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/obd_lov.h>
44 #include <linux/lustre_fsfilt.h>
45 #include <linux/lprocfs_status.h>
46 #include <linux/lustre_gs.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 struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
93 LASSERT(handle != NULL);
94 RETURN(class_handle2object(handle->cookie));
97 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);
115 /* Caller must hold mds->mds_epoch_sem */
116 static int mds_alloc_filterdata(struct inode *inode)
118 LASSERT(LUSTRE_FILTERDATA(inode) == NULL);
119 OBD_ALLOC(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data));
120 if (LUSTRE_FILTERDATA(inode) == 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(LUSTRE_FILTERDATA(inode) != NULL);
130 OBD_FREE(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data));
131 LUSTRE_FILTERDATA(inode) = 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);
158 if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
159 CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
160 MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
161 inode->i_generation);
165 if (MDS_FILTERDATA(inode) == NULL)
166 mds_alloc_filterdata(inode);
167 if (MDS_FILTERDATA(inode) == NULL) {
171 if (epoch > mds->mds_io_epoch) {
172 mds->mds_io_epoch = epoch;
173 CDEBUG(D_INODE, "repair MDS epoch "LPU64" for ino %lu/%u\n",
174 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
177 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
178 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
180 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
183 atomic_inc(&inode->i_writecount);
184 up(&mds->mds_epoch_sem);
188 /* Returns EAGAIN if the client needs to get size and/or cookies and close
189 * again -- which is never true if the file is about to be unlinked. Otherwise
190 * returns the number of remaining writers. */
191 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
192 struct mds_body *body, int unlinking)
197 down(&mds->mds_epoch_sem);
198 atomic_dec(&inode->i_writecount);
199 rc = atomic_read(&inode->i_writecount);
203 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
204 GOTO(out, rc = EAGAIN);
206 mds_free_filterdata(inode);
208 up(&mds->mds_epoch_sem);
212 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
215 down(&mds->mds_epoch_sem);
216 if (atomic_read(&inode->i_writecount) > 0) {
217 up(&mds->mds_epoch_sem);
220 atomic_dec(&inode->i_writecount);
221 up(&mds->mds_epoch_sem);
225 static void mds_allow_write_access(struct inode *inode)
228 atomic_inc(&inode->i_writecount);
231 int mds_query_write_access(struct inode *inode)
234 RETURN(atomic_read(&inode->i_writecount));
237 /* This replaces the VFS dentry_open, it manages mfd and writecount */
238 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
239 struct vfsmount *mnt, int flags,
240 struct ptlrpc_request *req,
241 struct mds_update_record *rec)
243 struct mds_export_data *med = &req->rq_export->exp_mds_data;
244 struct mds_obd *mds = mds_req2mds(req);
245 struct mds_file_data *mfd;
246 struct mds_body *body;
252 CERROR("mds: out of memory\n");
253 GOTO(cleanup_dentry, rc = -ENOMEM);
256 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
258 if (flags & FMODE_WRITE) {
259 /* FIXME: in recovery, need to pass old epoch here */
260 rc = mds_get_write_access(mds, dentry->d_inode, rec->ur_ioepoch);
262 GOTO(cleanup_mfd, rc);
263 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
264 } else if (flags & FMODE_EXEC) {
265 rc = mds_deny_write_access(mds, dentry->d_inode);
267 GOTO(cleanup_mfd, rc);
272 /* FIXME: invalidate update locks when first open for write comes in */
274 /* mark the file as open to handle open-unlink. */
275 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
276 mds_orphan_open_inc(dentry->d_inode);
277 UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
279 mfd->mfd_mode = flags;
280 mfd->mfd_dentry = dentry;
281 mfd->mfd_xid = req->rq_xid;
283 spin_lock(&med->med_open_lock);
284 list_add(&mfd->mfd_list, &med->med_open_head);
285 spin_unlock(&med->med_open_lock);
288 body->handle.cookie = mfd->mfd_handle.h_cookie;
292 mds_mfd_destroy(mfd);
298 mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
299 struct lov_desc *desc)
303 for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
304 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
305 le64_to_cpu(lmm->lmm_objects[i].l_object_id);
309 /* must be called with i_sem held */
311 mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req,
312 int offset, struct mds_update_record *rec,
313 struct dentry *dchild, void **handle,
316 struct inode *inode = dchild->d_inode;
317 struct mds_obd *mds = &obd->u.mds;
318 struct obd_trans_info oti = { 0 };
319 struct lov_stripe_md *lsm = NULL;
320 struct lov_mds_md *lmm = NULL;
321 int rc, lmm_bufsize, lmm_size;
322 struct obdo *oa = NULL;
323 struct mds_body *body;
327 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
328 !(rec->ur_flags & FMODE_WRITE))
331 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
333 if (!S_ISREG(inode->i_mode))
335 if (body->valid & OBD_MD_FLEASIZE)
338 OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
341 oti.oti_objid = *ids;
344 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
345 LASSERT(id_ino(rec->ur_id2));
346 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
347 lmm_size = rec->ur_eadatalen;
348 lmm = rec->ur_eadata;
352 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
353 if (IS_ERR(*handle)) {
354 rc = PTR_ERR(*handle);
359 mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
361 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
362 lmm_bufsize = req->rq_repmsg->buflens[offset];
363 LASSERT(lmm_buf != NULL);
364 LASSERT(lmm_bufsize >= lmm_size);
366 memcpy(lmm_buf, lmm, lmm_size);
367 rc = fsfilt_set_md(obd, inode, *handle, lmm,
370 CERROR("open replay failed to set md:%d\n", rc);
374 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
375 GOTO(out_ids, rc = -ENOMEM);
380 oa->o_mode = S_IFREG | 0600;
381 oa->o_id = inode->i_ino;
382 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
383 oa->o_generation = inode->i_generation;
384 oa->o_uid = 0; /* must have 0 uid / gid on OST */
386 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE |
387 OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
390 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
391 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
393 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
394 /* check if things like lfs setstripe are sending us the ea */
395 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
396 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
398 0, &lsm, rec->ur_eadata);
402 OBD_ALLOC(lmm, mds->mds_max_mdsize);
404 GOTO(out_oa, rc = -ENOMEM);
406 lmm_size = mds->mds_max_mdsize;
407 rc = mds_get_md(obd, dchild->d_parent->d_inode,
408 lmm, &lmm_size, 1, 0);
410 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
413 OBD_FREE(lmm, mds->mds_max_mdsize);
418 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
419 oti.oti_flags |= OBD_MODE_CROW;
420 rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
423 CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR),
424 "error creating objects for "
425 "inode %lu: rc = %d\n",
428 CERROR("obd_create returned invalid "
435 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
436 0, &lsm, rec->ur_eadata);
440 lsm->lsm_object_id = oa->o_id;
441 lsm->lsm_object_gr = oa->o_gr;
444 oa->o_size = inode->i_size;
445 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
446 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE);
448 /* pack lustre id to oss */
449 *(obdo_id(oa)) = body->id1;
450 oa->o_valid |= OBD_MD_FLIFID;
452 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti, NULL);
454 CERROR("error setting attrs for inode %lu: rc %d\n",
457 CERROR("obd_setattr returned bad rc %d\n", rc);
464 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
465 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
467 LASSERT(lsm && lsm->lsm_object_id);
469 rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
470 if (!id_ino(rec->ur_id2))
471 obd_free_memmd(mds->mds_dt_exp, &lsm);
473 CERROR("cannot pack lsm, err = %d\n", rc);
478 body->eadatasize = rc;
481 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
482 if (IS_ERR(*handle)) {
483 rc = PTR_ERR(*handle);
488 rc = fsfilt_set_md(obd, inode, *handle, lmm,
491 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
492 lmm_bufsize = req->rq_repmsg->buflens[offset];
494 LASSERT(lmm_bufsize >= lmm_size);
496 memcpy(lmm_buf, lmm, lmm_size);
497 obd_free_diskmd(mds->mds_dt_exp, &lmm);
499 oti_free_cookies(&oti);
503 OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
507 obd_free_memmd(mds->mds_dt_exp, &lsm);
512 mds_destroy_object(struct obd_device *obd,
513 struct inode *inode, int async)
515 struct mds_obd *mds = &obd->u.mds;
516 struct lov_mds_md *lmm = NULL;
520 LASSERT(inode != NULL);
522 if (inode->i_nlink != 0) {
523 CDEBUG(D_INODE, "attempt to destroy OSS object when "
524 "i_nlink == %d\n", (int)inode->i_nlink);
528 OBD_ALLOC(lmm, mds->mds_max_mdsize);
532 lmm_size = mds->mds_max_mdsize;
533 rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
535 CERROR("no stripe info for %lu/%lu inode\n",
536 (unsigned long)inode->i_ino,
537 (unsigned long)inode->i_generation);
538 GOTO(out_free_lmm, rc);
542 /* asynchronously unlink objecect on OSS */
543 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
546 CERROR("error unlinking object on OSS, "
548 GOTO(out_free_lmm, rc);
551 CDEBUG(D_INODE, "no stripping info found for inode "
552 "%lu/%lu\n", (unsigned long)inode->i_ino,
553 (unsigned long)inode->i_generation);
557 OBD_FREE(lmm, mds->mds_max_mdsize);
561 static void reconstruct_open(struct mds_update_record *rec, int offset,
562 struct ptlrpc_request *req,
563 struct lustre_handle *child_lockh)
565 struct mds_export_data *med = &req->rq_export->exp_mds_data;
566 struct mds_client_data *mcd = med->med_mcd;
567 struct mds_obd *mds = mds_req2mds(req);
568 struct mds_file_data *mfd;
569 struct obd_device *obd = req->rq_export->exp_obd;
570 struct dentry *parent = NULL, *dchild;
571 struct ldlm_reply *rep;
572 struct mds_body *body;
578 LASSERT(offset == 3); /* only called via intent */
579 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
580 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
582 /* copy rc, transno and disp; steal locks */
583 mds_req_from_mcd(req, mcd);
584 intent_set_disposition(rep, mcd->mcd_last_data);
586 /* Only replay if create or open actually happened. */
587 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
589 return; /* error looking up parent or child */
592 /* first, we try to open the file by fid. by the time of this
593 * request, inode can be an orphan and parent can disappear */
594 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
595 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT|REPLAY)\n",
597 dchild = mds_id2dentry(obd, rec->ur_id2, NULL);
598 } else if (rec->ur_namelen == 1) {
599 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
601 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
603 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
604 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
605 OLID4(rec->ur_id1), PTR_ERR(parent));
606 dchild = ll_lookup_one_len(rec->ur_name, parent,
607 rec->ur_namelen - 1);
608 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
609 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
612 if (!dchild->d_inode)
613 GOTO(out_dput, 0); /* child not present to open */
615 /* At this point, we know we have a child. We'll send
616 * it back _unless_ it not created and open failed. */
617 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
618 !intent_disposition(rep, DISP_OPEN_CREATE) &&
622 /* get lock (write for O_CREAT, read otherwise) */
623 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
624 if (S_ISREG(dchild->d_inode->i_mode)) {
625 struct lustre_capa capa = {
626 .lc_uid = rec->ur_uc.luc_uid,
627 .lc_op = capa_op(rec->ur_flags),
628 .lc_ino = dchild->d_inode->i_ino,
629 .lc_igen = dchild->d_inode->i_generation,
630 .lc_mdsid = mds->mds_num,
632 int off = 7; /* capa offset */
634 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
635 dchild->d_inode, 1, 0);
638 LASSERT(rc == req->rq_status);
640 /* If we have LOV EA data, the OST holds size, mtime */
641 if (!(body->valid & OBD_MD_FLEASIZE))
642 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
643 OBD_MD_FLATIME | OBD_MD_FLMTIME);
644 rc = mds_pack_capa(obd, med, NULL, &capa, req, &off, body);
647 /* If we have -EEXIST as the status, and we were asked to create
648 * exclusively, we can tell we failed because the file already existed.
650 if (req->rq_status == -EEXIST &&
651 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
652 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
656 /* If we didn't get as far as trying to open, then some locking thing
657 * probably went wrong, and we'll just bail here.
659 if (!intent_disposition(rep, DISP_OPEN_OPEN))
662 /* If we failed, then we must have failed opening, so don't look for
663 * file descriptor or anything, just give the client the bad news.
669 list_for_each(t, &med->med_open_head) {
670 mfd = list_entry(t, struct mds_file_data, mfd_list);
671 if (mfd->mfd_xid == req->rq_xid)
676 /* #warning "XXX fixme" bug 2991 */
677 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
678 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
679 * to detect a re-open */
681 mntget(mds->mds_vfsmnt);
682 CERROR("Re-opened file \n");
683 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
684 rec->ur_flags & ~MDS_OPEN_TRUNC, req, rec);
686 CERROR("mds: out of memory\n");
687 GOTO(out_dput, req->rq_status = -ENOMEM);
691 body->handle.cookie = mfd->mfd_handle.h_cookie;
692 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
693 mfd->mfd_handle.h_cookie);
704 /* do NOT or the MAY_*'s, you'll get the weakest */
705 int accmode(int flags)
709 if (flags & FMODE_READ)
711 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
713 if (flags & FMODE_EXEC)
718 /* Handles object creation, actual opening, and I/O epoch */
719 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
720 struct mds_body *body, int flags, void **handle,
721 struct mds_update_record *rec, struct ldlm_reply *rep)
723 struct obd_device *obd = req->rq_export->exp_obd;
724 struct mds_obd *mds = mds_req2mds(req);
725 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
726 struct mds_file_data *mfd = NULL;
729 int rc = 0, reply_off;
732 /* atomically create objects if necessary */
733 down(&dchild->d_inode->i_sem);
734 mode = dchild->d_inode->i_mode;
736 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
737 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
738 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
739 dchild->d_inode, 0, 0);
741 up(&dchild->d_inode->i_sem);
747 if ((body->valid & OBD_MD_FLEASIZE) &&
748 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
749 up(&dchild->d_inode->i_sem);
753 if (!(body->valid & OBD_MD_FLEASIZE)) {
754 /* no EA: create objects */
755 rc = mds_create_objects(obd, req, 2, rec,
756 dchild, handle, &ids);
758 CERROR("mds_create_object: rc = %d\n", rc);
759 up(&dchild->d_inode->i_sem);
764 if (S_ISREG(dchild->d_inode->i_mode) &&
765 (body->valid & OBD_MD_FLEASIZE)) {
766 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
769 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
770 dchild->d_inode, 0, 0);
772 up(&dchild->d_inode->i_sem);
779 rc = mds_pack_acl(req, reply_off, body, dchild->d_inode);
783 CERROR("pack posix acl: rc = %d\n", rc);
784 up(&dchild->d_inode->i_sem);
788 rc = mds_pack_gskey(obd, req->rq_repmsg, &reply_off, body,
791 CERROR("mds_pack_gskey: rc = %d\n", rc);
792 up(&dchild->d_inode->i_sem);
797 struct lustre_capa capa = {
798 .lc_uid = rec->ur_uc.luc_uid,
799 .lc_op = capa_op(rec->ur_flags),
800 .lc_ino = dchild->d_inode->i_ino,
801 .lc_igen = dchild->d_inode->i_generation,
802 .lc_mdsid = mds->mds_num,
805 rc = mds_pack_capa(obd, med, NULL, &capa, req,
808 CERROR("mds_pack_capa: rc = %d\n", rc);
809 up(&dchild->d_inode->i_sem);
816 /* If the inode has no EA data, then MDSs hold size, mtime */
817 if (S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) {
818 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
819 OBD_MD_FLATIME | OBD_MD_FLMTIME);
822 up(&dchild->d_inode->i_sem);
824 intent_set_disposition(rep, DISP_OPEN_OPEN);
825 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req, rec);
827 RETURN(PTR_ERR(mfd));
829 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
830 mfd->mfd_handle.h_cookie);
833 mds_dt_update_objids(obd, ids);
834 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
840 static int mds_open_by_id(struct ptlrpc_request *req,
841 struct lustre_id *id,
842 struct mds_body *body, int flags,
843 struct mds_update_record *rec,
844 struct ldlm_reply *rep)
846 struct mds_obd *mds = mds_req2mds(req);
847 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
848 struct dentry *dchild;
849 char idname[LL_ID_NAMELEN];
854 down(&pending_dir->i_sem);
855 idlen = ll_id2str(idname, id_ino(id), id_gen(id));
856 dchild = lookup_one_len(idname, mds->mds_pending_dir,
858 if (IS_ERR(dchild)) {
859 up(&pending_dir->i_sem);
860 rc = PTR_ERR(dchild);
861 CERROR("error looking up %s in PENDING: rc = %d\n",
866 up(&pending_dir->i_sem);
867 if (dchild->d_inode != NULL) {
868 mds_inode_set_orphan(dchild->d_inode);
869 mds_pack_inode2body(req2obd(req), body,
872 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
873 intent_set_disposition(rep, DISP_LOOKUP_POS);
874 CWARN("Orphan %s found and opened in PENDING directory\n",
880 /* we didn't find it in PENDING so it isn't an orphan. See if it was a
881 * regular inode that was previously created. */
882 dchild = mds_id2dentry(req2obd(req), id, NULL);
884 RETURN(PTR_ERR(dchild));
886 mds_pack_inode2body(req2obd(req), body,
889 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
890 intent_set_disposition(rep, DISP_LOOKUP_POS);
893 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
894 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
895 req, rc, rep ? rep->lock_policy_res1 : 0);
896 /* XXX what do we do here if mds_finish_transno itself failed? */
901 int mds_pin(struct ptlrpc_request *req, int offset)
903 struct obd_device *obd = req->rq_export->exp_obd;
904 struct mds_body *request_body, *reply_body;
905 struct lvfs_run_ctxt saved;
906 int rc, size = sizeof(*reply_body);
909 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
910 sizeof(*request_body));
912 rc = lustre_pack_reply(req, 1, &size, NULL);
915 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
917 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
918 rc = mds_open_by_id(req, &request_body->id1, reply_body,
919 request_body->flags, NULL, NULL);
920 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
926 * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
927 * child_lockh is NULL we just get the lock as a barrier to wait for other
928 * holders of this lock, and drop it right away again.
930 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
931 struct lustre_handle *child_lockh)
933 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
934 struct lustre_handle lockh;
935 int lock_flags = LDLM_FL_ATOMIC_CB;
939 if (child_lockh == NULL)
940 child_lockh = &lockh;
942 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
943 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
944 mds_blocking_ast, ldlm_completion_ast, NULL,
945 NULL, NULL, 0, NULL, child_lockh);
947 CERROR("ldlm_cli_enqueue: %d\n", rc);
948 else if (child_lockh == &lockh)
949 ldlm_lock_decref(child_lockh, LCK_EX);
954 int mds_open(struct mds_update_record *rec, int offset,
955 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
957 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
958 struct obd_device *obd = req->rq_export->exp_obd;
959 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
960 struct mds_obd *mds = mds_req2mds(req);
961 struct ldlm_reply *rep = NULL;
962 struct mds_body *body = NULL;
963 struct dentry *dchild = NULL, *dparent = NULL;
964 struct lustre_handle parent_lockh[2] = {{0}, {0}};
965 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
966 int parent_mode = LCK_PR;
968 struct dentry_params dp;
969 struct mea *mea = NULL;
970 int mea_size, update_mode;
971 int child_mode = LCK_PR;
972 struct lustre_id sid;
976 DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
977 OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
980 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
981 (obd_timeout + 1) / 4);
983 if (offset == 3) { /* intent */
984 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
985 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
986 } else if (offset == 1) { /* non-intent reint */
987 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
993 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
996 * Step 0: If we are passed a id, then we assume the client already
997 * opened this file and is only replaying the RPC, so we open the inode
998 * by fid (at some large expense in security).
1000 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1001 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
1002 OLID4(rec->ur_id2), rec->ur_namelen - 1,
1003 rec->ur_name, rec->ur_mode);
1005 LASSERT(id_ino(rec->ur_id2));
1007 rc = mds_open_by_id(req, rec->ur_id2, body,
1008 rec->ur_flags, rec, rep);
1009 if (rc != -ENOENT) {
1010 mds_body_do_reverse_map(med, body);
1014 /* We didn't find the correct inode on disk either, so we
1015 * need to re-create it via a regular replay. */
1016 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
1018 LASSERT(!id_ino(rec->ur_id2));
1021 LASSERT(offset == 3); /* If we got here, we must be called via intent */
1023 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
1024 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
1028 acc_mode = accmode(rec->ur_flags);
1030 /* Step 1: Find and lock the parent */
1031 if (rec->ur_flags & MDS_OPEN_CREAT) {
1032 /* XXX Well, in fact we only need this lock mode change if
1033 in addition to O_CREAT, the file does not exist.
1034 But we do not know if it exists or not yet */
1035 parent_mode = LCK_PW;
1038 if (rec->ur_namelen == 1) {
1039 /* client (LMV) wants to open the file by id */
1040 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
1041 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
1042 if (IS_ERR(dchild)) {
1043 rc = PTR_ERR(dparent);
1044 CERROR("child lookup by an id error %d\n", rc);
1051 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
1052 parent_lockh, &update_mode, rec->ur_name,
1053 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
1054 if (IS_ERR(dparent)) {
1055 rc = PTR_ERR(dparent);
1056 if (rc != -ENOENT) {
1057 CERROR("parent lookup for "DLID4" failed, error %d\n",
1058 OLID4(rec->ur_id1), rc);
1060 /* Just cannot find parent - make it look like
1061 usual negative lookup to avoid extra MDS RPC */
1062 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1063 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1067 LASSERT(dparent->d_inode != NULL);
1069 cleanup_phase = 1; /* parent dentry and lock */
1071 /* try to retrieve MEA data for this dir */
1072 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1076 if (mea != NULL && mea->mea_count) {
1078 * dir is already splitted, check is requested filename should *
1079 * live at this MDS or at another one.
1082 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1083 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1085 "%s: inapropriate MDS(%d) for %lu/%u:%s."
1086 " should be %lu(%d)\n", obd->obd_name,
1087 mea->mea_master, dparent->d_inode->i_ino,
1088 dparent->d_inode->i_generation, rec->ur_name,
1089 (unsigned long)id_group(&mea->mea_ids[i]), i);
1090 GOTO(cleanup, rc = -ERESTART);
1094 /* Step 2: Lookup the child */
1095 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1096 if (IS_ERR(dchild)) {
1097 rc = PTR_ERR(dchild);
1098 dchild = NULL; /* don't confuse mds_finish_transno() below */
1103 cleanup_phase = 2; /* child dentry */
1105 if (dchild->d_flags & DCACHE_CROSS_REF) {
1106 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1107 LASSERT(rec->ur_namelen > 1);
1109 /* we're gonna acquire LOOKUP lock on the child,
1110 * but we have already locked parent and our order
1111 * may conflict with enqueue_order_locks(). so,
1112 * drop parent lock and acquire both the locks in
1113 * common order. bug 6190 */
1115 if (parent_lockh[1].cookie != 0)
1116 ldlm_lock_decref(parent_lockh + 1, update_mode);
1118 ldlm_lock_decref(parent_lockh, parent_mode);
1123 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1124 parent_lockh, &dparent,
1125 LCK_PR, MDS_INODELOCK_UPDATE,
1126 &update_mode, rec->ur_name,
1127 rec->ur_namelen, child_lockh,
1129 MDS_INODELOCK_LOOKUP);
1135 if (parent_lockh[1].cookie != 0)
1136 ldlm_lock_decref(parent_lockh + 1, update_mode);
1138 ldlm_lock_decref(parent_lockh, LCK_PR);
1140 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1141 CDEBUG(D_OTHER, "race: name changed (%p)\n",
1143 if (dchild->d_inode)
1144 ldlm_lock_decref(child_lockh, LCK_PR);
1150 mds_pack_dentry2body(obd, body, dchild, 1);
1151 intent_set_disposition(rep, DISP_LOOKUP_POS);
1152 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1155 OBD_FREE(mea, mea_size);
1161 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1163 if (dchild->d_inode)
1164 intent_set_disposition(rep, DISP_LOOKUP_POS);
1166 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1168 /* Step 3: If the child was negative, and we're supposed to, create it.*/
1169 if (dchild->d_inode == NULL) {
1172 struct inode *inode;
1174 /* get parent id: ldlm lock on the parent protects ea */
1175 rc = mds_read_inode_sid(obd, dparent->d_inode, &sid);
1177 CERROR("can't read parent inode id. ino(%lu) rc(%d)\n",
1178 dparent->d_inode->i_ino, rc);
1182 ino = (unsigned long)id_ino(rec->ur_id2);
1184 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1187 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1188 obd->obd_name, dparent->d_inode->i_ino,
1189 dparent->d_inode->i_generation, rc);
1192 /* dir got splitted */
1193 GOTO(cleanup, rc = -ERESTART);
1194 } else if (rc < 0) {
1195 /* error happened during spitting */
1199 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1201 * dentry is negative and we weren't supposed to create
1202 * object, get out of here.
1204 GOTO(cleanup, rc = -ENOENT);
1207 intent_set_disposition(rep, DISP_OPEN_CREATE);
1208 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1210 if (IS_ERR(handle)) {
1211 rc = PTR_ERR(handle);
1215 if (id_fid(rec->ur_id2))
1216 fid = id_fid(rec->ur_id2);
1218 fid = mds_alloc_fid(obd);
1220 dchild->d_fsdata = (void *) &dp;
1225 dp.p_group = mds->mds_num;
1227 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1228 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1229 dchild->d_fsdata = NULL;
1232 CDEBUG(D_INODE, "error during create: %d\n", rc);
1236 inode = dchild->d_inode;
1239 LASSERT(ino == inode->i_ino);
1241 /* written as part of setattr */
1242 inode->i_generation = id_gen(rec->ur_id2);
1243 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1244 inode->i_ino, inode->i_generation);
1248 LTIME_S(iattr.ia_atime) = rec->ur_time;
1249 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1250 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1252 iattr.ia_uid = rec->ur_fsuid;
1253 if (dparent->d_inode->i_mode & S_ISGID)
1254 iattr.ia_gid = dparent->d_inode->i_gid;
1256 iattr.ia_gid = rec->ur_fsgid;
1258 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1259 ATTR_MTIME | ATTR_CTIME;
1261 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1263 CERROR("error on child setattr: rc = %d\n", rc);
1265 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1267 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1269 CERROR("error on parent setattr: rc = %d\n", rc);
1271 MD_COUNTER_INCREMENT(obd, create);
1274 mds_inode2id(obd, &body->id1, inode, fid);
1275 mds_update_inode_ids(obd, dchild->d_inode, handle, &body->id1, &sid);
1277 if ((rec->ur_flags & MDS_OPEN_HAS_KEY)) {
1278 rc = mds_set_gskey(obd, handle, dchild->d_inode,
1279 rec->ur_ea2data, rec->ur_ea2datalen,
1280 ATTR_KEY | ATTR_MAC);
1282 CERROR("error in set gs key rc %d\n", rc);
1286 rc = fsfilt_commit(obd, dchild->d_inode->i_sb, dchild->d_inode, handle, 0);
1288 acc_mode = 0; /* Don't check for permissions */
1290 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1292 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1293 "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1294 dchild->d_name.name, dchild, dchild->d_inode);
1296 if (S_ISREG(dchild->d_inode->i_mode)) {
1297 /* Check permissions etc */
1298 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1302 /* Can't write to a read-only file */
1303 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1304 GOTO(cleanup, rc = -EPERM);
1306 /* An append-only file must be opened in append mode for
1308 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1309 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1310 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1311 GOTO(cleanup, rc = -EPERM);
1314 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1315 (rec->ur_flags & MDS_OPEN_EXCL)) {
1316 /* File already exists, we didn't just create it, and we
1317 * were passed O_EXCL; err-or. */
1318 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1321 if (S_ISDIR(dchild->d_inode->i_mode)) {
1322 if (rec->ur_flags & MDS_OPEN_CREAT ||
1323 rec->ur_flags & FMODE_WRITE) {
1324 /* we are trying to create or write a exist dir */
1325 GOTO(cleanup, rc = -EISDIR);
1327 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1328 GOTO(cleanup, rc = -EACCES);
1331 * here was checking for possible GNS mount points to skip their
1332 * open. I removed it as its detection based only on SUID bit is
1333 * not reliable. Opening such a dirs should not cause any
1334 * problems, at least tests show that. --umka
1338 /* if we are following a symlink, don't open */
1339 if (S_ISLNK(dchild->d_inode->i_mode))
1340 GOTO(cleanup_no_trans, rc = 0);
1342 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1343 !S_ISDIR(dchild->d_inode->i_mode))
1344 GOTO(cleanup, rc = -ENOTDIR);
1346 /* check permission even it's special files */
1347 if (S_ISCHR(dchild->d_inode->i_mode) ||
1348 S_ISBLK(dchild->d_inode->i_mode) ||
1349 S_ISFIFO(dchild->d_inode->i_mode) ||
1350 S_ISSOCK(dchild->d_inode->i_mode)) {
1351 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1356 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1357 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1358 GOTO(cleanup, rc = -EAGAIN);
1362 /* We cannot use acc_mode here, because it is zeroed in case of
1363 creating a file, so we get wrong lockmode */
1364 if (accmode(rec->ur_flags) & MAY_WRITE)
1365 child_mode = LCK_CW;
1366 else if (accmode(rec->ur_flags) & MAY_EXEC)
1367 child_mode = LCK_PR;
1369 child_mode = LCK_CR;
1371 /* Always returning LOOKUP lock if open succesful to guard
1372 * dentry on client. In case of replay we do not get a lock
1373 * assuming that the caller has it already */
1374 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1375 struct ldlm_res_id child_res_id = { .name = {0}};
1376 ldlm_policy_data_t policy;
1377 int lock_flags = LDLM_FL_ATOMIC_CB;
1379 /* LOOKUP lock will protect dentry on client -bzzz */
1380 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1383 child_res_id.name[0] = id_fid(&body->id1);
1384 child_res_id.name[1] = id_group(&body->id1);
1386 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1387 child_res_id, LDLM_IBITS, &policy,
1388 child_mode, &lock_flags,
1389 mds_blocking_ast, ldlm_completion_ast,
1390 NULL, NULL, NULL, 0, NULL, child_lockh);
1397 /* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */
1400 /* Step 5: mds_open it */
1401 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1406 /* reintegration case */
1407 if ((rec->ur_flags & MDS_REINT_REQ)) {
1408 rc = mds_fidmap_add(obd, &body->id1);
1410 CERROR("can't create fid->ino mapping, err %d\n",
1417 /* if this is a writer, we have to invalidate client's
1418 * update locks in order to make sure they don't use
1419 * isize/iblocks from mds anymore.
1420 * FIXME: can cause a deadlock, use mds_get_parent_child_locked()
1421 * XXX: optimization is to do this for first writer only */
1422 if (accmode(rec->ur_flags) & MAY_WRITE) {
1423 struct ldlm_res_id child_res_id = { .name = {0}};
1424 ldlm_policy_data_t sz_policy;
1425 struct lustre_handle sz_lockh;
1426 int lock_flags = LDLM_FL_ATOMIC_CB;
1428 child_res_id.name[0] = id_fid(&body->id1);
1429 child_res_id.name[1] = id_group(&body->id1);
1430 sz_policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1432 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1433 child_res_id, LDLM_IBITS, &sz_policy,
1434 LCK_PW, &lock_flags, mds_blocking_ast,
1435 ldlm_completion_ast, NULL, NULL, NULL,
1436 0, NULL, &sz_lockh);
1438 ldlm_lock_decref(&sz_lockh, LCK_PW);
1440 CERROR("can't invalidate client's update locks\n");
1445 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1446 req, rc, rep ? rep->lock_policy_res1 : 0);
1449 switch (cleanup_phase) {
1452 ldlm_lock_decref(child_lockh, child_mode);
1453 child_lockh->cookie = 0;
1456 if (rc && created) {
1457 int err = vfs_unlink(dparent->d_inode, dchild);
1459 CERROR("unlink(%.*s) in error path: %d\n",
1460 dchild->d_name.len, dchild->d_name.name,
1463 } else if (created) {
1464 mds_lock_new_child(obd, dchild->d_inode, NULL);
1466 /* audit stuff for OPEN */
1468 mds_audit(req, dchild, rec->ur_name,
1469 rec->ur_namelen - 1, AUDIT_OPEN, rc);
1474 if (dparent == NULL)
1479 if (parent_lockh[1].cookie != 0)
1480 ldlm_lock_decref(parent_lockh + 1, update_mode);
1483 ldlm_lock_decref(parent_lockh, parent_mode);
1485 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1488 OBD_FREE(mea, mea_size);
1490 atomic_inc(&mds->mds_open_count);
1492 mds_body_do_reverse_map(med, body);
1495 * If we have not taken the "open" lock, we may not return 0 here,
1496 * because caller expects 0 to mean "lock is taken", and it needs
1497 * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1498 * client. Later caller should rewrite the return value back to zero
1499 * if it to be used any further.
1501 if ((cleanup_phase != 3) && !rc)
1507 /* Close a "file descriptor" and possibly unlink an orphan from the
1508 * PENDING directory. Caller must hold child->i_sem, this drops it.
1510 * If we are being called from mds_disconnect() because the client has
1511 * disappeared, then req == NULL and we do not update last_rcvd because
1512 * there is nothing that could be recovered by the client at this stage
1513 * (it will not even _have_ an entry in last_rcvd anymore).
1515 * Returns EAGAIN if the client needs to get more data and re-close. */
1516 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1517 struct obd_device *obd, struct mds_file_data *mfd,
1520 struct inode *inode = mfd->mfd_dentry->d_inode;
1521 char idname[LL_ID_NAMELEN];
1522 int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1523 struct dentry *pending_child = NULL;
1524 struct mds_obd *mds = &obd->u.mds;
1525 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1526 void *handle = NULL;
1527 struct mds_body *request_body = NULL, *reply_body = NULL;
1528 struct dentry_params dp;
1529 struct iattr iattr = { 0 };
1530 struct llog_create_locks *lcl = NULL;
1533 if (req && req->rq_reqmsg != NULL)
1534 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1535 sizeof(*request_body));
1536 if (req && req->rq_repmsg != NULL)
1537 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1538 sizeof(*reply_body));
1540 if (request_body && (request_body->valid & OBD_MD_FLSIZE)) {
1541 /* we set i_size/i_blocks here, nobody will see
1542 * them until all write references are dropped.
1543 * btw, we hold one reference */
1544 LASSERT(mfd->mfd_mode & FMODE_WRITE);
1545 LASSERT(request_body->valid & OBD_MD_FLEPOCH);
1546 LASSERT(MDS_FILTERDATA(inode));
1547 if (MDS_FILTERDATA(inode)->io_epoch != request_body->io_epoch)
1548 CDEBUG(D_ERROR, "try to update attr. for old epoch "
1549 LPD64" while current "LPD64"\n",
1550 MDS_FILTERDATA(inode)->io_epoch,
1551 request_body->io_epoch);
1552 i_size_write(inode, request_body->size);
1553 inode->i_blocks = request_body->blocks;
1554 LTIME_S(inode->i_mtime) = (request_body->mtime);
1556 LTIME_S(iattr.ia_mtime) = request_body->mtime;
1557 iattr.ia_size = inode->i_size;
1558 iattr.ia_valid |= ATTR_SIZE|ATTR_MTIME;
1559 mds_inode_unset_attrs_old(inode);
1562 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1563 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1564 idname, inode->i_nlink, mds_orphan_open_count(inode));
1566 last_orphan = mds_orphan_open_dec_test(inode) &&
1567 mds_inode_is_orphan(inode);
1568 UP_WRITE_I_ALLOC_SEM(inode);
1570 /* this is half of the actual "close" */
1571 if (mfd->mfd_mode & FMODE_WRITE) {
1572 rc = mds_put_write_access(mds, inode, request_body,
1573 last_orphan && unlink_orphan);
1574 } else if (mfd->mfd_mode & FMODE_EXEC) {
1575 mds_allow_write_access(inode);
1578 if (last_orphan && unlink_orphan) {
1579 struct lov_mds_md *lmm = NULL;
1580 int stripe_count = 0;
1581 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1583 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1585 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1586 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1587 CERROR("found \"orphan\" %s %s with link count %d\n",
1588 S_ISREG(inode->i_mode) ? "file" : "dir",
1589 idname, inode->i_nlink);
1591 * sadly, there is no easy way to save pending_child from
1592 * mds_reint_unlink() into mfd, so we need to re-lookup, but
1593 * normally it will still be in the dcache.
1595 down(&pending_dir->i_sem);
1596 cleanup_phase = 1; /* up(i_sem) when finished */
1597 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1599 if (IS_ERR(pending_child))
1600 GOTO(cleanup, rc = PTR_ERR(pending_child));
1601 if (pending_child->d_inode == NULL) {
1602 CERROR("orphan %s has been removed\n", idname);
1603 GOTO(cleanup, rc = 0);
1606 cleanup_phase = 2; /* dput(pending_child) when finished */
1607 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1608 rc = vfs_rmdir(pending_dir, pending_child);
1610 CERROR("error unlinking orphan dir %s: rc %d\n",
1615 if (req != NULL && req->rq_repmsg != NULL) {
1616 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1617 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1620 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1621 NULL, stripe_count);
1622 if (IS_ERR(handle)) {
1623 rc = PTR_ERR(handle);
1628 pending_child->d_fsdata = (void *) &dp;
1631 rc = vfs_unlink(pending_dir, pending_child);
1633 CERROR("error unlinking orphan %s: rc %d\n",
1636 if (req != NULL && req->rq_repmsg != NULL &&
1637 (reply_body->valid & OBD_MD_FLEASIZE) &&
1638 mds_log_op_unlink(obd, pending_child->d_inode,
1639 lmm, req->rq_repmsg->buflens[1],
1640 lustre_msg_buf(req->rq_repmsg, 2, 0),
1641 req->rq_repmsg->buflens[2], &lcl) > 0) {
1642 reply_body->valid |= OBD_MD_FLCOOKIE;
1645 rc = mds_destroy_object(obd, inode, 1);
1647 CERROR("cannot destroy OSS object on close, err %d\n",
1652 goto out; /* Don't bother updating attrs on unlinked inode */
1653 } else if ((mfd->mfd_mode & FMODE_WRITE) && rc == 0) {
1654 /* last writer closed file - let's update i_size/i_blocks */
1655 mds_validate_size(obd, inode, request_body, &iattr);
1659 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1660 /* Update the on-disk attributes if this was the last write
1661 * close, and all information was provided (i.e., rc == 0)
1663 * XXX this should probably be abstracted with mds_reint_setattr
1666 if (request_body->valid & OBD_MD_FLMTIME &&
1667 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1668 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1669 iattr.ia_valid |= ATTR_MTIME;
1671 if (request_body->valid & OBD_MD_FLCTIME &&
1672 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1673 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1674 iattr.ia_valid |= ATTR_CTIME;
1677 /* XXX can't set block count with fsfilt_setattr (!) */
1678 if (request_body->valid & OBD_MD_FLSIZE) {
1679 iattr.ia_valid |= ATTR_SIZE;
1680 iattr.ia_size = request_body->size;
1682 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1683 iattr.ia_valid |= ATTR_BLOCKS;
1684 iattr.ia_blocks = request_body->blocks
1689 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1690 /* Only start a transaction to write out only the atime if
1691 * it is more out-of-date than the specified limit. If we
1692 * are already going to write out the atime then do it anyway.
1694 if ((request_body->atime >
1695 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1696 (iattr.ia_valid != 0 &&
1697 request_body->atime > LTIME_S(inode->i_atime))) {
1698 LTIME_S(iattr.ia_atime) = request_body->atime;
1699 iattr.ia_valid |= ATTR_ATIME;
1703 if (iattr.ia_valid != 0) {
1705 handle = fsfilt_start(obd,inode,FSFILT_OP_SETATTR,NULL);
1707 GOTO(cleanup, rc = PTR_ERR(handle));
1708 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1710 CERROR("error in setattr(%s): rc %d\n", idname, rc);
1713 /* If other clients have this file open for write, rc will be > 0 */
1716 if (!obd->obd_recovering && mds_inode_has_old_attrs(inode)
1717 && !mds_inode_is_orphan(inode)
1718 && atomic_read(&inode->i_writecount) == 0
1719 && inode->i_nlink != 0) {
1720 CERROR("leave inode %lu/%u with old attributes (nlink = %d)\n",
1721 inode->i_ino, inode->i_generation, inode->i_nlink);
1723 l_dput(mfd->mfd_dentry);
1724 mds_mfd_destroy(mfd);
1727 atomic_dec(&mds->mds_open_count);
1728 if (req != NULL && reply_body != NULL) {
1729 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1730 } else if (handle) {
1731 int err, force_sync = 0;
1733 if (req && req->rq_export)
1734 force_sync = req->rq_export->exp_sync;
1736 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1739 CERROR("error committing close: %d\n", err);
1745 switch (cleanup_phase) {
1748 ptlrpc_save_llog_lock(req, lcl);
1749 dput(pending_child);
1751 up(&pending_dir->i_sem);
1756 static int mds_extent_lock_callback(struct ldlm_lock *lock,
1757 struct ldlm_lock_desc *new, void *data,
1760 struct lustre_handle lockh = { 0 };
1765 case LDLM_CB_BLOCKING:
1766 ldlm_lock2handle(lock, &lockh);
1767 rc = ldlm_cli_cancel(&lockh);
1769 CERROR("ldlm_cli_cancel failed: %d\n", rc);
1771 case LDLM_CB_CANCELING: {
1780 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
1781 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
1782 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
1784 int mds_validate_size(struct obd_device *obd, struct inode *inode,
1785 struct mds_body *body, struct iattr *iattr)
1787 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1788 struct lustre_handle lockh = { 0 };
1789 struct lov_stripe_md *lsm = NULL;
1794 /* we update i_size/i_blocks only for regular files */
1795 if (!S_ISREG(inode->i_mode))
1798 /* we shouldn't fetch size from OSTes during recovery - deadlock */
1799 if (obd->obd_recovering) {
1800 CERROR("size-on-mds has no support on OST yet\n");
1804 /* if nobody modified attrs. we're lucky */
1805 if (!mds_inode_has_old_attrs(inode))
1808 /* 1: client didn't send actual i_size/i_blocks
1809 * 2: we seem to be last writer
1810 * 3: the file doesn't look like to be disappeared
1811 * conclusion: we're gonna fetch them from OSSes */
1813 down(&inode->i_sem);
1814 len = fsfilt_get_md(obd, inode, NULL, 0, EA_LOV);
1818 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, len);
1819 GOTO(cleanup, rc = len);
1820 } else if (len == 0) {
1821 CDEBUG(D_INODE, "no LOV in inode %lu\n", inode->i_ino);
1822 GOTO(cleanup, rc = 0);
1825 OBD_ALLOC(lmm, len);
1827 CERROR("can't allocate memory\n");
1828 GOTO(cleanup, rc = -ENOMEM);
1831 down(&inode->i_sem);
1832 rc = fsfilt_get_md(obd, inode, lmm, len, EA_LOV);
1836 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1840 rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, lmm, len);
1842 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1846 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1848 flags = LDLM_FL_HAS_INTENT;
1849 rc = obd_enqueue(obd->u.mds.mds_dt_exp, lsm, LDLM_EXTENT, &policy,
1850 LCK_PR, &flags, mds_extent_lock_callback,
1851 ldlm_completion_ast, NULL, NULL,
1852 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1853 if (rc == -ENOENT) {
1854 /* while we were enqueueing lock on OST, another thread
1855 * unlinked the file and started OST object destoying.
1856 * it's safe to return 0 here */
1857 GOTO(cleanup, rc = 0);
1858 } else if (rc != 0) {
1859 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1863 CDEBUG(D_INODE, "LOV reports "LPD64"/%lu for "DLID4" [%s%s%s]\n",
1864 inode->i_size, inode->i_blocks, OLID4(&body->id1),
1865 atomic_read(&inode->i_writecount) > 1 ? "U" : "",
1866 mds_inode_has_old_attrs(inode) ? "D" : "",
1867 mds_inode_is_orphan(inode) ? "O" : "");
1869 i_size_write(inode, lov_merge_size(lsm, 0));
1870 inode->i_blocks = lov_merge_blocks(lsm);
1871 LTIME_S(inode->i_mtime) = lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
1872 iattr->ia_size = inode->i_size;
1873 LTIME_S(iattr->ia_mtime) = LTIME_S(inode->i_mtime);
1874 iattr->ia_valid |= ATTR_SIZE | ATTR_MTIME;
1876 DOWN_WRITE_I_ALLOC_SEM(inode);
1877 mds_inode_unset_attrs_old(inode);
1878 UP_WRITE_I_ALLOC_SEM(inode);
1880 obd_cancel(obd->u.mds.mds_dt_exp, lsm, LCK_PR, &lockh);
1884 obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
1890 int mds_close(struct ptlrpc_request *req, int offset)
1892 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1893 struct obd_device *obd = req->rq_export->exp_obd;
1894 struct mds_body *body;
1895 struct mds_file_data *mfd;
1896 struct lvfs_run_ctxt saved;
1897 struct inode *inode;
1898 int rc, repsize[3] = {sizeof(struct mds_body),
1899 obd->u.mds.mds_max_mdsize,
1900 obd->u.mds.mds_max_cookiesize};
1902 MD_COUNTER_INCREMENT(obd, close);
1904 rc = lustre_pack_reply(req, 3, repsize, NULL);
1906 CERROR("lustre_pack_reply: rc = %d\n", rc);
1907 req->rq_status = rc;
1909 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1912 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1913 DEBUG_REQ(D_HA, req, "close replay");
1914 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1915 lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1916 req->rq_repmsg->buflens[2]);
1919 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1920 lustre_swab_mds_body);
1922 CERROR("Can't unpack body\n");
1923 req->rq_status = -EFAULT;
1927 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1928 /* do some stuff */ ;
1930 mfd = mds_handle2mfd(&body->handle);
1932 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1933 ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1934 req->rq_status = -ESTALE;
1938 inode = mfd->mfd_dentry->d_inode;
1940 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1941 DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1943 if (body->flags & MDS_BFLAG_DIRTY_EPOCH) {
1944 /* the client modified data through the handle
1945 * we need to care about attrs. -bzzz */
1946 mds_inode_set_attrs_old(inode);
1949 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1950 struct mds_body *rep_body;
1952 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1953 sizeof (*rep_body));
1954 LASSERT(rep_body != NULL);
1956 mds_pack_inode2body(obd, rep_body, inode,
1957 (body->valid & OBD_MD_FID) ? 1 : 0);
1959 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1960 inode, MDS_PACK_MD_LOCK, 0);
1962 spin_lock(&med->med_open_lock);
1963 list_del(&mfd->mfd_list);
1964 spin_unlock(&med->med_open_lock);
1966 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1967 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1968 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1970 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1971 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1972 req->rq_status = -ENOMEM;
1982 int mds_done_writing(struct ptlrpc_request *req, int offset)
1984 struct mds_body *body;
1985 int rc, size = sizeof(struct mds_body);
1988 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1990 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1991 lustre_swab_mds_body);
1993 CERROR("Can't unpack body\n");
1994 req->rq_status = -EFAULT;
1998 rc = lustre_pack_reply(req, 1, &size, NULL);
2000 CERROR("lustre_pack_reply: rc = %d\n", rc);
2001 req->rq_status = rc;