1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mds/mds_open.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
40 * Author: Phil Schwan <phil@clusterfs.com>
41 * Author: Mike Shaver <shaver@clusterfs.com>
45 # define EXPORT_SYMTAB
47 #define DEBUG_SUBSYSTEM S_MDS
49 #include <linux/version.h>
50 #include <linux/module.h>
51 #include <linux/init.h>
52 #include <linux/version.h>
53 #include <linux/buffer_head.h>
54 #include <linux/workqueue.h>
56 #include <obd_class.h>
58 #include <lustre_fsfilt.h>
59 #include <lprocfs_status.h>
61 #include "mds_internal.h"
63 /* Exported function from this file are:
65 * mds_open - called by the intent handler
66 * mds_close - an rpc handling function
67 * mds_pin - an rpc handling function - which will go away
68 * mds_mfd_close - for force closing files when a client dies
72 * MDS file data handling: file data holds a handle for a file opened
76 static void mds_mfd_addref(void *mfdp)
78 struct mds_file_data *mfd = mfdp;
80 atomic_inc(&mfd->mfd_refcount);
81 CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
82 atomic_read(&mfd->mfd_refcount));
85 /* Create a new mds_file_data struct.
86 * One reference for handle+med_open_head list and dropped by mds_mfd_unlink(),
87 * one reference for the caller of this function. */
88 struct mds_file_data *mds_mfd_new(void)
90 struct mds_file_data *mfd;
92 OBD_ALLOC(mfd, sizeof *mfd);
94 CERROR("mds: out of memory\n");
98 atomic_set(&mfd->mfd_refcount, 2);
100 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
101 INIT_LIST_HEAD(&mfd->mfd_list);
102 class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
107 /* Get a new reference on the mfd pointed to by handle, if handle is still
108 * valid. Caller must drop reference with mds_mfd_put(). */
109 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
112 LASSERT(handle != NULL);
113 RETURN(class_handle2object(handle->cookie));
116 /* Drop mfd reference, freeing struct if this is the last one. */
117 static void mds_mfd_put(struct mds_file_data *mfd)
119 CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
120 atomic_read(&mfd->mfd_refcount) - 1);
121 LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
122 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
123 if (atomic_dec_and_test(&mfd->mfd_refcount)) {
124 OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
128 /* Remove the mfd handle so that it cannot be found by open/close again.
129 * Caller must hold med_open_lock for mfd_list manipulation. */
130 void mds_mfd_unlink(struct mds_file_data *mfd, int decref)
132 class_handle_unhash(&mfd->mfd_handle);
133 list_del_init(&mfd->mfd_list);
138 /* Caller must hold mds->mds_epoch_sem */
139 static int mds_alloc_filterdata(struct inode *inode)
141 LASSERT(inode->i_filterdata == NULL);
142 OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
143 if (inode->i_filterdata == NULL)
145 LASSERT(igrab(inode) == inode);
149 /* Caller must hold mds->mds_epoch_sem */
150 static void mds_free_filterdata(struct inode *inode)
152 LASSERT(inode->i_filterdata != NULL);
153 OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
154 inode->i_filterdata = NULL;
158 /* Write access to a file: executors cause a negative count,
159 * writers a positive count. The semaphore is needed to perform
160 * a check for the sign and then increment or decrement atomically.
162 * This code is closely tied to the allocation of the d_fsdata and the
163 * MDS epoch, so we use the same semaphore for the whole lot.
165 * We could use a different semaphore for each file, if it ever shows
166 * up in a profile, which it won't.
168 * epoch argument is nonzero during recovery */
169 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
174 down(&mds->mds_epoch_sem);
176 if (atomic_read(&inode->i_writecount) < 0) {
177 up(&mds->mds_epoch_sem);
182 if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
183 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
184 MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
185 inode->i_generation);
189 if (inode->i_filterdata == NULL)
190 mds_alloc_filterdata(inode);
191 if (inode->i_filterdata == NULL) {
195 if (epoch > mds->mds_io_epoch)
196 mds->mds_io_epoch = epoch;
199 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
200 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
201 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
204 atomic_inc(&inode->i_writecount);
205 up(&mds->mds_epoch_sem);
209 /* Returns EAGAIN if the client needs to get size and/or cookies and close
210 * again -- which is never true if the file is about to be unlinked. Otherwise
211 * returns the number of remaining writers. */
212 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
213 struct mds_body *body, int unlinking)
218 down(&mds->mds_epoch_sem);
219 atomic_dec(&inode->i_writecount);
220 rc = atomic_read(&inode->i_writecount);
224 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
225 GOTO(out, rc = EAGAIN);
227 mds_free_filterdata(inode);
229 up(&mds->mds_epoch_sem);
233 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
236 down(&mds->mds_epoch_sem);
237 if (atomic_read(&inode->i_writecount) > 0) {
238 up(&mds->mds_epoch_sem);
241 atomic_dec(&inode->i_writecount);
242 up(&mds->mds_epoch_sem);
246 static void mds_allow_write_access(struct inode *inode)
249 atomic_inc(&inode->i_writecount);
252 int mds_query_write_access(struct inode *inode)
255 RETURN(atomic_read(&inode->i_writecount));
258 /* This replaces the VFS dentry_open, it manages mfd and writecount */
259 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
260 struct vfsmount *mnt, int flags,
261 struct ptlrpc_request *req)
263 struct mds_export_data *med = &req->rq_export->exp_mds_data;
264 struct mds_obd *mds = mds_req2mds(req);
265 struct mds_file_data *mfd;
266 struct mds_body *body;
272 CERROR("mds: out of memory\n");
273 GOTO(cleanup_dentry, error = -ENOMEM);
276 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
278 if (flags & FMODE_WRITE) {
279 /* FIXME: in recovery, need to pass old epoch here */
280 error = mds_get_write_access(mds, dentry->d_inode, 0);
282 GOTO(cleanup_mfd, error);
283 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
284 } else if (flags & MDS_FMODE_EXEC) {
285 error = mds_deny_write_access(mds, dentry->d_inode);
287 GOTO(cleanup_mfd, error);
292 /* Mark the file as open to handle open-unlink. */
293 MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode);
294 mds_orphan_open_inc(dentry->d_inode);
295 MDS_UP_WRITE_ORPHAN_SEM(dentry->d_inode);
297 mfd->mfd_mode = flags;
298 mfd->mfd_dentry = dentry;
299 mfd->mfd_xid = req->rq_xid;
301 spin_lock(&med->med_open_lock);
302 list_add(&mfd->mfd_list, &med->med_open_head);
303 spin_unlock(&med->med_open_lock);
305 body->handle.cookie = mfd->mfd_handle.h_cookie;
311 mds_mfd_unlink(mfd, 1);
313 return ERR_PTR(error);
316 /* Must be called with i_mutex held */
317 static int mds_create_objects(struct ptlrpc_request *req, int offset,
318 struct mds_update_record *rec,
319 struct mds_obd *mds, struct obd_device *obd,
320 struct dentry *dchild, void **handle,
321 struct lov_mds_md **objid)
323 struct inode *inode = dchild->d_inode;
324 struct obd_trans_info oti = { 0 };
325 struct lov_mds_md *lmm = NULL;
327 struct mds_body *body;
328 struct obd_info oinfo = { { { 0 } } };
334 if (!S_ISREG(inode->i_mode))
336 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
337 !(rec->ur_flags & FMODE_WRITE))
340 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
342 if (body->valid & OBD_MD_FLEASIZE)
348 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
349 if (rec->ur_fid2->id == 0) {
350 DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
354 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
355 lmm_size = rec->ur_eadatalen;
356 lmm = rec->ur_eadata;
359 if (*handle == NULL) {
360 int stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
361 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_CREATE,
364 if (IS_ERR(*handle)) {
365 rc = PTR_ERR(*handle);
370 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
372 CERROR("open replay failed to set md:%d\n", rc);
374 /* for replay we not need send lmm to client, this not used now */
375 lustre_shrink_reply(req, offset, 0, 1);
381 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
382 GOTO(out_ids, rc = -ENOMEM);
384 OBDO_ALLOC(oinfo.oi_oa);
385 if (oinfo.oi_oa == NULL)
386 GOTO(out_ids, rc = -ENOMEM);
387 oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */
388 oinfo.oi_oa->o_gid = 0;
389 oinfo.oi_oa->o_mode = S_IFREG | 0600;
390 oinfo.oi_oa->o_id = inode->i_ino;
391 oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS |
392 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
393 oinfo.oi_oa->o_size = 0;
395 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
396 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
397 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
398 /* check if things like lfs setstripe are sending us the ea */
399 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
400 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
402 0, &oinfo.oi_md, rec->ur_eadata);
406 __u32 lmm_sz = mds->mds_max_mdsize;
407 OBD_ALLOC(lmm, lmm_sz);
409 GOTO(out_oa, rc = -ENOMEM);
412 rc = mds_get_md(obd, dchild->d_parent->d_inode,
413 lmm, &lmm_size, 1, 0,
414 req->rq_export->exp_connect_flags);
416 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
418 0, &oinfo.oi_md, lmm);
419 OBD_FREE(lmm, lmm_sz);
423 rc = obd_create(mds->mds_osc_exp, oinfo.oi_oa,
429 CDEBUG(level, "error creating objects for "
430 "inode %lu: rc = %d\n",
433 CERROR("obd_create returned invalid "
440 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
441 0, &oinfo.oi_md, rec->ur_eadata);
445 oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id;
447 if (i_size_read(inode)) {
448 oinfo.oi_oa->o_size = i_size_read(inode);
449 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE |
450 OBD_MD_FLATIME | OBD_MD_FLMTIME |
451 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
453 /* pack lustre id to OST */
454 oinfo.oi_oa->o_fid = body->fid1.id;
455 oinfo.oi_oa->o_generation = body->fid1.generation;
456 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
457 oinfo.oi_policy.l_extent.start = i_size_read(inode);
458 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
460 rc = obd_punch_rqset(mds->mds_osc_exp, &oinfo, &oti);
462 CERROR("error setting attrs for inode %lu: rc %d\n",
465 CERROR("obd_setattr_async returned bad rc %d\n",
473 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
474 obdo_refresh_inode(inode, oinfo.oi_oa, OBD_MD_FLBLKSZ);
476 LASSERT(oinfo.oi_md && oinfo.oi_md->lsm_object_id);
478 rc = obd_packmd(mds->mds_osc_exp, &lmm, oinfo.oi_md);
480 CERROR("cannot pack lsm, err = %d\n", rc);
484 body->eadatasize = rc;
487 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
488 if (IS_ERR(*handle)) {
489 rc = PTR_ERR(*handle);
491 GOTO(free_diskmd, rc);
494 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
495 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
497 memcpy(lmm_buf, lmm, lmm_size);
498 *objid = lmm_buf; /* save for mds_lov_update_objid */
501 obd_free_diskmd(mds->mds_osc_exp, &lmm);
503 oti_free_cookies(&oti);
504 OBDO_FREE(oinfo.oi_oa);
507 obd_free_memmd(mds->mds_osc_exp, &oinfo.oi_md);
511 static void reconstruct_open(struct mds_update_record *rec, int offset,
512 struct ptlrpc_request *req,
513 struct lustre_handle *child_lockh)
515 struct mds_export_data *med = &req->rq_export->exp_mds_data;
516 struct lsd_client_data *lcd = med->med_lcd;
517 struct mds_obd *mds = mds_req2mds(req);
518 struct mds_file_data *mfd;
519 struct obd_export *exp = req->rq_export;
520 struct obd_device *obd = exp->exp_obd;
521 struct dentry *parent, *dchild;
522 struct ldlm_reply *rep;
523 struct mds_body *body;
529 LASSERT(offset == DLM_INTENT_REC_OFF); /* only called via intent */
530 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
531 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
533 /* copy rc, transno and disp; steal locks */
534 mds_req_from_lcd(req, lcd);
535 intent_set_disposition(rep, le32_to_cpu(lcd->lcd_last_data));
537 /* Only replay if create or open actually happened. */
538 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
540 return; /* error looking up parent or child */
543 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
544 if (IS_ERR(parent)) {
545 rc = PTR_ERR(parent);
546 LCONSOLE_WARN("Parent "LPU64"/%u lookup error %d."
547 " Evicting client %s with export %s.\n",
548 rec->ur_fid1->id, rec->ur_fid1->generation, rc,
549 obd_uuid2str(&exp->exp_client_uuid),
550 obd_export_nid2str(exp));
551 mds_export_evict(exp);
556 dchild = mds_lookup(obd, rec->ur_name, parent, rec->ur_namelen - 1);
557 if (IS_ERR(dchild)) {
558 rc = PTR_ERR(dchild);
559 LCONSOLE_WARN("Child "LPU64"/%u lookup error %d."
560 " Evicting client %s with export %s.\n",
561 rec->ur_fid1->id, rec->ur_fid1->generation, rc,
562 obd_uuid2str(&exp->exp_client_uuid),
563 obd_export_nid2str(exp));
564 mds_export_evict(exp);
570 if (!dchild->d_inode)
571 GOTO(out_dput, 0); /* child not present to open */
573 /* At this point, we know we have a child. We'll send
574 * it back _unless_ it not created and open failed.
576 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
577 !intent_disposition(rep, DISP_OPEN_CREATE) &&
582 mds_pack_inode2body(body, dchild->d_inode);
583 if (S_ISREG(dchild->d_inode->i_mode)) {
584 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
585 body, dchild->d_inode, 1, 0,
586 req->rq_export->exp_connect_flags);
589 LASSERT(rc == req->rq_status);
591 /* If we have LOV EA data, the OST holds size, mtime */
592 if (!(body->valid & OBD_MD_FLEASIZE))
593 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
594 OBD_MD_FLATIME | OBD_MD_FLMTIME);
597 if (!(rec->ur_flags & MDS_OPEN_JOIN_FILE))
598 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
599 body->eadatasize, 0);
601 if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
602 !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
603 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
605 rc = mds_pack_acl(med, dchild->d_inode, req->rq_repmsg,
607 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
608 if (!req->rq_status && rc)
612 /* If we have -EEXIST as the status, and we were asked to create
613 * exclusively, we can tell we failed because the file already existed.
615 if (req->rq_status == -EEXIST &&
616 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
617 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
621 /* If we didn't get as far as trying to open, then some locking thing
622 * probably went wrong, and we'll just bail here.
624 if (!intent_disposition(rep, DISP_OPEN_OPEN))
627 /* If we failed, then we must have failed opening, so don't look for
628 * file descriptor or anything, just give the client the bad news.
634 spin_lock(&med->med_open_lock);
635 list_for_each(t, &med->med_open_head) {
636 mfd = list_entry(t, struct mds_file_data, mfd_list);
637 if (mfd->mfd_xid == req->rq_xid) {
643 spin_unlock(&med->med_open_lock);
645 /* #warning "XXX fixme" bug 2991 */
646 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
647 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
648 * to detect a re-open */
650 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
652 /* Allow file join in beta builds to allow debugging */
653 rc = mds_join_file(rec, req, dchild, NULL);
657 CWARN("file join is not supported in this version of "
659 GOTO(out_dput, req->rq_status = rc = -EOPNOTSUPP);
662 mntget(mds->mds_vfsmnt);
663 CERROR("Re-opened file \n");
664 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
665 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
666 mntput(mds->mds_vfsmnt);
668 req->rq_status = PTR_ERR(mfd);
670 CERROR("%s: opening inode "LPU64" failed: rc %d\n",
671 req->rq_export->exp_obd->obd_name,
672 (__u64)dchild->d_inode->i_ino, req->rq_status);
673 GOTO(out_dput, req->rq_status);
676 body->handle.cookie = mfd->mfd_handle.h_cookie;
677 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
678 mfd->mfd_handle.h_cookie);
690 /* if client disconnects during recovery it may resend opens which were replayed
691 * on server but their transno less then last_transno on server so they will not
692 * be detected as reconstructs */
693 static int open_replay_reconstruct(struct ptlrpc_request *req)
695 struct mds_export_data *med = &req->rq_export->exp_mds_data;
696 struct mds_file_data *mfd = NULL;
699 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
702 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
705 /* if mfd exists then replay was done already */
706 spin_lock(&med->med_open_lock);
707 list_for_each(t, &med->med_open_head) {
708 mfd = list_entry(t, struct mds_file_data, mfd_list);
709 if (mfd->mfd_xid == req->rq_xid) {
715 spin_unlock(&med->med_open_lock);
718 struct mds_body *body = lustre_msg_buf(req->rq_repmsg,
721 __u64 *pre_versions = lustre_msg_get_versions(req->rq_reqmsg);
723 body->handle.cookie = mfd->mfd_handle.h_cookie;
724 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
725 mfd->mfd_handle.h_cookie);
727 lustre_msg_set_versions(req->rq_repmsg, pre_versions);
728 lustre_msg_set_transno(req->rq_repmsg,
729 lustre_msg_get_transno(req->rq_reqmsg));
730 lustre_msg_set_status(req->rq_repmsg, 0);
736 /* do NOT or the MAY_*'s, you'll get the weakest */
737 static int accmode(struct inode *inode, int flags)
741 /* Sadly, NFSD reopens a file repeatedly during operation, so the
742 * "acc_mode = 0" allowance for newly-created files isn't honoured.
743 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
744 * owner can write to a file even if it is marked readonly to hide
745 * its brokenness. (bug 5781) */
746 if (flags & MDS_OPEN_OWNEROVERRIDE && inode->i_uid == current->fsuid)
749 if (flags & FMODE_READ)
751 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
753 if (flags & MDS_FMODE_EXEC)
758 /* Handles object creation, actual opening, and I/O epoch */
759 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
760 struct mds_body *body, int flags, void **handle,
761 struct mds_update_record *rec,struct ldlm_reply *rep,
762 struct lustre_handle *lockh)
764 struct mds_obd *mds = mds_req2mds(req);
765 struct obd_device *obd = req->rq_export->exp_obd;
766 struct mds_file_data *mfd = NULL;
767 struct lov_mds_md *lmm = NULL; /* object IDs created */
771 /* atomically create objects if necessary */
772 LOCK_INODE_MUTEX(dchild->d_inode);
774 if (S_ISREG(dchild->d_inode->i_mode) &&
775 !(body->valid & OBD_MD_FLEASIZE)) {
776 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
777 body, dchild->d_inode, 0, 0,
778 req->rq_export->exp_connect_flags);
780 UNLOCK_INODE_MUTEX(dchild->d_inode);
785 if ((body->valid & OBD_MD_FLEASIZE) &&
786 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
787 UNLOCK_INODE_MUTEX(dchild->d_inode);
790 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
792 /* Allow file join in beta builds to allow debugging */
793 UNLOCK_INODE_MUTEX(dchild->d_inode);
794 rc = mds_join_file(rec, req, dchild, lockh);
797 LOCK_INODE_MUTEX(dchild->d_inode);
799 CWARN("file join is not supported in this version of "
804 if (!(body->valid & OBD_MD_FLEASIZE) &&
805 !(body->valid & OBD_MD_FLMODEASIZE)) {
806 /* no EA: create objects */
807 rc = mds_create_objects(req, DLM_REPLY_REC_OFF + 1, rec,
808 mds, obd, dchild, handle, &lmm);
810 CERROR("mds_create_objects: rc = %d\n", rc);
811 UNLOCK_INODE_MUTEX(dchild->d_inode);
816 /* If the inode has no EA data, then MDS holds size, mtime */
817 if (S_ISREG(dchild->d_inode->i_mode) &&
818 !(body->valid & OBD_MD_FLEASIZE)) {
819 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
820 OBD_MD_FLATIME | OBD_MD_FLMTIME);
822 UNLOCK_INODE_MUTEX(dchild->d_inode);
824 if (rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE))
825 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
826 body->eadatasize, 0);
828 if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
829 rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
830 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
832 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
833 dchild->d_inode, req->rq_repmsg,
835 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
840 if ((rc = mds_lov_prepare_objids(obd,lmm)) != 0)
843 intent_set_disposition(rep, DISP_OPEN_OPEN);
844 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
846 RETURN(PTR_ERR(mfd));
848 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
849 mfd->mfd_handle.h_cookie);
851 mds_lov_update_objids(obd, lmm);
854 mds_mfd_unlink(mfd, 1);
860 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
861 struct mds_body *body, int flags,
862 struct mds_update_record *rec,struct ldlm_reply *rep)
864 struct obd_device *obd = req->rq_export->exp_obd;
865 struct mds_obd *mds = mds_req2mds(req);
866 struct dentry *dchild;
867 char fidname[LL_FID_NAMELEN];
872 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
873 dchild = mds_lookup(obd, fidname, mds->mds_pending_dir, fidlen);
874 if (IS_ERR(dchild)) {
875 rc = PTR_ERR(dchild);
876 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
880 if (dchild->d_inode != NULL) {
881 mds_inode_set_orphan(dchild->d_inode);
882 CWARN("Orphan %s found and opened in PENDING directory\n",
887 /* We didn't find it in PENDING so it isn't an orphan. See
888 * if it was a regular inode that was previously created. */
889 dchild = mds_fid2dentry(mds, fid, NULL);
891 RETURN(PTR_ERR(dchild));
894 mds_pack_inode2body(body, dchild->d_inode);
895 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
896 intent_set_disposition(rep, DISP_LOOKUP_POS);
898 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep, NULL);
899 rc = mds_finish_transno(mds, NULL, handle,
900 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
901 /* XXX what do we do here if mds_finish_transno itself failed? */
907 int mds_pin(struct ptlrpc_request *req, int offset)
909 struct obd_device *obd = req->rq_export->exp_obd;
910 struct mds_body *reqbody, *repbody;
911 struct lvfs_run_ctxt saved;
912 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
915 reqbody = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*reqbody));
917 rc = lustre_pack_reply(req, 2, size, NULL);
921 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
924 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
925 rc = mds_open_by_fid(req, &reqbody->fid1, repbody, reqbody->flags, NULL,
927 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
932 /* Get an internal lock on the inode number (but not generation) to sync
933 * new inode creation with inode unlink (bug 2029). If child_lockh is NULL
934 * we just get the lock as a barrier to wait for other holders of this lock,
935 * and drop it right away again. */
936 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
937 struct lustre_handle *child_lockh)
939 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
940 struct lustre_handle lockh;
941 int lock_flags = LDLM_FL_ATOMIC_CB;
944 if (child_lockh == NULL)
945 child_lockh = &lockh;
947 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
948 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
949 ldlm_blocking_ast, ldlm_completion_ast,
950 NULL, NULL, 0, NULL, child_lockh);
952 CERROR("ldlm_cli_enqueue_local: %d\n", rc);
953 else if (child_lockh == &lockh)
954 ldlm_lock_decref(child_lockh, LCK_EX);
959 int mds_open(struct mds_update_record *rec, int offset,
960 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
962 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
963 struct obd_device *obd = req->rq_export->exp_obd;
964 struct mds_obd *mds = mds_req2mds(req);
965 struct ldlm_reply *rep = NULL;
966 struct mds_body *body = NULL;
967 struct dentry *dchild = NULL, *dparent = NULL;
968 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
969 struct mds_export_data *med;
970 struct lustre_handle parent_lockh;
971 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
972 int parent_mode = LCK_CR;
974 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
975 unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
976 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
977 int child_mode = LCK_CR;
978 /* Always returning LOOKUP lock if open succesful to guard
982 int use_parent, need_open_lock;
983 unsigned int gid = current->fsgid;
986 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
988 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
989 (obd_timeout + 1) / 4);
991 CLASSERT(MAXQUOTAS < 4);
992 if (offset == DLM_INTENT_REC_OFF) { /* intent */
993 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
995 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
997 } else if (offset == REQ_REC_OFF) { /* non-intent reint */
998 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1000 LBUG(); /* XXX: not supported yet? */
1006 /* check the open resent|replay case */
1007 if (open_replay_reconstruct(req))
1010 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
1012 /* Step 0: If we are passed a fid, then we assume the client already
1013 * opened this file and is only replaying the RPC, so we open the
1014 * inode by fid (at some large expense in security). */
1015 /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */
1016 if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
1017 (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&&
1018 !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
1019 if (rec->ur_fid2->id == 0) {
1020 struct ldlm_lock *lock = ldlm_handle2lock(child_lockh);
1022 LDLM_ERROR(lock, "fid2 not set on open replay");
1023 LDLM_LOCK_PUT(lock);
1025 DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
1029 /** check there is no stale orphan with same inode number */
1030 if (rec->ur_flags & MDS_OPEN_CREAT) {
1031 rc = mds_check_stale_orphan(obd, rec->ur_fid2);
1036 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
1038 if (rc != -ENOENT) {
1039 if (req->rq_export->exp_libclient &&
1040 rec->ur_flags & MDS_OPEN_HAS_EA)
1046 /* We didn't find the correct inode on disk either, so we
1047 * need to re-create it via a regular replay. */
1048 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1049 DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay");
1052 } else if (rec->ur_fid2->id) {
1053 DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay",
1054 rec->ur_fid2->id, rec->ur_fid2->generation);
1058 /* If we got here, we must be called via intent */
1059 LASSERT(offset == DLM_INTENT_REC_OFF);
1061 med = &req->rq_export->exp_mds_data;
1062 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
1063 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
1067 if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE))
1068 parent_mode = LCK_EX;
1070 /* We cannot use acc_mode here, because it is zeroed in case of
1071 creating a file, so we get wrong lockmode */
1072 if (rec->ur_flags & FMODE_WRITE)
1073 child_mode = LCK_CW;
1074 else if (rec->ur_flags & MDS_FMODE_EXEC)
1075 child_mode = LCK_PR;
1077 child_mode = LCK_CR;
1079 /* join file and nfsd can't need lookup dchild as use parent for it */
1080 use_parent = (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1081 (rec->ur_flags & MDS_OPEN_LOCK) && (rec->ur_namelen == 1)) ||
1082 (rec->ur_flags & MDS_OPEN_JOIN_FILE);
1084 need_open_lock = !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1085 (rec->ur_flags & MDS_OPEN_LOCK);
1087 /* Try to lock both parent and child first. If child is not found,
1088 * return only locked parent. This is enough to prevent other
1089 * threads from changing this directory until creation is finished. */
1090 rc = mds_get_parent_child_locked(obd, &obd->u.mds,
1093 &dparent, parent_mode,
1094 MDS_INODELOCK_UPDATE,
1095 use_parent ? NULL : rec->ur_name,
1097 (rec->ur_flags & MDS_OPEN_LOCK) ?
1099 &dchild, child_mode,
1100 MDS_INODELOCK_LOOKUP |
1101 MDS_INODELOCK_OPEN);
1104 if (rc != -ENOENT) {
1105 CERROR("parent "LPU64"/%u lookup/take lock error %d\n",
1106 rec->ur_fid1->id, rec->ur_fid1->generation, rc);
1108 /* Just cannot find parent - make it look like
1109 * usual negative lookup to avoid extra MDS RPC */
1110 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1111 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1115 LASSERT(dparent->d_inode != NULL);
1117 cleanup_phase = 1; /* parent dentry and lock */
1120 dchild = dget(dparent);
1122 if (IS_ERR(dchild)) {
1123 rc = PTR_ERR(dchild);
1124 dchild = NULL; /* don't confuse mds_finish_transno() below */
1128 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
1129 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1130 GOTO(found_child, rc);
1133 cleanup_phase = 2; /* child dentry */
1135 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1136 if (dchild->d_inode)
1137 intent_set_disposition(rep, DISP_LOOKUP_POS);
1139 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1141 /*Step 3: If the child was negative, and we're supposed to, create it.*/
1142 if (dchild->d_inode == NULL) {
1143 unsigned long ino = rec->ur_fid2->id;
1145 struct inode *inode;
1147 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1148 /* It's negative and we weren't supposed to create it */
1149 GOTO(cleanup, rc = -ENOENT);
1152 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1153 GOTO(cleanup, rc = -EROFS);
1155 /** check there is no stale orphan with same inode number */
1156 rc = mds_check_stale_orphan(obd, rec->ur_fid2);
1160 /* version recovery check */
1161 rc = mds_version_get_check(req, dparent->d_inode, 0);
1163 GOTO(cleanup_no_trans, rc);
1165 if (dparent->d_inode->i_mode & S_ISGID)
1166 gid = dparent->d_inode->i_gid;
1168 gid = current->fsgid;
1170 /* we try to get enough quota to write here, and let ldiskfs
1171 * decide if it is out of quota or not b=14783
1172 * FIXME: after CMD is used, pointer to obd_trans_info* couldn't
1173 * be NULL, b=14840 */
1174 lquota_chkquota(mds_quota_interface_ref, obd,
1175 current->fsuid, gid, 1, &rec_pending,
1178 intent_set_disposition(rep, DISP_OPEN_CREATE);
1179 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1181 if (IS_ERR(handle)) {
1182 rc = PTR_ERR(handle);
1186 dchild->d_fsdata = (void *) &dp;
1190 LOCK_INODE_MUTEX(dparent->d_inode);
1191 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1192 UNLOCK_INODE_MUTEX(dparent->d_inode);
1193 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1194 dchild->d_fsdata = NULL;
1197 CDEBUG(D_INODE, "error during create: %d\n", rc);
1200 inode = dchild->d_inode;
1203 if (ino != inode->i_ino) {
1204 /* FID support is needed to replay this
1205 * correctly. Now fail gracefully like there is
1206 * version mismatch */
1207 if (req->rq_export->exp_delayed)
1213 /* Written as part of setattr */
1214 inode->i_generation = rec->ur_fid2->generation;
1215 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1216 inode->i_ino, inode->i_generation);
1219 LTIME_S(iattr.ia_atime) = rec->ur_time;
1220 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1221 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1223 iattr.ia_uid = current->fsuid; /* set by push_ctxt already */
1226 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1227 ATTR_MTIME | ATTR_CTIME;
1229 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1231 CERROR("error on child setattr: rc = %d\n", rc);
1233 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1235 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1237 CERROR("error on parent setattr: rc = %d\n", rc);
1239 rc = fsfilt_commit(obd, dchild->d_inode, handle, 0);
1241 acc_mode = 0; /* Don't check for permissions */
1243 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1244 /* Child previously existed so the lookup and lock is already
1245 * done, so no further locking is needed. */
1246 /* for nfs and join - we need two locks for same fid, but
1247 * with different mode */
1248 if (need_open_lock && !use_parent) {
1249 intent_set_disposition(rep, DISP_OPEN_LOCK);
1254 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1255 "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len,
1256 dchild->d_name.name, dchild, dchild->d_inode,
1257 dchild->d_inode->i_ino, dchild->d_inode->i_generation);
1260 mds_pack_inode2body(body, dchild->d_inode);
1262 if (S_ISREG(dchild->d_inode->i_mode)) {
1263 /* Check permissions etc */
1264 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1268 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) &&
1269 (acc_mode & MAY_WRITE))
1270 GOTO(cleanup, rc = -EROFS);
1272 /* An append-only file must be opened in append mode for
1274 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1275 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1276 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1277 GOTO(cleanup, rc = -EPERM);
1280 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1281 (rec->ur_flags & MDS_OPEN_EXCL)) {
1282 /* File already exists, we didn't just create it, and we
1283 * were passed O_EXCL; err-or. */
1284 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1287 /* if we are following a symlink, don't open */
1288 if (S_ISLNK(dchild->d_inode->i_mode))
1289 GOTO(cleanup_no_trans, rc = 0);
1291 if (S_ISDIR(dchild->d_inode->i_mode)) {
1292 if (rec->ur_flags & MDS_OPEN_CREAT ||
1293 rec->ur_flags & FMODE_WRITE) {
1294 /* we are trying to create or write a exist dir */
1295 GOTO(cleanup, rc = -EISDIR);
1297 if (rec->ur_flags & MDS_FMODE_EXEC) {
1298 /* we are trying to exec a directory */
1299 GOTO(cleanup, rc = -EACCES);
1301 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1302 intent_set_disposition(rep, DISP_OPEN_OPEN);
1303 GOTO(cleanup, rc = -EACCES);
1305 } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) {
1306 GOTO(cleanup, rc = -ENOTDIR);
1309 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1310 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1311 GOTO(cleanup, rc = -EAGAIN);
1314 if (need_open_lock) {
1315 ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN } };
1316 struct ldlm_res_id child_res_id;
1318 /* In case of replay we do not get a lock assuming that the
1319 caller has it already */
1320 child_res_id.name[0] = dchild->d_inode->i_ino;
1321 child_res_id.name[1] = dchild->d_inode->i_generation;
1323 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
1324 LDLM_IBITS, &policy, child_mode,
1325 &lock_flags, ldlm_blocking_ast,
1326 ldlm_completion_ast, NULL, NULL,
1327 0, NULL, child_lockh);
1331 /* Let mds_intent_policy know that we have a lock to return */
1332 intent_set_disposition(rep, DISP_OPEN_LOCK);
1335 if (!S_ISREG(dchild->d_inode->i_mode) &&
1336 !S_ISDIR(dchild->d_inode->i_mode) &&
1337 (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) {
1338 /* If client supports this, do not return open handle for
1339 * special device nodes */
1340 GOTO(cleanup_no_trans, rc = 0);
1343 /* Step 5: mds_open it */
1344 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1345 rep, &parent_lockh);
1349 inodes[0] = (!created || IS_ERR(dparent)) ? NULL : dparent->d_inode;
1350 inodes[1] = (created && dchild) ? dchild->d_inode : NULL;
1351 rc = mds_finish_transno(mds, inodes, handle, req, rc,
1352 rep ? rep->lock_policy_res1 : 0, 0);
1356 lquota_pending_commit(mds_quota_interface_ref, obd,
1357 current->fsuid, gid, rec_pending);
1358 switch (cleanup_phase) {
1360 if (rc && created) {
1362 LOCK_INODE_MUTEX(dparent->d_inode);
1363 err = ll_vfs_unlink(dparent->d_inode, dchild,
1365 UNLOCK_INODE_MUTEX(dparent->d_inode);
1367 CERROR("unlink(%.*s) in error path: %d\n",
1368 dchild->d_name.len, dchild->d_name.name,
1371 } else if (created) {
1372 mds_lock_new_child(obd, dchild->d_inode, NULL);
1373 /* save uid/gid for quota acquire/release */
1374 qpids[USRQUOTA] = dparent->d_inode->i_uid;
1375 qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1380 /* It is safe to leave IT_OPEN_LOCK set, if rc is not 0,
1381 * mds_intent_policy won't try to return any locks */
1382 if (rc && child_lockh->cookie)
1383 ldlm_lock_decref(child_lockh, child_mode);
1385 if (dparent == NULL)
1390 ldlm_lock_decref(&parent_lockh, parent_mode);
1392 ptlrpc_save_lock(req, &parent_lockh, parent_mode);
1394 /* trigger dqacq on the owner of child and parent */
1395 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1401 /* Close a "file descriptor" and possibly unlink an orphan from the
1402 * PENDING directory. Caller must hold child->i_mutex, this drops it.
1404 * If we are being called from mds_disconnect() because the client has
1405 * disappeared, then req == NULL and we do not update last_rcvd because
1406 * there is nothing that could be recovered by the client at this stage
1407 * (it will not even _have_ an entry in last_rcvd anymore).
1409 * Returns EAGAIN if the client needs to get more data and re-close. */
1410 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1411 struct obd_device *obd, struct mds_file_data *mfd,
1412 int unlink_orphan, struct lov_mds_md *lmm, int lmm_size,
1413 struct llog_cookie *logcookies, int cookies_size,
1416 struct inode *inode = mfd->mfd_dentry->d_inode;
1417 char fidname[LL_FID_NAMELEN];
1418 int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1419 struct dentry *pending_child = NULL;
1420 struct mds_obd *mds = &obd->u.mds;
1421 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1422 void *handle = NULL;
1423 struct mds_body *request_body = NULL, *reply_body = NULL;
1424 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1425 struct iattr iattr = { 0 };
1428 if (req && req->rq_reqmsg != NULL)
1429 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1430 sizeof(*request_body));
1431 if (req && req->rq_repmsg != NULL)
1432 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1433 sizeof(*reply_body));
1435 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1437 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname,
1438 inode->i_nlink, mds_orphan_open_count(inode));
1440 last_orphan = (mds_orphan_open_dec_test(inode) &&
1441 mds_inode_is_orphan(inode) &&
1442 !obd->obd_recovering);
1444 /* this is half of the actual "close" */
1445 if (mfd->mfd_mode & FMODE_WRITE) {
1446 rc = mds_put_write_access(mds, inode, request_body,
1447 last_orphan && unlink_orphan);
1448 } else if (mfd->mfd_mode & MDS_FMODE_EXEC) {
1449 mds_allow_write_access(inode);
1451 /* here writecount change also needs protection from orphan write sem.
1452 * so drop orphan write sem after mds_put_write_access, bz 12888. */
1453 MDS_UP_WRITE_ORPHAN_SEM(inode);
1455 if (last_orphan && unlink_orphan) {
1456 int stripe_count = 0;
1457 /* mds_put_write_access must have succeeded */
1458 LASSERTF(rc == 0, "inode %lu/%u: rc %d",
1459 inode->i_ino, inode->i_generation, rc);
1461 CDEBUG(D_INODE, "destroying orphan object %s\n", fidname);
1463 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1464 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1465 CERROR("found \"orphan\" %s %s with link count %d\n",
1466 S_ISREG(inode->i_mode) ? "file" : "dir",
1467 fidname, inode->i_nlink);
1469 /* Sadly, there is no easy way to save pending_child from
1470 * mds_reint_unlink() into mfd, so we need to re-lookup,
1471 * but normally it will still be in the dcache. */
1472 LOCK_INODE_MUTEX(pending_dir);
1473 cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */
1474 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1476 if (IS_ERR(pending_child))
1477 GOTO(cleanup, rc = PTR_ERR(pending_child));
1478 LASSERT(pending_child->d_inode != NULL);
1480 cleanup_phase = 2; /* dput(pending_child) when finished */
1481 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1482 rc = ll_vfs_rmdir(pending_dir, pending_child,
1485 CERROR("error unlinking orphan dir %s: rc %d\n",
1491 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1494 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1495 NULL, stripe_count);
1496 if (IS_ERR(handle)) {
1497 rc = PTR_ERR(handle);
1501 if (lmm != NULL && (*valid & OBD_MD_FLEASIZE) &&
1502 mds_log_op_unlink(obd, lmm, lmm_size,
1503 logcookies, cookies_size) > 0) {
1504 *valid |= OBD_MD_FLCOOKIE;
1509 pending_child->d_fsdata = (void *) &dp;
1510 rc = ll_vfs_unlink(pending_dir, pending_child, mds->mds_vfsmnt);
1512 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1514 goto out; /* Don't bother updating attrs on unlinked inode */
1517 if (request_body != NULL) {
1518 /* Only start a transaction to write out only the atime if it
1519 * is more out-of-date than the specified limit. If we are
1520 * already going to write out the atime then do it anyway. */
1521 if ((request_body->valid & OBD_MD_FLATIME) &&
1522 ((LTIME_S(iattr.ia_atime) >
1523 LTIME_S(inode->i_atime) + mds->mds_atime_diff) ||
1524 (iattr.ia_valid != 0 &&
1525 LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime)))) {
1526 LTIME_S(iattr.ia_atime) = request_body->atime;
1527 iattr.ia_valid |= ATTR_ATIME;
1530 /* Store a rough estimate of the file size on the MDS for
1531 * tools like e2scan and HSM that are just using this for *
1532 * rough decision making and will get the proper size later.
1533 * * This is NOT guaranteed to be correct with multiple *
1534 * writers, but is only needed until SOM is done. b=11063 */
1535 if ((request_body->valid & OBD_MD_FLSIZE) &&
1536 (iattr.ia_valid != 0)) {
1537 iattr.ia_size = request_body->size;
1538 iattr.ia_valid |= ATTR_SIZE;
1542 if (iattr.ia_valid != 0) {
1543 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1544 if (IS_ERR(handle)) {
1545 rc = PTR_ERR(handle);
1549 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1551 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1554 /* If other clients have this file open for write, rc will be > 0 */
1557 l_dput(mfd->mfd_dentry);
1561 if (req != NULL && reply_body != NULL) {
1562 rc = mds_finish_transno(mds, NULL, handle, req, rc, 0, 0);
1563 } else if (handle) {
1564 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1566 CERROR("error committing close: %d\n", err);
1572 switch (cleanup_phase) {
1574 dput(pending_child);
1576 UNLOCK_INODE_MUTEX(pending_dir);
1581 int mds_close(struct ptlrpc_request *req, int offset)
1583 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1584 struct obd_device *obd = req->rq_export->exp_obd;
1585 struct mds_body *body;
1586 struct mds_file_data *mfd;
1587 struct lvfs_run_ctxt saved;
1588 struct inode *inode;
1589 int rc, repsize[4] = { sizeof(struct ptlrpc_body),
1590 sizeof(struct mds_body),
1591 obd->u.mds.mds_max_mdsize,
1592 obd->u.mds.mds_max_cookiesize };
1593 struct mds_body *reply_body;
1594 struct lov_mds_md *lmm;
1596 struct llog_cookie *logcookies;
1600 rc = lustre_pack_reply(req, 4, repsize, NULL);
1602 req->rq_status = rc;
1603 /* continue on to drop local open even if we can't send reply */
1605 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1607 CDEBUG(D_INODE, "close req->rep_len %d mdsize %d cookiesize %d\n",
1609 obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize);
1610 mds_counter_incr(req->rq_export, LPROC_MDS_CLOSE);
1612 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1613 lustre_swab_mds_body);
1615 CERROR("Can't unpack body\n");
1616 req->rq_status = -EFAULT;
1620 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1621 /* do some stuff */ ;
1623 spin_lock(&med->med_open_lock);
1624 mfd = mds_handle2mfd(&body->handle);
1626 spin_unlock(&med->med_open_lock);
1627 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1628 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1629 req->rq_status = -ESTALE;
1632 /* Remove mfd handle so it can't be found again. We consume mfd_list
1633 * reference here, but still have mds_handle2mfd ref until mfd_close. */
1634 mds_mfd_unlink(mfd, 1);
1635 spin_unlock(&med->med_open_lock);
1637 inode = mfd->mfd_dentry->d_inode;
1638 /* child orphan sem protects orphan_dec_test && is_orphan race */
1639 MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */
1640 if (!obd->obd_recovering &&
1641 mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1642 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1644 LASSERT(body != NULL);
1646 mds_pack_inode2body(body, inode);
1647 mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode,
1648 MDS_PACK_MD_LOCK, 0,
1649 req->rq_export->exp_connect_flags);
1652 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1653 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body));
1654 lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, 0);
1655 lmm_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 1),
1656 logcookies = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 2, 0);
1657 cookies_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 2);
1658 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1,
1659 lmm, lmm_size, logcookies, cookies_size,
1660 &reply_body->valid);
1661 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1663 mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1);
1664 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1665 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1666 req->rq_status = -ENOMEM;
1673 int mds_done_writing(struct ptlrpc_request *req, int offset)
1675 struct mds_body *body;
1676 int rc, size[2] = { sizeof(struct ptlrpc_body),
1677 sizeof(struct mds_body) };
1680 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1682 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1683 lustre_swab_mds_body);
1685 CERROR("Can't unpack body\n");
1686 req->rq_status = -EFAULT;
1690 rc = lustre_pack_reply(req, 2, size, NULL);
1692 req->rq_status = rc;