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

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

 




On 2022/4/28 11:52, Baoquan He wrote:
> On 04/28/22 at 11:40am, Baoquan He wrote:
>> Hi Catalin, Zhen Lei,
>>
>> On 04/27/22 at 05:04pm, Catalin Marinas wrote:
>>> On Wed, Apr 27, 2022 at 09:49:20PM +0800, Leizhen (ThunderTown) wrote:
>>>> On 2022/4/27 20:32, Catalin Marinas wrote:
>>>>> 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.
>>>
>>> There will be some difference as the 4G limit doesn't always hold for
>>> arm64 (though it's true in most cases). Anyway, we can probably simplify
>>> things a bit while following the documented behaviour:
>>>
>>> 	crashkernel=Y		- current behaviour within ZONE_DMA
>>> 	crashkernel=Y,high	- allocate from above ZONE_DMA
>>> 	crashkernel=Y,low	- allocate within ZONE_DMA
>>>
>>> There is no fallback from crashkernel=Y.
>>>
>>> The question is whether we still want a default low allocation if
>>> crashkernel=Y,low is missing but 'high' is present. If we add this, I
>>> think we'd be consistent with kernel-parameters.txt for the 'low'
>>> description. A default 'low' is probably not that bad but I'm tempted to
>>> always mandate both 'high' and 'low'.
>>
>> Sorry to interrupt. Seems the ,high ,low and fallback are main concerns
>> about this version. And I have the same concerns about them which comes
>> from below points:
>> 1) we may need to take best effort to keep ,high, ,low behaviour
>> consistent on all ARCHes. Otherwise user/admin may be confused when they
>> deploy/configure kdump on different machines of different ARCHes in the
>> same LAB. I think we should try to avoid the confusion.

Yes, but for someone who is configuring crashkernel= for the first time, he
needs to read doc to understand how to configure it. The doc can show the
recommended default value of 'low' size.

After commit 94fb93341822 ("x86/crash: Allocate enough low memory when crashkernel=high"),
the default 'low' size doesn't make much sense anymore. The default size of swiotlb_size()
is 64M, far less than 256M. And if user specify "swiotlb=", he can also adjust crashkernel=Y,low.


+                * -swiotlb size: user-specified with swiotlb= or default.
-               low_size = swiotlb_size_or_default() + (8UL<<20);
+               low_size = max(swiotlb_size_or_default() + (8UL<<20), 256UL<<20);

That means all ARCHs can explicit configure crashkernel=256M,low, instead of
omitting it. This may be another way to avoid confusion. It's not hard for
programmer-turned-user/admin. However, this requires us to forgo backward
compatibility with the default size of 'low'.


>> 2) Fallback behaviour is important to our distros. The reason is we will
>> provide default value with crashkernel=xxxM along kernel of distros. In
>> this case, we hope the reservation will succeed by all means. The ,high
>> and ,low is an option if customer likes to take with expertise.

OK, I got it.

>>
>> After going through arm64 memory init code, I got below summary about
>> arm64_dma_phys_limit which is the first zone's upper limit. I think we
>> can make use of it to facilitate to simplify code.
>> ================================================================================
>>                         DMA                      DMA32                    NORMAL
>> 1)Raspberry Pi4         0~1G                     3G~4G                    (above 4G)
>> 2)Normal machine        0~4G                     0                        (above 4G)
>> 3)Special machine       (above 4G)~MAX
>> 4)No DMA|DMA32                                                            (above 4G)~MAX

arm64_memblock_init()
	reserve_crashkernel()        ---------------   0a30c53573b0 ("arm64: mm: Move reserve_crashkernel() into mem_init()")
paging_init()                                       |
	map_mem()                                   |
unflatten_device_tree or ACPI                       |  ----  //Raspberry Pi4 get dma zone base on dtb or ACPI
bootmem_init();                                     |      |
	zone_sizes_init()                           |      |
		of_dma_get_max_cpu_address          |  ----|
		//Update arm64_dma_phys_limit       |  ----|
	reserve_crashkernel()        <--------------  //Because we need arm64_dma_phys_limit to be updated above
request_standard_resources()

>>
>> -------------------------------------------
>>                       arm64_dma_phys_limit
>> 1)Raspberry Pi4         1G                     
>> 2)Normal machine        4G                     
>> 3)Special machine       MAX
>> 4)No DMA|DMA32          MAX
>>
>> Note: 3)Special machine means the machine's starting physical address is above 4G.
>> WHile 4)No DMA|DMA32 means kernel w/o CONFIG_ZONE_DMA|DMA32, and has
>> IOMMU hardware supporting.
>> ===================================================================================
>>
>> I made a draft patch based on this patchset, please feel free to check and
>> see if it's OK, or anything missing or wrongly understood. I removed
>> reserve_crashkernel_high() and only keep reserve_crashkernel() and
>> reserve_crashkernel_low() as the v21 did.
> 
> Sorry, forgot attaching the draft patch.
> 
> By the way, we can also have a simple version with basic ,high, ,low
> support, no fallback. We can add fallback and other optimization later.
> This can be plan B.

Yes, That's what Catalin suggested also.

Hi, Baoquan He:
  Without optimization, the whole Patch 3-4 and 6-7 can be dropped.

Process after abstraction:
	if (!IS_ENABLED(CONFIG_ZONE_DMA) && !IS_ENABLED(CONFIG_ZONE_DMA32)) {
		reserve_crashkernel()
		//block mapping
	} else {
		//page mapping
		reserve_crashkernel()
	}

------------ Simplified real-world process ---------
arm64_memblock_init()
	if (!IS_ENABLED(CONFIG_ZONE_DMA) && !IS_ENABLED(CONFIG_ZONE_DMA32))
		reserve_crashkernel()
paging_init()
	map_mem()
		if (!IS_ENABLED(CONFIG_ZONE_DMA) && !IS_ENABLED(CONFIG_ZONE_DMA32))
			//block mapping
		else
			//page mapping
unflatten_device_tree or ACPI
bootmem_init();
	zone_sizes_init()
		of_dma_get_max_cpu_address
		//Update arm64_dma_phys_limit
	if (IS_ENABLED(CONFIG_ZONE_DMA) || IS_ENABLED(CONFIG_ZONE_DMA32))
		reserve_crashkernel()


> 
> 

-- 
Regards,
  Zhen Lei



[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux