From: Jeff Layton Subject: Re: [PATCH] ceph: drop negtive child dentries before try pruning inode's alias Date: Thu, 30 Nov 2017 09:37:41 -0500 Message-ID: <1512052661.4132.4.camel@redhat.com> References: <20171130141259.8577-1-zyan@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Return-path: Received: from mail-qk0-f195.google.com ([209.85.220.195]:37457 "EHLO mail-qk0-f195.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750943AbdK3Oho (ORCPT ); Thu, 30 Nov 2017 09:37:44 -0500 Received: by mail-qk0-f195.google.com with SMTP id i130so9058736qke.4 for ; Thu, 30 Nov 2017 06:37:44 -0800 (PST) In-Reply-To: <20171130141259.8577-1-zyan@redhat.com> Sender: ceph-devel-owner@vger.kernel.org List-ID: To: "Yan, Zheng" , ceph-devel@vger.kernel.org On Thu, 2017-11-30 at 22:12 +0800, Yan, Zheng wrote: > negtive child dentry holds reference on inode's alias, it makes > d_prune_aliases() do nothing. > > Signed-off-by: "Yan, Zheng" > --- > fs/ceph/mds_client.c | 43 ++++++++++++++++++++++++++++++++++++------- > 1 file changed, 36 insertions(+), 7 deletions(-) > > diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c > index da181acd4a61..9e7bb5fa9295 100644 > --- a/fs/ceph/mds_client.c > +++ b/fs/ceph/mds_client.c > @@ -1440,6 +1440,30 @@ static int __close_session(struct ceph_mds_client *mdsc, > return request_close_session(mdsc, session); > } > > +static bool drop_negative_children(struct dentry *dentry) > +{ > + struct dentry *child; > + bool all_negtive = true; > + > + if (!d_is_dir(dentry)) > + goto out; > + > + spin_lock(&dentry->d_lock); > + list_for_each_entry(child, &dentry->d_subdirs, d_child) { > + if (d_really_is_positive(child)) { > + all_negtive = false; > + break; > + } > + } > + spin_unlock(&dentry->d_lock); > + > + if (all_negtive) > + shrink_dcache_parent(dentry); > +out: > + dput(dentry); > + return all_negtive; > +} > + > /* > * Trim old(er) caps. > * > @@ -1495,15 +1519,20 @@ static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg) > __ceph_remove_cap(cap, true); > session->s_trim_caps--; > } else { > - int refs; > + struct dentry *dentry; > /* try dropping referring dentries */ > spin_unlock(&ci->i_ceph_lock); > - d_prune_aliases(inode); > - refs = atomic_read(&inode->i_count); > - if (refs == 1) > - session->s_trim_caps--; > - dout("trim_caps_cb %p cap %p pruned, count now %d\n", > - inode, cap, refs); > + dentry = d_find_any_alias(inode); > + /* drop_negative_children() calls dput() */ > + if (dentry && drop_negative_children(dentry)) { > + int refs; > + d_prune_aliases(inode); > + refs = atomic_read(&inode->i_count); > + if (refs == 1) > + session->s_trim_caps--; > + dout("trim_caps_cb %p cap %p pruned, count now %d\n", > + inode, cap, refs); > + } > return 0; > } > Would it be better to do this at the VFS layer? Dropping negative children when we're going to prune dentries for an inode seems like it ought to be done as a matter of course in d_prune_aliases. -- Jeff Layton