On Wed, Nov 22, 2023 at 03:44:07PM +1100, Michael Ellerman wrote:

Naveen N Rao [off-list ref] writes:

quoted

On Tue, Nov 21, 2023 at 10:54:36AM +1100, Michael Ellerman wrote:

quoted

Building with GCC 13 (which has -array-bounds enabled) there are several

Thanks, gcc13 indeed helps reproduce the warnings.

Actually that part is no longer true since 0da6e5fd6c37 ("gcc: disable
'-Warray-bounds' for gcc-13 too").

quoted

quoted

warnings in sstep.c along the lines of:

  In function ‘do_byte_reverse’,
      inlined from ‘do_vec_load’ at arch/powerpc/lib/sstep.c:691:3,
      inlined from ‘emulate_loadstore’ at arch/powerpc/lib/sstep.c:3439:9:
  arch/powerpc/lib/sstep.c:289:23: error: array subscript 2 is outside array bounds of ‘u8[16]’ {aka ‘unsigned char[16]’} [-Werror=array-bounds=]
    289 |                 up[2] = byterev_8(up[1]);
        |                 ~~~~~~^~~~~~~~~~~~~~~~~~
  arch/powerpc/lib/sstep.c: In function ‘emulate_loadstore’:
  arch/powerpc/lib/sstep.c:681:11: note: at offset 16 into object ‘u’ of size 16
    681 |         } u = {};
        |           ^

do_byte_reverse() supports a size up to 32 bytes, but in these cases the
caller is only passing a 16 byte buffer. In practice there is no bug,
do_vec_load() is only called from the LOAD_VMX case in emulate_loadstore().
That in turn is only reached when analyse_instr() recognises VMX ops,
and in all cases the size is no greater than 16:

  $ git grep -w LOAD_VMX arch/powerpc/lib/sstep.c
  arch/powerpc/lib/sstep.c:                        op->type = MKOP(LOAD_VMX, 0, 1);
  arch/powerpc/lib/sstep.c:                        op->type = MKOP(LOAD_VMX, 0, 2);
  arch/powerpc/lib/sstep.c:                        op->type = MKOP(LOAD_VMX, 0, 4);
  arch/powerpc/lib/sstep.c:                        op->type = MKOP(LOAD_VMX, 0, 16);

Similarly for do_vec_store().

Although the warning is incorrect, the code would be safer if it clamped
the size from the caller to the known size of the buffer. Do that using
min_t().

But, do_vec_load() and do_vec_store() assume that the maximum size is 16 
(the address_ok() check as an example). So, should we be considering a 
bigger hammer to help detect future incorrect use?

Yeah true.

To be honest I don't know how paranoid we want to get, we could end up
putting WARN's all over the kernel :)

In this case I guess if the size is too large we overflow the buffer on
the kernel stack, so we should at least check the size.

But does it need a WARN? I'm not sure. If we had a case that was passing
a out-of-bound size hopefully we would notice in testing? :)

You're right, a simpler check should suffice. I will send an updated 
patch.

Thanks,
Naveen