X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_open.c;h=bcb7c2f8008cc88f8422a5d7be1c0a36a8c1f1db;hp=dcd18c5bc85fb9c32ed03be4f614ed7473697678;hb=8d27c92a66d63aaf8b8fbe1fc73e49263b5bed1e;hpb=f0c1b06f2418d2016e23eb6e8277be85ca13fc53 diff --git a/lustre/mdt/mdt_open.c b/lustre/mdt/mdt_open.c index dcd18c5..bcb7c2f 100644 --- a/lustre/mdt/mdt_open.c +++ b/lustre/mdt/mdt_open.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2013, Intel Corporation. + * Copyright (c) 2011, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -44,14 +40,16 @@ #include #include +#include #include "mdt_internal.h" +#include /* we do nothing because we do not have refcount now */ static void mdt_mfd_get(void *mfdp) { } -static struct portals_handle_ops mfd_handle_ops = { +static const struct portals_handle_ops mfd_open_handle_ops = { .hop_addref = mdt_mfd_get, .hop_free = NULL, }; @@ -65,10 +63,10 @@ struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med) OBD_ALLOC_PTR(mfd); if (mfd != NULL) { - CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link); - mfd->mfd_handle.h_owner = med; - CFS_INIT_LIST_HEAD(&mfd->mfd_list); - class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops); + INIT_LIST_HEAD_RCU(&mfd->mfd_open_handle.h_link); + mfd->mfd_owner = med; + INIT_LIST_HEAD(&mfd->mfd_list); + class_handle_hash(&mfd->mfd_open_handle, &mfd_open_handle_ops); } RETURN(mfd); @@ -80,19 +78,20 @@ struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med) * but mfd can be found in mfd list by that handle. * Callers need to be holding med_open_lock. */ -struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med, - const struct lustre_handle *handle, - bool is_replay_or_resent) +struct mdt_file_data *mdt_open_handle2mfd(struct mdt_export_data *med, + const struct lustre_handle *open_handle, + bool is_replay_or_resent) { struct mdt_file_data *mfd; ENTRY; - LASSERT(handle != NULL); - mfd = class_handle2object(handle->cookie, med); + LASSERT(open_handle != NULL); + mfd = class_handle2object(open_handle->cookie, &mfd_open_handle_ops); /* during dw/setattr replay the mfd can be found by old handle */ - if (mfd == NULL && is_replay_or_resent) { - cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) { - if (mfd->mfd_old_handle.cookie == handle->cookie) + if ((!mfd || mfd->mfd_owner != med) && is_replay_or_resent) { + list_for_each_entry(mfd, &med->med_open_head, mfd_list) { + if (mfd->mfd_open_handle_old.cookie == + open_handle->cookie) RETURN(mfd); } mfd = NULL; @@ -104,8 +103,8 @@ struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med, /* free mfd */ void mdt_mfd_free(struct mdt_file_data *mfd) { - LASSERT(cfs_list_empty(&mfd->mfd_list)); - OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle); + LASSERT(list_empty(&mfd->mfd_list)); + OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_open_handle); } static int mdt_create_data(struct mdt_thread_info *info, @@ -122,11 +121,7 @@ static int mdt_create_data(struct mdt_thread_info *info, ma->ma_need = MA_INODE | MA_LOV; ma->ma_valid = 0; mutex_lock(&o->mot_lov_mutex); - if (!(o->mot_flags & MOF_LOV_CREATED)) { - if (p != NULL && (fid_is_obf(mdt_object_fid(p)) || - fid_is_dot_lustre(mdt_object_fid(p)))) - GOTO(unlock, rc = -EPERM); - + if (!o->mot_lov_created) { rc = mdo_create_data(info->mti_env, p ? mdt_object_child(p) : NULL, mdt_object_child(o), spec, ma); @@ -134,437 +129,66 @@ static int mdt_create_data(struct mdt_thread_info *info, rc = mdt_attr_get_complex(info, o, ma); if (rc == 0 && ma->ma_valid & MA_LOV) - o->mot_flags |= MOF_LOV_CREATED; + o->mot_lov_created = 1; } -unlock: + mutex_unlock(&o->mot_lov_mutex); RETURN(rc); } -static int mdt_ioepoch_opened(struct mdt_object *mo) -{ - return mo->mot_ioepoch_count; -} - -int mdt_object_is_som_enabled(struct mdt_object *mo) -{ - return !mo->mot_ioepoch; -} - -/** - * Re-enable Size-on-MDS. - * Call under ->mot_ioepoch_mutex. - */ -static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch) -{ - if (ioepoch == mo->mot_ioepoch) { - LASSERT(!mdt_ioepoch_opened(mo)); - mo->mot_ioepoch = 0; - mo->mot_flags = 0; - } -} - -/** - * Open the IOEpoch. It is allowed if @writecount is not negative. - * The epoch and writecount handling is performed under the mot_ioepoch_mutex. - */ -int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o, - int created) -{ - struct mdt_device *mdt = info->mti_mdt; - int cancel = 0; - int rc = 0; - ENTRY; - - if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) || - !S_ISREG(lu_object_attr(&o->mot_obj))) - RETURN(0); - - mutex_lock(&o->mot_ioepoch_mutex); - if (mdt_ioepoch_opened(o)) { - /* Epoch continues even if there is no writers yet. */ - CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n", - o->mot_ioepoch, PFID(mdt_object_fid(o))); - } else { - /* XXX: ->mdt_ioepoch is not initialized at the mount */ - spin_lock(&mdt->mdt_ioepoch_lock); - if (mdt->mdt_ioepoch < info->mti_replayepoch) - mdt->mdt_ioepoch = info->mti_replayepoch; - - if (info->mti_replayepoch) - o->mot_ioepoch = info->mti_replayepoch; - else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL) - o->mot_ioepoch = ++mdt->mdt_ioepoch; - else - o->mot_ioepoch = mdt->mdt_ioepoch; - - spin_unlock(&mdt->mdt_ioepoch_lock); - - CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n", - o->mot_ioepoch, PFID(mdt_object_fid(o))); - if (created) - o->mot_flags |= MOF_SOM_CREATED; - cancel = 1; - } - o->mot_ioepoch_count++; - mutex_unlock(&o->mot_ioepoch_mutex); - - /* Cancel Size-on-MDS attributes cached on clients for the open case. - * In the truncate case, see mdt_reint_setattr(). */ - if (cancel && (info->mti_rr.rr_fid1 != NULL)) { - struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD]; - mdt_lock_reg_init(lh, LCK_EX); - rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE, - MDT_LOCAL_LOCK); - if (rc == 0) - mdt_object_unlock(info, o, lh, 1); - } - RETURN(rc); -} - -/** - * Update SOM on-disk attributes. - * If enabling, write update inodes and lustre-ea with the proper IOEpoch, - * mountid and attributes. If disabling, clean SOM xattr. - * Call under ->mot_ioepoch_mutex. - */ -static int mdt_som_attr_set(struct mdt_thread_info *info, - struct mdt_object *obj, __u64 ioepoch, bool enable) -{ - struct md_object *next = mdt_object_child(obj); - int rc; - ENTRY; - - CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64 - " on "DFID".\n", enable ? "update" : "disabling", - ioepoch, PFID(mdt_object_fid(obj))); - - if (enable) { - struct lu_buf *buf = &info->mti_buf; - struct som_attrs *attrs; - struct md_attr *ma = &info->mti_attr; - struct lu_attr *la = &ma->ma_attr; - struct obd_device *obd = info->mti_mdt->mdt_lut.lut_obd; - - attrs = (struct som_attrs *)info->mti_xattr_buf; - CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs)); - - /* pack SOM attributes */ - memset(attrs, 0, sizeof(*attrs)); - attrs->som_ioepoch = ioepoch; - attrs->som_mountid = obd->u.obt.obt_mount_count; - if ((la->la_valid & LA_SIZE) != 0) - attrs->som_size = la->la_size; - if ((la->la_valid & LA_BLOCKS) != 0) - attrs->som_blocks = la->la_blocks; - lustre_som_swab(attrs); - - /* update SOM attributes */ - buf->lb_buf = attrs; - buf->lb_len = sizeof(*attrs); - rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0); - } else { - /* delete SOM attributes */ - rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM); - } - - RETURN(rc); -} - -/** Perform the eviction specific actions on ioepoch close. */ -static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info, - struct mdt_object *o) -{ - int rc = 0; - - mutex_lock(&o->mot_ioepoch_mutex); - CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". " - "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)), - o->mot_ioepoch_count); - o->mot_ioepoch_count--; - - /* If eviction occured set MOF_SOM_RECOV, - * if no other epoch holders, disable SOM on disk. */ - o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV; - if (!mdt_ioepoch_opened(o)) { - rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE); - mdt_object_som_enable(o, o->mot_ioepoch); - } - mutex_unlock(&o->mot_ioepoch_mutex); - RETURN(rc); -} - -/** - * Perform the replay specific actions on ioepoch close. - * Skip SOM attribute update if obtained and just forget about the inode state - * for the last ioepoch holder. The SOM cache is invalidated on MDS failure. - */ -static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info, - struct mdt_object *o) -{ - int rc = MDT_IOEPOCH_CLOSED; - ENTRY; - - mutex_lock(&o->mot_ioepoch_mutex); - CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n", - o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count); - o->mot_ioepoch_count--; - - /* Get an info from the replayed request if client is supposed - * to send an Attibute Update, reconstruct @rc if so */ - if (info->mti_ioepoch->flags & MF_SOM_AU) - rc = MDT_IOEPOCH_GETATTR; - - if (!mdt_ioepoch_opened(o)) - mdt_object_som_enable(o, info->mti_ioepoch->ioepoch); - mutex_unlock(&o->mot_ioepoch_mutex); - - RETURN(rc); -} - -/** - * Regular file IOepoch close. - * Closes the ioepoch, checks the object state, apply obtained attributes and - * re-enable SOM on the object, if possible. Also checks if the recovery is - * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client - * to obtain SOM attributes under the server-side OST locks. - * - * Return value: - * MDT_IOEPOCH_CLOSED if ioepoch is closed. - * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed. - */ -static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info, - struct mdt_object *o) -{ - struct md_attr *tmp_ma; - struct lu_attr *la; - int achange, opened; - int recovery = 0; - int rc = 0, ret = MDT_IOEPOCH_CLOSED; - ENTRY; - - la = &info->mti_attr.ma_attr; - achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE); - - mutex_lock(&o->mot_ioepoch_mutex); - o->mot_ioepoch_count--; - - tmp_ma = &info->mti_u.som.attr; - tmp_ma->ma_lmm = info->mti_attr.ma_lmm; - tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size; - tmp_ma->ma_som = &info->mti_u.som.data; - tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM; - tmp_ma->ma_valid = 0; - rc = mdt_attr_get_complex(info, o, tmp_ma); - if (rc) - GOTO(error_up, rc); - - /* Check the on-disk SOM state. */ - if (o->mot_flags & MOF_SOM_RECOV) - recovery = 1; - else if (!(o->mot_flags & MOF_SOM_CREATED) && - !(tmp_ma->ma_valid & MA_SOM)) - recovery = 1; - - CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n", - o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count); - - opened = mdt_ioepoch_opened(o); - /** - * If IOEpoch is not opened, check if a Size-on-MDS update is needed. - * Skip the check for file with no LOV or for unlink files. - */ - if (!opened && tmp_ma->ma_valid & MA_LOV && - !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) { - if (recovery) - /* If some previous writer was evicted, re-ask the - * client for attributes. Even if attributes are - * provided, we cannot believe in them. - * Another use case is that there is no SOM cache on - * disk -- first access with SOM or there was an MDS - * failure. */ - ret = MDT_IOEPOCH_GETATTR; - else if (o->mot_flags & MOF_SOM_CHANGE) - /* Some previous writer changed the attribute. - * Do not believe to the current Size-on-MDS - * update, re-ask client. */ - ret = MDT_IOEPOCH_GETATTR; - else if (!(la->la_valid & LA_SIZE) && achange) - /* Attributes were changed by the last writer - * only but no Size-on-MDS update is received.*/ - ret = MDT_IOEPOCH_GETATTR; - } - - if (achange || ret == MDT_IOEPOCH_GETATTR) - o->mot_flags |= MOF_SOM_CHANGE; - - /* If epoch ends and relable SOM attributes are obtained, update them. - * Create SOM ea for new files even if there is no attributes obtained - * (0-length file). */ - if (ret == MDT_IOEPOCH_CLOSED && !opened) { - if (achange || o->mot_flags & MOF_SOM_CREATED) { - LASSERT(achange || !(la->la_valid & LA_SIZE)); - rc = mdt_som_attr_set(info, o, o->mot_ioepoch, - MDT_SOM_ENABLE); - /* Avoid the following setattrs of these attributes, - * e.g. for atime update. */ - info->mti_attr.ma_valid = 0; - } - mdt_object_som_enable(o, o->mot_ioepoch); - } - - mutex_unlock(&o->mot_ioepoch_mutex); - /* If recovery is needed, tell the client to perform GETATTR under - * the lock. */ - if (ret == MDT_IOEPOCH_GETATTR && recovery) { - struct mdt_body *rep; - rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); - rep->valid |= OBD_MD_FLGETATTRLOCK; - } - - RETURN(rc ? : ret); - -error_up: - mutex_unlock(&o->mot_ioepoch_mutex); - return rc; -} - -/** - * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if: - * - a client closes the IOEpoch; - * - a client eviction occured. - * Return values: - * MDT_IOEPOCH_OPENED if the client does not close IOEpoch. - * MDT_IOEPOCH_CLOSED if the client closes IOEpoch. - * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute - * update is needed. - */ -static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o) +int mdt_write_read(struct mdt_object *o) { - struct ptlrpc_request *req = mdt_info_req(info); + int rc = 0; ENTRY; - - if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) || - !S_ISREG(lu_object_attr(&o->mot_obj))) - RETURN(0); - - LASSERT(o->mot_ioepoch_count); - LASSERT(info->mti_ioepoch == NULL || - info->mti_ioepoch->ioepoch == o->mot_ioepoch); - - /* IOEpoch is closed only if client tells about it or eviction occures. - * In the replay case, always close the epoch. */ - if (req == NULL) - RETURN(mdt_ioepoch_close_on_eviction(info, o)); - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) - RETURN(mdt_ioepoch_close_on_replay(info, o)); - if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE)) - RETURN(mdt_ioepoch_close_reg(info, o)); - /* IO epoch is not closed. */ - RETURN(MDT_IOEPOCH_OPENED); + spin_lock(&o->mot_write_lock); + rc = o->mot_write_count; + spin_unlock(&o->mot_write_lock); + RETURN(rc); } -/** - * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting - * for attribute update. It happens if: - * - SOM Attribute Update is obtained; - * - the client failed to obtain it and informs MDS about it; - * - a client eviction occured. - * Apply obtained attributes for the 1st case, wipe out the on-disk SOM - * cache otherwise. - */ -int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o) +int mdt_write_get(struct mdt_object *o) { - struct ptlrpc_request *req = mdt_info_req(info); - __u64 ioepoch = 0; - int act = MDT_SOM_ENABLE; - int rc = 0; + int rc = 0; ENTRY; + spin_lock(&o->mot_write_lock); + if (o->mot_write_count < 0) + rc = -ETXTBSY; + else + o->mot_write_count++; + spin_unlock(&o->mot_write_lock); - LASSERT(!req || info->mti_ioepoch); - if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) || - !S_ISREG(lu_object_attr(&o->mot_obj))) - RETURN(0); - - /* No size whereas MF_SOM_CHANGE is set means client failed to - * obtain ost attributes, drop the SOM cache on disk if so. */ - if (!req || - (info->mti_ioepoch && - info->mti_ioepoch->flags & MF_SOM_CHANGE && - !(info->mti_attr.ma_attr.la_valid & LA_SIZE))) - act = MDT_SOM_DISABLE; - - mutex_lock(&o->mot_ioepoch_mutex); - /* Mark the object it is the recovery state if we failed to obtain - * SOM attributes. */ - if (act == MDT_SOM_DISABLE) - o->mot_flags |= MOF_SOM_RECOV; - - if (!mdt_ioepoch_opened(o)) { - ioepoch = info->mti_ioepoch ? - info->mti_ioepoch->ioepoch : o->mot_ioepoch; - - if (req != NULL - && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) - rc = mdt_som_attr_set(info, o, ioepoch, act); - mdt_object_som_enable(o, ioepoch); - } - mutex_unlock(&o->mot_ioepoch_mutex); - RETURN(rc); -} - -int mdt_write_read(struct mdt_object *o) -{ - int rc = 0; - ENTRY; - mutex_lock(&o->mot_ioepoch_mutex); - rc = o->mot_writecount; - mutex_unlock(&o->mot_ioepoch_mutex); - RETURN(rc); -} - -int mdt_write_get(struct mdt_object *o) -{ - int rc = 0; - ENTRY; - mutex_lock(&o->mot_ioepoch_mutex); - if (o->mot_writecount < 0) - rc = -ETXTBSY; - else - o->mot_writecount++; - mutex_unlock(&o->mot_ioepoch_mutex); - RETURN(rc); + RETURN(rc); } void mdt_write_put(struct mdt_object *o) { - ENTRY; - mutex_lock(&o->mot_ioepoch_mutex); - o->mot_writecount--; - mutex_unlock(&o->mot_ioepoch_mutex); - EXIT; + ENTRY; + spin_lock(&o->mot_write_lock); + o->mot_write_count--; + spin_unlock(&o->mot_write_lock); + EXIT; } static int mdt_write_deny(struct mdt_object *o) { - int rc = 0; - ENTRY; - mutex_lock(&o->mot_ioepoch_mutex); - if (o->mot_writecount > 0) - rc = -ETXTBSY; - else - o->mot_writecount--; - mutex_unlock(&o->mot_ioepoch_mutex); - RETURN(rc); + int rc = 0; + ENTRY; + spin_lock(&o->mot_write_lock); + if (o->mot_write_count > 0) + rc = -ETXTBSY; + else + o->mot_write_count--; + spin_unlock(&o->mot_write_lock); + RETURN(rc); } static void mdt_write_allow(struct mdt_object *o) { - ENTRY; - mutex_lock(&o->mot_ioepoch_mutex); - o->mot_writecount++; - mutex_unlock(&o->mot_ioepoch_mutex); - EXIT; + ENTRY; + spin_lock(&o->mot_write_lock); + o->mot_write_count++; + spin_unlock(&o->mot_write_lock); + EXIT; } /* there can be no real transaction so prepare the fake one */ @@ -574,21 +198,36 @@ static void mdt_empty_transno(struct mdt_thread_info *info, int rc) struct ptlrpc_request *req = mdt_info_req(info); struct tg_export_data *ted; struct lsd_client_data *lcd; - ENTRY; + + if (mdt_rdonly(req->rq_export)) + RETURN_EXIT; + /* transaction has occurred already */ if (lustre_msg_get_transno(req->rq_repmsg) != 0) RETURN_EXIT; + if (tgt_is_multimodrpcs_client(req->rq_export)) { + struct thandle *th; + + /* generate an empty transaction to get a transno + * and reply data */ + th = dt_trans_create(info->mti_env, mdt->mdt_bottom); + if (!IS_ERR(th)) { + rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th); + dt_trans_stop(info->mti_env, mdt->mdt_bottom, th); + } + RETURN_EXIT; + } + spin_lock(&mdt->mdt_lut.lut_translock); if (rc != 0) { if (info->mti_transno != 0) { struct obd_export *exp = req->rq_export; - CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n", + CERROR("%s: replay trans %llu NID %s: rc = %d\n", mdt_obd_name(mdt), info->mti_transno, - libcfs_nid2str(exp->exp_connection->c_peer.nid), - rc); + obd_export_nid2str(exp), rc); spin_unlock(&mdt->mdt_lut.lut_translock); RETURN_EXIT; } @@ -601,7 +240,7 @@ static void mdt_empty_transno(struct mdt_thread_info *info, int rc) } spin_unlock(&mdt->mdt_lut.lut_translock); - CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n", + CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n", info->mti_transno, req->rq_export->exp_obd->obd_last_committed); @@ -620,14 +259,13 @@ static void mdt_empty_transno(struct mdt_thread_info *info, int rc) * otherwise the following resend(after replay) can not * be checked correctly by xid */ mutex_unlock(&ted->ted_lcd_lock); - CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n", + CDEBUG(D_HA, "%s: transno = %llu < last_transno = %llu\n", mdt_obd_name(mdt), info->mti_transno, lcd->lcd_last_transno); RETURN_EXIT; } - if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE || - lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) { + if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) { if (info->mti_transno != 0) lcd->lcd_last_close_transno = info->mti_transno; lcd->lcd_last_close_xid = req->rq_xid; @@ -653,81 +291,123 @@ static void mdt_empty_transno(struct mdt_thread_info *info, int rc) EXIT; } -void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode) +void mdt_mfd_set_mode(struct mdt_file_data *mfd, u64 open_flags) { LASSERT(mfd != NULL); - CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n", - PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode); + CDEBUG(D_DENTRY, DFID " Change mfd open_flags %#llo -> %#llo.\n", + PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_open_flags, + open_flags); - mfd->mfd_mode = mode; + mfd->mfd_open_flags = open_flags; } -static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, - struct mdt_object *o, __u64 flags, int created) +/** + * prep ma_lmm/ma_lmv for md_attr from reply + */ +static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info, + struct mdt_object *obj, + struct md_attr *ma) { - struct ptlrpc_request *req = mdt_info_req(info); - struct mdt_export_data *med = &req->rq_export->exp_mdt_data; - struct mdt_file_data *mfd; - struct md_attr *ma = &info->mti_attr; - struct lu_attr *la = &ma->ma_attr; - struct mdt_body *repbody; - int rc = 0, isdir, isreg; - ENTRY; + LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL); + if (S_ISDIR(obj->mot_header.loh_attr)) { + ma->ma_lmv = req_capsule_server_get(info->mti_pill, + &RMF_MDT_MD); + ma->ma_lmv_size = req_capsule_get_size(info->mti_pill, + &RMF_MDT_MD, + RCL_SERVER); + if (ma->ma_lmv_size > 0) + ma->ma_need |= MA_LMV; + } else { + 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); + if (ma->ma_lmm_size > 0) + ma->ma_need |= MA_LOV; + } +} - repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); +static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, + struct mdt_object *o, u64 open_flags, int created, + struct ldlm_reply *rep) +{ + struct ptlrpc_request *req = mdt_info_req(info); + struct mdt_export_data *med = &req->rq_export->exp_mdt_data; + struct mdt_file_data *mfd; + struct md_attr *ma = &info->mti_attr; + struct lu_attr *la = &ma->ma_attr; + struct mdt_body *repbody; + bool isdir, isreg; + int rc = 0; - isreg = S_ISREG(la->la_mode); - isdir = S_ISDIR(la->la_mode); - if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) { - /* - * No EA, check whether it is will set regEA and dirEA since in - * above attr get, these size might be zero, so reset it, to - * retrieve the MD after create obj. - */ - ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, - &RMF_MDT_MD, - RCL_SERVER); - /* in replay case, p == NULL */ - rc = mdt_create_data(info, p, o); - if (rc) - RETURN(rc); - } + ENTRY; + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); - CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n", - ma->ma_valid, ma->ma_lmm_size); + isreg = S_ISREG(la->la_mode); + isdir = S_ISDIR(la->la_mode); + if (isreg && !(ma->ma_valid & MA_LOV) && + !(open_flags & MDS_OPEN_RELEASE)) { + /* + * No EA, check whether it is will set regEA and dirEA since in + * above attr get, these size might be zero, so reset it, to + * retrieve the MD after create obj. + */ + ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, + &RMF_MDT_MD, + RCL_SERVER); + /* in replay case, p == NULL */ + rc = mdt_create_data(info, p, o); + if (rc) + RETURN(rc); + + if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE) + mdt_set_disposition(info, rep, DISP_OPEN_STRIPE); + } - if (ma->ma_valid & MA_LOV) { - LASSERT(ma->ma_lmm_size != 0); - repbody->eadatasize = ma->ma_lmm_size; - if (isdir) - repbody->valid |= OBD_MD_FLDIREA; - else - repbody->valid |= OBD_MD_FLEASIZE; - } + CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n", + ma->ma_valid, ma->ma_lmm_size); - if (flags & FMODE_WRITE) { - rc = mdt_write_get(o); - if (rc == 0) { - mdt_ioepoch_open(info, o, created); - repbody->ioepoch = o->mot_ioepoch; - } - } else if (flags & MDS_FMODE_EXEC) { - /* if file is released, we can't deny write because we must - * restore (write) it to access it.*/ - if ((ma->ma_valid & MA_HSM) && - (ma->ma_hsm.mh_flags & HS_RELEASED)) - rc = 0; + if (ma->ma_valid & MA_LOV) { + LASSERT(ma->ma_lmm_size != 0); + repbody->mbo_eadatasize = ma->ma_lmm_size; + if (isdir) + repbody->mbo_valid |= OBD_MD_FLDIREA; else - rc = mdt_write_deny(o); - } - if (rc) - RETURN(rc); + repbody->mbo_valid |= OBD_MD_FLEASIZE; + } + + if (ma->ma_valid & MA_LMV) { + LASSERT(ma->ma_lmv_size != 0); + repbody->mbo_eadatasize = ma->ma_lmv_size; + LASSERT(isdir); + repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA; + } + + if (open_flags & MDS_FMODE_WRITE) + rc = mdt_write_get(o); + else if (open_flags & MDS_FMODE_EXEC) + rc = mdt_write_deny(o); - rc = mo_open(info->mti_env, mdt_object_child(o), - created ? flags | MDS_OPEN_CREATED : flags); - if (rc) - GOTO(err_out, rc); + if (rc) + RETURN(rc); + + rc = mo_open(info->mti_env, mdt_object_child(o), + created ? open_flags | MDS_OPEN_CREATED : open_flags, + &info->mti_spec); + if (rc != 0) { + /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the + * client does not know which suppgid should be sent to the MDS, + * or some other(s) changed the target file's GID after this RPC + * sent to the MDS with the suppgid as the original GID, then we + * give the client another chance to send the right suppgid. */ + if (rc == -EACCES && + allow_client_chgrp(info, lu_ucred(info->mti_env))) + mdt_set_disposition(info, rep, DISP_OPEN_DENY); + + GOTO(err_out, rc); + } mfd = mdt_mfd_new(med); if (mfd == NULL) @@ -742,16 +422,16 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, mfd->mfd_xid = req->rq_xid; /* - * @flags is always not zero. At least it should be FMODE_READ, + * @open_flags is always not zero. At least it should be FMODE_READ, * FMODE_WRITE or MDS_FMODE_EXEC. */ - LASSERT(flags != 0); + LASSERT(open_flags != 0); /* Open handling. */ - mdt_mfd_set_mode(mfd, flags); + mdt_mfd_set_mode(mfd, open_flags); atomic_inc(&o->mot_open_count); - if (flags & MDS_OPEN_LEASE) + if (open_flags & MDS_OPEN_LEASE) atomic_inc(&o->mot_lease_count); /* replay handle */ @@ -762,16 +442,17 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, * might be disconnected from server, and * restart replay, so there maybe some orphan * mfd here, we should remove them */ - LASSERT(info->mti_rr.rr_handle != NULL); + LASSERT(info->mti_rr.rr_open_handle != NULL); spin_lock(&med->med_open_lock); - old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true); + old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle, + true); if (old_mfd != NULL) { CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", " - "cookie = "LPX64"\n", mfd, + "cookie = %#llx\n", mfd, PFID(mdt_object_fid(mfd->mfd_object)), - info->mti_rr.rr_handle->cookie); - class_handle_unhash(&old_mfd->mfd_handle); - cfs_list_del_init(&old_mfd->mfd_list); + info->mti_rr.rr_open_handle->cookie); + class_handle_unhash(&old_mfd->mfd_open_handle); + list_del_init(&old_mfd->mfd_list); spin_unlock(&med->med_open_lock); /* no attr update for that close */ la->la_valid = 0; @@ -783,60 +464,60 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, } else { spin_unlock(&med->med_open_lock); CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", " - "cookie = "LPX64"\n", + "cookie = %#llx\n", PFID(mdt_object_fid(mfd->mfd_object)), - info->mti_rr.rr_handle->cookie); + info->mti_rr.rr_open_handle->cookie); } - CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n", - info->mti_rr.rr_handle->cookie); + CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n", + info->mti_rr.rr_open_handle->cookie); - mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie; + mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle; } - repbody->handle.cookie = mfd->mfd_handle.h_cookie; + repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie; if (req->rq_export->exp_disconnected) { spin_lock(&med->med_open_lock); - class_handle_unhash(&mfd->mfd_handle); - cfs_list_del_init(&mfd->mfd_list); + class_handle_unhash(&mfd->mfd_open_handle); + list_del_init(&mfd->mfd_list); spin_unlock(&med->med_open_lock); mdt_mfd_close(info, mfd); } else { spin_lock(&med->med_open_lock); - cfs_list_add(&mfd->mfd_list, &med->med_open_head); + list_add_tail(&mfd->mfd_list, &med->med_open_head); spin_unlock(&med->med_open_lock); } mdt_empty_transno(info, rc); - RETURN(rc); + RETURN(rc); err_out: - if (flags & FMODE_WRITE) - /* XXX We also need to close io epoch here. - * See LU-1220 - green */ - mdt_write_put(o); - else if (flags & FMODE_EXEC) - mdt_write_allow(o); - return rc; + if (open_flags & MDS_FMODE_WRITE) + mdt_write_put(o); + else if (open_flags & MDS_FMODE_EXEC) + mdt_write_allow(o); + + return rc; } -int mdt_finish_open(struct mdt_thread_info *info, - struct mdt_object *p, struct mdt_object *o, - __u64 flags, int created, struct ldlm_reply *rep) +static int mdt_finish_open(struct mdt_thread_info *info, + struct mdt_object *p, struct mdt_object *o, + u64 open_flags, int created, + struct ldlm_reply *rep) { - struct ptlrpc_request *req = mdt_info_req(info); - struct obd_export *exp = req->rq_export; - struct mdt_export_data *med = &req->rq_export->exp_mdt_data; - struct md_attr *ma = &info->mti_attr; - struct lu_attr *la = &ma->ma_attr; - struct mdt_file_data *mfd; - struct mdt_body *repbody; - int rc = 0; - int isreg, isdir, islnk; - cfs_list_t *t; - ENTRY; + struct ptlrpc_request *req = mdt_info_req(info); + struct obd_export *exp = req->rq_export; + struct mdt_export_data *med = &req->rq_export->exp_mdt_data; + struct md_attr *ma = &info->mti_attr; + struct lu_attr *la = &ma->ma_attr; + struct mdt_file_data *mfd; + struct mdt_body *repbody; + int rc = 0; + int isreg, isdir, islnk; + struct list_head *t; + ENTRY; LASSERT(ma->ma_valid & MA_INODE); @@ -847,194 +528,149 @@ int mdt_finish_open(struct mdt_thread_info *info, islnk = S_ISLNK(la->la_mode); mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o)); + /* compatibility check for 2.10 clients when it tries to open mirrored + * files. 2.10 clients don't verify overlapping components so they + * would read and write mirrored files just as if they were normal + * PFL files, which will cause the problem that sycned mirrors actually + * contain different data. + * Older clients are not a concern here because they don't even + * understand PFL layout. */ + if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV && + mdt_lmm_is_flr(ma->ma_lmm)) { + /* LU-10286: for simplicity clients who don't understand + * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't + * be able to open mirrored files */ + RETURN(-EOPNOTSUPP); + } + + /* Overstriped files can crash older clients */ + if (isreg && !exp_connect_overstriping(exp) && + mdt_lmm_is_overstriping(ma->ma_lmm)) + RETURN(-EOPNOTSUPP); + /* LU-2275, simulate broken behaviour (esp. prevalent in * pre-2.4 servers where a very strange reply is sent on error - * that looks like it was actually almost succesful and a failure at the - * same time */ + * that looks like it was actually almost successful and a + * failure at the same time.) */ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) { mdt_set_disposition(info, rep, DISP_OPEN_OPEN | DISP_LOOKUP_NEG | DISP_LOOKUP_POS); - if (flags & MDS_OPEN_LOCK) + if (open_flags & MDS_OPEN_LOCK) mdt_set_disposition(info, rep, DISP_OPEN_LOCK); RETURN(-ENOENT); } - if (exp_connect_rmtclient(exp)) { - void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL); - - rc = mdt_pack_remote_perm(info, o, buf); - if (rc) { - repbody->valid &= ~OBD_MD_FLRMTPERM; - repbody->aclsize = 0; - } else { - repbody->valid |= OBD_MD_FLRMTPERM; - repbody->aclsize = sizeof(struct mdt_remote_perm); - } - } #ifdef CONFIG_FS_POSIX_ACL - else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) { - const struct lu_env *env = info->mti_env; - struct md_object *next = mdt_object_child(o); - struct lu_buf *buf = &info->mti_buf; - - buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL); - buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL, - RCL_SERVER); - if (buf->lb_len > 0) { - rc = mo_xattr_get(env, next, buf, - XATTR_NAME_ACL_ACCESS); - if (rc < 0) { - if (rc == -ENODATA) { - repbody->aclsize = 0; - repbody->valid |= OBD_MD_FLACL; - rc = 0; - } else if (rc == -EOPNOTSUPP) { - rc = 0; - } else { - CERROR("got acl size: %d\n", rc); - } - } else { - repbody->aclsize = rc; - repbody->valid |= OBD_MD_FLACL; - rc = 0; - } - } - } + if (exp_connect_flags(exp) & OBD_CONNECT_ACL) { + struct lu_nodemap *nodemap = nodemap_get_from_exp(exp); + if (IS_ERR(nodemap)) + RETURN(PTR_ERR(nodemap)); + + rc = mdt_pack_acl2body(info, repbody, o, nodemap); + nodemap_putref(nodemap); + if (rc) + RETURN(rc); + } #endif - if (info->mti_mdt->mdt_lut.lut_mds_capa && - exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) { - struct lustre_capa *capa; - - capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1); - LASSERT(capa); - capa->lc_opc = CAPA_OPC_MDS_DEFAULT; - rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0); - if (rc) - RETURN(rc); - repbody->valid |= OBD_MD_FLMDSCAPA; - } - if (info->mti_mdt->mdt_lut.lut_oss_capa && - exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA && - S_ISREG(lu_object_attr(&o->mot_obj))) { - struct lustre_capa *capa; - - capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2); - LASSERT(capa); - capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags); - rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0); - if (rc) - RETURN(rc); - repbody->valid |= OBD_MD_FLOSSCAPA; - } - - /* - * If we are following a symlink, don't open; and do not return open - * handle for special nodes as client required. - */ + /* + * If we are following a symlink, don't open; and do not return open + * handle for special nodes as client required. + */ if (islnk || (!isreg && !isdir && (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) { lustre_msg_set_transno(req->rq_repmsg, 0); RETURN(0); } - /* - * We need to return the existing object's fid back, so it is done here, - * after preparing the reply. - */ - if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT)) - RETURN(-EEXIST); - - /* This can't be done earlier, we need to return reply body */ - if (isdir) { - if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) { - /* We are trying to create or write an existing dir. */ - RETURN(-EISDIR); - } - } else if (flags & MDS_OPEN_DIRECTORY) - RETURN(-ENOTDIR); + /* + * We need to return the existing object's fid back, so it is done here, + * after preparing the reply. + */ + if (!created && (open_flags & MDS_OPEN_EXCL) && + (open_flags & MDS_OPEN_CREAT)) + RETURN(-EEXIST); + + /* This can't be done earlier, we need to return reply body */ + if (isdir) { + if (open_flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) { + /* We are trying to create or write an existing dir. */ + RETURN(-EISDIR); + } + } else if (open_flags & MDS_OPEN_DIRECTORY) + RETURN(-ENOTDIR); - if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE, - OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) { - RETURN(-EAGAIN); - } + if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE, + OBD_FAIL_MDS_LDLM_REPLY_NET | OBD_FAIL_ONCE)) + RETURN(-EAGAIN); - mfd = NULL; - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) { + mfd = NULL; + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) { spin_lock(&med->med_open_lock); - cfs_list_for_each(t, &med->med_open_head) { - mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list); + list_for_each(t, &med->med_open_head) { + mfd = list_entry(t, struct mdt_file_data, mfd_list); if (mfd->mfd_xid == req->rq_xid) break; mfd = NULL; } spin_unlock(&med->med_open_lock); - if (mfd != NULL) { - repbody->handle.cookie = mfd->mfd_handle.h_cookie; - /*set repbody->ea_size for resent case*/ - if (ma->ma_valid & MA_LOV) { - LASSERT(ma->ma_lmm_size != 0); - repbody->eadatasize = ma->ma_lmm_size; - if (isdir) - repbody->valid |= OBD_MD_FLDIREA; - else - repbody->valid |= OBD_MD_FLEASIZE; + if (mfd != NULL) { + repbody->mbo_open_handle.cookie = + mfd->mfd_open_handle.h_cookie; + /* set repbody->ea_size for resent case */ + if (ma->ma_valid & MA_LOV) { + LASSERT(ma->ma_lmm_size != 0); + repbody->mbo_eadatasize = ma->ma_lmm_size; + if (isdir) + repbody->mbo_valid |= OBD_MD_FLDIREA; + else + repbody->mbo_valid |= OBD_MD_FLEASIZE; } mdt_set_disposition(info, rep, DISP_OPEN_OPEN); - RETURN(0); - } - } + RETURN(0); + } + } - rc = mdt_mfd_open(info, p, o, flags, created); + rc = mdt_mfd_open(info, p, o, open_flags, created, rep); if (!rc) mdt_set_disposition(info, rep, DISP_OPEN_OPEN); - RETURN(rc); + RETURN(rc); } -extern void mdt_req_from_lcd(struct ptlrpc_request *req, - struct lsd_client_data *lcd); - void mdt_reconstruct_open(struct mdt_thread_info *info, - struct mdt_lock_handle *lhc) + struct mdt_lock_handle *lhc) { - const struct lu_env *env = info->mti_env; - struct mdt_device *mdt = info->mti_mdt; - struct req_capsule *pill = info->mti_pill; - struct ptlrpc_request *req = mdt_info_req(info); - struct tg_export_data *ted = &req->rq_export->exp_target_data; - struct lsd_client_data *lcd = ted->ted_lcd; - struct md_attr *ma = &info->mti_attr; - struct mdt_reint_record *rr = &info->mti_rr; - __u64 flags = info->mti_spec.sp_cr_flags; - struct ldlm_reply *ldlm_rep; - struct mdt_object *parent; - struct mdt_object *child; - struct mdt_body *repbody; - int rc; - ENTRY; + const struct lu_env *env = info->mti_env; + struct mdt_device *mdt = info->mti_mdt; + struct req_capsule *pill = info->mti_pill; + struct ptlrpc_request *req = mdt_info_req(info); + struct md_attr *ma = &info->mti_attr; + struct mdt_reint_record *rr = &info->mti_rr; + u64 open_flags = info->mti_spec.sp_cr_flags; + struct ldlm_reply *ldlm_rep; + struct mdt_object *parent; + struct mdt_object *child; + struct mdt_body *repbody; + u64 opdata; + int rc; + ENTRY; - LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN); - ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP); - repbody = req_capsule_server_get(pill, &RMF_MDT_BODY); + LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN); + ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP); + repbody = req_capsule_server_get(pill, &RMF_MDT_BODY); - ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD); - ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD, - RCL_SERVER); ma->ma_need = MA_INODE | MA_HSM; - if (ma->ma_lmm_size > 0) - ma->ma_need |= MA_LOV; - - ma->ma_valid = 0; + ma->ma_valid = 0; - mdt_req_from_lcd(req, lcd); - mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data); + opdata = mdt_req_from_lrd(req, info->mti_reply_data); + mdt_set_disposition(info, ldlm_rep, opdata); - CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n", + CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n", ldlm_rep->lock_policy_res1, req->rq_status); if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) && @@ -1059,104 +695,98 @@ void mdt_reconstruct_open(struct mdt_thread_info *info, mdt_export_evict(exp); RETURN_EXIT; } - child = mdt_object_find(env, mdt, rr->rr_fid2); - if (IS_ERR(child)) { - rc = PTR_ERR(child); - LCONSOLE_WARN("Child "DFID" lookup error %d." - " Evicting client %s with export %s.\n", - PFID(mdt_object_fid(child)), rc, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp)); - mdt_object_put(env, parent); - mdt_export_evict(exp); - RETURN_EXIT; - } + + child = mdt_object_find(env, mdt, rr->rr_fid2); + if (IS_ERR(child)) { + rc = PTR_ERR(child); + LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; " + "evicting client %s with export %s\n", + PFID(rr->rr_fid2), rc, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp)); + mdt_object_put(env, parent); + mdt_export_evict(exp); + RETURN_EXIT; + } if (unlikely(mdt_object_remote(child))) { + mdt_object_put(env, parent); + mdt_object_put(env, child); /* the child object was created on remote server */ - if (!mdt_is_dne_client(exp)) { + if (!mdt_is_dne_client(exp)) /* Return -EIO for old client */ - mdt_object_put(env, parent); - mdt_object_put(env, child); GOTO(out, rc = -EIO); - } - repbody->fid1 = *rr->rr_fid2; - repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); - rc = 0; - } else { - if (mdt_object_exists(child)) { - mdt_set_capainfo(info, 1, rr->rr_fid2, - BYPASS_CAPA); - rc = mdt_attr_get_complex(info, child, ma); - if (rc == 0) - rc = mdt_finish_open(info, parent, - child, flags, - 1, ldlm_rep); - } else { - /* the child does not exist, we should do - * regular open */ - mdt_object_put(env, parent); - mdt_object_put(env, child); - GOTO(regular_open, 0); - } + repbody->mbo_fid1 = *rr->rr_fid2; + repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS); + GOTO(out, rc = 0); } - mdt_object_put(env, parent); - mdt_object_put(env, child); - GOTO(out, rc); - } else { -regular_open: - /* We did not try to create, so we are a pure open */ - rc = mdt_reint_open(info, lhc); - } - - EXIT; + if (mdt_object_exists(child)) { + mdt_prep_ma_buf_from_rep(info, child, ma); + rc = mdt_attr_get_complex(info, child, ma); + if (!rc) + rc = mdt_finish_open(info, parent, child, + open_flags, 1, ldlm_rep); + mdt_object_put(env, parent); + mdt_object_put(env, child); + if (!rc) + mdt_pack_size2body(info, rr->rr_fid2, + &lhc->mlh_reg_lh); + GOTO(out, rc); + } + /* the child does not exist, we should do regular open */ + mdt_object_put(env, parent); + mdt_object_put(env, child); + } + /* We did not try to create, so we are a pure open */ + rc = mdt_reint_open(info, lhc); + EXIT; out: - req->rq_status = rc; - lustre_msg_set_status(req->rq_repmsg, req->rq_status); - LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0)); + req->rq_status = rc; + lustre_msg_set_status(req->rq_repmsg, req->rq_status); + LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0)); } -int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep) +static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep) { - __u64 flags = info->mti_spec.sp_cr_flags; - struct mdt_reint_record *rr = &info->mti_rr; - struct md_attr *ma = &info->mti_attr; - struct mdt_object *o; - int rc; - ENTRY; + u64 open_flags = info->mti_spec.sp_cr_flags; + struct mdt_reint_record *rr = &info->mti_rr; + struct md_attr *ma = &info->mti_attr; + struct mdt_object *o; + int rc; - o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2); - if (IS_ERR(o)) - RETURN(rc = PTR_ERR(o)); + ENTRY; + o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2); + if (IS_ERR(o)) + RETURN(rc = PTR_ERR(o)); if (unlikely(mdt_object_remote(o))) { - /* the child object was created on remote server */ - struct mdt_body *repbody; + /* the child object was created on remote server */ + struct mdt_body *repbody; mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD | DISP_LOOKUP_POS)); - repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); - repbody->fid1 = *rr->rr_fid2; - repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); - rc = 0; + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + repbody->mbo_fid1 = *rr->rr_fid2; + repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS); + rc = 0; } else { if (mdt_object_exists(o)) { mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD | DISP_LOOKUP_POS)); - + mdt_prep_ma_buf_from_rep(info, o, ma); rc = mdt_attr_get_complex(info, o, ma); if (rc == 0) - rc = mdt_finish_open(info, NULL, o, flags, 0, - rep); + rc = mdt_finish_open(info, NULL, o, open_flags, + 0, rep); } else { rc = -ENOENT; } } - mdt_object_put(info->mti_env, o); - RETURN(rc); + mdt_object_put(info->mti_env, o); + RETURN(rc); } /* lock object for open */ @@ -1165,13 +795,17 @@ static int mdt_object_open_lock(struct mdt_thread_info *info, struct mdt_lock_handle *lhc, __u64 *ibits) { - struct md_attr *ma = &info->mti_attr; - __u64 open_flags = info->mti_spec.sp_cr_flags; - ldlm_mode_t lm = LCK_CR; - bool acq_lease = !!(open_flags & MDS_OPEN_LEASE); - bool try_layout = false; - bool create_layout = false; - int rc = 0; + struct md_attr *ma = &info->mti_attr; + __u64 open_flags = info->mti_spec.sp_cr_flags; + __u64 trybits = 0; + enum ldlm_mode lm = LCK_CR; + bool acq_lease = !!(open_flags & MDS_OPEN_LEASE); + bool try_layout = false; + bool create_layout = false; + int rc = 0; + int dom_stripes = LMM_NO_DOM; + bool dom_lock = false; + ENTRY; *ibits = 0; @@ -1187,6 +821,24 @@ static int mdt_object_open_lock(struct mdt_thread_info *info, if (exp_connect_layout(info->mti_exp) && !create_layout && ma->ma_need & MA_LOV) try_layout = true; + + /* DoM files can have just MDT stripe or combined MDT + OST + * stripes. + * - In the first case the open for read/write will do IO to + * the MDT stripe and it makes sense to take IO lock in + * advance along with OPEN even if it is blocking lock. + * - In the second case it is just size of MDT stripe and it + * is quite unlikely that client will write into it, though + * it may read it. So IO lock will be taken optionally if it + * is non-blocking one. + */ + if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL) + dom_stripes = mdt_lmm_dom_entry(ma->ma_lmm); + + if (dom_stripes == LMM_DOM_ONLY && + info->mti_mdt->mdt_opts.mo_dom_lock > 0 && + !mdt_dom_client_has_lock(info, mdt_object_fid(obj))) + dom_lock = true; } if (acq_lease) { @@ -1202,9 +854,10 @@ static int mdt_object_open_lock(struct mdt_thread_info *info, /* Lease must be with open lock */ if (!(open_flags & MDS_OPEN_LOCK)) { - CERROR("Request lease for file:"DFID ", but open lock " - "is missed, open_flags = "LPO64".\n", - PFID(mdt_object_fid(obj)), open_flags); + CERROR("%s: Request lease for file:"DFID ", but open lock " + "is missed, open_flags = %#llo : rc = %d\n", + mdt_obd_name(info->mti_mdt), + PFID(mdt_object_fid(obj)), open_flags, -EPROTO); GOTO(out, rc = -EPROTO); } @@ -1219,20 +872,16 @@ static int mdt_object_open_lock(struct mdt_thread_info *info, down_read(&obj->mot_open_sem); if (open_flags & MDS_OPEN_LOCK) { - if (open_flags & FMODE_WRITE) + if (open_flags & MDS_FMODE_WRITE) lm = LCK_CW; - /* if file is released, we can't deny write because we must - * restore (write) it to access it. */ - else if ((open_flags & MDS_FMODE_EXEC) && - !((ma->ma_valid & MA_HSM) && - (ma->ma_hsm.mh_flags & HS_RELEASED))) + else if (open_flags & MDS_FMODE_EXEC) lm = LCK_PR; else lm = LCK_CR; *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN; } else if (atomic_read(&obj->mot_lease_count) > 0) { - if (open_flags & FMODE_WRITE) + if (open_flags & MDS_FMODE_WRITE) lm = LCK_CW; else lm = LCK_CR; @@ -1242,51 +891,69 @@ static int mdt_object_open_lock(struct mdt_thread_info *info, try_layout = false; lhc = &info->mti_lh[MDT_LH_LOCAL]; + } else if (dom_lock) { + lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR; + if (info->mti_mdt->mdt_opts.mo_dom_lock == + TRYLOCK_DOM_ON_OPEN) { + trybits |= MDS_INODELOCK_DOM | + MDS_INODELOCK_LAYOUT; + } else { + /* mo_dom_lock == ALWAYS_DOM_LOCK_ON_OPEN */ + *ibits = MDS_INODELOCK_DOM; + if (info->mti_mdt->mdt_opts.mo_dom_read_open) { + trybits |= MDS_INODELOCK_LAYOUT; + } + } } + CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n", PFID(mdt_object_fid(obj)), - atomic_read(&obj->mot_open_count), lm); + atomic_read(&obj->mot_lease_count), lm); } mdt_lock_reg_init(lhc, lm); - /* one problem to return layout lock on open is that it may result - * in too many layout locks cached on the client side. */ + /* Return lookup lock to validate inode at the client side. + * This is pretty important otherwise MDT will return layout + * lock for each open. + * However this is a double-edged sword because changing + * permission will revoke a huge number of LOOKUP locks. + */ if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) { - /* return lookup lock to validate inode at the client side, - * this is pretty important otherwise mdt will return layout - * lock for each open. - * However this is a double-edged sword because changing - * permission will revoke huge # of LOOKUP locks. */ - *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP; - if (!mdt_object_lock_try(info, obj, lhc, *ibits, - MDT_CROSS_LOCK)) { - *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP); - if (*ibits != 0) - rc = mdt_object_lock(info, obj, lhc, *ibits, - MDT_CROSS_LOCK); - } - } else if (*ibits != 0) { - rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK); + if (!(*ibits & MDS_INODELOCK_LOOKUP)) + trybits |= MDS_INODELOCK_LOOKUP; + trybits |= MDS_INODELOCK_LAYOUT; } - CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64 - ", open_flags = "LPO64", try_layout = %d, rc = %d\n", - PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc); + if (*ibits | trybits) + rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false); + + CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx" + ", open_flags = %#llo, try_layout = %d : rc = %d\n", + mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)), + *ibits, trybits, open_flags, try_layout, rc); /* will change layout, revoke layout locks by enqueuing EX lock. */ if (rc == 0 && create_layout) { struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT]; CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID - ", open_flags = "LPO64"\n", + ", open_flags = %#llo\n", PFID(mdt_object_fid(obj)), open_flags); + /* We cannot enqueue another lock for the same resource we + * already have a lock for, due to mechanics of waiting list + * iterating in ldlm, see LU-3601. + * As such we'll drop the open lock we just got above here, + * it's ok not to have this open lock as it's main purpose is to + * flush unused cached client open handles. */ + if (lustre_handle_is_used(&lhc->mlh_reg_lh)) + mdt_object_unlock(info, obj, lhc, 1); + LASSERT(!try_layout); mdt_lock_handle_init(ll); mdt_lock_reg_init(ll, LCK_EX); - rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT, - MDT_LOCAL_LOCK); + rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT); OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2); } @@ -1311,11 +978,11 @@ static int mdt_object_open_lock(struct mdt_thread_info *info, is_replay_or_resent = req_is_replay(req) || lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT; - /* if the request is _not_ a replay request, rr_handle - * may be used to hold an openhandle which is issuing the + /* if the request is _not_ a replay request, rr_open_handle + * may be used to hold an open file handle which is issuing the * lease request, so that this openhandle doesn't count. */ - mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, - is_replay_or_resent); + mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle, + is_replay_or_resent); if (mfd != NULL) ++open_count; @@ -1362,21 +1029,23 @@ static void mdt_object_open_unlock(struct mdt_thread_info *info, up_read(&obj->mot_open_sem); /* Cross-ref case, the lock should be returned to the client */ - if (ibits == 0 || rc == -EREMOTE) + if (ibits == 0 || rc == -MDT_EREMOTE_OPEN) RETURN_EXIT; - if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) { + if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) && + !(ibits & MDS_INODELOCK_DOM)) { /* for the open request, the lock will only return to client * if open or layout lock is granted. */ rc = 1; } - if (rc != 0) { + if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) { struct ldlm_reply *ldlm_rep; ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP); mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK); - mdt_object_unlock(info, obj, lhc, 1); + if (lustre_handle_is_used(&lhc->mlh_reg_lh)) + mdt_object_unlock(info, obj, lhc, 1); } RETURN_EXIT; } @@ -1398,38 +1067,40 @@ static bool mdt_hsm_release_allow(const struct md_attr *ma) return true; } -int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep, - struct mdt_lock_handle *lhc) +static int mdt_open_by_fid_lock(struct mdt_thread_info *info, + struct ldlm_reply *rep, + struct mdt_lock_handle *lhc) { - const struct lu_env *env = info->mti_env; - struct mdt_device *mdt = info->mti_mdt; - __u64 flags = info->mti_spec.sp_cr_flags; - struct mdt_reint_record *rr = &info->mti_rr; - struct md_attr *ma = &info->mti_attr; - struct mdt_object *parent= NULL; - struct mdt_object *o; - int rc; - __u64 ibits = 0; - ENTRY; + const struct lu_env *env = info->mti_env; + struct mdt_device *mdt = info->mti_mdt; + u64 open_flags = info->mti_spec.sp_cr_flags; + struct mdt_reint_record *rr = &info->mti_rr; + struct md_attr *ma = &info->mti_attr; + struct mdt_object *parent = NULL; + struct mdt_object *o; + bool object_locked = false; + u64 ibits = 0; + int rc; - if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) { - if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) { - parent = mdt_object_find(env, mdt, rr->rr_fid1); - if (IS_ERR(parent)) { - CDEBUG(D_INODE, "Fail to find parent "DFID - " for anonymous created %ld, try to" - " use server-side parent.\n", - PFID(rr->rr_fid1), PTR_ERR(parent)); - parent = NULL; - } - } - if (parent == NULL) - ma->ma_need |= MA_PFID; - } + ENTRY; + if (md_should_create(open_flags)) { + if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) { + parent = mdt_object_find(env, mdt, rr->rr_fid1); + if (IS_ERR(parent)) { + CDEBUG(D_INODE, "Fail to find parent "DFID + " for anonymous created %ld, try to" + " use server-side parent.\n", + PFID(rr->rr_fid1), PTR_ERR(parent)); + parent = NULL; + } + } + if (parent == NULL) + ma->ma_need |= MA_PFID; + } - o = mdt_object_find(env, mdt, rr->rr_fid2); - if (IS_ERR(o)) - RETURN(rc = PTR_ERR(o)); + o = mdt_object_find(env, mdt, rr->rr_fid2); + if (IS_ERR(o)) + GOTO(out_parent_put, rc = PTR_ERR(o)); if (mdt_object_remote(o)) { CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n", @@ -1446,46 +1117,62 @@ int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep, mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD)); - if (flags & MDS_OPEN_RELEASE) + mdt_prep_ma_buf_from_rep(info, o, ma); + if (open_flags & MDS_OPEN_RELEASE) ma->ma_need |= MA_HSM; rc = mdt_attr_get_complex(info, o, ma); if (rc) GOTO(out, rc); - /* If a release request, check file flags are fine and ask for an + /* We should not change file's existing LOV EA */ + if (S_ISREG(lu_object_attr(&o->mot_obj)) && + open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV) + GOTO(out, rc = -EEXIST); + + /* If a release request, check file open flags are fine and ask for an * exclusive open access. */ - if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma)) + if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma)) GOTO(out, rc = -EPERM); - rc = mdt_object_open_lock(info, o, lhc, &ibits); - if (rc) - GOTO(out_unlock, rc); + rc = mdt_check_resent_lock(info, o, lhc); + if (rc < 0) { + GOTO(out, rc); + } else if (rc > 0) { + rc = mdt_object_open_lock(info, o, lhc, &ibits); + object_locked = true; + if (rc) + GOTO(out_unlock, rc); + } - if (ma->ma_valid & MA_PFID) { - parent = mdt_object_find(env, mdt, &ma->ma_pfid); - if (IS_ERR(parent)) { - CDEBUG(D_INODE, "Fail to find parent "DFID - " for anonymous created %ld, try to" - " use system default.\n", - PFID(&ma->ma_pfid), PTR_ERR(parent)); - parent = NULL; - } - } + if (ma->ma_valid & MA_PFID) { + parent = mdt_object_find(env, mdt, &ma->ma_pfid); + if (IS_ERR(parent)) { + CDEBUG(D_INODE, "Fail to find parent "DFID + " for anonymous created %ld, try to" + " use system default.\n", + PFID(&ma->ma_pfid), PTR_ERR(parent)); + parent = NULL; + } + } - rc = mdt_finish_open(info, parent, o, flags, 0, rep); + rc = mdt_finish_open(info, parent, o, open_flags, 0, rep); if (!rc) { mdt_set_disposition(info, rep, DISP_LOOKUP_POS); - if (flags & MDS_OPEN_LOCK) + if (open_flags & MDS_OPEN_LOCK) mdt_set_disposition(info, rep, DISP_OPEN_LOCK); - if (flags & MDS_OPEN_LEASE) + if (open_flags & MDS_OPEN_LEASE) mdt_set_disposition(info, rep, DISP_OPEN_LEASE); } GOTO(out_unlock, rc); out_unlock: - mdt_object_open_unlock(info, o, lhc, ibits, rc); + if (object_locked) + mdt_object_open_unlock(info, o, lhc, ibits, rc); out: mdt_object_put(env, o); + if (rc == 0) + mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh); +out_parent_put: if (parent != NULL) mdt_object_put(env, parent); return rc; @@ -1495,16 +1182,16 @@ out: static int mdt_cross_open(struct mdt_thread_info *info, const struct lu_fid *parent_fid, const struct lu_fid *fid, - struct ldlm_reply *rep, __u32 flags) + struct ldlm_reply *rep, u64 open_flags) { - struct md_attr *ma = &info->mti_attr; - struct mdt_object *o; - int rc; - ENTRY; + struct md_attr *ma = &info->mti_attr; + struct mdt_object *o; + int rc; + ENTRY; - o = mdt_object_find(info->mti_env, info->mti_mdt, fid); - if (IS_ERR(o)) - RETURN(rc = PTR_ERR(o)); + o = mdt_object_find(info->mti_env, info->mti_mdt, fid); + if (IS_ERR(o)) + RETURN(rc = PTR_ERR(o)); if (mdt_object_remote(o)) { /* Something is wrong here, the object is on another MDS! */ @@ -1512,30 +1199,30 @@ static int mdt_cross_open(struct mdt_thread_info *info, mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT); LU_OBJECT_DEBUG(D_WARNING, info->mti_env, &o->mot_obj, - "Object isn't on this server! FLD error?\n"); - rc = -EFAULT; + "Object isn't on this server! FLD error?"); + rc = -EFAULT; } else { if (mdt_object_exists(o)) { - /* Do permission check for cross-open. */ + int mask; + + /* Do permission check for cross-open after converting + * MDS_OPEN_* flags to MAY_* permission mask. + */ + mask = mds_accmode(open_flags); + rc = mo_permission(info->mti_env, NULL, - mdt_object_child(o), - NULL, flags | MDS_OPEN_CROSS); + mdt_object_child(o), NULL, mask); if (rc) goto out; - mdt_set_capainfo(info, 0, fid, BYPASS_CAPA); + mdt_prep_ma_buf_from_rep(info, o, ma); rc = mdt_attr_get_complex(info, o, ma); if (rc != 0) GOTO(out, rc); - /* Do not create lov object if the fid is opened - * under OBF */ - if (S_ISREG(ma->ma_attr.la_mode) && - !(ma->ma_valid & MA_LOV) && (flags & FMODE_WRITE) && - fid_is_obf(parent_fid)) - GOTO(out, rc = -EPERM); + mdt_pack_secctx_in_reply(info, o); - rc = mdt_finish_open(info, NULL, o, flags, 0, rep); + rc = mdt_finish_open(info, NULL, o, open_flags, 0, rep); } else { /* * Something is wrong here. lookup was positive but @@ -1545,220 +1232,294 @@ static int mdt_cross_open(struct mdt_thread_info *info, mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT); rc = -EFAULT; } - } + } out: - mdt_object_put(info->mti_env, o); - RETURN(rc); + mdt_object_put(info->mti_env, o); + + RETURN(rc); +} + +/* + * find root object and take its xattr lock if it's on remote MDT, later create + * may use fs default striping (which is stored in root xattr). + */ +static int mdt_lock_root_xattr(struct mdt_thread_info *info, + struct mdt_device *mdt) +{ + struct mdt_object *md_root = mdt->mdt_md_root; + struct lustre_handle lhroot; + int rc; + + if (md_root == NULL) { + lu_root_fid(&info->mti_tmp_fid1); + md_root = mdt_object_find(info->mti_env, mdt, + &info->mti_tmp_fid1); + if (IS_ERR(md_root)) + return PTR_ERR(md_root); + + spin_lock(&mdt->mdt_lock); + if (mdt->mdt_md_root != NULL) { + spin_unlock(&mdt->mdt_lock); + + LASSERTF(mdt->mdt_md_root == md_root, + "Different root object (" + DFID") instances, %p, %p\n", + PFID(&info->mti_tmp_fid1), + mdt->mdt_md_root, md_root); + LASSERT(atomic_read( + &md_root->mot_obj.lo_header->loh_ref) > 1); + + mdt_object_put(info->mti_env, md_root); + } else { + mdt->mdt_md_root = md_root; + spin_unlock(&mdt->mdt_lock); + } + } + + if (md_root->mot_cache_attr || !mdt_object_remote(md_root)) + return 0; + + rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root), + &lhroot, LCK_PR, MDS_INODELOCK_XATTR, + true); + if (rc < 0) + return rc; + + md_root->mot_cache_attr = 1; + + /* don't cancel this lock, so that we know the cached xattr is valid. */ + ldlm_lock_decref(&lhroot, LCK_PR); + + return 0; } int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) { - struct mdt_device *mdt = info->mti_mdt; - struct ptlrpc_request *req = mdt_info_req(info); - struct mdt_object *parent; - struct mdt_object *child; - struct mdt_lock_handle *lh; - struct ldlm_reply *ldlm_rep; - struct mdt_body *repbody; - struct lu_fid *child_fid = &info->mti_tmp_fid1; - struct md_attr *ma = &info->mti_attr; - __u64 create_flags = info->mti_spec.sp_cr_flags; - __u64 ibits = 0; - struct mdt_reint_record *rr = &info->mti_rr; - struct lu_name *lname; - int result, rc; - int created = 0; - __u32 msg_flags; - ENTRY; + struct mdt_device *mdt = info->mti_mdt; + struct ptlrpc_request *req = mdt_info_req(info); + struct mdt_object *parent; + struct mdt_object *child; + struct mdt_lock_handle *lh; + struct ldlm_reply *ldlm_rep; + struct mdt_body *repbody; + struct lu_fid *child_fid = &info->mti_tmp_fid1; + struct md_attr *ma = &info->mti_attr; + u64 open_flags = info->mti_spec.sp_cr_flags; + u64 ibits = 0; + struct mdt_reint_record *rr = &info->mti_rr; + int result, rc; + int created = 0; + int object_locked = 0; + u32 msg_flags; - OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE, - (obd_timeout + 1) / 4); + ENTRY; + OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE, + (obd_timeout + 1) / 4); mdt_counter_incr(req, LPROC_MDT_OPEN); - repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); - 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_need = MA_INODE; - if (ma->ma_lmm_size > 0) - ma->ma_need |= MA_LOV; + ma->ma_need = MA_INODE; + ma->ma_valid = 0; - ma->ma_valid = 0; + LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN); + ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP); - LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN); - ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP); + if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) { + CERROR("file join is not supported anymore.\n"); + GOTO(out, result = err_serious(-EOPNOTSUPP)); + } + msg_flags = lustre_msg_get_flags(req->rq_reqmsg); - if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) { - CERROR("file join is not supported anymore.\n"); - GOTO(out, result = err_serious(-EOPNOTSUPP)); - } - msg_flags = lustre_msg_get_flags(req->rq_reqmsg); + if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) && + info->mti_spec.u.sp_ea.eadata == NULL) + GOTO(out, result = err_serious(-EINVAL)); + + if (open_flags & MDS_FMODE_WRITE && + exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY) + GOTO(out, result = -EROFS); - if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) && - info->mti_spec.u.sp_ea.eadata == NULL) - GOTO(out, result = err_serious(-EINVAL)); + CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") " + "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n", + PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2), + open_flags, ma->ma_attr.la_mode, msg_flags); - CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") " - "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n", - PFID(rr->rr_fid1), rr->rr_name, - PFID(rr->rr_fid2), create_flags, - ma->ma_attr.la_mode, msg_flags); if (info->mti_cross_ref) { /* This is cross-ref open */ mdt_set_disposition(info, ldlm_rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD | DISP_LOOKUP_POS)); result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1, - ldlm_rep, create_flags); + ldlm_rep, open_flags); GOTO(out, result); - } else if (req_is_replay(req) || - (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) { - /* This is a replay request or from liblustre with ea. */ + } else if (req_is_replay(req)) { result = mdt_open_by_fid(info, ldlm_rep); - if (result != -ENOENT) { - if (req->rq_export->exp_libclient && - create_flags & MDS_OPEN_HAS_EA) - GOTO(out, result = 0); + if (result != -ENOENT) GOTO(out, result); - } + /* We didn't find the correct object, so we need to re-create it * via a regular replay. */ - if (!(create_flags & MDS_OPEN_CREAT)) { + if (!(open_flags & MDS_OPEN_CREAT)) { DEBUG_REQ(D_ERROR, req, "OPEN & CREAT not in open replay/by_fid."); GOTO(out, result = -EFAULT); } CDEBUG(D_INFO, "No object(1), continue as regular open.\n"); - } else if ((rr->rr_namelen == 0 && create_flags & MDS_OPEN_LOCK) || - (create_flags & MDS_OPEN_BY_FID)) { + } else if (open_flags & MDS_OPEN_BY_FID) { result = mdt_open_by_fid_lock(info, ldlm_rep, lhc); - /* If result is 0 then open by FID has found the file - * and there is nothing left for us to do here. More - * generally if it is anything other than -ENOENT or - * -EREMOTE then we return that now. If -ENOENT and - * MDS_OPEN_CREAT is set then we must create the file - * below. If -EREMOTE then we need to return a LOOKUP - * lock to the client, which we do below. Hence this - * odd looking condition. See LU-2523. */ - if (!(result == -ENOENT && (create_flags & MDS_OPEN_CREAT)) && - result != -EREMOTE) - GOTO(out, result); - - if (unlikely(rr->rr_namelen == 0)) - GOTO(out, result = -EINVAL); - - CDEBUG(D_INFO, "No object(2), continue as regular open.\n"); + if (result < 0) + CDEBUG(D_INFO, "no object for "DFID": %d\n", + PFID(rr->rr_fid2), result); + GOTO(out, result); } - if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) - GOTO(out, result = err_serious(-ENOMEM)); + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) + GOTO(out, result = err_serious(-ENOMEM)); - mdt_set_disposition(info, ldlm_rep, - (DISP_IT_EXECD | DISP_LOOKUP_EXECD)); + mdt_set_disposition(info, ldlm_rep, + (DISP_IT_EXECD | DISP_LOOKUP_EXECD)); - lh = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ? - LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen); + if (!lu_name_is_valid(&rr->rr_name)) + GOTO(out, result = -EPROTO); - parent = mdt_object_find_lock(info, rr->rr_fid1, lh, - MDS_INODELOCK_UPDATE); - if (IS_ERR(parent)) - GOTO(out, result = PTR_ERR(parent)); + result = mdt_lock_root_xattr(info, mdt); + if (result < 0) + GOTO(out, result); - /* get and check version of parent */ - result = mdt_version_get_check(info, parent, 0); - if (result) - GOTO(out_parent, result); +again: + lh = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_pdo_init(lh, (open_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR, + &rr->rr_name); - fid_zero(child_fid); + parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1); + if (IS_ERR(parent)) + GOTO(out, result = PTR_ERR(parent)); - lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen); - result = mdo_lookup(info->mti_env, mdt_object_child(parent), - lname, child_fid, &info->mti_spec); - LASSERTF(ergo(result == 0, fid_is_sane(child_fid)), - "looking for "DFID"/%s, result fid="DFID"\n", - PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid)); + result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE); + if (result != 0) { + mdt_object_put(info->mti_env, parent); + GOTO(out, result); + } - if (result != 0 && result != -ENOENT && result != -ESTALE) - GOTO(out_parent, result); + /* get and check version of parent */ + result = mdt_version_get_check(info, parent, 0); + if (result) + GOTO(out_parent, result); + + fid_zero(child_fid); + + result = -ENOENT; + if ((open_flags & MDS_OPEN_VOLATILE) == 0) + result = mdo_lookup(info->mti_env, mdt_object_child(parent), + &rr->rr_name, child_fid, &info->mti_spec); + + LASSERTF(ergo(result == 0, fid_is_sane(child_fid)), + "looking for "DFID"/"DNAME", found FID = "DFID"\n", + PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name), + PFID(child_fid)); + + if (result != 0 && result != -ENOENT && result != -ESTALE) + GOTO(out_parent, result); + + if (result == -ENOENT || result == -ESTALE) { + /* If the object is dead, let's check if the object + * is being migrated to a new object */ + if (result == -ESTALE) { + struct lu_buf lmv_buf; + + lmv_buf.lb_buf = info->mti_xattr_buf; + lmv_buf.lb_len = sizeof(info->mti_xattr_buf); + rc = mo_xattr_get(info->mti_env, + mdt_object_child(parent), + &lmv_buf, XATTR_NAME_LMV); + if (rc > 0) { + struct lmv_mds_md_v1 *lmv; + + lmv = lmv_buf.lb_buf; + if (le32_to_cpu(lmv->lmv_hash_type) & + LMV_HASH_FLAG_MIGRATION) { + /* Get the new parent FID and retry */ + mdt_object_unlock_put(info, parent, + lh, 1); + mdt_lock_handle_init(lh); + fid_le_to_cpu( + (struct lu_fid *)rr->rr_fid1, + &lmv->lmv_stripe_fids[1]); + goto again; + } + } + } - if (result == -ENOENT || result == -ESTALE) { - mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG); - if (result == -ESTALE) { - /* - * -ESTALE means the parent is a dead(unlinked) dir, so - * it should return -ENOENT to in accordance with the - * original mds implementaion. - */ - GOTO(out_parent, result = -ENOENT); - } - if (!(create_flags & MDS_OPEN_CREAT)) - GOTO(out_parent, result); - if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY) + mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG); + if (result == -ESTALE) { + /* + * -ESTALE means the parent is a dead(unlinked) dir, so + * it should return -ENOENT to in accordance with the + * original mds implementaion. + */ + GOTO(out_parent, result = -ENOENT); + } + + if (!(open_flags & MDS_OPEN_CREAT)) + GOTO(out_parent, result); + if (mdt_rdonly(req->rq_export)) GOTO(out_parent, result = -EROFS); - *child_fid = *info->mti_rr.rr_fid2; - LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n", - PFID(child_fid)); + *child_fid = *info->mti_rr.rr_fid2; + LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n", + PFID(child_fid)); /* In the function below, .hs_keycmp resolves to * lu_obj_hop_keycmp() */ /* coverity[overrun-buffer-val] */ child = mdt_object_new(info->mti_env, mdt, child_fid); } else { /* - * Check for O_EXCL is moved to the mdt_finish_open(), we need to - * return FID back in that case. + * Check for O_EXCL is moved to the mdt_finish_open(), + * we need to return FID back in that case. */ mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS); child = mdt_object_find(info->mti_env, mdt, child_fid); } - if (IS_ERR(child)) - GOTO(out_parent, result = PTR_ERR(child)); + if (IS_ERR(child)) + GOTO(out_parent, result = PTR_ERR(child)); - /** check version of child */ - rc = mdt_version_get_check(info, child, 1); - if (rc) - GOTO(out_child, result = rc); + /** check version of child */ + rc = mdt_version_get_check(info, child, 1); + if (rc) + GOTO(out_child, result = rc); - mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA); - if (result == -ENOENT) { + if (result == -ENOENT) { /* Create under OBF and .lustre is not permitted */ - if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1)) + if (!fid_is_md_operative(rr->rr_fid1) && + (open_flags & MDS_OPEN_VOLATILE) == 0) GOTO(out_child, result = -EPERM); - /* save versions in reply */ - mdt_version_get_save(info, parent, 0); - mdt_version_get_save(info, child, 1); + /* save versions in reply */ + mdt_version_get_save(info, parent, 0); + mdt_version_get_save(info, child, 1); - /* version of child will be changed */ - info->mti_mos = child; + /* version of child will be changed */ + tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child)); - /* Not found and with MDS_OPEN_CREAT: let's create it. */ - mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE); + /* Not found and with MDS_OPEN_CREAT: let's create it. */ + mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE); - /* Let lower layers know what is lock mode on directory. */ - info->mti_spec.sp_cr_mode = - mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode); + /* Let lower layers know what is lock mode on directory. */ + info->mti_spec.sp_cr_mode = + mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode); - /* - * 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; - - result = mdo_create(info->mti_env, - mdt_object_child(parent), - lname, - mdt_object_child(child), - &info->mti_spec, - &info->mti_attr); - if (result == -ERESTART) { - mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE); - GOTO(out_child, result); - } else { + /* Don't do lookup sanity check. We know name doesn't exist. */ + info->mti_spec.sp_cr_lookup = 0; + info->mti_spec.sp_feat = &dt_directory_features; + result = mdo_create(info->mti_env, mdt_object_child(parent), + &rr->rr_name, mdt_object_child(child), + &info->mti_spec, &info->mti_attr); + if (result == -ERESTART) { + mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE); + GOTO(out_child, result); + } else { + mdt_prep_ma_buf_from_rep(info, child, ma); /* XXX: we should call this once, see few lines below */ if (result == 0) result = mdt_attr_get_complex(info, child, ma); @@ -1766,7 +1527,8 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) if (result != 0) GOTO(out_child, result); } - created = 1; + created = 1; + mdt_counter_incr(req, LPROC_MDT_MKNOD); } else { /* * The object is on remote node, return its FID for remote open. @@ -1779,69 +1541,72 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) */ LASSERT(lhc != NULL); - if (lustre_handle_is_used(&lhc->mlh_reg_lh)) { - struct ldlm_lock *lock; - - LASSERT(msg_flags & MSG_RESENT); - - lock = ldlm_handle2lock(&lhc->mlh_reg_lh); - if (!lock) { - CERROR("Invalid lock handle "LPX64"\n", - lhc->mlh_reg_lh.cookie); - LBUG(); - } - LASSERT(fid_res_name_eq(mdt_object_fid(child), - &lock->l_resource->lr_name)); - LDLM_LOCK_PUT(lock); - rc = 0; - } else { - mdt_lock_handle_init(lhc); - mdt_lock_reg_init(lhc, LCK_PR); - - rc = mdt_object_lock(info, child, lhc, - MDS_INODELOCK_LOOKUP, - MDT_CROSS_LOCK); - } - repbody->fid1 = *mdt_object_fid(child); - repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); + rc = mdt_check_resent_lock(info, child, lhc); + if (rc < 0) { + GOTO(out_child, result = rc); + } else if (rc > 0) { + mdt_lock_handle_init(lhc); + mdt_lock_reg_init(lhc, LCK_PR); + + rc = mdt_object_lock(info, child, lhc, + MDS_INODELOCK_LOOKUP); + } + repbody->mbo_fid1 = *mdt_object_fid(child); + repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS); if (rc != 0) result = rc; else - result = -EREMOTE; + result = -MDT_EREMOTE_OPEN; GOTO(out_child, result); + } else if (mdt_object_exists(child)) { + /* We have to get attr & LOV EA & HSM for this + * object. */ + mdt_prep_ma_buf_from_rep(info, child, ma); + ma->ma_need |= MA_HSM; + result = mdt_attr_get_complex(info, child, ma); + if (result != 0) + GOTO(out_child, result); } else { - if (mdt_object_exists(child)) { - /* We have to get attr & LOV EA & HSM for this - * object */ - ma->ma_need |= MA_HSM; - result = mdt_attr_get_complex(info, child, ma); - } else { - /*object non-exist!!!*/ - LBUG(); - } + /* Object does not exist. Likely FS corruption. */ + CERROR("%s: name '"DNAME"' present, but FID " + DFID" is invalid\n", mdt_obd_name(info->mti_mdt), + PNAME(&rr->rr_name), PFID(child_fid)); + GOTO(out_child, result = -EIO); } - } + } - LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh)); + mdt_pack_secctx_in_reply(info, child); - /* get openlock if this is not replay and if a client requested it */ - if (!req_is_replay(req)) { - rc = mdt_object_open_lock(info, child, lhc, &ibits); - if (rc != 0) - GOTO(out_child_unlock, result = rc); - else if (create_flags & MDS_OPEN_LOCK) + rc = mdt_check_resent_lock(info, child, lhc); + if (rc < 0) { + GOTO(out_child, result = rc); + } else if (rc == 0) { + /* the open lock might already be gotten in + * ldlm_handle_enqueue() */ + LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT); + if (open_flags & MDS_OPEN_LOCK) mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK); + } else { + /* get openlock if this isn't replay and client requested it */ + if (!req_is_replay(req)) { + rc = mdt_object_open_lock(info, child, lhc, &ibits); + object_locked = 1; + if (rc != 0) + GOTO(out_child_unlock, result = rc); + else if (open_flags & MDS_OPEN_LOCK) + mdt_set_disposition(info, ldlm_rep, + DISP_OPEN_LOCK); + } } - /* Try to open it now. */ - rc = mdt_finish_open(info, parent, child, create_flags, + rc = mdt_finish_open(info, parent, child, open_flags, created, ldlm_rep); if (rc) { result = rc; - /* openlock will be released if mdt_finish_open failed */ + /* openlock will be released if mdt_finish_open() failed */ mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK); - if (created && create_flags & MDS_OPEN_VOLATILE) { + if (created && (open_flags & MDS_OPEN_VOLATILE)) { CERROR("%s: cannot open volatile file "DFID", orphan " "file will be left in PENDING directory until " "next reboot, rc = %d\n", mdt_obd_name(mdt), @@ -1852,11 +1617,10 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) if (created) { ma->ma_need = 0; ma->ma_valid = 0; - ma->ma_cookie_size = 0; rc = mdo_unlink(info->mti_env, mdt_object_child(parent), mdt_object_child(child), - lname, + &rr->rr_name, &info->mti_attr, 0); if (rc != 0) CERROR("%s: "DFID" cleanup of open: rc = %d\n", @@ -1867,9 +1631,12 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) } EXIT; out_child_unlock: - mdt_object_open_unlock(info, child, lhc, ibits, result); + if (object_locked) + mdt_object_open_unlock(info, child, lhc, ibits, result); out_child: mdt_object_put(info->mti_env, child); + if (result == 0) + mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh); out_parent: mdt_object_unlock_put(info, parent, lh, result || !created); out: @@ -1891,8 +1658,10 @@ static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info, struct lu_fid *local_root_fid = &info->mti_tmp_fid1; struct mdt_object *obj = NULL; struct mdt_object *local_root; - static const char name[] = "i_am_nobody"; - struct lu_name *lname; + static const struct lu_name lname = { + .ln_name = "i_am_nobody", + .ln_namelen = sizeof("i_am_nobody") - 1, + }; struct lu_ucred *uc; cfs_cap_t uc_cap_save; int rc; @@ -1922,12 +1691,10 @@ static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info, spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE; } - lname = mdt_name(env, (char *)name, sizeof(name) - 1); - uc = lu_ucred(env); uc_cap_save = uc->uc_cap; uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE; - rc = mdo_create(env, mdt_object_child(local_root), lname, + rc = mdo_create(env, mdt_object_child(local_root), &lname, mdt_object_child(obj), spec, attr); uc->uc_cap = uc_cap_save; if (rc < 0) { @@ -1936,7 +1703,7 @@ static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info, GOTO(out, rc); } - rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED); + rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED, spec); if (rc < 0) CERROR("%s: cannot open volatile file "DFID", orphan " "file will be left in PENDING directory until " @@ -1954,21 +1721,64 @@ out: return obj; } +/* XXX Look into layout in MDT layer. */ +static inline int mdt_hsm_set_released(struct lov_mds_md *lmm) +{ + struct lov_comp_md_v1 *comp_v1; + struct lov_mds_md *v1; + __u32 off; + int i; + + if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) { + comp_v1 = (struct lov_comp_md_v1 *)lmm; + + if (comp_v1->lcm_entry_count == 0) + return -EINVAL; + + for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) { + off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset); + v1 = (struct lov_mds_md *)((char *)comp_v1 + off); + v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED); + } + } else { + lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED); + } + return 0; +} + +static inline int mdt_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen) +{ + struct lov_comp_md_v1 *comp_v1; + + if (le32_to_cpu(lmm->lmm_magic == LOV_MAGIC_COMP_V1)) { + comp_v1 = (struct lov_comp_md_v1 *)lmm; + *gen = le32_to_cpu(comp_v1->lcm_layout_gen); + } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 || + le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) { + *gen = le16_to_cpu(lmm->lmm_layout_gen); + } else { + return -EINVAL; + } + return 0; +} + static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, struct md_attr *ma) { struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT]; + struct lu_ucred *uc = mdt_ucred(info); struct close_data *data; struct ldlm_lock *lease; struct mdt_object *orphan; struct md_attr *orp_ma; struct lu_buf *buf; + cfs_cap_t cap; bool lease_broken; int rc; int rc2; ENTRY; - if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY) + if (mdt_rdonly(info->mti_exp)) RETURN(-EROFS); data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA); @@ -1990,7 +1800,7 @@ static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, lease_broken = ldlm_is_cancel(lease); unlock_res_and_lock(lease); - LDLM_DEBUG(lease, DFID " lease broken? %d\n", + LDLM_DEBUG(lease, DFID " lease broken? %d", PFID(mdt_object_fid(o)), lease_broken); /* Cancel server side lease. Client side counterpart should @@ -2011,19 +1821,66 @@ static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, if (rc != 0) GOTO(out_unlock, rc); - if (!mdt_hsm_release_allow(ma)) - GOTO(out_unlock, rc = -EPERM); - - /* already released? */ - if (ma->ma_hsm.mh_flags & HS_RELEASED) - GOTO(out_unlock, rc = 0); + if (ma->ma_attr_flags & MDS_PCC_ATTACH) { + if (ma->ma_valid & MA_HSM) { + if (ma->ma_hsm.mh_flags & HS_RELEASED) + GOTO(out_unlock, rc = -EALREADY); + + if (ma->ma_hsm.mh_arch_id != data->cd_archive_id) + CDEBUG(D_CACHE, + DFID" archive id diff: %llu:%u\n", + PFID(mdt_object_fid(o)), + ma->ma_hsm.mh_arch_id, + data->cd_archive_id); + + if (!(ma->ma_hsm.mh_flags & HS_DIRTY) && + ma->ma_hsm.mh_arch_ver == data->cd_data_version) { + CDEBUG(D_CACHE, + DFID" data version matches: packed=%llu " + "and on-disk=%llu\n", + PFID(mdt_object_fid(o)), + data->cd_data_version, + ma->ma_hsm.mh_arch_ver); + ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS; + } - /* Compare on-disk and packed data_version */ - if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) { - CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64 - " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)), - data->cd_data_version, ma->ma_hsm.mh_arch_ver); - GOTO(out_unlock, rc = -EPERM); + if (ma->ma_hsm.mh_flags & HS_DIRTY) + ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS; + } else { + /* Set up HSM attribte for PCC archived object */ + BUILD_BUG_ON(sizeof(struct hsm_attrs) > + sizeof(info->mti_xattr_buf)); + buf = &info->mti_buf; + buf->lb_buf = info->mti_xattr_buf; + buf->lb_len = sizeof(struct hsm_attrs); + memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm)); + ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS; + ma->ma_hsm.mh_arch_id = data->cd_archive_id; + ma->ma_hsm.mh_arch_ver = data->cd_data_version; + lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm); + + rc = mo_xattr_set(info->mti_env, mdt_object_child(o), + buf, XATTR_NAME_HSM, 0); + if (rc) + GOTO(out_unlock, rc); + } + } else { + if (!mdt_hsm_release_allow(ma)) + GOTO(out_unlock, rc = -EPERM); + + /* already released? */ + if (ma->ma_hsm.mh_flags & HS_RELEASED) + GOTO(out_unlock, rc = 0); + + /* Compare on-disk and packed data_version */ + if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) { + CDEBUG(D_HSM, DFID" data_version mismatches: " + "packed=%llu and on-disk=%llu\n", + PFID(mdt_object_fid(o)), + data->cd_data_version, + ma->ma_hsm.mh_arch_ver); + GOTO(out_unlock, rc = -EPERM); + } } ma->ma_valid = MA_INODE; @@ -2032,6 +1889,17 @@ static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, if (rc < 0) GOTO(out_unlock, rc); + mutex_lock(&o->mot_som_mutex); + rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size, + ma->ma_attr.la_blocks); + mutex_unlock(&o->mot_som_mutex); + if (rc2 < 0) + CDEBUG(D_INODE, + "%s: File "DFID" SOM update failed: rc = %d\n", + mdt_obd_name(info->mti_mdt), + PFID(mdt_object_fid(o)), rc2); + + ma->ma_need = MA_INODE | MA_LOV; rc = mdt_attr_get_complex(info, o, ma); if (rc < 0) @@ -2040,36 +1908,38 @@ static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, if (!(ma->ma_valid & MA_LOV)) { /* Even empty file are released */ memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm)); - ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF); + ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED); ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0); ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE); ma->ma_lmm_size = sizeof(*ma->ma_lmm); } else { - /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret + /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret * ma_lmm as lov_user_md, then it will be confused by union of * layout_gen and stripe_offset. */ - if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1) - ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF); - else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3) - ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF); + if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) == + LOV_MAGIC_MAGIC) + ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED); else GOTO(out_unlock, rc = -EINVAL); } - /* Set file as released */ - ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED); + /* Set file as released. */ + rc = mdt_hsm_set_released(ma->ma_lmm); + if (rc) + GOTO(out_unlock, rc); - /* Hopefully it's not used in this call path */ - orp_ma = &info->mti_u.som.attr; + orp_ma = &info->mti_u.hsm.attr; orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR; - orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid; - orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid; + /* We use root ownership to bypass potential quota + * restrictions on the user and group of the file. */ + orp_ma->ma_attr.la_uid = 0; + orp_ma->ma_attr.la_gid = 0; orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID; orp_ma->ma_lmm = ma->ma_lmm; orp_ma->ma_lmm_size = ma->ma_lmm_size; orp_ma->ma_valid = MA_INODE | MA_LOV; orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma, - FMODE_WRITE); + MDS_FMODE_WRITE); if (IS_ERR(orphan)) { CERROR("%s: cannot open orphan file "DFID": rc = %ld\n", mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), @@ -2078,7 +1948,7 @@ static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, } /* Set up HSM attribute for orphan object */ - CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf)); + BUILD_BUG_ON(sizeof(struct hsm_attrs) > sizeof(info->mti_xattr_buf)); buf = &info->mti_buf; buf->lb_buf = info->mti_xattr_buf; buf->lb_len = sizeof(struct hsm_attrs); @@ -2088,47 +1958,67 @@ static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, mdt_lock_reg_init(lh, LCK_EX); rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT | - MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK); + MDS_INODELOCK_XATTR); if (rc != 0) GOTO(out_close, rc); + /* The orphan has root ownership so we need to raise + * CAP_FOWNER to set the HSM attributes. */ + cap = uc->uc_cap; + uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER); rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf, XATTR_NAME_HSM, 0); + uc->uc_cap = cap; + if (rc != 0) + GOTO(out_layout_lock, rc); - if (rc == 0) - /* Swap layout with orphan object */ - rc = mo_swap_layouts(info->mti_env, mdt_object_child(o), - mdt_object_child(orphan), - SWAP_LAYOUTS_MDS_HSM); + /* Swap layout with orphan objects. */ + rc = mo_swap_layouts(info->mti_env, mdt_object_child(o), + mdt_object_child(orphan), + SWAP_LAYOUTS_MDS_HSM); - /* Release exclusive LL */ - mdt_object_unlock(info, o, lh, 1); + if (!rc && ma->ma_attr_flags & MDS_PCC_ATTACH) { + ma->ma_need = MA_LOV; + rc = mdt_attr_get_complex(info, o, ma); + } EXIT; +out_layout_lock: + /* Release exclusive LL */ + mdt_object_unlock(info, o, lh, 1); out_close: /* Close orphan object anyway */ rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma, - FMODE_WRITE); + MDS_FMODE_WRITE); if (rc2 < 0) CERROR("%s: error closing volatile file "DFID": rc = %d\n", mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2); LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj, - "object closed\n"); + "object closed"); mdt_object_put(info->mti_env, orphan); out_unlock: up_write(&o->mot_open_sem); - if (rc == 0) { /* already released */ + /* already released */ + if (rc == 0) { struct mdt_body *repbody; + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); LASSERT(repbody != NULL); - repbody->valid |= OBD_MD_FLRELEASED; + repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED; + if (ma->ma_attr_flags & MDS_PCC_ATTACH) { + LASSERT(ma->ma_valid & MA_LOV); + rc = mdt_get_lmm_gen(ma->ma_lmm, + &repbody->mbo_layout_gen); + if (!rc) + repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION; + } } out_reprocess: - ldlm_reprocess_all(lease->l_resource); + ldlm_reprocess_all(lease->l_resource, lease); LDLM_LOCK_PUT(lease); ma->ma_valid = 0; @@ -2137,50 +2027,366 @@ out_reprocess: return rc; } -#define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \ - MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED) +int mdt_close_handle_layouts(struct mdt_thread_info *info, + struct mdt_object *o, struct md_attr *ma) +{ + struct mdt_lock_handle *lh1 = &info->mti_lh[MDT_LH_NEW]; + struct mdt_lock_handle *lh2 = &info->mti_lh[MDT_LH_OLD]; + struct close_data *data; + struct ldlm_lock *lease; + struct mdt_object *o1 = o, *o2; + bool lease_broken; + bool swap_objects; + int rc; + ENTRY; + + if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY) + RETURN(-EROFS); + + if (!S_ISREG(lu_object_attr(&o1->mot_obj))) + RETURN(-EINVAL); + + data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA); + if (data == NULL) + RETURN(-EPROTO); + + if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid)) + RETURN(-EINVAL); + + rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o)); + if (unlikely(rc == 0)) + RETURN(-EINVAL); + + /* Exchange o1 and o2, to enforce locking order */ + swap_objects = (rc < 0); + + lease = ldlm_handle2lock(&data->cd_handle); + if (lease == NULL) + RETURN(-ESTALE); + + o2 = mdt_object_find(info->mti_env, info->mti_mdt, &data->cd_fid); + if (IS_ERR(o2)) + GOTO(out_lease, rc = PTR_ERR(o2)); + + if (!S_ISREG(lu_object_attr(&o2->mot_obj))) { + swap_objects = false; /* not swapped yet */ + GOTO(out_obj, rc = -EINVAL); + } + + if (swap_objects) + swap(o1, o2); + + rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL, + MAY_WRITE); + if (rc < 0) + GOTO(out_obj, rc); + + rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL, + MAY_WRITE); + if (rc < 0) + GOTO(out_obj, rc); + + /* try to hold open_sem so that nobody else can open the file */ + if (!down_write_trylock(&o->mot_open_sem)) { + ldlm_lock_cancel(lease); + GOTO(out_obj, rc = -EBUSY); + } + + /* Check if the lease open lease has already canceled */ + lock_res_and_lock(lease); + lease_broken = ldlm_is_cancel(lease); + unlock_res_and_lock(lease); + + LDLM_DEBUG(lease, DFID " lease broken? %d", + PFID(mdt_object_fid(o)), lease_broken); + + /* Cancel server side lease. Client side counterpart should + * have been cancelled. It's okay to cancel it now as we've + * held mot_open_sem. */ + ldlm_lock_cancel(lease); + + if (lease_broken) + GOTO(out_unlock_sem, rc = -ESTALE); + + mdt_lock_reg_init(lh1, LCK_EX); + rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT | + MDS_INODELOCK_XATTR); + if (rc < 0) + GOTO(out_unlock_sem, rc); + + mdt_lock_reg_init(lh2, LCK_EX); + rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT | + MDS_INODELOCK_XATTR); + if (rc < 0) + GOTO(out_unlock1, rc); + + /* Swap layout with orphan object */ + if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) { + rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1), + mdt_object_child(o2), 0); + } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE || + ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) { + struct lu_buf *buf = &info->mti_buf; + struct md_rejig_data mrd; + + mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1); + if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) + mrd.mrd_mirror_id = data->cd_mirror_id; + + buf->lb_len = sizeof(mrd); + buf->lb_buf = &mrd; + rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf, + XATTR_LUSTRE_LOV, + ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ? + LU_XATTR_SPLIT : LU_XATTR_MERGE); + if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) { + int rc2; + + mutex_lock(&o->mot_som_mutex); + rc2 = mdt_set_som(info, o, SOM_FL_STRICT, + ma->ma_attr.la_size, + ma->ma_attr.la_blocks); + mutex_unlock(&o->mot_som_mutex); + if (rc2 < 0) + CERROR(DFID": Setting i_blocks error: %d, " + "i_blocks will be reported wrongly and " + "can only be fixed in next resync\n", + PFID(mdt_object_fid(o)), rc2); + } + } + if (rc < 0) + GOTO(out_unlock2, rc); + + EXIT; + +out_unlock2: + /* Release exclusive LL */ + mdt_object_unlock(info, o2, lh2, 1); + +out_unlock1: + mdt_object_unlock(info, o1, lh1, 1); + +out_unlock_sem: + up_write(&o->mot_open_sem); + + /* already swapped */ + if (rc == 0) { + struct mdt_body *repbody; + + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + LASSERT(repbody != NULL); + repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED; + } + +out_obj: + mdt_object_put(info->mti_env, swap_objects ? o1 : o2); + + ldlm_reprocess_all(lease->l_resource, lease); + +out_lease: + LDLM_LOCK_PUT(lease); + + if (ma != NULL) { + ma->ma_valid = 0; + ma->ma_need = 0; + } + + return rc; +} + +static int mdt_close_resync_done(struct mdt_thread_info *info, + struct mdt_object *o, struct md_attr *ma) +{ + struct close_data *data; + struct ldlm_lock *lease; + struct md_layout_change layout = { 0 }; + __u32 *resync_ids = NULL; + size_t resync_count = 0; + bool lease_broken; + int rc; + ENTRY; + + if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY) + RETURN(-EROFS); + + if (!S_ISREG(lu_object_attr(&o->mot_obj))) + RETURN(-EINVAL); + + data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA); + if (data == NULL) + RETURN(-EPROTO); + + if (ptlrpc_req_need_swab(mdt_info_req(info))) + lustre_swab_close_data_resync_done(&data->cd_resync); + + if (!fid_is_zero(&data->cd_fid)) + RETURN(-EPROTO); + + lease = ldlm_handle2lock(&data->cd_handle); + if (lease == NULL) + RETURN(-ESTALE); + + /* try to hold open_sem so that nobody else can open the file */ + if (!down_write_trylock(&o->mot_open_sem)) { + ldlm_lock_cancel(lease); + GOTO(out_reprocess, rc = -EBUSY); + } + + /* Check if the lease open lease has already canceled */ + lock_res_and_lock(lease); + lease_broken = ldlm_is_cancel(lease); + unlock_res_and_lock(lease); + + LDLM_DEBUG(lease, DFID " lease broken? %d\n", + PFID(mdt_object_fid(o)), lease_broken); + + /* Cancel server side lease. Client side counterpart should + * have been cancelled. It's okay to cancel it now as we've + * held mot_open_sem. */ + ldlm_lock_cancel(lease); + + if (lease_broken) /* don't perform release task */ + GOTO(out_unlock, rc = -ESTALE); + + resync_count = data->cd_resync.resync_count; + + if (resync_count > INLINE_RESYNC_ARRAY_SIZE) { + void *data; + + if (!req_capsule_has_field(info->mti_pill, &RMF_U32, + RCL_CLIENT)) + GOTO(out_unlock, rc = -EPROTO); + + OBD_ALLOC(resync_ids, resync_count * sizeof(__u32)); + if (!resync_ids) + GOTO(out_unlock, rc = -ENOMEM); + + data = req_capsule_client_get(info->mti_pill, &RMF_U32); + memcpy(resync_ids, data, resync_count * sizeof(__u32)); + + layout.mlc_resync_ids = resync_ids; + } else { + layout.mlc_resync_ids = data->cd_resync.resync_ids_inline; + } + + layout.mlc_opc = MD_LAYOUT_RESYNC_DONE; + layout.mlc_resync_count = resync_count; + if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) { + layout.mlc_som.lsa_valid = SOM_FL_STRICT; + layout.mlc_som.lsa_size = ma->ma_attr.la_size; + layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks; + } + rc = mdt_layout_change(info, o, &layout); + if (rc) + GOTO(out_unlock, rc); + + EXIT; + +out_unlock: + up_write(&o->mot_open_sem); + + /* already released */ + if (rc == 0) { + struct mdt_body *repbody; + + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + LASSERT(repbody != NULL); + repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED; + } + + if (resync_ids) + OBD_FREE(resync_ids, resync_count * sizeof(__u32)); + +out_reprocess: + ldlm_reprocess_all(lease->l_resource, lease); + LDLM_LOCK_PUT(lease); + + ma->ma_valid = 0; + ma->ma_need = 0; + + return rc; +} + +#define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED) static int mdt_mfd_closed(struct mdt_file_data *mfd) { - return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode)); + return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags)); } int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd) { - struct mdt_object *o = mfd->mfd_object; - struct md_object *next = mdt_object_child(o); - struct md_attr *ma = &info->mti_attr; - int ret = MDT_IOEPOCH_CLOSED; - int rc = 0; - __u64 mode; - ENTRY; + struct mdt_object *o = mfd->mfd_object; + struct md_object *next = mdt_object_child(o); + struct md_attr *ma = &info->mti_attr; + struct lu_fid *ofid = &info->mti_tmp_fid1; + int rc = 0; + u64 open_flags; + u64 intent; - mode = mfd->mfd_mode; + ENTRY; - if (ma->ma_attr_flags & MDS_HSM_RELEASE) { + open_flags = mfd->mfd_open_flags; + intent = ma->ma_attr_flags & MDS_CLOSE_INTENT; + *ofid = *mdt_object_fid(o); + + CDEBUG(D_INODE, "%s: close file "DFID" with intent: %llx\n", + mdt_obd_name(info->mti_mdt), PFID(ofid), intent); + + switch (intent) { + case MDS_HSM_RELEASE: { rc = mdt_hsm_release(info, o, ma); if (rc < 0) { CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n", - mdt_obd_name(info->mti_mdt), - PFID(mdt_object_fid(o)), rc); + mdt_obd_name(info->mti_mdt), + PFID(ofid), rc); /* continue to close even error occurred. */ } + break; + } + case MDS_CLOSE_LAYOUT_MERGE: + case MDS_CLOSE_LAYOUT_SPLIT: + case MDS_CLOSE_LAYOUT_SWAP: { + rc = mdt_close_handle_layouts(info, o, ma); + if (rc < 0) { + CDEBUG(D_INODE, + "%s: cannot swap layout of "DFID": rc = %d\n", + mdt_obd_name(info->mti_mdt), + PFID(ofid), rc); + /* continue to close even if error occurred. */ + } + break; + } + case MDS_CLOSE_RESYNC_DONE: + rc = mdt_close_resync_done(info, o, ma); + break; + default: + /* nothing */ + break; } - if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) { - mdt_write_put(o); - ret = mdt_ioepoch_close(info, o); - } else if (mode & MDS_FMODE_EXEC) { - mdt_write_allow(o); - } else if (mode & MDS_FMODE_EPOCH) { - ret = mdt_ioepoch_close(info, o); - } else if (mode & MDS_FMODE_SOM) { - ret = mdt_som_au_close(info, o); - } + if (S_ISREG(lu_object_attr(&o->mot_obj)) && + ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) { + int rc2; + + rc2 = mdt_lsom_update(info, o, false); + if (rc2 < 0) + CDEBUG(D_INODE, + "%s: File " DFID " LSOM failed: rc = %d\n", + mdt_obd_name(info->mti_mdt), + PFID(ofid), rc2); + /* continue to close even if error occured. */ + } + + if (open_flags & MDS_FMODE_WRITE) + mdt_write_put(o); + else if (open_flags & MDS_FMODE_EXEC) + mdt_write_allow(o); /* Update atime on close only. */ - if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE) - && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) { + if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ || + open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) && + (ma->ma_attr.la_valid & LA_ATIME)) { /* Set the atime only. */ ma->ma_valid = MA_INODE; ma->ma_attr.la_valid = LA_ATIME; @@ -2194,43 +2400,53 @@ int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd) ma->ma_need |= MA_INODE; ma->ma_valid &= ~MA_INODE; - if (!MFD_CLOSED(mode)) - rc = mo_close(info->mti_env, next, ma, mode); + LASSERT(atomic_read(&o->mot_open_count) > 0); + atomic_dec(&o->mot_open_count); + mdt_handle_last_unlink(info, o, ma); - if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) { - struct mdt_export_data *med; + if (!MFD_CLOSED(open_flags)) { + rc = mo_close(info->mti_env, next, ma, open_flags); + if (mdt_dom_check_for_discard(info, o)) + mdt_dom_discard_data(info, o); + } - /* The IOepoch is still opened or SOM update is needed. - * Put mfd back into the list. */ - LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM); - mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ? - MDS_FMODE_EPOCH : MDS_FMODE_SOM); + /* adjust open and lease count */ + if (open_flags & MDS_OPEN_LEASE) { + LASSERT(atomic_read(&o->mot_lease_count) > 0); + atomic_dec(&o->mot_lease_count); + } - LASSERT(mdt_info_req(info)); - med = &mdt_info_req(info)->rq_export->exp_mdt_data; - spin_lock(&med->med_open_lock); - cfs_list_add(&mfd->mfd_list, &med->med_open_head); - class_handle_hash_back(&mfd->mfd_handle); - spin_unlock(&med->med_open_lock); + mdt_mfd_free(mfd); + mdt_object_put(info->mti_env, o); - if (ret == MDT_IOEPOCH_OPENED) { - ret = 0; - } else { - ret = -EAGAIN; - CDEBUG(D_INODE, "Size-on-MDS attribute update is " - "needed on "DFID"\n", PFID(mdt_object_fid(o))); - } - } else { - /* adjust open and lease count */ - if (mode & MDS_OPEN_LEASE) { - LASSERT(atomic_read(&o->mot_lease_count) > 0); - atomic_dec(&o->mot_lease_count); - } - LASSERT(atomic_read(&o->mot_open_count) > 0); - atomic_dec(&o->mot_open_count); + RETURN(rc); +} - mdt_mfd_free(mfd); - mdt_object_put(info->mti_env, o); +int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req, + struct mdt_body *repbody) +{ + struct mdt_export_data *med; + struct mdt_file_data *mfd; + int ret = 0; + int rc = 0; + ENTRY; + + med = &req->rq_export->exp_mdt_data; + spin_lock(&med->med_open_lock); + mfd = mdt_open_handle2mfd(med, &info->mti_open_handle, + req_is_replay(req)); + if (mdt_mfd_closed(mfd)) { + spin_unlock(&med->med_open_lock); + CDEBUG(D_INODE, "no handle for file close: fid = "DFID + ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1), + info->mti_open_handle.cookie); + /** not serious error since bug 3633 */ + rc = -ESTALE; + } else { + class_handle_unhash(&mfd->mfd_open_handle); + list_del_init(&mfd->mfd_list); + spin_unlock(&med->med_open_lock); + ret = mdt_mfd_close(info, mfd); } RETURN(rc ? rc : ret); @@ -2240,9 +2456,6 @@ int mdt_close(struct tgt_session_info *tsi) { struct mdt_thread_info *info = tsi2mdt_info(tsi); struct ptlrpc_request *req = tgt_ses_req(tsi); - struct mdt_export_data *med; - struct mdt_file_data *mfd; - struct mdt_object *o; struct md_attr *ma = &info->mti_attr; struct mdt_body *repbody = NULL; int rc, ret = 0; @@ -2254,12 +2467,12 @@ int mdt_close(struct tgt_session_info *tsi) if (rc) GOTO(out, rc = err_serious(rc)); - LASSERT(info->mti_ioepoch); - + /* These fields are no longer used and are left for compatibility. + * size is always zero */ req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER, - info->mti_mdt->mdt_max_mdsize); + 0); req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER, - info->mti_mdt->mdt_max_cookiesize); + 0); rc = req_capsule_server_pack(info->mti_pill); if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) { mdt_client_compatibility(info); @@ -2278,43 +2491,17 @@ int mdt_close(struct tgt_session_info *tsi) 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; - repbody->eadatasize = 0; - repbody->aclsize = 0; - } else { - rc = err_serious(rc); - } - - med = &req->rq_export->exp_mdt_data; - spin_lock(&med->med_open_lock); - mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle, - req_is_replay(req)); - if (mdt_mfd_closed(mfd)) { - 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); - /** not serious error since bug 3633 */ - rc = -ESTALE; + ma->ma_need = MA_INODE | MA_LOV; + repbody->mbo_eadatasize = 0; + repbody->mbo_aclsize = 0; } else { - class_handle_unhash(&mfd->mfd_handle); - cfs_list_del_init(&mfd->mfd_list); - spin_unlock(&med->med_open_lock); + rc = err_serious(rc); + } + + rc = mdt_close_internal(info, req, repbody); + if (rc != -ESTALE) + mdt_empty_transno(info, rc); - /* Do not lose object before last unlink. */ - o = mfd->mfd_object; - mdt_object_get(info->mti_env, o); - ret = mdt_mfd_close(info, mfd); - if (repbody != NULL) - rc = mdt_handle_last_unlink(info, o, ma); - mdt_empty_transno(info, rc); - mdt_object_put(info->mti_env, o); - } if (repbody != NULL) { mdt_client_compatibility(info); rc = mdt_fix_reply(info); @@ -2331,86 +2518,3 @@ out: mdt_thread_info_fini(info); RETURN(rc ? rc : ret); } - -/** - * DONE_WRITING rpc handler. - * - * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()), - * only those DONE_WRITING rpc will be replayed which really wrote smth on disk, - * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just - * skip attributes and reconstruct the reply here. - */ -int mdt_done_writing(struct tgt_session_info *tsi) -{ - struct ptlrpc_request *req = tgt_ses_req(tsi); - struct mdt_thread_info *info = tsi2mdt_info(tsi); - struct mdt_body *repbody = NULL; - struct mdt_export_data *med; - struct mdt_file_data *mfd; - int rc; - ENTRY; - - rc = req_capsule_server_pack(tsi->tsi_pill); - if (rc) - GOTO(out, rc = err_serious(rc)); - - repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY); - repbody->eadatasize = 0; - repbody->aclsize = 0; - - /* Done Writing may come with the Size-on-MDS update. Unpack it. */ - rc = mdt_close_unpack(info); - if (rc) - GOTO(out, rc = err_serious(rc)); - - if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) { - mdt_exit_ucred(info); - GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg)); - } - - med = &info->mti_exp->exp_mdt_data; - 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 done write: fid = "DFID - ": cookie = "LPX64" ioepoch = "LPU64"\n", - PFID(info->mti_rr.rr_fid1), - info->mti_ioepoch->handle.cookie, - info->mti_ioepoch->ioepoch); - /* If this is a replay, reconstruct the transno. */ - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { - rc = info->mti_ioepoch->flags & MF_SOM_AU ? - -EAGAIN : 0; - mdt_empty_transno(info, rc); - } else - rc = -ESTALE; - GOTO(error_ucred, rc); - } - - LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH || - mfd->mfd_mode == MDS_FMODE_TRUNC); - class_handle_unhash(&mfd->mfd_handle); - cfs_list_del_init(&mfd->mfd_list); - spin_unlock(&med->med_open_lock); - - /* Set EPOCH CLOSE flag if not set by client. */ - info->mti_ioepoch->flags |= MF_EPOCH_CLOSE; - info->mti_attr.ma_valid = 0; - - info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize; - OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize); - if (info->mti_attr.ma_lmm == NULL) - GOTO(error_ucred, rc = -ENOMEM); - - rc = mdt_mfd_close(info, mfd); - - OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize); - mdt_empty_transno(info, rc); -error_ucred: - mdt_exit_ucred(info); -out: - mdt_thread_info_fini(info); - RETURN(rc); -}