X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_reint.c;h=617ac43a57f3c705280d2cc9ba64f285ff2e4a73;hp=ec2e5d82d01225f5836978cc68b79df84a6a320d;hb=f0c1b06f2418d2016e23eb6e8277be85ca13fc53;hpb=22464d1230ed58461f51d881f512d5e16644a735 diff --git a/lustre/mdt/mdt_reint.c b/lustre/mdt/mdt_reint.c index ec2e5d8..617ac43 100644 --- a/lustre/mdt/mdt_reint.c +++ b/lustre/mdt/mdt_reint.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -29,8 +27,7 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011 Whamcloud, Inc. - * + * Copyright (c) 2011, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -47,28 +44,14 @@ * Author: Yury Umanets */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_MDS #include "mdt_internal.h" -#include "lu_time.h" static inline void mdt_reint_init_ma(struct mdt_thread_info *info, struct md_attr *ma) { - ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD); - ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, - &RMF_MDT_MD, RCL_SERVER); - - ma->ma_cookie = req_capsule_server_get(info->mti_pill, - &RMF_LOGCOOKIES); - ma->ma_cookie_size = req_capsule_get_size(info->mti_pill, - &RMF_LOGCOOKIES, - RCL_SERVER); - - ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE; + ma->ma_need = MA_INODE; ma->ma_valid = 0; } @@ -83,8 +66,8 @@ static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc, if (repbody->valid & OBD_MD_FLMDSCAPA) RETURN(rc); - if (rc == 0 && info->mti_mdt->mdt_opts.mo_mds_capa && - info->mti_exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) { + if (rc == 0 && info->mti_mdt->mdt_lut.lut_mds_capa && + exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) { struct lustre_capa *capa; capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1); @@ -108,8 +91,8 @@ static void mdt_obj_version_get(struct mdt_thread_info *info, struct mdt_object *o, __u64 *version) { LASSERT(o); - LASSERT(mdt_object_exists(o) >= 0); - if (mdt_object_exists(o) > 0) + if (mdt_object_exists(o) && !mdt_object_remote(o) && + !fid_is_obf(mdt_object_fid(o))) *version = dt_version_get(info->mti_env, mdt_obj2dt(o)); else *version = ENOENT_VERSION; @@ -137,19 +120,19 @@ static int mdt_version_check(struct ptlrpc_request *req, /** Sanity check for malformed buffers */ if (pre_ver == NULL) { CERROR("No versions in request buffer\n"); - cfs_spin_lock(&req->rq_export->exp_lock); - req->rq_export->exp_vbr_failed = 1; - cfs_spin_unlock(&req->rq_export->exp_lock); - RETURN(-EOVERFLOW); - } else if (pre_ver[idx] != version) { - CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n", - pre_ver[idx], version); - cfs_spin_lock(&req->rq_export->exp_lock); - req->rq_export->exp_vbr_failed = 1; - cfs_spin_unlock(&req->rq_export->exp_lock); - RETURN(-EOVERFLOW); - } - RETURN(0); + spin_lock(&req->rq_export->exp_lock); + req->rq_export->exp_vbr_failed = 1; + spin_unlock(&req->rq_export->exp_lock); + RETURN(-EOVERFLOW); + } else if (pre_ver[idx] != version) { + CDEBUG(D_INODE, "Version mismatch "LPX64" != "LPX64"\n", + pre_ver[idx], version); + spin_lock(&req->rq_export->exp_lock); + req->rq_export->exp_vbr_failed = 1; + spin_unlock(&req->rq_export->exp_lock); + RETURN(-EOVERFLOW); + } + RETURN(0); } /** @@ -285,7 +268,10 @@ static int mdt_md_create(struct mdt_thread_info *info) DEBUG_REQ(D_INODE, mdt_info_req(info), "Create (%s->"DFID") in "DFID, rr->rr_name, PFID(rr->rr_fid2), PFID(rr->rr_fid1)); - repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1)) + RETURN(-EPERM); + + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); lh = &info->mti_lh[MDT_LH_PARENT]; mdt_lock_pdo_init(lh, LCK_PW, rr->rr_name, rr->rr_namelen); @@ -306,17 +292,52 @@ static int mdt_md_create(struct mdt_thread_info *info) lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); rc = mdt_lookup_version_check(info, parent, lname, &info->mti_tmp_fid1, 1); - /* -ENOENT is expected here */ - if (rc != 0 && rc != -ENOENT) - GOTO(out_put_parent, rc); + if (rc == 0) + GOTO(out_put_parent, rc = -EEXIST); - /* save version of file name for replay, it must be ENOENT here */ - mdt_enoent_version_save(info, 1); + /* -ENOENT is expected here */ + if (rc != -ENOENT) + GOTO(out_put_parent, rc); - child = mdt_object_find(info->mti_env, mdt, rr->rr_fid2); + /* save version of file name for replay, it must be ENOENT here */ + mdt_enoent_version_save(info, 1); + + child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2); if (likely(!IS_ERR(child))) { struct md_object *next = mdt_object_child(parent); + if (mdt_object_remote(child)) { + struct seq_server_site *ss; + struct lu_ucred *uc = mdt_ucred(info); + + if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) { + if (uc->uc_gid != + mdt->mdt_enable_remote_dir_gid && + mdt->mdt_enable_remote_dir_gid != -1) { + CERROR("%s: Creating remote dir is only" + " permitted for administrator or" + " set mdt_enable_remote_dir_gid:" + " rc = %d\n", + mdt_obd_name(mdt), -EPERM); + GOTO(out_put_child, rc = -EPERM); + } + } + + ss = mdt_seq_site(mdt); + if (ss->ss_node_id != 0 && + mdt->mdt_enable_remote_dir == 0) { + CERROR("%s: remote dir is only permitted on" + " MDT0 or set_param" + " mdt.*.enable_remote_dir=1\n", + mdt_obd_name(mdt)); + GOTO(out_put_child, rc = -EPERM); + } + if (!mdt_is_dne_client(mdt_info_req(info)->rq_export)) { + /* Return -EIO for old client */ + GOTO(out_put_child, rc = -EIO); + } + + } ma->ma_need = MA_INODE; ma->ma_valid = 0; /* capa for cross-ref will be stored here */ @@ -337,16 +358,19 @@ static int mdt_md_create(struct mdt_thread_info *info) info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode); - /* - * Do perform lookup sanity check. We do not know if name exists - * or not. - */ - info->mti_spec.sp_cr_lookup = 1; + /* + * Do not perform lookup sanity check. We know that name does + * not exist. + */ + info->mti_spec.sp_cr_lookup = 0; info->mti_spec.sp_feat = &dt_directory_features; rc = mdo_create(info->mti_env, next, lname, mdt_object_child(child), &info->mti_spec, ma); + if (rc == 0) + rc = mdt_attr_get_complex(info, child, ma); + if (rc == 0) { /* Return fid & attr to client. */ if (ma->ma_valid & MA_INODE) @@ -364,56 +388,6 @@ out_put_parent: RETURN(rc); } -/* Partial request to create object only */ -static int mdt_md_mkobj(struct mdt_thread_info *info) -{ - struct mdt_device *mdt = info->mti_mdt; - struct mdt_object *o; - struct mdt_body *repbody; - struct md_attr *ma = &info->mti_attr; - int rc; - ENTRY; - - DEBUG_REQ(D_INODE, mdt_info_req(info), "Partial create "DFID"", - PFID(info->mti_rr.rr_fid2)); - - repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); - - o = mdt_object_find(info->mti_env, mdt, info->mti_rr.rr_fid2); - if (!IS_ERR(o)) { - struct md_object *next = mdt_object_child(o); - - ma->ma_need = MA_INODE; - ma->ma_valid = 0; - - /* - * Cross-ref create can encounter already created obj in case of - * recovery, just get attr in that case. - */ - if (mdt_object_exists(o) == 1) { - rc = mo_attr_get(info->mti_env, next, ma); - } else { - /* - * Here, NO permission check for object_create, - * such check has been done on the original MDS. - */ - rc = mo_object_create(info->mti_env, next, - &info->mti_spec, ma); - } - if (rc == 0) { - /* Return fid & attr to client. */ - if (ma->ma_valid & MA_INODE) - mdt_pack_attr2body(info, repbody, &ma->ma_attr, - mdt_object_fid(o)); - } - mdt_object_put(info->mti_env, o); - } else - rc = PTR_ERR(o); - - mdt_create_pack_capa(info, rc, o, repbody); - RETURN(rc); -} - int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, struct md_attr *ma, int flags) { @@ -423,14 +397,18 @@ int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, int rc; ENTRY; - /* attr shouldn't be set on remote object */ - LASSERT(mdt_object_exists(mo) >= 0); + /* attr shouldn't be set on remote object */ + LASSERT(!mdt_object_remote(mo)); lh = &info->mti_lh[MDT_LH_PARENT]; mdt_lock_reg_init(lh, LCK_PW); - if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID)) - lockpart |= MDS_INODELOCK_LOOKUP; + /* Even though the new MDT will grant PERM lock to the old + * client, but the old client will almost ignore that during + * So it needs to revoke both LOOKUP and PERM lock here, so + * both new and old client can cancel the dcache */ + if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID)) + lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM; rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK); if (rc != 0) @@ -467,6 +445,50 @@ out_unlock: return rc; } +/** + * Check HSM flags and add HS_DIRTY flag if relevant. + * + * A file could be set dirty only if it has a copy in the backend (HS_EXISTS) + * and is not RELEASED. + */ +int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo, + struct md_attr *ma) +{ + int rc; + ENTRY; + + /* If the file was modified, add the dirty flag */ + ma->ma_need = MA_HSM; + rc = mdt_attr_get_complex(info, mo, ma); + if (rc) { + CERROR("file attribute read error for "DFID": %d.\n", + PFID(mdt_object_fid(mo)), rc); + RETURN(rc); + } + + /* If an up2date copy exists in the backend, add dirty flag */ + if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS) + && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) { + struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD]; + + ma->ma_hsm.mh_flags |= HS_DIRTY; + + mdt_lock_reg_init(lh, LCK_PW); + rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR, + MDT_LOCAL_LOCK); + if (rc != 0) + RETURN(rc); + + rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm); + if (rc) + CERROR("file attribute change error for "DFID": %d\n", + PFID(mdt_object_fid(mo)), rc); + mdt_object_unlock(info, mo, lh, rc); + } + + RETURN(rc); +} + static int mdt_reint_setattr(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) { @@ -476,9 +498,8 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, struct mdt_export_data *med = &req->rq_export->exp_mdt_data; struct mdt_file_data *mfd; struct mdt_object *mo; - struct md_object *next; struct mdt_body *repbody; - int som_au, rc; + int som_au, rc, rc2; ENTRY; DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1), @@ -487,7 +508,7 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, if (info->mti_dlm_req) ldlm_request_cancel(req, info->mti_dlm_req, 0); - repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); if (IS_ERR(mo)) GOTO(out, rc = PTR_ERR(mo)); @@ -507,7 +528,7 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, if (rc) GOTO(out_put, rc); - mfd = mdt_mfd_new(); + mfd = mdt_mfd_new(med); if (mfd == NULL) { mdt_write_put(mo); GOTO(out_put, rc = -ENOMEM); @@ -521,9 +542,9 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, mfd->mfd_object = mo; mfd->mfd_xid = req->rq_xid; - cfs_spin_lock(&med->med_open_lock); - cfs_list_add(&mfd->mfd_list, &med->med_open_head); - cfs_spin_unlock(&med->med_open_lock); + spin_lock(&med->med_open_lock); + cfs_list_add(&mfd->mfd_list, &med->med_open_head); + spin_unlock(&med->med_open_lock); repbody->handle.cookie = mfd->mfd_handle.h_cookie; } @@ -534,10 +555,11 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM); LASSERT(info->mti_ioepoch); - cfs_spin_lock(&med->med_open_lock); - mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle); - if (mfd == NULL) { - cfs_spin_unlock(&med->med_open_lock); + spin_lock(&med->med_open_lock); + mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle, + req_is_replay(req)); + if (mfd == NULL) { + spin_unlock(&med->med_open_lock); CDEBUG(D_INODE, "no handle for file close: " "fid = "DFID": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1), @@ -549,36 +571,42 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, class_handle_unhash(&mfd->mfd_handle); cfs_list_del_init(&mfd->mfd_list); - cfs_spin_unlock(&med->med_open_lock); - - /* Close the found mfd, update attributes. */ - ma->ma_lmm_size = info->mti_mdt->mdt_max_mdsize; - OBD_ALLOC_LARGE(ma->ma_lmm, info->mti_mdt->mdt_max_mdsize); - if (ma->ma_lmm == NULL) - GOTO(out_put, rc = -ENOMEM); + spin_unlock(&med->med_open_lock); mdt_mfd_close(info, mfd); - - OBD_FREE_LARGE(ma->ma_lmm, info->mti_mdt->mdt_max_mdsize); - } else { + } else if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) { + LASSERT((ma->ma_valid & MA_LOV) == 0); rc = mdt_attr_set(info, mo, ma, rr->rr_flags); if (rc) GOTO(out_put, rc); - } + } else if ((ma->ma_valid & MA_LOV) && (ma->ma_valid & MA_INODE)) { + struct lu_buf *buf = &info->mti_buf; + LASSERT(ma->ma_attr.la_valid == 0); + buf->lb_buf = ma->ma_lmm; + buf->lb_len = ma->ma_lmm_size; + rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), + buf, XATTR_NAME_LOV, 0); + if (rc) + GOTO(out_put, rc); + } else + LBUG(); + + /* If file data is modified, add the dirty flag */ + if (ma->ma_attr_flags & MDS_DATA_MODIFIED) + rc = mdt_add_dirty_flag(info, mo, ma); ma->ma_need = MA_INODE; ma->ma_valid = 0; - next = mdt_object_child(mo); - rc = mo_attr_get(info->mti_env, next, ma); - if (rc != 0) - GOTO(out_put, rc); + rc = mdt_attr_get_complex(info, mo, ma); + if (rc != 0) + GOTO(out_put, rc); - mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo)); + mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo)); - if (info->mti_mdt->mdt_opts.mo_oss_capa && - info->mti_exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA && - S_ISREG(lu_object_attr(&mo->mot_obj.mo_lu)) && - (ma->ma_attr.la_valid & LA_SIZE) && !som_au) { + if (info->mti_mdt->mdt_lut.lut_oss_capa && + exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA && + S_ISREG(lu_object_attr(&mo->mot_obj)) && + (ma->ma_attr.la_valid & LA_SIZE) && !som_au) { struct lustre_capa *capa; capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2); @@ -595,9 +623,12 @@ out_put: mdt_object_put(info->mti_env, mo); out: if (rc == 0) - mdt_counter_incr(req->rq_export, LPROC_MDT_SETATTR); + mdt_counter_incr(req, LPROC_MDT_SETATTR); - mdt_shrink_reply(info); + mdt_client_compatibility(info); + rc2 = mdt_fix_reply(info); + if (rc == 0) + rc = rc2; return rc; } @@ -614,35 +645,29 @@ static int mdt_reint_create(struct mdt_thread_info *info, if (info->mti_dlm_req) ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0); + LASSERT(info->mti_rr.rr_namelen > 0); switch (info->mti_attr.ma_attr.la_mode & S_IFMT) { - case S_IFDIR:{ - /* Cross-ref case. */ - /* TODO: we can add LPROC_MDT_CROSS for cross-ref stats */ - if (info->mti_cross_ref) { - rc = mdt_md_mkobj(info); - } else { - LASSERT(info->mti_rr.rr_namelen > 0); - mdt_counter_incr(req->rq_export, LPROC_MDT_MKDIR); - rc = mdt_md_create(info); - } - break; - } + case S_IFDIR: + mdt_counter_incr(req, LPROC_MDT_MKDIR); + break; case S_IFREG: case S_IFLNK: case S_IFCHR: case S_IFBLK: case S_IFIFO: - case S_IFSOCK:{ - /* Special file should stay on the same node as parent. */ - LASSERT(info->mti_rr.rr_namelen > 0); - mdt_counter_incr(req->rq_export, LPROC_MDT_MKNOD); - rc = mdt_md_create(info); - break; - } - default: - rc = err_serious(-EOPNOTSUPP); - } - RETURN(rc); + case S_IFSOCK: + /* Special file should stay on the same node as parent. */ + mdt_counter_incr(req, LPROC_MDT_MKNOD); + break; + default: + CERROR("%s: Unsupported mode %o\n", + mdt_obd_name(info->mti_mdt), + info->mti_attr.ma_attr.la_mode); + RETURN(err_serious(-EOPNOTSUPP)); + } + + rc = mdt_md_create(info); + RETURN(rc); } /* @@ -662,7 +687,8 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, struct mdt_lock_handle *child_lh; struct lu_name *lname; int rc; - ENTRY; + int no_name = 0; + ENTRY; DEBUG_REQ(D_INODE, req, "unlink "DFID"/%s", PFID(rr->rr_fid1), rr->rr_name); @@ -673,76 +699,143 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK)) RETURN(err_serious(-ENOENT)); + if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1)) + RETURN(-EPERM); /* - * step 1: lock the parent. Note, this may be child in case of - * remote operation denoted by ->mti_cross_ref flag. + * step 1: Found the parent. */ - parent_lh = &info->mti_lh[MDT_LH_PARENT]; - if (info->mti_cross_ref) { - /* - * Init reg lock for cross ref case when we need to do only - * ref del locally. - */ - mdt_lock_reg_init(parent_lh, LCK_PW); - } else { - mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name, - rr->rr_namelen); - } - mp = mdt_object_find_lock(info, rr->rr_fid1, parent_lh, - MDS_INODELOCK_UPDATE); - if (IS_ERR(mp)) { - rc = PTR_ERR(mp); - /* errors are possible here in cross-ref cases, see below */ - if (info->mti_cross_ref) - rc = 0; - GOTO(out, rc); - } - - rc = mdt_version_get_check_save(info, mp, 0); - if (rc) - GOTO(out_unlock_parent, rc); + mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1); + if (IS_ERR(mp)) { + rc = PTR_ERR(mp); + GOTO(out, rc); + } + + parent_lh = &info->mti_lh[MDT_LH_PARENT]; + lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); + if (mdt_object_remote(mp)) { + mdt_lock_reg_init(parent_lh, LCK_EX); + rc = mdt_remote_object_lock(info, mp, &parent_lh->mlh_rreg_lh, + parent_lh->mlh_rreg_mode, + MDS_INODELOCK_UPDATE); + if (rc != ELDLM_OK) + GOTO(put_parent, rc); + + } else { + mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name, + rr->rr_namelen); + rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE, + MDT_LOCAL_LOCK); + if (rc) + GOTO(put_parent, rc); + + rc = mdt_version_get_check_save(info, mp, 0); + if (rc) + GOTO(unlock_parent, rc); + } + + /* step 2: find & lock the child */ + /* lookup child object along with version checking */ + fid_zero(child_fid); + rc = mdt_lookup_version_check(info, mp, lname, child_fid, 1); + if (rc != 0) { + /* Name might not be able to find during resend of + * remote unlink, considering following case. + * dir_A is a remote directory, the name entry of + * dir_A is on MDT0, the directory is on MDT1, + * + * 1. client sends unlink req to MDT1. + * 2. MDT1 sends name delete update to MDT0. + * 3. name entry is being deleted in MDT0 synchronously. + * 4. MDT1 is restarted. + * 5. client resends unlink req to MDT1. So it can not + * find the name entry on MDT0 anymore. + * In this case, MDT1 only needs to destory the local + * directory. + * */ + if (mdt_object_remote(mp) && rc == -ENOENT && + !fid_is_zero(rr->rr_fid2) && + lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) { + no_name = 1; + *child_fid = *rr->rr_fid2; + } else { + GOTO(unlock_parent, rc); + } + } + + if (fid_is_obf(child_fid) || fid_is_dot_lustre(child_fid)) + GOTO(unlock_parent, rc = -EPERM); mdt_reint_init_ma(info, ma); - if (!ma->ma_lmm || !ma->ma_cookie) - GOTO(out_unlock_parent, rc = -EINVAL); - - if (info->mti_cross_ref) { - /* - * Remote partial operation. It is possible that replay may - * happen on parent MDT and this operation will be repeated. - * Therefore the object absense is allowed case and nothing - * should be done here. - */ - if (mdt_object_exists(mp) > 0) { - mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA); - rc = mo_ref_del(info->mti_env, - mdt_object_child(mp), ma); - if (rc == 0) - mdt_handle_last_unlink(info, mp, ma); - } else - rc = 0; - GOTO(out_unlock_parent, rc); - } - /* step 2: find & lock the child */ - lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); - /* lookup child object along with version checking */ - rc = mdt_lookup_version_check(info, mp, lname, child_fid, 1); - if (rc != 0) - GOTO(out_unlock_parent, rc); + /* We will lock the child regardless it is local or remote. No harm. */ + mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid); + if (IS_ERR(mc)) + GOTO(unlock_parent, rc = PTR_ERR(mc)); - /* We will lock the child regardless it is local or remote. No harm. */ - mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid); - if (IS_ERR(mc)) - GOTO(out_unlock_parent, rc = PTR_ERR(mc)); child_lh = &info->mti_lh[MDT_LH_CHILD]; mdt_lock_reg_init(child_lh, LCK_EX); - rc = mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_FULL, - MDT_CROSS_LOCK); - if (rc != 0) { - mdt_object_put(info->mti_env, mc); - GOTO(out_unlock_parent, rc); - } + if (mdt_object_remote(mc)) { + struct mdt_body *repbody; + + if (!fid_is_zero(rr->rr_fid2)) { + CDEBUG(D_INFO, "%s: name %s can not find "DFID"\n", + mdt_obd_name(info->mti_mdt), + (char *)rr->rr_name, PFID(mdt_object_fid(mc))); + GOTO(put_child, rc = -ENOENT); + } + CDEBUG(D_INFO, "%s: name %s: "DFID" is another MDT\n", + mdt_obd_name(info->mti_mdt), + (char *)rr->rr_name, PFID(mdt_object_fid(mc))); + + if (!mdt_is_dne_client(req->rq_export)) + /* Return -EIO for old client */ + GOTO(put_child, rc = -EIO); + + if (info->mti_spec.sp_rm_entry) { + struct lu_ucred *uc = mdt_ucred(info); + + if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) { + CERROR("%s: unlink remote entry is only " + "permitted for administrator: rc = %d\n", + mdt_obd_name(info->mti_mdt), + -EPERM); + GOTO(put_child, rc = -EPERM); + } + + ma->ma_need = MA_INODE; + ma->ma_valid = 0; + mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA); + rc = mdo_unlink(info->mti_env, mdt_object_child(mp), + NULL, lname, ma, no_name); + GOTO(put_child, rc); + } + /* Revoke the LOOKUP lock of the remote object granted by + * this MDT. Since the unlink will happen on another MDT, + * it will release the LOOKUP lock right away. Then What + * would happen if another client try to grab the LOOKUP + * lock at the same time with unlink XXX */ + mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP, + MDT_CROSS_LOCK); + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + LASSERT(repbody != NULL); + repbody->fid1 = *mdt_object_fid(mc); + repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); + GOTO(unlock_child, rc = -EREMOTE); + } else if (info->mti_spec.sp_rm_entry) { + rc = -EPERM; + CDEBUG(D_INFO, "%s: no rm_entry on local dir '%s': rc = %d\n", + mdt_obd_name(info->mti_mdt), (char *)rr->rr_name, rc); + GOTO(put_child, rc); + } + + /* We used to acquire MDS_INODELOCK_FULL here but we can't do + * this now because a running HSM restore on the child (unlink + * victim) will hold the layout lock. See LU-4002. */ + rc = mdt_object_lock(info, mc, child_lh, + MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE, + MDT_CROSS_LOCK); + if (rc != 0) + GOTO(put_child, rc); mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, OBD_FAIL_MDS_REINT_UNLINK_WRITE); @@ -755,15 +848,18 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, ma->ma_need = MA_INODE; ma->ma_valid = 0; mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA); - rc = mdo_unlink(info->mti_env, mdt_object_child(mp), - mdt_object_child(mc), lname, ma); - if (rc == 0) - mdt_handle_last_unlink(info, mc, ma); + + rc = mdo_unlink(info->mti_env, mdt_object_child(mp), + mdt_object_child(mc), lname, ma, no_name); + if (rc == 0 && !lu_object_is_dying(&mc->mot_header)) + rc = mdt_attr_get_complex(info, mc, ma); + if (rc == 0) + mdt_handle_last_unlink(info, mc, ma); if (ma->ma_valid & MA_INODE) { switch (ma->ma_attr.la_mode & S_IFMT) { case S_IFDIR: - mdt_counter_incr(req->rq_export, LPROC_MDT_RMDIR); + mdt_counter_incr(req, LPROC_MDT_RMDIR); break; case S_IFREG: case S_IFLNK: @@ -771,7 +867,7 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, case S_IFBLK: case S_IFIFO: case S_IFSOCK: - mdt_counter_incr(req->rq_export, LPROC_MDT_UNLINK); + mdt_counter_incr(req, LPROC_MDT_UNLINK); break; default: LASSERTF(0, "bad file type %o unlinking\n", @@ -780,10 +876,14 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, } EXIT; - - mdt_object_unlock_put(info, mc, child_lh, rc); -out_unlock_parent: - mdt_object_unlock_put(info, mp, parent_lh, rc); +unlock_child: + mdt_object_unlock(info, mc, child_lh, rc); +put_child: + mdt_object_put(info->mti_env, mc); +unlock_parent: + mdt_object_unlock(info, mp, parent_lh, rc); +put_parent: + mdt_object_put(info->mti_env, mp); out: return rc; } @@ -815,26 +915,15 @@ static int mdt_reint_link(struct mdt_thread_info *info, if (info->mti_dlm_req) ldlm_request_cancel(req, info->mti_dlm_req, 0); - if (info->mti_cross_ref) { - /* MDT holding name ask us to add ref. */ - lhs = &info->mti_lh[MDT_LH_CHILD]; - mdt_lock_reg_init(lhs, LCK_EX); - ms = mdt_object_find_lock(info, rr->rr_fid1, lhs, - MDS_INODELOCK_UPDATE); - if (IS_ERR(ms)) - RETURN(PTR_ERR(ms)); - - mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA); - rc = mo_ref_add(info->mti_env, mdt_object_child(ms), ma); - mdt_object_unlock_put(info, ms, lhs, rc); - RETURN(rc); - } - /* Invalid case so return error immediately instead of * processing it */ if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) RETURN(-EPERM); + if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) || + fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2)) + RETURN(-EPERM); + /* step 1: find & lock the target parent dir */ lhp = &info->mti_lh[MDT_LH_PARENT]; mdt_lock_pdo_init(lhp, LCK_PW, rr->rr_name, @@ -856,8 +945,16 @@ static int mdt_reint_link(struct mdt_thread_info *info, if (IS_ERR(ms)) GOTO(out_unlock_parent, rc = PTR_ERR(ms)); - rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE, - MDT_CROSS_LOCK); + if (mdt_object_remote(ms)) { + mdt_object_put(info->mti_env, ms); + CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n", + mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1), + PFID(rr->rr_fid2)); + GOTO(out_unlock_parent, rc = -EXDEV); + } + + rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE | + MDS_INODELOCK_XATTR, MDT_CROSS_LOCK); if (rc != 0) { mdt_object_put(info->mti_env, ms); GOTO(out_unlock_parent, rc); @@ -879,6 +976,11 @@ static int mdt_reint_link(struct mdt_thread_info *info, GOTO(out_unlock_child, rc); /* save version of file name for replay, it must be ENOENT here */ if (!req_is_replay(mdt_info_req(info))) { + if (rc != -ENOENT) { + CDEBUG(D_INFO, "link target %.*s existed!\n", + rr->rr_namelen, (char *)rr->rr_name); + GOTO(out_unlock_child, rc = -EEXIST); + } info->mti_ver[2] = ENOENT_VERSION; mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2); } @@ -887,7 +989,7 @@ static int mdt_reint_link(struct mdt_thread_info *info, mdt_object_child(ms), lname, ma); if (rc == 0) - mdt_counter_incr(req->rq_export, LPROC_MDT_LINK); + mdt_counter_incr(req, LPROC_MDT_LINK); EXIT; out_unlock_child: @@ -901,149 +1003,65 @@ out_unlock_parent: * (lh->mlh_pdo_hash) in parallel directory lock. */ static int mdt_pdir_hash_lock(struct mdt_thread_info *info, - struct mdt_lock_handle *lh, - struct mdt_object *obj, __u64 ibits) + struct mdt_lock_handle *lh, + struct mdt_object *obj, __u64 ibits) { - struct ldlm_res_id *res_id = &info->mti_res_id; - struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace; - ldlm_policy_data_t *policy = &info->mti_policy; - int rc; - - /* - * Finish res_id initializing by name hash marking part of - * directory which is taking modification. - */ - LASSERT(lh->mlh_pdo_hash != 0); - fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res_id); - memset(policy, 0, sizeof(*policy)); - policy->l_inodebits.bits = ibits; - /* - * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is - * going to be sent to client. If it is - mdt_intent_policy() path will - * fix it up and turn FL_LOCAL flag off. - */ - rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy, - res_id, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB, - &info->mti_exp->exp_handle.h_cookie); - return rc; -} - -/* partial operation for rename */ -static int mdt_reint_rename_tgt(struct mdt_thread_info *info) -{ - struct mdt_reint_record *rr = &info->mti_rr; - struct ptlrpc_request *req = mdt_info_req(info); - struct md_attr *ma = &info->mti_attr; - struct mdt_object *mtgtdir; - struct mdt_object *mtgt = NULL; - struct mdt_lock_handle *lh_tgtdir; - struct mdt_lock_handle *lh_tgt = NULL; - struct lu_fid *tgt_fid = &info->mti_tmp_fid1; - struct lu_name *lname; - int rc; - ENTRY; - - DEBUG_REQ(D_INODE, req, "rename_tgt: insert (%s->"DFID") in "DFID, - rr->rr_tgt, PFID(rr->rr_fid2), PFID(rr->rr_fid1)); - - /* step 1: lookup & lock the tgt dir. */ - lh_tgtdir = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_pdo_init(lh_tgtdir, LCK_PW, rr->rr_tgt, - rr->rr_tgtlen); - mtgtdir = mdt_object_find_lock(info, rr->rr_fid1, lh_tgtdir, - MDS_INODELOCK_UPDATE); - if (IS_ERR(mtgtdir)) - RETURN(PTR_ERR(mtgtdir)); - - /* step 2: find & lock the target object if exists. */ - mdt_set_capainfo(info, 0, rr->rr_fid1, BYPASS_CAPA); - lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen); - rc = mdo_lookup(info->mti_env, mdt_object_child(mtgtdir), - lname, tgt_fid, &info->mti_spec); - if (rc != 0 && rc != -ENOENT) { - GOTO(out_unlock_tgtdir, rc); - } else if (rc == 0) { - /* - * In case of replay that name can be already inserted, check - * that and do nothing if so. - */ - if (lu_fid_eq(tgt_fid, rr->rr_fid2)) - GOTO(out_unlock_tgtdir, rc); - - lh_tgt = &info->mti_lh[MDT_LH_CHILD]; - mdt_lock_reg_init(lh_tgt, LCK_EX); - - mtgt = mdt_object_find_lock(info, tgt_fid, lh_tgt, - MDS_INODELOCK_LOOKUP); - if (IS_ERR(mtgt)) - GOTO(out_unlock_tgtdir, rc = PTR_ERR(mtgt)); - - mdt_reint_init_ma(info, ma); - if (!ma->ma_lmm || !ma->ma_cookie) - GOTO(out_unlock_tgt, rc = -EINVAL); - - rc = mdo_rename_tgt(info->mti_env, mdt_object_child(mtgtdir), - mdt_object_child(mtgt), rr->rr_fid2, - lname, ma); - } else /* -ENOENT */ { - rc = mdo_name_insert(info->mti_env, mdt_object_child(mtgtdir), - lname, rr->rr_fid2, ma); - } - - /* handle last link of tgt object */ - if (rc == 0 && mtgt) - mdt_handle_last_unlink(info, mtgt, ma); - - EXIT; -out_unlock_tgt: - if (mtgt) - mdt_object_unlock_put(info, mtgt, lh_tgt, rc); -out_unlock_tgtdir: - mdt_object_unlock_put(info, mtgtdir, lh_tgtdir, rc); - return rc; + struct ldlm_res_id *res = &info->mti_res_id; + struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace; + ldlm_policy_data_t *policy = &info->mti_policy; + int rc; + + /* + * Finish res_id initializing by name hash marking part of + * directory which is taking modification. + */ + LASSERT(lh->mlh_pdo_hash != 0); + fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res); + memset(policy, 0, sizeof(*policy)); + policy->l_inodebits.bits = ibits; + /* + * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is + * going to be sent to client. If it is - mdt_intent_policy() path will + * fix it up and turn FL_LOCAL flag off. + */ + rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy, + res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB, + &info->mti_exp->exp_handle.h_cookie); + return rc; } static int mdt_rename_lock(struct mdt_thread_info *info, struct lustre_handle *lh) { - struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace; - ldlm_policy_data_t *policy = &info->mti_policy; - struct ldlm_res_id *res_id = &info->mti_res_id; - struct md_site *ms; - int rc; - ENTRY; - - ms = mdt_md_site(info->mti_mdt); - fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id); - - memset(policy, 0, sizeof *policy); - policy->l_inodebits.bits = MDS_INODELOCK_UPDATE; - - if (ms->ms_control_exp == NULL) { - int flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB; - - /* - * Current node is controller, that is mdt0, where we should - * take BFL lock. - */ - rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy, - LCK_EX, &flags, ldlm_blocking_ast, - ldlm_completion_ast, NULL, NULL, 0, - &info->mti_exp->exp_handle.h_cookie, - lh); - } else { - struct ldlm_enqueue_info einfo = { LDLM_IBITS, LCK_EX, - ldlm_blocking_ast, ldlm_completion_ast, NULL, NULL, NULL }; - int flags = 0; - - /* - * This is the case mdt0 is remote node, issue DLM lock like - * other clients. - */ - rc = ldlm_cli_enqueue(ms->ms_control_exp, NULL, &einfo, res_id, - policy, &flags, NULL, 0, lh, 0); - } - + struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace; + ldlm_policy_data_t *policy = &info->mti_policy; + struct ldlm_res_id *res_id = &info->mti_res_id; + __u64 flags = 0; + int rc; + ENTRY; + + fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id); + + memset(policy, 0, sizeof *policy); + policy->l_inodebits.bits = MDS_INODELOCK_UPDATE; +#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 5, 53, 0) + /* In phase I, we will not do cross-rename, so local BFL lock would + * be enough + */ + flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB; + /* + * Current node is controller, that is mdt0, where we should + * take BFL lock. + */ + rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy, + LCK_EX, &flags, ldlm_blocking_ast, + ldlm_completion_ast, NULL, NULL, 0, + LVB_T_NONE, + &info->mti_exp->exp_handle.h_cookie, + lh); +#else +#warning "Local rename lock is invalid for DNE phase II." +#endif RETURN(rc); } @@ -1096,6 +1114,17 @@ static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid) * 2 - src child; 3 - tgt child. * Update on disk version of src child. */ +/** + * For DNE phase I, only these renames are allowed + * mv src_p/src_c tgt_p/tgt_c + * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT. + * 2. src_p and tgt_p are same directory, and tgt_c does not + * exists. In this case, all of modification will happen + * in the MDT where ithesource parent is, only one remote + * update is needed, i.e. set c_time/m_time on the child. + * And tgt_c will be still in the same MDT as the original + * src_c. + */ static int mdt_reint_rename(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) { @@ -1121,20 +1150,19 @@ static int mdt_reint_rename(struct mdt_thread_info *info, if (info->mti_dlm_req) ldlm_request_cancel(req, info->mti_dlm_req, 0); - if (info->mti_cross_ref) { - rc = mdt_reint_rename_tgt(info); - RETURN(rc); - } - DEBUG_REQ(D_INODE, req, "rename "DFID"/%s to "DFID"/%s", PFID(rr->rr_fid1), rr->rr_name, PFID(rr->rr_fid2), rr->rr_tgt); - rc = mdt_rename_lock(info, &rename_lh); - if (rc) { - CERROR("Can't lock FS for rename, rc %d\n", rc); - RETURN(rc); - } + if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) || + fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2)) + RETURN(-EPERM); + + rc = mdt_rename_lock(info, &rename_lh); + if (rc) { + CERROR("Can't lock FS for rename, rc %d\n", rc); + RETURN(rc); + } lh_newp = &info->mti_lh[MDT_LH_NEW]; @@ -1176,24 +1204,31 @@ static int mdt_reint_rename(struct mdt_thread_info *info, if (rc) GOTO(out_put_target, rc); - rc = mdt_object_exists(mtgtdir); - if (rc == 0) { - GOTO(out_put_target, rc = -ESTALE); - } else if (rc > 0) { - /* we lock the target dir if it is local */ - rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp, - MDS_INODELOCK_UPDATE, - MDT_LOCAL_LOCK); - if (rc != 0) - GOTO(out_put_target, rc); - /* get and save correct version after locking */ - mdt_version_get_save(info, mtgtdir, 1); - } - } + if (unlikely(mdt_object_remote(mtgtdir))) { + CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID + "on different MDTs\n", PFID(rr->rr_fid1), + PFID(rr->rr_fid2)); + GOTO(out_put_target, rc = -EXDEV); + } else { + if (likely(mdt_object_exists(mtgtdir))) { + /* we lock the target dir if it is local */ + rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp, + MDS_INODELOCK_UPDATE, + MDT_LOCAL_LOCK); + if (rc != 0) + GOTO(out_put_target, rc); + /* get and save correct version after locking */ + mdt_version_get_save(info, mtgtdir, 1); + } else { + GOTO(out_put_target, rc = -ESTALE); + } + } + } /* step 3: find & lock the old object. */ lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); mdt_name_copy(&slname, lname); + fid_zero(old_fid); rc = mdt_lookup_version_check(info, msrcdir, &slname, old_fid, 2); if (rc != 0) GOTO(out_unlock_target, rc); @@ -1201,14 +1236,17 @@ static int mdt_reint_rename(struct mdt_thread_info *info, if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2)) GOTO(out_unlock_target, rc = -EINVAL); - mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid); - if (IS_ERR(mold)) - GOTO(out_unlock_target, rc = PTR_ERR(mold)); + if (fid_is_obf(old_fid) || fid_is_dot_lustre(old_fid)) + GOTO(out_unlock_target, rc = -EPERM); + + mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid); + if (IS_ERR(mold)) + GOTO(out_unlock_target, rc = PTR_ERR(mold)); lh_oldp = &info->mti_lh[MDT_LH_OLD]; mdt_lock_reg_init(lh_oldp, LCK_EX); - rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP, - MDT_CROSS_LOCK); + rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP | + MDS_INODELOCK_XATTR, MDT_CROSS_LOCK); if (rc != 0) { mdt_object_put(info->mti_env, mold); GOTO(out_unlock_target, rc); @@ -1223,6 +1261,7 @@ static int mdt_reint_rename(struct mdt_thread_info *info, /* new target object may not exist now */ lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen); /* lookup with version checking */ + fid_zero(new_fid); rc = mdt_lookup_version_check(info, mtgtdir, lname, new_fid, 3); if (rc == 0) { /* the new_fid should have been filled at this moment */ @@ -1233,13 +1272,35 @@ static int mdt_reint_rename(struct mdt_thread_info *info, lu_fid_eq(new_fid, rr->rr_fid2)) GOTO(out_unlock_old, rc = -EINVAL); + if (fid_is_obf(new_fid) || fid_is_dot_lustre(new_fid)) + GOTO(out_unlock_old, rc = -EPERM); + + if (mdt_object_remote(mold)) { + CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n", + PFID(old_fid)); + GOTO(out_unlock_old, rc = -EXDEV); + } + mdt_lock_reg_init(lh_newp, LCK_EX); mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid); if (IS_ERR(mnew)) GOTO(out_unlock_old, rc = PTR_ERR(mnew)); - rc = mdt_object_lock(info, mnew, lh_newp, - MDS_INODELOCK_FULL, MDT_CROSS_LOCK); + if (mdt_object_remote(mnew)) { + mdt_object_put(info->mti_env, mnew); + CDEBUG(D_INFO, "src child "DFID" is on another MDT\n", + PFID(new_fid)); + GOTO(out_unlock_old, rc = -EXDEV); + } + + /* We used to acquire MDS_INODELOCK_FULL here but we + * can't do this now because a running HSM restore on + * the rename onto victim will hold the layout + * lock. See LU-4002. */ + rc = mdt_object_lock(info, mnew, lh_newp, + MDS_INODELOCK_LOOKUP | + MDS_INODELOCK_UPDATE, + MDT_CROSS_LOCK); if (rc != 0) { mdt_object_put(info->mti_env, mnew); GOTO(out_unlock_old, rc); @@ -1250,13 +1311,18 @@ static int mdt_reint_rename(struct mdt_thread_info *info, } else if (rc != -EREMOTE && rc != -ENOENT) { GOTO(out_unlock_old, rc); } else { - mdt_enoent_version_save(info, 3); + /* If mnew does not exist and mold are remote directory, + * it only allows rename if they are under same directory */ + if (mtgtdir != msrcdir && mdt_object_remote(mold)) { + CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n", + PFID(old_fid)); + GOTO(out_unlock_old, rc = -EXDEV); + } + mdt_enoent_version_save(info, 3); } /* step 5: rename it */ mdt_reint_init_ma(info, ma); - if (!ma->ma_lmm || !ma->ma_cookie) - GOTO(out_unlock_new, rc = -EINVAL); mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, OBD_FAIL_MDS_REINT_RENAME_WRITE); @@ -1274,9 +1340,12 @@ static int mdt_reint_rename(struct mdt_thread_info *info, /* handle last link of tgt object */ if (rc == 0) { - mdt_counter_incr(req->rq_export, LPROC_MDT_RENAME); + mdt_counter_incr(req, LPROC_MDT_RENAME); if (mnew) mdt_handle_last_unlink(info, mnew, ma); + + mdt_rename_counter_tally(info, info->mti_mdt, req, + msrcdir, mtgtdir); } EXIT; @@ -1292,21 +1361,23 @@ out_put_target: out_unlock_source: mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc); out_rename_lock: - mdt_rename_unlock(&rename_lh); - return rc; + if (lustre_handle_is_used(&rename_lh)) + mdt_rename_unlock(&rename_lh); + return rc; } typedef int (*mdt_reinter)(struct mdt_thread_info *info, struct mdt_lock_handle *lhc); static mdt_reinter reinters[REINT_MAX] = { - [REINT_SETATTR] = mdt_reint_setattr, - [REINT_CREATE] = mdt_reint_create, - [REINT_LINK] = mdt_reint_link, - [REINT_UNLINK] = mdt_reint_unlink, - [REINT_RENAME] = mdt_reint_rename, - [REINT_OPEN] = mdt_reint_open, - [REINT_SETXATTR] = mdt_reint_setxattr + [REINT_SETATTR] = mdt_reint_setattr, + [REINT_CREATE] = mdt_reint_create, + [REINT_LINK] = mdt_reint_link, + [REINT_UNLINK] = mdt_reint_unlink, + [REINT_RENAME] = mdt_reint_rename, + [REINT_OPEN] = mdt_reint_open, + [REINT_SETXATTR] = mdt_reint_setxattr, + [REINT_RMENTRY] = mdt_reint_unlink }; int mdt_reint_rec(struct mdt_thread_info *info,