Re: [PATCH v3 net-next 0/7] net: ethernet: ti: cpsw: Add XDP support

From: Ivan Khoronzhuk
Date: Thu Jun 06 2019 - 09:28:34 EST

On Thu, Jun 06, 2019 at 10:08:50AM +0200, Jesper Dangaard Brouer wrote:
On Wed, 05 Jun 2019 12:14:50 -0700 (PDT)
David Miller <davem@xxxxxxxxxxxxx> wrote:

From: Ivan Khoronzhuk <ivan.khoronzhuk@xxxxxxxxxx>
Date: Wed, 5 Jun 2019 16:20:02 +0300

> This patchset adds XDP support for TI cpsw driver and base it on
> page_pool allocator. It was verified on af_xdp socket drop,
> af_xdp l2f, ebpf XDP_DROP, XDP_REDIRECT, XDP_PASS, XDP_TX.

Jesper et al., please give this a good once over.

The issue with merging this, is that I recently discovered two bug with
page_pool API, when using DMA-mappings, which result in missing
DMA-unmap's. These bugs are not "exposed" yet, but will get exposed
now with this drivers.

The two bugs are:

#1: in-flight packet-pages can still be on remote drivers TX queue,
while XDP RX driver manage to unregister the page_pool (waiting 1 RCU
period is not enough).

#2: this patchset also introduce page_pool_unmap_page(), which is
called before an XDP frame travel into networks stack (as no callback
exist, yet). But the CPUMAP redirect *also* needs to call this, else we
"leak"/miss DMA-unmap.

I do have a working prototype, that fixes these two bugs. I guess, I'm
under pressure to send this to the list soon...

In particular "cpsw" case no dma unmap issue and if no changes in page_pool
API then no changes to the driver required. page_pool_unmap_page() is
used here for consistency reasons with attention that it can be
inherited/reused by other SoCs for what it can be relevant.

One potential change as you mentioned is with dropping page_pool_destroy() that,
now, can look like:

@@ -571,7 +571,6 @@ static void cpsw_destroy_rx_pool(struct cpsw_priv *priv, int ch)

- page_pool_destroy(priv->page_pool[ch]);
priv->page_pool[ch] = NULL;

From what I know there is ongoing change for adding switchdev to cpsw that can
change a lot and can require more work to rebase / test this patchset, so I want
to believe it can be merged before this.

Ivan Khoronzhuk