On Tue, 2022-11-15 at 12:47 +0100, Peter Zijlstra wrote: > On Fri, Nov 04, 2022 at 03:35:42PM -0700, Rick Edgecombe wrote: > > @@ -1331,6 +1345,18 @@ void do_user_addr_fault(struct pt_regs > > *regs, > > > > perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address); > > > > + /* > > + * To service shadow stack read faults, unlike normal read > > faults, the > > + * fault handler needs to create a type of memory that will > > also be > > + * writable (with instructions that generate shadow stack > > writes). > > + * In the case of COW memory, the COW needs to take place > > even with > > + * a shadow stack read. Otherwise the shared page will be > > left (shadow > > + * stack) writable in userspace. So to trigger the > > appropriate behavior > > + * by setting FAULT_FLAG_WRITE for shadow stack accesses, > > even if the > > + * access was a shadow stack read. > > + */ > > Clear as mud... So SS pages are 'Write=0,Dirty=1', which, per > construction, lack a RW bit. And these pages are writable (WRUSS). > > pte_wrprotect() seems to do: _PAGE_DIRTY->_PAGE_COW (which is really > weird in this situation), resulting in: 'Write=0,Dirty=0,Cow=1'. > > That's regular RO memory and won't raise read-faults. > > But I'm thinking RET will trip #PF here when it tries to read the SS > because the SSP is not a proper shadow stack page? > > And in that case you want to tickle pte_mkwrite() to undo the > pte_wrprotect() above? > > So while the #PF is a 'read' fault due to RET not actually writing to > the shadow stack, you want to force a write fault so it will re- > instate > the SS page. > > Did I get that right? That's right. I think the assumption that needs to be broken in the readers head is that you can satisfy a read fault with read-only PTE. This is kind of baked in all over the place with the zero-pfn, COW, etc. Maybe I should try to start with that.
