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 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
338 if (IS_ERR(*handle)) {
339 rc = PTR_ERR(*handle);
345 if(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
346 LASSERT (rec->ur_fid2->id);
347 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
348 lmm_size = rec->ur_eadatalen;
349 lmm = rec->ur_eadata;
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);
366 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
367 GOTO(out_ids, rc = -ENOMEM);
371 GOTO(out_ids, rc = -ENOMEM);
372 oa->o_mode = S_IFREG | 0600;
373 oa->o_id = inode->i_ino;
374 oa->o_generation = inode->i_generation;
375 oa->o_uid = 0; /* must have 0 uid / gid on OST */
377 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
378 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
381 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
384 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
385 /* check if things like lstripe/lfs stripe are sending us the ea */
386 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
387 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
389 0, &lsm, rec->ur_eadata);
393 OBD_ALLOC(lmm, mds->mds_max_mdsize);
395 GOTO(out_oa, rc = -ENOMEM);
397 lmm_size = mds->mds_max_mdsize;
398 rc = mds_get_md(obd, dchild->d_parent->d_inode,
401 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
404 OBD_FREE(lmm, mds->mds_max_mdsize);
408 rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
413 CDEBUG(level, "error creating objects for "
414 "inode %lu: rc = %d\n",
417 CERROR("obd_create returned invalid "
424 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
425 0, &lsm, rec->ur_eadata);
429 lsm->lsm_object_id = oa->o_id;
432 oa->o_size = inode->i_size;
433 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
434 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
435 rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti);
437 CERROR("error setting attrs for inode %lu: rc %d\n",
440 CERROR("obd_setattr returned bad rc %d\n", rc);
447 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
448 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
450 LASSERT(lsm && lsm->lsm_object_id);
452 rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm);
453 if (!rec->ur_fid2->id)
454 obd_free_memmd(mds->mds_osc_exp, &lsm);
457 body->eadatasize = rc;
458 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
459 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
460 lmm_bufsize = req->rq_repmsg->buflens[offset];
462 LASSERT(lmm_bufsize >= lmm_size);
464 memcpy(lmm_buf, lmm, lmm_size);
465 obd_free_diskmd(mds->mds_osc_exp, &lmm);
467 oti_free_cookies(&oti);
471 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
477 static void reconstruct_open(struct mds_update_record *rec, int offset,
478 struct ptlrpc_request *req,
479 struct lustre_handle *child_lockh)
481 struct mds_export_data *med = &req->rq_export->exp_mds_data;
482 struct mds_client_data *mcd = med->med_mcd;
483 struct mds_obd *mds = mds_req2mds(req);
484 struct mds_file_data *mfd;
485 struct obd_device *obd = req->rq_export->exp_obd;
486 struct dentry *parent, *child;
487 struct ldlm_reply *rep;
488 struct mds_body *body;
494 LASSERT(offset == 2); /* only called via intent */
495 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
496 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
498 /* copy rc, transno and disp; steal locks */
499 mds_req_from_mcd(req, mcd);
500 intent_set_disposition(rep, mcd->mcd_last_data);
502 /* Only replay if create or open actually happened. */
503 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
505 return; /* error looking up parent or child */
508 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
509 LASSERT(!IS_ERR(parent));
511 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
512 LASSERT(!IS_ERR(child));
514 if (!child->d_inode) {
515 GOTO(out_dput, 0); /* child not present to open */
518 /* At this point, we know we have a child. We'll send
519 * it back _unless_ it not created and open failed.
521 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
522 !intent_disposition(rep, DISP_OPEN_CREATE) &&
527 /* get lock (write for O_CREAT, read otherwise) */
529 mds_pack_inode2fid(&body->fid1, child->d_inode);
530 mds_pack_inode2body(body, child->d_inode);
531 if (S_ISREG(child->d_inode->i_mode)) {
532 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
536 LASSERT(rc == req->rq_status);
538 /* If we have LOV EA data, the OST holds size, mtime */
539 if (!(body->valid & OBD_MD_FLEASIZE))
540 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
541 OBD_MD_FLATIME | OBD_MD_FLMTIME);
543 /* XXX need to check this case */
546 /* If we're opening a file without an EA, change to a write
547 lock (unless we already have one). */
549 /* If we have -EEXIST as the status, and we were asked to create
550 * exclusively, we can tell we failed because the file already existed.
552 if (req->rq_status == -EEXIST &&
553 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
554 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
558 /* If we didn't get as far as trying to open, then some locking thing
559 * probably went wrong, and we'll just bail here.
561 if (!intent_disposition(rep, DISP_OPEN_OPEN))
564 /* If we failed, then we must have failed opening, so don't look for
565 * file descriptor or anything, just give the client the bad news.
571 list_for_each(t, &med->med_open_head) {
572 mfd = list_entry(t, struct mds_file_data, mfd_list);
573 if (mfd->mfd_xid == req->rq_xid)
578 /* #warning "XXX fixme" bug 2991 */
579 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
580 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
581 * to detect a re-open */
583 mntget(mds->mds_vfsmnt);
584 CERROR("Re-opened file \n");
585 mfd = mds_dentry_open(child, mds->mds_vfsmnt,
586 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
588 CERROR("mds: out of memory\n");
589 GOTO(out_dput, req->rq_status = -ENOMEM);
601 /* do NOT or the MAY_*'s, you'll get the weakest */
602 static int accmode(int flags)
606 if (flags & FMODE_READ)
608 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
610 if (flags & FMODE_EXEC)
615 /* Handles object creation, actual opening, and I/O epoch */
616 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
617 struct mds_body *body, int flags, void **handle,
618 struct mds_update_record *rec,struct ldlm_reply *rep)
620 struct mds_obd *mds = mds_req2mds(req);
621 struct obd_device *obd = req->rq_export->exp_obd;
622 struct mds_file_data *mfd = NULL;
623 obd_id *ids = NULL; /* object IDs created */
627 /* atomically create objects if necessary */
628 down(&dchild->d_inode->i_sem);
629 if (S_ISREG(dchild->d_inode->i_mode) &&
630 !(body->valid & OBD_MD_FLEASIZE)) {
631 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
634 up(&dchild->d_inode->i_sem);
639 /* no EA: create objects */
640 rc = mds_create_objects(req, 2, rec, mds, obd,
641 dchild, handle, &ids);
643 CERROR("mds_create_objects: rc = %d\n", rc);
644 up(&dchild->d_inode->i_sem);
648 /* If the inode has EA data, then OSTs hold size, mtime */
649 if (S_ISREG(dchild->d_inode->i_mode) &&
650 !(body->valid & OBD_MD_FLEASIZE)) {
651 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
652 OBD_MD_FLATIME | OBD_MD_FLMTIME);
654 up(&dchild->d_inode->i_sem);
656 intent_set_disposition(rep, DISP_OPEN_OPEN);
657 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
659 RETURN(PTR_ERR(mfd));
661 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
662 mfd->mfd_handle.h_cookie);
664 mds_lov_update_objids(obd, ids);
665 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
668 // mds_mfd_destroy(mfd);
672 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
673 struct mds_body *body, int flags,
674 struct mds_update_record *rec,struct ldlm_reply *rep)
676 struct mds_obd *mds = mds_req2mds(req);
677 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
678 struct dentry *dchild;
679 char fidname[LL_FID_NAMELEN];
684 down(&pending_dir->i_sem);
685 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
686 dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
687 if (IS_ERR(dchild)) {
688 up(&pending_dir->i_sem);
689 rc = PTR_ERR(dchild);
690 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
694 if (dchild->d_inode != NULL) {
695 up(&pending_dir->i_sem);
696 mds_inode_set_orphan(dchild->d_inode);
697 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
698 mds_pack_inode2body(body, dchild->d_inode);
699 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
700 intent_set_disposition(rep, DISP_LOOKUP_POS);
701 CWARN("Orphan %s found and opened in PENDING directory\n",
706 up(&pending_dir->i_sem);
708 /* We didn't find it in PENDING so it isn't an orphan. See
709 * if it was a regular inode that was previously created. */
710 dchild = mds_fid2dentry(mds, fid, NULL);
712 RETURN(PTR_ERR(dchild));
714 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
715 mds_pack_inode2body(body, dchild->d_inode);
716 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
717 intent_set_disposition(rep, DISP_LOOKUP_POS);
720 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
721 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
722 req, rc, rep ? rep->lock_policy_res1 : 0);
723 /* XXX what do we do here if mds_finish_transno itself failed? */
729 int mds_pin(struct ptlrpc_request *req)
731 struct obd_device *obd = req->rq_export->exp_obd;
732 struct mds_body *request_body, *reply_body;
733 struct obd_run_ctxt saved;
734 int rc, size = sizeof(*reply_body);
737 request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body));
739 rc = lustre_pack_reply(req, 1, &size, NULL);
742 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
744 push_ctxt(&saved, &obd->obd_ctxt, NULL);
745 rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
746 request_body->flags, NULL, NULL);
747 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
752 /* Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
753 * If child_lockh is NULL we just get the lock as a barrier to wait for
754 * other holders of this lock, and drop it right away again. */
755 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
756 struct lustre_handle *child_lockh)
758 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
759 struct lustre_handle lockh;
763 if (child_lockh == NULL)
764 child_lockh = &lockh;
766 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
767 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
768 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
769 NULL, 0, NULL, child_lockh);
771 CERROR("ldlm_cli_enqueue: %d\n", rc);
772 else if (child_lockh == &lockh)
773 ldlm_lock_decref(child_lockh, LCK_EX);
778 int mds_open(struct mds_update_record *rec, int offset,
779 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
781 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
782 struct obd_device *obd = req->rq_export->exp_obd;
783 struct mds_obd *mds = mds_req2mds(req);
784 struct ldlm_reply *rep = NULL;
785 struct mds_body *body = NULL;
786 struct dentry *dchild = NULL, *dparent = NULL;
787 struct mds_export_data *med;
788 struct lustre_handle parent_lockh;
789 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
790 int parent_mode = LCK_PR;
792 struct dentry_params dp;
795 if (offset == 2) { /* intent */
796 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
797 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
798 } else if (offset == 0) { /* non-intent reint */
799 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
805 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
807 /* Step 0: If we are passed a fid, then we assume the client already
808 * opened this file and is only replaying the RPC, so we open the
809 * inode by fid (at some large expense in security). */
810 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
811 DEBUG_REQ(D_HA, req, "open replay, disp: "LPX64"\n",
812 rep->lock_policy_res1);
814 LASSERT(rec->ur_fid2->id);
816 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
820 /* We didn't find the correct inode on disk either, so we
821 * need to re-create it via a regular replay. */
822 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
824 LASSERT(!rec->ur_fid2->id);
827 LASSERT(offset == 2); /* If we got here, we must be called via intent */
829 med = &req->rq_export->exp_mds_data;
830 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
831 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
835 acc_mode = accmode(rec->ur_flags);
837 /* Step 1: Find and lock the parent */
838 if (rec->ur_flags & O_CREAT)
839 parent_mode = LCK_PW;
840 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
841 &parent_lockh, rec->ur_name,
842 rec->ur_namelen - 1);
843 if (IS_ERR(dparent)) {
844 rc = PTR_ERR(dparent);
845 CERROR("parent lookup error %d\n", rc);
848 LASSERT(dparent->d_inode != NULL);
850 cleanup_phase = 1; /* parent dentry and lock */
852 /* Step 2: Lookup the child */
853 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
854 if (IS_ERR(dchild)) {
855 rc = PTR_ERR(dchild);
856 dchild = NULL; /* don't confuse mds_finish_transno() below */
860 cleanup_phase = 2; /* child dentry */
862 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
864 intent_set_disposition(rep, DISP_LOOKUP_POS);
866 intent_set_disposition(rep, DISP_LOOKUP_NEG);
868 /*Step 3: If the child was negative, and we're supposed to, create it.*/
869 if (dchild->d_inode == NULL) {
870 unsigned long ino = rec->ur_fid2->id;
874 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
875 /* It's negative and we weren't supposed to create it */
876 GOTO(cleanup, rc = -ENOENT);
879 intent_set_disposition(rep, DISP_OPEN_CREATE);
880 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
882 if (IS_ERR(handle)) {
883 rc = PTR_ERR(handle);
887 dchild->d_fsdata = (void *) &dp;
891 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
892 if (dchild->d_fsdata == (void *)(unsigned long)ino)
893 dchild->d_fsdata = NULL;
896 CDEBUG(D_INODE, "error during create: %d\n", rc);
899 inode = dchild->d_inode;
901 LASSERT(ino == inode->i_ino);
902 /* Written as part of setattr */
903 inode->i_generation = rec->ur_fid2->generation;
904 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
905 inode->i_ino, inode->i_generation);
909 LTIME_S(iattr.ia_atime) = rec->ur_time;
910 LTIME_S(iattr.ia_ctime) = rec->ur_time;
911 LTIME_S(iattr.ia_mtime) = rec->ur_time;
913 iattr.ia_uid = rec->ur_fsuid;
914 if (dparent->d_inode->i_mode & S_ISGID)
915 iattr.ia_gid = dparent->d_inode->i_gid;
917 iattr.ia_gid = rec->ur_fsgid;
919 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
920 ATTR_MTIME | ATTR_CTIME;
922 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
924 CERROR("error on child setattr: rc = %d\n", rc);
926 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
928 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
930 CERROR("error on parent setattr: rc = %d\n", rc);
932 acc_mode = 0; /* Don't check for permissions */
935 LASSERT(!mds_inode_is_orphan(dchild->d_inode));
937 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
938 mds_pack_inode2body(body, dchild->d_inode);
940 if (S_ISREG(dchild->d_inode->i_mode)) {
941 /* Check permissions etc */
942 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
946 /* Can't write to a read-only file */
947 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
948 GOTO(cleanup, rc = -EPERM);
950 /* An append-only file must be opened in append mode for
952 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
953 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
954 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
955 GOTO(cleanup, rc = -EPERM);
958 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
959 (rec->ur_flags & MDS_OPEN_EXCL)) {
960 /* File already exists, we didn't just create it, and we
961 * were passed O_EXCL; err-or. */
962 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
965 /* if we are following a symlink, don't open */
966 if (S_ISLNK(dchild->d_inode->i_mode))
967 GOTO(cleanup, rc = 0);
969 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
970 !S_ISDIR(dchild->d_inode->i_mode))
971 GOTO(cleanup, rc = -ENOTDIR);
973 if (S_ISDIR(dchild->d_inode->i_mode)) {
974 if (rec->ur_flags & MDS_OPEN_CREAT || rec->ur_flags & FMODE_WRITE) {
975 /*we are tryying to create or write a exist dir*/
976 GOTO(cleanup, rc = -EISDIR);
978 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
979 GOTO(cleanup, rc = -EACCES);
983 /* Step 5: mds_open it */
984 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
989 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
990 req, rc, rep ? rep->lock_policy_res1 : 0);
992 switch (cleanup_phase) {
995 int err = vfs_unlink(dparent->d_inode, dchild);
997 CERROR("unlink(%*s) in error path: %d\n",
998 dchild->d_name.len, dchild->d_name.name,
1001 } else if (created) {
1002 mds_lock_new_child(obd, dchild->d_inode, NULL);
1006 if (dparent == NULL)
1011 ldlm_lock_decref(&parent_lockh, parent_mode);
1013 ptlrpc_save_lock (req, &parent_lockh, parent_mode);
1016 atomic_inc(&mds->mds_open_count);
1020 /* Close a "file descriptor" and possibly unlink an orphan from the
1021 * PENDING directory.
1023 * If we are being called from mds_disconnect() because the client has
1024 * disappeared, then req == NULL and we do not update last_rcvd because
1025 * there is nothing that could be recovered by the client at this stage
1026 * (it will not even _have_ an entry in last_rcvd anymore).
1028 * Returns EAGAIN if the client needs to get more data and re-close. */
1029 int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
1030 struct mds_file_data *mfd, int unlink_orphan)
1032 struct inode *inode = mfd->mfd_dentry->d_inode;
1033 char fidname[LL_FID_NAMELEN];
1034 int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1035 struct dentry *pending_child = NULL;
1036 struct mds_obd *mds = &obd->u.mds;
1037 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1038 void *handle = NULL;
1039 struct mds_body *request_body = NULL, *reply_body = NULL;
1040 struct dentry_params dp;
1044 request_body = lustre_msg_buf(req->rq_reqmsg, 0,
1045 sizeof(*request_body));
1046 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1047 sizeof(*reply_body));
1050 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1052 last_orphan = mds_open_orphan_dec_test(inode) &&
1053 mds_inode_is_orphan(inode);
1055 /* this is half of the actual "close" */
1056 if (mfd->mfd_mode & FMODE_WRITE) {
1057 rc = mds_put_write_access(mds, inode, request_body,
1058 last_orphan && unlink_orphan);
1059 } else if (mfd->mfd_mode & FMODE_EXEC) {
1060 mds_allow_write_access(inode);
1063 if (last_orphan && unlink_orphan) {
1064 int stripe_count = 0;
1065 struct lov_mds_md *lmm = NULL;
1066 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1068 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1070 /* Sadly, there is no easy way to save pending_child from
1071 * mds_reint_unlink() into mfd, so we need to re-lookup,
1072 * but normally it will still be in the dcache. */
1073 pending_dir = mds->mds_pending_dir->d_inode;
1074 down(&pending_dir->i_sem);
1075 cleanup_phase = 1; /* up(i_sem) when finished */
1076 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1078 if (IS_ERR(pending_child))
1079 GOTO(cleanup, rc = PTR_ERR(pending_child));
1080 LASSERT(pending_child->d_inode != NULL);
1082 cleanup_phase = 2; /* dput(pending_child) when finished */
1084 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1085 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1088 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1089 NULL, stripe_count);
1090 if (IS_ERR(handle)) {
1091 rc = PTR_ERR(handle);
1096 if (req != NULL && (reply_body->valid & OBD_MD_FLEASIZE) &&
1097 mds_log_op_unlink(obd, pending_child->d_inode, lmm,
1098 req->rq_repmsg->buflens[1],
1099 lustre_msg_buf(req->rq_repmsg, 2, 0),
1100 req->rq_repmsg->buflens[2]) > 0) {
1101 reply_body->valid |= OBD_MD_FLCOOKIE;
1104 pending_child->d_fsdata = (void *) &dp;
1107 if (S_ISDIR(pending_child->d_inode->i_mode))
1108 rc = vfs_rmdir(pending_dir, pending_child);
1110 rc = vfs_unlink(pending_dir, pending_child);
1112 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1113 } else if (mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1114 /* Update the on-disk attributes if this was the last write
1115 * close, and all information was provided (i.e., rc == 0)
1117 * XXX this should probably be abstracted with mds_reint_setattr
1122 /* XXX can't set block count with fsfilt_setattr (!) */
1123 iattr.ia_valid = ATTR_CTIME | ATTR_ATIME |
1124 ATTR_MTIME | ATTR_SIZE;
1125 iattr.ia_atime = request_body->atime;
1126 iattr.ia_ctime = request_body->ctime;
1127 iattr.ia_mtime = request_body->mtime;
1128 iattr.ia_size = request_body->size;
1129 /* iattr.ia_blocks = request_body->blocks */
1131 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1133 GOTO(cleanup, rc = PTR_ERR(handle));
1134 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1136 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1139 /* If other clients have this file open for write, rc will be > 0 */
1142 l_dput(mfd->mfd_dentry);
1143 mds_mfd_destroy(mfd);
1146 atomic_dec(&mds->mds_open_count);
1148 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1149 } else if (handle) {
1150 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1152 CERROR("error committing close: %d\n", err);
1158 switch (cleanup_phase) {
1160 dput(pending_child);
1162 up(&pending_dir->i_sem);
1167 int mds_close(struct ptlrpc_request *req)
1169 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1170 struct obd_device *obd = req->rq_export->exp_obd;
1171 struct mds_body *body;
1172 struct mds_file_data *mfd;
1173 struct obd_run_ctxt saved;
1174 struct inode *inode;
1175 int rc, repsize[3] = {sizeof(struct mds_body),
1176 obd->u.mds.mds_max_mdsize,
1177 obd->u.mds.mds_max_cookiesize};
1180 rc = lustre_pack_reply(req, 3, repsize, NULL);
1182 CERROR("lustre_pack_reply: rc = %d\n", rc);
1183 req->rq_status = rc;
1185 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1188 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1190 CERROR("Can't unpack body\n");
1191 req->rq_status = -EFAULT;
1195 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1196 /* do some stuff */ ;
1198 mfd = mds_handle2mfd(&body->handle);
1200 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1201 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1202 req->rq_status = -ESTALE;
1206 inode = mfd->mfd_dentry->d_inode;
1207 if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
1208 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1209 LASSERT(body != NULL);
1211 mds_pack_inode2fid(&body->fid1, inode);
1212 mds_pack_inode2body(body, inode);
1213 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
1215 spin_lock(&med->med_open_lock);
1216 list_del(&mfd->mfd_list);
1217 spin_unlock(&med->med_open_lock);
1219 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1220 req->rq_status = mds_mfd_close(rc ? NULL : req, obd, mfd, 1);
1221 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1223 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1224 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1225 req->rq_status = -ENOMEM;
1234 int mds_done_writing(struct ptlrpc_request *req)
1236 struct mds_body *body;
1237 int rc, size = sizeof(struct mds_body);
1240 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1242 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1244 CERROR("Can't unpack body\n");
1245 req->rq_status = -EFAULT;
1249 rc = lustre_pack_reply(req, 1, &size, NULL);
1251 CERROR("lustre_pack_reply: rc = %d\n", rc);
1252 req->rq_status = rc;