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 /* Per-directory striping default code removed, because
394 * it uses the same unnamed EA storage as the directory
395 * striping for CMD. -p */
397 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
398 rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
403 CDEBUG(level, "error creating objects for "
404 "inode %lu: rc = %d\n",
407 CERROR("obd_create returned invalid "
414 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
415 0, &lsm, rec->ur_eadata);
419 lsm->lsm_object_id = oa->o_id;
420 lsm->lsm_object_gr = oa->o_gr;
423 oa->o_size = inode->i_size;
424 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
425 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
426 rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti);
428 CERROR("error setting attrs for inode %lu: rc %d\n",
431 CERROR("obd_setattr returned bad rc %d\n", rc);
438 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
439 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
441 LASSERT(lsm && lsm->lsm_object_id);
443 rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm);
444 if (!rec->ur_fid2->id)
445 obd_free_memmd(mds->mds_osc_exp, &lsm);
448 body->eadatasize = rc;
449 rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
450 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
451 lmm_bufsize = req->rq_repmsg->buflens[offset];
453 LASSERT(lmm_bufsize >= lmm_size);
455 memcpy(lmm_buf, lmm, lmm_size);
456 obd_free_diskmd(mds->mds_osc_exp, &lmm);
458 oti_free_cookies(&oti);
462 OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
468 static void reconstruct_open(struct mds_update_record *rec, int offset,
469 struct ptlrpc_request *req,
470 struct lustre_handle *child_lockh)
472 struct mds_export_data *med = &req->rq_export->exp_mds_data;
473 struct mds_client_data *mcd = med->med_mcd;
474 struct mds_obd *mds = mds_req2mds(req);
475 struct mds_file_data *mfd;
476 struct obd_device *obd = req->rq_export->exp_obd;
477 struct dentry *parent, *child;
478 struct ldlm_reply *rep;
479 struct mds_body *body;
485 LASSERT(offset == 2); /* only called via intent */
486 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
487 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
489 /* copy rc, transno and disp; steal locks */
490 mds_req_from_mcd(req, mcd);
491 intent_set_disposition(rep, mcd->mcd_last_data);
493 /* Only replay if create or open actually happened. */
494 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
496 return; /* error looking up parent or child */
499 parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
500 LASSERT(!IS_ERR(parent));
502 child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
503 LASSERT(!IS_ERR(child));
505 if (!child->d_inode) {
506 GOTO(out_dput, 0); /* child not present to open */
509 /* At this point, we know we have a child. We'll send
510 * it back _unless_ it not created and open failed.
512 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
513 !intent_disposition(rep, DISP_OPEN_CREATE) &&
518 /* get lock (write for O_CREAT, read otherwise) */
520 mds_pack_inode2fid(obd, &body->fid1, child->d_inode);
521 mds_pack_inode2body(obd, body, child->d_inode);
522 if (S_ISREG(child->d_inode->i_mode)) {
523 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
527 LASSERT(rc == req->rq_status);
529 /* If we have LOV EA data, the OST holds size, mtime */
530 if (!(body->valid & OBD_MD_FLEASIZE))
531 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
532 OBD_MD_FLATIME | OBD_MD_FLMTIME);
534 /* XXX need to check this case */
537 /* If we're opening a file without an EA, change to a write
538 lock (unless we already have one). */
540 /* If we have -EEXIST as the status, and we were asked to create
541 * exclusively, we can tell we failed because the file already existed.
543 if (req->rq_status == -EEXIST &&
544 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
545 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
549 /* If we didn't get as far as trying to open, then some locking thing
550 * probably went wrong, and we'll just bail here.
552 if (!intent_disposition(rep, DISP_OPEN_OPEN))
555 /* If we failed, then we must have failed opening, so don't look for
556 * file descriptor or anything, just give the client the bad news.
562 list_for_each(t, &med->med_open_head) {
563 mfd = list_entry(t, struct mds_file_data, mfd_list);
564 if (mfd->mfd_xid == req->rq_xid)
569 /* #warning "XXX fixme" bug 2991 */
570 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
571 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
572 * to detect a re-open */
574 mntget(mds->mds_vfsmnt);
575 CERROR("Re-opened file \n");
576 mfd = mds_dentry_open(child, mds->mds_vfsmnt,
577 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
579 CERROR("mds: out of memory\n");
580 GOTO(out_dput, req->rq_status = -ENOMEM);
584 body->handle.cookie = mfd->mfd_handle.h_cookie;
585 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
586 mfd->mfd_handle.h_cookie);
596 /* do NOT or the MAY_*'s, you'll get the weakest */
597 static int accmode(int flags)
601 if (flags & FMODE_READ)
603 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
605 if (flags & FMODE_EXEC)
610 /* Handles object creation, actual opening, and I/O epoch */
611 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
612 struct mds_body *body, int flags, void **handle,
613 struct mds_update_record *rec,struct ldlm_reply *rep)
615 struct mds_obd *mds = mds_req2mds(req);
616 struct obd_device *obd = req->rq_export->exp_obd;
617 struct mds_file_data *mfd = NULL;
618 obd_id *ids = NULL; /* object IDs created */
623 /* atomically create objects if necessary */
624 down(&dchild->d_inode->i_sem);
625 mode = dchild->d_inode->i_mode;
626 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
627 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
628 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
631 up(&dchild->d_inode->i_sem);
636 /* no EA: create objects */
637 rc = mds_create_objects(req, 2, rec, mds, obd,
638 dchild, handle, &ids);
640 CERROR("mds_create_objects: rc = %d\n", rc);
641 up(&dchild->d_inode->i_sem);
644 if (S_ISREG(dchild->d_inode->i_mode) &&
645 (body->valid & OBD_MD_FLEASIZE)) {
646 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
649 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
652 up(&dchild->d_inode->i_sem);
657 /* If the inode has no EA data, then MDSs hold size, mtime */
658 if (S_ISREG(dchild->d_inode->i_mode) &&
659 !(body->valid & OBD_MD_FLEASIZE)) {
660 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
661 OBD_MD_FLATIME | OBD_MD_FLMTIME);
663 up(&dchild->d_inode->i_sem);
665 intent_set_disposition(rep, DISP_OPEN_OPEN);
666 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
668 RETURN(PTR_ERR(mfd));
670 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
671 mfd->mfd_handle.h_cookie);
673 mds_lov_update_objids(obd, ids);
674 OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
677 // mds_mfd_destroy(mfd);
681 static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
682 struct mds_body *body, int flags,
683 struct mds_update_record *rec,struct ldlm_reply *rep)
685 struct mds_obd *mds = mds_req2mds(req);
686 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
687 struct dentry *dchild;
688 char fidname[LL_FID_NAMELEN];
693 down(&pending_dir->i_sem);
694 fidlen = ll_fid2str(fidname, fid->id, fid->generation);
695 dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
696 if (IS_ERR(dchild)) {
697 up(&pending_dir->i_sem);
698 rc = PTR_ERR(dchild);
699 CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
703 if (dchild->d_inode != NULL) {
704 up(&pending_dir->i_sem);
705 mds_inode_set_orphan(dchild->d_inode);
706 mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
707 mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
708 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
709 intent_set_disposition(rep, DISP_LOOKUP_POS);
710 CWARN("Orphan %s found and opened in PENDING directory\n",
715 up(&pending_dir->i_sem);
717 /* We didn't find it in PENDING so it isn't an orphan. See
718 * if it was a regular inode that was previously created. */
719 dchild = mds_fid2dentry(mds, fid, NULL);
721 RETURN(PTR_ERR(dchild));
723 mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode);
724 mds_pack_inode2body(req2obd(req), body, dchild->d_inode);
725 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
726 intent_set_disposition(rep, DISP_LOOKUP_POS);
729 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
730 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
731 req, rc, rep ? rep->lock_policy_res1 : 0);
732 /* XXX what do we do here if mds_finish_transno itself failed? */
738 int mds_pin(struct ptlrpc_request *req)
740 struct obd_device *obd = req->rq_export->exp_obd;
741 struct mds_body *request_body, *reply_body;
742 struct lvfs_run_ctxt saved;
743 int rc, size = sizeof(*reply_body);
746 request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body));
748 rc = lustre_pack_reply(req, 1, &size, NULL);
751 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
753 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
754 rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
755 request_body->flags, NULL, NULL);
756 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
761 /* Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
762 * If child_lockh is NULL we just get the lock as a barrier to wait for
763 * other holders of this lock, and drop it right away again. */
764 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
765 struct lustre_handle *child_lockh)
767 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
768 struct lustre_handle lockh;
772 if (child_lockh == NULL)
773 child_lockh = &lockh;
775 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
776 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
777 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
778 NULL, 0, NULL, child_lockh);
780 CERROR("ldlm_cli_enqueue: %d\n", rc);
781 else if (child_lockh == &lockh)
782 ldlm_lock_decref(child_lockh, LCK_EX);
787 static int is_mount_object(struct dentry *dparent)
789 struct dentry *dchild;
791 if (!(dparent->d_inode->i_mode & S_ISUID))
794 dchild = lookup_one_len(".mntinfo", dparent, strlen(".mntinfo"));
795 if (IS_ERR(dchild) || dchild == NULL)
802 int mds_open(struct mds_update_record *rec, int offset,
803 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
805 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
806 struct obd_device *obd = req->rq_export->exp_obd;
807 struct mds_obd *mds = mds_req2mds(req);
808 struct ldlm_reply *rep = NULL;
809 struct mds_body *body = NULL;
810 struct dentry *dchild = NULL, *dparent = NULL;
811 struct mds_export_data *med;
812 struct lustre_handle parent_lockh[2];
813 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
814 int parent_mode = LCK_PR;
816 struct dentry_params dp;
817 struct mea *mea = NULL;
818 int mea_size, update_mode;
821 parent_lockh[0].cookie = 0;
822 parent_lockh[1].cookie = 0;
824 if (offset == 2) { /* intent */
825 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
826 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
827 } else if (offset == 0) { /* non-intent reint */
828 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
834 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
836 /* Step 0: If we are passed a fid, then we assume the client already
837 * opened this file and is only replaying the RPC, so we open the
838 * inode by fid (at some large expense in security). */
839 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
840 DEBUG_REQ(D_HA, req, "open replay, disp: "LPX64"\n",
841 rep->lock_policy_res1);
843 LASSERT(rec->ur_fid2->id);
845 rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
849 /* We didn't find the correct inode on disk either, so we
850 * need to re-create it via a regular replay. */
851 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
853 LASSERT(!rec->ur_fid2->id);
856 LASSERT(offset == 2); /* If we got here, we must be called via intent */
858 med = &req->rq_export->exp_mds_data;
859 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
860 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
864 acc_mode = accmode(rec->ur_flags);
866 /* Step 1: Find and lock the parent */
867 if (rec->ur_flags & O_CREAT) {
868 /* XXX Well, in fact we only need this lock mode change if
869 in addition to O_CREAT, the file does not exist.
870 But we do not know if it exists or not yet */
871 parent_mode = LCK_PW;
874 if (rec->ur_namelen == 1) {
875 /* client (LMV) wants to open the file by fid */
876 CDEBUG(D_OTHER, "OPEN by fid %u/%u/%u\n",
877 (unsigned) rec->ur_fid1->mds,
878 (unsigned) rec->ur_fid1->id,
879 (unsigned) rec->ur_fid1->generation);
880 dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL);
881 if (IS_ERR(dchild)) {
882 rc = PTR_ERR(dparent);
883 CERROR("child lookup by a fid error %d\n", rc);
889 dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
890 parent_lockh, &update_mode, rec->ur_name,
892 MDS_INODELOCK_UPDATE);
893 if (IS_ERR(dparent)) {
894 rc = PTR_ERR(dparent);
895 CERROR("parent lookup error %d\n", rc);
898 LASSERT(dparent->d_inode != NULL);
900 cleanup_phase = 1; /* parent dentry and lock */
902 /* try to retrieve MEA data for this dir */
903 rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size);
908 /* dir is already splitted, check is requested filename
909 * should live at this MDS or at another one */
911 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
912 if (mea->mea_master != mea->mea_fids[i].mds) {
913 CERROR("%s: inapropriate MDS(%d) for %lu/%u:%s."
914 " should be %d(%d)\n", obd->obd_name,
915 mea->mea_master, dparent->d_inode->i_ino,
916 dparent->d_inode->i_generation, rec->ur_name,
917 mea->mea_fids[i].mds, i);
918 GOTO(cleanup, rc = -ERESTART);
922 /* Step 2: Lookup the child */
923 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
924 if (IS_ERR(dchild)) {
925 rc = PTR_ERR(dchild);
926 dchild = NULL; /* don't confuse mds_finish_transno() below */
931 cleanup_phase = 2; /* child dentry */
933 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
935 if (dchild->d_flags & DCACHE_CROSS_REF) {
936 struct ldlm_res_id res_id = { . name = {0} };
937 ldlm_policy_data_t policy;
939 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
940 (unsigned long) dchild->d_mdsnum,
941 (unsigned long) dchild->d_inum,
942 (unsigned long) dchild->d_generation);
943 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
944 body->fid1.id = dchild->d_inum;
945 body->fid1.mds = dchild->d_mdsnum;
946 body->fid1.generation = dchild->d_generation;
947 intent_set_disposition(rep, DISP_LOOKUP_POS);
948 res_id.name[0] = dchild->d_inum;
949 res_id.name[1] = dchild->d_generation;
950 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
951 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
952 res_id, LDLM_IBITS, &policy,
955 ldlm_completion_ast, NULL, NULL,
956 NULL, 0, NULL, child_lockh);
958 if (parent_lockh[1].cookie != 0)
959 ldlm_lock_decref(parent_lockh + 1, update_mode);
961 ldlm_lock_decref(parent_lockh, parent_mode);
963 OBD_FREE(mea, mea_size);
970 intent_set_disposition(rep, DISP_LOOKUP_POS);
972 intent_set_disposition(rep, DISP_LOOKUP_NEG);
974 /*Step 3: If the child was negative, and we're supposed to, create it.*/
975 if (dchild->d_inode == NULL) {
976 unsigned long ino = rec->ur_fid2->id;
980 if ((rc = mds_try_to_split_dir(obd, dparent, &mea,
983 /* dir got splitted */
984 CERROR("%s: splitted %lu/%u - %d\n", obd->obd_name,
985 dparent->d_inode->i_ino,
986 dparent->d_inode->i_generation, rc);
987 GOTO(cleanup, rc = -ERESTART);
989 /* error happened during spitting */
994 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
995 /* It's negative and we weren't supposed to create it */
996 GOTO(cleanup, rc = -ENOENT);
999 intent_set_disposition(rep, DISP_OPEN_CREATE);
1000 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1002 if (IS_ERR(handle)) {
1003 rc = PTR_ERR(handle);
1007 dchild->d_fsdata = (void *) &dp;
1011 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
1012 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1013 dchild->d_fsdata = NULL;
1016 CDEBUG(D_INODE, "error during create: %d\n", rc);
1019 inode = dchild->d_inode;
1021 LASSERT(ino == inode->i_ino);
1022 /* Written as part of setattr */
1023 inode->i_generation = rec->ur_fid2->generation;
1024 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1025 inode->i_ino, inode->i_generation);
1029 LTIME_S(iattr.ia_atime) = LTIME_S(rec->ur_time);
1030 LTIME_S(iattr.ia_ctime) = LTIME_S(rec->ur_time);
1031 LTIME_S(iattr.ia_mtime) = LTIME_S(rec->ur_time);
1033 iattr.ia_uid = rec->_ur_fsuid;
1034 if (dparent->d_inode->i_mode & S_ISGID)
1035 iattr.ia_gid = dparent->d_inode->i_gid;
1037 iattr.ia_gid = rec->_ur_fsgid;
1039 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1040 ATTR_MTIME | ATTR_CTIME;
1042 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1044 CERROR("error on child setattr: rc = %d\n", rc);
1046 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1048 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1050 CERROR("error on parent setattr: rc = %d\n", rc);
1052 acc_mode = 0; /* Don't check for permissions */
1055 LASSERT(!mds_inode_is_orphan(dchild->d_inode));
1057 mds_pack_inode2fid(obd, &body->fid1, dchild->d_inode);
1058 mds_pack_inode2body(obd, body, dchild->d_inode);
1060 if (S_ISREG(dchild->d_inode->i_mode)) {
1061 /* Check permissions etc */
1062 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1066 /* Can't write to a read-only file */
1067 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1068 GOTO(cleanup, rc = -EPERM);
1070 /* An append-only file must be opened in append mode for
1072 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1073 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1074 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1075 GOTO(cleanup, rc = -EPERM);
1078 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1079 (rec->ur_flags & MDS_OPEN_EXCL)) {
1080 /* File already exists, we didn't just create it, and we
1081 * were passed O_EXCL; err-or. */
1082 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1085 /* if we are following a symlink, don't open */
1086 if (S_ISLNK(dchild->d_inode->i_mode))
1087 GOTO(cleanup, rc = 0);
1089 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1090 !S_ISDIR(dchild->d_inode->i_mode))
1091 GOTO(cleanup, rc = -ENOTDIR);
1093 if (S_ISDIR(dchild->d_inode->i_mode)) {
1094 if (rec->ur_flags & MDS_OPEN_CREAT ||
1095 rec->ur_flags & FMODE_WRITE) {
1096 /*we are trying to create or write a exist dir*/
1097 GOTO(cleanup, rc = -EISDIR);
1099 if (ll_permission(dchild->d_inode, acc_mode, NULL)) {
1100 GOTO(cleanup, rc = -EACCES);
1102 if (is_mount_object(dchild)) {
1103 CERROR("Found possible GNS mount object %*s; not "
1104 "opening.\n", dchild->d_name.len,
1105 dchild->d_name.name);
1106 GOTO(cleanup, rc = 0); // success, but don't really open
1110 /* Step 5: mds_open it */
1111 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
1116 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1117 req, rc, rep ? rep->lock_policy_res1 : 0);
1119 switch (cleanup_phase) {
1121 if (rc && created) {
1122 int err = vfs_unlink(dparent->d_inode, dchild);
1124 CERROR("unlink(%*s) in error path: %d\n",
1125 dchild->d_name.len, dchild->d_name.name,
1128 } else if (created) {
1129 mds_lock_new_child(obd, dchild->d_inode, NULL);
1133 if (dparent == NULL)
1138 if (parent_lockh[1].cookie != 0)
1139 ldlm_lock_decref(parent_lockh + 1, update_mode);
1142 ldlm_lock_decref(parent_lockh, parent_mode);
1144 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1147 atomic_inc(&mds->mds_open_count);
1149 OBD_FREE(mea, mea_size);
1153 /* Close a "file descriptor" and possibly unlink an orphan from the
1154 * PENDING directory.
1156 * If we are being called from mds_disconnect() because the client has
1157 * disappeared, then req == NULL and we do not update last_rcvd because
1158 * there is nothing that could be recovered by the client at this stage
1159 * (it will not even _have_ an entry in last_rcvd anymore).
1161 * Returns EAGAIN if the client needs to get more data and re-close. */
1162 int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
1163 struct mds_file_data *mfd, int unlink_orphan)
1165 struct inode *inode = mfd->mfd_dentry->d_inode;
1166 char fidname[LL_FID_NAMELEN];
1167 int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
1168 struct dentry *pending_child = NULL;
1169 struct mds_obd *mds = &obd->u.mds;
1170 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1171 void *handle = NULL;
1172 struct mds_body *request_body = NULL, *reply_body = NULL;
1173 struct dentry_params dp;
1174 struct iattr iattr = { 0 };
1177 if (req && req->rq_reqmsg != NULL)
1178 request_body = lustre_msg_buf(req->rq_reqmsg, 0,
1179 sizeof(*request_body));
1180 if (req && req->rq_repmsg != NULL)
1181 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1182 sizeof(*reply_body));
1184 fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
1186 last_orphan = mds_open_orphan_dec_test(inode) &&
1187 mds_inode_is_orphan(inode);
1189 /* this is half of the actual "close" */
1190 if (mfd->mfd_mode & FMODE_WRITE) {
1191 rc = mds_put_write_access(mds, inode, request_body,
1192 last_orphan && unlink_orphan);
1193 } else if (mfd->mfd_mode & FMODE_EXEC) {
1194 mds_allow_write_access(inode);
1197 if (last_orphan && unlink_orphan) {
1198 struct lov_mds_md *lmm = NULL;
1199 int stripe_count = 0;
1200 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1202 if (obd->obd_recovering) {
1203 CDEBUG(D_HA, "not remove orphan %s until recovery"
1204 " is over\n", fidname);
1208 CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
1210 /* Sadly, there is no easy way to save pending_child from
1211 * mds_reint_unlink() into mfd, so we need to re-lookup,
1212 * but normally it will still be in the dcache. */
1213 pending_dir = mds->mds_pending_dir->d_inode;
1214 down(&pending_dir->i_sem);
1215 cleanup_phase = 1; /* up(i_sem) when finished */
1216 pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
1218 if (IS_ERR(pending_child))
1219 GOTO(cleanup, rc = PTR_ERR(pending_child));
1220 LASSERT(pending_child->d_inode != NULL);
1222 cleanup_phase = 2; /* dput(pending_child) when finished */
1223 if (req != NULL && req->rq_repmsg != NULL) {
1224 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1225 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1228 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1229 NULL, stripe_count);
1230 if (IS_ERR(handle)) {
1231 rc = PTR_ERR(handle);
1236 if (req != NULL && req->rq_repmsg != NULL &&
1237 (reply_body->valid & OBD_MD_FLEASIZE) &&
1238 mds_log_op_unlink(obd, pending_child->d_inode, lmm,
1239 req->rq_repmsg->buflens[1],
1240 lustre_msg_buf(req->rq_repmsg, 2, 0),
1241 req->rq_repmsg->buflens[2]) > 0) {
1242 reply_body->valid |= OBD_MD_FLCOOKIE;
1245 pending_child->d_fsdata = (void *) &dp;
1248 if (S_ISDIR(pending_child->d_inode->i_mode))
1249 rc = vfs_rmdir(pending_dir, pending_child);
1251 rc = vfs_unlink(pending_dir, pending_child);
1253 CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
1255 goto out; /* Don't bother updating attrs on unlinked inode */
1258 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1259 /* Update the on-disk attributes if this was the last write
1260 * close, and all information was provided (i.e., rc == 0)
1262 * XXX this should probably be abstracted with mds_reint_setattr
1266 if (request_body->valid & OBD_MD_FLMTIME &&
1267 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1268 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1269 iattr.ia_valid |= ATTR_MTIME;
1271 if (request_body->valid & OBD_MD_FLCTIME &&
1272 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1273 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1274 iattr.ia_valid |= ATTR_CTIME;
1277 /* XXX can't set block count with fsfilt_setattr (!) */
1278 if (request_body->valid & OBD_MD_FLSIZE) {
1279 iattr.ia_valid |= ATTR_SIZE;
1280 iattr.ia_size = request_body->size;
1282 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1283 iattr.ia_valid |= ATTR_BLOCKS;
1284 iattr.ia_blocks = request_body->blocks
1290 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1291 /* Only start a transaction to write out only the atime if
1292 * it is more out-of-date than the specified limit. If we
1293 * are already going to write out the atime then do it anyway.
1295 if ((LTIME_S(request_body->atime) >
1296 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1297 (iattr.ia_valid != 0 &&
1298 LTIME_S(request_body->atime) > LTIME_S(inode->i_atime))) {
1299 LTIME_S(iattr.ia_atime) = LTIME_S(request_body->atime);
1300 iattr.ia_valid |= ATTR_ATIME;
1304 if (iattr.ia_valid != 0) {
1305 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1307 GOTO(cleanup, rc = PTR_ERR(handle));
1308 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1310 CERROR("error in setattr(%s): rc %d\n", fidname, rc);
1313 /* If other clients have this file open for write, rc will be > 0 */
1316 l_dput(mfd->mfd_dentry);
1317 mds_mfd_destroy(mfd);
1320 atomic_dec(&mds->mds_open_count);
1321 if (req != NULL && reply_body != NULL) {
1322 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1323 } else if (handle) {
1324 int err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, 0);
1326 CERROR("error committing close: %d\n", err);
1332 switch (cleanup_phase) {
1334 dput(pending_child);
1336 up(&pending_dir->i_sem);
1341 int mds_close(struct ptlrpc_request *req)
1343 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1344 struct obd_device *obd = req->rq_export->exp_obd;
1345 struct mds_body *body;
1346 struct mds_file_data *mfd;
1347 struct lvfs_run_ctxt saved;
1348 struct inode *inode;
1349 int rc, repsize[3] = {sizeof(struct mds_body),
1350 obd->u.mds.mds_max_mdsize,
1351 obd->u.mds.mds_max_cookiesize};
1354 rc = lustre_pack_reply(req, 3, repsize, NULL);
1356 CERROR("lustre_pack_reply: rc = %d\n", rc);
1357 req->rq_status = rc;
1359 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1362 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1364 CERROR("Can't unpack body\n");
1365 req->rq_status = -EFAULT;
1369 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1370 /* do some stuff */ ;
1372 mfd = mds_handle2mfd(&body->handle);
1374 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1375 ": cookie "LPX64, body->fid1.id, body->handle.cookie);
1376 req->rq_status = -ESTALE;
1380 inode = mfd->mfd_dentry->d_inode;
1381 if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
1382 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1383 LASSERT(body != NULL);
1385 mds_pack_inode2fid(obd, &body->fid1, inode);
1386 mds_pack_inode2body(obd, body, inode);
1387 mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
1389 spin_lock(&med->med_open_lock);
1390 list_del(&mfd->mfd_list);
1391 spin_unlock(&med->med_open_lock);
1393 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1394 req->rq_status = mds_mfd_close(req, obd, mfd, 1);
1395 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1397 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1398 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1399 req->rq_status = -ENOMEM;
1408 int mds_done_writing(struct ptlrpc_request *req)
1410 struct mds_body *body;
1411 int rc, size = sizeof(struct mds_body);
1414 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1416 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1418 CERROR("Can't unpack body\n");
1419 req->rq_status = -EFAULT;
1423 rc = lustre_pack_reply(req, 1, &size, NULL);
1425 CERROR("lustre_pack_reply: rc = %d\n", rc);
1426 req->rq_status = rc;