Re: Can we drop upstream Linux x32 support?

On Thu, Dec 13, 2018 at 12:40:25PM +0000, Catalin Marinas wrote:

On Wed, Dec 12, 2018 at 10:03:30AM -0800, Andy Lutomirski wrote:

quoted

On Wed, Dec 12, 2018 at 8:52 AM Rich Felker [off-list ref] wrote:

quoted

On Wed, Dec 12, 2018 at 08:39:53AM -0800, Andy Lutomirski wrote:

quoted

I'm proposing another alternative.  Given that x32 already proves that
the user bitness model doesn't have to match the kernel model (in x32,
user "long" is 32-bit but the kernel ABI "long" is 64-bit), I'm
proposing extending this to just make the kernel ABI be LP64.  So
__kernel_size_t would be 64-bit and pointers in kernel data structures
would be 64-bit.  In other words, most or all of the kernel ABI would
just match x86_64.

As far as I can tell, the only thing that really needs unusual
toolchain features here is that C doesn't have an extra-wide pointer
type.  The kernel headers would need a way to say "this pointer is
still logically a pointer, and user code may assume that it's 32 bits,
but it has 8-byte alignment."

None of this works on the userspace/C side, nor should any attempt be
made to make it work. Types fundamentally cannot have alignments
larger than their size. If you want to make the alignment of some
pointers 8, you have to make their size 8, and then you just have LP64
again if you did it for all pointers.

If on the other hand you tried to make just some pointers "wide
pointers", you'd also be completely breaking the specified API
contracts of standard interfaces. For example in struct iovec's
iov_base, &foo->iov_base is no longer a valid pointer to an object of
type void* that you can pass to interfaces expecting void**. Sloppy
misunderstandings like what you're making now are exactly why x32 is
already broken and buggy (&foo->tv_nsec already has wrong type for
struct timespec foo).

I don't think it's quite that broken.  For the struct iovec example,
we currently have:

           struct iovec {
               void  *iov_base;    /* Starting address */
               size_t iov_len;     /* Number of bytes to transfer */
           };

we could have, instead: (pardon any whitespace damage)

           struct iovec {
               void  *iov_base;    /* Starting address */
               uint32_t __pad0;
               size_t iov_len;     /* Number of bytes to transfer */
               uint32_t __pad1;
           } __attribute__((aligned(8));

or the same thing but where iov_len is uint64_t.  A pointer to
iov_base still works exactly as expected.  Something would need to be
done to ensure that the padding is all zeroed, which might be a real
problem.

We looked at this approach briefly for arm64/ILP32 and zeroing the pads
was the biggest problem. User programs would not explicitly zero the pad
and I'm not sure the compiler would be any smarter. This means it's the
kernel's responsibility to zero the pad (around get_user,
copy_from_user), so it doesn't actually simplify the kernel side of the
syscall interface.

If the data flow goes the other way (kernel to user), this approach
works fine.

quoted

No one wants to actually type all the macro gunk into the headers to
make this work, but this type of transformation is what I have in mind
when the compiler is asked to handle the headers.  Or there could
potentially be a tool that automatically consumes the uapi headers and
spits out modified headers like this.

If the compiler can handle the zeroing, that would be great, though not
sure how (some __attribute__((zero)) which generates a type constructor
for such structure; it kind of departs from what the C language offers).

The compiler fundamentally can't. At the very least it would require
effective type tracking, which requires shadow memory and is even more
controversial than -fstrict-aliasing (because in a sense it's a
stronger version thereof). But even effective type tracking would not
help, since you can have things like:

	struct iovec *iov = malloc(sizeof *iov);
	scanf("%p %zu", &iov->iov_base, &iov->iov_len);

where no store to the object via the struct type ever happens and the
only stores that do happen are invisible across translation unit
boundaries. (Ignore that scanf here is awful; it's just a canonical
example of a function that would store the members via pointers to
them.)

The kernel-side approach could work if the kernel had some markup for
fields that need to be zero- or sign-extended when copied from user in
a 32-bit process and applied them at copy time. That could also fix
the existing tv_nsec issue. I'm not sure how difficult/costly it would
be though.

Rich