[PATCH net] devlink: Fix devlink parallel commands processing

[PATCH net] devlink: Fix devlink parallel commands processing

[Shay Drory]

Commit 870c7ad4a52b ("devlink: protect devlink->dev by the instance
lock") added devlink instance locking inside a loop that iterates over
all the registered devlink instances on the machine in the pre-doit
phase. This can lead to serialization of devlink commands over
different devlink instances.

For example: While the first devlink instance is executing firmware
flash, all commands to other devlink instances on the machine are
forced to wait until the first devlink finishes.

Therefore, in the pre-doit phase, take the devlink instance lock only
for the devlink instance the command is targeting. Devlink layer is
taking a reference on the devlink instance, ensuring the devlink->dev
pointer is valid. This reference taking was introduced by commit
a380687200e0 ("devlink: take device reference for devlink object").
Without this commit, it would not be safe to access devlink->dev
lockless.

Fixes: 870c7ad4a52b ("devlink: protect devlink->dev by the instance lock")
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
---
 net/devlink/netlink.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

diff --git a/net/devlink/netlink.c b/net/devlink/netlink.c
index 499885c8b9ca..cffc7274de8c 100644
--- a/net/devlink/netlink.c
+++ b/net/devlink/netlink.c
@@ -193,15 +193,20 @@ devlink_get_from_attrs_lock(struct net *net, struct nlattr **attrs,
 	devname = nla_data(attrs[DEVLINK_ATTR_DEV_NAME]);
 
 	devlinks_xa_for_each_registered_get(net, index, devlink) {
-		devl_dev_lock(devlink, dev_lock);
-		if (devl_is_registered(devlink) &&
-		    strcmp(devlink->dev->bus->name, busname) == 0 &&
+		if (strcmp(devlink->dev->bus->name, busname) == 0 &&
 		    strcmp(dev_name(devlink->dev), devname) == 0)
-			return devlink;
-		devl_dev_unlock(devlink, dev_lock);
+			goto found;
 		devlink_put(devlink);
 	}
+	return ERR_PTR(-ENODEV);
+
+found:
+	devl_dev_lock(devlink, dev_lock);
+	if (devl_is_registered(devlink))
+		return devlink;
 
+	devl_dev_unlock(devlink, dev_lock);
+	devlink_put(devlink);
 	return ERR_PTR(-ENODEV);
 }
 
-- 
2.38.1

Re: [PATCH net] devlink: Fix devlink parallel commands processing

[Jakub Kicinski]

On Mon, 11 Mar 2024 10:57:26 +0200 Shay Drory wrote:
>  	devlinks_xa_for_each_registered_get(net, index, devlink) {
> -		devl_dev_lock(devlink, dev_lock);
> -		if (devl_is_registered(devlink) &&
> -		    strcmp(devlink->dev->bus->name, busname) == 0 &&
> +		if (strcmp(devlink->dev->bus->name, busname) == 0 &&
>  		    strcmp(dev_name(devlink->dev), devname) == 0)
> -			return devlink;
> -		devl_dev_unlock(devlink, dev_lock);
> +			goto found;

there's no need for a goto here:

		if (strcmp(devlink->dev->bus->name, busname) == 0 &&
		    strcmp(dev_name(devlink->dev), devname) == 0) {
			devl_dev_lock(devlink, dev_lock);
			if (devl_is_registered(devlink))
				return devlink;
			devl_dev_unlock(devlink, dev_lock);
		}

simpler, and also no change in behavior (in case some impossible
race happens and we have 2 devlinks with the same name, one already
unregistered and one registered).

>  		devlink_put(devlink);
>  	}
> +	return ERR_PTR(-ENODEV);
> +
> +found:
> +	devl_dev_lock(devlink, dev_lock);
> +	if (devl_is_registered(devlink))
> +		return devlink;
>  
> +	devl_dev_unlock(devlink, dev_lock);
> +	devlink_put(devlink);
>  	return ERR_PTR(-ENODEV);
-- 
pw-bot: cr