Re: PROBLEM: zstd bzImage decompression fails for some x86_32 config on 5.9-rc1

[Nick Terrell <terrelln@fb.com>]

> On Sep 28, 2020, at 11:02 AM, Nick Terrell <terrelln@fb.com> wrote:
> 
> 
> 
>> On Sep 28, 2020, at 1:55 AM, Feng Tang <feng.tang@intel.com> wrote:
>> 
>> Hi Nick,
>> 
>> 0day has found some kernel decomprssion failure case since 5.9-rc1 (X86_32
>> build), and it could be related with ZSTD code, though initially we bisected
>> to some other commits.
>> 
>> The error messages are: 
>> 	
>> 	early console in setup code
>> 	Wrong EFI loader signature.
>> 	early console in extract_kernel
>> 	input_data: 0x046f50b4
>> 	input_len: 0x01ebbeb6
>> 	output: 0x01000000
>> 	output_len: 0x04fc535c
>> 	kernel_total_size: 0x055f5000
>> 	needed_size: 0x055f5000
>> 	
>> 	Decompressing Linux...
>> 	
>> 	ZSTD-compressed data is corrupt
>> 
>> This could be reproduced by compiling the kernel with attached config,
>> and use QEMU to boot it.
>> 
>> We suspect it could be related with the kernel size, as we only see
>> it on big kernel, and some more info are:
>> 
>> * If we remove a lot of kernel config to build a much smaller kernel,
>> it will boot fine
>> 
>> * If we change the zstd algorithm from zstd22 to zstd19, the kernel will
>> boot fine with below patch
>> 
>> 	diff --git a/arch/x86/boot/compressed/Makefile b/arch/x86/boot/compressed/Makefile
>> 	index 3962f59..8fe71ba 100644
>> 	--- a/arch/x86/boot/compressed/Makefile
>> 	+++ b/arch/x86/boot/compressed/Makefile
>> 	@@ -147,7 +147,7 @@ $(obj)/vmlinux.bin.lzo: $(vmlinux.bin.all-y) FORCE
>> 	 $(obj)/vmlinux.bin.zst: $(vmlinux.bin.all-y) FORCE
>> 	-	$(call if_changed,zstd22)
>> 	+	$(call if_changed,zstd)
>> 
>> 
>> Please let me know if you need more info, and sorry for the late report
>> as we just tracked down to this point.
> 
> Thanks for the report, I will look into it today.

CC: Petr Malat

I’ve successfully reproduced, and found the issue. It turns out that this
patch [0] from Petr Malat fixes the issue. As I mentioned in that thread, his
fix corresponds to this upstream commit [1].

Can we get Petr's patch merged into v5.9?

This bug only happens when the window size is > 8 MB. A non-kernel workaround
would be to compress the kernel level 19 instead of level 22, which uses an
8 MB window size, instead of a 128 MB window size.

The reason it only shows up for large kernels, is that the code is only buggy
when an offset > 8 MB is used, so a kernel <= 8 MB can't trigger the bug.

Best,
Nick

[0] https://lkml.org/lkml/2020/9/14/94
[1] https://github.com/facebook/zstd/commit/8a5c0c98ae5a7884694589d7a69bc99011add94d

> Best,
> Nick
> 
>> Thanks,
>> Feng
>> 
>> 
>> 
>> <zstd_x86_32.config>