Re: [PATCH] swiotlb: Fix unexpected swiotlb_alloc_coherent() failures

[Takashi Iwai]

On Thu, 12 Apr 2018 12:32:54 +0200,
Robin Murphy wrote:
> 
> On 12/04/18 09:27, Takashi Iwai wrote:
> > On Thu, 12 Apr 2018 10:19:05 +0200,
> > Takashi Iwai wrote:
> >>
> >> On Thu, 12 Apr 2018 10:03:56 +0200,
> >> Takashi Iwai wrote:
> >>>
> >>> On Thu, 12 Apr 2018 08:02:27 +0200,
> >>> Christoph Hellwig wrote:
> >>>>
> >>>> On Wed, Apr 11, 2018 at 09:28:54AM +0200, Takashi Iwai wrote:
> >>>>>> But we should try a GFP_DMA32 allocation first, so this is a bit
> >>>>>> surprising.
> >>>>>
> >>>>> Hm, do we really try that?
> >>>>> Through a quick glance, dma_alloc_coherent_gfp_flags() gives GFP_DMA32
> >>>>> only when coherent mask <= DMA_BIT_MASK(32); in the case of iwlwifi,
> >>>>> it's 36bit, so GFP_DMA isn't set.
> >>>>
> >>>> Oh, yes - it is using an odd dma mask, and amdgpu seems to use an
> >>>> just as odd 40-bit dma mask.
> >>>>
> >>>>> We had a fallback allocation with GFP_DMA32 in the past, but this
> >>>>> seems gone long time ago along with cleanups (commit c647c3bb2d16).
> >>>>>
> >>>>> But I haven't followed about this topic for long time, so I might have
> >>>>> missed obviously...
> >>>>
> >>>> I think a fallback would be much better here rather than relying on the
> >>>> limited swiotlb buffer bool.  dma_direct_alloc (which in 4.17 is also
> >>>> used for x86) already has a GFP_DMA fallback, so extending this for
> >>>> GFP_DMA32 as well would seem reasonable.
> >>>>
> >>>> Any volunteers?
> >>>
> >>> Below is a quick attempt, totally untested.  Actually the retry with
> >>> GFP_DMA is superfluous for archs without it, so the first patch
> >>> corrects it.
> >>
> >> Gah, scratch this, it doesn't work.  A different check is needed...
> >
> > The v2 patches are below, replaced with IS_ENABLED(CONFIG_ZONE_DMA).
> 
> That looks pretty reasonable to me - I'd be tempted to try a factoring
> out a helper function to avoid the "goto again" logic, but that's
> hardly crucial.

Right, this change won't make the code flow more complex than before,
so I guess it's OK as a first step.

I'll submit this second patch as RFC while the first one as a safe
fix (in addition to another regression fix I stumbled on).

> What I'd really love to see is something like the alloc_pages_mask()
> proposal from years ago to come back such that GFP_DMA32 could die
> entirely, but I don't know anywhere near enough about the mm layer to
> consider taking that on myself :(

Yeah, it'd be lovely, but I'm afraid it's hard to fly high...


thanks,

Takashi