X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_reint.c;h=81c148fe573a2bbea9a126b3cc7893a0138b58ee;hp=686be9b2078b61336fb0bad8a2884ed12aa2ba32;hb=c1d0a355a6;hpb=9f1711f3d7d14cbe70fd058cbb59cda87d69eaa7 diff --git a/lustre/mdt/mdt_reint.c b/lustre/mdt/mdt_reint.c index 686be9b..81c148f 100644 --- a/lustre/mdt/mdt_reint.c +++ b/lustre/mdt/mdt_reint.c @@ -250,17 +250,25 @@ static int mdt_unlock_slaves(struct mdt_thread_info *mti, static inline int mdt_object_striped(struct mdt_thread_info *mti, struct mdt_object *obj) { + struct lu_device *bottom_dev; + struct lu_object *bottom_obj; int rc; if (!S_ISDIR(obj->mot_header.loh_attr)) return 0; - rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), &LU_BUF_NULL, + /* getxattr from bottom obj to avoid reading in shard FIDs */ + bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom); + bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev, + mdt_object_fid(obj), NULL); + if (IS_ERR(bottom_obj)) + return PTR_ERR(bottom_obj); + + rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL, XATTR_NAME_LMV); - if (rc <= 0) - return rc == -ENODATA ? 0 : rc; + lu_object_put(mti->mti_env, bottom_obj); - return 1; + return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc; } /** @@ -355,8 +363,8 @@ static int mdt_create(struct mdt_thread_info *info) int rc; ENTRY; - DEBUG_REQ(D_INODE, mdt_info_req(info), "Create ("DNAME"->"DFID") " - "in "DFID, + DEBUG_REQ(D_INODE, mdt_info_req(info), + "Create ("DNAME"->"DFID") in "DFID, PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1)); if (!fid_is_md_operative(rr->rr_fid1)) @@ -641,6 +649,8 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, if (info->mti_dlm_req) ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); + OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE); + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1); if (IS_ERR(mo)) @@ -681,13 +691,23 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, /* LU-10286: compatibility check for FLR. * Please check the comment in mdt_finish_open() for details */ - if (!exp_connect_flr(info->mti_exp)) { + if (!exp_connect_flr(info->mti_exp) || + !exp_connect_overstriping(info->mti_exp)) { rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV); if (rc < 0 && rc != -ENODATA) GOTO(out_put, rc); - if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm)) - GOTO(out_put, rc = -EOPNOTSUPP); + if (!exp_connect_flr(info->mti_exp)) { + if (rc > 0 && + mdt_lmm_is_flr(info->mti_big_lmm)) + GOTO(out_put, rc = -EOPNOTSUPP); + } + + if (!exp_connect_overstriping(info->mti_exp)) { + if (rc > 0 && + mdt_lmm_is_overstriping(info->mti_big_lmm)) + GOTO(out_put, rc = -EOPNOTSUPP); + } } /* For truncate, the file size sent from client @@ -718,6 +738,8 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, struct lu_buf *buf = &info->mti_buf; struct lu_ucred *uc = mdt_ucred(info); struct mdt_lock_handle *lh; + const char *name; + __u64 lockpart = MDS_INODELOCK_XATTR; /* reject if either remote or striped dir is disabled */ if (ma->ma_valid & MA_LMV) { @@ -731,27 +753,35 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, GOTO(out_put, rc = -EPERM); } + if (!S_ISDIR(lu_object_attr(&mo->mot_obj))) + GOTO(out_put, rc = -ENOTDIR); + if (ma->ma_attr.la_valid != 0) GOTO(out_put, rc = -EPROTO); - lh = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_reg_init(lh, LCK_PW); - - rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR); - if (rc != 0) - GOTO(out_put, rc); - if (ma->ma_valid & MA_LOV) { buf->lb_buf = ma->ma_lmm; buf->lb_len = ma->ma_lmm_size; + name = XATTR_NAME_LOV; } else { - buf->lb_buf = ma->ma_lmv; + struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md; + + buf->lb_buf = lmu; buf->lb_len = ma->ma_lmv_size; + name = XATTR_NAME_DEFAULT_LMV; + /* force client to update dir default layout */ + lockpart |= MDS_INODELOCK_LOOKUP; } + + lh = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_reg_init(lh, LCK_PW); + + rc = mdt_object_lock(info, mo, lh, lockpart); + if (rc != 0) + GOTO(out_put, rc); + rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf, - (ma->ma_valid & MA_LOV) ? - XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV, - 0); + name, 0); mdt_object_unlock(info, mo, lh, rc); if (rc) @@ -844,7 +874,7 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, struct mdt_lock_handle *child_lh; struct ldlm_enqueue_info *einfo = &info->mti_einfo[0]; __u64 lock_ibits; - bool cos_incompat = false, discard = false; + bool cos_incompat = false; int no_name = 0; int rc; @@ -921,7 +951,7 @@ relock: if (!cos_incompat) { rc = mdt_object_striped(info, mc); if (rc < 0) - GOTO(unlock_parent, rc); + GOTO(put_child, rc); cos_incompat = rc; if (cos_incompat) { @@ -1025,8 +1055,8 @@ relock: rc = mdt_attr_get_complex(info, mc, ma); if (rc) GOTO(out_stat, rc); - } else { - discard = mdt_dom_check_for_discard(info, mc); + } else if (mdt_dom_check_for_discard(info, mc)) { + mdt_dom_discard_data(info, mc); } mdt_handle_last_unlink(info, mc, ma); @@ -1060,13 +1090,7 @@ unlock_parent: mdt_object_unlock(info, mp, parent_lh, rc); put_parent: mdt_object_put(info->mti_env, mp); - - /* discard is just a PW DOM lock to drop the data on a client - * no need to keep objects being get and locked, do that after all. - */ - if (discard) - mdt_dom_discard_data(info, child_fid); - + CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE); return rc; } @@ -1094,6 +1118,11 @@ static int mdt_reint_link(struct mdt_thread_info *info, if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK)) RETURN(err_serious(-ENOENT)); + if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) { + req->rq_no_reply = 1; + RETURN(err_serious(-ENOENT)); + } + if (info->mti_dlm_req) ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); @@ -1639,6 +1668,9 @@ static int mdt_lock_remote_slaves(struct mdt_thread_info *info, for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) { fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]); + if (!fid_is_sane(fid)) + continue; + slave = mdt_object_find(info->mti_env, mdt, fid); if (IS_ERR(slave)) GOTO(out, rc = PTR_ERR(slave)); @@ -1915,7 +1947,7 @@ static int mdd_migrate_close(struct mdt_thread_info *info, * cancelled, it's okay to cancel it now as we've held mot_open_sem. */ ldlm_lock_cancel(lease); - ldlm_reprocess_all(lease->l_resource); + ldlm_reprocess_all(lease->l_resource, lease); LDLM_LOCK_PUT(lease); close: @@ -2139,7 +2171,11 @@ unlock_open_sem: if (open_sem_locked) up_write(&sobj->mot_open_sem); unlock_links: - mdt_unlock_list(info, &link_locks, do_sync ?: rc); + /* if we've got too many locks to save into RPC, + * then just commit before the locks are released */ + if (!rc && do_sync) + mdt_device_sync(env, mdt); + mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc); put_source: mdt_object_put(env, sobj); mdt_object_put(env, spobj); @@ -2152,9 +2188,6 @@ unlock_rename: if (lustre_handle_is_used(&rename_lh)) mdt_rename_unlock(&rename_lh); - if (!rc && do_sync) - mdt_device_sync(env, mdt); - return rc; } @@ -2273,20 +2306,10 @@ static int mdt_rename_determine_lock_order(struct mdt_thread_info *info, * 2 - srcdir child; 3 - tgtdir child. * Update on disk version of srcdir 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 *unused) { + struct mdt_device *mdt = info->mti_mdt; struct mdt_reint_record *rr = &info->mti_rr; struct md_attr *ma = &info->mti_attr; struct ptlrpc_request *req = mdt_info_req(info); @@ -2302,8 +2325,8 @@ static int mdt_reint_rename(struct mdt_thread_info *info, struct lu_fid *old_fid = &info->mti_tmp_fid1; struct lu_fid *new_fid = &info->mti_tmp_fid2; __u64 lock_ibits; - bool reverse = false; - bool cos_incompat, discard = false; + bool reverse = false, discard = false; + bool cos_incompat; int rc; ENTRY; @@ -2318,25 +2341,10 @@ static int mdt_reint_rename(struct mdt_thread_info *info, !fid_is_md_operative(rr->rr_fid2)) RETURN(-EPERM); - /* - * Note: do not enqueue rename lock for replay request, because - * if other MDT holds rename lock, but being blocked to wait for - * this MDT to finish its recovery, and the failover MDT can not - * get rename lock, which will cause deadlock. - */ - if (!req_is_replay(req)) { - rc = mdt_rename_lock(info, &rename_lh); - if (rc != 0) { - CERROR("%s: can't lock FS for rename: rc = %d\n", - mdt_obd_name(info->mti_mdt), rc); - RETURN(rc); - } - } - /* find both parents. */ msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0); if (IS_ERR(msrcdir)) - GOTO(out_unlock_rename, rc = PTR_ERR(msrcdir)); + RETURN(PTR_ERR(msrcdir)); OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5); @@ -2349,9 +2357,36 @@ static int mdt_reint_rename(struct mdt_thread_info *info, GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir)); } + /* + * Note: do not enqueue rename lock for replay request, because + * if other MDT holds rename lock, but being blocked to wait for + * this MDT to finish its recovery, and the failover MDT can not + * get rename lock, which will cause deadlock. + */ + if (!req_is_replay(req)) { + /* + * Normally rename RPC is handled on the MDT with the target + * directory (if target exists, it's on the MDT with the + * target), if the source directory is remote, it's a hint that + * source is remote too (this may not be true, but it won't + * cause any issue), return -EXDEV early to avoid taking + * rename_lock. + */ + if (!mdt->mdt_enable_remote_rename && + mdt_object_remote(msrcdir)) + GOTO(out_put_tgtdir, rc = -EXDEV); + + rc = mdt_rename_lock(info, &rename_lh); + if (rc != 0) { + CERROR("%s: can't lock FS for rename: rc = %d\n", + mdt_obd_name(mdt), rc); + GOTO(out_put_tgtdir, rc); + } + } + rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir); if (rc < 0) - GOTO(out_put_tgtdir, rc); + GOTO(out_unlock_rename, rc); reverse = rc; @@ -2376,7 +2411,7 @@ relock: rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1, cos_incompat); if (rc) - GOTO(out_put_tgtdir, rc); + GOTO(out_unlock_rename, rc); OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5); @@ -2384,13 +2419,13 @@ relock: cos_incompat); if (rc != 0) { mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc); - GOTO(out_put_tgtdir, rc); + GOTO(out_unlock_rename, rc); } } else { rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0, cos_incompat); if (rc) - GOTO(out_put_tgtdir, rc); + GOTO(out_unlock_rename, rc); OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5); @@ -2407,7 +2442,7 @@ relock: } if (rc != 0) { mdt_object_unlock(info, msrcdir, lh_srcdirp, rc); - GOTO(out_put_tgtdir, rc); + GOTO(out_unlock_rename, rc); } } @@ -2437,6 +2472,9 @@ relock: GOTO(out_put_old, rc = -ENOENT); } + if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename) + GOTO(out_put_old, rc = -EXDEV); + /* Check if @mtgtdir is subdir of @mold, before locking child * to avoid reverse locking. */ if (mtgtdir != msrcdir) { @@ -2611,7 +2649,6 @@ relock: mdt_handle_last_unlink(info, mnew, ma); discard = mdt_dom_check_for_discard(info, mnew); } - mdt_rename_counter_tally(info, info->mti_mdt, req, msrcdir, mtgtdir); } @@ -2622,28 +2659,30 @@ relock: out_unlock_old: mdt_object_unlock(info, mold, lh_oldp, rc); out_put_new: - if (mnew != NULL) + if (mnew && !discard) mdt_object_put(info->mti_env, mnew); out_put_old: mdt_object_put(info->mti_env, mold); out_unlock_parents: mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc); mdt_object_unlock(info, msrcdir, lh_srcdirp, rc); +out_unlock_rename: + if (lustre_handle_is_used(&rename_lh)) + mdt_rename_unlock(&rename_lh); out_put_tgtdir: mdt_object_put(info->mti_env, mtgtdir); out_put_srcdir: mdt_object_put(info->mti_env, msrcdir); -out_unlock_rename: - if (lustre_handle_is_used(&rename_lh)) - mdt_rename_unlock(&rename_lh); - /* If 'discard' is set then new_fid must exits. - * DOM data discard need neither object nor lock, - * so do this at the end. + /* The DoM discard can be done right in the place above where it is + * assigned, meanwhile it is done here after rename unlock due to + * compatibility with old clients, for them the discard blocks + * the main thread until completion. Check LU-11359 for details. */ - if (discard) - mdt_dom_discard_data(info, new_fid); - + if (discard) { + mdt_dom_discard_data(info, mnew); + mdt_object_put(info->mti_env, mnew); + } return rc; } @@ -2651,17 +2690,18 @@ static int mdt_reint_resync(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) { 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 *mo; - struct ldlm_lock *lease; - struct mdt_body *repbody; - struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id }; - bool lease_broken; - int rc, rc2; + struct ptlrpc_request *req = mdt_info_req(info); + struct md_attr *ma = &info->mti_attr; + struct mdt_object *mo; + struct ldlm_lock *lease; + struct mdt_body *repbody; + struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id }; + bool lease_broken; + int rc, rc2; + ENTRY; - DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1)); + DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1)); if (info->mti_dlm_req) ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP); @@ -2698,10 +2738,13 @@ static int mdt_reint_resync(struct mdt_thread_info *info, /* the file has yet opened by anyone else after we took the lease. */ layout.mlc_opc = MD_LAYOUT_RESYNC; - rc = mdt_layout_change(info, mo, &layout); + lhc = &info->mti_lh[MDT_LH_LOCAL]; + rc = mdt_layout_change(info, mo, lhc, &layout); if (rc) GOTO(out_unlock, rc); + mdt_object_unlock(info, mo, lhc, 0); + ma->ma_need = MA_INODE; ma->ma_valid = 0; rc = mdt_attr_get_complex(info, mo, ma);