X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fmds%2Fmds_open.c;h=43fa20c5a1bd622074a51b26ea781b0e4d884d2a;hb=c5050e412572b00cbe93d8517d2d1f767bebfa92;hp=c97c53e663c41e5523f4ba1d187711bd4062af4e;hpb=1d04a7876e12a33c65fdc89eb00ec344754a5d80;p=fs%2Flustre-release.git diff --git a/lustre/mds/mds_open.c b/lustre/mds/mds_open.c index c97c53e..43fa20c 100644 --- a/lustre/mds/mds_open.c +++ b/lustre/mds/mds_open.c @@ -23,54 +23,177 @@ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. */ -#define EXPORT_SYMTAB +#ifndef EXPORT_SYMTAB +# define EXPORT_SYMTAB +#endif #define DEBUG_SUBSYSTEM S_MDS #include -#include -#include #include -#include -#include #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)) # include # include #else # include #endif -#include + +#include #include +#include #include #include #include "mds_internal.h" -struct mds_file_data *mds_dentry_open(struct dentry *dentry, - struct vfsmount *mnt, - int flags, - struct ptlrpc_request *req) +/* Exported function from this file are: + * + * mds_open - called by the intent handler + * mds_close - an rpc handling function + * mds_pin - an rpc handling function - which will go away + * mds_mfd_close - for force closing files when a client dies + */ + +/* + * MDS file data handling: file data holds a handle for a file opened + * by a client. + */ + +static void mds_mfd_addref(void *mfdp) +{ + struct mds_file_data *mfd = mfdp; + + atomic_inc(&mfd->mfd_refcount); + CDEBUG(D_INFO, "GETting mfd %p : new refcount %d\n", mfd, + atomic_read(&mfd->mfd_refcount)); +} + +struct mds_file_data *mds_mfd_new(void) { - struct mds_export_data *med = &req->rq_export->exp_mds_data; - struct inode *inode; struct mds_file_data *mfd; - int mode, error; + OBD_ALLOC(mfd, sizeof *mfd); + if (mfd == NULL) { + CERROR("mds: out of memory\n"); + return NULL; + } + + atomic_set(&mfd->mfd_refcount, 2); + + INIT_LIST_HEAD(&mfd->mfd_handle.h_link); + class_handle_hash(&mfd->mfd_handle, mds_mfd_addref); + + return mfd; +} + +static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle) +{ + ENTRY; + LASSERT(handle != NULL); + RETURN(class_handle2object(handle->cookie)); +} + +static void mds_mfd_put(struct mds_file_data *mfd) +{ + CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd, + atomic_read(&mfd->mfd_refcount) - 1); + LASSERT(atomic_read(&mfd->mfd_refcount) > 0 && + atomic_read(&mfd->mfd_refcount) < 0x5a5a); + if (atomic_dec_and_test(&mfd->mfd_refcount)) { + LASSERT(list_empty(&mfd->mfd_handle.h_link)); + OBD_FREE(mfd, sizeof *mfd); + } +} + +static void mds_mfd_destroy(struct mds_file_data *mfd) +{ + class_handle_unhash(&mfd->mfd_handle); + mds_mfd_put(mfd); +} + + +/* + * Write access to a file: executors cause a negative count, + * writers a positive count. The spin lock is needed to perform + * a check for the sign and then increment or decrement atomically. + */ + +static spinlock_t mds_exec_lock = SPIN_LOCK_UNLOCKED; +static int mds_get_write_access(struct inode * inode) +{ + ENTRY; + spin_lock(&mds_exec_lock); + if (atomic_read(&inode->i_writecount) < 0) { + spin_unlock(&mds_exec_lock); + RETURN(-ETXTBSY); + } + atomic_inc(&inode->i_writecount); + spin_unlock(&mds_exec_lock); + RETURN(0); +} + +static int mds_deny_write_access(struct inode *inode) +{ + ENTRY; + spin_lock(&mds_exec_lock); + if (atomic_read(&inode->i_writecount) > 0) { + spin_unlock(&mds_exec_lock); + RETURN(-ETXTBSY); + } + atomic_dec(&inode->i_writecount); + spin_unlock(&mds_exec_lock); + RETURN(0); +} + +static void mds_put_write_access(struct inode * inode) +{ + ENTRY; + atomic_dec(&inode->i_writecount); +} + +static void mds_allow_write_access(struct inode *inode) +{ + ENTRY; + atomic_inc(&inode->i_writecount); +} + +int mds_query_write_access(struct inode *inode) +{ + ENTRY; + RETURN(atomic_read(&inode->i_writecount)); +} + +/* This replaces the VFS mds_dentry_open, it manages mfd and writecount */ +static struct mds_file_data *mds_dentry_open(struct dentry *dentry, + struct vfsmount *mnt, int flags, + struct ptlrpc_request *req) +{ + struct mds_export_data *med = &req->rq_export->exp_mds_data; + struct mds_file_data *mfd; + int error; + ENTRY; + mfd = mds_mfd_new(); if (mfd == NULL) { CERROR("mds: out of memory\n"); GOTO(cleanup_dentry, error = -ENOMEM); } - mode = (flags + 1) & O_ACCMODE; - inode = dentry->d_inode; - - if (mode & FMODE_WRITE) { - error = get_write_access(inode); + if (flags & FMODE_WRITE) { + error = mds_get_write_access(dentry->d_inode); + if (error) + GOTO(cleanup_mfd, error); + } else if (flags & FMODE_EXEC) { + error = mds_deny_write_access(dentry->d_inode); if (error) - goto cleanup_mfd; + GOTO(cleanup_mfd, error); } - mfd->mfd_mode = mode; + dget(dentry); + + /* Mark the file as open to handle open-unlink. */ + mds_open_orphan_inc(dentry->d_inode); + + mfd->mfd_mode = flags; mfd->mfd_dentry = dentry; mfd->mfd_xid = req->rq_xid; @@ -78,20 +201,243 @@ struct mds_file_data *mds_dentry_open(struct dentry *dentry, list_add(&mfd->mfd_list, &med->med_open_head); spin_unlock(&med->med_open_lock); mds_mfd_put(mfd); - return mfd; + RETURN(mfd); cleanup_mfd: mds_mfd_put(mfd); mds_mfd_destroy(mfd); cleanup_dentry: - dput(dentry); - mntput(mnt); return ERR_PTR(error); } -void reconstruct_open(struct mds_update_record *rec, int offset, - struct ptlrpc_request *req, - struct lustre_handle *child_lockh) +static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm, + struct lov_desc *desc) +{ + int i; + for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) { + ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] = + le64_to_cpu(lmm->lmm_objects[i].l_object_id); + } +} + +/* Must be called with i_sem held */ +static int mds_create_objects(struct ptlrpc_request *req, int offset, + struct mds_update_record *rec, + struct mds_obd *mds, struct obd_device *obd, + struct inode *inode, void **handle, obd_id **ids) +{ + struct obdo *oa; + struct obd_trans_info oti = { 0 }; + struct mds_body *body; + struct lov_stripe_md *lsm = NULL; + struct lov_mds_md *lmm = NULL; + void *lmm_buf; + int rc, lmm_bufsize, lmm_size; + ENTRY; + + if (rec->ur_flags & MDS_OPEN_DELAY_CREATE || + !(rec->ur_flags & FMODE_WRITE)) + RETURN(0); + + body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body)); + + if (!S_ISREG(inode->i_mode)) + RETURN(0); + if (body->valid & OBD_MD_FLEASIZE) + RETURN(0); + + OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids)); + if (*ids == NULL) + RETURN(-ENOMEM); + oti.oti_objid = *ids; + + if (*handle == NULL) + *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL); + if (IS_ERR(*handle)) { + rc = PTR_ERR(*handle); + *handle = NULL; + GOTO(out_ids, rc); + } + + /* replay case */ + if (rec->ur_fid2->id) { + body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE; + lmm_size = rec->ur_eadatalen; + lmm = rec->ur_eadata; + LASSERT(lmm); + + mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc); + + lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0); + lmm_bufsize = req->rq_repmsg->buflens[offset]; + LASSERT(lmm_buf); + LASSERT(lmm_bufsize >= lmm_size); + memcpy(lmm_buf, lmm, lmm_size); + rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size); + if (rc) + CERROR("open replay failed to set md:%d\n", rc); + RETURN(0); + } + + oa = obdo_alloc(); + if (oa == NULL) + GOTO(out_ids, rc = -ENOMEM); + oa->o_mode = S_IFREG | 0600; + oa->o_id = inode->i_ino; + oa->o_generation = inode->i_generation; + oa->o_uid = 0; /* must have 0 uid / gid on OST */ + oa->o_gid = 0; + oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE | + OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID; + oa->o_size = 0; + + obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME| + OBD_MD_FLCTIME); + + /* check if things like lstripe/lfs stripe are sending us the ea */ + if (rec->ur_flags & MDS_OPEN_HAS_EA) { + rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, mds->mds_osc_exp, + 0, &lsm, rec->ur_eadata); + if (rc) + GOTO(out_oa, rc); + } + + rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti); + if (rc) { + int level = D_ERROR; + if (rc == -ENOSPC) + level = D_INODE; + CDEBUG(level, "error creating objects for inode %lu: rc = %d\n", + inode->i_ino, rc); + if (rc > 0) { + CERROR("obd_create returned invalid rc %d\n", rc); + rc = -EIO; + } + GOTO(out_oa, rc); + } + + if (inode->i_size) { + oa->o_size = inode->i_size; + obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME| + OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE); + rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti); + if (rc) { + CERROR("error setting attrs for inode %lu: rc %d\n", + inode->i_ino, rc); + if (rc > 0) { + CERROR("obd_setattr returned bad rc %d\n", rc); + rc = -EIO; + } + GOTO(out_oa, rc); + } + } + + body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE; + obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ); + + LASSERT(lsm && lsm->lsm_object_id); + lmm = NULL; + rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm); + if (!rec->ur_fid2->id) + obd_free_memmd(mds->mds_osc_exp, &lsm); + LASSERT(rc >= 0); + lmm_size = rc; + body->eadatasize = rc; + rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size); + lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0); + lmm_bufsize = req->rq_repmsg->buflens[offset]; + LASSERT(lmm_buf); + LASSERT(lmm_bufsize >= lmm_size); + + memcpy(lmm_buf, lmm, lmm_size); + obd_free_diskmd(mds->mds_osc_exp, &lmm); + out_oa: + oti_free_cookies(&oti); + obdo_free(oa); + out_ids: + if (rc) { + OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids)); + *ids = NULL; + } + RETURN(rc); +} + +/* Caller must hold mds->mds_epoch_sem */ +static int mds_alloc_filterdata(struct inode *inode) +{ + LASSERT(inode->i_filterdata == NULL); + OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data)); + if (inode->i_filterdata == NULL) + return -ENOMEM; + LASSERT(igrab(inode) == inode); + return 0; +} + +/* Caller must hold mds->mds_epoch_sem */ +static void mds_free_filterdata(struct inode *inode) +{ + LASSERT(inode->i_filterdata != NULL); + OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data)); + inode->i_filterdata = NULL; + iput(inode); +} + +/* epoch argument is nonzero during recovery */ +static int mds_open_io_epoch(struct mds_obd *mds, struct inode *inode, + __u64 epoch) +{ + int rc = 0; + + down(&mds->mds_epoch_sem); + if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) { + CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n", + MDS_FILTERDATA(inode)->io_epoch, inode->i_ino, + inode->i_generation); + goto out; + } + + if (inode->i_filterdata == NULL) + mds_alloc_filterdata(inode); + if (inode->i_filterdata == NULL) { + rc = -ENOMEM; + goto out; + } + if (epoch > mds->mds_io_epoch) + mds->mds_io_epoch = epoch; + else + mds->mds_io_epoch++; + MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch; + CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n", + mds->mds_io_epoch, inode->i_ino, inode->i_generation); + out: + up(&mds->mds_epoch_sem); + return rc; +} + +/* Returns EAGAIN if the client needs to get size and/or cookies and close + * again -- which is never true if the file is about to be unlinked. */ +static int mds_close_io_epoch(struct mds_obd *mds, struct inode *inode, + struct mds_body *body, int unlinking) +{ + int rc = 0; + ENTRY; + + down(&mds->mds_epoch_sem); + if (mds_query_write_access(inode) > 0) + GOTO(out, rc); +#if 0 + if (!unlinking && !(body->valid & OBD_MD_FLSIZE)) + GOTO(out, rc = EAGAIN); +#endif + mds_free_filterdata(inode); + out: + up(&mds->mds_epoch_sem); + return rc; +} + +static void reconstruct_open(struct mds_update_record *rec, int offset, + struct ptlrpc_request *req, + struct lustre_handle *child_lockh) { struct ptlrpc_request *oldreq = req->rq_export->exp_outstanding_reply; struct mds_export_data *med = &req->rq_export->exp_mds_data; @@ -112,13 +458,9 @@ void reconstruct_open(struct mds_update_record *rec, int offset, body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body)); /* copy rc, transno and disp; steal locks */ - req->rq_transno = mcd->mcd_last_transno; - req->rq_status = mcd->mcd_last_result; + mds_req_from_mcd(req, mcd); intent_set_disposition(rep, mcd->mcd_last_data); - if (oldreq) - mds_steal_ack_locks(req->rq_export, req); - /* Only replay if create or open actually happened. */ if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) { EXIT; @@ -150,7 +492,7 @@ void reconstruct_open(struct mds_update_record *rec, int offset, mds_pack_inode2body(body, child->d_inode); if (S_ISREG(child->d_inode->i_mode)) { rc = mds_pack_md(obd, req->rq_repmsg, 2, body, - child->d_inode); + child->d_inode, 1); if (rc) LASSERT(rc == req->rq_status); @@ -170,7 +512,8 @@ void reconstruct_open(struct mds_update_record *rec, int offset, * exclusively, we can tell we failed because the file already existed. */ if (req->rq_status == -EEXIST && - ((rec->ur_flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))) { + ((rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_EXCL)) == + (MDS_OPEN_CREAT | MDS_OPEN_EXCL))) { GOTO(out_dput, 0); } @@ -194,14 +537,14 @@ void reconstruct_open(struct mds_update_record *rec, int offset, mfd = NULL; } - if (oldreq) { + if (oldreq != NULL) { /* if we're not recovering, it had better be found */ - LASSERT(mfd); + LASSERT(mfd != NULL); } else if (mfd == NULL) { mntget(mds->mds_vfsmnt); CERROR("Re-opened file \n"); mfd = mds_dentry_open(child, mds->mds_vfsmnt, - rec->ur_flags & ~(O_DIRECT | O_TRUNC), req); + rec->ur_flags & ~MDS_OPEN_TRUNC, req); if (!mfd) { CERROR("mds: out of memory\n"); GOTO(out_dput, req->rq_status = -ENOMEM); @@ -209,6 +552,9 @@ void reconstruct_open(struct mds_update_record *rec, int offset, put_child = 0; } + if ((rec->ur_flags & FMODE_WRITE) && + mds_open_io_epoch(mds, child->d_inode, 0) == 0) + body->io_epoch = MDS_FILTERDATA(child->d_inode)->io_epoch; body->handle.cookie = mfd->mfd_handle.h_cookie; out_dput: @@ -218,115 +564,177 @@ void reconstruct_open(struct mds_update_record *rec, int offset, EXIT; } -int mds_pin(struct ptlrpc_request *req) +/* do NOT or the MAY_*'s, you'll get the weakest */ +static int accmode(int flags) +{ + int res = 0; + + if (flags & FMODE_READ) + res = MAY_READ; + if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC)) + res |= MAY_WRITE; + if (flags & FMODE_EXEC) + res = MAY_EXEC; + return res; +} + +/* Handles object creation, actual opening, and I/O epoch */ +static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild, + struct mds_body *body, int flags, void **handle, + struct mds_update_record *rec,struct ldlm_reply *rep) { struct mds_obd *mds = mds_req2mds(req); - struct inode *pending_dir = mds->mds_pending_dir->d_inode; + struct obd_device *obd = req->rq_export->exp_obd; struct mds_file_data *mfd = NULL; - struct mds_body *body; + obd_id *ids = NULL; /* object IDs created */ + int rc; + ENTRY; + + /* atomically create objects if necessary */ + down(&dchild->d_inode->i_sem); + if (S_ISREG(dchild->d_inode->i_mode) && + !(body->valid & OBD_MD_FLEASIZE)) { + rc = mds_pack_md(obd, req->rq_repmsg, 2, body, + dchild->d_inode, 0); + if (rc) { + up(&dchild->d_inode->i_sem); + RETURN(rc); + } + } + if (rec != NULL) { + /* no EA: create objects */ + rc = mds_create_objects(req, 2, rec, mds, obd, + dchild->d_inode, handle, &ids); + if (rc) { + CERROR("mds_create_objects: rc = %d\n", rc); + up(&dchild->d_inode->i_sem); + RETURN(rc); + } + } + /* If the inode has EA data, then OSTs hold size, mtime */ + if (S_ISREG(dchild->d_inode->i_mode) && + !(body->valid & OBD_MD_FLEASIZE)) { + body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | + OBD_MD_FLATIME | OBD_MD_FLMTIME); + } + up(&dchild->d_inode->i_sem); + + intent_set_disposition(rep, DISP_OPEN_OPEN); + mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req); + if (IS_ERR(mfd)) + RETURN(PTR_ERR(mfd)); + + if ((rec->ur_flags & FMODE_WRITE) && + mds_open_io_epoch(mds, dchild->d_inode, 0) == 0) + body->io_epoch = MDS_FILTERDATA(dchild->d_inode)->io_epoch; + body->handle.cookie = mfd->mfd_handle.h_cookie; + + CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd, + mfd->mfd_handle.h_cookie); + if (ids != NULL) { + mds_lov_update_objids(obd, ids); + OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count); + } + if (rc) + mds_mfd_destroy(mfd); + RETURN(rc); +} + +static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid, + struct mds_body *body, int flags, + struct mds_update_record *rec,struct ldlm_reply *rep) +{ + struct mds_obd *mds = mds_req2mds(req); + struct inode *pending_dir = mds->mds_pending_dir->d_inode; struct dentry *dchild; - struct obd_run_ctxt saved; char fidname[LL_FID_NAMELEN]; - int fidlen = 0, rc, cleanup_phase = 0, size = sizeof(*body); + int fidlen = 0, rc; + void *handle = NULL; ENTRY; - body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body)); - down(&pending_dir->i_sem); - fidlen = ll_fid2str(fidname, body->fid1.id, body->fid1.generation); + fidlen = ll_fid2str(fidname, fid->id, fid->generation); dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen); if (IS_ERR(dchild)) { up(&pending_dir->i_sem); rc = PTR_ERR(dchild); - CERROR("error looking up %s in PENDING: rc = %d\n", - fidname, rc); + CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc); RETURN(rc); } - cleanup_phase = 2; - - if (dchild->d_inode) { + if (dchild->d_inode != NULL) { up(&pending_dir->i_sem); mds_inode_set_orphan(dchild->d_inode); mds_pack_inode2fid(&body->fid1, dchild->d_inode); mds_pack_inode2body(body, dchild->d_inode); - GOTO(openit, rc = 0); + intent_set_disposition(rep, DISP_LOOKUP_EXECD); + intent_set_disposition(rep, DISP_LOOKUP_POS); + CWARN("Orphan %s found and opened in PENDING directory\n", + fidname); + goto open; } dput(dchild); up(&pending_dir->i_sem); /* We didn't find it in PENDING so it isn't an orphan. See - * if it's a regular inode. */ - dchild = mds_fid2dentry(mds, &body->fid1, NULL); - if (!IS_ERR(dchild)) { - mds_pack_inode2fid(&body->fid1, dchild->d_inode); - mds_pack_inode2body(body, dchild->d_inode); - GOTO(openit, rc = 0); - } + * if it was a regular inode that was previously created. */ + dchild = mds_fid2dentry(mds, fid, NULL); + if (IS_ERR(dchild)) + RETURN(PTR_ERR(dchild)); - /* We didn't find this inode on disk, but we're trying to pin it. - * This should never happen. */ - CERROR("ENOENT during mds_pin for fid "LPU64"/%u\n", body->fid1.id, - body->fid1.generation); - RETURN(-ENOENT); + mds_pack_inode2fid(&body->fid1, dchild->d_inode); + mds_pack_inode2body(body, dchild->d_inode); + intent_set_disposition(rep, DISP_LOOKUP_EXECD); + intent_set_disposition(rep, DISP_LOOKUP_POS); - openit: - /* dentry_open does a dput(de) and mntput(mds->mds_vfsmnt) on error */ - mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, body->flags, req); - if (IS_ERR(mfd)) { - dchild = NULL; /* prevent a double dput in cleanup phase 2 */ - GOTO(cleanup, rc = PTR_ERR(mfd)); - } + open: + rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep); + rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle, + req, rc, rep ? rep->lock_policy_res1 : 0); + /* XXX what do we do here if mds_finish_transno itself failed? */ - rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg); - if (rc) { - CERROR("out of memoryK\n"); - GOTO(cleanup, rc); - } - body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body)); + l_dput(dchild); + RETURN(rc); +} - cleanup_phase = 4; /* mfd allocated */ - body->handle.cookie = mfd->mfd_handle.h_cookie; - CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd, - mfd->mfd_handle.h_cookie); - GOTO(cleanup, rc = 0); +int mds_pin(struct ptlrpc_request *req) +{ + struct obd_device *obd = req->rq_export->exp_obd; + struct mds_body *request_body, *reply_body; + struct obd_run_ctxt saved; + int rc, size = sizeof(*reply_body); + ENTRY; - cleanup: - push_ctxt(&saved, &mds->mds_ctxt, NULL); - rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, NULL, - req, rc, 0); - pop_ctxt(&saved, &mds->mds_ctxt, NULL); - /* XXX what do we do here if mds_finish_transno itself failed? */ - switch (cleanup_phase) { - case 4: - if (rc) - mds_mfd_destroy(mfd); - case 2: - if (rc || S_ISLNK(dchild->d_inode->i_mode)) - l_dput(dchild); - } - return rc; + request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body)); + + rc = lustre_pack_reply(req, 1, &size, NULL); + if (rc) + RETURN(rc); + reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body)); + + push_ctxt(&saved, &obd->obd_ctxt, NULL); + rc = mds_open_by_fid(req, &request_body->fid1, reply_body, + request_body->flags, NULL, NULL); + pop_ctxt(&saved, &obd->obd_ctxt, NULL); + + RETURN(rc); } int mds_open(struct mds_update_record *rec, int offset, struct ptlrpc_request *req, struct lustre_handle *child_lockh) { /* XXX ALLOCATE _something_ - 464 bytes on stack here */ - static const char acc_table [] = {[O_RDONLY] MAY_READ, - [O_WRONLY] MAY_WRITE, - [O_RDWR] MAY_READ | MAY_WRITE}; - struct mds_obd *mds = mds_req2mds(req); struct obd_device *obd = req->rq_export->exp_obd; + struct mds_obd *mds = mds_req2mds(req); struct ldlm_reply *rep = NULL; struct mds_body *body = NULL; struct dentry *dchild = NULL, *parent = NULL; struct mds_export_data *med; - struct mds_file_data *mfd = NULL; struct ldlm_res_id child_res_id = { .name = {0} }; struct lustre_handle parent_lockh; - int rc = 0, parent_mode, child_mode = LCK_PR, lock_flags, created = 0; - int cleanup_phase = 0, acc_mode; + int rc = 0, parent_mode = 0, cleanup_phase = 0, acc_mode, created = 0; void *handle = NULL; + struct dentry_params dp; ENTRY; if (offset == 2) { /* intent */ @@ -343,77 +751,37 @@ int mds_open(struct mds_update_record *rec, int offset, /* Step 0: If we are passed a fid, then we assume the client already * opened this file and is only replaying the RPC, so we open the - * inode by fid (at some large expense in security). - */ + * inode by fid (at some large expense in security). */ if (rec->ur_fid2->id) { - struct inode *pending_dir = mds->mds_pending_dir->d_inode; - char fidname[LL_FID_NAMELEN]; - int fidlen = 0; - - down(&pending_dir->i_sem); - fidlen = ll_fid2str(fidname, rec->ur_fid2->id, - rec->ur_fid2->generation); - dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen); - if (IS_ERR(dchild)) { - up(&pending_dir->i_sem); - rc = PTR_ERR(dchild); - CERROR("error looking up %s in PENDING: rc = %d\n", - fidname, rc); + rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags, + rec, rep); + if (rc != -ENOENT) RETURN(rc); - } - - if (dchild->d_inode) { - up(&pending_dir->i_sem); - mds_inode_set_orphan(dchild->d_inode); - mds_pack_inode2fid(&body->fid1, dchild->d_inode); - mds_pack_inode2body(body, dchild->d_inode); - cleanup_phase = 2; - GOTO(openit, rc = 0); - } - dput(dchild); - up(&pending_dir->i_sem); - - /* We didn't find it in PENDING so it isn't an orphan. See - * if it was a regular inode that was previously created. - */ - dchild = mds_fid2dentry(mds, rec->ur_fid2, NULL); - if (!IS_ERR(dchild)) { - mds_pack_inode2fid(&body->fid1, dchild->d_inode); - mds_pack_inode2body(body, dchild->d_inode); - cleanup_phase = 2; - GOTO(openit, rc = 0); - } - /* We didn't find the correct inode on disk either, so we - * need to re-create it via a regular replay. Do that below. - */ - LASSERT(rec->ur_flags & O_CREAT); + * need to re-create it via a regular replay. */ + LASSERT(rec->ur_flags & MDS_OPEN_CREAT); } LASSERT(offset == 2); /* If we got here, we must be called via intent */ med = &req->rq_export->exp_mds_data; if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) { CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n"); - req->rq_status = -ENOMEM; RETURN(-ENOMEM); } - if ((rec->ur_flags & O_ACCMODE) >= sizeof (acc_table)) - RETURN(-EINVAL); - acc_mode = acc_table[rec->ur_flags & O_ACCMODE]; - if ((rec->ur_flags & O_TRUNC) != 0) - acc_mode |= MAY_WRITE; + acc_mode = accmode(rec->ur_flags); /* Step 1: Find and lock the parent */ - parent_mode = (rec->ur_flags & O_CREAT) ? LCK_PW : LCK_PR; + parent_mode = (rec->ur_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR; parent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode, - &parent_lockh); + &parent_lockh, rec->ur_name, + rec->ur_namelen - 1); if (IS_ERR(parent)) { rc = PTR_ERR(parent); CERROR("parent lookup error %d\n", rc); GOTO(cleanup, rc); } - LASSERT(parent->d_inode); + LASSERT(parent->d_inode != NULL); cleanup_phase = 1; /* parent dentry and lock */ @@ -430,12 +798,13 @@ int mds_open(struct mds_update_record *rec, int offset, else intent_set_disposition(rep, DISP_LOOKUP_NEG); - /* Step 3: If the child was negative, and we're supposed to, - * create it. */ - if (!dchild->d_inode) { + /*Step 3: If the child was negative, and we're supposed to, create it.*/ + if (dchild->d_inode == NULL) { unsigned long ino = rec->ur_fid2->id; + struct iattr iattr; + struct inode *inode; - if (!(rec->ur_flags & O_CREAT)) { + if (!(rec->ur_flags & MDS_OPEN_CREAT)) { /* It's negative and we weren't supposed to create it */ GOTO(cleanup, rc = -ENOENT); } @@ -448,83 +817,62 @@ int mds_open(struct mds_update_record *rec, int offset, handle = NULL; GOTO(cleanup, rc); } - if (ino) - dchild->d_fsdata = (void *)(unsigned long)ino; + dchild->d_fsdata = (void *) &dp; + dp.p_ptr = req; + dp.p_inum = ino; - rc = vfs_create(parent->d_inode, dchild, rec->ur_mode); + rc = ll_vfs_create(parent->d_inode, dchild, rec->ur_mode, NULL); if (dchild->d_fsdata == (void *)(unsigned long)ino) dchild->d_fsdata = NULL; if (rc) { CDEBUG(D_INODE, "error during create: %d\n", rc); GOTO(cleanup, rc); - } else { - struct iattr iattr; - struct inode *inode = dchild->d_inode; - - if (ino) { - LASSERT(ino == inode->i_ino); - /* Written as part of setattr */ - inode->i_generation = rec->ur_fid2->generation; - CDEBUG(D_HA, "recreated ino %lu with gen %x\n", - inode->i_ino, inode->i_generation); - } + } + inode = dchild->d_inode; + if (ino) { + LASSERT(ino == inode->i_ino); + /* Written as part of setattr */ + inode->i_generation = rec->ur_fid2->generation; + CDEBUG(D_HA, "recreated ino %lu with gen %u\n", + inode->i_ino, inode->i_generation); + } - created = 1; - LTIME_S(iattr.ia_atime) = rec->ur_time; - LTIME_S(iattr.ia_ctime) = rec->ur_time; - LTIME_S(iattr.ia_mtime) = rec->ur_time; + created = 1; + LTIME_S(iattr.ia_atime) = rec->ur_time; + LTIME_S(iattr.ia_ctime) = rec->ur_time; + LTIME_S(iattr.ia_mtime) = rec->ur_time; - iattr.ia_uid = rec->ur_uid; - if (parent->d_inode->i_mode & S_ISGID) { - iattr.ia_gid = parent->d_inode->i_gid; - } else - iattr.ia_gid = rec->ur_gid; + iattr.ia_uid = rec->ur_fsuid; + if (parent->d_inode->i_mode & S_ISGID) + iattr.ia_gid = parent->d_inode->i_gid; + else + iattr.ia_gid = rec->ur_fsgid; - iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME | - ATTR_MTIME | ATTR_CTIME; + iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME | + ATTR_MTIME | ATTR_CTIME; - rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0); - if (rc) { - CERROR("error on setattr: rc = %d\n", rc); - /* XXX should we abort here in case of error? */ - } - } + rc = fsfilt_setattr(obd, dchild, handle, &iattr, 0); + if (rc) + CERROR("error on child setattr: rc = %d\n", rc); - child_mode = LCK_PW; - acc_mode = 0; /* Don't check for permissions */ - } + iattr.ia_valid = ATTR_MTIME | ATTR_CTIME; - LASSERT(!mds_inode_is_orphan(dchild->d_inode)); + rc = fsfilt_setattr(obd, parent, handle, &iattr, 0); + if (rc) + CERROR("error on parent setattr: rc = %d\n", rc); - /* Step 4: It's positive, so lock the child */ - child_res_id.name[0] = dchild->d_inode->i_ino; - child_res_id.name[1] = dchild->d_inode->i_generation; - reacquire: - lock_flags = 0; - /* For the open(O_CREAT) case, this would technically be a lock - * inversion (getting a VFS lock after starting a transaction), - * but in that case we cannot possibly block on this lock because - * we just created the child and also hold a write lock on the - * parent, so nobody could be holding the lock yet. - */ - rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL, - child_res_id, LDLM_PLAIN, NULL, 0, child_mode, - &lock_flags, ldlm_completion_ast, - mds_blocking_ast, NULL, child_lockh); - if (rc != ELDLM_OK) { - CERROR("ldlm_cli_enqueue: %d\n", rc); - GOTO(cleanup, rc = -EIO); + acc_mode = 0; /* Don't check for permissions */ } - cleanup_phase = 3; /* child lock */ + LASSERT(!mds_inode_is_orphan(dchild->d_inode)); mds_pack_inode2fid(&body->fid1, dchild->d_inode); mds_pack_inode2body(body, dchild->d_inode); if (S_ISREG(dchild->d_inode->i_mode)) { /* Check permissions etc */ - rc = permission(dchild->d_inode, acc_mode); + rc = ll_permission(dchild->d_inode, acc_mode, NULL); if (rc != 0) GOTO(cleanup, rc); @@ -535,90 +883,300 @@ int mds_open(struct mds_update_record *rec, int offset, /* An append-only file must be opened in append mode for * writing */ if (IS_APPEND(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0 && - ((rec->ur_flags & O_APPEND) == 0 || - (rec->ur_flags & O_TRUNC) != 0)) + ((rec->ur_flags & MDS_OPEN_APPEND) == 0 || + (rec->ur_flags & MDS_OPEN_TRUNC) != 0)) GOTO(cleanup, rc = -EPERM); - - rc = mds_pack_md(obd, req->rq_repmsg, 2, body, dchild->d_inode); - if (rc) - GOTO(cleanup, rc); - - /* If we have LOV EA data, the OST holds size, mtime */ - if (!(body->valid & OBD_MD_FLEASIZE)) - body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | - OBD_MD_FLATIME | OBD_MD_FLMTIME); } - if (!created && (rec->ur_flags & O_CREAT) && - (rec->ur_flags & O_EXCL)) { + if (!created && (rec->ur_flags & MDS_OPEN_CREAT) && + (rec->ur_flags & MDS_OPEN_EXCL)) { /* File already exists, we didn't just create it, and we * were passed O_EXCL; err-or. */ GOTO(cleanup, rc = -EEXIST); // returns a lock to the client } - /* If we're opening a file without an EA for write, the client needs - * a write lock. */ - if (S_ISREG(dchild->d_inode->i_mode) && (rec->ur_flags & O_ACCMODE) && - child_mode != LCK_PW && !(body->valid & OBD_MD_FLEASIZE)) { - ldlm_lock_decref(child_lockh, child_mode); - child_mode = LCK_PW; - goto reacquire; - } - /* if we are following a symlink, don't open */ if (S_ISLNK(dchild->d_inode->i_mode)) GOTO(cleanup, rc = 0); - if ((rec->ur_flags & O_DIRECTORY) && !S_ISDIR(dchild->d_inode->i_mode)) + if ((rec->ur_flags & MDS_OPEN_DIRECTORY) && + !S_ISDIR(dchild->d_inode->i_mode)) GOTO(cleanup, rc = -ENOTDIR); /* Step 5: mds_open it */ - intent_set_disposition(rep, DISP_OPEN_OPEN); - openit: - /* dentry_open does a dput(de) and mntput(mds->mds_vfsmnt) on error */ - mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, - rec->ur_flags & ~(O_DIRECT | O_TRUNC), req); - if (IS_ERR(mfd)) { - dchild = NULL; /* prevent a double dput in cleanup phase 2 */ - GOTO(cleanup, rc = PTR_ERR(mfd)); - } - - cleanup_phase = 4; /* mfd allocated */ - body->handle.cookie = mfd->mfd_handle.h_cookie; - CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd, - mfd->mfd_handle.h_cookie); - GOTO(cleanup, rc = 0); /* returns a lock to the client */ + rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec, + rep); + GOTO(cleanup, rc); cleanup: rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle, - req, rc, rep->lock_policy_res1); + req, rc, rep ? rep->lock_policy_res1 : 0); /* XXX what do we do here if mds_finish_transno itself failed? */ + if (created) { + /* Step 4: Lock new child to sync inode reuse (bug 2029) */ + int lock_flags = 0, err; + child_res_id.name[0] = dchild->d_inode->i_ino; + child_res_id.name[1] = 0; + err = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL, + child_res_id, LDLM_PLAIN, NULL, 0, + LCK_EX, &lock_flags, ldlm_completion_ast, + mds_blocking_ast, NULL, child_lockh); + if (err == ELDLM_OK) + ldlm_lock_decref(child_lockh, LCK_EX); + else + CERROR("ldlm_cli_enqueue: %d\n", err); + } + switch (cleanup_phase) { - case 4: - if (rc && !S_ISLNK(dchild->d_inode->i_mode)) - mds_mfd_destroy(mfd); - case 3: - /* This is the same logic as in the IT_OPEN part of - * ldlm_intent_policy: if we found the dentry, or we tried to - * open it (meaning that we created, if it wasn't found), then - * we return the lock to the caller and client. */ - if (intent_disposition(rep, DISP_LOOKUP_NEG) && - !intent_disposition(rep, DISP_OPEN_OPEN)) - ldlm_lock_decref(child_lockh, child_mode); case 2: - if (rc || S_ISLNK(dchild->d_inode->i_mode)) - l_dput(dchild); - case 1: - if (parent) { - l_dput(parent); - if (rc) { - ldlm_lock_decref(&parent_lockh, parent_mode); - } else { - memcpy(&req->rq_ack_locks[0].lock,&parent_lockh, - sizeof(parent_lockh)); - req->rq_ack_locks[0].mode = parent_mode; + if (rc && created) { + int err = vfs_unlink(parent->d_inode, dchild); + if (err) { + CERROR("unlink(%*s) in error path: %d\n", + dchild->d_name.len, dchild->d_name.name, + err); } } + l_dput(dchild); + case 1: + if (parent == NULL) + break; + + l_dput(parent); + if (rc) + ldlm_lock_decref(&parent_lockh, parent_mode); + else + ldlm_put_lock_into_req(req, &parent_lockh, parent_mode); } RETURN(rc); } + +/* Close a "file descriptor" and possibly unlink an orphan from the + * PENDING directory. + * + * If we are being called from mds_disconnect() because the client has + * disappeared, then req == NULL and we do not update last_rcvd because + * there is nothing that could be recovered by the client at this stage + * (it will not even _have_ an entry in last_rcvd anymore). + * + * Returns EAGAIN if the client needs to get more data and re-close. */ +int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd, + struct mds_file_data *mfd, int unlink_orphan) +{ + struct inode *inode = mfd->mfd_dentry->d_inode; + char fidname[LL_FID_NAMELEN]; + int last_orphan, fidlen, rc = 0, cleanup_phase = 0; + struct dentry *pending_child = NULL; + struct mds_obd *mds = &obd->u.mds; + struct inode *pending_dir = mds->mds_pending_dir->d_inode; + void *handle = NULL; + struct mds_body *request_body, *reply_body; + struct dentry_params dp; + ENTRY; + + if (req != NULL) { + request_body = lustre_msg_buf(req->rq_reqmsg, 0, + sizeof(*request_body)); + reply_body = lustre_msg_buf(req->rq_repmsg, 0, + sizeof(*reply_body)); + } + + fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation); + + last_orphan = mds_open_orphan_dec_test(inode) && + mds_inode_is_orphan(inode); + + /* this is the actual "close" */ + if (mfd->mfd_mode & FMODE_WRITE) { + mds_put_write_access(inode); + } else if (mfd->mfd_mode & FMODE_EXEC) { + mds_allow_write_access(inode); + } + + if (last_orphan && unlink_orphan) { + LASSERT(mds_query_write_access(inode) == 0); + if (mfd->mfd_mode & FMODE_WRITE) + rc = mds_close_io_epoch(mds, inode, request_body, 1); + + CWARN("destroying orphan object %s\n", fidname); + + /* Sadly, there is no easy way to save pending_child from + * mds_reint_unlink() into mfd, so we need to re-lookup, + * but normally it will still be in the dcache. */ + pending_dir = mds->mds_pending_dir->d_inode; + down(&pending_dir->i_sem); + cleanup_phase = 1; /* up(i_sem) when finished */ + pending_child = lookup_one_len(fidname, mds->mds_pending_dir, + fidlen); + if (IS_ERR(pending_child)) + GOTO(cleanup, rc = PTR_ERR(pending_child)); + LASSERT(pending_child->d_inode != NULL); + + cleanup_phase = 2; /* dput(pending_child) when finished */ + handle = fsfilt_start(obd, pending_dir, FSFILT_OP_UNLINK_LOG, + NULL); + if (IS_ERR(handle)) { + rc = PTR_ERR(handle); + handle = NULL; + GOTO(cleanup, rc); + } + +#ifdef ENABLE_ORPHANS + if (req != NULL && + (reply_body->valid & OBD_MD_FLEASIZE) && + mds_log_op_unlink(obd, pending_child->d_inode, + req->rq_repmsg, 1) > 0) { + reply_body->valid |= OBD_MD_FLCOOKIE; + } +#endif + pending_child->d_fsdata = (void *) &dp; + dp.p_inum = 0; + dp.p_ptr = req; + if (S_ISDIR(pending_child->d_inode->i_mode)) + rc = vfs_rmdir(pending_dir, pending_child); + else + rc = vfs_unlink(pending_dir, pending_child); + if (rc) + CERROR("error unlinking orphan %s: rc %d\n",fidname,rc); + } else if (mfd->mfd_mode & FMODE_WRITE && + mds_query_write_access(inode) == 0) { + // XXX this should probably be abstracted with mds_reint_setattr + struct iattr iattr; + + rc = mds_close_io_epoch(mds, inode, request_body, 0); + if (rc == EAGAIN) + goto dput; + +#if 0 + /* XXX can't set block count with fsfilt_setattr (!) */ + iattr.ia_valid = ATTR_CTIME | ATTR_ATIME | + ATTR_MTIME | ATTR_SIZE; + iattr.ia_atime = request_body->atime; + iattr.ia_ctime = request_body->ctime; + iattr.ia_mtime = request_body->mtime; + iattr.ia_size = request_body->size; + /* iattr.ia_blocks = request_body->blocks */ + + handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL); + if (IS_ERR(handle)) + GOTO(cleanup, rc = PTR_ERR(handle)); + rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0); + if (rc) + CERROR("error in setattr(%s): rc %d\n", fidname, rc); +#endif + } + dput: + l_dput(mfd->mfd_dentry); + mds_mfd_destroy(mfd); + + cleanup: + if (req) { + rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0); + } else if (handle) { + int err = fsfilt_commit(obd, pending_dir, handle, 0); + if (err) { + CERROR("error committing close: %d\n", err); + if (!rc) + rc = err; + } + } + + switch (cleanup_phase) { + case 2: + dput(pending_child); + case 1: + up(&pending_dir->i_sem); + } + RETURN(rc); +} + +int mds_close(struct ptlrpc_request *req) +{ + struct mds_export_data *med = &req->rq_export->exp_mds_data; + struct obd_device *obd = req->rq_export->exp_obd; + struct mds_body *body; + struct mds_file_data *mfd; + struct obd_run_ctxt saved; + struct inode *inode; + int rc, repsize[3] = {sizeof(struct mds_body), + obd->u.mds.mds_max_mdsize, + obd->u.mds.mds_max_cookiesize}; + ENTRY; + + MDS_CHECK_RESENT(req, mds_reconstruct_generic(req)); + + body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body); + if (body == NULL) { + CERROR("Can't unpack body\n"); + req->rq_status = -EFAULT; + RETURN(-EFAULT); + } + + if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES) + /* do some stuff */ ; + + mfd = mds_handle2mfd(&body->handle); + if (mfd == NULL) { + DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64 + ": cookie "LPX64, body->fid1.id, body->handle.cookie); + req->rq_status = -ESTALE; + RETURN(-ESTALE); + } + + rc = lustre_pack_reply(req, 3, repsize, NULL); + if (rc) { + CERROR("lustre_pack_reply: rc = %d\n", rc); + req->rq_status = rc; + } + + inode = mfd->mfd_dentry->d_inode; + if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) { + body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body)); + LASSERT(body != NULL); + + mds_pack_inode2fid(&body->fid1, inode); + mds_pack_inode2body(body, inode); + mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1); + } + spin_lock(&med->med_open_lock); + list_del(&mfd->mfd_list); + spin_unlock(&med->med_open_lock); + + push_ctxt(&saved, &obd->obd_ctxt, NULL); + req->rq_status = mds_mfd_close(rc ? NULL : req, obd, mfd, 1); + pop_ctxt(&saved, &obd->obd_ctxt, NULL); + + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) { + CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n"); + req->rq_status = -ENOMEM; + mds_mfd_put(mfd); + RETURN(-ENOMEM); + } + + mds_mfd_put(mfd); + RETURN(0); +} + +int mds_done_writing(struct ptlrpc_request *req) +{ + struct mds_body *body; + int rc, size = sizeof(struct mds_body); + ENTRY; + + MDS_CHECK_RESENT(req, mds_reconstruct_generic(req)); + + body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body); + if (body == NULL) { + CERROR("Can't unpack body\n"); + req->rq_status = -EFAULT; + RETURN(-EFAULT); + } + + rc = lustre_pack_reply(req, 1, &size, NULL); + if (rc) { + CERROR("lustre_pack_reply: rc = %d\n", rc); + req->rq_status = rc; + } + + RETURN(0); +}