Re: regression in ath10k dma allocation

[Nicolin Chen <nicoleotsuka@gmail.com>]

Hello Hillf,

On Mon, Aug 19, 2019 at 12:38:38AM +0200, Tobias Klausmann wrote:
> 
> On 18.08.19 05:13, Hillf Danton wrote:
> > On Sat, 17 Aug 2019 00:42:48 +0200 Tobias Klausmann wrote:
> > > Hi Nicolin,
> > > 
> > > On 17.08.19 00:25, Nicolin Chen wrote:
> > > > Hi Tobias
> > > > 
> > > > On Fri, Aug 16, 2019 at 10:16:45PM +0200, Tobias Klausmann wrote:
> > > > > > do you have CONFIG_DMA_CMA set in your config?  If not please make sure
> > > > > > you have this commit in your testing tree, and if the problem still
> > > > > > persists it would be a little odd and we'd have to dig deeper:
> > > > > > 
> > > > > > commit dd3dcede9fa0a0b661ac1f24843f4a1b1317fdb6
> > > > > > Author: Nicolin Chen <nicoleotsuka@gmail.com>
> > > > > > Date:   Wed May 29 17:54:25 2019 -0700
> > > > > > 
> > > > > >        dma-contiguous: fix !CONFIG_DMA_CMA version of dma_{alloc, free}_contiguous()
> > > > > yes CONFIG_DMA_CMA is set (=y, see attached config), the commit you mention
> > > > > above is included, if you have any hints how to go forward, please let me
> > > > > know!
> > > > For CONFIG_DMA_CMA=y, by judging the log with error code -12, I
> > > > feel this one should work for you. Would you please check if it
> > > > is included or try it out otherwise?
> > > > 
> > > > dma-contiguous: do not overwrite align in dma_alloc_contiguous()
> > > > https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/commit/?id=c6622a425acd1d2f3a443cd39b490a8777b622d7
> > > 
> > > Thanks for the hint, yet the commit is included and does not fix the
> > > problem!
> > > 
> Hi Hillf,
> 
> i just tested you first hunk (which comes from kernel/dma/direct.c if i'm
> not mistaken), it did not compile on its own, yet with a tiny bit of work it
> did, and it does indeed solve the regression. But if using that is the
> "right" way to do it, not sure, but its not on me to decide.
> 
> Anyway: Thanks for the hint,
> 
> Tobias
> 
> 
> > Hi Tobias
> > 
> > Two minor diffs below in hope that they might make sense.
> > 
> > 1, fallback unless dma coherent ok.
> > 
> > --- a/kernel/dma/contiguous.c
> > +++ b/kernel/dma/contiguous.c
> > @@ -246,6 +246,10 @@ struct page *dma_alloc_contiguous(struct
> >   		size_t cma_align = min_t(size_t, align, CONFIG_CMA_ALIGNMENT);
> >   		page = cma_alloc(cma, count, cma_align, gfp & __GFP_NOWARN);
> > +		if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
> > +			dma_free_contiguous(dev, page, size);
> > +			page = NULL;
> > +		}

Right...the condition was in-between. However, not every caller
of dma_alloc_contiguous() is supposed to have a coherent check.
So we either add a 'bool coherent_ok' to the API or revert the
dma-direct part back to the original. Probably former option is
better?

Thank you for the debugging. I have been a bit distracted, may
not be able to submit a fix very soon. Would you like to help?

Thanks!
Nicolin

> >   	}
> >   	/* Fallback allocation of normal pages */
> > --
> > 
> > 2, cleanup: cma unless contiguous
> > 
> > --- a/kernel/dma/contiguous.c
> > +++ b/kernel/dma/contiguous.c
> > @@ -234,18 +234,13 @@ struct page *dma_alloc_contiguous(struct
> >   	size_t count = PAGE_ALIGN(size) >> PAGE_SHIFT;
> >   	size_t align = get_order(PAGE_ALIGN(size));
> >   	struct page *page = NULL;
> > -	struct cma *cma = NULL;
> > -
> > -	if (dev && dev->cma_area)
> > -		cma = dev->cma_area;
> > -	else if (count > 1)
> > -		cma = dma_contiguous_default_area;
> >   	/* CMA can be used only in the context which permits sleeping */
> > -	if (cma && gfpflags_allow_blocking(gfp)) {
> > +	if (count > 1 && gfpflags_allow_blocking(gfp)) {
> >   		size_t cma_align = min_t(size_t, align, CONFIG_CMA_ALIGNMENT);
> > -		page = cma_alloc(cma, count, cma_align, gfp & __GFP_NOWARN);
> > +		page = cma_alloc(dev_get_cma_area(dev), count, cma_align,
> > +							gfp & __GFP_NOWARN);
> >   		if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
> >   			dma_free_contiguous(dev, page, size);
> >   			page = NULL;
> > --
> >