[PATCH v1 1/1] ARM: Select DMA_DIRECT_REMAP to fix restricted DMA

[Jim Quinlan]

On Fri, Sep 29, 2023 at 3:52 PM Arnd Bergmann <arnd at arndb.de> wrote:
>
> On Fri, Sep 29, 2023, at 15:24, Jim Quinlan wrote:
> > On Thu, Sep 28, 2023 at 11:17 AM Arnd Bergmann <arnd at arndb.de> wrote:
> >> On Thu, Sep 28, 2023, at 10:00, Jim Quinlan wrote:
> >
> > Our RC is definitely not coherent with the ARM/ARM64 caches.
>
> Ok, thanks for the confirmation.
>
> >> It's unlikely but not impossible, as the driver has some
> >> unusual constructs, using a lot of coherent mappings that
> >> might otherwise be streaming mappings, and relying on
> >> dma_sync_single_for_device(..., DMA_BIDIRECTIONAL) for other
> >> data, but without the corresponding dma_sync_single_for_cpu().
> >> If all the testing happens on x86, this might easily lead
> >> to a bug that only shows up on non-coherent systems but
> >> is never seen during testing.
> >>
> >> If the problem is not the "dma-coherent" property, can you
> >> double-check if using a different PCIe device works, or narrow
> >> down which specific buffer you saw get corrupted?
> >
> > I've done some testing, below are the results.  The new two devices, a
> > USB controller
> > and an M2 NVMe stick, behave the same as iwlwifi.
> >
> > Note that I'm not advocating that "select DMA_DIRECT_REMAP" is the
> > anser, I'm just showing that it fixes my examples.
>
> Ok, so I think we can stop looking at the device drivers for
> bugs then.
>
> > VER      PCI-DEV                       <--------- RESTRICTED DMA --------->
> >                       ARM64    ARM     ARM64    ARM    ARM+DMA_DIRECT_REMAP
> > 5.15     iwlwifi        P       P        P       F             P
> > 5.15     nvme           P       P        P       F             P
> > 5.15     usb            P       P        P       F             P
> >
> > 6.1      iwlwifi        P       P        P       F             P
> > 6.1      nvme           P       P        P       F             P
> > 6.1      usb            P        P       P       F             P
> >
> > Upstrm   iwlwifi        P       P        F       F             F
> > Upstrm   nvme           P       P        F       F             F
> > Upstrm   usb            P       P        F       F             F
> >                       ARM64    ARM     ARM64    ARM    ARM+DMA_DIRECT_REMAP
> > VER      PCI-DEV                       <--------- RESTRICTED DMA --------->
> >
> > LEGEND:
> >   P       := pass, driver probe and some functional test performed
> >   F       := fail, usually when probe is called; impossible to do
> > functional test
> >   Upstrm  := 633b47cb009d "Merge tag 'scsi-fixes' of
> > git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi"
> >
> >   iwlwifi := 7260 Wifi 8086:08b1
> >   nvme    := 1e95:1007
> >   usb     := Supahub, 1912:0014
>
> Thanks for the thorough testing, that looks very useful, even though
> I don't have an answer immediately. Maybe Robin can see something
> here that I don't.
>
> It's particularly interesting how arm64 only started failing
> on fairly recent kernels, so if nothing else helps you could
> always try bisecting the history between 6.1 and 633b47cb009d,
> hoping that the commit that broke it points us to the arm32
> problem.
>
> The only change I see in that time frame that seems related
> is 7bd6680b47fa ("Revert "Revert "arm64: dma: Drop cache
> invalidation from arch_dma_prep_coherent()"""), so you could
> start by reverting that. However, it's probably something
> else since this is for the coherent mappings, not the
> streaming ones.

Hi Arnd,

I started to do the bisection and it was clear early on that something
else was afoot --
it turns out that our 6.1 and 5.15 builds included a NAK'd upstream
commit which was required
by our legacy DT tree.  Once I updated our DT tree, the upstream results matched
those of 6.1 and 5.15.  I've included the new results below.

Long story short, it appears that restricted DMA does not work with ARCH=arm
platforms using non-dma-coherent devices.  AFAICT, the
dma_alloc_coherent() function
should return a pointer to an uncacheable page -- the only way our system can
have coherent memory -- but it is not.

The "select DMA_DIRECT_REMAP" makes things work in 5.15, 6.1 and upstream,
although I cannot really say if that is a viable solution.

Regards,
Jim Quinlan
Broadcom STB/CM

VER      PCI-DEV                       <--------- RESTRICTED DMA --------->
                      ARM64    ARM     ARM64    ARM    ARM+DMA_DIRECT_REMAP
5.15     iwlwifi        P       P        P       F             P
5.15     nvme           P       P        P       F             P
5.15     usb            P       P        P       F             P

6.1      iwlwifi        P       P        P       F             P
6.1      nvme           P       P        P       F             P
6.1      usb            P       P        P       F             P

Upstrm   iwlwifi        P       P        P       F             P
Upstrm   nvme           P       P        P       F             P
Upstrm   usb            P       P        P       F             P
                      ARM64    ARM     ARM64    ARM    ARM+DMA_DIRECT_REMAP
VER      PCI-DEV                       <--------- RESTRICTED DMA --------->

LEGEND:
  P       := pass, driver probe and some functional test performed
  F       := fail, usually when probe is called; impossible to do
functional test
  Upstrm  := 633b47cb009d "Merge tag 'scsi-fixes' of
git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi"

  iwlwifi := 7260 Wifi 8086:08b1
  nvme    := 1e95:1007
  usb     := Supahub, 1912:0014
>
>        Arnd
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 4210 bytes
Desc: S/MIME Cryptographic Signature
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20231001/88a5c4fb/attachment.p7s>