lkml.org 
[lkml]   [2016]   [Aug]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] dmaengine: qcom_hidma: release the descriptor before the callback
On Mon, Aug 08, 2016 at 02:25:28PM +0200, Lars-Peter Clausen wrote:
> On 08/08/2016 11:08 AM, Vinod Koul wrote:
> > On Thu, Aug 04, 2016 at 05:59:30PM +0200, Lars-Peter Clausen wrote:
> >> On 08/04/2016 05:38 PM, Russell King - ARM Linux wrote:
> >> [...]
> >>> What you instead need to do is to find some way to record in your
> >>> driver that transaction 2 failed, and when dma_cookie_status() says
> >>> that a transaction has DMA_COMPLETE status, you need to look up to
> >>> see whether it failed.
> >>
> >> In my opinion this is where the current API is broken by design. For each
> >> transfer that fails you need to store the cookie associated with that
> >> transfer in some kind of lookup table. Since there is no lifetime associated
> >> with a cookie entries in this table would need to be retained forever and it
> >> will grow unbound.
> >
> > And how many drivers can report errors? And how many drivers can guarantee
> > DMA_COMPLETE implies transaction was succesful.
>
> The former just a handful, the later hopefully all.
>
> >
> >> Ideally we'd mark error reporting through this interface as deprecated and
> >> discourage new users of the interface. As far as I can see most of the few
> >> drivers that do return DMA_ERROR get it wrong anyway, e.g. return it
> >> unconditionally for all cookies when an error occurred for any of them.
> >
> > Error reporting is quite tricky as detection is a problem. So yes if you
> > can do so, it is highly encouraged to report using new interface which is
> > better than client checking after callback.
> >
> > Btw what is the behaviour after error? I would think that client will see an
> > error and report to upper layer while initiaite closure of transaction. So
> > does driver need to keep the state for a longer time :-)
>
> The problem is that this is not really clearly defined.
>
> 1) What should be done when multiple descriptors are queued and an error is
> encountered on one of them. Should the descriptors that are after the one in
> the queue that caused the error be discarded or should they be executed as
> normal?

That is client's call.

But a reasonable way would be for client to propagate those errors up, so it
can terminate.

> 2) How long does a error result need to be retained. Can it be discarded
> when the terminate_all() is called, or can it be discarded when the next
> issue_pending() is called or should it be retained forever?

Uptill next terminate_all()

> Unless we can clearly define the semantics of error reporting it is very
> difficult for drivers to use it. Which is probably one of the reasons why
> there are only very few DMAengine consumers that do actual error checking.

Yes agreed, but most reasonable behaviour is to terminate. Also I would
expect the error reporting to be done thru new API and explcitly told that
we found error (if we can).

-
~Vinod

\
 
 \ /
  Last update: 2016-08-10 21:41    [W:0.090 / U:3.448 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site