Re: [PATCH] lib/ubsan: default UBSAN_ALIGNMENT to not set

[Arnd Bergmann]

On Wed, Jan 30, 2019 at 6:12 PM Andrey Ryabinin <aryabinin@xxxxxxxxxxxxx> wrote:
> On 1/30/19 4:36 AM, Andrew Morton wrote:
> > On Fri, 11 Jan 2019 22:01:16 +0300 Andrey Ryabinin <aryabinin@xxxxxxxxxxxxx> wrote:
> >> On 12/17/18 6:03 PM, Anders Roxell wrote:
> >>> When booting an allmodconfig kernel, there are a lot of false-positives.
> >>> With a message like this 'UBSAN: Undefined behaviour in...' with a call
> >>> trace that follows.
> >>>
> >>> Reworked so that when building a allmodconfig kernel that turns
> >>> everything into '=m' or '=y' will turn off UBSAN_ALIGNMENT.
> >>>
> >>> Suggested-by: Arnd Bergmann <arnd@xxxxxxxx>
> >>> Signed-off-by: Anders Roxell <anders.roxell@xxxxxxxxxx>
> >>> ---
> >>
> >> Acked-by: Andrey Ryabinin <aryabinin@xxxxxxxxxxxxx>
> >>
> >
> > Confused.  Why does allmodconfig result in UBSAN warnings?
> >
>
> UBSAN warnings is a result of enabling noisy CONFIG_UBSAN_ALIGNMENT which is disabled
> by default if HAVE_EFFICIENT_UNALIGNED_ACCESS=y.
> It's noisy even if don't have efficient unaligned access, e.g. people often add
>  __cacheline_aligned_in_smp in structs, but forget to align allocations of such struct
> (kmalloc() give 8-byte alignment in worst case).

That would be a fixable issue, but there is another one that I looked at
with Anders: We have a number of list_for_each_entry_safe() loops
that use a list_head on the stack and add objects of a type with
large alignment. When list_for_each_entry_safe() gets to the end of
that list, it sees a pointer to an unaligned structure starting  few bytes
before the list_head on stack. The only element of that structure that
is ever used is that list_head, so there is no actual problem, but the
warning triggers nonetheless.

      Arnd