Re: [PATCH v3 2/6] dmaengine: Add interleaved cyclic transaction type

[Laurent Pinchart <laurent.pinchart@ideasonboard.com>]

Hi Vinod,

On Wed, Mar 18, 2020 at 05:14:27PM +0200, Laurent Pinchart wrote:
> On Wed, Mar 11, 2020 at 05:52:48PM +0200, Laurent Pinchart wrote:
> > On Wed, Mar 04, 2020 at 09:54:26PM +0530, Vinod Koul wrote:
> >> On 04-03-20, 18:00, Laurent Pinchart wrote:
> >>> On Wed, Mar 04, 2020 at 09:07:18PM +0530, Vinod Koul wrote:
> >>>> On 04-03-20, 10:01, Laurent Pinchart wrote:
> >>>>> On Wed, Mar 04, 2020 at 10:43:01AM +0530, Vinod Koul wrote:
> >>>>>> On 03-03-20, 21:22, Laurent Pinchart wrote:
> >>>>>>> On Tue, Mar 03, 2020 at 10:02:54AM +0530, Vinod Koul wrote:
> >>>>>>>> On 02-03-20, 09:37, Laurent Pinchart wrote:
> >>>>>>>>>> I would be more comfortable in calling an API to do so :)
> >>>>>>>>>> The flow I am thinking is:
> >>>>>>>>>> 
> >>>>>>>>>> - prep cyclic1 txn
> >>>>>>>>>> - submit cyclic1 txn
> >>>>>>>>>> - call issue_pending() (cyclic one starts)
> >>>>>>>>>> 
> >>>>>>>>>> - prep cyclic2 txn
> >>>>>>>>>> - submit cyclic2 txn
> >>>>>>>>>> - signal_cyclic1_txn aka terminate_cookie()
> >>>>>>>>>> - cyclic1 completes, switch to cyclic2 (dmaengine driver)
> >>>>>>>>>> - get callback for cyclic1 (optional)
> >>>>>>>>>> 
> >>>>>>>>>> To check if hw supports terminate_cookie() or not we can check if the
> >>>>>>>>>> callback support is implemented
> >>>>>>>>> 
> >>>>>>>>> Two questions though:
> >>>>>>>>> 
> >>>>>>>>> - Where is .issue_pending() called for cyclic2 in your above sequence ?
> >>>>>>>>>   Surely it should be called somewhere, as the DMA engine API requires
> >>>>>>>>>   .issue_pending() to be called for a transfer to be executed, otherwise
> >>>>>>>>>   it stays in the submitted but not pending queue.
> >>>>>>>> 
> >>>>>>>> Sorry missed that one, I would do that after submit cyclic2 txn step and
> >>>>>>>> then signal signal_cyclic1_txn termination
> >>>>>>> 
> >>>>>>> OK, that matches my understanding, good :-)
> >>>>>>> 
> >>>>>>>>> - With the introduction of a new .terminate_cookie() operation, we need
> >>>>>>>>>   to specify that operation for all transfer types. What's its
> >>>>>>>> 
> >>>>>>>> Correct
> >>>>>>>> 
> >>>>>>>>>   envisioned semantics for non-cyclic transfers ? And how do DMA engine
> >>>>>>>>>   drivers report that they support .terminate_cookie() for cyclic
> >>>>>>>>>   transfers but not for other transfer types (the counterpart of
> >>>>>>>>>   reporting, in my proposition, that .issue_pending() isn't supported
> >>>>>>>>>   replace the current cyclic transfer) ?
> >>>>>>>> 
> >>>>>>>> Typically for dmaengine controller cyclic is *not* a special mode, only
> >>>>>>>> change is that a list provided to controller is circular.
> >>>>>>> 
> >>>>>>> I don't agree with this. For cyclic transfers to be replaceable in a
> >>>>>>> clean way, the feature must be specifically implemented at the hardware
> >>>>>>> level. A DMA engine that supports chaining transfers with an explicit
> >>>>>>> way to override that chaining, and without the logic to report if the
> >>>>>>> inherent race was lost or not, really can't support this API.
> >>>>>> 
> >>>>>> Well chaining is a typical feature in dmaengine and making last chain
> >>>>>> point to first makes it circular. I have seen couple of engines and this
> >>>>>> was the implementation in the hardware.
> >>>>>> 
> >>>>>> There can exist special hardware for this purposes as well, but the
> >>>>>> point is that the cyclic can be treated as circular list.
> >>>>>> 
> >>>>>>> Furthemore, for non-cyclic transfers, what would .terminate_cookie() do
> >>>>>>> ? I need it to be defined as terminating the current transfer when it
> >>>>>>> ends for the cyclic case, not terminating it immediately. All non-cyclic
> >>>>>>> transfers terminate by themselves when they end, so what would this new
> >>>>>>> operation do ?
> >>>>>> 
> >>>>>> I would use it for two purposes, cancelling txn but at the end of
> >>>>>> current txn. I have couple of usages where this would be helpful.
> >>>>> 
> >>>>> I fail to see how that would help. Non-cyclic transfers always stop at
> >>>>> the end of the transfer. "Cancelling txn but at the end of current txn"
> >>>>> is what DMA engine drivers already do if you call .terminate_cookie() on
> >>>>> the ongoing transfer. It would thus be a no-op.
> >>>> 
> >>>> Well that actually depends on the hardware, some of them support abort
> >>>> so people cancel it (terminate_all approach atm)
> >>> 
> >>> In that case it's not terminating at the end of the current transfer,
> >>> but terminating immediately (a.k.a. aborting), right ? Cancelling at the
> >>> end of the current transfer still seems to be a no-op to me for
> >>> non-cyclic transfers, as that's what they do on their own already.
> >> 
> >> Correct, it is abort for current txn.
> >> 
> >>>>>> Second in error handling where some engines do not support
> >>>>>> aborting (unless we reset the whole controller)
> >>>>> 
> >>>>> Could you explain that one ? I'm not sure to understand it.
> >>>> 
> >>>> So I have dma to a slow peripheral and it is stuck for some reason. I
> >>>> want to abort the cookie and let subsequent ones runs (btw this is for
> >>>> non cyclic case), so I would use that here. Today we terminate_all and
> >>>> then resubmit...
> >>> 
> >>> That's also for immediate abort, right ?
> >> 
> >> Right
> >> 
> >>> For this to work properly we need very accurate residue reporting, as
> >>> the client will usually need to know exactly what has been transferred.
> >>> The device would need to support DMA_RESIDUE_GRANULARITY_BURST when
> >>> aborting an ongoing transfer. What hardware supports this ?
> >> 
> >>  git grep DMA_RESIDUE_GRANULARITY_BURST drivers/dma/ |wc -l
> >> 27
> >> 
> >> So it seems many do support the burst reporting.
> > 
> > Yes, but not all of those may support aborting a transfer *and*
> > reporting the exact residue of cancelled transfers. We need both to
> > implement your proposal.
> > 
> >>>>>> But yes the .terminate_cookie() semantics should indicate if the
> >>>>>> termination should be immediate or end of current txn. I see people
> >>>>>> using it for both.
> >>>>> 
> >>>>> Immediate termination is *not* something I'll implement as I have no
> >>>>> good way to test that semantics. I assume you would be fine with leaving
> >>>>> that for later, when someone will need it ?
> >>>> 
> >>>> Sure, if you have hw to support please test. If not, you will not
> >>>> implement that.
> >>>> 
> >>>> The point is that API should support it and people can add support in
> >>>> the controllers and test :)
> >>> 
> >>> I still think this is a different API. We'll have
> >>> 
> >>> 1. Existing .issue_pending(), queueing the next transfer for non-cyclic
> >>>    cases, and being a no-op for cyclic cases.
> >>> 2. New .terminate_cookie(AT_END_OF_TRANSFER), being a no-op for
> >>>    non-cyclic cases, and moving to the next transfer for cyclic cases.
> >>> 3. New .terminate_cookie(ABORT_IMMEDIATELY), applicable to both cyclic
> >>>    and non-cyclic cases.
> >>> 
> >>> 3. is an API I don't need, and can't easily test. I agree that it can
> >>> have use cases (provided the DMA device can abort an ongoing transfer
> >>> *and* still support DMA_RESIDUE_GRANULARITY_BURST in that case).
> >>> 
> >>> I'm troubled by my inability to convince you that 1. and 2. are really
> >>> the same, with 1. addressing the non-cyclic case and 2. addressing the
> >>> cyclic case :-) This is why I think they should both be implemeted using
> >>> .issue_pending() (no other option for 1., that's what it uses today).
> >>> This wouldn't prevent implementing 3. with a new .terminate_cookie()
> >>> operation, that wouldn't need to take a flag as it would always operate
> >>> in ABORT_IMMEDIATELY mode. There would also be no need to report a new
> >>> capability for 3., as the presence of the .terminate_cookie() handler
> >>> would be enough to tell clients that the API is supported. Only a new
> >>> capability for 2. would be needed.
> >> 
> >> Well I agree 1 & 2 seem similar but I would like to define the behaviour
> >> not dependent on the txn being cyclic or not. That is my concern and
> >> hence the idea that:
> >> 
> >> 1. .issue_pending() will push txn to pending_queue, you may have a case
> >> where that is done only once (due to nature of txn), but no other
> >> implication
> >> 
> >> 2. .terminate_cookie(EOT) will abort the transfer at the end. Maybe not
> >> used for cyclic but irrespective of that, the behaviour would be abort
> >> at end of cyclic
> > 
> > Did you mean "maybe not used for non-cyclic" ?
> > 
> >> 3. .terminate_cookie(IMMEDIATE) will abort immediately. If there is
> >> anything in pending_queue that will get pushed to hardware.
> >> 
> >> 4. Cyclic by nature never completes
> >>    - as a consequence needs to be stopped by terminate_all/terminate_cookie
> >> 
> >> Does these rules make sense :)
> > 
> > It's a set of rules that I think can handle my use case, but I still
> > believe my proposal based on just .issue_pending() would be simpler, in
> > line with the existing API concepts, and wouldn't preclude the addition
> > of .terminate_cookie(IMMEDIATE) at a later point. It's your call though,
> > especially if you provide the implementation :-) When do you think you
> > will be able to do so ?
> 
> Gentle ping :-)

Any update ?

> >>>>>> And with this I think it would make sense to also add this to
> >>>>>> capabilities :)
> >>>>> 
> >>>>> I'll repeat the comment I made to Peter: you want me to implement a
> >>>>> feature that you think would be useful, but is completely unrelated to
> >>>>> my use case, while there's a more natural way to handle my issue with
> >>>>> the current API, without precluding in any way the addition of your new
> >>>>> feature in the future. Not fair.
> >>>> 
> >>>> So from API design pov, I would like this to support both the features.
> >>>> This helps us to not rework the API again for the immediate abort.
> >>>> 
> >>>> I am not expecting this to be implemented by you if your hw doesn't
> >>>> support it. The core changes are pretty minimal and callback in the
> >>>> driver is the one which does the job and yours wont do this
> >>> 
> >>> Xilinx DMA drivers don't support DMA_RESIDUE_GRANULARITY_BURST so I
> >>> can't test this indeed.
> >> 
> >> Sure I understand that! Am sure folks will respond to CFT and I guess
> >> Peter will also be interested in testing.
> > 
> > s/testing/implementing it/ :-)
> > 
> >>>>>>>> So, the .terminate_cookie() should be a feature for all type of txn's.
> >>>>>>>> If for some reason (dont discount what hw designers can do) a controller
> >>>>>>>> supports this for some specific type(s), then they should return
> >>>>>>>> -ENOTSUPP for cookies that do not support and let the caller know.
> >>>>>>> 
> >>>>>>> But then the caller can't know ahead of time, it will only find out when
> >>>>>>> it's too late, and can't decide not to use the DMA engine if it doesn't
> >>>>>>> support the feature. I don't think that's a very good option.
> >>>>>> 
> >>>>>> Agreed so lets go with adding these in caps.
> >>>>> 
> >>>>> So if there's a need for caps anyway, why not a cap that marks
> >>>>> .issue_pending() as moving from the current cyclic transfer to the next
> >>>>> one ? 
> >>>> 
> >>>> Is the overhead really too much on that :) If you like I can send the
> >>>> core patches and you would need to implement the driver side?
> >>> 
> >>> We can try that as a compromise. One of main concerns with developing
> >>> the core patches myself is that the .terminate_cookie() API still seems
> >>> ill-defined to me, so it would be much more efficient if you translate
> >> 
> >> yeah lets take a stab at defining this and see if we come up with
> >> something meaningful
> >> 
> >>> the idea you have in your idea into code than trying to communicate it
> >>> to me in all details (one of the grey areas is what should
> >>> .terminate_cookie() do if the cookie passed to the function corresponds
> >>> to an already terminated or, more tricky from a completion callback
> >>> point of view, an issued but not-yet-started transfer, or also a
> >>> submitted but not issued transfer). If you implement the core part, then
> >>> that problem will go away.
> >>> 
> >>> How about the implementation in virt-dma.[ch] by the way ?
> >> 
> >> It needs to be comprehended and tested as well.. since these are simple
> >> callbacks to driver, we should not need huge changes here (i need to
> >> double check though)

-- 
Regards,

Laurent Pinchart