1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Andreas Dilger <adilger@clusterfs.com>
7 * Author: Phil Schwan <phil@clusterfs.com>
8 * Author: Mike Shaver <shaver@clusterfs.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 # define EXPORT_SYMTAB
29 #define DEBUG_SUBSYSTEM S_MDS
31 #include <linux/version.h>
32 #include <linux/module.h>
33 #include <linux/init.h>
34 #include <linux/version.h>
35 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
36 # include <linux/buffer_head.h>
37 # include <linux/workqueue.h>
39 # include <linux/locks.h>
42 #include <linux/obd_class.h>
43 #include <linux/obd_lov.h>
44 #include <linux/lustre_fsfilt.h>
45 #include <linux/lprocfs_status.h>
47 #include "mds_internal.h"
49 /* Exported function from this file are:
51 * mds_open - called by the intent handler
52 * mds_close - an rpc handling function
53 * mds_pin - an rpc handling function - which will go away
54 * mds_mfd_close - for force closing files when a client dies
58 * MDS file data handling: file data holds a handle for a file opened
62 static void mds_mfd_addref(void *mfdp)
64 struct mds_file_data *mfd = mfdp;
66 atomic_inc(&mfd->mfd_refcount);
67 CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
68 atomic_read(&mfd->mfd_refcount));
71 struct mds_file_data *mds_mfd_new(void)
73 struct mds_file_data *mfd;
75 OBD_ALLOC(mfd, sizeof *mfd);
77 CERROR("mds: out of memory\n");
81 atomic_set(&mfd->mfd_refcount, 2);
83 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
84 class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
89 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
92 LASSERT(handle != NULL);
93 RETURN(class_handle2object(handle->cookie));
96 static void mds_mfd_put(struct mds_file_data *mfd)
98 CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
99 atomic_read(&mfd->mfd_refcount) - 1);
100 LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
101 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
102 if (atomic_dec_and_test(&mfd->mfd_refcount)) {
103 LASSERT(list_empty(&mfd->mfd_handle.h_link));
104 OBD_FREE(mfd, sizeof *mfd);
108 static void mds_mfd_destroy(struct mds_file_data *mfd)
110 class_handle_unhash(&mfd->mfd_handle);
115 /* Caller must hold mds->mds_epoch_sem */
116 static int mds_alloc_filterdata(struct inode *inode)
118 LASSERT(inode->i_filterdata == NULL);
119 OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
120 if (inode->i_filterdata == NULL)
122 LASSERT(igrab(inode) == inode);
126 /* Caller must hold mds->mds_epoch_sem */
127 static void mds_free_filterdata(struct inode *inode)
129 LASSERT(inode->i_filterdata != NULL);
130 OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
131 inode->i_filterdata = NULL;
135 /* Write access to a file: executors cause a negative count,
136 * writers a positive count. The semaphore is needed to perform
137 * a check for the sign and then increment or decrement atomically.
139 * This code is closely tied to the allocation of the d_fsdata and the
140 * MDS epoch, so we use the same semaphore for the whole lot.
142 * We could use a different semaphore for each file, if it ever shows
143 * up in a profile, which it won't.
145 * epoch argument is nonzero during recovery */
146 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
151 down(&mds->mds_epoch_sem);
153 if (atomic_read(&inode->i_writecount) < 0) {
154 up(&mds->mds_epoch_sem);
159 if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
160 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
161 MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
162 inode->i_generation);
166 if (inode->i_filterdata == NULL)
167 mds_alloc_filterdata(inode);
168 if (inode->i_filterdata == NULL) {
172 if (epoch > mds->mds_io_epoch)
173 mds->mds_io_epoch = epoch;
176 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
177 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
178 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
181 atomic_inc(&inode->i_writecount);
182 up(&mds->mds_epoch_sem);
186 /* Returns EAGAIN if the client needs to get size and/or cookies and close
187 * again -- which is never true if the file is about to be unlinked. Otherwise
188 * returns the number of remaining writers. */
189 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
190 struct mds_body *body, int unlinking)
195 down(&mds->mds_epoch_sem);
196 atomic_dec(&inode->i_writecount);
197 rc = atomic_read(&inode->i_writecount);
201 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
202 GOTO(out, rc = EAGAIN);
204 mds_free_filterdata(inode);
206 up(&mds->mds_epoch_sem);
210 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
213 down(&mds->mds_epoch_sem);
214 if (atomic_read(&inode->i_writecount) > 0) {
215 up(&mds->mds_epoch_sem);
218 atomic_dec(&inode->i_writecount);
219 up(&mds->mds_epoch_sem);
223 static void mds_allow_write_access(struct inode *inode)
226 atomic_inc(&inode->i_writecount);
229 int mds_query_write_access(struct inode *inode)
232 RETURN(atomic_read(&inode->i_writecount));
235 /* This replaces the VFS dentry_open, it manages mfd and writecount */
236 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
237 struct vfsmount *mnt, int flags,
238 struct ptlrpc_request *req)
240 struct mds_export_data *med = &req->rq_export->exp_mds_data;
241 struct mds_obd *mds = mds_req2mds(req);
242 struct mds_file_data *mfd;
243 struct mds_body *body;
249 CERROR("mds: out of memory\n");
250 GOTO(cleanup_dentry, error = -ENOMEM);
253 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
255 if (flags & FMODE_WRITE) {
256 /* FIXME: in recovery, need to pass old epoch here */
257 error = mds_get_write_access(mds, dentry->d_inode, 0);
259 GOTO(cleanup_mfd, error);
260 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
261 } else if (flags & FMODE_EXEC) {
262 error = mds_deny_write_access(mds, dentry->d_inode);
264 GOTO(cleanup_mfd, error);
269 /* Mark the file as open to handle open-unlink. */
271 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
272 mds_orphan_open_inc(dentry->d_inode);
273 UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
275 mfd->mfd_mode = flags;
276 mfd->mfd_dentry = dentry;
277 mfd->mfd_xid = req->rq_xid;
279 spin_lock(&med->med_open_lock);
280 list_add(&mfd->mfd_list, &med->med_open_head);
281 spin_unlock(&med->med_open_lock);
284 body->handle.cookie = mfd->mfd_handle.h_cookie;
290 mds_mfd_destroy(mfd);
292 return ERR_PTR(error);
295 static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
296 struct lov_desc *desc)
299 for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
300 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
301 le64_to_cpu(lmm->lmm_objects[i].l_object_id);
305 /* Must be called with i_sem held */
306 static int mds_create_objects(struct ptlrpc_request *req, int offset,
307 struct mds_update_record *rec,
308 struct mds_obd *mds, struct obd_device *obd,
309 struct dentry *dchild, void **handle,
312 struct obdo *oa = NULL;
313 struct obd_trans_info oti = { 0 };
314 struct mds_body *body;
315 struct lov_stripe_md *lsm = NULL;
316 struct lov_mds_md *lmm = NULL;
317 struct inode *inode = dchild->d_inode;
319 int rc, lmm_bufsize, lmm_size;
322 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
323 !(rec->ur_flags & FMODE_WRITE))
326 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
328 if (!S_ISREG(inode->i_mode))
330 if (body->valid & OBD_MD_FLEASIZE)
333 OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
336 oti.oti_objid = *ids;
339 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
340 LASSERT(id_ino(rec->ur_id2));
341 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
342 lmm_size = rec->ur_eadatalen;
343 lmm = rec->ur_eadata;
347 *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
348 if (IS_ERR(*handle)) {
349 rc = PTR_ERR(*handle);
354 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
356 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
357 lmm_bufsize = req->rq_repmsg->buflens[offset];
359 LASSERT(lmm_bufsize >= lmm_size);
360 memcpy(lmm_buf, lmm, lmm_size);
361 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
363 CERROR("open replay failed to set md:%d\n", rc);
367 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
368 GOTO(out_ids, rc = -ENOMEM);
372 GOTO(out_ids, rc = -ENOMEM);
373 oa->o_mode = S_IFREG | 0600;
374 oa->o_id = inode->i_ino;
375 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
376 oa->o_generation = inode->i_generation;
377 oa->o_uid = 0; /* must have 0 uid / gid on OST */
379 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
380 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
383 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
386 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
387 /* check if things like lfs setstripe are sending us the ea */
388 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
389 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
391 0, &lsm, rec->ur_eadata);
395 /* Per-directory striping default code removed, because
396 * it uses the same unnamed EA storage as the directory
397 * striping for CMD. -p */
399 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
400 rc = obd_create(mds->mds_lov_exp, oa, &lsm, &oti);
405 CDEBUG(level, "error creating objects for "
406 "inode %lu: rc = %d\n",
409 CERROR("obd_create returned invalid "
416 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_lov_exp,
417 0, &lsm, rec->ur_eadata);
421 lsm->lsm_object_id = oa->o_id;
422 lsm->lsm_object_gr = oa->o_gr;
425 oa->o_size = inode->i_size;
426 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
427 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
428 rc = obd_setattr(mds->mds_lov_exp, oa, lsm, &oti);
430 CERROR("error setting attrs for inode %lu: rc %d\n",
433 CERROR("obd_setattr returned bad rc %d\n", rc);
440 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
441 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
443 LASSERT(lsm && lsm->lsm_object_id);
445 rc = obd_packmd(mds->mds_lov_exp, &lmm, lsm);
446 if (!id_ino(rec->ur_id2))
447 obd_free_memmd(mds->mds_lov_exp, &lsm);
450 body->eadatasize = rc;
453 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
454 if (IS_ERR(*handle)) {
455 rc = PTR_ERR(*handle);
460 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
461 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
462 lmm_bufsize = req->rq_repmsg->buflens[offset];
464 LASSERT(lmm_bufsize >= lmm_size);
466 memcpy(lmm_buf, lmm, lmm_size);
467 obd_free_diskmd(mds->mds_lov_exp, &lmm);
469 oti_free_cookies(&oti);
473 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
479 static void reconstruct_open(struct mds_update_record *rec, int offset,
480 struct ptlrpc_request *req,
481 struct lustre_handle *child_lockh)
483 struct mds_export_data *med = &req->rq_export->exp_mds_data;
484 struct mds_client_data *mcd = med->med_mcd;
485 struct mds_obd *mds = mds_req2mds(req);
486 struct mds_file_data *mfd;
487 struct obd_device *obd = req->rq_export->exp_obd;
488 struct dentry *parent, *dchild;
489 struct ldlm_reply *rep;
490 struct mds_body *body;
496 LASSERT(offset == 3); /* only called via intent */
497 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
498 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
500 /* copy rc, transno and disp; steal locks */
501 mds_req_from_mcd(req, mcd);
502 intent_set_disposition(rep, mcd->mcd_last_data);
504 /* Only replay if create or open actually happened. */
505 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
507 return; /* error looking up parent or child */
510 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
511 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
512 OLID4(rec->ur_id1), PTR_ERR(parent));
514 dchild = ll_lookup_one_len(rec->ur_name, parent,
515 rec->ur_namelen - 1);
516 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
517 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
519 if (!dchild->d_inode)
520 GOTO(out_dput, 0); /* child not present to open */
522 /* At this point, we know we have a child. We'll send
523 * it back _unless_ it not created and open failed. */
524 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
525 !intent_disposition(rep, DISP_OPEN_CREATE) &&
529 /* get lock (write for O_CREAT, read otherwise) */
530 mds_pack_inode2body(obd, body, dchild->d_inode, 0);
531 if (S_ISREG(dchild->d_inode->i_mode)) {
532 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
536 LASSERT(rc == req->rq_status);
538 /* If we have LOV EA data, the OST holds size, mtime */
539 if (!(body->valid & OBD_MD_FLEASIZE))
540 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
541 OBD_MD_FLATIME | OBD_MD_FLMTIME);
544 /* If we have -EEXIST as the status, and we were asked to create
545 * exclusively, we can tell we failed because the file already existed.
547 if (req->rq_status == -EEXIST &&
548 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
549 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
553 /* If we didn't get as far as trying to open, then some locking thing
554 * probably went wrong, and we'll just bail here.
556 if (!intent_disposition(rep, DISP_OPEN_OPEN))
559 /* If we failed, then we must have failed opening, so don't look for
560 * file descriptor or anything, just give the client the bad news.
566 list_for_each(t, &med->med_open_head) {
567 mfd = list_entry(t, struct mds_file_data, mfd_list);
568 if (mfd->mfd_xid == req->rq_xid)
573 /* #warning "XXX fixme" bug 2991 */
574 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
575 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
576 * to detect a re-open */
578 mntget(mds->mds_vfsmnt);
579 CERROR("Re-opened file \n");
580 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
581 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
583 CERROR("mds: out of memory\n");
584 GOTO(out_dput, req->rq_status = -ENOMEM);
588 body->handle.cookie = mfd->mfd_handle.h_cookie;
589 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
590 mfd->mfd_handle.h_cookie);
600 /* do NOT or the MAY_*'s, you'll get the weakest */
601 static int accmode(int flags)
605 if (flags & FMODE_READ)
607 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
609 if (flags & FMODE_EXEC)
614 /* Handles object creation, actual opening, and I/O epoch */
615 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
616 struct mds_body *body, int flags, void **handle,
617 struct mds_update_record *rec,struct ldlm_reply *rep)
619 struct mds_obd *mds = mds_req2mds(req);
620 struct obd_device *obd = req->rq_export->exp_obd;
621 struct mds_file_data *mfd = NULL;
622 obd_id *ids = NULL; /* object IDs created */
627 /* atomically create objects if necessary */
628 down(&dchild->d_inode->i_sem);
629 mode = dchild->d_inode->i_mode;
630 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
631 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
632 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
635 up(&dchild->d_inode->i_sem);
640 /* no EA: create objects */
641 if ((body->valid & OBD_MD_FLEASIZE) &&
642 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
643 up(&dchild->d_inode->i_sem);
646 if (!(body->valid & OBD_MD_FLEASIZE)) {
647 /* no EA: create objects */
648 rc = mds_create_objects(req, 2, rec, mds, obd,
649 dchild, handle, &ids);
651 CERROR("mds_create_objects: rc = %d\n", rc);
652 up(&dchild->d_inode->i_sem);
656 if (S_ISREG(dchild->d_inode->i_mode) &&
657 (body->valid & OBD_MD_FLEASIZE)) {
658 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
661 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
664 up(&dchild->d_inode->i_sem);
669 /* If the inode has no EA data, then MDSs hold size, mtime */
670 if (S_ISREG(dchild->d_inode->i_mode) &&
671 !(body->valid & OBD_MD_FLEASIZE)) {
672 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
673 OBD_MD_FLATIME | OBD_MD_FLMTIME);
675 up(&dchild->d_inode->i_sem);
677 intent_set_disposition(rep, DISP_OPEN_OPEN);
678 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
680 RETURN(PTR_ERR(mfd));
682 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
683 mfd->mfd_handle.h_cookie);
685 mds_lov_update_objids(obd, ids);
686 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
689 // mds_mfd_destroy(mfd);
693 static int mds_open_by_id(struct ptlrpc_request *req,
694 struct lustre_id *id,
695 struct mds_body *body, int flags,
696 struct mds_update_record *rec,
697 struct ldlm_reply *rep)
699 struct mds_obd *mds = mds_req2mds(req);
700 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
701 struct dentry *dchild;
702 char idname[LL_ID_NAMELEN];
707 down(&pending_dir->i_sem);
709 idlen = ll_id2str(idname, id_ino(id), id_gen(id));
711 dchild = lookup_one_len(idname, mds->mds_pending_dir,
713 if (IS_ERR(dchild)) {
714 up(&pending_dir->i_sem);
715 rc = PTR_ERR(dchild);
716 CERROR("error looking up %s in PENDING: rc = %d\n",
721 if (dchild->d_inode != NULL) {
722 up(&pending_dir->i_sem);
723 mds_inode_set_orphan(dchild->d_inode);
724 mds_pack_inode2body(req2obd(req), body, dchild->d_inode, 0);
725 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
726 intent_set_disposition(rep, DISP_LOOKUP_POS);
727 CWARN("Orphan %s found and opened in PENDING directory\n",
732 up(&pending_dir->i_sem);
735 * we didn't find it in PENDING so it isn't an orphan. See if it was a
736 * regular inode that was previously created.
738 dchild = mds_id2dentry(req2obd(req), id, NULL);
740 RETURN(PTR_ERR(dchild));
742 mds_pack_inode2body(req2obd(req), body, dchild->d_inode, 0);
743 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
744 intent_set_disposition(rep, DISP_LOOKUP_POS);
747 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
748 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
749 req, rc, rep ? rep->lock_policy_res1 : 0);
750 /* XXX what do we do here if mds_finish_transno itself failed? */
756 int mds_pin(struct ptlrpc_request *req, int offset)
758 struct obd_device *obd = req->rq_export->exp_obd;
759 struct mds_body *request_body, *reply_body;
760 struct lvfs_run_ctxt saved;
761 int rc, size = sizeof(*reply_body);
764 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
765 sizeof(*request_body));
767 rc = lustre_pack_reply(req, 1, &size, NULL);
770 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
772 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
773 rc = mds_open_by_id(req, &request_body->id1, reply_body,
774 request_body->flags, NULL, NULL);
775 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
781 * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
782 * child_lockh is NULL we just get the lock as a barrier to wait for other
783 * holders of this lock, and drop it right away again.
785 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
786 struct lustre_handle *child_lockh)
788 struct ldlm_res_id child_res_id = { .name = { 0, 0, 0, 0 } };
789 struct lustre_handle lockh;
790 struct lustre_id sid;
795 rc = mds_read_inode_sid(obd, inode, &sid);
798 CERROR("Can't read inode self id, err = %d\n", rc);
802 child_res_id.name[0] = id_fid(&sid);
803 child_res_id.name[1] = id_group(&sid);
804 child_res_id.name[2] = 1;
806 if (child_lockh == NULL)
807 child_lockh = &lockh;
809 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
810 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
811 mds_blocking_ast, ldlm_completion_ast, NULL,
812 NULL, NULL, 0, NULL, child_lockh);
814 CERROR("ldlm_cli_enqueue: %d\n", rc);
815 else if (child_lockh == &lockh)
816 ldlm_lock_decref(child_lockh, LCK_EX);
821 static int is_mount_object(struct dentry *dparent)
823 struct dentry *dchild;
825 if (!(dparent->d_inode->i_mode & S_ISUID))
828 dchild = lookup_one_len(".mntinfo", dparent, strlen(".mntinfo"));
829 if (IS_ERR(dchild) || dchild == NULL)
836 int mds_open(struct mds_update_record *rec, int offset,
837 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
839 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
840 struct obd_device *obd = req->rq_export->exp_obd;
841 struct mds_obd *mds = mds_req2mds(req);
842 struct ldlm_reply *rep = NULL;
843 struct mds_body *body = NULL;
844 struct dentry *dchild = NULL, *dparent = NULL;
845 struct mds_export_data *med;
846 struct lustre_handle parent_lockh[2];
847 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
848 int parent_mode = LCK_PR;
850 struct dentry_params dp;
851 struct mea *mea = NULL;
852 int mea_size, update_mode;
855 DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
856 OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
859 parent_lockh[0].cookie = 0;
860 parent_lockh[1].cookie = 0;
862 if (offset == 3) { /* intent */
863 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
864 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
865 } else if (offset == 1) { /* non-intent reint */
866 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
872 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
873 MDS_UPDATE_COUNTER(mds, MDS_OPEN_COUNT);
876 * Step 0: If we are passed a id, then we assume the client already
877 * opened this file and is only replaying the RPC, so we open the inode
878 * by fid (at some large expense in security).
880 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
881 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
882 OLID4(rec->ur_id2), rec->ur_namelen - 1,
883 rec->ur_name, rec->ur_mode);
885 LASSERT(id_ino(rec->ur_id2));
887 rc = mds_open_by_id(req, rec->ur_id2, body, rec->ur_flags,
892 /* We didn't find the correct inode on disk either, so we
893 * need to re-create it via a regular replay. */
894 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
896 LASSERT(!id_ino(rec->ur_id2));
899 LASSERT(offset == 3); /* If we got here, we must be called via intent */
901 med = &req->rq_export->exp_mds_data;
902 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
903 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
907 acc_mode = accmode(rec->ur_flags);
909 /* Step 1: Find and lock the parent */
910 if (rec->ur_flags & O_CREAT) {
911 /* XXX Well, in fact we only need this lock mode change if
912 in addition to O_CREAT, the file does not exist.
913 But we do not know if it exists or not yet */
914 parent_mode = LCK_PW;
917 if (rec->ur_namelen == 1) {
918 /* client (LMV) wants to open the file by id */
919 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
920 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
921 if (IS_ERR(dchild)) {
922 rc = PTR_ERR(dparent);
923 CERROR("child lookup by an id error %d\n", rc);
929 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
930 parent_lockh, &update_mode, rec->ur_name,
931 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
932 if (IS_ERR(dparent)) {
933 rc = PTR_ERR(dparent);
935 CERROR("parent lookup for "DLID4" failed, error %d\n",
936 OLID4(rec->ur_id1), rc);
939 LASSERT(dparent->d_inode != NULL);
941 cleanup_phase = 1; /* parent dentry and lock */
943 /* try to retrieve MEA data for this dir */
944 rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
950 * dir is already splitted, check is requested filename should *
951 * live at this MDS or at another one.
954 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
955 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
957 "%s: inapropriate MDS(%d) for %lu/%u:%s."
958 " should be %lu(%d)\n", obd->obd_name,
959 mea->mea_master, dparent->d_inode->i_ino,
960 dparent->d_inode->i_generation, rec->ur_name,
961 (unsigned long)id_group(&mea->mea_ids[i]), i);
962 GOTO(cleanup, rc = -ERESTART);
966 /* Step 2: Lookup the child */
967 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
968 if (IS_ERR(dchild)) {
969 rc = PTR_ERR(dchild);
970 dchild = NULL; /* don't confuse mds_finish_transno() below */
975 cleanup_phase = 2; /* child dentry */
977 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
979 if (dchild->d_flags & DCACHE_CROSS_REF) {
980 struct ldlm_res_id res_id = { . name = {0} };
981 ldlm_policy_data_t policy;
984 mds_pack_dentry2body(obd, body, dchild, 1);
985 intent_set_disposition(rep, DISP_LOOKUP_POS);
987 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
990 res_id.name[0] = dchild->d_fid;
991 res_id.name[1] = dchild->d_mdsnum;
993 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
994 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
995 res_id, LDLM_IBITS, &policy,
996 LCK_PR, &flags, mds_blocking_ast,
997 ldlm_completion_ast, NULL, NULL,
998 NULL, 0, NULL, child_lockh);
1000 if (parent_lockh[1].cookie != 0)
1001 ldlm_lock_decref(parent_lockh + 1, update_mode);
1003 ldlm_lock_decref(parent_lockh, parent_mode);
1005 OBD_FREE(mea, mea_size);
1011 if (dchild->d_inode)
1012 intent_set_disposition(rep, DISP_LOOKUP_POS);
1014 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1016 /* Step 3: If the child was negative, and we're supposed to, create it.*/
1017 if (dchild->d_inode == NULL) {
1020 struct inode *inode;
1022 ino = (unsigned long)id_ino(rec->ur_id2);
1024 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1027 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1028 obd->obd_name, dparent->d_inode->i_ino,
1029 dparent->d_inode->i_generation, rc);
1032 /* dir got splitted */
1033 GOTO(cleanup, rc = -ERESTART);
1034 } else if (rc < 0) {
1035 /* error happened during spitting */
1039 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1041 * dentry is negative and we weren't supposed to create
1042 * object, get out of here.
1044 GOTO(cleanup, rc = -ENOENT);
1047 intent_set_disposition(rep, DISP_OPEN_CREATE);
1048 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1050 if (IS_ERR(handle)) {
1051 rc = PTR_ERR(handle);
1055 dchild->d_fsdata = (void *) &dp;
1059 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1060 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1061 dchild->d_fsdata = NULL;
1064 CDEBUG(D_INODE, "error during create: %d\n", rc);
1067 inode = dchild->d_inode;
1069 LASSERT(ino == inode->i_ino);
1071 /* written as part of setattr */
1072 inode->i_generation = id_gen(rec->ur_id2);
1073 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1074 inode->i_ino, inode->i_generation);
1078 LTIME_S(iattr.ia_atime) = rec->ur_time;
1079 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1080 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1082 iattr.ia_uid = rec->ur_fsuid;
1083 if (dparent->d_inode->i_mode & S_ISGID)
1084 iattr.ia_gid = dparent->d_inode->i_gid;
1086 iattr.ia_gid = rec->ur_fsgid;
1088 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1089 ATTR_MTIME | ATTR_CTIME;
1091 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1093 CERROR("error on child setattr: rc = %d\n", rc);
1095 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1097 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1099 CERROR("error on parent setattr: rc = %d\n", rc);
1101 MDS_UPDATE_COUNTER(mds, MDS_CREATE_COUNT);
1103 down(&dchild->d_inode->i_sem);
1105 rc = mds_update_inode_sid(obd, dchild->d_inode,
1106 handle, rec->ur_id2);
1108 CERROR("mds_update_inode_sid() failed, "
1111 id_assign_fid(&body->id1, rec->ur_id2);
1114 * make sure, that fid is up-to-date.
1116 mds_set_last_fid(obd, id_fid(rec->ur_id2));
1118 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1119 handle, &body->id1);
1121 CERROR("mds_alloc_inode_sid() failed, "
1125 up(&dchild->d_inode->i_sem);
1127 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1128 rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1129 dchild->d_inode, handle, 0);
1133 acc_mode = 0; /* Don't check for permissions */
1136 * we do not read fid from EA here, because it is already
1137 * updated and thus we avoid not needed IO, locking, etc.
1139 body->valid |= OBD_MD_FID;
1140 mds_pack_inode2body(obd, body, dchild->d_inode, 0);
1142 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1144 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1145 "dchild %*s (%p) inode %p\n", dchild->d_name.len,
1146 dchild->d_name.name, dchild, dchild->d_inode);
1148 if (S_ISREG(dchild->d_inode->i_mode)) {
1149 /* Check permissions etc */
1150 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1154 /* Can't write to a read-only file */
1155 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1156 GOTO(cleanup, rc = -EPERM);
1158 /* An append-only file must be opened in append mode for
1160 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1161 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1162 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1163 GOTO(cleanup, rc = -EPERM);
1166 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1167 (rec->ur_flags & MDS_OPEN_EXCL)) {
1168 /* File already exists, we didn't just create it, and we
1169 * were passed O_EXCL; err-or. */
1170 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1173 /* if we are following a symlink, don't open */
1174 if (S_ISLNK(dchild->d_inode->i_mode))
1175 GOTO(cleanup, rc = 0);
1177 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1178 !S_ISDIR(dchild->d_inode->i_mode))
1179 GOTO(cleanup, rc = -ENOTDIR);
1181 if (S_ISDIR(dchild->d_inode->i_mode)) {
1182 if (rec->ur_flags & MDS_OPEN_CREAT ||
1183 rec->ur_flags & FMODE_WRITE) {
1184 /*we are trying to create or write a exist dir*/
1185 GOTO(cleanup, rc = -EISDIR);
1187 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1188 GOTO(cleanup, rc = -EACCES);
1190 if (is_mount_object(dchild)) {
1191 CERROR("Found possible GNS mount object %*s; not "
1192 "opening.\n", dchild->d_name.len,
1193 dchild->d_name.name);
1194 GOTO(cleanup, rc = 0); // success, but don't really open
1198 /* Step 5: mds_open it */
1199 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1204 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1205 req, rc, rep ? rep->lock_policy_res1 : 0);
1207 switch (cleanup_phase) {
1209 if (rc && created) {
1210 int err = vfs_unlink(dparent->d_inode, dchild);
1212 CERROR("unlink(%*s) in error path: %d\n",
1213 dchild->d_name.len, dchild->d_name.name,
1216 } else if (created) {
1217 mds_lock_new_child(obd, dchild->d_inode, NULL);
1221 if (dparent == NULL)
1226 if (parent_lockh[1].cookie != 0)
1227 ldlm_lock_decref(parent_lockh + 1, update_mode);
1230 ldlm_lock_decref(parent_lockh, parent_mode);
1232 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1235 OBD_FREE(mea, mea_size);
1239 /* Close a "file descriptor" and possibly unlink an orphan from the
1240 * PENDING directory. Caller must hold child->i_sem, this drops it.
1242 * If we are being called from mds_disconnect() because the client has
1243 * disappeared, then req == NULL and we do not update last_rcvd because
1244 * there is nothing that could be recovered by the client at this stage
1245 * (it will not even _have_ an entry in last_rcvd anymore).
1247 * Returns EAGAIN if the client needs to get more data and re-close. */
1248 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1249 struct obd_device *obd, struct mds_file_data *mfd,
1252 struct inode *inode = mfd->mfd_dentry->d_inode;
1253 char idname[LL_ID_NAMELEN];
1254 int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1255 struct dentry *pending_child = NULL;
1256 struct mds_obd *mds = &obd->u.mds;
1257 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1258 void *handle = NULL;
1259 struct mds_body *request_body = NULL, *reply_body = NULL;
1260 struct dentry_params dp;
1261 struct iattr iattr = { 0 };
1262 struct llog_create_locks *lcl = NULL;
1265 if (req && req->rq_reqmsg != NULL)
1266 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1267 sizeof(*request_body));
1268 if (req && req->rq_repmsg != NULL)
1269 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1270 sizeof(*reply_body));
1272 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1273 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1274 idname, inode->i_nlink, mds_orphan_open_count(inode));
1276 last_orphan = mds_orphan_open_dec_test(inode) &&
1277 mds_inode_is_orphan(inode);
1278 UP_WRITE_I_ALLOC_SEM(inode);
1280 /* this is half of the actual "close" */
1281 if (mfd->mfd_mode & FMODE_WRITE) {
1282 rc = mds_put_write_access(mds, inode, request_body,
1283 last_orphan && unlink_orphan);
1284 } else if (mfd->mfd_mode & FMODE_EXEC) {
1285 mds_allow_write_access(inode);
1288 if (last_orphan && unlink_orphan) {
1289 struct lov_mds_md *lmm = NULL;
1290 int stripe_count = 0;
1291 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1293 if (obd->obd_recovering) {
1294 CDEBUG(D_HA, "not remove orphan %s until recovery"
1295 " is over\n", idname);
1299 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1301 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1302 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1303 CERROR("found \"orphan\" %s %s with link count %d\n",
1304 S_ISREG(inode->i_mode) ? "file" : "dir",
1305 idname, inode->i_nlink);
1307 * sadly, there is no easy way to save pending_child from
1308 * mds_reint_unlink() into mfd, so we need to re-lookup, but
1309 * normally it will still be in the dcache.
1311 down(&pending_dir->i_sem);
1312 cleanup_phase = 1; /* up(i_sem) when finished */
1313 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1315 if (IS_ERR(pending_child))
1316 GOTO(cleanup, rc = PTR_ERR(pending_child));
1317 LASSERT(pending_child->d_inode != NULL);
1319 cleanup_phase = 2; /* dput(pending_child) when finished */
1320 if (req != NULL && req->rq_repmsg != NULL) {
1321 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1322 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1325 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1326 NULL, stripe_count);
1327 if (IS_ERR(handle)) {
1328 rc = PTR_ERR(handle);
1333 pending_child->d_fsdata = (void *) &dp;
1336 if (S_ISDIR(pending_child->d_inode->i_mode))
1337 rc = vfs_rmdir(pending_dir, pending_child);
1339 rc = vfs_unlink(pending_dir, pending_child);
1341 CERROR("error unlinking orphan %s: rc %d\n",
1344 if (req != NULL && req->rq_repmsg != NULL &&
1345 (reply_body->valid & OBD_MD_FLEASIZE) &&
1346 mds_log_op_unlink(obd, pending_child->d_inode,
1347 lmm, req->rq_repmsg->buflens[1],
1348 lustre_msg_buf(req->rq_repmsg, 2, 0),
1349 req->rq_repmsg->buflens[2], &lcl) > 0) {
1350 reply_body->valid |= OBD_MD_FLCOOKIE;
1353 goto out; /* Don't bother updating attrs on unlinked inode */
1357 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1358 /* Update the on-disk attributes if this was the last write
1359 * close, and all information was provided (i.e., rc == 0)
1361 * XXX this should probably be abstracted with mds_reint_setattr
1364 if (request_body->valid & OBD_MD_FLMTIME &&
1365 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1366 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1367 iattr.ia_valid |= ATTR_MTIME;
1369 if (request_body->valid & OBD_MD_FLCTIME &&
1370 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1371 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1372 iattr.ia_valid |= ATTR_CTIME;
1375 /* XXX can't set block count with fsfilt_setattr (!) */
1376 if (request_body->valid & OBD_MD_FLSIZE) {
1377 iattr.ia_valid |= ATTR_SIZE;
1378 iattr.ia_size = request_body->size;
1380 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1381 iattr.ia_valid |= ATTR_BLOCKS;
1382 iattr.ia_blocks = request_body->blocks
1387 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1388 /* Only start a transaction to write out only the atime if
1389 * it is more out-of-date than the specified limit. If we
1390 * are already going to write out the atime then do it anyway.
1392 if ((request_body->atime >
1393 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1394 (iattr.ia_valid != 0 &&
1395 request_body->atime > LTIME_S(inode->i_atime))) {
1396 LTIME_S(iattr.ia_atime) = request_body->atime;
1397 iattr.ia_valid |= ATTR_ATIME;
1401 if (iattr.ia_valid != 0) {
1402 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1404 GOTO(cleanup, rc = PTR_ERR(handle));
1405 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1407 CERROR("error in setattr(%s): rc %d\n", idname, rc);
1410 /* If other clients have this file open for write, rc will be > 0 */
1413 l_dput(mfd->mfd_dentry);
1414 mds_mfd_destroy(mfd);
1417 if (req != NULL && reply_body != NULL) {
1418 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1419 } else if (handle) {
1420 int err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, 0);
1422 CERROR("error committing close: %d\n", err);
1428 switch (cleanup_phase) {
1431 ptlrpc_save_llog_lock(req, lcl);
1432 dput(pending_child);
1434 up(&pending_dir->i_sem);
1439 int mds_close(struct ptlrpc_request *req, int offset)
1441 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1442 struct obd_device *obd = req->rq_export->exp_obd;
1443 struct mds_body *body;
1444 struct mds_file_data *mfd;
1445 struct lvfs_run_ctxt saved;
1446 struct inode *inode;
1447 int rc, repsize[3] = {sizeof(struct mds_body),
1448 obd->u.mds.mds_max_mdsize,
1449 obd->u.mds.mds_max_cookiesize};
1452 MDS_UPDATE_COUNTER((&obd->u.mds), MDS_CLOSE_COUNT);
1454 rc = lustre_pack_reply(req, 3, repsize, NULL);
1456 CERROR("lustre_pack_reply: rc = %d\n", rc);
1457 req->rq_status = rc;
1459 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1462 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1463 DEBUG_REQ(D_HA, req, "close replay\n");
1464 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1465 lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1466 req->rq_repmsg->buflens[2]);
1470 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1471 lustre_swab_mds_body);
1473 CERROR("Can't unpack body\n");
1474 req->rq_status = -EFAULT;
1478 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1479 /* do some stuff */ ;
1481 mfd = mds_handle2mfd(&body->handle);
1483 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1484 ": cookie "LPX64, body->id1.li_stc.u.e3s.l3s_ino,
1485 body->handle.cookie);
1486 req->rq_status = -ESTALE;
1490 inode = mfd->mfd_dentry->d_inode;
1491 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1492 DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1493 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1494 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1495 LASSERT(body != NULL);
1497 mds_pack_inode2body(obd, body, inode, 0);
1498 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, MDS_PACK_MD_LOCK);
1500 spin_lock(&med->med_open_lock);
1501 list_del(&mfd->mfd_list);
1502 spin_unlock(&med->med_open_lock);
1504 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1505 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1506 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1508 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1509 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1510 req->rq_status = -ENOMEM;
1519 int mds_done_writing(struct ptlrpc_request *req, int offset)
1521 struct mds_body *body;
1522 int rc, size = sizeof(struct mds_body);
1525 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1527 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1528 lustre_swab_mds_body);
1530 CERROR("Can't unpack body\n");
1531 req->rq_status = -EFAULT;
1535 rc = lustre_pack_reply(req, 1, &size, NULL);
1537 CERROR("lustre_pack_reply: rc = %d\n", rc);
1538 req->rq_status = rc;