Re: [PATCH] mm/swap: Avoid undefined behavior in __swapoffset

[Michal Hocko]

On Tue 12-03-19 15:02:38, Pi-Hsun Shih wrote:
> On Thu, Mar 7, 2019 at 9:23 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> >
> > On Thu 07-03-19 20:47:52, Pi-Hsun Shih wrote:
> > > On Thu, Mar 7, 2019 at 8:23 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > > >
> > > > On Thu 07-03-19 17:46:50, Pi-Hsun Shih wrote:
> > > > > Use offsetof to calculate offset of a field to avoid UBSAN warning like:
> > > > >
> > > > > ===================================================================
> > > > > UBSAN: Undefined behaviour in mm/swapfile.c:3010:38
> > > > > member access within null pointer of type 'union swap_header'
> > > > > CPU: 6 PID: 1833 Comm: swapon Tainted: G S                4.19.23 #43
> > > > > Call trace:
> > > > >  dump_backtrace+0x0/0x194
> > > > >  show_stack+0x20/0x2c
> > > > >  __dump_stack+0x20/0x28
> > > > >  dump_stack+0x70/0x94
> > > > >  ubsan_epilogue+0x14/0x44
> > > > >  ubsan_type_mismatch_common+0xf4/0xfc
> > > > >  __ubsan_handle_type_mismatch_v1+0x34/0x54
> > > > >  __se_sys_swapon+0x654/0x1084
> > > > >  __arm64_sys_swapon+0x1c/0x24
> > > > >  el0_svc_common+0xa8/0x150
> > > > >  el0_svc_compat_handler+0x2c/0x38
> > > > >  el0_svc_compat+0x8/0x18
> > > > > ==================================================================
> > > >
> > > > Could you be more specific about what exactly is undefined here and
> > > > why offsetof is any better. AFAIR it uses the same construct unless a
> > > > compiler defines a built in.
> > > >
> > > > I do not object the change itself because it is cleaner to use the
> > > > existing helper but I am wondering why this is fixing ubsan. Is ubsan
> > > > defining the compiler variant and consider it safe?
> > > >
> > >
> > > The undefined behavior is from trying to accessing a member of NULL,
> > > even not using it value but only use the address.
> >
> > Hmm, we've been using this trick for ages and I do not remember any
> > compiler to complain as there is no real access. I am not sure what the
> > C standard has to tell about that but I presume reasonable compilers
> > will not abuse the UB here.
> >
> 
> Some more testing shows that GCC optimize the
> ((size_t)&((type*)0)->member) to a constant in the result binary, and
> never emit any UBSAN checks on the statement.
> Clang doesn't optimize it to a constant in -O0, optimize it to a
> constant in -O1 or above, and always emit the
> __ubsan_handle_type_mismatch check when "-fsanitize=undefined" is
> given.
> So this UBSAN warning only happens when kernel is compiled by clang, not GCC.
> 
> From what I've found, it's a UB from C standard view point
> (https://software.intel.com/en-us/blogs/2015/04/20/null-pointer-dereferencing-causes-undefined-behavior),
> but I agree that probably no reasonable compilers would abuse the UB
> here.

I really do not want to go and lawyering about the standard here but
getting an address of an offset based on NULL ptr is not really
dereferencing of a NULL ptr. At least this was not the case for ages
and no compiler can afford to change it because there is quite a lot of
userspace to rely on this construct.

But as I've said using offseoff is nicer so I completely support a patch
that get's read of a custom redefinition of it or open code directly.
But calling it an UB is a bit of stretch.
-- 
Michal Hocko
SUSE Labs