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/obd_ost.h>
53 #include <linux/lustre_mds.h>
54 #include <linux/lustre_fsfilt.h>
55 #include <linux/lustre_snap.h>
56 #include <linux/lprocfs_status.h>
57 #include <linux/lustre_commit_confd.h>
59 #include "mds_internal.h"
61 static int mds_intent_policy(struct ldlm_namespace *ns,
62 struct ldlm_lock **lockp, void *req_cookie,
63 ldlm_mode_t mode, int flags, void *data);
64 static int mds_postsetup(struct obd_device *obd);
65 static int mds_cleanup(struct obd_device *obd, int flags);
67 /* Assumes caller has already pushed into the kernel filesystem context */
68 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
69 loff_t offset, int count)
71 struct ptlrpc_bulk_desc *desc;
72 struct l_wait_info lwi;
74 int rc = 0, npages, i, tmpcount, tmpsize = 0;
77 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
79 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
80 OBD_ALLOC(pages, sizeof(*pages) * npages);
82 GOTO(out, rc = -ENOMEM);
84 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
87 GOTO(out_free, rc = -ENOMEM);
89 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
90 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
92 pages[i] = alloc_pages(GFP_KERNEL, 0);
94 GOTO(cleanup_buf, rc = -ENOMEM);
96 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
99 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
100 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
101 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
102 tmpsize, offset, file->f_dentry->d_inode->i_ino,
103 file->f_dentry->d_inode->i_size);
105 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
106 kmap(pages[i]), tmpsize, &offset);
110 GOTO(cleanup_buf, rc = -EIO);
113 LASSERT(desc->bd_nob == count);
115 rc = ptlrpc_start_bulk_transfer(desc);
117 GOTO(cleanup_buf, rc);
119 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
120 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
121 OBD_FAIL_MDS_SENDPAGE, rc);
122 GOTO(abort_bulk, rc);
125 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
126 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
127 LASSERT (rc == 0 || rc == -ETIMEDOUT);
130 if (desc->bd_success &&
131 desc->bd_nob_transferred == count)
132 GOTO(cleanup_buf, rc);
134 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
137 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
138 (rc == -ETIMEDOUT) ? "timeout" : "network error",
139 desc->bd_nob_transferred, count,
140 req->rq_export->exp_client_uuid.uuid,
141 req->rq_export->exp_connection->c_remote_uuid.uuid);
143 ptlrpc_fail_export(req->rq_export);
147 ptlrpc_abort_bulk (desc);
149 for (i = 0; i < npages; i++)
151 __free_pages(pages[i], 0);
153 ptlrpc_free_bulk(desc);
155 OBD_FREE(pages, sizeof(*pages) * npages);
160 int mds_lock_mode_for_dir(struct obd_device *obd,
161 struct dentry *dentry, int mode)
165 /* any dir access needs couple locks:
166 * 1) on part of dir we gonna lookup/modify in
167 * 2) on a whole dir to protect it from concurrent splitting
168 * and to flush client's cache for readdir()
169 * so, for a given mode and dentry this routine decides what
170 * lock mode to use for lock #2:
171 * 1) if caller's gonna lookup in dir then we need to protect
172 * dir from being splitted only - LCK_CR
173 * 2) if caller's gonna modify dir then we need to protect
174 * dir from being splitted and to flush cache - LCK_CW
175 * 3) if caller's gonna modify dir and that dir seems ready
176 * for splitting then we need to protect it from any
177 * type of access (lookup/modify/split) - LCK_EX -bzzz */
179 split = mds_splitting_expected(obd, dentry);
180 if (split == MDS_NO_SPLITTABLE) {
181 /* this inode won't be splitted. so we need not to protect from
182 * just flush client's cache on modification */
187 if (mode == LCK_PR) {
189 } else if (mode == LCK_PW) {
190 /* caller gonna modify directory.we use concurrent
191 write lock here to retract client's cache for readdir */
193 if (split == MDS_EXPECT_SPLIT) {
194 /* splitting possible. serialize any access
195 * the idea is that first one seen dir is
196 * splittable is given exclusive lock and
197 * split directory. caller passes lock mode
198 * to mds_try_to_split_dir() and splitting
199 * would be done with exclusive lock only -bzzz */
200 CDEBUG(D_OTHER, "%s: gonna split %u/%u\n",
202 (unsigned) dentry->d_inode->i_ino,
203 (unsigned) dentry->d_inode->i_generation);
207 CWARN("unexpected lock mode %d\n", mode);
214 /* only valid locked dentries or errors should be returned */
215 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
216 struct vfsmount **mnt, int lock_mode,
217 struct lustre_handle *lockh, int *mode,
218 char *name, int namelen, __u64 lockpart)
220 struct mds_obd *mds = &obd->u.mds;
221 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
222 struct ldlm_res_id res_id = { .name = {0} };
224 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
231 res_id.name[0] = de->d_inode->i_ino;
232 res_id.name[1] = de->d_inode->i_generation;
235 if (name && IS_PDIROPS(de->d_inode)) {
236 ldlm_policy_data_t cpolicy =
237 { .l_inodebits = { MDS_INODELOCK_UPDATE } };
238 LASSERT(mode != NULL);
239 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
241 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
243 &cpolicy, *mode, &flags,
245 ldlm_completion_ast, NULL, NULL,
246 NULL, 0, NULL, lockh + 1);
247 if (rc != ELDLM_OK) {
249 RETURN(ERR_PTR(-ENOLCK));
254 res_id.name[2] = full_name_hash(name, namelen);
256 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
257 de->d_inode->i_ino, de->d_inode->i_generation,
261 #warning "No PDIROPS support in the kernel"
263 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
264 LDLM_IBITS, &policy, lock_mode, &flags,
265 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
266 NULL, 0, NULL, lockh);
267 if (rc != ELDLM_OK) {
269 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
272 ldlm_lock_decref(lockh + 1, LCK_CW);
279 #ifndef DCACHE_DISCONNECTED
280 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
284 /* Look up an entry by inode number. */
285 /* this function ONLY returns valid dget'd dentries with an initialized inode
287 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
288 struct vfsmount **mnt)
291 unsigned long ino = fid->id;
292 __u32 generation = fid->generation;
294 struct dentry *result;
297 RETURN(ERR_PTR(-ESTALE));
299 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
301 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
302 ino, generation, mds->mds_sb);
304 /* under ext3 this is neither supposed to return bad inodes
306 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
310 inode = result->d_inode;
312 RETURN(ERR_PTR(-ENOENT));
314 /* here we disabled generation check, as root inode i_generation
315 * of cache mds and real mds are different. */
316 if (inode->i_ino != mds->mds_rootfid.id && generation &&
317 inode->i_generation != generation) {
318 /* we didn't find the right inode.. */
319 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
320 inode->i_ino, (unsigned long)inode->i_nlink,
321 atomic_read(&inode->i_count), inode->i_generation,
324 RETURN(ERR_PTR(-ENOENT));
328 *mnt = mds->mds_vfsmnt;
336 /* Establish a connection to the MDS.
338 * This will set up an export structure for the client to hold state data
339 * about that client, like open files, the last operation number it did
340 * on the server, etc.
342 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
343 struct obd_uuid *cluuid)
345 struct obd_export *exp;
346 struct mds_export_data *med; /* */
347 struct mds_client_data *mcd;
351 if (!conn || !obd || !cluuid)
354 /* XXX There is a small race between checking the list and adding a
355 * new connection for the same UUID, but the real threat (list
356 * corruption when multiple different clients connect) is solved.
358 * There is a second race between adding the export to the list,
359 * and filling in the client data below. Hence skipping the case
360 * of NULL mcd above. We should already be controlling multiple
361 * connects at the client, and we can't hold the spinlock over
362 * memory allocations without risk of deadlocking.
364 rc = class_connect(conn, obd, cluuid);
367 exp = class_conn2export(conn);
369 med = &exp->exp_mds_data;
371 OBD_ALLOC(mcd, sizeof(*mcd));
373 CERROR("mds: out of memory for client data\n");
374 GOTO(out, rc = -ENOMEM);
377 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
380 rc = mds_client_add(obd, &obd->u.mds, med, -1);
385 OBD_FREE(mcd, sizeof(*mcd));
386 class_disconnect(exp, 0);
388 class_export_put(exp);
393 static int mds_init_export(struct obd_export *exp)
395 struct mds_export_data *med = &exp->exp_mds_data;
397 INIT_LIST_HEAD(&med->med_open_head);
398 spin_lock_init(&med->med_open_lock);
402 static int mds_destroy_export(struct obd_export *export)
404 struct mds_export_data *med;
405 struct obd_device *obd = export->exp_obd;
406 struct lvfs_run_ctxt saved;
410 med = &export->exp_mds_data;
411 target_destroy_export(export);
413 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
416 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
418 /* Close any open files (which may also cause orphan unlinking). */
419 spin_lock(&med->med_open_lock);
420 while (!list_empty(&med->med_open_head)) {
421 struct list_head *tmp = med->med_open_head.next;
422 struct mds_file_data *mfd =
423 list_entry(tmp, struct mds_file_data, mfd_list);
424 BDEVNAME_DECLARE_STORAGE(btmp);
426 /* bug 1579: fix force-closing for 2.5 */
427 struct dentry *dentry = mfd->mfd_dentry;
429 list_del(&mfd->mfd_list);
430 spin_unlock(&med->med_open_lock);
432 /* If you change this message, be sure to update
433 * replay_single:test_46 */
434 CERROR("force closing client file handle for %*s (%s:%lu)\n",
435 dentry->d_name.len, dentry->d_name.name,
436 ll_bdevname(dentry->d_inode->i_sb, btmp),
437 dentry->d_inode->i_ino);
438 rc = mds_mfd_close(NULL, obd, mfd,
439 !(export->exp_flags & OBD_OPT_FAILOVER));
442 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
443 spin_lock(&med->med_open_lock);
445 spin_unlock(&med->med_open_lock);
446 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
449 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
454 static int mds_disconnect(struct obd_export *exp, int flags)
456 struct obd_device *obd;
458 unsigned long irqflags;
463 class_export_get(exp);
465 obd = class_exp2obd(exp);
467 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
468 exp->exp_handle.h_cookie);
473 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
474 && !atomic_read(&mds->mds_real_clients)) {
475 /* there was no client at all */
476 mds_lmv_disconnect(obd, flags);
479 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
480 && atomic_dec_and_test(&mds->mds_real_clients)) {
481 /* time to drop LMV connections */
482 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
483 "Disconnnect from LMV now\n",
484 obd->obd_name, exp->exp_client_uuid.uuid);
485 mds_lmv_disconnect(obd, flags);
488 spin_lock_irqsave(&exp->exp_lock, irqflags);
489 exp->exp_flags = flags;
490 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
492 /* Disconnect early so that clients can't keep using export */
493 rc = class_disconnect(exp, flags);
494 ldlm_cancel_locks_for_export(exp);
496 /* complete all outstanding replies */
497 spin_lock_irqsave(&exp->exp_lock, irqflags);
498 while (!list_empty(&exp->exp_outstanding_replies)) {
499 struct ptlrpc_reply_state *rs =
500 list_entry(exp->exp_outstanding_replies.next,
501 struct ptlrpc_reply_state, rs_exp_list);
502 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
504 spin_lock(&svc->srv_lock);
505 list_del_init(&rs->rs_exp_list);
506 ptlrpc_schedule_difficult_reply(rs);
507 spin_unlock(&svc->srv_lock);
509 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
511 class_export_put(exp);
515 static int mds_getstatus(struct ptlrpc_request *req)
517 struct mds_obd *mds = mds_req2mds(req);
518 struct mds_body *body;
519 int rc, size = sizeof(*body);
522 rc = lustre_pack_reply(req, 1, &size, NULL);
523 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
524 CERROR("mds: out of memory for message: size=%d\n", size);
525 req->rq_status = -ENOMEM; /* superfluous? */
529 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
530 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
532 /* the last_committed and last_xid fields are filled in for all
533 * replies already - no need to do so here also.
538 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
539 void *data, int flag)
544 if (flag == LDLM_CB_CANCELING) {
545 /* Don't need to do anything here. */
549 /* XXX layering violation! -phil */
550 l_lock(&lock->l_resource->lr_namespace->ns_lock);
551 /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
552 * such that mds_blocking_ast is called just before l_i_p takes the
553 * ns_lock, then by the time we get the lock, we might not be the
554 * correct blocking function anymore. So check, and return early, if
556 if (lock->l_blocking_ast != mds_blocking_ast) {
557 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
561 lock->l_flags |= LDLM_FL_CBPENDING;
562 do_ast = (!lock->l_readers && !lock->l_writers);
563 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
566 struct lustre_handle lockh;
569 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
570 ldlm_lock2handle(lock, &lockh);
571 rc = ldlm_cli_cancel(&lockh);
573 CERROR("ldlm_cli_cancel: %d\n", rc);
575 LDLM_DEBUG(lock, "Lock still has references, will be "
581 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
589 rc = fsfilt_get_md(obd, inode, md, *size);
594 CERROR("Error %d reading eadata for ino %lu\n",
599 if (S_ISREG(inode->i_mode))
600 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
601 if (S_ISDIR(inode->i_mode))
602 rc = mds_convert_mea_ea(obd, inode, md, lmm_size);
616 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
617 * Call with lock=0 if the caller has already taken the i_sem. */
618 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
619 struct mds_body *body, struct inode *inode, int lock)
621 struct mds_obd *mds = &obd->u.mds;
627 lmm = lustre_msg_buf(msg, offset, 0);
629 /* Some problem with getting eadata when I sized the reply
631 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
635 lmm_size = msg->buflens[offset];
637 /* I don't really like this, but it is a sanity check on the client
638 * MD request. However, if the client doesn't know how much space
639 * to reserve for the MD, it shouldn't be bad to have too much space.
641 if (lmm_size > mds->mds_max_mdsize) {
642 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
643 inode->i_ino, lmm_size, mds->mds_max_mdsize);
647 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
649 if (S_ISDIR(inode->i_mode))
650 body->valid |= OBD_MD_FLDIREA;
652 body->valid |= OBD_MD_FLEASIZE;
653 body->eadatasize = lmm_size;
660 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
661 struct ptlrpc_request *req,
662 struct mds_body *reqbody, int reply_off)
664 struct mds_body *body;
665 struct inode *inode = dentry->d_inode;
669 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
672 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
673 LASSERT(body != NULL); /* caller prepped reply */
675 if (dentry->d_flags & DCACHE_CROSS_REF) {
676 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
677 (unsigned long) dentry->d_mdsnum,
678 (unsigned long) dentry->d_inum,
679 (unsigned long) dentry->d_generation);
680 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
681 body->fid1.id = dentry->d_inum;
682 body->fid1.mds = dentry->d_mdsnum;
683 body->fid1.generation = dentry->d_generation;
686 mds_pack_inode2fid(obd, &body->fid1, inode);
687 mds_pack_inode2body(obd, body, inode);
689 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
690 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
691 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
694 /* If we have LOV EA data, the OST holds size, atime, mtime */
695 if (!(body->valid & OBD_MD_FLEASIZE) &&
696 !(body->valid & OBD_MD_FLDIREA))
697 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
698 OBD_MD_FLATIME | OBD_MD_FLMTIME);
699 } else if (S_ISLNK(inode->i_mode) &&
700 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
701 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
704 LASSERT (symname != NULL); /* caller prepped reply */
705 len = req->rq_repmsg->buflens[reply_off + 1];
707 rc = inode->i_op->readlink(dentry, symname, len);
709 CERROR("readlink failed: %d\n", rc);
710 } else if (rc != len - 1) {
711 CERROR ("Unexpected readlink rc %d: expecting %d\n",
715 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
716 body->valid |= OBD_MD_LINKNAME;
717 body->eadatasize = rc + 1;
718 symname[rc] = 0; /* NULL terminate */
726 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
727 struct dentry *dentry,
730 int rc = 0, size[1] = {sizeof(struct mds_body)};
733 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
734 CERROR("failed MDS_GETATTR_PACK test\n");
735 req->rq_status = -ENOMEM;
736 GOTO(out, rc = -ENOMEM);
739 rc = lustre_pack_reply(req, 1, size, NULL);
741 CERROR("out of memory\n");
742 GOTO(out, req->rq_status = rc);
750 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
753 struct mds_obd *mds = mds_req2mds(req);
754 struct mds_body *body;
755 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
758 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
759 LASSERT(body != NULL); /* checked by caller */
760 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
762 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
763 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
766 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
768 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
772 CERROR("error getting inode %lu MD: rc = %d\n",
775 } else if (rc > mds->mds_max_mdsize) {
777 CERROR("MD size %d larger than maximum possible %u\n",
778 rc, mds->mds_max_mdsize);
783 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
784 if (inode->i_size + 1 != body->eadatasize)
785 CERROR("symlink size: %Lu, reply space: %d\n",
786 inode->i_size + 1, body->eadatasize);
787 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
789 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
790 inode->i_size + 1, body->eadatasize);
793 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
794 CERROR("failed MDS_GETATTR_PACK test\n");
795 req->rq_status = -ENOMEM;
796 GOTO(out, rc = -ENOMEM);
799 rc = lustre_pack_reply(req, bufcount, size, NULL);
801 CERROR("out of memory\n");
802 GOTO(out, req->rq_status = rc);
810 int mds_check_mds_num(struct obd_device *obd, struct inode* inode,
811 char *name, int namelen)
813 struct mea *mea = NULL;
814 int mea_size, rc = 0;
817 rc = mds_get_lmv_attr(obd, inode, &mea, &mea_size);
821 /* dir is already splitted, check if requested filename
822 * should live at this MDS or at another one */
824 i = mea_name2idx(mea, name, namelen - 1);
825 if (mea->mea_master != mea->mea_fids[i].mds) {
827 "inapropriate MDS(%d) for %s. should be %d(%d)\n",
828 mea->mea_master, name, mea->mea_fids[i].mds, i);
834 OBD_FREE(mea, mea_size);
838 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
839 struct lustre_handle *child_lockh, int child_part)
841 struct obd_device *obd = req->rq_export->exp_obd;
842 struct ldlm_reply *rep = NULL;
843 struct lvfs_run_ctxt saved;
844 struct mds_body *body;
845 struct dentry *dparent = NULL, *dchild = NULL;
846 struct lvfs_ucred uc;
847 struct lustre_handle parent_lockh[2];
848 int namesize, update_mode;
849 int rc = 0, cleanup_phase = 0, resent_req = 0;
850 struct clonefs_info *clone_info = NULL;
854 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
856 /* Swab now, before anyone looks inside the request */
858 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
859 lustre_swab_mds_body);
861 CERROR("Can't swab mds_body\n");
862 GOTO(cleanup, rc = -EFAULT);
865 LASSERT_REQSWAB(req, offset + 1);
866 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
868 CERROR("Can't unpack name\n");
869 GOTO(cleanup, rc = -EFAULT);
872 clone_info = lustre_swab_reqbuf(req, offset + 2, sizeof(*clone_info),
873 lustre_swab_clonefs_info);
875 CDEBUG(D_INFO,"getattr name %s clone_info index %d \n", name,
876 clone_info->clone_index);
879 namesize = req->rq_reqmsg->buflens[offset + 1];
881 LASSERT (offset == 0 || offset == 2);
882 /* if requests were at offset 2, the getattr reply goes back at 1 */
884 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
888 uc.luc_fsuid = body->fsuid;
889 uc.luc_fsgid = body->fsgid;
890 uc.luc_cap = body->capability;
891 uc.luc_suppgid1 = body->suppgid;
892 uc.luc_suppgid2 = -1;
893 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
894 cleanup_phase = 1; /* kernel context */
895 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
897 LASSERT(namesize > 0);
899 /* we have no dentry here, drop LOOKUP bit */
900 child_part &= ~MDS_INODELOCK_LOOKUP;
901 CDEBUG(D_OTHER, "%s: request to retrieve attrs for %lu/%lu\n",
902 obd->obd_name, (unsigned long) body->fid1.id,
903 (unsigned long) body->fid1.generation);
904 dchild = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR,
905 parent_lockh, &update_mode,
906 NULL, 0, child_part);
907 if (IS_ERR(dchild)) {
908 CERROR("can't find inode: %d\n", (int) PTR_ERR(dchild));
909 GOTO(cleanup, rc = PTR_ERR(dchild));
911 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
913 if (parent_lockh[1].cookie)
914 ldlm_lock_decref(parent_lockh + 1, update_mode);
921 /* FIXME: handle raw lookup */
922 if (body->valid == OBD_MD_FLID) {
923 struct mds_obd *mds = &obd->u.mds;
924 struct mds_body *mds_reply;
925 int size = sizeof(*mds_reply);
928 dparent = mds_fid2dentry(mds, &body->fid1, NULL);
929 if (IS_ERR(dparent)) {
930 rc = PTR_ERR(dparent);
933 // The user requested ONLY the inode number, so do a raw lookup
934 rc = lustre_pack_reply(req, 1, &size, NULL);
936 CERROR("out of memory\n");
940 dir = dparent->d_inode;
941 LASSERT(dir->i_op->lookup_raw != NULL);
942 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
944 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
946 mds_reply->fid1.id = inum;
947 mds_reply->valid = OBD_MD_FLID;
952 if (child_lockh->cookie != 0) {
953 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
957 if (resent_req == 0) {
958 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
959 parent_lockh, &dparent,
960 LCK_PR, MDS_INODELOCK_LOOKUP,
961 &update_mode, name, namesize,
962 child_lockh, &dchild, LCK_PR,
963 child_part, clone_info);
967 struct ldlm_lock *granted_lock;
968 struct ll_fid child_fid;
969 struct ldlm_resource *res;
970 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
971 granted_lock = ldlm_handle2lock(child_lockh);
972 LASSERT(granted_lock);
974 res = granted_lock->l_resource;
975 child_fid.id = res->lr_name.name[0];
976 child_fid.generation = res->lr_name.name[1];
977 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
979 LDLM_LOCK_PUT(granted_lock);
982 cleanup_phase = 2; /* dchild, dparent, locks */
984 /* let's make sure this name should leave on this mds node */
985 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
991 if (!DENTRY_VALID(dchild)) {
992 intent_set_disposition(rep, DISP_LOOKUP_NEG);
993 /* in the intent case, the policy clears this error:
994 the disposition is enough */
998 intent_set_disposition(rep, DISP_LOOKUP_POS);
1001 if (req->rq_repmsg == NULL) {
1002 if (dchild->d_flags & DCACHE_CROSS_REF)
1003 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1005 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
1007 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1012 rc = mds_getattr_internal(obd, dchild, req, body, offset);
1013 GOTO(cleanup, rc); /* returns the lock to the client */
1016 switch (cleanup_phase) {
1018 if (resent_req == 0) {
1019 if (rc && DENTRY_VALID(dchild))
1020 ldlm_lock_decref(child_lockh, LCK_PR);
1022 ldlm_lock_decref(parent_lockh, LCK_PR);
1024 if (parent_lockh[1].cookie != 0)
1025 ldlm_lock_decref(parent_lockh + 1,
1034 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1040 static int mds_getattr(int offset, struct ptlrpc_request *req)
1042 struct mds_obd *mds = mds_req2mds(req);
1043 struct obd_device *obd = req->rq_export->exp_obd;
1044 struct lvfs_run_ctxt saved;
1046 struct mds_body *body;
1047 struct lvfs_ucred uc;
1051 body = lustre_swab_reqbuf (req, offset, sizeof (*body),
1052 lustre_swab_mds_body);
1054 CERROR ("Can't unpack body\n");
1058 uc.luc_fsuid = body->fsuid;
1059 uc.luc_fsgid = body->fsgid;
1060 uc.luc_cap = body->capability;
1061 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1062 de = mds_fid2dentry(mds, &body->fid1, NULL);
1064 rc = req->rq_status = -ENOENT;
1065 GOTO(out_pop, PTR_ERR(de));
1068 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1070 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1074 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
1079 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1084 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1085 unsigned long max_age)
1089 spin_lock(&obd->obd_osfs_lock);
1090 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1092 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1093 spin_unlock(&obd->obd_osfs_lock);
1098 static int mds_statfs(struct ptlrpc_request *req)
1100 struct obd_device *obd = req->rq_export->exp_obd;
1101 int rc, size = sizeof(struct obd_statfs);
1104 rc = lustre_pack_reply(req, 1, &size, NULL);
1105 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1106 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1110 /* We call this so that we can cache a bit - 1 jiffie worth */
1111 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1114 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1120 req->rq_status = rc;
1124 static int mds_sync(struct ptlrpc_request *req)
1126 struct obd_device *obd = req->rq_export->exp_obd;
1127 struct mds_obd *mds = &obd->u.mds;
1128 struct mds_body *body;
1129 int rc, size = sizeof(*body);
1132 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
1134 GOTO(out, rc = -EPROTO);
1136 rc = lustre_pack_reply(req, 1, &size, NULL);
1137 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1138 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1142 if (body->fid1.id == 0) {
1143 /* a fid of zero is taken to mean "sync whole filesystem" */
1144 rc = fsfilt_sync(obd, mds->mds_sb);
1148 /* just any file to grab fsync method - "file" arg unused */
1149 struct file *file = mds->mds_rcvd_filp;
1152 de = mds_fid2dentry(mds, &body->fid1, NULL);
1154 GOTO(out, rc = PTR_ERR(de));
1156 rc = file->f_op->fsync(NULL, de, 1);
1161 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1162 mds_pack_inode2fid(obd, &body->fid1, de->d_inode);
1163 mds_pack_inode2body(obd, body, de->d_inode);
1166 req->rq_status = rc;
1170 /* mds_readpage does not take a DLM lock on the inode, because the client must
1171 * already have a PR lock.
1173 * If we were to take another one here, a deadlock will result, if another
1174 * thread is already waiting for a PW lock. */
1175 static int mds_readpage(struct ptlrpc_request *req)
1177 struct obd_device *obd = req->rq_export->exp_obd;
1178 struct vfsmount *mnt;
1181 struct mds_body *body, *repbody;
1182 struct lvfs_run_ctxt saved;
1183 int rc, size = sizeof(*repbody);
1184 struct lvfs_ucred uc;
1187 rc = lustre_pack_reply(req, 1, &size, NULL);
1188 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1189 CERROR("mds: out of memory\n");
1190 GOTO(out, rc = -ENOMEM);
1193 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1195 GOTO (out, rc = -EFAULT);
1197 uc.luc_fsuid = body->fsuid;
1198 uc.luc_fsgid = body->fsgid;
1199 uc.luc_cap = body->capability;
1200 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1201 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1203 GOTO(out_pop, rc = PTR_ERR(de));
1205 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1207 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1208 /* note: in case of an error, dentry_open puts dentry */
1210 GOTO(out_pop, rc = PTR_ERR(file));
1212 /* body->size is actually the offset -eeb */
1213 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1214 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1215 body->size, de->d_inode->i_blksize);
1216 GOTO(out_file, rc = -EFAULT);
1219 /* body->nlink is actually the #bytes to read -eeb */
1220 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1221 CERROR("size %u is not multiple of blocksize %lu\n",
1222 body->nlink, de->d_inode->i_blksize);
1223 GOTO(out_file, rc = -EFAULT);
1226 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1227 repbody->size = file->f_dentry->d_inode->i_size;
1228 repbody->valid = OBD_MD_FLSIZE;
1230 /* to make this asynchronous make sure that the handling function
1231 doesn't send a reply when this function completes. Instead a
1232 callback function would send the reply */
1233 /* body->size is actually the offset -eeb */
1234 rc = mds_sendpage(req, file, body->size, body->nlink);
1237 filp_close(file, 0);
1239 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1241 req->rq_status = rc;
1245 int mds_reint(struct ptlrpc_request *req, int offset,
1246 struct lustre_handle *lockh)
1248 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
1252 OBD_ALLOC(rec, sizeof(*rec));
1256 rc = mds_update_unpack(req, offset, rec);
1257 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1258 CERROR("invalid record\n");
1259 GOTO(out, req->rq_status = -EINVAL);
1261 /* rc will be used to interrupt a for loop over multiple records */
1262 rc = mds_reint_rec(rec, offset, req, lockh);
1264 OBD_FREE(rec, sizeof(*rec));
1268 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1269 struct obd_device *obd, int *process)
1271 switch (req->rq_reqmsg->opc) {
1272 case MDS_CONNECT: /* This will never get here, but for completeness. */
1273 case OST_CONNECT: /* This will never get here, but for completeness. */
1274 case MDS_DISCONNECT:
1275 case OST_DISCONNECT:
1280 case MDS_SYNC: /* used in unmounting */
1285 *process = target_queue_recovery_request(req, obd);
1289 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1291 /* XXX what should we set rq_status to here? */
1292 req->rq_status = -EAGAIN;
1293 RETURN(ptlrpc_error(req));
1297 static char *reint_names[] = {
1298 [REINT_SETATTR] "setattr",
1299 [REINT_CREATE] "create",
1300 [REINT_LINK] "link",
1301 [REINT_UNLINK] "unlink",
1302 [REINT_RENAME] "rename",
1303 [REINT_OPEN] "open",
1306 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER |\
1307 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ|\
1308 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME|\
1311 static void reconstruct_create(struct ptlrpc_request *req)
1313 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1314 struct mds_client_data *mcd = med->med_mcd;
1315 struct ost_body *body;
1317 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1319 /* copy rc, transno and disp; steal locks */
1320 mds_req_from_mcd(req, mcd);
1321 CERROR("reconstruct reply for x"LPU64"\n", req->rq_xid);
1324 static int mdt_obj_create(struct ptlrpc_request *req)
1326 struct obd_device *obd = req->rq_export->exp_obd;
1327 struct ldlm_res_id res_id = { .name = {0} };
1328 struct mds_obd *mds = &obd->u.mds;
1329 struct ost_body *body, *repbody;
1330 char fidname[LL_FID_NAMELEN];
1331 struct inode *parent_inode;
1332 struct lustre_handle lockh;
1333 struct lvfs_run_ctxt saved;
1334 ldlm_policy_data_t policy;
1335 struct dentry *new = NULL;
1336 struct dentry_params dp;
1337 int mealen, flags = 0, rc, size = sizeof(*repbody), cleanup_phase = 0;
1338 unsigned int tmpname;
1339 struct lvfs_ucred uc;
1341 void *handle = NULL;
1344 DEBUG_REQ(D_HA, req, "create remote object");
1346 parent_inode = mds->mds_objects_dir->d_inode;
1348 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
1352 rc = lustre_pack_reply(req, 1, &size, NULL);
1356 MDS_CHECK_RESENT(req, reconstruct_create(req));
1358 uc.luc_fsuid = body->oa.o_uid;
1359 uc.luc_fsgid = body->oa.o_gid;
1361 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1363 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1365 if (body->oa.o_flags & OBD_FL_RECREATE_OBJS) {
1366 /* this is re-create request from MDS holding directory name.
1367 * we have to lookup given ino/generation first. if it exists
1368 * (good case) then there is nothing to do. if it does not
1369 * then we have to recreate it */
1371 fid.id = body->oa.o_id;
1372 fid.generation = body->oa.o_generation;
1373 new = mds_fid2dentry(mds, &fid, NULL);
1374 if (!IS_ERR(new) && new->d_inode) {
1375 CWARN("mkdir() repairing is on its way: %lu/%lu\n",
1376 (unsigned long) fid.id,
1377 (unsigned long) fid.generation);
1378 obdo_from_inode(&repbody->oa, new->d_inode,
1379 FILTER_VALID_FLAGS);
1380 repbody->oa.o_id = new->d_inode->i_ino;
1381 repbody->oa.o_generation = new->d_inode->i_generation;
1382 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1384 GOTO(cleanup, rc = 0);
1386 CWARN("hmm. for some reason dir %lu/%lu (or reply) got lost\n",
1387 (unsigned long) fid.id, (unsigned long) fid.generation);
1388 LASSERT(new->d_inode == NULL ||
1389 new->d_inode->i_generation != fid.generation);
1393 down(&parent_inode->i_sem);
1394 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1395 LASSERT(!IS_ERR(handle));
1396 cleanup_phase = 1; /* transaction */
1399 tmpname = ll_insecure_random_int();
1400 rc = sprintf(fidname, "%u", tmpname);
1401 new = lookup_one_len(fidname, mds->mds_objects_dir, rc);
1403 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1404 obd->obd_name, fidname, (int) PTR_ERR(new));
1405 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
1406 up(&parent_inode->i_sem);
1407 RETURN(PTR_ERR(new));
1408 } else if (new->d_inode) {
1409 CERROR("%s: name exists. repeat\n", obd->obd_name);
1413 new->d_fsdata = (void *) &dp;
1417 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1418 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1419 (unsigned long) body->oa.o_id,
1420 (unsigned long) body->oa.o_generation);
1421 dp.p_inum = body->oa.o_id;
1422 dp.p_generation = body->oa.o_generation;
1424 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1426 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1427 repbody->oa.o_id = new->d_inode->i_ino;
1428 repbody->oa.o_generation = new->d_inode->i_generation;
1429 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1431 rc = fsfilt_del_dir_entry(obd, new);
1432 up(&parent_inode->i_sem);
1435 CERROR("can't remove name for object: %d\n", rc);
1439 /* this lock should be taken to serialize MDS modifications
1440 * in failure case */
1441 res_id.name[0] = new->d_inode->i_ino;
1442 res_id.name[1] = new->d_inode->i_generation;
1443 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1444 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1445 res_id, LDLM_IBITS, &policy,
1446 LCK_EX, &flags, mds_blocking_ast,
1447 ldlm_completion_ast, NULL, NULL,
1448 NULL, 0, NULL, &lockh);
1451 cleanup_phase = 2; /* valid lockh */
1453 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
1454 (unsigned long) new->d_inode->i_ino,
1455 (unsigned long) new->d_inode->i_generation,
1456 (unsigned) new->d_inode->i_mode);
1458 up(&parent_inode->i_sem);
1459 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1463 if (body->oa.o_valid & OBD_MD_FLID) {
1464 /* this is new object for splitted dir. we have to
1465 * prevent recursive splitting on it -bzzz */
1466 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1467 OBD_ALLOC(mea, mealen);
1469 GOTO(cleanup, rc = -ENOMEM);
1470 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
1471 mea->mea_master = 0;
1473 down(&new->d_inode->i_sem);
1474 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1475 up(&new->d_inode->i_sem);
1476 OBD_FREE(mea, mealen);
1477 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
1478 obd->obd_name, new->d_inode->i_ino,
1479 new->d_inode->i_generation, flags);
1480 } else if (body->oa.o_easize) {
1481 /* we pass LCK_EX to split routine to signal that we have
1482 * exclusive access to the directory. simple because nobody
1483 * knows it already exists -bzzz */
1484 mds_try_to_split_dir(obd, new, NULL, body->oa.o_easize, LCK_EX);
1488 switch (cleanup_phase) {
1489 case 2: /* valid lockh */
1491 ptlrpc_save_lock(req, &lockh, LCK_EX);
1493 ldlm_lock_decref(&lockh, LCK_EX);
1494 case 1: /* transaction */
1495 rc = mds_finish_transno(mds, parent_inode, handle, req, rc, 0);
1499 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1503 static int mdt_get_info(struct ptlrpc_request *req)
1506 struct obd_export *exp = req->rq_export;
1507 int keylen, rc = 0, size = sizeof(obd_id);
1511 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1513 DEBUG_REQ(D_HA, req, "no get_info key");
1516 keylen = req->rq_reqmsg->buflens[0];
1518 if (keylen < strlen("mdsize") || memcmp(key, "mdsize", 6) != 0)
1521 rc = lustre_pack_reply(req, 1, &size, NULL);
1525 reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
1526 rc = obd_get_info(exp, keylen, key, &size, reply);
1527 req->rq_repmsg->status = 0;
1531 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1532 void *key, __u32 vallen, void *val)
1534 struct obd_device *obd;
1535 struct mds_obd *mds;
1539 obd = class_exp2obd(exp);
1541 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1542 exp->exp_handle.h_cookie);
1546 #define KEY_IS(str) \
1547 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
1550 if (KEY_IS("mds_num")) {
1553 CDEBUG(D_IOCTL, "set mds num %d\n", *(int*)val);
1554 mds->mds_num = *(int*)val;
1555 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
1556 valsize = sizeof(group);
1557 /*mds number has been changed, so the corresponding obdfilter exp
1558 *need to be changed too*/
1559 rc = obd_set_info(mds->mds_osc_exp, strlen("mds_conn"), "mds_conn",
1562 } else if (KEY_IS("client")) {
1563 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
1564 atomic_inc(&mds->mds_real_clients);
1565 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
1566 obd->obd_name, exp->exp_client_uuid.uuid,
1567 atomic_read(&mds->mds_real_clients));
1568 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
1570 if (mds->mds_lmv_name) {
1571 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1576 CDEBUG(D_IOCTL, "invalid key\n");
1580 static int mdt_set_info(struct ptlrpc_request *req)
1583 struct obd_export *exp = req->rq_export;
1584 int keylen, rc = 0, vallen;
1587 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1589 DEBUG_REQ(D_HA, req, "no set_info key");
1592 keylen = req->rq_reqmsg->buflens[0];
1594 if (keylen == strlen("mds_num") &&
1595 memcmp(key, "mds_num", keylen) == 0) {
1596 rc = lustre_pack_reply(req, 0, NULL, NULL);
1599 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
1601 vallen = req->rq_reqmsg->buflens[1];
1603 rc = obd_set_info(exp, keylen, key, vallen, val);
1604 req->rq_repmsg->status = 0;
1606 } else if (keylen == strlen("client") &&
1607 memcmp(key, "client", keylen) == 0) {
1608 rc = lustre_pack_reply(req, 0, NULL, NULL);
1611 rc = obd_set_info(exp, keylen, key, sizeof(obd_id), NULL);
1612 req->rq_repmsg->status = 0;
1615 CDEBUG(D_IOCTL, "invalid key\n");
1619 int mds_handle(struct ptlrpc_request *req)
1621 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1623 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1624 struct obd_device *obd = NULL;
1627 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1629 LASSERT(current->journal_info == NULL);
1630 /* XXX identical to OST */
1631 if (req->rq_reqmsg->opc != MDS_CONNECT) {
1632 struct mds_export_data *med;
1635 if (req->rq_export == NULL) {
1636 CERROR("lustre_mds: operation %d on unconnected MDS\n",
1637 req->rq_reqmsg->opc);
1638 req->rq_status = -ENOTCONN;
1639 GOTO(out, rc = -ENOTCONN);
1642 med = &req->rq_export->exp_mds_data;
1643 obd = req->rq_export->exp_obd;
1646 /* sanity check: if the xid matches, the request must
1647 * be marked as a resent or replayed */
1648 if (req->rq_xid == med->med_mcd->mcd_last_xid)
1649 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
1650 (MSG_RESENT | MSG_REPLAY),
1651 "rq_xid "LPU64" matches last_xid, "
1652 "expected RESENT flag\n",
1654 /* else: note the opposite is not always true; a
1655 * RESENT req after a failover will usually not match
1656 * the last_xid, since it was likely never
1657 * committed. A REPLAYed request will almost never
1658 * match the last xid, however it could for a
1659 * committed, but still retained, open. */
1661 spin_lock_bh(&obd->obd_processing_task_lock);
1662 recovering = obd->obd_recovering;
1663 spin_unlock_bh(&obd->obd_processing_task_lock);
1665 rc = mds_filter_recovery_request(req, obd,
1667 if (rc || should_process == 0) {
1669 } else if (should_process < 0) {
1670 req->rq_status = should_process;
1671 rc = ptlrpc_error(req);
1677 switch (req->rq_reqmsg->opc) {
1679 DEBUG_REQ(D_INODE, req, "connect");
1680 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1681 rc = target_handle_connect(req);
1683 /* Now that we have an export, set mds. */
1684 mds = mds_req2mds(req);
1687 case MDS_DISCONNECT:
1688 DEBUG_REQ(D_INODE, req, "disconnect");
1689 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1690 rc = target_handle_disconnect(req);
1691 req->rq_status = rc; /* superfluous? */
1695 DEBUG_REQ(D_INODE, req, "getstatus");
1696 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1697 rc = mds_getstatus(req);
1701 DEBUG_REQ(D_INODE, req, "getattr");
1702 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1703 rc = mds_getattr(0, req);
1706 case MDS_GETATTR_NAME: {
1707 struct lustre_handle lockh;
1708 DEBUG_REQ(D_INODE, req, "getattr_name");
1709 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1711 /* If this request gets a reconstructed reply, we won't be
1712 * acquiring any new locks in mds_getattr_name, so we don't
1716 rc = mds_getattr_name(0, req, &lockh, MDS_INODELOCK_UPDATE);
1717 /* this non-intent call (from an ioctl) is special */
1718 req->rq_status = rc;
1719 if (rc == 0 && lockh.cookie)
1720 ldlm_lock_decref(&lockh, LCK_PR);
1724 DEBUG_REQ(D_INODE, req, "statfs");
1725 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1726 rc = mds_statfs(req);
1730 DEBUG_REQ(D_INODE, req, "readpage");
1731 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1732 rc = mds_readpage(req);
1734 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1735 if (req->rq_reply_state) {
1736 lustre_free_reply_state (req->rq_reply_state);
1737 req->rq_reply_state = NULL;
1745 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1747 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1748 mds->mds_max_cookiesize};
1751 /* NB only peek inside req now; mds_reint() will swab it */
1753 CERROR ("Can't inspect opcode\n");
1758 if (lustre_msg_swabbed (req->rq_reqmsg))
1761 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1762 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1763 reint_names[opc] == NULL) ? reint_names[opc] :
1766 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1768 if (opc == REINT_UNLINK)
1770 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1775 rc = lustre_pack_reply(req, bufcount, size, NULL);
1779 rc = mds_reint(req, 0, NULL);
1780 fail = OBD_FAIL_MDS_REINT_NET_REP;
1785 DEBUG_REQ(D_INODE, req, "close");
1786 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1787 rc = mds_close(req);
1790 case MDS_DONE_WRITING:
1791 DEBUG_REQ(D_INODE, req, "done_writing");
1792 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1793 rc = mds_done_writing(req);
1797 DEBUG_REQ(D_INODE, req, "pin");
1798 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1803 DEBUG_REQ(D_INODE, req, "sync");
1804 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1809 DEBUG_REQ(D_INODE, req, "ping");
1810 rc = target_handle_ping(req);
1813 case OBD_LOG_CANCEL:
1814 CDEBUG(D_INODE, "log cancel\n");
1815 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1816 rc = -ENOTSUPP; /* la la la */
1820 DEBUG_REQ(D_INODE, req, "enqueue");
1821 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1822 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1823 ldlm_server_blocking_ast, NULL);
1826 DEBUG_REQ(D_INODE, req, "convert");
1827 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1828 rc = ldlm_handle_convert(req);
1830 case LDLM_BL_CALLBACK:
1831 case LDLM_CP_CALLBACK:
1832 DEBUG_REQ(D_INODE, req, "callback");
1833 CERROR("callbacks should not happen on MDS\n");
1835 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1837 case LLOG_ORIGIN_HANDLE_OPEN:
1838 DEBUG_REQ(D_INODE, req, "llog_init");
1839 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1840 rc = llog_origin_handle_open(req);
1842 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1843 DEBUG_REQ(D_INODE, req, "llog next block");
1844 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1845 rc = llog_origin_handle_next_block(req);
1847 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1848 DEBUG_REQ(D_INODE, req, "llog prev block");
1849 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1850 rc = llog_origin_handle_prev_block(req);
1852 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1853 DEBUG_REQ(D_INODE, req, "llog read header");
1854 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1855 rc = llog_origin_handle_read_header(req);
1857 case LLOG_ORIGIN_HANDLE_CLOSE:
1858 DEBUG_REQ(D_INODE, req, "llog close");
1859 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1860 rc = llog_origin_handle_close(req);
1863 DEBUG_REQ(D_INODE, req, "ost_create");
1864 rc = mdt_obj_create(req);
1867 DEBUG_REQ(D_INODE, req, "get_info");
1868 rc = mdt_get_info(req);
1871 DEBUG_REQ(D_INODE, req, "set_info");
1872 rc = mdt_set_info(req);
1875 CDEBUG(D_INODE, "write\n");
1876 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1877 rc = ost_brw_write(req, NULL);
1878 LASSERT(current->journal_info == NULL);
1879 /* mdt_brw sends its own replies */
1883 DEBUG_REQ(D_INODE, req, "llog catinfo");
1884 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1885 rc = llog_catinfo(req);
1888 req->rq_status = -ENOTSUPP;
1889 rc = ptlrpc_error(req);
1893 LASSERT(current->journal_info == NULL);
1897 /* If we're DISCONNECTing, the mds_export_data is already freed */
1898 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1899 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1900 struct obd_device *obd = list_entry(mds, struct obd_device,
1902 req->rq_repmsg->last_xid =
1903 le64_to_cpu(med->med_mcd->mcd_last_xid);
1905 if (!obd->obd_no_transno) {
1906 req->rq_repmsg->last_committed =
1907 obd->obd_last_committed;
1909 DEBUG_REQ(D_IOCTL, req,
1910 "not sending last_committed update");
1912 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1914 mds->mds_last_transno, obd->obd_last_committed,
1919 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1920 if (obd && obd->obd_recovering) {
1921 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1922 return target_queue_final_reply(req, rc);
1924 /* Lost a race with recovery; let the error path DTRT. */
1925 rc = req->rq_status = -ENOTCONN;
1928 target_send_reply(req, rc, fail);
1932 /* Update the server data on disk. This stores the new mount_count and
1933 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1934 * then the server last_rcvd value may be less than that of the clients.
1935 * This will alert us that we may need to do client recovery.
1937 * Also assumes for mds_last_transno that we are not modifying it (no locking).
1939 int mds_update_server_data(struct obd_device *obd, int force_sync)
1941 struct mds_obd *mds = &obd->u.mds;
1942 struct mds_server_data *msd = mds->mds_server_data;
1943 struct file *filp = mds->mds_rcvd_filp;
1944 struct lvfs_run_ctxt saved;
1949 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1950 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1952 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1953 mds->mds_mount_count, mds->mds_last_transno);
1954 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1956 CERROR("error writing MDS server data: rc = %d\n", rc);
1957 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1962 /* mount the file system (secretly) */
1963 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1965 struct lustre_cfg* lcfg = buf;
1966 struct mds_obd *mds = &obd->u.mds;
1967 char *options = NULL;
1968 struct vfsmount *mnt;
1973 dev_clear_rdonly(2);
1975 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1976 RETURN(rc = -EINVAL);
1978 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1979 if (IS_ERR(obd->obd_fsops))
1980 RETURN(rc = PTR_ERR(obd->obd_fsops));
1982 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1984 page = __get_free_page(GFP_KERNEL);
1988 options = (char *)page;
1989 memset(options, 0, PAGE_SIZE);
1991 /* here we use "iopen_nopriv" hardcoded, because it affects MDS utility
1992 * and the rest of options are passed by mount options. Probably this
1993 * should be moved to somewhere else like startup scripts or lconf. */
1994 sprintf(options, "iopen_nopriv");
1996 if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
1997 sprintf(options + strlen(options), ",%s",
1998 lcfg->lcfg_inlbuf4);
2000 /* we have to know mdsnum before touching underlying fs -bzzz */
2001 if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
2002 strcmp(lcfg->lcfg_inlbuf5, "dumb")) {
2005 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
2006 obd->obd_name, lcfg->lcfg_inlbuf5);
2008 generate_random_uuid(uuid);
2009 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
2011 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2012 if (mds->mds_lmv_name == NULL)
2013 RETURN(rc = -ENOMEM);
2015 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf5,
2016 lcfg->lcfg_inllen5);
2018 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
2020 OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2025 /* FIXME-WANGDI: this should be reworked when we will use lmv along
2026 * with cobd, because correct mdsnum is set in mds_lmv_connect(). */
2027 if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 && !mds->mds_lmv_obd &&
2028 strcmp(lcfg->lcfg_inlbuf6, "dumb")) {
2029 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master")) &&
2030 mds->mds_num == 0) {
2031 mds->mds_num = REAL_MDS_NUMBER;
2032 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache")) &&
2033 mds->mds_num == 0) {
2034 mds->mds_num = CACHE_MDS_NUMBER;
2038 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
2039 lcfg->lcfg_inlbuf1, options);
2045 CERROR("do_kern_mount failed: rc = %d\n", rc);
2049 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
2051 sema_init(&mds->mds_orphan_recovery_sem, 1);
2052 sema_init(&mds->mds_epoch_sem, 1);
2053 spin_lock_init(&mds->mds_transno_lock);
2054 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2055 atomic_set(&mds->mds_open_count, 0);
2056 atomic_set(&mds->mds_real_clients, 0);
2058 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
2059 LDLM_NAMESPACE_SERVER);
2060 if (obd->obd_namespace == NULL) {
2061 mds_cleanup(obd, 0);
2062 GOTO(err_put, rc = -ENOMEM);
2064 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
2066 rc = mds_fs_setup(obd, mnt);
2068 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
2072 rc = llog_start_commit_thread();
2076 /* this check for @dumb string is needed to handle mounting MDS with
2077 * smfs. Read lconf:MDSDEV.write_conf() for more details. */
2078 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2079 strcmp(lcfg->lcfg_inlbuf3, "dumb")) {
2082 generate_random_uuid(uuid);
2083 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
2085 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
2086 if (mds->mds_profile == NULL)
2087 GOTO(err_fs, rc = -ENOMEM);
2089 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
2090 lcfg->lcfg_inllen3);
2093 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2094 "mds_ldlm_client", &obd->obd_ldlm_client);
2095 obd->obd_replayable = 1;
2097 rc = mds_postsetup(obd);
2104 /* No extra cleanup needed for llog_init_commit_thread() */
2105 mds_fs_cleanup(obd, 0);
2107 ldlm_namespace_free(obd->obd_namespace, 0);
2108 obd->obd_namespace = NULL;
2111 mntput(mds->mds_vfsmnt);
2115 fsfilt_put_ops(obd->obd_fsops);
2119 static int mds_postsetup(struct obd_device *obd)
2121 struct mds_obd *mds = &obd->u.mds;
2125 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
2126 obd, 0, NULL, &llog_lvfs_ops);
2130 if (mds->mds_profile) {
2131 struct llog_ctxt *lgctxt;
2132 struct lvfs_run_ctxt saved;
2133 struct lustre_profile *lprof;
2134 struct config_llog_instance cfg;
2136 cfg.cfg_instance = NULL;
2137 cfg.cfg_uuid = mds->mds_lov_uuid;
2138 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2140 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
2142 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2143 GOTO(err_llog, rc = -EINVAL);
2146 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
2147 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2152 lprof = class_get_profile(mds->mds_profile);
2153 if (lprof == NULL) {
2154 CERROR("No profile found: %s\n", mds->mds_profile);
2155 GOTO(err_cleanup, rc = -ENOENT);
2157 rc = mds_lov_connect(obd, lprof->lp_osc);
2159 GOTO(err_cleanup, rc);
2161 rc = mds_lmv_postsetup(obd);
2163 GOTO(err_cleanup, rc);
2171 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
2172 LLOG_CONFIG_ORIG_CTXT));
2176 int mds_postrecov(struct obd_device *obd)
2178 struct mds_obd *mds = &obd->u.mds;
2179 struct llog_ctxt *ctxt;
2183 LASSERT(!obd->obd_recovering);
2184 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2185 LASSERT(ctxt != NULL);
2187 /* set nextid first, so we are sure it happens */
2188 rc = mds_lov_set_nextid(obd);
2190 CERROR("%s: mds_lov_set_nextid failed\n", obd->obd_name);
2194 /* clean PENDING dir */
2195 rc = mds_cleanup_orphans(obd);
2200 rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
2203 CERROR("%s: failed at llog_origin_connect: %d\n",
2208 /* remove the orphaned precreated objects */
2209 rc = mds_lov_clearorphans(mds, NULL /* all OSTs */);
2214 RETURN(rc < 0 ? rc : item);
2217 /* cleanup all llogging subsystems */
2218 rc = obd_llog_finish(obd, &obd->obd_llogs,
2219 mds->mds_lov_desc.ld_tgt_count);
2221 CERROR("%s: failed to cleanup llogging subsystems\n",
2226 int mds_lov_clean(struct obd_device *obd)
2228 struct mds_obd *mds = &obd->u.mds;
2230 if (mds->mds_profile) {
2232 struct config_llog_instance cfg;
2233 struct lvfs_run_ctxt saved;
2234 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
2236 OBD_ALLOC(cln_prof, len);
2237 sprintf(cln_prof, "%s-clean", mds->mds_profile);
2239 cfg.cfg_instance = NULL;
2240 cfg.cfg_uuid = mds->mds_lov_uuid;
2242 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2243 class_config_process_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT),
2245 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2247 OBD_FREE(cln_prof, len);
2248 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2249 mds->mds_profile = NULL;
2254 int mds_lmv_clean(struct obd_device *obd)
2256 struct mds_obd *mds = &obd->u.mds;
2258 if (mds->mds_lmv_name) {
2259 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
2260 mds->mds_lmv_name = NULL;
2265 static int mds_precleanup(struct obd_device *obd, int flags)
2271 mds_lov_disconnect(obd, flags);
2273 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2277 static int mds_cleanup(struct obd_device *obd, int flags)
2279 struct mds_obd *mds = &obd->u.mds;
2282 if (mds->mds_sb == NULL)
2285 mds_update_server_data(obd, 1);
2286 if (mds->mds_lov_objids != NULL) {
2287 OBD_FREE(mds->mds_lov_objids,
2288 mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
2290 mds_fs_cleanup(obd, flags);
2294 /* 2 seems normal on mds, (may_umount() also expects 2
2295 fwiw), but we only see 1 at this point in obdfilter. */
2296 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
2297 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
2298 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
2300 mntput(mds->mds_vfsmnt);
2304 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
2306 spin_lock_bh(&obd->obd_processing_task_lock);
2307 if (obd->obd_recovering) {
2308 target_cancel_recovery_timer(obd);
2309 obd->obd_recovering = 0;
2311 spin_unlock_bh(&obd->obd_processing_task_lock);
2314 dev_clear_rdonly(2);
2315 fsfilt_put_ops(obd->obd_fsops);
2320 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
2321 struct ldlm_lock *new_lock,
2322 struct lustre_handle *lockh)
2324 struct obd_export *exp = req->rq_export;
2325 struct obd_device *obd = exp->exp_obd;
2326 struct ldlm_request *dlmreq =
2327 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
2328 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
2329 struct list_head *iter;
2331 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2334 l_lock(&obd->obd_namespace->ns_lock);
2335 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2336 struct ldlm_lock *lock;
2337 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2338 if (lock == new_lock)
2340 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2341 lockh->cookie = lock->l_handle.h_cookie;
2342 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2344 l_unlock(&obd->obd_namespace->ns_lock);
2348 l_unlock(&obd->obd_namespace->ns_lock);
2350 /* If the xid matches, then we know this is a resent request,
2351 * and allow it. (It's probably an OPEN, for which we don't
2353 if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
2356 /* This remote handle isn't enqueued, so we never received or
2357 * processed this request. Clear MSG_RESENT, because it can
2358 * be handled like any normal request now. */
2360 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2362 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2366 int intent_disposition(struct ldlm_reply *rep, int flag)
2370 return (rep->lock_policy_res1 & flag);
2373 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2377 rep->lock_policy_res1 |= flag;
2380 static int mds_intent_policy(struct ldlm_namespace *ns,
2381 struct ldlm_lock **lockp, void *req_cookie,
2382 ldlm_mode_t mode, int flags, void *data)
2384 struct ptlrpc_request *req = req_cookie;
2385 struct ldlm_lock *lock = *lockp;
2386 struct ldlm_intent *it;
2387 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2388 struct ldlm_reply *rep;
2389 struct lustre_handle lockh = { 0 };
2390 struct ldlm_lock *new_lock;
2391 int getattr_part = MDS_INODELOCK_UPDATE;
2392 int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
2393 sizeof(struct mds_body),
2394 mds->mds_max_mdsize,
2395 mds->mds_max_cookiesize};
2398 LASSERT(req != NULL);
2400 if (req->rq_reqmsg->bufcount <= 1) {
2401 /* No intent was provided */
2402 int size = sizeof(struct ldlm_reply);
2403 rc = lustre_pack_reply(req, 1, &size, NULL);
2408 it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
2410 CERROR("Intent missing\n");
2411 RETURN(req->rq_status = -EFAULT);
2414 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2416 rc = lustre_pack_reply(req, 3, repsize, NULL);
2418 RETURN(req->rq_status = rc);
2420 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
2421 intent_set_disposition(rep, DISP_IT_EXECD);
2423 fixup_handle_for_resent_req(req, lock, &lockh);
2425 /* execute policy */
2426 switch ((long)it->opc) {
2428 case IT_CREAT|IT_OPEN:
2429 /* XXX swab here to assert that an mds_open reint
2430 * packet is following */
2431 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
2433 /* We abort the lock if the lookup was negative and
2434 * we did not make it to the OPEN portion */
2435 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2436 RETURN(ELDLM_LOCK_ABORTED);
2437 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2438 !intent_disposition(rep, DISP_OPEN_OPEN))
2440 /* IT_OPEN may return lock on cross-node dentry
2441 * that we want to hold during attr retrival -bzzz */
2442 if (rc != 0 || lockh.cookie == 0)
2443 RETURN(ELDLM_LOCK_ABORTED);
2446 getattr_part = MDS_INODELOCK_LOOKUP;
2449 getattr_part |= MDS_INODELOCK_LOOKUP;
2451 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh,
2453 /* FIXME: LDLM can set req->rq_status. MDS sets
2454 policy_res{1,2} with disposition and status.
2455 - replay: returns 0 & req->status is old status
2456 - otherwise: returns req->status */
2457 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2458 rep->lock_policy_res2 = 0;
2459 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2460 rep->lock_policy_res2)
2461 RETURN(ELDLM_LOCK_ABORTED);
2462 if (req->rq_status != 0) {
2464 rep->lock_policy_res2 = req->rq_status;
2465 RETURN(ELDLM_LOCK_ABORTED);
2469 rc = mds_lock_and_check_slave(offset, req, &lockh);
2470 if ((rep->lock_policy_res2 = rc)) {
2472 rep->lock_policy_res2 = 0;
2473 RETURN(ELDLM_LOCK_ABORTED);
2477 CERROR("Unhandled intent "LPD64"\n", it->opc);
2481 /* By this point, whatever function we called above must have either
2482 * filled in 'lockh', been an intent replay, or returned an error. We
2483 * want to allow replayed RPCs to not get a lock, since we would just
2484 * drop it below anyways because lock replay is done separately by the
2485 * client afterwards. For regular RPCs we want to give the new lock to
2486 * the client instead of whatever lock it was about to get. */
2487 new_lock = ldlm_handle2lock(&lockh);
2488 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2491 LASSERT(new_lock != NULL);
2493 /* If we've already given this lock to a client once, then we should
2494 * have no readers or writers. Otherwise, we should have one reader
2495 * _or_ writer ref (which will be zeroed below) before returning the
2496 * lock to a client. */
2497 if (new_lock->l_export == req->rq_export) {
2498 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2500 LASSERT(new_lock->l_export == NULL);
2501 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2506 if (new_lock->l_export == req->rq_export) {
2507 /* Already gave this to the client, which means that we
2508 * reconstructed a reply. */
2509 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2511 RETURN(ELDLM_LOCK_REPLACED);
2514 /* Fixup the lock to be given to the client */
2515 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
2516 new_lock->l_readers = 0;
2517 new_lock->l_writers = 0;
2519 new_lock->l_export = class_export_get(req->rq_export);
2520 list_add(&new_lock->l_export_chain,
2521 &new_lock->l_export->exp_ldlm_data.led_held_locks);
2523 new_lock->l_blocking_ast = lock->l_blocking_ast;
2524 new_lock->l_completion_ast = lock->l_completion_ast;
2526 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2527 sizeof(lock->l_remote_handle));
2529 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2531 LDLM_LOCK_PUT(new_lock);
2532 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
2534 RETURN(ELDLM_LOCK_REPLACED);
2537 int mds_attach(struct obd_device *dev, obd_count len, void *data)
2539 struct lprocfs_static_vars lvars;
2541 lprocfs_init_multi_vars(0, &lvars);
2542 return lprocfs_obd_attach(dev, lvars.obd_vars);
2545 int mds_detach(struct obd_device *dev)
2547 return lprocfs_obd_detach(dev);
2550 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
2552 struct lprocfs_static_vars lvars;
2554 lprocfs_init_multi_vars(1, &lvars);
2555 return lprocfs_obd_attach(dev, lvars.obd_vars);
2558 int mdt_detach(struct obd_device *dev)
2560 return lprocfs_obd_detach(dev);
2563 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
2565 struct mds_obd *mds = &obd->u.mds;
2570 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2571 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
2572 mds_handle, "mds", obd->obd_proc_entry);
2574 if (!mds->mds_service) {
2575 CERROR("failed to start service\n");
2579 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
2582 GOTO(err_thread, rc);
2584 mds->mds_setattr_service =
2585 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2586 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
2587 mds_handle, "mds_setattr",
2588 obd->obd_proc_entry);
2589 if (!mds->mds_setattr_service) {
2590 CERROR("failed to start getattr service\n");
2591 GOTO(err_thread, rc = -ENOMEM);
2594 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
2595 MDT_NUM_THREADS, "ll_mdt_attr");
2597 GOTO(err_thread2, rc);
2599 mds->mds_readpage_service =
2600 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2601 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
2602 mds_handle, "mds_readpage",
2603 obd->obd_proc_entry);
2604 if (!mds->mds_readpage_service) {
2605 CERROR("failed to start readpage service\n");
2606 GOTO(err_thread2, rc = -ENOMEM);
2609 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
2610 MDT_NUM_THREADS, "ll_mdt_rdpg");
2613 GOTO(err_thread3, rc);
2618 ptlrpc_unregister_service(mds->mds_readpage_service);
2620 ptlrpc_unregister_service(mds->mds_setattr_service);
2622 ptlrpc_unregister_service(mds->mds_service);
2626 static int mdt_cleanup(struct obd_device *obd, int flags)
2628 struct mds_obd *mds = &obd->u.mds;
2631 ptlrpc_stop_all_threads(mds->mds_readpage_service);
2632 ptlrpc_unregister_service(mds->mds_readpage_service);
2634 ptlrpc_stop_all_threads(mds->mds_setattr_service);
2635 ptlrpc_unregister_service(mds->mds_setattr_service);
2637 ptlrpc_stop_all_threads(mds->mds_service);
2638 ptlrpc_unregister_service(mds->mds_service);
2643 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2646 struct obd_device *obd = data;
2649 fid.generation = gen;
2650 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2653 static int mds_get_info(struct obd_export *exp, __u32 keylen,
2654 void *key, __u32 *vallen, void *val)
2656 struct obd_device *obd;
2657 struct mds_obd *mds;
2660 obd = class_exp2obd(exp);
2662 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2663 exp->exp_handle.h_cookie);
2667 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
2668 /*Get log_context handle*/
2669 unsigned long *llh_handle = val;
2670 *vallen = sizeof(unsigned long);
2671 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
2674 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
2675 /*Get log_context handle*/
2676 unsigned long *sb = val;
2677 *vallen = sizeof(unsigned long);
2678 *sb = (unsigned long)obd->u.mds.mds_sb;
2683 keylen == strlen("mdsize");
2684 if (keylen && memcmp(key, "mdsize", keylen) == 0) {
2685 __u32 *mdsize = val;
2686 *vallen = sizeof(*mdsize);
2687 *mdsize = mds->mds_max_mdsize;
2691 CDEBUG(D_IOCTL, "invalid key\n");
2695 struct lvfs_callback_ops mds_lvfs_ops = {
2696 l_fid2dentry: mds_lvfs_fid2dentry,
2699 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
2700 int objcount, struct obd_ioobj *obj,
2701 int niocount, struct niobuf_remote *nb,
2702 struct niobuf_local *res,
2703 struct obd_trans_info *oti);
2704 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
2705 int objcount, struct obd_ioobj *obj, int niocount,
2706 struct niobuf_local *res, struct obd_trans_info *oti,
2709 /* use obd ops to offer management infrastructure */
2710 static struct obd_ops mds_obd_ops = {
2711 .o_owner = THIS_MODULE,
2712 .o_attach = mds_attach,
2713 .o_detach = mds_detach,
2714 .o_connect = mds_connect,
2715 .o_init_export = mds_init_export,
2716 .o_destroy_export = mds_destroy_export,
2717 .o_disconnect = mds_disconnect,
2718 .o_setup = mds_setup,
2719 .o_precleanup = mds_precleanup,
2720 .o_cleanup = mds_cleanup,
2721 .o_postrecov = mds_postrecov,
2722 .o_statfs = mds_obd_statfs,
2723 .o_iocontrol = mds_iocontrol,
2724 .o_create = mds_obd_create,
2725 .o_destroy = mds_obd_destroy,
2726 .o_llog_init = mds_llog_init,
2727 .o_llog_finish = mds_llog_finish,
2728 .o_notify = mds_notify,
2729 .o_get_info = mds_get_info,
2730 .o_set_info = mds_set_info,
2731 .o_preprw = mds_preprw,
2732 .o_commitrw = mds_commitrw,
2735 static struct obd_ops mdt_obd_ops = {
2736 .o_owner = THIS_MODULE,
2737 .o_attach = mdt_attach,
2738 .o_detach = mdt_detach,
2739 .o_setup = mdt_setup,
2740 .o_cleanup = mdt_cleanup,
2741 .o_attach = mdt_attach,
2742 .o_detach = mdt_detach,
2745 static int __init mds_init(void)
2747 struct lprocfs_static_vars lvars;
2749 lprocfs_init_multi_vars(0, &lvars);
2750 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
2752 lprocfs_init_multi_vars(1, &lvars);
2753 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
2759 static void /*__exit*/ mds_exit(void)
2761 class_unregister_type(LUSTRE_MDS_NAME);
2762 class_unregister_type(LUSTRE_MDT_NAME);
2765 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2766 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2767 MODULE_LICENSE("GPL");
2769 module_init(mds_init);
2770 module_exit(mds_exit);