Hi, On Mon, Dec 11, 2023 at 06:18:17PM +0000, Catalin Marinas wrote: > On Fri, Nov 24, 2023 at 04:34:57PM +0000, Joey Gouly wrote: > > @@ -497,6 +498,23 @@ static void do_bad_area(unsigned long far, unsigned long esr, > > #define VM_FAULT_BADMAP ((__force vm_fault_t)0x010000) > > #define VM_FAULT_BADACCESS ((__force vm_fault_t)0x020000) > > > > +static bool fault_from_pkey(unsigned long esr, struct vm_area_struct *vma, > > + unsigned int mm_flags) > > +{ > > + unsigned long iss2 = ESR_ELx_ISS2(esr); > > + > > + if (!arch_pkeys_enabled()) > > + return false; > > + > > + if (iss2 & ESR_ELx_Overlay) > > + return true; > > + > > + return !arch_vma_access_permitted(vma, > > + mm_flags & FAULT_FLAG_WRITE, > > + mm_flags & FAULT_FLAG_INSTRUCTION, > > + mm_flags & FAULT_FLAG_REMOTE); > > +} > > Do we actually need this additional arch_vma_access_permitted() check? > The ESR should tell us if it was a POE fault. Permission overlay faults > have priority over the base permission faults, so we'd not need to fall > back to this additional checks. Well, see below, we could do something > slightly smarter here. We want this here as it follows other arch's which will fail with a pkey fault even if the page isn't actually mapped. If the paged isn't mapped we'd get a translation fault, but since we know the type of access and have the pkey in the VMA we check it here. > > I can see x86 and powerpc have similar checks (though at a different > point under the mmap lock) but I'm not familiar with their exception > model, exception priorities. > > > + > > static vm_fault_t __do_page_fault(struct mm_struct *mm, > > struct vm_area_struct *vma, unsigned long addr, > > unsigned int mm_flags, unsigned long vm_flags, > > @@ -688,9 +706,29 @@ static int __kprobes do_page_fault(unsigned long far, unsigned long esr, > > * Something tried to access memory that isn't in our memory > > * map. > > */ > > - arm64_force_sig_fault(SIGSEGV, > > - fault == VM_FAULT_BADACCESS ? SEGV_ACCERR : SEGV_MAPERR, > > - far, inf->name); > > + int fault_kind; > > + /* > > + * The pkey value that we return to userspace can be different > > + * from the pkey that caused the fault. > > + * > > + * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4); > > + * 2. T1 : set AMR to deny access to pkey=4, touches, page > > + * 3. T1 : faults... > > + * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5); > > + * 5. T1 : enters fault handler, takes mmap_lock, etc... > > + * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really > > + * faulted on a pte with its pkey=4. > > + */ > > + int pkey = vma_pkey(vma); > > Other than the vma_pkey() race, I'm more worried about the vma > completely disappearing, e.g. munmap() in another thread. We end up > dereferencing a free pointer here. We need to do this under the mmap > lock. > > Since we need to do this check under the mmap lock, we could add an > additional check to see if the pkey fault we got was a racy and just > restart the instruction if it no longer faults according to > por_el0_allows_pkey(). However, the code below is too late in the fault > handling to be able to do much other than SIGSEGV. After discussing in person, I agree with the assesment that this is the wrong place to do the check, and after looking at the x86 arch code, I see how they're doing it earlier. Thanks, Joey