Skip to contentSkip to conversation

NFSD: memory shrinker for NFSv4 clients

PATCH Superseded Dai Ngo 3 participants 5 messages Aug 30, 2022

Subsystems (2)

v3 of 7 Superseded Latest is v7 →

lore
Share & tools
lore

Dai Ngo

author reviewer
#0 lore raw

v3 cover NFSD: memory shrinker for NFSv4 clients

This patch series implements the memory shrinker for NFSv4 clients to react to system low memory condition.

The memory shrinker's count callback is used to trigger the laundromat. The actual work of destroying the expired clients is done by the laundromat itself. We can not destroying the expired clients on the memory shrinler's scan callback context to avoid possible deadlock.

By destroying the expired clients, all states associated with these clients are also released.

v2:
. fix kernel test robot errors in nfsd.h when CONFIG_NFSD_V4 not defined.

v3:
. add mod_delayed_work in nfsd_courtesy_client_scan to kick start
the laundromat.
---

Dai Ngo (2):
NFSD: keep track of the number of courtesy clients in the system
NFSD: add shrinker to reap courtesy clients on low memory condition

 fs/nfsd/netns.h     |  5 ++++
 fs/nfsd/nfs4state.c | 66 +++++++++++++++++++++++++++++++++++++++++++-----
 fs/nfsd/nfsctl.c    |  6 +++--
 fs/nfsd/nfsd.h      |  9 +++++--
 4 files changed, 75 insertions(+), 11 deletions(-)
author reviewer
#1 lore raw

PATCH v3 1/2 NFSD: keep track of the number of courtesy clients in the system

Add counter nfs4_courtesy_client_count to nfsd_net to keep track of the number of courtesy clients in the system.

---
 fs/nfsd/netns.h     |  2 ++
 fs/nfsd/nfs4state.c | 14 +++++++++++---
 2 files changed, 13 insertions(+), 3 deletions(-)
fs/nfsd/netns.h +2 -0
index ffe17743cc74..2695dff1378a 100644
@@ -192,6 +192,8 @@ struct nfsd_net {
 
 	atomic_t		nfs4_client_count;
 	int			nfs4_max_clients;
+
+	atomic_t		nfsd_courtesy_client_count;
 };
 
 /* Simple check to find out if a given net was properly initialized */
fs/nfsd/nfs4state.c +11 -3
index c5d199d7e6b4..9675b5d8f408 100644
@@ -169,7 +169,8 @@ static __be32 get_client_locked(struct nfs4_client *clp)
 	if (is_client_expired(clp))
 		return nfserr_expired;
 	atomic_inc(&clp->cl_rpc_users);
-	clp->cl_state = NFSD4_ACTIVE;
+	if (xchg(&clp->cl_state, NFSD4_ACTIVE) != NFSD4_ACTIVE)
+		atomic_add_unless(&nn->nfsd_courtesy_client_count, -1, 0);
 	return nfs_ok;
 }
 
@@ -190,7 +191,8 @@ renew_client_locked(struct nfs4_client *clp)
 
 	list_move_tail(&clp->cl_lru, &nn->client_lru);
 	clp->cl_time = ktime_get_boottime_seconds();
-	clp->cl_state = NFSD4_ACTIVE;
+	if (xchg(&clp->cl_state, NFSD4_ACTIVE) != NFSD4_ACTIVE)
+		atomic_add_unless(&nn->nfsd_courtesy_client_count, -1, 0);
 }
 
 static void put_client_renew_locked(struct nfs4_client *clp)
@@ -2233,6 +2235,8 @@ __destroy_client(struct nfs4_client *clp)
 	if (clp->cl_cb_conn.cb_xprt)
 		svc_xprt_put(clp->cl_cb_conn.cb_xprt);
 	atomic_add_unless(&nn->nfs4_client_count, -1, 0);
+	if (clp->cl_state != NFSD4_ACTIVE)
+		atomic_add_unless(&nn->nfsd_courtesy_client_count, -1, 0);
 	free_client(clp);
 	wake_up_all(&expiry_wq);
 }
@@ -4356,6 +4360,8 @@ void nfsd4_init_leases_net(struct nfsd_net *nn)
 	max_clients = (u64)si.totalram * si.mem_unit / (1024 * 1024 * 1024);
 	max_clients *= NFS4_CLIENTS_PER_GB;
 	nn->nfs4_max_clients = max_t(int, max_clients, NFS4_CLIENTS_PER_GB);
+
+	atomic_set(&nn->nfsd_courtesy_client_count, 0);
 }
 
 static void init_nfs4_replay(struct nfs4_replay *rp)
@@ -5879,7 +5885,9 @@ nfs4_get_client_reaplist(struct nfsd_net *nn, struct list_head *reaplist,
 		if (!state_expired(lt, clp->cl_time))
 			break;
 		if (!atomic_read(&clp->cl_rpc_users))
-			clp->cl_state = NFSD4_COURTESY;
+			if (xchg(&clp->cl_state, NFSD4_COURTESY) ==
+							NFSD4_ACTIVE)
+				atomic_inc(&nn->nfsd_courtesy_client_count);
 		if (!client_has_state(clp))
 			goto exp_client;
 		if (!nfs4_anylock_blockers(clp))
Signature · 1 line
2.9.5
maintainer
#2 lore raw

Re: [PATCH v3 1/2] NFSD: keep track of the number of courtesy clients in the system

Quoted 27 lines · Dai Ngo

Add counter nfs4_courtesy_client_count to nfsd_net to keep track of the number of courtesy clients in the system.

---
 fs/nfsd/netns.h     |  2 ++
 fs/nfsd/nfs4state.c | 14 +++++++++++---
 2 files changed, 13 insertions(+), 3 deletions(-)
fs/nfsd/netns.h +2 -0
index ffe17743cc74..2695dff1378a 100644
@@ -192,6 +192,8 @@ struct nfsd_net {
 
 	atomic_t		nfs4_client_count;
 	int			nfs4_max_clients;
+
+	atomic_t		nfsd_courtesy_client_count;
 };
 
 /* Simple check to find out if a given net was properly initialized */
fs/nfsd/nfs4state.c +0 -0
index c5d199d7e6b4..9675b5d8f408 100644

fs/nfsd/nfs4state.c static __be32 get_client_locked(struct nfs4_client *clp) · line 173

Quoted 6 lines
fs/nfsd/nfs4state.c +2 -1
@@ -169,7 +169,8 @@ static __be32 get_client_locked(struct nfs4_client *clp)
 	if (is_client_expired(clp))
 		return nfserr_expired;
 	atomic_inc(&clp->cl_rpc_users);
-	clp->cl_state = NFSD4_ACTIVE;
+	if (xchg(&clp->cl_state, NFSD4_ACTIVE) != NFSD4_ACTIVE)
+		atomic_add_unless(&nn->nfsd_courtesy_client_count, -1, 0);

I'd prefer this:

	if (clp->cl_state != NFSD4_ACTIVE)
		atomic_add_unless(...
	clp->cl_state = NFSD4_ACTIVE;

It's more lines, but it's easier to read. It's may also be more efficient since you're protected by a spinlock anyway.

I think it's also less deceptive when reading the code. When I see xchg() operations being used like that, my immediate thought is that they must be needed for synchronization purposes. It's not in this case.

Quoted 42 lines
fs/nfsd/nfs4state.c +9 -2
 	return nfs_ok;
 }
 
@@ -190,7 +191,8 @@ renew_client_locked(struct nfs4_client *clp)
 
 	list_move_tail(&clp->cl_lru, &nn->client_lru);
 	clp->cl_time = ktime_get_boottime_seconds();
-	clp->cl_state = NFSD4_ACTIVE;
+	if (xchg(&clp->cl_state, NFSD4_ACTIVE) != NFSD4_ACTIVE)
+		atomic_add_unless(&nn->nfsd_courtesy_client_count, -1, 0);
 }
 
 static void put_client_renew_locked(struct nfs4_client *clp)
@@ -2233,6 +2235,8 @@ __destroy_client(struct nfs4_client *clp)
 	if (clp->cl_cb_conn.cb_xprt)
 		svc_xprt_put(clp->cl_cb_conn.cb_xprt);
 	atomic_add_unless(&nn->nfs4_client_count, -1, 0);
+	if (clp->cl_state != NFSD4_ACTIVE)
+		atomic_add_unless(&nn->nfsd_courtesy_client_count, -1, 0);
 	free_client(clp);
 	wake_up_all(&expiry_wq);
 }
@@ -4356,6 +4360,8 @@ void nfsd4_init_leases_net(struct nfsd_net *nn)
 	max_clients = (u64)si.totalram * si.mem_unit / (1024 * 1024 * 1024);
 	max_clients *= NFS4_CLIENTS_PER_GB;
 	nn->nfs4_max_clients = max_t(int, max_clients, NFS4_CLIENTS_PER_GB);
+
+	atomic_set(&nn->nfsd_courtesy_client_count, 0);
 }
 
 static void init_nfs4_replay(struct nfs4_replay *rp)
@@ -5879,7 +5885,9 @@ nfs4_get_client_reaplist(struct nfsd_net *nn, struct list_head *reaplist,
 		if (!state_expired(lt, clp->cl_time))
 			break;
 		if (!atomic_read(&clp->cl_rpc_users))
-			clp->cl_state = NFSD4_COURTESY;
+			if (xchg(&clp->cl_state, NFSD4_COURTESY) ==
+							NFSD4_ACTIVE)
+				atomic_inc(&nn->nfsd_courtesy_client_count);
 		if (!client_has_state(clp))
 			goto exp_client;
 		if (!nfs4_anylock_blockers(clp))
Signature · 1 line
Jeff Layton <[email protected]>
#3 lore raw

Re: [PATCH v3 1/2] NFSD: keep track of the number of courtesy clients in the system

Quoted 45 lines · Jeff Layton
Quoted 27 lines · Dai Ngo

Add counter nfs4_courtesy_client_count to nfsd_net to keep track of the number of courtesy clients in the system.

---
fs/nfsd/netns.h | 2 ++
fs/nfsd/nfs4state.c | 14 +++++++++++---

2 files changed, 13 insertions(+), 3 deletions(-)
fs/nfsd/netns.h +2 -0
index ffe17743cc74..2695dff1378a 100644
@@ -192,6 +192,8 @@ struct nfsd_net {
 
 	atomic_t		nfs4_client_count;
 	int			nfs4_max_clients;
+
+	atomic_t		nfsd_courtesy_client_count;
 };
 
 /* Simple check to find out if a given net was properly initialized */
fs/nfsd/nfs4state.c +0 -0
index c5d199d7e6b4..9675b5d8f408 100644

fs/nfsd/nfs4state.c static __be32 get_client_locked(struct nfs4_client *clp) · line 173

Quoted 6 lines
fs/nfsd/nfs4state.c +2 -1
@@ -169,7 +169,8 @@ static __be32 get_client_locked(struct nfs4_client *clp)
 	if (is_client_expired(clp))
 		return nfserr_expired;
 	atomic_inc(&clp->cl_rpc_users);
-	clp->cl_state = NFSD4_ACTIVE;
+	if (xchg(&clp->cl_state, NFSD4_ACTIVE) != NFSD4_ACTIVE)
+		atomic_add_unless(&nn->nfsd_courtesy_client_count, -1, 0);

I'd prefer this:

	if (clp->cl_state != NFSD4_ACTIVE)
		atomic_add_unless(...
	clp->cl_state = NFSD4_ACTIVE;
Quoted 6 lines

It's more lines, but it's easier to read. It's may also be more efficient since you're protected by a spinlock anyway.

I think it's also less deceptive when reading the code. When I see xchg() operations being used like that, my immediate thought is that they must be needed for synchronization purposes. It's not in this case.

Agree: generally speaking, primitives like xchg() are used only when hardware-assisted synchronization is needed. These are expensive operations, to be used sparingly.

Let's replace these with plain vanilla "=", as Jeff suggested.

Likewise, I'm not totally convinced that nfsd_courtesy_client_count has to be an atomic.

Quoted 30 lines
Quoted 30 lines
fs/nfsd/nfs4state.c +6 -1
 	return nfs_ok;
 }
 
@@ -190,7 +191,8 @@ renew_client_locked(struct nfs4_client *clp)
 
 	list_move_tail(&clp->cl_lru, &nn->client_lru);
 	clp->cl_time = ktime_get_boottime_seconds();
-	clp->cl_state = NFSD4_ACTIVE;
+	if (xchg(&clp->cl_state, NFSD4_ACTIVE) != NFSD4_ACTIVE)
+		atomic_add_unless(&nn->nfsd_courtesy_client_count, -1, 0);
 }
 
 static void put_client_renew_locked(struct nfs4_client *clp)
@@ -2233,6 +2235,8 @@ __destroy_client(struct nfs4_client *clp)
 	if (clp->cl_cb_conn.cb_xprt)
 		svc_xprt_put(clp->cl_cb_conn.cb_xprt);
 	atomic_add_unless(&nn->nfs4_client_count, -1, 0);
+	if (clp->cl_state != NFSD4_ACTIVE)
+		atomic_add_unless(&nn->nfsd_courtesy_client_count, -1, 0);
 	free_client(clp);
 	wake_up_all(&expiry_wq);
 }
@@ -4356,6 +4360,8 @@ void nfsd4_init_leases_net(struct nfsd_net *nn)
 	max_clients = (u64)si.totalram * si.mem_unit / (1024 * 1024 * 1024);
 	max_clients *= NFS4_CLIENTS_PER_GB;
 	nn->nfs4_max_clients = max_t(int, max_clients, NFS4_CLIENTS_PER_GB);
+
+	atomic_set(&nn->nfsd_courtesy_client_count, 0);
 }
 
Quoted 15 lines
Quoted 12 lines
fs/nfsd/nfs4state.c +3 -1
 static void init_nfs4_replay(struct nfs4_replay *rp)
@@ -5879,7 +5885,9 @@ nfs4_get_client_reaplist(struct nfsd_net *nn, struct list_head *reaplist,
 		if (!state_expired(lt, clp->cl_time))
 			break;
 		if (!atomic_read(&clp->cl_rpc_users))
-			clp->cl_state = NFSD4_COURTESY;
+			if (xchg(&clp->cl_state, NFSD4_COURTESY) ==
+							NFSD4_ACTIVE)
+				atomic_inc(&nn->nfsd_courtesy_client_count);
 		if (!client_has_state(clp))
 			goto exp_client;
 		if (!nfs4_anylock_blockers(clp))
Signature · 1 line
Jeff Layton <[email protected]>

--
Chuck Lever

author reviewer
#4 lore raw

PATCH v3 2/2 NFSD: add shrinker to reap courtesy clients on low memory condition

In reply to Dai Ngo

Add the courtesy client shrinker to react to low memory condition triggered by the memory shrinker.

On the shrinker's count callback, we increment a callback counter and return the number of outstanding courtesy clients. When the laundromat runs, it checks if this counter is not zero and starts reaping old courtesy clients. The maximum number of clients to be reaped is limited to NFSD_CIENT_MAX_TRIM_PER_RUN (128). This limit is to prevent the laundromat from spending too much time reaping the clients and not processing other tasks in a timely manner.

The laundromat is rescheduled to run sooner if it detects low low memory condition and there are more clients to reap.

On the shrinker's scan callback, we return the number of clients That were reaped since the last scan callback. We can not reap the clients on the scan callback context since destroying the client might require call into the underlying filesystem or other subsystems which might allocate memory which can cause deadlock.

---
 fs/nfsd/netns.h     |  3 +++
 fs/nfsd/nfs4state.c | 52 ++++++++++++++++++++++++++++++++++++++++++++++++----
 fs/nfsd/nfsctl.c    |  6 ++++--
 fs/nfsd/nfsd.h      |  9 +++++++--
 4 files changed, 62 insertions(+), 8 deletions(-)
fs/nfsd/netns.h +3 -0
index 2695dff1378a..2a604951623f 100644
@@ -194,6 +194,9 @@ struct nfsd_net {
 	int			nfs4_max_clients;
 
 	atomic_t		nfsd_courtesy_client_count;
+	atomic_t		nfsd_client_shrinker_cb_count;
+	atomic_t		nfsd_client_shrinker_reapcount;
+	struct shrinker		nfsd_client_shrinker;
 };
 
 /* Simple check to find out if a given net was properly initialized */
fs/nfsd/nfs4state.c +48 -4

85 lines not on this page · raw message

fs/nfsd/nfsctl.c +4 -2
index 917fa1892fd2..597a26ad4183 100644
@@ -1481,11 +1481,12 @@ static __net_init int nfsd_init_net(struct net *net)
 		goto out_idmap_error;
 	nn->nfsd_versions = NULL;
 	nn->nfsd4_minorversions = NULL;
+	retval = nfsd4_init_leases_net(nn);
+	if (retval)
+		goto out_drc_error;
 	retval = nfsd_reply_cache_init(nn);
 	if (retval)
 		goto out_drc_error;
-	nfsd4_init_leases_net(nn);
-
 	get_random_bytes(&nn->siphash_key, sizeof(nn->siphash_key));
 	seqlock_init(&nn->writeverf_lock);
 
@@ -1507,6 +1508,7 @@ static __net_exit void nfsd_exit_net(struct net *net)
 	nfsd_idmap_shutdown(net);
 	nfsd_export_shutdown(net);
 	nfsd_netns_free_versions(net_generic(net, nfsd_net_id));
+	nfsd4_leases_net_shutdown(nn);
 }
 
 static struct pernet_operations nfsd_net_ops = {
fs/nfsd/nfsd.h +7 -2
index 57a468ed85c3..7e05ab7a3532 100644
@@ -498,7 +498,11 @@ extern void unregister_cld_notifier(void);
 extern void nfsd4_ssc_init_umount_work(struct nfsd_net *nn);
 #endif
 
-extern void nfsd4_init_leases_net(struct nfsd_net *nn);
+extern int nfsd4_init_leases_net(struct nfsd_net *nn);
+static inline void nfsd4_leases_net_shutdown(struct nfsd_net *nn)
+{
+	unregister_shrinker(&nn->nfsd_client_shrinker);
+};
 
 #else /* CONFIG_NFSD_V4 */
 static inline int nfsd4_is_junction(struct dentry *dentry)
@@ -506,7 +510,8 @@ static inline int nfsd4_is_junction(struct dentry *dentry)
 	return 0;
 }
 
-static inline void nfsd4_init_leases_net(struct nfsd_net *nn) {};
+static inline int nfsd4_init_leases_net(struct nfsd_net *nn) { return 0; };
+static inline void nfsd4_leases_net_shutdown(struct nfsd_net *nn) { };
 
 #define register_cld_notifier() 0
 #define unregister_cld_notifier() do { } while(0)
Signature · 1 line
2.9.5