Re: [RFC PATCH] arm: Don't trap conditional UDF instructions

[Robin Murphy <robin.murphy@arm.com>]

On 2020-05-13 7:56 pm, Fredrik Strupe wrote:
> On 13.05.2020 20:12, Russell King - ARM Linux admin wrote:
>> On Wed, May 13, 2020 at 05:41:58PM +0200, Fredrik Strupe wrote:
>>> Hi,
>>>
>>> This is more of a question than a patch, but I hope the attached patch makes
>>> the issue a bit clearer.
>>>
>>> The arm port of Linux supports hooking/trapping of undefined instructions. Some
>>> parts of the code use this to trap UDF instructions with certain immediates in
>>> order to use them for other purposes, like 'UDF #16' which is equivalent to a
>>> BKPT instruction in A32.
>>>
>>> Moreover, most of the undef hooks on UDF instructions assume that UDF is
>>> conditional and mask out the condition prefix during matching. The attached
>>> patch shows the locations where this happens. However, the Arm architecture
>>> reference manual explicitly states that UDF is *not* conditional, making
>>> any instruction encoding with a condition prefix other than 0xe (always
>>> execute) unallocated.
>>
>> The latest version of the ARM architecture reference manual may say
>> that, but earlier versions say different things. The latest reference
>> manual does not apply to earlier architectures, so if you're writing
>> code to cover multiple different architectures, you must have an
>> understanding of each of those architectures.
>>
>> So, from the code:
>>
>> 	ARM:   xxxx 0111 1111 xxxx xxxx xxxx 1111 xxxx
>>
>>  From DDI0100E:
>>
>> 3.13.1 Undefined instruction space
>>         Instructions with the following opcodes are undefined
>>         instruction space:
>>
>>         opcode[27:25] = 0b011
>>         opcode[4] = 1
>>
>>         31 28 27 26 25 24                                     5 4 3     0
>>         cond  0  1  1  x  x x x x x x x x x x x x x x x x x x x 1 x x x x
>>
>> So, in this version of the architecture, undefined instructions may
>> be conditional - and indeed that used to be the case.  The condition
>> code was always respected, and cond=1111 meant "never" (NV).
>>
>> Hence, trapping them if the condition code is not 1110 (AL) is
>> entirely reasonable, legal and safe.  If an ARM CPU defines an
>> instruction coding that matches the above, then it won't take the
>> undefined instruction trap, and we'll never see it.
>>
>> Now, as for UDF usage in the kernel, it may be quite correct that we
>> always use the AL condition code for them, but it would be very odd
>> for there to be an instruction implemented with a different (non-NV)
>> condition code that can't also have it's AL condition code encoding.
>> You could never execute such an instruction unconditionally.
>>
> 
> That makes sense. Thank you very much for a great answer!

The other subtlety is that UDF is a *mnemonic* for a specific encoding 
with cond=0b1110, thus "conditional UDF" is an oxymoron. You can still 
have a conditional instruction from the "permanently undefined" space 
(although on Armv7-A it's implementation-defined whether it causes an 
exception regardless of the condition), it's just not a UDF.

Robin.