[RFC6 PATCH v6 00/21] ILP32 for ARM64

[Catalin Marinas]

On Thu, May 12, 2016 at 05:24:57PM +0300, Yury Norov wrote:
> On Thu, May 12, 2016 at 03:07:35PM +0100, Catalin Marinas wrote:
> > On Thu, May 12, 2016 at 04:44:31PM +0300, Yury Norov wrote:
> > > On Thu, May 12, 2016 at 02:35:34PM +0100, Catalin Marinas wrote:
> > > > On Thu, May 12, 2016 at 03:20:00AM +0300, Yury Norov wrote:
> > > > > I debugged preadv02 and pwritev02 failures and found very weird bug.
> > > > > Test passes {iovec_base = 0xffffffff, iovec_len = 64} as one element
> > > > > of vector, and kernel reports successful read/write.
> > > > > 
> > > > > There are 2 problems:
> > > > > 1. How kernel allows such address to be passed to fs subsystem;
> > > > > 2. How fs successes to read/write at non-mapped, and in fact non-user
> > > > > address.
> > > > > 
> > > > > I don't know the answer on 2'nd question, and it might be something
> > > > > generic. But I investigated first problem.
> > > > > 
> > > > > The problem is that compat_rw_copy_check_uvector() uses access_ok() to
> > > > > validate user address, and on arm64 it ends up with checking buffer
> > > > > end against current_thread_info()->addr_limit.
> > > > > 
> > > > > current_thread_info()->addr_limit for ilp32, and most probably for
> > > > > aarch32 is equal to aarch64 one, and so adress_ok() doesn't fail.
> > > > > It happens because on thread creation we call flush_old_exec() to set 
> > > > > addr_limit, and completely ignore compat mode there.
> > 
> > [...]
> > 
> > > > > --- a/arch/arm64/kernel/binfmt_elf32.c
> > > > > +++ b/arch/arm64/kernel/binfmt_elf32.c
> > > > > @@ -12,6 +12,7 @@
> > > > >  do {						\
> > > > >  	clear_thread_flag(TIF_32BIT_AARCH64);	\
> > > > >  	set_thread_flag(TIF_32BIT);		\
> > > > > +	set_fs(TASK_SIZE_32);			\
> > > > >  } while (0)
> > > > >  
> > > > >  #define COMPAT_ARCH_DLINFO
> > > > > diff --git a/arch/arm64/kernel/binfmt_ilp32.c b/arch/arm64/kernel/binfmt_ilp32.c
> > > > > index a934fd4..a8599c6 100644
> > > > > --- a/arch/arm64/kernel/binfmt_ilp32.c
> > > > > +++ b/arch/arm64/kernel/binfmt_ilp32.c
> > > > > @@ -59,6 +59,7 @@ static void cputime_to_compat_timeval(const cputime_t cputime,
> > > > >  do {									\
> > > > >  	set_thread_flag(TIF_32BIT_AARCH64);				\
> > > > >  	clear_thread_flag(TIF_32BIT);					\
> > > > > +	set_fs(TASK_SIZE_32);						\
> > > > >  } while (0)
> > > > 
> > > > I don't think we need these two. AFAICT, flush_old_exec() takes care of
> > > > setting the USER_DS for the new thread.
> > > 
> > > That's true, but USER_DS depends on personality which is not set yet
> > > for new thread, as I wrote above. In fact, I tried correct USER_DS
> > > only, and it doesn't work
> > 
> > Ah, it looks like load_elf_binary() sets the personality after
> > flush_old_exec(). Looking at powerpc and x86, they set USER_DS to the
> > maximum 64-bit task value, so they should have a similar issue with
> > native 32-bit vs compat behaviour.
> 
> Hmmm. If so, it means we'd introduce generic fix. It would be removing 
> set_fs() from flush_old_exec() and appending it to load_elf_binary()
> after SET_PERSONALITY(). But I think it should be agreed with other
> arches developers.

The set_fs() in flush_old_exec() is probably fine, it may be meant to
re-set the USER_DS for the old thread.

It appears that at least powerpc and x86 don't have different USER_DS
setting for native and compat, so moving the set_fs() call further down
would not make any difference for them, nor will it fix the preadv02 LTP
test (if it fails for them, I haven't checked).

> I've sent standalone patch for aarch64 (you in CC) so let's move
> discussion there.

I've seen the patch but we would lose some discussion history here. I
think we should continue this thread and just summarise the conclusion
in reply to the other patch. This thread is also available on
linux-arch, in case other architecture maintainers follow it.

> > So what exactly is LTP complaining about? Is different error (like
> > EFAULT vs EINVAL) or not getting an error at all.
>  
> It should be EINVAL, but it succeed. The other problem is that
> following fs routines does not complain on wrong address.

I see. The test asks the kernel to write a single byte (out of maximum
64) to the user address 0xffffffff. In the absence of the access_ok()
check, this operation succeeds. If the preadv syscall gets 2 bytes as
the count, then it would fail with EFAULT.

While it's not really a bug, I agree that for matching the native 32-bit
behavior (basically for other syscalls like those involving vfs_read()),
the simplest fix would be to have a dynamic USER_DS.

-- 
Catalin