[PATCH v10 4/5] arm64: kdump: fix kdump broken with ZONE_DMA reintroduced

[Catalin Marinas]

Hi Chen,

On Wed, Jul 29, 2020 at 11:52:39AM +0800, chenzhou wrote:
> On 2020/7/28 1:30, Catalin Marinas wrote:
> > Anyway, there are two series solving slightly different issues with
> > kdump reservations:
> >
> > 1. This series which relaxes the crashkernel= allocation to go anywhere
> >    in the accessible space while having a dedicated crashkernel=X,low
> >    option for ZONE_DMA.
> >
> > 2. Bhupesh's series [1] forcing crashkernel=X allocations only from
> >    ZONE_DMA.
> >
> > For RPi4 support, we limited ZONE_DMA allocations to the 1st GB.
> > Existing crashkernel= uses may no longer work, depending on where the
> > allocation falls. Option (2) above is a quick fix assuming that the
> > crashkernel reservation is small enough. What's a typical crashkernel
> > option here? That series is probably more prone to reservation failures.
> >
> > Option (1), i.e. this series, doesn't solve the problem raised by
> > Bhupesh unless one uses the crashkernel=X,low argument. It can actually
> > make it worse even for ZONE_DMA32 since the allocation can go above 4G
> > (assuming that we change the ZONE_DMA configuration to only limit it to
> > 1GB on RPi4).
> >
> > I'm more inclined to keep the crashkernel= behaviour to ZONE_DMA
> > allocations. If this is too small for typical kdump, we can look into
> > expanding ZONE_DMA to 4G on non-RPi4 hardware (we had patches on the
> > list). In addition, if Chen thinks allocations above 4G are still needed
> > or if RPi4 needs a sufficiently large crashkernel=, I'd rather have a
> > ",high" option to explicitly require such access.
> 
> Thanks for your reply and exhaustive explanation.
> 
> In our ARM servers, we need to to reserve a large chunk for kdump(512M
> or 1G), there is no enough low memory. So we proposed this patch
> series "support reserving crashkernel above 4G on arm64 kdump" In
> April 2019.

Trying to go through the discussions last year, hopefully things get
clearer.

So prior to the ZONE_DMA change, you still couldn't reserve 1G in the
first 4GB? It shouldn't be sparsely populated during early boot.

> I introduce parameters "crashkernel=X,[high,low]" as x86_64 does in earlier versions.
> Suggested by James, to simplify, we call reserve_crashkernel_low() at the beginning of
> reserve_crashkernel() and then relax the arm64_dma32_phys_limit if reserve_crashkernel_low()
> allocated something.
> That is, just the parameter "crashkernel=X,low" is ok and i deleted "crashkernel=X,high".

The problem I see is that with your patches we diverge from x86
behaviour (and the arm64 behaviour prior to the ZONE_DMA reduction) as
we now require that crashkernel=X,low is always passed if you want
something in ZONE_DMA (and you do want, otherwise the crashdump kernel
fails to boot).

My main requirement is that crashkernel=X, without any suffix, still
works which I don't think is guaranteed with your patches (well,
ignoring RPi4 ZONE_DMA). Bhupesh's series is a quick fix but doesn't
solve your large allocation requirements (that may have worked prior to
the ZONE_DMA change).

> After the ZONE_DMA introduced in December 2019, the issue occurred as
> you said above. In fact, we didn't have RPi4 machine.

You don't even need to have a RPi4 machine, ZONE_DMA has been set to 1GB
unconditionally. And while we could move it back to 4GB on non-RPi4
hardware, I'd rather have a solution that fixes kdump for RPi4 as well.

> Originally, i suggested to fix this based on this patch series and
> used the dedicated option.
> 
> According to your clarify, for typical kdump, there are other
> solutions. In this case, "keep the crashkernel= behaviour to ZONE_DMA
> allocations" looks much better.
> 
> How about like this:
> 1. For ZONE_DMA issue, use Bhupesh's solution, keep the crashkernel=
>    behaviour to ZONE_DMA allocations.
> 2. For this patch series, make the reserve_crashkernel_low() to
>    ZONE_DMA allocations.

So you mean rebasing your series on top of Bhupesh's? I guess you can
combine the two, I really don't care which way as long as we fix both
issues and agree on the crashkernel= semantics. I think with some tweaks
we can go with your series alone.

IIUC from the x86 code (especially the part you #ifdef'ed out for
arm64), if ",low" is not passed (so just standard crashkernel=X), it
still allocates sufficient low memory for the swiotlb in ZONE_DMA. The
rest can go in a high region. Why can't we do something similar on
arm64? Of course, you can keep the ",low" argument for explicit
allocation but I don't want to mandate it.

So with an implicit ZONE_DMA allocation similar to the x86 one, we
probably don't need Bhupesh's series at all. In addition, we can limit
crashkernel= to the first 4G with a fall-back to high like x86 (not sure
if memblock_find_in_range() is guaranteed to search in ascending order).
I don't think we need an explicit ",high" annotation.

So with the above, just a crashkernel=1G gives you at least 256MB in
ZONE_DMA followed by the rest anywhere, with a preference for
ZONE_DMA32. This way we can also keep the reserve_crashkernel_low()
mostly intact from x86 (less #ifdef's).

Do I miss anything?

-- 
Catalin