1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
5 * Lustre Metadata Server (mds) request handler
7 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
8 * Author: Peter Braam <braam@clusterfs.com>
9 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
11 * Author: Mike Shaver <shaver@clusterfs.com>
13 * This file is part of Lustre, http://www.lustre.org.
15 * Lustre is free software; you can redistribute it and/or
16 * modify it under the terms of version 2 of the GNU General Public
17 * License as published by the Free Software Foundation.
19 * Lustre is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * GNU General Public License for more details.
24 * You should have received a copy of the GNU General Public License
25 * along with Lustre; if not, write to the Free Software
26 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
30 # define EXPORT_SYMTAB
32 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/module.h>
35 #include <linux/lustre_mds.h>
36 #include <linux/lustre_dlm.h>
37 #include <linux/init.h>
38 #include <linux/obd_class.h>
39 #include <linux/random.h>
41 #include <linux/jbd.h>
42 #include <linux/ext3_fs.h>
43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
44 # include <linux/smp_lock.h>
45 # include <linux/buffer_head.h>
46 # include <linux/workqueue.h>
47 # include <linux/mount.h>
49 # include <linux/locks.h>
51 #include <linux/obd_lov.h>
52 #include <linux/lustre_mds.h>
53 #include <linux/lustre_fsfilt.h>
54 #include <linux/lprocfs_status.h>
55 #include <linux/lustre_commit_confd.h>
57 #include "mds_internal.h"
59 static int mds_intent_policy(struct ldlm_namespace *ns,
60 struct ldlm_lock **lockp, void *req_cookie,
61 ldlm_mode_t mode, int flags, void *data);
62 static int mds_postsetup(struct obd_device *obd);
63 static int mds_cleanup(struct obd_device *obd, int flags);
65 /* Assumes caller has already pushed into the kernel filesystem context */
66 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
67 loff_t offset, int count)
69 struct ptlrpc_bulk_desc *desc;
70 struct l_wait_info lwi;
72 int rc = 0, npages, i, tmpcount, tmpsize = 0;
75 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
77 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
78 OBD_ALLOC(pages, sizeof(*pages) * npages);
80 GOTO(out, rc = -ENOMEM);
82 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
85 GOTO(out_free, rc = -ENOMEM);
87 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
88 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
90 pages[i] = alloc_pages(GFP_KERNEL, 0);
92 GOTO(cleanup_buf, rc = -ENOMEM);
94 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
97 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
98 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
99 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
100 tmpsize, offset, file->f_dentry->d_inode->i_ino,
101 file->f_dentry->d_inode->i_size);
103 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
104 kmap(pages[i]), tmpsize, &offset);
108 GOTO(cleanup_buf, rc = -EIO);
111 LASSERT(desc->bd_nob == count);
113 rc = ptlrpc_start_bulk_transfer(desc);
115 GOTO(cleanup_buf, rc);
117 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
118 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
119 OBD_FAIL_MDS_SENDPAGE, rc);
120 GOTO(abort_bulk, rc);
123 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
124 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
125 LASSERT (rc == 0 || rc == -ETIMEDOUT);
128 if (desc->bd_success &&
129 desc->bd_nob_transferred == count)
130 GOTO(cleanup_buf, rc);
132 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
135 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
136 (rc == -ETIMEDOUT) ? "timeout" : "network error",
137 desc->bd_nob_transferred, count,
138 req->rq_export->exp_client_uuid.uuid,
139 req->rq_export->exp_connection->c_remote_uuid.uuid);
141 ptlrpc_fail_export(req->rq_export);
145 ptlrpc_abort_bulk (desc);
147 for (i = 0; i < npages; i++)
149 __free_pages(pages[i], 0);
151 ptlrpc_free_bulk(desc);
153 OBD_FREE(pages, sizeof(*pages) * npages);
158 int mds_lock_mode_for_dir(struct obd_device *obd,
159 struct dentry *dentry, int mode)
163 /* any dir access needs couple locks:
164 * 1) on part of dir we gonna lookup/modify in
165 * 2) on a whole dir to protect it from concurrent splitting
166 * and to flush client's cache for readdir()
167 * so, for a given mode and dentry this routine decides what
168 * lock mode to use for lock #2:
169 * 1) if caller's gonna lookup in dir then we need to protect
170 * dir from being splitted only - LCK_CR
171 * 2) if caller's gonna modify dir then we need to protect
172 * dir from being splitted and to flush cache - LCK_CW
173 * 3) if caller's gonna modify dir and that dir seems ready
174 * for splitting then we need to protect it from any
175 * type of access (lookup/modify/split) - LCK_EX -bzzz */
177 split = mds_splitting_expected(obd, dentry);
178 if (split == MDS_NO_SPLITTABLE) {
179 /* this inode won't be splitted. so we need not to protect from
180 * just flush client's cache on modification */
185 if (mode == LCK_PR) {
187 } else if (mode == LCK_PW) {
188 /* caller gonna modify directory.we use concurrent
189 write lock here to retract client's cache for readdir */
191 if (split == MDS_EXPECT_SPLIT) {
192 /* splitting possible. serialize any access */
193 CDEBUG(D_OTHER, "%s: gonna split %u/%u\n",
195 (unsigned) dentry->d_inode->i_ino,
196 (unsigned) dentry->d_inode->i_generation);
200 CWARN("unexpected lock mode %d\n", mode);
207 /* only valid locked dentries or errors should be returned */
208 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
209 struct vfsmount **mnt, int lock_mode,
210 struct lustre_handle *lockh, int *mode,
211 char *name, int namelen, __u64 lockpart)
213 struct mds_obd *mds = &obd->u.mds;
214 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
215 struct ldlm_res_id res_id = { .name = {0} };
217 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
224 res_id.name[0] = de->d_inode->i_ino;
225 res_id.name[1] = de->d_inode->i_generation;
228 if (name && IS_PDIROPS(de->d_inode)) {
229 ldlm_policy_data_t cpolicy =
230 { .l_inodebits = { MDS_INODELOCK_UPDATE } };
231 LASSERT(mode != NULL);
232 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
234 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
236 &cpolicy, *mode, &flags,
238 ldlm_completion_ast, NULL, NULL,
239 NULL, 0, NULL, lockh + 1);
240 if (rc != ELDLM_OK) {
242 RETURN(ERR_PTR(-ENOLCK));
247 res_id.name[2] = full_name_hash(name, namelen);
249 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
250 de->d_inode->i_ino, de->d_inode->i_generation,
254 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
255 LDLM_IBITS, &policy, lock_mode, &flags,
256 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
257 NULL, 0, NULL, lockh);
258 if (rc != ELDLM_OK) {
260 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
263 ldlm_lock_decref(lockh + 1, LCK_CW);
270 #ifndef DCACHE_DISCONNECTED
271 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
275 /* Look up an entry by inode number. */
276 /* this function ONLY returns valid dget'd dentries with an initialized inode
278 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
279 struct vfsmount **mnt)
282 unsigned long ino = fid->id;
283 __u32 generation = fid->generation;
285 struct dentry *result;
288 RETURN(ERR_PTR(-ESTALE));
290 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
292 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
293 ino, generation, mds->mds_sb);
295 /* under ext3 this is neither supposed to return bad inodes
297 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
301 inode = result->d_inode;
303 RETURN(ERR_PTR(-ENOENT));
306 /* here we disabled generation check, as root inode i_generation
307 * of cache mds and real mds are different. */
308 if (generation && inode->i_generation != generation) {
309 /* we didn't find the right inode.. */
310 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
311 inode->i_ino, (unsigned long)inode->i_nlink,
312 atomic_read(&inode->i_count), inode->i_generation,
315 RETURN(ERR_PTR(-ENOENT));
320 *mnt = mds->mds_vfsmnt;
328 /* Establish a connection to the MDS.
330 * This will set up an export structure for the client to hold state data
331 * about that client, like open files, the last operation number it did
332 * on the server, etc.
334 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
335 struct obd_uuid *cluuid)
337 struct obd_export *exp;
338 struct mds_export_data *med; /* */
339 struct mds_client_data *mcd;
340 int rc, abort_recovery;
343 if (!conn || !obd || !cluuid)
346 /* Check for aborted recovery. */
347 spin_lock_bh(&obd->obd_processing_task_lock);
348 abort_recovery = obd->obd_abort_recovery;
349 spin_unlock_bh(&obd->obd_processing_task_lock);
351 target_abort_recovery(obd);
353 /* XXX There is a small race between checking the list and adding a
354 * new connection for the same UUID, but the real threat (list
355 * corruption when multiple different clients connect) is solved.
357 * There is a second race between adding the export to the list,
358 * and filling in the client data below. Hence skipping the case
359 * of NULL mcd above. We should already be controlling multiple
360 * connects at the client, and we can't hold the spinlock over
361 * memory allocations without risk of deadlocking.
363 rc = class_connect(conn, obd, cluuid);
366 exp = class_conn2export(conn);
368 med = &exp->exp_mds_data;
370 OBD_ALLOC(mcd, sizeof(*mcd));
372 CERROR("mds: out of memory for client data\n");
373 GOTO(out, rc = -ENOMEM);
376 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
379 rc = mds_client_add(obd, &obd->u.mds, med, -1);
384 OBD_FREE(mcd, sizeof(*mcd));
385 class_disconnect(exp, 0);
387 class_export_put(exp);
392 static int mds_init_export(struct obd_export *exp)
394 struct mds_export_data *med = &exp->exp_mds_data;
396 INIT_LIST_HEAD(&med->med_open_head);
397 spin_lock_init(&med->med_open_lock);
401 static int mds_destroy_export(struct obd_export *export)
403 struct mds_export_data *med;
404 struct obd_device *obd = export->exp_obd;
405 struct lvfs_run_ctxt saved;
409 med = &export->exp_mds_data;
410 target_destroy_export(export);
412 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
415 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
417 /* Close any open files (which may also cause orphan unlinking). */
418 spin_lock(&med->med_open_lock);
419 while (!list_empty(&med->med_open_head)) {
420 struct list_head *tmp = med->med_open_head.next;
421 struct mds_file_data *mfd =
422 list_entry(tmp, struct mds_file_data, mfd_list);
423 BDEVNAME_DECLARE_STORAGE(btmp);
425 /* bug 1579: fix force-closing for 2.5 */
426 struct dentry *dentry = mfd->mfd_dentry;
428 list_del(&mfd->mfd_list);
429 spin_unlock(&med->med_open_lock);
431 /* If you change this message, be sure to update
432 * replay_single:test_46 */
433 CERROR("force closing client file handle for %*s (%s:%lu)\n",
434 dentry->d_name.len, dentry->d_name.name,
435 ll_bdevname(dentry->d_inode->i_sb, btmp),
436 dentry->d_inode->i_ino);
437 rc = mds_mfd_close(NULL, obd, mfd,
438 !(export->exp_flags & OBD_OPT_FAILOVER));
441 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
442 spin_lock(&med->med_open_lock);
444 spin_unlock(&med->med_open_lock);
445 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
448 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
453 static int mds_disconnect(struct obd_export *exp, int flags)
455 struct obd_device *obd;
457 unsigned long irqflags;
462 class_export_get(exp);
464 obd = class_exp2obd(exp);
466 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
467 exp->exp_handle.h_cookie);
472 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
473 && !atomic_read(&mds->mds_real_clients)) {
474 /* there was no client at all */
475 mds_lmv_disconnect(obd, flags);
478 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
479 && atomic_dec_and_test(&mds->mds_real_clients)) {
480 /* time to drop LMV connections */
481 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
482 "Disconnnect from LMV now\n",
483 obd->obd_name, exp->exp_client_uuid.uuid);
484 mds_lmv_disconnect(obd, flags);
487 spin_lock_irqsave(&exp->exp_lock, irqflags);
488 exp->exp_flags = flags;
489 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
491 /* Disconnect early so that clients can't keep using export */
492 rc = class_disconnect(exp, flags);
493 ldlm_cancel_locks_for_export(exp);
495 /* complete all outstanding replies */
496 spin_lock_irqsave(&exp->exp_lock, irqflags);
497 while (!list_empty(&exp->exp_outstanding_replies)) {
498 struct ptlrpc_reply_state *rs =
499 list_entry(exp->exp_outstanding_replies.next,
500 struct ptlrpc_reply_state, rs_exp_list);
501 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
503 spin_lock(&svc->srv_lock);
504 list_del_init(&rs->rs_exp_list);
505 ptlrpc_schedule_difficult_reply(rs);
506 spin_unlock(&svc->srv_lock);
508 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
510 class_export_put(exp);
514 static int mds_getstatus(struct ptlrpc_request *req)
516 struct mds_obd *mds = mds_req2mds(req);
517 struct mds_body *body;
518 int rc, size = sizeof(*body);
521 rc = lustre_pack_reply(req, 1, &size, NULL);
522 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
523 CERROR("mds: out of memory for message: size=%d\n", size);
524 req->rq_status = -ENOMEM; /* superfluous? */
528 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
529 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
531 /* the last_committed and last_xid fields are filled in for all
532 * replies already - no need to do so here also.
537 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
538 void *data, int flag)
543 if (flag == LDLM_CB_CANCELING) {
544 /* Don't need to do anything here. */
548 /* XXX layering violation! -phil */
549 l_lock(&lock->l_resource->lr_namespace->ns_lock);
550 /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
551 * such that mds_blocking_ast is called just before l_i_p takes the
552 * ns_lock, then by the time we get the lock, we might not be the
553 * correct blocking function anymore. So check, and return early, if
555 if (lock->l_blocking_ast != mds_blocking_ast) {
556 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
560 lock->l_flags |= LDLM_FL_CBPENDING;
561 do_ast = (!lock->l_readers && !lock->l_writers);
562 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
565 struct lustre_handle lockh;
568 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
569 ldlm_lock2handle(lock, &lockh);
570 rc = ldlm_cli_cancel(&lockh);
572 CERROR("ldlm_cli_cancel: %d\n", rc);
574 LDLM_DEBUG(lock, "Lock still has references, will be "
580 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
588 rc = fsfilt_get_md(obd, inode, md, *size);
593 CERROR("Error %d reading eadata for ino %lu\n",
598 if (S_ISREG(inode->i_mode))
599 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
613 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
614 * Call with lock=0 if the caller has already taken the i_sem. */
615 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
616 struct mds_body *body, struct inode *inode, int lock)
618 struct mds_obd *mds = &obd->u.mds;
624 lmm = lustre_msg_buf(msg, offset, 0);
626 /* Some problem with getting eadata when I sized the reply
628 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
632 lmm_size = msg->buflens[offset];
634 /* I don't really like this, but it is a sanity check on the client
635 * MD request. However, if the client doesn't know how much space
636 * to reserve for the MD, it shouldn't be bad to have too much space.
638 if (lmm_size > mds->mds_max_mdsize) {
639 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
640 inode->i_ino, lmm_size, mds->mds_max_mdsize);
644 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
646 if (S_ISDIR(inode->i_mode))
647 body->valid |= OBD_MD_FLDIREA;
649 body->valid |= OBD_MD_FLEASIZE;
650 body->eadatasize = lmm_size;
657 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
658 struct ptlrpc_request *req,
659 struct mds_body *reqbody, int reply_off)
661 struct mds_body *body;
662 struct inode *inode = dentry->d_inode;
666 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
669 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
670 LASSERT(body != NULL); /* caller prepped reply */
672 if (dentry->d_flags & DCACHE_CROSS_REF) {
673 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
674 (unsigned long) dentry->d_mdsnum,
675 (unsigned long) dentry->d_inum,
676 (unsigned long) dentry->d_generation);
677 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
678 body->fid1.id = dentry->d_inum;
679 body->fid1.mds = dentry->d_mdsnum;
680 body->fid1.generation = dentry->d_generation;
683 mds_pack_inode2fid(obd, &body->fid1, inode);
684 mds_pack_inode2body(obd, body, inode);
686 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
687 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
688 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
691 /* If we have LOV EA data, the OST holds size, atime, mtime */
692 if (!(body->valid & OBD_MD_FLEASIZE) &&
693 !(body->valid & OBD_MD_FLDIREA))
694 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
695 OBD_MD_FLATIME | OBD_MD_FLMTIME);
696 } else if (S_ISLNK(inode->i_mode) &&
697 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
698 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
701 LASSERT (symname != NULL); /* caller prepped reply */
702 len = req->rq_repmsg->buflens[reply_off + 1];
704 rc = inode->i_op->readlink(dentry, symname, len);
706 CERROR("readlink failed: %d\n", rc);
707 } else if (rc != len - 1) {
708 CERROR ("Unexpected readlink rc %d: expecting %d\n",
712 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
713 body->valid |= OBD_MD_LINKNAME;
714 body->eadatasize = rc + 1;
715 symname[rc] = 0; /* NULL terminate */
723 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
724 struct dentry *dentry,
727 int rc = 0, size[1] = {sizeof(struct mds_body)};
730 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
731 CERROR("failed MDS_GETATTR_PACK test\n");
732 req->rq_status = -ENOMEM;
733 GOTO(out, rc = -ENOMEM);
736 rc = lustre_pack_reply(req, 1, size, NULL);
738 CERROR("out of memory\n");
739 GOTO(out, req->rq_status = rc);
747 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
750 struct mds_obd *mds = mds_req2mds(req);
751 struct mds_body *body;
752 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
755 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
756 LASSERT(body != NULL); /* checked by caller */
757 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
759 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
760 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
763 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
765 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
769 CERROR("error getting inode %lu MD: rc = %d\n",
772 } else if (rc > mds->mds_max_mdsize) {
774 CERROR("MD size %d larger than maximum possible %u\n",
775 rc, mds->mds_max_mdsize);
780 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
781 if (inode->i_size + 1 != body->eadatasize)
782 CERROR("symlink size: %Lu, reply space: %d\n",
783 inode->i_size + 1, body->eadatasize);
784 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
786 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
787 inode->i_size + 1, body->eadatasize);
790 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
791 CERROR("failed MDS_GETATTR_PACK test\n");
792 req->rq_status = -ENOMEM;
793 GOTO(out, rc = -ENOMEM);
796 rc = lustre_pack_reply(req, bufcount, size, NULL);
798 CERROR("out of memory\n");
799 GOTO(out, req->rq_status = rc);
807 int mds_check_mds_num(struct obd_device *obd, struct inode* inode,
808 char *name, int namelen)
810 struct mea *mea = NULL;
811 int mea_size, rc = 0;
814 rc = mds_get_lmv_attr(obd, inode, &mea, &mea_size);
818 /* dir is already splitted, check if requested filename
819 * should live at this MDS or at another one */
821 i = mea_name2idx(mea, name, namelen - 1);
822 if (mea->mea_master != i) {
823 CERROR("inapropriate MDS(%d) for %s. should be %d\n",
824 mea->mea_master, name, i);
830 OBD_FREE(mea, mea_size);
834 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
835 struct lustre_handle *child_lockh, int child_part)
837 struct obd_device *obd = req->rq_export->exp_obd;
838 struct ldlm_reply *rep = NULL;
839 struct lvfs_run_ctxt saved;
840 struct mds_body *body;
841 struct dentry *dparent = NULL, *dchild = NULL;
842 struct lvfs_ucred uc;
843 struct lustre_handle parent_lockh[2];
844 int namesize, update_mode;
845 int rc = 0, cleanup_phase = 0, resent_req = 0;
849 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
851 /* Swab now, before anyone looks inside the request */
853 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
854 lustre_swab_mds_body);
856 CERROR("Can't swab mds_body\n");
857 GOTO(cleanup, rc = -EFAULT);
860 LASSERT_REQSWAB(req, offset + 1);
861 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
863 CERROR("Can't unpack name\n");
864 GOTO(cleanup, rc = -EFAULT);
866 namesize = req->rq_reqmsg->buflens[offset + 1];
868 LASSERT (offset == 0 || offset == 2);
869 /* if requests were at offset 2, the getattr reply goes back at 1 */
871 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
875 uc.luc_fsuid = body->fsuid;
876 uc.luc_fsgid = body->fsgid;
877 uc.luc_cap = body->capability;
878 uc.luc_suppgid1 = body->suppgid;
879 uc.luc_suppgid2 = -1;
880 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
881 cleanup_phase = 1; /* kernel context */
882 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
884 LASSERT(namesize > 0);
886 /* we have no dentry here, drop LOOKUP bit */
887 child_part &= ~MDS_INODELOCK_LOOKUP;
888 CDEBUG(D_OTHER, "%s: request to retrieve attrs for %lu/%lu\n",
889 obd->obd_name, (unsigned long) body->fid1.id,
890 (unsigned long) body->fid1.generation);
891 dchild = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR,
892 parent_lockh, &update_mode,
893 NULL, 0, child_part);
894 if (IS_ERR(dchild)) {
895 CERROR("can't find inode: %d\n", (int) PTR_ERR(dchild));
896 GOTO(cleanup, rc = PTR_ERR(dchild));
898 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
900 if (parent_lockh[1].cookie)
901 ldlm_lock_decref(parent_lockh + 1, update_mode);
907 /* FIXME: handle raw lookup */
909 if (body->valid == OBD_MD_FLID) {
910 struct mds_body *mds_reply;
911 int size = sizeof(*mds_reply);
913 // The user requested ONLY the inode number, so do a raw lookup
914 rc = lustre_pack_reply(req, 1, &size, NULL);
916 CERROR("out of memory\n");
920 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
922 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
924 mds_reply->fid1.id = inum;
925 mds_reply->valid = OBD_MD_FLID;
930 if (child_lockh->cookie != 0) {
931 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
935 if (resent_req == 0) {
936 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
937 parent_lockh, &dparent,
938 LCK_PR, MDS_INODELOCK_LOOKUP,
939 &update_mode, name, namesize,
940 child_lockh, &dchild, LCK_PR,
945 struct ldlm_lock *granted_lock;
946 struct ll_fid child_fid;
947 struct ldlm_resource *res;
948 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
949 granted_lock = ldlm_handle2lock(child_lockh);
950 LASSERT(granted_lock);
952 res = granted_lock->l_resource;
953 child_fid.id = res->lr_name.name[0];
954 child_fid.generation = res->lr_name.name[1];
955 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
957 LDLM_LOCK_PUT(granted_lock);
960 cleanup_phase = 2; /* dchild, dparent, locks */
962 /* let's make sure this name should leave on this mds node */
963 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
969 if (!DENTRY_VALID(dchild)) {
970 intent_set_disposition(rep, DISP_LOOKUP_NEG);
971 /* in the intent case, the policy clears this error:
972 the disposition is enough */
976 intent_set_disposition(rep, DISP_LOOKUP_POS);
979 if (req->rq_repmsg == NULL) {
980 if (dchild->d_flags & DCACHE_CROSS_REF)
981 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
983 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
985 CERROR ("mds_getattr_pack_msg: %d\n", rc);
990 rc = mds_getattr_internal(obd, dchild, req, body, offset);
991 GOTO(cleanup, rc); /* returns the lock to the client */
994 switch (cleanup_phase) {
996 if (resent_req == 0) {
997 if (rc && DENTRY_VALID(dchild))
998 ldlm_lock_decref(child_lockh, LCK_PR);
1000 ldlm_lock_decref(parent_lockh, LCK_PR);
1002 if (parent_lockh[1].cookie != 0)
1003 ldlm_lock_decref(parent_lockh + 1,
1012 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1018 static int mds_getattr(int offset, struct ptlrpc_request *req)
1020 struct mds_obd *mds = mds_req2mds(req);
1021 struct obd_device *obd = req->rq_export->exp_obd;
1022 struct lvfs_run_ctxt saved;
1024 struct mds_body *body;
1025 struct lvfs_ucred uc;
1029 body = lustre_swab_reqbuf (req, offset, sizeof (*body),
1030 lustre_swab_mds_body);
1032 CERROR ("Can't unpack body\n");
1036 uc.luc_fsuid = body->fsuid;
1037 uc.luc_fsgid = body->fsgid;
1038 uc.luc_cap = body->capability;
1039 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1040 de = mds_fid2dentry(mds, &body->fid1, NULL);
1042 rc = req->rq_status = -ENOENT;
1043 GOTO(out_pop, PTR_ERR(de));
1046 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1048 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1052 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
1057 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1062 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1063 unsigned long max_age)
1067 spin_lock(&obd->obd_osfs_lock);
1068 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1070 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1071 spin_unlock(&obd->obd_osfs_lock);
1076 static int mds_statfs(struct ptlrpc_request *req)
1078 struct obd_device *obd = req->rq_export->exp_obd;
1079 int rc, size = sizeof(struct obd_statfs);
1082 rc = lustre_pack_reply(req, 1, &size, NULL);
1083 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1084 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1088 /* We call this so that we can cache a bit - 1 jiffie worth */
1089 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1092 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1098 req->rq_status = rc;
1102 static int mds_sync(struct ptlrpc_request *req)
1104 struct obd_device *obd = req->rq_export->exp_obd;
1105 struct mds_obd *mds = &obd->u.mds;
1106 struct mds_body *body;
1107 int rc, size = sizeof(*body);
1110 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
1112 GOTO(out, rc = -EPROTO);
1114 rc = lustre_pack_reply(req, 1, &size, NULL);
1115 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1116 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1120 if (body->fid1.id == 0) {
1121 /* a fid of zero is taken to mean "sync whole filesystem" */
1122 rc = fsfilt_sync(obd, mds->mds_sb);
1126 /* just any file to grab fsync method - "file" arg unused */
1127 struct file *file = mds->mds_rcvd_filp;
1130 de = mds_fid2dentry(mds, &body->fid1, NULL);
1132 GOTO(out, rc = PTR_ERR(de));
1134 rc = file->f_op->fsync(NULL, de, 1);
1139 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1140 mds_pack_inode2fid(obd, &body->fid1, de->d_inode);
1141 mds_pack_inode2body(obd, body, de->d_inode);
1144 req->rq_status = rc;
1148 /* mds_readpage does not take a DLM lock on the inode, because the client must
1149 * already have a PR lock.
1151 * If we were to take another one here, a deadlock will result, if another
1152 * thread is already waiting for a PW lock. */
1153 static int mds_readpage(struct ptlrpc_request *req)
1155 struct obd_device *obd = req->rq_export->exp_obd;
1156 struct vfsmount *mnt;
1159 struct mds_body *body, *repbody;
1160 struct lvfs_run_ctxt saved;
1161 int rc, size = sizeof(*repbody);
1162 struct lvfs_ucred uc;
1165 rc = lustre_pack_reply(req, 1, &size, NULL);
1166 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1167 CERROR("mds: out of memory\n");
1168 GOTO(out, rc = -ENOMEM);
1171 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1173 GOTO (out, rc = -EFAULT);
1175 uc.luc_fsuid = body->fsuid;
1176 uc.luc_fsgid = body->fsgid;
1177 uc.luc_cap = body->capability;
1178 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1179 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1181 GOTO(out_pop, rc = PTR_ERR(de));
1183 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1185 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1186 /* note: in case of an error, dentry_open puts dentry */
1188 GOTO(out_pop, rc = PTR_ERR(file));
1190 /* body->size is actually the offset -eeb */
1191 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1192 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1193 body->size, de->d_inode->i_blksize);
1194 GOTO(out_file, rc = -EFAULT);
1197 /* body->nlink is actually the #bytes to read -eeb */
1198 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1199 CERROR("size %u is not multiple of blocksize %lu\n",
1200 body->nlink, de->d_inode->i_blksize);
1201 GOTO(out_file, rc = -EFAULT);
1204 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1205 repbody->size = file->f_dentry->d_inode->i_size;
1206 repbody->valid = OBD_MD_FLSIZE;
1208 /* to make this asynchronous make sure that the handling function
1209 doesn't send a reply when this function completes. Instead a
1210 callback function would send the reply */
1211 /* body->size is actually the offset -eeb */
1212 rc = mds_sendpage(req, file, body->size, body->nlink);
1215 filp_close(file, 0);
1217 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1219 req->rq_status = rc;
1223 int mds_reint(struct ptlrpc_request *req, int offset,
1224 struct lustre_handle *lockh)
1226 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
1230 OBD_ALLOC(rec, sizeof(*rec));
1234 rc = mds_update_unpack(req, offset, rec);
1235 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1236 CERROR("invalid record\n");
1237 GOTO(out, req->rq_status = -EINVAL);
1239 /* rc will be used to interrupt a for loop over multiple records */
1240 rc = mds_reint_rec(rec, offset, req, lockh);
1242 OBD_FREE(rec, sizeof(*rec));
1246 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1247 struct obd_device *obd, int *process)
1249 switch (req->rq_reqmsg->opc) {
1250 case MDS_CONNECT: /* This will never get here, but for completeness. */
1251 case OST_CONNECT: /* This will never get here, but for completeness. */
1252 case MDS_DISCONNECT:
1253 case OST_DISCONNECT:
1258 case MDS_SYNC: /* used in unmounting */
1263 *process = target_queue_recovery_request(req, obd);
1267 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1269 /* XXX what should we set rq_status to here? */
1270 req->rq_status = -EAGAIN;
1271 RETURN(ptlrpc_error(req));
1275 static char *reint_names[] = {
1276 [REINT_SETATTR] "setattr",
1277 [REINT_CREATE] "create",
1278 [REINT_LINK] "link",
1279 [REINT_UNLINK] "unlink",
1280 [REINT_RENAME] "rename",
1281 [REINT_OPEN] "open",
1284 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER |\
1285 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ|\
1286 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME|\
1289 static void reconstruct_create(struct ptlrpc_request *req)
1291 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1292 struct mds_client_data *mcd = med->med_mcd;
1293 struct ost_body *body;
1295 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1297 /* copy rc, transno and disp; steal locks */
1298 mds_req_from_mcd(req, mcd);
1299 CERROR("reconstruct reply for x"LPU64"\n", req->rq_xid);
1302 static int mdt_obj_create(struct ptlrpc_request *req)
1304 struct obd_device *obd = req->rq_export->exp_obd;
1305 struct ldlm_res_id res_id = { .name = {0} };
1306 struct mds_obd *mds = &obd->u.mds;
1307 struct ost_body *body, *repbody;
1308 int rc, size = sizeof(*repbody);
1309 char fidname[LL_FID_NAMELEN];
1310 struct inode *parent_inode;
1311 struct lustre_handle lockh;
1312 struct lvfs_run_ctxt saved;
1313 ldlm_policy_data_t policy;
1314 struct dentry *new = NULL;
1315 struct dentry_params dp;
1316 int mealen, flags = 0;
1317 unsigned int tmpname;
1318 struct lvfs_ucred uc;
1323 DEBUG_REQ(D_HA, req, "create remote object");
1325 parent_inode = mds->mds_objects_dir->d_inode;
1327 body = lustre_swab_reqbuf(req, 0, sizeof(*body),
1328 lustre_swab_ost_body);
1332 MDS_CHECK_RESENT(req, reconstruct_create(req));
1334 uc.luc_fsuid = body->oa.o_uid;
1335 uc.luc_fsgid = body->oa.o_gid;
1337 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1339 rc = lustre_pack_reply(req, 1, &size, NULL);
1343 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1345 if (body->oa.o_flags & OBD_FL_RECREATE_OBJS) {
1346 /* this is re-create request from MDS holding directory name.
1347 * we have to lookup given ino/generation first. if it exists
1348 * (good case) then there is nothing to do. if it does not
1349 * then we have to recreate it */
1351 fid.id = body->oa.o_id;
1352 fid.generation = body->oa.o_generation;
1353 new = mds_fid2dentry(mds, &fid, NULL);
1354 if (!IS_ERR(new) && new->d_inode) {
1355 CWARN("mkdir() repairing is on its way: %lu/%lu\n",
1356 (unsigned long) fid.id,
1357 (unsigned long) fid.generation);
1358 obdo_from_inode(&repbody->oa, new->d_inode,
1359 FILTER_VALID_FLAGS);
1360 repbody->oa.o_id = new->d_inode->i_ino;
1361 repbody->oa.o_generation = new->d_inode->i_generation;
1362 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1363 GOTO(cleanup2, rc = 0);
1365 CWARN("hmm. for some reason dir %lu/%lu (or reply) got lost\n",
1366 (unsigned long) fid.id, (unsigned long) fid.generation);
1367 LASSERT(new->d_inode == NULL ||
1368 new->d_inode->i_generation != fid.generation);
1372 down(&parent_inode->i_sem);
1373 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1374 LASSERT(!IS_ERR(handle));
1377 tmpname = ll_insecure_random_int();
1378 rc = sprintf(fidname, "%u", tmpname);
1379 new = lookup_one_len(fidname, mds->mds_objects_dir, rc);
1381 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1382 obd->obd_name, fidname, (int) PTR_ERR(new));
1383 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
1384 up(&parent_inode->i_sem);
1385 RETURN(PTR_ERR(new));
1386 } else if (new->d_inode) {
1387 CERROR("%s: name exists. repeat\n", obd->obd_name);
1391 new->d_fsdata = (void *) &dp;
1395 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1396 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1397 (unsigned long) body->oa.o_id,
1398 (unsigned long) body->oa.o_generation);
1399 dp.p_inum = body->oa.o_id;
1400 dp.p_generation = body->oa.o_generation;
1402 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1404 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1405 repbody->oa.o_id = new->d_inode->i_ino;
1406 repbody->oa.o_generation = new->d_inode->i_generation;
1407 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1409 rc = fsfilt_del_dir_entry(obd, new);
1410 up(&parent_inode->i_sem);
1413 CERROR("can't remove name for object: %d\n", rc);
1417 /* this lock should be taken to serialize MDS modifications
1418 * in failure case */
1419 res_id.name[0] = new->d_inode->i_ino;
1420 res_id.name[1] = new->d_inode->i_generation;
1421 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1422 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1423 res_id, LDLM_IBITS, &policy,
1424 LCK_EX, &flags, mds_blocking_ast,
1425 ldlm_completion_ast, NULL, NULL,
1426 NULL, 0, NULL, &lockh);
1430 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
1431 (unsigned long) new->d_inode->i_ino,
1432 (unsigned long) new->d_inode->i_generation,
1433 (unsigned) new->d_inode->i_mode);
1435 up(&parent_inode->i_sem);
1436 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1439 if (rc == 0 && body->oa.o_valid & OBD_MD_FLID) {
1440 /* this is new object for splitted dir. we have to
1441 * prevent recursive splitting on it -bzzz */
1442 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1443 OBD_ALLOC(mea, mealen);
1445 GOTO(cleanup, rc = -ENOMEM);
1447 down(&new->d_inode->i_sem);
1448 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1449 up(&new->d_inode->i_sem);
1450 OBD_FREE(mea, mealen);
1454 rc = mds_finish_transno(mds, parent_inode, handle, req, rc, 0);
1456 ptlrpc_save_lock(req, &lockh, LCK_EX);
1458 ldlm_lock_decref(&lockh, LCK_EX);
1461 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1465 static int mdt_get_info(struct ptlrpc_request *req)
1468 struct obd_export *exp = req->rq_export;
1469 int keylen, rc = 0, size = sizeof(obd_id);
1473 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1475 DEBUG_REQ(D_HA, req, "no get_info key");
1478 keylen = req->rq_reqmsg->buflens[0];
1480 if (keylen < strlen("mdsize") || memcmp(key, "mdsize", 6) != 0)
1483 rc = lustre_pack_reply(req, 1, &size, NULL);
1487 reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
1488 rc = obd_get_info(exp, keylen, key, &size, reply);
1489 req->rq_repmsg->status = 0;
1493 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1494 void *key, __u32 vallen, void *val)
1496 struct obd_device *obd;
1497 struct mds_obd *mds;
1501 obd = class_exp2obd(exp);
1503 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1504 exp->exp_handle.h_cookie);
1508 #define KEY_IS(str) \
1509 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
1512 if (KEY_IS("mds_num")) {
1515 CDEBUG(D_IOCTL, "set mds num %d\n", *(int*)val);
1516 mds->mds_num = *(int*)val;
1517 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
1518 valsize = sizeof(group);
1519 /*mds number has been changed, so the corresponding obdfilter exp
1520 *need to be changed too*/
1521 rc = obd_set_info(mds->mds_osc_exp, strlen("mds_conn"), "mds_conn",
1524 } else if (KEY_IS("client")) {
1525 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
1526 atomic_inc(&mds->mds_real_clients);
1527 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
1528 obd->obd_name, exp->exp_client_uuid.uuid,
1529 atomic_read(&mds->mds_real_clients));
1530 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
1532 if (mds->mds_lmv_name) {
1533 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1538 CDEBUG(D_IOCTL, "invalid key\n");
1542 static int mdt_set_info(struct ptlrpc_request *req)
1545 struct obd_export *exp = req->rq_export;
1546 int keylen, rc = 0, vallen;
1549 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1551 DEBUG_REQ(D_HA, req, "no set_info key");
1554 keylen = req->rq_reqmsg->buflens[0];
1556 if (keylen == strlen("mds_num") &&
1557 memcmp(key, "mds_num", keylen) == 0) {
1558 rc = lustre_pack_reply(req, 0, NULL, NULL);
1561 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
1563 vallen = req->rq_reqmsg->buflens[1];
1565 rc = obd_set_info(exp, keylen, key, vallen, val);
1566 req->rq_repmsg->status = 0;
1568 } else if (keylen == strlen("client") &&
1569 memcmp(key, "client", keylen) == 0) {
1570 rc = lustre_pack_reply(req, 0, NULL, NULL);
1573 rc = obd_set_info(exp, keylen, key, sizeof(obd_id), NULL);
1574 req->rq_repmsg->status = 0;
1577 CDEBUG(D_IOCTL, "invalid key\n");
1581 extern int ost_brw_write(struct ptlrpc_request *, struct obd_trans_info *);
1582 int mds_handle(struct ptlrpc_request *req)
1584 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1586 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1587 struct obd_device *obd = NULL;
1590 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1592 LASSERT(current->journal_info == NULL);
1593 /* XXX identical to OST */
1594 if (req->rq_reqmsg->opc != MDS_CONNECT) {
1595 struct mds_export_data *med;
1596 int recovering, abort_recovery;
1598 if (req->rq_export == NULL) {
1599 CERROR("lustre_mds: operation %d on unconnected MDS\n",
1600 req->rq_reqmsg->opc);
1601 req->rq_status = -ENOTCONN;
1602 GOTO(out, rc = -ENOTCONN);
1605 med = &req->rq_export->exp_mds_data;
1606 obd = req->rq_export->exp_obd;
1609 /* sanity check: if the xid matches, the request must
1610 * be marked as a resent or replayed */
1611 if (req->rq_xid == med->med_mcd->mcd_last_xid)
1612 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
1613 (MSG_RESENT | MSG_REPLAY),
1614 "rq_xid "LPU64" matches last_xid, "
1615 "expected RESENT flag\n",
1617 /* else: note the opposite is not always true; a
1618 * RESENT req after a failover will usually not match
1619 * the last_xid, since it was likely never
1620 * committed. A REPLAYed request will almost never
1621 * match the last xid, however it could for a
1622 * committed, but still retained, open. */
1624 /* Check for aborted recovery. */
1625 spin_lock_bh(&obd->obd_processing_task_lock);
1626 abort_recovery = obd->obd_abort_recovery;
1627 recovering = obd->obd_recovering;
1628 spin_unlock_bh(&obd->obd_processing_task_lock);
1629 if (abort_recovery) {
1630 target_abort_recovery(obd);
1631 } else if (recovering) {
1632 rc = mds_filter_recovery_request(req, obd,
1634 if (rc || !should_process)
1639 switch (req->rq_reqmsg->opc) {
1641 DEBUG_REQ(D_INODE, req, "connect");
1642 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1643 rc = target_handle_connect(req, mds_handle);
1645 /* Now that we have an export, set mds. */
1646 mds = mds_req2mds(req);
1649 case MDS_DISCONNECT:
1650 DEBUG_REQ(D_INODE, req, "disconnect");
1651 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1652 rc = target_handle_disconnect(req);
1653 req->rq_status = rc; /* superfluous? */
1657 DEBUG_REQ(D_INODE, req, "getstatus");
1658 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1659 rc = mds_getstatus(req);
1663 DEBUG_REQ(D_INODE, req, "getattr");
1664 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1665 rc = mds_getattr(0, req);
1668 case MDS_GETATTR_NAME: {
1669 struct lustre_handle lockh;
1670 DEBUG_REQ(D_INODE, req, "getattr_name");
1671 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1673 /* If this request gets a reconstructed reply, we won't be
1674 * acquiring any new locks in mds_getattr_name, so we don't
1678 rc = mds_getattr_name(0, req, &lockh, MDS_INODELOCK_UPDATE);
1679 /* this non-intent call (from an ioctl) is special */
1680 req->rq_status = rc;
1681 if (rc == 0 && lockh.cookie)
1682 ldlm_lock_decref(&lockh, LCK_PR);
1686 DEBUG_REQ(D_INODE, req, "statfs");
1687 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1688 rc = mds_statfs(req);
1692 DEBUG_REQ(D_INODE, req, "readpage");
1693 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1694 rc = mds_readpage(req);
1696 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1697 if (req->rq_reply_state) {
1698 lustre_free_reply_state (req->rq_reply_state);
1699 req->rq_reply_state = NULL;
1707 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1709 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1710 mds->mds_max_cookiesize};
1713 /* NB only peek inside req now; mds_reint() will swab it */
1715 CERROR ("Can't inspect opcode\n");
1720 if (lustre_msg_swabbed (req->rq_reqmsg))
1723 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1724 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1725 reint_names[opc] == NULL) ? reint_names[opc] :
1728 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1730 if (opc == REINT_UNLINK)
1732 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1737 rc = lustre_pack_reply(req, bufcount, size, NULL);
1741 rc = mds_reint(req, 0, NULL);
1742 fail = OBD_FAIL_MDS_REINT_NET_REP;
1747 DEBUG_REQ(D_INODE, req, "close");
1748 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1749 rc = mds_close(req);
1752 case MDS_DONE_WRITING:
1753 DEBUG_REQ(D_INODE, req, "done_writing");
1754 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1755 rc = mds_done_writing(req);
1759 DEBUG_REQ(D_INODE, req, "pin");
1760 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1765 DEBUG_REQ(D_INODE, req, "sync");
1766 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1771 DEBUG_REQ(D_INODE, req, "ping");
1772 rc = target_handle_ping(req);
1775 case OBD_LOG_CANCEL:
1776 CDEBUG(D_INODE, "log cancel\n");
1777 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1778 rc = -ENOTSUPP; /* la la la */
1782 DEBUG_REQ(D_INODE, req, "enqueue");
1783 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1784 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1785 ldlm_server_blocking_ast, NULL);
1788 DEBUG_REQ(D_INODE, req, "convert");
1789 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1790 rc = ldlm_handle_convert(req);
1792 case LDLM_BL_CALLBACK:
1793 case LDLM_CP_CALLBACK:
1794 DEBUG_REQ(D_INODE, req, "callback");
1795 CERROR("callbacks should not happen on MDS\n");
1797 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1799 case LLOG_ORIGIN_HANDLE_OPEN:
1800 DEBUG_REQ(D_INODE, req, "llog_init");
1801 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1802 rc = llog_origin_handle_open(req);
1804 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1805 DEBUG_REQ(D_INODE, req, "llog next block");
1806 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1807 rc = llog_origin_handle_next_block(req);
1809 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1810 DEBUG_REQ(D_INODE, req, "llog prev block");
1811 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1812 rc = llog_origin_handle_prev_block(req);
1814 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1815 DEBUG_REQ(D_INODE, req, "llog read header");
1816 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1817 rc = llog_origin_handle_read_header(req);
1819 case LLOG_ORIGIN_HANDLE_CLOSE:
1820 DEBUG_REQ(D_INODE, req, "llog close");
1821 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1822 rc = llog_origin_handle_close(req);
1825 DEBUG_REQ(D_INODE, req, "ost_create");
1826 rc = mdt_obj_create(req);
1829 DEBUG_REQ(D_INODE, req, "get_info");
1830 rc = mdt_get_info(req);
1833 DEBUG_REQ(D_INODE, req, "set_info");
1834 rc = mdt_set_info(req);
1837 CDEBUG(D_INODE, "write\n");
1838 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1839 rc = ost_brw_write(req, NULL);
1840 LASSERT(current->journal_info == NULL);
1841 /* mdt_brw sends its own replies */
1845 DEBUG_REQ(D_INODE, req, "llog catinfo");
1846 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1847 rc = llog_catinfo(req);
1850 req->rq_status = -ENOTSUPP;
1851 rc = ptlrpc_error(req);
1855 LASSERT(current->journal_info == NULL);
1859 /* If we're DISCONNECTing, the mds_export_data is already freed */
1860 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1861 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1862 struct obd_device *obd = list_entry(mds, struct obd_device,
1864 req->rq_repmsg->last_xid =
1865 le64_to_cpu(med->med_mcd->mcd_last_xid);
1867 if (!obd->obd_no_transno) {
1868 req->rq_repmsg->last_committed =
1869 obd->obd_last_committed;
1871 DEBUG_REQ(D_IOCTL, req,
1872 "not sending last_committed update");
1874 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1876 mds->mds_last_transno, obd->obd_last_committed,
1881 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1882 if (obd && obd->obd_recovering) {
1883 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1884 return target_queue_final_reply(req, rc);
1886 /* Lost a race with recovery; let the error path DTRT. */
1887 rc = req->rq_status = -ENOTCONN;
1890 target_send_reply(req, rc, fail);
1894 /* Update the server data on disk. This stores the new mount_count and
1895 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1896 * then the server last_rcvd value may be less than that of the clients.
1897 * This will alert us that we may need to do client recovery.
1899 * Also assumes for mds_last_transno that we are not modifying it (no locking).
1901 int mds_update_server_data(struct obd_device *obd, int force_sync)
1903 struct mds_obd *mds = &obd->u.mds;
1904 struct mds_server_data *msd = mds->mds_server_data;
1905 struct file *filp = mds->mds_rcvd_filp;
1906 struct lvfs_run_ctxt saved;
1911 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1912 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1914 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1915 mds->mds_mount_count, mds->mds_last_transno);
1916 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1918 CERROR("error writing MDS server data: rc = %d\n", rc);
1919 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1924 /* mount the file system (secretly) */
1925 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1927 struct lustre_cfg* lcfg = buf;
1928 struct mds_obd *mds = &obd->u.mds;
1929 char *options = NULL;
1930 struct vfsmount *mnt;
1935 dev_clear_rdonly(2);
1937 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1938 RETURN(rc = -EINVAL);
1940 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1941 if (IS_ERR(obd->obd_fsops))
1942 RETURN(rc = PTR_ERR(obd->obd_fsops));
1944 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1946 page = __get_free_page(GFP_KERNEL);
1950 options = (char *)page;
1951 memset(options, 0, PAGE_SIZE);
1953 /* here we use "iopen_nopriv" hardcoded, because it affects MDS utility
1954 * and the rest of options are passed by mount options. Probably this
1955 * should be moved to somewhere else like startup scripts or lconf. */
1956 sprintf(options, "iopen_nopriv");
1958 if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
1959 sprintf(options + strlen(options), ",%s",
1960 lcfg->lcfg_inlbuf4);
1962 /* we have to know mdsnum before touching underlying fs -bzzz */
1963 if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
1964 strcmp(lcfg->lcfg_inlbuf5, "dumb")) {
1967 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
1968 obd->obd_name, lcfg->lcfg_inlbuf5);
1970 generate_random_uuid(uuid);
1971 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
1973 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen5);
1974 if (mds->mds_lmv_name == NULL)
1975 RETURN(rc = -ENOMEM);
1977 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf5,
1978 lcfg->lcfg_inllen5);
1980 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1982 OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen5);
1987 /* FIXME-WANGDI: this should be reworked when we will use lmv along
1988 * with cobd, because correct mdsnum is set in mds_lmv_connect(). */
1989 if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 && !mds->mds_lmv_obd &&
1990 strcmp(lcfg->lcfg_inlbuf6, "dumb")) {
1991 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master")) &&
1992 mds->mds_num == 0) {
1993 mds->mds_num = REAL_MDS_NUMBER;
1994 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache")) &&
1995 mds->mds_num == 0) {
1996 mds->mds_num = CACHE_MDS_NUMBER;
2000 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
2001 lcfg->lcfg_inlbuf1, options);
2007 CERROR("do_kern_mount failed: rc = %d\n", rc);
2011 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
2013 sema_init(&mds->mds_orphan_recovery_sem, 1);
2014 sema_init(&mds->mds_epoch_sem, 1);
2015 spin_lock_init(&mds->mds_transno_lock);
2016 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2017 atomic_set(&mds->mds_open_count, 0);
2018 atomic_set(&mds->mds_real_clients, 0);
2020 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
2021 LDLM_NAMESPACE_SERVER);
2022 if (obd->obd_namespace == NULL) {
2023 mds_cleanup(obd, 0);
2024 GOTO(err_put, rc = -ENOMEM);
2026 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
2028 rc = mds_fs_setup(obd, mnt);
2030 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
2034 rc = llog_start_commit_thread();
2038 /* this check for @dumb string is needed to handle mounting MDS with
2039 * smfs. Read lconf:MDSDEV.write_conf() for more details. */
2040 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2041 strcmp(lcfg->lcfg_inlbuf3, "dumb")) {
2044 generate_random_uuid(uuid);
2045 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
2047 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
2048 if (mds->mds_profile == NULL)
2049 GOTO(err_fs, rc = -ENOMEM);
2051 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
2052 lcfg->lcfg_inllen3);
2055 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2056 "mds_ldlm_client", &obd->obd_ldlm_client);
2057 obd->obd_replayable = 1;
2059 rc = mds_postsetup(obd);
2066 /* No extra cleanup needed for llog_init_commit_thread() */
2067 mds_fs_cleanup(obd, 0);
2069 ldlm_namespace_free(obd->obd_namespace, 0);
2070 obd->obd_namespace = NULL;
2073 mntput(mds->mds_vfsmnt);
2077 fsfilt_put_ops(obd->obd_fsops);
2081 static int mds_postsetup(struct obd_device *obd)
2083 struct mds_obd *mds = &obd->u.mds;
2087 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
2088 obd, 0, NULL, &llog_lvfs_ops);
2092 if (mds->mds_profile) {
2093 struct llog_ctxt *lgctxt;
2094 struct lvfs_run_ctxt saved;
2095 struct lustre_profile *lprof;
2096 struct config_llog_instance cfg;
2098 cfg.cfg_instance = NULL;
2099 cfg.cfg_uuid = mds->mds_lov_uuid;
2100 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2102 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
2104 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2105 GOTO(err_llog, rc = -EINVAL);
2108 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
2109 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2114 lprof = class_get_profile(mds->mds_profile);
2115 if (lprof == NULL) {
2116 CERROR("No profile found: %s\n", mds->mds_profile);
2117 GOTO(err_cleanup, rc = -ENOENT);
2119 rc = mds_lov_connect(obd, lprof->lp_osc);
2121 GOTO(err_cleanup, rc);
2123 rc = mds_lmv_postsetup(obd);
2125 GOTO(err_cleanup, rc);
2133 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
2134 LLOG_CONFIG_ORIG_CTXT));
2138 int mds_postrecov(struct obd_device *obd)
2140 struct mds_obd *mds = &obd->u.mds;
2141 struct llog_ctxt *ctxt;
2145 LASSERT(!obd->obd_recovering);
2146 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2147 LASSERT(ctxt != NULL);
2149 /* set nextid first, so we are sure it happens */
2150 rc = mds_lov_set_nextid(obd);
2152 CERROR("%s: mds_lov_set_nextid failed\n", obd->obd_name);
2156 /* clean PENDING dir */
2157 rc = mds_cleanup_orphans(obd);
2162 rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
2165 CERROR("%s: failed at llog_origin_connect: %d\n",
2170 /* remove the orphaned precreated objects */
2171 rc = mds_lov_clearorphans(mds, NULL /* all OSTs */);
2176 RETURN(rc < 0 ? rc : item);
2179 /* cleanup all llogging subsystems */
2180 rc = obd_llog_finish(obd, &obd->obd_llogs,
2181 mds->mds_lov_desc.ld_tgt_count);
2183 CERROR("%s: failed to cleanup llogging subsystems\n",
2188 int mds_lov_clean(struct obd_device *obd)
2190 struct mds_obd *mds = &obd->u.mds;
2192 if (mds->mds_profile) {
2194 struct config_llog_instance cfg;
2195 struct lvfs_run_ctxt saved;
2196 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
2198 OBD_ALLOC(cln_prof, len);
2199 sprintf(cln_prof, "%s-clean", mds->mds_profile);
2201 cfg.cfg_instance = NULL;
2202 cfg.cfg_uuid = mds->mds_lov_uuid;
2204 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2205 class_config_process_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT),
2207 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2209 OBD_FREE(cln_prof, len);
2210 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2211 mds->mds_profile = NULL;
2216 int mds_lmv_clean(struct obd_device *obd)
2218 struct mds_obd *mds = &obd->u.mds;
2220 if (mds->mds_lmv_name) {
2221 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
2222 mds->mds_lmv_name = NULL;
2227 static int mds_precleanup(struct obd_device *obd, int flags)
2233 mds_lov_disconnect(obd, flags);
2235 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2239 static int mds_cleanup(struct obd_device *obd, int flags)
2241 struct mds_obd *mds = &obd->u.mds;
2244 if (mds->mds_sb == NULL)
2247 mds_update_server_data(obd, 1);
2248 if (mds->mds_lov_objids != NULL) {
2249 OBD_FREE(mds->mds_lov_objids,
2250 mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
2252 mds_fs_cleanup(obd, flags);
2256 /* 2 seems normal on mds, (may_umount() also expects 2
2257 fwiw), but we only see 1 at this point in obdfilter. */
2258 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
2259 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
2260 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
2262 mntput(mds->mds_vfsmnt);
2266 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
2268 spin_lock_bh(&obd->obd_processing_task_lock);
2269 if (obd->obd_recovering) {
2270 target_cancel_recovery_timer(obd);
2271 obd->obd_recovering = 0;
2273 spin_unlock_bh(&obd->obd_processing_task_lock);
2276 dev_clear_rdonly(2);
2277 fsfilt_put_ops(obd->obd_fsops);
2282 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
2283 struct ldlm_lock *new_lock,
2284 struct lustre_handle *lockh)
2286 struct obd_export *exp = req->rq_export;
2287 struct obd_device *obd = exp->exp_obd;
2288 struct ldlm_request *dlmreq =
2289 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
2290 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
2291 struct list_head *iter;
2293 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2296 l_lock(&obd->obd_namespace->ns_lock);
2297 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2298 struct ldlm_lock *lock;
2299 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2300 if (lock == new_lock)
2302 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2303 lockh->cookie = lock->l_handle.h_cookie;
2304 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2306 l_unlock(&obd->obd_namespace->ns_lock);
2310 l_unlock(&obd->obd_namespace->ns_lock);
2312 /* If the xid matches, then we know this is a resent request,
2313 * and allow it. (It's probably an OPEN, for which we don't
2315 if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
2318 /* This remote handle isn't enqueued, so we never received or
2319 * processed this request. Clear MSG_RESENT, because it can
2320 * be handled like any normal request now. */
2322 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2324 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2328 int intent_disposition(struct ldlm_reply *rep, int flag)
2332 return (rep->lock_policy_res1 & flag);
2335 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2339 rep->lock_policy_res1 |= flag;
2342 static int mds_intent_policy(struct ldlm_namespace *ns,
2343 struct ldlm_lock **lockp, void *req_cookie,
2344 ldlm_mode_t mode, int flags, void *data)
2346 struct ptlrpc_request *req = req_cookie;
2347 struct ldlm_lock *lock = *lockp;
2348 struct ldlm_intent *it;
2349 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2350 struct ldlm_reply *rep;
2351 struct lustre_handle lockh = { 0 };
2352 struct ldlm_lock *new_lock;
2353 int getattr_part = MDS_INODELOCK_UPDATE;
2354 int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
2355 sizeof(struct mds_body),
2356 mds->mds_max_mdsize,
2357 mds->mds_max_cookiesize};
2360 LASSERT(req != NULL);
2362 if (req->rq_reqmsg->bufcount <= 1) {
2363 /* No intent was provided */
2364 int size = sizeof(struct ldlm_reply);
2365 rc = lustre_pack_reply(req, 1, &size, NULL);
2370 it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
2372 CERROR("Intent missing\n");
2373 RETURN(req->rq_status = -EFAULT);
2376 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2378 rc = lustre_pack_reply(req, 3, repsize, NULL);
2380 RETURN(req->rq_status = rc);
2382 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
2383 intent_set_disposition(rep, DISP_IT_EXECD);
2385 fixup_handle_for_resent_req(req, lock, &lockh);
2387 /* execute policy */
2388 switch ((long)it->opc) {
2390 case IT_CREAT|IT_OPEN:
2391 /* XXX swab here to assert that an mds_open reint
2392 * packet is following */
2393 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
2395 /* We abort the lock if the lookup was negative and
2396 * we did not make it to the OPEN portion */
2397 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2398 RETURN(ELDLM_LOCK_ABORTED);
2399 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2400 !intent_disposition(rep, DISP_OPEN_OPEN))
2402 /* IT_OPEN may return lock on cross-node dentry
2403 * that we want to hold during attr retrival -bzzz */
2404 if (rc != 0 || lockh.cookie == 0)
2405 RETURN(ELDLM_LOCK_ABORTED);
2408 getattr_part = MDS_INODELOCK_LOOKUP;
2411 getattr_part |= MDS_INODELOCK_LOOKUP;
2413 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh,
2415 /* FIXME: LDLM can set req->rq_status. MDS sets
2416 policy_res{1,2} with disposition and status.
2417 - replay: returns 0 & req->status is old status
2418 - otherwise: returns req->status */
2419 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2420 rep->lock_policy_res2 = 0;
2421 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2422 rep->lock_policy_res2)
2423 RETURN(ELDLM_LOCK_ABORTED);
2424 if (req->rq_status != 0) {
2426 rep->lock_policy_res2 = req->rq_status;
2427 RETURN(ELDLM_LOCK_ABORTED);
2431 rc = mds_lock_and_check_slave(offset, req, &lockh);
2432 if ((rep->lock_policy_res2 = rc)) {
2434 rep->lock_policy_res2 = 0;
2435 RETURN(ELDLM_LOCK_ABORTED);
2439 CERROR("Unhandled intent "LPD64"\n", it->opc);
2443 /* By this point, whatever function we called above must have either
2444 * filled in 'lockh', been an intent replay, or returned an error. We
2445 * want to allow replayed RPCs to not get a lock, since we would just
2446 * drop it below anyways because lock replay is done separately by the
2447 * client afterwards. For regular RPCs we want to give the new lock to
2448 * the client instead of whatever lock it was about to get. */
2449 new_lock = ldlm_handle2lock(&lockh);
2450 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2453 LASSERT(new_lock != NULL);
2455 /* If we've already given this lock to a client once, then we should
2456 * have no readers or writers. Otherwise, we should have one reader
2457 * _or_ writer ref (which will be zeroed below) before returning the
2458 * lock to a client. */
2459 if (new_lock->l_export == req->rq_export) {
2460 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2462 LASSERT(new_lock->l_export == NULL);
2463 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2468 if (new_lock->l_export == req->rq_export) {
2469 /* Already gave this to the client, which means that we
2470 * reconstructed a reply. */
2471 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2473 RETURN(ELDLM_LOCK_REPLACED);
2476 /* Fixup the lock to be given to the client */
2477 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
2478 new_lock->l_readers = 0;
2479 new_lock->l_writers = 0;
2481 new_lock->l_export = class_export_get(req->rq_export);
2482 list_add(&new_lock->l_export_chain,
2483 &new_lock->l_export->exp_ldlm_data.led_held_locks);
2485 new_lock->l_blocking_ast = lock->l_blocking_ast;
2486 new_lock->l_completion_ast = lock->l_completion_ast;
2488 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2489 sizeof(lock->l_remote_handle));
2491 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2493 LDLM_LOCK_PUT(new_lock);
2494 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
2496 RETURN(ELDLM_LOCK_REPLACED);
2499 int mds_attach(struct obd_device *dev, obd_count len, void *data)
2501 struct lprocfs_static_vars lvars;
2503 lprocfs_init_multi_vars(0, &lvars);
2504 return lprocfs_obd_attach(dev, lvars.obd_vars);
2507 int mds_detach(struct obd_device *dev)
2509 return lprocfs_obd_detach(dev);
2512 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
2514 struct lprocfs_static_vars lvars;
2516 lprocfs_init_multi_vars(1, &lvars);
2517 return lprocfs_obd_attach(dev, lvars.obd_vars);
2520 int mdt_detach(struct obd_device *dev)
2522 return lprocfs_obd_detach(dev);
2525 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
2527 struct mds_obd *mds = &obd->u.mds;
2532 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2533 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
2534 mds_handle, "mds", obd->obd_proc_entry);
2536 if (!mds->mds_service) {
2537 CERROR("failed to start service\n");
2541 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
2544 GOTO(err_thread, rc);
2546 mds->mds_setattr_service =
2547 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2548 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
2549 mds_handle, "mds_setattr",
2550 obd->obd_proc_entry);
2551 if (!mds->mds_setattr_service) {
2552 CERROR("failed to start getattr service\n");
2553 GOTO(err_thread, rc = -ENOMEM);
2556 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
2557 MDT_NUM_THREADS, "ll_mdt_attr");
2559 GOTO(err_thread2, rc);
2561 mds->mds_readpage_service =
2562 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2563 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
2564 mds_handle, "mds_readpage",
2565 obd->obd_proc_entry);
2566 if (!mds->mds_readpage_service) {
2567 CERROR("failed to start readpage service\n");
2568 GOTO(err_thread2, rc = -ENOMEM);
2571 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
2572 MDT_NUM_THREADS, "ll_mdt_rdpg");
2575 GOTO(err_thread3, rc);
2580 ptlrpc_unregister_service(mds->mds_readpage_service);
2582 ptlrpc_unregister_service(mds->mds_setattr_service);
2584 ptlrpc_unregister_service(mds->mds_service);
2588 static int mdt_cleanup(struct obd_device *obd, int flags)
2590 struct mds_obd *mds = &obd->u.mds;
2593 ptlrpc_stop_all_threads(mds->mds_readpage_service);
2594 ptlrpc_unregister_service(mds->mds_readpage_service);
2596 ptlrpc_stop_all_threads(mds->mds_setattr_service);
2597 ptlrpc_unregister_service(mds->mds_setattr_service);
2599 ptlrpc_stop_all_threads(mds->mds_service);
2600 ptlrpc_unregister_service(mds->mds_service);
2605 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2608 struct obd_device *obd = data;
2611 fid.generation = gen;
2612 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2615 static int mds_get_info(struct obd_export *exp, __u32 keylen,
2616 void *key, __u32 *vallen, void *val)
2618 struct obd_device *obd;
2619 struct mds_obd *mds;
2622 obd = class_exp2obd(exp);
2624 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2625 exp->exp_handle.h_cookie);
2629 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
2630 /*Get log_context handle*/
2631 unsigned long *llh_handle = val;
2632 *vallen = sizeof(unsigned long);
2633 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
2636 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
2637 /*Get log_context handle*/
2638 unsigned long *sb = val;
2639 *vallen = sizeof(unsigned long);
2640 *sb = (unsigned long)obd->u.mds.mds_sb;
2645 keylen == strlen("mdsize");
2646 if (keylen && memcmp(key, "mdsize", keylen) == 0) {
2647 __u32 *mdsize = val;
2648 *vallen = sizeof(*mdsize);
2649 *mdsize = mds->mds_max_mdsize;
2653 CDEBUG(D_IOCTL, "invalid key\n");
2657 struct lvfs_callback_ops mds_lvfs_ops = {
2658 l_fid2dentry: mds_lvfs_fid2dentry,
2661 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
2662 int objcount, struct obd_ioobj *obj,
2663 int niocount, struct niobuf_remote *nb,
2664 struct niobuf_local *res,
2665 struct obd_trans_info *oti);
2666 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
2667 int objcount, struct obd_ioobj *obj, int niocount,
2668 struct niobuf_local *res, struct obd_trans_info *oti,
2671 /* use obd ops to offer management infrastructure */
2672 static struct obd_ops mds_obd_ops = {
2673 .o_owner = THIS_MODULE,
2674 .o_attach = mds_attach,
2675 .o_detach = mds_detach,
2676 .o_connect = mds_connect,
2677 .o_init_export = mds_init_export,
2678 .o_destroy_export = mds_destroy_export,
2679 .o_disconnect = mds_disconnect,
2680 .o_setup = mds_setup,
2681 .o_precleanup = mds_precleanup,
2682 .o_cleanup = mds_cleanup,
2683 .o_postrecov = mds_postrecov,
2684 .o_statfs = mds_obd_statfs,
2685 .o_iocontrol = mds_iocontrol,
2686 .o_create = mds_obd_create,
2687 .o_destroy = mds_obd_destroy,
2688 .o_llog_init = mds_llog_init,
2689 .o_llog_finish = mds_llog_finish,
2690 .o_notify = mds_notify,
2691 .o_get_info = mds_get_info,
2692 .o_set_info = mds_set_info,
2693 .o_preprw = mds_preprw,
2694 .o_commitrw = mds_commitrw,
2697 static struct obd_ops mdt_obd_ops = {
2698 .o_owner = THIS_MODULE,
2699 .o_attach = mdt_attach,
2700 .o_detach = mdt_detach,
2701 .o_setup = mdt_setup,
2702 .o_cleanup = mdt_cleanup,
2703 .o_attach = mdt_attach,
2704 .o_detach = mdt_detach,
2707 static int __init mds_init(void)
2709 struct lprocfs_static_vars lvars;
2711 lprocfs_init_multi_vars(0, &lvars);
2712 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
2714 lprocfs_init_multi_vars(1, &lvars);
2715 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
2721 static void /*__exit*/ mds_exit(void)
2723 class_unregister_type(LUSTRE_MDS_NAME);
2724 class_unregister_type(LUSTRE_MDT_NAME);
2727 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2728 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2729 MODULE_LICENSE("GPL");
2731 module_init(mds_init);
2732 module_exit(mds_exit);