arm64: mm: kfence: only handle translation faults

Alexander noted that KFENCE only expects to handle faults from invalid page
table entries (i.e. translation faults), but arm64's fault handling logic will
call kfence_handle_page_fault() for other types of faults, including alignment
faults caused by unaligned atomics. This has the unfortunate property of
causing those other faults to be reported as "KFENCE: use-after-free",
which is misleading and hinders debugging.

Fix this by only forwarding unhandled translation faults to the KFENCE
code, similar to what x86 does already.

Alexander has verified that this passes all the tests in the KFENCE test
suite and avoids bogus reports on misaligned atomics.

Link: https://lore.kernel.org/all/20221102081620.1465154-1-zhongbaisong@huawei.com/
Fixes: 840b239 ("arm64, kfence: enable KFENCE for ARM64")
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Reviewed-by: Alexander Potapenko <glider@google.com>
Tested-by: Alexander Potapenko <glider@google.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Marco Elver <elver@google.com>
Cc: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20221114104411.2853040-1-mark.rutland@arm.com
Signed-off-by: Will Deacon <will@kernel.org>

arch/arm64/mm/fault.c

@@ -354,6 +354,11 @@ static bool is_el1_mte_sync_tag_check_fault(unsigned long esr)
 	return false;
 }
 
+static bool is_translation_fault(unsigned long esr)
+{
+	return (esr & ESR_ELx_FSC_TYPE) == ESR_ELx_FSC_FAULT;
+}
+
 static void __do_kernel_fault(unsigned long addr, unsigned long esr,
 			      struct pt_regs *regs)
 {
@@ -386,7 +391,8 @@ static void __do_kernel_fault(unsigned long addr, unsigned long esr,
 	} else if (addr < PAGE_SIZE) {
 		msg = "NULL pointer dereference";
 	} else {
-		if (kfence_handle_page_fault(addr, esr & ESR_ELx_WNR, regs))
+		if (is_translation_fault(esr) &&
+		    kfence_handle_page_fault(addr, esr & ESR_ELx_WNR, regs))
 			return;
 
 		msg = "paging request";