Re: Bug: mv643xxx fails with highmem

[Ezequiel Garcia <ezequiel.garcia@free-electrons.com>]

Russell, David:

On 12/11/2014 05:25 PM, Russell King - ARM Linux wrote:
> On Thu, Dec 11, 2014 at 03:10:55PM -0500, David Miller wrote:
>> From: Russell King - ARM Linux <linux@arm.linux.org.uk>
>> Date: Thu, 11 Dec 2014 19:49:20 +0000
>>
>>> Commit 69ad0dd7af22 removed skb_frag_dma_map() in favour of mapping
>>> all fragments with dma_map_single().  This fails when the driver is
>>> used in an environment with highmem.
>>
>> This change looks really buggy to me.
>>
>> Unfortunately, all the changes he subsequently makes for software TSO
>> support depend upon this :-/
>>
>> The change is definitely wrong.

I've been trying to find a fix for this issue, and also trying to
reproduce the bug.

As for the fix, we need to fix the non-TSO and TSO paths independently.
The former is fairly straightforward, but the latter might be a bit more
involved.

The problem is that the tso_t struct holds a pointer to the skb linear
and non-linear data.

struct tso_t {
        int next_frag_idx;
        void *data;
        size_t size;
        u16 ip_id;
        u32 tcp_seq;
};

Instead, we should deal with pages, and only map the non-linear skb with
skb_frag_dma_map().

On the other side, I haven't been able to reproduce this on my boards. I
did try to put a hack to hold most lowmem pages, but it didn't make a
difference. (In fact, I haven't been able to clearly see how the pages
for the skbuff are allocated from high memory.)

Russell, would you share any hints about your setup? I don't have access
to any Dove boards at the moment, but I do have Kirkwoods, Armadas and
i.MX6.

Thanks a lot for your report and help!
-- 
Ezequiel García, Free Electrons
Embedded Linux, Kernel and Android Engineering
http://free-electrons.com