X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_reint.c;h=60730a88455cd0fff7dea30fe2d4814322d41479;hb=3514e6f03bc08bf7c3e7746a54c8900a65fc0d74;hp=222c9ba3e854bb6a6a454b8b1a31ceba26dd1a86;hpb=6098ea781eca0a82ae3c707c265079ebcfb31392;p=fs%2Flustre-release.git diff --git a/lustre/mdt/mdt_reint.c b/lustre/mdt/mdt_reint.c index 222c9ba..60730a8 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. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2012, Whamcloud, Inc. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -44,9 +44,6 @@ * Author: Yury Umanets */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_MDS #include "mdt_internal.h" @@ -80,7 +77,8 @@ 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) { + if (rc == 0 && info->mti_mdt->mdt_opts.mo_mds_capa && + info->mti_exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) { struct lustre_capa *capa; capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1); @@ -95,6 +93,176 @@ static int mdt_create_pack_capa(struct mdt_thread_info *info, int rc, RETURN(rc); } +/** + * Get version of object by fid. + * + * Return real version or ENOENT_VERSION if object doesn't exist + */ +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) + *version = dt_version_get(info->mti_env, mdt_obj2dt(o)); + else + *version = ENOENT_VERSION; + CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n", + PFID(mdt_object_fid(o)), *version); +} + +/** + * Check version is correct. + * + * Should be called only during replay. + */ +static int mdt_version_check(struct ptlrpc_request *req, + __u64 version, int idx) +{ + __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg); + ENTRY; + + if (!exp_connect_vbr(req->rq_export)) + RETURN(0); + + LASSERT(req_is_replay(req)); + /** VBR: version is checked always because costs nothing */ + LASSERT(idx < PTLRPC_NUM_VERSIONS); + /** 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); +} + +/** + * Save pre-versions in reply. + */ +static void mdt_version_save(struct ptlrpc_request *req, __u64 version, + int idx) +{ + __u64 *reply_ver; + + if (!exp_connect_vbr(req->rq_export)) + return; + + LASSERT(!req_is_replay(req)); + LASSERT(req->rq_repmsg != NULL); + reply_ver = lustre_msg_get_versions(req->rq_repmsg); + if (reply_ver) + reply_ver[idx] = version; +} + +/** + * Save enoent version, it is needed when it is obvious that object doesn't + * exist, e.g. child during create. + */ +static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx) +{ + /* save version of file name for replay, it must be ENOENT here */ + if (!req_is_replay(mdt_info_req(info))) { + info->mti_ver[idx] = ENOENT_VERSION; + mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx); + } +} + +/** + * Get version from disk and save in reply buffer. + * + * Versions are saved in reply only during normal operations not replays. + */ +void mdt_version_get_save(struct mdt_thread_info *info, + struct mdt_object *mto, int idx) +{ + /* don't save versions during replay */ + if (!req_is_replay(mdt_info_req(info))) { + mdt_obj_version_get(info, mto, &info->mti_ver[idx]); + mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx); + } +} + +/** + * Get version from disk and check it, no save in reply. + */ +int mdt_version_get_check(struct mdt_thread_info *info, + struct mdt_object *mto, int idx) +{ + /* only check versions during replay */ + if (!req_is_replay(mdt_info_req(info))) + return 0; + + mdt_obj_version_get(info, mto, &info->mti_ver[idx]); + return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx); +} + +/** + * Get version from disk and check if recovery or just save. + */ +int mdt_version_get_check_save(struct mdt_thread_info *info, + struct mdt_object *mto, int idx) +{ + int rc = 0; + + mdt_obj_version_get(info, mto, &info->mti_ver[idx]); + if (req_is_replay(mdt_info_req(info))) + rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], + idx); + else + mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx); + return rc; +} + +/** + * Lookup with version checking. + * + * This checks version of 'name'. Many reint functions uses 'name' for child not + * FID, therefore we need to get object by name and check its version. + */ +int mdt_lookup_version_check(struct mdt_thread_info *info, + struct mdt_object *p, struct lu_name *lname, + struct lu_fid *fid, int idx) +{ + int rc, vbrc; + + rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid, + &info->mti_spec); + /* Check version only during replay */ + if (!req_is_replay(mdt_info_req(info))) + return rc; + + info->mti_ver[idx] = ENOENT_VERSION; + if (rc == 0) { + struct mdt_object *child; + child = mdt_object_find(info->mti_env, info->mti_mdt, fid); + if (likely(!IS_ERR(child))) { + mdt_obj_version_get(info, child, &info->mti_ver[idx]); + mdt_object_put(info->mti_env, child); + } + } + vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx); + return vbrc ? vbrc : rc; + +} + +/* + * VBR: we save three versions in reply: + * 0 - parent. Check that parent version is the same during replay. + * 1 - name. Version of 'name' if file exists with the same name or + * ENOENT_VERSION, it is needed because file may appear due to missed replays. + * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity + * check. + */ static int mdt_md_create(struct mdt_thread_info *info) { struct mdt_device *mdt = info->mti_mdt; @@ -121,7 +289,31 @@ static int mdt_md_create(struct mdt_thread_info *info) if (IS_ERR(parent)) RETURN(PTR_ERR(parent)); - child = mdt_object_find(info->mti_env, mdt, rr->rr_fid2); + if (mdt_object_obf(parent)) + GOTO(out_put_parent, rc = -EPERM); + + rc = mdt_version_get_check_save(info, parent, 0); + if (rc) + GOTO(out_put_parent, rc); + + /* + * Check child name version during replay. + * During create replay a file may exist with same name. + */ + 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); + if (rc == 0) + GOTO(out_put_parent, rc = -EEXIST); + + /* -ENOENT is expected here */ + if (rc != -ENOENT) + GOTO(out_put_parent, rc); + + /* 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); @@ -135,18 +327,23 @@ static int mdt_md_create(struct mdt_thread_info *info) mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, OBD_FAIL_MDS_REINT_CREATE_WRITE); + /* Version of child will be updated on disk. */ + info->mti_mos = child; + rc = mdt_version_get_check_save(info, child, 2); + if (rc) + GOTO(out_put_child, rc); + /* Let lower layer know current lock mode. */ 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; - lname = mdt_name(info->mti_env, (char *)rr->rr_name, - rr->rr_namelen); rc = mdo_create(info->mti_env, next, lname, mdt_object_child(child), &info->mti_spec, ma); @@ -156,11 +353,13 @@ static int mdt_md_create(struct mdt_thread_info *info) mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(child)); } +out_put_child: mdt_object_put(info->mti_env, child); - } else + } else { rc = PTR_ERR(child); - + } mdt_create_pack_capa(info, rc, child, repbody); +out_put_parent: mdt_object_unlock_put(info, parent, lh, rc); RETURN(rc); } @@ -215,48 +414,27 @@ static int mdt_md_mkobj(struct mdt_thread_info *info) RETURN(rc); } -/* In the raw-setattr case, we lock the child inode. - * In the write-back case or if being called from open, - * the client holds a lock already. - * We use the ATTR_FROM_OPEN (translated into MRF_SETATTR_LOCKED by - * mdt_setattr_unpack()) flag to tell these cases apart. */ -int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, int flags) +int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, + struct md_attr *ma, int flags) { - struct md_attr *ma = &info->mti_attr; struct mdt_lock_handle *lh; - int som_update = 0; + int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS); + __u64 lockpart = MDS_INODELOCK_UPDATE; int rc; ENTRY; /* attr shouldn't be set on remote object */ LASSERT(mdt_object_exists(mo) >= 0); - if (info->mti_epoch) - som_update = (info->mti_epoch->flags & MF_SOM_CHANGE); - - /* Try to avoid object_lock if another epoch has been started - * already. */ - if (som_update && (info->mti_epoch->ioepoch != mo->mot_ioepoch)) - RETURN(0); - lh = &info->mti_lh[MDT_LH_PARENT]; mdt_lock_reg_init(lh, LCK_PW); - if (!(flags & MRF_SETATTR_LOCKED)) { - __u64 lockpart = MDS_INODELOCK_UPDATE; - if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID)) - lockpart |= MDS_INODELOCK_LOOKUP; + if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID)) + lockpart |= MDS_INODELOCK_LOOKUP; - rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK); - if (rc != 0) - RETURN(rc); - } - - /* Setattrs are syncronized through dlm lock taken above. If another - * epoch started, its attributes may be already flushed on disk, - * skip setattr. */ - if (som_update && (info->mti_epoch->ioepoch != mo->mot_ioepoch)) - GOTO(out_unlock, rc = 0); + rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK); + if (rc != 0) + RETURN(rc); if (mdt_object_exists(mo) == 0) GOTO(out_unlock, rc = -ENOENT); @@ -269,19 +447,20 @@ int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo, int flags) if (unlikely(ma->ma_attr.la_valid == LA_CTIME)) ma->ma_attr_flags |= MDS_VTX_BYPASS; + /* VBR: update version if attr changed are important for recovery */ + if (do_vbr) { + /* update on-disk version of changed object */ + info->mti_mos = mo; + rc = mdt_version_get_check_save(info, mo, 0); + if (rc) + GOTO(out_unlock, rc); + } + /* all attrs are packed into mti_attr in unpack_setattr */ rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma); if (rc != 0) GOTO(out_unlock, rc); - /* Re-enable SIZEONMDS. */ - if (som_update) { - CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n", - mo->mot_ioepoch, PFID(mdt_object_fid(mo)), - mo->mot_epochcount); - mdt_sizeonmds_enable(info, mo); - } - EXIT; out_unlock: mdt_object_unlock(info, mo, lh, rc); @@ -291,7 +470,6 @@ out_unlock: static int mdt_reint_setattr(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) { - struct mdt_device *mdt = info->mti_mdt; struct md_attr *ma = &info->mti_attr; struct mdt_reint_record *rr = &info->mti_rr; struct ptlrpc_request *req = mdt_info_req(info); @@ -300,7 +478,7 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, struct mdt_object *mo; struct md_object *next; struct mdt_body *repbody; - int rc; + int som_au, rc, rc2; ENTRY; DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1), @@ -314,68 +492,78 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, if (IS_ERR(mo)) GOTO(out, rc = PTR_ERR(mo)); + /* start a log jounal handle if needed */ if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) { if ((ma->ma_attr.la_valid & LA_SIZE) || - (rr->rr_flags & MRF_SETATTR_LOCKED)) { + (rr->rr_flags & MRF_OPEN_TRUNC)) { /* Check write access for the O_TRUNC case */ - if (mdt_write_read(info->mti_mdt, mo) < 0) + if (mdt_write_read(mo) < 0) GOTO(out_put, rc = -ETXTBSY); } - } else if (info->mti_epoch && - (info->mti_epoch->flags & MF_EPOCH_OPEN)) { - /* Truncate case. */ - rc = mdt_write_get(info->mti_mdt, mo); + } else if (info->mti_ioepoch && + (info->mti_ioepoch->flags & MF_EPOCH_OPEN)) { + /* Truncate case. IOEpoch is opened. */ + rc = mdt_write_get(mo); if (rc) GOTO(out_put, rc); mfd = mdt_mfd_new(); - if (mfd == NULL) + if (mfd == NULL) { + mdt_write_put(mo); GOTO(out_put, rc = -ENOMEM); + } - mdt_epoch_open(info, mo); + mdt_ioepoch_open(info, mo, 0); repbody->ioepoch = mo->mot_ioepoch; mdt_object_get(info->mti_env, mo); - mdt_mfd_set_mode(mfd, FMODE_EPOCHLCK); + mdt_mfd_set_mode(mfd, MDS_FMODE_TRUNC); mfd->mfd_object = mo; mfd->mfd_xid = req->rq_xid; - spin_lock(&med->med_open_lock); - list_add(&mfd->mfd_list, &med->med_open_head); - spin_unlock(&med->med_open_lock); + cfs_spin_lock(&med->med_open_lock); + cfs_list_add(&mfd->mfd_list, &med->med_open_head); + cfs_spin_unlock(&med->med_open_lock); repbody->handle.cookie = mfd->mfd_handle.h_cookie; } - if (info->mti_epoch && (info->mti_epoch->flags & MF_SOM_CHANGE)) - ma->ma_attr_flags |= MDS_PERM_BYPASS | MDS_SOM; - - rc = mdt_attr_set(info, mo, rr->rr_flags); - if (rc) - GOTO(out_put, rc); - - if (info->mti_epoch && (info->mti_epoch->flags & MF_SOM_CHANGE)) { + som_au = info->mti_ioepoch && info->mti_ioepoch->flags & MF_SOM_CHANGE; + if (som_au) { + /* SOM Attribute update case. Find the proper mfd and update + * SOM attributes on the proper object. */ LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM); - LASSERT(info->mti_epoch); + LASSERT(info->mti_ioepoch); - spin_lock(&med->med_open_lock); - /* Size-on-MDS Update. Find and free mfd. */ - mfd = mdt_handle2mfd(info, &info->mti_epoch->handle); + cfs_spin_lock(&med->med_open_lock); + mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle); if (mfd == NULL) { - spin_unlock(&med->med_open_lock); - CDEBUG(D_INODE | D_ERROR, "no handle for file close: " - "fid = "DFID": cookie = "LPX64"\n", - PFID(info->mti_rr.rr_fid1), - info->mti_epoch->handle.cookie); + cfs_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), + info->mti_ioepoch->handle.cookie); GOTO(out_put, rc = -ESTALE); } - LASSERT(mfd->mfd_mode == FMODE_SOM); - LASSERT(!(info->mti_epoch->flags & MF_EPOCH_CLOSE)); + LASSERT(mfd->mfd_mode == MDS_FMODE_SOM); + LASSERT(!(info->mti_ioepoch->flags & MF_EPOCH_CLOSE)); class_handle_unhash(&mfd->mfd_handle); - list_del_init(&mfd->mfd_list); - spin_unlock(&med->med_open_lock); + 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); mdt_mfd_close(info, mfd); + + OBD_FREE_LARGE(ma->ma_lmm, info->mti_mdt->mdt_max_mdsize); + } else { + rc = mdt_attr_set(info, mo, ma, rr->rr_flags); + if (rc) + GOTO(out_put, rc); } ma->ma_need = MA_INODE; @@ -387,9 +575,10 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo)); - if (mdt->mdt_opts.mo_oss_capa && + 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)) { + (ma->ma_attr.la_valid & LA_SIZE) && !som_au) { struct lustre_capa *capa; capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2); @@ -405,14 +594,21 @@ static int mdt_reint_setattr(struct mdt_thread_info *info, out_put: mdt_object_put(info->mti_env, mo); out: - mdt_shrink_reply(info); + if (rc == 0) + mdt_counter_incr(req, LPROC_MDT_SETATTR); + + mdt_client_compatibility(info); + rc2 = mdt_fix_reply(info); + if (rc == 0) + rc = rc2; return rc; } static int mdt_reint_create(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) { - int rc; + struct ptlrpc_request *req = mdt_info_req(info); + int rc; ENTRY; if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) @@ -424,10 +620,15 @@ static int mdt_reint_create(struct mdt_thread_info *info, 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); - break; + } else { + LASSERT(info->mti_rr.rr_namelen > 0); + mdt_counter_incr(req, LPROC_MDT_MKDIR); + rc = mdt_md_create(info); } + break; } case S_IFREG: case S_IFLNK: @@ -437,6 +638,7 @@ static int mdt_reint_create(struct mdt_thread_info *info, case S_IFSOCK:{ /* Special file should stay on the same node as parent. */ LASSERT(info->mti_rr.rr_namelen > 0); + mdt_counter_incr(req, LPROC_MDT_MKNOD); rc = mdt_md_create(info); break; } @@ -446,6 +648,10 @@ static int mdt_reint_create(struct mdt_thread_info *info, RETURN(rc); } +/* + * VBR: save parent version in reply and child version getting by its name. + * Version of child is getting and checking during its lookup. If + */ static int mdt_reint_unlink(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) { @@ -470,9 +676,9 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK)) RETURN(err_serious(-ENOENT)); - /* + /* * step 1: lock the parent. Note, this may be child in case of - * remote operation denoted by ->mti_cross_ref flag. + * remote operation denoted by ->mti_cross_ref flag. */ parent_lh = &info->mti_lh[MDT_LH_PARENT]; if (info->mti_cross_ref) { @@ -495,6 +701,13 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, GOTO(out, rc); } + if (mdt_object_obf(mp)) + GOTO(out_unlock_parent, rc = -EPERM); + + rc = mdt_version_get_check_save(info, mp, 0); + if (rc) + GOTO(out_unlock_parent, rc); + mdt_reint_init_ma(info, ma); if (!ma->ma_lmm || !ma->ma_cookie) GOTO(out_unlock_parent, rc = -EINVAL); @@ -519,8 +732,9 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, /* step 2: find & lock the child */ lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); - rc = mdo_lookup(info->mti_env, mdt_object_child(mp), - lname, child_fid, &info->mti_spec); + /* 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) GOTO(out_unlock_parent, rc); @@ -539,7 +753,8 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, OBD_FAIL_MDS_REINT_UNLINK_WRITE); - + /* save version when object is locked */ + mdt_version_get_save(info, mc, 1); /* * Now we can only make sure we need MA_INODE, in mdd layer, will check * whether need MA_LOV and MA_COOKIE. @@ -552,7 +767,27 @@ static int mdt_reint_unlink(struct mdt_thread_info *info, 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, LPROC_MDT_RMDIR); + break; + case S_IFREG: + case S_IFLNK: + case S_IFCHR: + case S_IFBLK: + case S_IFIFO: + case S_IFSOCK: + mdt_counter_incr(req, LPROC_MDT_UNLINK); + break; + default: + LASSERTF(0, "bad file type %o unlinking\n", + ma->ma_attr.la_mode); + } + } + EXIT; + mdt_object_unlock_put(info, mc, child_lh, rc); out_unlock_parent: mdt_object_unlock_put(info, mp, parent_lh, rc); @@ -560,6 +795,10 @@ out: return rc; } +/* + * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by + * name. + */ static int mdt_reint_link(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) { @@ -605,13 +844,20 @@ static int mdt_reint_link(struct mdt_thread_info *info, /* step 1: find & lock the target parent dir */ lhp = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_pdo_init(lhp, LCK_EX, rr->rr_name, + mdt_lock_pdo_init(lhp, LCK_PW, rr->rr_name, rr->rr_namelen); mp = mdt_object_find_lock(info, rr->rr_fid2, lhp, MDS_INODELOCK_UPDATE); if (IS_ERR(mp)) RETURN(PTR_ERR(mp)); + if (mdt_object_obf(mp)) + GOTO(out_unlock_parent, rc = -EPERM); + + rc = mdt_version_get_check_save(info, mp, 0); + if (rc) + GOTO(out_unlock_parent, rc); + /* step 2: find & lock the source */ lhs = &info->mti_lh[MDT_LH_CHILD]; mdt_lock_reg_init(lhs, LCK_EX); @@ -631,16 +877,66 @@ static int mdt_reint_link(struct mdt_thread_info *info, mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom, OBD_FAIL_MDS_REINT_LINK_WRITE); + info->mti_mos = ms; + rc = mdt_version_get_check_save(info, ms, 1); + if (rc) + GOTO(out_unlock_child, rc); + lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); + /** check target version by name during replay */ + rc = mdt_lookup_version_check(info, mp, lname, &info->mti_tmp_fid1, 2); + if (rc != 0 && rc != -ENOENT) + 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))) { + info->mti_ver[2] = ENOENT_VERSION; + mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2); + } + rc = mdo_link(info->mti_env, mdt_object_child(mp), mdt_object_child(ms), lname, ma); + if (rc == 0) + mdt_counter_incr(req, LPROC_MDT_LINK); + EXIT; +out_unlock_child: mdt_object_unlock_put(info, ms, lhs, rc); out_unlock_parent: mdt_object_unlock_put(info, mp, lhp, rc); return rc; } +/** + * lock the part of the directory according to the hash of the name + * (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 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) @@ -723,25 +1019,17 @@ static int mdt_rename_lock(struct mdt_thread_info *info, 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 lu_site *ls; + struct md_site *ms; int rc; ENTRY; - /* - * Disable global rename BFL lock temporarily because - * when a mds do rename recoverying, which might enqueue - * BFL lock to the controller mds. and this req might be - * replay req for controller mds. but we did not have - * such handling in controller mds. XXX - */ - RETURN(0); - ls = info->mti_mdt->mdt_md_dev.md_lu_dev.ld_site; + 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 (ls->ls_control_exp == NULL) { + if (ms->ms_control_exp == NULL) { int flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB; /* @@ -751,7 +1039,8 @@ static int mdt_rename_lock(struct mdt_thread_info *info, rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy, LCK_EX, &flags, ldlm_blocking_ast, ldlm_completion_ast, NULL, NULL, 0, - NULL, lh); + &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 }; @@ -761,8 +1050,8 @@ static int mdt_rename_lock(struct mdt_thread_info *info, * This is the case mdt0 is remote node, issue DLM lock like * other clients. */ - rc = ldlm_cli_enqueue(ls->ls_control_exp, NULL, &einfo, res_id, - policy, &flags, NULL, 0, NULL, lh, 0); + rc = ldlm_cli_enqueue(ms->ms_control_exp, NULL, &einfo, res_id, + policy, &flags, NULL, 0, lh, 0); } RETURN(rc); @@ -771,9 +1060,6 @@ static int mdt_rename_lock(struct mdt_thread_info *info, static void mdt_rename_unlock(struct lustre_handle *lh) { ENTRY; - /* Disable global rename BFL lock temporarily. see above XXX*/ - EXIT; - return; LASSERT(lustre_handle_is_used(lh)); ldlm_lock_decref(lh, LCK_EX); EXIT; @@ -815,6 +1101,11 @@ static int mdt_rename_sanity(struct mdt_thread_info *info, struct lu_fid *fid) RETURN(rc); } +/* + * VBR: rename versions in reply: 0 - src parent; 1 - tgt parent; + * 2 - src child; 3 - tgt child. + * Update on disk version of src child. + */ static int mdt_reint_rename(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) { @@ -838,7 +1129,7 @@ static int mdt_reint_rename(struct mdt_thread_info *info, ENTRY; if (info->mti_dlm_req) - ldlm_request_cancel(mdt_info_req(info), info->mti_dlm_req, 0); + ldlm_request_cancel(req, info->mti_dlm_req, 0); if (info->mti_cross_ref) { rc = mdt_reint_rename_tgt(info); @@ -866,6 +1157,13 @@ static int mdt_reint_rename(struct mdt_thread_info *info, if (IS_ERR(msrcdir)) GOTO(out_rename_lock, rc = PTR_ERR(msrcdir)); + if (mdt_object_obf(msrcdir)) + GOTO(out_unlock_source, rc = -EPERM); + + rc = mdt_version_get_check_save(info, msrcdir, 0); + if (rc) + GOTO(out_unlock_source, rc); + /* step 2: find & lock the target dir. */ lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD]; mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, rr->rr_tgt, @@ -873,30 +1171,47 @@ static int mdt_reint_rename(struct mdt_thread_info *info, if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) { mdt_object_get(info->mti_env, msrcdir); mtgtdir = msrcdir; + if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) { + rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir, + MDS_INODELOCK_UPDATE); + if (rc) + GOTO(out_unlock_source, rc); + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10); + } } else { mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2); if (IS_ERR(mtgtdir)) GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir)); + if (mdt_object_obf(mtgtdir)) + GOTO(out_put_target, rc = -EPERM); + + /* check early, the real version will be saved after locking */ + rc = mdt_version_get_check(info, mtgtdir, 1); + if (rc) + GOTO(out_put_target, rc); + rc = mdt_object_exists(mtgtdir); - if (rc == 0) - GOTO(out_unlock_target, rc = -ESTALE); - else if (rc > 0) { + 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_unlock_target, rc); + GOTO(out_put_target, rc); + /* get and save correct version after locking */ + mdt_version_get_save(info, mtgtdir, 1); } } /* 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); - rc = mdo_lookup(info->mti_env, mdt_object_child(msrcdir), - &slname, old_fid, &info->mti_spec); + fid_zero(old_fid); + rc = mdt_lookup_version_check(info, msrcdir, &slname, old_fid, 2); if (rc != 0) GOTO(out_unlock_target, rc); @@ -915,13 +1230,18 @@ static int mdt_reint_rename(struct mdt_thread_info *info, mdt_object_put(info->mti_env, mold); GOTO(out_unlock_target, rc); } + + info->mti_mos = mold; + /* save version after locking */ + mdt_version_get_save(info, mold, 2); mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA); /* step 4: find & lock the new object. */ /* new target object may not exist now */ lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen); - rc = mdo_lookup(info->mti_env, mdt_object_child(mtgtdir), - lname, new_fid, &info->mti_spec); + /* 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 */ if (lu_fid_eq(old_fid, new_fid)) @@ -942,9 +1262,14 @@ static int mdt_reint_rename(struct mdt_thread_info *info, mdt_object_put(info->mti_env, mnew); GOTO(out_unlock_old, rc); } + /* get and save version after locking */ + mdt_version_get_save(info, mnew, 3); mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA); - } else if (rc != -EREMOTE && rc != -ENOENT) + } else if (rc != -EREMOTE && rc != -ENOENT) { GOTO(out_unlock_old, rc); + } else { + mdt_enoent_version_save(info, 3); + } /* step 5: rename it */ mdt_reint_init_ma(info, ma); @@ -966,8 +1291,14 @@ static int mdt_reint_rename(struct mdt_thread_info *info, lname, ma); /* handle last link of tgt object */ - if (rc == 0 && mnew) - mdt_handle_last_unlink(info, mnew, ma); + if (rc == 0) { + 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; out_unlock_new: @@ -976,7 +1307,9 @@ out_unlock_new: out_unlock_old: mdt_object_unlock_put(info, mold, lh_oldp, rc); out_unlock_target: - mdt_object_unlock_put(info, mtgtdir, lh_tgtdirp, rc); + mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc); +out_put_target: + mdt_object_put(info->mti_env, mtgtdir); out_unlock_source: mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc); out_rename_lock: