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 the Lustre file system, http://www.lustre.org
11 * Lustre is a trademark of Cluster File Systems, Inc.
13 * You may have signed or agreed to another license before downloading
14 * this software. If so, you are bound by the terms and conditions
15 * of that agreement, and the following does not apply to you. See the
16 * LICENSE file included with this distribution for more information.
18 * If you did not agree to a different license, then this copy of Lustre
19 * is open source software; you can redistribute it and/or modify it
20 * under the terms of version 2 of the GNU General Public License as
21 * published by the Free Software Foundation.
23 * In either case, Lustre is distributed in the hope that it will be
24 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
25 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
26 * license text for more details.
30 # define EXPORT_SYMTAB
32 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/version.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/version.h>
38 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
39 # include <linux/buffer_head.h>
40 # include <linux/workqueue.h>
42 # include <linux/locks.h>
45 #include <obd_class.h>
47 #include <lustre_fsfilt.h>
48 #include <lprocfs_status.h>
50 #include "mds_internal.h"
52 /* Exported function from this file are:
54 * mds_open - called by the intent handler
55 * mds_close - an rpc handling function
56 * mds_pin - an rpc handling function - which will go away
57 * mds_mfd_close - for force closing files when a client dies
61 * MDS file data handling: file data holds a handle for a file opened
65 static void mds_mfd_addref(void *mfdp)
67 struct mds_file_data *mfd = mfdp;
69 atomic_inc(&mfd->mfd_refcount);
70 CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
71 atomic_read(&mfd->mfd_refcount));
74 /* Create a new mds_file_data struct.
75 * One reference for handle+med_open_head list and dropped by mds_mfd_unlink(),
76 * one reference for the caller of this function. */
77 struct mds_file_data *mds_mfd_new(void)
79 struct mds_file_data *mfd;
81 OBD_ALLOC(mfd, sizeof *mfd);
83 CERROR("mds: out of memory\n");
87 atomic_set(&mfd->mfd_refcount, 2);
89 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
90 INIT_LIST_HEAD(&mfd->mfd_list);
91 class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
96 /* Get a new reference on the mfd pointed to by handle, if handle is still
97 * valid. Caller must drop reference with mds_mfd_put(). */
98 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
101 LASSERT(handle != NULL);
102 RETURN(class_handle2object(handle->cookie));
105 /* Drop mfd reference, freeing struct if this is the last one. */
106 static void mds_mfd_put(struct mds_file_data *mfd)
108 CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
109 atomic_read(&mfd->mfd_refcount) - 1);
110 LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
111 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
112 if (atomic_dec_and_test(&mfd->mfd_refcount)) {
113 LASSERT(list_empty(&mfd->mfd_handle.h_link));
114 OBD_FREE(mfd, sizeof *mfd);
118 /* Remove the mfd handle so that it cannot be found by open/close again.
119 * Caller must hold med_open_lock for mfd_list manipulation. */
120 void mds_mfd_unlink(struct mds_file_data *mfd, int decref)
122 class_handle_unhash(&mfd->mfd_handle);
123 list_del_init(&mfd->mfd_list);
128 /* Caller must hold mds->mds_epoch_sem */
129 static int mds_alloc_filterdata(struct inode *inode)
131 LASSERT(inode->i_filterdata == NULL);
132 OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
133 if (inode->i_filterdata == NULL)
135 LASSERT(igrab(inode) == inode);
139 /* Caller must hold mds->mds_epoch_sem */
140 static void mds_free_filterdata(struct inode *inode)
142 LASSERT(inode->i_filterdata != NULL);
143 OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
144 inode->i_filterdata = NULL;
148 /* Write access to a file: executors cause a negative count,
149 * writers a positive count. The semaphore is needed to perform
150 * a check for the sign and then increment or decrement atomically.
152 * This code is closely tied to the allocation of the d_fsdata and the
153 * MDS epoch, so we use the same semaphore for the whole lot.
155 * We could use a different semaphore for each file, if it ever shows
156 * up in a profile, which it won't.
158 * epoch argument is nonzero during recovery */
159 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
164 down(&mds->mds_epoch_sem);
166 if (atomic_read(&inode->i_writecount) < 0) {
167 up(&mds->mds_epoch_sem);
172 if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
173 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
174 MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
175 inode->i_generation);
179 if (inode->i_filterdata == NULL)
180 mds_alloc_filterdata(inode);
181 if (inode->i_filterdata == NULL) {
185 if (epoch > mds->mds_io_epoch)
186 mds->mds_io_epoch = epoch;
189 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
190 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
191 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
194 atomic_inc(&inode->i_writecount);
195 up(&mds->mds_epoch_sem);
199 /* Returns EAGAIN if the client needs to get size and/or cookies and close
200 * again -- which is never true if the file is about to be unlinked. Otherwise
201 * returns the number of remaining writers. */
202 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
203 struct mds_body *body, int unlinking)
208 down(&mds->mds_epoch_sem);
209 atomic_dec(&inode->i_writecount);
210 rc = atomic_read(&inode->i_writecount);
214 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
215 GOTO(out, rc = EAGAIN);
217 mds_free_filterdata(inode);
219 up(&mds->mds_epoch_sem);
223 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
226 down(&mds->mds_epoch_sem);
227 if (atomic_read(&inode->i_writecount) > 0) {
228 up(&mds->mds_epoch_sem);
231 atomic_dec(&inode->i_writecount);
232 up(&mds->mds_epoch_sem);
236 static void mds_allow_write_access(struct inode *inode)
239 atomic_inc(&inode->i_writecount);
242 int mds_query_write_access(struct inode *inode)
245 RETURN(atomic_read(&inode->i_writecount));
248 /* This replaces the VFS dentry_open, it manages mfd and writecount */
249 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
250 struct vfsmount *mnt, int flags,
251 struct ptlrpc_request *req)
253 struct mds_export_data *med = &req->rq_export->exp_mds_data;
254 struct mds_obd *mds = mds_req2mds(req);
255 struct mds_file_data *mfd;
256 struct mds_body *body;
262 CERROR("mds: out of memory\n");
263 GOTO(cleanup_dentry, error = -ENOMEM);
266 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
268 if (flags & FMODE_WRITE) {
269 /* FIXME: in recovery, need to pass old epoch here */
270 error = mds_get_write_access(mds, dentry->d_inode, 0);
272 GOTO(cleanup_mfd, error);
273 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
274 } else if (flags & MDS_FMODE_EXEC) {
275 error = mds_deny_write_access(mds, dentry->d_inode);
277 GOTO(cleanup_mfd, error);
282 /* Mark the file as open to handle open-unlink. */
283 MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode);
284 mds_orphan_open_inc(dentry->d_inode);
285 MDS_UP_WRITE_ORPHAN_SEM(dentry->d_inode);
287 mfd->mfd_mode = flags;
288 mfd->mfd_dentry = dentry;
289 mfd->mfd_xid = req->rq_xid;
291 spin_lock(&med->med_open_lock);
292 list_add(&mfd->mfd_list, &med->med_open_head);
293 spin_unlock(&med->med_open_lock);
295 body->handle.cookie = mfd->mfd_handle.h_cookie;
301 mds_mfd_unlink(mfd, 1);
303 return ERR_PTR(error);
306 /* Must be called with i_mutex held */
307 static int mds_create_objects(struct ptlrpc_request *req, int offset,
308 struct mds_update_record *rec,
309 struct mds_obd *mds, struct obd_device *obd,
310 struct dentry *dchild, void **handle,
313 struct inode *inode = dchild->d_inode;
314 struct obd_trans_info oti = { 0 };
315 struct lov_mds_md *lmm = NULL;
317 struct mds_body *body;
318 struct obd_info oinfo = { { { 0 } } };
322 if (!S_ISREG(inode->i_mode))
324 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
325 !(rec->ur_flags & FMODE_WRITE))
328 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
330 if (body->valid & OBD_MD_FLEASIZE)
333 OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
337 oti.oti_objid = *ids;
340 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
341 if (rec->ur_fid2->id == 0) {
342 DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
346 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
347 lmm_size = rec->ur_eadatalen;
348 lmm = rec->ur_eadata;
352 *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
353 if (IS_ERR(*handle)) {
354 rc = PTR_ERR(*handle);
359 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
361 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
363 CERROR("open replay failed to set md:%d\n", rc);
364 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
366 memcpy(lmm_buf, lmm, lmm_size);
371 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
372 GOTO(out_ids, rc = -ENOMEM);
374 oinfo.oi_oa = obdo_alloc();
375 if (oinfo.oi_oa == NULL)
376 GOTO(out_ids, rc = -ENOMEM);
377 oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */
378 oinfo.oi_oa->o_gid = 0;
379 oinfo.oi_oa->o_mode = S_IFREG | 0600;
380 oinfo.oi_oa->o_id = inode->i_ino;
381 oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS |
382 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
383 oinfo.oi_oa->o_size = 0;
385 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
386 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
388 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
389 /* check if things like lfs setstripe are sending us the ea */
390 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
391 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
393 0, &oinfo.oi_md, rec->ur_eadata);
397 OBD_ALLOC(lmm, mds->mds_max_mdsize);
399 GOTO(out_oa, rc = -ENOMEM);
401 lmm_size = mds->mds_max_mdsize;
402 rc = mds_get_md(obd, dchild->d_parent->d_inode,
405 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
407 0, &oinfo.oi_md, lmm);
408 OBD_FREE(lmm, mds->mds_max_mdsize);
412 rc = obd_create(mds->mds_osc_exp, oinfo.oi_oa,
418 CDEBUG(level, "error creating objects for "
419 "inode %lu: rc = %d\n",
422 CERROR("obd_create returned invalid "
429 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
430 0, &oinfo.oi_md, rec->ur_eadata);
434 oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id;
437 oinfo.oi_oa->o_size = inode->i_size;
438 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE |
439 OBD_MD_FLATIME | OBD_MD_FLMTIME |
440 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
442 /* pack lustre id to OST */
443 oinfo.oi_oa->o_fid = body->fid1.id;
444 oinfo.oi_oa->o_generation = body->fid1.generation;
445 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
447 rc = obd_setattr_rqset(mds->mds_osc_exp, &oinfo, &oti);
449 CERROR("error setting attrs for inode %lu: rc %d\n",
452 CERROR("obd_setattr_async returned bad rc %d\n",
460 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
461 obdo_refresh_inode(inode, oinfo.oi_oa, OBD_MD_FLBLKSZ);
463 LASSERT(oinfo.oi_md && oinfo.oi_md->lsm_object_id);
465 rc = obd_packmd(mds->mds_osc_exp, &lmm, oinfo.oi_md);
467 CERROR("cannot pack lsm, err = %d\n", rc);
471 body->eadatasize = rc;
474 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
475 if (IS_ERR(*handle)) {
476 rc = PTR_ERR(*handle);
481 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
482 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
484 memcpy(lmm_buf, lmm, lmm_size);
485 obd_free_diskmd(mds->mds_osc_exp, &lmm);
487 oti_free_cookies(&oti);
488 obdo_free(oinfo.oi_oa);
491 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
495 obd_free_memmd(mds->mds_osc_exp, &oinfo.oi_md);
499 static void reconstruct_open(struct mds_update_record *rec, int offset,
500 struct ptlrpc_request *req,
501 struct lustre_handle *child_lockh)
503 struct mds_export_data *med = &req->rq_export->exp_mds_data;
504 struct mds_client_data *mcd = med->med_mcd;
505 struct mds_obd *mds = mds_req2mds(req);
506 struct mds_file_data *mfd;
507 struct obd_device *obd = req->rq_export->exp_obd;
508 struct dentry *parent, *dchild;
509 struct ldlm_reply *rep;
510 struct mds_body *body;
516 LASSERT(offset == DLM_INTENT_REC_OFF); /* only called via intent */
517 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
518 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
520 /* copy rc, transno and disp; steal locks */
521 mds_req_from_mcd(req, mcd);
522 intent_set_disposition(rep, le32_to_cpu(mcd->mcd_last_data));
524 /* Only replay if create or open actually happened. */
525 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
527 return; /* error looking up parent or child */
530 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
531 LASSERT(!IS_ERR(parent));
533 dchild = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
534 LASSERT(!IS_ERR(dchild));
536 if (!dchild->d_inode)
537 GOTO(out_dput, 0); /* child not present to open */
539 /* At this point, we know we have a child. We'll send
540 * it back _unless_ it not created and open failed.
542 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
543 !intent_disposition(rep, DISP_OPEN_CREATE) &&
548 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
549 mds_pack_inode2body(body, dchild->d_inode);
550 if (S_ISREG(dchild->d_inode->i_mode)) {
551 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
552 body, dchild->d_inode, 1);
555 LASSERT(rc == req->rq_status);
557 /* If we have LOV EA data, the OST holds size, mtime */
558 if (!(body->valid & OBD_MD_FLEASIZE))
559 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
560 OBD_MD_FLATIME | OBD_MD_FLMTIME);
563 if (!(rec->ur_flags & MDS_OPEN_JOIN_FILE))
564 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
565 body->eadatasize, 0);
567 if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
568 !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
569 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
571 rc = mds_pack_acl(med, dchild->d_inode, req->rq_repmsg,
573 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
574 if (!req->rq_status && rc)
578 /* If we have -EEXIST as the status, and we were asked to create
579 * exclusively, we can tell we failed because the file already existed.
581 if (req->rq_status == -EEXIST &&
582 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
583 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
587 /* If we didn't get as far as trying to open, then some locking thing
588 * probably went wrong, and we'll just bail here.
590 if (!intent_disposition(rep, DISP_OPEN_OPEN))
593 /* If we failed, then we must have failed opening, so don't look for
594 * file descriptor or anything, just give the client the bad news.
600 spin_lock(&med->med_open_lock);
601 list_for_each(t, &med->med_open_head) {
602 mfd = list_entry(t, struct mds_file_data, mfd_list);
603 if (mfd->mfd_xid == req->rq_xid) {
609 spin_unlock(&med->med_open_lock);
611 /* #warning "XXX fixme" bug 2991 */
612 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
613 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
614 * to detect a re-open */
616 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
617 rc = mds_join_file(rec, req, dchild, NULL);
621 mntget(mds->mds_vfsmnt);
622 CERROR("Re-opened file \n");
623 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
624 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
626 CERROR("mds: out of memory\n");
627 GOTO(out_dput, req->rq_status = -ENOMEM);
631 body->handle.cookie = mfd->mfd_handle.h_cookie;
632 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
633 mfd->mfd_handle.h_cookie);
645 /* do NOT or the MAY_*'s, you'll get the weakest */
646 static int accmode(struct inode *inode, int flags)
650 /* Sadly, NFSD reopens a file repeatedly during operation, so the
651 * "acc_mode = 0" allowance for newly-created files isn't honoured.
652 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
653 * owner can write to a file even if it is marked readonly to hide
654 * its brokenness. (bug 5781) */
655 if (flags & MDS_OPEN_OWNEROVERRIDE && inode->i_uid == current->fsuid)
658 if (flags & FMODE_READ)
660 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
662 if (flags & MDS_FMODE_EXEC)
667 /* Handles object creation, actual opening, and I/O epoch */
668 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
669 struct mds_body *body, int flags, void **handle,
670 struct mds_update_record *rec,struct ldlm_reply *rep,
671 struct lustre_handle *lockh)
673 struct mds_obd *mds = mds_req2mds(req);
674 struct obd_device *obd = req->rq_export->exp_obd;
675 struct mds_file_data *mfd = NULL;
676 obd_id *ids = NULL; /* object IDs created */
680 /* atomically create objects if necessary */
681 LOCK_INODE_MUTEX(dchild->d_inode);
683 if (S_ISREG(dchild->d_inode->i_mode) &&
684 !(body->valid & OBD_MD_FLEASIZE)) {
685 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
686 body, dchild->d_inode, 0);
688 UNLOCK_INODE_MUTEX(dchild->d_inode);
693 if ((body->valid & OBD_MD_FLEASIZE) &&
694 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
695 UNLOCK_INODE_MUTEX(dchild->d_inode);
698 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
699 UNLOCK_INODE_MUTEX(dchild->d_inode);
700 rc = mds_join_file(rec, req, dchild, lockh);
703 LOCK_INODE_MUTEX(dchild->d_inode);
705 if (!(body->valid & OBD_MD_FLEASIZE) &&
706 !(body->valid & OBD_MD_FLMODEASIZE)) {
707 /* no EA: create objects */
708 rc = mds_create_objects(req, DLM_REPLY_REC_OFF + 1, rec,
709 mds, obd, dchild, handle, &ids);
711 CERROR("mds_create_objects: rc = %d\n", rc);
712 UNLOCK_INODE_MUTEX(dchild->d_inode);
717 /* If the inode has no EA data, then MDS holds size, mtime */
718 if (S_ISREG(dchild->d_inode->i_mode) &&
719 !(body->valid & OBD_MD_FLEASIZE)) {
720 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
721 OBD_MD_FLATIME | OBD_MD_FLMTIME);
723 UNLOCK_INODE_MUTEX(dchild->d_inode);
725 if (rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE))
726 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
727 body->eadatasize, 0);
729 if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
730 rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
731 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
733 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
734 dchild->d_inode, req->rq_repmsg,
736 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
741 intent_set_disposition(rep, DISP_OPEN_OPEN);
742 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
744 RETURN(PTR_ERR(mfd));
746 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
747 mfd->mfd_handle.h_cookie);
750 mds_lov_update_objids(obd, ids);
751 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
753 if (rc) /* coverity[deadcode] */
754 mds_mfd_unlink(mfd, 1);
760 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
761 struct mds_body *body, int flags,
762 struct mds_update_record *rec,struct ldlm_reply *rep)
764 struct mds_obd *mds = mds_req2mds(req);
765 struct dentry *dchild;
766 char fidname[LL_FID_NAMELEN];
771 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
772 dchild = ll_lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
773 if (IS_ERR(dchild)) {
774 rc = PTR_ERR(dchild);
775 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
779 if (dchild->d_inode != NULL) {
780 mds_inode_set_orphan(dchild->d_inode);
781 CWARN("Orphan %s found and opened in PENDING directory\n",
786 /* We didn't find it in PENDING so it isn't an orphan. See
787 * if it was a regular inode that was previously created. */
788 dchild = mds_fid2dentry(mds, fid, NULL);
790 RETURN(PTR_ERR(dchild));
793 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
794 mds_pack_inode2body(body, dchild->d_inode);
795 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
796 intent_set_disposition(rep, DISP_LOOKUP_POS);
798 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep, NULL);
799 rc = mds_finish_transno(mds, dchild->d_inode, handle,
800 req, rc, rep ? rep->lock_policy_res1 : 0);
801 /* XXX what do we do here if mds_finish_transno itself failed? */
807 int mds_pin(struct ptlrpc_request *req, int offset)
809 struct obd_device *obd = req->rq_export->exp_obd;
810 struct mds_body *reqbody, *repbody;
811 struct lvfs_run_ctxt saved;
812 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
815 reqbody = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*reqbody));
817 rc = lustre_pack_reply(req, 2, size, NULL);
820 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
823 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
824 rc = mds_open_by_fid(req, &reqbody->fid1, repbody, reqbody->flags, NULL,
826 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
831 /* Get an internal lock on the inode number (but not generation) to sync
832 * new inode creation with inode unlink (bug 2029). If child_lockh is NULL
833 * we just get the lock as a barrier to wait for other holders of this lock,
834 * and drop it right away again. */
835 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
836 struct lustre_handle *child_lockh)
838 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
839 struct lustre_handle lockh;
840 int lock_flags = LDLM_FL_ATOMIC_CB;
843 if (child_lockh == NULL)
844 child_lockh = &lockh;
846 rc = ldlm_cli_enqueue_local(obd->obd_namespace, child_res_id,
847 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
848 ldlm_blocking_ast, ldlm_completion_ast,
849 NULL, NULL, 0, NULL, child_lockh);
851 CERROR("ldlm_cli_enqueue_local: %d\n", rc);
852 else if (child_lockh == &lockh)
853 ldlm_lock_decref(child_lockh, LCK_EX);
858 int mds_open(struct mds_update_record *rec, int offset,
859 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
861 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
862 struct obd_device *obd = req->rq_export->exp_obd;
863 struct mds_obd *mds = mds_req2mds(req);
864 struct ldlm_reply *rep = NULL;
865 struct mds_body *body = NULL;
866 struct dentry *dchild = NULL, *dparent = NULL;
867 struct mds_export_data *med;
868 struct lustre_handle parent_lockh;
869 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
870 int parent_mode = LCK_CR;
872 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
873 unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
874 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
875 int child_mode = LCK_CR;
876 /* Always returning LOOKUP lock if open succesful to guard
878 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
879 struct ldlm_res_id child_res_id = { .name = {0}};
883 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
885 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
886 (obd_timeout + 1) / 4);
888 CLASSERT(MAXQUOTAS < 4);
889 if (offset == DLM_INTENT_REC_OFF) { /* intent */
890 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
892 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
894 } else if (offset == REQ_REC_OFF) { /* non-intent reint */
895 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
897 LBUG(); /* XXX: not supported yet? */
903 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
905 /* Step 0: If we are passed a fid, then we assume the client already
906 * opened this file and is only replaying the RPC, so we open the
907 * inode by fid (at some large expense in security). */
908 /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */
909 if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
910 (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&&
911 !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
912 if (rec->ur_fid2->id == 0) {
913 struct ldlm_lock *lock = ldlm_handle2lock(child_lockh);
915 LDLM_ERROR(lock, "fid2 not set on open replay");
918 DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
922 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
925 if (req->rq_export->exp_libclient &&
926 rec->ur_flags & MDS_OPEN_HAS_EA)
932 /* We didn't find the correct inode on disk either, so we
933 * need to re-create it via a regular replay. */
934 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
935 DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay");
938 } else if (rec->ur_fid2->id) {
939 DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay",
940 rec->ur_fid2->id, rec->ur_fid2->generation);
944 /* If we got here, we must be called via intent */
945 LASSERT(offset == DLM_INTENT_REC_OFF);
947 med = &req->rq_export->exp_mds_data;
948 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
949 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
953 /* Step 1: Find and lock the parent */
954 if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE))
955 parent_mode = LCK_EX;
956 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
957 &parent_lockh, rec->ur_name,
959 MDS_INODELOCK_UPDATE);
960 if (IS_ERR(dparent)) {
961 rc = PTR_ERR(dparent);
963 CERROR("parent "LPU64"/%u lookup error %d\n",
964 rec->ur_fid1->id, rec->ur_fid1->generation, rc);
966 /* Just cannot find parent - make it look like
967 * usual negative lookup to avoid extra MDS RPC */
968 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
969 intent_set_disposition(rep, DISP_LOOKUP_NEG);
973 LASSERT(dparent->d_inode != NULL);
975 cleanup_phase = 1; /* parent dentry and lock */
977 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
978 dchild = dget(dparent);
979 cleanup_phase = 2; /* child dentry */
980 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
981 GOTO(found_child, rc);
984 /* Step 2: Lookup the child */
985 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
986 if (IS_ERR(dchild)) {
987 rc = PTR_ERR(dchild);
988 dchild = NULL; /* don't confuse mds_finish_transno() below */
992 cleanup_phase = 2; /* child dentry */
994 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
996 intent_set_disposition(rep, DISP_LOOKUP_POS);
998 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1000 /*Step 3: If the child was negative, and we're supposed to, create it.*/
1001 if (dchild->d_inode == NULL) {
1002 unsigned long ino = rec->ur_fid2->id;
1004 struct inode *inode;
1006 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1007 /* It's negative and we weren't supposed to create it */
1008 GOTO(cleanup, rc = -ENOENT);
1011 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1012 GOTO(cleanup, rc = -EROFS);
1014 intent_set_disposition(rep, DISP_OPEN_CREATE);
1015 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1017 if (IS_ERR(handle)) {
1018 rc = PTR_ERR(handle);
1022 dchild->d_fsdata = (void *) &dp;
1026 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1027 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1028 dchild->d_fsdata = NULL;
1031 CDEBUG(D_INODE, "error during create: %d\n", rc);
1034 inode = dchild->d_inode;
1036 LASSERT(ino == inode->i_ino);
1037 /* Written as part of setattr */
1038 inode->i_generation = rec->ur_fid2->generation;
1039 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1040 inode->i_ino, inode->i_generation);
1044 LTIME_S(iattr.ia_atime) = rec->ur_time;
1045 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1046 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1048 iattr.ia_uid = current->fsuid; /* set by push_ctxt already */
1049 if (dparent->d_inode->i_mode & S_ISGID)
1050 iattr.ia_gid = dparent->d_inode->i_gid;
1052 iattr.ia_gid = current->fsgid;
1054 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1055 ATTR_MTIME | ATTR_CTIME;
1057 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1059 CERROR("error on child setattr: rc = %d\n", rc);
1061 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1063 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1065 CERROR("error on parent setattr: rc = %d\n", rc);
1067 rc = fsfilt_commit(obd, dchild->d_inode, handle, 0);
1069 acc_mode = 0; /* Don't check for permissions */
1071 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1074 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1075 "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len,
1076 dchild->d_name.name, dchild, dchild->d_inode,
1077 dchild->d_inode->i_ino, dchild->d_inode->i_generation);
1080 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
1081 mds_pack_inode2body(body, dchild->d_inode);
1083 if (S_ISREG(dchild->d_inode->i_mode)) {
1084 /* Check permissions etc */
1085 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1089 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) &&
1090 (acc_mode & MAY_WRITE))
1091 GOTO(cleanup, rc = -EROFS);
1093 /* An append-only file must be opened in append mode for
1095 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1096 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1097 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1098 GOTO(cleanup, rc = -EPERM);
1101 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1102 (rec->ur_flags & MDS_OPEN_EXCL)) {
1103 /* File already exists, we didn't just create it, and we
1104 * were passed O_EXCL; err-or. */
1105 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1108 /* if we are following a symlink, don't open */
1109 if (S_ISLNK(dchild->d_inode->i_mode))
1110 GOTO(cleanup_no_trans, rc = 0);
1112 if (S_ISDIR(dchild->d_inode->i_mode)) {
1113 if (rec->ur_flags & MDS_OPEN_CREAT ||
1114 rec->ur_flags & FMODE_WRITE) {
1115 /* we are trying to create or write a exist dir */
1116 GOTO(cleanup, rc = -EISDIR);
1118 if (rec->ur_flags & MDS_FMODE_EXEC) {
1119 /* we are trying to exec a directory */
1120 GOTO(cleanup, rc = -EACCES);
1122 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1123 intent_set_disposition(rep, DISP_OPEN_OPEN);
1124 GOTO(cleanup, rc = -EACCES);
1126 } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) {
1127 GOTO(cleanup, rc = -ENOTDIR);
1130 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1131 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1132 GOTO(cleanup, rc = -EAGAIN);
1135 /* Obtain OPEN lock as well */
1136 policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1138 /* We cannot use acc_mode here, because it is zeroed in case of
1139 creating a file, so we get wrong lockmode */
1140 if (accmode(dchild->d_inode, rec->ur_flags) & MAY_WRITE)
1141 child_mode = LCK_CW;
1142 else if (accmode(dchild->d_inode, rec->ur_flags) & MAY_EXEC)
1143 child_mode = LCK_PR;
1145 child_mode = LCK_CR;
1147 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1148 (rec->ur_flags & MDS_OPEN_LOCK)) {
1149 /* In case of replay we do not get a lock assuming that the
1150 caller has it already */
1151 child_res_id.name[0] = dchild->d_inode->i_ino;
1152 child_res_id.name[1] = dchild->d_inode->i_generation;
1154 rc = ldlm_cli_enqueue_local(obd->obd_namespace, child_res_id,
1155 LDLM_IBITS, &policy, child_mode,
1156 &lock_flags, ldlm_blocking_ast,
1157 ldlm_completion_ast, NULL, NULL,
1158 0, NULL, child_lockh);
1162 /* Let mds_intent_policy know that we have a lock to return */
1163 intent_set_disposition(rep, DISP_OPEN_LOCK);
1167 if (!S_ISREG(dchild->d_inode->i_mode) &&
1168 !S_ISDIR(dchild->d_inode->i_mode) &&
1169 (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) {
1170 /* If client supports this, do not return open handle for
1171 * special device nodes */
1172 GOTO(cleanup_no_trans, rc = 0);
1175 /* Step 5: mds_open it */
1176 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1177 rep, &parent_lockh);
1181 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1182 req, rc, rep ? rep->lock_policy_res1 : 0);
1185 switch (cleanup_phase) {
1188 /* It is safe to leave IT_OPEN_LOCK set, if rc is not 0,
1189 * mds_intent_policy won't try to return any locks */
1190 ldlm_lock_decref(child_lockh, child_mode);
1192 if (rc && created) {
1193 int err = vfs_unlink(dparent->d_inode, dchild);
1195 CERROR("unlink(%.*s) in error path: %d\n",
1196 dchild->d_name.len, dchild->d_name.name,
1199 } else if (created) {
1200 mds_lock_new_child(obd, dchild->d_inode, NULL);
1201 /* save uid/gid for quota acquire/release */
1202 qpids[USRQUOTA] = dparent->d_inode->i_uid;
1203 qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1207 if (dparent == NULL)
1212 ldlm_lock_decref(&parent_lockh, parent_mode);
1214 ptlrpc_save_lock(req, &parent_lockh, parent_mode);
1216 /* trigger dqacq on the owner of child and parent */
1217 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1222 /* Close a "file descriptor" and possibly unlink an orphan from the
1223 * PENDING directory. Caller must hold child->i_mutex, this drops it.
1225 * If we are being called from mds_disconnect() because the client has
1226 * disappeared, then req == NULL and we do not update last_rcvd because
1227 * there is nothing that could be recovered by the client at this stage
1228 * (it will not even _have_ an entry in last_rcvd anymore).
1230 * Returns EAGAIN if the client needs to get more data and re-close. */
1231 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1232 struct obd_device *obd, struct mds_file_data *mfd,
1233 int unlink_orphan, struct lov_mds_md *lmm, int lmm_size,
1234 struct llog_cookie *logcookies, int cookies_size,
1237 struct inode *inode = mfd->mfd_dentry->d_inode;
1238 char fidname[LL_FID_NAMELEN];
1239 int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1240 struct dentry *pending_child = NULL;
1241 struct mds_obd *mds = &obd->u.mds;
1242 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1243 void *handle = NULL;
1244 struct mds_body *request_body = NULL, *reply_body = NULL;
1245 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1246 struct iattr iattr = { 0 };
1249 if (req && req->rq_reqmsg != NULL)
1250 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1251 sizeof(*request_body));
1252 if (req && req->rq_repmsg != NULL)
1253 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1254 sizeof(*reply_body));
1256 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1258 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname,
1259 inode->i_nlink, mds_orphan_open_count(inode));
1261 last_orphan = mds_orphan_open_dec_test(inode) &&
1262 mds_inode_is_orphan(inode);
1263 MDS_UP_WRITE_ORPHAN_SEM(inode);
1265 /* this is half of the actual "close" */
1266 if (mfd->mfd_mode & FMODE_WRITE) {
1267 rc = mds_put_write_access(mds, inode, request_body,
1268 last_orphan && unlink_orphan);
1269 } else if (mfd->mfd_mode & MDS_FMODE_EXEC) {
1270 mds_allow_write_access(inode);
1273 if (last_orphan && unlink_orphan) {
1274 int stripe_count = 0;
1275 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1277 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1279 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1280 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1281 CERROR("found \"orphan\" %s %s with link count %d\n",
1282 S_ISREG(inode->i_mode) ? "file" : "dir",
1283 fidname, inode->i_nlink);
1285 /* Sadly, there is no easy way to save pending_child from
1286 * mds_reint_unlink() into mfd, so we need to re-lookup,
1287 * but normally it will still be in the dcache. */
1288 LOCK_INODE_MUTEX(pending_dir);
1289 cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */
1290 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1292 if (IS_ERR(pending_child))
1293 GOTO(cleanup, rc = PTR_ERR(pending_child));
1294 LASSERT(pending_child->d_inode != NULL);
1296 cleanup_phase = 2; /* dput(pending_child) when finished */
1297 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1298 rc = vfs_rmdir(pending_dir, pending_child);
1300 CERROR("error unlinking orphan dir %s: rc %d\n",
1306 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1309 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1310 NULL, stripe_count);
1311 if (IS_ERR(handle)) {
1312 rc = PTR_ERR(handle);
1316 if (lmm != NULL && (*valid & OBD_MD_FLEASIZE) &&
1317 mds_log_op_unlink(obd, lmm, lmm_size,
1318 logcookies, cookies_size) > 0) {
1319 *valid |= OBD_MD_FLCOOKIE;
1324 pending_child->d_fsdata = (void *) &dp;
1325 rc = vfs_unlink(pending_dir, pending_child);
1327 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1329 goto out; /* Don't bother updating attrs on unlinked inode */
1333 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1334 /* Update the on-disk attributes if this was the last write
1335 * close, and all information was provided (i.e., rc == 0)
1337 * XXX this should probably be abstracted with mds_reint_setattr
1340 if (request_body->valid & OBD_MD_FLMTIME &&
1341 LTIME_S(iattr.ia_mtime) > LTIME_S(inode->i_mtime)) {
1342 LTIME_S(iattr.ia_mtime) = request_body->mtime;
1343 iattr.ia_valid |= ATTR_MTIME;
1345 if (request_body->valid & OBD_MD_FLCTIME &&
1346 LTIME_S(iattr.ia_ctime) > LTIME_S(inode->i_ctime)) {
1347 LTIME_S(iattr.ia_ctime) = request_body->ctime;
1348 iattr.ia_valid |= ATTR_CTIME;
1351 /* XXX can't set block count with fsfilt_setattr (!) */
1352 if (request_body->valid & OBD_MD_FLSIZE) {
1353 iattr.ia_valid |= ATTR_SIZE;
1354 iattr.ia_size = request_body->size;
1356 /* iattr.ia_blocks = request_body->blocks */
1360 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1361 /* Only start a transaction to write out only the atime if
1362 * it is more out-of-date than the specified limit. If we
1363 * are already going to write out the atime then do it anyway.
1365 LTIME_S(iattr.ia_atime) = request_body->atime;
1366 if ((LTIME_S(iattr.ia_atime) >
1367 LTIME_S(inode->i_atime) + mds->mds_atime_diff) ||
1368 (iattr.ia_valid != 0 &&
1369 LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime)))
1370 iattr.ia_valid |= ATTR_ATIME;
1373 if (iattr.ia_valid != 0) {
1374 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1375 if (IS_ERR(handle)) {
1376 rc = PTR_ERR(handle);
1380 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1382 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1385 /* If other clients have this file open for write, rc will be > 0 */
1388 l_dput(mfd->mfd_dentry);
1392 if (req != NULL && reply_body != NULL) {
1393 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1394 } else if (handle) {
1395 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1397 CERROR("error committing close: %d\n", err);
1403 switch (cleanup_phase) {
1405 dput(pending_child);
1407 UNLOCK_INODE_MUTEX(pending_dir);
1412 int mds_close(struct ptlrpc_request *req, int offset)
1414 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1415 struct obd_device *obd = req->rq_export->exp_obd;
1416 struct mds_body *body;
1417 struct mds_file_data *mfd;
1418 struct lvfs_run_ctxt saved;
1419 struct inode *inode;
1420 int rc, repsize[4] = { sizeof(struct ptlrpc_body),
1421 sizeof(struct mds_body),
1422 obd->u.mds.mds_max_mdsize,
1423 obd->u.mds.mds_max_cookiesize };
1424 struct mds_body *reply_body;
1425 struct lov_mds_md *lmm;
1427 struct llog_cookie *logcookies;
1431 rc = lustre_pack_reply(req, 4, repsize, NULL);
1433 CERROR("lustre_pack_reply: rc = %d\n", rc);
1434 req->rq_status = rc;
1435 /* continue on to drop local open even if we can't send reply */
1437 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1440 CDEBUG(D_HA, "close req->rep_len %d mdsize %d cookiesize %d\n",
1442 obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize);
1443 mds_counter_incr(req->rq_export, LPROC_MDS_CLOSE);
1445 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1446 lustre_swab_mds_body);
1448 CERROR("Can't unpack body\n");
1449 req->rq_status = -EFAULT;
1453 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1454 /* do some stuff */ ;
1456 spin_lock(&med->med_open_lock);
1457 mfd = mds_handle2mfd(&body->handle);
1459 spin_unlock(&med->med_open_lock);
1460 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1461 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1462 req->rq_status = -ESTALE;
1465 /* Remove mfd handle so it can't be found again. We consume mfd_list
1466 * reference here, but still have mds_handle2mfd ref until mfd_close. */
1467 mds_mfd_unlink(mfd, 1);
1468 spin_unlock(&med->med_open_lock);
1470 inode = mfd->mfd_dentry->d_inode;
1471 /* child orphan sem protects orphan_dec_test && is_orphan race */
1472 MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */
1473 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1474 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1476 LASSERT(body != NULL);
1478 mds_pack_inode2fid(&body->fid1, inode);
1479 mds_pack_inode2body(body, inode);
1480 mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode,
1484 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1485 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body));
1486 lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, 0);
1487 lmm_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 1),
1488 logcookies = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 2, 0);
1489 cookies_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 2);
1490 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1,
1491 lmm, lmm_size, logcookies, cookies_size,
1492 &reply_body->valid);
1493 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1495 mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1);
1496 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1497 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1498 req->rq_status = -ENOMEM;
1505 int mds_done_writing(struct ptlrpc_request *req, int offset)
1507 struct mds_body *body;
1508 int rc, size[2] = { sizeof(struct ptlrpc_body),
1509 sizeof(struct mds_body) };
1512 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1514 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1515 lustre_swab_mds_body);
1517 CERROR("Can't unpack body\n");
1518 req->rq_status = -EFAULT;
1522 rc = lustre_pack_reply(req, 2, size, NULL);
1524 CERROR("lustre_pack_reply: rc = %d\n", rc);
1525 req->rq_status = rc;