From: Peng Fan <[email protected]>
ZONE_DMA should not be disabled, otherwise arm64_dma_phys_limit is
left uninitialized and cause swiotlb have IO TLB above 4GB which
might crash some platforms
Signed-off-by: Peng Fan <[email protected]>
---
Not sure whether need to address code to initialize the variables or
force select ZONE_DMA
arch/arm64/Kconfig | 1 +
1 file changed, 1 insertion(+)
diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index 33446269f692..377b391a41d6 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -206,6 +206,7 @@ config ARM64
select SWIOTLB
select SYSCTL_EXCEPTION_TRACE
select THREAD_INFO_IN_TASK
+ select ZONE_DMA
help
ARM 64-bit (AArch64) Linux support.
--
2.28.0
Hi Peng,
sorry for the inconvenience, this is most probably related to these changes:
https://lore.kernel.org/linux-arm-kernel/[email protected]/
On Thu, 2020-12-17 at 16:08 +0800, Peng Fan (OSS) wrote:
> From: Peng Fan <[email protected]>
>
> ZONE_DMA should not be disabled, otherwise arm64_dma_phys_limit is
> left uninitialized and cause swiotlb have IO TLB above 4GB which
> might crash some platforms
>
> Signed-off-by: Peng Fan <[email protected]>
> ---
>
> Not sure whether need to address code to initialize the variables or
> force select ZONE_DMA
What is the cause for the swiotlb related crashes? I assume it's DMA into an
address too high for the bus, but it might be something else.
I figure you have a setup with ZONE_DMA32, ZONE_NORMAL and !ZONE_DMA.
First of all, I'd suggest you try arm64's defaults (all zones enabled), the
series I mention above should fix most of the issues we've had with
ZONE_DMA/ZONE_DMA32 in the past. We now parse DT/ACPI and only create two
distinct DMA zones if really needed. Otherwise ZONE_DMA spans the whole 32 bit
address space.
That said, IMO we're not doing the right thing in the !ZONE_DMA && ZONE_DMA32
case, and this should fix it (I didn't test it):
- #define ARCH_LOW_ADDRESS_LIMIT (arm64_dma_phys_limit - 1)
+ #define ARCH_LOW_ADDRESS_LIMIT (arm64_dma_phys_limit ? : arm64_dma32_phys_limit)
Regards,
Nicolas
On Thu, Dec 17, 2020 at 09:44:31AM +0100, Nicolas Saenz Julienne wrote:
> On Thu, 2020-12-17 at 16:08 +0800, Peng Fan (OSS) wrote:
> > From: Peng Fan <[email protected]>
> >
> > ZONE_DMA should not be disabled, otherwise arm64_dma_phys_limit is
> > left uninitialized and cause swiotlb have IO TLB above 4GB which
> > might crash some platforms
Actually, arm64_dma_phys_limit would be zero-initialised. The
ARCH_LOW_ADDRESS_LIMIT becomes ~0UL, hence it covers the whole address
space.
> > Signed-off-by: Peng Fan <[email protected]>
> > ---
> >
> > Not sure whether need to address code to initialize the variables or
> > force select ZONE_DMA
>
> What is the cause for the swiotlb related crashes? I assume it's DMA into an
> address too high for the bus, but it might be something else.
I think that's the case, swiotlb is not within the low 32-bit of the
address space.
> I figure you have a setup with ZONE_DMA32, ZONE_NORMAL and !ZONE_DMA.
>
> First of all, I'd suggest you try arm64's defaults (all zones enabled), the
> series I mention above should fix most of the issues we've had with
> ZONE_DMA/ZONE_DMA32 in the past. We now parse DT/ACPI and only create two
> distinct DMA zones if really needed. Otherwise ZONE_DMA spans the whole 32 bit
> address space.
>
> That said, IMO we're not doing the right thing in the !ZONE_DMA && ZONE_DMA32
> case, and this should fix it (I didn't test it):
>
> - #define ARCH_LOW_ADDRESS_LIMIT (arm64_dma_phys_limit - 1)
> + #define ARCH_LOW_ADDRESS_LIMIT (arm64_dma_phys_limit ? : arm64_dma32_phys_limit)
Does this limit need to be inclusive?
--
Catalin
On Thu, 2020-12-17 at 12:43 +0000, Catalin Marinas wrote:
> On Thu, Dec 17, 2020 at 09:44:31AM +0100, Nicolas Saenz Julienne wrote:
> > On Thu, 2020-12-17 at 16:08 +0800, Peng Fan (OSS) wrote:
> > > From: Peng Fan <[email protected]>
> > >
> > > ZONE_DMA should not be disabled, otherwise arm64_dma_phys_limit is
> > > left uninitialized and cause swiotlb have IO TLB above 4GB which
> > > might crash some platforms
>
> Actually, arm64_dma_phys_limit would be zero-initialised. The
> ARCH_LOW_ADDRESS_LIMIT becomes ~0UL, hence it covers the whole address
> space.
>
> > > Signed-off-by: Peng Fan <[email protected]>
> > > ---
> > >
> > > Not sure whether need to address code to initialize the variables or
> > > force select ZONE_DMA
> >
> > What is the cause for the swiotlb related crashes? I assume it's DMA into an
> > address too high for the bus, but it might be something else.
>
> I think that's the case, swiotlb is not within the low 32-bit of the
> address space.
>
> > I figure you have a setup with ZONE_DMA32, ZONE_NORMAL and !ZONE_DMA.
> >
> > First of all, I'd suggest you try arm64's defaults (all zones enabled), the
> > series I mention above should fix most of the issues we've had with
> > ZONE_DMA/ZONE_DMA32 in the past. We now parse DT/ACPI and only create two
> > distinct DMA zones if really needed. Otherwise ZONE_DMA spans the whole 32 bit
> > address space.
> >
> > That said, IMO we're not doing the right thing in the !ZONE_DMA && ZONE_DMA32
> > case, and this should fix it (I didn't test it):
> >
> > - #define ARCH_LOW_ADDRESS_LIMIT (arm64_dma_phys_limit - 1)
> > + #define ARCH_LOW_ADDRESS_LIMIT (arm64_dma_phys_limit ? : arm64_dma32_phys_limit)
>
> Does this limit need to be inclusive?
Yes, I'm missing a '- 1'.
I'll send a fix if there is no further push back/comments.
Regards,
Nicolas