On Mon, Dec 04, 2023 at 11:11:26AM +0000, Ryan Roberts wrote:
> To be honest, while I understand pte_cont() and friends, I don't understand
> their relevance (or at least potential future relevance) to GUP?

GUP in general can be smarter to recognize if a pte/pmd is a cont_pte and
fetch the whole pte/pmd range if the caller specified.  Now it loops over
each pte/pmd.

Fast-gup is better as it at least doesn't take pgtable lock, for cont_pte
it looks inside gup_pte_range() which is good enough, but it'll still do
folio checks for each sub-pte, even though the 2nd+ folio checks should be
mostly the same (if to ignore races when the folio changed within the time
of processing the cont_pte chunk).

Slow-gup (as of what this series is about so far) doesn't do that either,
for each cont_pte whole entry it'll loop N times, frequently taking and
releasing the pgtable lock.  A smarter slow-gup can fundamentallly setup
follow_page_context.page_mask if it sees a cont_pte.  There might be a
challenge on whether holding the head page's refcount would stablize the
whole folio, but that may be another question to ask.

I think I also overlooked that PPC_8XX also has cont_pte support, so we
actually have three users indeed, if not counting potential future archs
adding support to also get that same tlb benefit.


Peter Xu

Reply via email to