X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmds%2Fhandler.c;h=dfdad070040a1e8adf7c782ded27c0fb213aeb2e;hp=fd7846d7723593c3073948dcf3125955b22c4e4c;hb=e44e9b278432a1df83482b1cd83b2081fabe94dc;hpb=090c677210ee2946d99c71412e4ff762bb300f4f diff --git a/lustre/mds/handler.c b/lustre/mds/handler.c index fd7846d..dfdad07 100644 --- a/lustre/mds/handler.c +++ b/lustre/mds/handler.c @@ -101,7 +101,8 @@ static int mds_sendpage(struct ptlrpc_request *req, struct file *file, file->f_dentry->d_inode->i_size); rc = fsfilt_readpage(req->rq_export->exp_obd, file, - page_address(pages[i]), tmpsize, &offset); + kmap(pages[i]), tmpsize, &offset); + kunmap(pages[i]); if (rc != tmpsize) GOTO(cleanup_buf, rc = -EIO); @@ -130,7 +131,7 @@ static int mds_sendpage(struct ptlrpc_request *req, struct file *file, rc = -ETIMEDOUT; /* XXX should this be a different errno? */ } - + DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n", (rc == -ETIMEDOUT) ? "timeout" : "network error", desc->bd_nob_transferred, count, @@ -154,10 +155,50 @@ static int mds_sendpage(struct ptlrpc_request *req, struct file *file, return rc; } +int mds_lock_mode_for_dir(struct obd_device *obd, + struct dentry *dentry, int mode) +{ + int ret_mode; + + /* any dir access needs couple locks: + * 1) on part of dir we gonna lookup/modify in + * 2) on a whole dir to protect it from concurrent splitting + * and to flush client's cache for readdir() + * so, for a given mode and dentry this routine decides what + * lock mode to use for lock #2: + * 1) if caller's gonna lookup in dir then we need to protect + * dir from being splitted only - LCK_CR + * 2) if caller's gonna modify dir then we need to protect + * dir from being splitted and to flush cache - LCK_CW + * 3) if caller's gonna modify dir and that dir seems ready + * for splitting then we need to protect it from any + * type of access (lookup/modify/split) - LCK_EX -bzzz */ + + if (mode == LCK_PR) { + ret_mode = LCK_CR; + } else if (mode == LCK_PW) { + /* caller gonna modify directory.we use concurrent + write lock here to retract client's cache for readdir */ + ret_mode = LCK_CW; + if (mds_splitting_expected(obd, dentry)) { + /* splitting possible. serialize any access */ + CERROR("%s: gonna split %lu/%lu\n", + obd->obd_name, + (unsigned long) dentry->d_inode->i_ino, + (unsigned long) dentry->d_inode->i_generation); + ret_mode = LCK_EX; + } + } else { + CWARN("unexpected lock mode %d\n", mode); + ret_mode = LCK_EX; + } + return ret_mode; +} + /* only valid locked dentries or errors should be returned */ struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid, struct vfsmount **mnt, int lock_mode, - struct lustre_handle *lockh, + struct lustre_handle *lockh, int *mode, char *name, int namelen, __u64 lockpart) { struct mds_obd *mds = &obd->u.mds; @@ -173,27 +214,27 @@ struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid, res_id.name[0] = de->d_inode->i_ino; res_id.name[1] = de->d_inode->i_generation; -#ifdef S_PDIROPS lockh[1].cookie = 0; +#ifdef S_PDIROPS if (name && IS_PDIROPS(de->d_inode)) { ldlm_policy_data_t cpolicy = { .l_inodebits = { MDS_INODELOCK_UPDATE } }; - /* lock just dir { ino, generation } to flush client cache */ - if (lock_mode == LCK_PW) { - rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, - res_id, LDLM_IBITS, - &cpolicy, LCK_CW, &flags, - mds_blocking_ast, - ldlm_completion_ast, NULL, NULL, - NULL, 0, NULL, lockh + 1); - if (rc != ELDLM_OK) { - l_dput(de); - RETURN(ERR_PTR(-ENOLCK)); - } - flags = 0; + LASSERT(mode != NULL); + *mode = mds_lock_mode_for_dir(obd, de, lock_mode); + rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, + res_id, LDLM_IBITS, + &cpolicy, *mode, &flags, + mds_blocking_ast, + ldlm_completion_ast, NULL, NULL, + NULL, 0, NULL, lockh + 1); + if (rc != ELDLM_OK) { + l_dput(de); + RETURN(ERR_PTR(-ENOLCK)); } + flags = 0; res_id.name[2] = full_name_hash(name, namelen); + CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n", de->d_inode->i_ino, de->d_inode->i_generation, res_id.name[2]); @@ -250,6 +291,9 @@ struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid, if (!inode) RETURN(ERR_PTR(-ENOENT)); +#if 0 + /* here we disabled generation check, as root inode i_generation + * of cache mds and real mds are different. */ if (generation && inode->i_generation != generation) { /* we didn't find the right inode.. */ CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n", @@ -259,6 +303,7 @@ struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid, dput(result); RETURN(ERR_PTR(-ENOENT)); } +#endif if (mnt) { *mnt = mds->mds_vfsmnt; @@ -333,7 +378,7 @@ out: return rc; } -static int mds_init_export(struct obd_export *exp) +static int mds_init_export(struct obd_export *exp) { struct mds_export_data *med = &exp->exp_mds_data; @@ -346,7 +391,7 @@ static int mds_destroy_export(struct obd_export *export) { struct mds_export_data *med; struct obd_device *obd = export->exp_obd; - struct obd_run_ctxt saved; + struct lvfs_run_ctxt saved; int rc = 0; ENTRY; @@ -356,7 +401,8 @@ static int mds_destroy_export(struct obd_export *export) if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid)) GOTO(out, 0); - push_ctxt(&saved, &obd->obd_ctxt, NULL); + push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); + /* Close any open files (which may also cause orphan unlinking). */ spin_lock(&med->med_open_lock); while (!list_empty(&med->med_open_head)) { @@ -371,11 +417,13 @@ static int mds_destroy_export(struct obd_export *export) list_del(&mfd->mfd_list); spin_unlock(&med->med_open_lock); + /* If you change this message, be sure to update + * replay_single:test_46 */ CERROR("force closing client file handle for %*s (%s:%lu)\n", dentry->d_name.len, dentry->d_name.name, ll_bdevname(dentry->d_inode->i_sb, btmp), dentry->d_inode->i_ino); - rc = mds_mfd_close(NULL, obd, mfd, + rc = mds_mfd_close(NULL, obd, mfd, !(export->exp_flags & OBD_OPT_FAILOVER)); if (rc) @@ -383,7 +431,7 @@ static int mds_destroy_export(struct obd_export *export) spin_lock(&med->med_open_lock); } spin_unlock(&med->med_open_lock); - pop_ctxt(&saved, &obd->obd_ctxt, NULL); + pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); out: mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER)); @@ -518,6 +566,39 @@ int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, RETURN(0); } +int mds_get_md(struct obd_device *obd, struct inode *inode, void *md, + int *size, int lock) +{ + int rc = 0; + int lmm_size; + + if (lock) + down(&inode->i_sem); + rc = fsfilt_get_md(obd, inode, md, *size); + if (lock) + up(&inode->i_sem); + + if (rc < 0) { + CERROR("Error %d reading eadata for ino %lu\n", + rc, inode->i_ino); + } else if (rc > 0) { + lmm_size = rc; + + if (S_ISREG(inode->i_mode)) + rc = mds_convert_lov_ea(obd, inode, md, lmm_size); + + if (rc == 0) { + *size = lmm_size; + rc = lmm_size; + } else if (rc > 0) { + *size = rc; + } + } + + RETURN (rc); +} + + /* Call with lock=1 if you want mds_pack_md to take the i_sem. * Call with lock=0 if the caller has already taken the i_sem. */ int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset, @@ -549,21 +630,12 @@ int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset, // RETURN(-EINVAL); } - if (lock) - down(&inode->i_sem); - rc = fsfilt_get_md(obd, inode, lmm, lmm_size); - if (lock) - up(&inode->i_sem); - if (rc < 0) { - CERROR("Error %d reading eadata for ino %lu\n", - rc, inode->i_ino); - } else if (rc > 0) { - lmm_size = rc; - if (S_ISREG(inode->i_mode)) - rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size); - if (rc > 0) - lmm_size = rc; - body->valid |= OBD_MD_FLEASIZE; + rc = mds_get_md(obd, inode, lmm, &lmm_size, lock); + if (rc > 0) { + if (S_ISDIR(inode->i_mode)) + body->valid |= OBD_MD_FLDIREA; + else + body->valid |= OBD_MD_FLEASIZE; body->eadatasize = lmm_size; rc = 0; } @@ -600,12 +672,14 @@ static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry, mds_pack_inode2fid(obd, &body->fid1, inode); mds_pack_inode2body(obd, body, inode); - if (S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE) != 0) { + if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) || + (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) { rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body, inode, 1); /* If we have LOV EA data, the OST holds size, atime, mtime */ - if (!(body->valid & OBD_MD_FLEASIZE)) + if (!(body->valid & OBD_MD_FLEASIZE) && + !(body->valid & OBD_MD_FLDIREA)) body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLATIME | OBD_MD_FLMTIME); } else if (S_ISLNK(inode->i_mode) && @@ -630,11 +704,7 @@ static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry, symname[rc] = 0; /* NULL terminate */ rc = 0; } - } else if (S_ISDIR(inode->i_mode) && - (reqbody->valid & OBD_MD_FLEASIZE) != 0) { - rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body, - inode, 1); - } + } RETURN(rc); } @@ -675,8 +745,8 @@ static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode, LASSERT(body != NULL); /* checked by caller */ LASSERT_REQSWABBED(req, offset); /* swabbed by caller */ - if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode)) - && (body->valid & OBD_MD_FLEASIZE)) { + if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) || + (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) { int rc; down(&inode->i_sem); rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0); @@ -723,25 +793,49 @@ static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode, return(rc); } -#define DENTRY_VALID(dentry) \ - ((dentry)->d_inode || ((dentry)->d_flags & DCACHE_CROSS_REF)) +int mds_check_mds_num(struct obd_device *obd, struct inode* inode, + char *name, int namelen) +{ + struct mea *mea = NULL; + int mea_size, rc = 0; + ENTRY; + + rc = mds_get_lmv_attr(obd, inode, &mea, &mea_size); + if (rc) + RETURN(rc); + if (mea != NULL) { + /* dir is already splitted, check is requested filename + * should live at this MDS or at another one */ + int i; + i = mea_name2idx(mea, name, namelen - 1); + if (mea->mea_master != i) { + CERROR("inapropriate MDS(%d) for %s. should be %d\n", + mea->mea_master, name, i); + rc = -ERESTART; + } + } + + if (mea) + OBD_FREE(mea, mea_size); + RETURN(rc); +} static int mds_getattr_name(int offset, struct ptlrpc_request *req, struct lustre_handle *child_lockh, int child_part) { struct obd_device *obd = req->rq_export->exp_obd; struct ldlm_reply *rep = NULL; - struct obd_run_ctxt saved; + struct lvfs_run_ctxt saved; struct mds_body *body; struct dentry *dparent = NULL, *dchild = NULL; - struct obd_ucred uc; + struct lvfs_ucred uc; struct lustre_handle parent_lockh[2]; - int namesize; + int namesize, update_mode; int rc = 0, cleanup_phase = 0, resent_req = 0; char *name; ENTRY; - LASSERT(!strcmp(obd->obd_type->typ_name, "mds")); + LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME)); /* Swab now, before anyone looks inside the request */ @@ -762,17 +856,17 @@ static int mds_getattr_name(int offset, struct ptlrpc_request *req, LASSERT (offset == 0 || offset == 2); /* if requests were at offset 2, the getattr reply goes back at 1 */ - if (offset) { + if (offset) { rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep)); offset = 1; } - uc.ouc_fsuid = body->fsuid; - uc.ouc_fsgid = body->fsgid; - uc.ouc_cap = body->capability; - uc.ouc_suppgid1 = body->suppgid; - uc.ouc_suppgid2 = -1; - push_ctxt(&saved, &obd->obd_ctxt, &uc); + uc.luc_fsuid = body->fsuid; + uc.luc_fsgid = body->fsgid; + uc.luc_cap = body->capability; + uc.luc_suppgid1 = body->suppgid; + uc.luc_suppgid2 = -1; + push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc); cleanup_phase = 1; /* kernel context */ intent_set_disposition(rep, DISP_LOOKUP_EXECD); @@ -784,7 +878,8 @@ static int mds_getattr_name(int offset, struct ptlrpc_request *req, obd->obd_name, (unsigned long) body->fid1.id, (unsigned long) body->fid1.generation); dchild = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR, - parent_lockh, NULL, 0, child_part); + parent_lockh, &update_mode, + NULL, 0, child_part); if (IS_ERR(dchild)) { CERROR("can't find inode: %d\n", (int) PTR_ERR(dchild)); GOTO(cleanup, rc = PTR_ERR(dchild)); @@ -792,7 +887,7 @@ static int mds_getattr_name(int offset, struct ptlrpc_request *req, memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0])); #ifdef S_PDIROPS if (parent_lockh[1].cookie) - ldlm_lock_decref(parent_lockh + 1, LCK_CW); + ldlm_lock_decref(parent_lockh + 1, update_mode); #endif cleanup_phase = 2; goto fill_inode; @@ -830,7 +925,7 @@ static int mds_getattr_name(int offset, struct ptlrpc_request *req, rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1, parent_lockh, &dparent, LCK_PR, MDS_INODELOCK_LOOKUP, - name, namesize, + &update_mode, name, namesize, child_lockh, &dchild, LCK_PR, child_part); if (rc) @@ -853,12 +948,19 @@ static int mds_getattr_name(int offset, struct ptlrpc_request *req, cleanup_phase = 2; /* dchild, dparent, locks */ + /* let's make sure this name should leave on this mds node */ + rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize); + if (rc) + GOTO(cleanup, rc); + fill_inode: + if (!DENTRY_VALID(dchild)) { intent_set_disposition(rep, DISP_LOOKUP_NEG); /* in the intent case, the policy clears this error: the disposition is enough */ - GOTO(cleanup, rc = -ENOENT); + rc = -ENOENT; + GOTO(cleanup, rc); } else { intent_set_disposition(rep, DISP_LOOKUP_POS); } @@ -888,7 +990,7 @@ fill_inode: #ifdef S_PDIROPS if (parent_lockh[1].cookie != 0) ldlm_lock_decref(parent_lockh + 1, - LCK_CW); + update_mode); #endif } if (dparent) @@ -896,7 +998,7 @@ fill_inode: } l_dput(dchild); case 1: - pop_ctxt(&saved, &obd->obd_ctxt, &uc); + pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc); default: ; } return rc; @@ -906,10 +1008,10 @@ static int mds_getattr(int offset, struct ptlrpc_request *req) { struct mds_obd *mds = mds_req2mds(req); struct obd_device *obd = req->rq_export->exp_obd; - struct obd_run_ctxt saved; + struct lvfs_run_ctxt saved; struct dentry *de; struct mds_body *body; - struct obd_ucred uc; + struct lvfs_ucred uc; int rc = 0; ENTRY; @@ -920,10 +1022,10 @@ static int mds_getattr(int offset, struct ptlrpc_request *req) RETURN (-EFAULT); } - uc.ouc_fsuid = body->fsuid; - uc.ouc_fsgid = body->fsgid; - uc.ouc_cap = body->capability; - push_ctxt(&saved, &obd->obd_ctxt, &uc); + uc.luc_fsuid = body->fsuid; + uc.luc_fsgid = body->fsgid; + uc.luc_cap = body->capability; + push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc); de = mds_fid2dentry(mds, &body->fid1, NULL); if (IS_ERR(de)) { rc = req->rq_status = -ENOENT; @@ -941,7 +1043,7 @@ static int mds_getattr(int offset, struct ptlrpc_request *req) l_dput(de); GOTO(out_pop, rc); out_pop: - pop_ctxt(&saved, &obd->obd_ctxt, &uc); + pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc); return rc; } @@ -1044,9 +1146,9 @@ static int mds_readpage(struct ptlrpc_request *req) struct dentry *de; struct file *file; struct mds_body *body, *repbody; - struct obd_run_ctxt saved; + struct lvfs_run_ctxt saved; int rc, size = sizeof(*repbody); - struct obd_ucred uc; + struct lvfs_ucred uc; ENTRY; rc = lustre_pack_reply(req, 1, &size, NULL); @@ -1059,10 +1161,10 @@ static int mds_readpage(struct ptlrpc_request *req) if (body == NULL) GOTO (out, rc = -EFAULT); - uc.ouc_fsuid = body->fsuid; - uc.ouc_fsgid = body->fsgid; - uc.ouc_cap = body->capability; - push_ctxt(&saved, &obd->obd_ctxt, &uc); + uc.luc_fsuid = body->fsuid; + uc.luc_fsgid = body->fsgid; + uc.luc_cap = body->capability; + push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc); de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt); if (IS_ERR(de)) GOTO(out_pop, rc = PTR_ERR(de)); @@ -1101,7 +1203,7 @@ static int mds_readpage(struct ptlrpc_request *req) out_file: filp_close(file, 0); out_pop: - pop_ctxt(&saved, &obd->obd_ctxt, &uc); + pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc); out: req->rq_status = rc; RETURN(0); @@ -1146,6 +1248,7 @@ static int mds_filter_recovery_request(struct ptlrpc_request *req, case OBD_PING: case MDS_REINT: case LDLM_ENQUEUE: + case OST_CREATE: *process = target_queue_recovery_request(req, obd); RETURN(0); @@ -1172,121 +1275,180 @@ static char *reint_names[] = { OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME|\ OBD_MD_FLID) +static void reconstruct_create(struct ptlrpc_request *req) +{ + struct mds_export_data *med = &req->rq_export->exp_mds_data; + struct mds_client_data *mcd = med->med_mcd; + struct ost_body *body; + + body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body)); + + /* copy rc, transno and disp; steal locks */ + mds_req_from_mcd(req, mcd); + CERROR("reconstruct reply for x"LPU64"\n", req->rq_xid); +} + static int mdt_obj_create(struct ptlrpc_request *req) { - struct obd_export *exp = req->rq_export; - struct obd_device *obd = exp->exp_obd; + struct obd_device *obd = req->rq_export->exp_obd; + struct ldlm_res_id res_id = { .name = {0} }; struct mds_obd *mds = &obd->u.mds; struct ost_body *body, *repbody; int rc, size = sizeof(*repbody); char fidname[LL_FID_NAMELEN]; struct inode *parent_inode; - struct obd_run_ctxt saved; - struct dentry *new_child; - int err, namelen, mealen; - struct obd_ucred uc; + struct lustre_handle lockh; + struct lvfs_run_ctxt saved; + ldlm_policy_data_t policy; + struct dentry *new = NULL; + struct dentry_params dp; + int mealen, flags = 0; + unsigned int tmpname; + struct lvfs_ucred uc; struct mea *mea; void *handle; ENTRY; + DEBUG_REQ(D_HA, req, "create remote object"); + parent_inode = mds->mds_objects_dir->d_inode; - body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body); + body = lustre_swab_reqbuf(req, 0, sizeof(*body), + lustre_swab_ost_body); if (body == NULL) RETURN(-EFAULT); - uc.ouc_fsuid = body->oa.o_uid; - uc.ouc_fsgid = body->oa.o_gid; + MDS_CHECK_RESENT(req, reconstruct_create(req)); + + uc.luc_fsuid = body->oa.o_uid; + uc.luc_fsgid = body->oa.o_gid; - push_ctxt(&saved, &obd->obd_ctxt, &uc); + push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc); rc = lustre_pack_reply(req, 1, &size, NULL); if (rc) RETURN(rc); repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody)); - memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); - - namelen = ll_fid2str(fidname, body->oa.o_id, body->oa.o_generation); - - down(&parent_inode->i_sem); - new_child = lookup_one_len(fidname, mds->mds_objects_dir, namelen); - if (new_child->d_inode != NULL) { - CERROR("impossible non-negative obj dentry " LPU64":%u!\n", - repbody->oa.o_id, repbody->oa.o_generation); - LBUG(); - } - handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode, - FSFILT_OP_MKDIR, NULL); - /* FIXME: error handling here */ - LASSERT(!IS_ERR(handle)); - rc = vfs_mkdir(parent_inode, new_child, body->oa.o_mode); - up(&parent_inode->i_sem); - /* FIXME: error handling here */ - if (rc) - CERROR("vfs_mkdir() returned %d\n", rc); - LASSERT(rc == 0); + if (body->oa.o_flags & OBD_FL_RECREATE_OBJS) { + /* this is re-create request from MDS holding directory name. + * we have to lookup given ino/generation first. if it exists + * (good case) then there is nothing to do. if it does not + * then we have to recreate it */ + struct ll_fid fid; + fid.id = body->oa.o_id; + fid.generation = body->oa.o_generation; + new = mds_fid2dentry(mds, &fid, NULL); + if (!IS_ERR(new) && new->d_inode) { + CWARN("mkdir() repairing is on its way: %lu/%lu\n", + (unsigned long) fid.id, + (unsigned long) fid.generation); + obdo_from_inode(&repbody->oa, new->d_inode, + FILTER_VALID_FLAGS); + repbody->oa.o_id = new->d_inode->i_ino; + repbody->oa.o_generation = new->d_inode->i_generation; + repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER; + GOTO(cleanup2, rc = 0); + } + CWARN("hmm. for some reason dir %lu/%lu (or reply) got lost\n", + (unsigned long) fid.id, (unsigned long) fid.generation); + LASSERT(new->d_inode == NULL || + new->d_inode->i_generation != fid.generation); + l_dput(new); + } - /* mark this object non-splittable */ - mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL); - OBD_ALLOC(mea, mealen); - LASSERT(mea != NULL); - mea->mea_count = 0; - down(&new_child->d_inode->i_sem); - handle = fsfilt_start(obd, new_child->d_inode, FSFILT_OP_SETATTR, NULL); + down(&parent_inode->i_sem); + handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL); LASSERT(!IS_ERR(handle)); - rc = fsfilt_set_md(obd, new_child->d_inode, handle, mea, mealen); - LASSERT(rc == 0); - fsfilt_commit(obd, new_child->d_inode, handle, 0); - LASSERT(rc == 0); - up(&new_child->d_inode->i_sem); - OBD_FREE(mea, mealen); - - err = fsfilt_commit(exp->exp_obd, mds->mds_objects_dir->d_inode, - handle, 0); - /* FIXME: error handling here */ - LASSERT(err == 0); - - obdo_from_inode(&repbody->oa, new_child->d_inode, FILTER_VALID_FLAGS); - repbody->oa.o_id = new_child->d_inode->i_ino; - repbody->oa.o_generation = new_child->d_inode->i_generation; - CDEBUG(D_OTHER, "created dirobj: %lu, %lu mode %o, uid %u, gid %u\n", - (unsigned long) repbody->oa.o_id, - (unsigned long) new_child->d_inode->i_ino, - (unsigned) new_child->d_inode->i_mode, - (unsigned) new_child->d_inode->i_uid, - (unsigned) new_child->d_inode->i_gid); - dput(new_child); - pop_ctxt(&saved, &obd->obd_ctxt, &uc); - RETURN(0); -} -static int mds_get_info(struct obd_export *exp, __u32 keylen, - void *key, __u32 *vallen, void *val) -{ - struct obd_device *obd; - struct mds_obd *mds; - ENTRY; +repeat: + tmpname = ll_insecure_random_int(); + rc = sprintf(fidname, "%u", tmpname); + new = lookup_one_len(fidname, mds->mds_objects_dir, rc); + if (IS_ERR(new)) { + CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n", + obd->obd_name, fidname, (int) PTR_ERR(new)); + fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0); + up(&parent_inode->i_sem); + RETURN(PTR_ERR(new)); + } else if (new->d_inode) { + CERROR("%s: name exists. repeat\n", obd->obd_name); + goto repeat; + } + + new->d_fsdata = (void *) &dp; + dp.p_inum = 0; + dp.p_ptr = req; + + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { + DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu", + (unsigned long) body->oa.o_id, + (unsigned long) body->oa.o_generation); + dp.p_inum = body->oa.o_id; + dp.p_generation = body->oa.o_generation; + } + rc = vfs_mkdir(parent_inode, new, body->oa.o_mode); + if (rc == 0) { + obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS); + repbody->oa.o_id = new->d_inode->i_ino; + repbody->oa.o_generation = new->d_inode->i_generation; + repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER; - obd = class_exp2obd(exp); - if (obd == NULL) { - CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n", - exp->exp_handle.h_cookie); - RETURN(-EINVAL); - } + rc = fsfilt_del_dir_entry(obd, new); + up(&parent_inode->i_sem); - mds = &obd->u.mds; - keylen == strlen("mdsize"); - if (keylen && memcmp(key, "mdsize", keylen) == 0) { - __u32 *mdsize = val; - *vallen = sizeof(*mdsize); - *mdsize = mds->mds_max_mdsize; - RETURN(0); - } + if (rc) { + CERROR("can't remove name for object: %d\n", rc); + GOTO(cleanup, rc); + } + + /* this lock should be taken to serialize MDS modifications + * in failure case */ + res_id.name[0] = new->d_inode->i_ino; + res_id.name[1] = new->d_inode->i_generation; + policy.l_inodebits.bits = MDS_INODELOCK_UPDATE; + rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, + res_id, LDLM_IBITS, &policy, + LCK_EX, &flags, mds_blocking_ast, + ldlm_completion_ast, NULL, NULL, + NULL, 0, NULL, &lockh); + if (rc != ELDLM_OK) + GOTO(cleanup, rc); - CDEBUG(D_IOCTL, "invalid key\n"); - RETURN(-EINVAL); + CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n", + (unsigned long) new->d_inode->i_ino, + (unsigned long) new->d_inode->i_generation, + (unsigned) new->d_inode->i_mode); + } else { + up(&parent_inode->i_sem); + CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc); + } + + if (rc == 0 && body->oa.o_valid & OBD_MD_FLID) { + /* this is new object for splitted dir. we have to + * prevent recursive splitting on it -bzzz */ + mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL); + OBD_ALLOC(mea, mealen); + if (mea == NULL) + GOTO(cleanup, rc = -ENOMEM); + mea->mea_count = 0; + down(&new->d_inode->i_sem); + rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen); + up(&new->d_inode->i_sem); + OBD_FREE(mea, mealen); + } + +cleanup: + rc = mds_finish_transno(mds, parent_inode, handle, req, rc, 0); + if (rc == 0) + ptlrpc_save_lock(req, &lockh, LCK_EX); + else + ldlm_lock_decref(&lockh, LCK_EX); +cleanup2: + l_dput(new); + pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc); + RETURN(rc); } static int mdt_get_info(struct ptlrpc_request *req) @@ -1318,10 +1480,11 @@ static int mdt_get_info(struct ptlrpc_request *req) } static int mds_set_info(struct obd_export *exp, __u32 keylen, - void *key, __u32 vallen, void *val) + void *key, __u32 vallen, void *val) { struct obd_device *obd; struct mds_obd *mds; + int rc = 0; ENTRY; obd = class_exp2obd(exp); @@ -1332,50 +1495,78 @@ static int mds_set_info(struct obd_export *exp, __u32 keylen, } mds = &obd->u.mds; - keylen == strlen("client"); - if (keylen && memcmp(key, "client", keylen) == 0) { + if (keylen == strlen("mds_num") && + memcmp(key, "mds_num", keylen) == 0) { + int valsize; + __u32 group; + CDEBUG(D_IOCTL, "set mds num %d\n", *(int*)val); + mds->mds_num = *(int*)val; + group = FILTER_GROUP_FIRST_MDS + mds->mds_num; + valsize = sizeof(group); + /*mds number has been changed, so the corresponding obdfilter exp + *need to be changed too*/ + rc = obd_set_info(mds->mds_osc_exp, strlen("mds_conn"), "mds_conn", + valsize, &group); + RETURN(rc); + } else if (keylen == strlen("client") && + memcmp(key, "client", keylen) == 0) { if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) { atomic_inc(&mds->mds_real_clients); - CDEBUG(D_OTHER, "%s: peer from %s is real client (%d)\n", - obd->obd_name, - exp->exp_client_uuid.uuid, - atomic_read(&mds->mds_real_clients)); + CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n", + obd->obd_name, exp->exp_client_uuid.uuid, + atomic_read(&mds->mds_real_clients)); exp->exp_flags |= OBD_OPT_REAL_CLIENT; } + if (mds->mds_lmv_name) { + rc = mds_lmv_connect(obd, mds->mds_lmv_name); + LASSERT(rc == 0); + } RETURN(0); } - CDEBUG(D_IOCTL, "invalid key\n"); RETURN(-EINVAL); } static int mdt_set_info(struct ptlrpc_request *req) { - char *key; + char *key, *val; struct obd_export *exp = req->rq_export; - int keylen, rc = 0, size = sizeof(obd_id); + int keylen, rc = 0, vallen; ENTRY; key = lustre_msg_buf(req->rq_reqmsg, 0, 1); if (key == NULL) { - DEBUG_REQ(D_HA, req, "no get_info key"); + DEBUG_REQ(D_HA, req, "no set_info key"); RETURN(-EFAULT); } keylen = req->rq_reqmsg->buflens[0]; - if ((keylen < strlen("client") || memcmp(key, "client", 6) != 0)) - RETURN(-EPROTO); + if (keylen == strlen("mds_num") && + memcmp(key, "mds_num", keylen) == 0) { + rc = lustre_pack_reply(req, 0, NULL, NULL); + if (rc) + RETURN(rc); + val = lustre_msg_buf(req->rq_reqmsg, 1, 0); - rc = lustre_pack_reply(req, 0, NULL, NULL); - if (rc) + vallen = req->rq_reqmsg->buflens[1]; + + rc = obd_set_info(exp, keylen, key, vallen, val); + req->rq_repmsg->status = 0; RETURN(rc); - rc = obd_set_info(exp, keylen, key, size, NULL); - req->rq_repmsg->status = 0; - RETURN(rc); + } else if (keylen == strlen("client") && + memcmp(key, "client", keylen) == 0) { + rc = lustre_pack_reply(req, 0, NULL, NULL); + if (rc) + RETURN(rc); + rc = obd_set_info(exp, keylen, key, sizeof(obd_id), NULL); + req->rq_repmsg->status = 0; + RETURN(rc); + } + CDEBUG(D_IOCTL, "invalid key\n"); + RETURN(-EINVAL); } extern int ost_brw_write(struct ptlrpc_request *, struct obd_trans_info *); - int mds_handle(struct ptlrpc_request *req) { int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET; @@ -1403,6 +1594,21 @@ int mds_handle(struct ptlrpc_request *req) obd = req->rq_export->exp_obd; mds = &obd->u.mds; + /* sanity check: if the xid matches, the request must + * be marked as a resent or replayed */ + if (req->rq_xid == med->med_mcd->mcd_last_xid) + LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) & + (MSG_RESENT | MSG_REPLAY), + "rq_xid "LPU64" matches last_xid, " + "expected RESENT flag\n", + req->rq_xid); + /* else: note the opposite is not always true; a + * RESENT req after a failover will usually not match + * the last_xid, since it was likely never + * committed. A REPLAYed request will almost never + * match the last xid, however it could for a + * committed, but still retained, open. */ + /* Check for aborted recovery. */ spin_lock_bh(&obd->obd_processing_task_lock); abort_recovery = obd->obd_abort_recovery; @@ -1588,6 +1794,11 @@ int mds_handle(struct ptlrpc_request *req) OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0); rc = llog_origin_handle_next_block(req); break; + case LLOG_ORIGIN_HANDLE_PREV_BLOCK: + DEBUG_REQ(D_INODE, req, "llog prev block"); + OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0); + rc = llog_origin_handle_prev_block(req); + break; case LLOG_ORIGIN_HANDLE_READ_HEADER: DEBUG_REQ(D_INODE, req, "llog read header"); OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0); @@ -1680,12 +1891,12 @@ int mds_update_server_data(struct obd_device *obd, int force_sync) struct mds_obd *mds = &obd->u.mds; struct mds_server_data *msd = mds->mds_server_data; struct file *filp = mds->mds_rcvd_filp; - struct obd_run_ctxt saved; + struct lvfs_run_ctxt saved; loff_t off = 0; int rc; ENTRY; - push_ctxt(&saved, &obd->obd_ctxt, NULL); + push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno); CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n", @@ -1693,20 +1904,20 @@ int mds_update_server_data(struct obd_device *obd, int force_sync) rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync); if (rc) CERROR("error writing MDS server data: rc = %d\n", rc); - pop_ctxt(&saved, &obd->obd_ctxt, NULL); + pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); RETURN(rc); } - /* mount the file system (secretly) */ static int mds_setup(struct obd_device *obd, obd_count len, void *buf) { struct lustre_cfg* lcfg = buf; struct mds_obd *mds = &obd->u.mds; + char *options = NULL; struct vfsmount *mnt; - int rc = 0; unsigned long page; + int rc = 0; ENTRY; dev_clear_rdonly(2); @@ -1720,38 +1931,65 @@ static int mds_setup(struct obd_device *obd, obd_count len, void *buf) mds->mds_max_mdsize = sizeof(struct lov_mds_md); + page = __get_free_page(GFP_KERNEL); + if (!page) + RETURN(-ENOMEM); + + options = (char *)page; + memset(options, 0, PAGE_SIZE); + + /* here we use "iopen_nopriv" hardcoded, because it affects MDS utility + * and the rest of options are passed by mount options. Probably this + * should be moved to somewhere else like startup scripts or lconf. */ + sprintf(options, "iopen_nopriv"); + + if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4) + sprintf(options + strlen(options), ",%s", + lcfg->lcfg_inlbuf4); + /* we have to know mdsnum before touching underlying fs -bzzz */ - if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4) { + if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 && + strcmp(lcfg->lcfg_inlbuf5, "dumb")) { class_uuid_t uuid; CDEBUG(D_OTHER, "MDS: %s is master for %s\n", - obd->obd_name, lcfg->lcfg_inlbuf4); + obd->obd_name, lcfg->lcfg_inlbuf5); generate_random_uuid(uuid); class_uuid_unparse(uuid, &mds->mds_lmv_uuid); - OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen4); + OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen5); if (mds->mds_lmv_name == NULL) RETURN(rc = -ENOMEM); - memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf4, - lcfg->lcfg_inllen4); + memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf5, + lcfg->lcfg_inllen5); + rc = mds_lmv_connect(obd, mds->mds_lmv_name); if (rc) { - OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen4); + OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen5); GOTO(err_ops, rc); } } + + /* FIXME-WANGDI: this should be reworked when we will use lmv along + * with cobd, because correct mdsnum is set in mds_lmv_connect(). */ + if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 && !mds->mds_lmv_obd && + strcmp(lcfg->lcfg_inlbuf6, "dumb")) { + if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master")) && + mds->mds_num == 0) { + mds->mds_num = REAL_MDS_NUMBER; + } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache")) && + mds->mds_num == 0) { + mds->mds_num = CACHE_MDS_NUMBER; + } + } + + mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0, + lcfg->lcfg_inlbuf1, options); - if (!(page = __get_free_page(GFP_KERNEL))) - RETURN(-ENOMEM); - - memset((void *)page, 0, PAGE_SIZE); - sprintf((char *)page, "iopen_nopriv"); - - mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0, - lcfg->lcfg_inlbuf1, (void *)page); free_page(page); + if (IS_ERR(mnt)) { rc = PTR_ERR(mnt); CERROR("do_kern_mount failed: rc = %d\n", rc); @@ -1784,7 +2022,6 @@ static int mds_setup(struct obd_device *obd, obd_count len, void *buf) rc = llog_start_commit_thread(); if (rc < 0) GOTO(err_fs, rc); - if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) { class_uuid_t uuid; @@ -1793,7 +2030,7 @@ static int mds_setup(struct obd_device *obd, obd_count len, void *buf) class_uuid_unparse(uuid, &mds->mds_lov_uuid); OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3); - if (mds->mds_profile == NULL) + if (mds->mds_profile == NULL) GOTO(err_fs, rc = -ENOMEM); memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3, @@ -1807,6 +2044,7 @@ static int mds_setup(struct obd_device *obd, obd_count len, void *buf) rc = mds_postsetup(obd); if (rc) GOTO(err_fs, rc); + RETURN(0); err_fs: @@ -1828,26 +2066,28 @@ err_ops: static int mds_postsetup(struct obd_device *obd) { struct mds_obd *mds = &obd->u.mds; - struct llog_ctxt *ctxt; int rc = 0; ENTRY; - rc = llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT, - obd, 0, NULL, &llog_lvfs_ops); + rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT, + obd, 0, NULL, &llog_lvfs_ops); if (rc) RETURN(rc); - if (mds->mds_profile) { - struct obd_run_ctxt saved; + /* This check for @dumb string is needed to handle mounting MDS + with smfs. Read lconf:MDSDEV.write_conf() for more detail + explanation. */ + if (mds->mds_profile && strcmp(mds->mds_profile, "dumb")) { + struct lvfs_run_ctxt saved; struct lustre_profile *lprof; struct config_llog_instance cfg; cfg.cfg_instance = NULL; cfg.cfg_uuid = mds->mds_lov_uuid; - push_ctxt(&saved, &obd->obd_ctxt, NULL); - ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT); - rc = class_config_parse_llog(ctxt, mds->mds_profile, &cfg); - pop_ctxt(&saved, &obd->obd_ctxt, NULL); + push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); + rc = class_config_parse_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT), + mds->mds_profile, &cfg); + pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); if (rc) GOTO(err_llog, rc); @@ -1870,45 +2110,69 @@ static int mds_postsetup(struct obd_device *obd) err_cleanup: mds_lov_clean(obd); err_llog: - ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT); - llog_cleanup(ctxt); + obd_llog_cleanup(llog_get_context(&obd->obd_llogs, + LLOG_CONFIG_ORIG_CTXT)); RETURN(rc); } -static int mds_postrecov(struct obd_device *obd) - +int mds_postrecov(struct obd_device *obd) { + struct mds_obd *mds = &obd->u.mds; struct llog_ctxt *ctxt; - int rc, rc2; + int rc, item = 0; ENTRY; LASSERT(!obd->obd_recovering); ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT); LASSERT(ctxt != NULL); + /* set nextid first, so we are sure it happens */ + rc = mds_lov_set_nextid(obd); + if (rc) { + CERROR("%s: mds_lov_set_nextid failed\n", obd->obd_name); + GOTO(out, rc); + } + + /* clean PENDING dir */ + rc = mds_cleanup_orphans(obd); + if (rc < 0) + GOTO(out, rc); + item = rc; + rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count, NULL, NULL, NULL); - if (rc != 0) { - CERROR("faild at llog_origin_connect: %d\n", rc); + if (rc) { + CERROR("%s: failed at llog_origin_connect: %d\n", + obd->obd_name, rc); + GOTO(out, rc); } - rc = mds_cleanup_orphans(obd); + /* remove the orphaned precreated objects */ + rc = mds_lov_clearorphans(mds, NULL /* all OSTs */); + if (rc) + GOTO(err_llog, rc); - rc2 = mds_lov_set_nextid(obd); - if (rc2 == 0) - rc2 = rc; - RETURN(rc2); +out: + RETURN(rc < 0 ? rc : item); + +err_llog: + /* cleanup all llogging subsystems */ + rc = obd_llog_finish(obd, &obd->obd_llogs, + mds->mds_lov_desc.ld_tgt_count); + if (rc) + CERROR("%s: failed to cleanup llogging subsystems\n", + obd->obd_name); + goto out; } int mds_lov_clean(struct obd_device *obd) { struct mds_obd *mds = &obd->u.mds; - struct llog_ctxt *ctxt; if (mds->mds_profile) { char * cln_prof; struct config_llog_instance cfg; - struct obd_run_ctxt saved; + struct lvfs_run_ctxt saved; int len = strlen(mds->mds_profile) + sizeof("-clean") + 1; OBD_ALLOC(cln_prof, len); @@ -1917,10 +2181,10 @@ int mds_lov_clean(struct obd_device *obd) cfg.cfg_instance = NULL; cfg.cfg_uuid = mds->mds_lov_uuid; - push_ctxt(&saved, &obd->obd_ctxt, NULL); - ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT); - class_config_parse_llog(ctxt, cln_prof, &cfg); - pop_ctxt(&saved, &obd->obd_ctxt, NULL); + push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); + class_config_parse_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT), + cln_prof, &cfg); + pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); OBD_FREE(cln_prof, len); OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1); @@ -1948,7 +2212,7 @@ static int mds_precleanup(struct obd_device *obd, int flags) mds_lmv_clean(obd); mds_lov_disconnect(obd, flags); mds_lov_clean(obd); - llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT)); + obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT)); RETURN(rc); } @@ -2025,12 +2289,18 @@ static void fixup_handle_for_resent_req(struct ptlrpc_request *req, } l_unlock(&obd->obd_namespace->ns_lock); + /* If the xid matches, then we know this is a resent request, + * and allow it. (It's probably an OPEN, for which we don't + * send a lock */ + if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid) + return; + /* This remote handle isn't enqueued, so we never received or * processed this request. Clear MSG_RESENT, because it can * be handled like any normal request now. */ lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT); - + DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64, remote_hdl.cookie); } @@ -2121,6 +2391,7 @@ static int mds_intent_policy(struct ldlm_namespace *ns, break; case IT_LOOKUP: getattr_part = MDS_INODELOCK_LOOKUP; + case IT_CHDIR: case IT_GETATTR: getattr_part |= MDS_INODELOCK_LOOKUP; case IT_READDIR: @@ -2128,11 +2399,11 @@ static int mds_intent_policy(struct ldlm_namespace *ns, getattr_part); /* FIXME: LDLM can set req->rq_status. MDS sets policy_res{1,2} with disposition and status. - - replay: returns 0 & req->status is old status + - replay: returns 0 & req->status is old status - otherwise: returns req->status */ if (intent_disposition(rep, DISP_LOOKUP_NEG)) rep->lock_policy_res2 = 0; - if (!intent_disposition(rep, DISP_LOOKUP_POS) || + if (!intent_disposition(rep, DISP_LOOKUP_POS) || rep->lock_policy_res2) RETURN(ELDLM_LOCK_ABORTED); if (req->rq_status != 0) { @@ -2228,24 +2499,23 @@ int mdt_detach(struct obd_device *dev) return lprocfs_obd_detach(dev); } -static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf) +static int mdt_setup(struct obd_device *obd, obd_count len, void *buf) { - struct mds_obd *mds = &obddev->u.mds; + struct mds_obd *mds = &obd->u.mds; int rc = 0; ENTRY; - mds->mds_service = + mds->mds_service = ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE, MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL, - mds_handle, "mds", - obddev->obd_proc_entry); + mds_handle, "mds", obd->obd_proc_entry); if (!mds->mds_service) { CERROR("failed to start service\n"); - RETURN(rc = -ENOMEM); + RETURN(-ENOMEM); } - rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS, + rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS, "ll_mdt"); if (rc) GOTO(err_thread, rc); @@ -2253,32 +2523,32 @@ static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf) mds->mds_setattr_service = ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE, MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL, - mds_handle, "mds_setattr", - obddev->obd_proc_entry); + mds_handle, "mds_setattr", + obd->obd_proc_entry); if (!mds->mds_setattr_service) { CERROR("failed to start getattr service\n"); GOTO(err_thread, rc = -ENOMEM); } - rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service, - MDT_NUM_THREADS, "ll_mdt_attr"); + rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service, + MDT_NUM_THREADS, "ll_mdt_attr"); if (rc) GOTO(err_thread2, rc); - + mds->mds_readpage_service = ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE, MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL, - mds_handle, "mds_readpage", - obddev->obd_proc_entry); + mds_handle, "mds_readpage", + obd->obd_proc_entry); if (!mds->mds_readpage_service) { CERROR("failed to start readpage service\n"); GOTO(err_thread2, rc = -ENOMEM); } - rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service, + rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service, MDT_NUM_THREADS, "ll_mdt_rdpg"); - if (rc) + if (rc) GOTO(err_thread3, rc); RETURN(0); @@ -2292,10 +2562,9 @@ err_thread: return rc; } - -static int mdt_cleanup(struct obd_device *obddev, int flags) +static int mdt_cleanup(struct obd_device *obd, int flags) { - struct mds_obd *mds = &obddev->u.mds; + struct mds_obd *mds = &obd->u.mds; ENTRY; ptlrpc_stop_all_threads(mds->mds_readpage_service); @@ -2310,7 +2579,8 @@ static int mdt_cleanup(struct obd_device *obddev, int flags) RETURN(0); } -static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data) +static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, + void *data) { struct obd_device *obd = data; struct ll_fid fid; @@ -2319,6 +2589,48 @@ static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *d return mds_fid2dentry(&obd->u.mds, &fid, NULL); } +static int mds_get_info(struct obd_export *exp, __u32 keylen, + void *key, __u32 *vallen, void *val) +{ + struct obd_device *obd; + struct mds_obd *mds; + ENTRY; + + obd = class_exp2obd(exp); + if (obd == NULL) { + CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n", + exp->exp_handle.h_cookie); + RETURN(-EINVAL); + } + + if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) { + /*Get log_context handle*/ + unsigned long *llh_handle = val; + *vallen = sizeof(unsigned long); + *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT]; + RETURN(0); + } + if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) { + /*Get log_context handle*/ + unsigned long *sb = val; + *vallen = sizeof(unsigned long); + *sb = (unsigned long)obd->u.mds.mds_sb; + RETURN(0); + } + + mds = &obd->u.mds; + keylen == strlen("mdsize"); + if (keylen && memcmp(key, "mdsize", keylen) == 0) { + __u32 *mdsize = val; + *vallen = sizeof(*mdsize); + *mdsize = mds->mds_max_mdsize; + RETURN(0); + } + + CDEBUG(D_IOCTL, "invalid key\n"); + RETURN(-EINVAL); + +} struct lvfs_callback_ops mds_lvfs_ops = { l_fid2dentry: mds_lvfs_fid2dentry, }; @@ -2335,36 +2647,38 @@ int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa, /* use obd ops to offer management infrastructure */ static struct obd_ops mds_obd_ops = { - o_owner: THIS_MODULE, - o_attach: mds_attach, - o_detach: mds_detach, - o_connect: mds_connect, - o_init_export: mds_init_export, - o_destroy_export: mds_destroy_export, - o_disconnect: mds_disconnect, - o_setup: mds_setup, - o_precleanup: mds_precleanup, - o_cleanup: mds_cleanup, - o_postrecov: mds_postrecov, - o_statfs: mds_obd_statfs, - o_iocontrol: mds_iocontrol, - o_create: mds_obd_create, - o_destroy: mds_obd_destroy, - o_llog_init: mds_llog_init, - o_llog_finish: mds_llog_finish, - o_notify: mds_notify, - o_get_info: mds_get_info, - o_set_info: mds_set_info, - o_preprw: mds_preprw, - o_commitrw: mds_commitrw, + .o_owner = THIS_MODULE, + .o_attach = mds_attach, + .o_detach = mds_detach, + .o_connect = mds_connect, + .o_init_export = mds_init_export, + .o_destroy_export = mds_destroy_export, + .o_disconnect = mds_disconnect, + .o_setup = mds_setup, + .o_precleanup = mds_precleanup, + .o_cleanup = mds_cleanup, + .o_postrecov = mds_postrecov, + .o_statfs = mds_obd_statfs, + .o_iocontrol = mds_iocontrol, + .o_create = mds_obd_create, + .o_destroy = mds_obd_destroy, + .o_llog_init = mds_llog_init, + .o_llog_finish = mds_llog_finish, + .o_notify = mds_notify, + .o_get_info = mds_get_info, + .o_set_info = mds_set_info, + .o_preprw = mds_preprw, + .o_commitrw = mds_commitrw, }; static struct obd_ops mdt_obd_ops = { - o_owner: THIS_MODULE, - o_attach: mdt_attach, - o_detach: mdt_detach, - o_setup: mdt_setup, - o_cleanup: mdt_cleanup, + .o_owner = THIS_MODULE, + .o_attach = mdt_attach, + .o_detach = mdt_detach, + .o_setup = mdt_setup, + .o_cleanup = mdt_cleanup, + .o_attach = mdt_attach, + .o_detach = mdt_detach, }; static int __init mds_init(void) @@ -2372,9 +2686,11 @@ static int __init mds_init(void) struct lprocfs_static_vars lvars; lprocfs_init_multi_vars(0, &lvars); - class_register_type(&mds_obd_ops, NULL, lvars.module_vars, LUSTRE_MDS_NAME); + class_register_type(&mds_obd_ops, NULL, lvars.module_vars, + LUSTRE_MDS_NAME); lprocfs_init_multi_vars(1, &lvars); - class_register_type(&mdt_obd_ops, NULL, lvars.module_vars, LUSTRE_MDT_NAME); + class_register_type(&mdt_obd_ops, NULL, lvars.module_vars, + LUSTRE_MDT_NAME); return 0; }