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 #include <linux/buffer_head.h>
39 #include <linux/workqueue.h>
41 #include <obd_class.h>
43 #include <lustre_fsfilt.h>
44 #include <lprocfs_status.h>
46 #include "mds_internal.h"
48 /* Exported function from this file are:
50 * mds_open - called by the intent handler
51 * mds_close - an rpc handling function
52 * mds_pin - an rpc handling function - which will go away
53 * mds_mfd_close - for force closing files when a client dies
57 * MDS file data handling: file data holds a handle for a file opened
61 static void mds_mfd_addref(void *mfdp)
63 struct mds_file_data *mfd = mfdp;
65 atomic_inc(&mfd->mfd_refcount);
66 CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
67 atomic_read(&mfd->mfd_refcount));
70 /* Create a new mds_file_data struct.
71 * One reference for handle+med_open_head list and dropped by mds_mfd_unlink(),
72 * one reference for the caller of this function. */
73 struct mds_file_data *mds_mfd_new(void)
75 struct mds_file_data *mfd;
77 OBD_ALLOC(mfd, sizeof *mfd);
79 CERROR("mds: out of memory\n");
83 atomic_set(&mfd->mfd_refcount, 2);
85 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
86 INIT_LIST_HEAD(&mfd->mfd_list);
87 class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
92 /* Get a new reference on the mfd pointed to by handle, if handle is still
93 * valid. Caller must drop reference with mds_mfd_put(). */
94 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
97 LASSERT(handle != NULL);
98 RETURN(class_handle2object(handle->cookie));
101 /* Drop mfd reference, freeing struct if this is the last one. */
102 static void mds_mfd_put(struct mds_file_data *mfd)
104 CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
105 atomic_read(&mfd->mfd_refcount) - 1);
106 LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
107 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
108 if (atomic_dec_and_test(&mfd->mfd_refcount)) {
109 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
113 /* Remove the mfd handle so that it cannot be found by open/close again.
114 * Caller must hold med_open_lock for mfd_list manipulation. */
115 void mds_mfd_unlink(struct mds_file_data *mfd, int decref)
117 class_handle_unhash(&mfd->mfd_handle);
118 list_del_init(&mfd->mfd_list);
123 /* Caller must hold mds->mds_epoch_sem */
124 static int mds_alloc_filterdata(struct inode *inode)
126 LASSERT(inode->i_filterdata == NULL);
127 OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
128 if (inode->i_filterdata == NULL)
130 LASSERT(igrab(inode) == inode);
134 /* Caller must hold mds->mds_epoch_sem */
135 static void mds_free_filterdata(struct inode *inode)
137 LASSERT(inode->i_filterdata != NULL);
138 OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
139 inode->i_filterdata = NULL;
143 /* Write access to a file: executors cause a negative count,
144 * writers a positive count. The semaphore is needed to perform
145 * a check for the sign and then increment or decrement atomically.
147 * This code is closely tied to the allocation of the d_fsdata and the
148 * MDS epoch, so we use the same semaphore for the whole lot.
150 * We could use a different semaphore for each file, if it ever shows
151 * up in a profile, which it won't.
153 * epoch argument is nonzero during recovery */
154 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
159 down(&mds->mds_epoch_sem);
161 if (atomic_read(&inode->i_writecount) < 0) {
162 up(&mds->mds_epoch_sem);
167 if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
168 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
169 MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
170 inode->i_generation);
174 if (inode->i_filterdata == NULL)
175 mds_alloc_filterdata(inode);
176 if (inode->i_filterdata == NULL) {
180 if (epoch > mds->mds_io_epoch)
181 mds->mds_io_epoch = epoch;
184 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
185 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
186 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
189 atomic_inc(&inode->i_writecount);
190 up(&mds->mds_epoch_sem);
194 /* Returns EAGAIN if the client needs to get size and/or cookies and close
195 * again -- which is never true if the file is about to be unlinked. Otherwise
196 * returns the number of remaining writers. */
197 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
198 struct mds_body *body, int unlinking)
203 down(&mds->mds_epoch_sem);
204 atomic_dec(&inode->i_writecount);
205 rc = atomic_read(&inode->i_writecount);
209 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
210 GOTO(out, rc = EAGAIN);
212 mds_free_filterdata(inode);
214 up(&mds->mds_epoch_sem);
218 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
221 down(&mds->mds_epoch_sem);
222 if (atomic_read(&inode->i_writecount) > 0) {
223 up(&mds->mds_epoch_sem);
226 atomic_dec(&inode->i_writecount);
227 up(&mds->mds_epoch_sem);
231 static void mds_allow_write_access(struct inode *inode)
234 atomic_inc(&inode->i_writecount);
237 int mds_query_write_access(struct inode *inode)
240 RETURN(atomic_read(&inode->i_writecount));
243 /* This replaces the VFS dentry_open, it manages mfd and writecount */
244 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
245 struct vfsmount *mnt, int flags,
246 struct ptlrpc_request *req)
248 struct mds_export_data *med = &req->rq_export->exp_mds_data;
249 struct mds_obd *mds = mds_req2mds(req);
250 struct mds_file_data *mfd;
251 struct mds_body *body;
257 CERROR("mds: out of memory\n");
258 GOTO(cleanup_dentry, error = -ENOMEM);
261 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
263 if (flags & FMODE_WRITE) {
264 /* FIXME: in recovery, need to pass old epoch here */
265 error = mds_get_write_access(mds, dentry->d_inode, 0);
267 GOTO(cleanup_mfd, error);
268 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
269 } else if (flags & MDS_FMODE_EXEC) {
270 error = mds_deny_write_access(mds, dentry->d_inode);
272 GOTO(cleanup_mfd, error);
277 /* Mark the file as open to handle open-unlink. */
278 MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode);
279 mds_orphan_open_inc(dentry->d_inode);
280 MDS_UP_WRITE_ORPHAN_SEM(dentry->d_inode);
282 mfd->mfd_mode = flags;
283 mfd->mfd_dentry = dentry;
284 mfd->mfd_xid = req->rq_xid;
286 spin_lock(&med->med_open_lock);
287 list_add(&mfd->mfd_list, &med->med_open_head);
288 spin_unlock(&med->med_open_lock);
290 body->handle.cookie = mfd->mfd_handle.h_cookie;
296 mds_mfd_unlink(mfd, 1);
298 return ERR_PTR(error);
301 /* Must be called with i_mutex held */
302 static int mds_create_objects(struct ptlrpc_request *req, int offset,
303 struct mds_update_record *rec,
304 struct mds_obd *mds, struct obd_device *obd,
305 struct dentry *dchild, void **handle,
308 struct inode *inode = dchild->d_inode;
309 struct obd_trans_info oti = { 0 };
310 struct lov_mds_md *lmm = NULL;
312 struct mds_body *body;
313 struct obd_info oinfo = { { { 0 } } };
317 if (!S_ISREG(inode->i_mode))
319 if (!md_should_create(rec->ur_flags))
322 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
324 if (body->valid & OBD_MD_FLEASIZE)
327 OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
331 oti.oti_objid = *ids;
334 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
335 if (rec->ur_fid2->id == 0) {
336 DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
340 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
341 lmm_size = rec->ur_eadatalen;
342 lmm = rec->ur_eadata;
346 *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
347 if (IS_ERR(*handle)) {
348 rc = PTR_ERR(*handle);
353 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
355 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
357 CERROR("open replay failed to set md:%d\n", rc);
358 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
360 memcpy(lmm_buf, lmm, lmm_size);
365 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
366 GOTO(out_ids, rc = -ENOMEM);
368 OBDO_ALLOC(oinfo.oi_oa);
369 if (oinfo.oi_oa == NULL)
370 GOTO(out_ids, rc = -ENOMEM);
371 oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */
372 oinfo.oi_oa->o_gid = 0;
373 oinfo.oi_oa->o_mode = S_IFREG | 0600;
374 oinfo.oi_oa->o_id = inode->i_ino;
375 oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS |
376 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
377 oinfo.oi_oa->o_size = 0;
379 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
380 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
382 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
383 /* check if things like lfs setstripe are sending us the ea */
384 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
385 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
387 0, &oinfo.oi_md, rec->ur_eadata);
391 OBD_ALLOC(lmm, mds->mds_max_mdsize);
393 GOTO(out_oa, rc = -ENOMEM);
395 lmm_size = mds->mds_max_mdsize;
396 rc = mds_get_md(obd, dchild->d_parent->d_inode,
399 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
401 0, &oinfo.oi_md, lmm);
402 OBD_FREE(lmm, mds->mds_max_mdsize);
406 rc = obd_create(mds->mds_osc_exp, oinfo.oi_oa,
412 CDEBUG(level, "error creating objects for "
413 "inode %lu: rc = %d\n",
416 CERROR("obd_create returned invalid "
423 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
424 0, &oinfo.oi_md, rec->ur_eadata);
428 oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id;
429 oinfo.oi_md->lsm_object_gr = oinfo.oi_oa->o_gr;
431 if (i_size_read(inode)) {
432 oinfo.oi_oa->o_size = i_size_read(inode);
433 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE |
434 OBD_MD_FLATIME | OBD_MD_FLMTIME |
435 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
437 /* pack lustre id to OST */
438 oinfo.oi_oa->o_fid = body->fid1.id;
439 oinfo.oi_oa->o_generation = body->fid1.generation;
440 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
442 rc = obd_setattr_rqset(mds->mds_osc_exp, &oinfo, &oti);
444 CERROR("error setting attrs for inode %lu: rc %d\n",
447 CERROR("obd_setattr_async returned bad rc %d\n",
455 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
456 obdo_refresh_inode(inode, oinfo.oi_oa, OBD_MD_FLBLKSZ);
458 LASSERT(oinfo.oi_md && oinfo.oi_md->lsm_object_id);
460 rc = obd_packmd(mds->mds_osc_exp, &lmm, oinfo.oi_md);
462 CERROR("cannot pack lsm, err = %d\n", rc);
466 body->eadatasize = rc;
469 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
470 if (IS_ERR(*handle)) {
471 rc = PTR_ERR(*handle);
473 GOTO(free_diskmd, rc);
476 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
477 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
479 memcpy(lmm_buf, lmm, lmm_size);
481 obd_free_diskmd(mds->mds_osc_exp, &lmm);
483 oti_free_cookies(&oti);
484 OBDO_FREE(oinfo.oi_oa);
487 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
491 obd_free_memmd(mds->mds_osc_exp, &oinfo.oi_md);
495 static void reconstruct_open(struct mds_update_record *rec, int offset,
496 struct ptlrpc_request *req,
497 struct lustre_handle *child_lockh)
499 struct mds_export_data *med = &req->rq_export->exp_mds_data;
500 struct mds_client_data *mcd = med->med_mcd;
501 struct mds_obd *mds = mds_req2mds(req);
502 struct mds_file_data *mfd;
503 struct obd_device *obd = req->rq_export->exp_obd;
504 struct dentry *parent, *dchild;
505 struct ldlm_reply *rep;
506 struct mds_body *body;
512 LASSERT(offset == DLM_INTENT_REC_OFF); /* only called via intent */
513 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
514 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
516 /* copy rc, transno and disp; steal locks */
517 mds_req_from_mcd(req, mcd);
518 intent_set_disposition(rep, le32_to_cpu(mcd->mcd_last_data));
520 /* Only replay if create or open actually happened. */
521 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
523 return; /* error looking up parent or child */
526 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
527 LASSERT(!IS_ERR(parent));
529 dchild = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
530 LASSERT(!IS_ERR(dchild));
532 if (!dchild->d_inode)
533 GOTO(out_dput, 0); /* child not present to open */
535 /* At this point, we know we have a child. We'll send
536 * it back _unless_ it not created and open failed.
538 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
539 !intent_disposition(rep, DISP_OPEN_CREATE) &&
544 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
545 mds_pack_inode2body(body, dchild->d_inode);
546 if (S_ISREG(dchild->d_inode->i_mode)) {
547 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
548 body, dchild->d_inode, 1);
551 LASSERT(rc == req->rq_status);
553 /* If we have LOV EA data, the OST holds size, mtime */
554 if (!(body->valid & OBD_MD_FLEASIZE))
555 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
556 OBD_MD_FLATIME | OBD_MD_FLMTIME);
559 if (!(rec->ur_flags & MDS_OPEN_JOIN_FILE))
560 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
561 body->eadatasize, 0);
563 if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
564 !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
565 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
567 rc = mds_pack_acl(med, dchild->d_inode, req->rq_repmsg,
569 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
570 if (!req->rq_status && rc)
574 /* If we have -EEXIST as the status, and we were asked to create
575 * exclusively, we can tell we failed because the file already existed.
577 if (req->rq_status == -EEXIST &&
578 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
579 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
583 /* If we didn't get as far as trying to open, then some locking thing
584 * probably went wrong, and we'll just bail here.
586 if (!intent_disposition(rep, DISP_OPEN_OPEN))
589 /* If we failed, then we must have failed opening, so don't look for
590 * file descriptor or anything, just give the client the bad news.
596 spin_lock(&med->med_open_lock);
597 list_for_each(t, &med->med_open_head) {
598 mfd = list_entry(t, struct mds_file_data, mfd_list);
599 if (mfd->mfd_xid == req->rq_xid) {
605 spin_unlock(&med->med_open_lock);
607 /* #warning "XXX fixme" bug 2991 */
608 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
609 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
610 * to detect a re-open */
612 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
613 rc = mds_join_file(rec, req, dchild, NULL);
617 mntget(mds->mds_vfsmnt);
618 CERROR("Re-opened file \n");
619 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
620 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
622 CERROR("mds: out of memory\n");
623 GOTO(out_dput, req->rq_status = -ENOMEM);
627 body->handle.cookie = mfd->mfd_handle.h_cookie;
628 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
629 mfd->mfd_handle.h_cookie);
641 /* do NOT or the MAY_*'s, you'll get the weakest */
642 static int accmode(struct inode *inode, int flags)
646 /* Sadly, NFSD reopens a file repeatedly during operation, so the
647 * "acc_mode = 0" allowance for newly-created files isn't honoured.
648 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
649 * owner can write to a file even if it is marked readonly to hide
650 * its brokenness. (bug 5781) */
651 if (flags & MDS_OPEN_OWNEROVERRIDE && inode->i_uid == current->fsuid)
654 if (flags & FMODE_READ)
656 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
658 if (flags & MDS_FMODE_EXEC)
663 /* Handles object creation, actual opening, and I/O epoch */
664 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
665 struct mds_body *body, int flags, void **handle,
666 struct mds_update_record *rec,struct ldlm_reply *rep,
667 struct lustre_handle *lockh)
669 struct mds_obd *mds = mds_req2mds(req);
670 struct obd_device *obd = req->rq_export->exp_obd;
671 struct mds_file_data *mfd = NULL;
672 obd_id *ids = NULL; /* object IDs created */
676 /* atomically create objects if necessary */
677 LOCK_INODE_MUTEX(dchild->d_inode);
679 if (S_ISREG(dchild->d_inode->i_mode) &&
680 !(body->valid & OBD_MD_FLEASIZE)) {
681 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
682 body, dchild->d_inode, 0);
684 UNLOCK_INODE_MUTEX(dchild->d_inode);
689 if ((body->valid & OBD_MD_FLEASIZE) &&
690 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
691 UNLOCK_INODE_MUTEX(dchild->d_inode);
694 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
695 UNLOCK_INODE_MUTEX(dchild->d_inode);
696 rc = mds_join_file(rec, req, dchild, lockh);
699 LOCK_INODE_MUTEX(dchild->d_inode);
701 if (!(body->valid & OBD_MD_FLEASIZE) &&
702 !(body->valid & OBD_MD_FLMODEASIZE)) {
703 /* no EA: create objects */
704 rc = mds_create_objects(req, DLM_REPLY_REC_OFF + 1, rec,
705 mds, obd, dchild, handle, &ids);
707 CERROR("mds_create_objects: rc = %d\n", rc);
708 UNLOCK_INODE_MUTEX(dchild->d_inode);
713 /* If the inode has no EA data, then MDS holds size, mtime */
714 if (S_ISREG(dchild->d_inode->i_mode) &&
715 !(body->valid & OBD_MD_FLEASIZE)) {
716 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
717 OBD_MD_FLATIME | OBD_MD_FLMTIME);
719 UNLOCK_INODE_MUTEX(dchild->d_inode);
721 if (rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE))
722 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
723 body->eadatasize, 0);
725 if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
726 rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
727 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
729 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
730 dchild->d_inode, req->rq_repmsg,
732 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
737 intent_set_disposition(rep, DISP_OPEN_OPEN);
738 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
740 RETURN(PTR_ERR(mfd));
742 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
743 mfd->mfd_handle.h_cookie);
746 mds_lov_update_objids(obd, ids);
747 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
749 if (rc) /* coverity[deadcode] */
750 mds_mfd_unlink(mfd, 1);
756 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
757 struct mds_body *body, int flags,
758 struct mds_update_record *rec,struct ldlm_reply *rep)
760 struct mds_obd *mds = mds_req2mds(req);
761 struct dentry *dchild;
762 char fidname[LL_FID_NAMELEN];
767 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
768 dchild = ll_lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
769 if (IS_ERR(dchild)) {
770 rc = PTR_ERR(dchild);
771 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
775 if (dchild->d_inode != NULL) {
776 mds_inode_set_orphan(dchild->d_inode);
777 CWARN("Orphan %s found and opened in PENDING directory\n",
782 /* We didn't find it in PENDING so it isn't an orphan. See
783 * if it was a regular inode that was previously created. */
784 dchild = mds_fid2dentry(mds, fid, NULL);
786 RETURN(PTR_ERR(dchild));
789 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
790 mds_pack_inode2body(body, dchild->d_inode);
791 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
792 intent_set_disposition(rep, DISP_LOOKUP_POS);
794 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep, NULL);
795 rc = mds_finish_transno(mds, dchild->d_inode, handle,
796 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
797 /* XXX what do we do here if mds_finish_transno itself failed? */
803 int mds_pin(struct ptlrpc_request *req, int offset)
805 struct obd_device *obd = req->rq_export->exp_obd;
806 struct mds_body *reqbody, *repbody;
807 struct lvfs_run_ctxt saved;
808 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
811 reqbody = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*reqbody));
813 rc = lustre_pack_reply(req, 2, size, NULL);
817 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
820 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
821 rc = mds_open_by_fid(req, &reqbody->fid1, repbody, reqbody->flags, NULL,
823 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
828 /* Get an internal lock on the inode number (but not generation) to sync
829 * new inode creation with inode unlink (bug 2029). If child_lockh is NULL
830 * we just get the lock as a barrier to wait for other holders of this lock,
831 * and drop it right away again. */
832 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
833 struct lustre_handle *child_lockh)
835 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
836 struct lustre_handle lockh;
837 int lock_flags = LDLM_FL_ATOMIC_CB;
840 if (child_lockh == NULL)
841 child_lockh = &lockh;
843 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
844 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
845 ldlm_blocking_ast, ldlm_completion_ast,
846 NULL, NULL, 0, NULL, child_lockh);
848 CERROR("ldlm_cli_enqueue_local: %d\n", rc);
849 else if (child_lockh == &lockh)
850 ldlm_lock_decref(child_lockh, LCK_EX);
855 int mds_open(struct mds_update_record *rec, int offset,
856 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
858 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
859 struct obd_device *obd = req->rq_export->exp_obd;
860 struct mds_obd *mds = mds_req2mds(req);
861 struct ldlm_reply *rep = NULL;
862 struct mds_body *body = NULL;
863 struct dentry *dchild = NULL, *dparent = NULL;
864 struct mds_export_data *med;
865 struct lustre_handle parent_lockh;
866 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
867 int parent_mode = LCK_CR;
869 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
870 unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
871 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
872 int child_mode = LCK_CR;
873 /* Always returning LOOKUP lock if open succesful to guard
875 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
876 struct ldlm_res_id child_res_id = { .name = {0}};
880 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
882 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
883 (obd_timeout + 1) / 4);
885 CLASSERT(MAXQUOTAS < 4);
886 if (offset == DLM_INTENT_REC_OFF) { /* intent */
887 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
889 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
891 } else if (offset == REQ_REC_OFF) { /* non-intent reint */
892 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
894 LBUG(); /* XXX: not supported yet? */
900 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
902 /* Step 0: If we are passed a fid, then we assume the client already
903 * opened this file and is only replaying the RPC, so we open the
904 * inode by fid (at some large expense in security). */
905 /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */
906 if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
907 (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&&
908 !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
909 if (rec->ur_fid2->id == 0) {
910 struct ldlm_lock *lock = ldlm_handle2lock(child_lockh);
912 LDLM_ERROR(lock, "fid2 not set on open replay");
915 DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
919 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
922 if (req->rq_export->exp_libclient &&
923 rec->ur_flags & MDS_OPEN_HAS_EA)
929 /* We didn't find the correct inode on disk either, so we
930 * need to re-create it via a regular replay. */
931 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
932 DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay");
935 } else if (rec->ur_fid2->id) {
936 DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay",
937 rec->ur_fid2->id, rec->ur_fid2->generation);
941 /* If we got here, we must be called via intent */
942 LASSERT(offset == DLM_INTENT_REC_OFF);
944 med = &req->rq_export->exp_mds_data;
945 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
946 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
950 /* Step 1: Find and lock the parent */
951 if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE))
952 parent_mode = LCK_EX;
953 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
954 &parent_lockh, MDS_INODELOCK_UPDATE);
955 if (IS_ERR(dparent)) {
956 rc = PTR_ERR(dparent);
958 CERROR("parent "LPU64"/%u lookup error %d\n",
959 rec->ur_fid1->id, rec->ur_fid1->generation, rc);
961 /* Just cannot find parent - make it look like
962 * usual negative lookup to avoid extra MDS RPC */
963 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
964 intent_set_disposition(rep, DISP_LOOKUP_NEG);
968 LASSERT(dparent->d_inode != NULL);
970 cleanup_phase = 1; /* parent dentry and lock */
972 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
973 dchild = dget(dparent);
974 cleanup_phase = 2; /* child dentry */
975 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
976 GOTO(found_child, rc);
979 /* Step 2: Lookup the child */
981 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
982 (rec->ur_flags & MDS_OPEN_LOCK) && (rec->ur_namelen == 1)) {
983 /* hack for nfsd with no_subtree_check, it will use anon
984 * dentry w/o filename to open the file. the anon dentry's
985 * parent was set to itself, so rec->ur_fid1 is the file.
986 * And in MDC it cannot derive the dentry's parent dentry,
987 * hence the file's name, so we hack here in MDS,
988 * refer to bug 13030. */
989 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
991 dchild = ll_lookup_one_len(rec->ur_name, dparent,
992 rec->ur_namelen - 1);
994 if (IS_ERR(dchild)) {
995 rc = PTR_ERR(dchild);
996 dchild = NULL; /* don't confuse mds_finish_transno() below */
1000 cleanup_phase = 2; /* child dentry */
1002 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1003 if (dchild->d_inode)
1004 intent_set_disposition(rep, DISP_LOOKUP_POS);
1006 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1008 /*Step 3: If the child was negative, and we're supposed to, create it.*/
1009 if (dchild->d_inode == NULL) {
1010 unsigned long ino = rec->ur_fid2->id;
1012 struct inode *inode;
1014 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1015 /* It's negative and we weren't supposed to create it */
1016 GOTO(cleanup, rc = -ENOENT);
1019 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1020 GOTO(cleanup, rc = -EROFS);
1022 intent_set_disposition(rep, DISP_OPEN_CREATE);
1023 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1025 if (IS_ERR(handle)) {
1026 rc = PTR_ERR(handle);
1030 dchild->d_fsdata = (void *) &dp;
1034 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1035 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1036 dchild->d_fsdata = NULL;
1039 CDEBUG(D_INODE, "error during create: %d\n", rc);
1042 inode = dchild->d_inode;
1044 LASSERT(ino == inode->i_ino);
1045 /* Written as part of setattr */
1046 inode->i_generation = rec->ur_fid2->generation;
1047 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1048 inode->i_ino, inode->i_generation);
1052 LTIME_S(iattr.ia_atime) = rec->ur_time;
1053 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1054 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1056 iattr.ia_uid = current->fsuid; /* set by push_ctxt already */
1057 if (dparent->d_inode->i_mode & S_ISGID)
1058 iattr.ia_gid = dparent->d_inode->i_gid;
1060 iattr.ia_gid = current->fsgid;
1062 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1063 ATTR_MTIME | ATTR_CTIME;
1065 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1067 CERROR("error on child setattr: rc = %d\n", rc);
1069 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1071 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1073 CERROR("error on parent setattr: rc = %d\n", rc);
1075 rc = fsfilt_commit(obd, dchild->d_inode, handle, 0);
1077 acc_mode = 0; /* Don't check for permissions */
1079 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1082 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1083 "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len,
1084 dchild->d_name.name, dchild, dchild->d_inode,
1085 dchild->d_inode->i_ino, dchild->d_inode->i_generation);
1088 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
1089 mds_pack_inode2body(body, dchild->d_inode);
1091 if (S_ISREG(dchild->d_inode->i_mode)) {
1092 /* Check permissions etc */
1093 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1097 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) &&
1098 (acc_mode & MAY_WRITE))
1099 GOTO(cleanup, rc = -EROFS);
1101 /* An append-only file must be opened in append mode for
1103 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1104 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1105 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1106 GOTO(cleanup, rc = -EPERM);
1109 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1110 (rec->ur_flags & MDS_OPEN_EXCL)) {
1111 /* File already exists, we didn't just create it, and we
1112 * were passed O_EXCL; err-or. */
1113 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1116 /* if we are following a symlink, don't open */
1117 if (S_ISLNK(dchild->d_inode->i_mode))
1118 GOTO(cleanup_no_trans, rc = 0);
1120 if (S_ISDIR(dchild->d_inode->i_mode)) {
1121 if (rec->ur_flags & MDS_OPEN_CREAT ||
1122 rec->ur_flags & FMODE_WRITE) {
1123 /* we are trying to create or write a exist dir */
1124 GOTO(cleanup, rc = -EISDIR);
1126 if (rec->ur_flags & MDS_FMODE_EXEC) {
1127 /* we are trying to exec a directory */
1128 GOTO(cleanup, rc = -EACCES);
1130 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1131 intent_set_disposition(rep, DISP_OPEN_OPEN);
1132 GOTO(cleanup, rc = -EACCES);
1134 } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) {
1135 GOTO(cleanup, rc = -ENOTDIR);
1138 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1139 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1140 GOTO(cleanup, rc = -EAGAIN);
1143 /* Obtain OPEN lock as well */
1144 policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1146 /* We cannot use acc_mode here, because it is zeroed in case of
1147 creating a file, so we get wrong lockmode */
1148 if (accmode(dchild->d_inode, rec->ur_flags) & MAY_WRITE)
1149 child_mode = LCK_CW;
1150 else if (accmode(dchild->d_inode, rec->ur_flags) & MAY_EXEC)
1151 child_mode = LCK_PR;
1153 child_mode = LCK_CR;
1155 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1156 (rec->ur_flags & MDS_OPEN_LOCK)) {
1157 /* In case of replay we do not get a lock assuming that the
1158 caller has it already */
1159 child_res_id.name[0] = dchild->d_inode->i_ino;
1160 child_res_id.name[1] = dchild->d_inode->i_generation;
1162 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
1163 LDLM_IBITS, &policy, child_mode,
1164 &lock_flags, ldlm_blocking_ast,
1165 ldlm_completion_ast, NULL, NULL,
1166 0, NULL, child_lockh);
1170 /* Let mds_intent_policy know that we have a lock to return */
1171 intent_set_disposition(rep, DISP_OPEN_LOCK);
1175 if (!S_ISREG(dchild->d_inode->i_mode) &&
1176 !S_ISDIR(dchild->d_inode->i_mode) &&
1177 (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) {
1178 /* If client supports this, do not return open handle for
1179 * special device nodes */
1180 GOTO(cleanup_no_trans, rc = 0);
1183 /* Step 5: mds_open it */
1184 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1185 rep, &parent_lockh);
1189 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1190 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
1193 switch (cleanup_phase) {
1196 /* It is safe to leave IT_OPEN_LOCK set, if rc is not 0,
1197 * mds_intent_policy won't try to return any locks */
1198 ldlm_lock_decref(child_lockh, child_mode);
1200 if (rc && created) {
1201 int err = vfs_unlink(dparent->d_inode, dchild);
1203 CERROR("unlink(%.*s) in error path: %d\n",
1204 dchild->d_name.len, dchild->d_name.name,
1207 } else if (created) {
1208 mds_lock_new_child(obd, dchild->d_inode, NULL);
1209 /* save uid/gid for quota acquire/release */
1210 qpids[USRQUOTA] = dparent->d_inode->i_uid;
1211 qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1215 if (dparent == NULL)
1220 ldlm_lock_decref(&parent_lockh, parent_mode);
1222 ptlrpc_save_lock(req, &parent_lockh, parent_mode);
1224 /* trigger dqacq on the owner of child and parent */
1225 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1228 /* If we have not taken the "open" lock, we may not return 0 here,
1229 because caller expects 0 to mean "lock is taken", and it needs
1230 nonzero return here for caller to return EDLM_LOCK_ABORTED to
1231 client. Later caller should rewrite the return value back to zero
1232 if it to be used any further
1237 /* Close a "file descriptor" and possibly unlink an orphan from the
1238 * PENDING directory. Caller must hold child->i_mutex, this drops it.
1240 * If we are being called from mds_disconnect() because the client has
1241 * disappeared, then req == NULL and we do not update last_rcvd because
1242 * there is nothing that could be recovered by the client at this stage
1243 * (it will not even _have_ an entry in last_rcvd anymore).
1245 * Returns EAGAIN if the client needs to get more data and re-close. */
1246 int mds_mfd_close(struct ptlrpc_request *req, int offset,struct obd_device *obd,
1247 struct mds_file_data *mfd, int unlink_orphan,
1248 struct lov_mds_md *lmm, int lmm_size,
1249 struct llog_cookie *logcookies, int cookies_size,
1252 struct inode *inode = mfd->mfd_dentry->d_inode;
1253 char fidname[LL_FID_NAMELEN];
1254 int last_orphan, fidlen, 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 lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1261 struct iattr iattr = { 0 };
1264 if (req && req->rq_reqmsg != NULL)
1265 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1266 sizeof(*request_body));
1267 if (req && req->rq_repmsg != NULL)
1268 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1269 sizeof(*reply_body));
1271 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1273 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname,
1274 inode->i_nlink, mds_orphan_open_count(inode));
1276 last_orphan = mds_orphan_open_dec_test(inode) &&
1277 mds_inode_is_orphan(inode);
1278 MDS_UP_WRITE_ORPHAN_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 & MDS_FMODE_EXEC) {
1285 mds_allow_write_access(inode);
1288 if (last_orphan && unlink_orphan) {
1289 int stripe_count = 0;
1290 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1292 CDEBUG(D_INODE, "destroying orphan object %s\n", fidname);
1294 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1295 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1296 CERROR("found \"orphan\" %s %s with link count %d\n",
1297 S_ISREG(inode->i_mode) ? "file" : "dir",
1298 fidname, inode->i_nlink);
1300 /* Sadly, there is no easy way to save pending_child from
1301 * mds_reint_unlink() into mfd, so we need to re-lookup,
1302 * but normally it will still be in the dcache. */
1303 LOCK_INODE_MUTEX(pending_dir);
1304 cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */
1305 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1307 if (IS_ERR(pending_child))
1308 GOTO(cleanup, rc = PTR_ERR(pending_child));
1309 LASSERT(pending_child->d_inode != NULL);
1311 cleanup_phase = 2; /* dput(pending_child) when finished */
1312 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1313 rc = vfs_rmdir(pending_dir, pending_child);
1315 CERROR("error unlinking orphan dir %s: rc %d\n",
1321 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1324 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1325 NULL, stripe_count);
1326 if (IS_ERR(handle)) {
1327 rc = PTR_ERR(handle);
1332 if (lmm != NULL && (*valid & OBD_MD_FLEASIZE) &&
1333 mds_log_op_unlink(obd, lmm, lmm_size,
1334 logcookies, cookies_size) > 0) {
1335 *valid |= OBD_MD_FLCOOKIE;
1340 pending_child->d_fsdata = (void *) &dp;
1341 rc = vfs_unlink(pending_dir, pending_child);
1343 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1345 goto out; /* Don't bother updating attrs on unlinked inode */
1349 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1350 /* Update the on-disk attributes if this was the last write
1351 * close, and all information was provided (i.e., rc == 0)
1353 * XXX this should probably be abstracted with mds_reint_setattr
1356 if (request_body->valid & OBD_MD_FLMTIME &&
1357 LTIME_S(iattr.ia_mtime) > LTIME_S(inode->i_mtime)) {
1358 LTIME_S(iattr.ia_mtime) = request_body->mtime;
1359 iattr.ia_valid |= ATTR_MTIME;
1361 if (request_body->valid & OBD_MD_FLCTIME &&
1362 LTIME_S(iattr.ia_ctime) > LTIME_S(inode->i_ctime)) {
1363 LTIME_S(iattr.ia_ctime) = request_body->ctime;
1364 iattr.ia_valid |= ATTR_CTIME;
1367 /* XXX can't set block count with fsfilt_setattr (!) */
1368 if (request_body->valid & OBD_MD_FLSIZE) {
1369 iattr.ia_valid |= ATTR_SIZE;
1370 iattr.ia_size = request_body->size;
1372 /* iattr.ia_blocks = request_body->blocks */
1376 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1377 /* Only start a transaction to write out only the atime if
1378 * it is more out-of-date than the specified limit. If we
1379 * are already going to write out the atime then do it anyway.
1381 LTIME_S(iattr.ia_atime) = request_body->atime;
1382 if ((LTIME_S(iattr.ia_atime) >
1383 LTIME_S(inode->i_atime) + mds->mds_atime_diff) ||
1384 (iattr.ia_valid != 0 &&
1385 LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime)))
1386 iattr.ia_valid |= ATTR_ATIME;
1389 if (iattr.ia_valid != 0) {
1390 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1391 if (IS_ERR(handle)) {
1392 rc = PTR_ERR(handle);
1396 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1398 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1401 /* If other clients have this file open for write, rc will be > 0 */
1404 l_dput(mfd->mfd_dentry);
1408 if (req != NULL && reply_body != NULL) {
1409 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0, 0);
1410 } else if (handle) {
1411 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1413 CERROR("error committing close: %d\n", err);
1419 switch (cleanup_phase) {
1421 dput(pending_child);
1423 UNLOCK_INODE_MUTEX(pending_dir);
1428 int mds_close(struct ptlrpc_request *req, int offset)
1430 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1431 struct obd_device *obd = req->rq_export->exp_obd;
1432 struct mds_body *body;
1433 struct mds_file_data *mfd;
1434 struct lvfs_run_ctxt saved;
1435 struct inode *inode;
1436 int rc, repsize[4] = { sizeof(struct ptlrpc_body),
1437 sizeof(struct mds_body),
1438 obd->u.mds.mds_max_mdsize,
1439 obd->u.mds.mds_max_cookiesize };
1440 struct mds_body *reply_body;
1441 struct lov_mds_md *lmm;
1443 struct llog_cookie *logcookies;
1447 rc = lustre_pack_reply(req, 4, repsize, NULL);
1449 req->rq_status = rc;
1450 /* continue on to drop local open even if we can't send reply */
1452 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1454 CDEBUG(D_INODE, "close req->rep_len %d mdsize %d cookiesize %d\n",
1456 obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize);
1457 mds_counter_incr(req->rq_export, LPROC_MDS_CLOSE);
1459 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1460 lustre_swab_mds_body);
1462 CERROR("Can't unpack body\n");
1463 req->rq_status = -EFAULT;
1467 spin_lock(&med->med_open_lock);
1468 mfd = mds_handle2mfd(&body->handle);
1470 spin_unlock(&med->med_open_lock);
1471 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1472 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1473 req->rq_status = -ESTALE;
1476 /* Remove mfd handle so it can't be found again. We consume mfd_list
1477 * reference here, but still have mds_handle2mfd ref until mfd_close. */
1478 mds_mfd_unlink(mfd, 1);
1479 spin_unlock(&med->med_open_lock);
1481 inode = mfd->mfd_dentry->d_inode;
1482 /* child orphan sem protects orphan_dec_test && is_orphan race */
1483 MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */
1484 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1485 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1487 LASSERT(body != NULL);
1489 mds_pack_inode2fid(&body->fid1, inode);
1490 mds_pack_inode2body(body, inode);
1491 mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode,
1495 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1496 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body));
1497 lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, 0);
1498 lmm_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 1),
1499 logcookies = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 2, 0);
1500 cookies_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 2);
1501 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1,
1502 lmm, lmm_size, logcookies, cookies_size,
1503 &reply_body->valid);
1504 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1506 mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1);
1507 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1508 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1509 req->rq_status = -ENOMEM;
1516 int mds_done_writing(struct ptlrpc_request *req, int offset)
1518 struct mds_body *body;
1519 int rc, size[2] = { sizeof(struct ptlrpc_body),
1520 sizeof(struct mds_body) };
1523 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1525 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1526 lustre_swab_mds_body);
1528 CERROR("Can't unpack body\n");
1529 req->rq_status = -EFAULT;
1533 rc = lustre_pack_reply(req, 2, size, NULL);
1535 req->rq_status = rc;