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, 1, BULK_PUT_SOURCE, MDS_BULK_PORTAL);
84 GOTO(out_free, rc = -ENOMEM);
86 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
87 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
89 pages[i] = alloc_pages(GFP_KERNEL, 0);
91 GOTO(cleanup_buf, rc = -ENOMEM);
93 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
96 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
97 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
98 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
99 tmpsize, offset, file->f_dentry->d_inode->i_ino,
100 file->f_dentry->d_inode->i_size);
102 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
103 page_address(pages[i]), tmpsize, &offset);
106 GOTO(cleanup_buf, rc = -EIO);
109 LASSERT(desc->bd_nob == count);
111 rc = ptlrpc_start_bulk_transfer(desc);
113 GOTO(cleanup_buf, rc);
115 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
116 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
117 OBD_FAIL_MDS_SENDPAGE, rc);
118 GOTO(abort_bulk, rc);
121 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
122 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
123 LASSERT (rc == 0 || rc == -ETIMEDOUT);
126 if (desc->bd_success &&
127 desc->bd_nob_transferred == count)
128 GOTO(cleanup_buf, rc);
130 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
133 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
134 (rc == -ETIMEDOUT) ? "timeout" : "network error",
135 desc->bd_nob_transferred, count,
136 req->rq_export->exp_client_uuid.uuid,
137 req->rq_export->exp_connection->c_remote_uuid.uuid);
139 ptlrpc_fail_export(req->rq_export);
143 ptlrpc_abort_bulk (desc);
145 for (i = 0; i < npages; i++)
147 __free_pages(pages[i], 0);
149 ptlrpc_free_bulk(desc);
151 OBD_FREE(pages, sizeof(*pages) * npages);
156 /* only valid locked dentries or errors should be returned */
157 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
158 struct vfsmount **mnt, int lock_mode,
159 struct lustre_handle *lockh,
160 char *name, int namelen)
162 struct mds_obd *mds = &obd->u.mds;
163 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
164 struct ldlm_res_id res_id = { .name = {0} };
171 res_id.name[0] = de->d_inode->i_ino;
172 res_id.name[1] = de->d_inode->i_generation;
173 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
174 LDLM_PLAIN, NULL, lock_mode, &flags,
175 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
176 NULL, 0, NULL, lockh);
177 if (rc != ELDLM_OK) {
179 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
185 #ifndef DCACHE_DISCONNECTED
186 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
190 /* Look up an entry by inode number. */
191 /* this function ONLY returns valid dget'd dentries with an initialized inode
193 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
194 struct vfsmount **mnt)
197 unsigned long ino = fid->id;
198 __u32 generation = fid->generation;
200 struct dentry *result;
203 RETURN(ERR_PTR(-ESTALE));
205 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
207 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
208 ino, generation, mds->mds_sb);
210 /* under ext3 this is neither supposed to return bad inodes
212 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
216 inode = result->d_inode;
218 RETURN(ERR_PTR(-ENOENT));
220 if (generation && inode->i_generation != generation) {
221 /* we didn't find the right inode.. */
222 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
223 inode->i_ino, (unsigned long)inode->i_nlink,
224 atomic_read(&inode->i_count), inode->i_generation,
227 RETURN(ERR_PTR(-ENOENT));
231 *mnt = mds->mds_vfsmnt;
239 /* Establish a connection to the MDS.
241 * This will set up an export structure for the client to hold state data
242 * about that client, like open files, the last operation number it did
243 * on the server, etc.
245 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
246 struct obd_uuid *cluuid)
248 struct obd_export *exp;
249 struct mds_export_data *med; /* */
250 struct mds_client_data *mcd;
251 int rc, abort_recovery;
254 if (!conn || !obd || !cluuid)
257 /* Check for aborted recovery. */
258 spin_lock_bh(&obd->obd_processing_task_lock);
259 abort_recovery = obd->obd_abort_recovery;
260 spin_unlock_bh(&obd->obd_processing_task_lock);
262 target_abort_recovery(obd);
264 /* XXX There is a small race between checking the list and adding a
265 * new connection for the same UUID, but the real threat (list
266 * corruption when multiple different clients connect) is solved.
268 * There is a second race between adding the export to the list,
269 * and filling in the client data below. Hence skipping the case
270 * of NULL mcd above. We should already be controlling multiple
271 * connects at the client, and we can't hold the spinlock over
272 * memory allocations without risk of deadlocking.
274 rc = class_connect(conn, obd, cluuid);
277 exp = class_conn2export(conn);
279 med = &exp->exp_mds_data;
281 OBD_ALLOC(mcd, sizeof(*mcd));
283 CERROR("mds: out of memory for client data\n");
284 GOTO(out, rc = -ENOMEM);
287 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
290 rc = mds_client_add(obd, &obd->u.mds, med, -1);
295 OBD_FREE(mcd, sizeof(*mcd));
296 class_disconnect(exp, 0);
298 class_export_put(exp);
303 static int mds_init_export(struct obd_export *exp)
305 struct mds_export_data *med = &exp->exp_mds_data;
307 INIT_LIST_HEAD(&med->med_open_head);
308 spin_lock_init(&med->med_open_lock);
312 static int mds_destroy_export(struct obd_export *export)
314 struct mds_export_data *med;
315 struct obd_device *obd = export->exp_obd;
316 struct obd_run_ctxt saved;
320 med = &export->exp_mds_data;
321 target_destroy_export(export);
323 push_ctxt(&saved, &obd->obd_ctxt, NULL);
324 /* Close any open files (which may also cause orphan unlinking). */
325 spin_lock(&med->med_open_lock);
326 while (!list_empty(&med->med_open_head)) {
327 struct list_head *tmp = med->med_open_head.next;
328 struct mds_file_data *mfd =
329 list_entry(tmp, struct mds_file_data, mfd_list);
330 BDEVNAME_DECLARE_STORAGE(btmp);
332 /* bug 1579: fix force-closing for 2.5 */
333 struct dentry *dentry = mfd->mfd_dentry;
335 list_del(&mfd->mfd_list);
336 spin_unlock(&med->med_open_lock);
338 CERROR("force closing client file handle for %*s (%s:%lu)\n",
339 dentry->d_name.len, dentry->d_name.name,
340 ll_bdevname(dentry->d_inode->i_sb, btmp),
341 dentry->d_inode->i_ino);
342 rc = mds_mfd_close(NULL, obd, mfd,
343 !(export->exp_flags & OBD_OPT_FAILOVER));
346 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
347 spin_lock(&med->med_open_lock);
349 spin_unlock(&med->med_open_lock);
350 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
352 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
357 static int mds_disconnect(struct obd_export *export, int flags)
359 unsigned long irqflags;
363 ldlm_cancel_locks_for_export(export);
365 /* complete all outstanding replies */
366 spin_lock_irqsave (&export->exp_lock, irqflags);
367 while (!list_empty (&export->exp_outstanding_replies)) {
368 struct ptlrpc_reply_state *rs =
369 list_entry (export->exp_outstanding_replies.next,
370 struct ptlrpc_reply_state, rs_exp_list);
371 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
373 spin_lock (&svc->srv_lock);
374 list_del_init (&rs->rs_exp_list);
375 ptlrpc_schedule_difficult_reply (rs);
376 spin_unlock (&svc->srv_lock);
378 spin_unlock_irqrestore (&export->exp_lock, irqflags);
380 spin_lock_irqsave(&export->exp_lock, irqflags);
381 export->exp_flags = flags;
382 spin_unlock_irqrestore(&export->exp_lock, irqflags);
384 rc = class_disconnect(export, flags);
388 static int mds_getstatus(struct ptlrpc_request *req)
390 struct mds_obd *mds = mds_req2mds(req);
391 struct mds_body *body;
392 int rc, size = sizeof(*body);
395 rc = lustre_pack_reply(req, 1, &size, NULL);
396 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
397 CERROR("mds: out of memory for message: size=%d\n", size);
398 req->rq_status = -ENOMEM; /* superfluous? */
402 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
403 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
405 /* the last_committed and last_xid fields are filled in for all
406 * replies already - no need to do so here also.
411 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
412 void *data, int flag)
417 if (flag == LDLM_CB_CANCELING) {
418 /* Don't need to do anything here. */
422 /* XXX layering violation! -phil */
423 l_lock(&lock->l_resource->lr_namespace->ns_lock);
424 /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
425 * such that mds_blocking_ast is called just before l_i_p takes the
426 * ns_lock, then by the time we get the lock, we might not be the
427 * correct blocking function anymore. So check, and return early, if
429 if (lock->l_blocking_ast != mds_blocking_ast) {
430 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
434 lock->l_flags |= LDLM_FL_CBPENDING;
435 do_ast = (!lock->l_readers && !lock->l_writers);
436 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
439 struct lustre_handle lockh;
442 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
443 ldlm_lock2handle(lock, &lockh);
444 rc = ldlm_cli_cancel(&lockh);
446 CERROR("ldlm_cli_cancel: %d\n", rc);
448 LDLM_DEBUG(lock, "Lock still has references, will be "
454 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
455 * Call with lock=0 if the caller has already taken the i_sem. */
456 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
457 struct mds_body *body, struct inode *inode, int lock)
459 struct mds_obd *mds = &obd->u.mds;
465 lmm = lustre_msg_buf(msg, offset, 0);
467 /* Some problem with getting eadata when I sized the reply
469 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
473 lmm_size = msg->buflens[offset];
475 /* I don't really like this, but it is a sanity check on the client
476 * MD request. However, if the client doesn't know how much space
477 * to reserve for the MD, it shouldn't be bad to have too much space.
479 if (lmm_size > mds->mds_max_mdsize) {
480 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
481 inode->i_ino, lmm_size, mds->mds_max_mdsize);
487 rc = fsfilt_get_md(obd, inode, lmm, lmm_size);
491 CERROR("Error %d reading eadata for ino %lu\n",
495 rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
499 body->valid |= OBD_MD_FLEASIZE;
500 body->eadatasize = lmm_size;
507 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
508 struct ptlrpc_request *req,
509 struct mds_body *reqbody, int reply_off)
511 struct mds_body *body;
512 struct inode *inode = dentry->d_inode;
519 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
520 LASSERT(body != NULL); /* caller prepped reply */
522 mds_pack_inode2fid(&body->fid1, inode);
523 mds_pack_inode2body(body, inode);
525 if (S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
526 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
529 /* If we have LOV EA data, the OST holds size, atime, mtime */
530 if (!(body->valid & OBD_MD_FLEASIZE))
531 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
532 OBD_MD_FLATIME | OBD_MD_FLMTIME);
533 } else if (S_ISLNK(inode->i_mode) &&
534 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
535 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
538 LASSERT (symname != NULL); /* caller prepped reply */
539 len = req->rq_repmsg->buflens[reply_off + 1];
541 rc = inode->i_op->readlink(dentry, symname, len);
543 CERROR("readlink failed: %d\n", rc);
544 } else if (rc != len - 1) {
545 CERROR ("Unexpected readlink rc %d: expecting %d\n",
549 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
550 body->valid |= OBD_MD_LINKNAME;
551 body->eadatasize = rc + 1;
552 symname[rc] = 0; /* NULL terminate */
560 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
563 struct mds_obd *mds = mds_req2mds(req);
564 struct mds_body *body;
565 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
568 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
569 LASSERT(body != NULL); /* checked by caller */
570 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
572 if (S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) {
575 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
577 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
581 CERROR("error getting inode %lu MD: rc = %d\n",
584 } else if (rc > mds->mds_max_mdsize) {
586 CERROR("MD size %d larger than maximum possible %u\n",
587 rc, mds->mds_max_mdsize);
592 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
593 if (inode->i_size + 1 != body->eadatasize)
594 CERROR("symlink size: %Lu, reply space: %d\n",
595 inode->i_size + 1, body->eadatasize);
596 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
598 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
599 inode->i_size + 1, body->eadatasize);
602 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
603 CERROR("failed MDS_GETATTR_PACK test\n");
604 req->rq_status = -ENOMEM;
605 GOTO(out, rc = -ENOMEM);
608 rc = lustre_pack_reply(req, bufcount, size, NULL);
610 CERROR("out of memory\n");
611 GOTO(out, req->rq_status = rc);
619 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
620 struct lustre_handle *child_lockh)
622 struct obd_device *obd = req->rq_export->exp_obd;
623 struct ldlm_reply *rep = NULL;
624 struct obd_run_ctxt saved;
625 struct mds_body *body;
626 struct dentry *dparent = NULL, *dchild = NULL;
628 struct lustre_handle parent_lockh;
630 int rc = 0, cleanup_phase = 0, resent_req = 0;
634 LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
636 /* Swab now, before anyone looks inside the request */
638 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
639 lustre_swab_mds_body);
641 CERROR("Can't swab mds_body\n");
642 GOTO(cleanup, rc = -EFAULT);
645 LASSERT_REQSWAB(req, offset + 1);
646 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
648 CERROR("Can't unpack name\n");
649 GOTO(cleanup, rc = -EFAULT);
651 namesize = req->rq_reqmsg->buflens[offset + 1];
653 LASSERT (offset == 0 || offset == 2);
654 /* if requests were at offset 2, the getattr reply goes back at 1 */
656 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
660 uc.ouc_fsuid = body->fsuid;
661 uc.ouc_fsgid = body->fsgid;
662 uc.ouc_cap = body->capability;
663 uc.ouc_suppgid1 = body->suppgid;
664 uc.ouc_suppgid2 = -1;
665 push_ctxt(&saved, &obd->obd_ctxt, &uc);
666 cleanup_phase = 1; /* kernel context */
667 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
669 /* FIXME: handle raw lookup */
671 if (body->valid == OBD_MD_FLID) {
672 struct mds_body *mds_reply;
673 int size = sizeof(*mds_reply);
675 // The user requested ONLY the inode number, so do a raw lookup
676 rc = lustre_pack_reply(req, 1, &size, NULL);
678 CERROR("out of memory\n");
682 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
684 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
686 mds_reply->fid1.id = inum;
687 mds_reply->valid = OBD_MD_FLID;
692 if (child_lockh->cookie != 0) {
693 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
697 if (resent_req == 0) {
698 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
699 &parent_lockh, &dparent,
700 LCK_PR, name, namesize,
701 child_lockh, &dchild, LCK_PR);
705 struct ldlm_lock *granted_lock;
706 struct ll_fid child_fid;
707 struct ldlm_resource *res;
708 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
709 granted_lock = ldlm_handle2lock(child_lockh);
710 LASSERT(granted_lock);
712 res = granted_lock->l_resource;
713 child_fid.id = res->lr_name.name[0];
714 child_fid.generation = res->lr_name.name[1];
715 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
717 LDLM_LOCK_PUT(granted_lock);
720 cleanup_phase = 2; /* dchild, dparent, locks */
722 if (dchild->d_inode == NULL) {
723 intent_set_disposition(rep, DISP_LOOKUP_NEG);
724 /* in the intent case, the policy clears this error:
725 the disposition is enough */
726 GOTO(cleanup, rc = -ENOENT);
728 intent_set_disposition(rep, DISP_LOOKUP_POS);
731 if (req->rq_repmsg == NULL) {
732 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
734 CERROR ("mds_getattr_pack_msg: %d\n", rc);
739 rc = mds_getattr_internal(obd, dchild, req, body, offset);
740 GOTO(cleanup, rc); /* returns the lock to the client */
743 switch (cleanup_phase) {
745 if (resent_req == 0) {
746 if (rc && dchild->d_inode)
747 ldlm_lock_decref(child_lockh, LCK_PR);
748 ldlm_lock_decref(&parent_lockh, LCK_PR);
753 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
759 static int mds_getattr(int offset, struct ptlrpc_request *req)
761 struct mds_obd *mds = mds_req2mds(req);
762 struct obd_device *obd = req->rq_export->exp_obd;
763 struct obd_run_ctxt saved;
765 struct mds_body *body;
770 body = lustre_swab_reqbuf (req, offset, sizeof (*body),
771 lustre_swab_mds_body);
773 CERROR ("Can't unpack body\n");
777 uc.ouc_fsuid = body->fsuid;
778 uc.ouc_fsgid = body->fsgid;
779 uc.ouc_cap = body->capability;
780 push_ctxt(&saved, &obd->obd_ctxt, &uc);
781 de = mds_fid2dentry(mds, &body->fid1, NULL);
783 rc = req->rq_status = -ENOENT;
784 GOTO(out_pop, PTR_ERR(de));
787 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
789 CERROR ("mds_getattr_pack_msg: %d\n", rc);
793 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
798 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
803 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
804 unsigned long max_age)
808 spin_lock(&obd->obd_osfs_lock);
809 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
811 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
812 spin_unlock(&obd->obd_osfs_lock);
817 static int mds_statfs(struct ptlrpc_request *req)
819 struct obd_device *obd = req->rq_export->exp_obd;
820 int rc, size = sizeof(struct obd_statfs);
823 rc = lustre_pack_reply(req, 1, &size, NULL);
824 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
825 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
829 /* We call this so that we can cache a bit - 1 jiffie worth */
830 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
833 CERROR("mds_obd_statfs failed: rc %d\n", rc);
843 static int mds_sync(struct ptlrpc_request *req)
845 struct obd_device *obd = req->rq_export->exp_obd;
846 struct mds_obd *mds = &obd->u.mds;
847 struct mds_body *body;
848 int rc, size = sizeof(*body);
851 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
853 GOTO(out, rc = -EPROTO);
855 rc = lustre_pack_reply(req, 1, &size, NULL);
856 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
857 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
861 if (body->fid1.id == 0) {
862 /* a fid of zero is taken to mean "sync whole filesystem" */
863 rc = fsfilt_sync(obd, mds->mds_sb);
867 /* just any file to grab fsync method - "file" arg unused */
868 struct file *file = mds->mds_rcvd_filp;
871 de = mds_fid2dentry(mds, &body->fid1, NULL);
873 GOTO(out, rc = PTR_ERR(de));
875 rc = file->f_op->fsync(NULL, de, 1);
880 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
881 mds_pack_inode2fid(&body->fid1, de->d_inode);
882 mds_pack_inode2body(body, de->d_inode);
889 /* mds_readpage does not take a DLM lock on the inode, because the client must
890 * already have a PR lock.
892 * If we were to take another one here, a deadlock will result, if another
893 * thread is already waiting for a PW lock. */
894 static int mds_readpage(struct ptlrpc_request *req)
896 struct obd_device *obd = req->rq_export->exp_obd;
897 struct vfsmount *mnt;
900 struct mds_body *body, *repbody;
901 struct obd_run_ctxt saved;
902 int rc, size = sizeof(*repbody);
906 rc = lustre_pack_reply(req, 1, &size, NULL);
907 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
908 CERROR("mds: out of memory\n");
909 GOTO(out, rc = -ENOMEM);
912 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
914 GOTO (out, rc = -EFAULT);
916 uc.ouc_fsuid = body->fsuid;
917 uc.ouc_fsgid = body->fsgid;
918 uc.ouc_cap = body->capability;
919 push_ctxt(&saved, &obd->obd_ctxt, &uc);
920 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
922 GOTO(out_pop, rc = PTR_ERR(de));
924 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
926 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
927 /* note: in case of an error, dentry_open puts dentry */
929 GOTO(out_pop, rc = PTR_ERR(file));
931 /* body->size is actually the offset -eeb */
932 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
933 CERROR("offset "LPU64" not on a block boundary of %lu\n",
934 body->size, de->d_inode->i_blksize);
935 GOTO(out_file, rc = -EFAULT);
938 /* body->nlink is actually the #bytes to read -eeb */
939 if (body->nlink & (de->d_inode->i_blksize - 1)) {
940 CERROR("size %u is not multiple of blocksize %lu\n",
941 body->nlink, de->d_inode->i_blksize);
942 GOTO(out_file, rc = -EFAULT);
945 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
946 repbody->size = file->f_dentry->d_inode->i_size;
947 repbody->valid = OBD_MD_FLSIZE;
949 /* to make this asynchronous make sure that the handling function
950 doesn't send a reply when this function completes. Instead a
951 callback function would send the reply */
952 /* body->size is actually the offset -eeb */
953 rc = mds_sendpage(req, file, body->size, body->nlink);
958 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
964 int mds_reint(struct ptlrpc_request *req, int offset,
965 struct lustre_handle *lockh)
967 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
970 OBD_ALLOC(rec, sizeof(*rec));
974 rc = mds_update_unpack(req, offset, rec);
975 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
976 CERROR("invalid record\n");
977 GOTO(out, req->rq_status = -EINVAL);
979 /* rc will be used to interrupt a for loop over multiple records */
980 rc = mds_reint_rec(rec, offset, req, lockh);
982 OBD_FREE(rec, sizeof(*rec));
986 static int mds_filter_recovery_request(struct ptlrpc_request *req,
987 struct obd_device *obd, int *process)
989 switch (req->rq_reqmsg->opc) {
990 case MDS_CONNECT: /* This will never get here, but for completeness. */
991 case OST_CONNECT: /* This will never get here, but for completeness. */
998 case MDS_SYNC: /* used in unmounting */
1002 *process = target_queue_recovery_request(req, obd);
1006 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1008 /* XXX what should we set rq_status to here? */
1009 req->rq_status = -EAGAIN;
1010 RETURN(ptlrpc_error(req));
1014 static char *reint_names[] = {
1015 [REINT_SETATTR] "setattr",
1016 [REINT_CREATE] "create",
1017 [REINT_LINK] "link",
1018 [REINT_UNLINK] "unlink",
1019 [REINT_RENAME] "rename",
1020 [REINT_OPEN] "open",
1023 int mds_handle(struct ptlrpc_request *req)
1025 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1027 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1028 struct obd_device *obd = NULL;
1031 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1033 LASSERT(current->journal_info == NULL);
1034 /* XXX identical to OST */
1035 if (req->rq_reqmsg->opc != MDS_CONNECT) {
1036 struct mds_export_data *med;
1037 int recovering, abort_recovery;
1039 if (req->rq_export == NULL) {
1040 CERROR("lustre_mds: operation %d on unconnected MDS\n",
1041 req->rq_reqmsg->opc);
1042 req->rq_status = -ENOTCONN;
1043 GOTO(out, rc = -ENOTCONN);
1046 med = &req->rq_export->exp_mds_data;
1047 obd = req->rq_export->exp_obd;
1050 /* Check for aborted recovery. */
1051 spin_lock_bh(&obd->obd_processing_task_lock);
1052 abort_recovery = obd->obd_abort_recovery;
1053 recovering = obd->obd_recovering;
1054 spin_unlock_bh(&obd->obd_processing_task_lock);
1055 if (abort_recovery) {
1056 target_abort_recovery(obd);
1057 } else if (recovering) {
1058 rc = mds_filter_recovery_request(req, obd,
1060 if (rc || !should_process)
1065 switch (req->rq_reqmsg->opc) {
1067 DEBUG_REQ(D_INODE, req, "connect");
1068 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1069 rc = target_handle_connect(req, mds_handle);
1071 /* Now that we have an export, set mds. */
1072 mds = mds_req2mds(req);
1075 case MDS_DISCONNECT:
1076 DEBUG_REQ(D_INODE, req, "disconnect");
1077 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1078 rc = target_handle_disconnect(req);
1079 req->rq_status = rc; /* superfluous? */
1083 DEBUG_REQ(D_INODE, req, "getstatus");
1084 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1085 rc = mds_getstatus(req);
1089 DEBUG_REQ(D_INODE, req, "getattr");
1090 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1091 rc = mds_getattr(0, req);
1094 case MDS_GETATTR_NAME: {
1095 struct lustre_handle lockh;
1096 DEBUG_REQ(D_INODE, req, "getattr_name");
1097 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1099 /* If this request gets a reconstructed reply, we won't be
1100 * acquiring any new locks in mds_getattr_name, so we don't
1104 rc = mds_getattr_name(0, req, &lockh);
1105 /* this non-intent call (from an ioctl) is special */
1106 req->rq_status = rc;
1107 if (rc == 0 && lockh.cookie)
1108 ldlm_lock_decref(&lockh, LCK_PR);
1112 DEBUG_REQ(D_INODE, req, "statfs");
1113 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1114 rc = mds_statfs(req);
1118 DEBUG_REQ(D_INODE, req, "readpage");
1119 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1120 rc = mds_readpage(req);
1122 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1123 if (req->rq_reply_state) {
1124 lustre_free_reply_state (req->rq_reply_state);
1125 req->rq_reply_state = NULL;
1133 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1135 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1136 mds->mds_max_cookiesize};
1139 /* NB only peek inside req now; mds_reint() will swab it */
1141 CERROR ("Can't inspect opcode\n");
1146 if (lustre_msg_swabbed (req->rq_reqmsg))
1149 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1150 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1151 reint_names[opc] == NULL) ? reint_names[opc] :
1154 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1156 if (opc == REINT_UNLINK)
1158 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1163 rc = lustre_pack_reply(req, bufcount, size, NULL);
1167 rc = mds_reint(req, 0, NULL);
1168 fail = OBD_FAIL_MDS_REINT_NET_REP;
1173 DEBUG_REQ(D_INODE, req, "close");
1174 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1175 rc = mds_close(req);
1178 case MDS_DONE_WRITING:
1179 DEBUG_REQ(D_INODE, req, "done_writing");
1180 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1181 rc = mds_done_writing(req);
1185 DEBUG_REQ(D_INODE, req, "pin");
1186 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1191 DEBUG_REQ(D_INODE, req, "sync");
1192 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1197 DEBUG_REQ(D_INODE, req, "ping");
1198 rc = target_handle_ping(req);
1201 case OBD_LOG_CANCEL:
1202 CDEBUG(D_INODE, "log cancel\n");
1203 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1204 rc = -ENOTSUPP; /* la la la */
1208 DEBUG_REQ(D_INODE, req, "enqueue");
1209 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1210 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1211 ldlm_server_blocking_ast, NULL);
1214 DEBUG_REQ(D_INODE, req, "convert");
1215 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1216 rc = ldlm_handle_convert(req);
1218 case LDLM_BL_CALLBACK:
1219 case LDLM_CP_CALLBACK:
1220 DEBUG_REQ(D_INODE, req, "callback");
1221 CERROR("callbacks should not happen on MDS\n");
1223 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1225 case LLOG_ORIGIN_HANDLE_CREATE:
1226 DEBUG_REQ(D_INODE, req, "llog_init");
1227 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1228 rc = llog_origin_handle_create(req);
1230 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1231 DEBUG_REQ(D_INODE, req, "llog next block");
1232 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1233 rc = llog_origin_handle_next_block(req);
1235 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1236 DEBUG_REQ(D_INODE, req, "llog read header");
1237 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1238 rc = llog_origin_handle_read_header(req);
1240 case LLOG_ORIGIN_HANDLE_CLOSE:
1241 DEBUG_REQ(D_INODE, req, "llog close");
1242 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1243 rc = llog_origin_handle_close(req);
1246 DEBUG_REQ(D_INODE, req, "llog catinfo");
1247 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1248 rc = llog_catinfo(req);
1251 req->rq_status = -ENOTSUPP;
1252 rc = ptlrpc_error(req);
1256 LASSERT(current->journal_info == NULL);
1260 /* If we're DISCONNECTing, the mds_export_data is already freed */
1261 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1262 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1263 struct obd_device *obd = list_entry(mds, struct obd_device,
1265 req->rq_repmsg->last_xid =
1266 le64_to_cpu(med->med_mcd->mcd_last_xid);
1268 if (!obd->obd_no_transno) {
1269 req->rq_repmsg->last_committed =
1270 obd->obd_last_committed;
1272 DEBUG_REQ(D_IOCTL, req,
1273 "not sending last_committed update");
1275 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1277 mds->mds_last_transno, obd->obd_last_committed,
1282 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1283 if (obd && obd->obd_recovering) {
1284 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1285 return target_queue_final_reply(req, rc);
1287 /* Lost a race with recovery; let the error path DTRT. */
1288 rc = req->rq_status = -ENOTCONN;
1291 target_send_reply(req, rc, fail);
1295 /* Update the server data on disk. This stores the new mount_count and
1296 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1297 * then the server last_rcvd value may be less than that of the clients.
1298 * This will alert us that we may need to do client recovery.
1300 * Also assumes for mds_last_transno that we are not modifying it (no locking).
1302 int mds_update_server_data(struct obd_device *obd, int force_sync)
1304 struct mds_obd *mds = &obd->u.mds;
1305 struct mds_server_data *msd = mds->mds_server_data;
1306 struct file *filp = mds->mds_rcvd_filp;
1307 struct obd_run_ctxt saved;
1312 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1313 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1315 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1316 mds->mds_mount_count, mds->mds_last_transno);
1317 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1319 CERROR("error writing MDS server data: rc = %d\n", rc);
1320 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1326 /* mount the file system (secretly) */
1327 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1329 struct lustre_cfg* lcfg = buf;
1330 struct mds_obd *mds = &obd->u.mds;
1331 struct vfsmount *mnt;
1336 dev_clear_rdonly(2);
1338 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1339 RETURN(rc = -EINVAL);
1341 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1342 if (IS_ERR(obd->obd_fsops))
1343 RETURN(rc = PTR_ERR(obd->obd_fsops));
1345 if (!(page = __get_free_page(GFP_KERNEL)))
1348 memset((void *)page, 0, PAGE_SIZE);
1349 sprintf((char *)page, "iopen_nopriv");
1351 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1352 lcfg->lcfg_inlbuf1, (void *)page);
1356 CERROR("do_kern_mount failed: rc = %d\n", rc);
1360 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1362 sema_init(&mds->mds_orphan_recovery_sem, 1);
1363 sema_init(&mds->mds_epoch_sem, 1);
1364 spin_lock_init(&mds->mds_transno_lock);
1365 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1366 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1367 atomic_set(&mds->mds_open_count, 0);
1369 obd->obd_namespace = ldlm_namespace_new("mds_server",
1370 LDLM_NAMESPACE_SERVER);
1371 if (obd->obd_namespace == NULL) {
1372 mds_cleanup(obd, 0);
1373 GOTO(err_put, rc = -ENOMEM);
1375 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1377 rc = mds_fs_setup(obd, mnt);
1379 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1383 rc = llog_start_commit_thread();
1388 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1391 generate_random_uuid(uuid);
1392 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1394 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1395 if (mds->mds_profile == NULL)
1396 GOTO(err_fs, rc = -ENOMEM);
1398 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1399 lcfg->lcfg_inllen3);
1403 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1404 "mds_ldlm_client", &obd->obd_ldlm_client);
1405 obd->obd_replayable = 1;
1407 rc = mds_postsetup(obd);
1413 /* No extra cleanup needed for llog_init_commit_thread() */
1414 mds_fs_cleanup(obd, 0);
1416 ldlm_namespace_free(obd->obd_namespace, 0);
1417 obd->obd_namespace = NULL;
1420 mntput(mds->mds_vfsmnt);
1424 fsfilt_put_ops(obd->obd_fsops);
1428 static int mds_postsetup(struct obd_device *obd)
1430 struct mds_obd *mds = &obd->u.mds;
1435 rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
1440 if (mds->mds_profile) {
1441 struct obd_run_ctxt saved;
1442 struct lustre_profile *lprof;
1443 struct config_llog_instance cfg;
1445 cfg.cfg_instance = NULL;
1446 cfg.cfg_uuid = mds->mds_lov_uuid;
1447 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1448 rc = class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1449 mds->mds_profile, &cfg);
1450 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1454 lprof = class_get_profile(mds->mds_profile);
1455 if (lprof == NULL) {
1456 CERROR("No profile found: %s\n", mds->mds_profile);
1457 GOTO(err_cleanup, rc = -ENOENT);
1459 rc = mds_lov_connect(obd, lprof->lp_osc);
1461 GOTO(err_cleanup, rc);
1469 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1473 static int mds_postrecov(struct obd_device *obd)
1478 LASSERT(!obd->obd_recovering);
1479 LASSERT(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT) != NULL);
1481 rc = llog_connect(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT),
1482 obd->u.mds.mds_lov_desc.ld_tgt_count, NULL, NULL);
1484 CERROR("faild at llog_origin_connect: %d\n", rc);
1487 rc = mds_cleanup_orphans(obd);
1489 rc2 = mds_lov_set_nextid(obd);
1495 int mds_lov_clean(struct obd_device *obd)
1497 struct mds_obd *mds = &obd->u.mds;
1499 if (mds->mds_profile) {
1501 struct config_llog_instance cfg;
1502 struct obd_run_ctxt saved;
1503 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1505 OBD_ALLOC(cln_prof, len);
1506 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1508 cfg.cfg_instance = NULL;
1509 cfg.cfg_uuid = mds->mds_lov_uuid;
1511 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1512 class_config_parse_llog(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT),
1514 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1516 OBD_FREE(cln_prof, len);
1517 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1518 mds->mds_profile = NULL;
1523 static int mds_precleanup(struct obd_device *obd, int flags)
1528 mds_lov_disconnect(obd, flags);
1530 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
1534 static int mds_cleanup(struct obd_device *obd, int flags)
1536 struct mds_obd *mds = &obd->u.mds;
1539 if (mds->mds_sb == NULL)
1542 mds_update_server_data(obd, 1);
1543 if (mds->mds_lov_objids != NULL) {
1544 OBD_FREE(mds->mds_lov_objids,
1545 mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
1547 mds_fs_cleanup(obd, flags);
1551 /* 2 seems normal on mds, (may_umount() also expects 2
1552 fwiw), but we only see 1 at this point in obdfilter. */
1553 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
1554 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
1555 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
1557 mntput(mds->mds_vfsmnt);
1561 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
1563 spin_lock_bh(&obd->obd_processing_task_lock);
1564 if (obd->obd_recovering) {
1565 target_cancel_recovery_timer(obd);
1566 obd->obd_recovering = 0;
1568 spin_unlock_bh(&obd->obd_processing_task_lock);
1571 dev_clear_rdonly(2);
1572 fsfilt_put_ops(obd->obd_fsops);
1577 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
1578 struct ldlm_lock *new_lock,
1579 struct lustre_handle *lockh)
1581 struct obd_export *exp = req->rq_export;
1582 struct obd_device *obd = exp->exp_obd;
1583 struct ldlm_request *dlmreq =
1584 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
1585 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
1586 struct list_head *iter;
1588 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
1591 l_lock(&obd->obd_namespace->ns_lock);
1592 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
1593 struct ldlm_lock *lock;
1594 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
1595 if (lock == new_lock)
1597 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
1598 lockh->cookie = lock->l_handle.h_cookie;
1599 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
1601 l_unlock(&obd->obd_namespace->ns_lock);
1605 l_unlock(&obd->obd_namespace->ns_lock);
1607 /* This remote handle isn't enqueued, so we never received or
1608 * processed this request. Clear MSG_RESENT, because it can
1609 * be handled like any normal request now. */
1611 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1613 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
1617 int intent_disposition(struct ldlm_reply *rep, int flag)
1621 return (rep->lock_policy_res1 & flag);
1624 void intent_set_disposition(struct ldlm_reply *rep, int flag)
1628 rep->lock_policy_res1 |= flag;
1631 static int mds_intent_policy(struct ldlm_namespace *ns,
1632 struct ldlm_lock **lockp, void *req_cookie,
1633 ldlm_mode_t mode, int flags, void *data)
1635 struct ptlrpc_request *req = req_cookie;
1636 struct ldlm_lock *lock = *lockp;
1637 struct ldlm_intent *it;
1638 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1639 struct ldlm_reply *rep;
1640 struct lustre_handle lockh = { 0 };
1641 struct ldlm_lock *new_lock;
1642 int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
1643 sizeof(struct mds_body),
1644 mds->mds_max_mdsize,
1645 mds->mds_max_cookiesize};
1648 LASSERT(req != NULL);
1650 if (req->rq_reqmsg->bufcount <= 1) {
1651 /* No intent was provided */
1652 int size = sizeof(struct ldlm_reply);
1653 rc = lustre_pack_reply(req, 1, &size, NULL);
1658 it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
1660 CERROR("Intent missing\n");
1661 RETURN(req->rq_status = -EFAULT);
1664 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
1666 rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3, repsize,
1669 RETURN(req->rq_status = rc);
1671 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1672 intent_set_disposition(rep, DISP_IT_EXECD);
1674 fixup_handle_for_resent_req(req, lock, &lockh);
1676 /* execute policy */
1677 switch ((long)it->opc) {
1679 case IT_CREAT|IT_OPEN:
1680 /* XXX swab here to assert that an mds_open reint
1681 * packet is following */
1682 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
1684 /* We abort the lock if the lookup was negative and
1685 * we did not make it to the OPEN portion */
1686 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
1687 RETURN(ELDLM_LOCK_ABORTED);
1688 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
1689 !intent_disposition(rep, DISP_OPEN_OPEN))
1691 RETURN(ELDLM_LOCK_ABORTED);
1696 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh);
1697 /* FIXME: LDLM can set req->rq_status. MDS sets
1698 policy_res{1,2} with disposition and status.
1699 - replay: returns 0 & req->status is old status
1700 - otherwise: returns req->status */
1701 if (intent_disposition(rep, DISP_LOOKUP_NEG))
1702 rep->lock_policy_res2 = 0;
1703 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
1704 rep->lock_policy_res2)
1705 RETURN(ELDLM_LOCK_ABORTED);
1706 if (req->rq_status != 0) {
1708 rep->lock_policy_res2 = req->rq_status;
1709 RETURN(ELDLM_LOCK_ABORTED);
1713 CERROR("Unhandled intent "LPD64"\n", it->opc);
1717 /* By this point, whatever function we called above must have either
1718 * filled in 'lockh', been an intent replay, or returned an error. We
1719 * want to allow replayed RPCs to not get a lock, since we would just
1720 * drop it below anyways because lock replay is done separately by the
1721 * client afterwards. For regular RPCs we want to give the new lock to
1722 * the client instead of whatever lock it was about to get. */
1723 new_lock = ldlm_handle2lock(&lockh);
1724 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
1727 LASSERT(new_lock != NULL);
1729 /* If we've already given this lock to a client once, then we should
1730 * have no readers or writers. Otherwise, we should have one reader
1731 * _or_ writer ref (which will be zeroed below) before returning the
1732 * 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);
1773 int mds_attach(struct obd_device *dev, obd_count len, void *data)
1775 struct lprocfs_static_vars lvars;
1777 lprocfs_init_multi_vars(0, &lvars);
1778 return lprocfs_obd_attach(dev, lvars.obd_vars);
1781 int mds_detach(struct obd_device *dev)
1783 return lprocfs_obd_detach(dev);
1786 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
1788 struct lprocfs_static_vars lvars;
1790 lprocfs_init_multi_vars(1, &lvars);
1791 return lprocfs_obd_attach(dev, lvars.obd_vars);
1794 int mdt_detach(struct obd_device *dev)
1796 return lprocfs_obd_detach(dev);
1799 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1801 struct mds_obd *mds = &obddev->u.mds;
1806 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1807 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1809 obddev->obd_proc_entry);
1811 if (!mds->mds_service) {
1812 CERROR("failed to start service\n");
1813 RETURN(rc = -ENOMEM);
1816 rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
1819 GOTO(err_thread, rc);
1821 mds->mds_setattr_service =
1822 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1823 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
1824 mds_handle, "mds_setattr",
1825 obddev->obd_proc_entry);
1826 if (!mds->mds_setattr_service) {
1827 CERROR("failed to start getattr service\n");
1828 GOTO(err_thread, rc = -ENOMEM);
1831 rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
1832 MDT_NUM_THREADS, "ll_mdt_attr");
1834 GOTO(err_thread2, rc);
1836 mds->mds_readpage_service =
1837 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
1838 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
1839 mds_handle, "mds_readpage",
1840 obddev->obd_proc_entry);
1841 if (!mds->mds_readpage_service) {
1842 CERROR("failed to start readpage service\n");
1843 GOTO(err_thread2, rc = -ENOMEM);
1846 rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
1847 MDT_NUM_THREADS, "ll_mdt_rdpg");
1850 GOTO(err_thread3, rc);
1855 ptlrpc_unregister_service(mds->mds_readpage_service);
1857 ptlrpc_unregister_service(mds->mds_setattr_service);
1859 ptlrpc_unregister_service(mds->mds_service);
1864 static int mdt_cleanup(struct obd_device *obddev, int flags)
1866 struct mds_obd *mds = &obddev->u.mds;
1869 ptlrpc_stop_all_threads(mds->mds_readpage_service);
1870 ptlrpc_unregister_service(mds->mds_readpage_service);
1872 ptlrpc_stop_all_threads(mds->mds_setattr_service);
1873 ptlrpc_unregister_service(mds->mds_setattr_service);
1875 ptlrpc_stop_all_threads(mds->mds_service);
1876 ptlrpc_unregister_service(mds->mds_service);
1881 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
1883 struct obd_device *obd = data;
1886 fid.generation = gen;
1887 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
1890 struct lvfs_callback_ops mds_lvfs_ops = {
1891 l_fid2dentry: mds_lvfs_fid2dentry,
1894 /* use obd ops to offer management infrastructure */
1895 static struct obd_ops mds_obd_ops = {
1896 o_owner: THIS_MODULE,
1897 o_attach: mds_attach,
1898 o_detach: mds_detach,
1899 o_connect: mds_connect,
1900 o_init_export: mds_init_export,
1901 o_destroy_export: mds_destroy_export,
1902 o_disconnect: mds_disconnect,
1904 o_precleanup: mds_precleanup,
1905 o_cleanup: mds_cleanup,
1906 o_postrecov: mds_postrecov,
1907 o_statfs: mds_obd_statfs,
1908 o_iocontrol: mds_iocontrol,
1909 o_create: mds_obd_create,
1910 o_destroy: mds_obd_destroy,
1911 o_llog_init: mds_llog_init,
1912 o_llog_finish: mds_llog_finish,
1913 o_notify: mds_notify,
1916 static struct obd_ops mdt_obd_ops = {
1917 o_owner: THIS_MODULE,
1918 o_attach: mdt_attach,
1919 o_detach: mdt_detach,
1921 o_cleanup: mdt_cleanup,
1924 static int __init mds_init(void)
1926 struct lprocfs_static_vars lvars;
1928 lprocfs_init_multi_vars(0, &lvars);
1929 class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
1930 lprocfs_init_multi_vars(1, &lvars);
1931 class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
1936 static void /*__exit*/ mds_exit(void)
1938 class_unregister_type(LUSTRE_MDS_NAME);
1939 class_unregister_type(LUSTRE_MDT_NAME);
1942 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1943 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
1944 MODULE_LICENSE("GPL");
1946 module_init(mds_init);
1947 module_exit(mds_exit);