Merge branch 'xdp-page_pool-fixes-and-in-flight-accounting'

Jesper Dangaard Brouer says:

====================
xdp: page_pool fixes and in-flight accounting

This patchset fix page_pool API and users, such that drivers can use it for
DMA-mapping. A number of places exist, where the DMA-mapping would not get
released/unmapped, all these are fixed. This occurs e.g. when an xdp_frame
gets converted to an SKB. As network stack doesn't have any callback for XDP
memory models.

The patchset also address a shutdown race-condition. Today removing a XDP
memory model, based on page_pool, is only delayed one RCU grace period. This
isn't enough as redirected xdp_frames can still be in-flight on different
queues (remote driver TX, cpumap or veth).

We stress that when drivers use page_pool for DMA-mapping, then they MUST
use one packet per page. This might change in the future, but more work lies
ahead, before we can lift this restriction.

This patchset change the page_pool API to be more strict, as in-flight page
accounting is added.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 3 insertions, 0 deletions

diff --git a/kernel/bpf/cpumap.c b/kernel/bpf/cpumap.c
index 8ee5532cf6a6..8dff08768087 100644
--- a/kernel/bpf/cpumap.c
+++ b/kernel/bpf/cpumap.c
@@ -208,6 +208,9 @@ static struct sk_buff *cpu_map_build_skb(struct bpf_cpu_map_entry *rcpu,
 	 * - RX ring dev queue index	(skb_record_rx_queue)
 	 */
 
+	/* Until page_pool get SKB return path, release DMA here */
+	xdp_release_frame(xdpf);
+
 	/* Allow SKB to reuse area used by xdp_frame */
 	xdp_scrub_frame(xdpf);