X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmds%2Fmds_open.c;h=9b613a14f085fbc10490fc0b10ceffffefb23e47;hp=5c7fdefc4fbe2ea2046babfc60d332eea9f91850;hb=7a3bf9e15dd53db6b4564a1e660e0d7584b487fc;hpb=51477d038f9901b8871089afa6084a5b12b23542 diff --git a/lustre/mds/mds_open.c b/lustre/mds/mds_open.c index 5c7fdef..9b613a1 100644 --- a/lustre/mds/mds_open.c +++ b/lustre/mds/mds_open.c @@ -7,20 +7,23 @@ * Author: Phil Schwan * Author: Mike Shaver * - * This file is part of Lustre, http://www.lustre.org. + * This file is part of the Lustre file system, http://www.lustre.org + * Lustre is a trademark of Cluster File Systems, Inc. * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * You may have signed or agreed to another license before downloading + * this software. If so, you are bound by the terms and conditions + * of that agreement, and the following does not apply to you. See the + * LICENSE file included with this distribution for more information. * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * If you did not agree to a different license, then this copy of Lustre + * is open source software; you can redistribute it and/or modify it + * under the terms of version 2 of the GNU General Public License as + * published by the Free Software Foundation. * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * In either case, Lustre is distributed in the hope that it will be + * useful, but WITHOUT ANY WARRANTY; without even the implied warranty + * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * license text for more details. */ #ifndef EXPORT_SYMTAB @@ -32,18 +35,13 @@ #include #include #include -#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)) -# include -# include -#else -# include -#endif +#include +#include -#include -#include -#include -#include -#include +#include +#include +#include +#include #include "mds_internal.h" @@ -69,6 +67,9 @@ static void mds_mfd_addref(void *mfdp) atomic_read(&mfd->mfd_refcount)); } +/* Create a new mds_file_data struct. + * One reference for handle+med_open_head list and dropped by mds_mfd_unlink(), + * one reference for the caller of this function. */ struct mds_file_data *mds_mfd_new(void) { struct mds_file_data *mfd; @@ -82,11 +83,14 @@ struct mds_file_data *mds_mfd_new(void) atomic_set(&mfd->mfd_refcount, 2); INIT_LIST_HEAD(&mfd->mfd_handle.h_link); + INIT_LIST_HEAD(&mfd->mfd_list); class_handle_hash(&mfd->mfd_handle, mds_mfd_addref); return mfd; } +/* Get a new reference on the mfd pointed to by handle, if handle is still + * valid. Caller must drop reference with mds_mfd_put(). */ static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle) { ENTRY; @@ -94,6 +98,7 @@ static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle) RETURN(class_handle2object(handle->cookie)); } +/* Drop mfd reference, freeing struct if this is the last one. */ static void mds_mfd_put(struct mds_file_data *mfd) { CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd, @@ -101,23 +106,26 @@ static void mds_mfd_put(struct mds_file_data *mfd) LASSERT(atomic_read(&mfd->mfd_refcount) > 0 && atomic_read(&mfd->mfd_refcount) < 0x5a5a); if (atomic_dec_and_test(&mfd->mfd_refcount)) { - LASSERT(list_empty(&mfd->mfd_handle.h_link)); - OBD_FREE(mfd, sizeof *mfd); + OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle); } } -static void mds_mfd_destroy(struct mds_file_data *mfd) +/* Remove the mfd handle so that it cannot be found by open/close again. + * Caller must hold med_open_lock for mfd_list manipulation. */ +void mds_mfd_unlink(struct mds_file_data *mfd, int decref) { class_handle_unhash(&mfd->mfd_handle); - mds_mfd_put(mfd); + list_del_init(&mfd->mfd_list); + if (decref) + mds_mfd_put(mfd); } /* Caller must hold mds->mds_epoch_sem */ static int mds_alloc_filterdata(struct inode *inode) { - LASSERT(LUSTRE_FILTERDATA(inode) == NULL); - OBD_ALLOC(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data)); - if (LUSTRE_FILTERDATA(inode) == NULL) + LASSERT(inode->i_filterdata == NULL); + OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data)); + if (inode->i_filterdata == NULL) return -ENOMEM; LASSERT(igrab(inode) == inode); return 0; @@ -126,9 +134,9 @@ static int mds_alloc_filterdata(struct inode *inode) /* Caller must hold mds->mds_epoch_sem */ static void mds_free_filterdata(struct inode *inode) { - LASSERT(LUSTRE_FILTERDATA(inode) != NULL); - OBD_FREE(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data)); - LUSTRE_FILTERDATA(inode) = NULL; + LASSERT(inode->i_filterdata != NULL); + OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data)); + inode->i_filterdata = NULL; iput(inode); } @@ -155,6 +163,7 @@ static int mds_get_write_access(struct mds_obd *mds, struct inode *inode, RETURN(-ETXTBSY); } + if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) { CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n", MDS_FILTERDATA(inode)->io_epoch, inode->i_ino, @@ -162,9 +171,9 @@ static int mds_get_write_access(struct mds_obd *mds, struct inode *inode, goto out; } - if (MDS_FILTERDATA(inode) == NULL) + if (inode->i_filterdata == NULL) mds_alloc_filterdata(inode); - if (MDS_FILTERDATA(inode) == NULL) { + if (inode->i_filterdata == NULL) { rc = -ENOMEM; goto out; } @@ -240,37 +249,35 @@ static struct mds_file_data *mds_dentry_open(struct dentry *dentry, struct mds_obd *mds = mds_req2mds(req); struct mds_file_data *mfd; struct mds_body *body; - int rc = 0; + int error; ENTRY; mfd = mds_mfd_new(); if (mfd == NULL) { CERROR("mds: out of memory\n"); - GOTO(cleanup_dentry, rc = -ENOMEM); + GOTO(cleanup_dentry, error = -ENOMEM); } - body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body)); + body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body)); if (flags & FMODE_WRITE) { /* FIXME: in recovery, need to pass old epoch here */ - rc = mds_get_write_access(mds, dentry->d_inode, 0); - if (rc) - GOTO(cleanup_mfd, rc); + error = mds_get_write_access(mds, dentry->d_inode, 0); + if (error) + GOTO(cleanup_mfd, error); body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch; - } else if (flags & FMODE_EXEC) { - rc = mds_deny_write_access(mds, dentry->d_inode); - if (rc) - GOTO(cleanup_mfd, rc); + } else if (flags & MDS_FMODE_EXEC) { + error = mds_deny_write_access(mds, dentry->d_inode); + if (error) + GOTO(cleanup_mfd, error); } dget(dentry); - /* FIXME: invalidate update locks when first open for write comes in */ - - /* mark the file as open to handle open-unlink. */ - DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode); + /* Mark the file as open to handle open-unlink. */ + MDS_DOWN_WRITE_ORPHAN_SEM(dentry->d_inode); mds_orphan_open_inc(dentry->d_inode); - UP_WRITE_I_ALLOC_SEM(dentry->d_inode); + MDS_UP_WRITE_ORPHAN_SEM(dentry->d_inode); mfd->mfd_mode = flags; mfd->mfd_dentry = dentry; @@ -279,119 +286,100 @@ static struct mds_file_data *mds_dentry_open(struct dentry *dentry, spin_lock(&med->med_open_lock); list_add(&mfd->mfd_list, &med->med_open_head); spin_unlock(&med->med_open_lock); - mds_mfd_put(mfd); body->handle.cookie = mfd->mfd_handle.h_cookie; + RETURN(mfd); + cleanup_mfd: mds_mfd_put(mfd); - mds_mfd_destroy(mfd); + mds_mfd_unlink(mfd, 1); cleanup_dentry: - return ERR_PTR(rc); -} - -static inline void -mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm, - struct lov_desc *desc) -{ - int i; - - for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) { - ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] = - le64_to_cpu(lmm->lmm_objects[i].l_object_id); - } + return ERR_PTR(error); } -/* must be called with i_sem held */ -int -mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req, - int offset, struct mds_update_record *rec, - struct dentry *dchild, void **handle, - obd_id **ids) +/* Must be called with i_mutex held */ +static int mds_create_objects(struct ptlrpc_request *req, int offset, + struct mds_update_record *rec, + struct mds_obd *mds, struct obd_device *obd, + struct dentry *dchild, void **handle, + struct lov_mds_md **objid) { struct inode *inode = dchild->d_inode; - struct mds_obd *mds = &obd->u.mds; struct obd_trans_info oti = { 0 }; - struct lov_stripe_md *lsm = NULL; struct lov_mds_md *lmm = NULL; - int rc, lmm_bufsize, lmm_size; - struct obdo *oa = NULL; + int rc, lmm_size; struct mds_body *body; + struct obd_info oinfo = { { { 0 } } }; void *lmm_buf; ENTRY; - if (rec->ur_flags & MDS_OPEN_DELAY_CREATE || - !(rec->ur_flags & FMODE_WRITE)) + if (!S_ISREG(inode->i_mode)) + RETURN(0); + if (!md_should_create(rec->ur_flags)) RETURN(0); - body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body)); + body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body)); - if (!S_ISREG(inode->i_mode)) - RETURN(0); if (body->valid & OBD_MD_FLEASIZE) RETURN(0); - OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids)); - if (*ids == NULL) - RETURN(-ENOMEM); - oti.oti_objid = *ids; - + oti_init(&oti, req); + /* replay case */ if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { - LASSERT(id_ino(rec->ur_id2)); + if (rec->ur_fid2->id == 0) { + DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay"); + RETURN(-EFAULT); + } + body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE; lmm_size = rec->ur_eadatalen; lmm = rec->ur_eadata; LASSERT(lmm); if (*handle == NULL) - *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL); + *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL); if (IS_ERR(*handle)) { rc = PTR_ERR(*handle); *handle = NULL; - RETURN(rc); + GOTO(out_ids, rc); } - mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc); - - lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0); - lmm_bufsize = req->rq_repmsg->buflens[offset]; - LASSERT(lmm_buf != NULL); - LASSERT(lmm_bufsize >= lmm_size); - - memcpy(lmm_buf, lmm, lmm_size); - rc = fsfilt_set_md(obd, inode, *handle, lmm, - lmm_size, EA_LOV); + rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov"); if (rc) CERROR("open replay failed to set md:%d\n", rc); - RETURN(0); + lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size); + LASSERT(lmm_buf); + memcpy(lmm_buf, lmm, lmm_size); + + *objid = lmm_buf; + RETURN(rc); } - if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO)) + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_ALLOC_OBDO)) GOTO(out_ids, rc = -ENOMEM); - oa = obdo_alloc(); - if (oa == NULL) - RETURN(-ENOMEM); - oa->o_mode = S_IFREG | 0600; - oa->o_id = inode->i_ino; - oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num; - oa->o_generation = inode->i_generation; - oa->o_uid = 0; /* must have 0 uid / gid on OST */ - oa->o_gid = 0; - oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE | - OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP; - oa->o_size = 0; - - obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME | + OBDO_ALLOC(oinfo.oi_oa); + if (oinfo.oi_oa == NULL) + GOTO(out_ids, rc = -ENOMEM); + oinfo.oi_oa->o_uid = 0; /* must have 0 uid / gid on OST */ + oinfo.oi_oa->o_gid = 0; + oinfo.oi_oa->o_mode = S_IFREG | 0600; + oinfo.oi_oa->o_id = inode->i_ino; + oinfo.oi_oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLFLAGS | + OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID; + oinfo.oi_oa->o_size = 0; + + obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME); if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) { /* check if things like lfs setstripe are sending us the ea */ if (rec->ur_flags & MDS_OPEN_HAS_EA) { rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, - mds->mds_dt_exp, - 0, &lsm, rec->ur_eadata); + mds->mds_osc_exp, + 0, &oinfo.oi_md, rec->ur_eadata); if (rc) GOTO(out_oa, rc); } else { @@ -401,24 +389,23 @@ mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req, lmm_size = mds->mds_max_mdsize; rc = mds_get_md(obd, dchild->d_parent->d_inode, - lmm, &lmm_size, 1, 0); + lmm, &lmm_size, 1); if (rc > 0) rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, - mds->mds_dt_exp, - 0, &lsm, lmm); + mds->mds_osc_exp, + 0, &oinfo.oi_md, lmm); OBD_FREE(lmm, mds->mds_max_mdsize); if (rc) GOTO(out_oa, rc); } - - LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS); - oti.oti_flags |= OBD_MODE_CROW; - rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti); - + rc = obd_create(mds->mds_osc_exp, oinfo.oi_oa, + &oinfo.oi_md, &oti); if (rc) { - CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR), - "error creating objects for " - "inode %lu: rc = %d\n", + int level = D_ERROR; + if (rc == -ENOSPC) + level = D_INODE; + CDEBUG(level, "error creating objects for " + "inode %lu: rc = %d\n", inode->i_ino, rc); if (rc > 0) { CERROR("obd_create returned invalid " @@ -428,25 +415,32 @@ mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req, GOTO(out_oa, rc); } } else { - rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp, - 0, &lsm, rec->ur_eadata); - if (rc) + rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp, + 0, &oinfo.oi_md, rec->ur_eadata); + if (rc) { GOTO(out_oa, rc); - - lsm->lsm_object_id = oa->o_id; - lsm->lsm_object_gr = oa->o_gr; + } + oinfo.oi_md->lsm_object_id = oinfo.oi_oa->o_id; + oinfo.oi_md->lsm_object_gr = oinfo.oi_oa->o_gr; } - if (inode->i_size) { - oa->o_size = inode->i_size; - obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME | - OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE); - - rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti); + if (i_size_read(inode)) { + oinfo.oi_oa->o_size = i_size_read(inode); + obdo_from_inode(oinfo.oi_oa, inode, OBD_MD_FLTYPE | + OBD_MD_FLATIME | OBD_MD_FLMTIME | + OBD_MD_FLCTIME | OBD_MD_FLSIZE); + + /* pack lustre id to OST */ + oinfo.oi_oa->o_fid = body->fid1.id; + oinfo.oi_oa->o_generation = body->fid1.generation; + oinfo.oi_oa->o_valid |= OBD_MD_FLFID | OBD_MD_FLGENER; + + rc = obd_setattr_rqset(mds->mds_osc_exp, &oinfo, &oti); if (rc) { CERROR("error setting attrs for inode %lu: rc %d\n", inode->i_ino, rc); if (rc > 0) { - CERROR("obd_setattr returned bad rc %d\n", rc); + CERROR("obd_setattr_async returned bad rc %d\n", + rc); rc = -EIO; } GOTO(out_oa, rc); @@ -454,18 +448,15 @@ mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req, } body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE; - obdo_refresh_inode(inode, oa, OBD_MD_FLBLKSZ); + obdo_refresh_inode(inode, oinfo.oi_oa, OBD_MD_FLBLKSZ); - LASSERT(lsm && lsm->lsm_object_id); + LASSERT(oinfo.oi_md && oinfo.oi_md->lsm_object_id); lmm = NULL; - rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm); - if (!id_ino(rec->ur_id2)) - obd_free_memmd(mds->mds_dt_exp, &lsm); + rc = obd_packmd(mds->mds_osc_exp, &lmm, oinfo.oi_md); if (rc < 0) { CERROR("cannot pack lsm, err = %d\n", rc); GOTO(out_oa, rc); } - lmm_size = rc; body->eadatasize = rc; @@ -474,82 +465,27 @@ mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req, if (IS_ERR(*handle)) { rc = PTR_ERR(*handle); *handle = NULL; - GOTO(out_oa, rc); + GOTO(free_diskmd, rc); } - rc = fsfilt_set_md(obd, inode, *handle, lmm, - lmm_size, EA_LOV); - - lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0); - lmm_bufsize = req->rq_repmsg->buflens[offset]; + rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size, "lov"); + lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, lmm_size); LASSERT(lmm_buf); - LASSERT(lmm_bufsize >= lmm_size); - memcpy(lmm_buf, lmm, lmm_size); - obd_free_diskmd(mds->mds_dt_exp, &lmm); + + *objid = lmm_buf; // save for mds_lov_update_objid + +free_diskmd: + obd_free_diskmd(mds->mds_osc_exp, &lmm); out_oa: oti_free_cookies(&oti); - obdo_free(oa); + OBDO_FREE(oinfo.oi_oa); out_ids: - if (rc) { - OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids)); - *ids = NULL; - } - if (lsm) - obd_free_memmd(mds->mds_dt_exp, &lsm); + if (oinfo.oi_md) + obd_free_memmd(mds->mds_osc_exp, &oinfo.oi_md); RETURN(rc); } -int -mds_destroy_object(struct obd_device *obd, - struct inode *inode, int async) -{ - struct mds_obd *mds = &obd->u.mds; - struct lov_mds_md *lmm = NULL; - int rc, lmm_size; - ENTRY; - - LASSERT(inode != NULL); - - if (inode->i_nlink != 0) { - CDEBUG(D_INODE, "attempt to destroy OSS object when " - "i_nlink == %d\n", (int)inode->i_nlink); - RETURN(0); - } - - OBD_ALLOC(lmm, mds->mds_max_mdsize); - if (lmm == NULL) - RETURN(-ENOMEM); - - lmm_size = mds->mds_max_mdsize; - rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0); - if (rc < 0) { - CERROR("no stripe info for %lu/%lu inode\n", - (unsigned long)inode->i_ino, - (unsigned long)inode->i_generation); - GOTO(out_free_lmm, rc); - } - - if (rc > 0) { - /* asynchronously unlink objecect on OSS */ - rc = mds_unlink_object(mds, inode, lmm, lmm_size, - NULL, 0, async); - if (rc) { - CERROR("error unlinking object on OSS, " - "err %d\n", rc); - GOTO(out_free_lmm, rc); - } - } else { - CDEBUG(D_INODE, "no stripping info found for inode " - "%lu/%lu\n", (unsigned long)inode->i_ino, - (unsigned long)inode->i_generation); - } - EXIT; -out_free_lmm: - OBD_FREE(lmm, mds->mds_max_mdsize); - return rc; -} - static void reconstruct_open(struct mds_update_record *rec, int offset, struct ptlrpc_request *req, struct lustre_handle *child_lockh) @@ -559,21 +495,21 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, struct mds_obd *mds = mds_req2mds(req); struct mds_file_data *mfd; struct obd_device *obd = req->rq_export->exp_obd; - struct dentry *parent = NULL, *dchild; + struct dentry *parent, *dchild; struct ldlm_reply *rep; struct mds_body *body; + int rc; struct list_head *t; int put_child = 1; - int rc; ENTRY; - LASSERT(offset == 3); /* only called via intent */ - rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep)); - body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body)); + LASSERT(offset == DLM_INTENT_REC_OFF); /* only called via intent */ + rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep)); + body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, sizeof(*body)); /* copy rc, transno and disp; steal locks */ mds_req_from_mcd(req, mcd); - intent_set_disposition(rep, mcd->mcd_last_data); + intent_set_disposition(rep, le32_to_cpu(mcd->mcd_last_data)); /* Only replay if create or open actually happened. */ if (!intent_disposition(rep, DISP_OPEN_CREATE | DISP_OPEN_OPEN) ) { @@ -581,35 +517,29 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, return; /* error looking up parent or child */ } - if (rec->ur_namelen == 1) { - CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n", - OLID4(rec->ur_id1)); - dchild = mds_id2dentry(obd, rec->ur_id1, NULL); - } else { - parent = mds_id2dentry(obd, rec->ur_id1, NULL); - LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n", - OLID4(rec->ur_id1), PTR_ERR(parent)); - dchild = ll_lookup_one_len(rec->ur_name, parent, - rec->ur_namelen - 1); - LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n", - OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild)); - } + parent = mds_fid2dentry(mds, rec->ur_fid1, NULL); + LASSERT(!IS_ERR(parent)); + + dchild = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1); + LASSERT(!IS_ERR(dchild)); if (!dchild->d_inode) GOTO(out_dput, 0); /* child not present to open */ /* At this point, we know we have a child. We'll send - * it back _unless_ it not created and open failed. */ + * it back _unless_ it not created and open failed. + */ if (intent_disposition(rep, DISP_OPEN_OPEN) && !intent_disposition(rep, DISP_OPEN_CREATE) && - req->rq_status) + req->rq_status) { GOTO(out_dput, 0); + } - /* get lock (write for O_CREAT, read otherwise) */ - mds_pack_inode2body(obd, body, dchild->d_inode, 1); + mds_pack_inode2fid(&body->fid1, dchild->d_inode); + mds_pack_inode2body(body, dchild->d_inode); if (S_ISREG(dchild->d_inode->i_mode)) { - rc = mds_pack_md(obd, req->rq_repmsg, 2, body, - dchild->d_inode, 1, 0); + rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1, + body, dchild->d_inode, 1); if (rc) LASSERT(rc == req->rq_status); @@ -619,7 +549,22 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLATIME | OBD_MD_FLMTIME); } - + + if (!(rec->ur_flags & MDS_OPEN_JOIN_FILE)) + lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1, + body->eadatasize, 0); + + if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL && + !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) { + int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1); + + rc = mds_pack_acl(med, dchild->d_inode, req->rq_repmsg, + body, acl_off); + lustre_shrink_reply(req, acl_off, body->aclsize, 0); + if (!req->rq_status && rc) + req->rq_status = rc; + } + /* If we have -EEXIST as the status, and we were asked to create * exclusively, we can tell we failed because the file already existed. */ @@ -642,18 +587,27 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, GOTO(out_dput, 0); mfd = NULL; + spin_lock(&med->med_open_lock); list_for_each(t, &med->med_open_head) { mfd = list_entry(t, struct mds_file_data, mfd_list); - if (mfd->mfd_xid == req->rq_xid) + if (mfd->mfd_xid == req->rq_xid) { + mds_mfd_addref(mfd); break; + } mfd = NULL; } + spin_unlock(&med->med_open_lock); /* #warning "XXX fixme" bug 2991 */ /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL. * Now that exp_outstanding_reply is a list, it's just using mfd != NULL * to detect a re-open */ if (mfd == NULL) { + if (rec->ur_flags & MDS_OPEN_JOIN_FILE) { + rc = mds_join_file(rec, req, dchild, NULL); + if (rc) + GOTO(out_dput, rc); + } mntget(mds->mds_vfsmnt); CERROR("Re-opened file \n"); mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, @@ -669,24 +623,33 @@ static void reconstruct_open(struct mds_update_record *rec, int offset, mfd->mfd_handle.h_cookie); } + mds_mfd_put(mfd); + out_dput: if (put_child) l_dput(dchild); - if (parent) - l_dput(parent); + l_dput(parent); EXIT; } /* do NOT or the MAY_*'s, you'll get the weakest */ -static int accmode(int flags) +static int accmode(struct inode *inode, int flags) { int res = 0; + /* Sadly, NFSD reopens a file repeatedly during operation, so the + * "acc_mode = 0" allowance for newly-created files isn't honoured. + * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file + * owner can write to a file even if it is marked readonly to hide + * its brokenness. (bug 5781) */ + if (flags & MDS_OPEN_OWNEROVERRIDE && inode->i_uid == current->fsuid) + return 0; + if (flags & FMODE_READ) res = MAY_READ; if (flags & (FMODE_WRITE|MDS_OPEN_TRUNC)) res |= MAY_WRITE; - if (flags & FMODE_EXEC) + if (flags & MDS_FMODE_EXEC) res = MAY_EXEC; return res; } @@ -694,86 +657,76 @@ static int accmode(int flags) /* Handles object creation, actual opening, and I/O epoch */ static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild, struct mds_body *body, int flags, void **handle, - struct mds_update_record *rec, struct ldlm_reply *rep) + struct mds_update_record *rec,struct ldlm_reply *rep, + struct lustre_handle *lockh) { - struct obd_device *obd = req->rq_export->exp_obd; struct mds_obd *mds = mds_req2mds(req); + struct obd_device *obd = req->rq_export->exp_obd; struct mds_file_data *mfd = NULL; - obd_id *ids = NULL; - unsigned mode; - int rc = 0, reply_off; + struct lov_mds_md *lmm; /* object IDs created */ + int rc = 0; ENTRY; /* atomically create objects if necessary */ - down(&dchild->d_inode->i_sem); - mode = dchild->d_inode->i_mode; + LOCK_INODE_MUTEX(dchild->d_inode); - if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) || - (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) { - rc = mds_pack_md(obd, req->rq_repmsg, 2, body, - dchild->d_inode, 0, 0); + if (S_ISREG(dchild->d_inode->i_mode) && + !(body->valid & OBD_MD_FLEASIZE)) { + rc = mds_pack_md(obd, req->rq_repmsg, DLM_REPLY_REC_OFF + 1, + body, dchild->d_inode, 0); if (rc) { - up(&dchild->d_inode->i_sem); + UNLOCK_INODE_MUTEX(dchild->d_inode); RETURN(rc); } } - if (rec != NULL) { if ((body->valid & OBD_MD_FLEASIZE) && (rec->ur_flags & MDS_OPEN_HAS_EA)) { - up(&dchild->d_inode->i_sem); + UNLOCK_INODE_MUTEX(dchild->d_inode); RETURN(-EEXIST); } - - if (!(body->valid & OBD_MD_FLEASIZE)) { - /* no EA: create objects */ - rc = mds_create_objects(obd, req, 2, rec, - dchild, handle, &ids); - if (rc) { - CERROR("mds_create_object: rc = %d\n", rc); - up(&dchild->d_inode->i_sem); + if (rec->ur_flags & MDS_OPEN_JOIN_FILE) { + UNLOCK_INODE_MUTEX(dchild->d_inode); + rc = mds_join_file(rec, req, dchild, lockh); + if (rc) RETURN(rc); - } + LOCK_INODE_MUTEX(dchild->d_inode); } - - if (S_ISREG(dchild->d_inode->i_mode) && - (body->valid & OBD_MD_FLEASIZE)) { - rc = mds_revalidate_lov_ea(obd, dchild->d_inode, - req->rq_repmsg, 2); - if (!rc) - rc = mds_pack_md(obd, req->rq_repmsg, 2, body, - dchild->d_inode, 0, 0); + if (!(body->valid & OBD_MD_FLEASIZE) && + !(body->valid & OBD_MD_FLMODEASIZE)) { + /* no EA: create objects */ + rc = mds_create_objects(req, DLM_REPLY_REC_OFF + 1, rec, + mds, obd, dchild, handle, &lmm); if (rc) { - up(&dchild->d_inode->i_sem); + CERROR("mds_create_objects: rc = %d\n", rc); + UNLOCK_INODE_MUTEX(dchild->d_inode); RETURN(rc); } } } - - reply_off = 3; - rc = mds_pack_acl(req, &reply_off, body, dchild->d_inode); - - if (rc < 0) { - CERROR("pack posix acl: rc = %d\n", rc); - up(&dchild->d_inode->i_sem); - RETURN(rc); - } - - rc = mds_pack_gskey(obd, req->rq_repmsg, &reply_off, body, - dchild->d_inode); - if (rc < 0) { - CERROR("mds_pack_gskey: rc = %d\n", rc); - up(&dchild->d_inode->i_sem); - RETURN(rc); - } - /* If the inode has no EA data, then MDSs hold size, mtime */ + /* If the inode has no EA data, then MDS holds size, mtime */ if (S_ISREG(dchild->d_inode->i_mode) && !(body->valid & OBD_MD_FLEASIZE)) { body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLATIME | OBD_MD_FLMTIME); } + UNLOCK_INODE_MUTEX(dchild->d_inode); + + if (rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) + lustre_shrink_reply(req, DLM_REPLY_REC_OFF + 1, + body->eadatasize, 0); + + if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL && + rec && !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) { + int acl_off = DLM_REPLY_REC_OFF + (body->eadatasize ? 2 : 1); - up(&dchild->d_inode->i_sem); + rc = mds_pack_acl(&req->rq_export->exp_mds_data, + dchild->d_inode, req->rq_repmsg, + body, acl_off); + lustre_shrink_reply(req, acl_off, body->aclsize, 0); + if (rc) + RETURN(rc); + } intent_set_disposition(rep, DISP_OPEN_OPEN); mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req); @@ -783,73 +736,59 @@ static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild, CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd, mfd->mfd_handle.h_cookie); - if (ids != NULL) { - mds_dt_update_objids(obd, ids); - OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count); - } + if (lmm != NULL) + mds_lov_update_objids(obd, lmm); + if (rc) /* coverity[deadcode] */ + mds_mfd_unlink(mfd, 1); + + mds_mfd_put(mfd); RETURN(rc); } -static int mds_open_by_id(struct ptlrpc_request *req, - struct lustre_id *id, - struct mds_body *body, int flags, - struct mds_update_record *rec, - struct ldlm_reply *rep) +static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid, + struct mds_body *body, int flags, + struct mds_update_record *rec,struct ldlm_reply *rep) { struct mds_obd *mds = mds_req2mds(req); - struct inode *pending_dir = mds->mds_pending_dir->d_inode; struct dentry *dchild; - char idname[LL_ID_NAMELEN]; - int idlen = 0, rc; + char fidname[LL_FID_NAMELEN]; + int fidlen = 0, rc; void *handle = NULL; ENTRY; - down(&pending_dir->i_sem); - idlen = ll_id2str(idname, id_ino(id), id_gen(id)); - dchild = lookup_one_len(idname, mds->mds_pending_dir, - idlen); + fidlen = ll_fid2str(fidname, fid->id, fid->generation); + dchild = ll_lookup_one_len(fidname, mds->mds_pending_dir, fidlen); if (IS_ERR(dchild)) { - up(&pending_dir->i_sem); rc = PTR_ERR(dchild); - CERROR("error looking up %s in PENDING: rc = %d\n", - idname, rc); + CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc); RETURN(rc); } - up(&pending_dir->i_sem); if (dchild->d_inode != NULL) { mds_inode_set_orphan(dchild->d_inode); - mds_pack_inode2body(req2obd(req), body, - dchild->d_inode, 1); - - intent_set_disposition(rep, DISP_LOOKUP_EXECD); - intent_set_disposition(rep, DISP_LOOKUP_POS); CWARN("Orphan %s found and opened in PENDING directory\n", - idname); - goto open; - } - l_dput(dchild); + fidname); + } else { + l_dput(dchild); - /* - * we didn't find it in PENDING so it isn't an orphan. See if it was a - * regular inode that was previously created. - */ - dchild = mds_id2dentry(req2obd(req), id, NULL); - if (IS_ERR(dchild)) - RETURN(PTR_ERR(dchild)); + /* We didn't find it in PENDING so it isn't an orphan. See + * if it was a regular inode that was previously created. */ + dchild = mds_fid2dentry(mds, fid, NULL); + if (IS_ERR(dchild)) + RETURN(PTR_ERR(dchild)); + } - mds_pack_inode2body(req2obd(req), body, - dchild->d_inode, 1); - + mds_pack_inode2fid(&body->fid1, dchild->d_inode); + mds_pack_inode2body(body, dchild->d_inode); intent_set_disposition(rep, DISP_LOOKUP_EXECD); intent_set_disposition(rep, DISP_LOOKUP_POS); - open: - rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep); - rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle, - req, rc, rep ? rep->lock_policy_res1 : 0); + rc = mds_finish_open(req, dchild, body, flags, &handle, rec, rep, NULL); + rc = mds_finish_transno(mds, dchild->d_inode, handle, + req, rc, rep ? rep->lock_policy_res1 : 0, 0); /* XXX what do we do here if mds_finish_transno itself failed? */ + l_dput(dchild); RETURN(rc); } @@ -857,32 +796,32 @@ static int mds_open_by_id(struct ptlrpc_request *req, int mds_pin(struct ptlrpc_request *req, int offset) { struct obd_device *obd = req->rq_export->exp_obd; - struct mds_body *request_body, *reply_body; + struct mds_body *reqbody, *repbody; struct lvfs_run_ctxt saved; - int rc, size = sizeof(*reply_body); + int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) }; ENTRY; - request_body = lustre_msg_buf(req->rq_reqmsg, offset, - sizeof(*request_body)); + reqbody = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*reqbody)); - rc = lustre_pack_reply(req, 1, &size, NULL); + rc = lustre_pack_reply(req, 2, size, NULL); if (rc) RETURN(rc); - reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body)); + + repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, + sizeof(*repbody)); push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - rc = mds_open_by_id(req, &request_body->id1, reply_body, - request_body->flags, NULL, NULL); + rc = mds_open_by_fid(req, &reqbody->fid1, repbody, reqbody->flags, NULL, + NULL); pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); RETURN(rc); } -/* - * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If - * child_lockh is NULL we just get the lock as a barrier to wait for other - * holders of this lock, and drop it right away again. - */ +/* Get an internal lock on the inode number (but not generation) to sync + * new inode creation with inode unlink (bug 2029). If child_lockh is NULL + * we just get the lock as a barrier to wait for other holders of this lock, + * and drop it right away again. */ int mds_lock_new_child(struct obd_device *obd, struct inode *inode, struct lustre_handle *child_lockh) { @@ -890,17 +829,16 @@ int mds_lock_new_child(struct obd_device *obd, struct inode *inode, struct lustre_handle lockh; int lock_flags = LDLM_FL_ATOMIC_CB; int rc; - ENTRY; if (child_lockh == NULL) child_lockh = &lockh; - rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id, - LDLM_PLAIN, NULL, LCK_EX, &lock_flags, - mds_blocking_ast, ldlm_completion_ast, NULL, - NULL, NULL, 0, NULL, child_lockh); + rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id, + LDLM_PLAIN, NULL, LCK_EX, &lock_flags, + ldlm_blocking_ast, ldlm_completion_ast, + NULL, NULL, 0, NULL, child_lockh); if (rc != ELDLM_OK) - CERROR("ldlm_cli_enqueue: %d\n", rc); + CERROR("ldlm_cli_enqueue_local: %d\n", rc); else if (child_lockh == &lockh) ldlm_lock_decref(child_lockh, LCK_EX); @@ -912,35 +850,41 @@ int mds_open(struct mds_update_record *rec, int offset, { /* XXX ALLOCATE _something_ - 464 bytes on stack here */ struct obd_device *obd = req->rq_export->exp_obd; - struct mds_export_data *med = &req->rq_export->u.eu_mds_data; struct mds_obd *mds = mds_req2mds(req); struct ldlm_reply *rep = NULL; struct mds_body *body = NULL; struct dentry *dchild = NULL, *dparent = NULL; - struct lustre_handle parent_lockh[2] = {{0}, {0}}; + struct mds_export_data *med; + struct lustre_handle parent_lockh; int rc = 0, cleanup_phase = 0, acc_mode, created = 0; - int parent_mode = LCK_PR; + int parent_mode = LCK_CR; void *handle = NULL; - struct dentry_params dp; - struct mea *mea = NULL; - int mea_size, update_mode; - int child_mode = LCK_PR; - struct lustre_id sid; - __u64 fid = 0; + struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT; + unsigned int qcids[MAXQUOTAS] = { current->fsuid, current->fsgid }; + unsigned int qpids[MAXQUOTAS] = { 0, 0 }; + int child_mode = LCK_CR; + /* Always returning LOOKUP lock if open succesful to guard + dentry on client. */ + ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}}; + struct ldlm_res_id child_res_id = { .name = {0}}; + int lock_flags = 0; ENTRY; - DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o", - OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name, - rec->ur_mode); - - OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PAUSE_OPEN | OBD_FAIL_ONCE, - (obd_timeout + 1) / 4); - - if (offset == 3) { /* intent */ - rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep)); - body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body)); - } else if (offset == 1) { /* non-intent reint */ - body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body)); + mds_counter_incr(req->rq_export, LPROC_MDS_OPEN); + + OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE, + (obd_timeout + 1) / 4); + + CLASSERT(MAXQUOTAS < 4); + if (offset == DLM_INTENT_REC_OFF) { /* intent */ + rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, + sizeof(*rep)); + body = lustre_msg_buf(req->rq_repmsg, DLM_REPLY_REC_OFF, + sizeof(*body)); + } else if (offset == REQ_REC_OFF) { /* non-intent reint */ + body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, + sizeof(*body)); + LBUG(); /* XXX: not supported yet? */ } else { body = NULL; LBUG(); @@ -948,73 +892,67 @@ int mds_open(struct mds_update_record *rec, int offset, MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh)); - /* - * Step 0: If we are passed a id, then we assume the client already - * opened this file and is only replaying the RPC, so we open the inode - * by fid (at some large expense in security). - */ - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { - CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n", - OLID4(rec->ur_id2), rec->ur_namelen - 1, - rec->ur_name, rec->ur_mode); - - LASSERT(id_ino(rec->ur_id2)); + /* Step 0: If we are passed a fid, then we assume the client already + * opened this file and is only replaying the RPC, so we open the + * inode by fid (at some large expense in security). */ + /*XXX liblustre use mds_open_by_fid to implement LL_IOC_LOV_SETSTRIPE */ + if (((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) || + (req->rq_export->exp_libclient && rec->ur_flags&MDS_OPEN_HAS_EA))&& + !(rec->ur_flags & MDS_OPEN_JOIN_FILE)) { + if (rec->ur_fid2->id == 0) { + struct ldlm_lock *lock = ldlm_handle2lock(child_lockh); + if (lock) { + LDLM_ERROR(lock, "fid2 not set on open replay"); + LDLM_LOCK_PUT(lock); + } + DEBUG_REQ(D_ERROR, req, "fid2 not set on open replay"); + RETURN(-EFAULT); + } - rc = mds_open_by_id(req, rec->ur_id2, body, - rec->ur_flags, rec, rep); + rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags, + rec, rep); if (rc != -ENOENT) { - mds_body_do_reverse_map(med, body); + if (req->rq_export->exp_libclient && + rec->ur_flags & MDS_OPEN_HAS_EA) + RETURN(0); + RETURN(rc); } /* We didn't find the correct inode on disk either, so we * need to re-create it via a regular replay. */ - LASSERT(rec->ur_flags & MDS_OPEN_CREAT); - } else { - LASSERT(!id_ino(rec->ur_id2)); + if (!(rec->ur_flags & MDS_OPEN_CREAT)) { + DEBUG_REQ(D_ERROR, req,"OPEN_CREAT not in open replay"); + RETURN(-EFAULT); + } + } else if (rec->ur_fid2->id) { + DEBUG_REQ(D_ERROR, req, "fid2 "LPU64"/%u on open non-replay", + rec->ur_fid2->id, rec->ur_fid2->generation); + RETURN(-EFAULT); } - LASSERT(offset == 3); /* If we got here, we must be called via intent */ + /* If we got here, we must be called via intent */ + LASSERT(offset == DLM_INTENT_REC_OFF); + med = &req->rq_export->exp_mds_data; if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) { CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n"); RETURN(-ENOMEM); } - acc_mode = accmode(rec->ur_flags); - /* Step 1: Find and lock the parent */ - if (rec->ur_flags & MDS_OPEN_CREAT) { - /* XXX Well, in fact we only need this lock mode change if - in addition to O_CREAT, the file does not exist. - But we do not know if it exists or not yet */ - parent_mode = LCK_PW; - } - - if (rec->ur_namelen == 1) { - /* client (LMV) wants to open the file by id */ - CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1)); - dchild = mds_id2dentry(obd, rec->ur_id1, NULL); - if (IS_ERR(dchild)) { - rc = PTR_ERR(dparent); - CERROR("child lookup by an id error %d\n", rc); - GOTO(cleanup, rc); - } - goto got_child; - } - -restart: - dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode, - parent_lockh, &update_mode, rec->ur_name, - rec->ur_namelen - 1, MDS_INODELOCK_UPDATE); + if (rec->ur_flags & (MDS_OPEN_CREAT | MDS_OPEN_JOIN_FILE)) + parent_mode = LCK_EX; + dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode, + &parent_lockh, MDS_INODELOCK_UPDATE); if (IS_ERR(dparent)) { rc = PTR_ERR(dparent); if (rc != -ENOENT) { - CERROR("parent lookup for "DLID4" failed, error %d\n", - OLID4(rec->ur_id1), rc); + CERROR("parent "LPU64"/%u lookup error %d\n", + rec->ur_fid1->id, rec->ur_fid1->generation, rc); } else { /* Just cannot find parent - make it look like - usual negative lookup to avoid extra MDS RPC */ + * usual negative lookup to avoid extra MDS RPC */ intent_set_disposition(rep, DISP_LOOKUP_EXECD); intent_set_disposition(rep, DISP_LOOKUP_NEG); } @@ -1024,142 +962,56 @@ restart: cleanup_phase = 1; /* parent dentry and lock */ - /* get parent id: ldlm lock on the parent protects ea */ - rc = mds_read_inode_sid(obd, dparent->d_inode, &sid); - if (rc) { - CERROR("can't read parent inode id. ino(%lu) rc(%d)\n", - dparent->d_inode->i_ino, rc); - GOTO(cleanup, rc); - } - - /* try to retrieve MEA data for this dir */ - rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size); - if (rc) - GOTO(cleanup, rc); - - if (mea != NULL && mea->mea_count) { - /* - * dir is already splitted, check is requested filename should * - * live at this MDS or at another one. - */ - int i; - i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1); - if (mea->mea_master != id_group(&mea->mea_ids[i])) { - CDEBUG(D_OTHER, - "%s: inapropriate MDS(%d) for %lu/%u:%s." - " should be %lu(%d)\n", obd->obd_name, - mea->mea_master, dparent->d_inode->i_ino, - dparent->d_inode->i_generation, rec->ur_name, - (unsigned long)id_group(&mea->mea_ids[i]), i); - GOTO(cleanup, rc = -ERESTART); - } + if (rec->ur_flags & MDS_OPEN_JOIN_FILE) { + dchild = dget(dparent); + cleanup_phase = 2; /* child dentry */ + acc_mode = accmode(dchild->d_inode, rec->ur_flags); + GOTO(found_child, rc); } /* Step 2: Lookup the child */ - dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1); + + if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) && + (rec->ur_flags & MDS_OPEN_LOCK) && (rec->ur_namelen == 1)) { + /* hack for nfsd with no_subtree_check, it will use anon + * dentry w/o filename to open the file. the anon dentry's + * parent was set to itself, so rec->ur_fid1 is the file. + * And in MDC it cannot derive the dentry's parent dentry, + * hence the file's name, so we hack here in MDS, + * refer to bug 13030. */ + dchild = mds_fid2dentry(mds, rec->ur_fid1, NULL); + } else { + dchild = ll_lookup_one_len(rec->ur_name, dparent, + rec->ur_namelen - 1); + } if (IS_ERR(dchild)) { rc = PTR_ERR(dchild); dchild = NULL; /* don't confuse mds_finish_transno() below */ GOTO(cleanup, rc); } -got_child: cleanup_phase = 2; /* child dentry */ - if (dchild->d_flags & DCACHE_CROSS_REF) { - CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1)); - LASSERT(rec->ur_namelen > 1); - - /* we're gonna acquire LOOKUP lock on the child, - * but we have already locked parent and our order - * may conflict with enqueue_order_locks(). so, - * drop parent lock and acquire both the locks in - * common order. bug 6190 */ -#ifdef S_PDIROPS - if (parent_lockh[1].cookie != 0) - ldlm_lock_decref(parent_lockh + 1, update_mode); -#endif - ldlm_lock_decref(parent_lockh, parent_mode); - l_dput(dchild); - l_dput(dparent); - cleanup_phase = 0; - - rc = mds_get_parent_child_locked(obd, mds, rec->ur_id1, - parent_lockh, &dparent, - LCK_PR, MDS_INODELOCK_UPDATE, - &update_mode, rec->ur_name, - rec->ur_namelen, child_lockh, - &dchild, LCK_PR, - MDS_INODELOCK_LOOKUP); - - if (rc) - GOTO(cleanup, rc); - -#ifdef S_PDIROPS - if (parent_lockh[1].cookie != 0) - ldlm_lock_decref(parent_lockh + 1, update_mode); -#endif - ldlm_lock_decref(parent_lockh, LCK_PR); - - if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) { - CDEBUG(D_OTHER, "race: name changed (%p)\n", - dchild->d_inode); - if (dchild->d_inode) - ldlm_lock_decref(child_lockh, LCK_PR); - l_dput(dchild); - l_dput(dparent); - GOTO(restart, rc); - } - - mds_pack_dentry2body(obd, body, dchild, 1); - intent_set_disposition(rep, DISP_LOOKUP_POS); - intent_set_disposition(rep, DISP_LOOKUP_EXECD); - - if (mea) - OBD_FREE(mea, mea_size); - l_dput(dchild); - l_dput(dparent); - RETURN(rc); - } - intent_set_disposition(rep, DISP_LOOKUP_EXECD); - if (dchild->d_inode) intent_set_disposition(rep, DISP_LOOKUP_POS); else intent_set_disposition(rep, DISP_LOOKUP_NEG); - /* Step 3: If the child was negative, and we're supposed to, create it.*/ + /*Step 3: If the child was negative, and we're supposed to, create it.*/ if (dchild->d_inode == NULL) { - unsigned long ino; + unsigned long ino = rec->ur_fid2->id; struct iattr iattr; struct inode *inode; - - ino = (unsigned long)id_ino(rec->ur_id2); - - rc = mds_try_to_split_dir(obd, dparent, &mea, 0, - update_mode); - - CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n", - obd->obd_name, dparent->d_inode->i_ino, - dparent->d_inode->i_generation, rc); - - if (rc > 0) { - /* dir got splitted */ - GOTO(cleanup, rc = -ERESTART); - } else if (rc < 0) { - /* error happened during spitting */ - GOTO(cleanup, rc); - } if (!(rec->ur_flags & MDS_OPEN_CREAT)) { - /* - * dentry is negative and we weren't supposed to create - * object, get out of here. - */ + /* It's negative and we weren't supposed to create it */ GOTO(cleanup, rc = -ENOENT); } + if (req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) + GOTO(cleanup, rc = -EROFS); + intent_set_disposition(rep, DISP_OPEN_CREATE); handle = fsfilt_start(obd, dparent->d_inode, FSFILT_OP_CREATE, NULL); @@ -1168,17 +1020,11 @@ got_child: handle = NULL; GOTO(cleanup, rc); } - if (ino) - fid = id_fid(rec->ur_id2); - else - fid = mds_alloc_fid(obd); dchild->d_fsdata = (void *) &dp; - dp.p_ptr = req; - dp.p_inum = ino; - dp.p_fid = fid; - dp.p_group = mds->mds_num; + dp.ldp_ptr = req; + dp.ldp_inum = ino; - rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL); + rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL); if (dchild->d_fsdata == (void *)(unsigned long)ino) dchild->d_fsdata = NULL; @@ -1189,9 +1035,8 @@ got_child: inode = dchild->d_inode; if (ino) { LASSERT(ino == inode->i_ino); - - /* written as part of setattr */ - inode->i_generation = id_gen(rec->ur_id2); + /* Written as part of setattr */ + inode->i_generation = rec->ur_fid2->generation; CDEBUG(D_HA, "recreated ino %lu with gen %u\n", inode->i_ino, inode->i_generation); } @@ -1201,11 +1046,11 @@ got_child: LTIME_S(iattr.ia_ctime) = rec->ur_time; LTIME_S(iattr.ia_mtime) = rec->ur_time; - iattr.ia_uid = rec->ur_fsuid; + iattr.ia_uid = current->fsuid; /* set by push_ctxt already */ if (dparent->d_inode->i_mode & S_ISGID) iattr.ia_gid = dparent->d_inode->i_gid; else - iattr.ia_gid = rec->ur_fsgid; + iattr.ia_gid = current->fsgid; iattr.ia_valid = ATTR_UID | ATTR_GID | ATTR_ATIME | ATTR_MTIME | ATTR_CTIME; @@ -1219,37 +1064,22 @@ got_child: rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0); if (rc) CERROR("error on parent setattr: rc = %d\n", rc); - else { - MD_COUNTER_INCREMENT(obd, create); - } - - mds_inode2id(obd, &body->id1, inode, fid); - mds_update_inode_ids(obd, dchild->d_inode, handle, &body->id1, &sid); - - if ((rec->ur_flags & MDS_OPEN_HAS_KEY) || - mds->mds_crypto_type == MKS_TYPE) { - rc = mds_set_gskey(obd, handle, dchild->d_inode, - rec->ur_ea2data, rec->ur_ea2datalen, - ATTR_KEY | ATTR_MAC); - if (rc) { - CERROR("error in set gs key rc %d\n", rc); - } - } - - if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */ - rc = fsfilt_commit(obd, dchild->d_inode->i_sb, - dchild->d_inode, handle, - req->rq_export->exp_sync); - handle = NULL; - } + rc = fsfilt_commit(obd, dchild->d_inode, handle, 0); + handle = NULL; acc_mode = 0; /* Don't check for permissions */ + } else { + acc_mode = accmode(dchild->d_inode, rec->ur_flags); } - mds_pack_inode2body(obd, body, dchild->d_inode, 1); - + LASSERTF(!mds_inode_is_orphan(dchild->d_inode), - "dchild %.*s (%p) inode %p\n", dchild->d_name.len, - dchild->d_name.name, dchild, dchild->d_inode); + "dchild %.*s (%p) inode %p/%lu/%u\n", dchild->d_name.len, + dchild->d_name.name, dchild, dchild->d_inode, + dchild->d_inode->i_ino, dchild->d_inode->i_generation); + +found_child: + mds_pack_inode2fid(&body->fid1, dchild->d_inode); + mds_pack_inode2body(body, dchild->d_inode); if (S_ISREG(dchild->d_inode->i_mode)) { /* Check permissions etc */ @@ -1257,9 +1087,9 @@ got_child: if (rc != 0) GOTO(cleanup, rc); - /* Can't write to a read-only file */ - if (IS_RDONLY(dchild->d_inode) && (acc_mode & MAY_WRITE) != 0) - GOTO(cleanup, rc = -EPERM); + if ((req->rq_export->exp_connect_flags & OBD_CONNECT_RDONLY) && + (acc_mode & MAY_WRITE)) + GOTO(cleanup, rc = -EROFS); /* An append-only file must be opened in append mode for * writing */ @@ -1276,130 +1106,89 @@ got_child: GOTO(cleanup, rc = -EEXIST); // returns a lock to the client } + /* if we are following a symlink, don't open */ + if (S_ISLNK(dchild->d_inode->i_mode)) + GOTO(cleanup_no_trans, rc = 0); + if (S_ISDIR(dchild->d_inode->i_mode)) { if (rec->ur_flags & MDS_OPEN_CREAT || rec->ur_flags & FMODE_WRITE) { /* we are trying to create or write a exist dir */ GOTO(cleanup, rc = -EISDIR); } - if (ll_permission(dchild->d_inode, acc_mode, NULL)) + if (rec->ur_flags & MDS_FMODE_EXEC) { + /* we are trying to exec a directory */ GOTO(cleanup, rc = -EACCES); - - /* - * here was checking for possible GNS mount points to skip their - * open. I removed it as its detection based only on SUID bit is - * not reliable. Opening such a dirs should not cause any - * problems, at least tests show that. --umka - */ - } - - /* if we are following a symlink, don't open */ - if (S_ISLNK(dchild->d_inode->i_mode)) - GOTO(cleanup_no_trans, rc = 0); - - if ((rec->ur_flags & MDS_OPEN_DIRECTORY) && - !S_ISDIR(dchild->d_inode->i_mode)) + } + if (ll_permission(dchild->d_inode, acc_mode, NULL)) { + intent_set_disposition(rep, DISP_OPEN_OPEN); + GOTO(cleanup, rc = -EACCES); + } + } else if (rec->ur_flags & MDS_OPEN_DIRECTORY) { GOTO(cleanup, rc = -ENOTDIR); + } - /* check permission even it's special files */ - if (S_ISCHR(dchild->d_inode->i_mode) || - S_ISBLK(dchild->d_inode->i_mode) || - S_ISFIFO(dchild->d_inode->i_mode) || - S_ISSOCK(dchild->d_inode->i_mode)) { - rc = ll_permission(dchild->d_inode, acc_mode, NULL); - if (rc != 0) - GOTO(cleanup, rc); + if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE, + OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) { + GOTO(cleanup, rc = -EAGAIN); } - if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) { - obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE; - GOTO(cleanup, rc = -EAGAIN); - } + /* Obtain OPEN lock as well */ + policy.l_inodebits.bits |= MDS_INODELOCK_OPEN; -#if 0 /* We cannot use acc_mode here, because it is zeroed in case of creating a file, so we get wrong lockmode */ - if (accmode(rec->ur_flags) & MAY_WRITE) - child_mode = LCK_CW; - else if (accmode(rec->ur_flags) & MAY_EXEC) + if (accmode(dchild->d_inode, rec->ur_flags) & MAY_WRITE) + child_mode = LCK_CW; + else if (accmode(dchild->d_inode, rec->ur_flags) & MAY_EXEC) child_mode = LCK_PR; else child_mode = LCK_CR; - /* Always returning LOOKUP lock if open succesful to guard - * dentry on client. In case of replay we do not get a lock - * assuming that the caller has it already */ - if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) { - struct ldlm_res_id child_res_id = { .name = {0}}; - ldlm_policy_data_t policy; - int lock_flags = LDLM_FL_ATOMIC_CB; - - /* LOOKUP lock will protect dentry on client -bzzz */ - policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP | - MDS_INODELOCK_OPEN; - - child_res_id.name[0] = id_fid(&body->id1); - child_res_id.name[1] = id_group(&body->id1); - - rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, - child_res_id, LDLM_IBITS, &policy, - child_mode, &lock_flags, - mds_blocking_ast, ldlm_completion_ast, - NULL, NULL, NULL, 0, NULL, child_lockh); + if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) && + (rec->ur_flags & MDS_OPEN_LOCK)) { + /* In case of replay we do not get a lock assuming that the + caller has it already */ + child_res_id.name[0] = dchild->d_inode->i_ino; + child_res_id.name[1] = dchild->d_inode->i_generation; + + rc = ldlm_cli_enqueue_local(obd->obd_namespace, &child_res_id, + LDLM_IBITS, &policy, child_mode, + &lock_flags, ldlm_blocking_ast, + ldlm_completion_ast, NULL, NULL, + 0, NULL, child_lockh); if (rc != ELDLM_OK) GOTO(cleanup, rc); + /* Let mds_intent_policy know that we have a lock to return */ + intent_set_disposition(rep, DISP_OPEN_LOCK); cleanup_phase = 3; } -#else -/* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */ -#warning "disable opencache lock for CMD2" -#endif - - /* Step 5: mds_open it */ - rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, - rec, rep); - if (rc) - GOTO(cleanup, rc); - /* if this is a writer, we have to invalidate client's - * update locks in order to make sure they don't use - * isize/iblocks from mds anymore. - * FIXME: can cause a deadlock, use mds_get_parent_child_locked() - * XXX: optimization is to do this for first writer only */ - if (accmode(rec->ur_flags) & MAY_WRITE) { - struct ldlm_res_id child_res_id = { .name = {0}}; - ldlm_policy_data_t sz_policy; - struct lustre_handle sz_lockh; - int lock_flags = 0; - - child_res_id.name[0] = id_fid(&body->id1); - child_res_id.name[1] = id_group(&body->id1); - sz_policy.l_inodebits.bits = MDS_INODELOCK_UPDATE; - - rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, - child_res_id, LDLM_IBITS, &sz_policy, - LCK_PW, &lock_flags, mds_blocking_ast, - ldlm_completion_ast, NULL, NULL, NULL, - 0, NULL, &sz_lockh); - if (rc == ELDLM_OK) - ldlm_lock_decref(&sz_lockh, LCK_PW); - else - CERROR("can't invalidate client's update locks\n"); + if (!S_ISREG(dchild->d_inode->i_mode) && + !S_ISDIR(dchild->d_inode->i_mode) && + (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH)) { + /* If client supports this, do not return open handle for + * special device nodes */ + GOTO(cleanup_no_trans, rc = 0); } - EXIT; -cleanup: + /* Step 5: mds_open it */ + rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec, + rep, &parent_lockh); + GOTO(cleanup, rc); + + cleanup: rc = mds_finish_transno(mds, dchild ? dchild->d_inode : NULL, handle, - req, rc, rep ? rep->lock_policy_res1 : 0); + req, rc, rep ? rep->lock_policy_res1 : 0, 0); -cleanup_no_trans: + cleanup_no_trans: switch (cleanup_phase) { case 3: - if (rc) { + if (rc) + /* It is safe to leave IT_OPEN_LOCK set, if rc is not 0, + * mds_intent_policy won't try to return any locks */ ldlm_lock_decref(child_lockh, child_mode); - child_lockh->cookie = 0; - } case 2: if (rc && created) { int err = vfs_unlink(dparent->d_inode, dchild); @@ -1410,6 +1199,9 @@ cleanup_no_trans: } } else if (created) { mds_lock_new_child(obd, dchild->d_inode, NULL); + /* save uid/gid for quota acquire/release */ + qpids[USRQUOTA] = dparent->d_inode->i_uid; + qpids[GRPQUOTA] = dparent->d_inode->i_gid; } l_dput(dchild); case 1: @@ -1417,37 +1209,26 @@ cleanup_no_trans: break; l_dput(dparent); -#ifdef S_PDIROPS - if (parent_lockh[1].cookie != 0) - ldlm_lock_decref(parent_lockh + 1, update_mode); -#endif if (rc) - ldlm_lock_decref(parent_lockh, parent_mode); + ldlm_lock_decref(&parent_lockh, parent_mode); else - ptlrpc_save_lock (req, parent_lockh, parent_mode); + ptlrpc_save_lock(req, &parent_lockh, parent_mode); } - if (mea) - OBD_FREE(mea, mea_size); - if (rc == 0) - atomic_inc(&mds->mds_open_count); + /* trigger dqacq on the owner of child and parent */ + lquota_adjust(mds_quota_interface_ref, obd, qcids, qpids, rc, + FSFILT_OP_CREATE); - mds_body_do_reverse_map(med, body); - - /* - * If we have not taken the "open" lock, we may not return 0 here, - * because caller expects 0 to mean "lock is taken", and it needs - * nonzero return here for caller to return EDLM_LOCK_ABORTED to - * client. Later caller should rewrite the return value back to zero - * if it to be used any further. + /* If we have not taken the "open" lock, we may not return 0 here, + because caller expects 0 to mean "lock is taken", and it needs + nonzero return here for caller to return EDLM_LOCK_ABORTED to + client. Later caller should rewrite the return value back to zero + if it to be used any further */ - if ((cleanup_phase != 3) && !rc) - rc = ENOLCK; - RETURN(rc); } /* Close a "file descriptor" and possibly unlink an orphan from the - * PENDING directory. Caller must hold child->i_sem, this drops it. + * PENDING directory. Caller must hold child->i_mutex, this drops it. * * If we are being called from mds_disconnect() because the client has * disappeared, then req == NULL and we do not update last_rcvd because @@ -1455,97 +1236,83 @@ cleanup_no_trans: * (it will not even _have_ an entry in last_rcvd anymore). * * Returns EAGAIN if the client needs to get more data and re-close. */ -int mds_mfd_close(struct ptlrpc_request *req, int offset, - struct obd_device *obd, struct mds_file_data *mfd, - int unlink_orphan) +int mds_mfd_close(struct ptlrpc_request *req, int offset,struct obd_device *obd, + struct mds_file_data *mfd, int unlink_orphan, + struct lov_mds_md *lmm, int lmm_size, + struct llog_cookie *logcookies, int cookies_size, + __u64 *valid) { struct inode *inode = mfd->mfd_dentry->d_inode; - char idname[LL_ID_NAMELEN]; - int last_orphan, idlen, rc = 0, cleanup_phase = 0; + char fidname[LL_FID_NAMELEN]; + int last_orphan, fidlen, rc = 0, cleanup_phase = 0; struct dentry *pending_child = NULL; struct mds_obd *mds = &obd->u.mds; struct inode *pending_dir = mds->mds_pending_dir->d_inode; void *handle = NULL; struct mds_body *request_body = NULL, *reply_body = NULL; - struct dentry_params dp; + struct lvfs_dentry_params dp = LVFS_DENTRY_PARAMS_INIT; struct iattr iattr = { 0 }; - struct llog_create_locks *lcl = NULL; ENTRY; if (req && req->rq_reqmsg != NULL) request_body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*request_body)); if (req && req->rq_repmsg != NULL) - reply_body = lustre_msg_buf(req->rq_repmsg, 0, + reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body)); - if (request_body && (request_body->valid & OBD_MD_FLSIZE)) { - /* we set i_size/i_blocks here, nobody will see - * them until all write references are dropped. - * btw, we hold one reference */ - LASSERT(mfd->mfd_mode & FMODE_WRITE); - i_size_write(inode, request_body->size); - inode->i_blocks = request_body->blocks; - iattr.ia_size = inode->i_size; - iattr.ia_valid |= ATTR_SIZE; - mds_inode_unset_attrs_old(inode); - } + fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation); - idlen = ll_id2str(idname, inode->i_ino, inode->i_generation); - CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, - idname, inode->i_nlink, mds_orphan_open_count(inode)); + CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, fidname, + inode->i_nlink, mds_orphan_open_count(inode)); last_orphan = mds_orphan_open_dec_test(inode) && - mds_inode_is_orphan(inode); - UP_WRITE_I_ALLOC_SEM(inode); + mds_inode_is_orphan(inode); /* this is half of the actual "close" */ if (mfd->mfd_mode & FMODE_WRITE) { rc = mds_put_write_access(mds, inode, request_body, last_orphan && unlink_orphan); - } else if (mfd->mfd_mode & FMODE_EXEC) { + } else if (mfd->mfd_mode & MDS_FMODE_EXEC) { mds_allow_write_access(inode); } + /* here writecount change also needs protection from orphan write sem. + * so drop orphan write sem after mds_put_write_access, bz 12888. */ + MDS_UP_WRITE_ORPHAN_SEM(inode); if (last_orphan && unlink_orphan) { - struct lov_mds_md *lmm = NULL; int stripe_count = 0; LASSERT(rc == 0); /* mds_put_write_access must have succeeded */ - CDEBUG(D_HA, "destroying orphan object %s\n", idname); - + CDEBUG(D_INODE, "destroying orphan object %s\n", fidname); + if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) || (S_ISDIR(inode->i_mode) && inode->i_nlink != 2)) CERROR("found \"orphan\" %s %s with link count %d\n", S_ISREG(inode->i_mode) ? "file" : "dir", - idname, inode->i_nlink); - /* - * sadly, there is no easy way to save pending_child from - * mds_reint_unlink() into mfd, so we need to re-lookup, but - * normally it will still be in the dcache. - */ - down(&pending_dir->i_sem); - cleanup_phase = 1; /* up(i_sem) when finished */ - pending_child = lookup_one_len(idname, mds->mds_pending_dir, - idlen); + fidname, inode->i_nlink); + + /* Sadly, there is no easy way to save pending_child from + * mds_reint_unlink() into mfd, so we need to re-lookup, + * but normally it will still be in the dcache. */ + LOCK_INODE_MUTEX(pending_dir); + cleanup_phase = 1; /* UNLOCK_INODE_MUTEX(pending_dir) when finished */ + pending_child = lookup_one_len(fidname, mds->mds_pending_dir, + fidlen); if (IS_ERR(pending_child)) GOTO(cleanup, rc = PTR_ERR(pending_child)); - if (pending_child->d_inode == NULL) { - CERROR("orphan %s has been removed\n", idname); - GOTO(cleanup, rc = 0); - } + LASSERT(pending_child->d_inode != NULL); cleanup_phase = 2; /* dput(pending_child) when finished */ if (S_ISDIR(pending_child->d_inode->i_mode)) { rc = vfs_rmdir(pending_dir, pending_child); if (rc) CERROR("error unlinking orphan dir %s: rc %d\n", - idname, rc); + fidname,rc); goto out; } - if (req != NULL && req->rq_repmsg != NULL) { - lmm = lustre_msg_buf(req->rq_repmsg, 1, 0); + if (lmm != NULL) { stripe_count = le32_to_cpu(lmm->lmm_stripe_count); } @@ -1557,34 +1324,20 @@ int mds_mfd_close(struct ptlrpc_request *req, int offset, GOTO(cleanup, rc); } + if (lmm != NULL && (*valid & OBD_MD_FLEASIZE) && + mds_log_op_unlink(obd, lmm, lmm_size, + logcookies, cookies_size) > 0) { + *valid |= OBD_MD_FLCOOKIE; + } + + dp.ldp_inum = 0; + dp.ldp_ptr = req; pending_child->d_fsdata = (void *) &dp; - dp.p_inum = 0; - dp.p_ptr = req; rc = vfs_unlink(pending_dir, pending_child); if (rc) - CERROR("error unlinking orphan %s: rc %d\n", - idname, rc); - - if (req != NULL && req->rq_repmsg != NULL && - (reply_body->valid & OBD_MD_FLEASIZE) && - mds_log_op_unlink(obd, pending_child->d_inode, - lmm, req->rq_repmsg->buflens[1], - lustre_msg_buf(req->rq_repmsg, 2, 0), - req->rq_repmsg->buflens[2], &lcl) > 0) { - reply_body->valid |= OBD_MD_FLCOOKIE; - } - - rc = mds_destroy_object(obd, inode, 1); - if (rc) { - CERROR("cannot destroy OSS object on close, err %d\n", - rc); - rc = 0; - } + CERROR("error unlinking orphan %s: rc %d\n",fidname,rc); goto out; /* Don't bother updating attrs on unlinked inode */ - } else if ((mfd->mfd_mode & FMODE_WRITE) && rc == 0) { - /* last writer closed file - let's update i_size/i_blocks */ - mds_validate_size(obd, inode, request_body, &iattr); } #if 0 @@ -1596,13 +1349,13 @@ int mds_mfd_close(struct ptlrpc_request *req, int offset, */ if (request_body->valid & OBD_MD_FLMTIME && - LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) { - LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime); + LTIME_S(iattr.ia_mtime) > LTIME_S(inode->i_mtime)) { + LTIME_S(iattr.ia_mtime) = request_body->mtime; iattr.ia_valid |= ATTR_MTIME; } if (request_body->valid & OBD_MD_FLCTIME && - LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) { - LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime); + LTIME_S(iattr.ia_ctime) > LTIME_S(inode->i_ctime)) { + LTIME_S(iattr.ia_ctime) = request_body->ctime; iattr.ia_valid |= ATTR_CTIME; } @@ -1611,10 +1364,7 @@ int mds_mfd_close(struct ptlrpc_request *req, int offset, iattr.ia_valid |= ATTR_SIZE; iattr.ia_size = request_body->size; } - /* if (request_body->valid & OBD_MD_FLBLOCKS) { - iattr.ia_valid |= ATTR_BLOCKS; - iattr.ia_blocks = request_body->blocks - } */ + /* iattr.ia_blocks = request_body->blocks */ } #endif @@ -1623,49 +1373,37 @@ int mds_mfd_close(struct ptlrpc_request *req, int offset, * it is more out-of-date than the specified limit. If we * are already going to write out the atime then do it anyway. * */ - if ((request_body->atime > - LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) || + LTIME_S(iattr.ia_atime) = request_body->atime; + if ((LTIME_S(iattr.ia_atime) > + LTIME_S(inode->i_atime) + mds->mds_atime_diff) || (iattr.ia_valid != 0 && - request_body->atime > LTIME_S(inode->i_atime))) { - LTIME_S(iattr.ia_atime) = request_body->atime; + LTIME_S(iattr.ia_atime) > LTIME_S(inode->i_atime))) iattr.ia_valid |= ATTR_ATIME; - } } if (iattr.ia_valid != 0) { - if (handle == NULL) - handle = fsfilt_start(obd,inode,FSFILT_OP_SETATTR,NULL); - if (IS_ERR(handle)) - GOTO(cleanup, rc = PTR_ERR(handle)); + handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL); + if (IS_ERR(handle)) { + rc = PTR_ERR(handle); + handle = NULL; + GOTO(cleanup, rc); + } rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0); if (rc) - CERROR("error in setattr(%s): rc %d\n", idname, rc); + CERROR("error in setattr(%s): rc %d\n", fidname, rc); } out: /* If other clients have this file open for write, rc will be > 0 */ if (rc > 0) rc = 0; - if (!obd->obd_recovering && mds_inode_has_old_attrs(inode) - && !mds_inode_is_orphan(inode) - && atomic_read(&inode->i_writecount) == 0) { - CERROR("leave inode %lu/%u with old attributes\n", - inode->i_ino, inode->i_generation); - } l_dput(mfd->mfd_dentry); - mds_mfd_destroy(mfd); + mds_mfd_put(mfd); cleanup: - atomic_dec(&mds->mds_open_count); if (req != NULL && reply_body != NULL) { - rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0); + rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0, 0); } else if (handle) { - int err, force_sync = 0; - - if (req && req->rq_export) - force_sync = req->rq_export->exp_sync; - - err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, - force_sync); + int err = fsfilt_commit(obd, pending_dir, handle, 0); if (err) { CERROR("error committing close: %d\n", err); if (!rc) @@ -1675,140 +1413,13 @@ out: switch (cleanup_phase) { case 2: - if (lcl != NULL) - ptlrpc_save_llog_lock(req, lcl); dput(pending_child); case 1: - up(&pending_dir->i_sem); + UNLOCK_INODE_MUTEX(pending_dir); } RETURN(rc); } -static int mds_extent_lock_callback(struct ldlm_lock *lock, - struct ldlm_lock_desc *new, void *data, - int flag) -{ - struct lustre_handle lockh = { 0 }; - int rc; - ENTRY; - - switch (flag) { - case LDLM_CB_BLOCKING: - ldlm_lock2handle(lock, &lockh); - rc = ldlm_cli_cancel(&lockh); - if (rc != ELDLM_OK) - CERROR("ldlm_cli_cancel failed: %d\n", rc); - break; - case LDLM_CB_CANCELING: { - break; - } - default: - LBUG(); - } - - RETURN(0); -} -__u64 lov_merge_size(struct lov_stripe_md *lsm, int kms); -__u64 lov_merge_blocks(struct lov_stripe_md *lsm); - -int mds_validate_size(struct obd_device *obd, struct inode *inode, - struct mds_body *body, struct iattr *iattr) -{ - ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } }; - struct lustre_handle lockh = { 0 }; - struct lov_stripe_md *lsm = NULL; - int rc, len, flags; - void *lmm = NULL; - ENTRY; - - /* we update i_size/i_blocks only for regular files */ - if (!S_ISREG(inode->i_mode)) - RETURN(0); - - /* we shouldn't fetch size from OSTes during recovery - deadlock */ - if (obd->obd_recovering) { - CERROR("size-on-mds has no support on OST yet\n"); - RETURN(0); - } - - /* if nobody modified attrs. we're lucky */ - if (!mds_inode_has_old_attrs(inode)) - RETURN(0); - - /* 1: client didn't send actual i_size/i_blocks - * 2: we seem to be last writer - * 3: the file doesn't look like to be disappeared - * conclusion: we're gonna fetch them from OSSes */ - - down(&inode->i_sem); - len = fsfilt_get_md(obd, inode, NULL, 0, EA_LOV); - up(&inode->i_sem); - - if (len < 0) { - CERROR("error getting inode %lu MD: %d\n", inode->i_ino, len); - GOTO(cleanup, rc = len); - } else if (len == 0) { - CDEBUG(D_INODE, "no LOV in inode %lu\n", inode->i_ino); - GOTO(cleanup, rc = 0); - } - - OBD_ALLOC(lmm, len); - if (lmm == NULL) { - CERROR("can't allocate memory\n"); - GOTO(cleanup, rc = -ENOMEM); - } - - down(&inode->i_sem); - rc = fsfilt_get_md(obd, inode, lmm, len, EA_LOV); - up(&inode->i_sem); - - if (rc < 0) { - CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc); - GOTO(cleanup, rc); - } - - rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, lmm, len); - if (rc < 0) { - CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc); - GOTO(cleanup, rc); - } - - CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino); - - flags = LDLM_FL_HAS_INTENT; - rc = obd_enqueue(obd->u.mds.mds_dt_exp, lsm, LDLM_EXTENT, &policy, - LCK_PR, &flags, mds_extent_lock_callback, - ldlm_completion_ast, NULL, NULL, - sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh); - if (rc != 0) { - CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc); - GOTO(cleanup, rc); - } - - CDEBUG(D_INODE, "LOV reports "LPD64"/%lu for "DLID4" [%s%s%s]\n", - inode->i_size, inode->i_blocks, OLID4(&body->id1), - atomic_read(&inode->i_writecount) > 1 ? "U" : "", - mds_inode_has_old_attrs(inode) ? "D" : "", - mds_inode_is_orphan(inode) ? "O" : ""); - - i_size_write(inode, lov_merge_size(lsm, 0)); - inode->i_blocks = lov_merge_blocks(lsm); - iattr->ia_size = inode->i_size; - iattr->ia_valid |= ATTR_SIZE; - DOWN_WRITE_I_ALLOC_SEM(inode); - mds_inode_unset_attrs_old(inode); - UP_WRITE_I_ALLOC_SEM(inode); - - obd_cancel(obd->u.mds.mds_dt_exp, lsm, LCK_PR, &lockh); - -cleanup: - if (lsm != NULL) - obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm); - if (lmm != NULL) - OBD_FREE(lmm, len); - RETURN(rc); -} - int mds_close(struct ptlrpc_request *req, int offset) { struct mds_export_data *med = &req->rq_export->exp_mds_data; @@ -1817,26 +1428,28 @@ int mds_close(struct ptlrpc_request *req, int offset) struct mds_file_data *mfd; struct lvfs_run_ctxt saved; struct inode *inode; - int rc, repsize[3] = {sizeof(struct mds_body), - obd->u.mds.mds_max_mdsize, - obd->u.mds.mds_max_cookiesize}; + int rc, repsize[4] = { sizeof(struct ptlrpc_body), + sizeof(struct mds_body), + obd->u.mds.mds_max_mdsize, + obd->u.mds.mds_max_cookiesize }; + struct mds_body *reply_body; + struct lov_mds_md *lmm; + int lmm_size; + struct llog_cookie *logcookies; + int cookies_size; ENTRY; - MD_COUNTER_INCREMENT(obd, close); - rc = lustre_pack_reply(req, 3, repsize, NULL); - if (rc) { - CERROR("lustre_pack_reply: rc = %d\n", rc); + rc = lustre_pack_reply(req, 4, repsize, NULL); + if (rc) req->rq_status = rc; - } else { + /* continue on to drop local open even if we can't send reply */ + else MDS_CHECK_RESENT(req, mds_reconstruct_generic(req)); - } - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { - DEBUG_REQ(D_HA, req, "close replay"); - memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0), - lustre_msg_buf(req->rq_reqmsg, offset + 1, 0), - req->rq_repmsg->buflens[2]); - } + CDEBUG(D_INODE, "close req->rep_len %d mdsize %d cookiesize %d\n", + req->rq_replen, + obd->u.mds.mds_max_mdsize, obd->u.mds.mds_max_cookiesize); + mds_counter_incr(req->rq_export, LPROC_MDS_CLOSE); body = lustre_swab_reqbuf(req, offset, sizeof(*body), lustre_swab_mds_body); @@ -1846,65 +1459,60 @@ int mds_close(struct ptlrpc_request *req, int offset) RETURN(-EFAULT); } - if (body->flags & MDS_BFLAG_UNCOMMITTED_WRITES) - /* do some stuff */ ; - + spin_lock(&med->med_open_lock); mfd = mds_handle2mfd(&body->handle); if (mfd == NULL) { + spin_unlock(&med->med_open_lock); DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64 - ": cookie "LPX64, id_ino(&body->id1), body->handle.cookie); + ": cookie "LPX64, body->fid1.id, body->handle.cookie); req->rq_status = -ESTALE; RETURN(-ESTALE); } + /* Remove mfd handle so it can't be found again. We consume mfd_list + * reference here, but still have mds_handle2mfd ref until mfd_close. */ + mds_mfd_unlink(mfd, 1); + spin_unlock(&med->med_open_lock); inode = mfd->mfd_dentry->d_inode; - - /* child i_alloc_sem protects orphan_dec_test && is_orphan race */ - DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */ - - if (body->flags & MDS_BFLAG_DIRTY_EPOCH) { - /* the client modified data through the handle - * we need to care about attrs. -bzzz */ - mds_inode_set_attrs_old(inode); - } - + /* child orphan sem protects orphan_dec_test && is_orphan race */ + MDS_DOWN_WRITE_ORPHAN_SEM(inode); /* mds_mfd_close drops this */ if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) { - struct mds_body *rep_body; - - rep_body = lustre_msg_buf(req->rq_repmsg, 0, - sizeof (*rep_body)); - LASSERT(rep_body != NULL); + body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, + sizeof(*body)); + LASSERT(body != NULL); - mds_pack_inode2body(obd, rep_body, inode, - (body->valid & OBD_MD_FID) ? 1 : 0); - - mds_pack_md(obd, req->rq_repmsg, 1, rep_body, - inode, MDS_PACK_MD_LOCK, 0); + mds_pack_inode2fid(&body->fid1, inode); + mds_pack_inode2body(body, inode); + mds_pack_md(obd, req->rq_repmsg, REPLY_REC_OFF + 1, body, inode, + MDS_PACK_MD_LOCK); } - spin_lock(&med->med_open_lock); - list_del(&mfd->mfd_list); - spin_unlock(&med->med_open_lock); push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); - req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1); + reply_body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*reply_body)); + lmm = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1, 0); + lmm_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 1), + logcookies = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 2, 0); + cookies_size = lustre_msg_buflen(req->rq_repmsg, REPLY_REC_OFF + 2); + req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1, + lmm, lmm_size, logcookies, cookies_size, + &reply_body->valid); pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL); + mds_shrink_reply(obd, req, body, REPLY_REC_OFF + 1); if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) { CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n"); req->rq_status = -ENOMEM; - mds_mfd_put(mfd); RETURN(-ENOMEM); } - mds_mfd_put(mfd); - RETURN(0); + RETURN(rc); } - int mds_done_writing(struct ptlrpc_request *req, int offset) { struct mds_body *body; - int rc, size = sizeof(struct mds_body); + int rc, size[2] = { sizeof(struct ptlrpc_body), + sizeof(struct mds_body) }; ENTRY; MDS_CHECK_RESENT(req, mds_reconstruct_generic(req)); @@ -1917,11 +1525,9 @@ int mds_done_writing(struct ptlrpc_request *req, int offset) RETURN(-EFAULT); } - rc = lustre_pack_reply(req, 1, &size, NULL); - if (rc) { - CERROR("lustre_pack_reply: rc = %d\n", rc); + rc = lustre_pack_reply(req, 2, size, NULL); + if (rc) req->rq_status = rc; - } RETURN(0); }