lkml.org 
[lkml]   [2020]   [Jul]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [BUG] XHCI getting ZONE_DMA32 memory > than its bus_dma_limit
On Fri, 3 Jul 2020, Robin Murphy wrote:

> > Just for the record the offending commit is: c84dc6e68a1d2 ("dma-pool: add
> > additional coherent pools to map to gfp mask").
> >
> > On Thu, 2020-07-02 at 12:49 -0500, Jeremy Linton wrote:
> > > Hi,
> > >
> > > Using 5.8rc3:
> > >
> > > The rpi4 has a 3G dev->bus_dma_limit on its XHCI controller. With a usb3
> > > hub, plus a few devices plugged in, randomly devices will fail
> > > operations. This appears to because xhci_alloc_container_ctx() is
> > > getting buffers > 3G via dma_pool_zalloc().
> > >
> > > Tracking that down, it seems to be caused by dma_alloc_from_pool() using
> > > dev_to_pool()->dma_direct_optimal_gfp_mask() to "optimistically" select
> > > the atomic_pool_dma32 but then failing to verify that the allocations in
> > > the pool are less than the dev bus_dma_limit.
> >
> > I can reproduce this too.
> >
> > The way I see it, dev_to_pool() wants a strict dma_direct_optimal_gfp_mask()
> > that is never wrong, since it's going to stick to that pool for the device's
> > lifetime. I've been looking at how to implement it, and it's not so trivial
> > as
> > I can't see a failproof way to make a distinction between who needs DMA32
> > and
> > who is OK with plain KERNEL memory.
> >
> > Otherwise, as Jeremy points out, the patch needs to implement allocations
> > with
> > an algorithm similar to __dma_direct_alloc_pages()'s, which TBH I don't know
> > if
> > it's a little overkill for the atomic context.
> >
> > Short of finding a fix in the coming rc's, I suggest we revert this.
>
> Or perhaps just get rid of atomic_pool_dma32 (and allocate atomic_pool_dma
> from ZONE_DMA32 if !ZONE_DMA). That should make it fall pretty much back in
> line while still preserving the potential benefit of the kernel pool for
> non-address-constrained devices.
>

I assume it depends on how often we have devices where
__dma_direct_alloc_pages() behavior is required, i.e. what requires the
dma_coherent_ok() checks and altering of the gfp flags to get memory that
works.

Is the idea that getting rid of atomic_pool_dma32 would use GFP_KERNEL
(and atomic_pool_kernel) as the default policy here? That doesn't do any
dma_coherent_ok() checks so dma_direct_alloc_pages would return from
ZONE_NORMAL without a < 3G check?

It *seems* like we want to check if dma_coherent_ok() succeeds for ret in
dma_direct_alloc_pages() when allocating from the atomic pool and, based
on criteria that allows fallback, just fall into
__dma_direct_alloc_pages()?

\
 
 \ /
  Last update: 2020-07-06 01:42    [W:0.046 / U:0.024 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site