tree f3232760a376c683a629c6c47e9ac4a9500967d1
parent f6e922365faf4cd576bd1cf3e64b58c8a32e1856
author Ilya Leoshkevich <iii@linux.ibm.com> 1710899652 +0100
committer Alexei Starovoitov <ast@kernel.org> 1710913963 -0700

s390/bpf: Fix bpf_plt pointer arithmetic

Kui-Feng Lee reported a crash on s390x triggered by the
dummy_st_ops/dummy_init_ptr_arg test [1]:

  [<0000000000000002>] 0x2
  [<00000000009d5cde>] bpf_struct_ops_test_run+0x156/0x250
  [<000000000033145a>] __sys_bpf+0xa1a/0xd00
  [<00000000003319dc>] __s390x_sys_bpf+0x44/0x50
  [<0000000000c4382c>] __do_syscall+0x244/0x300
  [<0000000000c59a40>] system_call+0x70/0x98

This is caused by GCC moving memcpy() after assignments in
bpf_jit_plt(), resulting in NULL pointers being written instead of
the return and the target addresses.

Looking at the GCC internals, the reordering is allowed because the
alias analysis thinks that the memcpy() destination and the assignments'
left-hand-sides are based on different objects: new_plt and
bpf_plt_ret/bpf_plt_target respectively, and therefore they cannot
alias.

This is in turn due to a violation of the C standard:

  When two pointers are subtracted, both shall point to elements of the
  same array object, or one past the last element of the array object
  ...

From the C's perspective, bpf_plt_ret and bpf_plt are distinct objects
and cannot be subtracted. In the practical terms, doing so confuses the
GCC's alias analysis.

The code was written this way in order to let the C side know a few
offsets defined in the assembly. While nice, this is by no means
necessary. Fix the noncompliance by hardcoding these offsets.

[1] https://lore.kernel.org/bpf/c9923c1d-971d-4022-8dc8-1364e929d34c@gmail.com/

Fixes: f1d5df84cd8c ("s390/bpf: Implement bpf_arch_text_poke()")
Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
Message-ID: <20240320015515.11883-1-iii@linux.ibm.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>