[Xen-devel] [patch] do_multicall_call abi violation? (x86_64)

["Scott Parish" <srparish@xxxxxxxxxx>]

Booting dom0 and running xend spews quite a number of:

 (XEN) (file=multicall.c, line=50) Error writing result back to multicall block.

To jump right to the end of a long story, i think that
do_multicall_call() is in x86_64 abi violation. 

   #define do_multicall_call(_call)                         \
       do {                                                 \
           __asm__ __volatile__ (                           \
               "movq  "STR(MULTICALL_op)"(%0),%%rax; "      \
     ...
               "movq  "STR(MULTICALL_arg4)"(%0),%%r8; "     \
               "callq *(%%rax); "                           \
               "movq  %%rax,"STR(MULTICALL_result)"(%0); "  \
               : : "b" (_call)                              \
               : "rax", "rdi", "rsi", "rdx", "rcx", "r8");  \
       } while ( 0 )

The abi says the following (v0.95, page 16):

       Registers %rbp, %rbx and %r12 through %r15 "belong" to the
       calling function and the called function is required to preserve
       their values. In other words, a called function must preserve
       these registers' values for its caller. Remaining registers
       "belong" to the called function. If a calling function wants to
       preserve such a register value across a function call, it must
       save the value in its local stack frame.

do_multicall_call() is doing a call, who does who knows what with
registers, but doesn't advertise clobbering all the abi's volatile
registers.

I tried searching for documentation, as it seems to me that gcc should
have some magic flag since it has to know the abi anyway, but i couldn't
find anything. Assuming that this isn't a compiler bug and there aren't
any such flags, the attached patch should make things abi compliant, and
fixes the problem i was seeing.

-----

If you're curious the exact connection between the abi violation and
the error messages...

do_multicall() looks somewhat like (pseudo-code):

   do_multicall(...) {
      if ( asm("inline-asm-1") )
         return;
 
      do_multicall_call();

      if ( asm("inline-asm-2") ) {
         printk("Error");
         return;
      }

the result of inline-asm-1 is in some register X. the "if" turns into a
jnz on X to a "ret". gcc realizes if we didn't jump, then X has to be 0,
so it's going to use that value for initialization of the __pu_err in
inline-asm-2.

the problem is that the call in do_multicall_call() stomps on X, as it
isn't advertised to clobber X.

inline-asm-2 is the __put_user code, which uses the fixup/exception-
table trick:

                "1:     mov"itype" %"rtype"1,%2\n"                      \
                "2:\n"                                                  \
                ".section .fixup,\"ax\"\n"                              \
                "3:     mov %3,%0\n"                                    \
                "       jmp 2b\n"                                       \
                ".previous\n"                                           \
                ".section __ex_table,\"a\"\n"                           \
                "       "__FIXUP_ALIGN"\n"                              \
                "       "__FIXUP_WORD" 1b,3b\n"                         \
                ".previous"                                             \
                : "=r"(err)                                             \
                : ltype (x), "m"(__m(addr)), "i"(errret), "0"(err))

so "err" (in register X) is uninitialized if we don't fault. finally
depending on what the call did with X if could be a non-zero value.


sRp

-- 
Scott Parish
Signed-off-by: srparish@xxxxxxxxxx

Attachment: register-ditty.diff
Description: Text document

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel