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);
300 /* this is object id allocation callback */
301 static int mds_obj_alloc(obd_id *objid)
304 LASSERT(objid != NULL);
309 mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
310 struct lov_desc *desc)
314 for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
315 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
316 le64_to_cpu(lmm->lmm_objects[i].l_object_id);
320 /* must be called with i_sem held */
322 mds_create_object(struct obd_device *obd, struct ptlrpc_request *req,
323 int offset, struct mds_update_record *rec,
324 struct dentry *dchild, void **handle,
327 struct inode *inode = dchild->d_inode;
328 struct mds_obd *mds = &obd->u.mds;
329 struct obd_trans_info oti = { 0 };
330 struct lov_stripe_md *lsm = NULL;
331 struct lov_mds_md *lmm = NULL;
332 int rc, lmm_bufsize, lmm_size;
333 struct obdo *oa = NULL;
334 struct mds_body *body;
338 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
339 !(rec->ur_flags & FMODE_WRITE))
342 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
344 if (!S_ISREG(inode->i_mode))
346 if (body->valid & OBD_MD_FLEASIZE)
352 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
353 LASSERT(id_ino(rec->ur_id2));
354 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
355 lmm_size = rec->ur_eadatalen;
356 lmm = rec->ur_eadata;
360 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
361 if (IS_ERR(*handle)) {
362 rc = PTR_ERR(*handle);
368 * FIXME: this is evil layering violation, all things related to
369 * stripping should be done by LOV. --umka.
371 mds_objids_from_lmm(ids, lmm, &mds->mds_dt_desc);
373 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
374 lmm_bufsize = req->rq_repmsg->buflens[offset];
376 LASSERT(lmm_buf != NULL);
377 LASSERT(lmm_bufsize >= lmm_size);
379 memcpy(lmm_buf, lmm, lmm_size);
380 rc = fsfilt_set_md(obd, inode, *handle, lmm,
383 CERROR("open replay failed to set md:%d\n", rc);
387 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
393 oa->o_mode = S_IFREG | 0600;
394 oa->o_id = inode->i_ino;
395 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
396 oa->o_generation = inode->i_generation;
397 oa->o_uid = 0; /* must have 0 uid / gid on OST */
399 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE |
400 OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
403 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
404 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
406 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
407 /* check if things like lfs setstripe are sending us the ea */
408 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
409 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
411 0, &lsm, rec->ur_eadata);
415 OBD_ALLOC(lmm, mds->mds_max_mdsize);
417 GOTO(out_oa, rc = -ENOMEM);
419 lmm_size = mds->mds_max_mdsize;
420 rc = mds_get_md(obd, dchild->d_parent->d_inode,
421 lmm, &lmm_size, 1, 0);
423 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
426 OBD_FREE(lmm, mds->mds_max_mdsize);
432 * create with CROW flag and base ids for allocating new ids on
435 oti.oti_flags |= OBD_MODE_CROW;
436 oti.oti_obj_alloc = mds_obj_alloc;
438 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
439 rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
445 CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR),
446 "error creating objects for "
447 "inode %lu: rc = %d\n",
450 CERROR("obd_create returned invalid "
457 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
458 0, &lsm, rec->ur_eadata);
462 lsm->lsm_object_id = oa->o_id;
463 lsm->lsm_object_gr = oa->o_gr;
466 oa->o_size = inode->i_size;
467 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
468 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE);
470 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
472 CERROR("error setting attrs for inode %lu: rc %d\n",
475 CERROR("obd_setattr returned bad rc %d\n", rc);
482 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
483 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
485 LASSERT(lsm && lsm->lsm_object_id);
487 rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
488 if (!id_ino(rec->ur_id2))
489 obd_free_memmd(mds->mds_dt_exp, &lsm);
491 CERROR("cannot pack lsm, err = %d\n", rc);
496 body->eadatasize = rc;
499 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
500 if (IS_ERR(*handle)) {
501 rc = PTR_ERR(*handle);
506 rc = fsfilt_set_md(obd, inode, *handle, lmm,
509 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
510 lmm_bufsize = req->rq_repmsg->buflens[offset];
512 LASSERT(lmm_bufsize >= lmm_size);
514 memcpy(lmm_buf, lmm, lmm_size);
515 obd_free_diskmd(mds->mds_dt_exp, &lmm);
517 oti_free_cookies(&oti);
521 obd_free_memmd(mds->mds_dt_exp, &lsm);
526 mds_destroy_object(struct obd_device *obd,
527 struct inode *inode, int async)
529 struct mds_obd *mds = &obd->u.mds;
530 struct lov_mds_md *lmm = NULL;
534 LASSERT(inode != NULL);
536 if (inode->i_nlink != 0) {
537 CDEBUG(D_INODE, "attempt to destroy OSS object when "
538 "i_nlink == %d\n", (int)inode->i_nlink);
542 OBD_ALLOC(lmm, mds->mds_max_mdsize);
546 lmm_size = mds->mds_max_mdsize;
547 rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
549 CERROR("no stripe info for %lu/%lu inode\n",
550 (unsigned long)inode->i_ino,
551 (unsigned long)inode->i_generation);
552 GOTO(out_free_lmm, rc);
556 /* asynchronously unlink objecect on OSS */
557 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
560 CERROR("error unlinking object on OSS, "
562 GOTO(out_free_lmm, rc);
565 CDEBUG(D_INODE, "no stripping info found for inode "
566 "%lu/%lu\n", (unsigned long)inode->i_ino,
567 (unsigned long)inode->i_generation);
571 OBD_FREE(lmm, mds->mds_max_mdsize);
575 static void reconstruct_open(struct mds_update_record *rec, int offset,
576 struct ptlrpc_request *req,
577 struct lustre_handle *child_lockh)
579 struct mds_export_data *med = &req->rq_export->exp_mds_data;
580 struct mds_client_data *mcd = med->med_mcd;
581 struct mds_obd *mds = mds_req2mds(req);
582 struct mds_file_data *mfd;
583 struct obd_device *obd = req->rq_export->exp_obd;
584 struct dentry *parent = NULL, *dchild;
585 struct ldlm_reply *rep;
586 struct mds_body *body;
592 LASSERT(offset == 3); /* only called via intent */
593 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
594 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
596 /* copy rc, transno and disp; steal locks */
597 mds_req_from_mcd(req, mcd);
598 intent_set_disposition(rep, mcd->mcd_last_data);
600 /* Only replay if create or open actually happened. */
601 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
603 return; /* error looking up parent or child */
606 if (rec->ur_namelen == 1) {
607 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
609 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
611 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
612 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
613 OLID4(rec->ur_id1), PTR_ERR(parent));
614 dchild = ll_lookup_one_len(rec->ur_name, parent,
615 rec->ur_namelen - 1);
616 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
617 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
620 if (!dchild->d_inode)
621 GOTO(out_dput, 0); /* child not present to open */
623 /* At this point, we know we have a child. We'll send
624 * it back _unless_ it not created and open failed. */
625 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
626 !intent_disposition(rep, DISP_OPEN_CREATE) &&
630 /* get lock (write for O_CREAT, read otherwise) */
631 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
632 if (S_ISREG(dchild->d_inode->i_mode)) {
633 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
634 dchild->d_inode, 1, 0);
637 LASSERT(rc == req->rq_status);
639 /* If we have LOV EA data, the OST holds size, mtime */
640 if (!(body->valid & OBD_MD_FLEASIZE))
641 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
642 OBD_MD_FLATIME | OBD_MD_FLMTIME);
645 /* If we have -EEXIST as the status, and we were asked to create
646 * exclusively, we can tell we failed because the file already existed.
648 if (req->rq_status == -EEXIST &&
649 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
650 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
654 /* If we didn't get as far as trying to open, then some locking thing
655 * probably went wrong, and we'll just bail here.
657 if (!intent_disposition(rep, DISP_OPEN_OPEN))
660 /* If we failed, then we must have failed opening, so don't look for
661 * file descriptor or anything, just give the client the bad news.
667 list_for_each(t, &med->med_open_head) {
668 mfd = list_entry(t, struct mds_file_data, mfd_list);
669 if (mfd->mfd_xid == req->rq_xid)
674 /* #warning "XXX fixme" bug 2991 */
675 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
676 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
677 * to detect a re-open */
679 mntget(mds->mds_vfsmnt);
680 CERROR("Re-opened file \n");
681 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
682 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
684 CERROR("mds: out of memory\n");
685 GOTO(out_dput, req->rq_status = -ENOMEM);
689 body->handle.cookie = mfd->mfd_handle.h_cookie;
690 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
691 mfd->mfd_handle.h_cookie);
702 /* do NOT or the MAY_*'s, you'll get the weakest */
703 static int accmode(int flags)
707 if (flags & FMODE_READ)
709 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
711 if (flags & FMODE_EXEC)
716 /* Handles object creation, actual opening, and I/O epoch */
717 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
718 struct mds_body *body, int flags, void **handle,
719 struct mds_update_record *rec, struct ldlm_reply *rep)
721 struct obd_device *obd = req->rq_export->exp_obd;
722 struct mds_obd *mds = mds_req2mds(req);
723 struct mds_file_data *mfd = NULL;
729 /* atomically create objects if necessary */
730 down(&dchild->d_inode->i_sem);
731 mode = dchild->d_inode->i_mode;
733 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
734 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
735 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
736 dchild->d_inode, 0, 0);
738 up(&dchild->d_inode->i_sem);
744 /* no EA: create objects */
745 if ((body->valid & OBD_MD_FLEASIZE) &&
746 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
747 up(&dchild->d_inode->i_sem);
751 if (!(body->valid & OBD_MD_FLEASIZE)) {
752 int ids_size = mds->mds_dt_desc.ld_tgt_count * sizeof(*ids);
754 OBD_ALLOC(ids, ids_size);
756 up(&dchild->d_inode->i_sem);
761 * synchronizing object creating to prevent another
762 * threads take the same base objid values.
764 down(&mds->mds_create_sem);
766 /* preparing base ids */
767 mds_dt_save_objids(obd, ids);
770 * create objects, @ids will contain new allocated obj
773 rc = mds_create_object(obd, req, 2, rec,
774 dchild, handle, ids);
776 CERROR("mds_create_object: rc = %d\n", rc);
777 up(&mds->mds_create_sem);
778 up(&dchild->d_inode->i_sem);
779 OBD_FREE(ids, ids_size);
784 * update MDS objids by new ones allocated in
785 * mds_create_object().
787 mds_dt_update_objids(obd, ids);
788 OBD_FREE(ids, ids_size);
790 up(&mds->mds_create_sem);
793 if (S_ISREG(dchild->d_inode->i_mode) &&
794 (body->valid & OBD_MD_FLEASIZE)) {
795 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
798 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
799 dchild->d_inode, 0, 0);
801 up(&dchild->d_inode->i_sem);
807 rc = mds_pack_acl(obd, req->rq_repmsg, 3, body, dchild->d_inode);
809 CERROR("mds_pack_acl: rc = %d\n", rc);
810 up(&dchild->d_inode->i_sem);
814 /* If the inode has no EA data, then MDSs hold size, mtime */
815 if (S_ISREG(dchild->d_inode->i_mode) &&
816 !(body->valid & OBD_MD_FLEASIZE)) {
817 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
818 OBD_MD_FLATIME | OBD_MD_FLMTIME);
821 up(&dchild->d_inode->i_sem);
823 intent_set_disposition(rep, DISP_OPEN_OPEN);
824 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
826 RETURN(PTR_ERR(mfd));
828 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
829 mfd->mfd_handle.h_cookie);
833 static int mds_open_by_id(struct ptlrpc_request *req,
834 struct lustre_id *id,
835 struct mds_body *body, int flags,
836 struct mds_update_record *rec,
837 struct ldlm_reply *rep)
839 struct mds_obd *mds = mds_req2mds(req);
840 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
841 struct dentry *dchild;
842 char idname[LL_ID_NAMELEN];
847 down(&pending_dir->i_sem);
848 idlen = ll_id2str(idname, id_ino(id), id_gen(id));
849 dchild = lookup_one_len(idname, mds->mds_pending_dir,
851 if (IS_ERR(dchild)) {
852 up(&pending_dir->i_sem);
853 rc = PTR_ERR(dchild);
854 CERROR("error looking up %s in PENDING: rc = %d\n",
859 up(&pending_dir->i_sem);
860 if (dchild->d_inode != NULL) {
861 mds_inode_set_orphan(dchild->d_inode);
862 mds_pack_inode2body(req2obd(req), body,
865 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
866 intent_set_disposition(rep, DISP_LOOKUP_POS);
867 CWARN("Orphan %s found and opened in PENDING directory\n",
874 * we didn't find it in PENDING so it isn't an orphan. See if it was a
875 * regular inode that was previously created.
877 dchild = mds_id2dentry(req2obd(req), id, NULL);
879 RETURN(PTR_ERR(dchild));
881 mds_pack_inode2body(req2obd(req), body,
884 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
885 intent_set_disposition(rep, DISP_LOOKUP_POS);
888 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
889 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
890 req, rc, rep ? rep->lock_policy_res1 : 0);
891 /* XXX what do we do here if mds_finish_transno itself failed? */
896 int mds_pin(struct ptlrpc_request *req, int offset)
898 struct obd_device *obd = req->rq_export->exp_obd;
899 struct mds_body *request_body, *reply_body;
900 struct lvfs_run_ctxt saved;
901 int rc, size = sizeof(*reply_body);
904 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
905 sizeof(*request_body));
907 rc = lustre_pack_reply(req, 1, &size, NULL);
910 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
912 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
913 rc = mds_open_by_id(req, &request_body->id1, reply_body,
914 request_body->flags, NULL, NULL);
915 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
921 * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
922 * child_lockh is NULL we just get the lock as a barrier to wait for other
923 * holders of this lock, and drop it right away again.
925 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
926 struct lustre_handle *child_lockh)
928 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
929 struct lustre_handle lockh;
934 if (child_lockh == NULL)
935 child_lockh = &lockh;
937 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
938 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
939 mds_blocking_ast, ldlm_completion_ast, NULL,
940 NULL, NULL, 0, NULL, child_lockh);
942 CERROR("ldlm_cli_enqueue: %d\n", rc);
943 else if (child_lockh == &lockh)
944 ldlm_lock_decref(child_lockh, LCK_EX);
949 int mds_open(struct mds_update_record *rec, int offset,
950 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
952 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
953 struct obd_device *obd = req->rq_export->exp_obd;
954 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
955 struct mds_obd *mds = mds_req2mds(req);
956 struct ldlm_reply *rep = NULL;
957 struct mds_body *body = NULL;
958 struct dentry *dchild = NULL, *dparent = NULL;
959 struct lustre_handle parent_lockh[2] = {{0}, {0}};
960 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
961 int parent_mode = LCK_PR;
963 struct dentry_params dp;
964 struct mea *mea = NULL;
965 int mea_size, update_mode;
966 int child_mode = LCK_PR;
967 /* Always returning LOOKUP lock if open succesful to guard
969 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
970 struct ldlm_res_id child_res_id = { .name = {0}};
974 DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
975 OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
978 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
979 (obd_timeout + 1) / 4);
981 if (offset == 3) { /* intent */
982 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
983 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
984 } else if (offset == 1) { /* non-intent reint */
985 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
991 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
994 * Step 0: If we are passed a id, then we assume the client already
995 * opened this file and is only replaying the RPC, so we open the inode
996 * by fid (at some large expense in security).
998 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
999 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
1000 OLID4(rec->ur_id2), rec->ur_namelen - 1,
1001 rec->ur_name, rec->ur_mode);
1003 LASSERT(id_ino(rec->ur_id2));
1005 rc = mds_open_by_id(req, rec->ur_id2, body,
1006 rec->ur_flags, rec, rep);
1007 if (rc != -ENOENT) {
1008 mds_body_do_reverse_map(med, body);
1012 /* We didn't find the correct inode on disk either, so we
1013 * need to re-create it via a regular replay. */
1014 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
1016 LASSERT(!id_ino(rec->ur_id2));
1019 LASSERT(offset == 3); /* If we got here, we must be called via intent */
1021 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
1022 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
1026 acc_mode = accmode(rec->ur_flags);
1028 /* Step 1: Find and lock the parent */
1029 if (rec->ur_flags & MDS_OPEN_CREAT) {
1030 /* XXX Well, in fact we only need this lock mode change if
1031 in addition to O_CREAT, the file does not exist.
1032 But we do not know if it exists or not yet */
1033 parent_mode = LCK_PW;
1036 if (rec->ur_namelen == 1) {
1037 /* client (LMV) wants to open the file by id */
1038 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
1039 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
1040 if (IS_ERR(dchild)) {
1041 rc = PTR_ERR(dparent);
1042 CERROR("child lookup by an id error %d\n", rc);
1049 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
1050 parent_lockh, &update_mode, rec->ur_name,
1051 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
1052 if (IS_ERR(dparent)) {
1053 rc = PTR_ERR(dparent);
1054 if (rc != -ENOENT) {
1055 CERROR("parent lookup for "DLID4" failed, error %d\n",
1056 OLID4(rec->ur_id1), rc);
1058 /* Just cannot find parent - make it look like
1059 usual negative lookup to avoid extra MDS RPC */
1060 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1061 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1065 LASSERT(dparent->d_inode != NULL);
1067 cleanup_phase = 1; /* parent dentry and lock */
1069 /* try to retrieve MEA data for this dir */
1070 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1076 * dir is already splitted, check is requested filename should *
1077 * live at this MDS or at another one.
1080 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1081 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1083 "%s: inapropriate MDS(%d) for %lu/%u:%s."
1084 " should be %lu(%d)\n", obd->obd_name,
1085 mea->mea_master, dparent->d_inode->i_ino,
1086 dparent->d_inode->i_generation, rec->ur_name,
1087 (unsigned long)id_group(&mea->mea_ids[i]), i);
1088 GOTO(cleanup, rc = -ERESTART);
1092 /* Step 2: Lookup the child */
1093 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1094 if (IS_ERR(dchild)) {
1095 rc = PTR_ERR(dchild);
1096 dchild = NULL; /* don't confuse mds_finish_transno() below */
1101 cleanup_phase = 2; /* child dentry */
1103 if (dchild->d_flags & DCACHE_CROSS_REF) {
1104 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1105 LASSERT(rec->ur_namelen > 1);
1107 /* we're gonna acquire LOOKUP lock on the child,
1108 * but we have already locked parent and our order
1109 * may conflict with enqueue_order_locks(). so,
1110 * drop parent lock and acquire both the locks in
1111 * common order. bug 6190 */
1113 if (parent_lockh[1].cookie != 0)
1114 ldlm_lock_decref(parent_lockh + 1, update_mode);
1116 ldlm_lock_decref(parent_lockh, parent_mode);
1121 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1122 parent_lockh, &dparent,
1123 LCK_PR, MDS_INODELOCK_UPDATE,
1124 &update_mode, rec->ur_name,
1125 rec->ur_namelen, child_lockh,
1127 MDS_INODELOCK_LOOKUP);
1133 if (parent_lockh[1].cookie != 0)
1134 ldlm_lock_decref(parent_lockh + 1, update_mode);
1136 ldlm_lock_decref(parent_lockh, LCK_PR);
1138 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1139 CDEBUG(D_OTHER, "race: name changed (%p)\n",
1141 if (dchild->d_inode)
1142 ldlm_lock_decref(child_lockh, LCK_PR);
1148 mds_pack_dentry2body(obd, body, dchild, 1);
1149 intent_set_disposition(rep, DISP_LOOKUP_POS);
1150 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1153 OBD_FREE(mea, mea_size);
1159 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1161 if (dchild->d_inode)
1162 intent_set_disposition(rep, DISP_LOOKUP_POS);
1164 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1166 /* Step 3: If the child was negative, and we're supposed to, create it.*/
1167 if (dchild->d_inode == NULL) {
1170 struct inode *inode;
1172 ino = (unsigned long)id_ino(rec->ur_id2);
1174 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1177 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1178 obd->obd_name, dparent->d_inode->i_ino,
1179 dparent->d_inode->i_generation, rc);
1182 /* dir got splitted */
1183 GOTO(cleanup, rc = -ERESTART);
1184 } else if (rc < 0) {
1185 /* error happened during spitting */
1189 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1191 * dentry is negative and we weren't supposed to create
1192 * object, get out of here.
1194 GOTO(cleanup, rc = -ENOENT);
1197 intent_set_disposition(rep, DISP_OPEN_CREATE);
1198 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1200 if (IS_ERR(handle)) {
1201 rc = PTR_ERR(handle);
1205 dchild->d_fsdata = (void *) &dp;
1209 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1210 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1211 dchild->d_fsdata = NULL;
1214 CDEBUG(D_INODE, "error during create: %d\n", rc);
1217 inode = dchild->d_inode;
1219 LASSERT(ino == inode->i_ino);
1221 /* written as part of setattr */
1222 inode->i_generation = id_gen(rec->ur_id2);
1223 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1224 inode->i_ino, inode->i_generation);
1228 LTIME_S(iattr.ia_atime) = rec->ur_time;
1229 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1230 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1232 iattr.ia_uid = rec->ur_fsuid;
1233 if (dparent->d_inode->i_mode & S_ISGID)
1234 iattr.ia_gid = dparent->d_inode->i_gid;
1236 iattr.ia_gid = rec->ur_fsgid;
1238 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1239 ATTR_MTIME | ATTR_CTIME;
1241 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1243 CERROR("error on child setattr: rc = %d\n", rc);
1245 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1247 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1249 CERROR("error on parent setattr: rc = %d\n", rc);
1251 MD_COUNTER_INCREMENT(obd, create);
1255 rc = mds_update_inode_sid(obd, dchild->d_inode,
1256 handle, rec->ur_id2);
1258 CERROR("mds_update_inode_sid() failed, "
1261 id_assign_fid(&body->id1, rec->ur_id2);
1264 * make sure, that fid is up-to-date.
1266 mds_set_last_fid(obd, id_fid(rec->ur_id2));
1268 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1269 handle, &body->id1);
1271 CERROR("mds_alloc_inode_sid() failed, "
1276 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1277 rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1278 dchild->d_inode, handle,
1279 req->rq_export->exp_sync);
1283 acc_mode = 0; /* Don't check for permissions */
1285 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1287 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1288 "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1289 dchild->d_name.name, dchild, dchild->d_inode);
1291 if (S_ISREG(dchild->d_inode->i_mode)) {
1292 /* Check permissions etc */
1293 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1297 /* Can't write to a read-only file */
1298 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1299 GOTO(cleanup, rc = -EPERM);
1301 /* An append-only file must be opened in append mode for
1303 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1304 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1305 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1306 GOTO(cleanup, rc = -EPERM);
1309 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1310 (rec->ur_flags & MDS_OPEN_EXCL)) {
1311 /* File already exists, we didn't just create it, and we
1312 * were passed O_EXCL; err-or. */
1313 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1316 if (S_ISDIR(dchild->d_inode->i_mode)) {
1317 if (rec->ur_flags & MDS_OPEN_CREAT ||
1318 rec->ur_flags & FMODE_WRITE) {
1319 /* we are trying to create or write a exist dir */
1320 GOTO(cleanup, rc = -EISDIR);
1322 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1323 GOTO(cleanup, rc = -EACCES);
1326 * here was checking for possible GNS mount points to skip their
1327 * open. I removed it as its detection based only on SUID bit is
1328 * not reliable. Opening such a dirs should not cause any
1329 * problems, at least tests show that. --umka
1333 /* if we are following a symlink, don't open */
1334 if (S_ISLNK(dchild->d_inode->i_mode))
1335 GOTO(cleanup_no_trans, rc = 0);
1337 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1338 !S_ISDIR(dchild->d_inode->i_mode))
1339 GOTO(cleanup, rc = -ENOTDIR);
1341 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1342 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1343 GOTO(cleanup, rc = -EAGAIN);
1346 #warning "disable opencache lock for CMD2"
1348 /* Obtain OPEN lock as well */
1349 policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1351 /* We cannot use acc_mode here, because it is zeroed in case of
1352 creating a file, so we get wrong lockmode */
1353 if (accmode(rec->ur_flags) & MAY_WRITE)
1354 child_mode = LCK_CW;
1355 else if (accmode(rec->ur_flags) & MAY_EXEC)
1356 child_mode = LCK_PR;
1358 child_mode = LCK_CR;
1360 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1361 /* In case of replay we do not get a lock assuming that the
1362 caller has it already */
1363 child_res_id.name[0] = id_fid(&body->id1);
1364 child_res_id.name[1] = id_group(&body->id1);
1366 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1367 child_res_id, LDLM_IBITS, &policy,
1368 child_mode, &lock_flags,
1369 mds_blocking_ast, ldlm_completion_ast,
1370 NULL, NULL, NULL, 0, NULL, child_lockh);
1378 /* Step 5: mds_open it */
1379 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1384 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1385 req, rc, rep ? rep->lock_policy_res1 : 0);
1388 switch (cleanup_phase) {
1391 ldlm_lock_decref(child_lockh, child_mode);
1392 child_lockh->cookie = 0;
1395 if (rc && created) {
1396 int err = vfs_unlink(dparent->d_inode, dchild);
1398 CERROR("unlink(%.*s) in error path: %d\n",
1399 dchild->d_name.len, dchild->d_name.name,
1402 } else if (created) {
1403 mds_lock_new_child(obd, dchild->d_inode, NULL);
1407 if (dparent == NULL)
1412 if (parent_lockh[1].cookie != 0)
1413 ldlm_lock_decref(parent_lockh + 1, update_mode);
1416 ldlm_lock_decref(parent_lockh, parent_mode);
1418 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1421 OBD_FREE(mea, mea_size);
1423 atomic_inc(&mds->mds_open_count);
1425 mds_body_do_reverse_map(med, body);
1428 * If we have not taken the "open" lock, we may not return 0 here,
1429 * because caller expects 0 to mean "lock is taken", and it needs
1430 * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1431 * client. Later caller should rewrite the return value back to zero
1432 * if it to be used any further.
1434 if ((cleanup_phase != 3) && !rc)
1440 /* Close a "file descriptor" and possibly unlink an orphan from the
1441 * PENDING directory. Caller must hold child->i_sem, this drops it.
1443 * If we are being called from mds_disconnect() because the client has
1444 * disappeared, then req == NULL and we do not update last_rcvd because
1445 * there is nothing that could be recovered by the client at this stage
1446 * (it will not even _have_ an entry in last_rcvd anymore).
1448 * Returns EAGAIN if the client needs to get more data and re-close. */
1449 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1450 struct obd_device *obd, struct mds_file_data *mfd,
1453 struct inode *inode = mfd->mfd_dentry->d_inode;
1454 char idname[LL_ID_NAMELEN];
1455 int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1456 struct dentry *pending_child = NULL;
1457 struct mds_obd *mds = &obd->u.mds;
1458 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1459 void *handle = NULL;
1460 struct mds_body *request_body = NULL, *reply_body = NULL;
1461 struct dentry_params dp;
1462 struct iattr iattr = { 0 };
1463 struct llog_create_locks *lcl = NULL;
1466 if (req && req->rq_reqmsg != NULL)
1467 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1468 sizeof(*request_body));
1469 if (req && req->rq_repmsg != NULL)
1470 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1471 sizeof(*reply_body));
1473 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1474 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1475 idname, inode->i_nlink, mds_orphan_open_count(inode));
1477 last_orphan = mds_orphan_open_dec_test(inode) &&
1478 mds_inode_is_orphan(inode);
1479 UP_WRITE_I_ALLOC_SEM(inode);
1481 /* this is half of the actual "close" */
1482 if (mfd->mfd_mode & FMODE_WRITE) {
1483 rc = mds_put_write_access(mds, inode, request_body,
1484 last_orphan && unlink_orphan);
1485 } else if (mfd->mfd_mode & FMODE_EXEC) {
1486 mds_allow_write_access(inode);
1489 if (last_orphan && unlink_orphan) {
1490 struct lov_mds_md *lmm = NULL;
1491 int stripe_count = 0;
1492 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1494 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1496 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1497 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1498 CERROR("found \"orphan\" %s %s with link count %d\n",
1499 S_ISREG(inode->i_mode) ? "file" : "dir",
1500 idname, inode->i_nlink);
1502 * sadly, there is no easy way to save pending_child from
1503 * mds_reint_unlink() into mfd, so we need to re-lookup, but
1504 * normally it will still be in the dcache.
1506 down(&pending_dir->i_sem);
1507 cleanup_phase = 1; /* up(i_sem) when finished */
1508 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1510 if (IS_ERR(pending_child))
1511 GOTO(cleanup, rc = PTR_ERR(pending_child));
1512 if (pending_child->d_inode == NULL) {
1513 CERROR("orphan %s has been removed\n", idname);
1514 GOTO(cleanup, rc = 0);
1517 cleanup_phase = 2; /* dput(pending_child) when finished */
1518 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1519 rc = vfs_rmdir(pending_dir, pending_child);
1521 CERROR("error unlinking orphan dir %s: rc %d\n",
1526 if (req != NULL && req->rq_repmsg != NULL) {
1527 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1528 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1531 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1532 NULL, stripe_count);
1533 if (IS_ERR(handle)) {
1534 rc = PTR_ERR(handle);
1539 pending_child->d_fsdata = (void *) &dp;
1542 rc = vfs_unlink(pending_dir, pending_child);
1544 CERROR("error unlinking orphan %s: rc %d\n",
1547 if (req != NULL && req->rq_repmsg != NULL &&
1548 (reply_body->valid & OBD_MD_FLEASIZE) &&
1549 mds_log_op_unlink(obd, pending_child->d_inode,
1550 lmm, req->rq_repmsg->buflens[1],
1551 lustre_msg_buf(req->rq_repmsg, 2, 0),
1552 req->rq_repmsg->buflens[2], &lcl) > 0) {
1553 reply_body->valid |= OBD_MD_FLCOOKIE;
1556 rc = mds_destroy_object(obd, inode, 1);
1558 CERROR("cannot destroy OSS object on close, err %d\n",
1563 goto out; /* Don't bother updating attrs on unlinked inode */
1567 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1568 /* Update the on-disk attributes if this was the last write
1569 * close, and all information was provided (i.e., rc == 0)
1571 * XXX this should probably be abstracted with mds_reint_setattr
1574 if (request_body->valid & OBD_MD_FLMTIME &&
1575 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1576 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1577 iattr.ia_valid |= ATTR_MTIME;
1579 if (request_body->valid & OBD_MD_FLCTIME &&
1580 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1581 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1582 iattr.ia_valid |= ATTR_CTIME;
1585 /* XXX can't set block count with fsfilt_setattr (!) */
1586 if (request_body->valid & OBD_MD_FLSIZE) {
1587 iattr.ia_valid |= ATTR_SIZE;
1588 iattr.ia_size = request_body->size;
1590 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1591 iattr.ia_valid |= ATTR_BLOCKS;
1592 iattr.ia_blocks = request_body->blocks
1597 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1598 /* Only start a transaction to write out only the atime if
1599 * it is more out-of-date than the specified limit. If we
1600 * are already going to write out the atime then do it anyway.
1602 if ((request_body->atime >
1603 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1604 (iattr.ia_valid != 0 &&
1605 request_body->atime > LTIME_S(inode->i_atime))) {
1606 LTIME_S(iattr.ia_atime) = request_body->atime;
1607 iattr.ia_valid |= ATTR_ATIME;
1611 if (iattr.ia_valid != 0) {
1612 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1614 GOTO(cleanup, rc = PTR_ERR(handle));
1615 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1617 CERROR("error in setattr(%s): rc %d\n", idname, rc);
1620 /* If other clients have this file open for write, rc will be > 0 */
1623 l_dput(mfd->mfd_dentry);
1624 mds_mfd_destroy(mfd);
1627 atomic_dec(&mds->mds_open_count);
1628 if (req != NULL && reply_body != NULL) {
1629 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1630 } else if (handle) {
1631 int err, force_sync = 0;
1633 if (req && req->rq_export)
1634 force_sync = req->rq_export->exp_sync;
1636 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1639 CERROR("error committing close: %d\n", err);
1645 switch (cleanup_phase) {
1648 ptlrpc_save_llog_lock(req, lcl);
1649 dput(pending_child);
1651 up(&pending_dir->i_sem);
1656 int mds_close(struct ptlrpc_request *req, int offset)
1658 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1659 struct obd_device *obd = req->rq_export->exp_obd;
1660 struct mds_body *body;
1661 struct mds_file_data *mfd;
1662 struct lvfs_run_ctxt saved;
1663 struct inode *inode;
1664 int rc, repsize[3] = {sizeof(struct mds_body),
1665 obd->u.mds.mds_max_mdsize,
1666 obd->u.mds.mds_max_cookiesize};
1668 MD_COUNTER_INCREMENT(obd, close);
1670 rc = lustre_pack_reply(req, 3, repsize, NULL);
1672 CERROR("lustre_pack_reply: rc = %d\n", rc);
1673 req->rq_status = rc;
1675 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1678 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1679 DEBUG_REQ(D_HA, req, "close replay");
1680 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1681 lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1682 req->rq_repmsg->buflens[2]);
1685 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1686 lustre_swab_mds_body);
1688 CERROR("Can't unpack body\n");
1689 req->rq_status = -EFAULT;
1693 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1694 /* do some stuff */ ;
1696 mfd = mds_handle2mfd(&body->handle);
1698 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1699 ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1700 req->rq_status = -ESTALE;
1704 inode = mfd->mfd_dentry->d_inode;
1705 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1706 DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1707 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1708 struct mds_body *rep_body;
1710 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1711 sizeof (*rep_body));
1712 LASSERT(rep_body != NULL);
1714 mds_pack_inode2body(obd, rep_body, inode,
1715 (body->valid & OBD_MD_FID) ? 1 : 0);
1717 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1718 inode, MDS_PACK_MD_LOCK, 0);
1720 spin_lock(&med->med_open_lock);
1721 list_del(&mfd->mfd_list);
1722 spin_unlock(&med->med_open_lock);
1724 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1725 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1726 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1728 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1729 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1730 req->rq_status = -ENOMEM;
1739 int mds_done_writing(struct ptlrpc_request *req, int offset)
1741 struct mds_body *body;
1742 int rc, size = sizeof(struct mds_body);
1745 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1747 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1748 lustre_swab_mds_body);
1750 CERROR("Can't unpack body\n");
1751 req->rq_status = -EFAULT;
1755 rc = lustre_pack_reply(req, 1, &size, NULL);
1757 CERROR("lustre_pack_reply: rc = %d\n", rc);
1758 req->rq_status = rc;