[v9,06/11] arm64: kexec_file: allow for loading Image-format kernel

Hi Akashi,

On 07/05/18 08:21, AKASHI Takahiro wrote:
> On Tue, May 01, 2018 at 06:46:11PM +0100, James Morse wrote:
>> On 25/04/18 07:26, AKASHI Takahiro wrote:
>>> This patch provides kexec_file_ops for "Image"-format kernel. In this
>>> implementation, a binary is always loaded with a fixed offset identified
>>> in text_offset field of its header.

>>> diff --git a/arch/arm64/include/asm/kexec.h b/arch/arm64/include/asm/kexec.h
>>> index e4de1223715f..3cba4161818a 100644
>>> --- a/arch/arm64/include/asm/kexec.h
>>> +++ b/arch/arm64/include/asm/kexec.h
>>> @@ -102,6 +102,56 @@ struct kimage_arch {
>>>  	void *dtb_buf;
>>>  };
>>>  
>>> +/**
>>> + * struct arm64_image_header - arm64 kernel image header
>>> + *
>>> + * @pe_sig: Optional PE format 'MZ' signature
>>> + * @branch_code: Instruction to branch to stext
>>> + * @text_offset: Image load offset, little endian
>>> + * @image_size: Effective image size, little endian
>>> + * @flags:
>>> + *	Bit 0: Kernel endianness. 0=little endian, 1=big endian
>>
>> Page size? What about 'phys_base'?, (whatever that is...)
>> Probably best to refer to Documentation/arm64/booting.txt here, its the
>> authoritative source of what these fields mean.
> 
> While we don't care other bit fields for now, I will add the reference
> to the Documentation file.

Thanks, I don't want to create a second, incomplete set of documentation!

>>> +	u64 reserved[3];
>>> +	u8 magic[4];
>>> +	u32 pe_header;
>>> +};
>>
>> I'm surprised we don't have a definition for this already, I guess its always
>> done in asm. We have kernel/image.h that holds some of this stuff, if we are
>> going to validate the flags, is it worth adding the code there, (and moving it
>> to include/asm)?
> 
> A comment at the beginning of this file says,
>     #ifndef LINKER_SCRIPT
>     #error This file should only be included in vmlinux.lds.S
>     #endif
> Let me think about.

Ah, I missed that.

Having two definitions of something makes me nervous that they can become
different... looks like that header belongs to the linker, and shouldn't be used
here then.

>> I guess you skip the MZ prefix as its not present for !EFI?
> 
> CONFIG_KEXEC_IMAGE_VERIFY_SIG depends on the fact that the file
> format is PE (that is, EFI is enabled).

So if the signature checking is enabled, its already been checked.

>> Could we check branch_code is non-zero, and text-offset points within image-size?
> 
> We could do it, but I don't think this check is very useful.
> 
>>
>> We could check that this platform supports the page-size/endian config that this
>> Image was built with... We get a message from the EFI stub if the page-size
>> can't be supported, it would be nice to do the same here (as we can).
> 
> There is no restriction on page-size or endianness for kexec.

No, but it won't boot if the hardware doesn't support it. The kernel will spin
at a magic address that is, difficult, to debug without JTAG. The bug report
will be "it didn't boot".

> What will be the purpose of this check?

These values are in the header so that the bootloader can check them, then print
a meaningful error. Here, kexec_file_load() is playing the part of the bootloader.

I'm assuming kexec_file_load() can only be used to kexec linux... unlike regular
kexec. Is this where I'm going wrong?

>>> diff --git a/arch/arm64/kernel/kexec_image.c b/arch/arm64/kernel/kexec_image.c
>>> new file mode 100644
>>> index 000000000000..4dd524ad6611
>>> --- /dev/null
>>> +++ b/arch/arm64/kernel/kexec_image.c
>>> @@ -0,0 +1,79 @@
>>
>>> +static void *image_load(struct kimage *image,
>>> +				char *kernel, unsigned long kernel_len,
>>> +				char *initrd, unsigned long initrd_len,
>>> +				char *cmdline, unsigned long cmdline_len)
>>> +{
>>> +	struct kexec_buf kbuf;
>>> +	struct arm64_image_header *h = (struct arm64_image_header *)kernel;
>>> +	unsigned long text_offset;
>>> +	int ret;
>>> +
>>> +	/* Load the kernel */
>>> +	kbuf.image = image;
>>> +	kbuf.buf_min = 0;
>>> +	kbuf.buf_max = ULONG_MAX;
>>> +	kbuf.top_down = false;
>>> +
>>> +	kbuf.buffer = kernel;
>>> +	kbuf.bufsz = kernel_len;
>>> +	kbuf.memsz = le64_to_cpu(h->image_size);
>>> +	text_offset = le64_to_cpu(h->text_offset);
>>> +	kbuf.buf_align = SZ_2M;
>>
>>> +	/* Adjust kernel segment with TEXT_OFFSET */
>>> +	kbuf.memsz += text_offset;
>>> +
>>> +	ret = kexec_add_buffer(&kbuf);
>>> +	if (ret)
>>> +		goto out;
>>> +
>>> +	image->arch.kern_segment = image->nr_segments - 1;
>>
>> You only seem to use kern_segment here, and in load_other_segments() called
>> below. Could it not be a local variable passed in? Instead of arch-specific data
>> we keep forever?
> 
> No, kern_segment is also used in load_other_segments() in machine_kexec_file.c.
> To optimize memory hole allocation logic in locate_mem_hole_callback(),
> we need to know the exact range of kernel image (start and end).

That's the second user. My badly-made point is one calls the other, but passes
the data via some until-kexec lifetime struct. (its not important, just an
indicator this worked differently in the past and hasn't been cleaned up).
I meant something like [0].

Thanks,

James

[0] a diff is worth a thousand words:
--------------------%<--------------------
--------------------%<--------------------

[v9,06/11] arm64: kexec_file: allow for loading Image-format kernel

Commit Message

Comments

Patch