Re: [PATCH] MIPS: MSA: misaligned support

[Leonid Yegoshin <Leonid.Yegoshin@imgtec.com>]

On 03/19/2015 02:51 AM, James Hogan wrote:
> On 18/03/15 23:25, Leonid Yegoshin wrote:
>> On 03/18/2015 03:12 PM, James Hogan wrote:
>>> Hi Leonid,
>>>
>>> On Wed, Mar 18, 2015 at 12:46:51PM -0700, Leonid Yegoshin wrote:
>>>
>>>> thread_msa_context_live() == check of TIF_MSA_CTX_LIVE == existence of
>>>> MSA context for thread.
>>>> It differs from MSA is owned by thread, it just says that thread has
>>>> already initialized MSA.
>>>>
>>>> Unfortunate choice of function name, I believe.
>>> Right (I mis-read when its cleared when i grepped). Still, that would
>>> make it even harder to hit since lose_fpu wouldn't clear it, and you
>>> already would've taken an MSA disabled exception first.
>> No, lose_fpu disables MSA now, saves MSA context and switches off
>> TIF_USEDMSA. See 33c771ba5c5d067f85a5a6c4b11047219b5b8f4e, "MIPS:
>> save/disable MSA in lose_fpu".
>>
>> However, a process still has MSA context initialized and it is indicated
>> by TIF_MSA_CTX_LIVE.
>> It should have it before it can get any AdE exception on MSA instruction.
> Yes, exactly.
>
>>> Anyway, my point was that there's nothing invalid about an unaligned
>>> load being the first MSA instruction. You might use it to load the
>>> initial vector state.
>> No, it is invalid. If MSA is disabled it should trigger "MSA Disabled"
>> exception.
> It's valid for the user to start their program with a ld.b.
> As you say, it'll raise an MSA disabled exception first though. The
> handler will own MSA, and set TIF_MSA_CTX_LIVE, which makes the check
> pointless?

"Unfortunately, some HW versions had AdE first and it may be logical 
from some HW point (if access is done before instruction is completely 
decoded). But that is wrong."

>
> I suppose an AdE from a normal unaligned load could still race with
> another thread modifying the instruction to an MSA ld.b, but even if it
> did, I don't think it would do any harm?
>
>> Unfortunately, some HW versions had AdE first and it may be logical from
>> some HW point (if access is done before instruction is completely
>> decoded). But that is wrong.
> Yes, MSA Disabled would clearly come under "Instruction Validity
> Exceptions", which is very sensibly higher priority than "Address error
> - Data access".
>
> Anyway, at the very least it needs a comment to justify what it is
> trying to catch and what harm it is trying to avoid, since it isn't
> obvious, and tbh seems pointless.
>
> Cheers
> James
>