mds_max_mdsize,
1713 mds->mds_max_cookiesize};
1716 LASSERT(req != NULL);
1718 if (req->rq_reqmsg->bufcount <= 1) {
1719 /* No intent was provided */
1720 int size = sizeof(struct ldlm_reply);
1721 rc = lustre_pack_reply(req, 1, &size, NULL);
1726 it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
1728 CERROR("Intent missing\n");
1729 RETURN(req->rq_status = -EFAULT);
1732 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
1734 rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3, repsize,
1737 RETURN(req->rq_status = rc);
1739 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1740 intent_set_disposition(rep, DISP_IT_EXECD);
1742 fixup_handle_for_resent_req(req, lock, &lockh);
1744 /* execute policy */
1745 switch ((long)it->opc) {
1747 case IT_CREAT|IT_OPEN:
1748 /* XXX swab here to assert that an mds_open reint
1749 * packet is following */
1750 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
1752 /* We abort the lock if the lookup was negative and
1753 * we did not make it to the OPEN portion */
1754 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
1755 RETURN(ELDLM_LOCK_ABORTED);
1756 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
1757 !intent_disposition(rep, DISP_OPEN_OPEN))
1759 RETURN(ELDLM_LOCK_ABORTED);
1762 getattr_part = MDS_INODELOCK_LOOKUP;
1764 getattr_part |= MDS_INODELOCK_LOOKUP;
1766 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh,
1768 /* FIXME: LDLM can set req->rq_status. MDS sets
1769 policy_res{1,2} with disposition and status.
1770 - replay: returns 0 & req->status is old status
1771 - otherwise: returns req->status */
1772 if (intent_disposition(rep, DISP_LOOKUP_NEG))
1773 rep->lock_policy_res2 = 0;
1774 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
1775 rep->lock_policy_res2)
1776 RETURN(ELDLM_LOCK_ABORTED);
1777 if (req->rq_status != 0) {
1779 rep->lock_policy_res2 = req->rq_status;
1780 RETURN(ELDLM_LOCK_ABORTED);
1784 CERROR("Unhandled intent "LPD64"\n", it->opc);
1788 /* By this point, whatever function we called above must have either
1789 * filled in 'lockh', been an intent replay, or returned an error. We
1790 * want to allow replayed RPCs to not get a lock, since we would just
1791 * drop it below anyways because lock replay is done separately by the
1792 * client afterwards. For regular RPCs we want to give the new lock to
1793 * the client instead of whatever lock it was about to get. */
1794 new_lock = ldlm_handle2lock(&lockh);
1795 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
1798 LASSERT(new_lock != NULL);
1800 /* If we've already given this lock to a client once, then we should
1801 * have no readers or writers. Otherwise, we should have one reader
1802 * _or_ writer ref (which will be zeroed below) before returning the
1803 * lock to a client. */
1804 if (new_lock->l_export == req->rq_export) {
1805 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
1807 LASSERT(new_lock->l_export == NULL);
1808 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
1813 if (new_lock->l_export == req->rq_export) {
1814 /* Already gave this to the client, which means that we
1815 * reconstructed a reply. */
1816 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
1818 RETURN(ELDLM_LOCK_REPLACED);
1821 /* Fixup the lock to be given to the client */
1822 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
1823 new_lock->l_readers = 0;
1824 new_lock->l_writers = 0;
1826 new_lock->l_export = class_export_get(req->rq_export);
1827 list_add(&new_lock->l_export_chain,
1828 &new_lock->l_export->exp_ldlm_data.led_held_locks);
1830 new_lock->l_blocking_ast = lock->l_blocking_ast;
1831 new_lock->l_completion_ast = lock->l_completion_ast;
1833 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
1834 sizeof(lock->l_remote_handle));
1836 new_lock->l_flags &= ~LDLM_FL_LOCAL;
1838 LDLM_LOCK_PUT(new_lock);
1839 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
1841 RETURN(ELDLM_LOCK_REPLACED);
1844 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
1846 struct mds_obd *mds = &obd->u.mds;
1847 struct lprocfs_static_vars lvars;
1851 lprocfs_init_vars(mdt, &lvars);
1852 lprocfs_obd_setup(obd, lvars.obd_vars);
1855 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1856 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1857 mds_handle, "mds", obd->obd_proc_entry);
1859 if (!mds->mds_service) {
1860 CERROR("failed to start service\n");
1861 GOTO(err_lprocfs, rc = -ENOMEM);
1864 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
1867 GOTO(err_thread, rc);
1869 mds->mds_setattr_service =
1870 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1871 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
1872 mds_handle, "mds_setattr",
1873 obd->obd_proc_entry);
1874 if (!mds->mds_setattr_service) {
1875 CERROR("failed to start getattr service\n");
1876 GOTO(err_thread, rc = -ENOMEM);
1879 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
1880 MDT_NUM_THREADS, "ll_mdt_attr");
1882 GOTO(err_thread2, rc);
1884 mds->mds_readpage_service =
1885 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1886 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
1887 mds_handle, "mds_readpage",
1888 obd->obd_proc_entry);
1889 if (!mds->mds_readpage_service) {
1890 CERROR("failed to start readpage service\n");
1891 GOTO(err_thread2, rc = -ENOMEM);
1894 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
1895 MDT_NUM_THREADS, "ll_mdt_rdpg");
1898 GOTO(err_thread3, rc);
1903 ptlrpc_unregister_service(mds->mds_readpage_service);
1905 ptlrpc_unregister_service(mds->mds_setattr_service);
1907 ptlrpc_unregister_service(mds->mds_service);
1909 lprocfs_obd_cleanup(obd);
1913 static int mdt_cleanup(struct obd_device *obd, int flags)
1915 struct mds_obd *mds = &obd->u.mds;
1918 ptlrpc_stop_all_threads(mds->mds_readpage_service);
1919 ptlrpc_unregister_service(mds->mds_readpage_service);
1921 ptlrpc_stop_all_threads(mds->mds_setattr_service);
1922 ptlrpc_unregister_service(mds->mds_setattr_service);
1924 ptlrpc_stop_all_threads(mds->mds_service);
1925 ptlrpc_unregister_service(mds->mds_service);
1927 lprocfs_obd_cleanup(obd);
1932 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
1935 struct obd_device *obd = data;
1938 fid.generation = gen;
1939 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
1942 struct lvfs_callback_ops mds_lvfs_ops = {
1943 l_fid2dentry: mds_lvfs_fid2dentry,
1946 /* use obd ops to offer management infrastructure */
1947 static struct obd_ops mds_obd_ops = {
1948 .o_owner = THIS_MODULE,
1949 .o_connect = mds_connect,
1950 .o_init_export = mds_init_export,
1951 .o_destroy_export = mds_destroy_export,
1952 .o_disconnect = mds_disconnect,
1953 .o_setup = mds_setup,
1954 .o_precleanup = mds_precleanup,
1955 .o_cleanup = mds_cleanup,
1956 .o_postrecov = mds_postrecov,
1957 .o_statfs = mds_obd_statfs,
1958 .o_iocontrol = mds_iocontrol,
1959 .o_create = mds_obd_create,
1960 .o_destroy = mds_obd_destroy,
1961 .o_llog_init = mds_llog_init,
1962 .o_llog_finish = mds_llog_finish,
1963 .o_notify = mds_notify,
1966 static struct obd_ops mdt_obd_ops = {
1967 .o_owner = THIS_MODULE,
1968 .o_setup = mdt_setup,
1969 .o_cleanup = mdt_cleanup,
1972 static int __init mds_init(void)
1974 struct lprocfs_static_vars lvars;
1976 lprocfs_init_vars(mds, &lvars);
1977 class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
1978 lprocfs_init_vars(mdt, &lvars);
1979 class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
1984 static void /*__exit*/ mds_exit(void)
1986 class_unregister_type(LUSTRE_MDS_NAME);
1987 class_unregister_type(LUSTRE_MDT_NAME);
1990 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1991 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
1992 MODULE_LICENSE("GPL");
1994 module_init(mds_init);
1995 module_exit(mds_exit);