X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Fdcache.c;h=5790afb3222a39423ab831c5b3c0e039bfcce401;hp=bafb2936c78bb7ac6a01daa0b5daffd811546f69;hb=32e547aa9cb913f5736ee3d58cb79f4e63ce2c0b;hpb=b7121b2fea17af5bc327df721af3348b3c97e413 diff --git a/lustre/llite/dcache.c b/lustre/llite/dcache.c index bafb293..5790afb 100644 --- a/lustre/llite/dcache.c +++ b/lustre/llite/dcache.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -17,17 +15,15 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -36,22 +32,23 @@ #include #include -#include #include +#include #define DEBUG_SUBSYSTEM S_LLITE #include -#include -#include #include -#include -//#include -//#include #include "llite_internal.h" -spinlock_t ll_lookup_lock = SPIN_LOCK_UNLOCKED; +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) @@ -60,115 +57,173 @@ static void ll_release(struct dentry *de) ENTRY; LASSERT(de != NULL); lld = ll_d2d(de); - if (lld == NULL) { /* NFS copies the de->d_op methods (bug 4655) */ - EXIT; - return; - } -#ifndef HAVE_VFS_INTENT_PATCHES + 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)); } -#endif - LASSERT(lld->lld_cwd_count == 0); - LASSERT(lld->lld_mnt_count == 0); - OBD_FREE(de->d_fsdata, sizeof(*lld)); - EXIT; + de->d_fsdata = NULL; + call_rcu(&lld->lld_rcu_head, free_dentry_data); + + EXIT; } -#ifdef DCACHE_LUSTRE_INVALID /* 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(). */ -int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name) + * lock_match() fails) and we spin around real_lookup(). + * + * This race doesn't apply to lookups in d_alloc_parallel(), and for + * those we want to ensure that only one dentry with a given name is + * in ll_lookup_nd() at a time. So allow invalid dentries to match + * while d_in_lookup(). We will be called again when the lookup + * completes, and can give a different answer then. + */ +#ifdef HAVE_D_COMPARE_7ARGS +static 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) +#elif defined(HAVE_D_COMPARE_5ARGS) +static int ll_dcompare(const struct dentry *parent, const struct dentry *dentry, + unsigned int len, const char *str, + const struct qstr *name) +#elif defined(HAVE_D_COMPARE_4ARGS) +static int ll_dcompare(const struct dentry *dentry, unsigned int len, + const char *str, const struct qstr *name) +#else +static int ll_dcompare(struct dentry *parent, struct qstr *d_name, + struct qstr *name) +#endif { - struct dentry *dchild; - ENTRY; +#if !defined(HAVE_D_COMPARE_7ARGS) && !defined(HAVE_D_COMPARE_5ARGS) && !defined(HAVE_D_COMPARE_4ARGS) + /* XXX: (ugh !) d_name must be in-dentry structure */ + struct dentry *dentry = container_of(d_name, struct dentry, d_name); + unsigned int len = d_name->len; + const char *str = d_name->name; +#endif + 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); - /* XXX: d_name must be in-dentry structure */ - 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); - RETURN(1); - } + CDEBUG(D_DENTRY, "found name %.*s(%p) flags %#x refc %d\n", + name->len, name->name, dentry, dentry->d_flags, + ll_d_count(dentry)); + + /* mountpoint is always valid */ + if (d_mountpoint((struct dentry *)dentry)) + RETURN(0); + + /* ensure exclusion against parallel lookup of the same name */ + if (d_in_lookup((struct dentry *)dentry)) + return 0; + + if (d_lustre_invalid(dentry)) + RETURN(1); - RETURN(0); + RETURN(0); } -#endif -/* should NOT be called with the dcache lock, see fs/dcache.c */ -static int ll_ddelete(struct dentry *de) +/** + * 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); -#ifndef DCACHE_LUSTRE_INVALID -#define DCACHE_LUSTRE_INVALID 0 -#endif + 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"); - 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"); -#if DCACHE_LUSTRE_INVALID == 0 -#undef DCACHE_LUSTRE_INVALID +#ifdef HAVE_DCACHE_LOCK + LASSERT(ll_d_count(de) == 0); +#else + /* kernel >= 2.6.38 last refcount is decreased after this function. */ + LASSERT(ll_d_count(de) == 1); #endif - RETURN(0); + if (d_lustre_invalid((struct dentry *)de)) + RETURN(1); + RETURN(0); } -void ll_set_dd(struct dentry *de) +int ll_d_init(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, - atomic_read(&de->d_count)); - - if (de->d_fsdata == NULL) { - struct ll_dentry_data *lld; - - OBD_ALLOC_PTR(lld); - if (likely(lld != NULL)) { - lock_dentry(de); - if (likely(de->d_fsdata == NULL)) - de->d_fsdata = lld; - else - OBD_FREE_PTR(lld); - unlock_dentry(de); - } - } - - EXIT; + 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, + ll_d_count(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)) { +#ifdef HAVE_DCACHE_LOCK + /* kernel >= 2.6.38 d_op is set in d_alloc() */ + de->d_op = &ll_d_ops; + smp_mb(); +#endif + de->d_fsdata = lld; + __d_lustre_invalidate(de); + } else { + OBD_FREE_PTR(lld); + } + spin_unlock(&de->d_lock); + } else { + RETURN(-ENOMEM); + } + } + LASSERT(de->d_op == &ll_d_ops); + + RETURN(0); } void ll_intent_drop_lock(struct lookup_intent *it) { - struct lustre_handle *handle; - - if (it->it_op && it->d.lustre.it_lock_mode) { - handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle; - CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64 - " 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 */ - it->d.lustre.it_lock_mode = 0; - } + if (it->it_op && it->it_lock_mode) { + struct lustre_handle handle; + + handle.cookie = it->it_lock_handle; + + CDEBUG(D_DLMTRACE, "releasing lock with cookie %#llx from it %p\n", + handle.cookie, it); + ldlm_lock_decref(&handle, it->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 */ + it->it_lock_mode = 0; + if (it->it_remote_lock_mode != 0) { + handle.cookie = it->it_remote_lock_handle; + + CDEBUG(D_DLMTRACE, "releasing remote lock with cookie" + "%#llx from it %p\n", handle.cookie, it); + ldlm_lock_decref(&handle, + it->it_remote_lock_mode); + it->it_remote_lock_mode = 0; + } + } } void ll_intent_release(struct lookup_intent *it) @@ -177,124 +232,39 @@ void ll_intent_release(struct lookup_intent *it) CDEBUG(D_INFO, "intent %p released\n", it); ll_intent_drop_lock(it); -#ifdef HAVE_VFS_INTENT_PATCHES - it->it_magic = 0; - it->it_op_release = 0; -#endif /* 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; -} + ptlrpc_req_finished(it->it_request); /* ll_file_open */ -/* Drop dentry if it is not used already, unhash otherwise. - Should be called with dcache lock held! - Returns: 1 if dentry was dropped, 0 if unhashed. */ -int ll_drop_dentry(struct dentry *dentry) -{ - lock_dentry(dentry); - 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); - unlock_dentry(dentry); - spin_unlock(&dcache_lock); - spin_unlock(&ll_lookup_lock); - dput(dentry); - spin_lock(&ll_lookup_lock); - spin_lock(&dcache_lock); - return 1; - } - /* disconected dentry can not be find without lookup, because we - * not need his to unhash or mark invalid. */ - if (dentry->d_flags & DCACHE_DISCONNECTED) { - unlock_dentry(dentry); - RETURN (0); - } + if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */ + ptlrpc_req_finished(it->it_request); -#ifdef DCACHE_LUSTRE_INVALID - if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) { -#else - if (!d_unhashed(dentry)) { -#endif - 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)); - /* actually we don't unhash the dentry, rather just - * mark it inaccessible for to __d_lookup(). otherwise - * sys_getcwd() could return -ENOENT -bzzz */ -#ifdef DCACHE_LUSTRE_INVALID - dentry->d_flags |= DCACHE_LUSTRE_INVALID; -#endif - if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode)) - __d_drop(dentry); - - } - unlock_dentry(dentry); - return 0; + it->it_disposition = 0; + it->it_request = NULL; + EXIT; } -void ll_unhash_aliases(struct inode *inode) +void ll_invalidate_aliases(struct inode *inode) { - struct list_head *tmp, *head; - ENTRY; + struct dentry *dentry; + DECLARE_LL_D_HLIST_NODE_PTR(p); + ENTRY; - if (inode == NULL) { - CERROR("unexpected NULL inode, tell phil\n"); - return; - } + LASSERT(inode != NULL); - CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n", - inode->i_ino, inode->i_generation, inode); - - head = &inode->i_dentry; - spin_lock(&ll_lookup_lock); - spin_lock(&dcache_lock); -restart: - tmp = head; - while ((tmp = tmp->next) != head) { - struct dentry *dentry = list_entry(tmp, struct 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); - } else if (d_mountpoint(dentry)) { - /* For mountpoints we skip removal of the dentry - which happens solely because we have a lock on it - obtained when this dentry was not a mountpoint yet */ - CDEBUG(D_DENTRY, "Skippind mountpoint dentry removal " - "%.*s (%p) parent %p\n", - dentry->d_name.len, - dentry->d_name.name, - dentry, dentry->d_parent); - - continue; - } - - if (ll_drop_dentry(dentry)) - goto restart; - } - spin_unlock(&dcache_lock); - spin_unlock(&ll_lookup_lock); + CDEBUG(D_INODE, "marking dentries for inode "DFID"(%p) invalid\n", + PFID(ll_inode2fid(inode)), inode); + + ll_lock_dcache(inode); + ll_d_hlist_for_each_entry(dentry, p, &inode->i_dentry) { + 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); + + d_lustre_invalidate(dentry, 0); + } + ll_unlock_dcache(inode); EXIT; } @@ -312,7 +282,7 @@ int ll_revalidate_it_finish(struct ptlrpc_request *request, if (it_disposition(it, DISP_LOOKUP_NEG)) RETURN(-ENOENT); - rc = ll_prep_inode(&de->d_inode, request, NULL); + rc = ll_prep_inode(&de->d_inode, request, NULL, it); RETURN(rc); } @@ -322,14 +292,13 @@ void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry) LASSERT(it != NULL); LASSERT(dentry != NULL); - if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) { + if (it->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); - md_set_lock_data(sbi->ll_md_exp, &it->d.lustre.it_lock_handle, - inode); + CDEBUG(D_DLMTRACE, "setting l_data to inode "DFID"(%p)\n", + PFID(ll_inode2fid(inode)), inode); + ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL); } /* drop lookup or getattr locks immediately */ @@ -341,468 +310,84 @@ void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry) } } -void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft) +static int ll_revalidate_dentry(struct dentry *dentry, + unsigned int lookup_flags) { - struct lookup_intent *it = *itp; -#ifdef HAVE_VFS_INTENT_PATCHES - if (it) { - LASSERTF(it->it_magic == INTENT_MAGIC, - "%p has bad intent magic: %x\n", - it, it->it_magic); - } -#endif + struct inode *dir = dentry->d_parent->d_inode; - if (!it || it->it_op == IT_GETXATTR) - it = *itp = deft; + /* If this is intermediate component path lookup and we were able to get + * to this dentry, then its lock has not been revoked and the + * path component is valid. */ + if (lookup_flags & (LOOKUP_CONTINUE | LOOKUP_PARENT)) + return 1; -#ifdef HAVE_VFS_INTENT_PATCHES - it->it_op_release = ll_intent_release; -#endif -} - -int ll_revalidate_it(struct dentry *de, int lookup_flags, - struct lookup_intent *it) -{ - 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 inode *parent; - int rc, first = 0; - - ENTRY; - CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name, - LL_IT2STR(it)); - - if (de->d_inode == NULL) { - /* 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); - -#ifdef DCACHE_LUSTRE_INVALID - if (de->d_flags & DCACHE_LUSTRE_INVALID) - RETURN(0); -#endif - - rc = ll_have_md_lock(de->d_parent->d_inode, - MDS_INODELOCK_UPDATE); - GOTO(out_sa, rc); - } - - exp = ll_i2mdexp(de->d_inode); - - /* Never execute intents for mount points. - * Attributes will be fixed up in ll_inode_revalidate_it */ - if (d_mountpoint(de)) - GOTO(out_sa, rc = 1); - - /* Root of the lustre tree. Always valid. - * Attributes will be fixed up in ll_inode_revalidate_it */ - if (de == de->d_sb->s_root) - GOTO(out_sa, rc = 1); - - OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5); - ll_frob_intent(&it, &lookup_it); - LASSERT(it); - parent = de->d_parent->d_inode; - - 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 ((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; - - /* - * 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 (it->it_flags & FMODE_WRITE) { - och_p = &lli->lli_mds_write_och; - och_usecount = &lli->lli_open_fd_write_count; - } else if (it->it_flags & FMODE_EXEC) { - och_p = &lli->lli_mds_exec_och; - och_usecount = &lli->lli_open_fd_exec_count; - } else { - och_p = &lli->lli_mds_read_och; - och_usecount = &lli->lli_open_fd_read_count; - } - /* Check for the proper lock. */ - if (!ll_have_md_lock(inode, MDS_INODELOCK_LOOKUP)) - goto do_lock; - down(&lli->lli_och_sem); - if (*och_p) { /* Everything is open already, do nothing */ - /*(*och_usecount)++; Do not let them steal our open - handle from under us */ - /* 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 - would decrement counter increased here. So we just - 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); - ll_finish_md_op_data(op_data); - RETURN(1); - } else { - up(&lli->lli_och_sem); - } - } - - if (it->it_op == IT_GETATTR) - first = ll_statahead_enter(de->d_parent->d_inode, &de, 0); - -do_lock: - it->it_create_mode &= ~current->fs->umask; - it->it_flags |= O_CHECK_STALE; - rc = md_intent_lock(exp, op_data, NULL, 0, it, - lookup_flags, - &req, ll_md_blocking_ast, 0); - it->it_flags &= ~O_CHECK_STALE; - ll_finish_md_op_data(op_data); - if (it->it_op == IT_GETATTR && !first) - /* If there are too many locks on client-side, then some - * locks taken by statahead maybe dropped automatically - * before the real "revalidate" using them. */ - ll_statahead_exit(de, req == NULL ? rc : 0); - else if (first == -EEXIST) - ll_statahead_mark(de); - - /* 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) - goto do_lookup; - GOTO(out, rc); - } - - if (rc < 0) { - if (rc != -ESTALE) { - 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 = ll_revalidate_it_finish(req, it, de); - if (rc != 0) { - if (rc != -ESTALE && rc != -ENOENT) - ll_intent_release(it); - GOTO(out, rc = 0); - } - - 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; - - /* unfortunately ll_intent_lock may cause a callback and revoke our - * dentry */ - spin_lock(&ll_lookup_lock); - spin_lock(&dcache_lock); - lock_dentry(de); - __d_drop(de); - unlock_dentry(de); - d_rehash_cond(de, 0); - spin_unlock(&dcache_lock); - spin_unlock(&ll_lookup_lock); - -out: - /* 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) { -#ifdef DCACHE_LUSTRE_INVALID - ll_unhash_aliases(de->d_inode); - /* done in ll_unhash_aliases() - dentry->d_flags |= DCACHE_LUSTRE_INVALID; */ + /* Symlink - always valid as long as the dentry was found */ +#ifdef HAVE_IOP_GET_LINK + if (dentry->d_inode && dentry->d_inode->i_op->get_link) #else - /* We do not want d_invalidate to kill all child dentries too */ - d_drop(de); + if (dentry->d_inode && dentry->d_inode->i_op->follow_link) #endif - } else { - 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)); - ll_lookup_finish_locks(it, de); -#ifdef DCACHE_LUSTRE_INVALID - lock_dentry(de); - de->d_flags &= ~DCACHE_LUSTRE_INVALID; - unlock_dentry(de); + return 1; + + /* + * VFS warns us that this is the second go around and previous + * operation failed (most likely open|creat), so this time + * we better talk to the server via the lookup path by name, + * not by fid. + */ + if (lookup_flags & LOOKUP_REVAL) + return 0; + +#ifndef HAVE_DCACHE_LOCK + if (lookup_flags & LOOKUP_RCU) + return -ECHILD; #endif - } - 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; - } - /* 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)); - - rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req, - ll_md_blocking_ast, 0); - if (rc >= 0) { - 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); - - 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 && it->it_op == IT_GETATTR && rc == 1) { - first = ll_statahead_enter(de->d_parent->d_inode, &de, 0); - if (!first) - ll_statahead_exit(de, 1); - else if (first == -EEXIST) - ll_statahead_mark(de); - } + if (dentry_may_statahead(dir, dentry)) + ll_statahead(dir, &dentry, dentry->d_inode == NULL); - return rc; + return 1; } -/*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag) -{ - 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; - struct obd_capa *oc; - 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; - } - - if (flag == 0 && (++ldd->lld_cwd_count) > 1) { - unlock_kernel(); - EXIT; - return; - } - unlock_kernel(); - - handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och; - oc = ll_mdscapa_get(inode); - rc = obd_pin(sbi->ll_md_exp, ll_inode2fid(inode), oc, handle, flag); - capa_put(oc); - if (rc) { - lock_kernel(); - memset(handle, 0, sizeof(*handle)); - if (flag == 0) - ldd->lld_cwd_count--; - else - ldd->lld_mnt_count--; - unlock_kernel(); - } - - EXIT; - return; -} - -/*static*/ void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag) +/* + * Always trust cached dentries. Update statahead window if necessary. + */ +#ifdef HAVE_IOP_ATOMIC_OPEN +static int ll_revalidate_nd(struct dentry *dentry, unsigned int flags) { - 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; - 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; - } + int rc; + ENTRY; - if (flag) - count = --ldd->lld_mnt_count; - else - count = --ldd->lld_cwd_count; - unlock_kernel(); + CDEBUG(D_VFSTRACE, "VFS Op:name=%s, flags=%u\n", + dentry->d_name.name, flags); - if (count != 0) { - EXIT; - return; - } - - rc = obd_unpin(sbi->ll_md_exp, &handle, flag); - EXIT; - return; + rc = ll_revalidate_dentry(dentry, flags); + RETURN(rc); } - -#ifdef HAVE_VFS_INTENT_PATCHES +#else static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd) { - int rc; - ENTRY; + int rc; + ENTRY; - if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST)) - rc = ll_revalidate_it(dentry, nd->flags, &nd->intent); - else - rc = ll_revalidate_it(dentry, 0, NULL); - - RETURN(rc); -} -#else -int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd) -{ - int rc; - ENTRY; + /* + * this is normally called from NFS export, and we don't know whether + * this is the last component. + */ + if (nd == NULL) + RETURN(1); - if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) { - struct lookup_intent *it; - it = ll_convert_intent(&nd->intent.open, nd->flags); - if (IS_ERR(it)) - RETURN(0); - if (it->it_op == (IT_OPEN|IT_CREAT)) - if (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*/ -#ifdef HAVE_FILE_IN_STRUCT_INTENT -// 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 (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,17)) -/* 2.6.1[456] have a bug in open_namei() that forgets to check - * nd->intent.open.file for error, so we need to return it as lookup's result - * instead */ - if (IS_ERR(filp)) - rc = 0; -#endif - } -#else - ll_release_openhandle(dentry, it); -#endif /* HAVE_FILE_IN_STRUCT_INTENT */ - } - 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 */ - } - -out_it: - if (it) { - ll_intent_release(it); - OBD_FREE(it, sizeof(*it)); - } - } else { - rc = ll_revalidate_it(dentry, 0, NULL); - } + CDEBUG(D_VFSTRACE, "VFS Op:name=%s, flags=%u\n", + dentry->d_name.name, nd->flags); - RETURN(rc); + rc = ll_revalidate_dentry(dentry, nd->flags); + RETURN(rc); } #endif -struct dentry_operations ll_d_ops = { +const struct dentry_operations ll_d_ops = { .d_revalidate = ll_revalidate_nd, .d_release = ll_release, - .d_delete = ll_ddelete, -#ifdef DCACHE_LUSTRE_INVALID + .d_delete = ll_ddelete, .d_compare = ll_dcompare, -#endif -#if 0 - .d_pin = ll_pin, - .d_unpin = ll_unpin, -#endif };