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);
365 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
366 GOTO(out_ids, rc = -ENOMEM);
370 GOTO(out_ids, rc = -ENOMEM);
371 oa->o_mode = S_IFREG | 0600;
372 oa->o_id = inode->i_ino;
373 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
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 | OBD_MD_FLGROUP;
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 lfs setstripe 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 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
394 rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
399 CDEBUG(level, "error creating objects for "
400 "inode %lu: rc = %d\n",
403 CERROR("obd_create returned invalid "
410 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
411 0, &lsm, rec->ur_eadata);
415 lsm->lsm_object_id = oa->o_id;
416 lsm->lsm_object_gr = oa->o_gr;
419 oa->o_size = inode->i_size;
420 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
421 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
422 rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti);
424 CERROR("error setting attrs for inode %lu: rc %d\n",
427 CERROR("obd_setattr returned bad rc %d\n", rc);
434 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
435 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
437 LASSERT(lsm && lsm->lsm_object_id);
439 rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm);
440 if (!rec->ur_fid2->id)
441 obd_free_memmd(mds->mds_osc_exp, &lsm);
444 body->eadatasize = rc;
445 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
446 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
447 lmm_bufsize = req->rq_repmsg->buflens[offset];
449 LASSERT(lmm_bufsize >= lmm_size);
451 memcpy(lmm_buf, lmm, lmm_size);
452 obd_free_diskmd(mds->mds_osc_exp, &lmm);
454 oti_free_cookies(&oti);
458 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
464 static void reconstruct_open(struct mds_update_record *rec, int offset,
465 struct ptlrpc_request *req,
466 struct lustre_handle *child_lockh)
468 struct mds_export_data *med = &req->rq_export->exp_mds_data;
469 struct mds_client_data *mcd = med->med_mcd;
470 struct mds_obd *mds = mds_req2mds(req);
471 struct mds_file_data *mfd;
472 struct obd_device *obd = req->rq_export->exp_obd;
473 struct dentry *parent, *child;
474 struct ldlm_reply *rep;
475 struct mds_body *body;
481 LASSERT(offset == 2); /* only called via intent */
482 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
483 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
485 /* copy rc, transno and disp; steal locks */
486 mds_req_from_mcd(req, mcd);
487 intent_set_disposition(rep, mcd->mcd_last_data);
489 /* Only replay if create or open actually happened. */
490 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
492 return; /* error looking up parent or child */
495 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
496 LASSERT(!IS_ERR(parent));
498 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
499 LASSERT(!IS_ERR(child));
501 if (!child->d_inode) {
502 GOTO(out_dput, 0); /* child not present to open */
505 /* At this point, we know we have a child. We'll send
506 * it back _unless_ it not created and open failed.
508 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
509 !intent_disposition(rep, DISP_OPEN_CREATE) &&
514 /* get lock (write for O_CREAT, read otherwise) */
516 mds_pack_inode2fid(obd, &body->fid1, child->d_inode);
517 mds_pack_inode2body(obd, body, child->d_inode);
518 if (S_ISREG(child->d_inode->i_mode)) {
519 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
523 LASSERT(rc == req->rq_status);
525 /* If we have LOV EA data, the OST holds size, mtime */
526 if (!(body->valid & OBD_MD_FLEASIZE))
527 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
528 OBD_MD_FLATIME | OBD_MD_FLMTIME);
530 /* XXX need to check this case */
533 /* If we're opening a file without an EA, change to a write
534 lock (unless we already have one). */
536 /* If we have -EEXIST as the status, and we were asked to create
537 * exclusively, we can tell we failed because the file already existed.
539 if (req->rq_status == -EEXIST &&
540 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
541 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
545 /* If we didn't get as far as trying to open, then some locking thing
546 * probably went wrong, and we'll just bail here.
548 if (!intent_disposition(rep, DISP_OPEN_OPEN))
551 /* If we failed, then we must have failed opening, so don't look for
552 * file descriptor or anything, just give the client the bad news.
558 list_for_each(t, &med->med_open_head) {
559 mfd = list_entry(t, struct mds_file_data, mfd_list);
560 if (mfd->mfd_xid == req->rq_xid)
565 /* #warning "XXX fixme" bug 2991 */
566 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
567 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
568 * to detect a re-open */
570 mntget(mds->mds_vfsmnt);
571 CERROR("Re-opened file \n");
572 mfd = mds_dentry_open(child, mds->mds_vfsmnt,
573 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
575 CERROR("mds: out of memory\n");
576 GOTO(out_dput, req->rq_status = -ENOMEM);
580 body->handle.cookie = mfd->mfd_handle.h_cookie;
581 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
582 mfd->mfd_handle.h_cookie);
592 /* do NOT or the MAY_*'s, you'll get the weakest */
593 static int accmode(int flags)
597 if (flags & FMODE_READ)
599 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
601 if (flags & FMODE_EXEC)
606 /* Handles object creation, actual opening, and I/O epoch */
607 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
608 struct mds_body *body, int flags, void **handle,
609 struct mds_update_record *rec,struct ldlm_reply *rep)
611 struct mds_obd *mds = mds_req2mds(req);
612 struct obd_device *obd = req->rq_export->exp_obd;
613 struct mds_file_data *mfd = NULL;
614 obd_id *ids = NULL; /* object IDs created */
619 /* atomically create objects if necessary */
620 down(&dchild->d_inode->i_sem);
621 mode = dchild->d_inode->i_mode;
622 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
623 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
624 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
627 up(&dchild->d_inode->i_sem);
632 /* no EA: create objects */
633 rc = mds_create_objects(req, 2, rec, mds, obd,
634 dchild, handle, &ids);
636 CERROR("mds_create_objects: rc = %d\n", rc);
637 up(&dchild->d_inode->i_sem);
640 if (S_ISREG(dchild->d_inode->i_mode) &&
641 (body->valid & OBD_MD_FLEASIZE)) {
642 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
645 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
648 up(&dchild->d_inode->i_sem);
653 /* If the inode has no EA data, then MDSs hold size, mtime */
654 if (S_ISREG(dchild->d_inode->i_mode) &&
655 !(body->valid & OBD_MD_FLEASIZE)) {
656 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
657 OBD_MD_FLATIME | OBD_MD_FLMTIME);
659 up(&dchild->d_inode->i_sem);
661 intent_set_disposition(rep, DISP_OPEN_OPEN);
662 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
664 RETURN(PTR_ERR(mfd));
666 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
667 mfd->mfd_handle.h_cookie);
669 mds_lov_update_objids(obd, ids);
670 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
673 // mds_mfd_destroy(mfd);
677 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
678 struct mds_body *body, int flags,
679 struct mds_update_record *rec,struct ldlm_reply *rep)
681 struct mds_obd *mds = mds_req2mds(req);
682 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
683 struct dentry *dchild;
684 char fidname[LL_FID_NAMELEN];
689 down(&pending_dir->i_sem);
690 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
691 dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
692 if (IS_ERR(dchild)) {
693 up(&pending_dir->i_sem);
694 rc = PTR_ERR(dchild);
695 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
699 if (dchild->d_inode != NULL) {
700 up(&pending_dir->i_sem);
701 mds_inode_set_orphan(dchild->d_inode);
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);
706 CWARN("Orphan %s found and opened in PENDING directory\n",
711 up(&pending_dir->i_sem);
713 /* We didn't find it in PENDING so it isn't an orphan. See
714 * if it was a regular inode that was previously created. */
715 dchild = mds_fid2dentry(mds, fid, NULL);
717 RETURN(PTR_ERR(dchild));
719 mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
720 mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
721 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
722 intent_set_disposition(rep, DISP_LOOKUP_POS);
725 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
726 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
727 req, rc, rep ? rep->lock_policy_res1 : 0);
728 /* XXX what do we do here if mds_finish_transno itself failed? */
734 int mds_pin(struct ptlrpc_request *req)
736 struct obd_device *obd = req->rq_export->exp_obd;
737 struct mds_body *request_body, *reply_body;
738 struct lvfs_run_ctxt saved;
739 int rc, size = sizeof(*reply_body);
742 request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body));
744 rc = lustre_pack_reply(req, 1, &size, NULL);
747 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
749 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
750 rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
751 request_body->flags, NULL, NULL);
752 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
757 /* Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
758 * If child_lockh is NULL we just get the lock as a barrier to wait for
759 * other holders of this lock, and drop it right away again. */
760 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
761 struct lustre_handle *child_lockh)
763 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
764 struct lustre_handle lockh;
768 if (child_lockh == NULL)
769 child_lockh = &lockh;
771 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
772 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
773 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
774 NULL, 0, NULL, child_lockh);
776 CERROR("ldlm_cli_enqueue: %d\n", rc);
777 else if (child_lockh == &lockh)
778 ldlm_lock_decref(child_lockh, LCK_EX);
783 static int is_mount_object(struct dentry *dparent)
785 struct dentry *dchild;
787 if (!(dparent->d_inode->i_mode & S_ISUID))
790 dchild = lookup_one_len(".mntinfo", dparent, strlen(".mntinfo"));
791 if (IS_ERR(dchild) || dchild == NULL)
798 int mds_open(struct mds_update_record *rec, int offset,
799 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
801 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
802 struct obd_device *obd = req->rq_export->exp_obd;
803 struct mds_obd *mds = mds_req2mds(req);
804 struct ldlm_reply *rep = NULL;
805 struct mds_body *body = NULL;
806 struct dentry *dchild = NULL, *dparent = NULL;
807 struct mds_export_data *med;
808 struct lustre_handle parent_lockh[2];
809 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
810 int parent_mode = LCK_PR;
812 struct dentry_params dp;
813 struct mea *mea = NULL;
814 int mea_size, update_mode;
817 parent_lockh[0].cookie = 0;
818 parent_lockh[1].cookie = 0;
820 if (offset == 2) { /* intent */
821 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
822 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
823 } else if (offset == 0) { /* non-intent reint */
824 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
830 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
832 /* Step 0: If we are passed a fid, then we assume the client already
833 * opened this file and is only replaying the RPC, so we open the
834 * inode by fid (at some large expense in security). */
835 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
836 DEBUG_REQ(D_HA, req, "open replay, disp: "LPX64"\n",
837 rep->lock_policy_res1);
839 LASSERT(rec->ur_fid2->id);
841 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
845 /* We didn't find the correct inode on disk either, so we
846 * need to re-create it via a regular replay. */
847 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
849 LASSERT(!rec->ur_fid2->id);
852 LASSERT(offset == 2); /* If we got here, we must be called via intent */
854 med = &req->rq_export->exp_mds_data;
855 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
856 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
860 acc_mode = accmode(rec->ur_flags);
862 /* Step 1: Find and lock the parent */
863 if (rec->ur_flags & O_CREAT) {
864 /* XXX Well, in fact we only need this lock mode change if
865 in addition to O_CREAT, the file does not exist.
866 But we do not know if it exists or not yet */
867 parent_mode = LCK_PW;
870 if (rec->ur_namelen == 1) {
871 /* client (LMV) wants to open the file by fid */
872 CDEBUG(D_OTHER, "OPEN by fid %u/%u/%u\n",
873 (unsigned) rec->ur_fid1->mds,
874 (unsigned) rec->ur_fid1->id,
875 (unsigned) rec->ur_fid1->generation);
876 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
877 if (IS_ERR(dchild)) {
878 rc = PTR_ERR(dparent);
879 CERROR("child lookup by a fid error %d\n", rc);
885 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
886 parent_lockh, &update_mode, rec->ur_name,
888 MDS_INODELOCK_UPDATE);
889 if (IS_ERR(dparent)) {
890 rc = PTR_ERR(dparent);
891 CERROR("parent lookup error %d\n", rc);
894 LASSERT(dparent->d_inode != NULL);
896 cleanup_phase = 1; /* parent dentry and lock */
898 /* try to retrieve MEA data for this dir */
899 rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
904 /* dir is already splitted, check is requested filename
905 * should live at this MDS or at another one */
907 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
908 if (mea->mea_master != i) {
909 CERROR("inapropriate MDS(%d) for %s. should be %d\n",
910 mea->mea_master, rec->ur_name, i);
911 GOTO(cleanup, rc = -ERESTART);
915 /* Step 2: Lookup the child */
916 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
917 if (IS_ERR(dchild)) {
918 rc = PTR_ERR(dchild);
919 dchild = NULL; /* don't confuse mds_finish_transno() below */
924 cleanup_phase = 2; /* child dentry */
926 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
928 if (dchild->d_flags & DCACHE_CROSS_REF) {
929 struct ldlm_res_id res_id = { . name = {0} };
930 ldlm_policy_data_t policy;
932 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
933 (unsigned long) dchild->d_mdsnum,
934 (unsigned long) dchild->d_inum,
935 (unsigned long) dchild->d_generation);
936 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
937 body->fid1.id = dchild->d_inum;
938 body->fid1.mds = dchild->d_mdsnum;
939 body->fid1.generation = dchild->d_generation;
940 intent_set_disposition(rep, DISP_LOOKUP_POS);
941 res_id.name[0] = dchild->d_inum;
942 res_id.name[1] = dchild->d_generation;
943 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
944 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
945 res_id, LDLM_IBITS, &policy,
948 ldlm_completion_ast, NULL, NULL,
949 NULL, 0, NULL, child_lockh);
951 if (parent_lockh[1].cookie != 0)
952 ldlm_lock_decref(parent_lockh + 1, update_mode);
954 ldlm_lock_decref(parent_lockh, parent_mode);
956 OBD_FREE(mea, mea_size);
963 intent_set_disposition(rep, DISP_LOOKUP_POS);
965 intent_set_disposition(rep, DISP_LOOKUP_NEG);
967 /*Step 3: If the child was negative, and we're supposed to, create it.*/
968 if (dchild->d_inode == NULL) {
969 unsigned long ino = rec->ur_fid2->id;
973 if ((rc = mds_try_to_split_dir(obd, dparent, &mea, 0))) {
975 /* dir got splitted */
976 GOTO(cleanup, rc = -ERESTART);
978 /* error happened during spitting */
983 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
984 /* It's negative and we weren't supposed to create it */
985 GOTO(cleanup, rc = -ENOENT);
988 intent_set_disposition(rep, DISP_OPEN_CREATE);
989 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
991 if (IS_ERR(handle)) {
992 rc = PTR_ERR(handle);
996 dchild->d_fsdata = (void *) &dp;
1000 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1001 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1002 dchild->d_fsdata = NULL;
1005 CDEBUG(D_INODE, "error during create: %d\n", rc);
1008 inode = dchild->d_inode;
1010 LASSERT(ino == inode->i_ino);
1011 /* Written as part of setattr */
1012 inode->i_generation = rec->ur_fid2->generation;
1013 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1014 inode->i_ino, inode->i_generation);
1018 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
1019 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1020 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1022 iattr.ia_uid = rec->_ur_fsuid;
1023 if (dparent->d_inode->i_mode & S_ISGID)
1024 iattr.ia_gid = dparent->d_inode->i_gid;
1026 iattr.ia_gid = rec->_ur_fsgid;
1028 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1029 ATTR_MTIME | ATTR_CTIME;
1031 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1033 CERROR("error on child setattr: rc = %d\n", rc);
1035 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1037 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1039 CERROR("error on parent setattr: rc = %d\n", rc);
1041 acc_mode = 0; /* Don't check for permissions */
1044 LASSERT(!mds_inode_is_orphan(dchild->d_inode));
1046 mds_pack_inode2fid(obd, &body->fid1, dchild->d_inode);
1047 mds_pack_inode2body(obd, body, dchild->d_inode);
1049 if (S_ISREG(dchild->d_inode->i_mode)) {
1050 /* Check permissions etc */
1051 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1055 /* Can't write to a read-only file */
1056 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1057 GOTO(cleanup, rc = -EPERM);
1059 /* An append-only file must be opened in append mode for
1061 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1062 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1063 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1064 GOTO(cleanup, rc = -EPERM);
1067 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1068 (rec->ur_flags & MDS_OPEN_EXCL)) {
1069 /* File already exists, we didn't just create it, and we
1070 * were passed O_EXCL; err-or. */
1071 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1074 /* if we are following a symlink, don't open */
1075 if (S_ISLNK(dchild->d_inode->i_mode))
1076 GOTO(cleanup, rc = 0);
1078 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1079 !S_ISDIR(dchild->d_inode->i_mode))
1080 GOTO(cleanup, rc = -ENOTDIR);
1082 if (S_ISDIR(dchild->d_inode->i_mode)) {
1083 if (rec->ur_flags & MDS_OPEN_CREAT ||
1084 rec->ur_flags & FMODE_WRITE) {
1085 /*we are trying to create or write a exist dir*/
1086 GOTO(cleanup, rc = -EISDIR);
1088 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1089 GOTO(cleanup, rc = -EACCES);
1091 if (is_mount_object(dchild)) {
1092 CERROR("Found possible GNS mount object %*s; not "
1093 "opening.\n", dchild->d_name.len,
1094 dchild->d_name.name);
1095 GOTO(cleanup, rc = 0); // success, but don't really open
1099 /* Step 5: mds_open it */
1100 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1105 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1106 req, rc, rep ? rep->lock_policy_res1 : 0);
1108 switch (cleanup_phase) {
1110 if (rc && created) {
1111 int err = vfs_unlink(dparent->d_inode, dchild);
1113 CERROR("unlink(%*s) in error path: %d\n",
1114 dchild->d_name.len, dchild->d_name.name,
1117 } else if (created) {
1118 mds_lock_new_child(obd, dchild->d_inode, NULL);
1122 if (dparent == NULL)
1127 if (parent_lockh[1].cookie != 0)
1128 ldlm_lock_decref(parent_lockh + 1, update_mode);
1131 ldlm_lock_decref(parent_lockh, parent_mode);
1133 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1136 atomic_inc(&mds->mds_open_count);
1138 OBD_FREE(mea, mea_size);
1142 /* Close a "file descriptor" and possibly unlink an orphan from the
1143 * PENDING directory.
1145 * If we are being called from mds_disconnect() because the client has
1146 * disappeared, then req == NULL and we do not update last_rcvd because
1147 * there is nothing that could be recovered by the client at this stage
1148 * (it will not even _have_ an entry in last_rcvd anymore).
1150 * Returns EAGAIN if the client needs to get more data and re-close. */
1151 int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
1152 struct mds_file_data *mfd, int unlink_orphan)
1154 struct inode *inode = mfd->mfd_dentry->d_inode;
1155 char fidname[LL_FID_NAMELEN];
1156 int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1157 struct dentry *pending_child = NULL;
1158 struct mds_obd *mds = &obd->u.mds;
1159 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1160 void *handle = NULL;
1161 struct mds_body *request_body = NULL, *reply_body = NULL;
1162 struct dentry_params dp;
1163 struct iattr iattr = { 0 };
1166 if (req && req->rq_reqmsg != NULL)
1167 request_body = lustre_msg_buf(req->rq_reqmsg, 0,
1168 sizeof(*request_body));
1169 if (req && req->rq_repmsg != NULL)
1170 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1171 sizeof(*reply_body));
1173 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1175 last_orphan = mds_open_orphan_dec_test(inode) &&
1176 mds_inode_is_orphan(inode);
1178 /* this is half of the actual "close" */
1179 if (mfd->mfd_mode & FMODE_WRITE) {
1180 rc = mds_put_write_access(mds, inode, request_body,
1181 last_orphan && unlink_orphan);
1182 } else if (mfd->mfd_mode & FMODE_EXEC) {
1183 mds_allow_write_access(inode);
1186 if (last_orphan && unlink_orphan) {
1187 struct lov_mds_md *lmm = NULL;
1188 int stripe_count = 0;
1189 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1191 if (obd->obd_recovering) {
1192 CDEBUG(D_HA, "not remove orphan %s until recovery"
1193 " is over\n", fidname);
1197 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1199 /* Sadly, there is no easy way to save pending_child from
1200 * mds_reint_unlink() into mfd, so we need to re-lookup,
1201 * but normally it will still be in the dcache. */
1202 pending_dir = mds->mds_pending_dir->d_inode;
1203 down(&pending_dir->i_sem);
1204 cleanup_phase = 1; /* up(i_sem) when finished */
1205 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1207 if (IS_ERR(pending_child))
1208 GOTO(cleanup, rc = PTR_ERR(pending_child));
1209 LASSERT(pending_child->d_inode != NULL);
1211 cleanup_phase = 2; /* dput(pending_child) when finished */
1212 if (req != NULL && req->rq_repmsg != NULL) {
1213 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1214 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1217 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1218 NULL, stripe_count);
1219 if (IS_ERR(handle)) {
1220 rc = PTR_ERR(handle);
1225 if (req != NULL && req->rq_repmsg != NULL &&
1226 (reply_body->valid & OBD_MD_FLEASIZE) &&
1227 mds_log_op_unlink(obd, pending_child->d_inode, lmm,
1228 req->rq_repmsg->buflens[1],
1229 lustre_msg_buf(req->rq_repmsg, 2, 0),
1230 req->rq_repmsg->buflens[2]) > 0) {
1231 reply_body->valid |= OBD_MD_FLCOOKIE;
1234 pending_child->d_fsdata = (void *) &dp;
1237 if (S_ISDIR(pending_child->d_inode->i_mode))
1238 rc = vfs_rmdir(pending_dir, pending_child);
1240 rc = vfs_unlink(pending_dir, pending_child);
1242 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1244 goto out; /* Don't bother updating attrs on unlinked inode */
1247 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1248 /* Update the on-disk attributes if this was the last write
1249 * close, and all information was provided (i.e., rc == 0)
1251 * XXX this should probably be abstracted with mds_reint_setattr
1255 if (request_body->valid & OBD_MD_FLMTIME &&
1256 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1257 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1258 iattr.ia_valid |= ATTR_MTIME;
1260 if (request_body->valid & OBD_MD_FLCTIME &&
1261 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1262 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1263 iattr.ia_valid |= ATTR_CTIME;
1266 /* XXX can't set block count with fsfilt_setattr (!) */
1267 if (request_body->valid & OBD_MD_FLSIZE) {
1268 iattr.ia_valid |= ATTR_SIZE;
1269 iattr.ia_size = request_body->size;
1271 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1272 iattr.ia_valid |= ATTR_BLOCKS;
1273 iattr.ia_blocks = request_body->blocks
1279 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1280 /* Only start a transaction to write out only the atime if
1281 * it is more out-of-date than the specified limit. If we
1282 * are already going to write out the atime then do it anyway.
1284 if ((LTIME_S(request_body->atime) >
1285 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1286 (iattr.ia_valid != 0 &&
1287 LTIME_S(request_body->atime) > LTIME_S(inode->i_atime))) {
1288 LTIME_S(iattr.ia_atime) = LTIME_S(request_body->atime);
1289 iattr.ia_valid |= ATTR_ATIME;
1293 if (iattr.ia_valid != 0) {
1294 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1296 GOTO(cleanup, rc = PTR_ERR(handle));
1297 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1299 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1302 /* If other clients have this file open for write, rc will be > 0 */
1305 l_dput(mfd->mfd_dentry);
1306 mds_mfd_destroy(mfd);
1309 atomic_dec(&mds->mds_open_count);
1310 if (req != NULL && reply_body != NULL) {
1311 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1312 } else if (handle) {
1313 int err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, 0);
1315 CERROR("error committing close: %d\n", err);
1321 switch (cleanup_phase) {
1323 dput(pending_child);
1325 up(&pending_dir->i_sem);
1330 int mds_close(struct ptlrpc_request *req)
1332 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1333 struct obd_device *obd = req->rq_export->exp_obd;
1334 struct mds_body *body;
1335 struct mds_file_data *mfd;
1336 struct lvfs_run_ctxt saved;
1337 struct inode *inode;
1338 int rc, repsize[3] = {sizeof(struct mds_body),
1339 obd->u.mds.mds_max_mdsize,
1340 obd->u.mds.mds_max_cookiesize};
1343 rc = lustre_pack_reply(req, 3, repsize, NULL);
1345 CERROR("lustre_pack_reply: rc = %d\n", rc);
1346 req->rq_status = rc;
1348 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1351 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1353 CERROR("Can't unpack body\n");
1354 req->rq_status = -EFAULT;
1358 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1359 /* do some stuff */ ;
1361 mfd = mds_handle2mfd(&body->handle);
1363 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1364 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1365 req->rq_status = -ESTALE;
1369 inode = mfd->mfd_dentry->d_inode;
1370 if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
1371 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1372 LASSERT(body != NULL);
1374 mds_pack_inode2fid(obd, &body->fid1, inode);
1375 mds_pack_inode2body(obd, body, inode);
1376 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
1378 spin_lock(&med->med_open_lock);
1379 list_del(&mfd->mfd_list);
1380 spin_unlock(&med->med_open_lock);
1382 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1383 req->rq_status = mds_mfd_close(req, obd, mfd, 1);
1384 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1386 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1387 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1388 req->rq_status = -ENOMEM;
1397 int mds_done_writing(struct ptlrpc_request *req)
1399 struct mds_body *body;
1400 int rc, size = sizeof(struct mds_body);
1403 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1405 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1407 CERROR("Can't unpack body\n");
1408 req->rq_status = -EFAULT;
1412 rc = lustre_pack_reply(req, 1, &size, NULL);
1414 CERROR("lustre_pack_reply: rc = %d\n", rc);
1415 req->rq_status = rc;