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 static void mds_mfd_put(struct mds_file_data *mfd)
99 CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
100 atomic_read(&mfd->mfd_refcount) - 1);
101 LASSERT(atomic_read(&mfd->mfd_refcount) > 0 &&
102 atomic_read(&mfd->mfd_refcount) < 0x5a5a);
103 if (atomic_dec_and_test(&mfd->mfd_refcount)) {
104 LASSERT(list_empty(&mfd->mfd_handle.h_link));
105 OBD_FREE(mfd, sizeof *mfd);
109 static void mds_mfd_destroy(struct mds_file_data *mfd)
111 class_handle_unhash(&mfd->mfd_handle);
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;
175 MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
176 CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
177 mds->mds_io_epoch, inode->i_ino, inode->i_generation);
180 atomic_inc(&inode->i_writecount);
181 up(&mds->mds_epoch_sem);
185 /* Returns EAGAIN if the client needs to get size and/or cookies and close
186 * again -- which is never true if the file is about to be unlinked. Otherwise
187 * returns the number of remaining writers. */
188 static int mds_put_write_access(struct mds_obd *mds, struct inode *inode,
189 struct mds_body *body, int unlinking)
194 down(&mds->mds_epoch_sem);
195 atomic_dec(&inode->i_writecount);
196 rc = atomic_read(&inode->i_writecount);
200 if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
201 GOTO(out, rc = EAGAIN);
203 mds_free_filterdata(inode);
205 up(&mds->mds_epoch_sem);
209 static int mds_deny_write_access(struct mds_obd *mds, struct inode *inode)
212 down(&mds->mds_epoch_sem);
213 if (atomic_read(&inode->i_writecount) > 0) {
214 up(&mds->mds_epoch_sem);
217 atomic_dec(&inode->i_writecount);
218 up(&mds->mds_epoch_sem);
222 static void mds_allow_write_access(struct inode *inode)
225 atomic_inc(&inode->i_writecount);
228 int mds_query_write_access(struct inode *inode)
231 RETURN(atomic_read(&inode->i_writecount));
234 /* This replaces the VFS dentry_open, it manages mfd and writecount */
235 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
236 struct vfsmount *mnt, int flags,
237 struct ptlrpc_request *req)
239 struct mds_export_data *med = &req->rq_export->exp_mds_data;
240 struct mds_obd *mds = mds_req2mds(req);
241 struct mds_file_data *mfd;
242 struct mds_body *body;
248 CERROR("mds: out of memory\n");
249 GOTO(cleanup_dentry, rc = -ENOMEM);
252 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
254 if (flags & FMODE_WRITE) {
255 /* FIXME: in recovery, need to pass old epoch here */
256 rc = mds_get_write_access(mds, dentry->d_inode, 0);
258 GOTO(cleanup_mfd, rc);
259 body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
260 } else if (flags & FMODE_EXEC) {
261 rc = mds_deny_write_access(mds, dentry->d_inode);
263 GOTO(cleanup_mfd, rc);
268 /* FIXME: invalidate update locks when first open for write comes in */
270 /* mark the file as open to handle open-unlink. */
271 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
272 mds_orphan_open_inc(dentry->d_inode);
273 UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
275 mfd->mfd_mode = flags;
276 mfd->mfd_dentry = dentry;
277 mfd->mfd_xid = req->rq_xid;
279 spin_lock(&med->med_open_lock);
280 list_add(&mfd->mfd_list, &med->med_open_head);
281 spin_unlock(&med->med_open_lock);
284 body->handle.cookie = mfd->mfd_handle.h_cookie;
288 mds_mfd_destroy(mfd);
294 mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
295 struct lov_desc *desc)
299 for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
300 ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
301 le64_to_cpu(lmm->lmm_objects[i].l_object_id);
305 /* must be called with i_sem held */
307 mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req,
308 int offset, struct mds_update_record *rec,
309 struct dentry *dchild, void **handle,
312 struct inode *inode = dchild->d_inode;
313 struct mds_obd *mds = &obd->u.mds;
314 struct obd_trans_info oti = { 0 };
315 struct lov_stripe_md *lsm = NULL;
316 struct lov_mds_md *lmm = NULL;
317 int rc, lmm_bufsize, lmm_size;
318 struct obdo *oa = NULL;
319 struct mds_body *body;
323 if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
324 !(rec->ur_flags & FMODE_WRITE))
327 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
329 if (!S_ISREG(inode->i_mode))
331 if (body->valid & OBD_MD_FLEASIZE)
334 OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
337 oti.oti_objid = *ids;
340 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
341 LASSERT(id_ino(rec->ur_id2));
342 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
343 lmm_size = rec->ur_eadatalen;
344 lmm = rec->ur_eadata;
348 *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
349 if (IS_ERR(*handle)) {
350 rc = PTR_ERR(*handle);
355 mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
357 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
358 lmm_bufsize = req->rq_repmsg->buflens[offset];
359 LASSERT(lmm_buf != NULL);
360 LASSERT(lmm_bufsize >= lmm_size);
362 memcpy(lmm_buf, lmm, lmm_size);
363 rc = fsfilt_set_md(obd, inode, *handle, lmm,
366 CERROR("open replay failed to set md:%d\n", rc);
370 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
371 GOTO(out_ids, rc = -ENOMEM);
376 oa->o_mode = S_IFREG | 0600;
377 oa->o_id = inode->i_ino;
378 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
379 oa->o_generation = inode->i_generation;
380 oa->o_uid = 0; /* must have 0 uid / gid on OST */
382 oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE |
383 OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
386 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
387 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
389 if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
390 /* check if things like lfs setstripe are sending us the ea */
391 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
392 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
394 0, &lsm, rec->ur_eadata);
398 OBD_ALLOC(lmm, mds->mds_max_mdsize);
400 GOTO(out_oa, rc = -ENOMEM);
402 lmm_size = mds->mds_max_mdsize;
403 rc = mds_get_md(obd, dchild->d_parent->d_inode,
404 lmm, &lmm_size, 1, 0);
406 rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
409 OBD_FREE(lmm, mds->mds_max_mdsize);
414 LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
415 oti.oti_flags |= OBD_MODE_CROW;
416 rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
419 CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR),
420 "error creating objects for "
421 "inode %lu: rc = %d\n",
424 CERROR("obd_create returned invalid "
431 rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
432 0, &lsm, rec->ur_eadata);
436 lsm->lsm_object_id = oa->o_id;
437 lsm->lsm_object_gr = oa->o_gr;
440 oa->o_size = inode->i_size;
441 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
442 OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE);
444 rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
446 CERROR("error setting attrs for inode %lu: rc %d\n",
449 CERROR("obd_setattr returned bad rc %d\n", rc);
456 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
457 obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ);
459 LASSERT(lsm && lsm->lsm_object_id);
461 rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
462 if (!id_ino(rec->ur_id2))
463 obd_free_memmd(mds->mds_dt_exp, &lsm);
465 CERROR("cannot pack lsm, err = %d\n", rc);
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_object(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 CDEBUG(D_INODE, "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 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 obd_device *obd = req->rq_export->exp_obd;
700 struct mds_obd *mds = mds_req2mds(req);
701 struct mds_file_data *mfd = NULL;
704 int rc = 0, reply_off;
707 /* atomically create objects if necessary */
708 down(&dchild->d_inode->i_sem);
709 mode = dchild->d_inode->i_mode;
711 if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
712 (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
713 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
714 dchild->d_inode, 0, 0);
716 up(&dchild->d_inode->i_sem);
722 if ((body->valid & OBD_MD_FLEASIZE) &&
723 (rec->ur_flags & MDS_OPEN_HAS_EA)) {
724 up(&dchild->d_inode->i_sem);
728 if (!(body->valid & OBD_MD_FLEASIZE)) {
729 /* no EA: create objects */
730 rc = mds_create_objects(obd, req, 2, rec,
731 dchild, handle, &ids);
733 CERROR("mds_create_object: rc = %d\n", rc);
734 up(&dchild->d_inode->i_sem);
739 if (S_ISREG(dchild->d_inode->i_mode) &&
740 (body->valid & OBD_MD_FLEASIZE)) {
741 rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
744 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
745 dchild->d_inode, 0, 0);
747 up(&dchild->d_inode->i_sem);
754 rc = mds_pack_acl(req, reply_off, body, dchild->d_inode);
758 CERROR("pack posix acl: rc = %d\n", rc);
759 up(&dchild->d_inode->i_sem);
763 rc = mds_pack_gskey(obd, req->rq_repmsg, &reply_off, body,
766 CERROR("mds_pack_gskey: rc = %d\n", rc);
767 up(&dchild->d_inode->i_sem);
772 struct lustre_capa capa = {
773 .lc_uid = rec->ur_uc.luc_uid,
774 .lc_op = capa_op(rec->ur_flags),
775 .lc_ino = dchild->d_inode->i_ino,
776 .lc_mdsid = mds->mds_num,
779 rc = mds_pack_capa(obd, NULL, &capa, req->rq_repmsg,
782 CERROR("mds_pack_capa: rc = %d\n", rc);
783 up(&dchild->d_inode->i_sem);
790 /* If the inode has no EA data, then MDSs hold size, mtime */
791 if (S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) {
792 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
793 OBD_MD_FLATIME | OBD_MD_FLMTIME);
796 up(&dchild->d_inode->i_sem);
798 intent_set_disposition(rep, DISP_OPEN_OPEN);
799 mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
801 RETURN(PTR_ERR(mfd));
803 CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
804 mfd->mfd_handle.h_cookie);
807 mds_dt_update_objids(obd, ids);
808 OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
814 static int mds_open_by_id(struct ptlrpc_request *req,
815 struct lustre_id *id,
816 struct mds_body *body, int flags,
817 struct mds_update_record *rec,
818 struct ldlm_reply *rep)
820 struct mds_obd *mds = mds_req2mds(req);
821 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
822 struct dentry *dchild;
823 char idname[LL_ID_NAMELEN];
828 down(&pending_dir->i_sem);
829 idlen = ll_id2str(idname, id_ino(id), id_gen(id));
830 dchild = lookup_one_len(idname, mds->mds_pending_dir,
832 if (IS_ERR(dchild)) {
833 up(&pending_dir->i_sem);
834 rc = PTR_ERR(dchild);
835 CERROR("error looking up %s in PENDING: rc = %d\n",
840 up(&pending_dir->i_sem);
841 if (dchild->d_inode != NULL) {
842 mds_inode_set_orphan(dchild->d_inode);
843 mds_pack_inode2body(req2obd(req), body,
846 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
847 intent_set_disposition(rep, DISP_LOOKUP_POS);
848 CWARN("Orphan %s found and opened in PENDING directory\n",
854 /* we didn't find it in PENDING so it isn't an orphan. See if it was a
855 * regular inode that was previously created. */
856 dchild = mds_id2dentry(req2obd(req), id, NULL);
858 RETURN(PTR_ERR(dchild));
860 mds_pack_inode2body(req2obd(req), body,
863 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
864 intent_set_disposition(rep, DISP_LOOKUP_POS);
867 rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep);
868 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
869 req, rc, rep ? rep->lock_policy_res1 : 0);
870 /* XXX what do we do here if mds_finish_transno itself failed? */
875 int mds_pin(struct ptlrpc_request *req, int offset)
877 struct obd_device *obd = req->rq_export->exp_obd;
878 struct mds_body *request_body, *reply_body;
879 struct lvfs_run_ctxt saved;
880 int rc, size = sizeof(*reply_body);
883 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
884 sizeof(*request_body));
886 rc = lustre_pack_reply(req, 1, &size, NULL);
889 reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
891 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
892 rc = mds_open_by_id(req, &request_body->id1, reply_body,
893 request_body->flags, NULL, NULL);
894 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
900 * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
901 * child_lockh is NULL we just get the lock as a barrier to wait for other
902 * holders of this lock, and drop it right away again.
904 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
905 struct lustre_handle *child_lockh)
907 struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
908 struct lustre_handle lockh;
909 int lock_flags = LDLM_FL_ATOMIC_CB;
913 if (child_lockh == NULL)
914 child_lockh = &lockh;
916 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
917 LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
918 mds_blocking_ast, ldlm_completion_ast, NULL,
919 NULL, NULL, 0, NULL, child_lockh);
921 CERROR("ldlm_cli_enqueue: %d\n", rc);
922 else if (child_lockh == &lockh)
923 ldlm_lock_decref(child_lockh, LCK_EX);
928 int mds_open(struct mds_update_record *rec, int offset,
929 struct ptlrpc_request *req, struct lustre_handle *child_lockh)
931 /* XXX ALLOCATE _something_ - 464 bytes on stack here */
932 struct obd_device *obd = req->rq_export->exp_obd;
933 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
934 struct mds_obd *mds = mds_req2mds(req);
935 struct ldlm_reply *rep = NULL;
936 struct mds_body *body = NULL;
937 struct dentry *dchild = NULL, *dparent = NULL;
938 struct lustre_handle parent_lockh[2] = {{0}, {0}};
939 int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
940 int parent_mode = LCK_PR;
942 struct dentry_params dp;
943 struct mea *mea = NULL;
944 int mea_size, update_mode;
945 int child_mode = LCK_PR;
946 struct lustre_id sid;
950 DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
951 OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
954 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE,
955 (obd_timeout + 1) / 4);
957 if (offset == 3) { /* intent */
958 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
959 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
960 } else if (offset == 1) { /* non-intent reint */
961 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
967 MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
970 * Step 0: If we are passed a id, then we assume the client already
971 * opened this file and is only replaying the RPC, so we open the inode
972 * by fid (at some large expense in security).
974 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
975 CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
976 OLID4(rec->ur_id2), rec->ur_namelen - 1,
977 rec->ur_name, rec->ur_mode);
979 LASSERT(id_ino(rec->ur_id2));
981 rc = mds_open_by_id(req, rec->ur_id2, body,
982 rec->ur_flags, rec, rep);
984 mds_body_do_reverse_map(med, body);
988 /* We didn't find the correct inode on disk either, so we
989 * need to re-create it via a regular replay. */
990 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
992 LASSERT(!id_ino(rec->ur_id2));
995 LASSERT(offset == 3); /* If we got here, we must be called via intent */
997 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
998 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
1002 acc_mode = accmode(rec->ur_flags);
1004 /* Step 1: Find and lock the parent */
1005 if (rec->ur_flags & MDS_OPEN_CREAT) {
1006 /* XXX Well, in fact we only need this lock mode change if
1007 in addition to O_CREAT, the file does not exist.
1008 But we do not know if it exists or not yet */
1009 parent_mode = LCK_PW;
1012 if (rec->ur_namelen == 1) {
1013 /* client (LMV) wants to open the file by id */
1014 CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
1015 dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
1016 if (IS_ERR(dchild)) {
1017 rc = PTR_ERR(dparent);
1018 CERROR("child lookup by an id error %d\n", rc);
1025 dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
1026 parent_lockh, &update_mode, rec->ur_name,
1027 rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
1028 if (IS_ERR(dparent)) {
1029 rc = PTR_ERR(dparent);
1030 if (rc != -ENOENT) {
1031 CERROR("parent lookup for "DLID4" failed, error %d\n",
1032 OLID4(rec->ur_id1), rc);
1034 /* Just cannot find parent - make it look like
1035 usual negative lookup to avoid extra MDS RPC */
1036 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1037 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1041 LASSERT(dparent->d_inode != NULL);
1043 cleanup_phase = 1; /* parent dentry and lock */
1045 /* try to retrieve MEA data for this dir */
1046 rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
1050 if (mea != NULL && mea->mea_count) {
1052 * dir is already splitted, check is requested filename should *
1053 * live at this MDS or at another one.
1056 i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
1057 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1059 "%s: inapropriate MDS(%d) for %lu/%u:%s."
1060 " should be %lu(%d)\n", obd->obd_name,
1061 mea->mea_master, dparent->d_inode->i_ino,
1062 dparent->d_inode->i_generation, rec->ur_name,
1063 (unsigned long)id_group(&mea->mea_ids[i]), i);
1064 GOTO(cleanup, rc = -ERESTART);
1068 /* Step 2: Lookup the child */
1069 dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
1070 if (IS_ERR(dchild)) {
1071 rc = PTR_ERR(dchild);
1072 dchild = NULL; /* don't confuse mds_finish_transno() below */
1077 cleanup_phase = 2; /* child dentry */
1079 if (dchild->d_flags & DCACHE_CROSS_REF) {
1080 CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
1081 LASSERT(rec->ur_namelen > 1);
1083 /* we're gonna acquire LOOKUP lock on the child,
1084 * but we have already locked parent and our order
1085 * may conflict with enqueue_order_locks(). so,
1086 * drop parent lock and acquire both the locks in
1087 * common order. bug 6190 */
1089 if (parent_lockh[1].cookie != 0)
1090 ldlm_lock_decref(parent_lockh + 1, update_mode);
1092 ldlm_lock_decref(parent_lockh, parent_mode);
1097 rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1,
1098 parent_lockh, &dparent,
1099 LCK_PR, MDS_INODELOCK_UPDATE,
1100 &update_mode, rec->ur_name,
1101 rec->ur_namelen, child_lockh,
1103 MDS_INODELOCK_LOOKUP);
1109 if (parent_lockh[1].cookie != 0)
1110 ldlm_lock_decref(parent_lockh + 1, update_mode);
1112 ldlm_lock_decref(parent_lockh, LCK_PR);
1114 if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
1115 CDEBUG(D_OTHER, "race: name changed (%p)\n",
1117 if (dchild->d_inode)
1118 ldlm_lock_decref(child_lockh, LCK_PR);
1124 mds_pack_dentry2body(obd, body, dchild, 1);
1125 intent_set_disposition(rep, DISP_LOOKUP_POS);
1126 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1129 OBD_FREE(mea, mea_size);
1135 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1137 if (dchild->d_inode)
1138 intent_set_disposition(rep, DISP_LOOKUP_POS);
1140 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1142 /* Step 3: If the child was negative, and we're supposed to, create it.*/
1143 if (dchild->d_inode == NULL) {
1146 struct inode *inode;
1148 /* get parent id: ldlm lock on the parent protects ea */
1149 rc = mds_read_inode_sid(obd, dparent->d_inode, &sid);
1151 CERROR("can't read parent inode id. ino(%lu) rc(%d)\n",
1152 dparent->d_inode->i_ino, rc);
1156 ino = (unsigned long)id_ino(rec->ur_id2);
1158 rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
1161 CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
1162 obd->obd_name, dparent->d_inode->i_ino,
1163 dparent->d_inode->i_generation, rc);
1166 /* dir got splitted */
1167 GOTO(cleanup, rc = -ERESTART);
1168 } else if (rc < 0) {
1169 /* error happened during spitting */
1173 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
1175 * dentry is negative and we weren't supposed to create
1176 * object, get out of here.
1178 GOTO(cleanup, rc = -ENOENT);
1181 intent_set_disposition(rep, DISP_OPEN_CREATE);
1182 handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE,
1184 if (IS_ERR(handle)) {
1185 rc = PTR_ERR(handle);
1189 if (id_fid(rec->ur_id2))
1190 fid = id_fid(rec->ur_id2);
1192 fid = mds_alloc_fid(obd);
1194 dchild->d_fsdata = (void *) &dp;
1199 dp.p_group = mds->mds_num;
1201 rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
1202 if (dchild->d_fsdata == (void *)(unsigned long)ino)
1203 dchild->d_fsdata = NULL;
1206 CDEBUG(D_INODE, "error during create: %d\n", rc);
1210 inode = dchild->d_inode;
1213 LASSERT(ino == inode->i_ino);
1215 /* written as part of setattr */
1216 inode->i_generation = id_gen(rec->ur_id2);
1217 CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
1218 inode->i_ino, inode->i_generation);
1222 LTIME_S(iattr.ia_atime) = rec->ur_time;
1223 LTIME_S(iattr.ia_ctime) = rec->ur_time;
1224 LTIME_S(iattr.ia_mtime) = rec->ur_time;
1226 iattr.ia_uid = rec->ur_fsuid;
1227 if (dparent->d_inode->i_mode & S_ISGID)
1228 iattr.ia_gid = dparent->d_inode->i_gid;
1230 iattr.ia_gid = rec->ur_fsgid;
1232 iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME |
1233 ATTR_MTIME | ATTR_CTIME;
1235 rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0);
1237 CERROR("error on child setattr: rc = %d\n", rc);
1239 iattr.ia_valid = ATTR_MTIME | ATTR_CTIME;
1241 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
1243 CERROR("error on parent setattr: rc = %d\n", rc);
1245 MD_COUNTER_INCREMENT(obd, create);
1248 mds_inode2id(obd, &body->id1, inode, fid);
1249 mds_update_inode_ids(obd, dchild->d_inode, handle, &body->id1, &sid);
1251 if ((rec->ur_flags & MDS_OPEN_HAS_KEY)) {
1252 rc = mds_set_gskey(obd, handle, dchild->d_inode,
1253 rec->ur_ea2data, rec->ur_ea2datalen,
1254 ATTR_KEY | ATTR_MAC);
1256 CERROR("error in set gs key rc %d\n", rc);
1260 if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
1261 rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
1262 dchild->d_inode, handle,
1263 req->rq_export->exp_sync);
1267 acc_mode = 0; /* Don't check for permissions */
1269 mds_pack_inode2body(obd, body, dchild->d_inode, 1);
1271 LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
1272 "dchild %.*s (%p) inode %p\n", dchild->d_name.len,
1273 dchild->d_name.name, dchild, dchild->d_inode);
1275 if (S_ISREG(dchild->d_inode->i_mode)) {
1276 /* Check permissions etc */
1277 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1281 /* Can't write to a read-only file */
1282 if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0)
1283 GOTO(cleanup, rc = -EPERM);
1285 /* An append-only file must be opened in append mode for
1287 if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 &&
1288 ((rec->ur_flags & MDS_OPEN_APPEND) == 0 ||
1289 (rec->ur_flags & MDS_OPEN_TRUNC) != 0))
1290 GOTO(cleanup, rc = -EPERM);
1293 if (!created && (rec->ur_flags & MDS_OPEN_CREAT) &&
1294 (rec->ur_flags & MDS_OPEN_EXCL)) {
1295 /* File already exists, we didn't just create it, and we
1296 * were passed O_EXCL; err-or. */
1297 GOTO(cleanup, rc = -EEXIST); // returns a lock to the client
1300 if (S_ISDIR(dchild->d_inode->i_mode)) {
1301 if (rec->ur_flags & MDS_OPEN_CREAT ||
1302 rec->ur_flags & FMODE_WRITE) {
1303 /* we are trying to create or write a exist dir */
1304 GOTO(cleanup, rc = -EISDIR);
1306 if (ll_permission(dchild->d_inode, acc_mode, NULL))
1307 GOTO(cleanup, rc = -EACCES);
1310 * here was checking for possible GNS mount points to skip their
1311 * open. I removed it as its detection based only on SUID bit is
1312 * not reliable. Opening such a dirs should not cause any
1313 * problems, at least tests show that. --umka
1317 /* if we are following a symlink, don't open */
1318 if (S_ISLNK(dchild->d_inode->i_mode))
1319 GOTO(cleanup_no_trans, rc = 0);
1321 if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
1322 !S_ISDIR(dchild->d_inode->i_mode))
1323 GOTO(cleanup, rc = -ENOTDIR);
1325 /* check permission even it's special files */
1326 if (S_ISCHR(dchild->d_inode->i_mode) ||
1327 S_ISBLK(dchild->d_inode->i_mode) ||
1328 S_ISFIFO(dchild->d_inode->i_mode) ||
1329 S_ISSOCK(dchild->d_inode->i_mode)) {
1330 rc = ll_permission(dchild->d_inode, acc_mode, NULL);
1335 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
1336 obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
1337 GOTO(cleanup, rc = -EAGAIN);
1341 /* We cannot use acc_mode here, because it is zeroed in case of
1342 creating a file, so we get wrong lockmode */
1343 if (accmode(rec->ur_flags) & MAY_WRITE)
1344 child_mode = LCK_CW;
1345 else if (accmode(rec->ur_flags) & MAY_EXEC)
1346 child_mode = LCK_PR;
1348 child_mode = LCK_CR;
1350 /* Always returning LOOKUP lock if open succesful to guard
1351 * dentry on client. In case of replay we do not get a lock
1352 * assuming that the caller has it already */
1353 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1354 struct ldlm_res_id child_res_id = { .name = {0}};
1355 ldlm_policy_data_t policy;
1356 int lock_flags = LDLM_FL_ATOMIC_CB;
1358 /* LOOKUP lock will protect dentry on client -bzzz */
1359 policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
1362 child_res_id.name[0] = id_fid(&body->id1);
1363 child_res_id.name[1] = id_group(&body->id1);
1365 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1366 child_res_id, LDLM_IBITS, &policy,
1367 child_mode, &lock_flags,
1368 mds_blocking_ast, ldlm_completion_ast,
1369 NULL, NULL, NULL, 0, NULL, child_lockh);
1376 /* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */
1377 #warning "disable opencache lock for CMD2"
1380 /* Step 5: mds_open it */
1381 rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
1386 /* reintegration case */
1387 if ((rec->ur_flags & MDS_REINT_REQ)) {
1388 rc = mds_fidmap_add(obd, &body->id1);
1390 CERROR("can't create fid->ino mapping, err %d\n",
1397 /* if this is a writer, we have to invalidate client's
1398 * update locks in order to make sure they don't use
1399 * isize/iblocks from mds anymore.
1400 * FIXME: can cause a deadlock, use mds_get_parent_child_locked()
1401 * XXX: optimization is to do this for first writer only */
1402 if (accmode(rec->ur_flags) & MAY_WRITE) {
1403 struct ldlm_res_id child_res_id = { .name = {0}};
1404 ldlm_policy_data_t sz_policy;
1405 struct lustre_handle sz_lockh;
1406 int lock_flags = LDLM_FL_ATOMIC_CB;
1408 child_res_id.name[0] = id_fid(&body->id1);
1409 child_res_id.name[1] = id_group(&body->id1);
1410 sz_policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1412 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1413 child_res_id, LDLM_IBITS, &sz_policy,
1414 LCK_PW, &lock_flags, mds_blocking_ast,
1415 ldlm_completion_ast, NULL, NULL, NULL,
1416 0, NULL, &sz_lockh);
1418 ldlm_lock_decref(&sz_lockh, LCK_PW);
1420 CERROR("can't invalidate client's update locks\n");
1425 rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle,
1426 req, rc, rep ? rep->lock_policy_res1 : 0);
1429 switch (cleanup_phase) {
1432 ldlm_lock_decref(child_lockh, child_mode);
1433 child_lockh->cookie = 0;
1436 if (rc && created) {
1437 int err = vfs_unlink(dparent->d_inode, dchild);
1439 CERROR("unlink(%.*s) in error path: %d\n",
1440 dchild->d_name.len, dchild->d_name.name,
1443 } else if (created) {
1444 mds_lock_new_child(obd, dchild->d_inode, NULL);
1446 /* audit stuff for OPEN */
1447 if (offset == 3 && (dchild->d_inode || dparent)) {
1448 struct lustre_id au_id;
1449 struct inode * au_inode = dchild->d_inode;
1451 if (au_inode == NULL)
1452 au_inode = dparent->d_inode;
1455 mds_inode2id(obd, &au_id, au_inode, fid);
1457 au_id = *(rec->ur_id1);
1458 mds_audit_open(req, &au_id, au_inode,
1459 rec->ur_name, rec->ur_namelen, rc);
1464 if (dparent == NULL)
1469 if (parent_lockh[1].cookie != 0)
1470 ldlm_lock_decref(parent_lockh + 1, update_mode);
1473 ldlm_lock_decref(parent_lockh, parent_mode);
1475 ptlrpc_save_lock (req, parent_lockh, parent_mode);
1478 OBD_FREE(mea, mea_size);
1480 atomic_inc(&mds->mds_open_count);
1482 mds_body_do_reverse_map(med, body);
1485 * If we have not taken the "open" lock, we may not return 0 here,
1486 * because caller expects 0 to mean "lock is taken", and it needs
1487 * nonzero return here for caller to return EDLM_LOCK_ABORTED to
1488 * client. Later caller should rewrite the return value back to zero
1489 * if it to be used any further.
1491 if ((cleanup_phase != 3) && !rc)
1497 /* Close a "file descriptor" and possibly unlink an orphan from the
1498 * PENDING directory. Caller must hold child->i_sem, this drops it.
1500 * If we are being called from mds_disconnect() because the client has
1501 * disappeared, then req == NULL and we do not update last_rcvd because
1502 * there is nothing that could be recovered by the client at this stage
1503 * (it will not even _have_ an entry in last_rcvd anymore).
1505 * Returns EAGAIN if the client needs to get more data and re-close. */
1506 int mds_mfd_close(struct ptlrpc_request *req, int offset,
1507 struct obd_device *obd, struct mds_file_data *mfd,
1510 struct inode *inode = mfd->mfd_dentry->d_inode;
1511 char idname[LL_ID_NAMELEN];
1512 int last_orphan, idlen, rc = 0, cleanup_phase = 0;
1513 struct dentry *pending_child = NULL;
1514 struct mds_obd *mds = &obd->u.mds;
1515 struct inode *pending_dir = mds->mds_pending_dir->d_inode;
1516 void *handle = NULL;
1517 struct mds_body *request_body = NULL, *reply_body = NULL;
1518 struct dentry_params dp;
1519 struct iattr iattr = { 0 };
1520 struct llog_create_locks *lcl = NULL;
1523 if (req && req->rq_reqmsg != NULL)
1524 request_body = lustre_msg_buf(req->rq_reqmsg, offset,
1525 sizeof(*request_body));
1526 if (req && req->rq_repmsg != NULL)
1527 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
1528 sizeof(*reply_body));
1530 if (request_body && (request_body->valid & OBD_MD_FLSIZE)) {
1531 /* we set i_size/i_blocks here, nobody will see
1532 * them until all write references are dropped.
1533 * btw, we hold one reference */
1534 LASSERT(mfd->mfd_mode & FMODE_WRITE);
1535 i_size_write(inode, request_body->size);
1536 inode->i_blocks = request_body->blocks;
1537 iattr.ia_size = inode->i_size;
1538 iattr.ia_valid |= ATTR_SIZE;
1539 mds_inode_unset_attrs_old(inode);
1542 idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
1543 CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
1544 idname, inode->i_nlink, mds_orphan_open_count(inode));
1546 last_orphan = mds_orphan_open_dec_test(inode) &&
1547 mds_inode_is_orphan(inode);
1548 UP_WRITE_I_ALLOC_SEM(inode);
1550 /* this is half of the actual "close" */
1551 if (mfd->mfd_mode & FMODE_WRITE) {
1552 rc = mds_put_write_access(mds, inode, request_body,
1553 last_orphan && unlink_orphan);
1554 } else if (mfd->mfd_mode & FMODE_EXEC) {
1555 mds_allow_write_access(inode);
1558 if (last_orphan && unlink_orphan) {
1559 struct lov_mds_md *lmm = NULL;
1560 int stripe_count = 0;
1561 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
1563 CDEBUG(D_HA, "destroying orphan object %s\n", idname);
1565 if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
1566 (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
1567 CERROR("found \"orphan\" %s %s with link count %d\n",
1568 S_ISREG(inode->i_mode) ? "file" : "dir",
1569 idname, inode->i_nlink);
1571 * sadly, there is no easy way to save pending_child from
1572 * mds_reint_unlink() into mfd, so we need to re-lookup, but
1573 * normally it will still be in the dcache.
1575 down(&pending_dir->i_sem);
1576 cleanup_phase = 1; /* up(i_sem) when finished */
1577 pending_child = lookup_one_len(idname, mds->mds_pending_dir,
1579 if (IS_ERR(pending_child))
1580 GOTO(cleanup, rc = PTR_ERR(pending_child));
1581 if (pending_child->d_inode == NULL) {
1582 CERROR("orphan %s has been removed\n", idname);
1583 GOTO(cleanup, rc = 0);
1586 cleanup_phase = 2; /* dput(pending_child) when finished */
1587 if (S_ISDIR(pending_child->d_inode->i_mode)) {
1588 rc = vfs_rmdir(pending_dir, pending_child);
1590 CERROR("error unlinking orphan dir %s: rc %d\n",
1595 if (req != NULL && req->rq_repmsg != NULL) {
1596 lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
1597 stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
1600 handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
1601 NULL, stripe_count);
1602 if (IS_ERR(handle)) {
1603 rc = PTR_ERR(handle);
1608 pending_child->d_fsdata = (void *) &dp;
1611 rc = vfs_unlink(pending_dir, pending_child);
1613 CERROR("error unlinking orphan %s: rc %d\n",
1616 if (req != NULL && req->rq_repmsg != NULL &&
1617 (reply_body->valid & OBD_MD_FLEASIZE) &&
1618 mds_log_op_unlink(obd, pending_child->d_inode,
1619 lmm, req->rq_repmsg->buflens[1],
1620 lustre_msg_buf(req->rq_repmsg, 2, 0),
1621 req->rq_repmsg->buflens[2], &lcl) > 0) {
1622 reply_body->valid |= OBD_MD_FLCOOKIE;
1625 rc = mds_destroy_object(obd, inode, 1);
1627 CERROR("cannot destroy OSS object on close, err %d\n",
1632 goto out; /* Don't bother updating attrs on unlinked inode */
1633 } else if ((mfd->mfd_mode & FMODE_WRITE) && rc == 0) {
1634 /* last writer closed file - let's update i_size/i_blocks */
1635 mds_validate_size(obd, inode, request_body, &iattr);
1639 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
1640 /* Update the on-disk attributes if this was the last write
1641 * close, and all information was provided (i.e., rc == 0)
1643 * XXX this should probably be abstracted with mds_reint_setattr
1646 if (request_body->valid & OBD_MD_FLMTIME &&
1647 LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
1648 LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
1649 iattr.ia_valid |= ATTR_MTIME;
1651 if (request_body->valid & OBD_MD_FLCTIME &&
1652 LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
1653 LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
1654 iattr.ia_valid |= ATTR_CTIME;
1657 /* XXX can't set block count with fsfilt_setattr (!) */
1658 if (request_body->valid & OBD_MD_FLSIZE) {
1659 iattr.ia_valid |= ATTR_SIZE;
1660 iattr.ia_size = request_body->size;
1662 /* if (request_body->valid & OBD_MD_FLBLOCKS) {
1663 iattr.ia_valid |= ATTR_BLOCKS;
1664 iattr.ia_blocks = request_body->blocks
1669 if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
1670 /* Only start a transaction to write out only the atime if
1671 * it is more out-of-date than the specified limit. If we
1672 * are already going to write out the atime then do it anyway.
1674 if ((request_body->atime >
1675 LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
1676 (iattr.ia_valid != 0 &&
1677 request_body->atime > LTIME_S(inode->i_atime))) {
1678 LTIME_S(iattr.ia_atime) = request_body->atime;
1679 iattr.ia_valid |= ATTR_ATIME;
1683 if (iattr.ia_valid != 0) {
1685 handle = fsfilt_start(obd,inode,FSFILT_OP_SETATTR,NULL);
1687 GOTO(cleanup, rc = PTR_ERR(handle));
1688 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
1690 CERROR("error in setattr(%s): rc %d\n", idname, rc);
1693 /* If other clients have this file open for write, rc will be > 0 */
1696 if (!obd->obd_recovering && mds_inode_has_old_attrs(inode)
1697 && !mds_inode_is_orphan(inode)
1698 && atomic_read(&inode->i_writecount) == 0
1699 && inode->i_nlink != 0) {
1700 CERROR("leave inode %lu/%u with old attributes (nlink = %d)\n",
1701 inode->i_ino, inode->i_generation, inode->i_nlink);
1703 l_dput(mfd->mfd_dentry);
1704 mds_mfd_destroy(mfd);
1707 atomic_dec(&mds->mds_open_count);
1708 if (req != NULL && reply_body != NULL) {
1709 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
1710 } else if (handle) {
1711 int err, force_sync = 0;
1713 if (req && req->rq_export)
1714 force_sync = req->rq_export->exp_sync;
1716 err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle,
1719 CERROR("error committing close: %d\n", err);
1725 switch (cleanup_phase) {
1728 ptlrpc_save_llog_lock(req, lcl);
1729 dput(pending_child);
1731 up(&pending_dir->i_sem);
1736 static int mds_extent_lock_callback(struct ldlm_lock *lock,
1737 struct ldlm_lock_desc *new, void *data,
1740 struct lustre_handle lockh = { 0 };
1745 case LDLM_CB_BLOCKING:
1746 ldlm_lock2handle(lock, &lockh);
1747 rc = ldlm_cli_cancel(&lockh);
1749 CERROR("ldlm_cli_cancel failed: %d\n", rc);
1751 case LDLM_CB_CANCELING: {
1760 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
1761 __u64 lov_merge_blocks(struct lov_stripe_md *lsm);
1763 int mds_validate_size(struct obd_device *obd, struct inode *inode,
1764 struct mds_body *body, struct iattr *iattr)
1766 ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
1767 struct lustre_handle lockh = { 0 };
1768 struct lov_stripe_md *lsm = NULL;
1773 /* we update i_size/i_blocks only for regular files */
1774 if (!S_ISREG(inode->i_mode))
1777 /* we shouldn't fetch size from OSTes during recovery - deadlock */
1778 if (obd->obd_recovering) {
1779 CERROR("size-on-mds has no support on OST yet\n");
1783 /* if nobody modified attrs. we're lucky */
1784 if (!mds_inode_has_old_attrs(inode))
1787 /* 1: client didn't send actual i_size/i_blocks
1788 * 2: we seem to be last writer
1789 * 3: the file doesn't look like to be disappeared
1790 * conclusion: we're gonna fetch them from OSSes */
1792 down(&inode->i_sem);
1793 len = fsfilt_get_md(obd, inode, NULL, 0, EA_LOV);
1797 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, len);
1798 GOTO(cleanup, rc = len);
1799 } else if (len == 0) {
1800 CDEBUG(D_INODE, "no LOV in inode %lu\n", inode->i_ino);
1801 GOTO(cleanup, rc = 0);
1804 OBD_ALLOC(lmm, len);
1806 CERROR("can't allocate memory\n");
1807 GOTO(cleanup, rc = -ENOMEM);
1810 down(&inode->i_sem);
1811 rc = fsfilt_get_md(obd, inode, lmm, len, EA_LOV);
1815 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1819 rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, lmm, len);
1821 CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
1825 CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
1827 flags = LDLM_FL_HAS_INTENT;
1828 rc = obd_enqueue(obd->u.mds.mds_dt_exp, lsm, LDLM_EXTENT, &policy,
1829 LCK_PR, &flags, mds_extent_lock_callback,
1830 ldlm_completion_ast, NULL, NULL,
1831 sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
1832 if (rc == -ENOENT) {
1833 /* while we were enqueueing lock on OST, another thread
1834 * unlinked the file and started OST object destoying.
1835 * it's safe to return 0 here */
1836 GOTO(cleanup, rc = 0);
1837 } else if (rc != 0) {
1838 CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
1842 CDEBUG(D_INODE, "LOV reports "LPD64"/%lu for "DLID4" [%s%s%s]\n",
1843 inode->i_size, inode->i_blocks, OLID4(&body->id1),
1844 atomic_read(&inode->i_writecount) > 1 ? "U" : "",
1845 mds_inode_has_old_attrs(inode) ? "D" : "",
1846 mds_inode_is_orphan(inode) ? "O" : "");
1848 i_size_write(inode, lov_merge_size(lsm, 0));
1849 inode->i_blocks = lov_merge_blocks(lsm);
1850 iattr->ia_size = inode->i_size;
1851 iattr->ia_valid |= ATTR_SIZE;
1852 DOWN_WRITE_I_ALLOC_SEM(inode);
1853 mds_inode_unset_attrs_old(inode);
1854 UP_WRITE_I_ALLOC_SEM(inode);
1856 obd_cancel(obd->u.mds.mds_dt_exp, lsm, LCK_PR, &lockh);
1860 obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
1866 int mds_close(struct ptlrpc_request *req, int offset)
1868 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1869 struct obd_device *obd = req->rq_export->exp_obd;
1870 struct mds_body *body;
1871 struct mds_file_data *mfd;
1872 struct lvfs_run_ctxt saved;
1873 struct inode *inode;
1874 int rc, repsize[3] = {sizeof(struct mds_body),
1875 obd->u.mds.mds_max_mdsize,
1876 obd->u.mds.mds_max_cookiesize};
1878 MD_COUNTER_INCREMENT(obd, close);
1880 rc = lustre_pack_reply(req, 3, repsize, NULL);
1882 CERROR("lustre_pack_reply: rc = %d\n", rc);
1883 req->rq_status = rc;
1885 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1888 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1889 DEBUG_REQ(D_HA, req, "close replay");
1890 memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
1891 lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
1892 req->rq_repmsg->buflens[2]);
1895 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1896 lustre_swab_mds_body);
1898 CERROR("Can't unpack body\n");
1899 req->rq_status = -EFAULT;
1903 if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES)
1904 /* do some stuff */ ;
1906 mfd = mds_handle2mfd(&body->handle);
1908 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
1909 ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie);
1910 req->rq_status = -ESTALE;
1914 inode = mfd->mfd_dentry->d_inode;
1916 /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
1917 DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
1919 if (body->flags & MDS_BFLAG_DIRTY_EPOCH) {
1920 /* the client modified data through the handle
1921 * we need to care about attrs. -bzzz */
1922 mds_inode_set_attrs_old(inode);
1925 if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
1926 struct mds_body *rep_body;
1928 rep_body = lustre_msg_buf(req->rq_repmsg, 0,
1929 sizeof (*rep_body));
1930 LASSERT(rep_body != NULL);
1932 mds_pack_inode2body(obd, rep_body, inode,
1933 (body->valid & OBD_MD_FID) ? 1 : 0);
1935 mds_pack_md(obd, req->rq_repmsg, 1, rep_body,
1936 inode, MDS_PACK_MD_LOCK, 0);
1938 spin_lock(&med->med_open_lock);
1939 list_del(&mfd->mfd_list);
1940 spin_unlock(&med->med_open_lock);
1942 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1943 req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
1944 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1946 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
1947 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
1948 req->rq_status = -ENOMEM;
1958 int mds_done_writing(struct ptlrpc_request *req, int offset)
1960 struct mds_body *body;
1961 int rc, size = sizeof(struct mds_body);
1964 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
1966 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1967 lustre_swab_mds_body);
1969 CERROR("Can't unpack body\n");
1970 req->rq_status = -EFAULT;
1974 rc = lustre_pack_reply(req, 1, &size, NULL);
1976 CERROR("lustre_pack_reply: rc = %d\n", rc);
1977 req->rq_status = rc;