X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_open.c;h=d59e92d25ef549101d408997283f7a0437678ab5;hp=934c6b880f5e0ddbe259c6852dc900172e0410cc;hb=da14903c4394cb476386c475ade5eda12b2a8aa7;hpb=89693927f0b065d44fdc496f6b49539118570104 diff --git a/lustre/mdt/mdt_open.c b/lustre/mdt/mdt_open.c index 934c6b8..d59e92d 100644 --- a/lustre/mdt/mdt_open.c +++ b/lustre/mdt/mdt_open.c @@ -23,11 +23,10 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2016, Intel Corporation. + * Copyright (c) 2011, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ - * Lustre is a trademark of Sun Microsystems, Inc. * * lustre/mdt/mdt_open.c * @@ -40,18 +39,11 @@ #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 = { - .hop_addref = mdt_mfd_get, - .hop_free = NULL, -}; +static const char mfd_open_handle_owner[] = "mdt"; /* Create a new mdt_file_data struct, initialize it, * and insert it to global hash table */ @@ -62,10 +54,11 @@ struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med) OBD_ALLOC_PTR(mfd); if (mfd != NULL) { - INIT_LIST_HEAD(&mfd->mfd_handle.h_link); - mfd->mfd_handle.h_owner = med; + refcount_set(&mfd->mfd_open_handle.h_ref, 1); + INIT_HLIST_NODE(&mfd->mfd_open_handle.h_link); + mfd->mfd_owner = med; INIT_LIST_HEAD(&mfd->mfd_list); - class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops); + class_handle_hash(&mfd->mfd_open_handle, mfd_open_handle_owner); } RETURN(mfd); @@ -77,19 +70,23 @@ 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_owner); + if (mfd) + refcount_dec(&mfd->mfd_open_handle.h_ref); + /* during dw/setattr replay the mfd can be found by old handle */ - if (mfd == NULL && is_replay_or_resent) { + if ((!mfd || mfd->mfd_owner != med) && is_replay_or_resent) { list_for_each_entry(mfd, &med->med_open_head, mfd_list) { - if (mfd->mfd_old_handle.cookie == handle->cookie) + if (mfd->mfd_open_handle_old.cookie == + open_handle->cookie) RETURN(mfd); } mfd = NULL; @@ -101,8 +98,10 @@ struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med, /* free mfd */ void mdt_mfd_free(struct mdt_file_data *mfd) { + LASSERT(refcount_read(&mfd->mfd_open_handle.h_ref) == 1); LASSERT(list_empty(&mfd->mfd_list)); - OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle); + OBD_FREE_PRE(mfd, sizeof(*mfd), "rcu"); + kfree_rcu(mfd, mfd_open_handle.h_rcu); } static int mdt_create_data(struct mdt_thread_info *info, @@ -225,8 +224,7 @@ static void mdt_empty_transno(struct mdt_thread_info *info, int rc) 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; } @@ -290,14 +288,15 @@ 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_DENTRY, DFID " Change mfd mode %#llo -> %#llo.\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; } /** @@ -307,7 +306,15 @@ static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info, struct mdt_object *obj, struct md_attr *ma) { - LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL); + if (ma->ma_lmv || ma->ma_lmm) { + CDEBUG(D_INFO, DFID " %s already set.\n", + PFID(mdt_object_fid(obj)), + ma->ma_lmv ? (ma->ma_lmm ? "ma_lmv and ma_lmm" + : "ma_lmv") + : "ma_lmm"); + return; + } + if (S_ISDIR(obj->mot_header.loh_attr)) { ma->ma_lmv = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD); @@ -328,45 +335,47 @@ static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info, } static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, - struct mdt_object *o, __u64 flags, int created, + 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; - int rc = 0, isdir, isreg; - ENTRY; + 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; - repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); + ENTRY; + repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); - 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); + 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); - } + } CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n", ma->ma_valid, ma->ma_lmm_size); - if (ma->ma_valid & MA_LOV) { - LASSERT(ma->ma_lmm_size != 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; @@ -381,16 +390,17 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA; } - if (flags & FMODE_WRITE) + if (open_flags & MDS_FMODE_WRITE) rc = mdt_write_get(o); - else if (flags & MDS_FMODE_EXEC) + else if (open_flags & MDS_FMODE_EXEC) rc = mdt_write_deny(o); - if (rc) - RETURN(rc); + if (rc) + RETURN(rc); - rc = mo_open(info->mti_env, mdt_object_child(o), - created ? flags | MDS_OPEN_CREATED : flags); + 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, @@ -417,16 +427,16 @@ 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 */ @@ -437,15 +447,16 @@ 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 = %#llx\n", mfd, PFID(mdt_object_fid(mfd->mfd_object)), - info->mti_rr.rr_handle->cookie); - class_handle_unhash(&old_mfd->mfd_handle); + 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 */ @@ -460,37 +471,37 @@ static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p, CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", " "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 %#llx in new mfd\n", - info->mti_rr.rr_handle->cookie); + 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->mbo_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); + 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); - 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) + if (open_flags & MDS_FMODE_WRITE) mdt_write_put(o); - else if (flags & MDS_FMODE_EXEC) + else if (open_flags & MDS_FMODE_EXEC) mdt_write_allow(o); return rc; @@ -498,7 +509,8 @@ err_out: static 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) + u64 open_flags, int created, + struct ldlm_reply *rep) { struct ptlrpc_request *req = mdt_info_req(info); struct obd_export *exp = req->rq_export; @@ -521,6 +533,26 @@ 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)); + /* 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 successful and a @@ -530,13 +562,13 @@ static int mdt_finish_open(struct mdt_thread_info *info, 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); } -#ifdef CONFIG_FS_POSIX_ACL +#ifdef CONFIG_LUSTRE_FS_POSIX_ACL if (exp_connect_flags(exp) & OBD_CONNECT_ACL) { struct lu_nodemap *nodemap = nodemap_get_from_exp(exp); if (IS_ERR(nodemap)) @@ -549,36 +581,36 @@ static int mdt_finish_open(struct mdt_thread_info *info, } #endif - /* - * 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); + /* + * 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); - /* 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); - - 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) { @@ -591,8 +623,9 @@ static int mdt_finish_open(struct mdt_thread_info *info, } spin_unlock(&med->med_open_lock); - if (mfd != NULL) { - repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie; + 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); @@ -607,7 +640,7 @@ static int mdt_finish_open(struct mdt_thread_info *info, } } - rc = mdt_mfd_open(info, p, o, flags, created, rep); + rc = mdt_mfd_open(info, p, o, open_flags, created, rep); if (!rc) mdt_set_disposition(info, rep, DISP_OPEN_OPEN); @@ -615,26 +648,26 @@ static int mdt_finish_open(struct mdt_thread_info *info, } 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 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; - __u64 opdata; + 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_need = MA_INODE | MA_HSM; ma->ma_valid = 0; @@ -682,89 +715,88 @@ void mdt_reconstruct_open(struct mdt_thread_info *info, } 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->mbo_fid1 = *rr->rr_fid2; repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS); - rc = 0; - } else { - if (mdt_object_exists(child)) { - mdt_prep_ma_buf_from_rep(info, child, ma); - 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); - } + 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)); } 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 = 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; + rc = 0; } else { if (mdt_object_exists(o)) { mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD | DISP_LOOKUP_POS)); + if ((open_flags & MDS_OPEN_EXCL) && + (open_flags & MDS_OPEN_CREAT)) + mdt_set_disposition(info, rep, + DISP_OPEN_CREATE); + 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 */ @@ -775,11 +807,16 @@ static int mdt_object_open_lock(struct mdt_thread_info *info, { 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; + __u32 dom_stripe = 0; + unsigned int dom_only = 0; + unsigned int dom_lock = 0; + ENTRY; *ibits = 0; @@ -795,6 +832,24 @@ 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 take IO lock at OPEN when it makes sense, + * check if file has DoM stripe and ask for lock if client + * no lock on that resource yet. + */ + if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL) + dom_stripe = mdt_lmm_dom_entry_check(ma->ma_lmm, + &dom_only); + /* If only DOM stripe is being used then we can expect IO + * to it after OPEN and will return corresponding DOM ibit + * using default strategy from mdt_opts.mo_dom_lock. + * Otherwise trylock mode is used always and DOM ibit will + * be returned optionally. + */ + if (dom_stripe && + !mdt_dom_client_has_lock(info, mdt_object_fid(obj))) + dom_lock = !dom_only ? TRYLOCK_DOM_ON_OPEN : + info->mti_mdt->mdt_opts.mo_dom_lock; } if (acq_lease) { @@ -817,8 +872,8 @@ static int mdt_object_open_lock(struct mdt_thread_info *info, GOTO(out, rc = -EPROTO); } - /* XXX: only exclusive open is supported. */ - lm = LCK_EX; + /* should conflict with new opens for write/execute */ + lm = LCK_PW; *ibits = MDS_INODELOCK_OPEN; /* never grant LCK_EX layout lock to client */ @@ -828,7 +883,7 @@ 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; else if (open_flags & MDS_FMODE_EXEC) lm = LCK_PR; @@ -837,7 +892,7 @@ static int mdt_object_open_lock(struct mdt_thread_info *info, *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; @@ -847,36 +902,37 @@ 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; + trybits |= MDS_INODELOCK_DOM | 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)) { - *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP); - if (*ibits != 0) - rc = mdt_object_lock(info, obj, lhc, *ibits); - } - } else if (*ibits != 0) { - rc = mdt_object_lock(info, obj, lhc, *ibits); + if (!(*ibits & MDS_INODELOCK_LOOKUP)) + trybits |= MDS_INODELOCK_LOOKUP; + trybits |= MDS_INODELOCK_LAYOUT; } - CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx" + 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, open_flags, try_layout, rc); + *ibits, trybits, open_flags, try_layout, rc); /* will change layout, revoke layout locks by enqueuing EX lock. */ if (rc == 0 && create_layout) { @@ -923,11 +979,11 @@ 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; @@ -935,8 +991,11 @@ static int mdt_object_open_lock(struct mdt_thread_info *info, PFID(mdt_object_fid(obj)), atomic_read(&obj->mot_open_count), open_count); - if (atomic_read(&obj->mot_open_count) > open_count) - GOTO(out, rc = -EBUSY); + if (atomic_read(&obj->mot_open_count) > open_count) { + /* fail if anyone *else* has opened file for write */ + if (mdt_write_read(obj) > 1) + GOTO(out, rc = -EBUSY); + } } GOTO(out, rc); @@ -977,7 +1036,8 @@ static void mdt_object_open_unlock(struct mdt_thread_info *info, 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; @@ -1015,32 +1075,32 @@ 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; - bool object_locked = false; - __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)) @@ -1062,7 +1122,7 @@ static int mdt_open_by_fid_lock(struct mdt_thread_info *info, mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD)); mdt_prep_ma_buf_from_rep(info, o, ma); - if (flags & MDS_OPEN_RELEASE) + if (open_flags & MDS_OPEN_RELEASE) ma->ma_need |= MA_HSM; rc = mdt_attr_get_complex(info, o, ma); if (rc) @@ -1070,12 +1130,12 @@ static int mdt_open_by_fid_lock(struct mdt_thread_info *info, /* We should not change file's existing LOV EA */ if (S_ISREG(lu_object_attr(&o->mot_obj)) && - flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV) + open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV) GOTO(out, rc = -EEXIST); - /* If a release request, check file flags are fine and ask for an + /* 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_check_resent_lock(info, o, lhc); @@ -1088,23 +1148,23 @@ static int mdt_open_by_fid_lock(struct mdt_thread_info *info, 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); @@ -1114,6 +1174,8 @@ out_unlock: 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); @@ -1124,16 +1186,16 @@ out_parent_put: 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! */ @@ -1142,13 +1204,18 @@ static int mdt_cross_open(struct mdt_thread_info *info, LU_OBJECT_DEBUG(D_WARNING, info->mti_env, &o->mot_obj, "Object isn't on this server! FLD error?"); - rc = -EFAULT; + 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; @@ -1157,7 +1224,15 @@ static int mdt_cross_open(struct mdt_thread_info *info, if (rc != 0) GOTO(out, rc); - rc = mdt_finish_open(info, NULL, o, flags, 0, rep); + rc = mdt_pack_secctx_in_reply(info, o); + if (unlikely(rc)) + GOTO(out, rc); + + rc = mdt_pack_encctx_in_reply(info, o); + if (unlikely(rc)) + GOTO(out, rc); + + rc = mdt_finish_open(info, NULL, o, open_flags, 0, rep); } else { /* * Something is wrong here. lookup was positive but @@ -1167,10 +1242,11 @@ 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); } /* @@ -1215,7 +1291,7 @@ static int mdt_lock_root_xattr(struct mdt_thread_info *info, rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root), &lhroot, LCK_PR, MDS_INODELOCK_XATTR, - false, true); + true); if (rc < 0) return rc; @@ -1229,55 +1305,55 @@ static int mdt_lock_root_xattr(struct mdt_thread_info *info, 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; - int result, rc; - int created = 0; - int object_locked = 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 = NULL; + 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; + enum ldlm_mode lock_mode = LCK_PR; + u32 msg_flags; + ktime_t kstart = ktime_get(); - 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_need = MA_INODE; - ma->ma_valid = 0; + ma->ma_need = MA_INODE; + 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(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 (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 ((create_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_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) && + info->mti_spec.u.sp_ea.eadata == NULL) + GOTO(out, result = err_serious(-EINVAL)); - if (create_flags & FMODE_WRITE && + if (open_flags & MDS_FMODE_WRITE && exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY) GOTO(out, result = -EROFS); 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), create_flags, - ma->ma_attr.la_mode, msg_flags); + PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2), + open_flags, ma->ma_attr.la_mode, msg_flags); if (info->mti_cross_ref) { /* This is cross-ref open */ @@ -1285,32 +1361,23 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) (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."); + "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 (create_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) { - /* - * MDS_OPEN_LOCK is checked for backward compatibility with 2.1 - * client. - */ + } else if (open_flags & MDS_OPEN_BY_FID) { result = mdt_open_by_fid_lock(info, ldlm_rep, lhc); if (result < 0) CDEBUG(D_INFO, "no object for "DFID": %d\n", @@ -1331,109 +1398,92 @@ int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc) if (result < 0) GOTO(out, result); -again: - lh = &info->mti_lh[MDT_LH_PARENT]; - mdt_lock_pdo_init(lh, - (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR, - &rr->rr_name); - parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1); if (IS_ERR(parent)) GOTO(out, result = PTR_ERR(parent)); - result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE); - if (result != 0) { + /* get and check version of parent */ + result = mdt_version_get_check(info, parent, 0); + if (result) { mdt_object_put(info->mti_env, parent); GOTO(out, result); } - /* get and check version of parent */ - result = mdt_version_get_check(info, parent, 0); - if (result) - GOTO(out_parent, result); + OBD_RACE(OBD_FAIL_MDS_REINT_OPEN); +again_pw: + fid_zero(child_fid); - fid_zero(child_fid); + if (open_flags & MDS_OPEN_VOLATILE) { + lh = NULL; + result = -ENOENT; + } else { + lh = &info->mti_lh[MDT_LH_PARENT]; + mdt_lock_pdo_init(lh, lock_mode, &rr->rr_name); + result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE); + if (result != 0) { + mdt_object_put(info->mti_env, parent); + GOTO(out, result); + } - result = mdo_lookup(info->mti_env, mdt_object_child(parent), - &rr->rr_name, child_fid, &info->mti_spec); + 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 != 0 && result != -ENOENT) + GOTO(out_parent, result); - 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); - } + OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2); - if (!(create_flags & MDS_OPEN_CREAT)) - GOTO(out_parent, result); + if (result == -ENOENT) { + mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG); + 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)); + + LASSERT(equi(lh == NULL, open_flags & MDS_OPEN_VOLATILE)); + + if (lh != NULL && lock_mode == LCK_PR) { + /* first pass: get write lock and restart */ + mdt_object_unlock(info, parent, lh, 1); + mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_NEG); + mdt_lock_handle_init(lh); + lock_mode = LCK_PW; + goto again_pw; + } + + *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); - if (result == -ENOENT) { + if (result == -ENOENT) { /* Create under OBF and .lustre is not permitted */ - if (!fid_is_md_operative(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 */ @@ -1443,30 +1493,20 @@ again: /* 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); + /* 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; - /* - * 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), - &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 { + 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) @@ -1476,7 +1516,8 @@ again: GOTO(out_child, result); } created = 1; - mdt_counter_incr(req, LPROC_MDT_MKNOD); + mdt_counter_incr(req, LPROC_MDT_MKNOD, + ktime_us_delta(ktime_get(), kstart)); } else { /* * The object is on remote node, return its FID for remote open. @@ -1507,6 +1548,14 @@ again: result = -MDT_EREMOTE_OPEN; GOTO(out_child, result); } else if (mdt_object_exists(child)) { + /* Check early for MDS_OPEN_DIRECTORY/O_DIRECTORY to + * avoid opening regular files from lfs getstripe + * since doing so breaks the leases used by lfs + * mirror. See LU-13693. */ + if (open_flags & MDS_OPEN_DIRECTORY && + S_ISREG(lu_object_attr(&child->mot_obj))) + GOTO(out_child, result = -ENOTDIR); + /* We have to get attr & LOV EA & HSM for this * object. */ mdt_prep_ma_buf_from_rep(info, child, ma); @@ -1521,7 +1570,18 @@ again: PNAME(&rr->rr_name), PFID(child_fid)); GOTO(out_child, result = -EIO); } - } + } + + repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize; + repbody->mbo_valid |= OBD_MD_FLMODEASIZE; + + rc = mdt_pack_secctx_in_reply(info, child); + if (unlikely(rc)) + GOTO(out_child, result = rc); + + rc = mdt_pack_encctx_in_reply(info, child); + if (unlikely(rc)) + GOTO(out_child, result = rc); rc = mdt_check_resent_lock(info, child, lhc); if (rc < 0) { @@ -1530,7 +1590,7 @@ again: /* the open lock might already be gotten in * ldlm_handle_enqueue() */ LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT); - if (create_flags & MDS_OPEN_LOCK) + 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 */ @@ -1539,20 +1599,20 @@ again: object_locked = 1; if (rc != 0) GOTO(out_child_unlock, result = rc); - else if (create_flags & MDS_OPEN_LOCK) + 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), @@ -1575,14 +1635,23 @@ again: mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE); } } + + mdt_counter_incr(req, LPROC_MDT_OPEN, + ktime_us_delta(ktime_get(), kstart)); + EXIT; out_child_unlock: 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); + if (lh != NULL) + mdt_object_unlock(info, parent, lh, result || !created); + + mdt_object_put(info->mti_env, parent); out: if (result) lustre_msg_set_transno(req->rq_repmsg, 0); @@ -1607,7 +1676,7 @@ static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info, .ln_namelen = sizeof("i_am_nobody") - 1, }; struct lu_ucred *uc; - cfs_cap_t uc_cap_save; + kernel_cap_t uc_cap_save; int rc; ENTRY; @@ -1625,7 +1694,6 @@ static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info, spec->sp_cr_lookup = 0; spec->sp_feat = &dt_directory_features; - spec->sp_cr_mode = MDL_MINMODE; /* no lock */ spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode; if (attr->ma_valid & MA_LOV) { spec->u.sp_ea.eadata = attr->ma_lmm; @@ -1637,7 +1705,7 @@ static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info, uc = lu_ucred(env); uc_cap_save = uc->uc_cap; - uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE; + cap_raise(uc->uc_cap, CAP_DAC_OVERRIDE); rc = mdo_create(env, mdt_object_child(local_root), &lname, mdt_object_child(obj), spec, attr); uc->uc_cap = uc_cap_save; @@ -1647,7 +1715,7 @@ 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 " @@ -1673,13 +1741,13 @@ static inline int mdt_hsm_set_released(struct lov_mds_md *lmm) __u32 off; int i; - if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEF)) { + 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 < le32_to_cpu(comp_v1->lcm_entry_count); i++) { + 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); @@ -1690,6 +1758,22 @@ static inline int mdt_hsm_set_released(struct lov_mds_md *lmm) 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) { @@ -1700,7 +1784,7 @@ static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, struct mdt_object *orphan; struct md_attr *orp_ma; struct lu_buf *buf; - cfs_cap_t cap; + kernel_cap_t cap; bool lease_broken; int rc; int rc2; @@ -1749,19 +1833,66 @@ 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=%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); + 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; @@ -1770,6 +1901,17 @@ 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) @@ -1778,17 +1920,17 @@ 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_*_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_MASK) == LOV_MAGIC_MAGIC) - ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEF); + ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED); else GOTO(out_unlock, rc = -EINVAL); } @@ -1809,7 +1951,7 @@ static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, 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), @@ -1818,7 +1960,7 @@ 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); @@ -1835,7 +1977,7 @@ static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, /* 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); + cap_raise(uc->uc_cap, CAP_FOWNER); rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf, XATTR_NAME_HSM, 0); uc->uc_cap = cap; @@ -1846,6 +1988,12 @@ static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o, rc = mo_swap_layouts(info->mti_env, mdt_object_child(o), mdt_object_child(orphan), SWAP_LAYOUTS_MDS_HSM); + + 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: @@ -1854,7 +2002,7 @@ out_layout_lock: 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); @@ -1872,10 +2020,18 @@ out_unlock: repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY); LASSERT(repbody != NULL); 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->l_policy_data.l_inodebits.bits); LDLM_LOCK_PUT(lease); ma->ma_valid = 0; @@ -1884,16 +2040,16 @@ out_reprocess: return rc; } -int mdt_close_swap_layouts(struct mdt_thread_info *info, - struct mdt_object *o, struct md_attr *ma) +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; + struct mdt_object *o1 = o, *o2 = NULL; bool lease_broken; - bool swap_objects; + bool swap_objects = false; int rc; ENTRY; @@ -1911,37 +2067,52 @@ int mdt_close_swap_layouts(struct mdt_thread_info *info, 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); + if (rc == 0) { + /** + * only MDS_CLOSE_LAYOUT_SPLIT use the same fid to indicate + * mirror deletion, so we'd zero cd_fid, and keeps o2 be NULL. + */ + if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)) + RETURN(-EINVAL); + + /* zero cd_fid to keeps o2 be NULL */ + fid_zero(&data->cd_fid); + } else if (rc < 0) { + /* Exchange o1 and o2, to enforce locking order */ + swap_objects = true; + } 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 (!fid_is_zero(&data->cd_fid)) { + 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 (!mdt_object_exists(o2)) + GOTO(out_obj, rc = -ENOENT); - if (swap_objects) - swap(o1, o2); + if (!S_ISREG(lu_object_attr(&o2->mot_obj))) + 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); + if (o2) { + 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)) { @@ -1971,15 +2142,72 @@ int mdt_close_swap_layouts(struct mdt_thread_info *info, 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); + if (o2) { + 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 */ - rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1), - mdt_object_child(o2), 0); + 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; + + if (o2) { + mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1); + } else { + if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)) { + /* paranoid check again */ + CERROR(DFID + ":only mirror split support NULL o2 object\n", + PFID(mdt_object_fid(o))); + GOTO(out_unlock1, rc = -EINVAL); + } + + /* set NULL mrd_obj for deleting mirror objects */ + mrd.mrd_obj = NULL; + } + + if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) { + mrd.mrd_mirror_id = data->cd_mirror_id; + /* set a small enough blocks in the SoM */ + ma->ma_attr.la_blocks >>= 1; + } + + 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 | + LA_LSIZE | LA_LBLOCKS)) { + int rc2; + enum lustre_som_flags lsf; + + if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) + lsf = SOM_FL_STRICT; + else + lsf = SOM_FL_LAZY; + + mutex_lock(&o->mot_som_mutex); + rc2 = mdt_set_som(info, o, lsf, + 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); @@ -1987,7 +2215,8 @@ int mdt_close_swap_layouts(struct mdt_thread_info *info, out_unlock2: /* Release exclusive LL */ - mdt_object_unlock(info, o2, lh2, 1); + if (o2) + mdt_object_unlock(info, o2, lh2, 1); out_unlock1: mdt_object_unlock(info, o1, lh1, 1); @@ -2005,9 +2234,15 @@ out_unlock_sem: } out_obj: - mdt_object_put(info->mti_env, swap_objects ? o1 : o2); + if (o1 != o) + /* the 2nd object has been used, and swapped to o1 */ + mdt_object_put(info->mti_env, o1); + else if (o2) + /* the 2nd object has been used, and not swapped */ + mdt_object_put(info->mti_env, o2); - ldlm_reprocess_all(lease->l_resource); + ldlm_reprocess_all(lease->l_resource, + lease->l_policy_data.l_inodebits.bits); out_lease: LDLM_LOCK_PUT(lease); @@ -2020,57 +2255,220 @@ out_lease: return rc; } -#define MFD_CLOSED(mode) ((mode) == MDS_FMODE_CLOSED) +static int mdt_close_resync_done(struct mdt_thread_info *info, + struct mdt_object *o, struct md_attr *ma) +{ + struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LOCAL]; + 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 (req_capsule_req_need_swab(info->mti_pill)) + 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_PTR_ARRAY(resync_ids, resync_count); + 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, lhc, &layout); + if (rc) + GOTO(out_unlock, rc); + + mdt_object_unlock(info, o, lhc, 0); + + 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_PTR_ARRAY(resync_ids, resync_count); + +out_reprocess: + ldlm_reprocess_all(lease->l_resource, + lease->l_policy_data.l_inodebits.bits); + 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 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; + + ENTRY; + + open_flags = mfd->mfd_open_flags; + intent = ma->ma_attr_flags & MDS_CLOSE_INTENT; + *ofid = *mdt_object_fid(o); - mode = mfd->mfd_mode; + /* the below message is checked in replay-single.sh test_46 */ + CDEBUG(D_INODE, "%s: %sclosing file handle "DFID" with intent: %llx\n", + mdt_obd_name(info->mti_mdt), + ma->ma_valid & MA_FORCE_LOG ? "force " : "", PFID(ofid), intent); - if (ma->ma_attr_flags & MDS_HSM_RELEASE) { + 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); + PFID(ofid), rc); /* continue to close even error occurred. */ } + break; } - - if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) { - rc = mdt_close_swap_layouts(info, o, ma); + 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", + "%s: cannot swap layout of "DFID": rc = %d\n", mdt_obd_name(info->mti_mdt), - PFID(mdt_object_fid(o)), rc); + 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) + 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 (mode & MDS_FMODE_EXEC) + 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)) { - /* 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); - } + /* Update atime|mtime|ctime on close. */ + 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 || + ma->ma_attr.la_valid & LA_MTIME || + ma->ma_attr.la_valid & LA_CTIME)) { + ma->ma_valid = MA_INODE; + ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES; + ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME); + + if (ma->ma_attr.la_valid & LA_MTIME) { + rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid); + if (!rc) + ma->ma_valid |= MA_PFID; + } + + rc = mo_attr_set(info->mti_env, next, ma); + } /* If file data is modified, add the dirty flag. */ if (ma->ma_attr_flags & MDS_DATA_MODIFIED) @@ -2079,17 +2477,22 @@ 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 (!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); + } /* adjust open and lease count */ - if (mode & MDS_OPEN_LEASE) { + if (open_flags & 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); mdt_mfd_free(mfd); mdt_object_put(info->mti_env, o); @@ -2101,34 +2504,26 @@ int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req, { struct mdt_export_data *med; struct mdt_file_data *mfd; - struct mdt_object *o; - struct md_attr *ma = &info->mti_attr; int ret = 0; int rc = 0; ENTRY; med = &req->rq_export->exp_mdt_data; spin_lock(&med->med_open_lock); - mfd = mdt_handle2mfd(med, &info->mti_close_handle, req_is_replay(req)); + 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_close_handle.cookie); + info->mti_open_handle.cookie); /** not serious error since bug 3633 */ rc = -ESTALE; } else { - class_handle_unhash(&mfd->mfd_handle); + class_handle_unhash(&mfd->mfd_open_handle); list_del_init(&mfd->mfd_list); spin_unlock(&med->med_open_lock); - - /* 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_object_put(info->mti_env, o); } RETURN(rc ? rc : ret); @@ -2140,10 +2535,10 @@ int mdt_close(struct tgt_session_info *tsi) struct ptlrpc_request *req = tgt_ses_req(tsi); struct md_attr *ma = &info->mti_attr; struct mdt_body *repbody = NULL; + ktime_t kstart = ktime_get(); 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) @@ -2176,9 +2571,9 @@ int mdt_close(struct tgt_session_info *tsi) ma->ma_need = MA_INODE | MA_LOV; repbody->mbo_eadatasize = 0; repbody->mbo_aclsize = 0; - } else { - rc = err_serious(rc); - } + } else { + rc = err_serious(rc); + } rc = mdt_close_internal(info, req, repbody); if (rc != -ESTALE) @@ -2198,5 +2593,8 @@ int mdt_close(struct tgt_session_info *tsi) tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP; out: mdt_thread_info_fini(info); + if (rc == 0) + mdt_counter_incr(req, LPROC_MDT_CLOSE, + ktime_us_delta(ktime_get(), kstart)); RETURN(rc ? rc : ret); }