On Fri, Jun 14, 2019 at 11:34:09AM +0200, Peter Zijlstra wrote: > On Wed, May 08, 2019 at 05:43:33PM +0300, Kirill A. Shutemov wrote: > > > +/* Prepare page to be used for encryption. Called from page allocator. */ > > +void __prep_encrypted_page(struct page *page, int order, int keyid, bool zero) > > +{ > > + int i; > > + > > + /* > > + * The hardware/CPU does not enforce coherency between mappings > > + * of the same physical page with different KeyIDs or > > + * encryption keys. We are responsible for cache management. > > + */ > > On alloc we should flush the unencrypted (key=0) range, while on free > (below) we should flush the encrypted (key!=0) range. > > But I seem to have missed where page_address() does the right thing > here. As you've seen by now, it will be addressed later in the patchset. I'll update the changelog to indicate that page_address() handles KeyIDs correctly. > > + clflush_cache_range(page_address(page), PAGE_SIZE * (1UL << order)); > > + > > + for (i = 0; i < (1 << order); i++) { > > + /* All pages coming out of the allocator should have KeyID 0 */ > > + WARN_ON_ONCE(lookup_page_ext(page)->keyid); > > + lookup_page_ext(page)->keyid = keyid; > > + > > So presumably page_address() is affected by this keyid, and the below > clear_highpage() then accesses the 'right' location? Yes. clear_highpage() -> kmap_atomic() -> page_address(). > > + /* Clear the page after the KeyID is set. */ > > + if (zero) > > + clear_highpage(page); > > + > > + page++; > > + } > > +} > > + > > +/* > > + * Handles freeing of encrypted page. > > + * Called from page allocator on freeing encrypted page. > > + */ > > +void free_encrypted_page(struct page *page, int order) > > +{ > > + int i; > > + > > + /* > > + * The hardware/CPU does not enforce coherency between mappings > > + * of the same physical page with different KeyIDs or > > + * encryption keys. We are responsible for cache management. > > + */ > > I still don't like that comment much; yes the hardware doesn't do it, > and yes we have to do it, but it doesn't explain the actual scheme > employed to do so. Fair enough. I'll do better. > > + clflush_cache_range(page_address(page), PAGE_SIZE * (1UL << order)); > > + > > + for (i = 0; i < (1 << order); i++) { > > + /* Check if the page has reasonable KeyID */ > > + WARN_ON_ONCE(lookup_page_ext(page)->keyid > mktme_nr_keyids); > > It should also check keyid > 0, so maybe: > > (unsigned)(keyid - 1) > keyids-1 > > instead? Makes sense. -- Kirill A. Shutemov