X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Fdcache.c;h=029e62120e2d0fd0d430779e62df13a3614419b4;hp=d74dc8b1ee9d6522e465d6ca72e31f03a4944e8a;hb=ea766bd959cc2347b223d597b3f8a31e010bed72;hpb=17b1b04c128e1823345020c969126ec10b4040c2 diff --git a/lustre/llite/dcache.c b/lustre/llite/dcache.c index d74dc8b..029e621 100644 --- a/lustre/llite/dcache.c +++ b/lustre/llite/dcache.c @@ -1,206 +1,345 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * Copyright (c) 2001-2003 Cluster File Systems, Inc. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * This file is part of Lustre, http://www.lustre.org. + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * 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. + * This program 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 version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * - * 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. + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf * - * 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. + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, + * CA 95054 USA or visit www.sun.com if you need additional information or + * have any questions. + * + * GPL HEADER END + */ +/* + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2011, 2013, Intel Corporation. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. */ #include #include -#include #include #define DEBUG_SUBSYSTEM S_LLITE -#include -#include -#include -#include -#include +#include +#include +#include +#include #include "llite_internal.h" +static void free_dentry_data(struct rcu_head *head) +{ + struct ll_dentry_data *lld; + + lld = container_of(head, struct ll_dentry_data, lld_rcu_head); + OBD_FREE_PTR(lld); +} + /* should NOT be called with the dcache lock, see fs/dcache.c */ static void ll_release(struct dentry *de) { struct ll_dentry_data *lld; ENTRY; LASSERT(de != NULL); - - CDEBUG(D_DENTRY, "releasing dentry %p\n", de); - lld = ll_d2d(de); - if (lld) { /* Root dentry does not have ll_dentry_data */ - LASSERT(lld->lld_cwd_count == 0); - LASSERT(lld->lld_mnt_count == 0); - OBD_FREE(de->d_fsdata, sizeof(struct ll_dentry_data)); + if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */ + RETURN_EXIT; + + if (lld->lld_it) { + ll_intent_release(lld->lld_it); + OBD_FREE(lld->lld_it, sizeof(*lld->lld_it)); } + LASSERT(lld->lld_cwd_count == 0); + LASSERT(lld->lld_mnt_count == 0); + de->d_fsdata = NULL; + call_rcu(&lld->lld_rcu_head, free_dentry_data); - EXIT; + EXIT; } /* Compare if two dentries are the same. Don't match if the existing dentry - * is marked DCACHE_LUSTRE_INVALID. Returns 1 if different, 0 if the same. + * is marked invalid. Returns 1 if different, 0 if the same. * * This avoids a race where ll_lookup_it() instantiates a dentry, but we get * an AST before calling d_revalidate_it(). The dentry still exists (marked * INVALID) so d_lookup() matches it, but we have no lock on it (so * lock_match() fails) and we spin around real_lookup(). */ -static int ll_dcompare(struct dentry *parent, struct qstr *d_name, - struct qstr *name){ - struct dentry *dchild; - ENTRY; +#ifdef HAVE_D_COMPARE_7ARGS +int ll_dcompare(const struct dentry *parent, const struct inode *pinode, + const struct dentry *dentry, const struct inode *inode, + unsigned int len, const char *str, const struct qstr *name) +#else +int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name) +#endif +{ +#ifdef HAVE_D_COMPARE_7ARGS + ENTRY; - if (d_name->len != name->len) - RETURN(1); + if (len != name->len) + RETURN(1); - if (memcmp(d_name->name, name->name, name->len)) - RETURN(1); + if (memcmp(str, name->name, len)) + RETURN(1); +#else + struct dentry *dentry; + ENTRY; + + if (d_name->len != name->len) + RETURN(1); + + if (memcmp(d_name->name, name->name, name->len)) + RETURN(1); + + /* XXX: d_name must be in-dentry structure */ + dentry = container_of(d_name, struct dentry, d_name); /* ugh */ +#endif + + CDEBUG(D_DENTRY, "found name %.*s(%p) flags %#x refc %d\n", + name->len, name->name, dentry, dentry->d_flags, + d_refcount(dentry)); - dchild = container_of(d_name, struct dentry, d_name); /* ugh */ - if (dchild->d_flags & DCACHE_LUSTRE_INVALID) { - CDEBUG(D_DENTRY,"INVALID dentry %p not matched, was bug 3784\n", - dchild); + /* mountpoint is always valid */ + if (d_mountpoint((struct dentry *)dentry)) + RETURN(0); + + if (d_lustre_invalid(dentry)) RETURN(1); - } RETURN(0); } -/* should NOT be called with the dcache lock, see fs/dcache.c */ -static int ll_ddelete(struct dentry *de) +static inline int return_if_equal(struct ldlm_lock *lock, void *data) { - ENTRY; - LASSERT(de); - CDEBUG(D_DENTRY, "%s dentry %*s (%p, parent %p, inode %p) %s%s\n", - (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"), - de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode, - d_unhashed(de) ? "" : "hashed,", - list_empty(&de->d_subdirs) ? "" : "subdirs"); - RETURN(0); + if ((lock->l_flags & + (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) == + (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) + return LDLM_ITER_CONTINUE; + return LDLM_ITER_STOP; } -void ll_set_dd(struct dentry *de) +/* find any ldlm lock of the inode in mdc and lov + * return 0 not find + * 1 find one + * < 0 error */ +static int find_cbdata(struct inode *inode) { + struct ll_sb_info *sbi = ll_i2sbi(inode); + struct lov_stripe_md *lsm; + int rc = 0; ENTRY; - LASSERT(de != NULL); - CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n", - de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode, - atomic_read(&de->d_count)); - lock_kernel(); - if (de->d_fsdata == NULL) { - OBD_ALLOC(de->d_fsdata, sizeof(struct ll_dentry_data)); + LASSERT(inode); + rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode), + return_if_equal, NULL); + if (rc != 0) + RETURN(rc); + + lsm = ccc_inode_lsm_get(inode); + if (lsm == NULL) + RETURN(rc); + + rc = obd_find_cbdata(sbi->ll_dt_exp, lsm, return_if_equal, NULL); + ccc_inode_lsm_put(inode, lsm); + + RETURN(rc); +} + +/** + * Called when last reference to a dentry is dropped and dcache wants to know + * whether or not it should cache it: + * - return 1 to delete the dentry immediately + * - return 0 to cache the dentry + * Should NOT be called with the dcache lock, see fs/dcache.c + */ +static int ll_ddelete(HAVE_D_DELETE_CONST struct dentry *de) +{ + ENTRY; + LASSERT(de); + + CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n", + d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping", + de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode, + d_unhashed((struct dentry *)de) ? "" : "hashed,", + list_empty(&de->d_subdirs) ? "" : "subdirs"); + +#ifdef HAVE_DCACHE_LOCK + LASSERT(d_refcount(de) == 0); +#else + /* kernel >= 2.6.38 last refcount is decreased after this function. */ + LASSERT(d_refcount(de) == 1); +#endif + + /* Disable this piece of code temproarily because this is called + * inside dcache_lock so it's not appropriate to do lots of work + * here. ATTENTION: Before this piece of code enabling, LU-2487 must be + * resolved. */ +#if 0 + /* if not ldlm lock for this inode, set i_nlink to 0 so that + * this inode can be recycled later b=20433 */ + if (de->d_inode && !find_cbdata(de->d_inode)) + clear_nlink(de->d_inode); +#endif + + if (d_lustre_invalid((struct dentry *)de)) + RETURN(1); + RETURN(0); +} + +static int ll_set_dd(struct dentry *de) +{ + ENTRY; + LASSERT(de != NULL); + + CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n", + de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode, + d_refcount(de)); + + if (de->d_fsdata == NULL) { + struct ll_dentry_data *lld; + + OBD_ALLOC_PTR(lld); + if (likely(lld != NULL)) { + spin_lock(&de->d_lock); + if (likely(de->d_fsdata == NULL)) + de->d_fsdata = lld; + else + OBD_FREE_PTR(lld); + spin_unlock(&de->d_lock); + } else { + RETURN(-ENOMEM); + } + } + + RETURN(0); +} + +int ll_dops_init(struct dentry *de, int block, int init_sa) +{ + struct ll_dentry_data *lld = ll_d2d(de); + int rc = 0; + + if (lld == NULL && block != 0) { + rc = ll_set_dd(de); + if (rc) + return rc; + + lld = ll_d2d(de); } - unlock_kernel(); - EXIT; + if (lld != NULL && init_sa != 0) + lld->lld_sa_generation = 0; + +#ifdef HAVE_DCACHE_LOCK + de->d_op = &ll_d_ops; +#else + /* kernel >= 2.6.38 d_op is set in d_alloc() */ + LASSERT(de->d_op == &ll_d_ops); +#endif + return rc; } void ll_intent_drop_lock(struct lookup_intent *it) { - struct lustre_handle *handle; - struct lustre_intent_data *itdata = LUSTRE_IT(it); + if (it->it_op && it->d.lustre.it_lock_mode) { + struct lustre_handle handle; + + handle.cookie = it->d.lustre.it_lock_handle; - if (it->it_op && itdata && itdata->it_lock_mode) { - handle = (struct lustre_handle *)&itdata->it_lock_handle; CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64 - " from it %p\n", handle->cookie, it); - ldlm_lock_decref(handle, itdata->it_lock_mode); + " from it %p\n", handle.cookie, it); + ldlm_lock_decref(&handle, it->d.lustre.it_lock_mode); /* bug 494: intent_release may be called multiple times, from * this thread and we don't want to double-decref this lock */ - itdata->it_lock_mode = 0; - } + it->d.lustre.it_lock_mode = 0; + if (it->d.lustre.it_remote_lock_mode != 0) { + handle.cookie = it->d.lustre.it_remote_lock_handle; + + CDEBUG(D_DLMTRACE, "releasing remote lock with cookie" + LPX64" from it %p\n", handle.cookie, it); + ldlm_lock_decref(&handle, + it->d.lustre.it_remote_lock_mode); + it->d.lustre.it_remote_lock_mode = 0; + } + } } void ll_intent_release(struct lookup_intent *it) { ENTRY; + CDEBUG(D_INFO, "intent %p released\n", it); ll_intent_drop_lock(it); - it->it_magic = 0; - it->it_op_release = 0; - ll_intent_free(it); + /* We are still holding extra reference on a request, need to free it */ + if (it_disposition(it, DISP_ENQ_OPEN_REF)) + ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */ + if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */ + ptlrpc_req_finished(it->d.lustre.it_data); + if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate + * to lookup */ + ptlrpc_req_finished(it->d.lustre.it_data); + + it->d.lustre.it_disposition = 0; + it->d.lustre.it_data = NULL; EXIT; } -void ll_intent_free(struct lookup_intent *it) -{ - if (it->d.fs_data) { - OBD_SLAB_FREE(it->d.fs_data, ll_intent_slab, - sizeof(struct lustre_intent_data)); - it->d.fs_data = NULL; - } -} - -void ll_unhash_aliases(struct inode *inode) +void ll_invalidate_aliases(struct inode *inode) { - struct list_head *tmp, *head; - struct ll_sb_info *sbi; - ENTRY; + struct dentry *dentry; + struct ll_d_hlist_node *p; + ENTRY; + + LASSERT(inode != NULL); + + CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n", + inode->i_ino, inode->i_generation, inode); + + ll_lock_dcache(inode); + ll_d_hlist_for_each_entry(dentry, p, &inode->i_dentry, d_alias) { + CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p " + "inode %p flags %d\n", dentry->d_name.len, + dentry->d_name.name, dentry, dentry->d_parent, + dentry->d_inode, dentry->d_flags); + + if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') { + CERROR("called on root (?) dentry=%p, inode=%p " + "ino=%lu\n", dentry, inode, inode->i_ino); + lustre_dump_dentry(dentry, 1); + libcfs_debug_dumpstack(NULL); + } - if (inode == NULL) { - CERROR("unexpected NULL inode, tell phil\n"); - EXIT; - return; - } + d_lustre_invalidate(dentry, 0); + } + ll_unlock_dcache(inode); - CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n", - inode->i_ino, inode->i_generation, inode); - - sbi = ll_i2sbi(inode); - head = &inode->i_dentry; -restart: - spin_lock(&dcache_lock); - tmp = head; - while ((tmp = tmp->next) != head) { - struct dentry *dentry = list_entry(tmp, struct dentry, d_alias); - if (atomic_read(&dentry->d_count) == 0) { - CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p " - "inode %p\n", dentry->d_name.len, - dentry->d_name.name, dentry, dentry->d_parent, - dentry->d_inode); - dget_locked(dentry); - __d_drop(dentry); - spin_unlock(&dcache_lock); - dput(dentry); - goto restart; - } else if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) { - CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p " - "inode %p refc %d\n", dentry->d_name.len, - dentry->d_name.name, dentry, dentry->d_parent, - dentry->d_inode, atomic_read(&dentry->d_count)); - hlist_del_init(&dentry->d_hash); - dentry->d_flags |= DCACHE_LUSTRE_INVALID; - hlist_add_head(&dentry->d_hash, - &sbi->ll_orphan_dentry_list); - } - } - spin_unlock(&dcache_lock); EXIT; } -extern struct dentry *ll_find_alias(struct inode *, struct dentry *); - -int revalidate_it_finish(struct ptlrpc_request *request, int offset, - struct lookup_intent *it, struct dentry *de) +int ll_revalidate_it_finish(struct ptlrpc_request *request, + struct lookup_intent *it, + struct dentry *de) { - struct ll_sb_info *sbi; int rc = 0; ENTRY; @@ -210,9 +349,7 @@ int revalidate_it_finish(struct ptlrpc_request *request, int offset, if (it_disposition(it, DISP_LOOKUP_NEG)) RETURN(-ENOENT); - sbi = ll_i2sbi(de->d_inode); - rc = ll_prep_inode(sbi->ll_dt_exp, sbi->ll_md_exp, - &de->d_inode, request, offset, NULL); + rc = ll_prep_inode(&de->d_inode, request, NULL, it); RETURN(rc); } @@ -222,26 +359,21 @@ void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry) LASSERT(it != NULL); LASSERT(dentry != NULL); - if (LUSTRE_IT(it)->it_lock_mode && dentry->d_inode != NULL) { + if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) { struct inode *inode = dentry->d_inode; + struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode); + CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n", inode, inode->i_ino, inode->i_generation); - mdc_set_lock_data(NULL, &LUSTRE_IT(it)->it_lock_handle, inode); + ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL); } /* drop lookup or getattr locks immediately */ - if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR || - it->it_op == IT_CHDIR) { -#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0)) - /* - * on 2.6 there are situations when several lookups and + if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) { + /* on 2.6 there are situation when several lookups and * revalidations may be requested during single operation. - * Therefore, we don't release intent here -bzzz - */ + * therefore, we don't release intent here -bzzz */ ll_intent_drop_lock(it); -#else - ll_intent_release(it); -#endif } } @@ -249,185 +381,103 @@ void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft) { struct lookup_intent *it = *itp; -#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0)) - if (it) { - LASSERTF(it->it_magic == INTENT_MAGIC, "bad intent magic: %x\n", - it->it_magic); - } -#endif - if (!it || it->it_op == IT_GETXATTR) it = *itp = deft; - if (it->d.fs_data) - return; - - if (ll_intent_alloc(it)) { - CERROR("Failed to allocate memory for lustre specific intent " - "data\n"); - /* XXX: we cannot return status just yet */ - LBUG(); - } } -int ll_intent_alloc(struct lookup_intent *it) -{ - if (it->d.fs_data) { - CERROR("Intent alloc on already allocated intent\n"); - return 0; - } - OBD_SLAB_ALLOC(it->d.fs_data, ll_intent_slab, SLAB_KERNEL, - sizeof(struct lustre_intent_data)); - if (!it->d.fs_data) { - CERROR("Failed to allocate memory for lustre specific intent " - "data\n"); - return -ENOMEM; - } - - it->it_op_release = ll_intent_release; - - return 0; -} - -int ll_revalidate_it(struct dentry *de, int flags, struct nameidata *nd, +int ll_revalidate_it(struct dentry *de, int lookup_flags, struct lookup_intent *it) { - struct lookup_intent lookup_it = { .it_op = IT_LOOKUP }; + struct md_op_data *op_data; struct ptlrpc_request *req = NULL; + struct lookup_intent lookup_it = { .it_op = IT_LOOKUP }; struct obd_export *exp; - struct it_cb_data icbd; - struct lustre_id pid; - struct lustre_id cid; - int orig_it, rc = 0; - ENTRY; + struct inode *parent = de->d_parent->d_inode; + int rc; - spin_lock(&de->d_lock); - - if ((de->d_flags & DCACHE_GNS_PENDING) && - !(de->d_flags & DCACHE_GNS_MOUNTING)) - { - spin_unlock(&de->d_lock); - - if (nd) { - int err = ll_gns_mount_object(de, nd->mnt); - if (err) - CERROR("can't mount %s, err = %d\n", - de->d_name.name, err); - } - RETURN(1); + ENTRY; + CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name, + LL_IT2STR(it)); + + if (de->d_inode == NULL) { + __u64 ibits; + + /* We can only use negative dentries if this is stat or lookup, + for opens and stuff we do need to query server. */ + /* If there is IT_CREAT in intent op set, then we must throw + away this negative dentry and actually do the request to + kernel to create whatever needs to be created (if possible)*/ + if (it && (it->it_op & IT_CREAT)) + RETURN(0); + + if (d_lustre_invalid(de)) + RETURN(0); + + ibits = MDS_INODELOCK_UPDATE; + rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE); + GOTO(out_sa, rc); } - spin_unlock(&de->d_lock); - - CDEBUG(D_VFSTRACE, "VFS Op:name=%s (%p), intent=%s\n", de->d_name.name, - de, LL_IT2STR(it)); - - /* Cached negative dentries are unsafe for now - look them up again */ - if (de->d_inode == NULL) - RETURN(0); - - /* Root of the tree is always valid, attributes would be fixed in - ll_inode_revalidate_it */ - if (de->d_sb->s_root == de) - RETURN(1); - - CDEBUG(D_INODE, "revalidate 0x%p: %*s -> %lu/%lu\n", - de, de->d_name.len, de->d_name.name, - (unsigned long) de->d_inode->i_ino, - (unsigned long) de->d_inode->i_generation); - - exp = ll_i2mdexp(de->d_inode); - ll_inode2id(&pid, de->d_parent->d_inode); - ll_inode2id(&cid, de->d_inode); - LASSERT(id_fid(&cid) != 0); - - icbd.icbd_parent = de->d_parent->d_inode; - icbd.icbd_childp = &de; - /* - * never execute intents for mount points. Attributes will be fixed up - * in ll_inode_revalidate_it(). - */ + /* Never execute intents for mount points. + * Attributes will be fixed up in ll_inode_revalidate_it */ if (d_mountpoint(de)) - RETURN(1); + GOTO(out_sa, rc = 1); + + /* need to get attributes in case root got changed from other client */ + if (de == de->d_sb->s_root) { + rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP); + if (rc == 0) + rc = 1; + GOTO(out_sa, rc); + } - if (nd != NULL) - nd->mnt->mnt_last_used = jiffies; + exp = ll_i2mdexp(de->d_inode); OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5); - orig_it = it ? it->it_op : IT_OPEN; ll_frob_intent(&it, &lookup_it); - LASSERT(it != NULL); + LASSERT(it); - if (it->it_op == IT_GETATTR) { /* We need to check for LOOKUP lock as - well */ - rc = ll_intent_alloc(&lookup_it); - if (rc) - LBUG(); /* Can't think of better idea just yet */ - - rc = md_intent_lock(exp, &pid, de->d_name.name, - de->d_name.len, NULL, 0, &cid, &lookup_it, - flags, &req, ll_mdc_blocking_ast); - /* If there was no lookup lock, no point in even checking for - UPDATE lock */ - if (!rc) { - it = &lookup_it; - if (!req) { - ll_intent_free(it); - goto do_lookup; - } - GOTO(out, rc); - } - if (it_disposition(&lookup_it, DISP_LOOKUP_NEG)) { - it = &lookup_it; - ll_intent_free(it); - GOTO(out, rc = 0); - } + if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de)) + RETURN(1); - if (req) - ptlrpc_req_finished(req); - req = NULL; - ll_lookup_finish_locks(&lookup_it, de); - /* XXX: on 2.6 ll_lookup_finish_locks does not call ll_intent_release */ - ll_intent_release(&lookup_it); - } + if (it->it_op == IT_OPEN) { + struct inode *inode = de->d_inode; + struct ll_inode_info *lli = ll_i2info(inode); + struct obd_client_handle **och_p; + __u64 *och_usecount; + __u64 ibits; + + /* + * We used to check for MDS_INODELOCK_OPEN here, but in fact + * just having LOOKUP lock is enough to justify inode is the + * same. And if inode is the same and we have suitable + * openhandle, then there is no point in doing another OPEN RPC + * just to throw away newly received openhandle. There are no + * security implications too, if file owner or access mode is + * change, LOOKUP lock is revoked. + */ -#if 1 - if ((it->it_op == IT_OPEN) && de->d_inode) { - struct inode *inode = de->d_inode; - struct ll_inode_info *lli = ll_i2info(inode); - struct obd_client_handle **och_p; - __u64 *och_usecount; - struct obd_device *obddev; - struct lustre_handle lockh; - int flags = LDLM_FL_BLOCK_GRANTED; - ldlm_policy_data_t policy = {.l_inodebits = {MDS_INODELOCK_OPEN}}; - struct ldlm_res_id file_res_id = {.name = {id_fid(&lli->lli_id), - id_group(&lli->lli_id)}}; - int lockmode; if (it->it_flags & FMODE_WRITE) { och_p = &lli->lli_mds_write_och; och_usecount = &lli->lli_open_fd_write_count; - lockmode = LCK_CW; } else if (it->it_flags & FMODE_EXEC) { och_p = &lli->lli_mds_exec_och; och_usecount = &lli->lli_open_fd_exec_count; - lockmode = LCK_PR; } else { och_p = &lli->lli_mds_read_och; och_usecount = &lli->lli_open_fd_read_count; - lockmode = LCK_CR; } - - /* Check for the proper lock */ - obddev = md_get_real_obd(exp, &lli->lli_id); - if (!ldlm_lock_match(obddev->obd_namespace, flags, &file_res_id, - LDLM_IBITS, &policy, lockmode, &lockh)) + /* Check for the proper lock. */ + ibits = MDS_INODELOCK_LOOKUP; + if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE)) goto do_lock; - down(&lli->lli_och_sem); + mutex_lock(&lli->lli_och_mutex); if (*och_p) { /* Everything is open already, do nothing */ /*(*och_usecount)++; Do not let them steal our open - handle from under us */ + handle from under us */ + SET_BUT_UNUSED(och_usecount); /* XXX The code above was my original idea, but in case we have the handle, but we cannot use it due to later checks (e.g. O_CREAT|O_EXCL flags set), nobody @@ -435,24 +485,38 @@ int ll_revalidate_it(struct dentry *de, int flags, struct nameidata *nd, hope the lock won't be invalidated in between. But if it would be, we'll reopen the open request to MDS later during file open path */ - up(&lli->lli_och_sem); - memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh, - sizeof(lockh)); - LUSTRE_IT(it)->it_lock_mode = lockmode; + mutex_unlock(&lli->lli_och_mutex); RETURN(1); } else { - /* Hm, interesting. Lock is present, but no open - handle? */ - up(&lli->lli_och_sem); - ldlm_lock_decref(&lockh, lockmode); + mutex_unlock(&lli->lli_och_mutex); } } -#endif + + if (it->it_op == IT_GETATTR) { + rc = ll_statahead_enter(parent, &de, 0); + if (rc == 1) + goto mark; + else if (rc != -EAGAIN && rc != 0) + GOTO(out, rc = 0); + } do_lock: - rc = md_intent_lock(exp, &pid, de->d_name.name, de->d_name.len, - NULL, 0, &cid, it, flags, &req, ll_mdc_blocking_ast); - /* If req is NULL, then md_intent_lock() only tried to do a lock match; + op_data = ll_prep_md_op_data(NULL, parent, de->d_inode, + de->d_name.name, de->d_name.len, + 0, LUSTRE_OPC_ANY, NULL); + if (IS_ERR(op_data)) + RETURN(PTR_ERR(op_data)); + + if (!IS_POSIXACL(parent) || !exp_connect_umask(exp)) + it->it_create_mode &= ~cfs_curproc_umask(); + it->it_create_mode |= M_CHECK_STALE; + rc = md_intent_lock(exp, op_data, NULL, 0, it, + lookup_flags, + &req, ll_md_blocking_ast, 0); + it->it_create_mode &= ~M_CHECK_STALE; + ll_finish_md_op_data(op_data); + + /* If req is NULL, then md_intent_lock only tried to do a lock match; * if all was well, it will return 1 if it found locks, 0 otherwise. */ if (req == NULL && rc >= 0) { if (!rc) @@ -462,228 +526,246 @@ do_lock: if (rc < 0) { if (rc != -ESTALE) { - CDEBUG(D_INFO, "ll_intent_lock(): rc %d : it->it_status " - "%d\n", rc, LUSTRE_IT(it)->it_status); + CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status " + "%d\n", rc, it->d.lustre.it_status); } GOTO(out, rc = 0); } + revalidate_finish: - rc = revalidate_it_finish(req, 1, it, de); + rc = ll_revalidate_it_finish(req, it, de); if (rc != 0) { - ll_intent_release(it); + if (rc != -ESTALE && rc != -ENOENT) + ll_intent_release(it); GOTO(out, rc = 0); } - rc = 1; - /* unfortunately ll_intent_lock may cause a callback and revoke our - dentry */ - spin_lock(&dcache_lock); - hlist_del_init(&de->d_hash); - __d_rehash(de); - spin_unlock(&dcache_lock); + if ((it->it_op & IT_OPEN) && de->d_inode && + !S_ISREG(de->d_inode->i_mode) && + !S_ISDIR(de->d_inode->i_mode)) { + ll_release_openhandle(de, it); + } + rc = 1; - GOTO(out, rc); out: - if (req != NULL && rc == 1) + /* We do not free request as it may be reused during following lookup + * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will + * be freed in ll_lookup_it or in ll_intent_release. But if + * request was not completed, we need to free it. (bug 5154, 9903) */ + if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE)) ptlrpc_req_finished(req); - if (rc == 0) { - if (it == &lookup_it) { - ll_intent_release(it); - if (req) /* Special case: We did lookup and it failed, - need to free request */ - ptlrpc_req_finished(req); - } - ll_unhash_aliases(de->d_inode); - return 0; + /* mdt may grant layout lock for the newly created file, so + * release the lock to avoid leaking */ + ll_intent_drop_lock(it); + ll_invalidate_aliases(de->d_inode); + } else { + __u64 bits = 0; + __u64 matched_bits = 0; + + CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p " + "inode %p refc %d\n", de->d_name.len, + de->d_name.name, de, de->d_parent, de->d_inode, + d_refcount(de)); + + ll_set_lock_data(exp, de->d_inode, it, &bits); + + /* Note: We have to match both LOOKUP and PERM lock + * here to make sure the dentry is valid and no one + * changing the permission. + * But if the client connects < 2.4 server, which will + * only grant LOOKUP lock, so we can only Match LOOKUP + * lock for old server */ + if (exp_connect_flags(ll_i2mdexp(de->d_inode)) && + OBD_CONNECT_LVB_TYPE) + matched_bits = + MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM; + else + matched_bits = MDS_INODELOCK_LOOKUP; + + if (((bits & matched_bits) == matched_bits) && + d_lustre_invalid(de)) + d_lustre_revalidate(de); + ll_lookup_finish_locks(it, de); + } + +mark: + if (it != NULL && it->it_op == IT_GETATTR && rc > 0) + ll_statahead_mark(parent, de); + RETURN(rc); + + /* + * This part is here to combat evil-evil race in real_lookup on 2.6 + * kernels. The race details are: We enter do_lookup() looking for some + * name, there is nothing in dcache for this name yet and d_lookup() + * returns NULL. We proceed to real_lookup(), and while we do this, + * another process does open on the same file we looking up (most simple + * reproducer), open succeeds and the dentry is added. Now back to + * us. In real_lookup() we do d_lookup() again and suddenly find the + * dentry, so we call d_revalidate on it, but there is no lock, so + * without this code we would return 0, but unpatched real_lookup just + * returns -ENOENT in such a case instead of retrying the lookup. Once + * this is dealt with in real_lookup(), all of this ugly mess can go and + * we can just check locks in ->d_revalidate without doing any RPCs + * ever. + */ +do_lookup: + if (it != &lookup_it) { + /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */ + if (it->it_op == IT_GETATTR) + lookup_it.it_op = IT_GETATTR; + ll_lookup_finish_locks(it, de); + it = &lookup_it; } - CDEBUG(D_DENTRY, "revalidated dentry %*s (%p) parent %p " - "inode %p refc %d\n", de->d_name.len, - de->d_name.name, de, de->d_parent, de->d_inode, - atomic_read(&de->d_count)); + /* Do real lookup here. */ + op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name, + de->d_name.len, 0, (it->it_op & IT_CREAT ? + LUSTRE_OPC_CREATE : + LUSTRE_OPC_ANY), NULL); + if (IS_ERR(op_data)) + RETURN(PTR_ERR(op_data)); - ll_lookup_finish_locks(it, de); - de->d_flags &= ~DCACHE_LUSTRE_INVALID; - if (it == &lookup_it) - ll_intent_release(it); - - if (!((de->d_inode->i_mode & S_ISUID) && S_ISDIR(de->d_inode->i_mode)) || - !(flags & LOOKUP_CONTINUE || (orig_it & (IT_CHDIR | IT_OPEN)))) - return rc; - - if (nd && !(de->d_flags & DCACHE_GNS_MOUNTING)) { - int err = ll_gns_mount_object(de, nd->mnt); - if (err) - CERROR("can't mount %s, err = %d\n", - de->d_name.name, err); - } - return rc; -do_lookup: - it = &lookup_it; - if (ll_intent_alloc(it)) - LBUG(); -// We did that already, right? ll_inode2id(&pid, de->d_parent->d_inode); - rc = md_intent_lock(exp, &pid, de->d_name.name, - de->d_name.len, NULL, 0, NULL, - it, 0, &req, ll_mdc_blocking_ast); + rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req, + ll_md_blocking_ast, 0); if (rc >= 0) { - struct mds_body *mds_body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*mds_body)); + struct mdt_body *mdt_body; + struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0}; + mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY); - /* See if we got same inode, if not - return error */ - if (id_equal_stc(&cid, &mds_body->id1)) + if (de->d_inode) + fid = *ll_inode2fid(de->d_inode); + + /* see if we got same inode, if not - return error */ + if (lu_fid_eq(&fid, &mdt_body->fid1)) { + ll_finish_md_op_data(op_data); + op_data = NULL; goto revalidate_finish; + } + ll_intent_release(it); } - + ll_finish_md_op_data(op_data); GOTO(out, rc = 0); + +out_sa: + /* + * For rc == 1 case, should not return directly to prevent losing + * statahead windows; for rc == 0 case, the "lookup" will be done later. + */ + if (it != NULL && it->it_op == IT_GETATTR && rc == 1) + ll_statahead_enter(parent, &de, 1); + goto mark; } -/*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag) +#ifdef HAVE_IOP_ATOMIC_OPEN +/* + * Always trust cached dentries. Update statahead window if necessary. + */ +int ll_revalidate_nd(struct dentry *dentry, unsigned int flags) { - struct inode *inode= de->d_inode; - struct ll_sb_info *sbi = ll_i2sbi(inode); - struct ll_dentry_data *ldd = ll_d2d(de); - struct obd_client_handle *handle; - int rc = 0; - ENTRY; - LASSERT(ldd); - - lock_kernel(); - /* Strictly speaking this introduces an additional race: the - * increments should wait until the rpc has returned. - * However, given that at present the function is void, this - * issue is moot. */ - if (flag == 1 && (++ldd->lld_mnt_count) > 1) { - unlock_kernel(); - EXIT; - return; - } + struct inode *parent = dentry->d_parent->d_inode; + int unplug = 0; - if (flag == 0 && (++ldd->lld_cwd_count) > 1) { - unlock_kernel(); - EXIT; - return; - } - unlock_kernel(); - - handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och; - rc = obd_pin(sbi->ll_md_exp, inode->i_ino, inode->i_generation, - inode->i_mode & S_IFMT, handle, flag); - - if (rc) { - lock_kernel(); - memset(handle, 0, sizeof(*handle)); - if (flag == 0) - ldd->lld_cwd_count--; - else - ldd->lld_mnt_count--; - unlock_kernel(); - } + ENTRY; + CDEBUG(D_VFSTRACE, "VFS Op:name=%s,flags=%u\n", + dentry->d_name.name, flags); - EXIT; - return; + if (!(flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE)) && + ll_need_statahead(parent, dentry) > 0) { + if (flags & LOOKUP_RCU) + RETURN(-ECHILD); + + if (dentry->d_inode == NULL) + unplug = 1; + do_statahead_enter(parent, &dentry, unplug); + ll_statahead_mark(parent, dentry); + } + + RETURN(1); } -/*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag) +#else /* !HAVE_IOP_ATOMIC_OPEN */ +int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd) { - struct ll_sb_info *sbi = ll_i2sbi(de->d_inode); - struct ll_dentry_data *ldd = ll_d2d(de); - struct obd_client_handle handle; - int count, rc = 0; + int rc; ENTRY; - LASSERT(ldd); - - lock_kernel(); - /* Strictly speaking this introduces an additional race: the - * increments should wait until the rpc has returned. - * However, given that at present the function is void, this - * issue is moot. */ - handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och; - if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) { - /* the "pin" failed */ - unlock_kernel(); - EXIT; - return; - } - if (flag) - count = --ldd->lld_mnt_count; - else - count = --ldd->lld_cwd_count; - unlock_kernel(); +#ifndef HAVE_DCACHE_LOCK + /* kernel >= 2.6.38 supports rcu-walk, but lustre doesn't. */ + if (nd->flags & LOOKUP_RCU) + return -ECHILD; +#endif - if (count != 0) { - EXIT; - return; - } + if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) { + struct lookup_intent *it; - rc = obd_unpin(sbi->ll_md_exp, &handle, flag); - EXIT; -} + it = ll_convert_intent(&nd->intent.open, nd->flags); + if (IS_ERR(it)) + RETURN(0); -#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0)) -static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd) -{ - int rc; - ENTRY; + if (it->it_op == (IT_OPEN|IT_CREAT) && + nd->intent.open.flags & O_EXCL) { + CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n"); + rc = 0; + goto out_it; + } + + rc = ll_revalidate_it(dentry, nd->flags, it); + + if (rc && (nd->flags & LOOKUP_OPEN) && + it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/ +// XXX Code duplication with ll_lookup_nd + if (S_ISFIFO(dentry->d_inode->i_mode)) { + // We cannot call open here as it would + // deadlock. + ptlrpc_req_finished( + (struct ptlrpc_request *) + it->d.lustre.it_data); + } else { + struct file *filp; + + nd->intent.open.file->private_data = it; + filp = lookup_instantiate_filp(nd, dentry,NULL); + if (IS_ERR(filp)) + rc = PTR_ERR(filp); + } + } + if (!rc && (nd->flags & LOOKUP_CREATE) && + it_disposition(it, DISP_OPEN_CREATE)) { + /* We created something but we may only return + * negative dentry here, so save request in dentry, + * if lookup will be called later on, it will + * pick the request, otherwise it would be freed + * with dentry */ + ll_d2d(dentry)->lld_it = it; + it = NULL; /* avoid freeing */ + } - if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST)) - rc = ll_revalidate_it(dentry, nd->flags, nd, &nd->intent.open); - else - rc = ll_revalidate_it(dentry, 0, nd, NULL); +out_it: + if (it) { + ll_intent_release(it); + OBD_FREE(it, sizeof(*it)); + } + } else { + rc = ll_revalidate_it(dentry, 0, NULL); + } RETURN(rc); } -#endif +#endif /* HAVE_IOP_ATOMIC_OPEN */ -#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)) -static void ll_dentry_iput(struct dentry *dentry, struct inode *inode) +void ll_d_iput(struct dentry *de, struct inode *inode) { - struct ll_sb_info *sbi = ll_i2sbi(inode); - struct lustre_id parent, child; - - LASSERT(dentry->d_parent && dentry->d_parent->d_inode); - ll_inode2id(&parent, dentry->d_parent->d_inode); - ll_inode2id(&child, inode); - md_change_cbdata_name(sbi->ll_md_exp, &parent, - (char *)dentry->d_name.name, - dentry->d_name.len, &child, - null_if_equal, inode); - iput(inode); + LASSERT(inode); + if (!find_cbdata(inode)) + clear_nlink(inode); + iput(inode); } -#else -static void ll_dentry_iput(struct dentry *dentry, struct inode *inode) -{ - struct ll_sb_info *sbi = ll_i2sbi(inode); - struct lustre_id parent, child; - - if (dentry->d_parent != dentry) { - /* Do not do this for root of the tree */ - LASSERT(dentry->d_parent && dentry->d_parent->d_inode); - ll_inode2id(&parent, dentry->d_parent->d_inode); - ll_inode2id(&child, inode); - md_change_cbdata_name(sbi->ll_md_exp, &parent, - (char *)dentry->d_name.name, - dentry->d_name.len, &child, - null_if_equal, inode); - } - iput(inode); - -} -#endif struct dentry_operations ll_d_ops = { -#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0)) .d_revalidate = ll_revalidate_nd, -#else - .d_revalidate_it = ll_revalidate_it, -#endif .d_release = ll_release, - .d_iput = ll_dentry_iput, - .d_delete = ll_ddelete, + .d_delete = ll_ddelete, + .d_iput = ll_d_iput, .d_compare = ll_dcompare, -#if 0 - .d_pin = ll_pin, - .d_unpin = ll_unpin, -#endif };