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/lustre_mds.h>
44 #include <linux/obd_lov.h>
45 #include <linux/lustre_fsfilt.h>
46 #include <linux/lprocfs_status.h>
48 #include "mds_internal.h"
50 /* Exported function from this file are:
52 * mds_open - called by the intent handler
53 * mds_close - an rpc handling function
54 * mds_pin - an rpc handling function - which will go away
55 * mds_mfd_close - for force closing files when a client dies
59 * MDS file data handling: file data holds a handle for a file opened
63 static void mds_mfd_addref(void *mfdp)
65 struct mds_file_data *mfd = mfdp;
67 atomic_inc(&mfd->mfd_refcount);
68 CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
69 atomic_read(&mfd->mfd_refcount));
72 struct mds_file_data *mds_mfd_new(void)
74 struct mds_file_data *mfd;
76 OBD_ALLOC(mfd, sizeof *mfd);
78 CERROR("mds: out of memory\n");
82 atomic_set(&mfd->mfd_refcount, 2);
84 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
85 class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
90 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
93 LASSERT(handle != NULL);
94 RETURN(class_handle2object(handle->cookie));
97 static void mds_mfd_put(struct mds_file_data *mfd)
99 CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
100 atomic_read(&mfd->mfd_refcount) - 1);
101 LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
102 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
103 if (atomic_dec_and_test(&mfd->mfd_refcount)) {
104 LASSERT(list_empty(&mfd->mfd_handle.h_link));
105 OBD_FREE(mfd, sizeof *mfd);
109 static void mds_mfd_destroy(struct mds_file_data *mfd)
111 class_handle_unhash(&mfd->mfd_handle);
116 /* Caller must hold mds->mds_epoch_sem */
117 static int mds_alloc_filterdata(struct inode *inode)
119 LASSERT(inode->i_filterdata == NULL);
120 OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
121 if (inode->i_filterdata == NULL)
123 LASSERT(igrab(inode) == inode);
127 /* Caller must hold mds->mds_epoch_sem */
128 static void mds_free_filterdata(struct inode *inode)
130 LASSERT(inode->i_filterdata != NULL);
131 OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
132 inode->i_filterdata = NULL;
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);
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 if (epoch > mds->mds_io_epoch)
174 mds->mds_io_epoch = epoch;
177 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
178 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
179 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
182 atomic_inc(&inode->i_writecount);
183 up(&mds->mds_epoch_sem);
187 /* Returns EAGAIN if the client needs to get size and/or cookies and close
188 * again -- which is never true if the file is about to be unlinked. Otherwise
189 * returns the number of remaining writers. */
190 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
191 struct mds_body *body, int unlinking)
196 down(&mds->mds_epoch_sem);
197 atomic_dec(&inode->i_writecount);
198 rc = atomic_read(&inode->i_writecount);
202 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
203 GOTO(out, rc = EAGAIN);
205 mds_free_filterdata(inode);
207 up(&mds->mds_epoch_sem);
211 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
214 down(&mds->mds_epoch_sem);
215 if (atomic_read(&inode->i_writecount) > 0) {
216 up(&mds->mds_epoch_sem);
219 atomic_dec(&inode->i_writecount);
220 up(&mds->mds_epoch_sem);
224 static void mds_allow_write_access(struct inode *inode)
227 atomic_inc(&inode->i_writecount);
230 int mds_query_write_access(struct inode *inode)
233 RETURN(atomic_read(&inode->i_writecount));
236 /* This replaces the VFS mds_dentry_open, it manages mfd and writecount */
237 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
238 struct vfsmount *mnt, int flags,
239 struct ptlrpc_request *req)
241 struct mds_export_data *med = &req->rq_export->exp_mds_data;
242 struct mds_obd *mds = mds_req2mds(req);
243 struct mds_file_data *mfd;
244 struct mds_body *body;
250 CERROR("mds: out of memory\n");
251 GOTO(cleanup_dentry, error = -ENOMEM);
254 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
256 if (flags & FMODE_WRITE) {
257 /* FIXME: in recovery, need to pass old epoch here */
258 error = mds_get_write_access(mds, dentry->d_inode, 0);
260 GOTO(cleanup_mfd, error);
261 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
262 } else if (flags & FMODE_EXEC) {
263 error = mds_deny_write_access(mds, dentry->d_inode);
265 GOTO(cleanup_mfd, error);
270 /* Mark the file as open to handle open-unlink. */
271 mds_open_orphan_inc(dentry->d_inode);
273 mfd->mfd_mode = flags;
274 mfd->mfd_dentry = dentry;
275 mfd->mfd_xid = req->rq_xid;
277 spin_lock(&med->med_open_lock);
278 list_add(&mfd->mfd_list, &med->med_open_head);
279 spin_unlock(&med->med_open_lock);
282 body->handle.cookie = mfd->mfd_handle.h_cookie;
288 mds_mfd_destroy(mfd);
290 return ERR_PTR(error);
293 static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
294 struct lov_desc *desc)
297 for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
298 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
299 le64_to_cpu(lmm->lmm_objects[i].l_object_id);
303 /* Must be called with i_sem held */
304 static int mds_create_objects(struct ptlrpc_request *req, int offset,
305 struct mds_update_record *rec,
306 struct mds_obd *mds, struct obd_device *obd,
307 struct dentry *dchild, void **handle,
311 struct obd_trans_info oti = { 0 };
312 struct mds_body *body;
313 struct lov_stripe_md *lsm = NULL;
314 struct lov_mds_md *lmm = NULL;
315 struct inode *inode = dchild->d_inode;
317 int rc, lmm_bufsize, lmm_size;
320 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
321 !(rec->ur_flags & FMODE_WRITE))
324 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
326 if (!S_ISREG(inode->i_mode))
328 if (body->valid & OBD_MD_FLEASIZE)
331 OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
334 oti.oti_objid = *ids;
337 if(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
338 LASSERT (rec->ur_fid2->id);
339 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
340 lmm_size = rec->ur_eadatalen;
341 lmm = rec->ur_eadata;
345 *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
346 if (IS_ERR(*handle)) {
347 rc = PTR_ERR(*handle);
352 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
354 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
355 lmm_bufsize = req->rq_repmsg->buflens[offset];
357 LASSERT(lmm_bufsize >= lmm_size);
358 memcpy(lmm_buf, lmm, lmm_size);
359 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
361 CERROR("open replay failed to set md:%d\n", rc);
365 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
366 GOTO(out_ids, rc = -ENOMEM);
370 GOTO(out_ids, rc = -ENOMEM);
371 oa->o_mode = S_IFREG | 0600;
372 oa->o_id = inode->i_ino;
373 oa->o_generation = inode->i_generation;
374 oa->o_uid = 0; /* must have 0 uid / gid on OST */
376 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
377 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
380 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
383 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
384 /* check if things like lfs setstripe are sending us the ea */
385 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
386 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
388 0, &lsm, rec->ur_eadata);
392 OBD_ALLOC(lmm, mds->mds_max_mdsize);
394 GOTO(out_oa, rc = -ENOMEM);
396 lmm_size = mds->mds_max_mdsize;
397 rc = mds_get_md(obd, dchild->d_parent->d_inode,
400 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
403 OBD_FREE(lmm, mds->mds_max_mdsize);
407 rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
412 CDEBUG(level, "error creating objects for "
413 "inode %lu: rc = %d\n",
416 CERROR("obd_create returned invalid "
423 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
424 0, &lsm, rec->ur_eadata);
428 lsm->lsm_object_id = oa->o_id;
431 oa->o_size = inode->i_size;
432 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
433 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
434 rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti);
436 CERROR("error setting attrs for inode %lu: rc %d\n",
439 CERROR("obd_setattr returned bad rc %d\n", rc);
446 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
447 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
449 LASSERT(lsm && lsm->lsm_object_id);
451 rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm);
452 if (!rec->ur_fid2->id)
453 obd_free_memmd(mds->mds_osc_exp, &lsm);
456 body->eadatasize = rc;
459 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
460 if (IS_ERR(*handle)) {
461 rc = PTR_ERR(*handle);
466 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
467 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
468 lmm_bufsize = req->rq_repmsg->buflens[offset];
470 LASSERT(lmm_bufsize >= lmm_size);
472 memcpy(lmm_buf, lmm, lmm_size);
473 obd_free_diskmd(mds->mds_osc_exp, &lmm);
475 oti_free_cookies(&oti);
479 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
485 static void reconstruct_open(struct mds_update_record *rec, int offset,
486 struct ptlrpc_request *req,
487 struct lustre_handle *child_lockh)
489 struct mds_export_data *med = &req->rq_export->exp_mds_data;
490 struct mds_client_data *mcd = med->med_mcd;
491 struct mds_obd *mds = mds_req2mds(req);
492 struct mds_file_data *mfd;
493 struct obd_device *obd = req->rq_export->exp_obd;
494 struct dentry *parent, *child;
495 struct ldlm_reply *rep;
496 struct mds_body *body;
502 LASSERT(offset == 2); /* only called via intent */
503 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
504 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
506 /* copy rc, transno and disp; steal locks */
507 mds_req_from_mcd(req, mcd);
508 intent_set_disposition(rep, mcd->mcd_last_data);
510 /* Only replay if create or open actually happened. */
511 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
513 return; /* error looking up parent or child */
516 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
517 LASSERT(!IS_ERR(parent));
519 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
520 LASSERT(!IS_ERR(child));
522 if (!child->d_inode) {
523 GOTO(out_dput, 0); /* child not present to open */
526 /* At this point, we know we have a child. We'll send
527 * it back _unless_ it not created and open failed.
529 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
530 !intent_disposition(rep, DISP_OPEN_CREATE) &&
535 /* get lock (write for O_CREAT, read otherwise) */
537 mds_pack_inode2fid(&body->fid1, child->d_inode);
538 mds_pack_inode2body(body, child->d_inode);
539 if (S_ISREG(child->d_inode->i_mode)) {
540 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
544 LASSERT(rc == req->rq_status);
546 /* If we have LOV EA data, the OST holds size, mtime */
547 if (!(body->valid & OBD_MD_FLEASIZE))
548 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
549 OBD_MD_FLATIME | OBD_MD_FLMTIME);
551 /* XXX need to check this case */
554 /* If we're opening a file without an EA, change to a write
555 lock (unless we already have one). */
557 /* If we have -EEXIST as the status, and we were asked to create
558 * exclusively, we can tell we failed because the file already existed.
560 if (req->rq_status == -EEXIST &&
561 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
562 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
566 /* If we didn't get as far as trying to open, then some locking thing
567 * probably went wrong, and we'll just bail here.
569 if (!intent_disposition(rep, DISP_OPEN_OPEN))
572 /* If we failed, then we must have failed opening, so don't look for
573 * file descriptor or anything, just give the client the bad news.
579 list_for_each(t, &med->med_open_head) {
580 mfd = list_entry(t, struct mds_file_data, mfd_list);
581 if (mfd->mfd_xid == req->rq_xid)
586 /* #warning "XXX fixme" bug 2991 */
587 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
588 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
589 * to detect a re-open */
591 mntget(mds->mds_vfsmnt);
592 CERROR("Re-opened file \n");
593 mfd = mds_dentry_open(child, mds->mds_vfsmnt,
594 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
596 CERROR("mds: out of memory\n");
597 GOTO(out_dput, req->rq_status = -ENOMEM);
601 body->handle.cookie = mfd->mfd_handle.h_cookie;
602 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
603 mfd->mfd_handle.h_cookie);
613 /* do NOT or the MAY_*'s, you'll get the weakest */
614 static int accmode(int flags)
618 if (flags & FMODE_READ)
620 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
622 if (flags & FMODE_EXEC)
627 /* Handles object creation, actual opening, and I/O epoch */
628 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
629 struct mds_body *body, int flags, void **handle,
630 struct mds_update_record *rec,struct ldlm_reply *rep)
632 struct mds_obd *mds = mds_req2mds(req);
633 struct obd_device *obd = req->rq_export->exp_obd;
634 struct mds_file_data *mfd = NULL;
635 obd_id *ids = NULL; /* object IDs created */
639 /* atomically create objects if necessary */
640 down(&dchild->d_inode->i_sem);
641 if (S_ISREG(dchild->d_inode->i_mode) &&
642 !(body->valid & OBD_MD_FLEASIZE)) {
643 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
646 up(&dchild->d_inode->i_sem);
651 /* no EA: create objects */
652 rc = mds_create_objects(req, 2, rec, mds, obd,
653 dchild, handle, &ids);
655 CERROR("mds_create_objects: rc = %d\n", rc);
656 up(&dchild->d_inode->i_sem);
660 /* If the inode has EA data, then OSTs hold size, mtime */
661 if (S_ISREG(dchild->d_inode->i_mode) &&
662 !(body->valid & OBD_MD_FLEASIZE)) {
663 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
664 OBD_MD_FLATIME | OBD_MD_FLMTIME);
666 up(&dchild->d_inode->i_sem);
668 intent_set_disposition(rep, DISP_OPEN_OPEN);
669 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
671 RETURN(PTR_ERR(mfd));
673 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
674 mfd->mfd_handle.h_cookie);
676 mds_lov_update_objids(obd, ids);
677 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
680 // mds_mfd_destroy(mfd);
684 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
685 struct mds_body *body, int flags,
686 struct mds_update_record *rec,struct ldlm_reply *rep)
688 struct mds_obd *mds = mds_req2mds(req);
689 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
690 struct dentry *dchild;
691 char fidname[LL_FID_NAMELEN];
696 down(&pending_dir->i_sem);
697 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
698 dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
699 if (IS_ERR(dchild)) {
700 up(&pending_dir->i_sem);
701 rc = PTR_ERR(dchild);
702 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
706 if (dchild->d_inode != NULL) {
707 up(&pending_dir->i_sem);
708 mds_inode_set_orphan(dchild->d_inode);
709 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
710 mds_pack_inode2body(body, dchild->d_inode);
711 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
712 intent_set_disposition(rep, DISP_LOOKUP_POS);
713 CWARN("Orphan %s found and opened in PENDING directory\n",
718 up(&pending_dir->i_sem);
720 /* We didn't find it in PENDING so it isn't an orphan. See
721 * if it was a regular inode that was previously created. */
722 dchild = mds_fid2dentry(mds, fid, NULL);
724 RETURN(PTR_ERR(dchild));
726 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
727 mds_pack_inode2body(body, dchild->d_inode);
728 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
729 intent_set_disposition(rep, DISP_LOOKUP_POS);
732 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
733 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
734 req, rc, rep ? rep->lock_policy_res1 : 0);
735 /* XXX what do we do here if mds_finish_transno itself failed? */
741 int mds_pin(struct ptlrpc_request *req)
743 struct obd_device *obd = req->rq_export->exp_obd;
744 struct mds_body *request_body, *reply_body;
745 struct lvfs_run_ctxt saved;
746 int rc, size = sizeof(*reply_body);
749 request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body));
751 rc = lustre_pack_reply(req, 1, &size, NULL);
754 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
756 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
757 rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
758 request_body->flags, NULL, NULL);
759 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
764 /* Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
765 * If child_lockh is NULL we just get the lock as a barrier to wait for
766 * other holders of this lock, and drop it right away again. */
767 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
768 struct lustre_handle *child_lockh)
770 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
771 struct lustre_handle lockh;
775 if (child_lockh == NULL)
776 child_lockh = &lockh;
778 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
779 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
780 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
781 NULL, 0, NULL, child_lockh);
783 CERROR("ldlm_cli_enqueue: %d\n", rc);
784 else if (child_lockh == &lockh)
785 ldlm_lock_decref(child_lockh, LCK_EX);
790 int mds_open(struct mds_update_record *rec, int offset,
791 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
793 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
794 struct obd_device *obd = req->rq_export->exp_obd;
795 struct mds_obd *mds = mds_req2mds(req);
796 struct ldlm_reply *rep = NULL;
797 struct mds_body *body = NULL;
798 struct dentry *dchild = NULL, *dparent = NULL;
799 struct mds_export_data *med;
800 struct lustre_handle parent_lockh;
801 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
802 int parent_mode = LCK_PR;
804 struct dentry_params dp;
807 if (offset == 2) { /* intent */
808 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
809 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
810 } else if (offset == 0) { /* non-intent reint */
811 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
817 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
819 /* Step 0: If we are passed a fid, then we assume the client already
820 * opened this file and is only replaying the RPC, so we open the
821 * inode by fid (at some large expense in security). */
822 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
823 DEBUG_REQ(D_HA, req, "open replay, disp: "LPX64"\n",
824 rep->lock_policy_res1);
826 LASSERT(rec->ur_fid2->id);
828 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
832 /* We didn't find the correct inode on disk either, so we
833 * need to re-create it via a regular replay. */
834 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
836 LASSERT(!rec->ur_fid2->id);
839 LASSERT(offset == 2); /* If we got here, we must be called via intent */
841 med = &req->rq_export->exp_mds_data;
842 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
843 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
847 acc_mode = accmode(rec->ur_flags);
849 /* Step 1: Find and lock the parent */
850 if (rec->ur_flags & O_CREAT) {
851 /* XXX Well, in fact we only need this lock mode change if
852 in addition to O_CREAT, the file does not exist.
853 But we do not know if it exists or not yet */
854 parent_mode = LCK_PW;
856 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
857 &parent_lockh, rec->ur_name,
859 MDS_INODELOCK_UPDATE);
860 if (IS_ERR(dparent)) {
861 rc = PTR_ERR(dparent);
862 CERROR("parent lookup error %d\n", rc);
865 LASSERT(dparent->d_inode != NULL);
867 cleanup_phase = 1; /* parent dentry and lock */
869 /* Step 2: Lookup the child */
870 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
871 if (IS_ERR(dchild)) {
872 rc = PTR_ERR(dchild);
873 dchild = NULL; /* don't confuse mds_finish_transno() below */
877 cleanup_phase = 2; /* child dentry */
879 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
881 intent_set_disposition(rep, DISP_LOOKUP_POS);
883 intent_set_disposition(rep, DISP_LOOKUP_NEG);
885 /*Step 3: If the child was negative, and we're supposed to, create it.*/
886 if (dchild->d_inode == NULL) {
887 unsigned long ino = rec->ur_fid2->id;
891 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
892 /* It's negative and we weren't supposed to create it */
893 GOTO(cleanup, rc = -ENOENT);
896 intent_set_disposition(rep, DISP_OPEN_CREATE);
897 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
899 if (IS_ERR(handle)) {
900 rc = PTR_ERR(handle);
904 dchild->d_fsdata = (void *) &dp;
908 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
909 if (dchild->d_fsdata == (void *)(unsigned long)ino)
910 dchild->d_fsdata = NULL;
913 CDEBUG(D_INODE, "error during create: %d\n", rc);
916 inode = dchild->d_inode;
918 LASSERT(ino == inode->i_ino);
919 /* Written as part of setattr */
920 inode->i_generation = rec->ur_fid2->generation;
921 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
922 inode->i_ino, inode->i_generation);
926 LTIME_S(iattr.ia_atime) = rec->ur_time;
927 LTIME_S(iattr.ia_ctime) = rec->ur_time;
928 LTIME_S(iattr.ia_mtime) = rec->ur_time;
930 iattr.ia_uid = rec->ur_fsuid;
931 if (dparent->d_inode->i_mode & S_ISGID)
932 iattr.ia_gid = dparent->d_inode->i_gid;
934 iattr.ia_gid = rec->ur_fsgid;
936 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
937 ATTR_MTIME | ATTR_CTIME;
939 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
941 CERROR("error on child setattr: rc = %d\n", rc);
943 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
945 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
947 CERROR("error on parent setattr: rc = %d\n", rc);
949 rc = mds_finish_transno(mds, dchild->d_inode, handle, req, 0,
950 rep ? rep->lock_policy_res1 : 0);
952 acc_mode = 0; /* Don't check for permissions */
955 LASSERT(!mds_inode_is_orphan(dchild->d_inode));
957 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
958 mds_pack_inode2body(body, dchild->d_inode);
960 if (S_ISREG(dchild->d_inode->i_mode)) {
961 /* Check permissions etc */
962 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
966 /* Can't write to a read-only file */
967 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
968 GOTO(cleanup, rc = -EPERM);
970 /* An append-only file must be opened in append mode for
972 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
973 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
974 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
975 GOTO(cleanup, rc = -EPERM);
978 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
979 (rec->ur_flags & MDS_OPEN_EXCL)) {
980 /* File already exists, we didn't just create it, and we
981 * were passed O_EXCL; err-or. */
982 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
985 /* if we are following a symlink, don't open */
986 if (S_ISLNK(dchild->d_inode->i_mode))
987 GOTO(cleanup, rc = 0);
989 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
990 !S_ISDIR(dchild->d_inode->i_mode))
991 GOTO(cleanup, rc = -ENOTDIR);
993 if (S_ISDIR(dchild->d_inode->i_mode)) {
994 if (rec->ur_flags & MDS_OPEN_CREAT ||
995 rec->ur_flags & FMODE_WRITE) {
996 /*we are tryying to create or write a exist dir*/
997 GOTO(cleanup, rc = -EISDIR);
999 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1000 intent_set_disposition(rep, DISP_OPEN_OPEN);
1001 GOTO(cleanup, rc = -EACCES);
1005 /* Step 5: mds_open it */
1006 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1011 if (dchild && dchild->d_inode)
1012 rc = mds_finish_transno(mds, dchild->d_inode, handle, req, rc,
1013 rep ? rep->lock_policy_res1 : 0);
1015 rc = mds_finish_transno(mds, !IS_ERR(dparent) ? dparent->d_inode :
1016 NULL, handle, req, rc,
1017 rep ? rep->lock_policy_res1 : 0);
1019 switch (cleanup_phase) {
1021 if (rc && created) {
1022 int err = vfs_unlink(dparent->d_inode, dchild);
1024 CERROR("unlink(%*s) in error path: %d\n",
1025 dchild->d_name.len, dchild->d_name.name,
1028 } else if (created) {
1029 mds_lock_new_child(obd, dchild->d_inode, NULL);
1033 if (dparent == NULL)
1038 ldlm_lock_decref(&parent_lockh, parent_mode);
1040 ptlrpc_save_lock (req, &parent_lockh, parent_mode);
1043 atomic_inc(&mds->mds_open_count);
1047 /* Close a "file descriptor" and possibly unlink an orphan from the
1048 * PENDING directory.
1050 * If we are being called from mds_disconnect() because the client has
1051 * disappeared, then req == NULL and we do not update last_rcvd because
1052 * there is nothing that could be recovered by the client at this stage
1053 * (it will not even _have_ an entry in last_rcvd anymore).
1055 * Returns EAGAIN if the client needs to get more data and re-close. */
1056 int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
1057 struct mds_file_data *mfd, int unlink_orphan)
1059 struct inode *inode = mfd->mfd_dentry->d_inode;
1060 char fidname[LL_FID_NAMELEN];
1061 int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1062 struct dentry *pending_child = NULL;
1063 struct mds_obd *mds = &obd->u.mds;
1064 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1065 void *handle = NULL;
1066 struct mds_body *request_body = NULL, *reply_body = NULL;
1067 struct dentry_params dp;
1068 struct iattr iattr = { 0 };
1071 if (req && req->rq_reqmsg != NULL)
1072 request_body = lustre_msg_buf(req->rq_reqmsg, 0,
1073 sizeof(*request_body));
1074 if (req && req->rq_repmsg != NULL)
1075 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1076 sizeof(*reply_body));
1078 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1080 last_orphan = mds_open_orphan_dec_test(inode) &&
1081 mds_inode_is_orphan(inode);
1083 /* this is half of the actual "close" */
1084 if (mfd->mfd_mode & FMODE_WRITE) {
1085 rc = mds_put_write_access(mds, inode, request_body,
1086 last_orphan && unlink_orphan);
1087 } else if (mfd->mfd_mode & FMODE_EXEC) {
1088 mds_allow_write_access(inode);
1091 if (last_orphan && unlink_orphan) {
1092 struct lov_mds_md *lmm = NULL;
1093 int stripe_count = 0;
1094 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1096 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1098 /* Sadly, there is no easy way to save pending_child from
1099 * mds_reint_unlink() into mfd, so we need to re-lookup,
1100 * but normally it will still be in the dcache. */
1101 pending_dir = mds->mds_pending_dir->d_inode;
1102 down(&pending_dir->i_sem);
1103 cleanup_phase = 1; /* up(i_sem) when finished */
1104 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1106 if (IS_ERR(pending_child))
1107 GOTO(cleanup, rc = PTR_ERR(pending_child));
1108 LASSERT(pending_child->d_inode != NULL);
1110 cleanup_phase = 2; /* dput(pending_child) when finished */
1111 if (req != NULL && req->rq_repmsg != NULL) {
1112 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1113 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1116 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1117 NULL, stripe_count);
1118 if (IS_ERR(handle)) {
1119 rc = PTR_ERR(handle);
1124 if (req != NULL && req->rq_repmsg != NULL &&
1125 (reply_body->valid & OBD_MD_FLEASIZE) &&
1126 mds_log_op_unlink(obd, pending_child->d_inode, lmm,
1127 req->rq_repmsg->buflens[1],
1128 lustre_msg_buf(req->rq_repmsg, 2, 0),
1129 req->rq_repmsg->buflens[2]) > 0) {
1130 reply_body->valid |= OBD_MD_FLCOOKIE;
1133 pending_child->d_fsdata = (void *) &dp;
1136 if (S_ISDIR(pending_child->d_inode->i_mode))
1137 rc = vfs_rmdir(pending_dir, pending_child);
1139 rc = vfs_unlink(pending_dir, pending_child);
1141 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1143 goto out; /* Don't bother updating attrs on unlinked inode */
1146 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1147 /* Update the on-disk attributes if this was the last write
1148 * close, and all information was provided (i.e., rc == 0)
1150 * XXX this should probably be abstracted with mds_reint_setattr
1154 if (request_body->valid & OBD_MD_FLMTIME &&
1155 request_body->mtime > LTIME_S(inode->i_mtime)) {
1156 LTIME_S(iattr.ia_mtime) = request_body->mtime;
1157 iattr.ia_valid |= ATTR_MTIME;
1159 if (request_body->valid & OBD_MD_FLCTIME &&
1160 request_body->ctime > LTIME_S(inode->i_ctime)) {
1161 LTIME_S(iattr.ia_ctime) = request_body->ctime;
1162 iattr.ia_valid |= ATTR_CTIME;
1165 /* XXX can't set block count with fsfilt_setattr (!) */
1166 if (request_body->valid & OBD_MD_FLSIZE) {
1167 iattr.ia_valid |= ATTR_SIZE;
1168 iattr.ia_size = request_body->size;
1170 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1171 iattr.ia_valid |= ATTR_BLOCKS;
1172 iattr.ia_blocks = request_body->blocks
1177 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1178 /* Only start a transaction to write out only the atime if
1179 * it is more out-of-date than the specified limit. If we
1180 * are already going to write out the atime then do it anyway.
1182 if ((request_body->atime >
1183 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1184 (iattr.ia_valid != 0 &&
1185 request_body->atime > LTIME_S(inode->i_atime))) {
1186 LTIME_S(iattr.ia_atime) = request_body->atime;
1187 iattr.ia_valid |= ATTR_ATIME;
1191 if (iattr.ia_valid != 0) {
1192 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1194 GOTO(cleanup, rc = PTR_ERR(handle));
1195 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1197 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1200 /* If other clients have this file open for write, rc will be > 0 */
1203 l_dput(mfd->mfd_dentry);
1204 mds_mfd_destroy(mfd);
1207 atomic_dec(&mds->mds_open_count);
1208 if (req != NULL && reply_body != NULL) {
1209 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1210 } else if (handle) {
1211 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1213 CERROR("error committing close: %d\n", err);
1219 switch (cleanup_phase) {
1221 dput(pending_child);
1223 up(&pending_dir->i_sem);
1228 int mds_close(struct ptlrpc_request *req)
1230 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1231 struct obd_device *obd = req->rq_export->exp_obd;
1232 struct mds_body *body;
1233 struct mds_file_data *mfd;
1234 struct lvfs_run_ctxt saved;
1235 struct inode *inode;
1236 int rc, repsize[3] = {sizeof(struct mds_body),
1237 obd->u.mds.mds_max_mdsize,
1238 obd->u.mds.mds_max_cookiesize};
1241 rc = lustre_pack_reply(req, 3, repsize, NULL);
1243 CERROR("lustre_pack_reply: rc = %d\n", rc);
1244 req->rq_status = rc;
1246 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1249 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1251 CERROR("Can't unpack body\n");
1252 req->rq_status = -EFAULT;
1256 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1257 /* do some stuff */ ;
1259 mfd = mds_handle2mfd(&body->handle);
1261 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1262 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1263 req->rq_status = -ESTALE;
1267 inode = mfd->mfd_dentry->d_inode;
1268 if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
1269 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1270 LASSERT(body != NULL);
1272 mds_pack_inode2fid(&body->fid1, inode);
1273 mds_pack_inode2body(body, inode);
1274 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
1276 spin_lock(&med->med_open_lock);
1277 list_del(&mfd->mfd_list);
1278 spin_unlock(&med->med_open_lock);
1280 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1281 req->rq_status = mds_mfd_close(req, obd, mfd, 1);
1282 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1285 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1286 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1287 req->rq_status = -ENOMEM;
1294 int mds_done_writing(struct ptlrpc_request *req)
1296 struct mds_body *body;
1297 int rc, size = sizeof(struct mds_body);
1300 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1302 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1304 CERROR("Can't unpack body\n");
1305 req->rq_status = -EFAULT;
1309 rc = lustre_pack_reply(req, 1, &size, NULL);
1311 CERROR("lustre_pack_reply: rc = %d\n", rc);
1312 req->rq_status = rc;