Re: [PATCH V3 6/6] crypto/nx: Add P9 NX support for 842 compression engine

[Haren Myneni <haren@linux.vnet.ibm.com>]

On 09/01/2017 04:29 AM, Michael Ellerman wrote:
> Hi Dan,
> 
> Thanks for reviewing this series.
> 
> Dan Streetman <ddstreet@ieee.org> writes:
>> On Tue, Aug 29, 2017 at 5:54 PM, Haren Myneni <haren@linux.vnet.ibm.com> wrote:
>>> On 08/29/2017 02:23 PM, Benjamin Herrenschmidt wrote:
>>>> On Tue, 2017-08-29 at 09:58 -0400, Dan Streetman wrote:
>>>>>> +
>>>>>> +       ret = -EINVAL;
>>>>>> +       if (coproc && coproc->vas.rxwin) {
>>>>>> +               wmem->txwin = nx842_alloc_txwin(coproc);
>>>>>
>>>>> this is wrong.  the workmem is scratch memory that's valid only for
>>>>> the duration of a single operation.
>>>
>>> Correct, workmem is used until crypto_free is called.
>>
>> that's not a 'single operation'.  a single operation is compress() or
>> decompress().
>>
>>>>>
>>>>> do you actually need a txwin per crypto transform?  or do you need a
>>>>> txwin per coprocessor?  or txwin per processor?  either per-coproc or
>>>>> per-cpu should be created at driver init and held separately
>>>>> (globally) instead of a per-transform txwin.  I really don't see why
>>>>> you would need a txwin per transform, because the coproc should not
>>>>> care how many different transforms there are.
>>>>
>>>> We should only need a single window for the whole kernel really, plus
>>>> one per user process who wants direct access but that's not relevant
>>>> here.
>>>
>>> Opening send window for each crypto transform (crypto_alloc,
>>> compression/decompression, ..., crypto_free) so that does not
>>> have to wait for the previous copy/paste complete.
>>> VAS will map send and receive windows, and can cache in send
>>> windows (up to 128). So I thought using the same send window
>>> (per chip) for more requests (say 1000) may be adding overhead.
>>>
>>> I will make changes if you prefer using 1 send window per chip.
>>
>> i don't have the spec, so i shouldn't be making the decision on it,
>> but i do know putting a persistent field into the workmem is the wrong
>> location.  If it's valid for the life of the transform, put it into
>> the transform context.  The workmem buffer is intended to be used only
>> during a single operation - it's "working memory" to perform each
>> individual crypto transformation.
> 
> I agree workmem isn't the right place for the txwin. But I don't believe
> it actually breaks anything to put txwin there.
> 
> So for now I'm going to merge this series as-is and I've asked Haren to
> send fixes as soon as he can to clean it up.

I will move txwin to nx842_crypto_ctx and send patch soon.  Even though nx842_crypto_ctx is also used for pseries, we can ignore txwin. 

Thanks
Haren


> 
> cheers
>