Date: Thu, 23 May 2024 15:13:23 -0700
From: Yang Shi <yang@...amperecomputing.com>
To: Catalin Marinas <catalin.marinas@....com>,
 "Christoph Lameter (Ampere)" <cl@...ux.com>
Cc: will@...nel.org, anshuman.khandual@....com, scott@...amperecomputing.com,
 linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org
Subject: Re: [v2 PATCH] arm64: mm: force write fault for atomic RMW
 instructions



On 5/23/24 2:34 PM, Catalin Marinas wrote:
> On Thu, May 23, 2024 at 12:43:34PM -0700, Christoph Lameter (Ampere) wrote:
>> On Thu, 23 May 2024, Catalin Marinas wrote:
>>>>> While this class includes all atomics that currently require write
>>>>> permission, there's some unallocated space in this range and we don't
>>>>> know what future architecture versions may introduce. Unfortunately we
>>>>> need to check each individual atomic op in this class (not sure what the
>>>>> overhead will be).
>>>> Can you tell us which bits or pattern is not allocated? Maybe we can exclude
>>>> that from the pattern.
>>> Yes, it may be easier to exclude those patterns. See the Arm ARM K.a
>>> section C4.1.94.29 (page 791).
>> Hmmm. We could consult an exception table once the pattern matches to reduce
>> the overhead.
> Yeah, check the atomic class first and then go into the finer-grained
> details. I think this would reduce the overhead for non-atomic
> instructions.

If I read the instruction encoding correctly, the unallocated 
instructions are decided by the below fields:

   - size
   - VAR
   - o3
   - opc

To exclude them I think we can do something like:

if atomic instructions {
     if V == 1
         return false;
     if o3 opc == 111x
         return false;
     switch VAR {
         000
             check o3 and opc
         001
             check 03 and opc
         010
             check o3 and opc
         011
             check o3 and opc
         default
             if size != 11
                 check o3 and opc
     }
}

So it may take 4 + the possible unallocated combos of o3 and opc 
branches for the worst case. I saw 5 different combos for o3 and opc, so 
9 branches for worst cases.

>
>> However, the harm done I think is acceptable even if we leave things as is.
>> In the worst case we create unnecesssary write fault processing for an
>> "atomic op" that does not need write access. Also: Why would it need to be
>> atomic if it does not write???
> I'm thinking of some conditional instruction that states no write if
> condition fails. But it could be even worse if the architects decide to
> reuse that unallocated space for some instructions that have nothing to
> do with the atomic accesses.

Even though the condition fails, forcing write fault still seems fine 
IIUC. I'm supposed the read fault will happen regardless of the 
condition. Then a page with all 0 content is installed. This is 
guaranteed. We just end up having write permission instead of read-only 
permission. We will also be in this state transiently with current 
supported atomic instructions.

But if they will be allocated to non-atomic instructions, we have to do 
fine-grained decoding, but it may be easier since we can just filter out 
those non-atomic instructions? Anyway it depends on how they will be 
used. Hopefully this won't happen.

>
> It's something we need to clarify with them but I'm about to go on
> holiday for a week, so I won't be able to check.

Have a good holiday.

>
>> The ultimate solution would be to change the spec so that arm processors can
>> skip useless read faults.
> I raised this already, waiting for feedback from the architects.

Thank you so much.

>

Powered by blists - more mailing lists