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 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
117 /* Remove the mfd handle so that it cannot be found by open/close again.
118 * Caller must hold med_open_lock for mfd_list manipulation. */
119 void mds_mfd_unlink(struct mds_file_data *mfd, int decref)
121 class_handle_unhash(&mfd->mfd_handle);
122 list_del_init(&mfd->mfd_list);
127 /* Caller must hold mds->mds_epoch_sem */
128 static int mds_alloc_filterdata(struct inode *inode)
130 LASSERT(inode->i_filterdata == NULL);
131 OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
132 if (inode->i_filterdata == NULL)
134 LASSERT(igrab(inode) == inode);
138 /* Caller must hold mds->mds_epoch_sem */
139 static void mds_free_filterdata(struct inode *inode)
141 LASSERT(inode->i_filterdata != NULL);
142 OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
143 inode->i_filterdata = NULL;
147 /* Write access to a file: executors cause a negative count,
148 * writers a positive count. The semaphore is needed to perform
149 * a check for the sign and then increment or decrement atomically.
151 * This code is closely tied to the allocation of the d_fsdata and the
152 * MDS epoch, so we use the same semaphore for the whole lot.
154 * We could use a different semaphore for each file, if it ever shows
155 * up in a profile, which it won't.
157 * epoch argument is nonzero during recovery */
158 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
163 down(&mds->mds_epoch_sem);
165 if (atomic_read(&inode->i_writecount) < 0) {
166 up(&mds->mds_epoch_sem);
171 if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
172 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
173 MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
174 inode->i_generation);
178 if (inode->i_filterdata == NULL)
179 mds_alloc_filterdata(inode);
180 if (inode->i_filterdata == NULL) {
184 if (epoch > mds->mds_io_epoch)
185 mds->mds_io_epoch = epoch;
188 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
189 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
190 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
193 atomic_inc(&inode->i_writecount);
194 up(&mds->mds_epoch_sem);
198 /* Returns EAGAIN if the client needs to get size and/or cookies and close
199 * again -- which is never true if the file is about to be unlinked. Otherwise
200 * returns the number of remaining writers. */
201 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
202 struct mds_body *body, int unlinking)
207 down(&mds->mds_epoch_sem);
208 atomic_dec(&inode->i_writecount);
209 rc = atomic_read(&inode->i_writecount);
213 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
214 GOTO(out, rc = EAGAIN);
216 mds_free_filterdata(inode);
218 up(&mds->mds_epoch_sem);
222 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
225 down(&mds->mds_epoch_sem);
226 if (atomic_read(&inode->i_writecount) > 0) {
227 up(&mds->mds_epoch_sem);
230 atomic_dec(&inode->i_writecount);
231 up(&mds->mds_epoch_sem);
235 static void mds_allow_write_access(struct inode *inode)
238 atomic_inc(&inode->i_writecount);
241 int mds_query_write_access(struct inode *inode)
244 RETURN(atomic_read(&inode->i_writecount));
247 /* This replaces the VFS dentry_open, it manages mfd and writecount */
248 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
249 struct vfsmount *mnt, int flags,
250 struct ptlrpc_request *req)
252 struct mds_export_data *med = &req->rq_export->exp_mds_data;
253 struct mds_obd *mds = mds_req2mds(req);
254 struct mds_file_data *mfd;
255 struct mds_body *body;
261 CERROR("mds: out of memory\n");
262 GOTO(cleanup_dentry, error = -ENOMEM);
265 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
267 if (flags & FMODE_WRITE) {
268 /* FIXME: in recovery, need to pass old epoch here */
269 error = mds_get_write_access(mds, dentry->d_inode, 0);
271 GOTO(cleanup_mfd, error);
272 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
273 } else if (flags & MDS_FMODE_EXEC) {
274 error = mds_deny_write_access(mds, dentry->d_inode);
276 GOTO(cleanup_mfd, error);
281 /* Mark the file as open to handle open-unlink. */
282 MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode);
283 mds_orphan_open_inc(dentry->d_inode);
284 MDS_UP_WRITE_ORPHAN_SEM(dentry->d_inode);
286 mfd->mfd_mode = flags;
287 mfd->mfd_dentry = dentry;
288 mfd->mfd_xid = req->rq_xid;
290 spin_lock(&med->med_open_lock);
291 list_add(&mfd->mfd_list, &med->med_open_head);
292 spin_unlock(&med->med_open_lock);
294 body->handle.cookie = mfd->mfd_handle.h_cookie;
300 mds_mfd_unlink(mfd, 1);
302 return ERR_PTR(error);
305 /* Must be called with i_mutex 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 inode *inode = dchild->d_inode;
313 struct obd_trans_info oti = { 0 };
314 struct lov_mds_md *lmm = NULL;
316 struct mds_body *body;
317 struct obd_info oinfo = { { { 0 } } };
321 if (!S_ISREG(inode->i_mode))
323 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
324 !(rec->ur_flags & FMODE_WRITE))
327 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
329 if (body->valid & OBD_MD_FLEASIZE)
332 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 if (rec->ur_fid2->id == 0) {
341 DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
345 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
346 lmm_size = rec->ur_eadatalen;
347 lmm = rec->ur_eadata;
351 *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
352 if (IS_ERR(*handle)) {
353 rc = PTR_ERR(*handle);
358 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
360 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
362 CERROR("open replay failed to set md:%d\n", rc);
363 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
365 memcpy(lmm_buf, lmm, lmm_size);
370 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
371 GOTO(out_ids, rc = -ENOMEM);
373 OBDO_ALLOC(oinfo.oi_oa);
374 if (oinfo.oi_oa == NULL)
375 GOTO(out_ids, rc = -ENOMEM);
376 oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */
377 oinfo.oi_oa->o_gid = 0;
378 oinfo.oi_oa->o_mode = S_IFREG | 0600;
379 oinfo.oi_oa->o_id = inode->i_ino;
380 oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS |
381 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
382 oinfo.oi_oa->o_size = 0;
384 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
385 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
387 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
388 /* check if things like lfs setstripe are sending us the ea */
389 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
390 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
392 0, &oinfo.oi_md, rec->ur_eadata);
396 OBD_ALLOC(lmm, mds->mds_max_mdsize);
398 GOTO(out_oa, rc = -ENOMEM);
400 lmm_size = mds->mds_max_mdsize;
401 rc = mds_get_md(obd, dchild->d_parent->d_inode,
404 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
406 0, &oinfo.oi_md, lmm);
407 OBD_FREE(lmm, mds->mds_max_mdsize);
411 rc = obd_create(mds->mds_osc_exp, oinfo.oi_oa,
417 CDEBUG(level, "error creating objects for "
418 "inode %lu: rc = %d\n",
421 CERROR("obd_create returned invalid "
428 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
429 0, &oinfo.oi_md, rec->ur_eadata);
433 oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id;
434 oinfo.oi_md->lsm_object_gr = oinfo.oi_oa->o_gr;
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);
478 GOTO(free_diskmd, rc);
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);
486 obd_free_diskmd(mds->mds_osc_exp, &lmm);
488 oti_free_cookies(&oti);
489 OBDO_FREE(oinfo.oi_oa);
492 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
496 obd_free_memmd(mds->mds_osc_exp, &oinfo.oi_md);
500 static void reconstruct_open(struct mds_update_record *rec, int offset,
501 struct ptlrpc_request *req,
502 struct lustre_handle *child_lockh)
504 struct mds_export_data *med = &req->rq_export->exp_mds_data;
505 struct mds_client_data *mcd = med->med_mcd;
506 struct mds_obd *mds = mds_req2mds(req);
507 struct mds_file_data *mfd;
508 struct obd_device *obd = req->rq_export->exp_obd;
509 struct dentry *parent, *dchild;
510 struct ldlm_reply *rep;
511 struct mds_body *body;
517 LASSERT(offset == DLM_INTENT_REC_OFF); /* only called via intent */
518 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
519 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
521 /* copy rc, transno and disp; steal locks */
522 mds_req_from_mcd(req, mcd);
523 intent_set_disposition(rep, le32_to_cpu(mcd->mcd_last_data));
525 /* Only replay if create or open actually happened. */
526 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
528 return; /* error looking up parent or child */
531 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
532 LASSERT(!IS_ERR(parent));
534 dchild = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
535 LASSERT(!IS_ERR(dchild));
537 if (!dchild->d_inode)
538 GOTO(out_dput, 0); /* child not present to open */
540 /* At this point, we know we have a child. We'll send
541 * it back _unless_ it not created and open failed.
543 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
544 !intent_disposition(rep, DISP_OPEN_CREATE) &&
549 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
550 mds_pack_inode2body(body, dchild->d_inode);
551 if (S_ISREG(dchild->d_inode->i_mode)) {
552 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
553 body, dchild->d_inode, 1);
556 LASSERT(rc == req->rq_status);
558 /* If we have LOV EA data, the OST holds size, mtime */
559 if (!(body->valid & OBD_MD_FLEASIZE))
560 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
561 OBD_MD_FLATIME | OBD_MD_FLMTIME);
564 if (!(rec->ur_flags & MDS_OPEN_JOIN_FILE))
565 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
566 body->eadatasize, 0);
568 if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
569 !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
570 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
572 rc = mds_pack_acl(med, dchild->d_inode, req->rq_repmsg,
574 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
575 if (!req->rq_status && rc)
579 /* If we have -EEXIST as the status, and we were asked to create
580 * exclusively, we can tell we failed because the file already existed.
582 if (req->rq_status == -EEXIST &&
583 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
584 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
588 /* If we didn't get as far as trying to open, then some locking thing
589 * probably went wrong, and we'll just bail here.
591 if (!intent_disposition(rep, DISP_OPEN_OPEN))
594 /* If we failed, then we must have failed opening, so don't look for
595 * file descriptor or anything, just give the client the bad news.
601 spin_lock(&med->med_open_lock);
602 list_for_each(t, &med->med_open_head) {
603 mfd = list_entry(t, struct mds_file_data, mfd_list);
604 if (mfd->mfd_xid == req->rq_xid) {
610 spin_unlock(&med->med_open_lock);
612 /* #warning "XXX fixme" bug 2991 */
613 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
614 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
615 * to detect a re-open */
617 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
618 rc = mds_join_file(rec, req, dchild, NULL);
622 mntget(mds->mds_vfsmnt);
623 CERROR("Re-opened file \n");
624 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
625 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
627 CERROR("mds: out of memory\n");
628 GOTO(out_dput, req->rq_status = -ENOMEM);
632 body->handle.cookie = mfd->mfd_handle.h_cookie;
633 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
634 mfd->mfd_handle.h_cookie);
646 /* do NOT or the MAY_*'s, you'll get the weakest */
647 static int accmode(struct inode *inode, int flags)
651 /* Sadly, NFSD reopens a file repeatedly during operation, so the
652 * "acc_mode = 0" allowance for newly-created files isn't honoured.
653 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
654 * owner can write to a file even if it is marked readonly to hide
655 * its brokenness. (bug 5781) */
656 if (flags & MDS_OPEN_OWNEROVERRIDE && inode->i_uid == current->fsuid)
659 if (flags & FMODE_READ)
661 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
663 if (flags & MDS_FMODE_EXEC)
668 /* Handles object creation, actual opening, and I/O epoch */
669 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
670 struct mds_body *body, int flags, void **handle,
671 struct mds_update_record *rec,struct ldlm_reply *rep,
672 struct lustre_handle *lockh)
674 struct mds_obd *mds = mds_req2mds(req);
675 struct obd_device *obd = req->rq_export->exp_obd;
676 struct mds_file_data *mfd = NULL;
677 obd_id *ids = NULL; /* object IDs created */
681 /* atomically create objects if necessary */
682 LOCK_INODE_MUTEX(dchild->d_inode);
684 if (S_ISREG(dchild->d_inode->i_mode) &&
685 !(body->valid & OBD_MD_FLEASIZE)) {
686 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
687 body, dchild->d_inode, 0);
689 UNLOCK_INODE_MUTEX(dchild->d_inode);
694 if ((body->valid & OBD_MD_FLEASIZE) &&
695 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
696 UNLOCK_INODE_MUTEX(dchild->d_inode);
699 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
700 UNLOCK_INODE_MUTEX(dchild->d_inode);
701 rc = mds_join_file(rec, req, dchild, lockh);
704 LOCK_INODE_MUTEX(dchild->d_inode);
706 if (!(body->valid & OBD_MD_FLEASIZE) &&
707 !(body->valid & OBD_MD_FLMODEASIZE)) {
708 /* no EA: create objects */
709 rc = mds_create_objects(req, DLM_REPLY_REC_OFF + 1, rec,
710 mds, obd, dchild, handle, &ids);
712 CERROR("mds_create_objects: rc = %d\n", rc);
713 UNLOCK_INODE_MUTEX(dchild->d_inode);
718 /* If the inode has no EA data, then MDS holds size, mtime */
719 if (S_ISREG(dchild->d_inode->i_mode) &&
720 !(body->valid & OBD_MD_FLEASIZE)) {
721 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
722 OBD_MD_FLATIME | OBD_MD_FLMTIME);
724 UNLOCK_INODE_MUTEX(dchild->d_inode);
726 if (rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE))
727 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
728 body->eadatasize, 0);
730 if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
731 rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
732 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
734 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
735 dchild->d_inode, req->rq_repmsg,
737 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
742 intent_set_disposition(rep, DISP_OPEN_OPEN);
743 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
745 RETURN(PTR_ERR(mfd));
747 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
748 mfd->mfd_handle.h_cookie);
751 mds_lov_update_objids(obd, ids);
752 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
754 if (rc) /* coverity[deadcode] */
755 mds_mfd_unlink(mfd, 1);
761 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
762 struct mds_body *body, int flags,
763 struct mds_update_record *rec,struct ldlm_reply *rep)
765 struct mds_obd *mds = mds_req2mds(req);
766 struct dentry *dchild;
767 char fidname[LL_FID_NAMELEN];
772 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
773 dchild = ll_lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
774 if (IS_ERR(dchild)) {
775 rc = PTR_ERR(dchild);
776 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
780 if (dchild->d_inode != NULL) {
781 mds_inode_set_orphan(dchild->d_inode);
782 CWARN("Orphan %s found and opened in PENDING directory\n",
787 /* We didn't find it in PENDING so it isn't an orphan. See
788 * if it was a regular inode that was previously created. */
789 dchild = mds_fid2dentry(mds, fid, NULL);
791 RETURN(PTR_ERR(dchild));
794 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
795 mds_pack_inode2body(body, dchild->d_inode);
796 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
797 intent_set_disposition(rep, DISP_LOOKUP_POS);
799 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep, NULL);
800 rc = mds_finish_transno(mds, dchild->d_inode, handle,
801 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
802 /* XXX what do we do here if mds_finish_transno itself failed? */
808 int mds_pin(struct ptlrpc_request *req, int offset)
810 struct obd_device *obd = req->rq_export->exp_obd;
811 struct mds_body *reqbody, *repbody;
812 struct lvfs_run_ctxt saved;
813 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
816 reqbody = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*reqbody));
818 rc = lustre_pack_reply(req, 2, size, NULL);
821 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
824 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
825 rc = mds_open_by_fid(req, &reqbody->fid1, repbody, reqbody->flags, NULL,
827 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
832 /* Get an internal lock on the inode number (but not generation) to sync
833 * new inode creation with inode unlink (bug 2029). If child_lockh is NULL
834 * we just get the lock as a barrier to wait for other holders of this lock,
835 * and drop it right away again. */
836 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
837 struct lustre_handle *child_lockh)
839 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
840 struct lustre_handle lockh;
841 int lock_flags = LDLM_FL_ATOMIC_CB;
844 if (child_lockh == NULL)
845 child_lockh = &lockh;
847 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
848 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
849 ldlm_blocking_ast, ldlm_completion_ast,
850 NULL, NULL, 0, NULL, child_lockh);
852 CERROR("ldlm_cli_enqueue_local: %d\n", rc);
853 else if (child_lockh == &lockh)
854 ldlm_lock_decref(child_lockh, LCK_EX);
859 int mds_open(struct mds_update_record *rec, int offset,
860 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
862 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
863 struct obd_device *obd = req->rq_export->exp_obd;
864 struct mds_obd *mds = mds_req2mds(req);
865 struct ldlm_reply *rep = NULL;
866 struct mds_body *body = NULL;
867 struct dentry *dchild = NULL, *dparent = NULL;
868 struct mds_export_data *med;
869 struct lustre_handle parent_lockh;
870 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
871 int parent_mode = LCK_CR;
873 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
874 unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
875 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
876 int child_mode = LCK_CR;
877 /* Always returning LOOKUP lock if open succesful to guard
879 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
880 struct ldlm_res_id child_res_id = { .name = {0}};
884 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
886 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
887 (obd_timeout + 1) / 4);
889 CLASSERT(MAXQUOTAS < 4);
890 if (offset == DLM_INTENT_REC_OFF) { /* intent */
891 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
893 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
895 } else if (offset == REQ_REC_OFF) { /* non-intent reint */
896 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
898 LBUG(); /* XXX: not supported yet? */
904 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
906 /* Step 0: If we are passed a fid, then we assume the client already
907 * opened this file and is only replaying the RPC, so we open the
908 * inode by fid (at some large expense in security). */
909 /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */
910 if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
911 (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&&
912 !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
913 if (rec->ur_fid2->id == 0) {
914 struct ldlm_lock *lock = ldlm_handle2lock(child_lockh);
916 LDLM_ERROR(lock, "fid2 not set on open replay");
919 DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
923 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
926 if (req->rq_export->exp_libclient &&
927 rec->ur_flags & MDS_OPEN_HAS_EA)
933 /* We didn't find the correct inode on disk either, so we
934 * need to re-create it via a regular replay. */
935 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
936 DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay");
939 } else if (rec->ur_fid2->id) {
940 DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay",
941 rec->ur_fid2->id, rec->ur_fid2->generation);
945 /* If we got here, we must be called via intent */
946 LASSERT(offset == DLM_INTENT_REC_OFF);
948 med = &req->rq_export->exp_mds_data;
949 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
950 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
954 /* Step 1: Find and lock the parent */
955 if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE))
956 parent_mode = LCK_EX;
957 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
958 &parent_lockh, MDS_INODELOCK_UPDATE);
959 if (IS_ERR(dparent)) {
960 rc = PTR_ERR(dparent);
962 CERROR("parent "LPU64"/%u lookup error %d\n",
963 rec->ur_fid1->id, rec->ur_fid1->generation, rc);
965 /* Just cannot find parent - make it look like
966 * usual negative lookup to avoid extra MDS RPC */
967 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
968 intent_set_disposition(rep, DISP_LOOKUP_NEG);
972 LASSERT(dparent->d_inode != NULL);
974 cleanup_phase = 1; /* parent dentry and lock */
976 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
977 dchild = dget(dparent);
978 cleanup_phase = 2; /* child dentry */
979 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
980 GOTO(found_child, rc);
983 /* Step 2: Lookup the child */
985 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
986 (rec->ur_flags & MDS_OPEN_LOCK) && (rec->ur_namelen == 1)) {
987 /* hack for nfsd with no_subtree_check, it will use anon
988 * dentry w/o filename to open the file. the anon dentry's
989 * parent was set to itself, so rec->ur_fid1 is the file.
990 * And in MDC it cannot derive the dentry's parent dentry,
991 * hence the file's name, so we hack here in MDS,
992 * refer to bug 13030. */
993 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
995 dchild = ll_lookup_one_len(rec->ur_name, dparent,
996 rec->ur_namelen - 1);
998 if (IS_ERR(dchild)) {
999 rc = PTR_ERR(dchild);
1000 dchild = NULL; /* don't confuse mds_finish_transno() below */
1004 cleanup_phase = 2; /* child dentry */
1006 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1007 if (dchild->d_inode)
1008 intent_set_disposition(rep, DISP_LOOKUP_POS);
1010 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1012 /*Step 3: If the child was negative, and we're supposed to, create it.*/
1013 if (dchild->d_inode == NULL) {
1014 unsigned long ino = rec->ur_fid2->id;
1016 struct inode *inode;
1018 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1019 /* It's negative and we weren't supposed to create it */
1020 GOTO(cleanup, rc = -ENOENT);
1023 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1024 GOTO(cleanup, rc = -EROFS);
1026 intent_set_disposition(rep, DISP_OPEN_CREATE);
1027 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1029 if (IS_ERR(handle)) {
1030 rc = PTR_ERR(handle);
1034 dchild->d_fsdata = (void *) &dp;
1038 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1039 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1040 dchild->d_fsdata = NULL;
1043 CDEBUG(D_INODE, "error during create: %d\n", rc);
1046 inode = dchild->d_inode;
1048 LASSERT(ino == inode->i_ino);
1049 /* Written as part of setattr */
1050 inode->i_generation = rec->ur_fid2->generation;
1051 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1052 inode->i_ino, inode->i_generation);
1056 LTIME_S(iattr.ia_atime) = rec->ur_time;
1057 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1058 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1060 iattr.ia_uid = current->fsuid; /* set by push_ctxt already */
1061 if (dparent->d_inode->i_mode & S_ISGID)
1062 iattr.ia_gid = dparent->d_inode->i_gid;
1064 iattr.ia_gid = current->fsgid;
1066 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1067 ATTR_MTIME | ATTR_CTIME;
1069 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1071 CERROR("error on child setattr: rc = %d\n", rc);
1073 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1075 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1077 CERROR("error on parent setattr: rc = %d\n", rc);
1079 rc = fsfilt_commit(obd, dchild->d_inode, handle, 0);
1081 acc_mode = 0; /* Don't check for permissions */
1083 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1086 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1087 "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len,
1088 dchild->d_name.name, dchild, dchild->d_inode,
1089 dchild->d_inode->i_ino, dchild->d_inode->i_generation);
1092 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
1093 mds_pack_inode2body(body, dchild->d_inode);
1095 if (S_ISREG(dchild->d_inode->i_mode)) {
1096 /* Check permissions etc */
1097 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1101 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) &&
1102 (acc_mode & MAY_WRITE))
1103 GOTO(cleanup, rc = -EROFS);
1105 /* An append-only file must be opened in append mode for
1107 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1108 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1109 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1110 GOTO(cleanup, rc = -EPERM);
1113 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1114 (rec->ur_flags & MDS_OPEN_EXCL)) {
1115 /* File already exists, we didn't just create it, and we
1116 * were passed O_EXCL; err-or. */
1117 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1120 /* if we are following a symlink, don't open */
1121 if (S_ISLNK(dchild->d_inode->i_mode))
1122 GOTO(cleanup_no_trans, rc = 0);
1124 if (S_ISDIR(dchild->d_inode->i_mode)) {
1125 if (rec->ur_flags & MDS_OPEN_CREAT ||
1126 rec->ur_flags & FMODE_WRITE) {
1127 /* we are trying to create or write a exist dir */
1128 GOTO(cleanup, rc = -EISDIR);
1130 if (rec->ur_flags & MDS_FMODE_EXEC) {
1131 /* we are trying to exec a directory */
1132 GOTO(cleanup, rc = -EACCES);
1134 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1135 intent_set_disposition(rep, DISP_OPEN_OPEN);
1136 GOTO(cleanup, rc = -EACCES);
1138 } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) {
1139 GOTO(cleanup, rc = -ENOTDIR);
1142 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1143 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1144 GOTO(cleanup, rc = -EAGAIN);
1147 /* Obtain OPEN lock as well */
1148 policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1150 /* We cannot use acc_mode here, because it is zeroed in case of
1151 creating a file, so we get wrong lockmode */
1152 if (accmode(dchild->d_inode, rec->ur_flags) & MAY_WRITE)
1153 child_mode = LCK_CW;
1154 else if (accmode(dchild->d_inode, rec->ur_flags) & MAY_EXEC)
1155 child_mode = LCK_PR;
1157 child_mode = LCK_CR;
1159 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1160 (rec->ur_flags & MDS_OPEN_LOCK)) {
1161 /* In case of replay we do not get a lock assuming that the
1162 caller has it already */
1163 child_res_id.name[0] = dchild->d_inode->i_ino;
1164 child_res_id.name[1] = dchild->d_inode->i_generation;
1166 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
1167 LDLM_IBITS, &policy, child_mode,
1168 &lock_flags, ldlm_blocking_ast,
1169 ldlm_completion_ast, NULL, NULL,
1170 0, NULL, child_lockh);
1174 /* Let mds_intent_policy know that we have a lock to return */
1175 intent_set_disposition(rep, DISP_OPEN_LOCK);
1179 if (!S_ISREG(dchild->d_inode->i_mode) &&
1180 !S_ISDIR(dchild->d_inode->i_mode) &&
1181 (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) {
1182 /* If client supports this, do not return open handle for
1183 * special device nodes */
1184 GOTO(cleanup_no_trans, rc = 0);
1187 /* Step 5: mds_open it */
1188 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1189 rep, &parent_lockh);
1193 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1194 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
1197 switch (cleanup_phase) {
1200 /* It is safe to leave IT_OPEN_LOCK set, if rc is not 0,
1201 * mds_intent_policy won't try to return any locks */
1202 ldlm_lock_decref(child_lockh, child_mode);
1204 if (rc && created) {
1205 int err = vfs_unlink(dparent->d_inode, dchild);
1207 CERROR("unlink(%.*s) in error path: %d\n",
1208 dchild->d_name.len, dchild->d_name.name,
1211 } else if (created) {
1212 mds_lock_new_child(obd, dchild->d_inode, NULL);
1213 /* save uid/gid for quota acquire/release */
1214 qpids[USRQUOTA] = dparent->d_inode->i_uid;
1215 qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1219 if (dparent == NULL)
1224 ldlm_lock_decref(&parent_lockh, parent_mode);
1226 ptlrpc_save_lock(req, &parent_lockh, parent_mode);
1228 /* trigger dqacq on the owner of child and parent */
1229 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1232 /* If we have not taken the "open" lock, we may not return 0 here,
1233 because caller expects 0 to mean "lock is taken", and it needs
1234 nonzero return here for caller to return EDLM_LOCK_ABORTED to
1235 client. Later caller should rewrite the return value back to zero
1236 if it to be used any further
1241 /* Close a "file descriptor" and possibly unlink an orphan from the
1242 * PENDING directory. Caller must hold child->i_mutex, this drops it.
1244 * If we are being called from mds_disconnect() because the client has
1245 * disappeared, then req == NULL and we do not update last_rcvd because
1246 * there is nothing that could be recovered by the client at this stage
1247 * (it will not even _have_ an entry in last_rcvd anymore).
1249 * Returns EAGAIN if the client needs to get more data and re-close. */
1250 int mds_mfd_close(struct ptlrpc_request *req, int offset,struct obd_device *obd,
1251 struct mds_file_data *mfd, int unlink_orphan,
1252 struct lov_mds_md *lmm, int lmm_size,
1253 struct llog_cookie *logcookies, int cookies_size,
1256 struct inode *inode = mfd->mfd_dentry->d_inode;
1257 char fidname[LL_FID_NAMELEN];
1258 int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1259 struct dentry *pending_child = NULL;
1260 struct mds_obd *mds = &obd->u.mds;
1261 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1262 void *handle = NULL;
1263 struct mds_body *request_body = NULL, *reply_body = NULL;
1264 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1265 struct iattr iattr = { 0 };
1268 if (req && req->rq_reqmsg != NULL)
1269 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1270 sizeof(*request_body));
1271 if (req && req->rq_repmsg != NULL)
1272 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1273 sizeof(*reply_body));
1275 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1277 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname,
1278 inode->i_nlink, mds_orphan_open_count(inode));
1280 last_orphan = mds_orphan_open_dec_test(inode) &&
1281 mds_inode_is_orphan(inode);
1282 MDS_UP_WRITE_ORPHAN_SEM(inode);
1284 /* this is half of the actual "close" */
1285 if (mfd->mfd_mode & FMODE_WRITE) {
1286 rc = mds_put_write_access(mds, inode, request_body,
1287 last_orphan && unlink_orphan);
1288 } else if (mfd->mfd_mode & MDS_FMODE_EXEC) {
1289 mds_allow_write_access(inode);
1292 if (last_orphan && unlink_orphan) {
1293 int stripe_count = 0;
1294 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1296 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1298 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1299 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1300 CERROR("found \"orphan\" %s %s with link count %d\n",
1301 S_ISREG(inode->i_mode) ? "file" : "dir",
1302 fidname, inode->i_nlink);
1304 /* Sadly, there is no easy way to save pending_child from
1305 * mds_reint_unlink() into mfd, so we need to re-lookup,
1306 * but normally it will still be in the dcache. */
1307 LOCK_INODE_MUTEX(pending_dir);
1308 cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */
1309 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1311 if (IS_ERR(pending_child))
1312 GOTO(cleanup, rc = PTR_ERR(pending_child));
1313 LASSERT(pending_child->d_inode != NULL);
1315 cleanup_phase = 2; /* dput(pending_child) when finished */
1316 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1317 rc = vfs_rmdir(pending_dir, pending_child);
1319 CERROR("error unlinking orphan dir %s: rc %d\n",
1325 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1328 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1329 NULL, stripe_count);
1330 if (IS_ERR(handle)) {
1331 rc = PTR_ERR(handle);
1336 if (lmm != NULL && (*valid & OBD_MD_FLEASIZE) &&
1337 mds_log_op_unlink(obd, lmm, lmm_size,
1338 logcookies, cookies_size) > 0) {
1339 *valid |= OBD_MD_FLCOOKIE;
1344 pending_child->d_fsdata = (void *) &dp;
1345 rc = vfs_unlink(pending_dir, pending_child);
1347 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1349 goto out; /* Don't bother updating attrs on unlinked inode */
1353 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1354 /* Update the on-disk attributes if this was the last write
1355 * close, and all information was provided (i.e., rc == 0)
1357 * XXX this should probably be abstracted with mds_reint_setattr
1360 if (request_body->valid & OBD_MD_FLMTIME &&
1361 LTIME_S(iattr.ia_mtime) > LTIME_S(inode->i_mtime)) {
1362 LTIME_S(iattr.ia_mtime) = request_body->mtime;
1363 iattr.ia_valid |= ATTR_MTIME;
1365 if (request_body->valid & OBD_MD_FLCTIME &&
1366 LTIME_S(iattr.ia_ctime) > LTIME_S(inode->i_ctime)) {
1367 LTIME_S(iattr.ia_ctime) = request_body->ctime;
1368 iattr.ia_valid |= ATTR_CTIME;
1371 /* XXX can't set block count with fsfilt_setattr (!) */
1372 if (request_body->valid & OBD_MD_FLSIZE) {
1373 iattr.ia_valid |= ATTR_SIZE;
1374 iattr.ia_size = request_body->size;
1376 /* iattr.ia_blocks = request_body->blocks */
1380 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1381 /* Only start a transaction to write out only the atime if
1382 * it is more out-of-date than the specified limit. If we
1383 * are already going to write out the atime then do it anyway.
1385 LTIME_S(iattr.ia_atime) = request_body->atime;
1386 if ((LTIME_S(iattr.ia_atime) >
1387 LTIME_S(inode->i_atime) + mds->mds_atime_diff) ||
1388 (iattr.ia_valid != 0 &&
1389 LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime)))
1390 iattr.ia_valid |= ATTR_ATIME;
1393 if (iattr.ia_valid != 0) {
1394 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1395 if (IS_ERR(handle)) {
1396 rc = PTR_ERR(handle);
1400 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1402 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1405 /* If other clients have this file open for write, rc will be > 0 */
1408 l_dput(mfd->mfd_dentry);
1412 if (req != NULL && reply_body != NULL) {
1413 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0, 0);
1414 } else if (handle) {
1415 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1417 CERROR("error committing close: %d\n", err);
1423 switch (cleanup_phase) {
1425 dput(pending_child);
1427 UNLOCK_INODE_MUTEX(pending_dir);
1432 int mds_close(struct ptlrpc_request *req, int offset)
1434 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1435 struct obd_device *obd = req->rq_export->exp_obd;
1436 struct mds_body *body;
1437 struct mds_file_data *mfd;
1438 struct lvfs_run_ctxt saved;
1439 struct inode *inode;
1440 int rc, repsize[4] = { sizeof(struct ptlrpc_body),
1441 sizeof(struct mds_body),
1442 obd->u.mds.mds_max_mdsize,
1443 obd->u.mds.mds_max_cookiesize };
1444 struct mds_body *reply_body;
1445 struct lov_mds_md *lmm;
1447 struct llog_cookie *logcookies;
1451 rc = lustre_pack_reply(req, 4, repsize, NULL);
1453 CERROR("lustre_pack_reply: rc = %d\n", rc);
1454 req->rq_status = rc;
1455 /* continue on to drop local open even if we can't send reply */
1457 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1460 CDEBUG(D_HA, "close req->rep_len %d mdsize %d cookiesize %d\n",
1462 obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize);
1463 mds_counter_incr(req->rq_export, LPROC_MDS_CLOSE);
1465 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1466 lustre_swab_mds_body);
1468 CERROR("Can't unpack body\n");
1469 req->rq_status = -EFAULT;
1473 spin_lock(&med->med_open_lock);
1474 mfd = mds_handle2mfd(&body->handle);
1476 spin_unlock(&med->med_open_lock);
1477 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1478 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1479 req->rq_status = -ESTALE;
1482 /* Remove mfd handle so it can't be found again. We consume mfd_list
1483 * reference here, but still have mds_handle2mfd ref until mfd_close. */
1484 mds_mfd_unlink(mfd, 1);
1485 spin_unlock(&med->med_open_lock);
1487 inode = mfd->mfd_dentry->d_inode;
1488 /* child orphan sem protects orphan_dec_test && is_orphan race */
1489 MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */
1490 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1491 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1493 LASSERT(body != NULL);
1495 mds_pack_inode2fid(&body->fid1, inode);
1496 mds_pack_inode2body(body, inode);
1497 mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode,
1501 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1502 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body));
1503 lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, 0);
1504 lmm_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 1),
1505 logcookies = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 2, 0);
1506 cookies_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 2);
1507 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1,
1508 lmm, lmm_size, logcookies, cookies_size,
1509 &reply_body->valid);
1510 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1512 mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1);
1513 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1514 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1515 req->rq_status = -ENOMEM;
1522 int mds_done_writing(struct ptlrpc_request *req, int offset)
1524 struct mds_body *body;
1525 int rc, size[2] = { sizeof(struct ptlrpc_body),
1526 sizeof(struct mds_body) };
1529 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1531 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1532 lustre_swab_mds_body);
1534 CERROR("Can't unpack body\n");
1535 req->rq_status = -EFAULT;
1539 rc = lustre_pack_reply(req, 2, size, NULL);
1541 CERROR("lustre_pack_reply: rc = %d\n", rc);
1542 req->rq_status = rc;