Re: [PATCH net-next v1 06/11] page_pool: Coordinate refill with memory providers

[Stanislav Fomichev]

On 11/10, Björn Töpel wrote:
> A UMEM has a fixed number of chunks. When a refill finds none, and the
> chunks come back later, nothing schedules the queue again: RX stalls.
> Kicks after each queue operation are easy to miss.
> 
> Drivers report the end of each refill with page_pool_refill_done().
> The provider then asks for another poll, or wakes the queue later
> through the NAPI that the core publishes while the queue runs. Socket
> wakeups use that NAPI too, so such drivers need no ndo_xsk_wakeup.
> 
> With XDP_USE_NEED_WAKEUP, the FILL flag stays set: a flag set only
> when starved can miss an entry posted at the same time.
> 
> Signed-off-by: Björn Töpel <bjorn@xxxxxxxxxx>
> ---
>  Documentation/networking/netmem.rst     | 10 +++
>  include/net/netdev_rx_queue.h           |  6 ++
>  include/net/page_pool/helpers.h         | 24 ++++++
>  include/net/page_pool/memory_provider.h |  2 +
>  include/net/page_pool/types.h           |  6 ++
>  net/core/dev.c                          |  4 +
>  net/core/dev.h                          |  4 +
>  net/core/netdev_rx_queue.c              | 97 ++++++++++++++++++++++++-
>  net/xdp/xsk.c                           |  4 +
>  net/xdp/xsk.h                           |  1 +
>  net/xdp/xsk_pp.c                        | 73 ++++++++++++++++++-
>  11 files changed, 226 insertions(+), 5 deletions(-)
> 
> diff --git a/Documentation/networking/netmem.rst b/Documentation/networking/netmem.rst
> index 39cb75339704..74c73d7d126e 100644
> --- a/Documentation/networking/netmem.rst
> +++ b/Documentation/networking/netmem.rst
> @@ -136,6 +136,16 @@ Driver RX Requirements
>     circulation times, such as when userspace holds a reference in zerocopy
>     scenarios.
>  
> +8. After each RX ring refill, call page_pool_refill_done() once per page
> +   pool, with PP_REFILL_FULL only if all rings that use the pool are full.
> +   Rings of a queue that take buffers from one such provider share one
> +   page pool. Keep buffers in each of them, first in a ring that every
> +   packet needs, like a header ring, so that a provider with few buffers
> +   does not leave them all in one ring. If the call returns false, poll
> +   again soon, for example by returning the full NAPI budget. Register the
> +   NAPI of each RX queue with netif_queue_set_napi(); the core schedules
> +   it when provider buffers become available.
> +
>  Driver TX Requirements
>  ======================
>  
> diff --git a/include/net/netdev_rx_queue.h b/include/net/netdev_rx_queue.h
> index 989120cbf452..ca12cbf7b825 100644
> --- a/include/net/netdev_rx_queue.h
> +++ b/include/net/netdev_rx_queue.h
> @@ -33,6 +33,12 @@ struct netdev_rx_queue {
>  	struct pp_memory_provider_params mp_params;
>  	/* stamps the page pools of each queue allocation */
>  	u32				mp_gen;
> +	/* NAPI of the running queue, for memory provider wakeups. Set while
> +	 * mp_armed, and if the provider has a refill_done op. Written under
> +	 * the ops lock, read under RCU, see netif_mp_wake().
> +	 */
> +	struct napi_struct __rcu	*mp_napi;
> +	bool				mp_armed;

The comment should explain why mp_napi != napi? (I do not understand why)