Re: [PATCH 3/3 v3] x86/kdump: clean up all the code related to the backup region

[Eric W. Biederman]

lijiang <lijiang@xxxxxxxxxx> writes:

> å 2019å10æ13æ 11:54, Eric W. Biederman åé:
>> Dave Young <dyoung@xxxxxxxxxx> writes:
>> 
>>> Hi Eric,
>>>
>>> On 10/12/19 at 06:26am, Eric W. Biederman wrote:
>>>> Lianbo Jiang <lijiang@xxxxxxxxxx> writes:
>>>>
>>>>> When the crashkernel kernel command line option is specified, the
>>>>> low 1MiB memory will always be reserved, which makes that the memory
>>>>> allocated later won't fall into the low 1MiB area, thereby, it's not
>>>>> necessary to create a backup region and also no need to copy the first
>>>>> 640k content to a backup region.
>>>>>
>>>>> Currently, the code related to the backup region can be safely removed,
>>>>> so lets clean up.
>>>>>
>>>>> Signed-off-by: Lianbo Jiang <lijiang@xxxxxxxxxx>
>>>>> ---
>>>>
>>>>> diff --git a/arch/x86/kernel/crash.c b/arch/x86/kernel/crash.c
>>>>> index eb651fbde92a..cc5774fc84c0 100644
>>>>> --- a/arch/x86/kernel/crash.c
>>>>> +++ b/arch/x86/kernel/crash.c
>>>>> @@ -173,8 +173,6 @@ void native_machine_crash_shutdown(struct pt_regs *regs)
>>>>>  
>>>>>  #ifdef CONFIG_KEXEC_FILE
>>>>>  
>>>>> -static unsigned long crash_zero_bytes;
>>>>> -
>>>>>  static int get_nr_ram_ranges_callback(struct resource *res, void *arg)
>>>>>  {
>>>>>  	unsigned int *nr_ranges = arg;
>>>>> @@ -234,9 +232,15 @@ static int prepare_elf64_ram_headers_callback(struct resource *res, void *arg)
>>>>>  {
>>>>>  	struct crash_mem *cmem = arg;
>>>>>  
>>>>> -	cmem->ranges[cmem->nr_ranges].start = res->start;
>>>>> -	cmem->ranges[cmem->nr_ranges].end = res->end;
>>>>> -	cmem->nr_ranges++;
>>>>> +	if (res->start >= SZ_1M) {
>>>>> +		cmem->ranges[cmem->nr_ranges].start = res->start;
>>>>> +		cmem->ranges[cmem->nr_ranges].end = res->end;
>>>>> +		cmem->nr_ranges++;
>>>>> +	} else if (res->end > SZ_1M) {
>>>>> +		cmem->ranges[cmem->nr_ranges].start = SZ_1M;
>>>>> +		cmem->ranges[cmem->nr_ranges].end = res->end;
>>>>> +		cmem->nr_ranges++;
>>>>> +	}
>>>>
>>>> What is going on with this chunk?  I can guess but this needs a clear
>>>> comment.
>>>
>>> Indeed it needs some code comment, this is based on some offline
>>> discussion.  cat /proc/vmcore will give a warning because ioremap is
>>> mapping the system ram.
>>>
>>> We pass the first 1M to kdump kernel in e820 as system ram so that 2nd
>>> kernel can use the low 1M memory because for example the trampoline
>>> code.
>>>
>>>>
>>>>>  
>>>>>  	return 0;
>>>>>  }
>>>>
>>>>> @@ -356,9 +337,12 @@ int crash_setup_memmap_entries(struct kimage *image, struct boot_params *params)
>>>>>  	memset(&cmd, 0, sizeof(struct crash_memmap_data));
>>>>>  	cmd.params = params;
>>>>>  
>>>>> -	/* Add first 640K segment */
>>>>> -	ei.addr = image->arch.backup_src_start;
>>>>> -	ei.size = image->arch.backup_src_sz;
>>>>> +	/*
>>>>> +	 * Add the low memory range[0x1000, SZ_1M], skip
>>>>> +	 * the first zero page.
>>>>> +	 */
>>>>> +	ei.addr = PAGE_SIZE;
>>>>> +	ei.size = SZ_1M - PAGE_SIZE;
>>>>>  	ei.type = E820_TYPE_RAM;
>>>>>  	add_e820_entry(params, &ei);
>>>>
>>>> Likewise here.  Why do we need a special case?
>>>> Why the magic with PAGE_SIZE?
>>>
>>> Good catch, the zero page part is useless, I think no other special
>>> reason, just assumed zero page is not usable, but it should be ok to
>>> remove the special handling, just pass 0 - 1M is good enough.
>> 
>> But if we have stopped special casing the low 1M.  Why do we need a
>> special case here at all?
>> 
> Here, need to pass the low memory range to kdump kernel, which will guarantee
> the availability of low memory in kdump kernel, otherwise, kdump kernel won't
> use the low memory region.
>
>> If you need the special case it is almost certainly wrong to say you
>> have ram above 640KiB and below 1MiB.  That is the legacy ROM and video
>> MMIO area.
>> 
>> There is a reason the original code said 640KiB.
>> 
> Do you mean that the 640k region is good enough here instead of 1MiB?

Reading through the code of crash_setup_memap_entries I see that what
the code is doing now.  The code is repeating the e820 memory map with
the memory areas that were not reserved for the crash kernel removed.

In which case what the code needs to be doing something like:

	cmd.type = E820_TYPE_RAM;
	flags = IORESOURCE_MEM;
	walk_iomem_res_desc(IORES_DESC_RESERVED, flags, 0, 1024*1024, &cmd,
			       memmap_entry_callback);

Depending on which bugs exist it might make sense to limit this to
the low 640KiB.  But finding something the kernel already recognizes
as RAM should prevent most of those problems already.  Barring bugs
I admit it doesn't make sense to repeat the work that someone else
has already done.

This bit:
	/* Add e820 reserved ranges */
	cmd.type = E820_TYPE_RESERVED;
	flags = IORESOURCE_MEM;
	walk_iomem_res_desc(IORES_DESC_RESERVED, flags, 0, -1, &cmd,
			   memmap_entry_callback);

Should probably start at 1MiB instead of 0.  Just so we don't report the
memory below 1MiB as unconditionally reserved.   I don't properly
understand the IORES_DESC_RESERVED flag, and how that differs from
flags.  So please test my suggestions to verify the code works as
expected.

Eric