Re: [PATCH] powerpc/bug: Cast to unsigned long before passing to inline asm

From
Michael Ellerman <mpe@ellerman.id.au>
Date
2021-09-01 07:18:24
Reply to
Segher Boessenkool, 2021-08-31
Segher Boessenkool [off-list ref] writes:
On Tue, Aug 31, 2021 at 11:27:20PM +1000, Michael Ellerman wrote:
quoted
Nathan filed an LLVM bug [2], in which Eli Friedman explained that "if
you pass a value of a type that's narrower than a register to an inline
asm, the high bits are undefined". In this case we are passing a bool
to the inline asm, which is a single bit value, and so the compiler
thinks it can leave the high bits of r30 unmasked, because only the
value of bit 0 matters.

Because the inline asm compares the full width of the register (32 or
64-bit) we need to ensure the value passed to the inline asm has all
bits defined. So fix it by casting to long.

We also already cast to long for the similar case in BUG_ENTRY(), which
it turns out was added to fix a similar bug in 2005 in commit
32818c2eb6b8 ("[PATCH] ppc64: Fix issue with gcc 4.0 compiled kernels").
That points to <https://gcc.gnu.org/PR23422>, which shows the correct
explanation.
That's a pity because I don't understand that explanation ^_^

Why does sign extension matter when we're comparing against zero?
The code as it was did **not** pass a bool.  It perhaps passed an int
(so many macros in play, it is hard to tell for sure, but it is int or
long int, perhaps unsigned (which does not change anything here).
I don't understand that. It definitely is passing a bool at the source
level. Are you saying it's getting promoted somehow?

It expands to:

      asm goto(
              "1:   "
              "tdnei"
              "
              " " % 4,
              0 " "\n " ".section __ex_table,\"a\";"
                                              " "
                                              ".balign 4;"
                                              " "
                                              ".long (1b) - . ;"
                                              " "
                                              ".long (%l[__label_warn_on]) - . ;"
                                              " "
                                              ".previous"
                                              " "
                                              ".section __bug_table,\"aw\"\n"
                                              "2:\t.4byte 1b - 2b, %0 - 2b\n"
                                              "\t.short %1, %2\n"
                                              ".org 2b+%3\n"
                                              ".previous\n"
              :
              : "i"("lib/klist.c"), "i"(62),
                "i"((1 << 0) | ((9) << 8)),
                "i"(sizeof(struct bug_entry)),
                "r"(knode_dead(knode))
                ^^^^^^^^^^^^^^^^^^^^^
              :
              : __label_warn_on);

And knode_dead() returns bool:

  static bool knode_dead(struct klist_node *knode)
  {
  	return (unsigned long)knode->n_klist & KNODE_DEAD;
  }

So in my book that means the type there is bool. But I'm not a compiler
guy so I guessing I'm missing something.
But td wants a 64-bit register, not a 32-bit one (which are the only two
possibilities for the "r" constraint on PowerPC).  The cast to "long" is
fine for powerpc64, but it has nothing to do with "narrower than a
register".
If it's 32-bit vs 64-bit, and the clang explanation is correct, then
we'd expect the low 32-bits of the value passed to the asm to have the
correct value, ie. have been masked with KNODE_DEAD.
If this is not the correct explanation for LLVM, it needs to solve some
other bug.
OK. I just need to get this fixed in the kernel, so I'll do a new
version with a changelog that is ~= "shrug not sure what's going on" and
merge that. Then we can argue about what is really going on later :)

cheeers
Thread (5 messages) 5 messages, 3 authors, last reply 2021-09-01
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help