Re: [PATCH] arm64: armv8_deprecated: Fix undef_hook mask for thumb setend

From: Fredrik Strupe
Date: Tue Apr 07 2020 - 07:18:27 EST

On 07.04.2020 12:47, Suzuki K Poulose wrote:

On 04/07/2020 10:27 AM, Catalin Marinas wrote:
On Mon, Apr 06, 2020 at 04:16:05PM +0200, Fredrik Strupe wrote:
Use a full 32-bit mask to prevent accidental matchings of thumb32
instructions where the second half-word is equal to the thumb16 setend
encoding.

This fixes the same problem as the following patch:

ÂÂÂÂ https://lkml.org/lkml/2020/3/16/341

This link is not guaranteed to be stable and the commit should have the
full description rather than referring to another email.

but for setend emulation instead.

Signed-off-by: Fredrik Strupe <fredrik@xxxxxxxxxx>

It also needs Fixes: and Cc: stable tags.

---
 arch/arm64/kernel/armv8_deprecated.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/arm64/kernel/armv8_deprecated.c b/arch/arm64/kernel/armv8_deprecated.c
index 9d3442d62..8c06dfee0 100644
--- a/arch/arm64/kernel/armv8_deprecated.c
+++ b/arch/arm64/kernel/armv8_deprecated.c
@@ -609,7 +609,7 @@ static struct undef_hook setend_hooks[] = {
ÂÂÂÂÂ },
ÂÂÂÂÂ {
ÂÂÂÂÂÂÂÂÂ /* Thumb mode */
-ÂÂÂÂÂÂÂ .instr_maskÂÂÂ = 0x0000fff7,
+ÂÂÂÂÂÂÂ .instr_maskÂÂÂ = 0xfffffff7,
ÂÂÂÂÂÂÂÂÂ .instr_valÂÂÂ = 0x0000b650,

I can see how this could happen but it would be useful to provide a
concrete example in the commit log.

The instruction opcode built by call_undef_hook() first reads a u16 as a
T16 instruction and the above should be fine. However, if this looks
like a T32 opcode, it reads a subsequent u16 which becomes the lowest
half-word and the above mask/val may inadvertently match it.


Thanks for the feedback. I have updated the patch with the requested changes.
We also do a check on the pstate_val, along with the instr_val to
confirm the mode. So this should be fine as it is ?

Suzuki
pstate_val only indicates thumb execution, not whether the current instruction
is a T16 or T32 instruction.

(Sorry for duplicate, something went terribly wrong with the formatting of the
previous email.)

Fredrik

---
For thumb instructions, call_undef_hook() in traps.c first reads a u16,
and if the u16 indicates a T32 instruction (u16 >= 0xe800), a second
u16 is read, which then makes up the the lower half-word of a T32
instruction. For T16 instructions, the second u16 is not read,
which makes the resulting u32 opcode always have the upper half set to
0.

However, having the upper half of instr_mask in the undef_hook set to 0
masks out the upper half of all thumb instructions - both T16 and T32.
This results in trapped T32 instructions with the lower half-word equal
to the T16 encoding of setend (b650) being matched, even though the upper
half-word is not 0000 and thus indicates a T32 opcode.

An example of such a T32 instruction is eaa0b650, which should raise a
SIGILL since T32 instructions with an eaa prefix are unallocated as per
Arm ARM, but instead works as a SETEND because the second half-word is set
to b650.

This patch fixes the issue by extending instr_mask to include the
upper u32 half, which will still match T16 instructions where the upper
half is 0, but not T32 instructions.

Signed-off-by: Fredrik Strupe <fredrik@xxxxxxxxxx>
Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
Cc: Will Deacon <will.deacon@xxxxxxx>
Fixes: 2d888f48e056 ("arm64: Emulate SETEND for AArch32 tasks")
---
Âarch/arm64/kernel/armv8_deprecated.c | 2 +-
Â1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/arm64/kernel/armv8_deprecated.c b/arch/arm64/kernel/armv8_deprecated.c
index 9d3442d62..8c06dfee0 100644
--- a/arch/arm64/kernel/armv8_deprecated.c
+++ b/arch/arm64/kernel/armv8_deprecated.c
@@ -609,7 +609,7 @@ static struct undef_hook setend_hooks[] = {
ÂÂÂÂ },
ÂÂÂÂ {
ÂÂÂÂ ÂÂÂ /* Thumb mode */
-ÂÂÂ ÂÂÂ .instr_maskÂÂÂ = 0x0000fff7,
+ÂÂÂ ÂÂÂ .instr_maskÂÂÂ = 0xfffffff7,
ÂÂÂÂ ÂÂÂ .instr_valÂÂÂ = 0x0000b650,
ÂÂÂÂ ÂÂÂ .pstate_maskÂÂÂ = (PSR_AA32_T_BIT | PSR_AA32_MODE_MASK),
ÂÂÂÂ ÂÂÂ .pstate_valÂÂÂ = (PSR_AA32_T_BIT | PSR_AA32_MODE_USR),
--
2.20.1