On 7/15/25 02:47, Jakub Kicinski wrote:
On Mon, 14 Jul 2025 22:58:31 +0900 Harry Yoo wrote:
Could you please share your thoughts on why it's hard to judge them and
what's missing from the series, such as in the comments, changelog, or
the cover letter?
My main concern (as shared on earlier revisions) is the type hierarchy
Quick overview. struct netmem_desc is a common denominator b/w pp pages
and net_iov, and contains fields used by the pp generic path, e.g.
refcount, dma_addr. Before, pp was using type casting hacks to keep
code generic with some overhead on bit masking, now it'll be able to
look up netmem_desc from a netmem and use it directly.
That's pretty much what I was suggesting niov / page aliasing to be
1+ years ago, but unfortunately that didn't happen. It definitely
removes some type casting hackiness.
exposed to the drivers.
v10 adds a bunch of "pp_page_to_nmdesc(page)->pp" in the drivers,
Not sure I have a strong opinion, but it can be turned into a helper.
Converting things back and forth or blindly
downcasting to netmem and upcasting back to the CPU-readable type is
no good.
--
Pavel Begunkov
