Marco Elver <elver@xxxxxxxxxx> writes:
> On Mon, May 03, 2021 at 09:03PM -0700, Peter Collingbourne wrote:
>> On Mon, May 3, 2021 at 8:42 PM Eric W. Biederman <ebiederm@xxxxxxxxxxxx> wrote:
>> > Marco Elver <elver@xxxxxxxxxx> writes:
>> > > On Mon, 3 May 2021 at 23:04, Eric W. Biederman <ebiederm@xxxxxxxxxxxx> wrote:
>> > >> "Eric W. Beiderman" <ebiederm@xxxxxxxxxxxx> writes:
>> > >> > From: "Eric W. Biederman" <ebiederm@xxxxxxxxxxxx>
>> > >> >
>> > >> > The si_perf code really wants to add a u64 field. This change enables
>> > >> > that by reorganizing the definition of siginfo_t, so that a 64bit
>> > >> > field can be added without increasing the alignment of other fields.
>> > >
>> > > If you can, it'd be good to have an explanation for this, because it's
>> > > not at all obvious -- some future archeologist will wonder how we ever
>> > > came up with this definition of siginfo...
>> > >
>> > > (I see the trick here is that before the union would have changed
>> > > alignment, introducing padding after the 3 ints -- but now because the
>> > > 3 ints are inside the union the union's padding no longer adds padding
>> > > for these ints. Perhaps you can explain it better than I can. Also
>> > > see below.)
>> >
>> > Yes. The big idea is adding a 64bit field into the second union
>> > in the _sigfault case will increase the alignment of that second
>> > union to 64bit.
>> >
>> > In the 64bit case the alignment is already 64bit so it is not an
>> > issue.
>> >
>> > In the 32bit case there are 3 ints followed by a pointer. When the
>> > 64bit member is added the alignment of _segfault becomes 64bit. That
>> > 64bit alignment after 3 ints changes the location of the 32bit pointer.
>> >
>> > By moving the 3 preceding ints into _segfault that does not happen.
>> >
>> >
>> >
>> > There remains one very subtle issue that I think isn't a problem
>> > but I would appreciate someone else double checking me.
>> >
>> >
>> > The old definition of siginfo_t on 32bit almost certainly had 32bit
>> > alignment. With the addition of a 64bit member siginfo_t gains 64bit
>> > alignment. This difference only matters if the 64bit field is accessed.
>> > Accessing a 64bit field with 32bit alignment will cause unaligned access
>> > exceptions on some (most?) architectures.
>> >
>> > For the 64bit field to be accessed the code needs to be recompiled with
>> > the new headers. Which implies that when everything is recompiled
>> > siginfo_t will become 64bit aligned.
>> >
>> >
>> > So the change should be safe unless someone is casting something with
>> > 32bit alignment into siginfo_t.
>>
>> How about if someone has a field of type siginfo_t as an element of a
>> struct? For example:
>>
>> struct foo {
>> int x;
>> siginfo_t y;
>> };
>>
>> With this change wouldn't the y field move from offset 4 to offset 8?
>
> This is a problem if such a struct is part of the ABI -- in the kernel I
> found these that might be problematic:
>
> | arch/csky/kernel/signal.c:struct rt_sigframe {
> | arch/csky/kernel/signal.c- /*
> | arch/csky/kernel/signal.c- * pad[3] is compatible with the same struct defined in
> | arch/csky/kernel/signal.c- * gcc/libgcc/config/csky/linux-unwind.h
> | arch/csky/kernel/signal.c- */
> | arch/csky/kernel/signal.c- int pad[3];
> | arch/csky/kernel/signal.c- struct siginfo info;
> | arch/csky/kernel/signal.c- struct ucontext uc;
> | arch/csky/kernel/signal.c-};
> | [...]
> | arch/parisc/include/asm/rt_sigframe.h-#define SIGRETURN_TRAMP 4
> | arch/parisc/include/asm/rt_sigframe.h-#define SIGRESTARTBLOCK_TRAMP 5
> | arch/parisc/include/asm/rt_sigframe.h-#define TRAMP_SIZE (SIGRETURN_TRAMP + SIGRESTARTBLOCK_TRAMP)
> | arch/parisc/include/asm/rt_sigframe.h-
> | arch/parisc/include/asm/rt_sigframe.h:struct rt_sigframe {
> | arch/parisc/include/asm/rt_sigframe.h- /* XXX: Must match trampoline size in arch/parisc/kernel/signal.c
> | arch/parisc/include/asm/rt_sigframe.h- Secondary to that it must protect the ERESTART_RESTARTBLOCK
> | arch/parisc/include/asm/rt_sigframe.h- trampoline we left on the stack (we were bad and didn't
> | arch/parisc/include/asm/rt_sigframe.h- change sp so we could run really fast.) */
> | arch/parisc/include/asm/rt_sigframe.h- unsigned int tramp[TRAMP_SIZE];
> | arch/parisc/include/asm/rt_sigframe.h- struct siginfo info;
> | [..]
> | arch/parisc/kernel/signal32.h-#define COMPAT_SIGRETURN_TRAMP 4
> | arch/parisc/kernel/signal32.h-#define COMPAT_SIGRESTARTBLOCK_TRAMP 5
> | arch/parisc/kernel/signal32.h-#define COMPAT_TRAMP_SIZE (COMPAT_SIGRETURN_TRAMP + \
> | arch/parisc/kernel/signal32.h- COMPAT_SIGRESTARTBLOCK_TRAMP)
> | arch/parisc/kernel/signal32.h-
> | arch/parisc/kernel/signal32.h:struct compat_rt_sigframe {
> | arch/parisc/kernel/signal32.h- /* XXX: Must match trampoline size in arch/parisc/kernel/signal.c
> | arch/parisc/kernel/signal32.h- Secondary to that it must protect the ERESTART_RESTARTBLOCK
> | arch/parisc/kernel/signal32.h- trampoline we left on the stack (we were bad and didn't
> | arch/parisc/kernel/signal32.h- change sp so we could run really fast.) */
> | arch/parisc/kernel/signal32.h- compat_uint_t tramp[COMPAT_TRAMP_SIZE];
> | arch/parisc/kernel/signal32.h- compat_siginfo_t info;
>
> Adding these static asserts to parisc shows the problem:
>
> | diff --git a/arch/parisc/kernel/signal.c b/arch/parisc/kernel/signal.c
> | index fb1e94a3982b..0be582fb81be 100644
> | --- a/arch/parisc/kernel/signal.c
> | +++ b/arch/parisc/kernel/signal.c
> | @@ -610,3 +610,6 @@ void do_notify_resume(struct pt_regs *regs, long in_syscall)
> | if (test_thread_flag(TIF_NOTIFY_RESUME))
> | tracehook_notify_resume(regs);
> | }
> | +
> | +static_assert(sizeof(unsigned long) == 4); // 32 bit build
> | +static_assert(offsetof(struct rt_sigframe, info) == 9 * 4);
>
> This passes without the siginfo rework in this patch. With it:
>
> | ./include/linux/build_bug.h:78:41: error: static assertion failed: "offsetof(struct rt_sigframe, info) == 9 * 4"
>
> As sad as it is, I don't think we can have our cake and eat it, too. :-(
>
> Unless you see why this is fine, I think we need to drop this patch and
> go back to the simpler version you had.
No. I really can't. I think we are stuck with 32bit alignment on 32bit
architectures at this point. Which precludes 32bit architectures from
including a 64bit field.
The variant of this that concerns me the most is siginfo_t embedded in a
structure in a library combined with code that is compiled with new
headers. The offset of the embedded siginfo_t could very easily change
and break things.
That makes the alignment an ABI property we can't mess with. Shame.
I will figure out some static asserts to verify this property remains
on 32bit and respin this series.
Eric
