On Tue, Dec 24, 2024 at 10:51:19AM +0100, Martin Pieuchot wrote:
> When the kernel decides to move a page to the inactive list, with
> uvm_pagedeactivate(), it also remove its pmap references to ensure
> the next access will generate a fault.
> 
> Diff below moves the pmap_page_protect(9) call inside uvm_pagedeactivate().
> 
> This is a baby step towards reducing the contention on the `pageqlock'.
> 
> I'm aware that many more improvements are possible.  I'm trying to move
> very slowly to not introduce too many changes in behavior and expose
> existing bugs.  That's why I'm keeping the existing behavior of always
> calling pmap_page_protect(9) inside uvm_pagedeactivate().

Makes sense

> Note that currently, one call to uvm_pagedeactivate() is not coupled with
> a pmap_page_protect(), the diff below fixes that/makes it coherent.

Looks correct.

ok tb

You could remove the braces in uvm_fault.c and uvm_vnode.c as style(9)
incoherently requires and makes optional at the same time.