1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Andreas Dilger <adilger@clusterfs.com>
7 * Author: Phil Schwan <phil@clusterfs.com>
8 * Author: Mike Shaver <shaver@clusterfs.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27 # define EXPORT_SYMTAB
29 #define DEBUG_SUBSYSTEM S_MDS
31 #include <linux/version.h>
32 #include <linux/module.h>
33 #include <linux/init.h>
34 #include <linux/version.h>
35 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
36 # include <linux/buffer_head.h>
37 # include <linux/workqueue.h>
39 # include <linux/locks.h>
42 #include <linux/obd_class.h>
43 #include <linux/obd_lov.h>
44 #include <linux/lustre_fsfilt.h>
45 #include <linux/lprocfs_status.h>
47 #include "mds_internal.h"
49 /* Exported function from this file are:
51 * mds_open - called by the intent handler
52 * mds_close - an rpc handling function
53 * mds_pin - an rpc handling function - which will go away
54 * mds_mfd_close - for force closing files when a client dies
58 * MDS file data handling: file data holds a handle for a file opened
62 static void mds_mfd_addref(void *mfdp)
64 struct mds_file_data *mfd = mfdp;
66 atomic_inc(&mfd->mfd_refcount);
67 CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
68 atomic_read(&mfd->mfd_refcount));
71 struct mds_file_data *mds_mfd_new(void)
73 struct mds_file_data *mfd;
75 OBD_ALLOC(mfd, sizeof *mfd);
77 CERROR("mds: out of memory\n");
81 atomic_set(&mfd->mfd_refcount, 2);
83 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
84 class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
89 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
92 LASSERT(handle != NULL);
93 RETURN(class_handle2object(handle->cookie));
96 static void mds_mfd_put(struct mds_file_data *mfd)
98 CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
99 atomic_read(&mfd->mfd_refcount) - 1);
100 LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
101 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
102 if (atomic_dec_and_test(&mfd->mfd_refcount)) {
103 LASSERT(list_empty(&mfd->mfd_handle.h_link));
104 OBD_FREE(mfd, sizeof *mfd);
108 static void mds_mfd_destroy(struct mds_file_data *mfd)
110 class_handle_unhash(&mfd->mfd_handle);
114 #ifdef IFILTERDATA_ACTUALLY_USED
115 /* Caller must hold mds->mds_epoch_sem */
116 static int mds_alloc_filterdata(struct inode *inode)
118 LASSERT(inode->i_filterdata == NULL);
119 OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
120 if (inode->i_filterdata == NULL)
122 LASSERT(igrab(inode) == inode);
126 /* Caller must hold mds->mds_epoch_sem */
127 static void mds_free_filterdata(struct inode *inode)
129 LASSERT(inode->i_filterdata != NULL);
130 OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
131 inode->i_filterdata = NULL;
134 #endif /*IFILTERDATA_ACTUALLY_USED*/
136 /* Write access to a file: executors cause a negative count,
137 * writers a positive count. The semaphore is needed to perform
138 * a check for the sign and then increment or decrement atomically.
140 * This code is closely tied to the allocation of the d_fsdata and the
141 * MDS epoch, so we use the same semaphore for the whole lot.
143 * We could use a different semaphore for each file, if it ever shows
144 * up in a profile, which it won't.
146 * epoch argument is nonzero during recovery */
147 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
152 down(&mds->mds_epoch_sem);
154 if (atomic_read(&inode->i_writecount) < 0) {
155 up(&mds->mds_epoch_sem);
159 #ifdef IFILTERDATA_ACTUALLY_USED
160 if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
161 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
162 MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
163 inode->i_generation);
167 if (inode->i_filterdata == NULL)
168 mds_alloc_filterdata(inode);
169 if (inode->i_filterdata == NULL) {
173 #endif /*IFILTERDATA_ACTUALLY_USED*/
174 if (epoch > mds->mds_io_epoch)
175 mds->mds_io_epoch = epoch;
178 #ifdef IFILTERDATA_ACTUALLY_USED
179 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
180 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
181 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
183 #endif /*IFILTERDATA_ACTUALLY_USED*/
185 atomic_inc(&inode->i_writecount);
186 up(&mds->mds_epoch_sem);
190 /* Returns EAGAIN if the client needs to get size and/or cookies and close
191 * again -- which is never true if the file is about to be unlinked. Otherwise
192 * returns the number of remaining writers. */
193 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
194 struct mds_body *body, int unlinking)
199 down(&mds->mds_epoch_sem);
200 atomic_dec(&inode->i_writecount);
201 rc = atomic_read(&inode->i_writecount);
205 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
206 GOTO(out, rc = EAGAIN);
208 #ifdef IFILTERDATA_ACTUALLY_USED
209 mds_free_filterdata(inode);
212 up(&mds->mds_epoch_sem);
216 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
219 down(&mds->mds_epoch_sem);
220 if (atomic_read(&inode->i_writecount) > 0) {
221 up(&mds->mds_epoch_sem);
224 atomic_dec(&inode->i_writecount);
225 up(&mds->mds_epoch_sem);
229 static void mds_allow_write_access(struct inode *inode)
232 atomic_inc(&inode->i_writecount);
235 int mds_query_write_access(struct inode *inode)
238 RETURN(atomic_read(&inode->i_writecount));
241 /* This replaces the VFS dentry_open, it manages mfd and writecount */
242 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
243 struct vfsmount *mnt, int flags,
244 struct ptlrpc_request *req)
246 struct mds_export_data *med = &req->rq_export->exp_mds_data;
247 struct mds_obd *mds = mds_req2mds(req);
248 struct mds_file_data *mfd;
249 struct mds_body *body;
255 CERROR("mds: out of memory\n");
256 GOTO(cleanup_dentry, error = -ENOMEM);
259 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
261 if (flags & FMODE_WRITE) {
262 /* FIXME: in recovery, need to pass old epoch here */
263 error = mds_get_write_access(mds, dentry->d_inode, 0);
265 GOTO(cleanup_mfd, error);
266 #ifdef IFILTERDATA_ACTUALLY_USED
267 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
268 #endif /*IFILTERDATA_ACTUALLY_USED*/
269 } else if (flags & FMODE_EXEC) {
270 error = mds_deny_write_access(mds, dentry->d_inode);
272 GOTO(cleanup_mfd, error);
277 /* Mark the file as open to handle open-unlink. */
279 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
280 mds_orphan_open_inc(dentry->d_inode);
281 UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
283 mfd->mfd_mode = flags;
284 mfd->mfd_dentry = dentry;
285 mfd->mfd_xid = req->rq_xid;
287 spin_lock(&med->med_open_lock);
288 list_add(&mfd->mfd_list, &med->med_open_head);
289 spin_unlock(&med->med_open_lock);
292 body->handle.cookie = mfd->mfd_handle.h_cookie;
298 mds_mfd_destroy(mfd);
300 return ERR_PTR(error);
303 static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
304 struct lov_desc *desc)
307 for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
308 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
309 le64_to_cpu(lmm->lmm_objects[i].l_object_id);
313 /* Must be called with i_sem held */
314 static int mds_create_objects(struct ptlrpc_request *req, int offset,
315 struct mds_update_record *rec,
316 struct mds_obd *mds, struct obd_device *obd,
317 struct dentry *dchild, void **handle,
320 struct obdo *oa = NULL;
321 struct obd_trans_info oti = { 0 };
322 struct mds_body *body;
323 struct lov_stripe_md *lsm = NULL;
324 struct lov_mds_md *lmm = NULL;
325 struct inode *inode = dchild->d_inode;
327 int rc, lmm_bufsize, lmm_size;
330 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
331 !(rec->ur_flags & FMODE_WRITE))
334 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
336 if (!S_ISREG(inode->i_mode))
338 if (body->valid & OBD_MD_FLEASIZE)
341 OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
344 oti.oti_objid = *ids;
347 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
348 LASSERT(id_ino(rec->ur_id2));
349 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
350 lmm_size = rec->ur_eadatalen;
351 lmm = rec->ur_eadata;
355 *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
356 if (IS_ERR(*handle)) {
357 rc = PTR_ERR(*handle);
362 mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
364 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
365 lmm_bufsize = req->rq_repmsg->buflens[offset];
367 LASSERT(lmm_bufsize >= lmm_size);
368 memcpy(lmm_buf, lmm, lmm_size);
369 rc = fsfilt_set_md(obd, inode, *handle, lmm,
372 CERROR("open replay failed to set md:%d\n", rc);
376 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
377 GOTO(out_ids, rc = -ENOMEM);
381 GOTO(out_ids, rc = -ENOMEM);
382 oa->o_mode = S_IFREG | 0600;
383 oa->o_id = inode->i_ino;
384 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
385 oa->o_generation = inode->i_generation;
386 oa->o_uid = 0; /* must have 0 uid / gid on OST */
388 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
389 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
392 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
395 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
396 /* check if things like lfs setstripe are sending us the ea */
397 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
398 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
400 0, &lsm, rec->ur_eadata);
404 OBD_ALLOC(lmm, mds->mds_max_mdsize);
406 GOTO(out_oa, rc = -ENOMEM);
408 lmm_size = mds->mds_max_mdsize;
409 rc = mds_get_md(obd, dchild->d_parent->d_inode,
410 lmm, &lmm_size, 1, 0);
412 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
415 OBD_FREE(lmm, mds->mds_max_mdsize);
419 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
420 rc = obd_create(mds->mds_dt_exp, oa, &lsm, &oti);
425 CDEBUG(level, "error creating objects for "
426 "inode %lu: rc = %d\n",
429 CERROR("obd_create returned invalid "
436 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
437 0, &lsm, rec->ur_eadata);
441 lsm->lsm_object_id = oa->o_id;
442 lsm->lsm_object_gr = oa->o_gr;
445 oa->o_size = inode->i_size;
446 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
447 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
448 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
450 CERROR("error setting attrs for inode %lu: rc %d\n",
453 CERROR("obd_setattr returned bad rc %d\n", rc);
460 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
461 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
463 LASSERT(lsm && lsm->lsm_object_id);
465 rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
466 if (!id_ino(rec->ur_id2))
467 obd_free_memmd(mds->mds_dt_exp, &lsm);
470 body->eadatasize = rc;
473 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
474 if (IS_ERR(*handle)) {
475 rc = PTR_ERR(*handle);
480 rc = fsfilt_set_md(obd, inode, *handle, lmm,
483 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
484 lmm_bufsize = req->rq_repmsg->buflens[offset];
486 LASSERT(lmm_bufsize >= lmm_size);
488 memcpy(lmm_buf, lmm, lmm_size);
489 obd_free_diskmd(mds->mds_dt_exp, &lmm);
491 oti_free_cookies(&oti);
495 OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
499 obd_free_memmd(mds->mds_dt_exp, &lsm);
503 static void reconstruct_open(struct mds_update_record *rec, int offset,
504 struct ptlrpc_request *req,
505 struct lustre_handle *child_lockh)
507 struct mds_export_data *med = &req->rq_export->exp_mds_data;
508 struct mds_client_data *mcd = med->med_mcd;
509 struct mds_obd *mds = mds_req2mds(req);
510 struct mds_file_data *mfd;
511 struct obd_device *obd = req->rq_export->exp_obd;
512 struct dentry *parent, *dchild;
513 struct ldlm_reply *rep;
514 struct mds_body *body;
520 LASSERT(offset == 3); /* only called via intent */
521 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
522 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
524 /* copy rc, transno and disp; steal locks */
525 mds_req_from_mcd(req, mcd);
526 intent_set_disposition(rep, mcd->mcd_last_data);
528 /* Only replay if create or open actually happened. */
529 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
531 return; /* error looking up parent or child */
534 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
535 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
536 OLID4(rec->ur_id1), PTR_ERR(parent));
538 dchild = ll_lookup_one_len(rec->ur_name, parent,
539 rec->ur_namelen - 1);
540 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
541 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
543 if (!dchild->d_inode)
544 GOTO(out_dput, 0); /* child not present to open */
546 /* At this point, we know we have a child. We'll send
547 * it back _unless_ it not created and open failed. */
548 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
549 !intent_disposition(rep, DISP_OPEN_CREATE) &&
553 /* get lock (write for O_CREAT, read otherwise) */
554 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
555 if (S_ISREG(dchild->d_inode->i_mode)) {
556 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
557 dchild->d_inode, 1, 0);
560 LASSERT(rc == req->rq_status);
562 /* If we have LOV EA data, the OST holds size, mtime */
563 if (!(body->valid & OBD_MD_FLEASIZE))
564 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
565 OBD_MD_FLATIME | OBD_MD_FLMTIME);
568 /* If we have -EEXIST as the status, and we were asked to create
569 * exclusively, we can tell we failed because the file already existed.
571 if (req->rq_status == -EEXIST &&
572 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
573 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
577 /* If we didn't get as far as trying to open, then some locking thing
578 * probably went wrong, and we'll just bail here.
580 if (!intent_disposition(rep, DISP_OPEN_OPEN))
583 /* If we failed, then we must have failed opening, so don't look for
584 * file descriptor or anything, just give the client the bad news.
590 list_for_each(t, &med->med_open_head) {
591 mfd = list_entry(t, struct mds_file_data, mfd_list);
592 if (mfd->mfd_xid == req->rq_xid)
597 /* #warning "XXX fixme" bug 2991 */
598 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
599 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
600 * to detect a re-open */
602 mntget(mds->mds_vfsmnt);
603 CERROR("Re-opened file \n");
604 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
605 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
607 CERROR("mds: out of memory\n");
608 GOTO(out_dput, req->rq_status = -ENOMEM);
612 body->handle.cookie = mfd->mfd_handle.h_cookie;
613 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
614 mfd->mfd_handle.h_cookie);
624 /* do NOT or the MAY_*'s, you'll get the weakest */
625 static int accmode(int flags)
629 if (flags & FMODE_READ)
631 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
633 if (flags & FMODE_EXEC)
638 /* Handles object creation, actual opening, and I/O epoch */
639 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
640 struct mds_body *body, int flags, void **handle,
641 struct mds_update_record *rec, struct ldlm_reply *rep)
643 struct mds_obd *mds = mds_req2mds(req);
644 struct obd_device *obd = req->rq_export->exp_obd;
645 struct mds_file_data *mfd = NULL;
646 obd_id *ids = NULL; /* object IDs created */
651 /* atomically create objects if necessary */
652 down(&dchild->d_inode->i_sem);
653 mode = dchild->d_inode->i_mode;
654 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
655 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
656 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
657 dchild->d_inode, 0, 0);
659 up(&dchild->d_inode->i_sem);
664 /* no EA: create objects */
665 if ((body->valid & OBD_MD_FLEASIZE) &&
666 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
667 up(&dchild->d_inode->i_sem);
670 if (!(body->valid & OBD_MD_FLEASIZE)) {
671 /* no EA: create objects */
672 rc = mds_create_objects(req, 2, rec, mds, obd,
673 dchild, handle, &ids);
675 CERROR("mds_create_objects: rc = %d\n", rc);
676 up(&dchild->d_inode->i_sem);
680 if (S_ISREG(dchild->d_inode->i_mode) &&
681 (body->valid & OBD_MD_FLEASIZE)) {
682 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
685 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
686 dchild->d_inode, 0, 0);
688 up(&dchild->d_inode->i_sem);
693 rc = mds_pack_acl(obd, req->rq_repmsg, 3, body, dchild->d_inode);
695 CERROR("mds_pack_acl: rc = %d\n", rc);
696 up(&dchild->d_inode->i_sem);
700 /* If the inode has no EA data, then MDSs hold size, mtime */
701 if (S_ISREG(dchild->d_inode->i_mode) &&
702 !(body->valid & OBD_MD_FLEASIZE)) {
703 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
704 OBD_MD_FLATIME | OBD_MD_FLMTIME);
706 up(&dchild->d_inode->i_sem);
708 intent_set_disposition(rep, DISP_OPEN_OPEN);
709 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
711 RETURN(PTR_ERR(mfd));
713 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
714 mfd->mfd_handle.h_cookie);
716 mds_dt_update_objids(obd, ids);
717 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
720 // mds_mfd_destroy(mfd);
724 static int mds_open_by_id(struct ptlrpc_request *req,
725 struct lustre_id *id,
726 struct mds_body *body, int flags,
727 struct mds_update_record *rec,
728 struct ldlm_reply *rep)
730 struct mds_obd *mds = mds_req2mds(req);
731 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
732 struct dentry *dchild;
733 char idname[LL_ID_NAMELEN];
738 down(&pending_dir->i_sem);
739 idlen = ll_id2str(idname, id_ino(id), id_gen(id));
740 dchild = lookup_one_len(idname, mds->mds_pending_dir,
742 if (IS_ERR(dchild)) {
743 up(&pending_dir->i_sem);
744 rc = PTR_ERR(dchild);
745 CERROR("error looking up %s in PENDING: rc = %d\n",
750 up(&pending_dir->i_sem);
751 if (dchild->d_inode != NULL) {
752 mds_inode_set_orphan(dchild->d_inode);
753 mds_pack_inode2body(req2obd(req), body,
756 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
757 intent_set_disposition(rep, DISP_LOOKUP_POS);
758 CWARN("Orphan %s found and opened in PENDING directory\n",
765 * we didn't find it in PENDING so it isn't an orphan. See if it was a
766 * regular inode that was previously created.
768 dchild = mds_id2dentry(req2obd(req), id, NULL);
770 RETURN(PTR_ERR(dchild));
772 mds_pack_inode2body(req2obd(req), body,
775 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
776 intent_set_disposition(rep, DISP_LOOKUP_POS);
779 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
780 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
781 req, rc, rep ? rep->lock_policy_res1 : 0);
782 /* XXX what do we do here if mds_finish_transno itself failed? */
787 int mds_pin(struct ptlrpc_request *req, int offset)
789 struct obd_device *obd = req->rq_export->exp_obd;
790 struct mds_body *request_body, *reply_body;
791 struct lvfs_run_ctxt saved;
792 int rc, size = sizeof(*reply_body);
795 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
796 sizeof(*request_body));
798 rc = lustre_pack_reply(req, 1, &size, NULL);
801 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
803 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
804 rc = mds_open_by_id(req, &request_body->id1, reply_body,
805 request_body->flags, NULL, NULL);
806 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
812 * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
813 * child_lockh is NULL we just get the lock as a barrier to wait for other
814 * holders of this lock, and drop it right away again.
816 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
817 struct lustre_handle *child_lockh)
819 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
820 struct lustre_handle lockh;
825 if (child_lockh == NULL)
826 child_lockh = &lockh;
828 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
829 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
830 mds_blocking_ast, ldlm_completion_ast, NULL,
831 NULL, NULL, 0, NULL, child_lockh);
833 CERROR("ldlm_cli_enqueue: %d\n", rc);
834 else if (child_lockh == &lockh)
835 ldlm_lock_decref(child_lockh, LCK_EX);
840 int mds_open(struct mds_update_record *rec, int offset,
841 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
843 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
844 struct obd_device *obd = req->rq_export->exp_obd;
845 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
846 struct mds_obd *mds = mds_req2mds(req);
847 struct ldlm_reply *rep = NULL;
848 struct mds_body *body = NULL;
849 struct dentry *dchild = NULL, *dparent = NULL;
850 struct lustre_handle parent_lockh[2] = {{0}, {0}};
851 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
852 int parent_mode = LCK_PR;
854 struct dentry_params dp;
855 struct mea *mea = NULL;
856 int mea_size, update_mode;
857 int child_mode = LCK_PR;
858 /* Always returning LOOKUP lock if open succesful to guard
860 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
861 struct ldlm_res_id child_res_id = { .name = {0}};
865 DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
866 OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
869 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
870 (obd_timeout + 1) / 4);
872 if (offset == 3) { /* intent */
873 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
874 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
875 } else if (offset == 1) { /* non-intent reint */
876 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
882 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
885 * Step 0: If we are passed a id, then we assume the client already
886 * opened this file and is only replaying the RPC, so we open the inode
887 * by fid (at some large expense in security).
889 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
890 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
891 OLID4(rec->ur_id2), rec->ur_namelen - 1,
892 rec->ur_name, rec->ur_mode);
894 LASSERT(id_ino(rec->ur_id2));
896 rc = mds_open_by_id(req, rec->ur_id2, body, rec->ur_flags,
899 mds_body_do_reverse_map(med, body);
903 /* We didn't find the correct inode on disk either, so we
904 * need to re-create it via a regular replay. */
905 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
907 LASSERT(!id_ino(rec->ur_id2));
910 LASSERT(offset == 3); /* If we got here, we must be called via intent */
912 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
913 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
917 acc_mode = accmode(rec->ur_flags);
919 /* Step 1: Find and lock the parent */
920 if (rec->ur_flags & MDS_OPEN_CREAT) {
921 /* XXX Well, in fact we only need this lock mode change if
922 in addition to O_CREAT, the file does not exist.
923 But we do not know if it exists or not yet */
924 parent_mode = LCK_PW;
927 if (rec->ur_namelen == 1) {
928 /* client (LMV) wants to open the file by id */
929 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
930 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
931 if (IS_ERR(dchild)) {
932 rc = PTR_ERR(dparent);
933 CERROR("child lookup by an id error %d\n", rc);
940 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
941 parent_lockh, &update_mode, rec->ur_name,
942 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
943 if (IS_ERR(dparent)) {
944 rc = PTR_ERR(dparent);
946 CERROR("parent lookup for "DLID4" failed, error %d\n",
947 OLID4(rec->ur_id1), rc);
949 /* Just cannot find parent - make it look like
950 usual negative lookup to avoid extra MDS RPC */
951 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
952 intent_set_disposition(rep, DISP_LOOKUP_NEG);
956 LASSERT(dparent->d_inode != NULL);
958 cleanup_phase = 1; /* parent dentry and lock */
960 /* try to retrieve MEA data for this dir */
961 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
967 * dir is already splitted, check is requested filename should *
968 * live at this MDS or at another one.
971 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
972 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
974 "%s: inapropriate MDS(%d) for %lu/%u:%s."
975 " should be %lu(%d)\n", obd->obd_name,
976 mea->mea_master, dparent->d_inode->i_ino,
977 dparent->d_inode->i_generation, rec->ur_name,
978 (unsigned long)id_group(&mea->mea_ids[i]), i);
979 GOTO(cleanup, rc = -ERESTART);
983 /* Step 2: Lookup the child */
984 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
985 if (IS_ERR(dchild)) {
986 rc = PTR_ERR(dchild);
987 dchild = NULL; /* don't confuse mds_finish_transno() below */
992 cleanup_phase = 2; /* child dentry */
994 if (dchild->d_flags & DCACHE_CROSS_REF) {
995 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
997 /* we're gonna acquire LOOKUP lock on the child,
998 * but we have already locked parent and our order
999 * may conflict with enqueue_order_locks(). so,
1000 * drop parent lock and acquire both the locks in
1001 * common order. bug 6190 */
1002 ldlm_lock_decref(parent_lockh, parent_mode);
1007 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1008 parent_lockh, &dparent,
1009 LCK_PR, MDS_INODELOCK_UPDATE,
1010 &update_mode, rec->ur_name,
1011 rec->ur_namelen, child_lockh,
1013 MDS_INODELOCK_LOOKUP);
1018 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1019 /* wow! someone unlink and create new one yet */
1020 CDEBUG(D_OTHER, "nice race, repeat lookup\n");
1021 ldlm_lock_decref(parent_lockh, parent_mode);
1022 ldlm_lock_decref(child_lockh, LCK_PR);
1025 LASSERT(rec->ur_namelen > 1);
1029 mds_pack_dentry2body(obd, body, dchild, 1);
1030 intent_set_disposition(rep, DISP_LOOKUP_POS);
1031 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1034 if (parent_lockh[1].cookie != 0)
1035 ldlm_lock_decref(parent_lockh + 1, update_mode);
1037 ldlm_lock_decref(parent_lockh, parent_mode);
1039 OBD_FREE(mea, mea_size);
1045 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1047 if (dchild->d_inode)
1048 intent_set_disposition(rep, DISP_LOOKUP_POS);
1050 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1052 /* Step 3: If the child was negative, and we're supposed to, create it.*/
1053 if (dchild->d_inode == NULL) {
1056 struct inode *inode;
1058 ino = (unsigned long)id_ino(rec->ur_id2);
1060 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1063 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1064 obd->obd_name, dparent->d_inode->i_ino,
1065 dparent->d_inode->i_generation, rc);
1068 /* dir got splitted */
1069 GOTO(cleanup, rc = -ERESTART);
1070 } else if (rc < 0) {
1071 /* error happened during spitting */
1075 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1077 * dentry is negative and we weren't supposed to create
1078 * object, get out of here.
1080 GOTO(cleanup, rc = -ENOENT);
1083 intent_set_disposition(rep, DISP_OPEN_CREATE);
1084 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1086 if (IS_ERR(handle)) {
1087 rc = PTR_ERR(handle);
1091 dchild->d_fsdata = (void *) &dp;
1095 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1096 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1097 dchild->d_fsdata = NULL;
1100 CDEBUG(D_INODE, "error during create: %d\n", rc);
1103 inode = dchild->d_inode;
1105 LASSERT(ino == inode->i_ino);
1107 /* written as part of setattr */
1108 inode->i_generation = id_gen(rec->ur_id2);
1109 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1110 inode->i_ino, inode->i_generation);
1114 LTIME_S(iattr.ia_atime) = rec->ur_time;
1115 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1116 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1118 iattr.ia_uid = rec->ur_fsuid;
1119 if (dparent->d_inode->i_mode & S_ISGID)
1120 iattr.ia_gid = dparent->d_inode->i_gid;
1122 iattr.ia_gid = rec->ur_fsgid;
1124 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1125 ATTR_MTIME | ATTR_CTIME;
1127 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1129 CERROR("error on child setattr: rc = %d\n", rc);
1131 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1133 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1135 CERROR("error on parent setattr: rc = %d\n", rc);
1137 MD_COUNTER_INCREMENT(obd, create);
1140 down(&dchild->d_inode->i_sem);
1142 rc = mds_update_inode_sid(obd, dchild->d_inode,
1143 handle, rec->ur_id2);
1145 CERROR("mds_update_inode_sid() failed, "
1148 id_assign_fid(&body->id1, rec->ur_id2);
1151 * make sure, that fid is up-to-date.
1153 mds_set_last_fid(obd, id_fid(rec->ur_id2));
1155 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1156 handle, &body->id1);
1158 CERROR("mds_alloc_inode_sid() failed, "
1162 up(&dchild->d_inode->i_sem);
1164 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1165 rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1166 dchild->d_inode, handle,
1167 req->rq_export->exp_sync);
1171 acc_mode = 0; /* Don't check for permissions */
1173 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1175 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1176 "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1177 dchild->d_name.name, dchild, dchild->d_inode);
1179 if (S_ISREG(dchild->d_inode->i_mode)) {
1180 /* Check permissions etc */
1181 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1185 /* Can't write to a read-only file */
1186 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1187 GOTO(cleanup, rc = -EPERM);
1189 /* An append-only file must be opened in append mode for
1191 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1192 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1193 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1194 GOTO(cleanup, rc = -EPERM);
1197 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1198 (rec->ur_flags & MDS_OPEN_EXCL)) {
1199 /* File already exists, we didn't just create it, and we
1200 * were passed O_EXCL; err-or. */
1201 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1204 if (S_ISDIR(dchild->d_inode->i_mode)) {
1205 if (rec->ur_flags & MDS_OPEN_CREAT ||
1206 rec->ur_flags & FMODE_WRITE) {
1207 /* we are trying to create or write a exist dir */
1208 GOTO(cleanup, rc = -EISDIR);
1210 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1211 GOTO(cleanup, rc = -EACCES);
1214 * here was checking for possible GNS mount points to skip their
1215 * open. I removed it as its detection based only on SUID bit is
1216 * not reliable. Opening such a dirs should not cause any
1217 * problems, at least tests show that. --umka
1221 /* if we are following a symlink, don't open */
1222 if (S_ISLNK(dchild->d_inode->i_mode))
1223 GOTO(cleanup_no_trans, rc = 0);
1225 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1226 !S_ISDIR(dchild->d_inode->i_mode))
1227 GOTO(cleanup, rc = -ENOTDIR);
1229 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1230 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1231 GOTO(cleanup, rc = -EAGAIN);
1234 /* Obtain OPEN lock as well */
1235 policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1237 /* We cannot use acc_mode here, because it is zeroed in case of
1238 creating a file, so we get wrong lockmode */
1239 if (accmode(rec->ur_flags) & MAY_WRITE)
1240 child_mode = LCK_CW;
1241 else if (accmode(rec->ur_flags) & MAY_EXEC)
1242 child_mode = LCK_PR;
1244 child_mode = LCK_CR;
1246 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1247 struct lustre_id sid;
1249 down(&dchild->d_inode->i_sem);
1250 rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
1251 up(&dchild->d_inode->i_sem);
1253 CERROR("Can't read inode self id, "
1254 "inode %lu, rc %d\n",
1255 dchild->d_inode->i_ino, rc);
1259 /* In case of replay we do not get a lock assuming that the
1260 caller has it already */
1261 child_res_id.name[0] = id_fid(&sid);
1262 child_res_id.name[1] = id_group(&sid);
1264 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1265 child_res_id, LDLM_IBITS, &policy,
1266 child_mode, &lock_flags,
1267 mds_blocking_ast, ldlm_completion_ast,
1268 NULL, NULL, NULL, 0, NULL, child_lockh);
1275 /* Step 5: mds_open it */
1276 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1281 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1282 req, rc, rep ? rep->lock_policy_res1 : 0);
1285 switch (cleanup_phase) {
1288 ldlm_lock_decref(child_lockh, child_mode);
1289 child_lockh->cookie = 0;
1292 if (rc && created) {
1293 int err = vfs_unlink(dparent->d_inode, dchild);
1295 CERROR("unlink(%.*s) in error path: %d\n",
1296 dchild->d_name.len, dchild->d_name.name,
1299 } else if (created) {
1300 mds_lock_new_child(obd, dchild->d_inode, NULL);
1304 if (dparent == NULL)
1309 if (parent_lockh[1].cookie != 0)
1310 ldlm_lock_decref(parent_lockh + 1, update_mode);
1313 ldlm_lock_decref(parent_lockh, parent_mode);
1315 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1318 OBD_FREE(mea, mea_size);
1320 atomic_inc(&mds->mds_open_count);
1322 mds_body_do_reverse_map(med, body);
1325 * If we have not taken the "open" lock, we may not return 0 here,
1326 * because caller expects 0 to mean "lock is taken", and it needs
1327 * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1328 * client. Later caller should rewrite the return value back to zero
1329 * if it to be used any further.
1331 if ((cleanup_phase != 3) && !rc)
1337 /* Close a "file descriptor" and possibly unlink an orphan from the
1338 * PENDING directory. Caller must hold child->i_sem, this drops it.
1340 * If we are being called from mds_disconnect() because the client has
1341 * disappeared, then req == NULL and we do not update last_rcvd because
1342 * there is nothing that could be recovered by the client at this stage
1343 * (it will not even _have_ an entry in last_rcvd anymore).
1345 * Returns EAGAIN if the client needs to get more data and re-close. */
1346 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1347 struct obd_device *obd, struct mds_file_data *mfd,
1350 struct inode *inode = mfd->mfd_dentry->d_inode;
1351 char idname[LL_ID_NAMELEN];
1352 int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1353 struct dentry *pending_child = NULL;
1354 struct mds_obd *mds = &obd->u.mds;
1355 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1356 void *handle = NULL;
1357 struct mds_body *request_body = NULL, *reply_body = NULL;
1358 struct dentry_params dp;
1359 struct iattr iattr = { 0 };
1360 struct llog_create_locks *lcl = NULL;
1363 if (req && req->rq_reqmsg != NULL)
1364 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1365 sizeof(*request_body));
1366 if (req && req->rq_repmsg != NULL)
1367 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1368 sizeof(*reply_body));
1370 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1371 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1372 idname, inode->i_nlink, mds_orphan_open_count(inode));
1374 last_orphan = mds_orphan_open_dec_test(inode) &&
1375 mds_inode_is_orphan(inode);
1376 UP_WRITE_I_ALLOC_SEM(inode);
1378 /* this is half of the actual "close" */
1379 if (mfd->mfd_mode & FMODE_WRITE) {
1380 rc = mds_put_write_access(mds, inode, request_body,
1381 last_orphan && unlink_orphan);
1382 } else if (mfd->mfd_mode & FMODE_EXEC) {
1383 mds_allow_write_access(inode);
1386 if (last_orphan && unlink_orphan) {
1387 struct lov_mds_md *lmm = NULL;
1388 int stripe_count = 0;
1389 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1391 if (obd->obd_recovering) {
1392 CDEBUG(D_HA, "not remove orphan %s until recovery"
1393 " is over\n", idname);
1397 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1399 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1400 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1401 CERROR("found \"orphan\" %s %s with link count %d\n",
1402 S_ISREG(inode->i_mode) ? "file" : "dir",
1403 idname, inode->i_nlink);
1405 * sadly, there is no easy way to save pending_child from
1406 * mds_reint_unlink() into mfd, so we need to re-lookup, but
1407 * normally it will still be in the dcache.
1409 down(&pending_dir->i_sem);
1410 cleanup_phase = 1; /* up(i_sem) when finished */
1411 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1413 if (IS_ERR(pending_child))
1414 GOTO(cleanup, rc = PTR_ERR(pending_child));
1415 LASSERT(pending_child->d_inode != NULL);
1417 cleanup_phase = 2; /* dput(pending_child) when finished */
1418 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1419 rc = vfs_rmdir(pending_dir, pending_child);
1421 CERROR("error unlinking orphan dir %s: rc %d\n",
1426 if (req != NULL && req->rq_repmsg != NULL) {
1427 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1428 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1431 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1432 NULL, stripe_count);
1433 if (IS_ERR(handle)) {
1434 rc = PTR_ERR(handle);
1439 pending_child->d_fsdata = (void *) &dp;
1442 rc = vfs_unlink(pending_dir, pending_child);
1444 CERROR("error unlinking orphan %s: rc %d\n",
1447 if (req != NULL && req->rq_repmsg != NULL &&
1448 (reply_body->valid & OBD_MD_FLEASIZE) &&
1449 mds_log_op_unlink(obd, pending_child->d_inode,
1450 lmm, req->rq_repmsg->buflens[1],
1451 lustre_msg_buf(req->rq_repmsg, 2, 0),
1452 req->rq_repmsg->buflens[2], &lcl) > 0) {
1453 reply_body->valid |= OBD_MD_FLCOOKIE;
1456 goto out; /* Don't bother updating attrs on unlinked inode */
1460 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1461 /* Update the on-disk attributes if this was the last write
1462 * close, and all information was provided (i.e., rc == 0)
1464 * XXX this should probably be abstracted with mds_reint_setattr
1467 if (request_body->valid & OBD_MD_FLMTIME &&
1468 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1469 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1470 iattr.ia_valid |= ATTR_MTIME;
1472 if (request_body->valid & OBD_MD_FLCTIME &&
1473 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1474 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1475 iattr.ia_valid |= ATTR_CTIME;
1478 /* XXX can't set block count with fsfilt_setattr (!) */
1479 if (request_body->valid & OBD_MD_FLSIZE) {
1480 iattr.ia_valid |= ATTR_SIZE;
1481 iattr.ia_size = request_body->size;
1483 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1484 iattr.ia_valid |= ATTR_BLOCKS;
1485 iattr.ia_blocks = request_body->blocks
1490 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1491 /* Only start a transaction to write out only the atime if
1492 * it is more out-of-date than the specified limit. If we
1493 * are already going to write out the atime then do it anyway.
1495 if ((request_body->atime >
1496 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1497 (iattr.ia_valid != 0 &&
1498 request_body->atime > LTIME_S(inode->i_atime))) {
1499 LTIME_S(iattr.ia_atime) = request_body->atime;
1500 iattr.ia_valid |= ATTR_ATIME;
1504 if (iattr.ia_valid != 0) {
1505 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1507 GOTO(cleanup, rc = PTR_ERR(handle));
1508 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1510 CERROR("error in setattr(%s): rc %d\n", idname, rc);
1513 /* If other clients have this file open for write, rc will be > 0 */
1516 l_dput(mfd->mfd_dentry);
1517 mds_mfd_destroy(mfd);
1520 atomic_dec(&mds->mds_open_count);
1521 if (req != NULL && reply_body != NULL) {
1522 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1523 } else if (handle) {
1524 int err, force_sync = 0;
1526 if (req && req->rq_export)
1527 force_sync = req->rq_export->exp_sync;
1529 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1532 CERROR("error committing close: %d\n", err);
1538 switch (cleanup_phase) {
1541 ptlrpc_save_llog_lock(req, lcl);
1542 dput(pending_child);
1544 up(&pending_dir->i_sem);
1549 int mds_close(struct ptlrpc_request *req, int offset)
1551 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1552 struct obd_device *obd = req->rq_export->exp_obd;
1553 struct mds_body *body;
1554 struct mds_file_data *mfd;
1555 struct lvfs_run_ctxt saved;
1556 struct inode *inode;
1557 int rc, repsize[3] = {sizeof(struct mds_body),
1558 obd->u.mds.mds_max_mdsize,
1559 obd->u.mds.mds_max_cookiesize};
1561 MD_COUNTER_INCREMENT(obd, close);
1563 rc = lustre_pack_reply(req, 3, repsize, NULL);
1565 CERROR("lustre_pack_reply: rc = %d\n", rc);
1566 req->rq_status = rc;
1568 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1571 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1572 DEBUG_REQ(D_HA, req, "close replay\n");
1573 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1574 lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1575 req->rq_repmsg->buflens[2]);
1578 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1579 lustre_swab_mds_body);
1581 CERROR("Can't unpack body\n");
1582 req->rq_status = -EFAULT;
1586 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1587 /* do some stuff */ ;
1589 mfd = mds_handle2mfd(&body->handle);
1591 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1592 ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1593 req->rq_status = -ESTALE;
1597 inode = mfd->mfd_dentry->d_inode;
1598 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1599 DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1600 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1601 struct mds_body *rep_body;
1603 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1604 sizeof (*rep_body));
1605 LASSERT(rep_body != NULL);
1607 mds_pack_inode2body(obd, rep_body, inode,
1608 (body->valid & OBD_MD_FID) ? 1 : 0);
1610 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1611 inode, MDS_PACK_MD_LOCK, 0);
1613 spin_lock(&med->med_open_lock);
1614 list_del(&mfd->mfd_list);
1615 spin_unlock(&med->med_open_lock);
1617 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1618 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1619 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1621 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1622 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1623 req->rq_status = -ENOMEM;
1632 int mds_done_writing(struct ptlrpc_request *req, int offset)
1634 struct mds_body *body;
1635 int rc, size = sizeof(struct mds_body);
1638 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1640 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1641 lustre_swab_mds_body);
1643 CERROR("Can't unpack body\n");
1644 req->rq_status = -EFAULT;
1648 rc = lustre_pack_reply(req, 1, &size, NULL);
1650 CERROR("lustre_pack_reply: rc = %d\n", rc);
1651 req->rq_status = rc;