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_PRIVATE_DATA(inode) == NULL);
142 OBD_ALLOC(INODE_PRIVATE_DATA(inode), sizeof(struct mds_filter_data));
143 if (INODE_PRIVATE_DATA(inode) == 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_PRIVATE_DATA(inode) != NULL);
153 OBD_FREE(INODE_PRIVATE_DATA(inode), sizeof(struct mds_filter_data));
154 INODE_PRIVATE_DATA(inode) = 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 (MDS_FILTERDATA(inode) == NULL)
190 mds_alloc_filterdata(inode);
191 if (MDS_FILTERDATA(inode) == 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;
300 body->handle.cookie = mfd->mfd_handle.h_cookie;
302 if (req->rq_export->exp_disconnected) {
303 mds_mfd_unlink(mfd, 0);
304 MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode);
305 mds_mfd_close(NULL, REQ_REC_OFF, req->rq_export->exp_obd,
306 mfd, 0, NULL, 0, NULL, 0, NULL);
308 spin_lock(&med->med_open_lock);
309 list_add(&mfd->mfd_list, &med->med_open_head);
310 spin_unlock(&med->med_open_lock);
317 mds_mfd_unlink(mfd, 1);
319 return ERR_PTR(error);
322 /* Must be called with i_mutex held */
323 static int mds_create_objects(struct ptlrpc_request *req, int offset,
324 struct mds_update_record *rec,
325 struct mds_obd *mds, struct obd_device *obd,
326 struct dentry *dchild, void **handle,
327 struct lov_mds_md **objid)
329 struct inode *inode = dchild->d_inode;
330 struct obd_trans_info oti = { 0 };
331 struct lov_mds_md *lmm = NULL;
333 struct mds_body *body;
334 struct obd_info oinfo = { { { 0 } } };
340 if (!S_ISREG(inode->i_mode))
342 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
343 !(rec->ur_flags & FMODE_WRITE))
346 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
348 if (body->valid & OBD_MD_FLEASIZE)
354 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
355 if (rec->ur_fid2->id == 0) {
356 DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
360 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
361 lmm_size = rec->ur_eadatalen;
362 lmm = rec->ur_eadata;
364 if (lov_mds_md_size(lmm->lmm_stripe_count,
365 lmm->lmm_magic) != lmm_size)
366 CWARN("Bad lmm_size during open replay for inode "
367 "%lu\n", inode->i_ino);
369 if (*handle == NULL) {
370 int stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
371 *handle = fsfilt_start_log(obd, inode, FSFILT_OP_CREATE,
374 if (IS_ERR(*handle)) {
375 rc = PTR_ERR(*handle);
380 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
382 CERROR("open replay failed to set md:%d\n", rc);
384 /* for replay we not need send lmm to client, this not used now */
385 lustre_shrink_reply(req, offset, 0, 1);
391 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
392 GOTO(out_ids, rc = -ENOMEM);
394 OBDO_ALLOC(oinfo.oi_oa);
395 if (oinfo.oi_oa == NULL)
396 GOTO(out_ids, rc = -ENOMEM);
397 oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */
398 oinfo.oi_oa->o_gid = 0;
399 oinfo.oi_oa->o_mode = S_IFREG | 0600;
400 oinfo.oi_oa->o_id = inode->i_ino;
401 oinfo.oi_oa->o_gr = 0;
402 oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS |
403 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID| OBD_MD_FLGROUP;
404 oinfo.oi_oa->o_size = 0;
406 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
407 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
408 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
409 /* check if things like lfs setstripe are sending us the ea */
410 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
411 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
413 0, &oinfo.oi_md, rec->ur_eadata);
417 __u32 lmm_sz = mds->mds_max_mdsize;
418 OBD_ALLOC(lmm, lmm_sz);
420 GOTO(out_oa, rc = -ENOMEM);
423 rc = mds_get_md(obd, dchild->d_parent->d_inode,
424 lmm, &lmm_size, 1, 0,
425 req->rq_export->exp_connect_flags);
427 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
429 0, &oinfo.oi_md, lmm);
430 OBD_FREE(lmm, lmm_sz);
434 rc = obd_create(mds->mds_lov_exp, oinfo.oi_oa,
440 CDEBUG(level, "error creating objects for "
441 "inode %lu: rc = %d\n",
444 CERROR("obd_create returned invalid "
451 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_lov_exp,
452 0, &oinfo.oi_md, rec->ur_eadata);
456 oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id;
458 if (i_size_read(inode)) {
459 oinfo.oi_oa->o_size = i_size_read(inode);
460 obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE |
461 OBD_MD_FLATIME | OBD_MD_FLMTIME |
462 OBD_MD_FLCTIME | OBD_MD_FLSIZE);
464 /* pack lustre id to OST */
465 oinfo.oi_oa->o_fid = body->fid1.id;
466 oinfo.oi_oa->o_generation = body->fid1.generation;
467 oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER;
468 oinfo.oi_policy.l_extent.start = i_size_read(inode);
469 oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
471 rc = obd_punch_rqset(mds->mds_lov_exp, &oinfo, &oti);
473 CERROR("error setting attrs for inode %lu: rc %d\n",
476 CERROR("obd_setattr_async returned bad rc %d\n",
484 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
485 obdo_refresh_inode(inode, oinfo.oi_oa, OBD_MD_FLBLKSZ);
487 LASSERT(oinfo.oi_md && oinfo.oi_md->lsm_object_id);
489 rc = obd_packmd(mds->mds_lov_exp, &lmm, oinfo.oi_md);
491 CERROR("cannot pack lsm, err = %d\n", rc);
495 body->eadatasize = rc;
498 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
499 if (IS_ERR(*handle)) {
500 rc = PTR_ERR(*handle);
502 GOTO(free_diskmd, rc);
505 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov");
506 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size);
508 memcpy(lmm_buf, lmm, lmm_size);
509 *objid = lmm_buf; /* save for mds_lov_update_objid */
512 obd_free_diskmd(mds->mds_lov_exp, &lmm);
514 oti_free_cookies(&oti);
515 OBDO_FREE(oinfo.oi_oa);
518 obd_free_memmd(mds->mds_lov_exp, &oinfo.oi_md);
522 static void reconstruct_open(struct mds_update_record *rec, int offset,
523 struct ptlrpc_request *req,
524 struct lustre_handle *child_lockh)
526 struct mds_export_data *med = &req->rq_export->exp_mds_data;
527 struct lsd_client_data *lcd = med->med_lcd;
528 struct mds_obd *mds = mds_req2mds(req);
529 struct mds_file_data *mfd = NULL;
530 struct obd_export *exp = req->rq_export;
531 struct obd_device *obd = exp->exp_obd;
532 struct dentry *parent, *dchild;
533 struct ldlm_reply *rep;
534 struct mds_body *body;
540 LASSERT(offset == DLM_INTENT_REC_OFF); /* only called via intent */
541 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
542 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body));
544 /* copy rc, transno and disp; steal locks */
545 mds_req_from_lcd(req, lcd);
546 ldlm_reply_set_disposition(rep, le32_to_cpu(lcd->lcd_last_data));
548 /* Only replay if create or open actually happened. */
549 if (!ldlm_reply_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
551 return; /* error looking up parent or child */
554 /* If we failed, then we must have failed opening, so don't look for
555 * file descriptor or anything, just give the client the bad news.
557 if (req->rq_status) {
562 /* Now let's see if we have file descriptor present.
563 * No need to lookup child as it could be already deleted by another
564 * thread (bug 15010) */
565 spin_lock(&med->med_open_lock);
566 list_for_each(t, &med->med_open_head) {
567 mfd = list_entry(t, struct mds_file_data, mfd_list);
568 if (mfd->mfd_xid == req->rq_xid) {
574 spin_unlock(&med->med_open_lock);
576 if (mfd && mfd->mfd_dentry && mfd->mfd_dentry->d_inode) {
577 dchild = mfd->mfd_dentry;
580 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
581 if (IS_ERR(parent)) {
582 rc = PTR_ERR(parent);
583 LCONSOLE_WARN("Parent "LPU64"/%u lookup error %d."
584 " Evicting client %s with export %s.\n",
585 rec->ur_fid1->id,rec->ur_fid1->generation,
586 rc, obd_uuid2str(&exp->exp_client_uuid),
587 obd_export_nid2str(exp));
588 mds_export_evict(exp);
593 dchild = mds_lookup(obd, rec->ur_name, parent, rec->ur_namelen - 1);
595 if (IS_ERR(dchild)) {
596 rc = PTR_ERR(dchild);
597 LCONSOLE_WARN("Child "LPU64"/%u lookup error %d."
598 " Evicting client %s with export %s.\n",
599 rec->ur_fid1->id,rec->ur_fid1->generation,
600 rc, obd_uuid2str(&exp->exp_client_uuid),
601 obd_export_nid2str(exp));
602 mds_export_evict(exp);
608 if (!dchild->d_inode)
609 GOTO(out_dput, 0); /* child not present to open */
611 /* At this point, we know we have a child. We'll send
612 * it back _unless_ it not created and open failed.
614 if (ldlm_reply_disposition(rep, DISP_OPEN_OPEN) &&
615 !ldlm_reply_disposition(rep, DISP_OPEN_CREATE) &&
620 mds_pack_inode2body(body, dchild->d_inode);
621 if (S_ISREG(dchild->d_inode->i_mode)) {
622 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
623 body, dchild->d_inode, 1, 0,
624 req->rq_export->exp_connect_flags);
627 LASSERT(rc == req->rq_status);
629 /* If we have LOV EA data, the OST holds size, mtime */
630 if (!(body->valid & OBD_MD_FLEASIZE))
631 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
632 OBD_MD_FLATIME | OBD_MD_FLMTIME);
635 if (!(rec->ur_flags & MDS_OPEN_JOIN_FILE))
636 lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1,
637 body->eadatasize, 0);
639 if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
640 !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
641 int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1);
643 rc = mds_pack_acl(med, dchild->d_inode, req->rq_repmsg,
645 lustre_shrink_reply(req, acl_off, body->aclsize, 0);
646 if (!req->rq_status && rc)
650 /* #warning "XXX fixme" bug 2991 */
651 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
652 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
653 * to detect a re-open */
655 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
657 /* Allow file join in beta builds to allow debugging */
658 rc = mds_join_file(rec, req, dchild, NULL);
662 CWARN("file join is not supported in this version of "
664 GOTO(out_dput, req->rq_status = rc = -EOPNOTSUPP);
667 mntget(mds->mds_vfsmnt);
668 CERROR("Re-opened file \n");
669 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
670 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
671 mntput(mds->mds_vfsmnt);
673 req->rq_status = PTR_ERR(mfd);
675 CERROR("%s: opening inode "LPU64" failed: rc %d\n",
676 req->rq_export->exp_obd->obd_name,
677 (__u64)dchild->d_inode->i_ino, req->rq_status);
678 GOTO(out_dput, req->rq_status);
681 body->handle.cookie = mfd->mfd_handle.h_cookie;
682 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
683 mfd->mfd_handle.h_cookie);
686 if (!ldlm_reply_disposition(rep, DISP_OPEN_LOCK))
689 /* child_lockh has been set in fixup_handle_for_resent_req called
690 * in mds_intent_policy for resent request */
691 if (child_lockh == NULL || !lustre_handle_is_used(child_lockh)) {
692 /* the lock is already canceled! clear DISP_OPEN_LOCK */
693 ldlm_reply_clear_disposition(rep, DISP_OPEN_LOCK);
705 /* if client disconnects during recovery it may resend opens which were replayed
706 * on server but their transno less then last_transno on server so they will not
707 * be detected as reconstructs */
708 static int open_replay_reconstruct(struct ptlrpc_request *req)
710 struct mds_export_data *med = &req->rq_export->exp_mds_data;
711 struct mds_file_data *mfd = NULL;
714 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
717 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
720 /* if mfd exists then replay was done already */
721 spin_lock(&med->med_open_lock);
722 list_for_each(t, &med->med_open_head) {
723 mfd = list_entry(t, struct mds_file_data, mfd_list);
724 if (mfd->mfd_xid == req->rq_xid) {
730 spin_unlock(&med->med_open_lock);
733 struct mds_body *body = lustre_msg_buf(req->rq_repmsg,
736 __u64 *pre_versions = lustre_msg_get_versions(req->rq_reqmsg);
738 body->handle.cookie = mfd->mfd_handle.h_cookie;
739 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
740 mfd->mfd_handle.h_cookie);
742 lustre_msg_set_versions(req->rq_repmsg, pre_versions);
743 lustre_msg_set_transno(req->rq_repmsg,
744 lustre_msg_get_transno(req->rq_reqmsg));
745 lustre_msg_set_status(req->rq_repmsg, 0);
751 /* do NOT or the MAY_*'s, you'll get the weakest */
752 static int accmode(struct inode *inode, int flags)
756 /* Sadly, NFSD reopens a file repeatedly during operation, so the
757 * "acc_mode = 0" allowance for newly-created files isn't honoured.
758 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
759 * owner can write to a file even if it is marked readonly to hide
760 * its brokenness. (bug 5781) */
761 if (flags & MDS_OPEN_OWNEROVERRIDE && inode->i_uid == current->fsuid)
764 if (flags & FMODE_READ)
766 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
768 if (flags & MDS_FMODE_EXEC)
773 /* Handles object creation, actual opening, and I/O epoch */
774 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
775 struct mds_body *body, int flags, void **handle,
776 struct mds_update_record *rec,struct ldlm_reply *rep,
777 struct lustre_handle *lockh)
779 struct mds_obd *mds = mds_req2mds(req);
780 struct obd_device *obd = req->rq_export->exp_obd;
781 struct mds_file_data *mfd = NULL;
782 struct lov_mds_md *lmm = NULL; /* object IDs created */
786 /* atomically create objects if necessary */
787 LOCK_INODE_MUTEX(dchild->d_inode);
789 if (S_ISREG(dchild->d_inode->i_mode) &&
790 !(body->valid & OBD_MD_FLEASIZE)) {
791 rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1,
792 body, dchild->d_inode, 0, 0,
793 req->rq_export->exp_connect_flags);
795 UNLOCK_INODE_MUTEX(dchild->d_inode);
800 if ((body->valid & OBD_MD_FLEASIZE) &&
801 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
802 UNLOCK_INODE_MUTEX(dchild->d_inode);
805 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
807 /* Allow file join in beta builds to allow debugging */
808 UNLOCK_INODE_MUTEX(dchild->d_inode);
809 rc = mds_join_file(rec, req, dchild, lockh);
812 LOCK_INODE_MUTEX(dchild->d_inode);
814 CWARN("file join is not supported in this version of "
819 if (!(body->valid & OBD_MD_FLEASIZE) &&
820 !(body->valid & OBD_MD_FLMODEASIZE)) {
821 /* split open transactions here */
822 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_SPLIT_OPEN, 10);
823 /* no EA: create objects */
824 rc = mds_create_objects(req, DLM_REPLY_REC_OFF + 1, rec,
825 mds, obd, dchild, handle, &lmm);
827 CERROR("mds_create_objects: rc = %d\n", rc);
828 UNLOCK_INODE_MUTEX(dchild->d_inode);
833 /* If the inode has no EA data, then MDS holds size, mtime */
834 if (S_ISREG(dchild->d_inode->i_mode) &&
835 !(body->valid & OBD_MD_FLEASIZE)) {
836 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
837 OBD_MD_FLATIME | OBD_MD_FLMTIME);
839 UNLOCK_INODE_MUTEX(dchild->d_inode);
841 if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL &&
842 rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
843 int acl_off = DLM_REPLY_REC_OFF + 2;
845 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
846 dchild->d_inode, req->rq_repmsg,
852 if ((rc = mds_lov_prepare_objids(obd,lmm)) != 0)
855 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
857 RETURN(PTR_ERR(mfd));
859 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
860 mfd->mfd_handle.h_cookie);
862 mds_lov_update_objids(obd, lmm);
865 mds_mfd_unlink(mfd, 1);
871 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
872 struct mds_body *body, int flags,
873 struct mds_update_record *rec,struct ldlm_reply *rep)
875 struct obd_device *obd = req->rq_export->exp_obd;
876 struct mds_obd *mds = mds_req2mds(req);
877 struct dentry *dchild, *dparent = NULL;
878 char fidname[LL_FID_NAMELEN];
881 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
884 ldlm_reply_set_disposition(rep, DISP_LOOKUP_EXECD);
885 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
886 dchild = mds_lookup(obd, fidname, mds->mds_pending_dir, fidlen);
887 if (IS_ERR(dchild)) {
888 rc = PTR_ERR(dchild);
889 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
893 if (dchild->d_inode != NULL) {
894 mds_inode_set_orphan(dchild->d_inode);
895 CWARN("Orphan %s found and opened in PENDING directory\n",
898 __u64 *pre_versions = lustre_msg_get_versions(req->rq_reqmsg);
901 /* We didn't find it in PENDING so it isn't an orphan. See
902 * if it was a regular inode that was previously created. */
903 dchild = mds_fid2dentry(mds, fid, NULL);
905 RETURN(PTR_ERR(dchild));
908 * this can be replay of partially committed open|create,
909 * the create itself was committed while LOV EA weren't
910 * We need to set versions again if conditions are:
912 * - the transaction is greater than last_committed
913 * - this was open|create
914 * - there was real create so parent pre_version was saved
916 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
917 (lustre_msg_get_transno(req->rq_reqmsg) >
918 req->rq_export->exp_last_committed) &&
919 (rec->ur_flags & MDS_OPEN_CREAT) &&
920 (pre_versions && pre_versions[0] != 0)) {
921 /* need parent to set version */
922 dparent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
923 if (IS_ERR(dparent)) {
924 CERROR("Can't find parent for open replay\n");
926 RETURN(PTR_ERR(dparent));
928 /* though file was created, the versions were not
929 * changed yet, need to replay that too */
930 inodes[0] = dparent->d_inode;
931 inodes[1] = dchild->d_inode;
935 mds_pack_inode2body(body, dchild->d_inode);
936 ldlm_reply_set_disposition(rep, DISP_LOOKUP_POS);
937 ldlm_reply_set_disposition(rep, DISP_OPEN_OPEN);
939 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep, NULL);
940 rc = mds_finish_transno(mds, inodes, handle,
941 req, rc, rep ? rep->lock_policy_res1 : 0, 0);
942 /* XXX what do we do here if mds_finish_transno itself failed? */
949 int mds_pin(struct ptlrpc_request *req, int offset)
951 struct obd_device *obd = req->rq_export->exp_obd;
952 struct mds_body *reqbody, *repbody;
953 struct lvfs_run_ctxt saved;
954 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
957 reqbody = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*reqbody));
959 rc = lustre_pack_reply(req, 2, size, NULL);
963 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
966 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
967 rc = mds_open_by_fid(req, &reqbody->fid1, repbody, reqbody->flags, NULL,
969 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
974 /* Get an internal lock on the inode number (but not generation) to sync
975 * new inode creation with inode unlink (bug 2029). If child_lockh is NULL
976 * we just get the lock as a barrier to wait for other holders of this lock,
977 * and drop it right away again. */
978 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
979 struct lustre_handle *child_lockh)
981 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
982 struct lustre_handle lockh;
983 int lock_flags = LDLM_FL_ATOMIC_CB;
986 if (child_lockh == NULL)
987 child_lockh = &lockh;
989 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
990 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
991 ldlm_blocking_ast, ldlm_completion_ast,
992 NULL, NULL, 0, NULL, child_lockh);
994 CERROR("ldlm_cli_enqueue_local: %d\n", rc);
995 else if (child_lockh == &lockh)
996 ldlm_lock_decref(child_lockh, LCK_EX);
1001 int mds_open(struct mds_update_record *rec, int offset,
1002 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
1004 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
1005 struct obd_device *obd = req->rq_export->exp_obd;
1006 struct mds_obd *mds = mds_req2mds(req);
1007 struct ldlm_reply *rep = NULL;
1008 struct mds_body *body = NULL;
1009 struct dentry *dchild = NULL, *dparent = NULL;
1010 struct inode *inodes[PTLRPC_NUM_VERSIONS] = { NULL };
1011 struct mds_export_data *med;
1012 struct lustre_handle parent_lockh;
1013 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
1014 int parent_mode = LCK_CR;
1015 void *handle = NULL;
1016 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1017 unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid };
1018 unsigned int qpids[MAXQUOTAS] = { 0, 0 };
1019 unsigned int ids[MAXQUOTAS] = { 0, 0 };
1020 int child_mode = LCK_CR;
1021 /* Always returning LOOKUP lock if open succesful to guard
1022 dentry on client. */
1024 int quota_pending[2] = {0, 0};
1025 int use_parent, need_open_lock;
1026 unsigned int gid = current->fsgid;
1029 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
1031 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
1032 (obd_timeout + 1) / 4);
1034 CLASSERT(MAXQUOTAS < 4);
1035 if (offset == DLM_INTENT_REC_OFF) { /* intent */
1036 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
1038 body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF,
1040 } else if (offset == REQ_REC_OFF) { /* non-intent reint */
1041 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1043 LBUG(); /* XXX: not supported yet? */
1049 /* check the open resent|replay case */
1050 if (open_replay_reconstruct(req))
1053 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
1055 /* Step 0: If we are passed a fid, then we assume the client already
1056 * opened this file and is only replaying the RPC, so we open the
1057 * inode by fid (at some large expense in security). */
1058 /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */
1059 if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
1060 (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&&
1061 !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) {
1062 if (rec->ur_fid2->id == 0) {
1063 struct ldlm_lock *lock = ldlm_handle2lock(child_lockh);
1065 LDLM_ERROR(lock, "fid2 not set on open replay");
1066 LDLM_LOCK_PUT(lock);
1068 DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay");
1072 /** check there is no stale orphan with same inode number */
1073 if (rec->ur_flags & MDS_OPEN_CREAT) {
1074 rc = mds_check_stale_orphan(obd, rec->ur_fid2);
1079 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
1081 if (rc != -ENOENT) {
1082 if (req->rq_export->exp_libclient &&
1083 rec->ur_flags & MDS_OPEN_HAS_EA)
1089 /* We didn't find the correct inode on disk either, so we
1090 * need to re-create it via a regular replay. */
1091 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1092 DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay");
1095 } else if (rec->ur_fid2->id) {
1096 DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay",
1097 rec->ur_fid2->id, rec->ur_fid2->generation);
1101 /* If we got here, we must be called via intent */
1102 LASSERT(offset == DLM_INTENT_REC_OFF);
1104 med = &req->rq_export->exp_mds_data;
1105 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
1106 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
1110 if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE))
1111 parent_mode = LCK_EX;
1113 /* We cannot use acc_mode here, because it is zeroed in case of
1114 creating a file, so we get wrong lockmode */
1115 if (rec->ur_flags & FMODE_WRITE)
1116 child_mode = LCK_CW;
1117 else if (rec->ur_flags & MDS_FMODE_EXEC)
1118 child_mode = LCK_PR;
1120 child_mode = LCK_CR;
1122 /* join file and nfsd can't need lookup dchild as use parent for it */
1123 use_parent = (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1124 (rec->ur_flags & MDS_OPEN_LOCK) && (rec->ur_namelen == 1)) ||
1125 (rec->ur_flags & MDS_OPEN_JOIN_FILE);
1127 need_open_lock = !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) &&
1128 (rec->ur_flags & MDS_OPEN_LOCK);
1130 /* Try to lock both parent and child first. If child is not found,
1131 * return only locked parent. This is enough to prevent other
1132 * threads from changing this directory until creation is finished. */
1133 rc = mds_get_parent_child_locked(obd, &obd->u.mds,
1136 &dparent, parent_mode,
1137 MDS_INODELOCK_UPDATE,
1138 use_parent ? NULL : rec->ur_name,
1140 (rec->ur_flags & MDS_OPEN_LOCK) ?
1142 &dchild, child_mode,
1143 MDS_INODELOCK_LOOKUP |
1144 MDS_INODELOCK_OPEN);
1147 if (rc != -ENOENT) {
1148 CERROR("parent "LPU64"/%u lookup/take lock error %d\n",
1149 rec->ur_fid1->id, rec->ur_fid1->generation, rc);
1151 /* Just cannot find parent - make it look like
1152 * usual negative lookup to avoid extra MDS RPC */
1153 ldlm_reply_set_disposition(rep, DISP_LOOKUP_EXECD);
1154 ldlm_reply_set_disposition(rep, DISP_LOOKUP_NEG);
1158 LASSERT(dparent->d_inode != NULL);
1160 cleanup_phase = 1; /* parent dentry and lock */
1163 dchild = dget(dparent);
1165 if (rec->ur_flags & MDS_OPEN_JOIN_FILE) {
1166 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1167 GOTO(found_child, rc);
1170 cleanup_phase = 2; /* child dentry */
1172 ldlm_reply_set_disposition(rep, DISP_LOOKUP_EXECD);
1173 if (dchild->d_inode)
1174 ldlm_reply_set_disposition(rep, DISP_LOOKUP_POS);
1176 ldlm_reply_set_disposition(rep, DISP_LOOKUP_NEG);
1178 /*Step 3: If the child was negative, and we're supposed to, create it.*/
1179 if (dchild->d_inode == NULL) {
1180 unsigned long ino = rec->ur_fid2->id;
1182 struct inode *inode;
1184 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1185 /* It's negative and we weren't supposed to create it */
1186 GOTO(cleanup, rc = -ENOENT);
1189 if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY)
1190 GOTO(cleanup, rc = -EROFS);
1192 /** check there is no stale orphan with same inode number */
1193 rc = mds_check_stale_orphan(obd, rec->ur_fid2);
1197 /* version recovery check */
1198 rc = mds_version_get_check(req, dparent->d_inode, 0);
1200 GOTO(cleanup_no_trans, rc);
1202 if (dparent->d_inode->i_mode & S_ISGID)
1203 gid = dparent->d_inode->i_gid;
1205 gid = current->fsgid;
1207 /* we try to get enough quota to write here, and let ldiskfs
1208 * decide if it is out of quota or not b=14783
1209 * FIXME: after CMD is used, pointer to obd_trans_info* couldn't
1210 * be NULL, b=14840 */
1211 ids[0] = current->fsuid;
1213 lquota_chkquota(mds_quota_interface_ref, req->rq_export,
1214 ids[0], ids[1], 1, quota_pending,
1217 ldlm_reply_set_disposition(rep, DISP_OPEN_CREATE);
1218 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1220 if (IS_ERR(handle)) {
1221 rc = PTR_ERR(handle);
1225 dchild->d_fsdata = (void *) &dp;
1229 LOCK_INODE_MUTEX(dparent->d_inode);
1230 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1231 UNLOCK_INODE_MUTEX(dparent->d_inode);
1232 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1233 dchild->d_fsdata = NULL;
1236 CDEBUG(D_INODE, "error during create: %d\n", rc);
1239 inode = dchild->d_inode;
1242 if (ino != inode->i_ino) {
1243 /* FID support is needed to replay this
1244 * correctly. Now fail gracefully like there is
1245 * version mismatch */
1246 if (req->rq_export->exp_delayed)
1252 /* Written as part of setattr */
1253 inode->i_generation = rec->ur_fid2->generation;
1254 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1255 inode->i_ino, inode->i_generation);
1258 LTIME_S(iattr.ia_atime) = rec->ur_time;
1259 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1260 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1262 iattr.ia_uid = current->fsuid; /* set by push_ctxt already */
1265 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1266 ATTR_MTIME | ATTR_CTIME;
1268 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1270 CERROR("error on child setattr: rc = %d\n", rc);
1272 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1274 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1276 CERROR("error on parent setattr: rc = %d\n", rc);
1278 rc = fsfilt_commit(obd, dchild->d_inode, handle, 0);
1280 acc_mode = 0; /* Don't check for permissions */
1282 acc_mode = accmode(dchild->d_inode, rec->ur_flags);
1283 /* Child previously existed so the lookup and lock is already
1284 * done, so no further locking is needed. */
1285 /* for nfs and join - we need two locks for same fid, but
1286 * with different mode */
1287 if (need_open_lock && !use_parent) {
1288 ldlm_reply_set_disposition(rep, DISP_OPEN_LOCK);
1293 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1294 "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len,
1295 dchild->d_name.name, dchild, dchild->d_inode,
1296 dchild->d_inode->i_ino, dchild->d_inode->i_generation);
1299 mds_pack_inode2body(body, dchild->d_inode);
1300 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1301 (rec->ur_flags & MDS_OPEN_EXCL)) {
1302 /* File already exists, we didn't just create it, and we
1303 * were passed O_EXCL; err-or. */
1304 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1307 /* if we are following special file, don't open */
1308 if (!S_ISREG(dchild->d_inode->i_mode) &&
1309 !S_ISDIR(dchild->d_inode->i_mode))
1310 GOTO(cleanup_no_trans, rc = 0);
1312 ldlm_reply_set_disposition(rep, DISP_OPEN_OPEN);
1313 if (S_ISREG(dchild->d_inode->i_mode)) {
1314 /* Check permissions etc */
1315 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1319 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) &&
1320 (acc_mode & MAY_WRITE))
1321 GOTO(cleanup, rc = -EROFS);
1323 /* An append-only file must be opened in append mode for
1325 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1326 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1327 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1328 GOTO(cleanup, rc = -EPERM);
1330 if (S_ISDIR(dchild->d_inode->i_mode)) {
1331 if (rec->ur_flags & MDS_OPEN_CREAT ||
1332 rec->ur_flags & FMODE_WRITE) {
1333 /* we are trying to create or write a exist dir*/
1334 GOTO(cleanup, rc = -EISDIR);
1336 if (rec->ur_flags & MDS_FMODE_EXEC) {
1337 /* we are trying to exec a directory */
1338 GOTO(cleanup, rc = -EACCES);
1340 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1341 GOTO(cleanup, rc = -EACCES);
1342 } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) {
1343 GOTO(cleanup, rc = -ENOTDIR);
1347 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1348 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1349 GOTO(cleanup, rc = -EAGAIN);
1352 if (need_open_lock) {
1353 ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN } };
1354 struct ldlm_res_id child_res_id;
1356 /* In case of replay we do not get a lock assuming that the
1357 caller has it already */
1358 child_res_id.name[0] = dchild->d_inode->i_ino;
1359 child_res_id.name[1] = dchild->d_inode->i_generation;
1361 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id,
1362 LDLM_IBITS, &policy, child_mode,
1363 &lock_flags, ldlm_blocking_ast,
1364 ldlm_completion_ast, NULL, NULL,
1365 0, NULL, child_lockh);
1369 /* Let mds_intent_policy know that we have a lock to return */
1370 ldlm_reply_set_disposition(rep, DISP_OPEN_LOCK);
1373 if (!S_ISREG(dchild->d_inode->i_mode) &&
1374 !S_ISDIR(dchild->d_inode->i_mode) &&
1375 (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) {
1376 /* If client supports this, do not return open handle for
1377 * special device nodes */
1378 GOTO(cleanup_no_trans, rc = 0);
1381 /* Step 5: mds_open it */
1382 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1383 rep, &parent_lockh);
1387 inodes[0] = (!created || IS_ERR(dparent)) ? NULL : dparent->d_inode;
1388 inodes[1] = (created && dchild) ? dchild->d_inode : NULL;
1389 rc = mds_finish_transno(mds, inodes, handle, req, rc,
1390 rep ? rep->lock_policy_res1 : 0, 0);
1393 if (quota_pending[0] || quota_pending[1])
1394 lquota_pending_commit(mds_quota_interface_ref, obd,
1395 ids[0], ids[1], quota_pending);
1396 switch (cleanup_phase) {
1398 if (rc && created) {
1400 LOCK_INODE_MUTEX(dparent->d_inode);
1401 err = ll_vfs_unlink(dparent->d_inode, dchild,
1403 UNLOCK_INODE_MUTEX(dparent->d_inode);
1405 CERROR("unlink(%.*s) in error path: %d\n",
1406 dchild->d_name.len, dchild->d_name.name,
1409 } else if (created) {
1410 mds_lock_new_child(obd, dchild->d_inode, NULL);
1411 /* save uid/gid for quota acquire/release */
1412 qpids[USRQUOTA] = dparent->d_inode->i_uid;
1413 qpids[GRPQUOTA] = dparent->d_inode->i_gid;
1418 /* It is safe to leave IT_OPEN_LOCK set, if rc is not 0,
1419 * mds_intent_policy won't try to return any locks */
1420 if (rc && child_lockh->cookie)
1421 ldlm_lock_decref(child_lockh, child_mode);
1423 if (dparent == NULL)
1428 ldlm_lock_decref(&parent_lockh, parent_mode);
1430 ptlrpc_save_lock(req, &parent_lockh, parent_mode);
1432 /* trigger dqacq on the owner of child and parent */
1433 lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc,
1439 /* Close a "file descriptor" and possibly unlink an orphan from the
1440 * PENDING directory. Caller must hold child->i_mutex, this drops it.
1442 * If we are being called from mds_disconnect() because the client has
1443 * disappeared, then req == NULL and we do not update last_rcvd because
1444 * there is nothing that could be recovered by the client at this stage
1445 * (it will not even _have_ an entry in last_rcvd anymore). */
1446 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1447 struct obd_device *obd, struct mds_file_data *mfd,
1448 int unlink_orphan, struct lov_mds_md *lmm, int lmm_size,
1449 struct llog_cookie *logcookies, int cookies_size,
1452 struct inode *inode = mfd->mfd_dentry->d_inode;
1453 char fidname[LL_FID_NAMELEN];
1454 int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1455 struct dentry *pending_child = NULL;
1456 struct mds_obd *mds = &obd->u.mds;
1457 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1458 void *handle = NULL;
1459 struct mds_body *request_body = NULL, *reply_body = NULL;
1460 struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT;
1461 struct iattr iattr = { 0 };
1464 if (req && req->rq_reqmsg != NULL)
1465 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1466 sizeof(*request_body));
1467 if (req && req->rq_repmsg != NULL)
1468 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1469 sizeof(*reply_body));
1471 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1473 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname,
1474 inode->i_nlink, mds_orphan_open_count(inode));
1476 last_orphan = (mds_orphan_open_dec_test(inode) &&
1477 mds_inode_is_orphan(inode) &&
1478 !obd->obd_recovering);
1480 /* this is half of the actual "close" */
1481 if (mfd->mfd_mode & FMODE_WRITE) {
1482 rc = mds_put_write_access(mds, inode, request_body,
1483 last_orphan && unlink_orphan);
1484 } else if (mfd->mfd_mode & MDS_FMODE_EXEC) {
1485 mds_allow_write_access(inode);
1487 /* here writecount change also needs protection from orphan write sem.
1488 * so drop orphan write sem after mds_put_write_access, bz 12888. */
1489 MDS_UP_WRITE_ORPHAN_SEM(inode);
1491 if (last_orphan && unlink_orphan) {
1492 int stripe_count = 0;
1493 /* mds_put_write_access must have succeeded */
1494 LASSERTF(rc == 0, "inode %lu/%u: rc %d",
1495 inode->i_ino, inode->i_generation, rc);
1497 CDEBUG(D_INODE, "destroying orphan object %s\n", fidname);
1499 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1500 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1501 CERROR("found \"orphan\" %s %s with link count %d\n",
1502 S_ISREG(inode->i_mode) ? "file" : "dir",
1503 fidname, inode->i_nlink);
1505 /* Sadly, there is no easy way to save pending_child from
1506 * mds_reint_unlink() into mfd, so we need to re-lookup,
1507 * but normally it will still be in the dcache. */
1508 LOCK_INODE_MUTEX(pending_dir);
1509 cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */
1510 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1512 if (IS_ERR(pending_child))
1513 GOTO(cleanup, rc = PTR_ERR(pending_child));
1514 LASSERT(pending_child->d_inode != NULL);
1516 cleanup_phase = 2; /* dput(pending_child) when finished */
1517 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1518 rc = ll_vfs_rmdir(pending_dir, pending_child,
1521 CERROR("error unlinking orphan dir %s: rc %d\n",
1527 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1530 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1531 NULL, stripe_count);
1532 if (IS_ERR(handle)) {
1533 rc = PTR_ERR(handle);
1537 if (lmm != NULL && (*valid & OBD_MD_FLEASIZE) &&
1538 mds_log_op_unlink(obd, lmm, lmm_size,
1539 logcookies, cookies_size) > 0) {
1540 *valid |= OBD_MD_FLCOOKIE;
1545 pending_child->d_fsdata = (void *) &dp;
1546 rc = ll_vfs_unlink(pending_dir, pending_child, mds->mds_vfsmnt);
1548 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1550 goto out; /* Don't bother updating attrs on unlinked inode */
1553 if (request_body != NULL) {
1554 /* Only start a transaction to write out only the atime if it
1555 * is more out-of-date than the specified limit. If we are
1556 * already going to write out the atime then do it anyway. */
1557 if ((request_body->valid & OBD_MD_FLATIME) &&
1558 ((LTIME_S(iattr.ia_atime) >
1559 LTIME_S(inode->i_atime) + mds->mds_atime_diff) ||
1560 (iattr.ia_valid != 0 &&
1561 LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime)))) {
1562 LTIME_S(iattr.ia_atime) = request_body->atime;
1563 iattr.ia_valid |= ATTR_ATIME;
1566 /* Store a rough estimate of the file size on the MDS for
1567 * tools like e2scan and HSM that are just using this for *
1568 * rough decision making and will get the proper size later.
1569 * * This is NOT guaranteed to be correct with multiple *
1570 * writers, but is only needed until SOM is done. b=11063 */
1571 if (S_ISREG(inode->i_mode) &&
1572 (request_body->valid & OBD_MD_FLSIZE) &&
1573 (iattr.ia_valid != 0)) {
1574 iattr.ia_size = request_body->size;
1575 iattr.ia_valid |= ATTR_SIZE;
1579 if (iattr.ia_valid != 0) {
1580 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1581 if (IS_ERR(handle)) {
1582 rc = PTR_ERR(handle);
1586 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1588 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1591 /* If other clients have this file open for write, rc will be > 0 */
1596 l_dput(mfd->mfd_dentry);
1598 if (req != NULL && reply_body != NULL) {
1599 rc = mds_finish_transno(mds, NULL, handle, req, rc, 0, 0);
1600 } else if (handle) {
1601 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1603 CERROR("error committing close: %d\n", err);
1609 switch (cleanup_phase) {
1611 dput(pending_child);
1613 UNLOCK_INODE_MUTEX(pending_dir);
1618 int mds_close(struct ptlrpc_request *req, int offset)
1620 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1621 struct obd_device *obd = req->rq_export->exp_obd;
1622 struct mds_body *body;
1623 struct mds_file_data *mfd;
1624 struct lvfs_run_ctxt saved;
1625 struct inode *inode;
1626 int rc, repsize[4] = { sizeof(struct ptlrpc_body),
1627 sizeof(struct mds_body),
1628 obd->u.mds.mds_max_mdsize,
1629 obd->u.mds.mds_max_cookiesize };
1630 struct mds_body *reply_body;
1631 struct lov_mds_md *lmm;
1633 struct llog_cookie *logcookies;
1637 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1638 lustre_swab_mds_body);
1640 CERROR("Can't unpack body\n");
1641 req->rq_status = -EFAULT;
1642 GOTO(cleanup, rc = -EFAULT);
1644 /*XXX need indicase - close is need to return LOV EA */
1645 body->valid |= OBD_MD_FLEASIZE;
1647 rc = lustre_pack_reply(req, 4, repsize, NULL);
1649 req->rq_status = rc;
1650 /* continue on to drop local open even if we can't send reply */
1652 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1654 CDEBUG(D_INODE, "close req->rep_len %d mdsize %d cookiesize %d\n",
1656 obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize);
1657 mds_counter_incr(req->rq_export, LPROC_MDS_CLOSE);
1659 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1660 /* do some stuff */ ;
1662 spin_lock(&med->med_open_lock);
1663 mfd = mds_handle2mfd(&body->handle);
1665 spin_unlock(&med->med_open_lock);
1666 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1667 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1668 req->rq_status = -ESTALE;
1669 GOTO(cleanup, rc = -ESTALE);
1671 /* Remove mfd handle so it can't be found again. We consume mfd_list
1672 * reference here, but still have mds_handle2mfd ref until mfd_close. */
1673 mds_mfd_unlink(mfd, 1);
1674 spin_unlock(&med->med_open_lock);
1676 inode = mfd->mfd_dentry->d_inode;
1677 /* child orphan sem protects orphan_dec_test && is_orphan race */
1678 MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */
1679 if (!obd->obd_recovering &&
1680 mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1681 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1683 LASSERT(body != NULL);
1685 mds_pack_inode2body(body, inode);
1686 mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode,
1687 MDS_PACK_MD_LOCK, 0,
1688 req->rq_export->exp_connect_flags);
1691 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1692 reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body));
1693 lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, 0);
1694 lmm_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 1),
1695 logcookies = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 2, 0);
1696 cookies_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 2);
1697 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1,
1698 lmm, lmm_size, logcookies, cookies_size,
1699 &reply_body->valid);
1700 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1703 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1704 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1705 req->rq_status = -ENOMEM;
1712 int mds_done_writing(struct ptlrpc_request *req, int offset)
1714 struct mds_body *body;
1715 int rc, size[2] = { sizeof(struct ptlrpc_body),
1716 sizeof(struct mds_body) };
1719 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1721 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1722 lustre_swab_mds_body);
1724 CERROR("Can't unpack body\n");
1725 req->rq_status = -EFAULT;
1729 rc = lustre_pack_reply(req, 2, size, NULL);
1731 req->rq_status = rc;