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_postsetup(struct obd_device *obd);
60 static int mds_cleanup(struct obd_device *obd, int flags);
62 /* Assumes caller has already pushed into the kernel filesystem context */
63 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
64 loff_t offset, int count)
66 struct ptlrpc_bulk_desc *desc;
67 struct l_wait_info lwi;
69 int rc = 0, npages, i, tmpcount, tmpsize = 0;
72 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
74 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
75 OBD_ALLOC(pages, sizeof(*pages) * npages);
77 GOTO(out, rc = -ENOMEM);
79 desc = ptlrpc_prep_bulk_exp (req, 1, BULK_PUT_SOURCE, MDS_BULK_PORTAL);
81 GOTO(out_free, rc = -ENOMEM);
83 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
84 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
86 pages[i] = alloc_pages(GFP_KERNEL, 0);
88 GOTO(cleanup_buf, rc = -ENOMEM);
90 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
93 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
94 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
95 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
96 tmpsize, offset, file->f_dentry->d_inode->i_ino,
97 file->f_dentry->d_inode->i_size);
99 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
100 page_address(pages[i]), tmpsize, &offset);
103 GOTO(cleanup_buf, rc = -EIO);
106 LASSERT(desc->bd_nob == count);
108 rc = ptlrpc_start_bulk_transfer(desc);
110 GOTO(cleanup_buf, rc);
112 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
113 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
114 OBD_FAIL_MDS_SENDPAGE, rc);
115 GOTO(abort_bulk, rc);
118 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
119 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
120 LASSERT (rc == 0 || rc == -ETIMEDOUT);
123 if (desc->bd_success &&
124 desc->bd_nob_transferred == count)
125 GOTO(cleanup_buf, rc);
127 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
130 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
131 (rc == -ETIMEDOUT) ? "timeout" : "network error",
132 desc->bd_nob_transferred, count,
133 req->rq_export->exp_client_uuid.uuid,
134 req->rq_export->exp_connection->c_remote_uuid.uuid);
136 ptlrpc_fail_export(req->rq_export);
140 ptlrpc_abort_bulk (desc);
142 for (i = 0; i < npages; i++)
144 __free_pages(pages[i], 0);
146 ptlrpc_free_bulk(desc);
148 OBD_FREE(pages, sizeof(*pages) * npages);
153 /* only valid locked dentries or errors should be returned */
154 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
155 struct vfsmount **mnt, int lock_mode,
156 struct lustre_handle *lockh,
157 char *name, int namelen)
159 struct mds_obd *mds = &obd->u.mds;
160 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
161 struct ldlm_res_id res_id = { .name = {0} };
168 res_id.name[0] = de->d_inode->i_ino;
169 res_id.name[1] = de->d_inode->i_generation;
170 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
171 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
172 &flags, ldlm_completion_ast,
173 mds_blocking_ast, NULL, lockh);
174 if (rc != ELDLM_OK) {
176 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
182 #ifndef DCACHE_DISCONNECTED
183 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
187 /* Look up an entry by inode number. */
188 /* this function ONLY returns valid dget'd dentries with an initialized inode
190 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
191 struct vfsmount **mnt)
194 unsigned long ino = fid->id;
195 __u32 generation = fid->generation;
197 struct dentry *result;
200 RETURN(ERR_PTR(-ESTALE));
202 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
204 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
205 ino, generation, mds->mds_sb);
207 /* under ext3 this is neither supposed to return bad inodes
209 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
213 inode = result->d_inode;
215 RETURN(ERR_PTR(-ENOENT));
217 if (generation && inode->i_generation != generation) {
218 /* we didn't find the right inode.. */
219 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
220 inode->i_ino, (unsigned long)inode->i_nlink,
221 atomic_read(&inode->i_count), inode->i_generation,
224 RETURN(ERR_PTR(-ENOENT));
228 *mnt = mds->mds_vfsmnt;
236 /* Establish a connection to the MDS.
238 * This will set up an export structure for the client to hold state data
239 * about that client, like open files, the last operation number it did
240 * on the server, etc.
242 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
243 struct obd_uuid *cluuid)
245 struct obd_export *exp;
246 struct mds_export_data *med; /* */
247 struct mds_client_data *mcd;
248 int rc, abort_recovery;
251 if (!conn || !obd || !cluuid)
254 /* Check for aborted recovery. */
255 spin_lock_bh(&obd->obd_processing_task_lock);
256 abort_recovery = obd->obd_abort_recovery;
257 spin_unlock_bh(&obd->obd_processing_task_lock);
259 target_abort_recovery(obd);
261 /* XXX There is a small race between checking the list and adding a
262 * new connection for the same UUID, but the real threat (list
263 * corruption when multiple different clients connect) is solved.
265 * There is a second race between adding the export to the list,
266 * and filling in the client data below. Hence skipping the case
267 * of NULL mcd above. We should already be controlling multiple
268 * connects at the client, and we can't hold the spinlock over
269 * memory allocations without risk of deadlocking.
271 rc = class_connect(conn, obd, cluuid);
274 exp = class_conn2export(conn);
276 med = &exp->exp_mds_data;
278 OBD_ALLOC(mcd, sizeof(*mcd));
280 CERROR("mds: out of memory for client data\n");
281 GOTO(out, rc = -ENOMEM);
284 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
287 rc = mds_client_add(obd, &obd->u.mds, med, -1);
292 OBD_FREE(mcd, sizeof(*mcd));
293 class_disconnect(exp, 0);
295 class_export_put(exp);
300 static int mds_init_export(struct obd_export *exp)
302 struct mds_export_data *med = &exp->exp_mds_data;
304 INIT_LIST_HEAD(&med->med_open_head);
305 spin_lock_init(&med->med_open_lock);
309 static int mds_destroy_export(struct obd_export *export)
311 struct mds_export_data *med;
312 struct obd_device *obd = export->exp_obd;
313 struct obd_run_ctxt saved;
317 med = &export->exp_mds_data;
318 target_destroy_export(export);
320 push_ctxt(&saved, &obd->obd_ctxt, NULL);
321 /* Close any open files (which may also cause orphan unlinking). */
322 spin_lock(&med->med_open_lock);
323 while (!list_empty(&med->med_open_head)) {
324 struct list_head *tmp = med->med_open_head.next;
325 struct mds_file_data *mfd =
326 list_entry(tmp, struct mds_file_data, mfd_list);
327 BDEVNAME_DECLARE_STORAGE(btmp);
329 /* bug 1579: fix force-closing for 2.5 */
330 struct dentry *dentry = mfd->mfd_dentry;
332 list_del(&mfd->mfd_list);
333 spin_unlock(&med->med_open_lock);
335 CERROR("force closing client file handle for %*s (%s:%lu)\n",
336 dentry->d_name.len, dentry->d_name.name,
337 ll_bdevname(dentry->d_inode->i_sb, btmp),
338 dentry->d_inode->i_ino);
339 rc = mds_mfd_close(NULL, obd, mfd,
340 !(export->exp_flags & OBD_OPT_FAILOVER));
343 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
344 spin_lock(&med->med_open_lock);
346 spin_unlock(&med->med_open_lock);
347 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
349 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
354 static int mds_disconnect(struct obd_export *export, int flags)
356 unsigned long irqflags;
360 ldlm_cancel_locks_for_export(export);
362 /* complete all outstanding replies */
363 spin_lock_irqsave (&export->exp_lock, irqflags);
364 while (!list_empty (&export->exp_outstanding_replies)) {
365 struct ptlrpc_reply_state *rs =
366 list_entry (export->exp_outstanding_replies.next,
367 struct ptlrpc_reply_state, rs_exp_list);
368 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
370 spin_lock (&svc->srv_lock);
371 list_del_init (&rs->rs_exp_list);
372 ptlrpc_schedule_difficult_reply (rs);
373 spin_unlock (&svc->srv_lock);
375 spin_unlock_irqrestore (&export->exp_lock, irqflags);
377 spin_lock_irqsave(&export->exp_lock, irqflags);
378 export->exp_flags = flags;
379 spin_unlock_irqrestore(&export->exp_lock, irqflags);
381 rc = class_disconnect(export, flags);
385 static int mds_getstatus(struct ptlrpc_request *req)
387 struct mds_obd *mds = mds_req2mds(req);
388 struct mds_body *body;
389 int rc, size = sizeof(*body);
392 rc = lustre_pack_reply(req, 1, &size, NULL);
393 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
394 CERROR("mds: out of memory for message: size=%d\n", size);
395 req->rq_status = -ENOMEM; /* superfluous? */
399 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
400 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
402 /* the last_committed and last_xid fields are filled in for all
403 * replies already - no need to do so here also.
408 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
409 void *data, int flag)
414 if (flag == LDLM_CB_CANCELING) {
415 /* Don't need to do anything here. */
419 /* XXX layering violation! -phil */
420 l_lock(&lock->l_resource->lr_namespace->ns_lock);
421 /* Get this: if mds_blocking_ast is racing with ldlm_intent_policy,
422 * such that mds_blocking_ast is called just before l_i_p takes the
423 * ns_lock, then by the time we get the lock, we might not be the
424 * correct blocking function anymore. So check, and return early, if
426 if (lock->l_blocking_ast != mds_blocking_ast) {
427 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
431 lock->l_flags |= LDLM_FL_CBPENDING;
432 do_ast = (!lock->l_readers && !lock->l_writers);
433 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
436 struct lustre_handle lockh;
439 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
440 ldlm_lock2handle(lock, &lockh);
441 rc = ldlm_cli_cancel(&lockh);
443 CERROR("ldlm_cli_cancel: %d\n", rc);
445 LDLM_DEBUG(lock, "Lock still has references, will be "
451 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
452 * Call with lock=0 if the caller has already taken the i_sem. */
453 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
454 struct mds_body *body, struct inode *inode, int lock)
456 struct mds_obd *mds = &obd->u.mds;
462 lmm = lustre_msg_buf(msg, offset, 0);
464 /* Some problem with getting eadata when I sized the reply
466 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
470 lmm_size = msg->buflens[offset];
472 /* I don't really like this, but it is a sanity check on the client
473 * MD request. However, if the client doesn't know how much space
474 * to reserve for the MD, it shouldn't be bad to have too much space.
476 if (lmm_size > mds->mds_max_mdsize) {
477 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
478 inode->i_ino, lmm_size, mds->mds_max_mdsize);
484 rc = fsfilt_get_md(obd, inode, lmm, lmm_size);
488 CERROR("Error %d reading eadata for ino %lu\n",
492 rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
496 body->valid |= OBD_MD_FLEASIZE;
497 body->eadatasize = lmm_size;
504 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
505 struct ptlrpc_request *req,
506 struct mds_body *reqbody, int reply_off)
508 struct mds_body *body;
509 struct inode *inode = dentry->d_inode;
516 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
517 LASSERT(body != NULL); /* caller prepped reply */
519 mds_pack_inode2fid(&body->fid1, inode);
520 mds_pack_inode2body(body, inode);
522 if (S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
523 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
526 /* If we have LOV EA data, the OST holds size, atime, mtime */
527 if (!(body->valid & OBD_MD_FLEASIZE))
528 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
529 OBD_MD_FLATIME | OBD_MD_FLMTIME);
530 } else if (S_ISLNK(inode->i_mode) &&
531 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
532 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
535 LASSERT (symname != NULL); /* caller prepped reply */
536 len = req->rq_repmsg->buflens[reply_off + 1];
538 rc = inode->i_op->readlink(dentry, symname, len);
540 CERROR("readlink failed: %d\n", rc);
541 } else if (rc != len - 1) {
542 CERROR ("Unexpected readlink rc %d: expecting %d\n",
546 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
547 body->valid |= OBD_MD_LINKNAME;
548 body->eadatasize = rc + 1;
549 symname[rc] = 0; /* NULL terminate */
557 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
560 struct mds_obd *mds = mds_req2mds(req);
561 struct mds_body *body;
562 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
565 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
566 LASSERT(body != NULL); /* checked by caller */
567 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
569 if (S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) {
572 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
574 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
578 CERROR("error getting inode %lu MD: rc = %d\n",
581 } else if (rc > mds->mds_max_mdsize) {
583 CERROR("MD size %d larger than maximum possible %u\n",
584 rc, mds->mds_max_mdsize);
589 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
590 if (inode->i_size + 1 != body->eadatasize)
591 CERROR("symlink size: %Lu, reply space: %d\n",
592 inode->i_size + 1, body->eadatasize);
593 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
595 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
596 inode->i_size + 1, body->eadatasize);
599 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
600 CERROR("failed MDS_GETATTR_PACK test\n");
601 req->rq_status = -ENOMEM;
602 GOTO(out, rc = -ENOMEM);
605 rc = lustre_pack_reply(req, bufcount, size, NULL);
607 CERROR("out of memory\n");
608 GOTO(out, req->rq_status = rc);
616 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
617 struct lustre_handle *child_lockh)
619 struct obd_device *obd = req->rq_export->exp_obd;
620 struct ldlm_reply *rep = NULL;
621 struct obd_run_ctxt saved;
622 struct mds_body *body;
623 struct dentry *dparent = NULL, *dchild = NULL;
625 struct lustre_handle parent_lockh;
627 int rc = 0, cleanup_phase = 0, resent_req = 0;
631 LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
633 /* Swab now, before anyone looks inside the request */
635 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
636 lustre_swab_mds_body);
638 CERROR("Can't swab mds_body\n");
639 GOTO(cleanup, rc = -EFAULT);
642 LASSERT_REQSWAB(req, offset + 1);
643 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
645 CERROR("Can't unpack name\n");
646 GOTO(cleanup, rc = -EFAULT);
648 namesize = req->rq_reqmsg->buflens[offset + 1];
650 LASSERT (offset == 0 || offset == 2);
651 /* if requests were at offset 2, the getattr reply goes back at 1 */
653 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
657 uc.ouc_fsuid = body->fsuid;
658 uc.ouc_fsgid = body->fsgid;
659 uc.ouc_cap = body->capability;
660 uc.ouc_suppgid1 = body->suppgid;
661 uc.ouc_suppgid2 = -1;
662 push_ctxt(&saved, &obd->obd_ctxt, &uc);
663 cleanup_phase = 1; /* kernel context */
664 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
666 /* FIXME: handle raw lookup */
668 if (body->valid == OBD_MD_FLID) {
669 struct mds_body *mds_reply;
670 int size = sizeof(*mds_reply);
672 // The user requested ONLY the inode number, so do a raw lookup
673 rc = lustre_pack_reply(req, 1, &size, NULL);
675 CERROR("out of memory\n");
679 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
681 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
683 mds_reply->fid1.id = inum;
684 mds_reply->valid = OBD_MD_FLID;
689 if (child_lockh->cookie != 0) {
690 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
694 if (resent_req == 0) {
695 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
696 &parent_lockh, &dparent,
697 LCK_PR, name, namesize,
698 child_lockh, &dchild, LCK_PR);
702 struct ldlm_lock *granted_lock;
703 struct ll_fid child_fid;
704 struct ldlm_resource *res;
705 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
706 granted_lock = ldlm_handle2lock(child_lockh);
707 LASSERT(granted_lock);
709 res = granted_lock->l_resource;
710 child_fid.id = res->lr_name.name[0];
711 child_fid.generation = res->lr_name.name[1];
712 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
714 LDLM_LOCK_PUT(granted_lock);
717 cleanup_phase = 2; /* dchild, dparent, locks */
719 if (dchild->d_inode == NULL) {
720 intent_set_disposition(rep, DISP_LOOKUP_NEG);
721 /* in the intent case, the policy clears this error:
722 the disposition is enough */
723 GOTO(cleanup, rc = -ENOENT);
725 intent_set_disposition(rep, DISP_LOOKUP_POS);
728 if (req->rq_repmsg == NULL) {
729 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
731 CERROR ("mds_getattr_pack_msg: %d\n", rc);
736 rc = mds_getattr_internal(obd, dchild, req, body, offset);
737 GOTO(cleanup, rc); /* returns the lock to the client */
740 switch (cleanup_phase) {
742 if (resent_req == 0) {
743 if (rc && dchild->d_inode)
744 ldlm_lock_decref(child_lockh, LCK_PR);
745 ldlm_lock_decref(&parent_lockh, LCK_PR);
750 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
756 static int mds_getattr(int offset, struct ptlrpc_request *req)
758 struct mds_obd *mds = mds_req2mds(req);
759 struct obd_device *obd = req->rq_export->exp_obd;
760 struct obd_run_ctxt saved;
762 struct mds_body *body;
767 body = lustre_swab_reqbuf (req, offset, sizeof (*body),
768 lustre_swab_mds_body);
770 CERROR ("Can't unpack body\n");
774 uc.ouc_fsuid = body->fsuid;
775 uc.ouc_fsgid = body->fsgid;
776 uc.ouc_cap = body->capability;
777 push_ctxt(&saved, &obd->obd_ctxt, &uc);
778 de = mds_fid2dentry(mds, &body->fid1, NULL);
780 rc = req->rq_status = -ENOENT;
781 GOTO(out_pop, PTR_ERR(de));
784 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
786 CERROR ("mds_getattr_pack_msg: %d\n", rc);
790 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
795 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
800 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
801 unsigned long max_age)
805 spin_lock(&obd->obd_osfs_lock);
806 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
808 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
809 spin_unlock(&obd->obd_osfs_lock);
814 static int mds_statfs(struct ptlrpc_request *req)
816 struct obd_device *obd = req->rq_export->exp_obd;
817 int rc, size = sizeof(struct obd_statfs);
820 rc = lustre_pack_reply(req, 1, &size, NULL);
821 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
822 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
826 /* We call this so that we can cache a bit - 1 jiffie worth */
827 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
830 CERROR("mds_obd_statfs failed: rc %d\n", rc);
840 static int mds_sync(struct ptlrpc_request *req)
842 struct obd_device *obd = req->rq_export->exp_obd;
843 struct mds_obd *mds = &obd->u.mds;
844 struct mds_body *body;
845 int rc, size = sizeof(*body);
848 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
850 GOTO(out, rc = -EPROTO);
852 rc = lustre_pack_reply(req, 1, &size, NULL);
853 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
854 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
858 if (body->fid1.id == 0) {
859 /* a fid of zero is taken to mean "sync whole filesystem" */
860 rc = fsfilt_sync(obd, mds->mds_sb);
864 /* just any file to grab fsync method - "file" arg unused */
865 struct file *file = mds->mds_rcvd_filp;
868 de = mds_fid2dentry(mds, &body->fid1, NULL);
870 GOTO(out, rc = PTR_ERR(de));
872 rc = file->f_op->fsync(NULL, de, 1);
877 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
878 mds_pack_inode2fid(&body->fid1, de->d_inode);
879 mds_pack_inode2body(body, de->d_inode);
886 /* mds_readpage does not take a DLM lock on the inode, because the client must
887 * already have a PR lock.
889 * If we were to take another one here, a deadlock will result, if another
890 * thread is already waiting for a PW lock. */
891 static int mds_readpage(struct ptlrpc_request *req)
893 struct obd_device *obd = req->rq_export->exp_obd;
894 struct vfsmount *mnt;
897 struct mds_body *body, *repbody;
898 struct obd_run_ctxt saved;
899 int rc, size = sizeof(*repbody);
903 rc = lustre_pack_reply(req, 1, &size, NULL);
904 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
905 CERROR("mds: out of memory\n");
906 GOTO(out, rc = -ENOMEM);
909 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
911 GOTO (out, rc = -EFAULT);
913 uc.ouc_fsuid = body->fsuid;
914 uc.ouc_fsgid = body->fsgid;
915 uc.ouc_cap = body->capability;
916 push_ctxt(&saved, &obd->obd_ctxt, &uc);
917 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
919 GOTO(out_pop, rc = PTR_ERR(de));
921 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
923 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
924 /* note: in case of an error, dentry_open puts dentry */
926 GOTO(out_pop, rc = PTR_ERR(file));
928 /* body->size is actually the offset -eeb */
929 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
930 CERROR("offset "LPU64" not on a block boundary of %lu\n",
931 body->size, de->d_inode->i_blksize);
932 GOTO(out_file, rc = -EFAULT);
935 /* body->nlink is actually the #bytes to read -eeb */
936 if (body->nlink & (de->d_inode->i_blksize - 1)) {
937 CERROR("size %u is not multiple of blocksize %lu\n",
938 body->nlink, de->d_inode->i_blksize);
939 GOTO(out_file, rc = -EFAULT);
942 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
943 repbody->size = file->f_dentry->d_inode->i_size;
944 repbody->valid = OBD_MD_FLSIZE;
946 /* to make this asynchronous make sure that the handling function
947 doesn't send a reply when this function completes. Instead a
948 callback function would send the reply */
949 /* body->size is actually the offset -eeb */
950 rc = mds_sendpage(req, file, body->size, body->nlink);
955 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
961 int mds_reint(struct ptlrpc_request *req, int offset,
962 struct lustre_handle *lockh)
964 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
967 OBD_ALLOC(rec, sizeof(*rec));
971 rc = mds_update_unpack(req, offset, rec);
972 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
973 CERROR("invalid record\n");
974 GOTO(out, req->rq_status = -EINVAL);
976 /* rc will be used to interrupt a for loop over multiple records */
977 rc = mds_reint_rec(rec, offset, req, lockh);
979 OBD_FREE(rec, sizeof(*rec));
983 static int mds_filter_recovery_request(struct ptlrpc_request *req,
984 struct obd_device *obd, int *process)
986 switch (req->rq_reqmsg->opc) {
987 case MDS_CONNECT: /* This will never get here, but for completeness. */
988 case OST_CONNECT: /* This will never get here, but for completeness. */
995 case MDS_SYNC: /* used in unmounting */
999 *process = target_queue_recovery_request(req, obd);
1003 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1005 /* XXX what should we set rq_status to here? */
1006 req->rq_status = -EAGAIN;
1007 RETURN(ptlrpc_error(req));
1011 static char *reint_names[] = {
1012 [REINT_SETATTR] "setattr",
1013 [REINT_CREATE] "create",
1014 [REINT_LINK] "link",
1015 [REINT_UNLINK] "unlink",
1016 [REINT_RENAME] "rename",
1017 [REINT_OPEN] "open",
1020 int mds_handle(struct ptlrpc_request *req)
1022 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1024 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1025 struct obd_device *obd = NULL;
1028 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1030 LASSERT(current->journal_info == NULL);
1031 /* XXX identical to OST */
1032 if (req->rq_reqmsg->opc != MDS_CONNECT) {
1033 struct mds_export_data *med;
1034 int recovering, abort_recovery;
1036 if (req->rq_export == NULL) {
1037 CERROR("lustre_mds: operation %d on unconnected MDS\n",
1038 req->rq_reqmsg->opc);
1039 req->rq_status = -ENOTCONN;
1040 GOTO(out, rc = -ENOTCONN);
1043 med = &req->rq_export->exp_mds_data;
1044 obd = req->rq_export->exp_obd;
1047 /* Check for aborted recovery. */
1048 spin_lock_bh(&obd->obd_processing_task_lock);
1049 abort_recovery = obd->obd_abort_recovery;
1050 recovering = obd->obd_recovering;
1051 spin_unlock_bh(&obd->obd_processing_task_lock);
1052 if (abort_recovery) {
1053 target_abort_recovery(obd);
1054 } else if (recovering) {
1055 rc = mds_filter_recovery_request(req, obd,
1057 if (rc || !should_process)
1062 switch (req->rq_reqmsg->opc) {
1064 DEBUG_REQ(D_INODE, req, "connect");
1065 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1066 rc = target_handle_connect(req, mds_handle);
1068 /* Now that we have an export, set mds. */
1069 mds = mds_req2mds(req);
1072 case MDS_DISCONNECT:
1073 DEBUG_REQ(D_INODE, req, "disconnect");
1074 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1075 rc = target_handle_disconnect(req);
1076 req->rq_status = rc; /* superfluous? */
1080 DEBUG_REQ(D_INODE, req, "getstatus");
1081 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1082 rc = mds_getstatus(req);
1086 DEBUG_REQ(D_INODE, req, "getattr");
1087 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1088 rc = mds_getattr(0, req);
1091 case MDS_GETATTR_NAME: {
1092 struct lustre_handle lockh;
1093 DEBUG_REQ(D_INODE, req, "getattr_name");
1094 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1096 /* If this request gets a reconstructed reply, we won't be
1097 * acquiring any new locks in mds_getattr_name, so we don't
1101 rc = mds_getattr_name(0, req, &lockh);
1102 /* this non-intent call (from an ioctl) is special */
1103 req->rq_status = rc;
1104 if (rc == 0 && lockh.cookie)
1105 ldlm_lock_decref(&lockh, LCK_PR);
1109 DEBUG_REQ(D_INODE, req, "statfs");
1110 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1111 rc = mds_statfs(req);
1115 DEBUG_REQ(D_INODE, req, "readpage");
1116 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1117 rc = mds_readpage(req);
1119 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1120 if (req->rq_reply_state) {
1121 lustre_free_reply_state (req->rq_reply_state);
1122 req->rq_reply_state = NULL;
1130 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1132 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1133 mds->mds_max_cookiesize};
1136 /* NB only peek inside req now; mds_reint() will swab it */
1138 CERROR ("Can't inspect opcode\n");
1143 if (lustre_msg_swabbed (req->rq_reqmsg))
1146 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1147 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1148 reint_names[opc] == NULL) ? reint_names[opc] :
1151 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1153 if (opc == REINT_UNLINK)
1155 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1160 rc = lustre_pack_reply(req, bufcount, size, NULL);
1164 rc = mds_reint(req, 0, NULL);
1165 fail = OBD_FAIL_MDS_REINT_NET_REP;
1170 DEBUG_REQ(D_INODE, req, "close");
1171 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1172 rc = mds_close(req);
1175 case MDS_DONE_WRITING:
1176 DEBUG_REQ(D_INODE, req, "done_writing");
1177 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1178 rc = mds_done_writing(req);
1182 DEBUG_REQ(D_INODE, req, "pin");
1183 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1188 DEBUG_REQ(D_INODE, req, "sync");
1189 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1194 DEBUG_REQ(D_INODE, req, "ping");
1195 rc = target_handle_ping(req);
1198 case OBD_LOG_CANCEL:
1199 CDEBUG(D_INODE, "log cancel\n");
1200 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1201 rc = -ENOTSUPP; /* la la la */
1205 DEBUG_REQ(D_INODE, req, "enqueue");
1206 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1207 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1208 ldlm_server_blocking_ast);
1211 DEBUG_REQ(D_INODE, req, "convert");
1212 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1213 rc = ldlm_handle_convert(req);
1215 case LDLM_BL_CALLBACK:
1216 case LDLM_CP_CALLBACK:
1217 DEBUG_REQ(D_INODE, req, "callback");
1218 CERROR("callbacks should not happen on MDS\n");
1220 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1222 case LLOG_ORIGIN_HANDLE_CREATE:
1223 DEBUG_REQ(D_INODE, req, "llog_init");
1224 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1225 rc = llog_origin_handle_create(req);
1227 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1228 DEBUG_REQ(D_INODE, req, "llog next block");
1229 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1230 rc = llog_origin_handle_next_block(req);
1232 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1233 DEBUG_REQ(D_INODE, req, "llog read header");
1234 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1235 rc = llog_origin_handle_read_header(req);
1237 case LLOG_ORIGIN_HANDLE_CLOSE:
1238 DEBUG_REQ(D_INODE, req, "llog close");
1239 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1240 rc = llog_origin_handle_close(req);
1243 DEBUG_REQ(D_INODE, req, "llog catinfo");
1244 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1245 rc = llog_catinfo(req);
1248 req->rq_status = -ENOTSUPP;
1249 rc = ptlrpc_error(req);
1253 LASSERT(current->journal_info == NULL);
1257 /* If we're DISCONNECTing, the mds_export_data is already freed */
1258 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1259 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1260 struct obd_device *obd = list_entry(mds, struct obd_device,
1262 req->rq_repmsg->last_xid =
1263 le64_to_cpu(med->med_mcd->mcd_last_xid);
1265 if (!obd->obd_no_transno) {
1266 req->rq_repmsg->last_committed =
1267 obd->obd_last_committed;
1269 DEBUG_REQ(D_IOCTL, req,
1270 "not sending last_committed update");
1272 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1274 mds->mds_last_transno, obd->obd_last_committed,
1279 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1280 if (obd && obd->obd_recovering) {
1281 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1282 return target_queue_final_reply(req, rc);
1284 /* Lost a race with recovery; let the error path DTRT. */
1285 rc = req->rq_status = -ENOTCONN;
1288 target_send_reply(req, rc, fail);
1292 /* Update the server data on disk. This stores the new mount_count and
1293 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1294 * then the server last_rcvd value may be less than that of the clients.
1295 * This will alert us that we may need to do client recovery.
1297 * Also assumes for mds_last_transno that we are not modifying it (no locking).
1299 int mds_update_server_data(struct obd_device *obd, int force_sync)
1301 struct mds_obd *mds = &obd->u.mds;
1302 struct mds_server_data *msd = mds->mds_server_data;
1303 struct file *filp = mds->mds_rcvd_filp;
1304 struct obd_run_ctxt saved;
1309 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1310 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1312 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1313 mds->mds_mount_count, mds->mds_last_transno);
1314 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1316 CERROR("error writing MDS server data: rc = %d\n", rc);
1317 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1323 /* mount the file system (secretly) */
1324 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1326 struct lustre_cfg* lcfg = buf;
1327 struct mds_obd *mds = &obd->u.mds;
1328 struct vfsmount *mnt;
1333 dev_clear_rdonly(2);
1335 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1336 RETURN(rc = -EINVAL);
1338 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1339 if (IS_ERR(obd->obd_fsops))
1340 RETURN(rc = PTR_ERR(obd->obd_fsops));
1342 if (!(page = __get_free_page(GFP_KERNEL)))
1345 memset((void *)page, 0, PAGE_SIZE);
1346 sprintf((char *)page, "iopen_nopriv");
1348 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1349 lcfg->lcfg_inlbuf1, (void *)page);
1353 CERROR("do_kern_mount failed: rc = %d\n", rc);
1357 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1359 sema_init(&mds->mds_orphan_recovery_sem, 1);
1360 sema_init(&mds->mds_epoch_sem, 1);
1361 spin_lock_init(&mds->mds_transno_lock);
1362 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1363 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1364 atomic_set(&mds->mds_open_count, 0);
1366 obd->obd_namespace = ldlm_namespace_new("mds_server",
1367 LDLM_NAMESPACE_SERVER);
1368 if (obd->obd_namespace == NULL) {
1369 mds_cleanup(obd, 0);
1370 GOTO(err_put, rc = -ENOMEM);
1373 rc = mds_fs_setup(obd, mnt);
1375 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1379 rc = llog_start_commit_thread();
1384 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1387 generate_random_uuid(uuid);
1388 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1390 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1391 if (mds->mds_profile == NULL)
1392 GOTO(err_fs, rc = -ENOMEM);
1394 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1395 lcfg->lcfg_inllen3);
1399 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1400 "mds_ldlm_client", &obd->obd_ldlm_client);
1401 obd->obd_replayable = 1;
1403 rc = mds_postsetup(obd);
1409 /* No extra cleanup needed for llog_init_commit_thread() */
1410 mds_fs_cleanup(obd, 0);
1412 ldlm_namespace_free(obd->obd_namespace, 0);
1413 obd->obd_namespace = NULL;
1416 mntput(mds->mds_vfsmnt);
1420 fsfilt_put_ops(obd->obd_fsops);
1424 static int mds_postsetup(struct obd_device *obd)
1426 struct mds_obd *mds = &obd->u.mds;
1431 rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
1436 if (mds->mds_profile) {
1437 struct obd_run_ctxt saved;
1438 struct lustre_profile *lprof;
1439 struct config_llog_instance cfg;
1441 cfg.cfg_instance = NULL;
1442 cfg.cfg_uuid = mds->mds_lov_uuid;
1443 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1444 rc = class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1445 mds->mds_profile, &cfg);
1446 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1450 lprof = class_get_profile(mds->mds_profile);
1451 if (lprof == NULL) {
1452 CERROR("No profile found: %s\n", mds->mds_profile);
1453 GOTO(err_cleanup, rc = -ENOENT);
1455 rc = mds_lov_connect(obd, lprof->lp_osc);
1457 GOTO(err_cleanup, rc);
1465 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1469 static int mds_postrecov(struct obd_device *obd)
1474 LASSERT(!obd->obd_recovering);
1475 LASSERT(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT) != NULL);
1477 rc = llog_connect(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT),
1478 obd->u.mds.mds_lov_desc.ld_tgt_count, NULL, NULL);
1480 CERROR("faild at llog_origin_connect: %d\n", rc);
1483 rc = mds_cleanup_orphans(obd);
1485 rc2 = mds_lov_set_nextid(obd);
1491 int mds_lov_clean(struct obd_device *obd)
1493 struct mds_obd *mds = &obd->u.mds;
1495 if (mds->mds_profile) {
1497 struct config_llog_instance cfg;
1498 struct obd_run_ctxt saved;
1499 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1501 OBD_ALLOC(cln_prof, len);
1502 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1504 cfg.cfg_instance = NULL;
1505 cfg.cfg_uuid = mds->mds_lov_uuid;
1507 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1508 class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1510 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1512 OBD_FREE(cln_prof, len);
1513 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1514 mds->mds_profile = NULL;
1519 static int mds_precleanup(struct obd_device *obd, int flags)
1524 mds_lov_disconnect(obd, flags);
1526 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1530 static int mds_cleanup(struct obd_device *obd, int flags)
1532 struct mds_obd *mds = &obd->u.mds;
1535 if (mds->mds_sb == NULL)
1538 mds_update_server_data(obd, 1);
1539 if (mds->mds_lov_objids != NULL) {
1540 OBD_FREE(mds->mds_lov_objids,
1541 mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1543 mds_fs_cleanup(obd, flags);
1547 /* 2 seems normal on mds, (may_umount() also expects 2
1548 fwiw), but we only see 1 at this point in obdfilter. */
1549 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1550 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1551 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1553 mntput(mds->mds_vfsmnt);
1557 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1559 spin_lock_bh(&obd->obd_processing_task_lock);
1560 if (obd->obd_recovering) {
1561 target_cancel_recovery_timer(obd);
1562 obd->obd_recovering = 0;
1564 spin_unlock_bh(&obd->obd_processing_task_lock);
1567 dev_clear_rdonly(2);
1568 fsfilt_put_ops(obd->obd_fsops);
1573 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
1574 struct ldlm_lock *new_lock,
1575 struct lustre_handle *lockh)
1577 struct obd_export *exp = req->rq_export;
1578 struct obd_device *obd = exp->exp_obd;
1579 struct ldlm_request *dlmreq =
1580 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
1581 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
1582 struct list_head *iter;
1584 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
1587 l_lock(&obd->obd_namespace->ns_lock);
1588 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
1589 struct ldlm_lock *lock;
1590 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1591 if (lock == new_lock)
1593 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
1594 lockh->cookie = lock->l_handle.h_cookie;
1595 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
1597 l_unlock(&obd->obd_namespace->ns_lock);
1601 l_unlock(&obd->obd_namespace->ns_lock);
1603 /* This remote handle isn't enqueued, so we never received or
1604 * processed this request. Clear MSG_RESENT, because it can
1605 * be handled like any normal request now. */
1607 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1609 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
1613 int intent_disposition(struct ldlm_reply *rep, int flag)
1617 return (rep->lock_policy_res1 & flag);
1620 void intent_set_disposition(struct ldlm_reply *rep, int flag)
1624 rep->lock_policy_res1 |= flag;
1627 static int ldlm_intent_policy(struct ldlm_namespace *ns,
1628 struct ldlm_lock **lockp, void *req_cookie,
1629 ldlm_mode_t mode, int flags, void *data)
1631 struct ptlrpc_request *req = req_cookie;
1632 struct ldlm_lock *lock = *lockp;
1639 if (req->rq_reqmsg->bufcount > 1) {
1640 /* an intent needs to be considered */
1641 struct ldlm_intent *it;
1642 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1643 struct ldlm_reply *rep;
1644 struct lustre_handle lockh = { 0 };
1645 struct ldlm_lock *new_lock;
1646 int offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
1647 sizeof(struct mds_body),
1648 mds->mds_max_mdsize,
1649 mds->mds_max_cookiesize};
1651 it = lustre_swab_reqbuf(req, 1, sizeof (*it),
1652 lustre_swab_ldlm_intent);
1654 CERROR ("Intent missing\n");
1655 req->rq_status = -EFAULT;
1656 RETURN(req->rq_status);
1659 LDLM_DEBUG(lock, "intent policy, opc: %s",
1660 ldlm_it2str(it->opc));
1662 rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3,
1665 RETURN(req->rq_status = rc);
1667 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1668 intent_set_disposition(rep, DISP_IT_EXECD);
1670 fixup_handle_for_resent_req(req, lock, &lockh);
1672 /* execute policy */
1673 switch ((long)it->opc) {
1675 case IT_CREAT|IT_OPEN:
1676 /* XXX swab here to assert that an mds_open reint
1677 * packet is following */
1678 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
1680 /* We abort the lock if the lookup was negative and
1681 * we did not make it to the OPEN portion */
1682 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
1683 RETURN(ELDLM_LOCK_ABORTED);
1684 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
1685 !intent_disposition(rep, DISP_OPEN_OPEN))
1687 RETURN(ELDLM_LOCK_ABORTED);
1692 rep->lock_policy_res2 = mds_getattr_name(offset, req,
1694 /* FIXME: LDLM can set req->rq_status. MDS sets
1695 policy_res{1,2} with disposition and status.
1696 - replay: returns 0 & req->status is old status
1697 - otherwise: returns req->status */
1698 if (intent_disposition(rep, DISP_LOOKUP_NEG))
1699 rep->lock_policy_res2 = 0;
1700 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
1701 rep->lock_policy_res2)
1702 RETURN(ELDLM_LOCK_ABORTED);
1703 if (req->rq_status != 0) {
1705 rep->lock_policy_res2 = req->rq_status;
1706 RETURN(ELDLM_LOCK_ABORTED);
1710 CERROR("Unhandled intent "LPD64"\n", it->opc);
1714 /* By this point, whatever function we called above must have
1715 * either filled in 'lockh', been an intent replay, or returned
1716 * an error. We want to allow replayed RPCs to not get a lock,
1717 * since we would just drop it below anyways because lock replay
1718 * is done separately by the client afterwards. For regular
1719 * RPCs we want to give the new lock to the client instead of
1720 * whatever lock it was about to get.
1722 new_lock = ldlm_handle2lock(&lockh);
1723 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
1726 LASSERT(new_lock != NULL);
1728 /* If we've already given this lock to a client once, then we
1729 * should have no readers or writers. Otherwise, we should
1730 * have one reader _or_ writer ref (which will be zeroed below)
1731 * before returning the lock to a client.
1733 if (new_lock->l_export == req->rq_export) {
1734 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
1736 LASSERT(new_lock->l_export == NULL);
1737 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
1742 if (new_lock->l_export == req->rq_export) {
1743 /* Already gave this to the client, which means that we
1744 * reconstructed a reply. */
1745 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
1747 RETURN(ELDLM_LOCK_REPLACED);
1750 /* Fixup the lock to be given to the client */
1751 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
1752 new_lock->l_readers = 0;
1753 new_lock->l_writers = 0;
1755 new_lock->l_export = class_export_get(req->rq_export);
1756 list_add(&new_lock->l_export_chain,
1757 &new_lock->l_export->exp_ldlm_data.led_held_locks);
1759 new_lock->l_blocking_ast = lock->l_blocking_ast;
1760 new_lock->l_completion_ast = lock->l_completion_ast;
1762 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
1763 sizeof(lock->l_remote_handle));
1765 new_lock->l_flags &= ~LDLM_FL_LOCAL;
1767 LDLM_LOCK_PUT(new_lock);
1768 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
1770 RETURN(ELDLM_LOCK_REPLACED);
1772 int size = sizeof(struct ldlm_reply);
1773 rc = lustre_pack_reply(req, 1, &size, NULL);
1782 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1784 struct lprocfs_static_vars lvars;
1786 lprocfs_init_multi_vars(0, &lvars);
1787 return lprocfs_obd_attach(dev, lvars.obd_vars);
1790 int mds_detach(struct obd_device *dev)
1792 return lprocfs_obd_detach(dev);
1795 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
1797 struct lprocfs_static_vars lvars;
1799 lprocfs_init_multi_vars(1, &lvars);
1800 return lprocfs_obd_attach(dev, lvars.obd_vars);
1803 int mdt_detach(struct obd_device *dev)
1805 return lprocfs_obd_detach(dev);
1808 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1810 struct mds_obd *mds = &obddev->u.mds;
1815 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1816 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1818 obddev->obd_proc_entry);
1820 if (!mds->mds_service) {
1821 CERROR("failed to start service\n");
1822 RETURN(rc = -ENOMEM);
1825 rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
1828 GOTO(err_thread, rc);
1830 mds->mds_setattr_service =
1831 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1832 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
1833 mds_handle, "mds_setattr",
1834 obddev->obd_proc_entry);
1835 if (!mds->mds_setattr_service) {
1836 CERROR("failed to start getattr service\n");
1837 GOTO(err_thread, rc = -ENOMEM);
1840 rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
1841 MDT_NUM_THREADS, "ll_mdt_attr");
1843 GOTO(err_thread2, rc);
1845 mds->mds_readpage_service =
1846 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1847 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
1848 mds_handle, "mds_readpage",
1849 obddev->obd_proc_entry);
1850 if (!mds->mds_readpage_service) {
1851 CERROR("failed to start readpage service\n");
1852 GOTO(err_thread2, rc = -ENOMEM);
1855 rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
1856 MDT_NUM_THREADS, "ll_mdt_rdpg");
1859 GOTO(err_thread3, rc);
1864 ptlrpc_unregister_service(mds->mds_readpage_service);
1866 ptlrpc_unregister_service(mds->mds_setattr_service);
1868 ptlrpc_unregister_service(mds->mds_service);
1873 static int mdt_cleanup(struct obd_device *obddev, int flags)
1875 struct mds_obd *mds = &obddev->u.mds;
1878 ptlrpc_stop_all_threads(mds->mds_readpage_service);
1879 ptlrpc_unregister_service(mds->mds_readpage_service);
1881 ptlrpc_stop_all_threads(mds->mds_setattr_service);
1882 ptlrpc_unregister_service(mds->mds_setattr_service);
1884 ptlrpc_stop_all_threads(mds->mds_service);
1885 ptlrpc_unregister_service(mds->mds_service);
1890 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
1892 struct obd_device *obd = data;
1895 fid.generation = gen;
1896 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
1899 struct lvfs_callback_ops mds_lvfs_ops = {
1900 l_fid2dentry: mds_lvfs_fid2dentry,
1903 /* use obd ops to offer management infrastructure */
1904 static struct obd_ops mds_obd_ops = {
1905 o_owner: THIS_MODULE,
1906 o_attach: mds_attach,
1907 o_detach: mds_detach,
1908 o_connect: mds_connect,
1909 o_init_export: mds_init_export,
1910 o_destroy_export: mds_destroy_export,
1911 o_disconnect: mds_disconnect,
1913 o_precleanup: mds_precleanup,
1914 o_cleanup: mds_cleanup,
1915 o_postrecov: mds_postrecov,
1916 o_statfs: mds_obd_statfs,
1917 o_iocontrol: mds_iocontrol,
1918 o_create: mds_obd_create,
1919 o_destroy: mds_obd_destroy,
1920 o_llog_init: mds_llog_init,
1921 o_llog_finish: mds_llog_finish,
1922 o_notify: mds_notify,
1925 static struct obd_ops mdt_obd_ops = {
1926 o_owner: THIS_MODULE,
1927 o_attach: mdt_attach,
1928 o_detach: mdt_detach,
1930 o_cleanup: mdt_cleanup,
1933 static int __init mds_init(void)
1935 struct lprocfs_static_vars lvars;
1937 lprocfs_init_multi_vars(0, &lvars);
1938 class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
1939 lprocfs_init_multi_vars(1, &lvars);
1940 class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
1941 ldlm_register_intent(ldlm_intent_policy);
1946 static void /*__exit*/ mds_exit(void)
1948 ldlm_unregister_intent();
1949 class_unregister_type(LUSTRE_MDS_NAME);
1950 class_unregister_type(LUSTRE_MDT_NAME);
1953 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1954 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
1955 MODULE_LICENSE("GPL");
1957 module_init(mds_init);
1958 module_exit(mds_exit);