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)
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} };
172 res_id.name[0] = de->d_inode->i_ino;
173 res_id.name[1] = de->d_inode->i_generation;
174 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
175 LDLM_PLAIN, NULL, lock_mode, &flags,
176 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
177 NULL, 0, NULL, lockh);
178 if (rc != ELDLM_OK) {
180 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
186 #ifndef DCACHE_DISCONNECTED
187 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
191 /* Look up an entry by inode number. */
192 /* this function ONLY returns valid dget'd dentries with an initialized inode
194 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
195 struct vfsmount **mnt)
198 unsigned long ino = fid->id;
199 __u32 generation = fid->generation;
201 struct dentry *result;
204 RETURN(ERR_PTR(-ESTALE));
206 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
208 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
209 ino, generation, mds->mds_sb);
211 /* under ext3 this is neither supposed to return bad inodes
213 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
217 inode = result->d_inode;
219 RETURN(ERR_PTR(-ENOENT));
221 if (generation && inode->i_generation != generation) {
222 /* we didn't find the right inode.. */
223 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
224 inode->i_ino, (unsigned long)inode->i_nlink,
225 atomic_read(&inode->i_count), inode->i_generation,
228 RETURN(ERR_PTR(-ENOENT));
232 *mnt = mds->mds_vfsmnt;
240 /* Establish a connection to the MDS.
242 * This will set up an export structure for the client to hold state data
243 * about that client, like open files, the last operation number it did
244 * on the server, etc.
246 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
247 struct obd_uuid *cluuid)
249 struct obd_export *exp;
250 struct mds_export_data *med; /* */
251 struct mds_client_data *mcd;
252 int rc, abort_recovery;
255 if (!conn || !obd || !cluuid)
258 /* Check for aborted recovery. */
259 spin_lock_bh(&obd->obd_processing_task_lock);
260 abort_recovery = obd->obd_abort_recovery;
261 spin_unlock_bh(&obd->obd_processing_task_lock);
263 target_abort_recovery(obd);
265 /* XXX There is a small race between checking the list and adding a
266 * new connection for the same UUID, but the real threat (list
267 * corruption when multiple different clients connect) is solved.
269 * There is a second race between adding the export to the list,
270 * and filling in the client data below. Hence skipping the case
271 * of NULL mcd above. We should already be controlling multiple
272 * connects at the client, and we can't hold the spinlock over
273 * memory allocations without risk of deadlocking.
275 rc = class_connect(conn, obd, cluuid);
278 exp = class_conn2export(conn);
280 med = &exp->exp_mds_data;
282 OBD_ALLOC(mcd, sizeof(*mcd));
284 CERROR("mds: out of memory for client data\n");
285 GOTO(out, rc = -ENOMEM);
288 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
291 rc = mds_client_add(obd, &obd->u.mds, med, -1);
296 OBD_FREE(mcd, sizeof(*mcd));
297 class_disconnect(exp, 0);
299 class_export_put(exp);
304 static int mds_init_export(struct obd_export *exp)
306 struct mds_export_data *med = &exp->exp_mds_data;
308 INIT_LIST_HEAD(&med->med_open_head);
309 spin_lock_init(&med->med_open_lock);
313 static int mds_destroy_export(struct obd_export *export)
315 struct mds_export_data *med;
316 struct obd_device *obd = export->exp_obd;
317 struct obd_run_ctxt saved;
321 med = &export->exp_mds_data;
322 target_destroy_export(export);
324 push_ctxt(&saved, &obd->obd_ctxt, NULL);
325 /* Close any open files (which may also cause orphan unlinking). */
326 spin_lock(&med->med_open_lock);
327 while (!list_empty(&med->med_open_head)) {
328 struct list_head *tmp = med->med_open_head.next;
329 struct mds_file_data *mfd =
330 list_entry(tmp, struct mds_file_data, mfd_list);
331 BDEVNAME_DECLARE_STORAGE(btmp);
333 /* bug 1579: fix force-closing for 2.5 */
334 struct dentry *dentry = mfd->mfd_dentry;
336 list_del(&mfd->mfd_list);
337 spin_unlock(&med->med_open_lock);
339 CERROR("force closing client file handle for %*s (%s:%lu)\n",
340 dentry->d_name.len, dentry->d_name.name,
341 ll_bdevname(dentry->d_inode->i_sb, btmp),
342 dentry->d_inode->i_ino);
343 rc = mds_mfd_close(NULL, obd, mfd,
344 !(export->exp_flags & OBD_OPT_FAILOVER));
347 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
348 spin_lock(&med->med_open_lock);
350 spin_unlock(&med->med_open_lock);
351 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
353 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
358 static int mds_disconnect(struct obd_export *export, int flags)
360 unsigned long irqflags;
364 ldlm_cancel_locks_for_export(export);
366 /* complete all outstanding replies */
367 spin_lock_irqsave (&export->exp_lock, irqflags);
368 while (!list_empty (&export->exp_outstanding_replies)) {
369 struct ptlrpc_reply_state *rs =
370 list_entry (export->exp_outstanding_replies.next,
371 struct ptlrpc_reply_state, rs_exp_list);
372 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
374 spin_lock (&svc->srv_lock);
375 list_del_init (&rs->rs_exp_list);
376 ptlrpc_schedule_difficult_reply (rs);
377 spin_unlock (&svc->srv_lock);
379 spin_unlock_irqrestore (&export->exp_lock, irqflags);
381 spin_lock_irqsave(&export->exp_lock, irqflags);
382 export->exp_flags = flags;
383 spin_unlock_irqrestore(&export->exp_lock, irqflags);
385 rc = class_disconnect(export, flags);
389 static int mds_getstatus(struct ptlrpc_request *req)
391 struct mds_obd *mds = mds_req2mds(req);
392 struct mds_body *body;
393 int rc, size = sizeof(*body);
396 rc = lustre_pack_reply(req, 1, &size, NULL);
397 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
398 CERROR("mds: out of memory for message: size=%d\n", size);
399 req->rq_status = -ENOMEM; /* superfluous? */
403 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
404 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
406 /* the last_committed and last_xid fields are filled in for all
407 * replies already - no need to do so here also.
412 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
413 void *data, int flag)
418 if (flag == LDLM_CB_CANCELING) {
419 /* Don't need to do anything here. */
423 /* XXX layering violation! -phil */
424 l_lock(&lock->l_resource->lr_namespace->ns_lock);
425 /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
426 * such that mds_blocking_ast is called just before l_i_p takes the
427 * ns_lock, then by the time we get the lock, we might not be the
428 * correct blocking function anymore. So check, and return early, if
430 if (lock->l_blocking_ast != mds_blocking_ast) {
431 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
435 lock->l_flags |= LDLM_FL_CBPENDING;
436 do_ast = (!lock->l_readers && !lock->l_writers);
437 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
440 struct lustre_handle lockh;
443 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
444 ldlm_lock2handle(lock, &lockh);
445 rc = ldlm_cli_cancel(&lockh);
447 CERROR("ldlm_cli_cancel: %d\n", rc);
449 LDLM_DEBUG(lock, "Lock still has references, will be "
455 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
456 * Call with lock=0 if the caller has already taken the i_sem. */
457 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
458 struct mds_body *body, struct inode *inode, int lock)
460 struct mds_obd *mds = &obd->u.mds;
466 lmm = lustre_msg_buf(msg, offset, 0);
468 /* Some problem with getting eadata when I sized the reply
470 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
474 lmm_size = msg->buflens[offset];
476 /* I don't really like this, but it is a sanity check on the client
477 * MD request. However, if the client doesn't know how much space
478 * to reserve for the MD, it shouldn't be bad to have too much space.
480 if (lmm_size > mds->mds_max_mdsize) {
481 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
482 inode->i_ino, lmm_size, mds->mds_max_mdsize);
488 rc = fsfilt_get_md(obd, inode, lmm, lmm_size);
492 CERROR("Error %d reading eadata for ino %lu\n",
496 rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
500 body->valid |= OBD_MD_FLEASIZE;
501 body->eadatasize = lmm_size;
508 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
509 struct ptlrpc_request *req,
510 struct mds_body *reqbody, int reply_off)
512 struct mds_body *body;
513 struct inode *inode = dentry->d_inode;
520 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
521 LASSERT(body != NULL); /* caller prepped reply */
523 mds_pack_inode2fid(&body->fid1, inode);
524 mds_pack_inode2body(body, inode);
526 if (S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
527 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
530 /* If we have LOV EA data, the OST holds size, atime, mtime */
531 if (!(body->valid & OBD_MD_FLEASIZE))
532 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
533 OBD_MD_FLATIME | OBD_MD_FLMTIME);
534 } else if (S_ISLNK(inode->i_mode) &&
535 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
536 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
539 LASSERT (symname != NULL); /* caller prepped reply */
540 len = req->rq_repmsg->buflens[reply_off + 1];
542 rc = inode->i_op->readlink(dentry, symname, len);
544 CERROR("readlink failed: %d\n", rc);
545 } else if (rc != len - 1) {
546 CERROR ("Unexpected readlink rc %d: expecting %d\n",
550 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
551 body->valid |= OBD_MD_LINKNAME;
552 body->eadatasize = rc + 1;
553 symname[rc] = 0; /* NULL terminate */
561 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
564 struct mds_obd *mds = mds_req2mds(req);
565 struct mds_body *body;
566 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
569 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
570 LASSERT(body != NULL); /* checked by caller */
571 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
573 if (S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) {
576 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
578 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
582 CERROR("error getting inode %lu MD: rc = %d\n",
585 } else if (rc > mds->mds_max_mdsize) {
587 CERROR("MD size %d larger than maximum possible %u\n",
588 rc, mds->mds_max_mdsize);
593 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
594 if (inode->i_size + 1 != body->eadatasize)
595 CERROR("symlink size: %Lu, reply space: %d\n",
596 inode->i_size + 1, body->eadatasize);
597 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
599 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
600 inode->i_size + 1, body->eadatasize);
603 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
604 CERROR("failed MDS_GETATTR_PACK test\n");
605 req->rq_status = -ENOMEM;
606 GOTO(out, rc = -ENOMEM);
609 rc = lustre_pack_reply(req, bufcount, size, NULL);
611 CERROR("out of memory\n");
612 GOTO(out, req->rq_status = rc);
620 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
621 struct lustre_handle *child_lockh)
623 struct obd_device *obd = req->rq_export->exp_obd;
624 struct ldlm_reply *rep = NULL;
625 struct obd_run_ctxt saved;
626 struct mds_body *body;
627 struct dentry *dparent = NULL, *dchild = NULL;
629 struct lustre_handle parent_lockh;
631 int rc = 0, cleanup_phase = 0, resent_req = 0;
635 LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
637 /* Swab now, before anyone looks inside the request */
639 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
640 lustre_swab_mds_body);
642 CERROR("Can't swab mds_body\n");
643 GOTO(cleanup, rc = -EFAULT);
646 LASSERT_REQSWAB(req, offset + 1);
647 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
649 CERROR("Can't unpack name\n");
650 GOTO(cleanup, rc = -EFAULT);
652 namesize = req->rq_reqmsg->buflens[offset + 1];
654 LASSERT (offset == 0 || offset == 2);
655 /* if requests were at offset 2, the getattr reply goes back at 1 */
657 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
661 uc.ouc_fsuid = body->fsuid;
662 uc.ouc_fsgid = body->fsgid;
663 uc.ouc_cap = body->capability;
664 uc.ouc_suppgid1 = body->suppgid;
665 uc.ouc_suppgid2 = -1;
666 push_ctxt(&saved, &obd->obd_ctxt, &uc);
667 cleanup_phase = 1; /* kernel context */
668 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
670 /* FIXME: handle raw lookup */
672 if (body->valid == OBD_MD_FLID) {
673 struct mds_body *mds_reply;
674 int size = sizeof(*mds_reply);
676 // The user requested ONLY the inode number, so do a raw lookup
677 rc = lustre_pack_reply(req, 1, &size, NULL);
679 CERROR("out of memory\n");
683 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
685 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
687 mds_reply->fid1.id = inum;
688 mds_reply->valid = OBD_MD_FLID;
693 if (child_lockh->cookie != 0) {
694 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
698 if (resent_req == 0) {
699 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
700 &parent_lockh, &dparent,
701 LCK_PR, name, namesize,
702 child_lockh, &dchild, LCK_PR);
706 struct ldlm_lock *granted_lock;
707 struct ll_fid child_fid;
708 struct ldlm_resource *res;
709 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
710 granted_lock = ldlm_handle2lock(child_lockh);
711 LASSERT(granted_lock);
713 res = granted_lock->l_resource;
714 child_fid.id = res->lr_name.name[0];
715 child_fid.generation = res->lr_name.name[1];
716 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
718 LDLM_LOCK_PUT(granted_lock);
721 cleanup_phase = 2; /* dchild, dparent, locks */
723 if (dchild->d_inode == NULL) {
724 intent_set_disposition(rep, DISP_LOOKUP_NEG);
725 /* in the intent case, the policy clears this error:
726 the disposition is enough */
727 GOTO(cleanup, rc = -ENOENT);
729 intent_set_disposition(rep, DISP_LOOKUP_POS);
732 if (req->rq_repmsg == NULL) {
733 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
735 CERROR ("mds_getattr_pack_msg: %d\n", rc);
740 rc = mds_getattr_internal(obd, dchild, req, body, offset);
741 GOTO(cleanup, rc); /* returns the lock to the client */
744 switch (cleanup_phase) {
746 if (resent_req == 0) {
747 if (rc && dchild->d_inode)
748 ldlm_lock_decref(child_lockh, LCK_PR);
749 ldlm_lock_decref(&parent_lockh, LCK_PR);
754 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
760 static int mds_getattr(int offset, struct ptlrpc_request *req)
762 struct mds_obd *mds = mds_req2mds(req);
763 struct obd_device *obd = req->rq_export->exp_obd;
764 struct obd_run_ctxt saved;
766 struct mds_body *body;
771 body = lustre_swab_reqbuf (req, offset, sizeof (*body),
772 lustre_swab_mds_body);
774 CERROR ("Can't unpack body\n");
778 uc.ouc_fsuid = body->fsuid;
779 uc.ouc_fsgid = body->fsgid;
780 uc.ouc_cap = body->capability;
781 push_ctxt(&saved, &obd->obd_ctxt, &uc);
782 de = mds_fid2dentry(mds, &body->fid1, NULL);
784 rc = req->rq_status = -ENOENT;
785 GOTO(out_pop, PTR_ERR(de));
788 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
790 CERROR ("mds_getattr_pack_msg: %d\n", rc);
794 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
799 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
804 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
805 unsigned long max_age)
809 spin_lock(&obd->obd_osfs_lock);
810 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
812 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
813 spin_unlock(&obd->obd_osfs_lock);
818 static int mds_statfs(struct ptlrpc_request *req)
820 struct obd_device *obd = req->rq_export->exp_obd;
821 int rc, size = sizeof(struct obd_statfs);
824 rc = lustre_pack_reply(req, 1, &size, NULL);
825 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
826 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
830 /* We call this so that we can cache a bit - 1 jiffie worth */
831 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
834 CERROR("mds_obd_statfs failed: rc %d\n", rc);
844 static int mds_sync(struct ptlrpc_request *req)
846 struct obd_device *obd = req->rq_export->exp_obd;
847 struct mds_obd *mds = &obd->u.mds;
848 struct mds_body *body;
849 int rc, size = sizeof(*body);
852 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
854 GOTO(out, rc = -EPROTO);
856 rc = lustre_pack_reply(req, 1, &size, NULL);
857 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
858 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
862 if (body->fid1.id == 0) {
863 /* a fid of zero is taken to mean "sync whole filesystem" */
864 rc = fsfilt_sync(obd, mds->mds_sb);
868 /* just any file to grab fsync method - "file" arg unused */
869 struct file *file = mds->mds_rcvd_filp;
872 de = mds_fid2dentry(mds, &body->fid1, NULL);
874 GOTO(out, rc = PTR_ERR(de));
876 rc = file->f_op->fsync(NULL, de, 1);
881 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
882 mds_pack_inode2fid(&body->fid1, de->d_inode);
883 mds_pack_inode2body(body, de->d_inode);
890 /* mds_readpage does not take a DLM lock on the inode, because the client must
891 * already have a PR lock.
893 * If we were to take another one here, a deadlock will result, if another
894 * thread is already waiting for a PW lock. */
895 static int mds_readpage(struct ptlrpc_request *req)
897 struct obd_device *obd = req->rq_export->exp_obd;
898 struct vfsmount *mnt;
901 struct mds_body *body, *repbody;
902 struct obd_run_ctxt saved;
903 int rc, size = sizeof(*repbody);
907 rc = lustre_pack_reply(req, 1, &size, NULL);
908 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
909 CERROR("mds: out of memory\n");
910 GOTO(out, rc = -ENOMEM);
913 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
915 GOTO (out, rc = -EFAULT);
917 uc.ouc_fsuid = body->fsuid;
918 uc.ouc_fsgid = body->fsgid;
919 uc.ouc_cap = body->capability;
920 push_ctxt(&saved, &obd->obd_ctxt, &uc);
921 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
923 GOTO(out_pop, rc = PTR_ERR(de));
925 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
927 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
928 /* note: in case of an error, dentry_open puts dentry */
930 GOTO(out_pop, rc = PTR_ERR(file));
932 /* body->size is actually the offset -eeb */
933 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
934 CERROR("offset "LPU64" not on a block boundary of %lu\n",
935 body->size, de->d_inode->i_blksize);
936 GOTO(out_file, rc = -EFAULT);
939 /* body->nlink is actually the #bytes to read -eeb */
940 if (body->nlink & (de->d_inode->i_blksize - 1)) {
941 CERROR("size %u is not multiple of blocksize %lu\n",
942 body->nlink, de->d_inode->i_blksize);
943 GOTO(out_file, rc = -EFAULT);
946 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
947 repbody->size = file->f_dentry->d_inode->i_size;
948 repbody->valid = OBD_MD_FLSIZE;
950 /* to make this asynchronous make sure that the handling function
951 doesn't send a reply when this function completes. Instead a
952 callback function would send the reply */
953 /* body->size is actually the offset -eeb */
954 rc = mds_sendpage(req, file, body->size, body->nlink);
959 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
965 int mds_reint(struct ptlrpc_request *req, int offset,
966 struct lustre_handle *lockh)
968 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
971 OBD_ALLOC(rec, sizeof(*rec));
975 rc = mds_update_unpack(req, offset, rec);
976 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
977 CERROR("invalid record\n");
978 GOTO(out, req->rq_status = -EINVAL);
980 /* rc will be used to interrupt a for loop over multiple records */
981 rc = mds_reint_rec(rec, offset, req, lockh);
983 OBD_FREE(rec, sizeof(*rec));
987 static int mds_filter_recovery_request(struct ptlrpc_request *req,
988 struct obd_device *obd, int *process)
990 switch (req->rq_reqmsg->opc) {
991 case MDS_CONNECT: /* This will never get here, but for completeness. */
992 case OST_CONNECT: /* This will never get here, but for completeness. */
999 case MDS_SYNC: /* used in unmounting */
1003 *process = target_queue_recovery_request(req, obd);
1007 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1009 /* XXX what should we set rq_status to here? */
1010 req->rq_status = -EAGAIN;
1011 RETURN(ptlrpc_error(req));
1015 static char *reint_names[] = {
1016 [REINT_SETATTR] "setattr",
1017 [REINT_CREATE] "create",
1018 [REINT_LINK] "link",
1019 [REINT_UNLINK] "unlink",
1020 [REINT_RENAME] "rename",
1021 [REINT_OPEN] "open",
1024 int mds_handle(struct ptlrpc_request *req)
1026 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1028 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1029 struct obd_device *obd = NULL;
1032 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1034 LASSERT(current->journal_info == NULL);
1035 /* XXX identical to OST */
1036 if (req->rq_reqmsg->opc != MDS_CONNECT) {
1037 struct mds_export_data *med;
1038 int recovering, abort_recovery;
1040 if (req->rq_export == NULL) {
1041 CERROR("lustre_mds: operation %d on unconnected MDS\n",
1042 req->rq_reqmsg->opc);
1043 req->rq_status = -ENOTCONN;
1044 GOTO(out, rc = -ENOTCONN);
1047 med = &req->rq_export->exp_mds_data;
1048 obd = req->rq_export->exp_obd;
1051 /* Check for aborted recovery. */
1052 spin_lock_bh(&obd->obd_processing_task_lock);
1053 abort_recovery = obd->obd_abort_recovery;
1054 recovering = obd->obd_recovering;
1055 spin_unlock_bh(&obd->obd_processing_task_lock);
1056 if (abort_recovery) {
1057 target_abort_recovery(obd);
1058 } else if (recovering) {
1059 rc = mds_filter_recovery_request(req, obd,
1061 if (rc || !should_process)
1066 switch (req->rq_reqmsg->opc) {
1068 DEBUG_REQ(D_INODE, req, "connect");
1069 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1070 rc = target_handle_connect(req, mds_handle);
1072 /* Now that we have an export, set mds. */
1073 mds = mds_req2mds(req);
1076 case MDS_DISCONNECT:
1077 DEBUG_REQ(D_INODE, req, "disconnect");
1078 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1079 rc = target_handle_disconnect(req);
1080 req->rq_status = rc; /* superfluous? */
1084 DEBUG_REQ(D_INODE, req, "getstatus");
1085 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1086 rc = mds_getstatus(req);
1090 DEBUG_REQ(D_INODE, req, "getattr");
1091 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1092 rc = mds_getattr(0, req);
1095 case MDS_GETATTR_NAME: {
1096 struct lustre_handle lockh;
1097 DEBUG_REQ(D_INODE, req, "getattr_name");
1098 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1100 /* If this request gets a reconstructed reply, we won't be
1101 * acquiring any new locks in mds_getattr_name, so we don't
1105 rc = mds_getattr_name(0, req, &lockh);
1106 /* this non-intent call (from an ioctl) is special */
1107 req->rq_status = rc;
1108 if (rc == 0 && lockh.cookie)
1109 ldlm_lock_decref(&lockh, LCK_PR);
1113 DEBUG_REQ(D_INODE, req, "statfs");
1114 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1115 rc = mds_statfs(req);
1119 DEBUG_REQ(D_INODE, req, "readpage");
1120 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1121 rc = mds_readpage(req);
1123 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1124 if (req->rq_reply_state) {
1125 lustre_free_reply_state (req->rq_reply_state);
1126 req->rq_reply_state = NULL;
1134 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1136 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1137 mds->mds_max_cookiesize};
1140 /* NB only peek inside req now; mds_reint() will swab it */
1142 CERROR ("Can't inspect opcode\n");
1147 if (lustre_msg_swabbed (req->rq_reqmsg))
1150 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1151 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1152 reint_names[opc] == NULL) ? reint_names[opc] :
1155 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1157 if (opc == REINT_UNLINK)
1159 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1164 rc = lustre_pack_reply(req, bufcount, size, NULL);
1168 rc = mds_reint(req, 0, NULL);
1169 fail = OBD_FAIL_MDS_REINT_NET_REP;
1174 DEBUG_REQ(D_INODE, req, "close");
1175 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1176 rc = mds_close(req);
1179 case MDS_DONE_WRITING:
1180 DEBUG_REQ(D_INODE, req, "done_writing");
1181 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1182 rc = mds_done_writing(req);
1186 DEBUG_REQ(D_INODE, req, "pin");
1187 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1192 DEBUG_REQ(D_INODE, req, "sync");
1193 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1198 DEBUG_REQ(D_INODE, req, "ping");
1199 rc = target_handle_ping(req);
1202 case OBD_LOG_CANCEL:
1203 CDEBUG(D_INODE, "log cancel\n");
1204 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1205 rc = -ENOTSUPP; /* la la la */
1209 DEBUG_REQ(D_INODE, req, "enqueue");
1210 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1211 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1212 ldlm_server_blocking_ast, NULL);
1215 DEBUG_REQ(D_INODE, req, "convert");
1216 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1217 rc = ldlm_handle_convert(req);
1219 case LDLM_BL_CALLBACK:
1220 case LDLM_CP_CALLBACK:
1221 DEBUG_REQ(D_INODE, req, "callback");
1222 CERROR("callbacks should not happen on MDS\n");
1224 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1226 case LLOG_ORIGIN_HANDLE_CREATE:
1227 DEBUG_REQ(D_INODE, req, "llog_init");
1228 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1229 rc = llog_origin_handle_create(req);
1231 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1232 DEBUG_REQ(D_INODE, req, "llog next block");
1233 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1234 rc = llog_origin_handle_next_block(req);
1236 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1237 DEBUG_REQ(D_INODE, req, "llog read header");
1238 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1239 rc = llog_origin_handle_read_header(req);
1241 case LLOG_ORIGIN_HANDLE_CLOSE:
1242 DEBUG_REQ(D_INODE, req, "llog close");
1243 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1244 rc = llog_origin_handle_close(req);
1247 DEBUG_REQ(D_INODE, req, "llog catinfo");
1248 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1249 rc = llog_catinfo(req);
1252 req->rq_status = -ENOTSUPP;
1253 rc = ptlrpc_error(req);
1257 LASSERT(current->journal_info == NULL);
1261 /* If we're DISCONNECTing, the mds_export_data is already freed */
1262 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1263 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1264 struct obd_device *obd = list_entry(mds, struct obd_device,
1266 req->rq_repmsg->last_xid =
1267 le64_to_cpu(med->med_mcd->mcd_last_xid);
1269 if (!obd->obd_no_transno) {
1270 req->rq_repmsg->last_committed =
1271 obd->obd_last_committed;
1273 DEBUG_REQ(D_IOCTL, req,
1274 "not sending last_committed update");
1276 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1278 mds->mds_last_transno, obd->obd_last_committed,
1283 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1284 if (obd && obd->obd_recovering) {
1285 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1286 return target_queue_final_reply(req, rc);
1288 /* Lost a race with recovery; let the error path DTRT. */
1289 rc = req->rq_status = -ENOTCONN;
1292 target_send_reply(req, rc, fail);
1296 /* Update the server data on disk. This stores the new mount_count and
1297 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1298 * then the server last_rcvd value may be less than that of the clients.
1299 * This will alert us that we may need to do client recovery.
1301 * Also assumes for mds_last_transno that we are not modifying it (no locking).
1303 int mds_update_server_data(struct obd_device *obd, int force_sync)
1305 struct mds_obd *mds = &obd->u.mds;
1306 struct mds_server_data *msd = mds->mds_server_data;
1307 struct file *filp = mds->mds_rcvd_filp;
1308 struct obd_run_ctxt saved;
1313 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1314 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1316 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1317 mds->mds_mount_count, mds->mds_last_transno);
1318 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1320 CERROR("error writing MDS server data: rc = %d\n", rc);
1321 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1327 /* mount the file system (secretly) */
1328 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1330 struct lustre_cfg* lcfg = buf;
1331 struct mds_obd *mds = &obd->u.mds;
1332 struct vfsmount *mnt;
1337 dev_clear_rdonly(2);
1339 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1340 RETURN(rc = -EINVAL);
1342 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1343 if (IS_ERR(obd->obd_fsops))
1344 RETURN(rc = PTR_ERR(obd->obd_fsops));
1346 if (!(page = __get_free_page(GFP_KERNEL)))
1349 memset((void *)page, 0, PAGE_SIZE);
1350 sprintf((char *)page, "iopen_nopriv");
1352 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1353 lcfg->lcfg_inlbuf1, (void *)page);
1357 CERROR("do_kern_mount failed: rc = %d\n", rc);
1361 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1363 sema_init(&mds->mds_orphan_recovery_sem, 1);
1364 sema_init(&mds->mds_epoch_sem, 1);
1365 spin_lock_init(&mds->mds_transno_lock);
1366 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1367 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1368 atomic_set(&mds->mds_open_count, 0);
1370 obd->obd_namespace = ldlm_namespace_new("mds_server",
1371 LDLM_NAMESPACE_SERVER);
1372 if (obd->obd_namespace == NULL) {
1373 mds_cleanup(obd, 0);
1374 GOTO(err_put, rc = -ENOMEM);
1376 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1378 rc = mds_fs_setup(obd, mnt);
1380 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1384 rc = llog_start_commit_thread();
1389 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1392 generate_random_uuid(uuid);
1393 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1395 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1396 if (mds->mds_profile == NULL)
1397 GOTO(err_fs, rc = -ENOMEM);
1399 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1400 lcfg->lcfg_inllen3);
1404 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1405 "mds_ldlm_client", &obd->obd_ldlm_client);
1406 obd->obd_replayable = 1;
1408 rc = mds_postsetup(obd);
1414 /* No extra cleanup needed for llog_init_commit_thread() */
1415 mds_fs_cleanup(obd, 0);
1417 ldlm_namespace_free(obd->obd_namespace, 0);
1418 obd->obd_namespace = NULL;
1421 mntput(mds->mds_vfsmnt);
1425 fsfilt_put_ops(obd->obd_fsops);
1429 static int mds_postsetup(struct obd_device *obd)
1431 struct mds_obd *mds = &obd->u.mds;
1436 rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
1441 if (mds->mds_profile) {
1442 struct obd_run_ctxt saved;
1443 struct lustre_profile *lprof;
1444 struct config_llog_instance cfg;
1446 cfg.cfg_instance = NULL;
1447 cfg.cfg_uuid = mds->mds_lov_uuid;
1448 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1449 rc = class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1450 mds->mds_profile, &cfg);
1451 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1455 lprof = class_get_profile(mds->mds_profile);
1456 if (lprof == NULL) {
1457 CERROR("No profile found: %s\n", mds->mds_profile);
1458 GOTO(err_cleanup, rc = -ENOENT);
1460 rc = mds_lov_connect(obd, lprof->lp_osc);
1462 GOTO(err_cleanup, rc);
1470 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1474 static int mds_postrecov(struct obd_device *obd)
1479 LASSERT(!obd->obd_recovering);
1480 LASSERT(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT) != NULL);
1482 rc = llog_connect(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT),
1483 obd->u.mds.mds_lov_desc.ld_tgt_count, NULL, NULL);
1485 CERROR("faild at llog_origin_connect: %d\n", rc);
1488 rc = mds_cleanup_orphans(obd);
1490 rc2 = mds_lov_set_nextid(obd);
1496 int mds_lov_clean(struct obd_device *obd)
1498 struct mds_obd *mds = &obd->u.mds;
1500 if (mds->mds_profile) {
1502 struct config_llog_instance cfg;
1503 struct obd_run_ctxt saved;
1504 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1506 OBD_ALLOC(cln_prof, len);
1507 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1509 cfg.cfg_instance = NULL;
1510 cfg.cfg_uuid = mds->mds_lov_uuid;
1512 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1513 class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1515 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1517 OBD_FREE(cln_prof, len);
1518 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1519 mds->mds_profile = NULL;
1524 static int mds_precleanup(struct obd_device *obd, int flags)
1529 mds_lov_disconnect(obd, flags);
1531 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1535 static int mds_cleanup(struct obd_device *obd, int flags)
1537 struct mds_obd *mds = &obd->u.mds;
1540 if (mds->mds_sb == NULL)
1543 mds_update_server_data(obd, 1);
1544 if (mds->mds_lov_objids != NULL) {
1545 OBD_FREE(mds->mds_lov_objids,
1546 mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1548 mds_fs_cleanup(obd, flags);
1552 /* 2 seems normal on mds, (may_umount() also expects 2
1553 fwiw), but we only see 1 at this point in obdfilter. */
1554 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1555 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1556 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1558 mntput(mds->mds_vfsmnt);
1562 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1564 spin_lock_bh(&obd->obd_processing_task_lock);
1565 if (obd->obd_recovering) {
1566 target_cancel_recovery_timer(obd);
1567 obd->obd_recovering = 0;
1569 spin_unlock_bh(&obd->obd_processing_task_lock);
1572 dev_clear_rdonly(2);
1573 fsfilt_put_ops(obd->obd_fsops);
1578 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
1579 struct ldlm_lock *new_lock,
1580 struct lustre_handle *lockh)
1582 struct obd_export *exp = req->rq_export;
1583 struct obd_device *obd = exp->exp_obd;
1584 struct ldlm_request *dlmreq =
1585 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
1586 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
1587 struct list_head *iter;
1589 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
1592 l_lock(&obd->obd_namespace->ns_lock);
1593 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
1594 struct ldlm_lock *lock;
1595 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1596 if (lock == new_lock)
1598 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
1599 lockh->cookie = lock->l_handle.h_cookie;
1600 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
1602 l_unlock(&obd->obd_namespace->ns_lock);
1606 l_unlock(&obd->obd_namespace->ns_lock);
1608 /* This remote handle isn't enqueued, so we never received or
1609 * processed this request. Clear MSG_RESENT, because it can
1610 * be handled like any normal request now. */
1612 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1614 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
1618 int intent_disposition(struct ldlm_reply *rep, int flag)
1622 return (rep->lock_policy_res1 & flag);
1625 void intent_set_disposition(struct ldlm_reply *rep, int flag)
1629 rep->lock_policy_res1 |= flag;
1632 static int mds_intent_policy(struct ldlm_namespace *ns,
1633 struct ldlm_lock **lockp, void *req_cookie,
1634 ldlm_mode_t mode, int flags, void *data)
1636 struct ptlrpc_request *req = req_cookie;
1637 struct ldlm_lock *lock = *lockp;
1638 struct ldlm_intent *it;
1639 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1640 struct ldlm_reply *rep;
1641 struct lustre_handle lockh = { 0 };
1642 struct ldlm_lock *new_lock;
1643 int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
1644 sizeof(struct mds_body),
1645 mds->mds_max_mdsize,
1646 mds->mds_max_cookiesize};
1649 LASSERT(req != NULL);
1651 if (req->rq_reqmsg->bufcount <= 1) {
1652 /* No intent was provided */
1653 int size = sizeof(struct ldlm_reply);
1654 rc = lustre_pack_reply(req, 1, &size, NULL);
1659 it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
1661 CERROR("Intent missing\n");
1662 RETURN(req->rq_status = -EFAULT);
1665 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
1667 rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3, repsize,
1670 RETURN(req->rq_status = rc);
1672 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1673 intent_set_disposition(rep, DISP_IT_EXECD);
1675 fixup_handle_for_resent_req(req, lock, &lockh);
1677 /* execute policy */
1678 switch ((long)it->opc) {
1680 case IT_CREAT|IT_OPEN:
1681 /* XXX swab here to assert that an mds_open reint
1682 * packet is following */
1683 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
1685 /* We abort the lock if the lookup was negative and
1686 * we did not make it to the OPEN portion */
1687 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
1688 RETURN(ELDLM_LOCK_ABORTED);
1689 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
1690 !intent_disposition(rep, DISP_OPEN_OPEN))
1692 RETURN(ELDLM_LOCK_ABORTED);
1697 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh);
1698 /* FIXME: LDLM can set req->rq_status. MDS sets
1699 policy_res{1,2} with disposition and status.
1700 - replay: returns 0 & req->status is old status
1701 - otherwise: returns req->status */
1702 if (intent_disposition(rep, DISP_LOOKUP_NEG))
1703 rep->lock_policy_res2 = 0;
1704 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
1705 rep->lock_policy_res2)
1706 RETURN(ELDLM_LOCK_ABORTED);
1707 if (req->rq_status != 0) {
1709 rep->lock_policy_res2 = req->rq_status;
1710 RETURN(ELDLM_LOCK_ABORTED);
1714 CERROR("Unhandled intent "LPD64"\n", it->opc);
1718 /* By this point, whatever function we called above must have either
1719 * filled in 'lockh', been an intent replay, or returned an error. We
1720 * want to allow replayed RPCs to not get a lock, since we would just
1721 * drop it below anyways because lock replay is done separately by the
1722 * client afterwards. For regular RPCs we want to give the new lock to
1723 * the client instead of whatever lock it was about to get. */
1724 new_lock = ldlm_handle2lock(&lockh);
1725 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
1728 LASSERT(new_lock != NULL);
1730 /* If we've already given this lock to a client once, then we should
1731 * have no readers or writers. Otherwise, we should have one reader
1732 * _or_ writer ref (which will be zeroed below) before returning the
1733 * lock to a client. */
1734 if (new_lock->l_export == req->rq_export) {
1735 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
1737 LASSERT(new_lock->l_export == NULL);
1738 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
1743 if (new_lock->l_export == req->rq_export) {
1744 /* Already gave this to the client, which means that we
1745 * reconstructed a reply. */
1746 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
1748 RETURN(ELDLM_LOCK_REPLACED);
1751 /* Fixup the lock to be given to the client */
1752 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
1753 new_lock->l_readers = 0;
1754 new_lock->l_writers = 0;
1756 new_lock->l_export = class_export_get(req->rq_export);
1757 list_add(&new_lock->l_export_chain,
1758 &new_lock->l_export->exp_ldlm_data.led_held_locks);
1760 new_lock->l_blocking_ast = lock->l_blocking_ast;
1761 new_lock->l_completion_ast = lock->l_completion_ast;
1763 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
1764 sizeof(lock->l_remote_handle));
1766 new_lock->l_flags &= ~LDLM_FL_LOCAL;
1768 LDLM_LOCK_PUT(new_lock);
1769 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
1771 RETURN(ELDLM_LOCK_REPLACED);
1774 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1776 struct lprocfs_static_vars lvars;
1778 lprocfs_init_multi_vars(0, &lvars);
1779 return lprocfs_obd_attach(dev, lvars.obd_vars);
1782 int mds_detach(struct obd_device *dev)
1784 return lprocfs_obd_detach(dev);
1787 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
1789 struct lprocfs_static_vars lvars;
1791 lprocfs_init_multi_vars(1, &lvars);
1792 return lprocfs_obd_attach(dev, lvars.obd_vars);
1795 int mdt_detach(struct obd_device *dev)
1797 return lprocfs_obd_detach(dev);
1800 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1802 struct mds_obd *mds = &obddev->u.mds;
1807 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1808 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1810 obddev->obd_proc_entry);
1812 if (!mds->mds_service) {
1813 CERROR("failed to start service\n");
1814 RETURN(rc = -ENOMEM);
1817 rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
1820 GOTO(err_thread, rc);
1822 mds->mds_setattr_service =
1823 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1824 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
1825 mds_handle, "mds_setattr",
1826 obddev->obd_proc_entry);
1827 if (!mds->mds_setattr_service) {
1828 CERROR("failed to start getattr service\n");
1829 GOTO(err_thread, rc = -ENOMEM);
1832 rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
1833 MDT_NUM_THREADS, "ll_mdt_attr");
1835 GOTO(err_thread2, rc);
1837 mds->mds_readpage_service =
1838 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1839 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
1840 mds_handle, "mds_readpage",
1841 obddev->obd_proc_entry);
1842 if (!mds->mds_readpage_service) {
1843 CERROR("failed to start readpage service\n");
1844 GOTO(err_thread2, rc = -ENOMEM);
1847 rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
1848 MDT_NUM_THREADS, "ll_mdt_rdpg");
1851 GOTO(err_thread3, rc);
1856 ptlrpc_unregister_service(mds->mds_readpage_service);
1858 ptlrpc_unregister_service(mds->mds_setattr_service);
1860 ptlrpc_unregister_service(mds->mds_service);
1865 static int mdt_cleanup(struct obd_device *obddev, int flags)
1867 struct mds_obd *mds = &obddev->u.mds;
1870 ptlrpc_stop_all_threads(mds->mds_readpage_service);
1871 ptlrpc_unregister_service(mds->mds_readpage_service);
1873 ptlrpc_stop_all_threads(mds->mds_setattr_service);
1874 ptlrpc_unregister_service(mds->mds_setattr_service);
1876 ptlrpc_stop_all_threads(mds->mds_service);
1877 ptlrpc_unregister_service(mds->mds_service);
1882 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
1884 struct obd_device *obd = data;
1887 fid.generation = gen;
1888 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
1891 struct lvfs_callback_ops mds_lvfs_ops = {
1892 l_fid2dentry: mds_lvfs_fid2dentry,
1895 /* use obd ops to offer management infrastructure */
1896 static struct obd_ops mds_obd_ops = {
1897 o_owner: THIS_MODULE,
1898 o_attach: mds_attach,
1899 o_detach: mds_detach,
1900 o_connect: mds_connect,
1901 o_init_export: mds_init_export,
1902 o_destroy_export: mds_destroy_export,
1903 o_disconnect: mds_disconnect,
1905 o_precleanup: mds_precleanup,
1906 o_cleanup: mds_cleanup,
1907 o_postrecov: mds_postrecov,
1908 o_statfs: mds_obd_statfs,
1909 o_iocontrol: mds_iocontrol,
1910 o_create: mds_obd_create,
1911 o_destroy: mds_obd_destroy,
1912 o_llog_init: mds_llog_init,
1913 o_llog_finish: mds_llog_finish,
1914 o_notify: mds_notify,
1917 static struct obd_ops mdt_obd_ops = {
1918 o_owner: THIS_MODULE,
1919 o_attach: mdt_attach,
1920 o_detach: mdt_detach,
1922 o_cleanup: mdt_cleanup,
1925 static int __init mds_init(void)
1927 struct lprocfs_static_vars lvars;
1929 lprocfs_init_multi_vars(0, &lvars);
1930 class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
1931 lprocfs_init_multi_vars(1, &lvars);
1932 class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
1937 static void /*__exit*/ mds_exit(void)
1939 class_unregister_type(LUSTRE_MDS_NAME);
1940 class_unregister_type(LUSTRE_MDT_NAME);
1943 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1944 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
1945 MODULE_LICENSE("GPL");
1947 module_init(mds_init);
1948 module_exit(mds_exit);