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>
47 #include "mds_internal.h"
49 /* Exported function from this file are:
51 * mds_open - called by the intent handler
52 * mds_close - an rpc handling function
53 * mds_pin - an rpc handling function - which will go away
54 * mds_mfd_close - for force closing files when a client dies
58 * MDS file data handling: file data holds a handle for a file opened
62 static void mds_mfd_addref(void *mfdp)
64 struct mds_file_data *mfd = mfdp;
66 atomic_inc(&mfd->mfd_refcount);
67 CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd,
68 atomic_read(&mfd->mfd_refcount));
71 struct mds_file_data *mds_mfd_new(void)
73 struct mds_file_data *mfd;
75 OBD_ALLOC(mfd, sizeof *mfd);
77 CERROR("mds: out of memory\n");
81 atomic_set(&mfd->mfd_refcount, 2);
83 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
84 class_handle_hash(&mfd->mfd_handle, mds_mfd_addref);
89 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
92 LASSERT(handle != NULL);
93 RETURN(class_handle2object(handle->cookie));
96 static void mds_mfd_put(struct mds_file_data *mfd)
98 CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
99 atomic_read(&mfd->mfd_refcount) - 1);
100 LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
101 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
102 if (atomic_dec_and_test(&mfd->mfd_refcount)) {
103 LASSERT(list_empty(&mfd->mfd_handle.h_link));
104 OBD_FREE(mfd, sizeof *mfd);
108 static void mds_mfd_destroy(struct mds_file_data *mfd)
110 class_handle_unhash(&mfd->mfd_handle);
114 #ifdef IFILTERDATA_ACTUALLY_USED
115 /* Caller must hold mds->mds_epoch_sem */
116 static int mds_alloc_filterdata(struct inode *inode)
118 LASSERT(inode->i_filterdata == NULL);
119 OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
120 if (inode->i_filterdata == NULL)
122 LASSERT(igrab(inode) == inode);
126 /* Caller must hold mds->mds_epoch_sem */
127 static void mds_free_filterdata(struct inode *inode)
129 LASSERT(inode->i_filterdata != NULL);
130 OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
131 inode->i_filterdata = NULL;
134 #endif /*IFILTERDATA_ACTUALLY_USED*/
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);
159 #ifdef IFILTERDATA_ACTUALLY_USED
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 #endif /*IFILTERDATA_ACTUALLY_USED*/
174 if (epoch > mds->mds_io_epoch)
175 mds->mds_io_epoch = epoch;
178 #ifdef IFILTERDATA_ACTUALLY_USED
179 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
180 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
181 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
183 #endif /*IFILTERDATA_ACTUALLY_USED*/
185 atomic_inc(&inode->i_writecount);
186 up(&mds->mds_epoch_sem);
190 /* Returns EAGAIN if the client needs to get size and/or cookies and close
191 * again -- which is never true if the file is about to be unlinked. Otherwise
192 * returns the number of remaining writers. */
193 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
194 struct mds_body *body, int unlinking)
199 down(&mds->mds_epoch_sem);
200 atomic_dec(&inode->i_writecount);
201 rc = atomic_read(&inode->i_writecount);
205 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
206 GOTO(out, rc = EAGAIN);
208 #ifdef IFILTERDATA_ACTUALLY_USED
209 mds_free_filterdata(inode);
212 up(&mds->mds_epoch_sem);
216 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
219 down(&mds->mds_epoch_sem);
220 if (atomic_read(&inode->i_writecount) > 0) {
221 up(&mds->mds_epoch_sem);
224 atomic_dec(&inode->i_writecount);
225 up(&mds->mds_epoch_sem);
229 static void mds_allow_write_access(struct inode *inode)
232 atomic_inc(&inode->i_writecount);
235 int mds_query_write_access(struct inode *inode)
238 RETURN(atomic_read(&inode->i_writecount));
241 /* This replaces the VFS dentry_open, it manages mfd and writecount */
242 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
243 struct vfsmount *mnt, int flags,
244 struct ptlrpc_request *req)
246 struct mds_export_data *med = &req->rq_export->exp_mds_data;
247 struct mds_obd *mds = mds_req2mds(req);
248 struct mds_file_data *mfd;
249 struct mds_body *body;
255 CERROR("mds: out of memory\n");
256 GOTO(cleanup_dentry, error = -ENOMEM);
259 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
261 if (flags & FMODE_WRITE) {
262 /* FIXME: in recovery, need to pass old epoch here */
263 error = mds_get_write_access(mds, dentry->d_inode, 0);
265 GOTO(cleanup_mfd, error);
266 #ifdef IFILTERDATA_ACTUALLY_USED
267 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
268 #endif /*IFILTERDATA_ACTUALLY_USED*/
269 } else if (flags & FMODE_EXEC) {
270 error = mds_deny_write_access(mds, dentry->d_inode);
272 GOTO(cleanup_mfd, error);
277 /* Mark the file as open to handle open-unlink. */
279 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
280 mds_orphan_open_inc(dentry->d_inode);
281 UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
283 mfd->mfd_mode = flags;
284 mfd->mfd_dentry = dentry;
285 mfd->mfd_xid = req->rq_xid;
287 spin_lock(&med->med_open_lock);
288 list_add(&mfd->mfd_list, &med->med_open_head);
289 spin_unlock(&med->med_open_lock);
292 body->handle.cookie = mfd->mfd_handle.h_cookie;
298 mds_mfd_destroy(mfd);
300 return ERR_PTR(error);
303 static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
304 struct lov_desc *desc)
307 for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
308 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
309 le64_to_cpu(lmm->lmm_objects[i].l_object_id);
313 /* Must be called with i_sem held */
314 static int mds_create_objects(struct ptlrpc_request *req, int offset,
315 struct mds_update_record *rec,
316 struct mds_obd *mds, struct obd_device *obd,
317 struct dentry *dchild, void **handle,
320 struct obdo *oa = NULL;
321 struct obd_trans_info oti = { 0 };
322 struct mds_body *body;
323 struct lov_stripe_md *lsm = NULL;
324 struct lov_mds_md *lmm = NULL;
325 struct inode *inode = dchild->d_inode;
327 int rc, lmm_bufsize, lmm_size;
330 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
331 !(rec->ur_flags & FMODE_WRITE))
334 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
336 if (!S_ISREG(inode->i_mode))
338 if (body->valid & OBD_MD_FLEASIZE)
341 OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
344 oti.oti_objid = *ids;
347 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
348 LASSERT(id_ino(rec->ur_id2));
349 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
350 lmm_size = rec->ur_eadatalen;
351 lmm = rec->ur_eadata;
355 *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
356 if (IS_ERR(*handle)) {
357 rc = PTR_ERR(*handle);
362 mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
364 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
365 lmm_bufsize = req->rq_repmsg->buflens[offset];
367 LASSERT(lmm_bufsize >= lmm_size);
368 memcpy(lmm_buf, lmm, lmm_size);
369 rc = fsfilt_set_md(obd, inode, *handle, lmm,
372 CERROR("open replay failed to set md:%d\n", rc);
376 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
377 GOTO(out_ids, rc = -ENOMEM);
381 GOTO(out_ids, rc = -ENOMEM);
382 oa->o_mode = S_IFREG | 0600;
383 oa->o_id = inode->i_ino;
384 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
385 oa->o_generation = inode->i_generation;
386 oa->o_uid = 0; /* must have 0 uid / gid on OST */
388 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
389 OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
392 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
395 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
396 /* check if things like lfs setstripe are sending us the ea */
397 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
398 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
400 0, &lsm, rec->ur_eadata);
404 OBD_ALLOC(lmm, mds->mds_max_mdsize);
406 GOTO(out_oa, rc = -ENOMEM);
408 lmm_size = mds->mds_max_mdsize;
409 rc = mds_get_md(obd, dchild->d_parent->d_inode,
410 lmm, &lmm_size, 1, 0);
412 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
415 OBD_FREE(lmm, mds->mds_max_mdsize);
419 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
420 rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
425 CDEBUG(level, "error creating objects for "
426 "inode %lu: rc = %d\n",
429 CERROR("obd_create returned invalid "
436 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
437 0, &lsm, rec->ur_eadata);
441 lsm->lsm_object_id = oa->o_id;
442 lsm->lsm_object_gr = oa->o_gr;
445 oa->o_size = inode->i_size;
446 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
447 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
448 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
450 CERROR("error setting attrs for inode %lu: rc %d\n",
453 CERROR("obd_setattr returned bad rc %d\n", rc);
460 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
461 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
463 LASSERT(lsm && lsm->lsm_object_id);
465 rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
466 if (!id_ino(rec->ur_id2))
467 obd_free_memmd(mds->mds_dt_exp, &lsm);
470 body->eadatasize = rc;
473 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
474 if (IS_ERR(*handle)) {
475 rc = PTR_ERR(*handle);
480 rc = fsfilt_set_md(obd, inode, *handle, lmm,
483 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
484 lmm_bufsize = req->rq_repmsg->buflens[offset];
486 LASSERT(lmm_bufsize >= lmm_size);
488 memcpy(lmm_buf, lmm, lmm_size);
489 obd_free_diskmd(mds->mds_dt_exp, &lmm);
491 oti_free_cookies(&oti);
495 OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
499 obd_free_memmd(mds->mds_dt_exp, &lsm);
504 mds_destroy_objects(struct obd_device *obd,
505 struct inode *inode, int async)
507 struct mds_obd *mds = &obd->u.mds;
508 struct lov_mds_md *lmm = NULL;
512 LASSERT(inode != NULL);
514 if (inode->i_nlink != 0) {
515 CWARN("attempt to destroy OSS object when "
516 "i_nlink == %d\n", (int)inode->i_nlink);
520 OBD_ALLOC(lmm, mds->mds_max_mdsize);
524 lmm_size = mds->mds_max_mdsize;
525 rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
527 CERROR("no stripe info for %lu/%lu inode\n",
528 (unsigned long)inode->i_ino,
529 (unsigned long)inode->i_generation);
530 GOTO(out_free_lmm, rc);
534 /* asynchronously unlink objecect on OSS */
535 rc = mds_unlink_object(mds, inode, lmm, lmm_size,
538 CERROR("error unlinking object on OSS, "
540 GOTO(out_free_lmm, rc);
543 CDEBUG(D_INODE, "no stripping info found for inode "
544 "%lu/%lu\n", (unsigned long)inode->i_ino,
545 (unsigned long)inode->i_generation);
549 OBD_FREE(lmm, mds->mds_max_mdsize);
553 static void reconstruct_open(struct mds_update_record *rec, int offset,
554 struct ptlrpc_request *req,
555 struct lustre_handle *child_lockh)
557 struct mds_export_data *med = &req->rq_export->exp_mds_data;
558 struct mds_client_data *mcd = med->med_mcd;
559 struct mds_obd *mds = mds_req2mds(req);
560 struct mds_file_data *mfd;
561 struct obd_device *obd = req->rq_export->exp_obd;
562 struct dentry *parent = NULL, *dchild;
563 struct ldlm_reply *rep;
564 struct mds_body *body;
570 LASSERT(offset == 3); /* only called via intent */
571 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
572 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
574 /* copy rc, transno and disp; steal locks */
575 mds_req_from_mcd(req, mcd);
576 intent_set_disposition(rep, mcd->mcd_last_data);
578 /* Only replay if create or open actually happened. */
579 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
581 return; /* error looking up parent or child */
584 if (rec->ur_namelen == 1) {
585 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
587 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
589 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
590 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
591 OLID4(rec->ur_id1), PTR_ERR(parent));
592 dchild = ll_lookup_one_len(rec->ur_name, parent,
593 rec->ur_namelen - 1);
594 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
595 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
598 if (!dchild->d_inode)
599 GOTO(out_dput, 0); /* child not present to open */
601 /* At this point, we know we have a child. We'll send
602 * it back _unless_ it not created and open failed. */
603 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
604 !intent_disposition(rep, DISP_OPEN_CREATE) &&
608 /* get lock (write for O_CREAT, read otherwise) */
609 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
610 if (S_ISREG(dchild->d_inode->i_mode)) {
611 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
612 dchild->d_inode, 1, 0);
615 LASSERT(rc == req->rq_status);
617 /* If we have LOV EA data, the OST holds size, mtime */
618 if (!(body->valid & OBD_MD_FLEASIZE))
619 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
620 OBD_MD_FLATIME | OBD_MD_FLMTIME);
623 /* If we have -EEXIST as the status, and we were asked to create
624 * exclusively, we can tell we failed because the file already existed.
626 if (req->rq_status == -EEXIST &&
627 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
628 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
632 /* If we didn't get as far as trying to open, then some locking thing
633 * probably went wrong, and we'll just bail here.
635 if (!intent_disposition(rep, DISP_OPEN_OPEN))
638 /* If we failed, then we must have failed opening, so don't look for
639 * file descriptor or anything, just give the client the bad news.
645 list_for_each(t, &med->med_open_head) {
646 mfd = list_entry(t, struct mds_file_data, mfd_list);
647 if (mfd->mfd_xid == req->rq_xid)
652 /* #warning "XXX fixme" bug 2991 */
653 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
654 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
655 * to detect a re-open */
657 mntget(mds->mds_vfsmnt);
658 CERROR("Re-opened file \n");
659 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
660 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
662 CERROR("mds: out of memory\n");
663 GOTO(out_dput, req->rq_status = -ENOMEM);
667 body->handle.cookie = mfd->mfd_handle.h_cookie;
668 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
669 mfd->mfd_handle.h_cookie);
680 /* do NOT or the MAY_*'s, you'll get the weakest */
681 static int accmode(int flags)
685 if (flags & FMODE_READ)
687 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
689 if (flags & FMODE_EXEC)
694 /* Handles object creation, actual opening, and I/O epoch */
695 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
696 struct mds_body *body, int flags, void **handle,
697 struct mds_update_record *rec, struct ldlm_reply *rep)
699 struct mds_obd *mds = mds_req2mds(req);
700 struct obd_device *obd = req->rq_export->exp_obd;
701 struct mds_file_data *mfd = NULL;
702 obd_id *ids = NULL; /* object IDs created */
707 /* atomically create objects if necessary */
708 down(&dchild->d_inode->i_sem);
709 mode = dchild->d_inode->i_mode;
710 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
711 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
712 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
713 dchild->d_inode, 0, 0);
715 up(&dchild->d_inode->i_sem);
720 /* no EA: create objects */
721 if ((body->valid & OBD_MD_FLEASIZE) &&
722 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
723 up(&dchild->d_inode->i_sem);
726 if (!(body->valid & OBD_MD_FLEASIZE)) {
727 /* no EA: create objects */
728 rc = mds_create_objects(req, 2, rec, mds, obd,
729 dchild, handle, &ids);
731 CERROR("mds_create_objects: rc = %d\n", rc);
732 up(&dchild->d_inode->i_sem);
736 if (S_ISREG(dchild->d_inode->i_mode) &&
737 (body->valid & OBD_MD_FLEASIZE)) {
738 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
741 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
742 dchild->d_inode, 0, 0);
744 up(&dchild->d_inode->i_sem);
749 rc = mds_pack_acl(obd, req->rq_repmsg, 3, body, dchild->d_inode);
751 CERROR("mds_pack_acl: rc = %d\n", rc);
752 up(&dchild->d_inode->i_sem);
756 /* If the inode has no EA data, then MDSs hold size, mtime */
757 if (S_ISREG(dchild->d_inode->i_mode) &&
758 !(body->valid & OBD_MD_FLEASIZE)) {
759 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
760 OBD_MD_FLATIME | OBD_MD_FLMTIME);
762 up(&dchild->d_inode->i_sem);
764 intent_set_disposition(rep, DISP_OPEN_OPEN);
765 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
767 RETURN(PTR_ERR(mfd));
769 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
770 mfd->mfd_handle.h_cookie);
772 mds_dt_update_objids(obd, ids);
773 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
776 // mds_mfd_destroy(mfd);
780 static int mds_open_by_id(struct ptlrpc_request *req,
781 struct lustre_id *id,
782 struct mds_body *body, int flags,
783 struct mds_update_record *rec,
784 struct ldlm_reply *rep)
786 struct mds_obd *mds = mds_req2mds(req);
787 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
788 struct dentry *dchild;
789 char idname[LL_ID_NAMELEN];
794 down(&pending_dir->i_sem);
795 idlen = ll_id2str(idname, id_ino(id), id_gen(id));
796 dchild = lookup_one_len(idname, mds->mds_pending_dir,
798 if (IS_ERR(dchild)) {
799 up(&pending_dir->i_sem);
800 rc = PTR_ERR(dchild);
801 CERROR("error looking up %s in PENDING: rc = %d\n",
806 up(&pending_dir->i_sem);
807 if (dchild->d_inode != NULL) {
808 mds_inode_set_orphan(dchild->d_inode);
809 mds_pack_inode2body(req2obd(req), body,
812 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
813 intent_set_disposition(rep, DISP_LOOKUP_POS);
814 CWARN("Orphan %s found and opened in PENDING directory\n",
821 * we didn't find it in PENDING so it isn't an orphan. See if it was a
822 * regular inode that was previously created.
824 dchild = mds_id2dentry(req2obd(req), id, NULL);
826 RETURN(PTR_ERR(dchild));
828 mds_pack_inode2body(req2obd(req), body,
831 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
832 intent_set_disposition(rep, DISP_LOOKUP_POS);
835 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
836 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
837 req, rc, rep ? rep->lock_policy_res1 : 0);
838 /* XXX what do we do here if mds_finish_transno itself failed? */
843 int mds_pin(struct ptlrpc_request *req, int offset)
845 struct obd_device *obd = req->rq_export->exp_obd;
846 struct mds_body *request_body, *reply_body;
847 struct lvfs_run_ctxt saved;
848 int rc, size = sizeof(*reply_body);
851 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
852 sizeof(*request_body));
854 rc = lustre_pack_reply(req, 1, &size, NULL);
857 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
859 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
860 rc = mds_open_by_id(req, &request_body->id1, reply_body,
861 request_body->flags, NULL, NULL);
862 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
868 * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
869 * child_lockh is NULL we just get the lock as a barrier to wait for other
870 * holders of this lock, and drop it right away again.
872 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
873 struct lustre_handle *child_lockh)
875 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
876 struct lustre_handle lockh;
881 if (child_lockh == NULL)
882 child_lockh = &lockh;
884 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
885 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
886 mds_blocking_ast, ldlm_completion_ast, NULL,
887 NULL, NULL, 0, NULL, child_lockh);
889 CERROR("ldlm_cli_enqueue: %d\n", rc);
890 else if (child_lockh == &lockh)
891 ldlm_lock_decref(child_lockh, LCK_EX);
896 int mds_open(struct mds_update_record *rec, int offset,
897 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
899 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
900 struct obd_device *obd = req->rq_export->exp_obd;
901 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
902 struct mds_obd *mds = mds_req2mds(req);
903 struct ldlm_reply *rep = NULL;
904 struct mds_body *body = NULL;
905 struct dentry *dchild = NULL, *dparent = NULL;
906 struct lustre_handle parent_lockh[2] = {{0}, {0}};
907 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
908 int parent_mode = LCK_PR;
910 struct dentry_params dp;
911 struct mea *mea = NULL;
912 int mea_size, update_mode;
913 int child_mode = LCK_PR;
914 /* Always returning LOOKUP lock if open succesful to guard
916 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
917 struct ldlm_res_id child_res_id = { .name = {0}};
921 DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
922 OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
925 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
926 (obd_timeout + 1) / 4);
928 if (offset == 3) { /* intent */
929 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
930 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
931 } else if (offset == 1) { /* non-intent reint */
932 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
938 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
941 * Step 0: If we are passed a id, then we assume the client already
942 * opened this file and is only replaying the RPC, so we open the inode
943 * by fid (at some large expense in security).
945 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
946 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
947 OLID4(rec->ur_id2), rec->ur_namelen - 1,
948 rec->ur_name, rec->ur_mode);
950 LASSERT(id_ino(rec->ur_id2));
952 rc = mds_open_by_id(req, rec->ur_id2, body, rec->ur_flags,
955 mds_body_do_reverse_map(med, body);
959 /* We didn't find the correct inode on disk either, so we
960 * need to re-create it via a regular replay. */
961 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
963 LASSERT(!id_ino(rec->ur_id2));
966 LASSERT(offset == 3); /* If we got here, we must be called via intent */
968 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
969 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
973 acc_mode = accmode(rec->ur_flags);
975 /* Step 1: Find and lock the parent */
976 if (rec->ur_flags & MDS_OPEN_CREAT) {
977 /* XXX Well, in fact we only need this lock mode change if
978 in addition to O_CREAT, the file does not exist.
979 But we do not know if it exists or not yet */
980 parent_mode = LCK_PW;
983 if (rec->ur_namelen == 1) {
984 /* client (LMV) wants to open the file by id */
985 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
986 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
987 if (IS_ERR(dchild)) {
988 rc = PTR_ERR(dparent);
989 CERROR("child lookup by an id error %d\n", rc);
996 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
997 parent_lockh, &update_mode, rec->ur_name,
998 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
999 if (IS_ERR(dparent)) {
1000 rc = PTR_ERR(dparent);
1001 if (rc != -ENOENT) {
1002 CERROR("parent lookup for "DLID4" failed, error %d\n",
1003 OLID4(rec->ur_id1), rc);
1005 /* Just cannot find parent - make it look like
1006 usual negative lookup to avoid extra MDS RPC */
1007 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1008 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1012 LASSERT(dparent->d_inode != NULL);
1014 cleanup_phase = 1; /* parent dentry and lock */
1016 /* try to retrieve MEA data for this dir */
1017 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1023 * dir is already splitted, check is requested filename should *
1024 * live at this MDS or at another one.
1027 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1028 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1030 "%s: inapropriate MDS(%d) for %lu/%u:%s."
1031 " should be %lu(%d)\n", obd->obd_name,
1032 mea->mea_master, dparent->d_inode->i_ino,
1033 dparent->d_inode->i_generation, rec->ur_name,
1034 (unsigned long)id_group(&mea->mea_ids[i]), i);
1035 GOTO(cleanup, rc = -ERESTART);
1039 /* Step 2: Lookup the child */
1040 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1041 if (IS_ERR(dchild)) {
1042 rc = PTR_ERR(dchild);
1043 dchild = NULL; /* don't confuse mds_finish_transno() below */
1048 cleanup_phase = 2; /* child dentry */
1050 if (dchild->d_flags & DCACHE_CROSS_REF) {
1051 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1052 LASSERT(rec->ur_namelen > 1);
1054 /* we're gonna acquire LOOKUP lock on the child,
1055 * but we have already locked parent and our order
1056 * may conflict with enqueue_order_locks(). so,
1057 * drop parent lock and acquire both the locks in
1058 * common order. bug 6190 */
1060 if (parent_lockh[1].cookie != 0)
1061 ldlm_lock_decref(parent_lockh + 1, update_mode);
1063 ldlm_lock_decref(parent_lockh, parent_mode);
1068 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1069 parent_lockh, &dparent,
1070 LCK_PR, MDS_INODELOCK_UPDATE,
1071 &update_mode, rec->ur_name,
1072 rec->ur_namelen, child_lockh,
1074 MDS_INODELOCK_LOOKUP);
1080 if (parent_lockh[1].cookie != 0)
1081 ldlm_lock_decref(parent_lockh + 1, update_mode);
1083 ldlm_lock_decref(parent_lockh, LCK_PR);
1085 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1086 CDEBUG(D_OTHER, "race: name changed (%p)\n",
1088 if (dchild->d_inode)
1089 ldlm_lock_decref(child_lockh, LCK_PR);
1095 mds_pack_dentry2body(obd, body, dchild, 1);
1096 intent_set_disposition(rep, DISP_LOOKUP_POS);
1097 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1100 OBD_FREE(mea, mea_size);
1106 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1108 if (dchild->d_inode)
1109 intent_set_disposition(rep, DISP_LOOKUP_POS);
1111 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1113 /* Step 3: If the child was negative, and we're supposed to, create it.*/
1114 if (dchild->d_inode == NULL) {
1117 struct inode *inode;
1119 ino = (unsigned long)id_ino(rec->ur_id2);
1121 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1124 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1125 obd->obd_name, dparent->d_inode->i_ino,
1126 dparent->d_inode->i_generation, rc);
1129 /* dir got splitted */
1130 GOTO(cleanup, rc = -ERESTART);
1131 } else if (rc < 0) {
1132 /* error happened during spitting */
1136 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1138 * dentry is negative and we weren't supposed to create
1139 * object, get out of here.
1141 GOTO(cleanup, rc = -ENOENT);
1144 intent_set_disposition(rep, DISP_OPEN_CREATE);
1145 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1147 if (IS_ERR(handle)) {
1148 rc = PTR_ERR(handle);
1152 dchild->d_fsdata = (void *) &dp;
1156 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1157 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1158 dchild->d_fsdata = NULL;
1161 CDEBUG(D_INODE, "error during create: %d\n", rc);
1164 inode = dchild->d_inode;
1166 LASSERT(ino == inode->i_ino);
1168 /* written as part of setattr */
1169 inode->i_generation = id_gen(rec->ur_id2);
1170 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1171 inode->i_ino, inode->i_generation);
1175 LTIME_S(iattr.ia_atime) = rec->ur_time;
1176 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1177 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1179 iattr.ia_uid = rec->ur_fsuid;
1180 if (dparent->d_inode->i_mode & S_ISGID)
1181 iattr.ia_gid = dparent->d_inode->i_gid;
1183 iattr.ia_gid = rec->ur_fsgid;
1185 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1186 ATTR_MTIME | ATTR_CTIME;
1188 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1190 CERROR("error on child setattr: rc = %d\n", rc);
1192 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1194 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1196 CERROR("error on parent setattr: rc = %d\n", rc);
1198 MD_COUNTER_INCREMENT(obd, create);
1201 down(&dchild->d_inode->i_sem);
1203 rc = mds_update_inode_sid(obd, dchild->d_inode,
1204 handle, rec->ur_id2);
1206 CERROR("mds_update_inode_sid() failed, "
1209 id_assign_fid(&body->id1, rec->ur_id2);
1212 * make sure, that fid is up-to-date.
1214 mds_set_last_fid(obd, id_fid(rec->ur_id2));
1216 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1217 handle, &body->id1);
1219 CERROR("mds_alloc_inode_sid() failed, "
1223 up(&dchild->d_inode->i_sem);
1225 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1226 rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1227 dchild->d_inode, handle,
1228 req->rq_export->exp_sync);
1232 acc_mode = 0; /* Don't check for permissions */
1234 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1236 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1237 "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1238 dchild->d_name.name, dchild, dchild->d_inode);
1240 if (S_ISREG(dchild->d_inode->i_mode)) {
1241 /* Check permissions etc */
1242 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1246 /* Can't write to a read-only file */
1247 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1248 GOTO(cleanup, rc = -EPERM);
1250 /* An append-only file must be opened in append mode for
1252 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1253 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1254 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1255 GOTO(cleanup, rc = -EPERM);
1258 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1259 (rec->ur_flags & MDS_OPEN_EXCL)) {
1260 /* File already exists, we didn't just create it, and we
1261 * were passed O_EXCL; err-or. */
1262 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1265 if (S_ISDIR(dchild->d_inode->i_mode)) {
1266 if (rec->ur_flags & MDS_OPEN_CREAT ||
1267 rec->ur_flags & FMODE_WRITE) {
1268 /* we are trying to create or write a exist dir */
1269 GOTO(cleanup, rc = -EISDIR);
1271 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1272 GOTO(cleanup, rc = -EACCES);
1275 * here was checking for possible GNS mount points to skip their
1276 * open. I removed it as its detection based only on SUID bit is
1277 * not reliable. Opening such a dirs should not cause any
1278 * problems, at least tests show that. --umka
1282 /* if we are following a symlink, don't open */
1283 if (S_ISLNK(dchild->d_inode->i_mode))
1284 GOTO(cleanup_no_trans, rc = 0);
1286 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1287 !S_ISDIR(dchild->d_inode->i_mode))
1288 GOTO(cleanup, rc = -ENOTDIR);
1290 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1291 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1292 GOTO(cleanup, rc = -EAGAIN);
1295 /* Obtain OPEN lock as well */
1296 policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1298 /* We cannot use acc_mode here, because it is zeroed in case of
1299 creating a file, so we get wrong lockmode */
1300 if (accmode(rec->ur_flags) & MAY_WRITE)
1301 child_mode = LCK_CW;
1302 else if (accmode(rec->ur_flags) & MAY_EXEC)
1303 child_mode = LCK_PR;
1305 child_mode = LCK_CR;
1307 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1308 struct lustre_id sid;
1310 down(&dchild->d_inode->i_sem);
1311 rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
1312 up(&dchild->d_inode->i_sem);
1314 CERROR("Can't read inode self id, "
1315 "inode %lu, rc %d\n",
1316 dchild->d_inode->i_ino, rc);
1320 /* In case of replay we do not get a lock assuming that the
1321 caller has it already */
1322 child_res_id.name[0] = id_fid(&sid);
1323 child_res_id.name[1] = id_group(&sid);
1325 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1326 child_res_id, LDLM_IBITS, &policy,
1327 child_mode, &lock_flags,
1328 mds_blocking_ast, ldlm_completion_ast,
1329 NULL, NULL, NULL, 0, NULL, child_lockh);
1336 /* Step 5: mds_open it */
1337 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1342 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1343 req, rc, rep ? rep->lock_policy_res1 : 0);
1346 switch (cleanup_phase) {
1349 ldlm_lock_decref(child_lockh, child_mode);
1350 child_lockh->cookie = 0;
1353 if (rc && created) {
1354 int err = vfs_unlink(dparent->d_inode, dchild);
1356 CERROR("unlink(%.*s) in error path: %d\n",
1357 dchild->d_name.len, dchild->d_name.name,
1360 } else if (created) {
1361 mds_lock_new_child(obd, dchild->d_inode, NULL);
1365 if (dparent == NULL)
1370 if (parent_lockh[1].cookie != 0)
1371 ldlm_lock_decref(parent_lockh + 1, update_mode);
1374 ldlm_lock_decref(parent_lockh, parent_mode);
1376 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1379 OBD_FREE(mea, mea_size);
1381 atomic_inc(&mds->mds_open_count);
1383 mds_body_do_reverse_map(med, body);
1386 * If we have not taken the "open" lock, we may not return 0 here,
1387 * because caller expects 0 to mean "lock is taken", and it needs
1388 * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1389 * client. Later caller should rewrite the return value back to zero
1390 * if it to be used any further.
1392 if ((cleanup_phase != 3) && !rc)
1398 /* Close a "file descriptor" and possibly unlink an orphan from the
1399 * PENDING directory. Caller must hold child->i_sem, this drops it.
1401 * If we are being called from mds_disconnect() because the client has
1402 * disappeared, then req == NULL and we do not update last_rcvd because
1403 * there is nothing that could be recovered by the client at this stage
1404 * (it will not even _have_ an entry in last_rcvd anymore).
1406 * Returns EAGAIN if the client needs to get more data and re-close. */
1407 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1408 struct obd_device *obd, struct mds_file_data *mfd,
1411 struct inode *inode = mfd->mfd_dentry->d_inode;
1412 char idname[LL_ID_NAMELEN];
1413 int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1414 struct dentry *pending_child = NULL;
1415 struct mds_obd *mds = &obd->u.mds;
1416 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1417 void *handle = NULL;
1418 struct mds_body *request_body = NULL, *reply_body = NULL;
1419 struct dentry_params dp;
1420 struct iattr iattr = { 0 };
1421 struct llog_create_locks *lcl = NULL;
1424 if (req && req->rq_reqmsg != NULL)
1425 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1426 sizeof(*request_body));
1427 if (req && req->rq_repmsg != NULL)
1428 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1429 sizeof(*reply_body));
1431 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1432 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1433 idname, inode->i_nlink, mds_orphan_open_count(inode));
1435 last_orphan = mds_orphan_open_dec_test(inode) &&
1436 mds_inode_is_orphan(inode);
1437 if (last_orphan && unlink_orphan)
1438 mds_inode_unset_orphan(inode);
1439 UP_WRITE_I_ALLOC_SEM(inode);
1441 /* this is half of the actual "close" */
1442 if (mfd->mfd_mode & FMODE_WRITE) {
1443 rc = mds_put_write_access(mds, inode, request_body,
1444 last_orphan && unlink_orphan);
1445 } else if (mfd->mfd_mode & FMODE_EXEC) {
1446 mds_allow_write_access(inode);
1449 if (last_orphan && unlink_orphan) {
1450 struct lov_mds_md *lmm = NULL;
1451 int stripe_count = 0;
1452 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1454 if (obd->obd_recovering) {
1455 CDEBUG(D_HA, "not remove orphan %s until recovery"
1456 " is over\n", idname);
1460 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1462 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1463 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1464 CERROR("found \"orphan\" %s %s with link count %d\n",
1465 S_ISREG(inode->i_mode) ? "file" : "dir",
1466 idname, inode->i_nlink);
1468 * sadly, there is no easy way to save pending_child from
1469 * mds_reint_unlink() into mfd, so we need to re-lookup, but
1470 * normally it will still be in the dcache.
1472 down(&pending_dir->i_sem);
1473 cleanup_phase = 1; /* up(i_sem) when finished */
1474 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1476 if (IS_ERR(pending_child))
1477 GOTO(cleanup, rc = PTR_ERR(pending_child));
1478 LASSERT(pending_child->d_inode != NULL);
1480 cleanup_phase = 2; /* dput(pending_child) when finished */
1481 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1482 rc = vfs_rmdir(pending_dir, pending_child);
1484 CERROR("error unlinking orphan dir %s: rc %d\n",
1489 if (req != NULL && req->rq_repmsg != NULL) {
1490 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1491 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1494 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1495 NULL, stripe_count);
1496 if (IS_ERR(handle)) {
1497 rc = PTR_ERR(handle);
1502 pending_child->d_fsdata = (void *) &dp;
1505 rc = vfs_unlink(pending_dir, pending_child);
1507 CERROR("error unlinking orphan %s: rc %d\n",
1510 if (req != NULL && req->rq_repmsg != NULL &&
1511 (reply_body->valid & OBD_MD_FLEASIZE) &&
1512 mds_log_op_unlink(obd, pending_child->d_inode,
1513 lmm, req->rq_repmsg->buflens[1],
1514 lustre_msg_buf(req->rq_repmsg, 2, 0),
1515 req->rq_repmsg->buflens[2], &lcl) > 0) {
1516 reply_body->valid |= OBD_MD_FLCOOKIE;
1519 rc = mds_destroy_objects(obd, inode, 1);
1521 CERROR("cannot destroy OSS object on close, err %d\n",
1526 goto out; /* Don't bother updating attrs on unlinked inode */
1530 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1531 /* Update the on-disk attributes if this was the last write
1532 * close, and all information was provided (i.e., rc == 0)
1534 * XXX this should probably be abstracted with mds_reint_setattr
1537 if (request_body->valid & OBD_MD_FLMTIME &&
1538 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1539 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1540 iattr.ia_valid |= ATTR_MTIME;
1542 if (request_body->valid & OBD_MD_FLCTIME &&
1543 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1544 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1545 iattr.ia_valid |= ATTR_CTIME;
1548 /* XXX can't set block count with fsfilt_setattr (!) */
1549 if (request_body->valid & OBD_MD_FLSIZE) {
1550 iattr.ia_valid |= ATTR_SIZE;
1551 iattr.ia_size = request_body->size;
1553 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1554 iattr.ia_valid |= ATTR_BLOCKS;
1555 iattr.ia_blocks = request_body->blocks
1560 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1561 /* Only start a transaction to write out only the atime if
1562 * it is more out-of-date than the specified limit. If we
1563 * are already going to write out the atime then do it anyway.
1565 if ((request_body->atime >
1566 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1567 (iattr.ia_valid != 0 &&
1568 request_body->atime > LTIME_S(inode->i_atime))) {
1569 LTIME_S(iattr.ia_atime) = request_body->atime;
1570 iattr.ia_valid |= ATTR_ATIME;
1574 if (iattr.ia_valid != 0) {
1575 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1577 GOTO(cleanup, rc = PTR_ERR(handle));
1578 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1580 CERROR("error in setattr(%s): rc %d\n", idname, rc);
1583 /* If other clients have this file open for write, rc will be > 0 */
1586 l_dput(mfd->mfd_dentry);
1587 mds_mfd_destroy(mfd);
1590 atomic_dec(&mds->mds_open_count);
1591 if (req != NULL && reply_body != NULL) {
1592 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1593 } else if (handle) {
1594 int err, force_sync = 0;
1596 if (req && req->rq_export)
1597 force_sync = req->rq_export->exp_sync;
1599 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1602 CERROR("error committing close: %d\n", err);
1608 switch (cleanup_phase) {
1611 ptlrpc_save_llog_lock(req, lcl);
1612 dput(pending_child);
1614 up(&pending_dir->i_sem);
1619 int mds_close(struct ptlrpc_request *req, int offset)
1621 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1622 struct obd_device *obd = req->rq_export->exp_obd;
1623 struct mds_body *body;
1624 struct mds_file_data *mfd;
1625 struct lvfs_run_ctxt saved;
1626 struct inode *inode;
1627 int rc, repsize[3] = {sizeof(struct mds_body),
1628 obd->u.mds.mds_max_mdsize,
1629 obd->u.mds.mds_max_cookiesize};
1631 MD_COUNTER_INCREMENT(obd, close);
1633 rc = lustre_pack_reply(req, 3, repsize, NULL);
1635 CERROR("lustre_pack_reply: rc = %d\n", rc);
1636 req->rq_status = rc;
1638 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1641 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1642 DEBUG_REQ(D_HA, req, "close replay");
1643 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1644 lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1645 req->rq_repmsg->buflens[2]);
1648 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1649 lustre_swab_mds_body);
1651 CERROR("Can't unpack body\n");
1652 req->rq_status = -EFAULT;
1656 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1657 /* do some stuff */ ;
1659 mfd = mds_handle2mfd(&body->handle);
1661 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1662 ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1663 req->rq_status = -ESTALE;
1667 inode = mfd->mfd_dentry->d_inode;
1668 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1669 DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1670 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1671 struct mds_body *rep_body;
1673 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1674 sizeof (*rep_body));
1675 LASSERT(rep_body != NULL);
1677 mds_pack_inode2body(obd, rep_body, inode,
1678 (body->valid & OBD_MD_FID) ? 1 : 0);
1680 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1681 inode, MDS_PACK_MD_LOCK, 0);
1683 spin_lock(&med->med_open_lock);
1684 list_del(&mfd->mfd_list);
1685 spin_unlock(&med->med_open_lock);
1687 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1688 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1689 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1691 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1692 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1693 req->rq_status = -ENOMEM;
1702 int mds_done_writing(struct ptlrpc_request *req, int offset)
1704 struct mds_body *body;
1705 int rc, size = sizeof(struct mds_body);
1708 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1710 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1711 lustre_swab_mds_body);
1713 CERROR("Can't unpack body\n");
1714 req->rq_status = -EFAULT;
1718 rc = lustre_pack_reply(req, 1, &size, NULL);
1720 CERROR("lustre_pack_reply: rc = %d\n", rc);
1721 req->rq_status = rc;