Re: [PATCH 1/2] um: mark rodata read-only and implement _nofault accesses

[Johannes Berg <johannes@sipsolutions.net>]

On Thu, 2025-04-03 at 12:19 -0700, Nathan Chancellor wrote:
> 
> Thanks, I applied that change, which shows a slightly different crash
> message now:

Pretty sure it's all just a bug in my inline assembly, and clang
allocates registers differently:

#define ___backtrack_faulted(_faulted)                                  \
        asm volatile (                                                  \
                "mov $0, %0\n"                                          \
                "movq $__get_kernel_nofault_faulted_%=,%1\n"            \
                "jmp _end_%=\n"                                         \
                "__get_kernel_nofault_faulted_%=:\n"                    \
                "mov $1, %0;"                                           \
                "_end_%=:"                                              \
                : "=r" (_faulted),                                      \
                  "=m" (current->thread.segv_continue) ::               \
        )


It _looks_ as though both %0 and %1 are output only, but clang compiles
it to:

  51:   48 83 fb 08             cmp    $0x8,%rbx
  55:   72 44                   jb     9b <_end_0+0x2a>
  57:   48 8b 01                mov    (%rcx),%rax

// start inline assembly ---vvv--- //
  5a:   b8 00 00 00 00          mov    $0x0,%eax
  5f:   48 c7 80 90 07 00 00    movq   $0x0,0x790(%rax)       // crash
  66:   00 00 00 00 
                        66: R_X86_64_32S        .text+0x6c
  6a:   eb 05                   jmp    71 <_end_0>

000000000000006c <__get_kernel_nofault_faulted_0>:
  6c:   b8 01 00 00 00          mov    $0x1,%eax
// end inline assembly ---^^^--- //

0000000000000071 <_end_0>:
  71:   85 c0                   test   %eax,%eax
  73:   75 56                   jne    cb <_end_1+0x10>


which clearly cannot work? I must be missing something. Switching the
first two instructions fixes it, of course, but right now I can't see
what I forgot in terms of constraints to make the compiler not do that.
Probably trivial to someone more familiar with inline assembly.
Modifying the _faulted to be +r instead of =r also fixes it.

johannes