Re: [PATCH v9 07/11] arm64: kexec_file: add crash dump support

[James Morse]

Hi Akashi,

On 18/05/18 11:39, AKASHI Takahiro wrote:
> On Tue, May 15, 2018 at 06:11:15PM +0100, James Morse wrote:
>> On 25/04/18 07:26, AKASHI Takahiro wrote:
>>> Enabling crash dump (kdump) includes
>>> * prepare contents of ELF header of a core dump file, /proc/vmcore,
>>>   using crash_prepare_elf64_headers(), and
>>> * add two device tree properties, "linux,usable-memory-range" and
>>>   "linux,elfcorehdr", which represent repsectively a memory range

>>> diff --git a/arch/arm64/kernel/machine_kexec_file.c b/arch/arm64/kernel/machine_kexec_file.c
>>> index 37c0a9dc2e47..ec674f4d267c 100644
>>> --- a/arch/arm64/kernel/machine_kexec_file.c
>>> +++ b/arch/arm64/kernel/machine_kexec_file.c

>>> +static void fill_property(void *buf, u64 val64, int cells)
>>> +{
>>> +	u32 val32;
>>> +
>>> +	if (cells == 1) {
>>> +		val32 = cpu_to_fdt32((u32)val64);
>>> +		memcpy(buf, &val32, sizeof(val32));
>>> +	} else {
>>
>>> +		memset(buf, 0, cells * sizeof(u32) - sizeof(u64));
>>> +		buf += cells * sizeof(u32) - sizeof(u64);
>>
>> Is this trying to clear the 'top' cells and shuffle the pointer to point at the
>> 'bottom' 2? I'm pretty sure this isn't endian safe.
>>
>> Do we really expect a system to have #address-cells > 2?
> 
> I don't know, but just for safety.

Okay, so this is aiming to be a cover-all-cases library function.


>>> +		val64 = cpu_to_fdt64(val64);
>>> +		memcpy(buf, &val64, sizeof(val64));
>>> +	}
>>> +}
>>> +
>>> +static int fdt_setprop_range(void *fdt, int nodeoffset, const char *name,
>>> +				unsigned long addr, unsigned long size)
>>
>> (the device-tree spec describes a 'ranges' property, which had me confused. This
>> is encoding a prop-encoded-array)
> 
> Should we rename it to, say, fdt_setprop_reg()?

Sure, but I'd really like this code to come from libfdt. I'm hoping for some
temporary workaround, lets see what the DT folk say.


>>> +	if (!buf)
>>> +		return -ENOMEM;
>>> +
>>> +	fill_property(prop, addr, __dt_root_addr_cells);
>>> +	prop += __dt_root_addr_cells * sizeof(u32);
>>> +
>>> +	fill_property(prop, size, __dt_root_size_cells);
>>> +
>>> +	result = fdt_setprop(fdt, nodeoffset, name, buf, buf_size);
>>> +
>>> +	vfree(buf);
>>> +
>>> +	return result;
>>> +}
>>
>> Doesn't this stuff belong in libfdt? I guess there is no 'add array element' api
>> because this the first time we've wanted to create a node with more than
>> key=fixed-size-value.
>>
>> I don't think this belongs in arch C code. Do we have a plan for getting libfdt
>> to support encoding prop-arrays? Can we put it somewhere anyone else duplicating
>> this will find it, until we can (re)move it?
> 
> I will temporarily move all fdt-related stuff to a separate file, but
> 
>> I have no idea how that happens... it looks like the devicetree list is the
>> place to ask.
> 
> should we always sync with the original dtc/libfdt repository?

I thought so, libfdt is one of those external libraries that the kernel
consumes, like acpica. For acpica at least the rule is changes go upstream, then
get sync'd back.


>>>  static int setup_dtb(struct kimage *image,
>>>  		unsigned long initrd_load_addr, unsigned long initrd_len,
>>>  		char *cmdline, unsigned long cmdline_len,
>>> @@ -88,10 +165,26 @@ static int setup_dtb(struct kimage *image,
>>>  	int range_len;
>>>  	int ret;
>>>  
>>> +	/* check ranges against root's #address-cells and #size-cells */
>>> +	if (image->type == KEXEC_TYPE_CRASH &&
>>> +		(!cells_size_fitted(image->arch.elf_load_addr,
>>> +				image->arch.elf_headers_sz) ||
>>> +		 !cells_size_fitted(crashk_res.start,
>>> +				crashk_res.end - crashk_res.start + 1))) {
>>> +		pr_err("Crash memory region doesn't fit into DT's root cell sizes.\n");
>>> +		ret = -EINVAL;
>>> +		goto out_err;
>>> +	}
>>
>> To check I've understood this properly: This can happen if the firmware provided
>> a DTB with 32bit address/size cells, but at least some of the memory requires 64
>> bit address/size cells. This could only happen on a UEFI system where the
>> firmware-DTB doesn't describe memory. ACPI-only systems would have the EFIstub DT.
> 
> Probably, yes. I assumed the case where #address-cells and #size-cells
> were just missing in fdt.

Ah, that's another one. I just wanted to check we could boot on a system where
this can happen.


>>>  	/* duplicate dt blob */
>>>  	buf_size = fdt_totalsize(initial_boot_params);
>>>  	range_len = (__dt_root_addr_cells + __dt_root_size_cells) * sizeof(u32);
>>>  
>>> +	if (image->type == KEXEC_TYPE_CRASH)
>>> +		buf_size += fdt_prop_len("linux,elfcorehdr", range_len)
>>> +				+ fdt_prop_len("linux,usable-memory-range",
>>> +								range_len);

>                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[...]

>> Don't you need to add "linux,usable-memory-range" to the buf_size estimate?
> 
> I think the code exists. See above.

Sorry, turns out I can't read!


>>> +		if (ret)
>>> +			goto out_err;
>>> +	}
>>
>>> @@ -148,17 +258,109 @@ static int setup_dtb(struct kimage *image,
>>
>>> +static struct crash_mem *get_crash_memory_ranges(void)
>>> +{
>>> +	unsigned int nr_ranges;
>>> +	struct crash_mem *cmem;
>>> +
>>> +	nr_ranges = 1; /* for exclusion of crashkernel region */
>>> +	walk_system_ram_res(0, -1, &nr_ranges, get_nr_ranges_callback);
>>> +
>>> +	cmem = vmalloc(sizeof(struct crash_mem) +
>>> +			sizeof(struct crash_mem_range) * nr_ranges);
>>> +	if (!cmem)
>>> +		return NULL;
>>> +
>>> +	cmem->max_nr_ranges = nr_ranges;
>>> +	cmem->nr_ranges = 0;
>>> +	walk_system_ram_res(0, -1, cmem, add_mem_range_callback);
>>> +
>>> +	/* Exclude crashkernel region */
>>> +	if (crash_exclude_mem_range(cmem, crashk_res.start, crashk_res.end)) {
>>> +		vfree(cmem);
>>> +		return NULL;
>>> +	}
>>> +
>>> +	return cmem;
>>> +}
>>
>> Could this function be included in prepare_elf_headers() so that the alloc() and
>> free() occur together.
> 
> Or aiming that arm64 and x86 have similar-look code?

What's the advantage in things looking the same? If they are the same, it
probably shouldn't be in per-arch code. Otherwise it should be as simple as
possible, otherwise we can't spot the bugs/leaks.

But I think walking memblock here will remove all 'looks the same' properties here.


>>> +static int prepare_elf_headers(void **addr, unsigned long *sz)
>>> +{
>>> +	struct crash_mem *cmem;
>>> +	int ret = 0;
>>> +
>>> +	cmem = get_crash_memory_ranges();
>>> +	if (!cmem)
>>> +		return -ENOMEM;
>>> +
>>> +	ret =  crash_prepare_elf64_headers(cmem, true, addr, sz);
>>> +
>>> +	vfree(cmem);
>>
>>> +	return ret;
>>> +}
>>
>> All this is moving memory-range information from core-code's
>> walk_system_ram_res() into core-code's struct crash_mem, and excluding
>> crashk_res, which again is accessible to the core code.
>>
>> It looks like this is duplicated in arch/x86 and arch/arm64 because arm64
>> doesn't have a second 'crashk_low_res' region, and always wants elf64, instead
>> of when IS_ENABLED(CONFIG_X86_64).
>> If we can abstract just those two, more of this could be moved to core code
>> where powerpc can make use of it if they want to support kdump with
>> kexec_file_load().
>>
>> But, its getting late for cross-architecture dependencies, lets put that on the
>> for-later list. (assuming there isn't a powerpc-kdump series out there adding a
>> third copy of this)
> 
> Sure. X86 code has so many exceptional lines in the code :)

They also pass the e820 'usable-memory' map on the cmdline...


Thanks,

James