arm: alignment trap tweaks

Hi all,

Revisiting the alignment trap again thanks to an obscure corner case with NEON alignment hints - patch 1 is a repost from a while back as a ping, because modern userspaces really should just get a resounding "Don't do that!" if they manage to step outside the architecture. Patch 2 is the fix for cases where fixup does need to be turned back on for whatever reason.

The following testcase compiled with -marm illustrates the problem:
with fixup enabled, when the VLD1 instruction with the alignment hint set faults by performing a misaligned access, it fails to load the NEON registers as expected and eventually writes back nonsense to the base register, resulting in a misleading segfault on the next iteration.

--->8---

#include <stdint.h>
#include <stdio.h>

static uint8_t buffer[256], got[256];

int main() {
	uint8_t *base = (uint8_t *)((intptr_t)buffer + 0x7 & ~0x7);
	void *p = base, *q = got;
	int i;

	for (i=0; i<64; i++)
		base[i] = i;
	for (i=0; i<4; i++) {
		asm volatile ("vld1.8 {d0,d1}, [%0]\n vst1.8 {d0,d1}, [%1]\n"
				: "=r"(p),"=r"(q) : "0"(p),"1"(q) : "d0","d1");
		printf("unaligned: %p [%d,%d,%d,...]\n", p++, got[0], got[1], got[2]);
	}
	p = base;
	for (i=0; i<4; i++) {
		asm volatile ("vld1.8 {d0,d1}, [%0:64]\n vst1.8 {d0,d1}, [%1]\n"
				: "=r"(p),"=r"(q) : "0"(p),"1"(q) : "d0","d1");
		printf("misaligned: %p [%d,%d,%d,...]\n", p++, got[0], got[1], got[2]);
	}
	return 0;
}

--->8---

Robin Murphy (2):
arm: SIGBUS on unsupported ARMv6 unaligned accesses
arm: don't break misaligned NEON load/store

 arch/arm/mm/alignment.c |    9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

--
1.7.9.5

This patch changes the default behaviour for userspace alignment faults on v6 from silent fixup to SIGBUS. This only affects code that violates the v6 unaligned access model - bad assembly/JIT code or high-level language code in violation of the relevant language spec - which should be corrected rather than unwittingly relying on performance-degrading fixups. Fixup behaviour can still be controlled from boot parameters or at runtime for compatibility with existing incorrect software.

There will be no impact for v5 and earlier.

---

This one is already in the patch system as 7944/1, since nobody objected to the original RFC or patch postings.
Russell: any news on whether you're happy to take this?

 arch/arm/mm/alignment.c |    6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

index 9240364..9a93315 100644
@@ -106,10 +106,10 @@ static int safe_usermode(int new_usermode, bool warn)
 	 * making any progress.
 	 */
 	if (cpu_is_v6_unaligned() && !(new_usermode & (UM_FIXUP | UM_SIGNAL))) {
-		new_usermode |= UM_FIXUP;
+		new_usermode |= UM_SIGNAL;
 
 		if (warn)
-			printk(KERN_WARNING "alignment: ignoring faults is unsafe on this CPU.  Defaulting to fixup mode.\n");
+			printk(KERN_WARNING "alignment: ignoring faults is unsafe on this CPU.  Defaulting to signal mode.\n");
 	}
 
 	return new_usermode;

@@ -971,7 +971,7 @@ static int __init alignment_init(void)
 		cr_alignment &= ~CR_A;
 		cr_no_alignment &= ~CR_A;
 		set_cr(cr_alignment);
-		ai_usermode = safe_usermode(ai_usermode, false);
+		ai_usermode = safe_usermode(ai_usermode | UM_WARN, false);
 	}
 #endif
 

1.7.9.5

The alignment fixup incorrectly decodes faulting ARM VLDn/VSTn instructions as LDR/STR, leading to register corruption. Detect these and correctly treat them as unhandled.

---
 arch/arm/mm/alignment.c |    3 +++
 1 file changed, 3 insertions(+)

index 9a93315..5812e0c 100644
@@ -40,6 +40,7 @@
  * This code is not portable to processors with late data abort handling.
  */
 #define CODING_BITS(i)	(i & 0x0e000000)
+#define COND_BITS(i)	(i & 0xf0000000)
 
 #define LDST_I_BIT(i)	(i & (1 << 26))		/* Immediate constant	*/
 #define LDST_P_BIT(i)	(i & (1 << 24))		/* Preindex		*/

@@ -817,6 +818,8 @@ do_alignment(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
 		break;
 
 	case 0x04000000:	/* ldr or str immediate */
+		if (COND_BITS(instr) == 0xf0000000) /* NEON VLDn, VSTn */
+			goto bad;
 		offset.un = OFFSET_BITS(instr);
 		handler = do_alignment_ldrstr;
 		break;

--
1.7.9.5

Ping. Any comment on these? Admittedly it's a very hard-to-hit bug, but it's a bug all the same.

Thanks,
Robin.