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 inode *inode, void **handle, obd_id **ids)
310 struct obd_trans_info oti = { 0 };
311 struct mds_body *body;
312 struct lov_stripe_md *lsm = NULL;
313 struct lov_mds_md *lmm = NULL;
315 int rc, lmm_bufsize, lmm_size;
318 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
319 !(rec->ur_flags & FMODE_WRITE))
322 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
324 if (!S_ISREG(inode->i_mode))
326 if (body->valid & OBD_MD_FLEASIZE)
329 OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
332 oti.oti_objid = *ids;
335 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
336 if (IS_ERR(*handle)) {
337 rc = PTR_ERR(*handle);
343 if(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
344 LASSERT (rec->ur_fid2->id);
345 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
346 lmm_size = rec->ur_eadatalen;
347 lmm = rec->ur_eadata;
350 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
352 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
353 lmm_bufsize = req->rq_repmsg->buflens[offset];
355 LASSERT(lmm_bufsize >= lmm_size);
356 memcpy(lmm_buf, lmm, lmm_size);
357 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
359 CERROR("open replay failed to set md:%d\n", rc);
364 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
365 GOTO(out_ids, rc = -ENOMEM);
369 GOTO(out_ids, rc = -ENOMEM);
370 oa->o_mode = S_IFREG | 0600;
371 oa->o_id = inode->i_ino;
372 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
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 | OBD_MD_FLGROUP;
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 lstripe/lfs stripe 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 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
393 rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
398 CDEBUG(level, "error creating objects for "
399 "inode %lu: rc = %d\n",
402 CERROR("obd_create returned invalid "
409 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
410 0, &lsm, rec->ur_eadata);
414 lsm->lsm_object_id = oa->o_id;
415 lsm->lsm_object_gr = oa->o_gr;
418 oa->o_size = inode->i_size;
419 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
420 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
421 rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti);
423 CERROR("error setting attrs for inode %lu: rc %d\n",
426 CERROR("obd_setattr returned bad rc %d\n", rc);
433 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
434 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
436 LASSERT(lsm && lsm->lsm_object_id);
438 rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm);
439 if (!rec->ur_fid2->id)
440 obd_free_memmd(mds->mds_osc_exp, &lsm);
443 body->eadatasize = rc;
444 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
445 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
446 lmm_bufsize = req->rq_repmsg->buflens[offset];
448 LASSERT(lmm_bufsize >= lmm_size);
450 memcpy(lmm_buf, lmm, lmm_size);
451 obd_free_diskmd(mds->mds_osc_exp, &lmm);
453 oti_free_cookies(&oti);
457 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
463 static void reconstruct_open(struct mds_update_record *rec, int offset,
464 struct ptlrpc_request *req,
465 struct lustre_handle *child_lockh)
467 struct mds_export_data *med = &req->rq_export->exp_mds_data;
468 struct mds_client_data *mcd = med->med_mcd;
469 struct mds_obd *mds = mds_req2mds(req);
470 struct mds_file_data *mfd;
471 struct obd_device *obd = req->rq_export->exp_obd;
472 struct dentry *parent, *child;
473 struct ldlm_reply *rep;
474 struct mds_body *body;
480 LASSERT(offset == 2); /* only called via intent */
481 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
482 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
484 /* copy rc, transno and disp; steal locks */
485 mds_req_from_mcd(req, mcd);
486 intent_set_disposition(rep, mcd->mcd_last_data);
488 /* Only replay if create or open actually happened. */
489 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
491 return; /* error looking up parent or child */
494 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
495 LASSERT(!IS_ERR(parent));
497 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
498 LASSERT(!IS_ERR(child));
500 if (!child->d_inode) {
501 GOTO(out_dput, 0); /* child not present to open */
504 /* At this point, we know we have a child. We'll send
505 * it back _unless_ it not created and open failed.
507 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
508 !intent_disposition(rep, DISP_OPEN_CREATE) &&
513 /* get lock (write for O_CREAT, read otherwise) */
515 mds_pack_inode2fid(obd, &body->fid1, child->d_inode);
516 mds_pack_inode2body(obd, body, child->d_inode);
517 if (S_ISREG(child->d_inode->i_mode)) {
518 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
522 LASSERT(rc == req->rq_status);
524 /* If we have LOV EA data, the OST holds size, mtime */
525 if (!(body->valid & OBD_MD_FLEASIZE))
526 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
527 OBD_MD_FLATIME | OBD_MD_FLMTIME);
529 /* XXX need to check this case */
532 /* If we're opening a file without an EA, change to a write
533 lock (unless we already have one). */
535 /* If we have -EEXIST as the status, and we were asked to create
536 * exclusively, we can tell we failed because the file already existed.
538 if (req->rq_status == -EEXIST &&
539 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
540 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
544 /* If we didn't get as far as trying to open, then some locking thing
545 * probably went wrong, and we'll just bail here.
547 if (!intent_disposition(rep, DISP_OPEN_OPEN))
550 /* If we failed, then we must have failed opening, so don't look for
551 * file descriptor or anything, just give the client the bad news.
557 list_for_each(t, &med->med_open_head) {
558 mfd = list_entry(t, struct mds_file_data, mfd_list);
559 if (mfd->mfd_xid == req->rq_xid)
565 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
566 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
567 * to detect a re-open */
569 mntget(mds->mds_vfsmnt);
570 CERROR("Re-opened file \n");
571 mfd = mds_dentry_open(child, mds->mds_vfsmnt,
572 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
574 CERROR("mds: out of memory\n");
575 GOTO(out_dput, req->rq_status = -ENOMEM);
587 /* do NOT or the MAY_*'s, you'll get the weakest */
588 static int accmode(int flags)
592 if (flags & FMODE_READ)
594 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
596 if (flags & FMODE_EXEC)
601 /* Handles object creation, actual opening, and I/O epoch */
602 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
603 struct mds_body *body, int flags, void **handle,
604 struct mds_update_record *rec,struct ldlm_reply *rep)
606 struct mds_obd *mds = mds_req2mds(req);
607 struct obd_device *obd = req->rq_export->exp_obd;
608 struct mds_file_data *mfd = NULL;
609 obd_id *ids = NULL; /* object IDs created */
614 /* atomically create objects if necessary */
615 down(&dchild->d_inode->i_sem);
616 mode = dchild->d_inode->i_mode;
617 if ((S_ISREG(mode) || S_ISDIR(mode)) &&
618 !(body->valid & OBD_MD_FLEASIZE)) {
619 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
622 up(&dchild->d_inode->i_sem);
627 /* no EA: create objects */
628 rc = mds_create_objects(req, 2, rec, mds, obd,
629 dchild->d_inode, handle, &ids);
631 CERROR("mds_create_objects: rc = %d\n", rc);
632 up(&dchild->d_inode->i_sem);
636 /* If the inode has EA data, then OSTs hold size, mtime */
637 if (S_ISREG(dchild->d_inode->i_mode) &&
638 !(body->valid & OBD_MD_FLEASIZE)) {
639 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
640 OBD_MD_FLATIME | OBD_MD_FLMTIME);
642 up(&dchild->d_inode->i_sem);
644 intent_set_disposition(rep, DISP_OPEN_OPEN);
645 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
647 RETURN(PTR_ERR(mfd));
649 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
650 mfd->mfd_handle.h_cookie);
652 mds_lov_update_objids(obd, ids);
653 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
656 // mds_mfd_destroy(mfd);
660 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
661 struct mds_body *body, int flags,
662 struct mds_update_record *rec,struct ldlm_reply *rep)
664 struct mds_obd *mds = mds_req2mds(req);
665 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
666 struct dentry *dchild;
667 char fidname[LL_FID_NAMELEN];
672 down(&pending_dir->i_sem);
673 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
674 dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
675 if (IS_ERR(dchild)) {
676 up(&pending_dir->i_sem);
677 rc = PTR_ERR(dchild);
678 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
682 if (dchild->d_inode != NULL) {
683 up(&pending_dir->i_sem);
684 mds_inode_set_orphan(dchild->d_inode);
685 mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
686 mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
687 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
688 intent_set_disposition(rep, DISP_LOOKUP_POS);
689 CWARN("Orphan %s found and opened in PENDING directory\n",
694 up(&pending_dir->i_sem);
696 /* We didn't find it in PENDING so it isn't an orphan. See
697 * if it was a regular inode that was previously created. */
698 dchild = mds_fid2dentry(mds, fid, NULL);
700 RETURN(PTR_ERR(dchild));
702 mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
703 mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
704 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
705 intent_set_disposition(rep, DISP_LOOKUP_POS);
708 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
709 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
710 req, rc, rep ? rep->lock_policy_res1 : 0);
711 /* XXX what do we do here if mds_finish_transno itself failed? */
717 int mds_pin(struct ptlrpc_request *req)
719 struct obd_device *obd = req->rq_export->exp_obd;
720 struct mds_body *request_body, *reply_body;
721 struct obd_run_ctxt saved;
722 int rc, size = sizeof(*reply_body);
725 request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body));
727 rc = lustre_pack_reply(req, 1, &size, NULL);
730 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
732 push_ctxt(&saved, &obd->obd_ctxt, NULL);
733 rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
734 request_body->flags, NULL, NULL);
735 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
740 /* Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
741 * If child_lockh is NULL we just get the lock as a barrier to wait for
742 * other holders of this lock, and drop it right away again. */
743 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
744 struct lustre_handle *child_lockh)
746 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
747 struct lustre_handle lockh;
751 if (child_lockh == NULL)
752 child_lockh = &lockh;
754 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
755 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
756 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
757 NULL, 0, NULL, child_lockh);
759 CERROR("ldlm_cli_enqueue: %d\n", rc);
760 else if (child_lockh == &lockh)
761 ldlm_lock_decref(child_lockh, LCK_EX);
766 int mds_open(struct mds_update_record *rec, int offset,
767 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
769 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
770 struct obd_device *obd = req->rq_export->exp_obd;
771 struct mds_obd *mds = mds_req2mds(req);
772 struct ldlm_reply *rep = NULL;
773 struct mds_body *body = NULL;
774 struct dentry *dchild = NULL, *dparent = NULL;
775 struct mds_export_data *med;
776 struct lustre_handle parent_lockh[2];
777 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
778 int parent_mode = LCK_PR;
780 struct dentry_params dp;
781 struct mea *mea = NULL;
785 parent_lockh[0].cookie = 0;
786 parent_lockh[1].cookie = 0;
788 if (offset == 2) { /* intent */
789 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
790 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
791 } else if (offset == 0) { /* non-intent reint */
792 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
798 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
800 /* Step 0: If we are passed a fid, then we assume the client already
801 * opened this file and is only replaying the RPC, so we open the
802 * inode by fid (at some large expense in security). */
803 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
804 DEBUG_REQ(D_HA, req, "open replay, disp: "LPX64"\n",
805 rep->lock_policy_res1);
807 LASSERT(rec->ur_fid2->id);
809 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
813 /* We didn't find the correct inode on disk either, so we
814 * need to re-create it via a regular replay. */
815 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
817 LASSERT(!rec->ur_fid2->id);
820 LASSERT(offset == 2); /* If we got here, we must be called via intent */
822 med = &req->rq_export->exp_mds_data;
823 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
824 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
828 acc_mode = accmode(rec->ur_flags);
830 /* Step 1: Find and lock the parent */
831 if (rec->ur_flags & O_CREAT) {
832 /* XXX Well, in fact we only need this lock mode change if
833 in addition to O_CREAT, the file does not exist.
834 But we do not know if it exists or not yet */
835 parent_mode = LCK_PW;
838 if (rec->ur_namelen == 1) {
839 /* client (LMV) wants to open the file by fid */
840 CDEBUG(D_OTHER, "OPEN by fid %u/%u/%u\n",
841 (unsigned) rec->ur_fid1->mds,
842 (unsigned) rec->ur_fid1->id,
843 (unsigned) rec->ur_fid1->generation);
844 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
845 if (IS_ERR(dchild)) {
846 rc = PTR_ERR(dparent);
847 CERROR("child lookup by a fid error %d\n", rc);
853 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
854 parent_lockh, rec->ur_name,
856 MDS_INODELOCK_UPDATE);
857 if (IS_ERR(dparent)) {
858 rc = PTR_ERR(dparent);
859 CERROR("parent lookup error %d\n", rc);
862 LASSERT(dparent->d_inode != NULL);
864 cleanup_phase = 1; /* parent dentry and lock */
866 /* try to retrieve MEA data for this dir */
867 rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
870 /* dir is already splitted, check is requested filename
871 * should live at this MDS or at another one */
873 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
874 if (mea->mea_master != i) {
875 CERROR("inapropriate MDS(%d) for %s. should be %d\n",
876 mea->mea_master, rec->ur_name, i);
877 GOTO(cleanup, rc = -ESTALE);
881 /* Step 2: Lookup the child */
882 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
883 if (IS_ERR(dchild)) {
884 rc = PTR_ERR(dchild);
885 dchild = NULL; /* don't confuse mds_finish_transno() below */
890 cleanup_phase = 2; /* child dentry */
892 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
894 if (dchild->d_flags & DCACHE_CROSS_REF) {
895 struct ldlm_res_id res_id = { . name = {0} };
896 ldlm_policy_data_t policy;
898 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
899 (unsigned long) dchild->d_mdsnum,
900 (unsigned long) dchild->d_inum,
901 (unsigned long) dchild->d_generation);
902 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
903 body->fid1.id = dchild->d_inum;
904 body->fid1.mds = dchild->d_mdsnum;
905 body->fid1.generation = dchild->d_generation;
906 intent_set_disposition(rep, DISP_LOOKUP_POS);
907 res_id.name[0] = dchild->d_inum;
908 res_id.name[1] = dchild->d_generation;
909 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
910 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
911 res_id, LDLM_IBITS, &policy,
914 ldlm_completion_ast, NULL, NULL,
915 NULL, 0, NULL, child_lockh);
917 if (parent_lockh[1].cookie != 0)
918 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
920 ldlm_lock_decref(parent_lockh, parent_mode);
922 OBD_FREE(mea, mea_size);
929 intent_set_disposition(rep, DISP_LOOKUP_POS);
931 intent_set_disposition(rep, DISP_LOOKUP_NEG);
933 /*Step 3: If the child was negative, and we're supposed to, create it.*/
934 if (dchild->d_inode == NULL) {
935 unsigned long ino = rec->ur_fid2->id;
939 if ((rc = mds_try_to_split_dir(obd, dparent, &mea, 0))) {
941 /* dir got splitted */
942 GOTO(cleanup, rc = -ESTALE);
944 /* error happened during spitting */
949 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
950 /* It's negative and we weren't supposed to create it */
951 GOTO(cleanup, rc = -ENOENT);
954 intent_set_disposition(rep, DISP_OPEN_CREATE);
955 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
957 if (IS_ERR(handle)) {
958 rc = PTR_ERR(handle);
962 dchild->d_fsdata = (void *) &dp;
966 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
967 if (dchild->d_fsdata == (void *)(unsigned long)ino)
968 dchild->d_fsdata = NULL;
971 CDEBUG(D_INODE, "error during create: %d\n", rc);
974 inode = dchild->d_inode;
976 LASSERT(ino == inode->i_ino);
977 /* Written as part of setattr */
978 inode->i_generation = rec->ur_fid2->generation;
979 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
980 inode->i_ino, inode->i_generation);
984 LTIME_S(iattr.ia_atime) = rec->ur_time;
985 LTIME_S(iattr.ia_ctime) = rec->ur_time;
986 LTIME_S(iattr.ia_mtime) = rec->ur_time;
988 iattr.ia_uid = rec->ur_fsuid;
989 if (dparent->d_inode->i_mode & S_ISGID)
990 iattr.ia_gid = dparent->d_inode->i_gid;
992 iattr.ia_gid = rec->ur_fsgid;
994 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
995 ATTR_MTIME | ATTR_CTIME;
997 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
999 CERROR("error on child setattr: rc = %d\n", rc);
1001 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1003 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1005 CERROR("error on parent setattr: rc = %d\n", rc);
1007 acc_mode = 0; /* Don't check for permissions */
1010 LASSERT(!mds_inode_is_orphan(dchild->d_inode));
1012 mds_pack_inode2fid(obd, &body->fid1, dchild->d_inode);
1013 mds_pack_inode2body(obd, body, dchild->d_inode);
1015 if (S_ISREG(dchild->d_inode->i_mode)) {
1016 /* Check permissions etc */
1017 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1021 /* Can't write to a read-only file */
1022 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1023 GOTO(cleanup, rc = -EPERM);
1025 /* An append-only file must be opened in append mode for
1027 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1028 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1029 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1030 GOTO(cleanup, rc = -EPERM);
1033 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1034 (rec->ur_flags & MDS_OPEN_EXCL)) {
1035 /* File already exists, we didn't just create it, and we
1036 * were passed O_EXCL; err-or. */
1037 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1040 /* if we are following a symlink, don't open */
1041 if (S_ISLNK(dchild->d_inode->i_mode))
1042 GOTO(cleanup, rc = 0);
1044 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1045 !S_ISDIR(dchild->d_inode->i_mode))
1046 GOTO(cleanup, rc = -ENOTDIR);
1048 if (S_ISDIR(dchild->d_inode->i_mode)) {
1049 if (rec->ur_flags & MDS_OPEN_CREAT || rec->ur_flags & FMODE_WRITE) {
1050 /*we are tryying to create or write a exist dir*/
1051 GOTO(cleanup, rc = -EISDIR);
1053 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1054 GOTO(cleanup, rc = -EACCES);
1059 struct ldlm_res_id res_id = { . name = {0} };
1060 ldlm_policy_data_t policy;
1062 res_id.name[0] = dchild->d_inode->i_ino;
1063 res_id.name[1] = dchild->d_inode->i_generation;
1064 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1065 MDS_INODELOCK_UPDATE;
1066 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1067 res_id, LDLM_IBITS, &policy,
1070 ldlm_completion_ast, NULL, NULL,
1071 NULL, 0, NULL, child_lockh);
1076 /* Step 5: mds_open it */
1077 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1082 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1083 req, rc, rep ? rep->lock_policy_res1 : 0);
1086 switch (cleanup_phase) {
1089 ldlm_lock_decref(child_lockh, LCK_PR);
1091 if (rc && created) {
1092 int err = vfs_unlink(dparent->d_inode, dchild);
1094 CERROR("unlink(%*s) in error path: %d\n",
1095 dchild->d_name.len, dchild->d_name.name,
1098 } else if (created) {
1099 mds_lock_new_child(obd, dchild->d_inode, NULL);
1103 if (dparent == NULL)
1108 if (parent_lockh[1].cookie != 0)
1109 ldlm_lock_decref(parent_lockh + 1, LCK_CW);
1112 ldlm_lock_decref(parent_lockh, parent_mode);
1114 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1117 atomic_inc(&mds->mds_open_count);
1119 OBD_FREE(mea, mea_size);
1120 if ((cleanup_phase != 3) && !rc)
1125 /* Close a "file descriptor" and possibly unlink an orphan from the
1126 * PENDING directory.
1128 * If we are being called from mds_disconnect() because the client has
1129 * disappeared, then req == NULL and we do not update last_rcvd because
1130 * there is nothing that could be recovered by the client at this stage
1131 * (it will not even _have_ an entry in last_rcvd anymore).
1133 * Returns EAGAIN if the client needs to get more data and re-close. */
1134 int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
1135 struct mds_file_data *mfd, int unlink_orphan)
1137 struct inode *inode = mfd->mfd_dentry->d_inode;
1138 char fidname[LL_FID_NAMELEN];
1139 int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1140 struct dentry *pending_child = NULL;
1141 struct mds_obd *mds = &obd->u.mds;
1142 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1143 void *handle = NULL;
1144 struct mds_body *request_body = NULL, *reply_body = NULL;
1145 struct dentry_params dp;
1146 struct lov_mds_md *lmm;
1150 request_body = lustre_msg_buf(req->rq_reqmsg, 0,
1151 sizeof(*request_body));
1152 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1153 sizeof(*reply_body));
1156 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1158 last_orphan = mds_open_orphan_dec_test(inode) &&
1159 mds_inode_is_orphan(inode);
1161 /* this is half of the actual "close" */
1162 if (mfd->mfd_mode & FMODE_WRITE) {
1163 rc = mds_put_write_access(mds, inode, request_body,
1164 last_orphan && unlink_orphan);
1165 } else if (mfd->mfd_mode & FMODE_EXEC) {
1166 mds_allow_write_access(inode);
1169 if (last_orphan && unlink_orphan) {
1170 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1172 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1174 /* Sadly, there is no easy way to save pending_child from
1175 * mds_reint_unlink() into mfd, so we need to re-lookup,
1176 * but normally it will still be in the dcache. */
1177 pending_dir = mds->mds_pending_dir->d_inode;
1178 down(&pending_dir->i_sem);
1179 cleanup_phase = 1; /* up(i_sem) when finished */
1180 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1182 if (IS_ERR(pending_child))
1183 GOTO(cleanup, rc = PTR_ERR(pending_child));
1184 LASSERT(pending_child->d_inode != NULL);
1186 cleanup_phase = 2; /* dput(pending_child) when finished */
1187 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1188 handle = fsfilt_start_log(obd, pending_dir,
1189 FSFILT_OP_UNLINK, NULL,
1190 le32_to_cpu(lmm->lmm_stripe_count));
1191 if (IS_ERR(handle)) {
1192 rc = PTR_ERR(handle);
1198 (reply_body->valid & OBD_MD_FLEASIZE) &&
1199 mds_log_op_unlink(obd, pending_child->d_inode,
1200 lustre_msg_buf(req->rq_repmsg, 1, 0),
1201 req->rq_repmsg->buflens[1],
1202 lustre_msg_buf(req->rq_repmsg, 2, 0),
1203 req->rq_repmsg->buflens[2]) > 0) {
1204 reply_body->valid |= OBD_MD_FLCOOKIE;
1207 pending_child->d_fsdata = (void *) &dp;
1210 if (S_ISDIR(pending_child->d_inode->i_mode))
1211 rc = vfs_rmdir(pending_dir, pending_child);
1213 rc = vfs_unlink(pending_dir, pending_child);
1215 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1216 } else if (mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1217 /* Update the on-disk attributes if this was the last write
1218 * close, and all information was provided (i.e., rc == 0)
1220 * XXX this should probably be abstracted with mds_reint_setattr
1225 /* XXX can't set block count with fsfilt_setattr (!) */
1226 iattr.ia_valid = ATTR_CTIME | ATTR_ATIME |
1227 ATTR_MTIME | ATTR_SIZE;
1228 iattr.ia_atime = request_body->atime;
1229 iattr.ia_ctime = request_body->ctime;
1230 iattr.ia_mtime = request_body->mtime;
1231 iattr.ia_size = request_body->size;
1232 /* iattr.ia_blocks = request_body->blocks */
1234 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1236 GOTO(cleanup, rc = PTR_ERR(handle));
1237 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1239 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1242 /* If other clients have this file open for write, rc will be > 0 */
1245 l_dput(mfd->mfd_dentry);
1246 mds_mfd_destroy(mfd);
1249 atomic_dec(&mds->mds_open_count);
1251 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1252 } else if (handle) {
1253 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1255 CERROR("error committing close: %d\n", err);
1261 switch (cleanup_phase) {
1263 dput(pending_child);
1265 up(&pending_dir->i_sem);
1270 int mds_close(struct ptlrpc_request *req)
1272 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1273 struct obd_device *obd = req->rq_export->exp_obd;
1274 struct mds_body *body;
1275 struct mds_file_data *mfd;
1276 struct obd_run_ctxt saved;
1277 struct inode *inode;
1278 int rc, repsize[3] = {sizeof(struct mds_body),
1279 obd->u.mds.mds_max_mdsize,
1280 obd->u.mds.mds_max_cookiesize};
1283 rc = lustre_pack_reply(req, 3, repsize, NULL);
1285 CERROR("lustre_pack_reply: rc = %d\n", rc);
1286 req->rq_status = rc;
1288 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1291 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1293 CERROR("Can't unpack body\n");
1294 req->rq_status = -EFAULT;
1298 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1299 /* do some stuff */ ;
1301 mfd = mds_handle2mfd(&body->handle);
1303 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1304 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1305 req->rq_status = -ESTALE;
1309 inode = mfd->mfd_dentry->d_inode;
1310 if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
1311 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1312 LASSERT(body != NULL);
1314 mds_pack_inode2fid(obd, &body->fid1, inode);
1315 mds_pack_inode2body(obd, body, inode);
1316 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
1318 spin_lock(&med->med_open_lock);
1319 list_del(&mfd->mfd_list);
1320 spin_unlock(&med->med_open_lock);
1322 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1323 req->rq_status = mds_mfd_close(rc ? NULL : req, obd, mfd, 1);
1324 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1326 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1327 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1328 req->rq_status = -ENOMEM;
1337 int mds_done_writing(struct ptlrpc_request *req)
1339 struct mds_body *body;
1340 int rc, size = sizeof(struct mds_body);
1343 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1345 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1347 CERROR("Can't unpack body\n");
1348 req->rq_status = -EFAULT;
1352 rc = lustre_pack_reply(req, 1, &size, NULL);
1354 CERROR("lustre_pack_reply: rc = %d\n", rc);
1355 req->rq_status = rc;