Skip to contentSkip to conversation

devlink: keep the instance mutex alive until references are gone

PATCH Applied Jakub Kicinski 4 participants 5 messages Jan 11, 2023 – Jan 12, 2023

Subsystems (3)
lore
Share & tools
lore

Jakub Kicinski

author maintainer
#0 lore raw

PATCH 1/1 devlink: keep the instance mutex alive until references are gone

The reference needs to keep the instance memory around, but also the instance lock must remain valid. Users will take the lock, check registration status and release the lock. mutex_destroy() etc. belong in the same place as the freeing of the memory.

Unfortunately lockdep_unregister_key() sleeps so we need to switch the an rcu_work.

Note that the problem is a bit hard to repro, because devlink_pernet_pre_exit() iterates over registered instances. AFAIU the instances must get devlink_free()d concurrently with the namespace getting deleted for the problem to occur.

Reported-by [email protected]
Fixes 9053637e0da7 ("devlink: remove the registration guarantee of references")

---
Jiri, this will likely conflict with your series, sorry :(

---
 net/devlink/core.c          | 16 +++++++++++++---
 net/devlink/devl_internal.h |  3 ++-
 2 files changed, 15 insertions(+), 4 deletions(-)
net/devlink/core.c +13 -3
index a31a317626d7..60beca2df7cc 100644
@@ -83,10 +83,21 @@ struct devlink *__must_check devlink_try_get(struct devlink *devlink)
 	return NULL;
 }
 
+static void devlink_release(struct work_struct *work)
+{
+	struct devlink *devlink;
+
+	devlink = container_of(to_rcu_work(work), struct devlink, rwork);
+
+	mutex_destroy(&devlink->lock);
+	lockdep_unregister_key(&devlink->lock_key);
+	kfree(devlink);
+}
+
 void devlink_put(struct devlink *devlink)
 {
 	if (refcount_dec_and_test(&devlink->refcount))
-		kfree_rcu(devlink, rcu);
+		queue_rcu_work(system_wq, &devlink->rwork);
 }
 
 struct devlink *devlinks_xa_find_get(struct net *net, unsigned long *indexp)
@@ -231,6 +242,7 @@ struct devlink *devlink_alloc_ns(const struct devlink_ops *ops,
 	INIT_LIST_HEAD(&devlink->trap_list);
 	INIT_LIST_HEAD(&devlink->trap_group_list);
 	INIT_LIST_HEAD(&devlink->trap_policer_list);
+	INIT_RCU_WORK(&devlink->rwork, devlink_release);
 	lockdep_register_key(&devlink->lock_key);
 	mutex_init(&devlink->lock);
 	lockdep_set_class(&devlink->lock, &devlink->lock_key);
@@ -259,8 +271,6 @@ void devlink_free(struct devlink *devlink)
 
 	mutex_destroy(&devlink->linecards_lock);
 	mutex_destroy(&devlink->reporters_lock);
-	mutex_destroy(&devlink->lock);
-	lockdep_unregister_key(&devlink->lock_key);
 	WARN_ON(!list_empty(&devlink->trap_policer_list));
 	WARN_ON(!list_empty(&devlink->trap_group_list));
 	WARN_ON(!list_empty(&devlink->trap_list));
net/devlink/devl_internal.h +2 -1
index 5d2bbe295659..e724e4c2a4ff 100644
@@ -7,6 +7,7 @@
 #include <linux/netdevice.h>
 #include <linux/notifier.h>
 #include <linux/types.h>
+#include <linux/workqueue.h>
 #include <linux/xarray.h>
 #include <net/devlink.h>
 #include <net/net_namespace.h>
@@ -51,7 +52,7 @@ struct devlink {
 	struct lock_class_key lock_key;
 	u8 reload_failed:1;
 	refcount_t refcount;
-	struct rcu_head rcu;
+	struct rcu_work rwork;
 	struct notifier_block netdevice_nb;
 	char priv[] __aligned(NETDEV_ALIGN);
 };
Signature · 1 line
2.38.1
maintainer Reviewed-by on v1 1/1
#1 lore raw
Quoted 8 lines · Wed, Jan 11, 2023 at 05:29:08AM CET, [email protected] wrote:

The reference needs to keep the instance memory around, but also the instance lock must remain valid. Users will take the lock, check registration status and release the lock. mutex_destroy() etc. belong in the same place as the freeing of the memory.

Unfortunately lockdep_unregister_key() sleeps so we need to switch the an rcu_work.

Quoted 8 lines

Note that the problem is a bit hard to repro, because devlink_pernet_pre_exit() iterates over registered instances. AFAIU the instances must get devlink_free()d concurrently with the namespace getting deleted for the problem to occur.

Reported-by [email protected]
Fixes 9053637e0da7 ("devlink: remove the registration guarantee of references")
Quoted 2 lines

---
Jiri, this will likely conflict with your series, sorry :(

Yeah, I will rebase and send v5 after this is applied, no worries.

maintainer
#2 lore raw
In reply to Jakub Kicinski
Quoted 8 lines · Wed, Jan 11, 2023 at 05:29:08AM CET, [email protected] wrote:

The reference needs to keep the instance memory around, but also the instance lock must remain valid. Users will take the lock, check registration status and release the lock. mutex_destroy() etc. belong in the same place as the freeing of the memory.

Unfortunately lockdep_unregister_key() sleeps so we need to switch the an rcu_work.

Quoted 8 lines

Note that the problem is a bit hard to repro, because devlink_pernet_pre_exit() iterates over registered instances. AFAIU the instances must get devlink_free()d concurrently with the namespace getting deleted for the problem to occur.

Reported-by [email protected]
Fixes 9053637e0da7 ("devlink: remove the registration guarantee of references")
Reported-by [email protected]
Reviewed-by on v1 1/1
#3 lore raw
In reply to Jakub Kicinski
Quoted 8 lines · Jakub Kicinski

The reference needs to keep the instance memory around, but also the instance lock must remain valid. Users will take the lock, check registration status and release the lock. mutex_destroy() etc. belong in the same place as the freeing of the memory.

Unfortunately lockdep_unregister_key() sleeps so we need to switch the an rcu_work.

Quoted 8 lines

Note that the problem is a bit hard to repro, because devlink_pernet_pre_exit() iterates over registered instances. AFAIU the instances must get devlink_free()d concurrently with the namespace getting deleted for the problem to occur.

Reported-by [email protected]
Fixes 9053637e0da7 ("devlink: remove the registration guarantee of references")

Makes sense to me.

Quoted 23 lines

---
Jiri, this will likely conflict with your series, sorry :(

---
 net/devlink/core.c          | 16 +++++++++++++---
 net/devlink/devl_internal.h |  3 ++-
 2 files changed, 15 insertions(+), 4 deletions(-)
net/devlink/core.c +8 -0
index a31a317626d7..60beca2df7cc 100644
@@ -83,10 +83,21 @@ struct devlink *__must_check devlink_try_get(struct devlink *devlink)
 	return NULL;
 }
 
+static void devlink_release(struct work_struct *work)
+{
+	struct devlink *devlink;
+
+	devlink = container_of(to_rcu_work(work), struct devlink, rwork);
+
+	mutex_destroy(&devlink->lock);
+	lockdep_unregister_key(&devlink->lock_key);

net/devlink/core.c struct devlink *__must_check devlink_try_get(struct devlink *devlink) · line 102

Quoted 10 lines
net/devlink/core.c +4 -1
+	kfree(devlink);
+}
+
 void devlink_put(struct devlink *devlink)
 {
 	if (refcount_dec_and_test(&devlink->refcount))
-		kfree_rcu(devlink, rcu);
+		queue_rcu_work(system_wq, &devlink->rwork);
 }
 

You can't directly call devlink_release because callers of devlink_put shouldn't sleep. Ok so instead we queue RCU work to do it later. Makes sense. I was thinking if we'd used kref instead of raw refcount we could just kref_put, except that just directly calls the release function and we'd have to queue_rcu_work anyways.

Ok.

Quoted 10 lines
net/devlink/core.c +1 -0
 struct devlink *devlinks_xa_find_get(struct net *net, unsigned long *indexp)
@@ -231,6 +242,7 @@ struct devlink *devlink_alloc_ns(const struct devlink_ops *ops,
 	INIT_LIST_HEAD(&devlink->trap_list);
 	INIT_LIST_HEAD(&devlink->trap_group_list);
 	INIT_LIST_HEAD(&devlink->trap_policer_list);
+	INIT_RCU_WORK(&devlink->rwork, devlink_release);
 	lockdep_register_key(&devlink->lock_key);
 	mutex_init(&devlink->lock);
 	lockdep_set_class(&devlink->lock, &devlink->lock_key);

net/devlink/core.c void devlink_free(struct devlink *devlink) · line 263

Quoted 5 lines
net/devlink/core.c +0 -2
@@ -259,8 +271,6 @@ void devlink_free(struct devlink *devlink)
 
 	mutex_destroy(&devlink->linecards_lock);
 	mutex_destroy(&devlink->reporters_lock);
-	mutex_destroy(&devlink->lock);
-	lockdep_unregister_key(&devlink->lock_key);

It seems like we probably would want to move linecards_lock and reporters_lock too, except we know that these will be removed soon anyways. Ok.

Quoted 24 lines
net/devlink/core.c +0 -0
 	WARN_ON(!list_empty(&devlink->trap_policer_list));
 	WARN_ON(!list_empty(&devlink->trap_group_list));
 	WARN_ON(!list_empty(&devlink->trap_list));
net/devlink/devl_internal.h +2 -1
index 5d2bbe295659..e724e4c2a4ff 100644
@@ -7,6 +7,7 @@
 #include <linux/netdevice.h>
 #include <linux/notifier.h>
 #include <linux/types.h>
+#include <linux/workqueue.h>
 #include <linux/xarray.h>
 #include <net/devlink.h>
 #include <net/net_namespace.h>
@@ -51,7 +52,7 @@ struct devlink {
 	struct lock_class_key lock_key;
 	u8 reload_failed:1;
 	refcount_t refcount;
-	struct rcu_head rcu;
+	struct rcu_work rwork;
 	struct notifier_block netdevice_nb;
 	char priv[] __aligned(NETDEV_ALIGN);
 };
#4 lore raw
In reply to Jakub Kicinski
This patch was applied to netdev/net-next.git (master) by Jakub Kicinski <[email protected]>:

Hello:

This patch was applied to netdev/net-next.git (master) by Jakub Kicinski <[email protected]>:

Quoted 9 lines · On Tue, 10 Jan 2023 20:29:08 -0800 you wrote:

The reference needs to keep the instance memory around, but also the instance lock must remain valid. Users will take the lock, check registration status and release the lock. mutex_destroy() etc. belong in the same place as the freeing of the memory.

Unfortunately lockdep_unregister_key() sleeps so we need to switch the an rcu_work.

[...]

Here is the summary with links:
- [net-next] devlink: keep the instance mutex alive until references are gone

    https://git.kernel.org/netdev/net-next/c/93e71edfd90c

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html