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/module.h>
32 #include <linux/init.h>
33 #include <linux/version.h>
34 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
35 # include <linux/buffer_head.h>
36 # include <linux/workqueue.h>
38 # include <linux/locks.h>
41 #include <linux/obd_class.h>
42 #include <linux/lustre_mds.h>
43 #include <linux/obd_lov.h>
44 #include <linux/lustre_fsfilt.h>
45 #include <linux/lprocfs_status.h>
47 #include "mds_internal.h"
49 /* Exported function from this file are:
51 * mds_open - called by the intent handler
52 * mds_close - an rpc handling function
53 * mds_pin - an rpc handling function - which will go away
54 * mds_mfd_close - for force closing files when a client dies
58 * MDS file data handling: file data holds a handle for a file opened
62 static void mds_mfd_addref(void *mfdp)
64 struct mds_file_data *mfd = mfdp;
66 atomic_inc(&mfd->mfd_refcount);
67 CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
68 atomic_read(&mfd->mfd_refcount));
71 struct mds_file_data *mds_mfd_new(void)
73 struct mds_file_data *mfd;
75 OBD_ALLOC(mfd, sizeof *mfd);
77 CERROR("mds: out of memory\n");
81 atomic_set(&mfd->mfd_refcount, 2);
83 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
84 class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
89 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
92 LASSERT(handle != NULL);
93 RETURN(class_handle2object(handle->cookie));
96 static void mds_mfd_put(struct mds_file_data *mfd)
98 CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
99 atomic_read(&mfd->mfd_refcount) - 1);
100 LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
101 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
102 if (atomic_dec_and_test(&mfd->mfd_refcount)) {
103 LASSERT(list_empty(&mfd->mfd_handle.h_link));
104 OBD_FREE(mfd, sizeof *mfd);
108 static void mds_mfd_destroy(struct mds_file_data *mfd)
110 class_handle_unhash(&mfd->mfd_handle);
115 /* Caller must hold mds->mds_epoch_sem */
116 static int mds_alloc_filterdata(struct inode *inode)
118 LASSERT(inode->i_filterdata == NULL);
119 OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
120 if (inode->i_filterdata == NULL)
122 LASSERT(igrab(inode) == inode);
126 /* Caller must hold mds->mds_epoch_sem */
127 static void mds_free_filterdata(struct inode *inode)
129 LASSERT(inode->i_filterdata != NULL);
130 OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
131 inode->i_filterdata = NULL;
135 /* Write access to a file: executors cause a negative count,
136 * writers a positive count. The semaphore is needed to perform
137 * a check for the sign and then increment or decrement atomically.
139 * This code is closely tied to the allocation of the d_fsdata and the
140 * MDS epoch, so we use the same semaphore for the whole lot.
142 * We could use a different semaphore for each file, if it ever shows
143 * up in a profile, which it won't.
145 * epoch argument is nonzero during recovery */
146 static int mds_get_write_access(struct mds_obd *mds, struct inode *inode,
151 down(&mds->mds_epoch_sem);
153 if (atomic_read(&inode->i_writecount) < 0) {
154 up(&mds->mds_epoch_sem);
159 if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
160 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
161 MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
162 inode->i_generation);
166 if (inode->i_filterdata == NULL)
167 mds_alloc_filterdata(inode);
168 if (inode->i_filterdata == NULL) {
172 if (epoch > mds->mds_io_epoch)
173 mds->mds_io_epoch = epoch;
176 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
177 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
178 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
181 atomic_inc(&inode->i_writecount);
182 up(&mds->mds_epoch_sem);
186 /* Returns EAGAIN if the client needs to get size and/or cookies and close
187 * again -- which is never true if the file is about to be unlinked. Otherwise
188 * returns the number of remaining writers. */
189 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
190 struct mds_body *body, int unlinking)
195 down(&mds->mds_epoch_sem);
196 atomic_dec(&inode->i_writecount);
197 rc = atomic_read(&inode->i_writecount);
201 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
202 GOTO(out, rc = EAGAIN);
204 mds_free_filterdata(inode);
206 up(&mds->mds_epoch_sem);
210 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
213 down(&mds->mds_epoch_sem);
214 if (atomic_read(&inode->i_writecount) > 0) {
215 up(&mds->mds_epoch_sem);
218 atomic_dec(&inode->i_writecount);
219 up(&mds->mds_epoch_sem);
223 static void mds_allow_write_access(struct inode *inode)
226 atomic_inc(&inode->i_writecount);
229 int mds_query_write_access(struct inode *inode)
232 RETURN(atomic_read(&inode->i_writecount));
235 /* This replaces the VFS mds_dentry_open, it manages mfd and writecount */
236 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
237 struct vfsmount *mnt, int flags,
238 struct ptlrpc_request *req)
240 struct mds_export_data *med = &req->rq_export->exp_mds_data;
241 struct mds_obd *mds = mds_req2mds(req);
242 struct mds_file_data *mfd;
243 struct mds_body *body;
249 CERROR("mds: out of memory\n");
250 GOTO(cleanup_dentry, error = -ENOMEM);
253 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
255 if (flags & FMODE_WRITE) {
256 /* FIXME: in recovery, need to pass old epoch here */
257 error = mds_get_write_access(mds, dentry->d_inode, 0);
259 GOTO(cleanup_mfd, error);
260 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
261 } else if (flags & FMODE_EXEC) {
262 error = mds_deny_write_access(mds, dentry->d_inode);
264 GOTO(cleanup_mfd, error);
269 /* Mark the file as open to handle open-unlink. */
270 mds_open_orphan_inc(dentry->d_inode);
272 mfd->mfd_mode = flags;
273 mfd->mfd_dentry = dentry;
274 mfd->mfd_xid = req->rq_xid;
276 spin_lock(&med->med_open_lock);
277 list_add(&mfd->mfd_list, &med->med_open_head);
278 spin_unlock(&med->med_open_lock);
281 body->handle.cookie = mfd->mfd_handle.h_cookie;
287 mds_mfd_destroy(mfd);
289 return ERR_PTR(error);
292 static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
293 struct lov_desc *desc)
296 for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
297 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
298 le64_to_cpu(lmm->lmm_objects[i].l_object_id);
302 /* Must be called with i_sem held */
303 static int mds_create_objects(struct ptlrpc_request *req, int offset,
304 struct mds_update_record *rec,
305 struct mds_obd *mds, struct obd_device *obd,
306 struct inode *inode, void **handle, obd_id **ids)
309 struct obd_trans_info oti = { 0 };
310 struct mds_body *body;
311 struct lov_stripe_md *lsm = NULL;
312 struct lov_mds_md *lmm = NULL;
314 int rc, lmm_bufsize, lmm_size;
317 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
318 !(rec->ur_flags & FMODE_WRITE))
321 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
323 if (!S_ISREG(inode->i_mode))
325 if (body->valid & OBD_MD_FLEASIZE)
328 OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
331 oti.oti_objid = *ids;
334 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
335 if (IS_ERR(*handle)) {
336 rc = PTR_ERR(*handle);
342 if(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
343 LASSERT (rec->ur_fid2->id);
344 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
345 lmm_size = rec->ur_eadatalen;
346 lmm = rec->ur_eadata;
349 mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
351 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
352 lmm_bufsize = req->rq_repmsg->buflens[offset];
354 LASSERT(lmm_bufsize >= lmm_size);
355 memcpy(lmm_buf, lmm, lmm_size);
356 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
358 CERROR("open replay failed to set md:%d\n", rc);
363 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
364 GOTO(out_ids, rc = -ENOMEM);
368 GOTO(out_ids, rc = -ENOMEM);
369 oa->o_mode = S_IFREG | 0600;
370 oa->o_id = inode->i_ino;
371 oa->o_generation = inode->i_generation;
372 oa->o_uid = 0; /* must have 0 uid / gid on OST */
374 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
375 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
378 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
381 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
382 /* check if things like lstripe/lfs stripe are sending us the ea */
383 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
384 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
386 0, &lsm, rec->ur_eadata);
390 rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
395 CDEBUG(level, "error creating objects for "
396 "inode %lu: rc = %d\n",
399 CERROR("obd_create returned invalid "
406 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
407 0, &lsm, rec->ur_eadata);
411 lsm->lsm_object_id = oa->o_id;
414 oa->o_size = inode->i_size;
415 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
416 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
417 rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti);
419 CERROR("error setting attrs for inode %lu: rc %d\n",
422 CERROR("obd_setattr returned bad rc %d\n", rc);
429 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
430 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
432 LASSERT(lsm && lsm->lsm_object_id);
434 rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm);
435 if (!rec->ur_fid2->id)
436 obd_free_memmd(mds->mds_osc_exp, &lsm);
439 body->eadatasize = rc;
440 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
441 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
442 lmm_bufsize = req->rq_repmsg->buflens[offset];
444 LASSERT(lmm_bufsize >= lmm_size);
446 memcpy(lmm_buf, lmm, lmm_size);
447 obd_free_diskmd(mds->mds_osc_exp, &lmm);
449 oti_free_cookies(&oti);
453 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
459 static void reconstruct_open(struct mds_update_record *rec, int offset,
460 struct ptlrpc_request *req,
461 struct lustre_handle *child_lockh)
463 struct mds_export_data *med = &req->rq_export->exp_mds_data;
464 struct mds_client_data *mcd = med->med_mcd;
465 struct mds_obd *mds = mds_req2mds(req);
466 struct mds_file_data *mfd;
467 struct obd_device *obd = req->rq_export->exp_obd;
468 struct dentry *parent, *child;
469 struct ldlm_reply *rep;
470 struct mds_body *body;
476 LASSERT(offset == 2); /* only called via intent */
477 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
478 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
480 /* copy rc, transno and disp; steal locks */
481 mds_req_from_mcd(req, mcd);
482 intent_set_disposition(rep, mcd->mcd_last_data);
484 /* Only replay if create or open actually happened. */
485 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
487 return; /* error looking up parent or child */
490 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
491 LASSERT(!IS_ERR(parent));
493 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
494 LASSERT(!IS_ERR(child));
496 if (!child->d_inode) {
497 GOTO(out_dput, 0); /* child not present to open */
500 /* At this point, we know we have a child. We'll send
501 * it back _unless_ it not created and open failed.
503 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
504 !intent_disposition(rep, DISP_OPEN_CREATE) &&
509 /* get lock (write for O_CREAT, read otherwise) */
511 mds_pack_inode2fid(&body->fid1, child->d_inode);
512 mds_pack_inode2body(body, child->d_inode);
513 if (S_ISREG(child->d_inode->i_mode)) {
514 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
518 LASSERT(rc == req->rq_status);
520 /* If we have LOV EA data, the OST holds size, mtime */
521 if (!(body->valid & OBD_MD_FLEASIZE))
522 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
523 OBD_MD_FLATIME | OBD_MD_FLMTIME);
525 /* XXX need to check this case */
528 /* If we're opening a file without an EA, change to a write
529 lock (unless we already have one). */
531 /* If we have -EEXIST as the status, and we were asked to create
532 * exclusively, we can tell we failed because the file already existed.
534 if (req->rq_status == -EEXIST &&
535 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
536 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
540 /* If we didn't get as far as trying to open, then some locking thing
541 * probably went wrong, and we'll just bail here.
543 if (!intent_disposition(rep, DISP_OPEN_OPEN))
546 /* If we failed, then we must have failed opening, so don't look for
547 * file descriptor or anything, just give the client the bad news.
553 list_for_each(t, &med->med_open_head) {
554 mfd = list_entry(t, struct mds_file_data, mfd_list);
555 if (mfd->mfd_xid == req->rq_xid)
561 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
562 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
563 * to detect a re-open */
565 mntget(mds->mds_vfsmnt);
566 CERROR("Re-opened file \n");
567 mfd = mds_dentry_open(child, mds->mds_vfsmnt,
568 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
570 CERROR("mds: out of memory\n");
571 GOTO(out_dput, req->rq_status = -ENOMEM);
583 /* do NOT or the MAY_*'s, you'll get the weakest */
584 static int accmode(int flags)
588 if (flags & FMODE_READ)
590 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
592 if (flags & FMODE_EXEC)
597 /* Handles object creation, actual opening, and I/O epoch */
598 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
599 struct mds_body *body, int flags, void **handle,
600 struct mds_update_record *rec,struct ldlm_reply *rep)
602 struct mds_obd *mds = mds_req2mds(req);
603 struct obd_device *obd = req->rq_export->exp_obd;
604 struct mds_file_data *mfd = NULL;
605 obd_id *ids = NULL; /* object IDs created */
609 /* atomically create objects if necessary */
610 down(&dchild->d_inode->i_sem);
611 if (S_ISREG(dchild->d_inode->i_mode) &&
612 !(body->valid & OBD_MD_FLEASIZE)) {
613 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
616 up(&dchild->d_inode->i_sem);
621 /* no EA: create objects */
622 rc = mds_create_objects(req, 2, rec, mds, obd,
623 dchild->d_inode, handle, &ids);
625 CERROR("mds_create_objects: rc = %d\n", rc);
626 up(&dchild->d_inode->i_sem);
630 /* If the inode has EA data, then OSTs hold size, mtime */
631 if (S_ISREG(dchild->d_inode->i_mode) &&
632 !(body->valid & OBD_MD_FLEASIZE)) {
633 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
634 OBD_MD_FLATIME | OBD_MD_FLMTIME);
636 up(&dchild->d_inode->i_sem);
638 intent_set_disposition(rep, DISP_OPEN_OPEN);
639 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
641 RETURN(PTR_ERR(mfd));
643 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
644 mfd->mfd_handle.h_cookie);
646 mds_lov_update_objids(obd, ids);
647 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
650 // mds_mfd_destroy(mfd);
654 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
655 struct mds_body *body, int flags,
656 struct mds_update_record *rec,struct ldlm_reply *rep)
658 struct mds_obd *mds = mds_req2mds(req);
659 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
660 struct dentry *dchild;
661 char fidname[LL_FID_NAMELEN];
666 down(&pending_dir->i_sem);
667 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
668 dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
669 if (IS_ERR(dchild)) {
670 up(&pending_dir->i_sem);
671 rc = PTR_ERR(dchild);
672 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
676 if (dchild->d_inode != NULL) {
677 up(&pending_dir->i_sem);
678 mds_inode_set_orphan(dchild->d_inode);
679 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
680 mds_pack_inode2body(body, dchild->d_inode);
681 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
682 intent_set_disposition(rep, DISP_LOOKUP_POS);
683 CWARN("Orphan %s found and opened in PENDING directory\n",
688 up(&pending_dir->i_sem);
690 /* We didn't find it in PENDING so it isn't an orphan. See
691 * if it was a regular inode that was previously created. */
692 dchild = mds_fid2dentry(mds, fid, NULL);
694 RETURN(PTR_ERR(dchild));
696 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
697 mds_pack_inode2body(body, dchild->d_inode);
698 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
699 intent_set_disposition(rep, DISP_LOOKUP_POS);
702 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
703 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
704 req, rc, rep ? rep->lock_policy_res1 : 0);
705 /* XXX what do we do here if mds_finish_transno itself failed? */
711 int mds_pin(struct ptlrpc_request *req)
713 struct obd_device *obd = req->rq_export->exp_obd;
714 struct mds_body *request_body, *reply_body;
715 struct obd_run_ctxt saved;
716 int rc, size = sizeof(*reply_body);
719 request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body));
721 rc = lustre_pack_reply(req, 1, &size, NULL);
724 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
726 push_ctxt(&saved, &obd->obd_ctxt, NULL);
727 rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
728 request_body->flags, NULL, NULL);
729 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
734 /* Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
735 * If child_lockh is NULL we just get the lock as a barrier to wait for
736 * other holders of this lock, and drop it right away again. */
737 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
738 struct lustre_handle *child_lockh)
740 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
741 struct lustre_handle lockh;
745 if (child_lockh == NULL)
746 child_lockh = &lockh;
748 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
749 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
750 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
751 NULL, 0, NULL, child_lockh);
753 CERROR("ldlm_cli_enqueue: %d\n", rc);
754 else if (child_lockh == &lockh)
755 ldlm_lock_decref(child_lockh, LCK_EX);
760 int mds_open(struct mds_update_record *rec, int offset,
761 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
763 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
764 struct obd_device *obd = req->rq_export->exp_obd;
765 struct mds_obd *mds = mds_req2mds(req);
766 struct ldlm_reply *rep = NULL;
767 struct mds_body *body = NULL;
768 struct dentry *dchild = NULL, *dparent = NULL;
769 struct mds_export_data *med;
770 struct lustre_handle parent_lockh;
771 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
772 int parent_mode = LCK_PR;
774 struct dentry_params dp;
777 if (offset == 2) { /* intent */
778 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
779 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
780 } else if (offset == 0) { /* non-intent reint */
781 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
787 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
789 /* Step 0: If we are passed a fid, then we assume the client already
790 * opened this file and is only replaying the RPC, so we open the
791 * inode by fid (at some large expense in security). */
792 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
793 DEBUG_REQ(D_HA, req, "open replay, disp: "LPX64"\n",
794 rep->lock_policy_res1);
796 LASSERT(rec->ur_fid2->id);
798 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
802 /* We didn't find the correct inode on disk either, so we
803 * need to re-create it via a regular replay. */
804 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
806 LASSERT(!rec->ur_fid2->id);
809 LASSERT(offset == 2); /* If we got here, we must be called via intent */
811 med = &req->rq_export->exp_mds_data;
812 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
813 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
817 acc_mode = accmode(rec->ur_flags);
819 /* Step 1: Find and lock the parent */
820 if (rec->ur_flags & O_CREAT)
821 parent_mode = LCK_PW;
822 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
823 &parent_lockh, rec->ur_name,
824 rec->ur_namelen - 1);
825 if (IS_ERR(dparent)) {
826 rc = PTR_ERR(dparent);
827 CERROR("parent lookup error %d\n", rc);
830 LASSERT(dparent->d_inode != NULL);
832 cleanup_phase = 1; /* parent dentry and lock */
834 /* Step 2: Lookup the child */
835 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
836 if (IS_ERR(dchild)) {
837 rc = PTR_ERR(dchild);
838 dchild = NULL; /* don't confuse mds_finish_transno() below */
842 cleanup_phase = 2; /* child dentry */
844 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
846 intent_set_disposition(rep, DISP_LOOKUP_POS);
848 intent_set_disposition(rep, DISP_LOOKUP_NEG);
850 /*Step 3: If the child was negative, and we're supposed to, create it.*/
851 if (dchild->d_inode == NULL) {
852 unsigned long ino = rec->ur_fid2->id;
856 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
857 /* It's negative and we weren't supposed to create it */
858 GOTO(cleanup, rc = -ENOENT);
861 intent_set_disposition(rep, DISP_OPEN_CREATE);
862 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
864 if (IS_ERR(handle)) {
865 rc = PTR_ERR(handle);
869 dchild->d_fsdata = (void *) &dp;
873 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
874 if (dchild->d_fsdata == (void *)(unsigned long)ino)
875 dchild->d_fsdata = NULL;
878 CDEBUG(D_INODE, "error during create: %d\n", rc);
881 inode = dchild->d_inode;
883 LASSERT(ino == inode->i_ino);
884 /* Written as part of setattr */
885 inode->i_generation = rec->ur_fid2->generation;
886 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
887 inode->i_ino, inode->i_generation);
891 LTIME_S(iattr.ia_atime) = rec->ur_time;
892 LTIME_S(iattr.ia_ctime) = rec->ur_time;
893 LTIME_S(iattr.ia_mtime) = rec->ur_time;
895 iattr.ia_uid = rec->ur_fsuid;
896 if (dparent->d_inode->i_mode & S_ISGID)
897 iattr.ia_gid = dparent->d_inode->i_gid;
899 iattr.ia_gid = rec->ur_fsgid;
901 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
902 ATTR_MTIME | ATTR_CTIME;
904 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
906 CERROR("error on child setattr: rc = %d\n", rc);
908 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
910 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
912 CERROR("error on parent setattr: rc = %d\n", rc);
914 acc_mode = 0; /* Don't check for permissions */
917 LASSERT(!mds_inode_is_orphan(dchild->d_inode));
919 mds_pack_inode2fid(&body->fid1, dchild->d_inode);
920 mds_pack_inode2body(body, dchild->d_inode);
922 if (S_ISREG(dchild->d_inode->i_mode)) {
923 /* Check permissions etc */
924 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
928 /* Can't write to a read-only file */
929 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
930 GOTO(cleanup, rc = -EPERM);
932 /* An append-only file must be opened in append mode for
934 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
935 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
936 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
937 GOTO(cleanup, rc = -EPERM);
940 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
941 (rec->ur_flags & MDS_OPEN_EXCL)) {
942 /* File already exists, we didn't just create it, and we
943 * were passed O_EXCL; err-or. */
944 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
947 /* if we are following a symlink, don't open */
948 if (S_ISLNK(dchild->d_inode->i_mode))
949 GOTO(cleanup, rc = 0);
951 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
952 !S_ISDIR(dchild->d_inode->i_mode))
953 GOTO(cleanup, rc = -ENOTDIR);
955 if (S_ISDIR(dchild->d_inode->i_mode)) {
956 if (rec->ur_flags & MDS_OPEN_CREAT || rec->ur_flags & FMODE_WRITE) {
957 /*we are tryying to create or write a exist dir*/
958 GOTO(cleanup, rc = -EISDIR);
960 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
961 GOTO(cleanup, rc = -EACCES);
965 /* Step 5: mds_open it */
966 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
971 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
972 req, rc, rep ? rep->lock_policy_res1 : 0);
974 switch (cleanup_phase) {
977 int err = vfs_unlink(dparent->d_inode, dchild);
979 CERROR("unlink(%*s) in error path: %d\n",
980 dchild->d_name.len, dchild->d_name.name,
983 } else if (created) {
984 mds_lock_new_child(obd, dchild->d_inode, NULL);
993 ldlm_lock_decref(&parent_lockh, parent_mode);
995 ptlrpc_save_lock (req, &parent_lockh, parent_mode);
998 atomic_inc(&mds->mds_open_count);
1002 /* Close a "file descriptor" and possibly unlink an orphan from the
1003 * PENDING directory.
1005 * If we are being called from mds_disconnect() because the client has
1006 * disappeared, then req == NULL and we do not update last_rcvd because
1007 * there is nothing that could be recovered by the client at this stage
1008 * (it will not even _have_ an entry in last_rcvd anymore).
1010 * Returns EAGAIN if the client needs to get more data and re-close. */
1011 int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
1012 struct mds_file_data *mfd, int unlink_orphan)
1014 struct inode *inode = mfd->mfd_dentry->d_inode;
1015 char fidname[LL_FID_NAMELEN];
1016 int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1017 struct dentry *pending_child = NULL;
1018 struct mds_obd *mds = &obd->u.mds;
1019 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1020 void *handle = NULL;
1021 struct mds_body *request_body = NULL, *reply_body = NULL;
1022 struct dentry_params dp;
1026 request_body = lustre_msg_buf(req->rq_reqmsg, 0,
1027 sizeof(*request_body));
1028 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1029 sizeof(*reply_body));
1032 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1034 last_orphan = mds_open_orphan_dec_test(inode) &&
1035 mds_inode_is_orphan(inode);
1037 /* this is half of the actual "close" */
1038 if (mfd->mfd_mode & FMODE_WRITE) {
1039 rc = mds_put_write_access(mds, inode, request_body,
1040 last_orphan && unlink_orphan);
1041 } else if (mfd->mfd_mode & FMODE_EXEC) {
1042 mds_allow_write_access(inode);
1045 if (last_orphan && unlink_orphan) {
1046 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1048 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1050 /* Sadly, there is no easy way to save pending_child from
1051 * mds_reint_unlink() into mfd, so we need to re-lookup,
1052 * but normally it will still be in the dcache. */
1053 pending_dir = mds->mds_pending_dir->d_inode;
1054 down(&pending_dir->i_sem);
1055 cleanup_phase = 1; /* up(i_sem) when finished */
1056 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1058 if (IS_ERR(pending_child))
1059 GOTO(cleanup, rc = PTR_ERR(pending_child));
1060 LASSERT(pending_child->d_inode != NULL);
1062 cleanup_phase = 2; /* dput(pending_child) when finished */
1063 handle = fsfilt_start(obd, pending_dir, FSFILT_OP_UNLINK_LOG,
1065 if (IS_ERR(handle)) {
1066 rc = PTR_ERR(handle);
1072 (reply_body->valid & OBD_MD_FLEASIZE) &&
1073 mds_log_op_unlink(obd, pending_child->d_inode,
1074 lustre_msg_buf(req->rq_repmsg, 1, 0),
1075 req->rq_repmsg->buflens[1],
1076 lustre_msg_buf(req->rq_repmsg, 2, 0),
1077 req->rq_repmsg->buflens[2]) > 0) {
1078 reply_body->valid |= OBD_MD_FLCOOKIE;
1081 pending_child->d_fsdata = (void *) &dp;
1084 if (S_ISDIR(pending_child->d_inode->i_mode))
1085 rc = vfs_rmdir(pending_dir, pending_child);
1087 rc = vfs_unlink(pending_dir, pending_child);
1089 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1090 } else if (mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1091 /* Update the on-disk attributes if this was the last write
1092 * close, and all information was provided (i.e., rc == 0)
1094 * XXX this should probably be abstracted with mds_reint_setattr
1099 /* XXX can't set block count with fsfilt_setattr (!) */
1100 iattr.ia_valid = ATTR_CTIME | ATTR_ATIME |
1101 ATTR_MTIME | ATTR_SIZE;
1102 iattr.ia_atime = request_body->atime;
1103 iattr.ia_ctime = request_body->ctime;
1104 iattr.ia_mtime = request_body->mtime;
1105 iattr.ia_size = request_body->size;
1106 /* iattr.ia_blocks = request_body->blocks */
1108 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1110 GOTO(cleanup, rc = PTR_ERR(handle));
1111 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1113 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1116 /* If other clients have this file open for write, rc will be > 0 */
1119 l_dput(mfd->mfd_dentry);
1120 mds_mfd_destroy(mfd);
1123 atomic_dec(&mds->mds_open_count);
1125 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1126 } else if (handle) {
1127 int err = fsfilt_commit(obd, pending_dir, handle, 0);
1129 CERROR("error committing close: %d\n", err);
1135 switch (cleanup_phase) {
1137 dput(pending_child);
1139 up(&pending_dir->i_sem);
1144 int mds_close(struct ptlrpc_request *req)
1146 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1147 struct obd_device *obd = req->rq_export->exp_obd;
1148 struct mds_body *body;
1149 struct mds_file_data *mfd;
1150 struct obd_run_ctxt saved;
1151 struct inode *inode;
1152 int rc, repsize[3] = {sizeof(struct mds_body),
1153 obd->u.mds.mds_max_mdsize,
1154 obd->u.mds.mds_max_cookiesize};
1157 rc = lustre_pack_reply(req, 3, repsize, NULL);
1159 CERROR("lustre_pack_reply: rc = %d\n", rc);
1160 req->rq_status = rc;
1162 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1165 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1167 CERROR("Can't unpack body\n");
1168 req->rq_status = -EFAULT;
1172 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1173 /* do some stuff */ ;
1175 mfd = mds_handle2mfd(&body->handle);
1177 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1178 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1179 req->rq_status = -ESTALE;
1183 inode = mfd->mfd_dentry->d_inode;
1184 if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
1185 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1186 LASSERT(body != NULL);
1188 mds_pack_inode2fid(&body->fid1, inode);
1189 mds_pack_inode2body(body, inode);
1190 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
1192 spin_lock(&med->med_open_lock);
1193 list_del(&mfd->mfd_list);
1194 spin_unlock(&med->med_open_lock);
1196 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1197 req->rq_status = mds_mfd_close(rc ? NULL : req, obd, mfd, 1);
1198 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1200 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1201 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1202 req->rq_status = -ENOMEM;
1211 int mds_done_writing(struct ptlrpc_request *req)
1213 struct mds_body *body;
1214 int rc, size = sizeof(struct mds_body);
1217 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1219 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1221 CERROR("Can't unpack body\n");
1222 req->rq_status = -EFAULT;
1226 rc = lustre_pack_reply(req, 1, &size, NULL);
1228 CERROR("lustre_pack_reply: rc = %d\n", rc);
1229 req->rq_status = rc;