X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fmds%2Fmds_open.c;h=f00cac4f2d314bf32b04783e28912330c7122d01;hb=628ac7d56843566cf7b84e6ff794e6dfd016dee8;hp=07ea3d861834602510d2e1c2c392604186e0e330;hpb=c79f7a9cdeeef1e421e47cc853482b247d8bd969;p=fs%2Flustre-release.git diff --git a/lustre/mds/mds_open.c b/lustre/mds/mds_open.c index 07ea3d8..f00cac4 100644 --- a/lustre/mds/mds_open.c +++ b/lustre/mds/mds_open.c @@ -40,10 +40,10 @@ #endif #include -#include #include #include #include +#include #include "mds_internal.h" @@ -87,14 +87,14 @@ struct mds_file_data *mds_mfd_new(void) return mfd; } -static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle) +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) +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); @@ -112,13 +112,12 @@ static void mds_mfd_destroy(struct mds_file_data *mfd) mds_mfd_put(mfd); } - /* 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) + LASSERT(LUSTRE_FILTERDATA(inode) == NULL); + OBD_ALLOC(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data)); + if (LUSTRE_FILTERDATA(inode) == NULL) return -ENOMEM; LASSERT(igrab(inode) == inode); return 0; @@ -127,9 +126,9 @@ static int mds_alloc_filterdata(struct inode *inode) /* 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; + LASSERT(LUSTRE_FILTERDATA(inode) != NULL); + OBD_FREE(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data)); + LUSTRE_FILTERDATA(inode) = NULL; iput(inode); } @@ -156,7 +155,6 @@ static int mds_get_write_access(struct mds_obd *mds, struct inode *inode, RETURN(-ETXTBSY); } - 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, @@ -164,19 +162,22 @@ static int mds_get_write_access(struct mds_obd *mds, struct inode *inode, goto out; } - if (inode->i_filterdata == NULL) + if (MDS_FILTERDATA(inode) == NULL) mds_alloc_filterdata(inode); - if (inode->i_filterdata == NULL) { + if (MDS_FILTERDATA(inode) == NULL) { rc = -ENOMEM; goto out; } - if (epoch > mds->mds_io_epoch) + if (epoch > mds->mds_io_epoch) { mds->mds_io_epoch = epoch; - else + CDEBUG(D_INODE, "repair MDS epoch "LPU64" for ino %lu/%u\n", + mds->mds_io_epoch, inode->i_ino, inode->i_generation); + } else { 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); + } 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: if (rc == 0) atomic_inc(&inode->i_writecount); @@ -233,42 +234,47 @@ int mds_query_write_access(struct inode *inode) RETURN(atomic_read(&inode->i_writecount)); } -/* This replaces the VFS mds_dentry_open, it manages mfd and writecount */ +/* This replaces the VFS 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 ptlrpc_request *req, + struct mds_update_record *rec) { struct mds_export_data *med = &req->rq_export->exp_mds_data; struct mds_obd *mds = mds_req2mds(req); struct mds_file_data *mfd; struct mds_body *body; - int error; + int rc = 0; ENTRY; mfd = mds_mfd_new(); if (mfd == NULL) { CERROR("mds: out of memory\n"); - GOTO(cleanup_dentry, error = -ENOMEM); + GOTO(cleanup_dentry, rc = -ENOMEM); } body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body)); if (flags & FMODE_WRITE) { /* FIXME: in recovery, need to pass old epoch here */ - error = mds_get_write_access(mds, dentry->d_inode, 0); - if (error) - GOTO(cleanup_mfd, error); + rc = mds_get_write_access(mds, dentry->d_inode, rec->ur_ioepoch); + if (rc) + GOTO(cleanup_mfd, rc); body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch; } else if (flags & FMODE_EXEC) { - error = mds_deny_write_access(mds, dentry->d_inode); - if (error) - GOTO(cleanup_mfd, error); + rc = mds_deny_write_access(mds, dentry->d_inode); + if (rc) + GOTO(cleanup_mfd, rc); } dget(dentry); - /* Mark the file as open to handle open-unlink. */ - mds_open_orphan_inc(dentry->d_inode); + /* FIXME: invalidate update locks when first open for write comes in */ + + /* mark the file as open to handle open-unlink. */ + DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode); + mds_orphan_open_inc(dentry->d_inode); + UP_WRITE_I_ALLOC_SEM(dentry->d_inode); mfd->mfd_mode = flags; mfd->mfd_dentry = dentry; @@ -280,41 +286,42 @@ static struct mds_file_data *mds_dentry_open(struct dentry *dentry, mds_mfd_put(mfd); body->handle.cookie = mfd->mfd_handle.h_cookie; - RETURN(mfd); - cleanup_mfd: mds_mfd_put(mfd); mds_mfd_destroy(mfd); cleanup_dentry: - return ERR_PTR(error); + return ERR_PTR(rc); } -static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm, - struct lov_desc *desc) +static inline 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 dentry *dchild, void **handle, - obd_id **ids) +/* must be called with i_sem held */ +int +mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req, + int offset, struct mds_update_record *rec, + struct dentry *dchild, void **handle, + obd_id **ids) { - struct obdo *oa; + struct inode *inode = dchild->d_inode; + struct mds_obd *mds = &obd->u.mds; struct obd_trans_info oti = { 0 }; - struct mds_body *body; struct lov_stripe_md *lsm = NULL; struct lov_mds_md *lmm = NULL; - struct inode *inode = dchild->d_inode; - void *lmm_buf; int rc, lmm_bufsize, lmm_size; + struct obdo *oa = NULL; + struct mds_body *body; + void *lmm_buf; ENTRY; if (rec->ur_flags & MDS_OPEN_DELAY_CREATE || @@ -328,35 +335,37 @@ static int mds_create_objects(struct ptlrpc_request *req, int offset, if (body->valid & OBD_MD_FLEASIZE) RETURN(0); - OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids)); + OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids)); if (*ids == NULL) RETURN(-ENOMEM); oti.oti_objid = *ids; - + /* replay case */ - if(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { - LASSERT (rec->ur_fid2->id); + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { + LASSERT(id_ino(rec->ur_id2)); body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE; lmm_size = rec->ur_eadatalen; lmm = rec->ur_eadata; LASSERT(lmm); if (*handle == NULL) - *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL); + *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL); if (IS_ERR(*handle)) { rc = PTR_ERR(*handle); *handle = NULL; - GOTO(out_ids, rc); + RETURN(rc); } - mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc); + mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc); lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0); lmm_bufsize = req->rq_repmsg->buflens[offset]; - LASSERT(lmm_buf); + LASSERT(lmm_buf != NULL); LASSERT(lmm_bufsize >= lmm_size); + memcpy(lmm_buf, lmm, lmm_size); - rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size); + rc = fsfilt_set_md(obd, inode, *handle, lmm, + lmm_size, EA_LOV); if (rc) CERROR("open replay failed to set md:%d\n", rc); RETURN(0); @@ -367,41 +376,53 @@ static int mds_create_objects(struct ptlrpc_request *req, int offset, oa = obdo_alloc(); if (oa == NULL) - GOTO(out_ids, rc = -ENOMEM); + RETURN(-ENOMEM); oa->o_mode = S_IFREG | 0600; oa->o_id = inode->i_ino; oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num; 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 | OBD_MD_FLGROUP; + oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE | + OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP; oa->o_size = 0; - obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME| - OBD_MD_FLCTIME); + obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME | + OBD_MD_FLMTIME | OBD_MD_FLCTIME); if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) { /* check if things like lfs setstripe are sending us the ea */ if (rec->ur_flags & MDS_OPEN_HAS_EA) { rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, - mds->mds_osc_exp, + mds->mds_dt_exp, 0, &lsm, rec->ur_eadata); if (rc) GOTO(out_oa, rc); } else { - /* Per-directory striping default code removed, because - * it uses the same unnamed EA storage as the directory - * striping for CMD. -p */ - } + OBD_ALLOC(lmm, mds->mds_max_mdsize); + if (lmm == NULL) + GOTO(out_oa, rc = -ENOMEM); + + lmm_size = mds->mds_max_mdsize; + rc = mds_get_md(obd, dchild->d_parent->d_inode, + lmm, &lmm_size, 1, 0); + if (rc > 0) + rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, + mds->mds_dt_exp, + 0, &lsm, lmm); + OBD_FREE(lmm, mds->mds_max_mdsize); + if (rc) + GOTO(out_oa, rc); + } + LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS); - rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti); + oti.oti_flags |= OBD_MODE_CROW; + rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &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", + CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR), + "error creating objects for " + "inode %lu: rc = %d\n", inode->i_ino, rc); if (rc > 0) { CERROR("obd_create returned invalid " @@ -411,19 +432,24 @@ static int mds_create_objects(struct ptlrpc_request *req, int offset, GOTO(out_oa, rc); } } else { - rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp, + rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp, 0, &lsm, rec->ur_eadata); - if (rc) { + if (rc) GOTO(out_oa, rc); - } + lsm->lsm_object_id = oa->o_id; lsm->lsm_object_gr = oa->o_gr; } 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); + obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME | + OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE); + + /* pack lustre id to oss */ + *(obdo_id(oa)) = body->id1; + oa->o_valid |= OBD_MD_FLIFID; + + rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti, NULL); if (rc) { CERROR("error setting attrs for inode %lu: rc %d\n", inode->i_ino, rc); @@ -440,10 +466,14 @@ static int mds_create_objects(struct ptlrpc_request *req, int offset, 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); + rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm); + if (!id_ino(rec->ur_id2)) + obd_free_memmd(mds->mds_dt_exp, &lsm); + if (rc < 0) { + CERROR("cannot pack lsm, err = %d\n", rc); + GOTO(out_oa, rc); + } + lmm_size = rc; body->eadatasize = rc; @@ -452,28 +482,82 @@ static int mds_create_objects(struct ptlrpc_request *req, int offset, if (IS_ERR(*handle)) { rc = PTR_ERR(*handle); *handle = NULL; - GOTO(out_ids, rc); + GOTO(out_oa, rc); } - rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size); + rc = fsfilt_set_md(obd, inode, *handle, lmm, + lmm_size, EA_LOV); + 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: + obd_free_diskmd(mds->mds_dt_exp, &lmm); +out_oa: oti_free_cookies(&oti); obdo_free(oa); - out_ids: +out_ids: if (rc) { - OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids)); + OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids)); *ids = NULL; } + if (lsm) + obd_free_memmd(mds->mds_dt_exp, &lsm); RETURN(rc); } +int +mds_destroy_object(struct obd_device *obd, + struct inode *inode, int async) +{ + struct mds_obd *mds = &obd->u.mds; + struct lov_mds_md *lmm = NULL; + int rc, lmm_size; + ENTRY; + + LASSERT(inode != NULL); + + if (inode->i_nlink != 0) { + CDEBUG(D_INODE, "attempt to destroy OSS object when " + "i_nlink == %d\n", (int)inode->i_nlink); + RETURN(0); + } + + OBD_ALLOC(lmm, mds->mds_max_mdsize); + if (lmm == NULL) + RETURN(-ENOMEM); + + lmm_size = mds->mds_max_mdsize; + rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0); + if (rc < 0) { + CERROR("no stripe info for %lu/%lu inode\n", + (unsigned long)inode->i_ino, + (unsigned long)inode->i_generation); + GOTO(out_free_lmm, rc); + } + + if (rc > 0) { + /* asynchronously unlink objecect on OSS */ + rc = mds_unlink_object(mds, inode, lmm, lmm_size, + NULL, 0, async); + if (rc) { + CERROR("error unlinking object on OSS, " + "err %d\n", rc); + GOTO(out_free_lmm, rc); + } + } else { + CDEBUG(D_INODE, "no stripping info found for inode " + "%lu/%lu\n", (unsigned long)inode->i_ino, + (unsigned long)inode->i_generation); + } + EXIT; +out_free_lmm: + OBD_FREE(lmm, mds->mds_max_mdsize); + return rc; +} + static void reconstruct_open(struct mds_update_record *rec, int offset, struct ptlrpc_request *req, struct lustre_handle *child_lockh) @@ -483,15 +567,15 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, struct mds_obd *mds = mds_req2mds(req); struct mds_file_data *mfd; struct obd_device *obd = req->rq_export->exp_obd; - struct dentry *parent, *child; + struct dentry *parent = NULL, *dchild; struct ldlm_reply *rep; struct mds_body *body; - int rc; struct list_head *t; int put_child = 1; + int rc; ENTRY; - LASSERT(offset == 2); /* only called via intent */ + LASSERT(offset == 3); /* only called via intent */ rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep)); body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body)); @@ -505,16 +589,27 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, return; /* error looking up parent or child */ } - parent = mds_fid2dentry(mds, rec->ur_fid1, NULL); - LASSERTF(!IS_ERR(parent), "fid "LPU64"/%u rc %ld\n", rec->ur_fid1->id, - rec->ur_fid1->generation, PTR_ERR(parent)); - - child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1); - LASSERTF(!IS_ERR(child), "parent "LPU64"/%u child %s rc %ld\n", - rec->ur_fid1->id, rec->ur_fid1->generation, rec->ur_name, - PTR_ERR(child)); + /* first, we try to open the file by fid. by the time of this + * request, inode can be an orphan and parent can disappear */ + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { + CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT|REPLAY)\n", + OLID4(rec->ur_id2)); + dchild = mds_id2dentry(obd, rec->ur_id2, NULL); + } else if (rec->ur_namelen == 1) { + CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n", + OLID4(rec->ur_id1)); + dchild = mds_id2dentry(obd, rec->ur_id1, NULL); + } else { + parent = mds_id2dentry(obd, rec->ur_id1, NULL); + LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n", + OLID4(rec->ur_id1), PTR_ERR(parent)); + dchild = ll_lookup_one_len(rec->ur_name, parent, + rec->ur_namelen - 1); + LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n", + OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild)); + } - if (!child->d_inode) + if (!dchild->d_inode) GOTO(out_dput, 0); /* child not present to open */ /* At this point, we know we have a child. We'll send @@ -524,11 +619,11 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, req->rq_status) GOTO(out_dput, 0); - mds_pack_inode2fid(obd, &body->fid1, child->d_inode); - mds_pack_inode2body(obd, body, child->d_inode); - if (S_ISREG(child->d_inode->i_mode)) { + /* get lock (write for O_CREAT, read otherwise) */ + mds_pack_inode2body(obd, body, dchild->d_inode, 1); + if (S_ISREG(dchild->d_inode->i_mode)) { rc = mds_pack_md(obd, req->rq_repmsg, 2, body, - child->d_inode, 1); + dchild->d_inode, 1, 0); if (rc) LASSERT(rc == req->rq_status); @@ -537,13 +632,9 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, if (!(body->valid & OBD_MD_FLEASIZE)) body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLATIME | OBD_MD_FLMTIME); - } else { - /* XXX need to check this case */ + DEBUG_REQ(D_ERROR, req, "no capa for "DLID4, OLID4(&body->id1)); } - - /* If we're opening a file without an EA, change to a write - lock (unless we already have one). */ - + /* If we have -EEXIST as the status, and we were asked to create * exclusively, we can tell we failed because the file already existed. */ @@ -580,8 +671,8 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, if (mfd == NULL) { mntget(mds->mds_vfsmnt); CERROR("Re-opened file \n"); - mfd = mds_dentry_open(child, mds->mds_vfsmnt, - rec->ur_flags & ~MDS_OPEN_TRUNC, req); + mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, + rec->ur_flags & ~MDS_OPEN_TRUNC, req, rec); if (!mfd) { CERROR("mds: out of memory\n"); GOTO(out_dput, req->rq_status = -ENOMEM); @@ -595,13 +686,14 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, out_dput: if (put_child) - l_dput(child); - l_dput(parent); + l_dput(dchild); + if (parent) + l_dput(parent); EXIT; } /* do NOT or the MAY_*'s, you'll get the weakest */ -static int accmode(int flags) +int accmode(int flags) { int res = 0; @@ -617,118 +709,174 @@ static int accmode(int flags) /* 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_update_record *rec, struct ldlm_reply *rep) { - struct mds_obd *mds = mds_req2mds(req); struct obd_device *obd = req->rq_export->exp_obd; + struct mds_obd *mds = mds_req2mds(req); + struct mds_export_data *med = &req->rq_export->u.eu_mds_data; struct mds_file_data *mfd = NULL; - obd_id *ids = NULL; /* object IDs created */ + obd_id *ids = NULL; unsigned mode; - int rc = 0; + int rc = 0, reply_off; ENTRY; /* atomically create objects if necessary */ down(&dchild->d_inode->i_sem); mode = dchild->d_inode->i_mode; + if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) || (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) { rc = mds_pack_md(obd, req->rq_repmsg, 2, body, - dchild->d_inode, 0); + dchild->d_inode, 0, 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, handle, &ids); - if (rc) { - CERROR("mds_create_objects: rc = %d\n", rc); + if ((body->valid & OBD_MD_FLEASIZE) && + (rec->ur_flags & MDS_OPEN_HAS_EA)) { up(&dchild->d_inode->i_sem); - RETURN(rc); + RETURN(-EEXIST); + } + + if (!(body->valid & OBD_MD_FLEASIZE)) { + /* no EA: create objects */ + rc = mds_create_objects(obd, req, 2, rec, + dchild, handle, &ids); + if (rc) { + CERROR("mds_create_object: rc = %d\n", rc); + up(&dchild->d_inode->i_sem); + RETURN(rc); + } } + if (S_ISREG(dchild->d_inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) { rc = mds_revalidate_lov_ea(obd, dchild->d_inode, req->rq_repmsg, 2); if (!rc) rc = mds_pack_md(obd, req->rq_repmsg, 2, body, - dchild->d_inode, 0); + dchild->d_inode, 0, 0); if (rc) { up(&dchild->d_inode->i_sem); RETURN(rc); } } } + + reply_off = 3; + rc = mds_pack_acl(req, reply_off, body, dchild->d_inode); + reply_off += 2; + + if (rc < 0) { + CERROR("pack posix acl: rc = %d\n", rc); + up(&dchild->d_inode->i_sem); + RETURN(rc); + } + + rc = mds_pack_gskey(obd, req->rq_repmsg, &reply_off, body, + dchild->d_inode); + if (rc < 0) { + CERROR("mds_pack_gskey: rc = %d\n", rc); + up(&dchild->d_inode->i_sem); + RETURN(rc); + } + + if (S_ISREG(mode)) { + struct lustre_capa capa = { + .lc_uid = rec->ur_uc.luc_uid, + .lc_op = capa_op(rec->ur_flags), + .lc_ino = dchild->d_inode->i_ino, + .lc_igen = dchild->d_inode->i_generation, + .lc_mdsid = mds->mds_num, + }; + + rc = mds_pack_capa(obd, med, NULL, &capa, req, + &reply_off, body); + if (rc < 0) { + CERROR("mds_pack_capa: rc = %d\n", rc); + up(&dchild->d_inode->i_sem); + RETURN(rc); + } + } else { + reply_off++; + } + /* If the inode has no EA data, then MDSs hold size, mtime */ - if (S_ISREG(dchild->d_inode->i_mode) && - !(body->valid & OBD_MD_FLEASIZE)) { + if (S_ISREG(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); + mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req, rec); if (IS_ERR(mfd)) RETURN(PTR_ERR(mfd)); 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); + mds_dt_update_objids(obd, ids); + OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_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) +static int mds_open_by_id(struct ptlrpc_request *req, + struct lustre_id *id, + 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; - char fidname[LL_FID_NAMELEN]; - int fidlen = 0, rc; + char idname[LL_ID_NAMELEN]; + int idlen = 0, rc; void *handle = NULL; ENTRY; down(&pending_dir->i_sem); - fidlen = ll_fid2str(fidname, fid->id, fid->generation); - dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen); + idlen = ll_id2str(idname, id_ino(id), id_gen(id)); + dchild = lookup_one_len(idname, mds->mds_pending_dir, + idlen); 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", + idname, rc); RETURN(rc); } + up(&pending_dir->i_sem); if (dchild->d_inode != NULL) { - up(&pending_dir->i_sem); mds_inode_set_orphan(dchild->d_inode); - mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode); - mds_pack_inode2body(req2obd(req), body, dchild->d_inode); + mds_pack_inode2body(req2obd(req), body, + dchild->d_inode, 1); + 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); + idname); goto open; } - dput(dchild); - up(&pending_dir->i_sem); + l_dput(dchild); - /* 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, fid, NULL); + /* 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_id2dentry(req2obd(req), id, NULL); if (IS_ERR(dchild)) RETURN(PTR_ERR(dchild)); - mds_pack_inode2fid(req2obd(req), &body->fid1, dchild->d_inode); - mds_pack_inode2body(req2obd(req), body, dchild->d_inode); + mds_pack_inode2body(req2obd(req), body, + dchild->d_inode, 1); + intent_set_disposition(rep, DISP_LOOKUP_EXECD); intent_set_disposition(rep, DISP_LOOKUP_POS); @@ -737,12 +885,11 @@ static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid, 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? */ - l_dput(dchild); RETURN(rc); } -int mds_pin(struct ptlrpc_request *req) +int mds_pin(struct ptlrpc_request *req, int offset) { struct obd_device *obd = req->rq_export->exp_obd; struct mds_body *request_body, *reply_body; @@ -750,7 +897,8 @@ int mds_pin(struct ptlrpc_request *req) int rc, size = sizeof(*reply_body); ENTRY; - request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body)); + request_body = lustre_msg_buf(req->rq_reqmsg, offset, + sizeof(*request_body)); rc = lustre_pack_reply(req, 1, &size, NULL); if (rc) @@ -758,31 +906,34 @@ int mds_pin(struct ptlrpc_request *req) reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body)); push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - rc = mds_open_by_fid(req, &request_body->fid1, reply_body, - request_body->flags, NULL, NULL); + rc = mds_open_by_id(req, &request_body->id1, reply_body, + request_body->flags, NULL, NULL); pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); RETURN(rc); } -/* Get a lock on the ino to sync with creation WRT inode reuse (bug 2029). - * If child_lockh is NULL we just get the lock as a barrier to wait for - * other holders of this lock, and drop it right away again. */ +/* + * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If + * child_lockh is NULL we just get the lock as a barrier to wait for other + * holders of this lock, and drop it right away again. + */ int mds_lock_new_child(struct obd_device *obd, struct inode *inode, struct lustre_handle *child_lockh) { struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } }; struct lustre_handle lockh; - int lock_flags = 0; + int lock_flags = LDLM_FL_ATOMIC_CB; int rc; + ENTRY; if (child_lockh == NULL) child_lockh = &lockh; rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id, LDLM_PLAIN, NULL, LCK_EX, &lock_flags, - mds_blocking_ast, ldlm_completion_ast, NULL, NULL, - NULL, 0, NULL, child_lockh); + mds_blocking_ast, ldlm_completion_ast, NULL, + NULL, NULL, 0, NULL, child_lockh); if (rc != ELDLM_OK) CERROR("ldlm_cli_enqueue: %d\n", rc); else if (child_lockh == &lockh) @@ -791,51 +942,39 @@ int mds_lock_new_child(struct obd_device *obd, struct inode *inode, RETURN(rc); } -static int is_mount_object(struct dentry *dparent) -{ - struct dentry *dchild; - - if (!(dparent->d_inode->i_mode & S_ISUID)) - return 0; - - dchild = lookup_one_len(".mntinfo", dparent, strlen(".mntinfo")); - if (IS_ERR(dchild) || dchild == NULL) - return 0; - - dput(dchild); - return 1; -} - 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 */ struct obd_device *obd = req->rq_export->exp_obd; + struct mds_export_data *med = &req->rq_export->u.eu_mds_data; struct mds_obd *mds = mds_req2mds(req); struct ldlm_reply *rep = NULL; struct mds_body *body = NULL; struct dentry *dchild = NULL, *dparent = NULL; - struct mds_export_data *med; - struct lustre_handle parent_lockh[2]; + struct lustre_handle parent_lockh[2] = {{0}, {0}}; int rc = 0, cleanup_phase = 0, acc_mode, created = 0; int parent_mode = LCK_PR; void *handle = NULL; struct dentry_params dp; struct mea *mea = NULL; int mea_size, update_mode; + int child_mode = LCK_PR; + struct lustre_id sid; + __u64 fid = 0; ENTRY; - DEBUG_REQ(D_INODE, req, "parent "LPU64"/%u name %*s mode %o", - rec->ur_fid1->id, rec->ur_fid1->generation, - rec->ur_namelen - 1, rec->ur_name, rec->ur_mode); + DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o", + OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name, + rec->ur_mode); - parent_lockh[0].cookie = 0; - parent_lockh[1].cookie = 0; + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE, + (obd_timeout + 1) / 4); - if (offset == 2) { /* intent */ + if (offset == 3) { /* intent */ rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep)); body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body)); - } else if (offset == 0) { /* non-intent reint */ + } else if (offset == 1) { /* non-intent reint */ body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body)); } else { body = NULL; @@ -844,33 +983,34 @@ int mds_open(struct mds_update_record *rec, int offset, MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh)); - MDS_UPDATE_COUNTER(mds, MDS_OPEN_COUNT); - - /* 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). */ + /* + * Step 0: If we are passed a id, 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). + */ if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { - DEBUG_REQ(D_HA, req, "open replay"); - CDEBUG(D_HA, "open fid "LPU64"/%u name %*s mode %o\n", - rec->ur_fid2->id, rec->ur_fid2->generation, - rec->ur_namelen - 1, rec->ur_name, rec->ur_mode); + CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n", + OLID4(rec->ur_id2), rec->ur_namelen - 1, + rec->ur_name, rec->ur_mode); - LASSERT(rec->ur_fid2->id); + LASSERT(id_ino(rec->ur_id2)); - rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags, - rec, rep); - if (rc != -ENOENT) + rc = mds_open_by_id(req, rec->ur_id2, body, + rec->ur_flags, rec, rep); + if (rc != -ENOENT) { + mds_body_do_reverse_map(med, body); RETURN(rc); + } + /* We didn't find the correct inode on disk either, so we * need to re-create it via a regular replay. */ LASSERT(rec->ur_flags & MDS_OPEN_CREAT); } else { - LASSERT(!rec->ur_fid2->id); + LASSERT(!id_ino(rec->ur_id2)); } - LASSERT(offset == 2); /* If we got here, we must be called via intent */ + LASSERT(offset == 3); /* 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"); RETURN(-ENOMEM); @@ -879,7 +1019,7 @@ int mds_open(struct mds_update_record *rec, int offset, acc_mode = accmode(rec->ur_flags); /* Step 1: Find and lock the parent */ - if (rec->ur_flags & O_CREAT) { + if (rec->ur_flags & MDS_OPEN_CREAT) { /* XXX Well, in fact we only need this lock mode change if in addition to O_CREAT, the file does not exist. But we do not know if it exists or not yet */ @@ -887,27 +1027,32 @@ int mds_open(struct mds_update_record *rec, int offset, } if (rec->ur_namelen == 1) { - /* client (LMV) wants to open the file by fid */ - CDEBUG(D_OTHER, "OPEN by fid %u/%u/%u\n", - (unsigned) rec->ur_fid1->mds, - (unsigned) rec->ur_fid1->id, - (unsigned) rec->ur_fid1->generation); - dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL); + /* client (LMV) wants to open the file by id */ + CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1)); + dchild = mds_id2dentry(obd, rec->ur_id1, NULL); if (IS_ERR(dchild)) { rc = PTR_ERR(dparent); - CERROR("child lookup by a fid error %d\n", rc); + CERROR("child lookup by an id error %d\n", rc); GOTO(cleanup, rc); } goto got_child; } - - dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode, - parent_lockh, &update_mode, rec->ur_name, - rec->ur_namelen - 1, - MDS_INODELOCK_UPDATE); + +restart: + dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode, + parent_lockh, &update_mode, rec->ur_name, + rec->ur_namelen - 1, MDS_INODELOCK_UPDATE); if (IS_ERR(dparent)) { rc = PTR_ERR(dparent); - CERROR("parent lookup error %d\n", rc); + if (rc != -ENOENT) { + CERROR("parent lookup for "DLID4" failed, error %d\n", + OLID4(rec->ur_id1), rc); + } else { + /* Just cannot find parent - make it look like + usual negative lookup to avoid extra MDS RPC */ + intent_set_disposition(rep, DISP_LOOKUP_EXECD); + intent_set_disposition(rep, DISP_LOOKUP_NEG); + } GOTO(cleanup, rc); } LASSERT(dparent->d_inode != NULL); @@ -915,22 +1060,24 @@ int mds_open(struct mds_update_record *rec, int offset, cleanup_phase = 1; /* parent dentry and lock */ /* try to retrieve MEA data for this dir */ - rc = mds_get_lmv_attr(obd, dparent->d_inode, &mea, &mea_size); + rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size); if (rc) GOTO(cleanup, rc); - if (mea != NULL) { - /* dir is already splitted, check is requested filename - * should live at this MDS or at another one */ + if (mea != NULL && mea->mea_count) { + /* + * dir is already splitted, check is requested filename should * + * live at this MDS or at another one. + */ int i; i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1); - if (mea->mea_master != mea->mea_fids[i].mds) { + if (mea->mea_master != id_group(&mea->mea_ids[i])) { CDEBUG(D_OTHER, "%s: inapropriate MDS(%d) for %lu/%u:%s." - " should be %d(%d)\n", obd->obd_name, + " should be %lu(%d)\n", obd->obd_name, mea->mea_master, dparent->d_inode->i_ino, dparent->d_inode->i_generation, rec->ur_name, - mea->mea_fids[i].mds, i); + (unsigned long)id_group(&mea->mea_ids[i]), i); GOTO(cleanup, rc = -ERESTART); } } @@ -946,35 +1093,55 @@ int mds_open(struct mds_update_record *rec, int offset, got_child: cleanup_phase = 2; /* child dentry */ - intent_set_disposition(rep, DISP_LOOKUP_EXECD); - if (dchild->d_flags & DCACHE_CROSS_REF) { - struct ldlm_res_id res_id = { . name = {0} }; - ldlm_policy_data_t policy; - int flags = 0; - CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n", - (unsigned long) dchild->d_mdsnum, - (unsigned long) dchild->d_inum, - (unsigned long) dchild->d_generation); - body->valid |= OBD_MD_FLID | OBD_MD_MDS; - body->fid1.id = dchild->d_inum; - body->fid1.mds = dchild->d_mdsnum; - body->fid1.generation = dchild->d_generation; - intent_set_disposition(rep, DISP_LOOKUP_POS); - res_id.name[0] = dchild->d_inum; - res_id.name[1] = dchild->d_generation; - policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP; - rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, - res_id, LDLM_IBITS, &policy, - LCK_PR, &flags, - mds_blocking_ast, - ldlm_completion_ast, NULL, NULL, - NULL, 0, NULL, child_lockh); + CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1)); + LASSERT(rec->ur_namelen > 1); + + /* we're gonna acquire LOOKUP lock on the child, + * but we have already locked parent and our order + * may conflict with enqueue_order_locks(). so, + * drop parent lock and acquire both the locks in + * common order. bug 6190 */ #ifdef S_PDIROPS if (parent_lockh[1].cookie != 0) ldlm_lock_decref(parent_lockh + 1, update_mode); #endif ldlm_lock_decref(parent_lockh, parent_mode); + l_dput(dchild); + l_dput(dparent); + cleanup_phase = 0; + + rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1, + parent_lockh, &dparent, + LCK_PR, MDS_INODELOCK_UPDATE, + &update_mode, rec->ur_name, + rec->ur_namelen, child_lockh, + &dchild, LCK_PR, + MDS_INODELOCK_LOOKUP); + + if (rc) + GOTO(cleanup, rc); + +#ifdef S_PDIROPS + if (parent_lockh[1].cookie != 0) + ldlm_lock_decref(parent_lockh + 1, update_mode); +#endif + ldlm_lock_decref(parent_lockh, LCK_PR); + + if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) { + CDEBUG(D_OTHER, "race: name changed (%p)\n", + dchild->d_inode); + if (dchild->d_inode) + ldlm_lock_decref(child_lockh, LCK_PR); + l_dput(dchild); + l_dput(dparent); + GOTO(restart, rc); + } + + mds_pack_dentry2body(obd, body, dchild, 1); + intent_set_disposition(rep, DISP_LOOKUP_POS); + intent_set_disposition(rep, DISP_LOOKUP_EXECD); + if (mea) OBD_FREE(mea, mea_size); l_dput(dchild); @@ -982,20 +1149,36 @@ got_child: RETURN(rc); } + intent_set_disposition(rep, DISP_LOOKUP_EXECD); + if (dchild->d_inode) intent_set_disposition(rep, DISP_LOOKUP_POS); else intent_set_disposition(rep, DISP_LOOKUP_NEG); - /*Step 3: If the child was negative, and we're supposed to, create it.*/ + /* 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; + unsigned long ino; struct iattr iattr; struct inode *inode; - rc = mds_try_to_split_dir(obd, dparent, &mea, 0, update_mode); + + /* get parent id: ldlm lock on the parent protects ea */ + rc = mds_read_inode_sid(obd, dparent->d_inode, &sid); + if (rc) { + CERROR("can't read parent inode id. ino(%lu) rc(%d)\n", + dparent->d_inode->i_ino, rc); + GOTO(cleanup, rc); + } + + ino = (unsigned long)id_ino(rec->ur_id2); + + rc = mds_try_to_split_dir(obd, dparent, &mea, 0, + update_mode); + CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n", obd->obd_name, dparent->d_inode->i_ino, dparent->d_inode->i_generation, rc); + if (rc > 0) { /* dir got splitted */ GOTO(cleanup, rc = -ERESTART); @@ -1005,7 +1188,10 @@ got_child: } if (!(rec->ur_flags & MDS_OPEN_CREAT)) { - /* It's negative and we weren't supposed to create it */ + /* + * dentry is negative and we weren't supposed to create + * object, get out of here. + */ GOTO(cleanup, rc = -ENOENT); } @@ -1017,11 +1203,19 @@ got_child: handle = NULL; GOTO(cleanup, rc); } + if (id_fid(rec->ur_id2)) + fid = id_fid(rec->ur_id2); + else + fid = mds_alloc_fid(obd); + dchild->d_fsdata = (void *) &dp; dp.p_ptr = req; dp.p_inum = ino; + + dp.p_fid = fid; + dp.p_group = mds->mds_num; - rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL); + rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL); if (dchild->d_fsdata == (void *)(unsigned long)ino) dchild->d_fsdata = NULL; @@ -1029,11 +1223,14 @@ got_child: CDEBUG(D_INODE, "error during create: %d\n", rc); GOTO(cleanup, rc); } + inode = dchild->d_inode; + if (ino) { LASSERT(ino == inode->i_ino); - /* Written as part of setattr */ - inode->i_generation = rec->ur_fid2->generation; + + /* written as part of setattr */ + inode->i_generation = id_gen(rec->ur_id2); CDEBUG(D_HA, "recreated ino %lu with gen %u\n", inode->i_ino, inode->i_generation); } @@ -1043,11 +1240,11 @@ got_child: LTIME_S(iattr.ia_ctime) = rec->ur_time; LTIME_S(iattr.ia_mtime) = rec->ur_time; - iattr.ia_uid = rec->_ur_fsuid; + iattr.ia_uid = rec->ur_fsuid; if (dparent->d_inode->i_mode & S_ISGID) iattr.ia_gid = dparent->d_inode->i_gid; else - iattr.ia_gid = rec->_ur_fsgid; + iattr.ia_gid = rec->ur_fsgid; iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME | ATTR_MTIME | ATTR_CTIME; @@ -1061,22 +1258,31 @@ got_child: rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0); if (rc) CERROR("error on parent setattr: rc = %d\n", rc); - else - MDS_UPDATE_COUNTER(mds, MDS_CREATE_COUNT); + else { + MD_COUNTER_INCREMENT(obd, create); + } - if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */ - rc = fsfilt_commit(obd, dchild->d_inode->i_sb, - dchild->d_inode, handle, 0); - handle = NULL; + mds_inode2id(obd, &body->id1, inode, fid); + mds_update_inode_ids(obd, dchild->d_inode, handle, &body->id1, &sid); + + if ((rec->ur_flags & MDS_OPEN_HAS_KEY)) { + rc = mds_set_gskey(obd, handle, dchild->d_inode, + rec->ur_ea2data, rec->ur_ea2datalen, + ATTR_KEY | ATTR_MAC); + if (rc) { + CERROR("error in set gs key rc %d\n", rc); + } } + rc = fsfilt_commit(obd, dchild->d_inode->i_sb, dchild->d_inode, handle, 0); + handle = NULL; acc_mode = 0; /* Don't check for permissions */ } - - LASSERT(!mds_inode_is_orphan(dchild->d_inode)); - - mds_pack_inode2fid(obd, &body->fid1, dchild->d_inode); - mds_pack_inode2body(obd, body, dchild->d_inode); + mds_pack_inode2body(obd, body, dchild->d_inode, 1); + + LASSERTF(!mds_inode_is_orphan(dchild->d_inode), + "dchild %.*s (%p) inode %p\n", dchild->d_name.len, + dchild->d_name.name, dchild, dchild->d_inode); if (S_ISREG(dchild->d_inode->i_mode)) { /* Check permissions etc */ @@ -1103,54 +1309,157 @@ got_child: GOTO(cleanup, rc = -EEXIST); // returns a lock to the client } + if (S_ISDIR(dchild->d_inode->i_mode)) { + if (rec->ur_flags & MDS_OPEN_CREAT || + rec->ur_flags & FMODE_WRITE) { + /* we are trying to create or write a exist dir */ + GOTO(cleanup, rc = -EISDIR); + } + if (ll_permission(dchild->d_inode, acc_mode, NULL)) + GOTO(cleanup, rc = -EACCES); + + /* + * here was checking for possible GNS mount points to skip their + * open. I removed it as its detection based only on SUID bit is + * not reliable. Opening such a dirs should not cause any + * problems, at least tests show that. --umka + */ + } + /* if we are following a symlink, don't open */ if (S_ISLNK(dchild->d_inode->i_mode)) - GOTO(cleanup, rc = 0); + GOTO(cleanup_no_trans, rc = 0); if ((rec->ur_flags & MDS_OPEN_DIRECTORY) && !S_ISDIR(dchild->d_inode->i_mode)) GOTO(cleanup, rc = -ENOTDIR); - if (S_ISDIR(dchild->d_inode->i_mode)) { - if (rec->ur_flags & MDS_OPEN_CREAT || - rec->ur_flags & FMODE_WRITE) { - /*we are trying to create or write a exist dir*/ - GOTO(cleanup, rc = -EISDIR); - } - if (ll_permission(dchild->d_inode, acc_mode, NULL)) { - GOTO(cleanup, rc = -EACCES); - } - if (is_mount_object(dchild)) { - CERROR("Found possible GNS mount object %*s; not " - "opening.\n", dchild->d_name.len, - dchild->d_name.name); - GOTO(cleanup, rc = 0); // success, but don't really open - } + /* check permission even it's special files */ + if (S_ISCHR(dchild->d_inode->i_mode) || + S_ISBLK(dchild->d_inode->i_mode) || + S_ISFIFO(dchild->d_inode->i_mode) || + S_ISSOCK(dchild->d_inode->i_mode)) { + rc = ll_permission(dchild->d_inode, acc_mode, NULL); + if (rc != 0) + GOTO(cleanup, rc); } + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) { + obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE; + GOTO(cleanup, rc = -EAGAIN); + } + +#if 0 + /* We cannot use acc_mode here, because it is zeroed in case of + creating a file, so we get wrong lockmode */ + if (accmode(rec->ur_flags) & MAY_WRITE) + child_mode = LCK_CW; + else if (accmode(rec->ur_flags) & MAY_EXEC) + child_mode = LCK_PR; + else + child_mode = LCK_CR; + + /* Always returning LOOKUP lock if open succesful to guard + * dentry on client. In case of replay we do not get a lock + * assuming that the caller has it already */ + if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) { + struct ldlm_res_id child_res_id = { .name = {0}}; + ldlm_policy_data_t policy; + int lock_flags = LDLM_FL_ATOMIC_CB; + + /* LOOKUP lock will protect dentry on client -bzzz */ + policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP | + MDS_INODELOCK_OPEN; + + child_res_id.name[0] = id_fid(&body->id1); + child_res_id.name[1] = id_group(&body->id1); + + rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, + child_res_id, LDLM_IBITS, &policy, + child_mode, &lock_flags, + mds_blocking_ast, ldlm_completion_ast, + NULL, NULL, NULL, 0, NULL, child_lockh); + if (rc != ELDLM_OK) + GOTO(cleanup, rc); + + cleanup_phase = 3; + } +#else +/* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */ +#endif + /* Step 5: mds_open it */ - rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec, - rep); - GOTO(cleanup, rc); + rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, + rec, rep); + if (rc) + GOTO(cleanup, rc); - cleanup: + /* reintegration case */ + if ((rec->ur_flags & MDS_REINT_REQ)) { + rc = mds_fidmap_add(obd, &body->id1); + if (rc < 0) { + CERROR("can't create fid->ino mapping, err %d\n", + rc); + } else { + rc = 0; + } + } + + /* if this is a writer, we have to invalidate client's + * update locks in order to make sure they don't use + * isize/iblocks from mds anymore. + * FIXME: can cause a deadlock, use mds_get_parent_child_locked() + * XXX: optimization is to do this for first writer only */ + if (accmode(rec->ur_flags) & MAY_WRITE) { + struct ldlm_res_id child_res_id = { .name = {0}}; + ldlm_policy_data_t sz_policy; + struct lustre_handle sz_lockh; + int lock_flags = LDLM_FL_ATOMIC_CB; + + child_res_id.name[0] = id_fid(&body->id1); + child_res_id.name[1] = id_group(&body->id1); + sz_policy.l_inodebits.bits = MDS_INODELOCK_UPDATE; + + rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, + child_res_id, LDLM_IBITS, &sz_policy, + LCK_PW, &lock_flags, mds_blocking_ast, + ldlm_completion_ast, NULL, NULL, NULL, + 0, NULL, &sz_lockh); + if (rc == ELDLM_OK) + ldlm_lock_decref(&sz_lockh, LCK_PW); + else + CERROR("can't invalidate client's update locks\n"); + } + + EXIT; +cleanup: rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle, req, rc, rep ? rep->lock_policy_res1 : 0); +cleanup_no_trans: switch (cleanup_phase) { + case 3: + if (rc) { + ldlm_lock_decref(child_lockh, child_mode); + child_lockh->cookie = 0; + } case 2: if (rc && created) { int err = vfs_unlink(dparent->d_inode, dchild); if (err) { - CERROR("unlink(%*s) in error path: %d\n", + CERROR("unlink(%.*s) in error path: %d\n", dchild->d_name.len, dchild->d_name.name, err); } } else if (created) { -#if 0 mds_lock_new_child(obd, dchild->d_inode, NULL); -#endif } + /* audit stuff for OPEN */ + if (offset == 3) { + mds_audit(req, dchild, rec->ur_name, + rec->ur_namelen - 1, AUDIT_OPEN, rc); + } + l_dput(dchild); case 1: if (dparent == NULL) @@ -1168,11 +1477,26 @@ got_child: } if (mea) OBD_FREE(mea, mea_size); + if (rc == 0) + atomic_inc(&mds->mds_open_count); + + mds_body_do_reverse_map(med, body); + + /* + * If we have not taken the "open" lock, we may not return 0 here, + * because caller expects 0 to mean "lock is taken", and it needs + * nonzero return here for caller to return EDLM_LOCK_ABORTED to + * client. Later caller should rewrite the return value back to zero + * if it to be used any further. + */ + if ((cleanup_phase != 3) && !rc) + rc = ENOLCK; + RETURN(rc); } /* Close a "file descriptor" and possibly unlink an orphan from the - * PENDING directory. + * PENDING directory. Caller must hold child->i_sem, this drops it. * * If we are being called from mds_disconnect() because the client has * disappeared, then req == NULL and we do not update last_rcvd because @@ -1180,12 +1504,13 @@ got_child: * (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) +int mds_mfd_close(struct ptlrpc_request *req, int offset, + 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; + char idname[LL_ID_NAMELEN]; + int last_orphan, idlen, 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; @@ -1197,16 +1522,41 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd, ENTRY; if (req && req->rq_reqmsg != NULL) - request_body = lustre_msg_buf(req->rq_reqmsg, 0, + request_body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*request_body)); if (req && req->rq_repmsg != NULL) reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body)); - fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation); + if (request_body && (request_body->valid & OBD_MD_FLSIZE)) { + /* we set i_size/i_blocks here, nobody will see + * them until all write references are dropped. + * btw, we hold one reference */ + LASSERT(mfd->mfd_mode & FMODE_WRITE); + LASSERT(request_body->valid & OBD_MD_FLEPOCH); + LASSERT(MDS_FILTERDATA(inode)); + if (MDS_FILTERDATA(inode)->io_epoch != request_body->io_epoch) + CDEBUG(D_ERROR, "try to update attr. for old epoch " + LPD64" while current "LPD64"\n", + MDS_FILTERDATA(inode)->io_epoch, + request_body->io_epoch); + i_size_write(inode, request_body->size); + inode->i_blocks = request_body->blocks; + LTIME_S(inode->i_mtime) = (request_body->mtime); + + LTIME_S(iattr.ia_mtime) = request_body->mtime; + iattr.ia_size = inode->i_size; + iattr.ia_valid |= ATTR_SIZE|ATTR_MTIME; + mds_inode_unset_attrs_old(inode); + } + + idlen = ll_id2str(idname, inode->i_ino, inode->i_generation); + CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, + idname, inode->i_nlink, mds_orphan_open_count(inode)); - last_orphan = mds_open_orphan_dec_test(inode) && + last_orphan = mds_orphan_open_dec_test(inode) && mds_inode_is_orphan(inode); + UP_WRITE_I_ALLOC_SEM(inode); /* this is half of the actual "close" */ if (mfd->mfd_mode & FMODE_WRITE) { @@ -1221,27 +1571,38 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd, int stripe_count = 0; LASSERT(rc == 0); /* mds_put_write_access must have succeeded */ - if (obd->obd_recovering) { - CDEBUG(D_HA, "not remove orphan %s until recovery" - " is over\n", fidname); - GOTO(out, rc); - } - - CDEBUG(D_HA, "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; + CDEBUG(D_HA, "destroying orphan object %s\n", idname); + + if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) || + (S_ISDIR(inode->i_mode) && inode->i_nlink != 2)) + CERROR("found \"orphan\" %s %s with link count %d\n", + S_ISREG(inode->i_mode) ? "file" : "dir", + idname, inode->i_nlink); + /* + * 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. + */ down(&pending_dir->i_sem); cleanup_phase = 1; /* up(i_sem) when finished */ - pending_child = lookup_one_len(fidname, mds->mds_pending_dir, - fidlen); + pending_child = lookup_one_len(idname, mds->mds_pending_dir, + idlen); if (IS_ERR(pending_child)) GOTO(cleanup, rc = PTR_ERR(pending_child)); - LASSERT(pending_child->d_inode != NULL); + if (pending_child->d_inode == NULL) { + CERROR("orphan %s has been removed\n", idname); + GOTO(cleanup, rc = 0); + } cleanup_phase = 2; /* dput(pending_child) when finished */ + if (S_ISDIR(pending_child->d_inode->i_mode)) { + rc = vfs_rmdir(pending_dir, pending_child); + if (rc) + CERROR("error unlinking orphan dir %s: rc %d\n", + idname, rc); + goto out; + } + if (req != NULL && req->rq_repmsg != NULL) { lmm = lustre_msg_buf(req->rq_repmsg, 1, 0); stripe_count = le32_to_cpu(lmm->lmm_stripe_count); @@ -1258,12 +1619,10 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd, 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); + rc = vfs_unlink(pending_dir, pending_child); if (rc) - CERROR("error unlinking orphan %s: rc %d\n",fidname,rc); + CERROR("error unlinking orphan %s: rc %d\n", + idname, rc); if (req != NULL && req->rq_repmsg != NULL && (reply_body->valid & OBD_MD_FLEASIZE) && @@ -1273,10 +1632,21 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd, req->rq_repmsg->buflens[2], &lcl) > 0) { reply_body->valid |= OBD_MD_FLCOOKIE; } + + rc = mds_destroy_object(obd, inode, 1); + if (rc) { + CERROR("cannot destroy OSS object on close, err %d\n", + rc); + rc = 0; + } goto out; /* Don't bother updating attrs on unlinked inode */ + } else if ((mfd->mfd_mode & FMODE_WRITE) && rc == 0) { + /* last writer closed file - let's update i_size/i_blocks */ + mds_validate_size(obd, inode, request_body, &iattr); } +#if 0 if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) { /* Update the on-disk attributes if this was the last write * close, and all information was provided (i.e., rc == 0) @@ -1284,7 +1654,6 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd, * XXX this should probably be abstracted with mds_reint_setattr */ -#if 0 if (request_body->valid & OBD_MD_FLMTIME && LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) { LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime); @@ -1306,9 +1675,8 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd, iattr.ia_blocks = request_body->blocks } */ -#endif } - +#endif if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) { /* Only start a transaction to write out only the atime if * it is more out-of-date than the specified limit. If we @@ -1324,25 +1692,40 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd, } if (iattr.ia_valid != 0) { - handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL); + if (handle == NULL) + 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); + CERROR("error in setattr(%s): rc %d\n", idname, rc); } out: /* If other clients have this file open for write, rc will be > 0 */ if (rc > 0) rc = 0; + if (!obd->obd_recovering && mds_inode_has_old_attrs(inode) + && !mds_inode_is_orphan(inode) + && atomic_read(&inode->i_writecount) == 0 + && inode->i_nlink != 0) { + CERROR("leave inode %lu/%u with old attributes (nlink = %d)\n", + inode->i_ino, inode->i_generation, inode->i_nlink); + } l_dput(mfd->mfd_dentry); mds_mfd_destroy(mfd); cleanup: + atomic_dec(&mds->mds_open_count); if (req != NULL && reply_body != NULL) { rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0); } else if (handle) { - int err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, 0); + int err, force_sync = 0; + + if (req && req->rq_export) + force_sync = req->rq_export->exp_sync; + + err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, + force_sync); if (err) { CERROR("error committing close: %d\n", err); if (!rc) @@ -1361,7 +1744,141 @@ out: RETURN(rc); } -int mds_close(struct ptlrpc_request *req) +static int mds_extent_lock_callback(struct ldlm_lock *lock, + struct ldlm_lock_desc *new, void *data, + int flag) +{ + struct lustre_handle lockh = { 0 }; + int rc; + ENTRY; + + switch (flag) { + case LDLM_CB_BLOCKING: + ldlm_lock2handle(lock, &lockh); + rc = ldlm_cli_cancel(&lockh); + if (rc != ELDLM_OK) + CERROR("ldlm_cli_cancel failed: %d\n", rc); + break; + case LDLM_CB_CANCELING: { + break; + } + default: + LBUG(); + } + + RETURN(0); +} +__u64 lov_merge_size(struct lov_stripe_md *lsm, int kms); +__u64 lov_merge_blocks(struct lov_stripe_md *lsm); +__u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time); + +int mds_validate_size(struct obd_device *obd, struct inode *inode, + struct mds_body *body, struct iattr *iattr) +{ + ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } }; + struct lustre_handle lockh = { 0 }; + struct lov_stripe_md *lsm = NULL; + int rc, len, flags; + void *lmm = NULL; + ENTRY; + + /* we update i_size/i_blocks only for regular files */ + if (!S_ISREG(inode->i_mode)) + RETURN(0); + + /* we shouldn't fetch size from OSTes during recovery - deadlock */ + if (obd->obd_recovering) { + CERROR("size-on-mds has no support on OST yet\n"); + RETURN(0); + } + + /* if nobody modified attrs. we're lucky */ + if (!mds_inode_has_old_attrs(inode)) + RETURN(0); + + /* 1: client didn't send actual i_size/i_blocks + * 2: we seem to be last writer + * 3: the file doesn't look like to be disappeared + * conclusion: we're gonna fetch them from OSSes */ + + down(&inode->i_sem); + len = fsfilt_get_md(obd, inode, NULL, 0, EA_LOV); + up(&inode->i_sem); + + if (len < 0) { + CERROR("error getting inode %lu MD: %d\n", inode->i_ino, len); + GOTO(cleanup, rc = len); + } else if (len == 0) { + CDEBUG(D_INODE, "no LOV in inode %lu\n", inode->i_ino); + GOTO(cleanup, rc = 0); + } + + OBD_ALLOC(lmm, len); + if (lmm == NULL) { + CERROR("can't allocate memory\n"); + GOTO(cleanup, rc = -ENOMEM); + } + + down(&inode->i_sem); + rc = fsfilt_get_md(obd, inode, lmm, len, EA_LOV); + up(&inode->i_sem); + + if (rc < 0) { + CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc); + GOTO(cleanup, rc); + } + + rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, lmm, len); + if (rc < 0) { + CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc); + GOTO(cleanup, rc); + } + + CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino); + + flags = LDLM_FL_HAS_INTENT; + rc = obd_enqueue(obd->u.mds.mds_dt_exp, lsm, LDLM_EXTENT, &policy, + LCK_PR, &flags, mds_extent_lock_callback, + ldlm_completion_ast, NULL, NULL, + sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh); + if (rc == -ENOENT) { + /* while we were enqueueing lock on OST, another thread + * unlinked the file and started OST object destoying. + * it's safe to return 0 here */ + GOTO(cleanup, rc = 0); + } else if (rc != 0) { + CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc); + GOTO(cleanup, rc); + } + + CDEBUG(D_INODE, "LOV reports "LPD64"/%lu for "DLID4" [%s%s%s]\n", + inode->i_size, inode->i_blocks, OLID4(&body->id1), + atomic_read(&inode->i_writecount) > 1 ? "U" : "", + mds_inode_has_old_attrs(inode) ? "D" : "", + mds_inode_is_orphan(inode) ? "O" : ""); + + i_size_write(inode, lov_merge_size(lsm, 0)); + inode->i_blocks = lov_merge_blocks(lsm); + LTIME_S(inode->i_mtime) = lov_merge_mtime(lsm, LTIME_S(inode->i_mtime)); + iattr->ia_size = inode->i_size; + LTIME_S(iattr->ia_mtime) = LTIME_S(inode->i_mtime); + iattr->ia_valid |= ATTR_SIZE | ATTR_MTIME; + + DOWN_WRITE_I_ALLOC_SEM(inode); + mds_inode_unset_attrs_old(inode); + UP_WRITE_I_ALLOC_SEM(inode); + + obd_cancel(obd->u.mds.mds_dt_exp, lsm, LCK_PR, &lockh); + +cleanup: + if (lsm != NULL) + obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm); + if (lmm != NULL) + OBD_FREE(lmm, len); + RETURN(rc); +} + +int mds_close(struct ptlrpc_request *req, int offset) { struct mds_export_data *med = &req->rq_export->exp_mds_data; struct obd_device *obd = req->rq_export->exp_obd; @@ -1373,8 +1890,7 @@ int mds_close(struct ptlrpc_request *req) obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize}; ENTRY; - - MDS_UPDATE_COUNTER((&obd->u.mds), MDS_CLOSE_COUNT); + MD_COUNTER_INCREMENT(obd, close); rc = lustre_pack_reply(req, 3, repsize, NULL); if (rc) { @@ -1385,14 +1901,14 @@ int mds_close(struct ptlrpc_request *req) } if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { - DEBUG_REQ(D_HA, req, "close replay\n"); + DEBUG_REQ(D_HA, req, "close replay"); memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0), - lustre_msg_buf(req->rq_reqmsg, 1, 0), + lustre_msg_buf(req->rq_reqmsg, offset + 1, 0), req->rq_repmsg->buflens[2]); } - - body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body); + body = lustre_swab_reqbuf(req, offset, sizeof(*body), + lustre_swab_mds_body); if (body == NULL) { CERROR("Can't unpack body\n"); req->rq_status = -EFAULT; @@ -1405,26 +1921,41 @@ int mds_close(struct ptlrpc_request *req) 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); + ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie); req->rq_status = -ESTALE; RETURN(-ESTALE); } 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(obd, &body->fid1, inode); - mds_pack_inode2body(obd, body, inode); - mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1); + /* child i_alloc_sem protects orphan_dec_test && is_orphan race */ + DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */ + + if (body->flags & MDS_BFLAG_DIRTY_EPOCH) { + /* the client modified data through the handle + * we need to care about attrs. -bzzz */ + mds_inode_set_attrs_old(inode); + } + + if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) { + struct mds_body *rep_body; + + rep_body = lustre_msg_buf(req->rq_repmsg, 0, + sizeof (*rep_body)); + LASSERT(rep_body != NULL); + + mds_pack_inode2body(obd, rep_body, inode, + (body->valid & OBD_MD_FID) ? 1 : 0); + + mds_pack_md(obd, req->rq_repmsg, 1, rep_body, + inode, MDS_PACK_MD_LOCK, 0); } spin_lock(&med->med_open_lock); list_del(&mfd->mfd_list); spin_unlock(&med->med_open_lock); push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - req->rq_status = mds_mfd_close(req, obd, mfd, 1); + req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1); pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) { @@ -1438,7 +1969,8 @@ int mds_close(struct ptlrpc_request *req) RETURN(0); } -int mds_done_writing(struct ptlrpc_request *req) + +int mds_done_writing(struct ptlrpc_request *req, int offset) { struct mds_body *body; int rc, size = sizeof(struct mds_body); @@ -1446,7 +1978,8 @@ int mds_done_writing(struct ptlrpc_request *req) MDS_CHECK_RESENT(req, mds_reconstruct_generic(req)); - body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body); + body = lustre_swab_reqbuf(req, offset, sizeof(*body), + lustre_swab_mds_body); if (body == NULL) { CERROR("Can't unpack body\n"); req->rq_status = -EFAULT;