arm64: Fix potential race with hardware DBM in ptep_set_access_flags()

In a system with DBM (dirty bit management) capable agents there is a
possible race between a CPU executing ptep_set_access_flags() (maybe
non-DBM capable) and a hardware update of the dirty state (clearing of
PTE_RDONLY). The scenario:

a) the pte is writable (PTE_WRITE set), clean (PTE_RDONLY set) and old
   (PTE_AF clear)
b) ptep_set_access_flags() is called as a result of a read access and it
   needs to set the pte to writable, clean and young (PTE_AF set)
c) a DBM-capable agent, as a result of a different write access, is
   marking the entry as young (setting PTE_AF) and dirty (clearing
   PTE_RDONLY)

The current ptep_set_access_flags() implementation would set the
PTE_RDONLY bit in the resulting value overriding the DBM update and
losing the dirty state.

This patch fixes such race by setting PTE_RDONLY to the most permissive
(lowest value) of the current entry and the new one.

Fixes: 66dbd6e6 ("arm64: Implement ptep_set_access_flags() for hardware AF/DBM")
Cc: Will Deacon <will.deacon@arm.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Acked-by: Steve Capper <steve.capper@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>

arch/arm64/mm/fault.c

@@ -163,26 +163,27 @@ int ptep_set_access_flags(struct vm_area_struct *vma,
 	/* only preserve the access flags and write permission */
 	pte_val(entry) &= PTE_AF | PTE_WRITE | PTE_DIRTY;
 
-	/*
-	 * PTE_RDONLY is cleared by default in the asm below, so set it in
-	 * back if necessary (read-only or clean PTE).
-	 */
+	/* set PTE_RDONLY if actual read-only or clean PTE */
 	if (!pte_write(entry) || !pte_sw_dirty(entry))
 		pte_val(entry) |= PTE_RDONLY;
 
 	/*
 	 * Setting the flags must be done atomically to avoid racing with the
-	 * hardware update of the access/dirty state.
+	 * hardware update of the access/dirty state. The PTE_RDONLY bit must
+	 * be set to the most permissive (lowest value) of *ptep and entry
+	 * (calculated as: a & b == ~(~a | ~b)).
 	 */
+	pte_val(entry) ^= PTE_RDONLY;
 	asm volatile("//	ptep_set_access_flags\n"
 	"	prfm	pstl1strm, %2\n"
 	"1:	ldxr	%0, %2\n"
-	"	and	%0, %0, %3		// clear PTE_RDONLY\n"
+	"	eor	%0, %0, %3		// negate PTE_RDONLY in *ptep\n"
 	"	orr	%0, %0, %4		// set flags\n"
+	"	eor	%0, %0, %3		// negate final PTE_RDONLY\n"
 	"	stxr	%w1, %0, %2\n"
 	"	cbnz	%w1, 1b\n"
 	: "=&r" (old_pteval), "=&r" (tmp), "+Q" (pte_val(*ptep))
-	: "L" (~PTE_RDONLY), "r" (pte_val(entry)));
+	: "L" (PTE_RDONLY), "r" (pte_val(entry)));
 
 	flush_tlb_fix_spurious_fault(vma, address);
 	return 1;