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, rc = -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 rc = mds_get_write_access(mds, dentry->d_inode, 0);
265 GOTO(cleanup_mfd, rc);
266 #ifdef IFILTERDATA_ACTUALLY_USED
267 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
269 } else if (flags & FMODE_EXEC) {
270 rc = mds_deny_write_access(mds, dentry->d_inode);
272 GOTO(cleanup_mfd, rc);
277 /* mark the file as open to handle open-unlink. */
278 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
279 mds_orphan_open_inc(dentry->d_inode);
280 UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
282 mfd->mfd_mode = flags;
283 mfd->mfd_dentry = dentry;
284 mfd->mfd_xid = req->rq_xid;
286 spin_lock(&med->med_open_lock);
287 list_add(&mfd->mfd_list, &med->med_open_head);
288 spin_unlock(&med->med_open_lock);
291 body->handle.cookie = mfd->mfd_handle.h_cookie;
295 mds_mfd_destroy(mfd);
301 mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
302 struct lov_desc *desc)
306 for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
307 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
308 le64_to_cpu(lmm->lmm_objects[i].l_object_id);
312 /* must be called with i_sem held */
314 mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req,
315 int offset, struct mds_update_record *rec,
316 struct dentry *dchild, void **handle,
319 struct inode *inode = dchild->d_inode;
320 struct mds_obd *mds = &obd->u.mds;
321 struct obd_trans_info oti = { 0 };
322 struct lov_stripe_md *lsm = NULL;
323 struct lov_mds_md *lmm = NULL;
324 int rc, lmm_bufsize, lmm_size;
325 struct obdo *oa = NULL;
326 struct mds_body *body;
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];
366 LASSERT(lmm_buf != NULL);
367 LASSERT(lmm_bufsize >= lmm_size);
369 memcpy(lmm_buf, lmm, lmm_size);
370 rc = fsfilt_set_md(obd, inode, *handle, lmm,
373 CERROR("open replay failed to set md:%d\n", rc);
377 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
383 oa->o_mode = S_IFREG | 0600;
384 oa->o_id = inode->i_ino;
385 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
386 oa->o_generation = inode->i_generation;
387 oa->o_uid = 0; /* must have 0 uid / gid on OST */
389 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE |
390 OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
393 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
394 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
396 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
397 /* check if things like lfs setstripe are sending us the ea */
398 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
399 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
401 0, &lsm, rec->ur_eadata);
405 OBD_ALLOC(lmm, mds->mds_max_mdsize);
407 GOTO(out_oa, rc = -ENOMEM);
409 lmm_size = mds->mds_max_mdsize;
410 rc = mds_get_md(obd, dchild->d_parent->d_inode,
411 lmm, &lmm_size, 1, 0);
413 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
416 OBD_FREE(lmm, mds->mds_max_mdsize);
421 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
422 oti.oti_flags |= OBD_MODE_CROW;
423 rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
429 CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR),
430 "error creating objects for "
431 "inode %lu: rc = %d\n",
434 CERROR("obd_create returned invalid "
441 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
442 0, &lsm, rec->ur_eadata);
446 lsm->lsm_object_id = oa->o_id;
447 lsm->lsm_object_gr = oa->o_gr;
450 oa->o_size = inode->i_size;
451 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
452 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE);
454 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
456 CERROR("error setting attrs for inode %lu: rc %d\n",
459 CERROR("obd_setattr returned bad rc %d\n", rc);
466 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
467 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
469 LASSERT(lsm && lsm->lsm_object_id);
471 rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
472 if (!id_ino(rec->ur_id2))
473 obd_free_memmd(mds->mds_dt_exp, &lsm);
475 CERROR("cannot pack lsm, err = %d\n", rc);
480 body->eadatasize = rc;
483 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
484 if (IS_ERR(*handle)) {
485 rc = PTR_ERR(*handle);
490 rc = fsfilt_set_md(obd, inode, *handle, lmm,
493 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
494 lmm_bufsize = req->rq_repmsg->buflens[offset];
496 LASSERT(lmm_bufsize >= lmm_size);
498 memcpy(lmm_buf, lmm, lmm_size);
499 obd_free_diskmd(mds->mds_dt_exp, &lmm);
501 oti_free_cookies(&oti);
505 obd_free_memmd(mds->mds_dt_exp, &lsm);
510 mds_destroy_object(struct obd_device *obd,
511 struct inode *inode, int async)
513 struct mds_obd *mds = &obd->u.mds;
514 struct lov_mds_md *lmm = NULL;
518 LASSERT(inode != NULL);
520 if (inode->i_nlink != 0) {
521 CDEBUG(D_INODE, "attempt to destroy OSS object when "
522 "i_nlink == %d\n", (int)inode->i_nlink);
526 OBD_ALLOC(lmm, mds->mds_max_mdsize);
530 lmm_size = mds->mds_max_mdsize;
531 rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
533 CERROR("no stripe info for %lu/%lu inode\n",
534 (unsigned long)inode->i_ino,
535 (unsigned long)inode->i_generation);
536 GOTO(out_free_lmm, rc);
540 /* asynchronously unlink objecect on OSS */
541 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
544 CERROR("error unlinking object on OSS, "
546 GOTO(out_free_lmm, rc);
549 CDEBUG(D_INODE, "no stripping info found for inode "
550 "%lu/%lu\n", (unsigned long)inode->i_ino,
551 (unsigned long)inode->i_generation);
555 OBD_FREE(lmm, mds->mds_max_mdsize);
559 static void reconstruct_open(struct mds_update_record *rec, int offset,
560 struct ptlrpc_request *req,
561 struct lustre_handle *child_lockh)
563 struct mds_export_data *med = &req->rq_export->exp_mds_data;
564 struct mds_client_data *mcd = med->med_mcd;
565 struct mds_obd *mds = mds_req2mds(req);
566 struct mds_file_data *mfd;
567 struct obd_device *obd = req->rq_export->exp_obd;
568 struct dentry *parent = NULL, *dchild;
569 struct ldlm_reply *rep;
570 struct mds_body *body;
576 LASSERT(offset == 3); /* only called via intent */
577 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
578 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
580 /* copy rc, transno and disp; steal locks */
581 mds_req_from_mcd(req, mcd);
582 intent_set_disposition(rep, mcd->mcd_last_data);
584 /* Only replay if create or open actually happened. */
585 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
587 return; /* error looking up parent or child */
590 if (rec->ur_namelen == 1) {
591 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
593 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
595 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
596 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
597 OLID4(rec->ur_id1), PTR_ERR(parent));
598 dchild = ll_lookup_one_len(rec->ur_name, parent,
599 rec->ur_namelen - 1);
600 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
601 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
604 if (!dchild->d_inode)
605 GOTO(out_dput, 0); /* child not present to open */
607 /* At this point, we know we have a child. We'll send
608 * it back _unless_ it not created and open failed. */
609 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
610 !intent_disposition(rep, DISP_OPEN_CREATE) &&
614 /* get lock (write for O_CREAT, read otherwise) */
615 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
616 if (S_ISREG(dchild->d_inode->i_mode)) {
617 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
618 dchild->d_inode, 1, 0);
621 LASSERT(rc == req->rq_status);
623 /* If we have LOV EA data, the OST holds size, mtime */
624 if (!(body->valid & OBD_MD_FLEASIZE))
625 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
626 OBD_MD_FLATIME | OBD_MD_FLMTIME);
629 /* If we have -EEXIST as the status, and we were asked to create
630 * exclusively, we can tell we failed because the file already existed.
632 if (req->rq_status == -EEXIST &&
633 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
634 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
638 /* If we didn't get as far as trying to open, then some locking thing
639 * probably went wrong, and we'll just bail here.
641 if (!intent_disposition(rep, DISP_OPEN_OPEN))
644 /* If we failed, then we must have failed opening, so don't look for
645 * file descriptor or anything, just give the client the bad news.
651 list_for_each(t, &med->med_open_head) {
652 mfd = list_entry(t, struct mds_file_data, mfd_list);
653 if (mfd->mfd_xid == req->rq_xid)
658 /* #warning "XXX fixme" bug 2991 */
659 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
660 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
661 * to detect a re-open */
663 mntget(mds->mds_vfsmnt);
664 CERROR("Re-opened file \n");
665 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
666 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
668 CERROR("mds: out of memory\n");
669 GOTO(out_dput, req->rq_status = -ENOMEM);
673 body->handle.cookie = mfd->mfd_handle.h_cookie;
674 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
675 mfd->mfd_handle.h_cookie);
686 /* do NOT or the MAY_*'s, you'll get the weakest */
687 static int accmode(int flags)
691 if (flags & FMODE_READ)
693 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
695 if (flags & FMODE_EXEC)
700 /* Handles object creation, actual opening, and I/O epoch */
701 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
702 struct mds_body *body, int flags, void **handle,
703 struct mds_update_record *rec, struct ldlm_reply *rep)
705 struct obd_device *obd = req->rq_export->exp_obd;
706 struct mds_obd *mds = mds_req2mds(req);
707 struct mds_file_data *mfd = NULL;
713 /* atomically create objects if necessary */
714 down(&dchild->d_inode->i_sem);
715 mode = dchild->d_inode->i_mode;
717 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
718 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
719 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
720 dchild->d_inode, 0, 0);
722 up(&dchild->d_inode->i_sem);
728 if ((body->valid & OBD_MD_FLEASIZE) &&
729 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
730 up(&dchild->d_inode->i_sem);
734 if (!(body->valid & OBD_MD_FLEASIZE)) {
735 /* no EA: create objects */
736 rc = mds_create_objects(obd, req, 2, rec,
737 dchild, handle, &ids);
739 CERROR("mds_create_object: rc = %d\n", rc);
740 up(&dchild->d_inode->i_sem);
745 if (S_ISREG(dchild->d_inode->i_mode) &&
746 (body->valid & OBD_MD_FLEASIZE)) {
747 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
750 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
751 dchild->d_inode, 0, 0);
753 up(&dchild->d_inode->i_sem);
759 rc = mds_pack_acl(obd, req->rq_repmsg, 3, body, dchild->d_inode);
761 CERROR("mds_pack_acl: rc = %d\n", rc);
762 up(&dchild->d_inode->i_sem);
766 /* If the inode has no EA data, then MDSs hold size, mtime */
767 if (S_ISREG(dchild->d_inode->i_mode) &&
768 !(body->valid & OBD_MD_FLEASIZE)) {
769 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
770 OBD_MD_FLATIME | OBD_MD_FLMTIME);
773 up(&dchild->d_inode->i_sem);
775 intent_set_disposition(rep, DISP_OPEN_OPEN);
776 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
778 RETURN(PTR_ERR(mfd));
780 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
781 mfd->mfd_handle.h_cookie);
784 mds_dt_update_objids(obd, ids);
785 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
791 static int mds_open_by_id(struct ptlrpc_request *req,
792 struct lustre_id *id,
793 struct mds_body *body, int flags,
794 struct mds_update_record *rec,
795 struct ldlm_reply *rep)
797 struct mds_obd *mds = mds_req2mds(req);
798 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
799 struct dentry *dchild;
800 char idname[LL_ID_NAMELEN];
805 down(&pending_dir->i_sem);
806 idlen = ll_id2str(idname, id_ino(id), id_gen(id));
807 dchild = lookup_one_len(idname, mds->mds_pending_dir,
809 if (IS_ERR(dchild)) {
810 up(&pending_dir->i_sem);
811 rc = PTR_ERR(dchild);
812 CERROR("error looking up %s in PENDING: rc = %d\n",
817 up(&pending_dir->i_sem);
818 if (dchild->d_inode != NULL) {
819 mds_inode_set_orphan(dchild->d_inode);
820 mds_pack_inode2body(req2obd(req), body,
823 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
824 intent_set_disposition(rep, DISP_LOOKUP_POS);
825 CWARN("Orphan %s found and opened in PENDING directory\n",
832 * we didn't find it in PENDING so it isn't an orphan. See if it was a
833 * regular inode that was previously created.
835 dchild = mds_id2dentry(req2obd(req), id, NULL);
837 RETURN(PTR_ERR(dchild));
839 mds_pack_inode2body(req2obd(req), body,
842 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
843 intent_set_disposition(rep, DISP_LOOKUP_POS);
846 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
847 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
848 req, rc, rep ? rep->lock_policy_res1 : 0);
849 /* XXX what do we do here if mds_finish_transno itself failed? */
854 int mds_pin(struct ptlrpc_request *req, int offset)
856 struct obd_device *obd = req->rq_export->exp_obd;
857 struct mds_body *request_body, *reply_body;
858 struct lvfs_run_ctxt saved;
859 int rc, size = sizeof(*reply_body);
862 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
863 sizeof(*request_body));
865 rc = lustre_pack_reply(req, 1, &size, NULL);
868 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
870 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
871 rc = mds_open_by_id(req, &request_body->id1, reply_body,
872 request_body->flags, NULL, NULL);
873 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
879 * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
880 * child_lockh is NULL we just get the lock as a barrier to wait for other
881 * holders of this lock, and drop it right away again.
883 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
884 struct lustre_handle *child_lockh)
886 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
887 struct lustre_handle lockh;
892 if (child_lockh == NULL)
893 child_lockh = &lockh;
895 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
896 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
897 mds_blocking_ast, ldlm_completion_ast, NULL,
898 NULL, NULL, 0, NULL, child_lockh);
900 CERROR("ldlm_cli_enqueue: %d\n", rc);
901 else if (child_lockh == &lockh)
902 ldlm_lock_decref(child_lockh, LCK_EX);
907 int mds_open(struct mds_update_record *rec, int offset,
908 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
910 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
911 struct obd_device *obd = req->rq_export->exp_obd;
912 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
913 struct mds_obd *mds = mds_req2mds(req);
914 struct ldlm_reply *rep = NULL;
915 struct mds_body *body = NULL;
916 struct dentry *dchild = NULL, *dparent = NULL;
917 struct lustre_handle parent_lockh[2] = {{0}, {0}};
918 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
919 int parent_mode = LCK_PR;
921 struct dentry_params dp;
922 struct mea *mea = NULL;
923 int mea_size, update_mode;
924 int child_mode = LCK_PR;
927 DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
928 OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
931 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
932 (obd_timeout + 1) / 4);
934 if (offset == 3) { /* intent */
935 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
936 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
937 } else if (offset == 1) { /* non-intent reint */
938 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
944 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
947 * Step 0: If we are passed a id, then we assume the client already
948 * opened this file and is only replaying the RPC, so we open the inode
949 * by fid (at some large expense in security).
951 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
952 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
953 OLID4(rec->ur_id2), rec->ur_namelen - 1,
954 rec->ur_name, rec->ur_mode);
956 LASSERT(id_ino(rec->ur_id2));
958 rc = mds_open_by_id(req, rec->ur_id2, body,
959 rec->ur_flags, rec, rep);
961 mds_body_do_reverse_map(med, body);
965 /* We didn't find the correct inode on disk either, so we
966 * need to re-create it via a regular replay. */
967 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
969 LASSERT(!id_ino(rec->ur_id2));
972 LASSERT(offset == 3); /* If we got here, we must be called via intent */
974 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
975 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
979 acc_mode = accmode(rec->ur_flags);
981 /* Step 1: Find and lock the parent */
982 if (rec->ur_flags & MDS_OPEN_CREAT) {
983 /* XXX Well, in fact we only need this lock mode change if
984 in addition to O_CREAT, the file does not exist.
985 But we do not know if it exists or not yet */
986 parent_mode = LCK_PW;
989 if (rec->ur_namelen == 1) {
990 /* client (LMV) wants to open the file by id */
991 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
992 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
993 if (IS_ERR(dchild)) {
994 rc = PTR_ERR(dparent);
995 CERROR("child lookup by an id error %d\n", rc);
1002 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
1003 parent_lockh, &update_mode, rec->ur_name,
1004 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
1005 if (IS_ERR(dparent)) {
1006 rc = PTR_ERR(dparent);
1007 if (rc != -ENOENT) {
1008 CERROR("parent lookup for "DLID4" failed, error %d\n",
1009 OLID4(rec->ur_id1), rc);
1011 /* Just cannot find parent - make it look like
1012 usual negative lookup to avoid extra MDS RPC */
1013 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1014 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1018 LASSERT(dparent->d_inode != NULL);
1020 cleanup_phase = 1; /* parent dentry and lock */
1022 /* try to retrieve MEA data for this dir */
1023 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1029 * dir is already splitted, check is requested filename should *
1030 * live at this MDS or at another one.
1033 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1034 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1036 "%s: inapropriate MDS(%d) for %lu/%u:%s."
1037 " should be %lu(%d)\n", obd->obd_name,
1038 mea->mea_master, dparent->d_inode->i_ino,
1039 dparent->d_inode->i_generation, rec->ur_name,
1040 (unsigned long)id_group(&mea->mea_ids[i]), i);
1041 GOTO(cleanup, rc = -ERESTART);
1045 /* Step 2: Lookup the child */
1046 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1047 if (IS_ERR(dchild)) {
1048 rc = PTR_ERR(dchild);
1049 dchild = NULL; /* don't confuse mds_finish_transno() below */
1054 cleanup_phase = 2; /* child dentry */
1056 if (dchild->d_flags & DCACHE_CROSS_REF) {
1057 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1058 LASSERT(rec->ur_namelen > 1);
1060 /* we're gonna acquire LOOKUP lock on the child,
1061 * but we have already locked parent and our order
1062 * may conflict with enqueue_order_locks(). so,
1063 * drop parent lock and acquire both the locks in
1064 * common order. bug 6190 */
1066 if (parent_lockh[1].cookie != 0)
1067 ldlm_lock_decref(parent_lockh + 1, update_mode);
1069 ldlm_lock_decref(parent_lockh, parent_mode);
1074 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1075 parent_lockh, &dparent,
1076 LCK_PR, MDS_INODELOCK_UPDATE,
1077 &update_mode, rec->ur_name,
1078 rec->ur_namelen, child_lockh,
1080 MDS_INODELOCK_LOOKUP);
1086 if (parent_lockh[1].cookie != 0)
1087 ldlm_lock_decref(parent_lockh + 1, update_mode);
1089 ldlm_lock_decref(parent_lockh, LCK_PR);
1091 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1092 CDEBUG(D_OTHER, "race: name changed (%p)\n",
1094 if (dchild->d_inode)
1095 ldlm_lock_decref(child_lockh, LCK_PR);
1101 mds_pack_dentry2body(obd, body, dchild, 1);
1102 intent_set_disposition(rep, DISP_LOOKUP_POS);
1103 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1106 OBD_FREE(mea, mea_size);
1112 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1114 if (dchild->d_inode)
1115 intent_set_disposition(rep, DISP_LOOKUP_POS);
1117 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1119 /* Step 3: If the child was negative, and we're supposed to, create it.*/
1120 if (dchild->d_inode == NULL) {
1123 struct inode *inode;
1125 ino = (unsigned long)id_ino(rec->ur_id2);
1127 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1130 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1131 obd->obd_name, dparent->d_inode->i_ino,
1132 dparent->d_inode->i_generation, rc);
1135 /* dir got splitted */
1136 GOTO(cleanup, rc = -ERESTART);
1137 } else if (rc < 0) {
1138 /* error happened during spitting */
1142 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1144 * dentry is negative and we weren't supposed to create
1145 * object, get out of here.
1147 GOTO(cleanup, rc = -ENOENT);
1150 intent_set_disposition(rep, DISP_OPEN_CREATE);
1151 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1153 if (IS_ERR(handle)) {
1154 rc = PTR_ERR(handle);
1158 dchild->d_fsdata = (void *) &dp;
1162 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1163 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1164 dchild->d_fsdata = NULL;
1167 CDEBUG(D_INODE, "error during create: %d\n", rc);
1170 inode = dchild->d_inode;
1172 LASSERT(ino == inode->i_ino);
1174 /* written as part of setattr */
1175 inode->i_generation = id_gen(rec->ur_id2);
1176 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1177 inode->i_ino, inode->i_generation);
1181 LTIME_S(iattr.ia_atime) = rec->ur_time;
1182 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1183 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1185 iattr.ia_uid = rec->ur_fsuid;
1186 if (dparent->d_inode->i_mode & S_ISGID)
1187 iattr.ia_gid = dparent->d_inode->i_gid;
1189 iattr.ia_gid = rec->ur_fsgid;
1191 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1192 ATTR_MTIME | ATTR_CTIME;
1194 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1196 CERROR("error on child setattr: rc = %d\n", rc);
1198 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1200 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1202 CERROR("error on parent setattr: rc = %d\n", rc);
1204 MD_COUNTER_INCREMENT(obd, create);
1208 rc = mds_update_inode_sid(obd, dchild->d_inode,
1209 handle, rec->ur_id2);
1211 CERROR("mds_update_inode_sid() failed, "
1214 id_assign_fid(&body->id1, rec->ur_id2);
1217 * make sure, that fid is up-to-date.
1219 mds_set_last_fid(obd, id_fid(rec->ur_id2));
1221 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1222 handle, &body->id1);
1224 CERROR("mds_alloc_inode_sid() failed, "
1229 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1230 rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1231 dchild->d_inode, handle,
1232 req->rq_export->exp_sync);
1236 acc_mode = 0; /* Don't check for permissions */
1238 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1240 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1241 "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1242 dchild->d_name.name, dchild, dchild->d_inode);
1244 if (S_ISREG(dchild->d_inode->i_mode)) {
1245 /* Check permissions etc */
1246 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1250 /* Can't write to a read-only file */
1251 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1252 GOTO(cleanup, rc = -EPERM);
1254 /* An append-only file must be opened in append mode for
1256 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1257 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1258 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1259 GOTO(cleanup, rc = -EPERM);
1262 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1263 (rec->ur_flags & MDS_OPEN_EXCL)) {
1264 /* File already exists, we didn't just create it, and we
1265 * were passed O_EXCL; err-or. */
1266 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1269 if (S_ISDIR(dchild->d_inode->i_mode)) {
1270 if (rec->ur_flags & MDS_OPEN_CREAT ||
1271 rec->ur_flags & FMODE_WRITE) {
1272 /* we are trying to create or write a exist dir */
1273 GOTO(cleanup, rc = -EISDIR);
1275 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1276 GOTO(cleanup, rc = -EACCES);
1279 * here was checking for possible GNS mount points to skip their
1280 * open. I removed it as its detection based only on SUID bit is
1281 * not reliable. Opening such a dirs should not cause any
1282 * problems, at least tests show that. --umka
1286 /* if we are following a symlink, don't open */
1287 if (S_ISLNK(dchild->d_inode->i_mode))
1288 GOTO(cleanup_no_trans, rc = 0);
1290 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1291 !S_ISDIR(dchild->d_inode->i_mode))
1292 GOTO(cleanup, rc = -ENOTDIR);
1294 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1295 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1296 GOTO(cleanup, rc = -EAGAIN);
1299 #warning "disable opencache lock for CMD2"
1301 /* We cannot use acc_mode here, because it is zeroed in case of
1302 creating a file, so we get wrong lockmode */
1303 if (accmode(rec->ur_flags) & MAY_WRITE)
1304 child_mode = LCK_CW;
1305 else if (accmode(rec->ur_flags) & MAY_EXEC)
1306 child_mode = LCK_PR;
1308 child_mode = LCK_CR;
1310 /* Always returning LOOKUP lock if open succesful to guard
1311 * dentry on client. In case of replay we do not get a lock
1312 * assuming that the caller has it already */
1313 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1314 struct ldlm_res_id child_res_id = { .name = {0}};
1315 ldlm_policy_data_t policy;
1318 /* LOOKUP lock will protect dentry on client -bzzz */
1319 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1322 child_res_id.name[0] = id_fid(&body->id1);
1323 child_res_id.name[1] = id_group(&body->id1);
1325 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1326 child_res_id, LDLM_IBITS, &policy,
1327 child_mode, &lock_flags,
1328 mds_blocking_ast, ldlm_completion_ast,
1329 NULL, NULL, NULL, 0, NULL, child_lockh);
1337 /* Step 5: mds_open it */
1338 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1343 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1344 req, rc, rep ? rep->lock_policy_res1 : 0);
1347 switch (cleanup_phase) {
1350 ldlm_lock_decref(child_lockh, child_mode);
1351 child_lockh->cookie = 0;
1354 if (rc && created) {
1355 int err = vfs_unlink(dparent->d_inode, dchild);
1357 CERROR("unlink(%.*s) in error path: %d\n",
1358 dchild->d_name.len, dchild->d_name.name,
1361 } else if (created) {
1362 mds_lock_new_child(obd, dchild->d_inode, NULL);
1366 if (dparent == NULL)
1371 if (parent_lockh[1].cookie != 0)
1372 ldlm_lock_decref(parent_lockh + 1, update_mode);
1375 ldlm_lock_decref(parent_lockh, parent_mode);
1377 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1380 OBD_FREE(mea, mea_size);
1382 atomic_inc(&mds->mds_open_count);
1384 mds_body_do_reverse_map(med, body);
1387 * If we have not taken the "open" lock, we may not return 0 here,
1388 * because caller expects 0 to mean "lock is taken", and it needs
1389 * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1390 * client. Later caller should rewrite the return value back to zero
1391 * if it to be used any further.
1393 if ((cleanup_phase != 3) && !rc)
1399 /* Close a "file descriptor" and possibly unlink an orphan from the
1400 * PENDING directory. Caller must hold child->i_sem, this drops it.
1402 * If we are being called from mds_disconnect() because the client has
1403 * disappeared, then req == NULL and we do not update last_rcvd because
1404 * there is nothing that could be recovered by the client at this stage
1405 * (it will not even _have_ an entry in last_rcvd anymore).
1407 * Returns EAGAIN if the client needs to get more data and re-close. */
1408 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1409 struct obd_device *obd, struct mds_file_data *mfd,
1412 struct inode *inode = mfd->mfd_dentry->d_inode;
1413 char idname[LL_ID_NAMELEN];
1414 int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1415 struct dentry *pending_child = NULL;
1416 struct mds_obd *mds = &obd->u.mds;
1417 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1418 void *handle = NULL;
1419 struct mds_body *request_body = NULL, *reply_body = NULL;
1420 struct dentry_params dp;
1421 struct iattr iattr = { 0 };
1422 struct llog_create_locks *lcl = NULL;
1425 if (req && req->rq_reqmsg != NULL)
1426 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1427 sizeof(*request_body));
1428 if (req && req->rq_repmsg != NULL)
1429 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1430 sizeof(*reply_body));
1432 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1433 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1434 idname, inode->i_nlink, mds_orphan_open_count(inode));
1436 last_orphan = mds_orphan_open_dec_test(inode) &&
1437 mds_inode_is_orphan(inode);
1438 UP_WRITE_I_ALLOC_SEM(inode);
1440 /* this is half of the actual "close" */
1441 if (mfd->mfd_mode & FMODE_WRITE) {
1442 rc = mds_put_write_access(mds, inode, request_body,
1443 last_orphan && unlink_orphan);
1444 } else if (mfd->mfd_mode & FMODE_EXEC) {
1445 mds_allow_write_access(inode);
1448 if (last_orphan && unlink_orphan) {
1449 struct lov_mds_md *lmm = NULL;
1450 int stripe_count = 0;
1451 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1453 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1455 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1456 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1457 CERROR("found \"orphan\" %s %s with link count %d\n",
1458 S_ISREG(inode->i_mode) ? "file" : "dir",
1459 idname, inode->i_nlink);
1461 * sadly, there is no easy way to save pending_child from
1462 * mds_reint_unlink() into mfd, so we need to re-lookup, but
1463 * normally it will still be in the dcache.
1465 down(&pending_dir->i_sem);
1466 cleanup_phase = 1; /* up(i_sem) when finished */
1467 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1469 if (IS_ERR(pending_child))
1470 GOTO(cleanup, rc = PTR_ERR(pending_child));
1471 if (pending_child->d_inode == NULL) {
1472 CERROR("orphan %s has been removed\n", idname);
1473 GOTO(cleanup, rc = 0);
1476 cleanup_phase = 2; /* dput(pending_child) when finished */
1477 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1478 rc = vfs_rmdir(pending_dir, pending_child);
1480 CERROR("error unlinking orphan dir %s: rc %d\n",
1485 if (req != NULL && req->rq_repmsg != NULL) {
1486 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1487 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1490 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1491 NULL, stripe_count);
1492 if (IS_ERR(handle)) {
1493 rc = PTR_ERR(handle);
1498 pending_child->d_fsdata = (void *) &dp;
1501 rc = vfs_unlink(pending_dir, pending_child);
1503 CERROR("error unlinking orphan %s: rc %d\n",
1506 if (req != NULL && req->rq_repmsg != NULL &&
1507 (reply_body->valid & OBD_MD_FLEASIZE) &&
1508 mds_log_op_unlink(obd, pending_child->d_inode,
1509 lmm, req->rq_repmsg->buflens[1],
1510 lustre_msg_buf(req->rq_repmsg, 2, 0),
1511 req->rq_repmsg->buflens[2], &lcl) > 0) {
1512 reply_body->valid |= OBD_MD_FLCOOKIE;
1515 rc = mds_destroy_object(obd, inode, 1);
1517 CERROR("cannot destroy OSS object on close, err %d\n",
1522 goto out; /* Don't bother updating attrs on unlinked inode */
1526 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1527 /* Update the on-disk attributes if this was the last write
1528 * close, and all information was provided (i.e., rc == 0)
1530 * XXX this should probably be abstracted with mds_reint_setattr
1533 if (request_body->valid & OBD_MD_FLMTIME &&
1534 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1535 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1536 iattr.ia_valid |= ATTR_MTIME;
1538 if (request_body->valid & OBD_MD_FLCTIME &&
1539 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1540 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1541 iattr.ia_valid |= ATTR_CTIME;
1544 /* XXX can't set block count with fsfilt_setattr (!) */
1545 if (request_body->valid & OBD_MD_FLSIZE) {
1546 iattr.ia_valid |= ATTR_SIZE;
1547 iattr.ia_size = request_body->size;
1549 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1550 iattr.ia_valid |= ATTR_BLOCKS;
1551 iattr.ia_blocks = request_body->blocks
1556 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1557 /* Only start a transaction to write out only the atime if
1558 * it is more out-of-date than the specified limit. If we
1559 * are already going to write out the atime then do it anyway.
1561 if ((request_body->atime >
1562 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1563 (iattr.ia_valid != 0 &&
1564 request_body->atime > LTIME_S(inode->i_atime))) {
1565 LTIME_S(iattr.ia_atime) = request_body->atime;
1566 iattr.ia_valid |= ATTR_ATIME;
1570 if (iattr.ia_valid != 0) {
1571 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1573 GOTO(cleanup, rc = PTR_ERR(handle));
1574 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1576 CERROR("error in setattr(%s): rc %d\n", idname, rc);
1579 /* If other clients have this file open for write, rc will be > 0 */
1582 l_dput(mfd->mfd_dentry);
1583 mds_mfd_destroy(mfd);
1586 atomic_dec(&mds->mds_open_count);
1587 if (req != NULL && reply_body != NULL) {
1588 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1589 } else if (handle) {
1590 int err, force_sync = 0;
1592 if (req && req->rq_export)
1593 force_sync = req->rq_export->exp_sync;
1595 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1598 CERROR("error committing close: %d\n", err);
1604 switch (cleanup_phase) {
1607 ptlrpc_save_llog_lock(req, lcl);
1608 dput(pending_child);
1610 up(&pending_dir->i_sem);
1615 int mds_close(struct ptlrpc_request *req, int offset)
1617 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1618 struct obd_device *obd = req->rq_export->exp_obd;
1619 struct mds_body *body;
1620 struct mds_file_data *mfd;
1621 struct lvfs_run_ctxt saved;
1622 struct inode *inode;
1623 int rc, repsize[3] = {sizeof(struct mds_body),
1624 obd->u.mds.mds_max_mdsize,
1625 obd->u.mds.mds_max_cookiesize};
1627 MD_COUNTER_INCREMENT(obd, close);
1629 rc = lustre_pack_reply(req, 3, repsize, NULL);
1631 CERROR("lustre_pack_reply: rc = %d\n", rc);
1632 req->rq_status = rc;
1634 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1637 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1638 DEBUG_REQ(D_HA, req, "close replay");
1639 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1640 lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1641 req->rq_repmsg->buflens[2]);
1644 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1645 lustre_swab_mds_body);
1647 CERROR("Can't unpack body\n");
1648 req->rq_status = -EFAULT;
1652 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1653 /* do some stuff */ ;
1655 mfd = mds_handle2mfd(&body->handle);
1657 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1658 ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1659 req->rq_status = -ESTALE;
1663 inode = mfd->mfd_dentry->d_inode;
1664 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1665 DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1666 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1667 struct mds_body *rep_body;
1669 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1670 sizeof (*rep_body));
1671 LASSERT(rep_body != NULL);
1673 mds_pack_inode2body(obd, rep_body, inode,
1674 (body->valid & OBD_MD_FID) ? 1 : 0);
1676 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1677 inode, MDS_PACK_MD_LOCK, 0);
1679 spin_lock(&med->med_open_lock);
1680 list_del(&mfd->mfd_list);
1681 spin_unlock(&med->med_open_lock);
1683 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1684 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1685 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1687 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1688 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1689 req->rq_status = -ENOMEM;
1698 int mds_done_writing(struct ptlrpc_request *req, int offset)
1700 struct mds_body *body;
1701 int rc, size = sizeof(struct mds_body);
1704 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1706 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1707 lustre_swab_mds_body);
1709 CERROR("Can't unpack body\n");
1710 req->rq_status = -EFAULT;
1714 rc = lustre_pack_reply(req, 1, &size, NULL);
1716 CERROR("lustre_pack_reply: rc = %d\n", rc);
1717 req->rq_status = rc;