This MDS reconnect error path leaves s_cap_reconnect set.
send_mds_reconnect() sets the bit at the beginning of the reconnect,
but the first failing operation after that, ceph_pagelist_encode_32(),
can jump to `fail:` without clearing it.

__ceph_remove_cap() consults that flag to decide whether cap releases
should be queued. A reconnect-preparation failure therefore leaves the
session in reconnect mode from the cap-release path's point of view
and can strand release work until some later state transition repairs
it.

Signed-off-by: Max Kellermann <max.kellermann@ionos.com>
---
 fs/ceph/mds_client.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index b1746273f186..4fa471d9b3b2 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -4956,7 +4956,7 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
 	/* placeholder for nr_caps */
 	err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
 	if (err)
-		goto fail;
+		goto fail_clear_cap_reconnect;
 
 	if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
 		recon_state.msg_version = 3;
@@ -5046,6 +5046,10 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
 	ceph_pagelist_release(recon_state.pagelist);
 	return;
 
+fail_clear_cap_reconnect:
+	spin_lock(&session->s_cap_lock);
+	session->s_cap_reconnect = 0;
+	spin_unlock(&session->s_cap_lock);
 fail:
 	ceph_msg_put(reply);
 	up_read(&mdsc->snap_rwsem);
-- 
2.47.3

On Mon, 2026-03-30 at 10:43 +0200, Max Kellermann wrote:
> This MDS reconnect error path leaves s_cap_reconnect set.
> send_mds_reconnect() sets the bit at the beginning of the reconnect,
> but the first failing operation after that, ceph_pagelist_encode_32(),
> can jump to `fail:` without clearing it.
> 
> __ceph_remove_cap() consults that flag to decide whether cap releases
> should be queued. A reconnect-preparation failure therefore leaves the
> session in reconnect mode from the cap-release path's point of view
> and can strand release work until some later state transition repairs
> it.
> 
> Signed-off-by: Max Kellermann <max.kellermann@ionos.com>
> ---
>  fs/ceph/mds_client.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
> index b1746273f186..4fa471d9b3b2 100644
> --- a/fs/ceph/mds_client.c
> +++ b/fs/ceph/mds_client.c
> @@ -4956,7 +4956,7 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
>  	/* placeholder for nr_caps */
>  	err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
>  	if (err)
> -		goto fail;
> +		goto fail_clear_cap_reconnect;
>  
>  	if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
>  		recon_state.msg_version = 3;
> @@ -5046,6 +5046,10 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
>  	ceph_pagelist_release(recon_state.pagelist);
>  	return;
>  
> +fail_clear_cap_reconnect:
> +	spin_lock(&session->s_cap_lock);
> +	session->s_cap_reconnect = 0;
> +	spin_unlock(&session->s_cap_lock);
>  fail:
>  	ceph_msg_put(reply);
>  	up_read(&mdsc->snap_rwsem);

Makes sense. Nice fix. But, maybe, we need to consider refactoring of this
function because of such issue.

Reviewed-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>

Thanks,
Slava.

On Mon, 2026-03-30 at 17:29 +0000, Viacheslav Dubeyko wrote:
> On Mon, 2026-03-30 at 10:43 +0200, Max Kellermann wrote:
> > This MDS reconnect error path leaves s_cap_reconnect set.
> > send_mds_reconnect() sets the bit at the beginning of the reconnect,
> > but the first failing operation after that, ceph_pagelist_encode_32(),
> > can jump to `fail:` without clearing it.
> > 
> > __ceph_remove_cap() consults that flag to decide whether cap releases
> > should be queued. A reconnect-preparation failure therefore leaves the
> > session in reconnect mode from the cap-release path's point of view
> > and can strand release work until some later state transition repairs
> > it.
> > 
> > Signed-off-by: Max Kellermann <max.kellermann@ionos.com>
> > ---
> >  fs/ceph/mds_client.c | 6 +++++-
> >  1 file changed, 5 insertions(+), 1 deletion(-)
> > 
> > diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
> > index b1746273f186..4fa471d9b3b2 100644
> > --- a/fs/ceph/mds_client.c
> > +++ b/fs/ceph/mds_client.c
> > @@ -4956,7 +4956,7 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
> >  	/* placeholder for nr_caps */
> >  	err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
> >  	if (err)
> > -		goto fail;
> > +		goto fail_clear_cap_reconnect;
> >  
> >  	if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
> >  		recon_state.msg_version = 3;
> > @@ -5046,6 +5046,10 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
> >  	ceph_pagelist_release(recon_state.pagelist);
> >  	return;
> >  
> > +fail_clear_cap_reconnect:
> > +	spin_lock(&session->s_cap_lock);
> > +	session->s_cap_reconnect = 0;
> > +	spin_unlock(&session->s_cap_lock);
> >  fail:
> >  	ceph_msg_put(reply);
> >  	up_read(&mdsc->snap_rwsem);
> 
> Makes sense. Nice fix. But, maybe, we need to consider refactoring of this
> function because of such issue.
> 
> Reviewed-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>
> 
> 

The xfstests run on 7.0-rc6 didn't reveal any new issues related to the patch.

Tested-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>

Thanks,
Slava.  

On Thu, 2026-04-02 at 18:12 +0000, Viacheslav Dubeyko wrote:
> On Mon, 2026-03-30 at 17:29 +0000, Viacheslav Dubeyko wrote:
> > On Mon, 2026-03-30 at 10:43 +0200, Max Kellermann wrote:
> > > This MDS reconnect error path leaves s_cap_reconnect set.
> > > send_mds_reconnect() sets the bit at the beginning of the reconnect,
> > > but the first failing operation after that, ceph_pagelist_encode_32(),
> > > can jump to `fail:` without clearing it.
> > > 
> > > __ceph_remove_cap() consults that flag to decide whether cap releases
> > > should be queued. A reconnect-preparation failure therefore leaves the
> > > session in reconnect mode from the cap-release path's point of view
> > > and can strand release work until some later state transition repairs
> > > it.
> > > 
> > > Signed-off-by: Max Kellermann <max.kellermann@ionos.com>
> > > ---
> > >  fs/ceph/mds_client.c | 6 +++++-
> > >  1 file changed, 5 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
> > > index b1746273f186..4fa471d9b3b2 100644
> > > --- a/fs/ceph/mds_client.c
> > > +++ b/fs/ceph/mds_client.c
> > > @@ -4956,7 +4956,7 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
> > >  	/* placeholder for nr_caps */
> > >  	err = ceph_pagelist_encode_32(recon_state.pagelist, 0);
> > >  	if (err)
> > > -		goto fail;
> > > +		goto fail_clear_cap_reconnect;
> > >  
> > >  	if (test_bit(CEPHFS_FEATURE_MULTI_RECONNECT, &session->s_features)) {
> > >  		recon_state.msg_version = 3;
> > > @@ -5046,6 +5046,10 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
> > >  	ceph_pagelist_release(recon_state.pagelist);
> > >  	return;
> > >  
> > > +fail_clear_cap_reconnect:
> > > +	spin_lock(&session->s_cap_lock);
> > > +	session->s_cap_reconnect = 0;
> > > +	spin_unlock(&session->s_cap_lock);
> > >  fail:
> > >  	ceph_msg_put(reply);
> > >  	up_read(&mdsc->snap_rwsem);
> > 
> > Makes sense. Nice fix. But, maybe, we need to consider refactoring of this
> > function because of such issue.
> > 
> > Reviewed-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>
> > 
> > 
> 
> The xfstests run on 7.0-rc6 didn't reveal any new issues related to the patch.
> 
> Tested-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>
> 

Applied on testing branch of CephFS kernel client.

Thanks,
Slava.