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/lprocfs_status.h>
56 #include <linux/lustre_commit_confd.h>
59 #include <linux/lustre_smfs.h>
60 #include <linux/lustre_snap.h>
62 #include "mds_internal.h"
64 static int mds_intent_policy(struct ldlm_namespace *ns,
65 struct ldlm_lock **lockp, void *req_cookie,
66 ldlm_mode_t mode, int flags, void *data);
67 static int mds_postsetup(struct obd_device *obd);
68 static int mds_cleanup(struct obd_device *obd, int flags);
71 /* Assumes caller has already pushed into the kernel filesystem context */
72 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
73 loff_t offset, int count)
75 struct ptlrpc_bulk_desc *desc;
76 struct l_wait_info lwi;
78 int rc = 0, npages, i, tmpcount, tmpsize = 0;
81 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
83 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
84 OBD_ALLOC(pages, sizeof(*pages) * npages);
86 GOTO(out, rc = -ENOMEM);
88 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
91 GOTO(out_free, rc = -ENOMEM);
93 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
94 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
96 pages[i] = alloc_pages(GFP_KERNEL, 0);
98 GOTO(cleanup_buf, rc = -ENOMEM);
100 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
103 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
104 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
105 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
106 tmpsize, offset, file->f_dentry->d_inode->i_ino,
107 file->f_dentry->d_inode->i_size);
109 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
110 kmap(pages[i]), tmpsize, &offset);
114 GOTO(cleanup_buf, rc = -EIO);
117 LASSERT(desc->bd_nob == count);
119 rc = ptlrpc_start_bulk_transfer(desc);
121 GOTO(cleanup_buf, rc);
123 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
124 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
125 OBD_FAIL_MDS_SENDPAGE, rc);
126 GOTO(abort_bulk, rc);
129 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
130 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
131 LASSERT (rc == 0 || rc == -ETIMEDOUT);
134 if (desc->bd_success &&
135 desc->bd_nob_transferred == count)
136 GOTO(cleanup_buf, rc);
138 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
141 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
142 (rc == -ETIMEDOUT) ? "timeout" : "network error",
143 desc->bd_nob_transferred, count,
144 req->rq_export->exp_client_uuid.uuid,
145 req->rq_export->exp_connection->c_remote_uuid.uuid);
147 ptlrpc_fail_export(req->rq_export);
151 ptlrpc_abort_bulk (desc);
153 for (i = 0; i < npages; i++)
155 __free_pages(pages[i], 0);
157 ptlrpc_free_bulk(desc);
159 OBD_FREE(pages, sizeof(*pages) * npages);
164 int mds_lock_mode_for_dir(struct obd_device *obd,
165 struct dentry *dentry, int mode)
169 /* any dir access needs couple locks:
170 * 1) on part of dir we gonna lookup/modify in
171 * 2) on a whole dir to protect it from concurrent splitting
172 * and to flush client's cache for readdir()
173 * so, for a given mode and dentry this routine decides what
174 * lock mode to use for lock #2:
175 * 1) if caller's gonna lookup in dir then we need to protect
176 * dir from being splitted only - LCK_CR
177 * 2) if caller's gonna modify dir then we need to protect
178 * dir from being splitted and to flush cache - LCK_CW
179 * 3) if caller's gonna modify dir and that dir seems ready
180 * for splitting then we need to protect it from any
181 * type of access (lookup/modify/split) - LCK_EX -bzzz */
183 split = mds_splitting_expected(obd, dentry);
185 /* it is important to check here only for MDS_NO_SPLITTABLE.
186 * The reason is that MDS_NO_SPLITTABLE means dir is not splittable
187 * in principle and another thread will not split it on the quiet.
188 * But if we have MDS_NO_SPLIT_EXPECTED, this means, that dir may be
189 * splitted anytime, but not now (forcurrent thread) and we should
190 * consider that it can happen soon and go that branch which can yield
191 * LCK_EX to protect from possible splitting. */
192 if (split == MDS_NO_SPLITTABLE) {
193 /* this inode won't be splitted. so we need not to protect from
194 * just flush client's cache on modification */
202 if (mode == LCK_PR) {
204 } else if (mode == LCK_PW) {
205 /* caller gonna modify directory.we use concurrent
206 write lock here to retract client's cache for readdir */
208 if (split == MDS_EXPECT_SPLIT) {
209 /* splitting possible. serialize any access
210 * the idea is that first one seen dir is
211 * splittable is given exclusive lock and
212 * split directory. caller passes lock mode
213 * to mds_try_to_split_dir() and splitting
214 * would be done with exclusive lock only -bzzz */
215 CDEBUG(D_OTHER, "%s: gonna split %u/%u\n",
217 (unsigned) dentry->d_inode->i_ino,
218 (unsigned) dentry->d_inode->i_generation);
226 /* only valid locked dentries or errors should be returned */
227 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
228 struct vfsmount **mnt, int lock_mode,
229 struct lustre_handle *lockh, int *mode,
230 char *name, int namelen, __u64 lockpart)
232 struct mds_obd *mds = &obd->u.mds;
233 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
234 struct ldlm_res_id res_id = { .name = {0} };
236 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
243 res_id.name[0] = de->d_inode->i_ino;
244 res_id.name[1] = de->d_inode->i_generation;
247 if (name && IS_PDIROPS(de->d_inode)) {
248 ldlm_policy_data_t cpolicy =
249 { .l_inodebits = { MDS_INODELOCK_UPDATE } };
250 LASSERT(mode != NULL);
251 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
253 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
255 &cpolicy, *mode, &flags,
257 ldlm_completion_ast, NULL, NULL,
258 NULL, 0, NULL, lockh + 1);
259 if (rc != ELDLM_OK) {
261 RETURN(ERR_PTR(-ENOLCK));
266 res_id.name[2] = full_name_hash(name, namelen);
268 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
269 de->d_inode->i_ino, de->d_inode->i_generation,
273 #warning "No PDIROPS support in the kernel"
275 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
276 LDLM_IBITS, &policy, lock_mode, &flags,
277 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
278 NULL, 0, NULL, lockh);
279 if (rc != ELDLM_OK) {
281 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
284 ldlm_lock_decref(lockh + 1, LCK_CW);
291 #ifndef DCACHE_DISCONNECTED
292 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
296 /* Look up an entry by inode number. */
297 /* this function ONLY returns valid dget'd dentries with an initialized inode
299 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
300 struct vfsmount **mnt)
303 unsigned long ino = fid->id;
304 __u32 generation = fid->generation;
306 struct dentry *result;
309 RETURN(ERR_PTR(-ESTALE));
311 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
313 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
314 ino, generation, mds->mds_sb);
316 /* under ext3 this is neither supposed to return bad inodes
318 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
322 inode = result->d_inode;
324 RETURN(ERR_PTR(-ENOENT));
326 /* here we disabled generation check, as root inode i_generation
327 * of cache mds and real mds are different. */
328 if (inode->i_ino != mds->mds_rootfid.id && generation &&
329 inode->i_generation != generation) {
330 /* we didn't find the right inode.. */
331 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
332 inode->i_ino, (unsigned long)inode->i_nlink,
333 atomic_read(&inode->i_count), inode->i_generation,
336 RETURN(ERR_PTR(-ENOENT));
340 *mnt = mds->mds_vfsmnt;
348 /* Establish a connection to the MDS.
350 * This will set up an export structure for the client to hold state data
351 * about that client, like open files, the last operation number it did
352 * on the server, etc.
354 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
355 struct obd_uuid *cluuid, unsigned long connect_flags)
357 struct obd_export *exp;
358 struct mds_export_data *med; /* */
359 struct mds_client_data *mcd;
363 if (!conn || !obd || !cluuid)
366 /* XXX There is a small race between checking the list and adding a
367 * new connection for the same UUID, but the real threat (list
368 * corruption when multiple different clients connect) is solved.
370 * There is a second race between adding the export to the list,
371 * and filling in the client data below. Hence skipping the case
372 * of NULL mcd above. We should already be controlling multiple
373 * connects at the client, and we can't hold the spinlock over
374 * memory allocations without risk of deadlocking.
376 rc = class_connect(conn, obd, cluuid);
379 exp = class_conn2export(conn);
381 med = &exp->exp_mds_data;
383 OBD_ALLOC(mcd, sizeof(*mcd));
385 CERROR("mds: out of memory for client data\n");
386 GOTO(out, rc = -ENOMEM);
389 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
392 rc = mds_client_add(obd, &obd->u.mds, med, -1);
396 if (!(connect_flags & OBD_OPT_MDS_CONNECTION)) {
397 struct mds_obd *mds = &obd->u.mds;
398 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
399 atomic_inc(&mds->mds_real_clients);
400 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
401 obd->obd_name, exp->exp_client_uuid.uuid,
402 atomic_read(&mds->mds_real_clients));
403 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
405 if (mds->mds_lmv_name)
406 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
411 OBD_FREE(mcd, sizeof(*mcd));
412 class_disconnect(exp, 0);
414 class_export_put(exp);
419 static int mds_init_export(struct obd_export *exp)
421 struct mds_export_data *med = &exp->exp_mds_data;
423 INIT_LIST_HEAD(&med->med_open_head);
424 spin_lock_init(&med->med_open_lock);
428 static int mds_destroy_export(struct obd_export *export)
430 struct mds_export_data *med;
431 struct obd_device *obd = export->exp_obd;
432 struct lvfs_run_ctxt saved;
436 med = &export->exp_mds_data;
437 target_destroy_export(export);
439 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
442 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
444 /* Close any open files (which may also cause orphan unlinking). */
445 spin_lock(&med->med_open_lock);
446 while (!list_empty(&med->med_open_head)) {
447 struct list_head *tmp = med->med_open_head.next;
448 struct mds_file_data *mfd =
449 list_entry(tmp, struct mds_file_data, mfd_list);
450 BDEVNAME_DECLARE_STORAGE(btmp);
452 /* bug 1579: fix force-closing for 2.5 */
453 struct dentry *dentry = mfd->mfd_dentry;
455 list_del(&mfd->mfd_list);
456 spin_unlock(&med->med_open_lock);
458 /* If you change this message, be sure to update
459 * replay_single:test_46 */
460 CERROR("force closing client file handle for %*s (%s:%lu)\n",
461 dentry->d_name.len, dentry->d_name.name,
462 ll_bdevname(dentry->d_inode->i_sb, btmp),
463 dentry->d_inode->i_ino);
464 rc = mds_mfd_close(NULL, obd, mfd,
465 !(export->exp_flags & OBD_OPT_FAILOVER));
468 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
469 spin_lock(&med->med_open_lock);
471 spin_unlock(&med->med_open_lock);
472 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
475 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
480 static int mds_disconnect(struct obd_export *exp, int flags)
482 struct obd_device *obd;
484 unsigned long irqflags;
489 class_export_get(exp);
491 obd = class_exp2obd(exp);
493 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
494 exp->exp_handle.h_cookie);
499 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
500 && !atomic_read(&mds->mds_real_clients)) {
501 /* there was no client at all */
502 mds_lmv_disconnect(obd, flags);
505 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
506 && atomic_dec_and_test(&mds->mds_real_clients)) {
507 /* time to drop LMV connections */
508 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
509 "Disconnnect from LMV now\n",
510 obd->obd_name, exp->exp_client_uuid.uuid);
511 mds_lmv_disconnect(obd, flags);
514 spin_lock_irqsave(&exp->exp_lock, irqflags);
515 exp->exp_flags = flags;
516 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
518 /* Disconnect early so that clients can't keep using export */
519 rc = class_disconnect(exp, flags);
520 ldlm_cancel_locks_for_export(exp);
522 /* complete all outstanding replies */
523 spin_lock_irqsave(&exp->exp_lock, irqflags);
524 while (!list_empty(&exp->exp_outstanding_replies)) {
525 struct ptlrpc_reply_state *rs =
526 list_entry(exp->exp_outstanding_replies.next,
527 struct ptlrpc_reply_state, rs_exp_list);
528 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
530 spin_lock(&svc->srv_lock);
531 list_del_init(&rs->rs_exp_list);
532 ptlrpc_schedule_difficult_reply(rs);
533 spin_unlock(&svc->srv_lock);
535 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
537 class_export_put(exp);
541 static int mds_getstatus(struct ptlrpc_request *req)
543 struct mds_obd *mds = mds_req2mds(req);
544 struct mds_body *body;
545 int rc, size = sizeof(*body);
548 rc = lustre_pack_reply(req, 1, &size, NULL);
549 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
550 CERROR("mds: out of memory for message: size=%d\n", size);
551 req->rq_status = -ENOMEM; /* superfluous? */
555 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
556 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
558 /* the last_committed and last_xid fields are filled in for all
559 * replies already - no need to do so here also.
564 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
565 void *data, int flag)
570 if (flag == LDLM_CB_CANCELING) {
571 /* Don't need to do anything here. */
575 /* XXX layering violation! -phil */
576 l_lock(&lock->l_resource->lr_namespace->ns_lock);
577 /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
578 * such that mds_blocking_ast is called just before l_i_p takes the
579 * ns_lock, then by the time we get the lock, we might not be the
580 * correct blocking function anymore. So check, and return early, if
582 if (lock->l_blocking_ast != mds_blocking_ast) {
583 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
587 lock->l_flags |= LDLM_FL_CBPENDING;
588 do_ast = (!lock->l_readers && !lock->l_writers);
589 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
592 struct lustre_handle lockh;
595 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
596 ldlm_lock2handle(lock, &lockh);
597 rc = ldlm_cli_cancel(&lockh);
599 CERROR("ldlm_cli_cancel: %d\n", rc);
601 LDLM_DEBUG(lock, "Lock still has references, will be "
607 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
615 rc = fsfilt_get_md(obd, inode, md, *size);
620 CERROR("Error %d reading eadata for ino %lu\n",
625 if (S_ISREG(inode->i_mode))
626 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
627 if (S_ISDIR(inode->i_mode))
628 rc = mds_convert_mea_ea(obd, inode, md, lmm_size);
642 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
643 * Call with lock=0 if the caller has already taken the i_sem. */
644 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
645 struct mds_body *body, struct inode *inode, int lock)
647 struct mds_obd *mds = &obd->u.mds;
653 lmm = lustre_msg_buf(msg, offset, 0);
655 /* Some problem with getting eadata when I sized the reply
657 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
661 lmm_size = msg->buflens[offset];
663 /* I don't really like this, but it is a sanity check on the client
664 * MD request. However, if the client doesn't know how much space
665 * to reserve for the MD, it shouldn't be bad to have too much space.
667 if (lmm_size > mds->mds_max_mdsize) {
668 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
669 inode->i_ino, lmm_size, mds->mds_max_mdsize);
673 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
675 if (S_ISDIR(inode->i_mode))
676 body->valid |= OBD_MD_FLDIREA;
678 body->valid |= OBD_MD_FLEASIZE;
679 body->eadatasize = lmm_size;
686 static void mds_root_squash_rec(struct mds_obd *mds, ptl_nid_t *peernid,
687 struct mds_update_record *rec)
689 if (!mds->mds_squash_uid || rec->_ur_fsuid)
692 if (*peernid == mds->mds_nosquash_nid)
695 CDEBUG(D_OTHER, "squash req from 0x%llx, (%d:%d/%x)=>(%d:%d/%x)\n",
697 rec->_ur_fsuid, rec->_ur_fsgid, rec->_ur_cap,
698 mds->mds_squash_uid, mds->mds_squash_gid,
699 (rec->_ur_cap & ~CAP_FS_MASK));
701 rec->_ur_fsuid = mds->mds_squash_uid;
702 rec->_ur_fsgid = mds->mds_squash_gid;
703 /* XXX should we remove all capabilities? */
704 rec->_ur_cap &= ~CAP_FS_MASK;
705 rec->_ur_suppgid1 = -1;
706 rec->_ur_suppgid2 = -1;
708 static void mds_root_squash_body(struct mds_obd *mds, ptl_nid_t *peernid,
709 struct mds_body *body)
711 if (!mds->mds_squash_uid || body->fsuid)
714 if (*peernid == mds->mds_nosquash_nid)
717 CDEBUG(D_OTHER, "squash req from 0x%llx, (%d:%d/%x)=>(%d:%d/%x)\n",
719 body->fsuid, body->fsgid, body->capability,
720 mds->mds_squash_uid, mds->mds_squash_gid,
721 (body->capability & ~CAP_FS_MASK));
723 body->fsuid = mds->mds_squash_uid;
724 body->fsgid = mds->mds_squash_gid;
725 /* XXX should we remove all capabilities? */
726 body->capability &= ~CAP_FS_MASK;
729 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
730 struct ptlrpc_request *req,
731 struct mds_body *reqbody, int reply_off)
733 struct mds_body *body;
734 struct inode *inode = dentry->d_inode;
738 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
741 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
742 LASSERT(body != NULL); /* caller prepped reply */
744 if (dentry->d_flags & DCACHE_CROSS_REF) {
745 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
746 (unsigned long) dentry->d_mdsnum,
747 (unsigned long) dentry->d_inum,
748 (unsigned long) dentry->d_generation);
749 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
750 body->fid1.id = dentry->d_inum;
751 body->fid1.mds = dentry->d_mdsnum;
752 body->fid1.generation = dentry->d_generation;
755 mds_pack_inode2fid(obd, &body->fid1, inode);
756 mds_pack_inode2body(obd, body, inode);
758 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
759 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
760 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
763 /* If we have LOV EA data, the OST holds size, atime, mtime */
764 if (!(body->valid & OBD_MD_FLEASIZE) &&
765 !(body->valid & OBD_MD_FLDIREA))
766 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
767 OBD_MD_FLATIME | OBD_MD_FLMTIME);
768 } else if (S_ISLNK(inode->i_mode) &&
769 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
770 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
773 LASSERT (symname != NULL); /* caller prepped reply */
774 len = req->rq_repmsg->buflens[reply_off + 1];
776 rc = inode->i_op->readlink(dentry, symname, len);
778 CERROR("readlink failed: %d\n", rc);
779 } else if (rc != len - 1) {
780 CERROR ("Unexpected readlink rc %d: expecting %d\n",
784 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
785 body->valid |= OBD_MD_LINKNAME;
786 body->eadatasize = rc + 1;
787 symname[rc] = 0; /* NULL terminate */
795 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
796 struct dentry *dentry,
799 int rc = 0, size[1] = {sizeof(struct mds_body)};
802 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
803 CERROR("failed MDS_GETATTR_PACK test\n");
804 req->rq_status = -ENOMEM;
805 GOTO(out, rc = -ENOMEM);
808 rc = lustre_pack_reply(req, 1, size, NULL);
810 CERROR("out of memory\n");
811 GOTO(out, req->rq_status = rc);
819 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
822 struct mds_obd *mds = mds_req2mds(req);
823 struct mds_body *body;
824 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
827 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
828 LASSERT(body != NULL); /* checked by caller */
829 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
831 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
832 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
835 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
837 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
841 CERROR("error getting inode %lu MD: rc = %d\n",
844 } else if (rc > mds->mds_max_mdsize) {
846 CERROR("MD size %d larger than maximum possible %u\n",
847 rc, mds->mds_max_mdsize);
852 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
853 if (inode->i_size + 1 != body->eadatasize)
854 CERROR("symlink size: %Lu, reply space: %d\n",
855 inode->i_size + 1, body->eadatasize);
856 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
858 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
859 inode->i_size + 1, body->eadatasize);
862 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
863 CERROR("failed MDS_GETATTR_PACK test\n");
864 req->rq_status = -ENOMEM;
865 GOTO(out, rc = -ENOMEM);
868 rc = lustre_pack_reply(req, bufcount, size, NULL);
870 CERROR("out of memory\n");
871 GOTO(out, req->rq_status = rc);
879 int mds_check_mds_num(struct obd_device *obd, struct inode* inode,
880 char *name, int namelen)
882 struct mea *mea = NULL;
883 int mea_size, rc = 0;
886 rc = mds_get_lmv_attr(obd, inode, &mea, &mea_size);
890 /* dir is already splitted, check if requested filename
891 * should live at this MDS or at another one */
893 i = mea_name2idx(mea, name, namelen - 1);
894 if (mea->mea_master != mea->mea_fids[i].mds) {
896 "inapropriate MDS(%d) for %s. should be %d(%d)\n",
897 mea->mea_master, name, mea->mea_fids[i].mds, i);
903 OBD_FREE(mea, mea_size);
907 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
908 struct lustre_handle *child_lockh, int child_part)
910 struct obd_device *obd = req->rq_export->exp_obd;
911 struct ldlm_reply *rep = NULL;
912 struct lvfs_run_ctxt saved;
913 struct mds_body *body;
914 struct dentry *dparent = NULL, *dchild = NULL;
915 struct lvfs_ucred uc;
916 struct lustre_handle parent_lockh[2];
917 int namesize, update_mode;
918 int rc = 0, cleanup_phase = 0, resent_req = 0;
919 struct clonefs_info *clone_info = NULL;
923 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
925 MDS_UPDATE_COUNTER((&obd->u.mds), MDS_GETATTR_NAME_COUNT);
927 /* Swab now, before anyone looks inside the request */
928 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
929 lustre_swab_mds_body);
931 CERROR("Can't swab mds_body\n");
932 GOTO(cleanup, rc = -EFAULT);
935 LASSERT_REQSWAB(req, offset + 1);
936 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
938 CERROR("Can't unpack name\n");
939 GOTO(cleanup, rc = -EFAULT);
941 namesize = req->rq_reqmsg->buflens[offset + 1];
943 LASSERT (offset == 0 || offset == 2);
944 /* if requests were at offset 2, the getattr reply goes back at 1 */
946 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
949 mds_root_squash_body(&obd->u.mds, &req->rq_peer.peer_nid, body);
950 uc.luc_fsuid = body->fsuid;
951 uc.luc_fsgid = body->fsgid;
952 uc.luc_cap = body->capability;
953 uc.luc_suppgid1 = body->suppgid;
954 uc.luc_suppgid2 = -1;
955 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
956 cleanup_phase = 1; /* kernel context */
957 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
959 LASSERT(namesize > 0);
961 /* we have no dentry here, drop LOOKUP bit */
962 child_part &= ~MDS_INODELOCK_LOOKUP;
963 CDEBUG(D_OTHER, "%s: request to retrieve attrs for %lu/%lu\n",
964 obd->obd_name, (unsigned long) body->fid1.id,
965 (unsigned long) body->fid1.generation);
966 dchild = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR,
967 parent_lockh, &update_mode,
968 NULL, 0, child_part);
969 if (IS_ERR(dchild)) {
970 CERROR("can't find inode: %d\n", (int) PTR_ERR(dchild));
971 GOTO(cleanup, rc = PTR_ERR(dchild));
973 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
975 if (parent_lockh[1].cookie)
976 ldlm_lock_decref(parent_lockh + 1, update_mode);
983 /* FIXME: handle raw lookup */
984 if (body->valid == OBD_MD_FLID) {
985 struct mds_obd *mds = &obd->u.mds;
986 struct mds_body *mds_reply;
987 int size = sizeof(*mds_reply);
990 dparent = mds_fid2dentry(mds, &body->fid1, NULL);
991 if (IS_ERR(dparent)) {
992 rc = PTR_ERR(dparent);
995 // The user requested ONLY the inode number, so do a raw lookup
996 rc = lustre_pack_reply(req, 1, &size, NULL);
998 CERROR("out of memory\n");
1002 dir = dparent->d_inode;
1003 LASSERT(dir->i_op->lookup_raw != NULL);
1004 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1006 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
1007 sizeof(*mds_reply));
1008 mds_reply->fid1.id = inum;
1009 mds_reply->valid = OBD_MD_FLID;
1014 if (child_lockh->cookie != 0) {
1015 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1019 if (resent_req == 0) {
1020 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
1021 parent_lockh, &dparent,
1022 LCK_PR, MDS_INODELOCK_LOOKUP,
1023 &update_mode, name, namesize,
1024 child_lockh, &dchild, LCK_PR,
1025 child_part, clone_info);
1029 /* let's make sure this name should leave on this mds node */
1030 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1034 struct ldlm_lock *granted_lock;
1035 struct ll_fid child_fid;
1036 struct ldlm_resource *res;
1037 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1038 granted_lock = ldlm_handle2lock(child_lockh);
1039 LASSERT(granted_lock);
1041 res = granted_lock->l_resource;
1042 child_fid.id = res->lr_name.name[0];
1043 child_fid.generation = res->lr_name.name[1];
1044 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
1046 LDLM_LOCK_PUT(granted_lock);
1049 cleanup_phase = 2; /* dchild, dparent, locks */
1053 if (!DENTRY_VALID(dchild)) {
1054 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1055 /* in the intent case, the policy clears this error:
1056 the disposition is enough */
1060 intent_set_disposition(rep, DISP_LOOKUP_POS);
1063 if (req->rq_repmsg == NULL) {
1064 if (dchild->d_flags & DCACHE_CROSS_REF)
1065 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1067 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
1069 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1074 rc = mds_getattr_internal(obd, dchild, req, body, offset);
1075 GOTO(cleanup, rc); /* returns the lock to the client */
1078 switch (cleanup_phase) {
1080 if (resent_req == 0) {
1081 if (rc && DENTRY_VALID(dchild))
1082 ldlm_lock_decref(child_lockh, LCK_PR);
1084 ldlm_lock_decref(parent_lockh, LCK_PR);
1086 if (parent_lockh[1].cookie != 0)
1087 ldlm_lock_decref(parent_lockh + 1,
1096 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1102 static int mds_getattr(int offset, struct ptlrpc_request *req)
1104 struct mds_obd *mds = mds_req2mds(req);
1105 struct obd_device *obd = req->rq_export->exp_obd;
1106 struct lvfs_run_ctxt saved;
1108 struct mds_body *body;
1109 struct lvfs_ucred uc;
1113 body = lustre_swab_reqbuf (req, offset, sizeof (*body),
1114 lustre_swab_mds_body);
1116 CERROR ("Can't unpack body\n");
1120 MDS_UPDATE_COUNTER(mds, MDS_GETATTR_COUNT);
1122 uc.luc_fsuid = body->fsuid;
1123 uc.luc_fsgid = body->fsgid;
1124 uc.luc_cap = body->capability;
1125 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1126 de = mds_fid2dentry(mds, &body->fid1, NULL);
1128 rc = req->rq_status = -ENOENT;
1129 GOTO(out_pop, PTR_ERR(de));
1132 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1134 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1138 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
1143 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1148 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1149 unsigned long max_age)
1153 spin_lock(&obd->obd_osfs_lock);
1154 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1156 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1157 spin_unlock(&obd->obd_osfs_lock);
1162 static int mds_statfs(struct ptlrpc_request *req)
1164 struct obd_device *obd = req->rq_export->exp_obd;
1165 int rc, size = sizeof(struct obd_statfs);
1168 rc = lustre_pack_reply(req, 1, &size, NULL);
1169 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1170 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1174 MDS_UPDATE_COUNTER((&obd->u.mds), MDS_STATFS_COUNT);
1176 /* We call this so that we can cache a bit - 1 jiffie worth */
1177 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1180 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1186 req->rq_status = rc;
1190 static int mds_sync(struct ptlrpc_request *req)
1192 struct obd_device *obd = req->rq_export->exp_obd;
1193 struct mds_obd *mds = &obd->u.mds;
1194 struct mds_body *body;
1195 int rc, size = sizeof(*body);
1198 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
1200 GOTO(out, rc = -EPROTO);
1202 rc = lustre_pack_reply(req, 1, &size, NULL);
1203 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1204 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1208 if (body->fid1.id == 0) {
1209 /* a fid of zero is taken to mean "sync whole filesystem" */
1210 rc = fsfilt_sync(obd, mds->mds_sb);
1214 /* just any file to grab fsync method - "file" arg unused */
1215 struct file *file = mds->mds_rcvd_filp;
1218 de = mds_fid2dentry(mds, &body->fid1, NULL);
1220 GOTO(out, rc = PTR_ERR(de));
1222 rc = file->f_op->fsync(NULL, de, 1);
1227 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1228 mds_pack_inode2fid(obd, &body->fid1, de->d_inode);
1229 mds_pack_inode2body(obd, body, de->d_inode);
1232 req->rq_status = rc;
1236 /* mds_readpage does not take a DLM lock on the inode, because the client must
1237 * already have a PR lock.
1239 * If we were to take another one here, a deadlock will result, if another
1240 * thread is already waiting for a PW lock. */
1241 static int mds_readpage(struct ptlrpc_request *req)
1243 struct obd_device *obd = req->rq_export->exp_obd;
1244 struct vfsmount *mnt;
1247 struct mds_body *body, *repbody;
1248 struct lvfs_run_ctxt saved;
1249 int rc, size = sizeof(*repbody);
1250 struct lvfs_ucred uc;
1253 rc = lustre_pack_reply(req, 1, &size, NULL);
1254 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1255 CERROR("mds: out of memory\n");
1256 GOTO(out, rc = -ENOMEM);
1259 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1261 GOTO (out, rc = -EFAULT);
1263 mds_root_squash_body(&obd->u.mds, &req->rq_peer.peer_nid, body);
1264 uc.luc_fsuid = body->fsuid;
1265 uc.luc_fsgid = body->fsgid;
1266 uc.luc_cap = body->capability;
1267 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1268 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1270 GOTO(out_pop, rc = PTR_ERR(de));
1272 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1274 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1275 /* note: in case of an error, dentry_open puts dentry */
1277 GOTO(out_pop, rc = PTR_ERR(file));
1279 /* body->size is actually the offset -eeb */
1280 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1281 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1282 body->size, de->d_inode->i_blksize);
1283 GOTO(out_file, rc = -EFAULT);
1286 /* body->nlink is actually the #bytes to read -eeb */
1287 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1288 CERROR("size %u is not multiple of blocksize %lu\n",
1289 body->nlink, de->d_inode->i_blksize);
1290 GOTO(out_file, rc = -EFAULT);
1293 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1294 repbody->size = file->f_dentry->d_inode->i_size;
1295 repbody->valid = OBD_MD_FLSIZE;
1297 /* to make this asynchronous make sure that the handling function
1298 doesn't send a reply when this function completes. Instead a
1299 callback function would send the reply */
1300 /* body->size is actually the offset -eeb */
1301 rc = mds_sendpage(req, file, body->size, body->nlink);
1304 filp_close(file, 0);
1306 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1308 req->rq_status = rc;
1312 int mds_reint(struct ptlrpc_request *req, int offset,
1313 struct lustre_handle *lockh)
1315 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
1319 OBD_ALLOC(rec, sizeof(*rec));
1323 rc = mds_update_unpack(req, offset, rec);
1324 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1325 CERROR("invalid record\n");
1326 GOTO(out, req->rq_status = -EINVAL);
1328 mds_root_squash_rec(&req->rq_export->exp_obd->u.mds,
1329 &req->rq_peer.peer_nid, rec);
1330 /* rc will be used to interrupt a for loop over multiple records */
1331 rc = mds_reint_rec(rec, offset, req, lockh);
1333 OBD_FREE(rec, sizeof(*rec));
1337 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1338 struct obd_device *obd, int *process)
1340 switch (req->rq_reqmsg->opc) {
1341 case MDS_CONNECT: /* This will never get here, but for completeness. */
1342 case OST_CONNECT: /* This will never get here, but for completeness. */
1343 case MDS_DISCONNECT:
1344 case OST_DISCONNECT:
1349 case MDS_SYNC: /* used in unmounting */
1354 *process = target_queue_recovery_request(req, obd);
1358 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1360 /* XXX what should we set rq_status to here? */
1361 req->rq_status = -EAGAIN;
1362 RETURN(ptlrpc_error(req));
1366 static char *reint_names[] = {
1367 [REINT_SETATTR] "setattr",
1368 [REINT_CREATE] "create",
1369 [REINT_LINK] "link",
1370 [REINT_UNLINK] "unlink",
1371 [REINT_RENAME] "rename",
1372 [REINT_OPEN] "open",
1375 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER |\
1376 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ|\
1377 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME|\
1380 static void reconstruct_create(struct ptlrpc_request *req)
1382 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1383 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1384 struct mds_client_data *mcd = med->med_mcd;
1385 struct dentry *dentry;
1386 struct ost_body *body;
1390 /* copy rc, transno and disp; steal locks */
1391 mds_req_from_mcd(req, mcd);
1392 if (req->rq_status) {
1398 fid.id = mcd->mcd_last_data;
1401 LASSERT(fid.id != 0);
1403 dentry = mds_fid2dentry(mds, &fid, NULL);
1404 if (IS_ERR(dentry)) {
1405 CERROR("can't find inode "LPU64"\n", fid.id);
1406 req->rq_status = PTR_ERR(dentry);
1411 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1412 req->rq_xid, fid.id, dentry->d_inode->i_ino,
1413 dentry->d_inode->i_generation);
1415 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1416 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
1417 body->oa.o_id = dentry->d_inode->i_ino;
1418 body->oa.o_generation = dentry->d_inode->i_generation;
1419 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1426 static int mdt_obj_create(struct ptlrpc_request *req)
1428 struct obd_device *obd = req->rq_export->exp_obd;
1429 struct mds_obd *mds = &obd->u.mds;
1430 struct ost_body *body, *repbody;
1431 char fidname[LL_FID_NAMELEN];
1432 struct inode *parent_inode;
1433 struct lvfs_run_ctxt saved;
1434 struct dentry *new = NULL;
1435 struct dentry_params dp;
1436 int mealen, flags = 0, rc, size = sizeof(*repbody), cleanup_phase = 0;
1437 struct lvfs_ucred uc;
1439 void *handle = NULL;
1440 unsigned long cr_inum = 0;
1443 DEBUG_REQ(D_HA, req, "create remote object");
1445 parent_inode = mds->mds_unnamed_dir->d_inode;
1447 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
1451 rc = lustre_pack_reply(req, 1, &size, NULL);
1455 MDS_CHECK_RESENT(req, reconstruct_create(req));
1457 uc.luc_fsuid = body->oa.o_uid;
1458 uc.luc_fsgid = body->oa.o_gid;
1460 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1462 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1464 if (body->oa.o_flags & OBD_FL_RECREATE_OBJS) {
1465 /* this is re-create request from MDS holding directory name.
1466 * we have to lookup given ino/generation first. if it exists
1467 * (good case) then there is nothing to do. if it does not
1468 * then we have to recreate it */
1470 fid.id = body->oa.o_id;
1471 fid.generation = body->oa.o_generation;
1472 new = mds_fid2dentry(mds, &fid, NULL);
1473 if (!IS_ERR(new) && new->d_inode) {
1474 CDEBUG(D_HA, "mkdir() repairing is on its way: %lu/%lu\n",
1475 (unsigned long) fid.id,
1476 (unsigned long) fid.generation);
1477 obdo_from_inode(&repbody->oa, new->d_inode,
1478 FILTER_VALID_FLAGS);
1479 repbody->oa.o_id = new->d_inode->i_ino;
1480 repbody->oa.o_generation = new->d_inode->i_generation;
1481 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1483 cr_inum = new->d_inode->i_ino;
1484 GOTO(cleanup, rc = 0);
1486 CWARN("hmm. for some reason dir %lu/%lu (or reply) got lost\n",
1487 (unsigned long) fid.id, (unsigned long) fid.generation);
1488 LASSERT(new->d_inode == NULL ||
1489 new->d_inode->i_generation != fid.generation);
1493 down(&parent_inode->i_sem);
1494 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1495 if (IS_ERR(handle)) {
1496 up(&parent_inode->i_sem);
1497 GOTO(cleanup, rc = PTR_ERR(handle));
1499 cleanup_phase = 1; /* transaction */
1502 rc = sprintf(fidname, "%u.%u", ll_insecure_random_int(), current->pid);
1503 new = lookup_one_len(fidname, mds->mds_unnamed_dir, rc);
1505 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1506 obd->obd_name, fidname, (int) PTR_ERR(new));
1507 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
1508 up(&parent_inode->i_sem);
1509 RETURN(PTR_ERR(new));
1510 } else if (new->d_inode) {
1511 CERROR("%s: name exists. repeat\n", obd->obd_name);
1515 new->d_fsdata = (void *) &dp;
1519 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1520 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1521 (unsigned long) body->oa.o_id,
1522 (unsigned long) body->oa.o_generation);
1523 dp.p_inum = body->oa.o_id;
1525 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1527 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1528 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
1529 "BUG 3550: failed to recreate obj "
1531 body->oa.o_id, new->d_inode->i_ino);
1532 LASSERTF(body->oa.o_generation ==
1533 new->d_inode->i_generation,
1534 "BUG 3550: failed to recreate obj/gen "
1535 LPU64"/%u -> %lu/%u\n",
1536 body->oa.o_id, body->oa.o_generation,
1537 new->d_inode->i_ino,
1538 new->d_inode->i_generation);
1540 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1541 new->d_inode->i_generation = body->oa.o_generation;
1542 mark_inode_dirty(new->d_inode);
1545 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1546 repbody->oa.o_id = new->d_inode->i_ino;
1547 repbody->oa.o_generation = new->d_inode->i_generation;
1548 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1550 rc = fsfilt_del_dir_entry(obd, new);
1551 up(&parent_inode->i_sem);
1554 CERROR("can't remove name for object: %d\n", rc);
1558 cleanup_phase = 2; /* created directory object */
1560 cr_inum = new->d_inode->i_ino;
1561 CDEBUG(D_OTHER, "created dirobj: %lu/%u mode %o\n",
1562 new->d_inode->i_ino, new->d_inode->i_generation,
1563 new->d_inode->i_mode);
1565 up(&parent_inode->i_sem);
1566 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1570 if (body->oa.o_valid & OBD_MD_FLID) {
1571 /* this is new object for splitted dir. we have to
1572 * prevent recursive splitting on it -bzzz */
1573 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1574 OBD_ALLOC(mea, mealen);
1576 GOTO(cleanup, rc = -ENOMEM);
1577 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
1578 mea->mea_master = 0;
1580 down(&new->d_inode->i_sem);
1581 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1582 up(&new->d_inode->i_sem);
1583 OBD_FREE(mea, mealen);
1584 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
1585 obd->obd_name, new->d_inode->i_ino,
1586 new->d_inode->i_generation, flags);
1587 } else if (body->oa.o_easize) {
1588 /* we pass LCK_EX to split routine to signal that we have
1589 * exclusive access to the directory. simple because nobody
1590 * knows it already exists -bzzz */
1591 mds_try_to_split_dir(obd, new, NULL, body->oa.o_easize, LCK_EX);
1595 switch (cleanup_phase) {
1596 case 2: /* object has been created, but we'll may want to replay it later */
1598 ptlrpc_require_repack(req);
1599 case 1: /* transaction */
1600 rc = mds_finish_transno(mds, parent_inode, handle, req, rc, cr_inum);
1604 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1608 static int mdt_get_info(struct ptlrpc_request *req)
1611 struct obd_export *exp = req->rq_export;
1612 int keylen, rc = 0, size = sizeof(obd_id);
1616 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1618 DEBUG_REQ(D_HA, req, "no get_info key");
1621 keylen = req->rq_reqmsg->buflens[0];
1623 if (keylen < strlen("mdsize") || memcmp(key, "mdsize", 6) != 0)
1626 rc = lustre_pack_reply(req, 1, &size, NULL);
1630 reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
1631 rc = obd_get_info(exp, keylen, key, &size, reply);
1632 req->rq_repmsg->status = 0;
1636 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1637 void *key, __u32 vallen, void *val)
1639 struct obd_device *obd;
1640 struct mds_obd *mds;
1644 obd = class_exp2obd(exp);
1646 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1647 exp->exp_handle.h_cookie);
1651 #define KEY_IS(str) \
1652 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
1655 if (KEY_IS("mds_num")) {
1658 CDEBUG(D_IOCTL, "set mds num %d\n", *(int*)val);
1659 mds->mds_num = *(int*)val;
1660 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
1661 valsize = sizeof(group);
1662 /*mds number has been changed, so the corresponding obdfilter exp
1663 *need to be changed too*/
1664 rc = obd_set_info(mds->mds_osc_exp, strlen("mds_conn"), "mds_conn",
1668 CDEBUG(D_IOCTL, "invalid key\n");
1672 static int mdt_set_info(struct ptlrpc_request *req)
1675 struct obd_export *exp = req->rq_export;
1676 int keylen, rc = 0, vallen;
1679 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1681 DEBUG_REQ(D_HA, req, "no set_info key");
1684 keylen = req->rq_reqmsg->buflens[0];
1686 if (keylen == strlen("mds_num") &&
1687 memcmp(key, "mds_num", keylen) == 0) {
1688 rc = lustre_pack_reply(req, 0, NULL, NULL);
1691 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
1693 vallen = req->rq_reqmsg->buflens[1];
1695 rc = obd_set_info(exp, keylen, key, vallen, val);
1696 req->rq_repmsg->status = 0;
1698 } else if (keylen == strlen("client") &&
1699 memcmp(key, "client", keylen) == 0) {
1700 rc = lustre_pack_reply(req, 0, NULL, NULL);
1703 rc = obd_set_info(exp, keylen, key, sizeof(obd_id), NULL);
1704 req->rq_repmsg->status = 0;
1707 CDEBUG(D_IOCTL, "invalid key\n");
1711 int mds_handle(struct ptlrpc_request *req)
1713 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1715 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1716 struct obd_device *obd = NULL;
1719 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1721 LASSERT(current->journal_info == NULL);
1722 /* XXX identical to OST */
1723 if (req->rq_reqmsg->opc != MDS_CONNECT) {
1724 struct mds_export_data *med;
1727 if (req->rq_export == NULL) {
1728 CERROR("lustre_mds: operation %d on unconnected MDS\n",
1729 req->rq_reqmsg->opc);
1730 req->rq_status = -ENOTCONN;
1731 GOTO(out, rc = -ENOTCONN);
1734 med = &req->rq_export->exp_mds_data;
1735 obd = req->rq_export->exp_obd;
1738 /* sanity check: if the xid matches, the request must
1739 * be marked as a resent or replayed */
1740 if (req->rq_xid == med->med_mcd->mcd_last_xid)
1741 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
1742 (MSG_RESENT | MSG_REPLAY),
1743 "rq_xid "LPU64" matches last_xid, "
1744 "expected RESENT flag\n",
1746 /* else: note the opposite is not always true; a
1747 * RESENT req after a failover will usually not match
1748 * the last_xid, since it was likely never
1749 * committed. A REPLAYed request will almost never
1750 * match the last xid, however it could for a
1751 * committed, but still retained, open. */
1753 spin_lock_bh(&obd->obd_processing_task_lock);
1754 recovering = obd->obd_recovering;
1755 spin_unlock_bh(&obd->obd_processing_task_lock);
1757 rc = mds_filter_recovery_request(req, obd,
1759 if (rc || should_process == 0) {
1761 } else if (should_process < 0) {
1762 req->rq_status = should_process;
1763 rc = ptlrpc_error(req);
1769 switch (req->rq_reqmsg->opc) {
1771 DEBUG_REQ(D_INODE, req, "connect");
1772 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1773 rc = target_handle_connect(req);
1775 /* Now that we have an export, set mds. */
1776 mds = mds_req2mds(req);
1779 case MDS_DISCONNECT:
1780 DEBUG_REQ(D_INODE, req, "disconnect");
1781 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1782 rc = target_handle_disconnect(req);
1783 req->rq_status = rc; /* superfluous? */
1787 DEBUG_REQ(D_INODE, req, "getstatus");
1788 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1789 rc = mds_getstatus(req);
1793 DEBUG_REQ(D_INODE, req, "getattr");
1794 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1795 rc = mds_getattr(0, req);
1798 case MDS_GETATTR_NAME: {
1799 struct lustre_handle lockh;
1800 DEBUG_REQ(D_INODE, req, "getattr_name");
1801 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1803 /* If this request gets a reconstructed reply, we won't be
1804 * acquiring any new locks in mds_getattr_name, so we don't
1808 rc = mds_getattr_name(0, req, &lockh, MDS_INODELOCK_UPDATE);
1809 /* this non-intent call (from an ioctl) is special */
1810 req->rq_status = rc;
1811 if (rc == 0 && lockh.cookie)
1812 ldlm_lock_decref(&lockh, LCK_PR);
1816 DEBUG_REQ(D_INODE, req, "statfs");
1817 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1818 rc = mds_statfs(req);
1822 DEBUG_REQ(D_INODE, req, "readpage");
1823 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1824 rc = mds_readpage(req);
1826 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1827 if (req->rq_reply_state) {
1828 lustre_free_reply_state (req->rq_reply_state);
1829 req->rq_reply_state = NULL;
1837 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1839 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1840 mds->mds_max_cookiesize};
1843 /* NB only peek inside req now; mds_reint() will swab it */
1845 CERROR ("Can't inspect opcode\n");
1850 if (lustre_msg_swabbed (req->rq_reqmsg))
1853 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1854 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1855 reint_names[opc] == NULL) ? reint_names[opc] :
1858 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1860 if (opc == REINT_UNLINK || opc == REINT_RENAME)
1862 else if (opc == REINT_OPEN)
1867 rc = lustre_pack_reply(req, bufcount, size, NULL);
1871 rc = mds_reint(req, 0, NULL);
1872 fail = OBD_FAIL_MDS_REINT_NET_REP;
1877 DEBUG_REQ(D_INODE, req, "close");
1878 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1879 rc = mds_close(req);
1882 case MDS_DONE_WRITING:
1883 DEBUG_REQ(D_INODE, req, "done_writing");
1884 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1885 rc = mds_done_writing(req);
1889 DEBUG_REQ(D_INODE, req, "pin");
1890 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1895 DEBUG_REQ(D_INODE, req, "sync");
1896 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1901 DEBUG_REQ(D_INODE, req, "ping");
1902 rc = target_handle_ping(req);
1905 case OBD_LOG_CANCEL:
1906 CDEBUG(D_INODE, "log cancel\n");
1907 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1908 rc = -ENOTSUPP; /* la la la */
1912 DEBUG_REQ(D_INODE, req, "enqueue");
1913 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1914 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1915 ldlm_server_blocking_ast, NULL);
1918 DEBUG_REQ(D_INODE, req, "convert");
1919 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1920 rc = ldlm_handle_convert(req);
1922 case LDLM_BL_CALLBACK:
1923 case LDLM_CP_CALLBACK:
1924 DEBUG_REQ(D_INODE, req, "callback");
1925 CERROR("callbacks should not happen on MDS\n");
1927 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1929 case LLOG_ORIGIN_HANDLE_OPEN:
1930 DEBUG_REQ(D_INODE, req, "llog_init");
1931 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1932 rc = llog_origin_handle_open(req);
1934 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1935 DEBUG_REQ(D_INODE, req, "llog next block");
1936 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1937 rc = llog_origin_handle_next_block(req);
1939 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1940 DEBUG_REQ(D_INODE, req, "llog prev block");
1941 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1942 rc = llog_origin_handle_prev_block(req);
1944 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1945 DEBUG_REQ(D_INODE, req, "llog read header");
1946 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1947 rc = llog_origin_handle_read_header(req);
1949 case LLOG_ORIGIN_HANDLE_CLOSE:
1950 DEBUG_REQ(D_INODE, req, "llog close");
1951 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1952 rc = llog_origin_handle_close(req);
1955 DEBUG_REQ(D_INODE, req, "ost_create");
1956 rc = mdt_obj_create(req);
1959 DEBUG_REQ(D_INODE, req, "get_info");
1960 rc = mdt_get_info(req);
1963 DEBUG_REQ(D_INODE, req, "set_info");
1964 rc = mdt_set_info(req);
1967 CDEBUG(D_INODE, "write\n");
1968 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1969 rc = ost_brw_write(req, NULL);
1970 LASSERT(current->journal_info == NULL);
1971 /* mdt_brw sends its own replies */
1975 DEBUG_REQ(D_INODE, req, "llog catinfo");
1976 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1977 rc = llog_catinfo(req);
1980 req->rq_status = -ENOTSUPP;
1981 rc = ptlrpc_error(req);
1985 LASSERT(current->journal_info == NULL);
1989 /* If we're DISCONNECTing, the mds_export_data is already freed */
1990 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1991 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1992 struct obd_device *obd = list_entry(mds, struct obd_device,
1994 req->rq_repmsg->last_xid =
1995 le64_to_cpu(med->med_mcd->mcd_last_xid);
1997 if (!obd->obd_no_transno) {
1998 req->rq_repmsg->last_committed =
1999 obd->obd_last_committed;
2001 DEBUG_REQ(D_IOCTL, req,
2002 "not sending last_committed update");
2004 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2006 mds->mds_last_transno, obd->obd_last_committed,
2011 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
2012 if (obd && obd->obd_recovering) {
2013 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
2014 return target_queue_final_reply(req, rc);
2016 /* Lost a race with recovery; let the error path DTRT. */
2017 rc = req->rq_status = -ENOTCONN;
2020 target_send_reply(req, rc, fail);
2024 /* Update the server data on disk. This stores the new mount_count and
2025 * also the last_rcvd value to disk. If we don't have a clean shutdown,
2026 * then the server last_rcvd value may be less than that of the clients.
2027 * This will alert us that we may need to do client recovery.
2029 * Also assumes for mds_last_transno that we are not modifying it (no locking).
2031 int mds_update_server_data(struct obd_device *obd, int force_sync)
2033 struct mds_obd *mds = &obd->u.mds;
2034 struct mds_server_data *msd = mds->mds_server_data;
2035 struct file *filp = mds->mds_rcvd_filp;
2036 struct lvfs_run_ctxt saved;
2041 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2042 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2044 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2045 mds->mds_mount_count, mds->mds_last_transno);
2046 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
2048 CERROR("error writing MDS server data: rc = %d\n", rc);
2049 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2054 /* mount the file system (secretly) */
2055 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
2057 struct lustre_cfg* lcfg = buf;
2058 struct mds_obd *mds = &obd->u.mds;
2059 char *options = NULL;
2060 struct vfsmount *mnt;
2065 dev_clear_rdonly(2);
2067 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
2068 RETURN(rc = -EINVAL);
2070 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
2071 if (IS_ERR(obd->obd_fsops))
2072 RETURN(rc = PTR_ERR(obd->obd_fsops));
2074 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2076 page = __get_free_page(GFP_KERNEL);
2080 options = (char *)page;
2081 memset(options, 0, PAGE_SIZE);
2083 /* here we use "iopen_nopriv" hardcoded, because it affects MDS utility
2084 * and the rest of options are passed by mount options. Probably this
2085 * should be moved to somewhere else like startup scripts or lconf. */
2086 sprintf(options, "iopen_nopriv");
2088 if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
2089 sprintf(options + strlen(options), ",%s",
2090 lcfg->lcfg_inlbuf4);
2092 /* we have to know mdsnum before touching underlying fs -bzzz */
2093 sema_init(&mds->mds_lmv_sem, 1);
2094 mds->mds_lmv_connected = 0;
2095 if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
2096 strcmp(lcfg->lcfg_inlbuf5, "dumb")) {
2099 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
2100 obd->obd_name, lcfg->lcfg_inlbuf5);
2102 generate_random_uuid(uuid);
2103 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
2105 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2106 if (mds->mds_lmv_name == NULL)
2107 RETURN(rc = -ENOMEM);
2109 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf5,
2110 lcfg->lcfg_inllen5);
2112 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
2114 OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2119 /* FIXME-WANGDI: this should be reworked when we will use lmv along
2120 * with cobd, because correct mdsnum is set in mds_lmv_connect(). */
2121 if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 && !mds->mds_lmv_obd &&
2122 strcmp(lcfg->lcfg_inlbuf6, "dumb")) {
2123 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master")) &&
2124 mds->mds_num == 0) {
2125 mds->mds_num = REAL_MDS_NUMBER;
2126 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache")) &&
2127 mds->mds_num == 0) {
2128 mds->mds_num = CACHE_MDS_NUMBER;
2132 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
2133 lcfg->lcfg_inlbuf1, options);
2139 CERROR("do_kern_mount failed: rc = %d\n", rc);
2143 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
2145 sema_init(&mds->mds_orphan_recovery_sem, 1);
2146 sema_init(&mds->mds_epoch_sem, 1);
2147 spin_lock_init(&mds->mds_transno_lock);
2148 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2149 atomic_set(&mds->mds_real_clients, 0);
2151 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
2152 LDLM_NAMESPACE_SERVER);
2153 if (obd->obd_namespace == NULL) {
2154 mds_cleanup(obd, 0);
2155 GOTO(err_put, rc = -ENOMEM);
2157 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
2159 rc = mds_fs_setup(obd, mnt);
2161 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
2165 rc = llog_start_commit_thread();
2169 /* this check for @dumb string is needed to handle mounting MDS with
2170 * smfs. Read lconf:MDSDEV.write_conf() for more details. */
2171 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2172 strcmp(lcfg->lcfg_inlbuf3, "dumb")) {
2175 generate_random_uuid(uuid);
2176 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
2178 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
2179 if (mds->mds_profile == NULL)
2180 GOTO(err_fs, rc = -ENOMEM);
2182 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
2183 lcfg->lcfg_inllen3);
2186 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2187 "mds_ldlm_client", &obd->obd_ldlm_client);
2188 obd->obd_replayable = 1;
2190 mds->mds_counters = lprocfs_alloc_mds_counters();
2192 rc = mds_postsetup(obd);
2199 /* No extra cleanup needed for llog_init_commit_thread() */
2200 mds_fs_cleanup(obd, 0);
2202 ldlm_namespace_free(obd->obd_namespace, 0);
2203 obd->obd_namespace = NULL;
2206 mntput(mds->mds_vfsmnt);
2210 fsfilt_put_ops(obd->obd_fsops);
2214 static int mds_postsetup(struct obd_device *obd)
2216 struct mds_obd *mds = &obd->u.mds;
2220 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
2221 obd, 0, NULL, &llog_lvfs_ops);
2225 if (mds->mds_profile) {
2226 struct llog_ctxt *lgctxt;
2227 struct lvfs_run_ctxt saved;
2228 struct lustre_profile *lprof;
2229 struct config_llog_instance cfg;
2231 cfg.cfg_instance = NULL;
2232 cfg.cfg_uuid = mds->mds_lov_uuid;
2233 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2235 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
2237 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2238 GOTO(err_llog, rc = -EINVAL);
2241 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
2242 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2247 lprof = class_get_profile(mds->mds_profile);
2248 if (lprof == NULL) {
2249 CERROR("No profile found: %s\n", mds->mds_profile);
2250 GOTO(err_cleanup, rc = -ENOENT);
2252 rc = mds_lov_connect(obd, lprof->lp_osc);
2254 GOTO(err_cleanup, rc);
2256 rc = mds_lmv_postsetup(obd);
2258 GOTO(err_cleanup, rc);
2266 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
2267 LLOG_CONFIG_ORIG_CTXT));
2271 int mds_postrecov(struct obd_device *obd)
2273 struct mds_obd *mds = &obd->u.mds;
2274 struct llog_ctxt *ctxt;
2278 LASSERT(!obd->obd_recovering);
2279 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2280 LASSERT(ctxt != NULL);
2282 /* set nextid first, so we are sure it happens */
2283 rc = mds_lov_set_nextid(obd);
2285 CERROR("%s: mds_lov_set_nextid failed\n", obd->obd_name);
2289 /* clean PENDING dir */
2290 rc = mds_cleanup_orphans(obd);
2295 rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
2298 CERROR("%s: failed at llog_origin_connect: %d\n",
2303 /* remove the orphaned precreated objects */
2304 rc = mds_lov_clearorphans(mds, NULL /* all OSTs */);
2309 RETURN(rc < 0 ? rc : item);
2312 /* cleanup all llogging subsystems */
2313 rc = obd_llog_finish(obd, &obd->obd_llogs,
2314 mds->mds_lov_desc.ld_tgt_count);
2316 CERROR("%s: failed to cleanup llogging subsystems\n",
2321 int mds_lov_clean(struct obd_device *obd)
2323 struct mds_obd *mds = &obd->u.mds;
2325 if (mds->mds_profile) {
2327 struct config_llog_instance cfg;
2328 struct lvfs_run_ctxt saved;
2329 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
2331 OBD_ALLOC(cln_prof, len);
2332 sprintf(cln_prof, "%s-clean", mds->mds_profile);
2334 cfg.cfg_instance = NULL;
2335 cfg.cfg_uuid = mds->mds_lov_uuid;
2337 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2338 class_config_process_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT),
2340 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2342 OBD_FREE(cln_prof, len);
2343 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2344 mds->mds_profile = NULL;
2349 int mds_lmv_clean(struct obd_device *obd)
2351 struct mds_obd *mds = &obd->u.mds;
2353 if (mds->mds_lmv_name) {
2354 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
2355 mds->mds_lmv_name = NULL;
2360 static int mds_precleanup(struct obd_device *obd, int flags)
2366 mds_lov_disconnect(obd, flags);
2368 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2372 static int mds_cleanup(struct obd_device *obd, int flags)
2374 struct mds_obd *mds = &obd->u.mds;
2377 if (mds->mds_sb == NULL)
2380 mds_update_server_data(obd, 1);
2381 if (mds->mds_lov_objids != NULL) {
2382 OBD_FREE(mds->mds_lov_objids,
2383 mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
2385 mds_fs_cleanup(obd, flags);
2389 /* 2 seems normal on mds, (may_umount() also expects 2
2390 fwiw), but we only see 1 at this point in obdfilter. */
2391 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
2392 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
2393 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
2395 mntput(mds->mds_vfsmnt);
2399 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
2401 if (mds->mds_counters) {
2402 lprocfs_free_mds_counters(mds->mds_counters);
2405 spin_lock_bh(&obd->obd_processing_task_lock);
2406 if (obd->obd_recovering) {
2407 target_cancel_recovery_timer(obd);
2408 obd->obd_recovering = 0;
2410 spin_unlock_bh(&obd->obd_processing_task_lock);
2413 dev_clear_rdonly(2);
2414 fsfilt_put_ops(obd->obd_fsops);
2419 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
2420 struct ldlm_lock *new_lock,
2421 struct lustre_handle *lockh)
2423 struct obd_export *exp = req->rq_export;
2424 struct obd_device *obd = exp->exp_obd;
2425 struct ldlm_request *dlmreq =
2426 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
2427 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
2428 struct list_head *iter;
2430 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2433 l_lock(&obd->obd_namespace->ns_lock);
2434 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2435 struct ldlm_lock *lock;
2436 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2437 if (lock == new_lock)
2439 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2440 lockh->cookie = lock->l_handle.h_cookie;
2441 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2443 l_unlock(&obd->obd_namespace->ns_lock);
2447 l_unlock(&obd->obd_namespace->ns_lock);
2449 /* If the xid matches, then we know this is a resent request,
2450 * and allow it. (It's probably an OPEN, for which we don't
2452 if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
2455 /* This remote handle isn't enqueued, so we never received or
2456 * processed this request. Clear MSG_RESENT, because it can
2457 * be handled like any normal request now. */
2459 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2461 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2465 int intent_disposition(struct ldlm_reply *rep, int flag)
2469 return (rep->lock_policy_res1 & flag);
2472 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2476 rep->lock_policy_res1 |= flag;
2479 static int mds_intent_policy(struct ldlm_namespace *ns,
2480 struct ldlm_lock **lockp, void *req_cookie,
2481 ldlm_mode_t mode, int flags, void *data)
2483 struct ptlrpc_request *req = req_cookie;
2484 struct ldlm_lock *lock = *lockp;
2485 struct ldlm_intent *it;
2486 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2487 struct ldlm_reply *rep;
2488 struct lustre_handle lockh = { 0 };
2489 struct ldlm_lock *new_lock;
2490 int getattr_part = MDS_INODELOCK_UPDATE;
2491 int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
2492 sizeof(struct mds_body),
2493 mds->mds_max_mdsize,
2494 mds->mds_max_cookiesize};
2497 LASSERT(req != NULL);
2499 if (req->rq_reqmsg->bufcount <= 1) {
2500 /* No intent was provided */
2501 int size = sizeof(struct ldlm_reply);
2502 rc = lustre_pack_reply(req, 1, &size, NULL);
2507 it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
2509 CERROR("Intent missing\n");
2510 RETURN(req->rq_status = -EFAULT);
2513 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2515 rc = lustre_pack_reply(req, 3, repsize, NULL);
2517 RETURN(req->rq_status = rc);
2519 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
2520 intent_set_disposition(rep, DISP_IT_EXECD);
2522 fixup_handle_for_resent_req(req, lock, &lockh);
2524 /* execute policy */
2525 switch ((long)it->opc) {
2527 case IT_CREAT|IT_OPEN:
2528 /* XXX swab here to assert that an mds_open reint
2529 * packet is following */
2530 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
2532 /* We abort the lock if the lookup was negative and
2533 * we did not make it to the OPEN portion */
2534 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2535 RETURN(ELDLM_LOCK_ABORTED);
2536 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2537 !intent_disposition(rep, DISP_OPEN_OPEN))
2539 /* IT_OPEN may return lock on cross-node dentry
2540 * that we want to hold during attr retrival -bzzz */
2541 if (rc != 0 || lockh.cookie == 0)
2542 RETURN(ELDLM_LOCK_ABORTED);
2545 getattr_part = MDS_INODELOCK_LOOKUP;
2548 getattr_part |= MDS_INODELOCK_LOOKUP;
2550 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh,
2552 /* FIXME: LDLM can set req->rq_status. MDS sets
2553 policy_res{1,2} with disposition and status.
2554 - replay: returns 0 & req->status is old status
2555 - otherwise: returns req->status */
2556 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2557 rep->lock_policy_res2 = 0;
2558 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2559 rep->lock_policy_res2)
2560 RETURN(ELDLM_LOCK_ABORTED);
2561 if (req->rq_status != 0) {
2563 rep->lock_policy_res2 = req->rq_status;
2564 RETURN(ELDLM_LOCK_ABORTED);
2568 rc = mds_lock_and_check_slave(offset, req, &lockh);
2569 if ((rep->lock_policy_res2 = rc)) {
2571 rep->lock_policy_res2 = 0;
2572 RETURN(ELDLM_LOCK_ABORTED);
2576 CERROR("Unhandled intent "LPD64"\n", it->opc);
2580 /* By this point, whatever function we called above must have either
2581 * filled in 'lockh', been an intent replay, or returned an error. We
2582 * want to allow replayed RPCs to not get a lock, since we would just
2583 * drop it below anyways because lock replay is done separately by the
2584 * client afterwards. For regular RPCs we want to give the new lock to
2585 * the client instead of whatever lock it was about to get. */
2586 new_lock = ldlm_handle2lock(&lockh);
2587 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2590 LASSERT(new_lock != NULL);
2592 /* If we've already given this lock to a client once, then we should
2593 * have no readers or writers. Otherwise, we should have one reader
2594 * _or_ writer ref (which will be zeroed below) before returning the
2595 * lock to a client. */
2596 if (new_lock->l_export == req->rq_export) {
2597 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2599 LASSERT(new_lock->l_export == NULL);
2600 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2605 if (new_lock->l_export == req->rq_export) {
2606 /* Already gave this to the client, which means that we
2607 * reconstructed a reply. */
2608 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2610 RETURN(ELDLM_LOCK_REPLACED);
2613 /* Fixup the lock to be given to the client */
2614 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
2615 new_lock->l_readers = 0;
2616 new_lock->l_writers = 0;
2618 new_lock->l_export = class_export_get(req->rq_export);
2619 list_add(&new_lock->l_export_chain,
2620 &new_lock->l_export->exp_ldlm_data.led_held_locks);
2622 new_lock->l_blocking_ast = lock->l_blocking_ast;
2623 new_lock->l_completion_ast = lock->l_completion_ast;
2625 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2626 sizeof(lock->l_remote_handle));
2628 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2630 LDLM_LOCK_PUT(new_lock);
2631 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
2633 RETURN(ELDLM_LOCK_REPLACED);
2636 int mds_attach(struct obd_device *dev, obd_count len, void *data)
2638 struct lprocfs_static_vars lvars;
2640 lprocfs_init_multi_vars(0, &lvars);
2641 return lprocfs_obd_attach(dev, lvars.obd_vars);
2644 int mds_detach(struct obd_device *dev)
2646 return lprocfs_obd_detach(dev);
2649 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
2651 struct lprocfs_static_vars lvars;
2653 lprocfs_init_multi_vars(1, &lvars);
2654 return lprocfs_obd_attach(dev, lvars.obd_vars);
2657 int mdt_detach(struct obd_device *dev)
2659 return lprocfs_obd_detach(dev);
2662 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
2664 struct mds_obd *mds = &obd->u.mds;
2669 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2670 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
2671 mds_handle, "mds", obd->obd_proc_entry);
2673 if (!mds->mds_service) {
2674 CERROR("failed to start service\n");
2678 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
2681 GOTO(err_thread, rc);
2683 mds->mds_setattr_service =
2684 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2685 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
2686 mds_handle, "mds_setattr",
2687 obd->obd_proc_entry);
2688 if (!mds->mds_setattr_service) {
2689 CERROR("failed to start getattr service\n");
2690 GOTO(err_thread, rc = -ENOMEM);
2693 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
2694 MDT_NUM_THREADS, "ll_mdt_attr");
2696 GOTO(err_thread2, rc);
2698 mds->mds_readpage_service =
2699 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2700 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
2701 mds_handle, "mds_readpage",
2702 obd->obd_proc_entry);
2703 if (!mds->mds_readpage_service) {
2704 CERROR("failed to start readpage service\n");
2705 GOTO(err_thread2, rc = -ENOMEM);
2708 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
2709 MDT_NUM_THREADS, "ll_mdt_rdpg");
2712 GOTO(err_thread3, rc);
2717 ptlrpc_unregister_service(mds->mds_readpage_service);
2719 ptlrpc_unregister_service(mds->mds_setattr_service);
2721 ptlrpc_unregister_service(mds->mds_service);
2725 static int mdt_cleanup(struct obd_device *obd, int flags)
2727 struct mds_obd *mds = &obd->u.mds;
2730 ptlrpc_stop_all_threads(mds->mds_readpage_service);
2731 ptlrpc_unregister_service(mds->mds_readpage_service);
2733 ptlrpc_stop_all_threads(mds->mds_setattr_service);
2734 ptlrpc_unregister_service(mds->mds_setattr_service);
2736 ptlrpc_stop_all_threads(mds->mds_service);
2737 ptlrpc_unregister_service(mds->mds_service);
2742 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2745 struct obd_device *obd = data;
2748 fid.generation = gen;
2749 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2752 static int mds_get_info(struct obd_export *exp, __u32 keylen,
2753 void *key, __u32 *vallen, void *val)
2755 struct obd_device *obd;
2756 struct mds_obd *mds;
2759 obd = class_exp2obd(exp);
2761 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2762 exp->exp_handle.h_cookie);
2766 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
2767 /*Get log_context handle*/
2768 unsigned long *llh_handle = val;
2769 *vallen = sizeof(unsigned long);
2770 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
2773 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
2774 /*Get log_context handle*/
2775 unsigned long *sb = val;
2776 *vallen = sizeof(unsigned long);
2777 *sb = (unsigned long)obd->u.mds.mds_sb;
2782 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
2783 __u32 *mdsize = val;
2784 *vallen = sizeof(*mdsize);
2785 *mdsize = mds->mds_max_mdsize;
2789 CDEBUG(D_IOCTL, "invalid key\n");
2793 struct lvfs_callback_ops mds_lvfs_ops = {
2794 l_fid2dentry: mds_lvfs_fid2dentry,
2797 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
2798 int objcount, struct obd_ioobj *obj,
2799 int niocount, struct niobuf_remote *nb,
2800 struct niobuf_local *res,
2801 struct obd_trans_info *oti);
2802 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
2803 int objcount, struct obd_ioobj *obj, int niocount,
2804 struct niobuf_local *res, struct obd_trans_info *oti,
2807 /* use obd ops to offer management infrastructure */
2808 static struct obd_ops mds_obd_ops = {
2809 .o_owner = THIS_MODULE,
2810 .o_attach = mds_attach,
2811 .o_detach = mds_detach,
2812 .o_connect = mds_connect,
2813 .o_init_export = mds_init_export,
2814 .o_destroy_export = mds_destroy_export,
2815 .o_disconnect = mds_disconnect,
2816 .o_setup = mds_setup,
2817 .o_precleanup = mds_precleanup,
2818 .o_cleanup = mds_cleanup,
2819 .o_postrecov = mds_postrecov,
2820 .o_statfs = mds_obd_statfs,
2821 .o_iocontrol = mds_iocontrol,
2822 .o_create = mds_obd_create,
2823 .o_destroy = mds_obd_destroy,
2824 .o_llog_init = mds_llog_init,
2825 .o_llog_finish = mds_llog_finish,
2826 .o_notify = mds_notify,
2827 .o_get_info = mds_get_info,
2828 .o_set_info = mds_set_info,
2829 .o_preprw = mds_preprw,
2830 .o_commitrw = mds_commitrw,
2833 static struct obd_ops mdt_obd_ops = {
2834 .o_owner = THIS_MODULE,
2835 .o_attach = mdt_attach,
2836 .o_detach = mdt_detach,
2837 .o_setup = mdt_setup,
2838 .o_cleanup = mdt_cleanup,
2839 .o_attach = mdt_attach,
2840 .o_detach = mdt_detach,
2843 static int __init mds_init(void)
2845 struct lprocfs_static_vars lvars;
2847 lprocfs_init_multi_vars(0, &lvars);
2848 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
2850 lprocfs_init_multi_vars(1, &lvars);
2851 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
2857 static void /*__exit*/ mds_exit(void)
2859 class_unregister_type(LUSTRE_MDS_NAME);
2860 class_unregister_type(LUSTRE_MDT_NAME);
2863 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2864 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2865 MODULE_LICENSE("GPL");
2867 module_init(mds_init);
2868 module_exit(mds_exit);