From: Chuck Lever Down the road we want to eliminate the use of the global kernel lock entirely from the NFS client. To do this, we need to protect the fields in the nfs_inode structure adequately. Start by serializing updates to the "cache_validity" field. Note this change addresses an SMP hang found by njw@osdl.org, where processes deadlock because nfs_end_data_update and nfs_revalidate_mapping update the "cache_validity" field without proper serialization. Test plan: Millions of fsx ops on SMP clients. Run Nick Wilson's breaknfs program on large SMP clients. Signed-off-by: Chuck Lever Signed-off-by: Andrew Morton --- fs/nfs/dir.c | 6 ++++++ fs/nfs/inode.c | 37 +++++++++++++++++++++++++++++++++---- fs/nfs/nfs3acl.c | 2 ++ fs/nfs/read.c | 4 ++++ include/linux/nfs_fs.h | 5 ++++- 5 files changed, 49 insertions(+), 5 deletions(-) diff -puN fs/nfs/dir.c~nfs-introduce-the-use-of-inode-i_lock-to-protect-fields-in-nfsi fs/nfs/dir.c --- devel/fs/nfs/dir.c~nfs-introduce-the-use-of-inode-i_lock-to-protect-fields-in-nfsi 2005-07-26 00:38:30.000000000 -0700 +++ devel-akpm/fs/nfs/dir.c 2005-07-26 00:38:30.000000000 -0700 @@ -189,7 +189,9 @@ int nfs_readdir_filler(nfs_readdir_descr goto error; } SetPageUptodate(page); + spin_lock(&inode->i_lock); NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME; + spin_unlock(&inode->i_lock); /* Ensure consistent page alignment of the data. * Note: assumes we have exclusive access to this mapping either * through inode->i_sem or some other mechanism. @@ -462,7 +464,9 @@ int uncached_readdir(nfs_readdir_descrip page, NFS_SERVER(inode)->dtsize, desc->plus); + spin_lock(&inode->i_lock); NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME; + spin_unlock(&inode->i_lock); desc->page = page; desc->ptr = kmap(page); /* matching kunmap in nfs_do_filldir */ if (desc->error >= 0) { @@ -1596,7 +1600,9 @@ void nfs_access_add_cache(struct inode * put_rpccred(cache->cred); cache->cred = get_rpccred(set->cred); } + spin_lock(&inode->i_lock); nfsi->cache_validity &= ~NFS_INO_INVALID_ACCESS; + spin_unlock(&inode->i_lock); cache->jiffies = set->jiffies; cache->mask = set->mask; } diff -puN fs/nfs/inode.c~nfs-introduce-the-use-of-inode-i_lock-to-protect-fields-in-nfsi fs/nfs/inode.c --- devel/fs/nfs/inode.c~nfs-introduce-the-use-of-inode-i_lock-to-protect-fields-in-nfsi 2005-07-26 00:38:30.000000000 -0700 +++ devel-akpm/fs/nfs/inode.c 2005-07-26 00:38:30.000000000 -0700 @@ -617,6 +617,8 @@ nfs_zap_caches(struct inode *inode) struct nfs_inode *nfsi = NFS_I(inode); int mode = inode->i_mode; + spin_lock(&inode->i_lock); + NFS_ATTRTIMEO(inode) = NFS_MINATTRTIMEO(inode); NFS_ATTRTIMEO_UPDATE(inode) = jiffies; @@ -625,6 +627,8 @@ nfs_zap_caches(struct inode *inode) nfsi->cache_validity |= NFS_INO_INVALID_ATTR|NFS_INO_INVALID_DATA|NFS_INO_INVALID_ACCESS|NFS_INO_INVALID_ACL|NFS_INO_REVAL_PAGECACHE; else nfsi->cache_validity |= NFS_INO_INVALID_ATTR|NFS_INO_INVALID_ACCESS|NFS_INO_INVALID_ACL|NFS_INO_REVAL_PAGECACHE; + + spin_unlock(&inode->i_lock); } static void nfs_zap_acl_cache(struct inode *inode) @@ -634,7 +638,9 @@ static void nfs_zap_acl_cache(struct ino clear_acl_cache = NFS_PROTO(inode)->clear_acl_cache; if (clear_acl_cache != NULL) clear_acl_cache(inode); + spin_lock(&inode->i_lock); NFS_I(inode)->cache_validity &= ~NFS_INO_INVALID_ACL; + spin_unlock(&inode->i_lock); } /* @@ -833,8 +839,11 @@ nfs_setattr(struct dentry *dentry, struc vmtruncate(inode, attr->ia_size); } } - if ((attr->ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID)) != 0) + if ((attr->ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID)) != 0) { + spin_lock(&inode->i_lock); NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ACCESS|NFS_INO_INVALID_ACL; + spin_unlock(&inode->i_lock); + } nfs_end_data_update(inode); unlock_kernel(); return error; @@ -1073,6 +1082,7 @@ __nfs_revalidate_inode(struct nfs_server (long long)NFS_FILEID(inode), status); goto out; } + spin_lock(&inode->i_lock); cache_validity = nfsi->cache_validity; nfsi->cache_validity &= ~NFS_INO_REVAL_PAGECACHE; @@ -1082,6 +1092,7 @@ __nfs_revalidate_inode(struct nfs_server */ if (verifier == nfsi->cache_change_attribute) nfsi->cache_validity &= ~(NFS_INO_INVALID_ATTR|NFS_INO_INVALID_ATIME); + spin_unlock(&inode->i_lock); nfs_revalidate_mapping(inode, inode->i_mapping); @@ -1140,12 +1151,16 @@ void nfs_revalidate_mapping(struct inode nfs_wb_all(inode); } invalidate_inode_pages2(mapping); + + spin_lock(&inode->i_lock); nfsi->cache_validity &= ~NFS_INO_INVALID_DATA; if (S_ISDIR(inode->i_mode)) { memset(nfsi->cookieverf, 0, sizeof(nfsi->cookieverf)); /* This ensures we revalidate child dentries */ nfsi->cache_change_attribute++; } + spin_unlock(&inode->i_lock); + dfprintk(PAGECACHE, "NFS: (%s/%Ld) data cache invalidated\n", inode->i_sb->s_id, (long long)NFS_FILEID(inode)); @@ -1175,10 +1190,12 @@ void nfs_end_data_update(struct inode *i if (!nfs_have_delegation(inode, FMODE_READ)) { /* Mark the attribute cache for revalidation */ + spin_lock(&inode->i_lock); nfsi->cache_validity |= NFS_INO_INVALID_ATTR; /* Directories and symlinks: invalidate page cache too */ if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode)) nfsi->cache_validity |= NFS_INO_INVALID_DATA; + spin_unlock(&inode->i_lock); } nfsi->cache_change_attribute ++; atomic_dec(&nfsi->data_updates); @@ -1203,6 +1220,8 @@ int nfs_refresh_inode(struct inode *inod if (nfs_have_delegation(inode, FMODE_READ)) return 0; + spin_lock(&inode->i_lock); + /* Are we in the process of updating data on the server? */ data_unstable = nfs_caches_unstable(inode); @@ -1217,13 +1236,17 @@ int nfs_refresh_inode(struct inode *inod } } - if ((fattr->valid & NFS_ATTR_FATTR) == 0) + if ((fattr->valid & NFS_ATTR_FATTR) == 0) { + spin_unlock(&inode->i_lock); return 0; + } /* Has the inode gone and changed behind our back? */ if (nfsi->fileid != fattr->fileid - || (inode->i_mode & S_IFMT) != (fattr->mode & S_IFMT)) + || (inode->i_mode & S_IFMT) != (fattr->mode & S_IFMT)) { + spin_unlock(&inode->i_lock); return -EIO; + } cur_size = i_size_read(inode); new_isize = nfs_size_to_loff_t(fattr->size); @@ -1262,6 +1285,7 @@ int nfs_refresh_inode(struct inode *inod nfsi->cache_validity |= NFS_INO_INVALID_ATIME; nfsi->read_cache_jiffies = fattr->timestamp; + spin_unlock(&inode->i_lock); return 0; } @@ -1300,11 +1324,15 @@ static int nfs_update_inode(struct inode goto out_err; } + spin_lock(&inode->i_lock); + /* * Make sure the inode's type hasn't changed. */ - if ((inode->i_mode & S_IFMT) != (fattr->mode & S_IFMT)) + if ((inode->i_mode & S_IFMT) != (fattr->mode & S_IFMT)) { + spin_unlock(&inode->i_lock); goto out_changed; + } /* * Update the read time so we don't revalidate too often. @@ -1397,6 +1425,7 @@ static int nfs_update_inode(struct inode if (!nfs_have_delegation(inode, FMODE_READ)) nfsi->cache_validity |= invalid; + spin_unlock(&inode->i_lock); return 0; out_changed: /* diff -puN fs/nfs/nfs3acl.c~nfs-introduce-the-use-of-inode-i_lock-to-protect-fields-in-nfsi fs/nfs/nfs3acl.c --- devel/fs/nfs/nfs3acl.c~nfs-introduce-the-use-of-inode-i_lock-to-protect-fields-in-nfsi 2005-07-26 00:38:30.000000000 -0700 +++ devel-akpm/fs/nfs/nfs3acl.c 2005-07-26 00:38:30.000000000 -0700 @@ -308,7 +308,9 @@ static int nfs3_proc_setacls(struct inod nfs_begin_data_update(inode); status = rpc_call(server->client_acl, ACLPROC3_SETACL, &args, &fattr, 0); + spin_lock(&inode->i_lock); NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ACCESS; + spin_unlock(&inode->i_lock); nfs_end_data_update(inode); dprintk("NFS reply setacl: %d\n", status); diff -puN fs/nfs/read.c~nfs-introduce-the-use-of-inode-i_lock-to-protect-fields-in-nfsi fs/nfs/read.c --- devel/fs/nfs/read.c~nfs-introduce-the-use-of-inode-i_lock-to-protect-fields-in-nfsi 2005-07-26 00:38:30.000000000 -0700 +++ devel-akpm/fs/nfs/read.c 2005-07-26 00:38:30.000000000 -0700 @@ -140,7 +140,9 @@ static int nfs_readpage_sync(struct nfs_ if (rdata->res.eof != 0 || result == 0) break; } while (count); + spin_lock(&inode->i_lock); NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATIME; + spin_unlock(&inode->i_lock); if (count) memclear_highpage_flush(page, rdata->args.pgbase, count); @@ -472,7 +474,9 @@ void nfs_readpage_result(struct rpc_task } task->tk_status = -EIO; } + spin_lock(&data->inode->i_lock); NFS_I(data->inode)->cache_validity |= NFS_INO_INVALID_ATIME; + spin_unlock(&data->inode->i_lock); data->complete(data, status); } diff -puN include/linux/nfs_fs.h~nfs-introduce-the-use-of-inode-i_lock-to-protect-fields-in-nfsi include/linux/nfs_fs.h --- devel/include/linux/nfs_fs.h~nfs-introduce-the-use-of-inode-i_lock-to-protect-fields-in-nfsi 2005-07-26 00:38:30.000000000 -0700 +++ devel-akpm/include/linux/nfs_fs.h 2005-07-26 00:38:30.000000000 -0700 @@ -238,8 +238,11 @@ static inline int nfs_caches_unstable(st static inline void NFS_CACHEINV(struct inode *inode) { - if (!nfs_caches_unstable(inode)) + if (!nfs_caches_unstable(inode)) { + spin_lock(&inode->i_lock); NFS_I(inode)->cache_validity |= NFS_INO_INVALID_ATTR | NFS_INO_INVALID_ACCESS; + spin_unlock(&inode->i_lock); + } } static inline int nfs_server_capable(struct inode *inode, int cap) _