On 2020/9/4 12:16, Dave Young wrote:
> On 09/04/20 at 12:02pm, chenzhou wrote:
>>
>> On 2020/9/4 11:10, Dave Young wrote:
>>> On 09/04/20 at 11:04am, Dave Young wrote:
>>>> On 09/03/20 at 07:26pm, chenzhou wrote:
>>>>> Hi Catalin,
>>>>>
>>>>>
>>>>> On 2020/9/3 1:09, Catalin Marinas wrote:
>>>>>> On Sat, Aug 01, 2020 at 09:08:54PM +0800, Chen Zhou wrote:
>>>>>>> There are following issues in arm64 kdump:
>>>>>>> 1. We use crashkernel=X to reserve crashkernel below 4G, which
>>>>>>> will fail when there is no enough low memory.
>>>>>>> 2. If reserving crashkernel above 4G, in this case, crash dump
>>>>>>> kernel will boot failure because there is no low memory available
>>>>>>> for allocation.
>>>>>>> 3. Since commit 1a8e1cef7603 ("arm64: use both ZONE_DMA and ZONE_DMA32"),
>>>>>>> if the memory reserved for crash dump kernel falled in ZONE_DMA32,
>>>>>>> the devices in crash dump kernel need to use ZONE_DMA will alloc
>>>>>>> fail.
>>>>>>>
>>>>>>> To solve these issues, change the behavior of crashkernel=X.
>>>>>>> crashkernel=X tries low allocation in ZONE_DMA, and fall back to
>>>>>>> high allocation if it fails.
>>>>>>>
>>>>>>> If requized size X is too large and leads to very little free memory
>>>>>>> in ZONE_DMA after low allocation, the system may not work normally.
>>>>>>> So add a threshold and go for high allocation directly if the required
>>>>>>> size is too large. The value of threshold is set as the half of
>>>>>>> the low memory.
>>>>>>>
>>>>>>> If crash_base is outside ZONE_DMA, try to allocate at least 256M in
>>>>>>> ZONE_DMA automatically. "crashkernel=Y,low" can be used to allocate
>>>>>>> specified size low memory.
>>>>>> Except for the threshold to keep zone ZONE_DMA memory,
>>>>>> reserve_crashkernel() looks very close to the x86 version. Shall we try
>>>>>> to make this generic as well? In the first instance, you could avoid the
>>>>>> threshold check if it takes an explicit ",high" option.
>>>>> Ok, i will try to do this.
>>>>>
>>>>> I look into the function reserve_crashkernel() of x86 and found the start address is
>>>>> CRASH_ALIGN in function memblock_find_in_range(), which is different with arm64.
>>>>>
>>>>> I don't figure out why is CRASH_ALIGN in x86, is there any specific reason?
>>>> Hmm, took another look at the option CONFIG_PHYSICAL_ALIGN
>>>> config PHYSICAL_ALIGN
>>>> hex "Alignment value to which kernel should be aligned"
>>>> default "0x200000"
>>>> range 0x2000 0x1000000 if X86_32
>>>> range 0x200000 0x1000000 if X86_64
>>>>
>>>> According to above, I think the 16M should come from the largest value
>>>> But the default value is 2M, with smaller value reservation can have
>>>> more chance to succeed.
>>>>
>>>> It seems we still need arch specific CRASH_ALIGN, but the initial
>>>> version you added the #ifdef for different arches, can you move the
>>>> macro to arch specific headers?
>>> And just keep the x86 align value as is, I can try to change the x86
>>> value later to CONFIG_PHYSICAL_ALIGN, in this way this series can be
>>> cleaner.
>> Ok. I have no question about the value of macro CRASH_ALIGN,
>> instead the lower bound of memblock_find_in_range().
>>
>> For x86, in reserve_crashkernel()，restrict the lower bound of the range to CRASH_ALIGN,
>> ...
>> crash_base = memblock_find_in_range(CRASH_ALIGN,
>> CRASH_ADDR_LOW_MAX,
>> crash_size, CRASH_ALIGN);
>> ...
>>
>> in reserve_crashkernel_low()，with no this restriction.
>> ...
>> low_base = memblock_find_in_range(0, 1ULL << 32, low_size, CRASH_ALIGN);
>> ...
>>
>> How about all making memblock_find_in_range() search from the start of memory?
>> If it is ok, i will do like this in the generic version.
> I feel starting with CRASH_ALIGN sounds better, can you just search from
> CRASH_ALIGN in generic version?
ok.
>
> Thanks
> Dave
>
>
> .
>
