[PATCH v4] arm64: mm: fix linear mem mapping access performance degradation

[guanghui.fgh]

在 2022/7/5 20:56, Mike Rapoport 写道:
> On Tue, Jul 05, 2022 at 01:11:16PM +0100, Will Deacon wrote:
>> On Tue, Jul 05, 2022 at 08:07:07PM +0800, guanghui.fgh wrote:
>>>
>>> 1.The rodata full is harm to the performance and has been disabled in-house.
>>>
>>> 2.When using crashkernel with rodata non full, the kernel also will use non
>>> block/section mapping which cause high d-TLB miss and degrade performance
>>> greatly.
>>> This patch fix it to use block/section mapping as far as possible.
>>>
>>> bool can_set_direct_map(void)
>>> {
>>> 	return rodata_full || debug_pagealloc_enabled();
>>> }
>>>
>>> map_mem:
>>> if (can_set_direct_map() || IS_ENABLED(CONFIG_KFENCE))
>>> 	flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
>>>
>>> 3.When rodata full is disabled, crashkernel also need protect(keep
>>> arch_kexec_[un]protect_crashkres using).
>>> I think crashkernel should't depend on radata full(Maybe other architecture
>>> don't support radata full now).
>>
>> I think this is going round in circles :/
>>
>> As a first step, can we please leave the crashkernel mapped unless
>> rodata=full? It should be a much simpler patch to write, review and maintain
>> and it gives you the performance you want when crashkernel is being used.
> 
> Since we are talking about large systems, what do you think about letting
> them set CRASH_ALIGN to PUD_SIZE, then
> 
> 	unmap(crashkernel);
> 	__create_pgd_mapping(crashkernel, NO_BLOCK_MAPPINGS);
> 
> should be enough to make crash kernel mapped with base pages.
>   
>> Will
> 
Thanks.

1.When kernel boots with crashkernel, the crashkernel parameters format is:
0M-2G:0M,2G-256G:256M,256G-1024G:320M,1024G-:384M which is self-adaption 
to multiple system.

2.As mentioned above, the crashkernel mem size maybe less than 
PUD_SIZE(Not multiple time to PUD_SIZE).
So, maybe we also need use some non block/section mappings.