Re: 4879b7ae05 ("Merge tag 'dmaengine-4.12-rc1' of .."): WARNING: kernel stack regs at bd92bc2e in 01-cpu-hotplug:3811 has bad 'bp' value 000001be

From: Linus Torvalds
Date: Mon Oct 02 2017 - 17:58:17 EST


On Mon, Oct 2, 2017 at 2:26 PM, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
>
> The bisect is pointing to a commit which is almost 5 months old, so this
> is pre-ORC. Kallsyms *is* enabled, but the unwinder dump isn't smart
> enough to realize it's dumping misaligned stack addresses:

Ahh, I didn't pick up on that "esp isn't aligned" part.

That said, if %esp gets unaligned at some point, it's not clear
exactly when we should align it. An unaligned stack pointer will
continue to _work_ just potentially perform fairly badly.

But more likely, we picked the wrong frame value to begin with.

For example, maybe that decode_frame_pointer() logic really should
check not that the low bit in bp is set, but instead check that it's a
valid "unsigned long *" that has the low bit set.

IOW, the difference would be that instead of checking

if (!(regs & 0x1))
return NULL;

if would check

if ((regs & (sizeof(unsigned long)-1)) != 1)
return NULL;

but also maybe add logic to simply not trust a next frame pointer that
isn't appropriately aligned.

So I think adding PTR_ALIGN() there in the unwind dumper might be a
bit late. By that time it has already accepted what looks like a
garbage frame. No?

Linus