Re: [PATCH v22 5/9] arm64: kdump: Reimplement crashkernel=X

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




On 2022/4/27 20:32, Catalin Marinas wrote:
> On Wed, Apr 27, 2022 at 02:54:52PM +0800, Leizhen (ThunderTown) wrote:
>> On 2022/4/27 2:02, Catalin Marinas wrote:
>>> On Thu, Apr 14, 2022 at 07:57:16PM +0800, Zhen Lei wrote:
>>>>  /*
>>>>   * reserve_crashkernel() - reserves memory for crash kernel
>>>>   *
>>>>   * This function reserves memory area given in "crashkernel=" kernel command
>>>>   * line parameter. The memory reserved is used by dump capture kernel when
>>>>   * primary kernel is crashing.
>>>> + *
>>>> + * NOTE: Reservation of crashkernel,low is special since its existence
>>>> + * is not independent, need rely on the existence of crashkernel,high.
>>>> + * Here, four cases of crashkernel low memory reservation are summarized:
>>>> + * 1) crashkernel=Y,low is specified explicitly, the size of crashkernel low
>>>> + *    memory takes Y;
>>>> + * 2) crashkernel=,low is not given, while crashkernel=,high is specified,
>>>> + *    take the default crashkernel low memory size;
>>>> + * 3) crashkernel=X is specified, while fallback to get a memory region
>>>> + *    in high memory, take the default crashkernel low memory size;
>>>> + * 4) crashkernel='invalid value',low is specified, failed the whole
>>>> + *    crashkernel reservation and bail out.
>>>
>>> Following the x86 behaviour made sense when we were tried to get that
>>> code generic. Now that we moved the logic under arch/arm64, we can
>>> diverge a bit. I lost track of the original (v1/v2) proposal but I
>>> wonder whether we still need the fallback to high for crashkernel=Y.
>>
>> I don't think anyone has raised this demand yet! If it weren't for the
>> fact that crashkernel=X appeared earlier, it would probably have been
>> enough for a combination of crashkernel=X,high and crashkernel=Y,low.
>>
>> In fact, I also tend not to support "fallback to high for crashkernel=Y".
>> I took over this from Chen Zhou. In the absence of any objection, I had
>> to inherit. Now that you've brought it up, I'm happy to delete it.
>> Supporting this feature complicates the code logic a lot. The point is,
>> it's not fully backwards compatible yet. For example, someone may want
>> crashkernel=3G to report failure, but the the new support make it work.
> 
> BTW, prior to v20, this patch had this line:
> 
> 	crashk_low_res.name = "Crash kernel (low)";
> 
> I can't find it anymore. Do the kexec tools need to distinguish between
> low and high or they can cope with multiple "Crash kernel" entries?

Yes, I've updated the kexec tool patch based on Borislav Petkov's comments
to keep it the same as x86. And this patch has been merged into kexec v2.0.24.
b5a34a2 arm64: support more than one crash kernel regions

The kexec tool will first sorted all "Crash kernel" memory range in ascending
order, then choose the last one (the memory range with the highest address) to
store the second kernel's Image,dtb,initrd.

> 
>>> Maybe simpler, no fallbacks:
>>>
>>> 	crashkernel=Y - keep the current behaviour, ignore high,low
>>> 	crashkernel=Y,high - allocate above ZONE_DMA
>>> 	crashkernel=Y,low - allocate within ZONE_DMA
>>>
>>> From your proposal, the difference is that the Y,high option won't
>>> have any default ZONE_DMA fallback, one would have to explicitly pass
>>> the Y,low option if needed.
>>
>> I agree with you. Now we don't need code generic, so there is no need to
>> carry the historical burden of other ARCHs. arm64 does not need to delve
>> into that empirical value(the default size of crash low memory).
>>
>>> Just a thought, maybe it makes the code simpler. But I'm open to
>>> discussion if there are good arguments for the proposed (x86-like)
>>> behaviour. One argument could be for crashkernel=Y to fall back to high
>>> if distros don't want to bother with high/low settings.
>>
>> I think distros should take precedence over "crashkernel=Y,high". After all,
>> ZONE_DMA memory is more valuable than high memory.
> 
> My point is whether an admin configuring the kernel command line needs
> to know the layout of ZONE_DMA etc. to figure out how much to pass in

No need.

> high and low. The fallbacks in this case have some value but they also
> complicate the code logic. The 4GB limit does not always make sense
> either for some platforms (RPi4 has a ZONE_DMA limit of 1GB).
> 
> I think one could always pass a default command line like:
> 
> 	crashkernel=1G,high crashkernel=128M,low
> 
> without much knowledge of the SoC memory layout.

Yes, that's what the end result is. The user specify crashkernel=128M,low
and the implementation ensure the 128M low memory is allocated from DMA zone.
We use arm64_dma_phys_limit as the upper limit for crash low memory.

+#define CRASH_ADDR_LOW_MAX             arm64_dma_phys_limit
+       unsigned long long crash_max = CRASH_ADDR_LOW_MAX;
+       crash_base = memblock_phys_alloc_range(crash_size, CRASH_ALIGN,
                                               crash_base, crash_max);

> 
> Another option is to only introduce crashkernel=Y,low and, when that is
> passed, crashkernel=Y can go above arm64_dma_phys_limit. We won't need a
> 'high' option at all:
> 
> 	crashkernel=1G				- all within ZONE_DMA
> 	crashkernel=1G crashkernel=128M,low	- 128M in ZONE_DMA
> 						  1G above ZONE_DMA
> 
> If ZONE_DMA is not present or it extends to the whole RAM, we can ignore
> the 'low' option.

I think although the code is hard to make generic, the interface is better to
be relatively uniform. A user might have to maintain both x86 and arm64, and
so on. It's not a good thing that the difference is too big.

A well had already been dug by the forefathers, and it seemed unnecessary to
dig another well beside it if there was no particular advantage.

> 

-- 
Regards,
  Zhen Lei

_______________________________________________
kexec mailing list
kexec@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/kexec



[Index of Archives]     [LM Sensors]     [Linux Sound]     [ALSA Users]     [ALSA Devel]     [Linux Audio Users]     [Linux Media]     [Kernel]     [Gimp]     [Yosemite News]     [Linux Media]

  Powered by Linux