Re: [net RFC] net/mlx4_en: Use frag stride in crossing page boundary condition

On Thu, 2018-06-14 at 20:53 +0000, Saeed Mahameed wrote:

On Thu, 2018-06-14 at 13:47 -0700, saeedm@mellanox.com wrote:

quoted

On Thu, 2018-06-14 at 12:12 -0700, Eric Dumazet wrote:

quoted

On 06/14/2018 11:56 AM, Saeed Mahameed wrote:

quoted

Interestingly for this exact frag_stride we don't have an issue
:)
since it goes through a different condition branch
(the page flipping thing):

if (frag_info->frag_stride == PAGE_SIZE / 2) {
    frags->page_offset ^= PAGE_SIZE / 2;
	release = page_count(page) != 1 ||
		  page_is_pfmemalloc(page) ||
		  page_to_nid(page) != numa_mem_id();

I guess you forgot to test on PowerPC where PAGE_SIZE=65536 ?

On PowerPC, the first branch is never taken.

Oh, sorry i see where you are going with this, you mean for xdp in
the
other branch we need to take care of PAGE_SIZE > 4k .. 

I was looking at the code without my fix :)

with my fix:
release = frags->page_offset + frag_info->frag_stride > PAGE_SIZE;

for XDP: frag_info->frag_stride is PAGE_SIZE, so release will always be
true regardless of PAGE_SIZE.

So i guess i didn't quite understand your PowerPC concern.. can you
elaborate ?

You are totally right, good catch! I will have to figure this out :)
..

quoted

This code is already in the driver, i guess as optimization for
when
a
page can be reused only twice (PAGE_SIZE=4K, strides_size = 2k).
frag_stride is never more than 2k.

for power PC we should always take the other branch which will
reuse
as
much 2k strides as possible in a page regardless of PAGE_SIZE.

so are we good with my change ?