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);
503 static void reconstruct_open(struct mds_update_record *rec, int offset,
504 struct ptlrpc_request *req,
505 struct lustre_handle *child_lockh)
507 struct mds_export_data *med = &req->rq_export->exp_mds_data;
508 struct mds_client_data *mcd = med->med_mcd;
509 struct mds_obd *mds = mds_req2mds(req);
510 struct mds_file_data *mfd;
511 struct obd_device *obd = req->rq_export->exp_obd;
512 struct dentry *parent = NULL, *dchild;
513 struct ldlm_reply *rep;
514 struct mds_body *body;
520 LASSERT(offset == 3); /* only called via intent */
521 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
522 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
524 /* copy rc, transno and disp; steal locks */
525 mds_req_from_mcd(req, mcd);
526 intent_set_disposition(rep, mcd->mcd_last_data);
528 /* Only replay if create or open actually happened. */
529 if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) {
531 return; /* error looking up parent or child */
534 if (rec->ur_namelen == 1) {
535 CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
537 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
539 parent = mds_id2dentry(obd, rec->ur_id1, NULL);
540 LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
541 OLID4(rec->ur_id1), PTR_ERR(parent));
542 dchild = ll_lookup_one_len(rec->ur_name, parent,
543 rec->ur_namelen - 1);
544 LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
545 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
548 if (!dchild->d_inode)
549 GOTO(out_dput, 0); /* child not present to open */
551 /* At this point, we know we have a child. We'll send
552 * it back _unless_ it not created and open failed. */
553 if (intent_disposition(rep, DISP_OPEN_OPEN) &&
554 !intent_disposition(rep, DISP_OPEN_CREATE) &&
558 /* get lock (write for O_CREAT, read otherwise) */
559 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
560 if (S_ISREG(dchild->d_inode->i_mode)) {
561 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
562 dchild->d_inode, 1, 0);
565 LASSERT(rc == req->rq_status);
567 /* If we have LOV EA data, the OST holds size, mtime */
568 if (!(body->valid & OBD_MD_FLEASIZE))
569 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
570 OBD_MD_FLATIME | OBD_MD_FLMTIME);
573 /* If we have -EEXIST as the status, and we were asked to create
574 * exclusively, we can tell we failed because the file already existed.
576 if (req->rq_status == -EEXIST &&
577 ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) ==
578 (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) {
582 /* If we didn't get as far as trying to open, then some locking thing
583 * probably went wrong, and we'll just bail here.
585 if (!intent_disposition(rep, DISP_OPEN_OPEN))
588 /* If we failed, then we must have failed opening, so don't look for
589 * file descriptor or anything, just give the client the bad news.
595 list_for_each(t, &med->med_open_head) {
596 mfd = list_entry(t, struct mds_file_data, mfd_list);
597 if (mfd->mfd_xid == req->rq_xid)
602 /* #warning "XXX fixme" bug 2991 */
603 /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
604 * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
605 * to detect a re-open */
607 mntget(mds->mds_vfsmnt);
608 CERROR("Re-opened file \n");
609 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
610 rec->ur_flags & ~MDS_OPEN_TRUNC, req);
612 CERROR("mds: out of memory\n");
613 GOTO(out_dput, req->rq_status = -ENOMEM);
617 body->handle.cookie = mfd->mfd_handle.h_cookie;
618 CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
619 mfd->mfd_handle.h_cookie);
630 /* do NOT or the MAY_*'s, you'll get the weakest */
631 static int accmode(int flags)
635 if (flags & FMODE_READ)
637 if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC))
639 if (flags & FMODE_EXEC)
644 /* Handles object creation, actual opening, and I/O epoch */
645 static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
646 struct mds_body *body, int flags, void **handle,
647 struct mds_update_record *rec, struct ldlm_reply *rep)
649 struct mds_obd *mds = mds_req2mds(req);
650 struct obd_device *obd = req->rq_export->exp_obd;
651 struct mds_file_data *mfd = NULL;
652 obd_id *ids = NULL; /* object IDs created */
657 /* atomically create objects if necessary */
658 down(&dchild->d_inode->i_sem);
659 mode = dchild->d_inode->i_mode;
660 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
661 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
662 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
663 dchild->d_inode, 0, 0);
665 up(&dchild->d_inode->i_sem);
670 /* no EA: create objects */
671 if ((body->valid & OBD_MD_FLEASIZE) &&
672 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
673 up(&dchild->d_inode->i_sem);
676 if (!(body->valid & OBD_MD_FLEASIZE)) {
677 /* no EA: create objects */
678 rc = mds_create_objects(req, 2, rec, mds, obd,
679 dchild, handle, &ids);
681 CERROR("mds_create_objects: rc = %d\n", rc);
682 up(&dchild->d_inode->i_sem);
686 if (S_ISREG(dchild->d_inode->i_mode) &&
687 (body->valid & OBD_MD_FLEASIZE)) {
688 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
691 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
692 dchild->d_inode, 0, 0);
694 up(&dchild->d_inode->i_sem);
699 rc = mds_pack_acl(obd, req->rq_repmsg, 3, body, dchild->d_inode);
701 CERROR("mds_pack_acl: rc = %d\n", rc);
702 up(&dchild->d_inode->i_sem);
706 /* If the inode has no EA data, then MDSs hold size, mtime */
707 if (S_ISREG(dchild->d_inode->i_mode) &&
708 !(body->valid & OBD_MD_FLEASIZE)) {
709 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
710 OBD_MD_FLATIME | OBD_MD_FLMTIME);
712 up(&dchild->d_inode->i_sem);
714 intent_set_disposition(rep, DISP_OPEN_OPEN);
715 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
717 RETURN(PTR_ERR(mfd));
719 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
720 mfd->mfd_handle.h_cookie);
722 mds_dt_update_objids(obd, ids);
723 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
726 // mds_mfd_destroy(mfd);
730 static int mds_open_by_id(struct ptlrpc_request *req,
731 struct lustre_id *id,
732 struct mds_body *body, int flags,
733 struct mds_update_record *rec,
734 struct ldlm_reply *rep)
736 struct mds_obd *mds = mds_req2mds(req);
737 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
738 struct dentry *dchild;
739 char idname[LL_ID_NAMELEN];
744 down(&pending_dir->i_sem);
745 idlen = ll_id2str(idname, id_ino(id), id_gen(id));
746 dchild = lookup_one_len(idname, mds->mds_pending_dir,
748 if (IS_ERR(dchild)) {
749 up(&pending_dir->i_sem);
750 rc = PTR_ERR(dchild);
751 CERROR("error looking up %s in PENDING: rc = %d\n",
756 up(&pending_dir->i_sem);
757 if (dchild->d_inode != NULL) {
758 mds_inode_set_orphan(dchild->d_inode);
759 mds_pack_inode2body(req2obd(req), body,
762 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
763 intent_set_disposition(rep, DISP_LOOKUP_POS);
764 CWARN("Orphan %s found and opened in PENDING directory\n",
771 * we didn't find it in PENDING so it isn't an orphan. See if it was a
772 * regular inode that was previously created.
774 dchild = mds_id2dentry(req2obd(req), id, NULL);
776 RETURN(PTR_ERR(dchild));
778 mds_pack_inode2body(req2obd(req), body,
781 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
782 intent_set_disposition(rep, DISP_LOOKUP_POS);
785 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
786 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
787 req, rc, rep ? rep->lock_policy_res1 : 0);
788 /* XXX what do we do here if mds_finish_transno itself failed? */
793 int mds_pin(struct ptlrpc_request *req, int offset)
795 struct obd_device *obd = req->rq_export->exp_obd;
796 struct mds_body *request_body, *reply_body;
797 struct lvfs_run_ctxt saved;
798 int rc, size = sizeof(*reply_body);
801 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
802 sizeof(*request_body));
804 rc = lustre_pack_reply(req, 1, &size, NULL);
807 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
809 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
810 rc = mds_open_by_id(req, &request_body->id1, reply_body,
811 request_body->flags, NULL, NULL);
812 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
818 * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
819 * child_lockh is NULL we just get the lock as a barrier to wait for other
820 * holders of this lock, and drop it right away again.
822 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
823 struct lustre_handle *child_lockh)
825 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
826 struct lustre_handle lockh;
831 if (child_lockh == NULL)
832 child_lockh = &lockh;
834 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
835 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
836 mds_blocking_ast, ldlm_completion_ast, NULL,
837 NULL, NULL, 0, NULL, child_lockh);
839 CERROR("ldlm_cli_enqueue: %d\n", rc);
840 else if (child_lockh == &lockh)
841 ldlm_lock_decref(child_lockh, LCK_EX);
846 int mds_open(struct mds_update_record *rec, int offset,
847 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
849 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
850 struct obd_device *obd = req->rq_export->exp_obd;
851 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
852 struct mds_obd *mds = mds_req2mds(req);
853 struct ldlm_reply *rep = NULL;
854 struct mds_body *body = NULL;
855 struct dentry *dchild = NULL, *dparent = NULL;
856 struct lustre_handle parent_lockh[2] = {{0}, {0}};
857 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
858 int parent_mode = LCK_PR;
860 struct dentry_params dp;
861 struct mea *mea = NULL;
862 int mea_size, update_mode;
863 int child_mode = LCK_PR;
864 /* Always returning LOOKUP lock if open succesful to guard
866 ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
867 struct ldlm_res_id child_res_id = { .name = {0}};
871 DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
872 OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
875 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
876 (obd_timeout + 1) / 4);
878 if (offset == 3) { /* intent */
879 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
880 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
881 } else if (offset == 1) { /* non-intent reint */
882 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
888 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
891 * Step 0: If we are passed a id, then we assume the client already
892 * opened this file and is only replaying the RPC, so we open the inode
893 * by fid (at some large expense in security).
895 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
896 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
897 OLID4(rec->ur_id2), rec->ur_namelen - 1,
898 rec->ur_name, rec->ur_mode);
900 LASSERT(id_ino(rec->ur_id2));
902 rc = mds_open_by_id(req, rec->ur_id2, body, rec->ur_flags,
905 mds_body_do_reverse_map(med, body);
909 /* We didn't find the correct inode on disk either, so we
910 * need to re-create it via a regular replay. */
911 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
913 LASSERT(!id_ino(rec->ur_id2));
916 LASSERT(offset == 3); /* If we got here, we must be called via intent */
918 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
919 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
923 acc_mode = accmode(rec->ur_flags);
925 /* Step 1: Find and lock the parent */
926 if (rec->ur_flags & MDS_OPEN_CREAT) {
927 /* XXX Well, in fact we only need this lock mode change if
928 in addition to O_CREAT, the file does not exist.
929 But we do not know if it exists or not yet */
930 parent_mode = LCK_PW;
933 if (rec->ur_namelen == 1) {
934 /* client (LMV) wants to open the file by id */
935 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
936 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
937 if (IS_ERR(dchild)) {
938 rc = PTR_ERR(dparent);
939 CERROR("child lookup by an id error %d\n", rc);
946 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
947 parent_lockh, &update_mode, rec->ur_name,
948 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
949 if (IS_ERR(dparent)) {
950 rc = PTR_ERR(dparent);
952 CERROR("parent lookup for "DLID4" failed, error %d\n",
953 OLID4(rec->ur_id1), rc);
955 /* Just cannot find parent - make it look like
956 usual negative lookup to avoid extra MDS RPC */
957 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
958 intent_set_disposition(rep, DISP_LOOKUP_NEG);
962 LASSERT(dparent->d_inode != NULL);
964 cleanup_phase = 1; /* parent dentry and lock */
966 /* try to retrieve MEA data for this dir */
967 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
973 * dir is already splitted, check is requested filename should *
974 * live at this MDS or at another one.
977 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
978 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
980 "%s: inapropriate MDS(%d) for %lu/%u:%s."
981 " should be %lu(%d)\n", obd->obd_name,
982 mea->mea_master, dparent->d_inode->i_ino,
983 dparent->d_inode->i_generation, rec->ur_name,
984 (unsigned long)id_group(&mea->mea_ids[i]), i);
985 GOTO(cleanup, rc = -ERESTART);
989 /* Step 2: Lookup the child */
990 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
991 if (IS_ERR(dchild)) {
992 rc = PTR_ERR(dchild);
993 dchild = NULL; /* don't confuse mds_finish_transno() below */
998 cleanup_phase = 2; /* child dentry */
1000 if (dchild->d_flags & DCACHE_CROSS_REF) {
1001 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1002 LASSERT(rec->ur_namelen > 1);
1004 /* we're gonna acquire LOOKUP lock on the child,
1005 * but we have already locked parent and our order
1006 * may conflict with enqueue_order_locks(). so,
1007 * drop parent lock and acquire both the locks in
1008 * common order. bug 6190 */
1010 if (parent_lockh[1].cookie != 0)
1011 ldlm_lock_decref(parent_lockh + 1, update_mode);
1013 ldlm_lock_decref(parent_lockh, parent_mode);
1018 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1019 parent_lockh, &dparent,
1020 LCK_PR, MDS_INODELOCK_UPDATE,
1021 &update_mode, rec->ur_name,
1022 rec->ur_namelen, child_lockh,
1024 MDS_INODELOCK_LOOKUP);
1030 if (parent_lockh[1].cookie != 0)
1031 ldlm_lock_decref(parent_lockh + 1, update_mode);
1033 ldlm_lock_decref(parent_lockh, parent_mode);
1035 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1036 CDEBUG(D_OTHER, "race: name changed (%p)\n",
1038 if (dchild->d_inode)
1039 ldlm_lock_decref(child_lockh, LCK_PR);
1045 mds_pack_dentry2body(obd, body, dchild, 1);
1046 intent_set_disposition(rep, DISP_LOOKUP_POS);
1047 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1050 OBD_FREE(mea, mea_size);
1056 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1058 if (dchild->d_inode)
1059 intent_set_disposition(rep, DISP_LOOKUP_POS);
1061 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1063 /* Step 3: If the child was negative, and we're supposed to, create it.*/
1064 if (dchild->d_inode == NULL) {
1067 struct inode *inode;
1069 ino = (unsigned long)id_ino(rec->ur_id2);
1071 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1074 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1075 obd->obd_name, dparent->d_inode->i_ino,
1076 dparent->d_inode->i_generation, rc);
1079 /* dir got splitted */
1080 GOTO(cleanup, rc = -ERESTART);
1081 } else if (rc < 0) {
1082 /* error happened during spitting */
1086 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1088 * dentry is negative and we weren't supposed to create
1089 * object, get out of here.
1091 GOTO(cleanup, rc = -ENOENT);
1094 intent_set_disposition(rep, DISP_OPEN_CREATE);
1095 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1097 if (IS_ERR(handle)) {
1098 rc = PTR_ERR(handle);
1102 dchild->d_fsdata = (void *) &dp;
1106 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1107 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1108 dchild->d_fsdata = NULL;
1111 CDEBUG(D_INODE, "error during create: %d\n", rc);
1114 inode = dchild->d_inode;
1116 LASSERT(ino == inode->i_ino);
1118 /* written as part of setattr */
1119 inode->i_generation = id_gen(rec->ur_id2);
1120 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1121 inode->i_ino, inode->i_generation);
1125 LTIME_S(iattr.ia_atime) = rec->ur_time;
1126 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1127 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1129 iattr.ia_uid = rec->ur_fsuid;
1130 if (dparent->d_inode->i_mode & S_ISGID)
1131 iattr.ia_gid = dparent->d_inode->i_gid;
1133 iattr.ia_gid = rec->ur_fsgid;
1135 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1136 ATTR_MTIME | ATTR_CTIME;
1138 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1140 CERROR("error on child setattr: rc = %d\n", rc);
1142 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1144 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1146 CERROR("error on parent setattr: rc = %d\n", rc);
1148 MD_COUNTER_INCREMENT(obd, create);
1151 down(&dchild->d_inode->i_sem);
1153 rc = mds_update_inode_sid(obd, dchild->d_inode,
1154 handle, rec->ur_id2);
1156 CERROR("mds_update_inode_sid() failed, "
1159 id_assign_fid(&body->id1, rec->ur_id2);
1162 * make sure, that fid is up-to-date.
1164 mds_set_last_fid(obd, id_fid(rec->ur_id2));
1166 rc = mds_alloc_inode_sid(obd, dchild->d_inode,
1167 handle, &body->id1);
1169 CERROR("mds_alloc_inode_sid() failed, "
1173 up(&dchild->d_inode->i_sem);
1175 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1176 rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1177 dchild->d_inode, handle,
1178 req->rq_export->exp_sync);
1182 acc_mode = 0; /* Don't check for permissions */
1184 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1186 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1187 "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1188 dchild->d_name.name, dchild, dchild->d_inode);
1190 if (S_ISREG(dchild->d_inode->i_mode)) {
1191 /* Check permissions etc */
1192 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1196 /* Can't write to a read-only file */
1197 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1198 GOTO(cleanup, rc = -EPERM);
1200 /* An append-only file must be opened in append mode for
1202 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1203 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1204 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1205 GOTO(cleanup, rc = -EPERM);
1208 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1209 (rec->ur_flags & MDS_OPEN_EXCL)) {
1210 /* File already exists, we didn't just create it, and we
1211 * were passed O_EXCL; err-or. */
1212 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1215 if (S_ISDIR(dchild->d_inode->i_mode)) {
1216 if (rec->ur_flags & MDS_OPEN_CREAT ||
1217 rec->ur_flags & FMODE_WRITE) {
1218 /* we are trying to create or write a exist dir */
1219 GOTO(cleanup, rc = -EISDIR);
1221 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1222 GOTO(cleanup, rc = -EACCES);
1225 * here was checking for possible GNS mount points to skip their
1226 * open. I removed it as its detection based only on SUID bit is
1227 * not reliable. Opening such a dirs should not cause any
1228 * problems, at least tests show that. --umka
1232 /* if we are following a symlink, don't open */
1233 if (S_ISLNK(dchild->d_inode->i_mode))
1234 GOTO(cleanup_no_trans, rc = 0);
1236 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1237 !S_ISDIR(dchild->d_inode->i_mode))
1238 GOTO(cleanup, rc = -ENOTDIR);
1240 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1241 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1242 GOTO(cleanup, rc = -EAGAIN);
1245 /* Obtain OPEN lock as well */
1246 policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
1248 /* We cannot use acc_mode here, because it is zeroed in case of
1249 creating a file, so we get wrong lockmode */
1250 if (accmode(rec->ur_flags) & MAY_WRITE)
1251 child_mode = LCK_CW;
1252 else if (accmode(rec->ur_flags) & MAY_EXEC)
1253 child_mode = LCK_PR;
1255 child_mode = LCK_CR;
1257 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1258 struct lustre_id sid;
1260 down(&dchild->d_inode->i_sem);
1261 rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
1262 up(&dchild->d_inode->i_sem);
1264 CERROR("Can't read inode self id, "
1265 "inode %lu, rc %d\n",
1266 dchild->d_inode->i_ino, rc);
1270 /* In case of replay we do not get a lock assuming that the
1271 caller has it already */
1272 child_res_id.name[0] = id_fid(&sid);
1273 child_res_id.name[1] = id_group(&sid);
1275 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1276 child_res_id, LDLM_IBITS, &policy,
1277 child_mode, &lock_flags,
1278 mds_blocking_ast, ldlm_completion_ast,
1279 NULL, NULL, NULL, 0, NULL, child_lockh);
1286 /* Step 5: mds_open it */
1287 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1292 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1293 req, rc, rep ? rep->lock_policy_res1 : 0);
1296 switch (cleanup_phase) {
1299 ldlm_lock_decref(child_lockh, child_mode);
1300 child_lockh->cookie = 0;
1303 if (rc && created) {
1304 int err = vfs_unlink(dparent->d_inode, dchild);
1306 CERROR("unlink(%.*s) in error path: %d\n",
1307 dchild->d_name.len, dchild->d_name.name,
1310 } else if (created) {
1311 mds_lock_new_child(obd, dchild->d_inode, NULL);
1315 if (dparent == NULL)
1320 if (parent_lockh[1].cookie != 0)
1321 ldlm_lock_decref(parent_lockh + 1, update_mode);
1324 ldlm_lock_decref(parent_lockh, parent_mode);
1326 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1329 OBD_FREE(mea, mea_size);
1331 atomic_inc(&mds->mds_open_count);
1333 mds_body_do_reverse_map(med, body);
1336 * If we have not taken the "open" lock, we may not return 0 here,
1337 * because caller expects 0 to mean "lock is taken", and it needs
1338 * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1339 * client. Later caller should rewrite the return value back to zero
1340 * if it to be used any further.
1342 if ((cleanup_phase != 3) && !rc)
1348 /* Close a "file descriptor" and possibly unlink an orphan from the
1349 * PENDING directory. Caller must hold child->i_sem, this drops it.
1351 * If we are being called from mds_disconnect() because the client has
1352 * disappeared, then req == NULL and we do not update last_rcvd because
1353 * there is nothing that could be recovered by the client at this stage
1354 * (it will not even _have_ an entry in last_rcvd anymore).
1356 * Returns EAGAIN if the client needs to get more data and re-close. */
1357 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1358 struct obd_device *obd, struct mds_file_data *mfd,
1361 struct inode *inode = mfd->mfd_dentry->d_inode;
1362 char idname[LL_ID_NAMELEN];
1363 int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1364 struct dentry *pending_child = NULL;
1365 struct mds_obd *mds = &obd->u.mds;
1366 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1367 void *handle = NULL;
1368 struct mds_body *request_body = NULL, *reply_body = NULL;
1369 struct dentry_params dp;
1370 struct iattr iattr = { 0 };
1371 struct llog_create_locks *lcl = NULL;
1374 if (req && req->rq_reqmsg != NULL)
1375 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1376 sizeof(*request_body));
1377 if (req && req->rq_repmsg != NULL)
1378 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1379 sizeof(*reply_body));
1381 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1382 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1383 idname, inode->i_nlink, mds_orphan_open_count(inode));
1385 last_orphan = mds_orphan_open_dec_test(inode) &&
1386 mds_inode_is_orphan(inode);
1387 UP_WRITE_I_ALLOC_SEM(inode);
1389 /* this is half of the actual "close" */
1390 if (mfd->mfd_mode & FMODE_WRITE) {
1391 rc = mds_put_write_access(mds, inode, request_body,
1392 last_orphan && unlink_orphan);
1393 } else if (mfd->mfd_mode & FMODE_EXEC) {
1394 mds_allow_write_access(inode);
1397 if (last_orphan && unlink_orphan) {
1398 struct lov_mds_md *lmm = NULL;
1399 int stripe_count = 0;
1400 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1402 if (obd->obd_recovering) {
1403 CDEBUG(D_HA, "not remove orphan %s until recovery"
1404 " is over\n", idname);
1408 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1410 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1411 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1412 CERROR("found \"orphan\" %s %s with link count %d\n",
1413 S_ISREG(inode->i_mode) ? "file" : "dir",
1414 idname, inode->i_nlink);
1416 * sadly, there is no easy way to save pending_child from
1417 * mds_reint_unlink() into mfd, so we need to re-lookup, but
1418 * normally it will still be in the dcache.
1420 down(&pending_dir->i_sem);
1421 cleanup_phase = 1; /* up(i_sem) when finished */
1422 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1424 if (IS_ERR(pending_child))
1425 GOTO(cleanup, rc = PTR_ERR(pending_child));
1426 LASSERT(pending_child->d_inode != NULL);
1428 cleanup_phase = 2; /* dput(pending_child) when finished */
1429 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1430 rc = vfs_rmdir(pending_dir, pending_child);
1432 CERROR("error unlinking orphan dir %s: rc %d\n",
1437 if (req != NULL && req->rq_repmsg != NULL) {
1438 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1439 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1442 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1443 NULL, stripe_count);
1444 if (IS_ERR(handle)) {
1445 rc = PTR_ERR(handle);
1450 pending_child->d_fsdata = (void *) &dp;
1453 rc = vfs_unlink(pending_dir, pending_child);
1455 CERROR("error unlinking orphan %s: rc %d\n",
1458 if (req != NULL && req->rq_repmsg != NULL &&
1459 (reply_body->valid & OBD_MD_FLEASIZE) &&
1460 mds_log_op_unlink(obd, pending_child->d_inode,
1461 lmm, req->rq_repmsg->buflens[1],
1462 lustre_msg_buf(req->rq_repmsg, 2, 0),
1463 req->rq_repmsg->buflens[2], &lcl) > 0) {
1464 reply_body->valid |= OBD_MD_FLCOOKIE;
1467 goto out; /* Don't bother updating attrs on unlinked inode */
1471 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1472 /* Update the on-disk attributes if this was the last write
1473 * close, and all information was provided (i.e., rc == 0)
1475 * XXX this should probably be abstracted with mds_reint_setattr
1478 if (request_body->valid & OBD_MD_FLMTIME &&
1479 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1480 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1481 iattr.ia_valid |= ATTR_MTIME;
1483 if (request_body->valid & OBD_MD_FLCTIME &&
1484 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1485 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1486 iattr.ia_valid |= ATTR_CTIME;
1489 /* XXX can't set block count with fsfilt_setattr (!) */
1490 if (request_body->valid & OBD_MD_FLSIZE) {
1491 iattr.ia_valid |= ATTR_SIZE;
1492 iattr.ia_size = request_body->size;
1494 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1495 iattr.ia_valid |= ATTR_BLOCKS;
1496 iattr.ia_blocks = request_body->blocks
1501 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1502 /* Only start a transaction to write out only the atime if
1503 * it is more out-of-date than the specified limit. If we
1504 * are already going to write out the atime then do it anyway.
1506 if ((request_body->atime >
1507 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1508 (iattr.ia_valid != 0 &&
1509 request_body->atime > LTIME_S(inode->i_atime))) {
1510 LTIME_S(iattr.ia_atime) = request_body->atime;
1511 iattr.ia_valid |= ATTR_ATIME;
1515 if (iattr.ia_valid != 0) {
1516 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
1518 GOTO(cleanup, rc = PTR_ERR(handle));
1519 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1521 CERROR("error in setattr(%s): rc %d\n", idname, rc);
1524 /* If other clients have this file open for write, rc will be > 0 */
1527 l_dput(mfd->mfd_dentry);
1528 mds_mfd_destroy(mfd);
1531 atomic_dec(&mds->mds_open_count);
1532 if (req != NULL && reply_body != NULL) {
1533 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1534 } else if (handle) {
1535 int err, force_sync = 0;
1537 if (req && req->rq_export)
1538 force_sync = req->rq_export->exp_sync;
1540 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1543 CERROR("error committing close: %d\n", err);
1549 switch (cleanup_phase) {
1552 ptlrpc_save_llog_lock(req, lcl);
1553 dput(pending_child);
1555 up(&pending_dir->i_sem);
1560 int mds_close(struct ptlrpc_request *req, int offset)
1562 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1563 struct obd_device *obd = req->rq_export->exp_obd;
1564 struct mds_body *body;
1565 struct mds_file_data *mfd;
1566 struct lvfs_run_ctxt saved;
1567 struct inode *inode;
1568 int rc, repsize[3] = {sizeof(struct mds_body),
1569 obd->u.mds.mds_max_mdsize,
1570 obd->u.mds.mds_max_cookiesize};
1572 MD_COUNTER_INCREMENT(obd, close);
1574 rc = lustre_pack_reply(req, 3, repsize, NULL);
1576 CERROR("lustre_pack_reply: rc = %d\n", rc);
1577 req->rq_status = rc;
1579 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1582 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1583 DEBUG_REQ(D_HA, req, "close replay\n");
1584 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1585 lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1586 req->rq_repmsg->buflens[2]);
1589 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1590 lustre_swab_mds_body);
1592 CERROR("Can't unpack body\n");
1593 req->rq_status = -EFAULT;
1597 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1598 /* do some stuff */ ;
1600 mfd = mds_handle2mfd(&body->handle);
1602 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1603 ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1604 req->rq_status = -ESTALE;
1608 inode = mfd->mfd_dentry->d_inode;
1609 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1610 DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1611 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1612 struct mds_body *rep_body;
1614 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1615 sizeof (*rep_body));
1616 LASSERT(rep_body != NULL);
1618 mds_pack_inode2body(obd, rep_body, inode,
1619 (body->valid & OBD_MD_FID) ? 1 : 0);
1621 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1622 inode, MDS_PACK_MD_LOCK, 0);
1624 spin_lock(&med->med_open_lock);
1625 list_del(&mfd->mfd_list);
1626 spin_unlock(&med->med_open_lock);
1628 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1629 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1630 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1632 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1633 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1634 req->rq_status = -ENOMEM;
1643 int mds_done_writing(struct ptlrpc_request *req, int offset)
1645 struct mds_body *body;
1646 int rc, size = sizeof(struct mds_body);
1649 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1651 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1652 lustre_swab_mds_body);
1654 CERROR("Can't unpack body\n");
1655 req->rq_status = -EFAULT;
1659 rc = lustre_pack_reply(req, 1, &size, NULL);
1661 CERROR("lustre_pack_reply: rc = %d\n", rc);
1662 req->rq_status = rc;