lkml.org 
[lkml]   [2020]   [Jul]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v10 4/5] arm64: kdump: fix kdump broken with ZONE_DMA reintroduced
From
Date
Hi Catalin,


On 2020/7/29 23:20, Catalin Marinas wrote:
> On Wed, Jul 29, 2020 at 10:14:32PM +0800, chenzhou wrote:
>> On 2020/7/29 19:58, Catalin Marinas wrote:
>>> On Wed, Jul 29, 2020 at 11:52:39AM +0800, chenzhou wrote:
>>>> 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.
>> It is a good idea to combine the two.
>>
>> For parameter crashkernel=X, we do like this:
>> 1. allocate some low memory in ZONE_DMA(or ZONE_DMA32 if CONFIG_ZONE_DMA=n)
>> 2. allocate X size memory in a high region
>>
>> ",low" argument can be used to specify the low memory.
>>
>> Do i understand correctly?
> Yes, although we could follow the x86 approach:
>
> 1. Try low (ZONE_DMA for arm64) allocation, fallback to high allocation
> if it fails.
>
> 2. If crash_base is outside ZONE_DMA, call reserve_crashkernel_low()
> which either honours the ,low option or allocates some small amount
> in ZONE_DMA.
>
> If at some point we have platforms failing step 2, we'll look at
> changing ZONE_DMA to the full 4GB on non-RPi4 platforms.
>
> It looks to me like x86 ignores the ,low option if the first step
> managed to get some low memory. Shall we do the same on arm64?
Yes, we could do like this.
>
>>> 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).
>> Yes. We can let crashkernel=X try to reserve low memory and fall back to use high memory
>> if failing to find a low range.
> The only question is whether we need to preserve some more ZONE_DMA on
> the current system. If for example we pass a crashkernel=512M and some
> cma=, we may end up with very little free memory in ZONE_DMA. That's
> mostly an issue for RPi4 since other platforms would work with
> ZONE_DMA32. We could add a threshold and go for high allocation directly
> if the required size is too large.
Ok. I will think about the threshold in the next version and make the value be 1/2 or 1/3 of the ZONE_DMA.
>
>> About the function reserve_crashkernel_low(), if we put it in arch/arm64, there is some common
>> code with x86_64. Some suggestions about this?
> If we can use this function almost intact, just move it in a common
> place. But if it gets sprinkled with #ifdef CONFIG_ARM64, I'd rather
> duplicate it. I'd still prefer to move it to a common place if possible.
>
> You can go a step further and also move the x86 reserve_crashkernel() to
> common code. I don't think there a significant difference between arm64
> and x86 here. You'd have to define arch-specific specific
> CRASH_ADDR_LOW_MAX etc.
I will take these into account and send the next version recently.
>
> Also patches moving code should not have any functional change. The
> CRASH_ALIGN change from 16M to 2M on x86 should be a separate patch as
> it needs to be acked by the x86 maintainers (IIRC, Ingo only acked the
> function move if there was no functional change; CRASH_ALIGN is used for
> the start address, not just alignment, on x86).
>
Thanks,
Chen Zhou

\
 
 \ /
  Last update: 2020-07-30 10:23    [W:0.043 / U:3.580 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site