On Fri, May 01, 2020 at 04:20:20AM +0100, Al Viro wrote:
> On Fri, May 01, 2020 at 03:37:34AM +0100, Al Viro wrote:
> > On Thu, Apr 30, 2020 at 01:38:44PM -0700, ira.weiny@xxxxxxxxx wrote:
> >
> > > -static inline void *kmap_atomic(struct page *page)
> > > +static inline void *kmap_atomic_prot(struct page *page, pgprot_t prot)
> > > {
> > > preempt_disable();
> > > pagefault_disable();
> > > if (!PageHighMem(page))
> > > return page_address(page);
> > > - return kmap_atomic_high(page);
> > > + return kmap_atomic_high_prot(page, prot);
> > > }
> > > +#define kmap_atomic(page) kmap_atomic_prot(page, kmap_prot)
> >
> > OK, so it *was* just a bisect hazard - you return to original semantics
> > wrt preempt_disable()...
>
> FWIW, how about doing the following: just before #5/10 have a patch
> that would touch only microblaze, ppc and x86 splitting their
> kmap_atomic_prot() into an inline helper + kmap_atomic_high_prot().
> Then your #5 would leave their kmap_atomic_prot() as-is (it would
> use kmap_atomic_prot_high() instead). The rest of the series plays
> out pretty much the same way it does now, and wrappers on those
> 3 architectures would go away when an identical generic one is
> introduced in this commit (#9/10).
>
> AFAICS, that would avoid the bisect hazard and might even end
> up with less noise in the patches...
This works. V2 coming out shortly.
Thanks for catching this,
Ira
