X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_open.c;h=e1717ddfccc8721f52ca9d599af1a24178436a0f;hp=9720d90692084b68b36f49903e8755671457164d;hb=9ddec96478d83afccec1d09266e063b2a5a0f71a;hpb=eea39e21a6d57db2a2ebadef93c1cd5a8b6d1ccf diff --git a/lustre/mdt/mdt_open.c b/lustre/mdt/mdt_open.c index 9720d90..e1717dd 100644 --- a/lustre/mdt/mdt_open.c +++ b/lustre/mdt/mdt_open.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2012, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -40,9 +40,6 @@ * Author: Huang Hua */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_MDS #include @@ -54,6 +51,11 @@ static void mdt_mfd_get(void *mfdp) { } +static struct portals_handle_ops mfd_handle_ops = { + .hop_addref = mdt_mfd_get, + .hop_free = NULL, +}; + /* Create a new mdt_file_data struct, initialize it, * and insert it to global hash table */ struct mdt_file_data *mdt_mfd_new(void) @@ -65,7 +67,7 @@ struct mdt_file_data *mdt_mfd_new(void) if (mfd != NULL) { CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link); CFS_INIT_LIST_HEAD(&mfd->mfd_list); - class_handle_hash(&mfd->mfd_handle, mdt_mfd_get); + class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops); } RETURN(mfd); } @@ -85,10 +87,9 @@ struct mdt_file_data *mdt_handle2mfd(struct mdt_thread_info *info, LASSERT(handle != NULL); mfd = class_handle2object(handle->cookie); /* during dw/setattr replay the mfd can be found by old handle */ - if (mfd == NULL && - lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { + if (mfd == NULL && req_is_replay(req)) { struct mdt_export_data *med = &req->rq_export->exp_mdt_data; - list_for_each_entry(mfd, &med->med_open_head, mfd_list) { + cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) { if (mfd->mfd_old_handle.cookie == handle->cookie) RETURN (mfd); } @@ -100,55 +101,72 @@ struct mdt_file_data *mdt_handle2mfd(struct mdt_thread_info *info, /* free mfd */ void mdt_mfd_free(struct mdt_file_data *mfd) { - LASSERT(list_empty(&mfd->mfd_list)); + LASSERT(cfs_list_empty(&mfd->mfd_list)); OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle); } static int mdt_create_data(struct mdt_thread_info *info, struct mdt_object *p, struct mdt_object *o) { - struct md_op_spec *spec = &info->mti_spec; - struct md_attr *ma = &info->mti_attr; - int rc; - ENTRY; - - if (!md_should_create(spec->sp_cr_flags)) - RETURN(0); - - ma->ma_need = MA_INODE | MA_LOV; - ma->ma_valid = 0; - rc = mdo_create_data(info->mti_env, - p ? mdt_object_child(p) : NULL, - mdt_object_child(o), spec, ma); - RETURN(rc); + struct md_op_spec *spec = &info->mti_spec; + struct md_attr *ma = &info->mti_attr; + int rc = 0; + ENTRY; + + if (!md_should_create(spec->sp_cr_flags)) + RETURN(0); + + 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); + + rc = mdo_create_data(info->mti_env, + p ? mdt_object_child(p) : NULL, + mdt_object_child(o), spec, ma); + if (rc == 0) + rc = mdt_attr_get_complex(info, o, ma); + + if (rc == 0 && ma->ma_valid & MA_LOV) + o->mot_flags |= MOF_LOV_CREATED; + } +unlock: + mutex_unlock(&o->mot_lov_mutex); + RETURN(rc); } -static int mdt_epoch_opened(struct mdt_object *mo) +static int mdt_ioepoch_opened(struct mdt_object *mo) { - return mo->mot_epochcount; + return mo->mot_ioepoch_count; } -int mdt_sizeonmds_enabled(struct mdt_object *mo) +int mdt_object_is_som_enabled(struct mdt_object *mo) { return !mo->mot_ioepoch; } -/* Re-enable Size-on-MDS. */ -void mdt_sizeonmds_enable(struct mdt_thread_info *info, - struct mdt_object *mo) +/** + * Re-enable Size-on-MDS. + * Call under ->mot_ioepoch_mutex. + */ +static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch) { - spin_lock(&info->mti_mdt->mdt_ioepoch_lock); - if (info->mti_epoch->ioepoch == mo->mot_ioepoch) { - LASSERT(!mdt_epoch_opened(mo)); + if (ioepoch == mo->mot_ioepoch) { + LASSERT(!mdt_ioepoch_opened(mo)); mo->mot_ioepoch = 0; mo->mot_flags = 0; - } - spin_unlock(&info->mti_mdt->mdt_ioepoch_lock); + } } -/* Open the epoch. Epoch open is allowed if @writecount is not negative. - * The epoch and writecount handling is performed under the mdt_ioepoch_lock. */ -int mdt_epoch_open(struct mdt_thread_info *info, struct mdt_object *o) +/** + * 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; @@ -159,27 +177,37 @@ int mdt_epoch_open(struct mdt_thread_info *info, struct mdt_object *o) !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) RETURN(0); - spin_lock(&mdt->mdt_ioepoch_lock); - if (mdt_epoch_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 { - if (info->mti_replayepoch > mdt->mdt_ioepoch) + 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 - mdt->mdt_ioepoch++; - o->mot_ioepoch = info->mti_replayepoch ? - info->mti_replayepoch : mdt->mdt_ioepoch; - CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n", - mdt->mdt_ioepoch, PFID(mdt_object_fid(o))); - cancel = 1; - } - o->mot_epochcount++; - spin_unlock(&mdt->mdt_ioepoch_lock); + o->mot_ioepoch = mdt->mdt_ioepoch; - /* Cancel Size-on-MDS attributes on clients if not truncate. - * In the later case, mdt_reint_setattr will do it. */ + 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); @@ -191,181 +219,433 @@ int mdt_epoch_open(struct mdt_thread_info *info, struct mdt_object *o) RETURN(rc); } -/* Update the on-disk attributes if needed and re-enable Size-on-MDS caching. */ -static int mdt_sizeonmds_update(struct mdt_thread_info *info, - struct mdt_object *o) +/** + * 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, "Closing epoch "LPU64" on "DFID". Count %d\n", - o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_epochcount); - - if (info->mti_attr.ma_attr.la_valid & LA_SIZE) { - /* Do Size-on-MDS attribute update. - * Size-on-MDS is re-enabled inside. */ - /* XXX: since we have opened the file, it is unnecessary - * to check permission when close it. Between the "open" - * and "close", maybe someone has changed the file mode - * or flags, or the file created mode do not permit wirte, - * and so on. Just set MDS_PERM_BYPASS for all the cases. */ - info->mti_attr.ma_attr_flags |= MDS_PERM_BYPASS | MDS_SOM; - info->mti_attr.ma_attr.la_valid &= LA_SIZE | LA_BLOCKS | - LA_ATIME | LA_MTIME | LA_CTIME; - RETURN(mdt_attr_set(info, o, 0)); - } else - mdt_sizeonmds_enable(info, o); - RETURN(0); + 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); } -/* Epoch closes. - * Returns 1 if epoch does not close. - * Returns 0 if epoch closes. - * Returns -EAGAIN if epoch closes but an Size-on-MDS Update is still needed - * from the client. */ -static int mdt_epoch_close(struct mdt_thread_info *info, struct mdt_object *o) +/** 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 eviction = (mdt_info_req(info) == NULL ? 1 : 0); - struct lu_attr *la = &info->mti_attr.ma_attr; - int achange = 0; - int opened; - int rc = 1; + 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; - if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) || - !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) - RETURN(0); + 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--; - spin_lock(&info->mti_mdt->mdt_ioepoch_lock); + /* 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; - /* Epoch closes only if client tells about it or eviction occures. */ - if (eviction || (info->mti_epoch->flags & MF_EPOCH_CLOSE)) { - LASSERT(o->mot_epochcount); - o->mot_epochcount--; + if (!mdt_ioepoch_opened(o)) + mdt_object_som_enable(o, info->mti_ioepoch->ioepoch); + mutex_unlock(&o->mot_ioepoch_mutex); - CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n", - o->mot_ioepoch, PFID(mdt_object_fid(o)), - o->mot_epochcount); + RETURN(rc); +} - if (!eviction) - achange = (info->mti_epoch->flags & MF_SOM_CHANGE); +/** + * 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; - rc = 0; - if (!eviction && !mdt_epoch_opened(o)) { - /* Epoch ends. Is an Size-on-MDS update needed? */ - if (o->mot_flags & MF_SOM_CHANGE) { - /* Some previous writer changed the attribute. - * Do not believe to the current Size-on-MDS - * update, re-ask client. */ - rc = -EAGAIN; - } else if (!(la->la_valid & LA_SIZE) && achange) { - /* Attributes were changed by the last writer - * only but no Size-on-MDS update is received.*/ - rc = -EAGAIN; - } + 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); + } - if (achange || eviction) - o->mot_flags |= MF_SOM_CHANGE; + 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; } - opened = mdt_epoch_opened(o); - spin_unlock(&info->mti_mdt->mdt_ioepoch_lock); + RETURN(rc ? : ret); - /* If eviction occurred, do nothing. */ - if ((rc == 0) && !opened && !eviction) { - /* Epoch ends and wanted Size-on-MDS update is obtained. */ - rc = mdt_sizeonmds_update(info, o); - /* Avoid the following setattrs of these attributes, e.g. - * for atime update. */ - info->mti_attr.ma_valid = 0; +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) +{ + struct ptlrpc_request *req = mdt_info_req(info); + ENTRY; + + if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) || + !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) + 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->flags & MF_EPOCH_CLOSE) + RETURN(mdt_ioepoch_close_reg(info, o)); + /* IO epoch is not closed. */ + RETURN(MDT_IOEPOCH_OPENED); +} + +/** + * 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) +{ + struct ptlrpc_request *req = mdt_info_req(info); + __u64 ioepoch = 0; + int act = MDT_SOM_ENABLE; + int rc = 0; + ENTRY; + + LASSERT(!req || info->mti_ioepoch); + if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) || + !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) + 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 (!(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_device *mdt, struct mdt_object *o) +int mdt_write_read(struct mdt_object *o) { int rc = 0; ENTRY; - spin_lock(&mdt->mdt_ioepoch_lock); + mutex_lock(&o->mot_ioepoch_mutex); rc = o->mot_writecount; - spin_unlock(&mdt->mdt_ioepoch_lock); + mutex_unlock(&o->mot_ioepoch_mutex); RETURN(rc); } -int mdt_write_get(struct mdt_device *mdt, struct mdt_object *o) +int mdt_write_get(struct mdt_object *o) { int rc = 0; ENTRY; - spin_lock(&mdt->mdt_ioepoch_lock); + mutex_lock(&o->mot_ioepoch_mutex); if (o->mot_writecount < 0) rc = -ETXTBSY; else o->mot_writecount++; - spin_unlock(&mdt->mdt_ioepoch_lock); + mutex_unlock(&o->mot_ioepoch_mutex); RETURN(rc); } -static void mdt_write_put(struct mdt_device *mdt, struct mdt_object *o) +void mdt_write_put(struct mdt_object *o) { ENTRY; - spin_lock(&mdt->mdt_ioepoch_lock); + mutex_lock(&o->mot_ioepoch_mutex); o->mot_writecount--; - spin_unlock(&mdt->mdt_ioepoch_lock); + mutex_unlock(&o->mot_ioepoch_mutex); EXIT; } -static int mdt_write_deny(struct mdt_device *mdt, struct mdt_object *o) +static int mdt_write_deny(struct mdt_object *o) { int rc = 0; ENTRY; - spin_lock(&mdt->mdt_ioepoch_lock); + mutex_lock(&o->mot_ioepoch_mutex); if (o->mot_writecount > 0) rc = -ETXTBSY; else o->mot_writecount--; - spin_unlock(&mdt->mdt_ioepoch_lock); + mutex_unlock(&o->mot_ioepoch_mutex); RETURN(rc); } -static void mdt_write_allow(struct mdt_device *mdt, struct mdt_object *o) +static void mdt_write_allow(struct mdt_object *o) { ENTRY; - spin_lock(&mdt->mdt_ioepoch_lock); + mutex_lock(&o->mot_ioepoch_mutex); o->mot_writecount++; - spin_unlock(&mdt->mdt_ioepoch_lock); + mutex_unlock(&o->mot_ioepoch_mutex); EXIT; } /* there can be no real transaction so prepare the fake one */ -static void mdt_empty_transno(struct mdt_thread_info* info) +static void mdt_empty_transno(struct mdt_thread_info *info, int rc) { - struct mdt_device *mdt = info->mti_mdt; - struct ptlrpc_request *req = mdt_info_req(info); + struct mdt_device *mdt = info->mti_mdt; + struct ptlrpc_request *req = mdt_info_req(info); + struct tg_export_data *ted; + struct lsd_client_data *lcd; ENTRY; - /* transaction is occured already */ - if (lustre_msg_get_transno(req->rq_repmsg) != 0) { - EXIT; - return; - } - - spin_lock(&mdt->mdt_transno_lock); - if (info->mti_transno == 0) { - info->mti_transno = ++ mdt->mdt_last_transno; - } else { - /* should be replay */ - if (info->mti_transno > mdt->mdt_last_transno) - mdt->mdt_last_transno = info->mti_transno; - } - spin_unlock(&mdt->mdt_transno_lock); - - CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n", + /* transaction has occurred already */ + if (lustre_msg_get_transno(req->rq_repmsg) != 0) + 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", + mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name, + info->mti_transno, + libcfs_nid2str(exp->exp_connection->c_peer.nid), + rc); + RETURN_EXIT; + } + } else if (info->mti_transno == 0) { + info->mti_transno = ++mdt->mdt_lut.lut_last_transno; + } else { + /* should be replay */ + if (info->mti_transno > mdt->mdt_lut.lut_last_transno) + mdt->mdt_lut.lut_last_transno = info->mti_transno; + } + spin_unlock(&mdt->mdt_lut.lut_translock); + + CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n", info->mti_transno, req->rq_export->exp_obd->obd_last_committed); req->rq_transno = info->mti_transno; lustre_msg_set_transno(req->rq_repmsg, info->mti_transno); - lustre_msg_set_last_xid(req->rq_repmsg, req->rq_xid); + + /* update lcd in memory only for resent cases */ + ted = &req->rq_export->exp_target_data; + LASSERT(ted); + mutex_lock(&ted->ted_lcd_lock); + lcd = ted->ted_lcd; + if (info->mti_transno < lcd->lcd_last_transno && + info->mti_transno != 0) { + /* This should happen during replay. Do not update + * last rcvd info if replay req transno < last transno, + * 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", + mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name, + 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 (info->mti_transno != 0) + lcd->lcd_last_close_transno = info->mti_transno; + lcd->lcd_last_close_xid = req->rq_xid; + lcd->lcd_last_close_result = rc; + } else { + /* VBR: save versions in last_rcvd for reconstruct. */ + __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg); + if (pre_versions) { + lcd->lcd_pre_versions[0] = pre_versions[0]; + lcd->lcd_pre_versions[1] = pre_versions[1]; + lcd->lcd_pre_versions[2] = pre_versions[2]; + lcd->lcd_pre_versions[3] = pre_versions[3]; + } + if (info->mti_transno != 0) + lcd->lcd_last_transno = info->mti_transno; + + lcd->lcd_last_xid = req->rq_xid; + lcd->lcd_last_result = rc; + lcd->lcd_last_data = info->mti_opdata; + } + mutex_unlock(&ted->ted_lcd_lock); + EXIT; } @@ -380,7 +660,7 @@ void mdt_mfd_set_mode(struct mdt_file_data *mfd, int mode) } static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, - struct mdt_object *o, int flags, int created) + struct mdt_object *o, __u64 flags, int created) { struct ptlrpc_request *req = mdt_info_req(info); struct mdt_export_data *med = &req->rq_export->exp_mdt_data; @@ -395,7 +675,7 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, isreg = S_ISREG(la->la_mode); isdir = S_ISDIR(la->la_mode); - if ((isreg && !(ma->ma_valid & MA_LOV))) { + if (isreg && !(ma->ma_valid & MA_LOV)) { /* * 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 @@ -423,13 +703,19 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, } if (flags & FMODE_WRITE) { - rc = mdt_write_get(info->mti_mdt, o); + rc = mdt_write_get(o); if (rc == 0) { - mdt_epoch_open(info, o); + mdt_ioepoch_open(info, o, created); repbody->ioepoch = o->mot_ioepoch; } } else if (flags & MDS_FMODE_EXEC) { - rc = mdt_write_deny(info->mti_mdt, o); + /* 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; + else + rc = mdt_write_deny(o); } if (rc) RETURN(rc); @@ -437,7 +723,7 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, rc = mo_open(info->mti_env, mdt_object_child(o), created ? flags | MDS_OPEN_CREATED : flags); if (rc) - RETURN(rc); + GOTO(err_out, rc); mfd = mdt_mfd_new(); if (mfd != NULL) { @@ -449,7 +735,7 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, /* * @flags is always not zero. At least it should be FMODE_READ, - * FMODE_WRITE or FMODE_EXEC. + * FMODE_WRITE or MDS_FMODE_EXEC. */ LASSERT(flags != 0); @@ -460,7 +746,7 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, mfd->mfd_xid = req->rq_xid; /* replay handle */ - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { + if (req_is_replay(req)) { struct mdt_file_data *old_mfd; /* Check wheather old cookie already exist in * the list, becasue when do recovery, client @@ -470,46 +756,72 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, LASSERT(info->mti_rr.rr_handle != NULL); old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle); if (old_mfd) { - CDEBUG(D_HA, "del orph mfd %p cookie" LPX64"\n", - mfd, info->mti_rr.rr_handle->cookie); - spin_lock(&med->med_open_lock); - class_handle_unhash(&old_mfd->mfd_handle); - list_del_init(&old_mfd->mfd_list); - spin_unlock(&med->med_open_lock); - mdt_mfd_free(old_mfd); + CDEBUG(D_HA, "del orph mfd %p fid=("DFID") " + "cookie=" LPX64"\n", mfd, + PFID(mdt_object_fid(mfd->mfd_object)), + info->mti_rr.rr_handle->cookie); + spin_lock(&med->med_open_lock); + class_handle_unhash(&old_mfd->mfd_handle); + cfs_list_del_init(&old_mfd->mfd_list); + spin_unlock(&med->med_open_lock); + /* no attr update for that close */ + la->la_valid = 0; + ma->ma_valid |= MA_FLAGS; + ma->ma_attr_flags |= MDS_RECOV_OPEN; + mdt_mfd_close(info, old_mfd); + ma->ma_attr_flags &= ~MDS_RECOV_OPEN; + ma->ma_valid &= ~MA_FLAGS; } CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n", info->mti_rr.rr_handle->cookie); mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie; } - spin_lock(&med->med_open_lock); - list_add(&mfd->mfd_list, &med->med_open_head); - spin_unlock(&med->med_open_lock); - repbody->handle.cookie = mfd->mfd_handle.h_cookie; - mdt_empty_transno(info); - } else - rc = -ENOMEM; + + 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); + 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); + spin_unlock(&med->med_open_lock); + } + + mdt_empty_transno(info, rc); + } else { + GOTO(err_out, rc = -ENOMEM); + } 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; +} -static int mdt_finish_open(struct mdt_thread_info *info, - struct mdt_object *p, struct mdt_object *o, - int flags, int created, struct ldlm_reply *rep) +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) { 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 mdt_device *mdt = info->mti_mdt; 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; + cfs_list_t *t; ENTRY; LASSERT(ma->ma_valid & MA_INODE); @@ -521,7 +833,22 @@ static 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)); - if (med->med_rmtclient) { + /* 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 */ + 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) + 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); @@ -534,7 +861,7 @@ static int mdt_finish_open(struct mdt_thread_info *info, } } #ifdef CONFIG_FS_POSIX_ACL - else if (req->rq_export->exp_connect_flags & OBD_CONNECT_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; @@ -564,26 +891,26 @@ static int mdt_finish_open(struct mdt_thread_info *info, } #endif - if (mdt->mdt_opts.mo_mds_capa) { + if (info->mti_mdt->mdt_opts.mo_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; - capa->lc_uid = 0; rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0); if (rc) RETURN(rc); repbody->valid |= OBD_MD_FLMDSCAPA; } - if (mdt->mdt_opts.mo_oss_capa && - S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) { + if (info->mti_mdt->mdt_opts.mo_oss_capa && + exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA && + S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) { 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); - capa->lc_uid = 0; rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0); if (rc) RETURN(rc); @@ -594,13 +921,11 @@ static int mdt_finish_open(struct mdt_thread_info *info, * 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 && - (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) { - lustre_msg_set_transno(req->rq_repmsg, 0); - RETURN(0); - } - - mdt_set_disposition(info, rep, DISP_OPEN_OPEN); + 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, @@ -625,15 +950,14 @@ static int mdt_finish_open(struct mdt_thread_info *info, mfd = NULL; if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) { - spin_lock(&med->med_open_lock); - 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); + 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); + 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; @@ -646,11 +970,15 @@ static int mdt_finish_open(struct mdt_thread_info *info, else repbody->valid |= OBD_MD_FLEASIZE; } + mdt_set_disposition(info, rep, DISP_OPEN_OPEN); RETURN(0); } } rc = mdt_mfd_open(info, p, o, flags, created); + if (!rc) + mdt_set_disposition(info, rep, DISP_OPEN_OPEN); + RETURN(rc); } @@ -664,8 +992,8 @@ void mdt_reconstruct_open(struct mdt_thread_info *info, struct mdt_device *mdt = info->mti_mdt; struct req_capsule *pill = info->mti_pill; struct ptlrpc_request *req = mdt_info_req(info); - struct mdt_export_data *med = &req->rq_export->exp_mdt_data; - struct lsd_client_data *lcd = med->med_lcd; + 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; __u32 flags = info->mti_spec.sp_cr_flags; @@ -683,14 +1011,17 @@ void mdt_reconstruct_open(struct mdt_thread_info *info, 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_LOV; + ma->ma_need = MA_INODE | MA_HSM; + if (ma->ma_lmm_size > 0) + ma->ma_need |= MA_LOV; + ma->ma_valid = 0; - mdt_req_from_lcd(req, med->med_lcd); + mdt_req_from_lcd(req, lcd); mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data); - CERROR("This is reconstruct open: disp="LPX64", result=%d\n", - ldlm_rep->lock_policy_res1, req->rq_status); + CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n", + ldlm_rep->lock_policy_res1, req->rq_status); if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) && req->rq_status != 0) @@ -708,46 +1039,53 @@ void mdt_reconstruct_open(struct mdt_thread_info *info, rc = PTR_ERR(parent); LCONSOLE_WARN("Parent "DFID" lookup error %d." " Evicting client %s with export %s.\n", - PFID(mdt_object_fid(parent)), rc, + PFID(rr->rr_fid1), rc, obd_uuid2str(&exp->exp_client_uuid), obd_export_nid2str(exp)); mdt_export_evict(exp); - EXIT; - return; + RETURN_EXIT; } child = mdt_object_find(env, mdt, rr->rr_fid2); if (IS_ERR(child)) { - rc = PTR_ERR(parent); + 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_export_evict(exp); - EXIT; - return; - } - rc = mdt_object_exists(child); - if (rc > 0) { - struct md_object *next; - - mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA); - next = mdt_object_child(child); - rc = mo_attr_get(env, next, ma); - if (rc == 0) - rc = mdt_finish_open(info, parent, child, - flags, 1, ldlm_rep); - } else if (rc < 0) { - /* the child object was created on remote server */ - repbody->fid1 = *rr->rr_fid2; - repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); - rc = 0; - } else if (rc == 0) { - /* the child does not exist, we should do regular open */ mdt_object_put(env, parent); - mdt_object_put(env, child); - GOTO(regular_open, 0); + mdt_export_evict(exp); + RETURN_EXIT; } + + if (unlikely(mdt_object_remote(child))) { + /* the child object was created on remote server */ + 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); + } + } mdt_object_put(env, parent); mdt_object_put(env, child); GOTO(out, rc); @@ -764,10 +1102,9 @@ out: LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0)); } -static int mdt_open_by_fid(struct mdt_thread_info* info, - struct ldlm_reply *rep) +int mdt_open_by_fid(struct mdt_thread_info* info, + struct ldlm_reply *rep) { - const struct lu_env *env = info->mti_env; __u32 flags = info->mti_spec.sp_cr_flags; struct mdt_reint_record *rr = &info->mti_rr; struct md_attr *ma = &info->mti_attr; @@ -779,30 +1116,239 @@ static int mdt_open_by_fid(struct mdt_thread_info* info, if (IS_ERR(o)) RETURN(rc = PTR_ERR(o)); - rc = mdt_object_exists(o); - if (rc > 0) { - mdt_set_disposition(info, rep, (DISP_IT_EXECD | - DISP_LOOKUP_EXECD | - DISP_LOOKUP_POS)); - - rc = mo_attr_get(env, mdt_object_child(o), ma); - if (rc == 0) - rc = mdt_finish_open(info, NULL, o, flags, 0, rep); - } else if (rc == 0) { - rc = -ENOENT; - } else { + if (unlikely(mdt_object_remote(o))) { /* 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; - } + } else { + if (mdt_object_exists(o)) { + mdt_set_disposition(info, rep, (DISP_IT_EXECD | + DISP_LOOKUP_EXECD | + DISP_LOOKUP_POS)); + + rc = mdt_attr_get_complex(info, o, ma); + if (rc == 0) + rc = mdt_finish_open(info, NULL, o, flags, 0, + rep); + } else { + rc = -ENOENT; + } + } mdt_object_put(info->mti_env, o); RETURN(rc); } +/* lock object for open */ +static int mdt_object_open_lock(struct mdt_thread_info *info, + struct mdt_object *obj, + 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 try_layout = false; + bool create_layout = false; + int rc = 0; + ENTRY; + + *ibits = 0; + if (open_flags & MDS_OPEN_LOCK) { + if (open_flags & 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))) + lm = LCK_PR; + else + lm = LCK_CR; + + *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN; + } + + if (S_ISREG(lu_object_attr(&obj->mot_obj.mo_lu))) { + if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) && + md_should_create(open_flags)) + create_layout = true; + if (exp_connect_layout(info->mti_exp) && !create_layout && + ma->ma_need & MA_LOV) + try_layout = true; + } + + mdt_lock_handle_init(lhc); + 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. */ + 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); + } + + 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); + + /* 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", + PFID(mdt_object_fid(obj)), open_flags); + + 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); + + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2); + } + + RETURN(rc); +} + +static void mdt_object_open_unlock(struct mdt_thread_info *info, + struct mdt_object *obj, + struct mdt_lock_handle *lhc, + __u64 ibits, int rc) +{ + __u64 open_flags = info->mti_spec.sp_cr_flags; + struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT]; + + /* Release local layout lock - the layout lock put in MDT_LH_LAYOUT + * will never return to client side. */ + if (lustre_handle_is_used(&ll->mlh_reg_lh)) { + LASSERT(!(ibits & MDS_INODELOCK_LAYOUT)); + mdt_object_unlock(info, obj, ll, 1); + } + + /* Cross-ref case, the lock should be returned to the client */ + if (ibits == 0 || rc == -EREMOTE) + return; + + if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) { + /* for the open request, the lock will only return to client + * if open or layout lock is granted. */ + rc = 1; + } + + if (rc != 0) { + 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); + } +} + +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; + __u32 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; + ENTRY; + + 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; + } + + o = mdt_object_find(env, mdt, rr->rr_fid2); + if (IS_ERR(o)) + RETURN(rc = PTR_ERR(o)); + + if (mdt_object_remote(o)) { + CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n", + info->mti_mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name, + PFID(rr->rr_fid2)); + GOTO(out, rc = -EREMOTE); + } else if (!mdt_object_exists(o)) { + mdt_set_disposition(info, rep, + DISP_IT_EXECD | + DISP_LOOKUP_EXECD | + DISP_LOOKUP_NEG); + GOTO(out, rc = -ENOENT); + } + + mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD)); + + rc = mdt_attr_get_complex(info, o, ma); + if (rc) + GOTO(out, rc); + + rc = mdt_object_open_lock(info, o, lhc, &ibits); + if (rc) + GOTO(out, 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; + } + } + + rc = mdt_finish_open(info, parent, o, flags, 0, rep); + if (!rc) { + mdt_set_disposition(info, rep, DISP_LOOKUP_POS); + if (flags & MDS_OPEN_LOCK) + mdt_set_disposition(info, rep, DISP_OPEN_LOCK); + } + GOTO(out, rc); + +out: + mdt_object_open_unlock(info, o, lhc, ibits, rc); + mdt_object_put(env, o); + if (parent != NULL) + mdt_object_put(env, parent); + return rc; +} + int mdt_pin(struct mdt_thread_info* info) { ENTRY; @@ -810,9 +1356,10 @@ int mdt_pin(struct mdt_thread_info* info) } /* Cross-ref request. Currently it can only be a pure open (w/o create) */ -static int mdt_cross_open(struct mdt_thread_info* info, - const struct lu_fid *fid, - struct ldlm_reply *rep, __u32 flags) +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 md_attr *ma = &info->mti_attr; struct mdt_object *o; @@ -823,35 +1370,46 @@ static int mdt_cross_open(struct mdt_thread_info* info, if (IS_ERR(o)) RETURN(rc = PTR_ERR(o)); - rc = mdt_object_exists(o); - if (rc > 0) { - /* Do permission check for cross-open. */ - rc = mo_permission(info->mti_env, NULL, mdt_object_child(o), - NULL, flags | MDS_OPEN_CROSS); - if (rc) - goto out; - - mdt_set_capainfo(info, 0, fid, BYPASS_CAPA); - rc = mo_attr_get(info->mti_env, mdt_object_child(o), ma); - if (rc == 0) - rc = mdt_finish_open(info, NULL, o, flags, 0, rep); - } else if (rc == 0) { - /* - * Something is wrong here. lookup was positive but there is - * no object! - */ - CERROR("Cross-ref object doesn't exist!\n"); - rc = -EFAULT; - } else { - /* Something is wrong here, the object is on another MDS! */ - CERROR("The object isn't on this server! FLD error?\n"); - LU_OBJECT_DEBUG(D_WARNING, info->mti_env, - &o->mot_obj.mo_lu, - "Object isn't on this server! FLD error?\n"); - + if (mdt_object_remote(o)) { + /* Something is wrong here, the object is on another MDS! */ + CERROR("%s: "DFID" isn't on this server!: rc = %d\n", + mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT); + LU_OBJECT_DEBUG(D_WARNING, info->mti_env, + &o->mot_obj.mo_lu, + "Object isn't on this server! FLD error?\n"); rc = -EFAULT; + } else { + if (mdt_object_exists(o)) { + /* Do permission check for cross-open. */ + rc = mo_permission(info->mti_env, NULL, + mdt_object_child(o), + NULL, flags | MDS_OPEN_CROSS); + if (rc) + goto out; + + mdt_set_capainfo(info, 0, fid, BYPASS_CAPA); + 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); + + rc = mdt_finish_open(info, NULL, o, flags, 0, rep); + } else { + /* + * Something is wrong here. lookup was positive but + * there is no object! + */ + CERROR("%s: "DFID" doesn't exist!: rc = %d\n", + mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT); + rc = -EFAULT; + } } - out: mdt_object_put(info->mti_env, o); RETURN(rc); @@ -868,7 +1426,8 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) struct mdt_body *repbody; struct lu_fid *child_fid = &info->mti_tmp_fid1; struct md_attr *ma = &info->mti_attr; - __u32 create_flags = info->mti_spec.sp_cr_flags; + __u64 create_flags = info->mti_spec.sp_cr_flags; + __u64 ibits; struct mdt_reint_record *rr = &info->mti_rr; struct lu_name *lname; int result, rc; @@ -879,52 +1438,83 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) 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); 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 | MA_LOV; + ma->ma_need = MA_INODE; + if (ma->ma_lmm_size > 0) + ma->ma_need |= MA_LOV; + 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); - /* TODO: JOIN file */ - if (create_flags & MDS_OPEN_JOIN_FILE) { - CERROR("JOIN file will be supported soon\n"); + 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 ((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"/(%s->"DFID") " - "cr_flag=0%o mode=0%06o msg_flag=0x%x\n", + "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 (msg_flags & MSG_REPLAY || - (req->rq_export->exp_libclient && create_flags&MDS_OPEN_HAS_EA)) { - /* This is a replay request or from liblustre with ea. */ - 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); - 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)) { - DEBUG_REQ(D_ERROR, req,"OPEN & CREAT not in open replay."); - GOTO(out, result = -EFAULT); - } - CDEBUG(D_INFO, "Open replay did find object, continue as " - "regular open\n"); - } + 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); + 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. */ + 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); + 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)) { + 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)) { + 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 (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) GOTO(out, result = err_serious(-ENOMEM)); @@ -932,14 +1522,6 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) mdt_set_disposition(info, ldlm_rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD)); - if (info->mti_cross_ref) { - /* This is cross-ref open */ - mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS); - result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep, - create_flags); - GOTO(out, result); - } - 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); @@ -949,10 +1531,14 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) if (IS_ERR(parent)) GOTO(out, result = PTR_ERR(parent)); + /* get and check version of parent */ + result = mdt_version_get_check(info, parent, 0); + if (result) + GOTO(out_parent, result); + fid_zero(child_fid); 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)), @@ -977,20 +1563,39 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) *child_fid = *info->mti_rr.rr_fid2; LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n", PFID(child_fid)); - } else { - /* - * 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); + /* 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. + */ + 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)); + /** 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) { + /* Create under OBF and .lustre is not permitted */ + if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1)) + GOTO(out_child, result = -EPERM); + + /* 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; + /* Not found and with MDS_OPEN_CREAT: let's create it. */ mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE); @@ -1003,6 +1608,7 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) * 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), @@ -1014,18 +1620,20 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE); GOTO(out_child, result); } else { + + /* XXX: we should call this once, see few lines below */ + if (result == 0) + result = mdt_attr_get_complex(info, child, ma); + if (result != 0) GOTO(out_child, result); } created = 1; } else { - /* We have to get attr & lov ea for this object */ - result = mo_attr_get(info->mti_env, mdt_object_child(child), - ma); /* * The object is on remote node, return its FID for remote open. */ - if (result == -EREMOTE) { + if (mdt_object_remote(child)) { /* * Check if this lock already was sent to client and * this is resent case. For resent case do not take lock @@ -1060,73 +1668,70 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS); if (rc != 0) result = rc; + else + result = -EREMOTE; 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(); + } + } } LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh)); - /* get openlock if this is not replay and if a client requested it */ - if (!(msg_flags & MSG_REPLAY) && create_flags & MDS_OPEN_LOCK) { - ldlm_mode_t lm; - - LASSERT(!created); - if (create_flags & FMODE_WRITE) - lm = LCK_CW; - else if (create_flags & MDS_FMODE_EXEC) - lm = LCK_PR; - else - lm = LCK_CR; - mdt_lock_handle_init(lhc); - mdt_lock_reg_init(lhc, lm); - rc = mdt_object_lock(info, child, lhc, - MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN, - MDT_CROSS_LOCK); - if (rc) { - result = rc; - GOTO(out_child, result); - } else { - result = -EREMOTE; - mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK); - } - } - - /* Try to open it now. */ - rc = mdt_finish_open(info, parent, child, create_flags, - created, ldlm_rep); - if (rc) { - result = rc; - if (lustre_handle_is_used(&lhc->mlh_reg_lh)) - /* openlock was acquired and mdt_finish_open failed - - drop the openlock */ - mdt_object_unlock(info, child, lhc, 1); - if (created) { - ma->ma_need = 0; - ma->ma_valid = 0; - ma->ma_cookie_size = 0; - info->mti_no_need_trans = 1; - rc = mdo_unlink(info->mti_env, - mdt_object_child(parent), - mdt_object_child(child), - lname, - &info->mti_attr); - if (rc != 0) - CERROR("Error in cleanup of open\n"); - } - } + /* 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, result = rc); + else if (create_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, + created, ldlm_rep); + if (rc) { + result = rc; + /* openlock will be released if mdt_finish_open failed */ + mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK); + 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, + &info->mti_attr, 0); + if (rc != 0) + CERROR("%s: "DFID" cleanup of open: rc = %d\n", + mdt_obd_name(info->mti_mdt), + PFID(mdt_object_fid(child)), rc); + mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE); + } + } EXIT; out_child: + mdt_object_open_unlock(info, child, lhc, ibits, result); mdt_object_put(info->mti_env, child); out_parent: - mdt_object_unlock_put(info, parent, lh, result); + mdt_object_unlock_put(info, parent, lh, result || !created); out: - if (result && result != -EREMOTE) - lustre_msg_set_transno(req->rq_repmsg, 0); - return result; + if (result) + lustre_msg_set_transno(req->rq_repmsg, 0); + return result; } -#define MFD_CLOSED(mode) (((mode) & ~(FMODE_EPOCH | FMODE_SOM | \ - FMODE_EPOCHLCK)) == FMODE_CLOSED) +#define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \ + MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED) static int mdt_mfd_closed(struct mdt_file_data *mfd) { @@ -1138,61 +1743,77 @@ 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 rc = 0, ret = 0; + int ret = MDT_IOEPOCH_CLOSED; + int rc = 0; int mode; ENTRY; mode = mfd->mfd_mode; - if ((mode & FMODE_WRITE) || (mode & FMODE_EPOCHLCK)) { - mdt_write_put(info->mti_mdt, o); - ret = mdt_epoch_close(info, o); + 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(info->mti_mdt, o); - } else if (mode & FMODE_EPOCH) { - ret = mdt_epoch_close(info, o); + 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); } /* 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)) { /* Set the atime only. */ + ma->ma_valid = MA_INODE; ma->ma_attr.la_valid = LA_ATIME; rc = mo_attr_set(info->mti_env, next, ma); } + /* If file data is modified, add the dirty flag. + * + * If MDS_CLOSE_CLEANUP is set, this file is being closed due to an + * eviction, file could have been modified and now dirty + * regarding to HSM archive, check this! + * The logic here is to mark a file dirty if there's a chance it was + * dirtied before the client was evicted, so that we don't have to wait + * for a release attempt before finding out the file was actually dirty + * and fail the release. Aggressively marking it dirty here will cause + * the policy engine to attempt to re-archive it; when rearchiving, we + * can compare the current version to the LMA data_version and make the + * archive request into a noop if it's not actually dirty. + */ + if ((ma->ma_attr_flags & MDS_DATA_MODIFIED) || + ((ma->ma_attr_flags & MDS_CLOSE_CLEANUP) && + (mode & (FMODE_WRITE|MDS_FMODE_TRUNC)))) + rc = mdt_add_dirty_flag(info, o, ma); + ma->ma_need |= MA_INODE; - ma->ma_valid = 0; + ma->ma_valid &= ~MA_INODE; if (!MFD_CLOSED(mode)) - rc = mo_close(info->mti_env, next, ma); - else if (ret == -EAGAIN) - rc = mo_attr_get(info->mti_env, next, ma); - - /* If the object is unlinked, do not try to re-enable SIZEONMDS */ - if ((ret == -EAGAIN) && (ma->ma_valid & MA_INODE) && - (ma->ma_attr.la_nlink == 0)) { - ret = 0; - } + rc = mo_close(info->mti_env, next, ma, mode); - if ((ret == -EAGAIN) || (ret == 1)) { + if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) { struct mdt_export_data *med; - /* The epoch has not closed or Size-on-MDS update is needed. + /* 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 == 1 ? FMODE_EPOCH : FMODE_SOM)); + mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ? + MDS_FMODE_EPOCH : MDS_FMODE_SOM); LASSERT(mdt_info_req(info)); med = &mdt_info_req(info)->rq_export->exp_mdt_data; - spin_lock(&med->med_open_lock); - list_add(&mfd->mfd_list, &med->med_open_head); - class_handle_hash_back(&mfd->mfd_handle); - spin_unlock(&med->med_open_lock); + 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); - if (ret == 1) { + 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))); } @@ -1215,20 +1836,26 @@ int mdt_close(struct mdt_thread_info *info) int rc, ret = 0; ENTRY; + mdt_counter_incr(req, LPROC_MDT_CLOSE); /* Close may come with the Size-on-MDS update. Unpack it. */ rc = mdt_close_unpack(info); if (rc) RETURN(err_serious(rc)); - LASSERT(info->mti_epoch); + LASSERT(info->mti_ioepoch); req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER, info->mti_mdt->mdt_max_mdsize); req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER, info->mti_mdt->mdt_max_cookiesize); rc = req_capsule_server_pack(info->mti_pill); - if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) + if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) { + mdt_client_compatibility(info); + if (rc == 0) + mdt_fix_reply(info); + mdt_exit_ucred(info); RETURN(lustre_msg_get_status(req->rq_repmsg)); + } /* Continue to close handle even if we can not pack reply */ if (rc == 0) { @@ -1252,18 +1879,19 @@ int mdt_close(struct mdt_thread_info *info) } med = &req->rq_export->exp_mdt_data; - spin_lock(&med->med_open_lock); - mfd = mdt_handle2mfd(info, &info->mti_epoch->handle); - 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_epoch->handle.cookie); - rc = err_serious(-ESTALE); - } else { - class_handle_unhash(&mfd->mfd_handle); - list_del_init(&mfd->mfd_list); - spin_unlock(&med->med_open_lock); + spin_lock(&med->med_open_lock); + mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle); + 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; + } else { + class_handle_unhash(&mfd->mfd_handle); + cfs_list_del_init(&mfd->mfd_list); + spin_unlock(&med->med_open_lock); /* Do not lose object before last unlink. */ o = mfd->mfd_object; @@ -1271,12 +1899,15 @@ int mdt_close(struct mdt_thread_info *info) ret = mdt_mfd_close(info, mfd); if (repbody != NULL) rc = mdt_handle_last_unlink(info, o, ma); - mdt_empty_transno(info); + mdt_empty_transno(info, rc); mdt_object_put(info->mti_env, o); } - if (repbody != NULL) - mdt_shrink_reply(info); + if (repbody != NULL) { + mdt_client_compatibility(info); + rc = mdt_fix_reply(info); + } + mdt_exit_ucred(info); if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) RETURN(err_serious(-ENOMEM)); @@ -1286,8 +1917,17 @@ int mdt_close(struct mdt_thread_info *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 mdt_thread_info *info) { + struct ptlrpc_request *req = mdt_info_req(info); struct mdt_body *repbody = NULL; struct mdt_export_data *med; struct mdt_file_data *mfd; @@ -1308,30 +1948,51 @@ int mdt_done_writing(struct mdt_thread_info *info) if (rc) RETURN(err_serious(rc)); - if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) - RETURN(lustre_msg_get_status(mdt_info_req(info)->rq_repmsg)); + if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) { + mdt_exit_ucred(info); + RETURN(lustre_msg_get_status(req->rq_repmsg)); + } med = &info->mti_exp->exp_mdt_data; - spin_lock(&med->med_open_lock); - mfd = mdt_handle2mfd(info, &info->mti_epoch->handle); - if (mfd == NULL) { - spin_unlock(&med->med_open_lock); + spin_lock(&med->med_open_lock); + mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle); + if (mfd == NULL) { + spin_unlock(&med->med_open_lock); CDEBUG(D_INODE, "no handle for done write: fid = "DFID - ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1), - info->mti_epoch->handle.cookie); - RETURN(-ESTALE); + ": 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 == FMODE_EPOCH || - mfd->mfd_mode == FMODE_EPOCHLCK); + LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH || + mfd->mfd_mode == MDS_FMODE_TRUNC); class_handle_unhash(&mfd->mfd_handle); - list_del_init(&mfd->mfd_list); - spin_unlock(&med->med_open_lock); + cfs_list_del_init(&mfd->mfd_list); + spin_unlock(&med->med_open_lock); /* Set EPOCH CLOSE flag if not set by client. */ - info->mti_epoch->flags |= MF_EPOCH_CLOSE; + 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); - mdt_empty_transno(info); - RETURN(rc); + + 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); + RETURN(rc); }