diff mbox

ceph: drop negtive child dentries before try pruning inode's alias

Message ID 20171130141259.8577-1-zyan@redhat.com (mailing list archive)
State New, archived
Headers show

Commit Message

Yan, Zheng Nov. 30, 2017, 2:12 p.m. UTC
negtive child dentry holds reference on inode's alias, it makes
d_prune_aliases() do nothing.

Signed-off-by: "Yan, Zheng" <zyan@redhat.com>
---
 fs/ceph/mds_client.c | 43 ++++++++++++++++++++++++++++++++++++-------
 1 file changed, 36 insertions(+), 7 deletions(-)

Comments

Jeff Layton Nov. 30, 2017, 2:37 p.m. UTC | #1
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" <zyan@redhat.com>
> ---
>  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.
Yan, Zheng Dec. 1, 2017, 1:07 a.m. UTC | #2
On Thu, Nov 30, 2017 at 10:37 PM, Jeff Layton <jlayton@redhat.com> wrote:
> 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" <zyan@redhat.com>
>> ---
>>  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.
>

I checked other users of d_prune_aliases. Some of them call
d_invalidate/shrink_dcache_parent before calling d_prune_aliases, some
of them only call d_prune_aliases for non-directory inode. It seems
only ceph has this requirement (trim unused inodes so that MDS can
trim corresponding objects from its cache)

Regards
Yan, Zheng


> --
> Jeff Layton <jlayton@redhat.com>
> --
> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Jeff Layton Dec. 4, 2017, 3:25 p.m. UTC | #3
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" <zyan@redhat.com>
> ---
>  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);

It's sort of nasty to drop the reference here. It makes it harder to
track how that gets done when you acquire and drop references across a
function boundary like this. There's also no reason for it since you
always do it last here anyway.

It'd be nicer to just move that into the caller.

> +	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;
>  	}
>  

I think the rest looks fine though.
Ilya Dryomov Dec. 7, 2017, 1:21 p.m. UTC | #4
On Mon, Dec 4, 2017 at 4:25 PM, Jeff Layton <jlayton@redhat.com> wrote:
> 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" <zyan@redhat.com>
>> ---
>>  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);
>
> It's sort of nasty to drop the reference here. It makes it harder to
> track how that gets done when you acquire and drop references across a
> function boundary like this. There's also no reason for it since you
> always do it last here anyway.
>
> It'd be nicer to just move that into the caller.
>
>> +     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;
>>       }
>>
>
> I think the rest looks fine though.

Hi Jeff, Zheng,

What is the status of this patch?  It is marked for stable in the
testing branch and Zheng force-pushed it two days ago, addressing
comments.  Jeff, can you take a look?

Zheng, negtive and all_negtive typos need fixing.

Also, would it make sense to merge these two patches ("ceph: decease
session->s_trim_caps only after caps get trimmed" and "ceph: drop
negtive child dentries before try pruning inode's alias") into one?
They seem to fix the same bug and the second almost entirely redoes the
first.

Thanks,

                Ilya
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Jeff Layton Dec. 7, 2017, 8:42 p.m. UTC | #5
On Thu, 2017-12-07 at 14:21 +0100, Ilya Dryomov wrote:
> On Mon, Dec 4, 2017 at 4:25 PM, Jeff Layton <jlayton@redhat.com> wrote:
> > 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" <zyan@redhat.com>
> > > ---
> > >  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);
> > 
> > It's sort of nasty to drop the reference here. It makes it harder to
> > track how that gets done when you acquire and drop references across a
> > function boundary like this. There's also no reason for it since you
> > always do it last here anyway.
> > 
> > It'd be nicer to just move that into the caller.
> > 
> > > +     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;
> > >       }
> > > 
> > 
> > I think the rest looks fine though.
> 
> Hi Jeff, Zheng,
> 
> What is the status of this patch?  It is marked for stable in the
> testing branch and Zheng force-pushed it two days ago, addressing
> comments.  Jeff, can you take a look?
> 
> Zheng, negtive and all_negtive typos need fixing.
> 
> Also, would it make sense to merge these two patches ("ceph: decease
> session->s_trim_caps only after caps get trimmed" and "ceph: drop
> negtive child dentries before try pruning inode's alias") into one?
> They seem to fix the same bug and the second almost entirely redoes the
> first.
> 


Yep, that one (56b5c1eed996) looks good to me.
Yan, Zheng Dec. 8, 2017, 7:32 a.m. UTC | #6
On Thu, Dec 7, 2017 at 9:21 PM, Ilya Dryomov <idryomov@gmail.com> wrote:
> On Mon, Dec 4, 2017 at 4:25 PM, Jeff Layton <jlayton@redhat.com> wrote:
>> 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" <zyan@redhat.com>
>>> ---
>>>  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);
>>
>> It's sort of nasty to drop the reference here. It makes it harder to
>> track how that gets done when you acquire and drop references across a
>> function boundary like this. There's also no reason for it since you
>> always do it last here anyway.
>>
>> It'd be nicer to just move that into the caller.
>>
>>> +     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;
>>>       }
>>>
>>
>> I think the rest looks fine though.
>
> Hi Jeff, Zheng,
>
> What is the status of this patch?  It is marked for stable in the
> testing branch and Zheng force-pushed it two days ago, addressing
> comments.  Jeff, can you take a look?
>
> Zheng, negtive and all_negtive typos need fixing.
>
> Also, would it make sense to merge these two patches ("ceph: decease
> session->s_trim_caps only after caps get trimmed" and "ceph: drop
> negtive child dentries before try pruning inode's alias") into one?
> They seem to fix the same bug and the second almost entirely redoes the
> first.
>

I merged them and re-push the new one to testing branch.

Thank all of you.


> Thanks,
>
>                 Ilya
> --
> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

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;
 	}