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);
426 CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR),
427 "error creating objects for "
428 "inode %lu: rc = %d\n",
431 CERROR("obd_create returned invalid "
438 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
439 0, &lsm, rec->ur_eadata);
443 lsm->lsm_object_id = oa->o_id;
444 lsm->lsm_object_gr = oa->o_gr;
447 oa->o_size = inode->i_size;
448 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
449 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE);
451 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
453 CERROR("error setting attrs for inode %lu: rc %d\n",
456 CERROR("obd_setattr returned bad rc %d\n", rc);
463 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
464 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
466 LASSERT(lsm && lsm->lsm_object_id);
468 rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
469 if (!id_ino(rec->ur_id2))
470 obd_free_memmd(mds->mds_dt_exp, &lsm);
472 CERROR("cannot pack lsm, err = %d\n", rc);
477 body->eadatasize = rc;
480 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
481 if (IS_ERR(*handle)) {
482 rc = PTR_ERR(*handle);
487 rc = fsfilt_set_md(obd, inode, *handle, lmm,
490 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
491 lmm_bufsize = req->rq_repmsg->buflens[offset];
493 LASSERT(lmm_bufsize >= lmm_size);
495 memcpy(lmm_buf, lmm, lmm_size);
496 obd_free_diskmd(mds->mds_dt_exp, &lmm);
498 oti_free_cookies(&oti);
502 obd_free_memmd(mds->mds_dt_exp, &lsm);
507 mds_destroy_object(struct obd_device *obd,
508 struct inode *inode, int async)
510 struct mds_obd *mds = &obd->u.mds;
511 struct lov_mds_md *lmm = NULL;
515 LASSERT(inode != NULL);
517 if (inode->i_nlink != 0) {
518 CDEBUG(D_INODE, "attempt to destroy OSS object when "
519 "i_nlink == %d\n", (int)inode->i_nlink);
523 OBD_ALLOC(lmm, mds->mds_max_mdsize);
527 lmm_size = mds->mds_max_mdsize;
528 rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
530 CERROR("no stripe info for %lu/%lu inode\n",
531 (unsigned long)inode->i_ino,
532 (unsigned long)inode->i_generation);
533 GOTO(out_free_lmm, rc);
537 /* asynchronously unlink objecect on OSS */
538 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
541 CERROR("error unlinking object on OSS, "
543 GOTO(out_free_lmm, rc);
546 CDEBUG(D_INODE, "no stripping info found for inode "
547 "%lu/%lu\n", (unsigned long)inode->i_ino,
548 (unsigned long)inode->i_generation);
552 OBD_FREE(lmm, mds->mds_max_mdsize);
556 static void reconstruct_open(struct mds_update_record *rec, int offset,
557 struct ptlrpc_request *req,
558 struct lustre_handle *child_lockh)
560 struct mds_export_data *med = &req->rq_export->exp_mds_data;
561 struct mds_client_data *mcd = med->med_mcd;
562 struct mds_obd *mds = mds_req2mds(req);
563 struct mds_file_data *mfd;
564 struct obd_device *obd = req->rq_export->exp_obd;
565 struct dentry *parent = NULL, *dchild;
566 struct ldlm_reply *rep;
567 struct mds_body *body;
573 LASSERT(offset == 3); /* only called via intent */
574 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
575 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
577 /* copy rc, transno and disp; steal locks */
578 mds_req_from_mcd(req, mcd);
579 intent_set_disposition(rep, mcd->mcd_last_data);
581 /* Only replay if create or open actually happened. */
582 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
584 return; /* error looking up parent or child */
587 if (rec->ur_namelen == 1) {
588 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
590 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
592 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
593 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
594 OLID4(rec->ur_id1), PTR_ERR(parent));
595 dchild = ll_lookup_one_len(rec->ur_name, parent,
596 rec->ur_namelen - 1);
597 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
598 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
601 if (!dchild->d_inode)
602 GOTO(out_dput, 0); /* child not present to open */
604 /* At this point, we know we have a child. We'll send
605 * it back _unless_ it not created and open failed. */
606 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
607 !intent_disposition(rep, DISP_OPEN_CREATE) &&
611 /* get lock (write for O_CREAT, read otherwise) */
612 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
613 if (S_ISREG(dchild->d_inode->i_mode)) {
614 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
615 dchild->d_inode, 1, 0);
618 LASSERT(rc == req->rq_status);
620 /* If we have LOV EA data, the OST holds size, mtime */
621 if (!(body->valid & OBD_MD_FLEASIZE))
622 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
623 OBD_MD_FLATIME | OBD_MD_FLMTIME);
626 /* If we have -EEXIST as the status, and we were asked to create
627 * exclusively, we can tell we failed because the file already existed.
629 if (req->rq_status == -EEXIST &&
630 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
631 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
635 /* If we didn't get as far as trying to open, then some locking thing
636 * probably went wrong, and we'll just bail here.
638 if (!intent_disposition(rep, DISP_OPEN_OPEN))
641 /* If we failed, then we must have failed opening, so don't look for
642 * file descriptor or anything, just give the client the bad news.
648 list_for_each(t, &med->med_open_head) {
649 mfd = list_entry(t, struct mds_file_data, mfd_list);
650 if (mfd->mfd_xid == req->rq_xid)
655 /* #warning "XXX fixme" bug 2991 */
656 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
657 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
658 * to detect a re-open */
660 mntget(mds->mds_vfsmnt);
661 CERROR("Re-opened file \n");
662 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
663 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
665 CERROR("mds: out of memory\n");
666 GOTO(out_dput, req->rq_status = -ENOMEM);
670 body->handle.cookie = mfd->mfd_handle.h_cookie;
671 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
672 mfd->mfd_handle.h_cookie);
683 /* do NOT or the MAY_*'s, you'll get the weakest */
684 static int accmode(int flags)
688 if (flags & FMODE_READ)
690 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
692 if (flags & FMODE_EXEC)
697 /* Handles object creation, actual opening, and I/O epoch */
698 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
699 struct mds_body *body, int flags, void **handle,
700 struct mds_update_record *rec, struct ldlm_reply *rep)
702 struct obd_device *obd = req->rq_export->exp_obd;
703 struct mds_obd *mds = mds_req2mds(req);
704 struct mds_file_data *mfd = NULL;
710 /* atomically create objects if necessary */
711 down(&dchild->d_inode->i_sem);
712 mode = dchild->d_inode->i_mode;
714 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
715 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
716 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
717 dchild->d_inode, 0, 0);
719 up(&dchild->d_inode->i_sem);
725 if ((body->valid & OBD_MD_FLEASIZE) &&
726 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
727 up(&dchild->d_inode->i_sem);
731 if (!(body->valid & OBD_MD_FLEASIZE)) {
732 /* no EA: create objects */
733 rc = mds_create_objects(obd, req, 2, rec,
734 dchild, handle, &ids);
736 CERROR("mds_create_object: rc = %d\n", rc);
737 up(&dchild->d_inode->i_sem);
742 if (S_ISREG(dchild->d_inode->i_mode) &&
743 (body->valid & OBD_MD_FLEASIZE)) {
744 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
747 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
748 dchild->d_inode, 0, 0);
750 up(&dchild->d_inode->i_sem);
756 rc = mds_pack_acl(obd, req->rq_repmsg, 3, body, dchild->d_inode);
758 CERROR("mds_pack_acl: rc = %d\n", rc);
759 up(&dchild->d_inode->i_sem);
763 /* If the inode has no EA data, then MDSs hold size, mtime */
764 if (S_ISREG(dchild->d_inode->i_mode) &&
765 !(body->valid & OBD_MD_FLEASIZE)) {
766 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
767 OBD_MD_FLATIME | OBD_MD_FLMTIME);
770 up(&dchild->d_inode->i_sem);
772 intent_set_disposition(rep, DISP_OPEN_OPEN);
773 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
775 RETURN(PTR_ERR(mfd));
777 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
778 mfd->mfd_handle.h_cookie);
781 mds_dt_update_objids(obd, ids);
782 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
788 static int mds_open_by_id(struct ptlrpc_request *req,
789 struct lustre_id *id,
790 struct mds_body *body, int flags,
791 struct mds_update_record *rec,
792 struct ldlm_reply *rep)
794 struct mds_obd *mds = mds_req2mds(req);
795 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
796 struct dentry *dchild;
797 char idname[LL_ID_NAMELEN];
802 down(&pending_dir->i_sem);
803 idlen = ll_id2str(idname, id_ino(id), id_gen(id));
804 dchild = lookup_one_len(idname, mds->mds_pending_dir,
806 if (IS_ERR(dchild)) {
807 up(&pending_dir->i_sem);
808 rc = PTR_ERR(dchild);
809 CERROR("error looking up %s in PENDING: rc = %d\n",
814 up(&pending_dir->i_sem);
815 if (dchild->d_inode != NULL) {
816 mds_inode_set_orphan(dchild->d_inode);
817 mds_pack_inode2body(req2obd(req), body,
820 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
821 intent_set_disposition(rep, DISP_LOOKUP_POS);
822 CWARN("Orphan %s found and opened in PENDING directory\n",
829 * we didn't find it in PENDING so it isn't an orphan. See if it was a
830 * regular inode that was previously created.
832 dchild = mds_id2dentry(req2obd(req), id, NULL);
834 RETURN(PTR_ERR(dchild));
836 mds_pack_inode2body(req2obd(req), body,
839 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
840 intent_set_disposition(rep, DISP_LOOKUP_POS);
843 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
844 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
845 req, rc, rep ? rep->lock_policy_res1 : 0);
846 /* XXX what do we do here if mds_finish_transno itself failed? */
851 int mds_pin(struct ptlrpc_request *req, int offset)
853 struct obd_device *obd = req->rq_export->exp_obd;
854 struct mds_body *request_body, *reply_body;
855 struct lvfs_run_ctxt saved;
856 int rc, size = sizeof(*reply_body);
859 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
860 sizeof(*request_body));
862 rc = lustre_pack_reply(req, 1, &size, NULL);
865 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
867 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
868 rc = mds_open_by_id(req, &request_body->id1, reply_body,
869 request_body->flags, NULL, NULL);
870 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
876 * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
877 * child_lockh is NULL we just get the lock as a barrier to wait for other
878 * holders of this lock, and drop it right away again.
880 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
881 struct lustre_handle *child_lockh)
883 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
884 struct lustre_handle lockh;
889 if (child_lockh == NULL)
890 child_lockh = &lockh;
892 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
893 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
894 mds_blocking_ast, ldlm_completion_ast, NULL,
895 NULL, NULL, 0, NULL, child_lockh);
897 CERROR("ldlm_cli_enqueue: %d\n", rc);
898 else if (child_lockh == &lockh)
899 ldlm_lock_decref(child_lockh, LCK_EX);
904 int mds_open(struct mds_update_record *rec, int offset,
905 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
907 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
908 struct obd_device *obd = req->rq_export->exp_obd;
909 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
910 struct mds_obd *mds = mds_req2mds(req);
911 struct ldlm_reply *rep = NULL;
912 struct mds_body *body = NULL;
913 struct dentry *dchild = NULL, *dparent = NULL;
914 struct lustre_handle parent_lockh[2] = {{0}, {0}};
915 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
916 int parent_mode = LCK_PR;
918 struct dentry_params dp;
919 struct mea *mea = NULL;
920 int mea_size, update_mode;
921 int child_mode = LCK_PR;
924 DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
925 OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
928 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
929 (obd_timeout + 1) / 4);
931 if (offset == 3) { /* intent */
932 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
933 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
934 } else if (offset == 1) { /* non-intent reint */
935 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
941 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
944 * Step 0: If we are passed a id, then we assume the client already
945 * opened this file and is only replaying the RPC, so we open the inode
946 * by fid (at some large expense in security).
948 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
949 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
950 OLID4(rec->ur_id2), rec->ur_namelen - 1,
951 rec->ur_name, rec->ur_mode);
953 LASSERT(id_ino(rec->ur_id2));
955 rc = mds_open_by_id(req, rec->ur_id2, body,
956 rec->ur_flags, rec, rep);
958 mds_body_do_reverse_map(med, body);
962 /* We didn't find the correct inode on disk either, so we
963 * need to re-create it via a regular replay. */
964 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
966 LASSERT(!id_ino(rec->ur_id2));
969 LASSERT(offset == 3); /* If we got here, we must be called via intent */
971 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
972 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
976 acc_mode = accmode(rec->ur_flags);
978 /* Step 1: Find and lock the parent */
979 if (rec->ur_flags & MDS_OPEN_CREAT) {
980 /* XXX Well, in fact we only need this lock mode change if
981 in addition to O_CREAT, the file does not exist.
982 But we do not know if it exists or not yet */
983 parent_mode = LCK_PW;
986 if (rec->ur_namelen == 1) {
987 /* client (LMV) wants to open the file by id */
988 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
989 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
990 if (IS_ERR(dchild)) {
991 rc = PTR_ERR(dparent);
992 CERROR("child lookup by an id error %d\n", rc);
999 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
1000 parent_lockh, &update_mode, rec->ur_name,
1001 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
1002 if (IS_ERR(dparent)) {
1003 rc = PTR_ERR(dparent);
1004 if (rc != -ENOENT) {
1005 CERROR("parent lookup for "DLID4" failed, error %d\n",
1006 OLID4(rec->ur_id1), rc);
1008 /* Just cannot find parent - make it look like
1009 usual negative lookup to avoid extra MDS RPC */
1010 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1011 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1015 LASSERT(dparent->d_inode != NULL);
1017 cleanup_phase = 1; /* parent dentry and lock */
1019 /* try to retrieve MEA data for this dir */
1020 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1026 * dir is already splitted, check is requested filename should *
1027 * live at this MDS or at another one.
1030 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1031 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1033 "%s: inapropriate MDS(%d) for %lu/%u:%s."
1034 " should be %lu(%d)\n", obd->obd_name,
1035 mea->mea_master, dparent->d_inode->i_ino,
1036 dparent->d_inode->i_generation, rec->ur_name,
1037 (unsigned long)id_group(&mea->mea_ids[i]), i);
1038 GOTO(cleanup, rc = -ERESTART);
1042 /* Step 2: Lookup the child */
1043 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1044 if (IS_ERR(dchild)) {
1045 rc = PTR_ERR(dchild);
1046 dchild = NULL; /* don't confuse mds_finish_transno() below */
1051 cleanup_phase = 2; /* child dentry */
1053 if (dchild->d_flags & DCACHE_CROSS_REF) {
1054 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1055 LASSERT(rec->ur_namelen > 1);
1057 /* we're gonna acquire LOOKUP lock on the child,
1058 * but we have already locked parent and our order
1059 * may conflict with enqueue_order_locks(). so,
1060 * drop parent lock and acquire both the locks in
1061 * common order. bug 6190 */
1063 if (parent_lockh[1].cookie != 0)
1064 ldlm_lock_decref(parent_lockh + 1, update_mode);
1066 ldlm_lock_decref(parent_lockh, parent_mode);
1071 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1072 parent_lockh, &dparent,
1073 LCK_PR, MDS_INODELOCK_UPDATE,
1074 &update_mode, rec->ur_name,
1075 rec->ur_namelen, child_lockh,
1077 MDS_INODELOCK_LOOKUP);
1083 if (parent_lockh[1].cookie != 0)
1084 ldlm_lock_decref(parent_lockh + 1, update_mode);
1086 ldlm_lock_decref(parent_lockh, LCK_PR);
1088 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1089 CDEBUG(D_OTHER, "race: name changed (%p)\n",
1091 if (dchild->d_inode)
1092 ldlm_lock_decref(child_lockh, LCK_PR);
1098 mds_pack_dentry2body(obd, body, dchild, 1);
1099 intent_set_disposition(rep, DISP_LOOKUP_POS);
1100 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1103 OBD_FREE(mea, mea_size);
1109 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1111 if (dchild->d_inode)
1112 intent_set_disposition(rep, DISP_LOOKUP_POS);
1114 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1116 /* Step 3: If the child was negative, and we're supposed to, create it.*/
1117 if (dchild->d_inode == NULL) {
1120 struct inode *inode;
1122 ino = (unsigned long)id_ino(rec->ur_id2);
1124 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1127 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1128 obd->obd_name, dparent->d_inode->i_ino,
1129 dparent->d_inode->i_generation, rc);
1132 /* dir got splitted */
1133 GOTO(cleanup, rc = -ERESTART);
1134 } else if (rc < 0) {
1135 /* error happened during spitting */
1139 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1141 * dentry is negative and we weren't supposed to create
1142 * object, get out of here.
1144 GOTO(cleanup, rc = -ENOENT);
1147 intent_set_disposition(rep, DISP_OPEN_CREATE);
1148 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1150 if (IS_ERR(handle)) {
1151 rc = PTR_ERR(handle);
1155 dchild->d_fsdata = (void *) &dp;
1159 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1160 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1161 dchild->d_fsdata = NULL;
1164 CDEBUG(D_INODE, "error during create: %d\n", rc);
1167 inode = dchild->d_inode;
1169 LASSERT(ino == inode->i_ino);
1171 /* written as part of setattr */
1172 inode->i_generation = id_gen(rec->ur_id2);
1173 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1174 inode->i_ino, inode->i_generation);
1178 LTIME_S(iattr.ia_atime) = rec->ur_time;
1179 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1180 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1182 iattr.ia_uid = rec->ur_fsuid;
1183 if (dparent->d_inode->i_mode & S_ISGID)
1184 iattr.ia_gid = dparent->d_inode->i_gid;
1186 iattr.ia_gid = rec->ur_fsgid;
1188 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1189 ATTR_MTIME | ATTR_CTIME;
1191 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1193 CERROR("error on child setattr: rc = %d\n", rc);
1195 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1197 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1199 CERROR("error on parent setattr: rc = %d\n", rc);
1201 MD_COUNTER_INCREMENT(obd, create);
1205 rc = mds_update_inode_sid(obd, dchild->d_inode,
1206 handle, rec->ur_id2);
1208 CERROR("mds_update_inode_sid() failed, "
1211 id_assign_fid(&body->id1, rec->ur_id2);
1214 * make sure, that fid is up-to-date.
1216 mds_set_last_fid(obd, id_fid(rec->ur_id2));
1218 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1219 handle, &body->id1);
1221 CERROR("mds_alloc_inode_sid() failed, "
1226 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1227 rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1228 dchild->d_inode, handle,
1229 req->rq_export->exp_sync);
1233 acc_mode = 0; /* Don't check for permissions */
1235 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1237 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1238 "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1239 dchild->d_name.name, dchild, dchild->d_inode);
1241 if (S_ISREG(dchild->d_inode->i_mode)) {
1242 /* Check permissions etc */
1243 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1247 /* Can't write to a read-only file */
1248 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1249 GOTO(cleanup, rc = -EPERM);
1251 /* An append-only file must be opened in append mode for
1253 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1254 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1255 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1256 GOTO(cleanup, rc = -EPERM);
1259 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1260 (rec->ur_flags & MDS_OPEN_EXCL)) {
1261 /* File already exists, we didn't just create it, and we
1262 * were passed O_EXCL; err-or. */
1263 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1266 if (S_ISDIR(dchild->d_inode->i_mode)) {
1267 if (rec->ur_flags & MDS_OPEN_CREAT ||
1268 rec->ur_flags & FMODE_WRITE) {
1269 /* we are trying to create or write a exist dir */
1270 GOTO(cleanup, rc = -EISDIR);
1272 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1273 GOTO(cleanup, rc = -EACCES);
1276 * here was checking for possible GNS mount points to skip their
1277 * open. I removed it as its detection based only on SUID bit is
1278 * not reliable. Opening such a dirs should not cause any
1279 * problems, at least tests show that. --umka
1283 /* if we are following a symlink, don't open */
1284 if (S_ISLNK(dchild->d_inode->i_mode))
1285 GOTO(cleanup_no_trans, rc = 0);
1287 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1288 !S_ISDIR(dchild->d_inode->i_mode))
1289 GOTO(cleanup, rc = -ENOTDIR);
1291 /* check permission even it's special files */
1292 if (S_ISCHR(dchild->d_inode->i_mode) ||
1293 S_ISBLK(dchild->d_inode->i_mode) ||
1294 S_ISFIFO(dchild->d_inode->i_mode) ||
1295 S_ISSOCK(dchild->d_inode->i_mode)) {
1296 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1301 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1302 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1303 GOTO(cleanup, rc = -EAGAIN);
1307 /* We cannot use acc_mode here, because it is zeroed in case of
1308 creating a file, so we get wrong lockmode */
1309 if (accmode(rec->ur_flags) & MAY_WRITE)
1310 child_mode = LCK_CW;
1311 else if (accmode(rec->ur_flags) & MAY_EXEC)
1312 child_mode = LCK_PR;
1314 child_mode = LCK_CR;
1316 /* Always returning LOOKUP lock if open succesful to guard
1317 * dentry on client. In case of replay we do not get a lock
1318 * assuming that the caller has it already */
1319 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1320 struct ldlm_res_id child_res_id = { .name = {0}};
1321 ldlm_policy_data_t policy;
1324 /* LOOKUP lock will protect dentry on client -bzzz */
1325 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1328 child_res_id.name[0] = id_fid(&body->id1);
1329 child_res_id.name[1] = id_group(&body->id1);
1331 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1332 child_res_id, LDLM_IBITS, &policy,
1333 child_mode, &lock_flags,
1334 mds_blocking_ast, ldlm_completion_ast,
1335 NULL, NULL, NULL, 0, NULL, child_lockh);
1342 /* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */
1343 #warning "disable opencache lock for CMD2"
1346 /* Step 5: mds_open it */
1347 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1352 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1353 req, rc, rep ? rep->lock_policy_res1 : 0);
1356 switch (cleanup_phase) {
1359 ldlm_lock_decref(child_lockh, child_mode);
1360 child_lockh->cookie = 0;
1363 if (rc && created) {
1364 int err = vfs_unlink(dparent->d_inode, dchild);
1366 CERROR("unlink(%.*s) in error path: %d\n",
1367 dchild->d_name.len, dchild->d_name.name,
1370 } else if (created) {
1371 mds_lock_new_child(obd, dchild->d_inode, NULL);
1375 if (dparent == NULL)
1380 if (parent_lockh[1].cookie != 0)
1381 ldlm_lock_decref(parent_lockh + 1, update_mode);
1384 ldlm_lock_decref(parent_lockh, parent_mode);
1386 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1389 OBD_FREE(mea, mea_size);
1391 atomic_inc(&mds->mds_open_count);
1393 mds_body_do_reverse_map(med, body);
1396 * If we have not taken the "open" lock, we may not return 0 here,
1397 * because caller expects 0 to mean "lock is taken", and it needs
1398 * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1399 * client. Later caller should rewrite the return value back to zero
1400 * if it to be used any further.
1402 if ((cleanup_phase != 3) && !rc)
1408 /* Close a "file descriptor" and possibly unlink an orphan from the
1409 * PENDING directory. Caller must hold child->i_sem, this drops it.
1411 * If we are being called from mds_disconnect() because the client has
1412 * disappeared, then req == NULL and we do not update last_rcvd because
1413 * there is nothing that could be recovered by the client at this stage
1414 * (it will not even _have_ an entry in last_rcvd anymore).
1416 * Returns EAGAIN if the client needs to get more data and re-close. */
1417 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1418 struct obd_device *obd, struct mds_file_data *mfd,
1421 struct inode *inode = mfd->mfd_dentry->d_inode;
1422 char idname[LL_ID_NAMELEN];
1423 int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1424 struct dentry *pending_child = NULL;
1425 struct mds_obd *mds = &obd->u.mds;
1426 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1427 void *handle = NULL;
1428 struct mds_body *request_body = NULL, *reply_body = NULL;
1429 struct dentry_params dp;
1430 struct iattr iattr = { 0 };
1431 struct llog_create_locks *lcl = NULL;
1434 if (req && req->rq_reqmsg != NULL)
1435 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1436 sizeof(*request_body));
1437 if (req && req->rq_repmsg != NULL)
1438 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1439 sizeof(*reply_body));
1441 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1442 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1443 idname, inode->i_nlink, mds_orphan_open_count(inode));
1445 last_orphan = mds_orphan_open_dec_test(inode) &&
1446 mds_inode_is_orphan(inode);
1447 UP_WRITE_I_ALLOC_SEM(inode);
1449 /* this is half of the actual "close" */
1450 if (mfd->mfd_mode & FMODE_WRITE) {
1451 rc = mds_put_write_access(mds, inode, request_body,
1452 last_orphan && unlink_orphan);
1453 } else if (mfd->mfd_mode & FMODE_EXEC) {
1454 mds_allow_write_access(inode);
1457 if (last_orphan && unlink_orphan) {
1458 struct lov_mds_md *lmm = NULL;
1459 int stripe_count = 0;
1460 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1462 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1464 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1465 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1466 CERROR("found \"orphan\" %s %s with link count %d\n",
1467 S_ISREG(inode->i_mode) ? "file" : "dir",
1468 idname, inode->i_nlink);
1470 * sadly, there is no easy way to save pending_child from
1471 * mds_reint_unlink() into mfd, so we need to re-lookup, but
1472 * normally it will still be in the dcache.
1474 down(&pending_dir->i_sem);
1475 cleanup_phase = 1; /* up(i_sem) when finished */
1476 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1478 if (IS_ERR(pending_child))
1479 GOTO(cleanup, rc = PTR_ERR(pending_child));
1480 if (pending_child->d_inode == NULL) {
1481 CERROR("orphan %s has been removed\n", idname);
1482 GOTO(cleanup, rc = 0);
1485 cleanup_phase = 2; /* dput(pending_child) when finished */
1486 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1487 rc = vfs_rmdir(pending_dir, pending_child);
1489 CERROR("error unlinking orphan dir %s: rc %d\n",
1494 if (req != NULL && req->rq_repmsg != NULL) {
1495 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1496 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1499 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1500 NULL, stripe_count);
1501 if (IS_ERR(handle)) {
1502 rc = PTR_ERR(handle);
1507 pending_child->d_fsdata = (void *) &dp;
1510 rc = vfs_unlink(pending_dir, pending_child);
1512 CERROR("error unlinking orphan %s: rc %d\n",
1515 if (req != NULL && req->rq_repmsg != NULL &&
1516 (reply_body->valid & OBD_MD_FLEASIZE) &&
1517 mds_log_op_unlink(obd, pending_child->d_inode,
1518 lmm, req->rq_repmsg->buflens[1],
1519 lustre_msg_buf(req->rq_repmsg, 2, 0),
1520 req->rq_repmsg->buflens[2], &lcl) > 0) {
1521 reply_body->valid |= OBD_MD_FLCOOKIE;
1524 rc = mds_destroy_object(obd, inode, 1);
1526 CERROR("cannot destroy OSS object on close, err %d\n",
1531 goto out; /* Don't bother updating attrs on unlinked inode */
1535 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1536 /* Update the on-disk attributes if this was the last write
1537 * close, and all information was provided (i.e., rc == 0)
1539 * XXX this should probably be abstracted with mds_reint_setattr
1542 if (request_body->valid & OBD_MD_FLMTIME &&
1543 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1544 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1545 iattr.ia_valid |= ATTR_MTIME;
1547 if (request_body->valid & OBD_MD_FLCTIME &&
1548 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1549 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1550 iattr.ia_valid |= ATTR_CTIME;
1553 /* XXX can't set block count with fsfilt_setattr (!) */
1554 if (request_body->valid & OBD_MD_FLSIZE) {
1555 iattr.ia_valid |= ATTR_SIZE;
1556 iattr.ia_size = request_body->size;
1558 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1559 iattr.ia_valid |= ATTR_BLOCKS;
1560 iattr.ia_blocks = request_body->blocks
1565 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1566 /* Only start a transaction to write out only the atime if
1567 * it is more out-of-date than the specified limit. If we
1568 * are already going to write out the atime then do it anyway.
1570 if ((request_body->atime >
1571 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1572 (iattr.ia_valid != 0 &&
1573 request_body->atime > LTIME_S(inode->i_atime))) {
1574 LTIME_S(iattr.ia_atime) = request_body->atime;
1575 iattr.ia_valid |= ATTR_ATIME;
1579 if (iattr.ia_valid != 0) {
1580 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1582 GOTO(cleanup, rc = PTR_ERR(handle));
1583 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1585 CERROR("error in setattr(%s): rc %d\n", idname, rc);
1588 /* If other clients have this file open for write, rc will be > 0 */
1591 l_dput(mfd->mfd_dentry);
1592 mds_mfd_destroy(mfd);
1595 atomic_dec(&mds->mds_open_count);
1596 if (req != NULL && reply_body != NULL) {
1597 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1598 } else if (handle) {
1599 int err, force_sync = 0;
1601 if (req && req->rq_export)
1602 force_sync = req->rq_export->exp_sync;
1604 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1607 CERROR("error committing close: %d\n", err);
1613 switch (cleanup_phase) {
1616 ptlrpc_save_llog_lock(req, lcl);
1617 dput(pending_child);
1619 up(&pending_dir->i_sem);
1624 int mds_close(struct ptlrpc_request *req, int offset)
1626 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1627 struct obd_device *obd = req->rq_export->exp_obd;
1628 struct mds_body *body;
1629 struct mds_file_data *mfd;
1630 struct lvfs_run_ctxt saved;
1631 struct inode *inode;
1632 int rc, repsize[3] = {sizeof(struct mds_body),
1633 obd->u.mds.mds_max_mdsize,
1634 obd->u.mds.mds_max_cookiesize};
1636 MD_COUNTER_INCREMENT(obd, close);
1638 rc = lustre_pack_reply(req, 3, repsize, NULL);
1640 CERROR("lustre_pack_reply: rc = %d\n", rc);
1641 req->rq_status = rc;
1643 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1646 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1647 DEBUG_REQ(D_HA, req, "close replay");
1648 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1649 lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1650 req->rq_repmsg->buflens[2]);
1653 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1654 lustre_swab_mds_body);
1656 CERROR("Can't unpack body\n");
1657 req->rq_status = -EFAULT;
1661 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1662 /* do some stuff */ ;
1664 mfd = mds_handle2mfd(&body->handle);
1666 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1667 ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1668 req->rq_status = -ESTALE;
1672 inode = mfd->mfd_dentry->d_inode;
1673 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1674 DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1675 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1676 struct mds_body *rep_body;
1678 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1679 sizeof (*rep_body));
1680 LASSERT(rep_body != NULL);
1682 mds_pack_inode2body(obd, rep_body, inode,
1683 (body->valid & OBD_MD_FID) ? 1 : 0);
1685 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1686 inode, MDS_PACK_MD_LOCK, 0);
1688 spin_lock(&med->med_open_lock);
1689 list_del(&mfd->mfd_list);
1690 spin_unlock(&med->med_open_lock);
1692 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1693 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1694 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1696 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1697 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1698 req->rq_status = -ENOMEM;
1707 int mds_done_writing(struct ptlrpc_request *req, int offset)
1709 struct mds_body *body;
1710 int rc, size = sizeof(struct mds_body);
1713 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1715 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1716 lustre_swab_mds_body);
1718 CERROR("Can't unpack body\n");
1719 req->rq_status = -EFAULT;
1723 rc = lustre_pack_reply(req, 1, &size, NULL);
1725 CERROR("lustre_pack_reply: rc = %d\n", rc);
1726 req->rq_status = rc;