On 29/05/14 20:04, Eric Dumazet wrote:

On Thu, 2014-05-29 at 18:12 +0100, Jim Baxter wrote:

quoted

This fixes a problem with dropped packets over 16k CDC-NCM
when the connection is being heavily used.

The issue was that the skb truesize for the unpacked NCM
packets was too high after they were cloned from the 16k
skb, this lead to the potential memory calculated by the
Kernel running out of memory earlier then it should.

Signed-off-by: Jim Baxter <redacted>
---

Note the patch is OK, but changelog a bit misleading ;)

Kernel was not running out of memory, because truesize was correct.

The problem here is that a frame was consuming more kernel memory than
really needed, so chances of hitting socket sk_rcvbuf limit was high.

BTW :
#define NTB_OUT_SIZE              16384

alloc_skb(size) ->
	kmalloc(16384 + sizeof(struct skb_shared_info)) ->
		roundup() => 32768

So truesize of the skb was infact ~32KB, which is really insane indeed.
After your patch, its back to ~2KB

Acked-by: Eric Dumazet <edumazet@google.com>

Thank you, I will clarify the description.

Jim