Re: [PATCH v17 08/15] arm64: kexec: configure EL2 vectors for kexec

[Will Deacon]

On Wed, Sep 29, 2021 at 11:54:55PM -0400, Pasha Tatashin wrote:
> > > +/* Allocates pages for kexec page table */
> > > +static void *kexec_page_alloc(void *arg)
> > > +{
> > > +     struct kimage *kimage = (struct kimage *)arg;
> > > +     struct page *page = kimage_alloc_control_pages(kimage, 0);
> > > +
> > > +     if (!page)
> > > +             return NULL;
> > > +
> > > +     memset(page_address(page), 0, PAGE_SIZE);
> >
> > Hmm, I think we might be missing barriers here to ensure that the zeroes
> > are visible to the page-table walker before we plumb the page into the
> > page-table.
> >
> > Usually, that's taken care of by the smp_wmb() in __pXX_alloc() but I
> > can't see that here. Is it hiding?
> 
> Based on the comment in __pte_alloc() that smp_wmb() is needed in
> order to synchronize pte setup with other cpus prior to making it
> visible to them. This is not needed here. First, by the time these
> page tables are used the other cpus are offlined (kexec reboot code is
> single threaded). Second, we never insert any entry into a page table
> that is actively used by any cpu.

I think the comment there is wrong, but the barrier is still necessary.
How else do you guarantee that the page-table walker reads the zeroes from
the memset?

Will