Re: [PATCH net-next v1 3/4] net: devmem: support leased queues on netdev unregister

[Stanislav Fomichev <sdf.kernel@gmail.com>]

On 11/03, Mina Almasry wrote:
> Previously, bind_dev was only populated when bind-tx was called on a
> virtual netdev (where map_dev != netdev), and leased RX bindings stored
> the virtual netdev in map_dev without tracking the physical backing
> netdev at all. Unregistering either the virtual device or the physical
> backing device while a leased RX or TX devmem binding was active failed
> to detach the dma-buf cleanly.
> 
> Always populate both device pointers on struct
> net_devmem_dmabuf_binding:
> - map_dev: the physical net_device whose dma_dev holds the DMA mapping
>   (resolved from leased RX queues in netdev_nl_get_dma_dev(), with its
>   netdev_lock held across net_devmem_bind_dmabuf()).
> - bind_dev: the net_device the user called bind-rx or bind-tx on (equal
>   to map_dev for direct bindings, or the virtual device for leased
>   bindings).
> 
> With both pointers tracked on all bindings, handle socket teardown,
> queue uninstall, and netdev unregistration in either order (bind_dev
> first or map_dev first):
> - Lock &binding->bind_dev during socket teardown in
>   netdev_nl_sock_priv_destroy(), close and erase bound_rxqs via
>   binding->bind_dev in net_devmem_unbind_dmabuf(), and require
>   binding->map_dev to remain non-NULL in net_devmem_get_binding().
> - Clear binding->bind_dev in mp_dmabuf_devmem_uninstall() once
>   bound_rxqs is empty, leaving unleased bindings mapped until the
>   binding is freed or map_dev is unregistered.
> - In net_devmem_dev_unregister(dev), if binding->bind_dev == dev, clear
>   NET_DEVMEM_BINDING_LIVE and clear binding->bind_dev if bound_rxqs is
>   empty or dev is not a queue leasee; if binding->map_dev == dev, clear
>   NET_DEVMEM_BINDING_LIVE, unmap the dmabuf, and clear binding->map_dev.
> - Reject new bindings and RX queue opens in net_devmem_bind_dmabuf() and
>   __netif_mp_open_rxq() if bind_dev or map_dev is no longer
>   NETREG_REGISTERED.
> 
> Signed-off-by: Mina Almasry <almasrymina@google.com>
> ---
>  net/core/devmem.c          | 85 +++++++++++++++++++++-----------------
>  net/core/devmem.h          |  8 ++--
>  net/core/netdev-genl.c     | 33 ++++++++++-----
>  net/core/netdev_rx_queue.c |  5 +++
>  4 files changed, 79 insertions(+), 52 deletions(-)
> 
> diff --git a/net/core/devmem.c b/net/core/devmem.c
> index 6446355e787d4..1d6b89f196e02 100644
> --- a/net/core/devmem.c
> +++ b/net/core/devmem.c
> @@ -86,7 +86,7 @@ net_devmem_dmabuf_binding_unmap(struct net_devmem_dmabuf_binding *binding)
>  
>  /**
>   * net_devmem_lock_dev() - lock a binding's net_device pointer if still alive
> - * @dev_ptr:		pointer to the net_device field on a devmem binding
> + * @dev_ptr:		pointer to &binding->map_dev or &binding->bind_dev
>   * @dev_tracker:	reftracker for the held net_device reference
>   *
>   * Safely dereference @dev_ptr, acquire netdev_lock() on the device, and
> @@ -95,17 +95,18 @@ net_devmem_dmabuf_binding_unmap(struct net_devmem_dmabuf_binding *binding)
>   * Invariant: *@dev_ptr is set once at bind time and is only ever cleared to
>   * NULL while holding netdev_lock() on that device (in
>   * mp_dmabuf_devmem_uninstall(), net_devmem_unbind_dmabuf(), or
> - * net_devmem_dev_unregister() during NETDEV_UNREGISTER before
> - * synchronize_net() and netdev_run_todo() free the net_device).
> + * net_devmem_dev_unregister(), all before synchronize_net() and
> + * netdev_run_todo() free the net_device).
>   *
>   * Taking rcu_read_lock() guarantees that a non-NULL device read from
>   * *@dev_ptr has not yet completed unregister_netdevice_many_notify()
>   * (which waits for an RCU grace period via synchronize_net() after
> - * NETDEV_UNREGISTER before freeing the net_device), so netdev_hold() can
> - * safely acquire a reference before sleeping on netdev_lock(). Once
> - * netdev_lock() is held, re-checking that READ_ONCE(*dev_ptr) == dev
> - * guarantees that the device has not yet been detached from the binding and
> - * cannot finish unregistering until net_devmem_unlock_dev().
> + * NETDEV_UNREGISTER and ndo_uninit before freeing the net_device), so
> + * netdev_hold() can safely acquire a reference before sleeping on
> + * netdev_lock(). Once netdev_lock() is held, re-checking that
> + * READ_ONCE(*dev_ptr) == dev guarantees that the device has not yet been
> + * detached from the binding and cannot finish unregistering until
> + * net_devmem_unlock_dev().
>   *
>   * Return: The locked and held net_device if *@dev_ptr is still non-NULL,
>   * or NULL if already cleared.
> @@ -219,22 +220,24 @@ void net_devmem_unbind_dmabuf(struct net_devmem_dmabuf_binding *binding)
>  	if (binding->list.next)
>  		list_del(&binding->list);
>  
> -	xa_for_each(&binding->bound_rxqs, xa_idx, rxq) {
> -		const struct pp_memory_provider_params mp_params = {
> -			.mp_priv	= binding,
> -			.mp_ops		= &dmabuf_devmem_ops,
> -		};
> +	if (binding->bind_dev) {
> +		xa_for_each(&binding->bound_rxqs, xa_idx, rxq) {
> +			const struct pp_memory_provider_params mp_params = {
> +				.mp_priv	= binding,
> +				.mp_ops		= &dmabuf_devmem_ops,
> +			};
>  
> -		rxq_idx = get_netdev_rx_queue_index(rxq);
> +			xa_erase(&binding->bound_rxqs, xa_idx);

What's the reason we start xa_erase'ing here?

> +			rxq_idx = get_netdev_rx_queue_index(rxq);
>  
> -		if (netif_mp_close_rxq(binding->map_dev, rxq_idx, &mp_params))
> -			busy = true;
> -	}
> +			if (netif_mp_close_rxq(binding->bind_dev, rxq_idx,
> +					       &mp_params))
> +				busy = true;
> +		}
>  
> -	if (binding->direction == DMA_FROM_DEVICE && binding->map_dev &&
> -	    !busy) {
> -		net_devmem_dmabuf_binding_unmap(binding);
> -		WRITE_ONCE(binding->map_dev, NULL);
> +		if (binding->direction == DMA_FROM_DEVICE && !busy)
> +			net_devmem_dmabuf_binding_unmap(binding);
> +		WRITE_ONCE(binding->bind_dev, NULL);
>  	}
>  
>  	percpu_ref_kill(&binding->ref);
> @@ -252,24 +255,20 @@ void net_devmem_dev_unregister(struct net_device *dev)
>  
>  	xa_lock(&net_devmem_dmabuf_bindings);
>  	xa_for_each(&net_devmem_dmabuf_bindings, xa_idx, binding) {
> -		if (binding->direction == DMA_FROM_DEVICE) {
> -			if (READ_ONCE(binding->map_dev) != dev ||
> -			    xa_get_mark(&net_devmem_dmabuf_bindings, xa_idx,
> -					NET_DEVMEM_BINDING_LIVE))
> -				continue;
> -		} else {
> -			if (READ_ONCE(binding->bind_dev) == dev)
> -				WRITE_ONCE(binding->bind_dev, NULL);
> -			if (READ_ONCE(binding->map_dev) != dev)
> -				continue;
> -			WRITE_ONCE(binding->bind_dev, NULL);
> +		if (READ_ONCE(binding->bind_dev) == dev) {
>  			__xa_clear_mark(&net_devmem_dmabuf_bindings, xa_idx,
>  					NET_DEVMEM_BINDING_LIVE);
> +			if (xa_empty(&binding->bound_rxqs) ||
> +			    !netif_is_queue_leasee(dev))
> +				WRITE_ONCE(binding->bind_dev, NULL);
>  		}
> +		if (READ_ONCE(binding->map_dev) != dev)
> +			continue;
>  
> +		__xa_clear_mark(&net_devmem_dmabuf_bindings, xa_idx,
> +				NET_DEVMEM_BINDING_LIVE);
>  		xa_unlock(&net_devmem_dmabuf_bindings);
> -		if (!netif_is_queue_leasee(dev))
> -			net_devmem_dmabuf_binding_unmap(binding);
> +		net_devmem_dmabuf_binding_unmap(binding);
>  		xa_lock(&net_devmem_dmabuf_bindings);
>  		WRITE_ONCE(binding->map_dev, NULL);
>  	}
> @@ -324,6 +323,15 @@ net_devmem_bind_dmabuf(struct net_device *map_dev,
>  	size_t i;
>  	long err;
>  
> +	netdev_assert_locked(bind_dev);
> +	netdev_assert_locked(map_dev);
> +

[..]

> +	if (bind_dev->reg_state != NETREG_REGISTERED ||
> +	    map_dev->reg_state != NETREG_REGISTERED) {
> +		NL_SET_ERR_MSG(extack, "Device is being unregistered");
> +		return ERR_PTR(-ENODEV);
> +	}

I saw Dragos asking about these, I'm confused as well on why we need them.

> +
>  	if (!dma_dev) {
>  		NL_SET_ERR_MSG(extack, "Device doesn't support DMA");
>  		return ERR_PTR(-EOPNOTSUPP);
> @@ -560,9 +568,9 @@ struct net_devmem_dmabuf_binding *net_devmem_get_binding(struct sock *sk,
>  	 * net_device.
>  	 */
>  	dst_dev = dst_dev_rcu(dst);
> -	if (unlikely(!dst_dev) ||
> -	    unlikely(dst_dev != READ_ONCE(binding->map_dev) &&
> -		     dst_dev != READ_ONCE(binding->bind_dev))) {
> +	if (unlikely(!dst_dev || !READ_ONCE(binding->map_dev) ||
> +		     (dst_dev != READ_ONCE(binding->map_dev) &&
> +		      dst_dev != READ_ONCE(binding->bind_dev)))) {

Now we READ_ONCE(map_dev) twice? Time to add a map_dev var and do it once?