[PATCH] nfsd: prevent a double put of revoked layout stateids

[PATCH] nfsd: prevent a double put of revoked layout stateids

[Jinpyo Lee]

nfsd4_free_stateid() finds an admin-revoked layout stateid with
find_stateid_locked(), which does not acquire a reference. Since the
layout case does not mark the stateid closed before dropping cl_lock,
two FREE_STATEID operations in separate session slots can both select
the same stateid and drop a reference they do not own.

With fence-work teardown in place, the second caller can also wait in
nfsd4_stop_layout_fence() until the first caller frees the stateid, then
resume through its freed fence mutex and layout lock.

Mark the layout stateid CLOSED under cl_lock and ls_lock so only one
disposal caller proceeds. Take a temporary reference before dropping
cl_lock and hold it while stopping fence work and draining the remaining
layout segments, then release that reference with the final put.

The layout stateid IDR does not own a separate reference. The initial
allocation reference is released when LAYOUTGET completes, and each
layout segment holds its own reference. Serialize segment draining with
in-flight layout operations using ls_mutex, and recheck CLOSED after
acquiring that mutex so a stale LAYOUTGET cannot add a segment after the
drain. Also make the newly allocated layout stateid the error-path put
target.

Read CLOSED under ls_lock in FREE_STATEID, and restrict
nfs40_clean_admin_revoked() to NFSv4.0 clients. Session clients dispose
of revoked stateids through FREE_STATEID and must not be processed by
the NFSv4.0 cleanup while layout disposal is in progress.

The vulnerability research and validation were conducted by members of
the Tobabz team as part of the Best of the Best 15th program.

Fixes: 86b9898920a6 ("NFSD: Handle layout stid in nfsd4_drop_revoked_stid()")
Assisted-by: LLM
Signed-off-by: Jinpyo Lee <bint4b13@gmail.com>
---
Validated on nfsd-testing d918c664efd5ec6a7309e98f567fb018bcde0e11
(2026-10-12):
- Applied cleanly and passed strict checkpatch with no errors or warnings.
- Built and booted an x86_64 kernel and modules with GCC 13.3, W=1,
  Generic KASAN, lockdep, and pNFS block and SCSI layouts enabled.
- NFSv4.2, NFSv4.0, and NFSv3 read/write, fio, and unmount tests passed.
- On a real loopback pNFS SCSI layout, administrator revocation followed
  by one FREE_STATEID removed the layout stateid and reduced the
  admin-revoked count from 2 to 1. The sequence passed twice.
- Tracing observed nfsd4_stop_layout_fence() in both revocation and
  FREE_STATEID processing, followed by nfsd_layoutstate_free.
- No KASAN, lockdep, refcount, list-corruption, or Oops report occurred;
  kernel taint remained zero.

The full xfstests suite and a forced concurrent two-slot FREE_STATEID
race were not run on this revision.

 fs/nfsd/nfs4layouts.c | 24 ++++++++++++++++++++++++
 fs/nfsd/nfs4state.c   | 37 ++++++++++++++++++++++++++++++++++++-
 fs/nfsd/pnfs.h        |  6 ++++++
 3 files changed, 66 insertions(+), 1 deletion(-)

diff --git a/fs/nfsd/nfs4layouts.c b/fs/nfsd/nfs4layouts.c
index 898776c9d..74c39e216 100644
--- a/fs/nfsd/nfs4layouts.c
+++ b/fs/nfsd/nfs4layouts.c
@@ -312,6 +312,7 @@ nfsd4_preprocess_layout_stateid(struct svc_rqst *rqstp,
 		status = nfserr_jukebox;
 		if (!ls)
 			goto out;
+		stid = &ls->ls_stid;
 		mutex_lock(&ls->ls_mutex);
 	} else {
 		ls = container_of(stid, struct nfs4_layout_stateid, ls_stid);
@@ -324,6 +325,15 @@ nfsd4_preprocess_layout_stateid(struct svc_rqst *rqstp,
 			goto out_unlock_stid;
 	}
 
+	/* A new or looked-up stateid may have been closed before ls_mutex. */
+	spin_lock(&ls->ls_lock);
+	if (ls->ls_stid.sc_status & SC_STATUS_CLOSED) {
+		spin_unlock(&ls->ls_lock);
+		status = nfserr_bad_stateid;
+		goto out_unlock_stid;
+	}
+	spin_unlock(&ls->ls_lock);
+
 	*lsp = ls;
 	return 0;
 
@@ -594,6 +604,20 @@ nfsd4_return_all_layouts(struct nfs4_layout_stateid *ls,
 	spin_unlock(&ls->ls_lock);
 }
 
+void nfsd4_return_revoked_layout(struct nfs4_layout_stateid *ls)
+{
+	LIST_HEAD(reaplist);
+
+	/*
+	 * The caller holds a stateid reference until segment draining ends.
+	 * Wait for any LAYOUTGET that passed the CLOSED check.
+	 */
+	mutex_lock(&ls->ls_mutex);
+	nfsd4_return_all_layouts(ls, &reaplist);
+	mutex_unlock(&ls->ls_mutex);
+	nfsd4_free_layouts(&reaplist);
+}
+
 void
 nfsd4_return_all_client_layouts(struct nfs4_client *clp)
 {
diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index 920967dd6..02bfbd242 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -5837,6 +5837,24 @@ nfsd4_find_existing_open(struct nfs4_file *fp, struct nfsd4_open *open)
 	return ret;
 }
 
+/* cl_lock is held; layout stateid status also requires ls_lock. */
+static bool nfsd4_stid_is_closed(struct nfs4_stid *s)
+{
+	struct nfs4_layout_stateid *ls;
+	bool closed;
+
+	lockdep_assert_held(&s->sc_client->cl_lock);
+
+	if (s->sc_type != SC_TYPE_LAYOUT)
+		return s->sc_status & SC_STATUS_CLOSED;
+
+	ls = layoutstateid(s);
+	spin_lock(&ls->ls_lock);
+	closed = s->sc_status & SC_STATUS_CLOSED;
+	spin_unlock(&ls->ls_lock);
+	return closed;
+}
+
 static void nfsd4_drop_revoked_stid(struct nfs4_stid *s)
 	__releases(&s->sc_client->cl_lock)
 {
@@ -5871,9 +5889,24 @@ static void nfsd4_drop_revoked_stid(struct nfs4_stid *s)
 		break;
 	case SC_TYPE_LAYOUT:
 		ls = layoutstateid(s);
+		spin_lock(&ls->ls_lock);
+		if (s->sc_status & SC_STATUS_CLOSED) {
+			spin_unlock(&ls->ls_lock);
+			spin_unlock(&cl->cl_lock);
+			break;
+		}
+		s->sc_status |= SC_STATUS_CLOSED;
+		spin_unlock(&ls->ls_lock);
+		/*
+		 * The IDR does not own a layout stateid reference. Keep ls
+		 * alive while draining worker and segment references below;
+		 * the final put releases this temporary reference.
+		 */
+		refcount_inc(&s->sc_count);
 		list_del_init(&ls->ls_perclnt);
 		spin_unlock(&cl->cl_lock);
 		nfsd4_stop_layout_fence(ls);
+		nfsd4_return_revoked_layout(ls);
 		nfs4_put_stid(s);
 		break;
 	default:
@@ -7899,6 +7932,8 @@ static void nfs40_clean_admin_revoked(struct nfsd_net *nn,
 		unsigned long id, tmp;
 		struct nfs4_stid *stid;
 
+		if (clp->cl_minorversion != 0)
+			continue;
 		if (atomic_read(&clp->cl_admin_revoked) == 0)
 			continue;
 		if (is_client_expired(clp))
@@ -8698,7 +8733,7 @@ nfsd4_free_stateid(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
 
 	spin_lock(&cl->cl_lock);
 	s = find_stateid_locked(cl, stateid);
-	if (!s || s->sc_status & SC_STATUS_CLOSED)
+	if (!s || nfsd4_stid_is_closed(s))
 		goto out_unlock;
 	if (s->sc_status & SC_STATUS_ADMIN_REVOKED) {
 		nfsd4_drop_revoked_stid(s);
diff --git a/fs/nfsd/pnfs.h b/fs/nfsd/pnfs.h
index 559df0e5b..9b0f17a4d 100644
--- a/fs/nfsd/pnfs.h
+++ b/fs/nfsd/pnfs.h
@@ -75,6 +75,7 @@ struct nfsd4_deviceid_map *nfsd4_find_devid_map(int idx);
 #ifdef CONFIG_NFSD_PNFS
 void nfsd4_setup_layout_type(struct svc_export *exp);
 void nfsd4_return_all_client_layouts(struct nfs4_client *);
+void nfsd4_return_revoked_layout(struct nfs4_layout_stateid *ls);
 void nfsd4_return_all_file_layouts(struct nfs4_client *clp,
 		struct nfs4_file *fp);
 void nfsd4_close_layout(struct nfs4_layout_stateid *ls);
@@ -93,6 +94,11 @@ static inline void nfsd4_setup_layout_type(struct svc_export *exp)
 static inline void nfsd4_return_all_client_layouts(struct nfs4_client *clp)
 {
 }
+
+static inline void nfsd4_return_revoked_layout(struct nfs4_layout_stateid *ls)
+{
+}
+
 static inline void nfsd4_return_all_file_layouts(struct nfs4_client *clp,
 		struct nfs4_file *fp)
 {

base-commit: d918c664efd5ec6a7309e98f567fb018bcde0e11
-- 
2.50.1 (Apple Git-155)

Re: [PATCH] nfsd: prevent a double put of revoked layout stateids

[Jeff Layton]

On Mon, 2026-10-12 at 17:47 +0900, Jinpyo Lee wrote:
> nfsd4_free_stateid() finds an admin-revoked layout stateid with
> find_stateid_locked(), which does not acquire a reference. Since the
> layout case does not mark the stateid closed before dropping cl_lock,
> two FREE_STATEID operations in separate session slots can both select
> the same stateid and drop a reference they do not own.
> 
> With fence-work teardown in place, the second caller can also wait in
> nfsd4_stop_layout_fence() until the first caller frees the stateid, then
> resume through its freed fence mutex and layout lock.
> 
> Mark the layout stateid CLOSED under cl_lock and ls_lock so only one
> disposal caller proceeds. Take a temporary reference before dropping
> cl_lock and hold it while stopping fence work and draining the remaining
> layout segments, then release that reference with the final put.
> 
> The layout stateid IDR does not own a separate reference. The initial
> allocation reference is released when LAYOUTGET completes, and each
> layout segment holds its own reference. Serialize segment draining with
> in-flight layout operations using ls_mutex, and recheck CLOSED after
> acquiring that mutex so a stale LAYOUTGET cannot add a segment after the
> drain. Also make the newly allocated layout stateid the error-path put
> target.
> 
> Read CLOSED under ls_lock in FREE_STATEID, and restrict
> nfs40_clean_admin_revoked() to NFSv4.0 clients. Session clients dispose
> of revoked stateids through FREE_STATEID and must not be processed by
> the NFSv4.0 cleanup while layout disposal is in progress.
> 
> The vulnerability research and validation were conducted by members of
> the Tobabz team as part of the Best of the Best 15th program.
> 
> Fixes: 86b9898920a6 ("NFSD: Handle layout stid in nfsd4_drop_revoked_stid()")
> Assisted-by: LLM
> Signed-off-by: Jinpyo Lee <bint4b13@gmail.com>
> ---
> Validated on nfsd-testing d918c664efd5ec6a7309e98f567fb018bcde0e11
> (2026-10-12):
> - Applied cleanly and passed strict checkpatch with no errors or warnings.
> - Built and booted an x86_64 kernel and modules with GCC 13.3, W=1,
>   Generic KASAN, lockdep, and pNFS block and SCSI layouts enabled.
> - NFSv4.2, NFSv4.0, and NFSv3 read/write, fio, and unmount tests passed.
> - On a real loopback pNFS SCSI layout, administrator revocation followed
>   by one FREE_STATEID removed the layout stateid and reduced the
>   admin-revoked count from 2 to 1. The sequence passed twice.
> - Tracing observed nfsd4_stop_layout_fence() in both revocation and
>   FREE_STATEID processing, followed by nfsd_layoutstate_free.
> - No KASAN, lockdep, refcount, list-corruption, or Oops report occurred;
>   kernel taint remained zero.
> 
> The full xfstests suite and a forced concurrent two-slot FREE_STATEID
> race were not run on this revision.
> 
>  fs/nfsd/nfs4layouts.c | 24 ++++++++++++++++++++++++
>  fs/nfsd/nfs4state.c   | 37 ++++++++++++++++++++++++++++++++++++-
>  fs/nfsd/pnfs.h        |  6 ++++++
>  3 files changed, 66 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/nfsd/nfs4layouts.c b/fs/nfsd/nfs4layouts.c
> index 898776c9d..74c39e216 100644
> --- a/fs/nfsd/nfs4layouts.c
> +++ b/fs/nfsd/nfs4layouts.c
> @@ -312,6 +312,7 @@ nfsd4_preprocess_layout_stateid(struct svc_rqst *rqstp,
>  		status = nfserr_jukebox;
>  		if (!ls)
>  			goto out;
> +		stid = &ls->ls_stid;
>  		mutex_lock(&ls->ls_mutex);
>  	} else {
>  		ls = container_of(stid, struct nfs4_layout_stateid, ls_stid);
> @@ -324,6 +325,15 @@ nfsd4_preprocess_layout_stateid(struct svc_rqst *rqstp,
>  			goto out_unlock_stid;
>  	}
>  
> +	/* A new or looked-up stateid may have been closed before ls_mutex. */
> +	spin_lock(&ls->ls_lock);
> +	if (ls->ls_stid.sc_status & SC_STATUS_CLOSED) {
> +		spin_unlock(&ls->ls_lock);
> +		status = nfserr_bad_stateid;
> +		goto out_unlock_stid;
> +	}
> +	spin_unlock(&ls->ls_lock);
> +
>  	*lsp = ls;
>  	return 0;
>  
> @@ -594,6 +604,20 @@ nfsd4_return_all_layouts(struct nfs4_layout_stateid *ls,
>  	spin_unlock(&ls->ls_lock);
>  }
>  
> +void nfsd4_return_revoked_layout(struct nfs4_layout_stateid *ls)
> +{
> +	LIST_HEAD(reaplist);
> +
> +	/*
> +	 * The caller holds a stateid reference until segment draining ends.
> +	 * Wait for any LAYOUTGET that passed the CLOSED check.
> +	 */
> +	mutex_lock(&ls->ls_mutex);
> +	nfsd4_return_all_layouts(ls, &reaplist);
> +	mutex_unlock(&ls->ls_mutex);
> +	nfsd4_free_layouts(&reaplist);
> +}
> +
>  void
>  nfsd4_return_all_client_layouts(struct nfs4_client *clp)
>  {
> diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
> index 920967dd6..02bfbd242 100644
> --- a/fs/nfsd/nfs4state.c
> +++ b/fs/nfsd/nfs4state.c
> @@ -5837,6 +5837,24 @@ nfsd4_find_existing_open(struct nfs4_file *fp, struct nfsd4_open *open)
>  	return ret;
>  }
>  
> +/* cl_lock is held; layout stateid status also requires ls_lock. */
> +static bool nfsd4_stid_is_closed(struct nfs4_stid *s)
> +{
> +	struct nfs4_layout_stateid *ls;
> +	bool closed;
> +
> +	lockdep_assert_held(&s->sc_client->cl_lock);
> +
> +	if (s->sc_type != SC_TYPE_LAYOUT)
> +		return s->sc_status & SC_STATUS_CLOSED;
> +
> +	ls = layoutstateid(s);
> +	spin_lock(&ls->ls_lock);
> +	closed = s->sc_status & SC_STATUS_CLOSED;
> +	spin_unlock(&ls->ls_lock);
> +	return closed;
> +}
> +

^^^ 
I must say that I'm not thrilled with the different locking rules for
SC_STATUS_CLOSED with different types of stateids. That has a strong
code smell in my opinion and will be difficult (for humans) to keep
straight.

I don't have a better suggestion for the moment, but I'd strongly
prefer something that didn't add these extra rules around a flag common
to all stateids.

>  static void nfsd4_drop_revoked_stid(struct nfs4_stid *s)
>  	__releases(&s->sc_client->cl_lock)
>  {
> @@ -5871,9 +5889,24 @@ static void nfsd4_drop_revoked_stid(struct nfs4_stid *s)
>  		break;
>  	case SC_TYPE_LAYOUT:
>  		ls = layoutstateid(s);
> +		spin_lock(&ls->ls_lock);
> +		if (s->sc_status & SC_STATUS_CLOSED) {
> +			spin_unlock(&ls->ls_lock);
> +			spin_unlock(&cl->cl_lock);
> +			break;
> +		}
> +		s->sc_status |= SC_STATUS_CLOSED;
> +		spin_unlock(&ls->ls_lock);
> +		/*
> +		 * The IDR does not own a layout stateid reference. Keep ls
> +		 * alive while draining worker and segment references below;
> +		 * the final put releases this temporary reference.
> +		 */
> +		refcount_inc(&s->sc_count);
>  		list_del_init(&ls->ls_perclnt);
>  		spin_unlock(&cl->cl_lock);
>  		nfsd4_stop_layout_fence(ls);
> +		nfsd4_return_revoked_layout(ls);
>  		nfs4_put_stid(s);
>  		break;
>  	default:
> @@ -7899,6 +7932,8 @@ static void nfs40_clean_admin_revoked(struct nfsd_net *nn,
>  		unsigned long id, tmp;
>  		struct nfs4_stid *stid;
>  
> +		if (clp->cl_minorversion != 0)
> +			continue;
>  		if (atomic_read(&clp->cl_admin_revoked) == 0)
>  			continue;
>  		if (is_client_expired(clp))
> @@ -8698,7 +8733,7 @@ nfsd4_free_stateid(struct svc_rqst *rqstp, struct nfsd4_compound_state *cstate,
>  
>  	spin_lock(&cl->cl_lock);
>  	s = find_stateid_locked(cl, stateid);
> -	if (!s || s->sc_status & SC_STATUS_CLOSED)
> +	if (!s || nfsd4_stid_is_closed(s))
>  		goto out_unlock;
>  	if (s->sc_status & SC_STATUS_ADMIN_REVOKED) {
>  		nfsd4_drop_revoked_stid(s);
> diff --git a/fs/nfsd/pnfs.h b/fs/nfsd/pnfs.h
> index 559df0e5b..9b0f17a4d 100644
> --- a/fs/nfsd/pnfs.h
> +++ b/fs/nfsd/pnfs.h
> @@ -75,6 +75,7 @@ struct nfsd4_deviceid_map *nfsd4_find_devid_map(int idx);
>  #ifdef CONFIG_NFSD_PNFS
>  void nfsd4_setup_layout_type(struct svc_export *exp);
>  void nfsd4_return_all_client_layouts(struct nfs4_client *);
> +void nfsd4_return_revoked_layout(struct nfs4_layout_stateid *ls);
>  void nfsd4_return_all_file_layouts(struct nfs4_client *clp,
>  		struct nfs4_file *fp);
>  void nfsd4_close_layout(struct nfs4_layout_stateid *ls);
> @@ -93,6 +94,11 @@ static inline void nfsd4_setup_layout_type(struct svc_export *exp)
>  static inline void nfsd4_return_all_client_layouts(struct nfs4_client *clp)
>  {
>  }
> +
> +static inline void nfsd4_return_revoked_layout(struct nfs4_layout_stateid *ls)
> +{
> +}
> +
>  static inline void nfsd4_return_all_file_layouts(struct nfs4_client *clp,
>  		struct nfs4_file *fp)
>  {
> 
> base-commit: d918c664efd5ec6a7309e98f567fb018bcde0e11

-- 
Jeff Layton <jlayton@kernel.org>