Re: [PATCH 3/4] dt-binding: remoteproc: venus rproc dt binding document

[Marek Szyprowski]

Hi Bjorn,


On 2016-09-02 22:12, Bjorn Andersson wrote:
> On Fri 02 Sep 04:52 PDT 2016, Marek Szyprowski wrote:
>
> > On 2016-09-01 16:58, Stanimir Varbanov wrote:
> > ...
> > > > > > But I presume we have the implementation issue of dma_alloc_coherent()
> > > > > > failing in either case with the 5MB size. I think we need to look into
> > > > > I'd be good to include Marek Szyprowski? At least he will know what
> > > > > design restrictions there are.
> > > > Please do. The more I look at this the more I think we must use the
> > > > existing infrastructure for allocating "dma memory". Getting
> > > > dma_alloc_coherent() supporting non-power-of-2 memory regions would
> > > Just to be precise it should be dma_alloc_from_coherent().
> > >
> > > Marek, what is your opinion on that, can we make dma_alloc_from_coherent
> > > able to allocate memory for sizes with bigger granularity.
> > >
> > > For your convenience here [1] is the mail thread.
> > There should be no technical restrictions to add support for bigger
> > granularity than power-of-2. dma_alloc_from_coherent uses standard
> > bitmap based allocator, so it already support tracking allocations of
> > arbitrary size.
> I believe we should be able to change the parameter of
> bitmap_{find_free,release,allocate}_region() to take a size rather than
> an order.
>
> The mask used in __reg_op() is an unsigned long, that is stamped over
> the region to be masked or cleared, so there are some clear restrictions
> in what parameters we can pass there - without having to break this
> operation up in steps.
>
> But if drive the offset by taking the next power-of-two of the size and
> then align the number of bits to min(count, BITS_PER_LONG) we should
> retain the performance characteristics and requirements of __reg_op().
>
> > However for the small allocations (smaller than 64KiB?, 512KiB?) it
> > would make sense to keep nearest-power-of-2 round up to prevent memory
> > fragmentation.
> But in our case each bit matches a single page, so by making sure the
> mask always fills the unsigned long in the larger cases we would end up
> with having to align things to 128kb (or 256kb if unsigned long is 64
> bit).

By preventing memory fragmentation I wanted to align small allocations (less
than the mentioned 64KiB or 512KiB) to nearest-power-of-2 of their size just
like it is done now.

Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland