Re: [PATCH v2] x86: fix alternative_callN usage of ALTERNATIVE asm macro

["Roger Pau Monné" <roger.pau@citrix.com>]

On Thu, May 23, 2019 at 03:41:23AM -0600, Jan Beulich wrote:
> >>> On 22.05.19 at 18:45, <roger.pau@citrix.com> wrote:
> > alternative_callN using inline assembly to generate the alternative
> > patch sites should be using the ALTERNATIVE C preprocessor macro
> > rather than the ALTERNATIVE assembly macro,
> 
> Why? See INDIRECT_{CALL,JMP}. My goal, as said on irc, would be
> to eventually eliminate the redundant C macros, in favor of just using
> the assembler ones.

Using the current assembly macros for inline asm alternatives would
regress the build on llvm based toolchains. If that's indeed the path
forward I will have to look into making those work in inline assembly
instances.

> > the more that using the
> > assembly macro in an inline assembly instance causes the following
> > error on llvm based toolchains:
> > 
> > <instantiation>:1:1: error: invalid symbol redefinition
> > .L0_orig_s: call *genapic+64(%rip); .L0_orig_e: .L0_diff = (.L0_repl_e1 - 
> > .L0_repl_s1) - (...
> 
> The understanding I get is that clang doesn't properly support the
> \@ construct, expanding it to zero every time.

Yes, that's my understanding also. I've already filled a bug report:

https://bugs.llvm.org/show_bug.cgi?id=42034

> That's a clang bug
> imo, and hence the wording here should reflect this, rather than
> suggesting the code is broken.(I seem to vaguely recall an issue
> with clang instantiating a new assembly environment every time
> it encounters an asm().)

IIRC I've fixed that one upstream quite some time ago, and should be
fixed in versions >= 6.

> Without clang fixed, and with us wanting
> to be able to continue to build with clang, this then voids the entire
> purpose of f850619692 ("x86/alternatives: allow using assembler
> macros in favor of C ones"), which irc was originally part of the
> series, but went in much ahead of it.

I can look into workarounds to this, the one that comes to mind is
using .altmacro and LOCAL in order to create unique labels in the
macro. I can test if such approach would work if the plan is to only
rely on the assembly alternative code.

> 
> > --- a/xen/include/asm-x86/alternative.h
> > +++ b/xen/include/asm-x86/alternative.h
> > @@ -202,9 +202,8 @@ extern void alternative_branches(void);
> >      rettype ret_;                                                  \
> >      register unsigned long r10_ asm("r10");                        \
> >      register unsigned long r11_ asm("r11");                        \
> > -    asm volatile (__stringify(ALTERNATIVE "call *%c[addr](%%rip)", \
> > -                                          "call .",                \
> > -                                          X86_FEATURE_ALWAYS)      \
> > +    asm volatile (ALTERNATIVE("call *%c[addr](%%rip)", "call .",   \
> > +                              X86_FEATURE_ALWAYS)                  \
> >                    : ALT_CALL ## n ## _OUT, "=a" (ret_),            \
> >                      "=r" (r10_), "=r" (r11_) ASM_CALL_CONSTRAINT   \
> >                    : [addr] "i" (&(func)), "g" (func)               \
> 
> Okay, luckily the code change itself is simple enough, so it really
> wasn't that I had to use the variant used to make things work at
> all.

Since the only change requested is related to the commit message,
would you be OK to update the commit message to:

---8<---
x86: remove alternative_callN usage of ALTERNATIVE asm macro

alternative_callN using inline assembly to generate the alternative
patch sites should be using the ALTERNATIVE C preprocessor macro
rather than the ALTERNATIVE assembly macro, the more that using the
assembly macro in an inline assembly instance triggers the following
bug on llvm based toolchains:

<instantiation>:1:1: error: invalid symbol redefinition
.L0_orig_s: call *genapic+64(%rip); .L0_orig_e: .L0_diff = (.L0_repl_e1 - .L0_repl_s1) - (...
^
<instantiation>:1:37: error: invalid symbol redefinition
.L0_orig_s: call *genapic+64(%rip); .L0_orig_e: .L0_diff = (.L0_repl_e1 - .L0_repl_s1) - (...
                                    ^
<instantiation>:1:60: error: invalid reassignment of non-absolute variable '.L0_diff'
.L0_orig_s: call *genapic+64(%rip); .L0_orig_e: .L0_diff = (.L0_repl_e1 - .L0_repl_s1) - (...
                                                           ^
<inline asm>:1:2: note: while in macro instantiation
        ALTERNATIVE "call *genapic+64(%rip)", "call .", X86_FEATURE_LM
        ^
<instantiation>:1:156: error: invalid symbol redefinition
  ...- (.L0_orig_e - .L0_orig_s); mknops ((-(.L0_diff > 0)) * .L0_diff); .L0_orig_p:
                                                                         ^
<instantiation>:18:5: error: invalid symbol redefinition
    .L0_repl_s1: call .; .L0_repl_e1:
    ^
<instantiation>:18:26: error: invalid symbol redefinition
    .L0_repl_s1: call .; .L0_repl_e1:
                         ^
<instantiation>:1:1: error: invalid symbol redefinition
.L0_orig_s: call *genapic+64(%rip); .L0_orig_e: .L0_diff = (.L0_repl_e1 - .L0_repl_s1) - (...
^
<instantiation>:1:37: error: invalid symbol redefinition
.L0_orig_s: call *genapic+64(%rip); .L0_orig_e: .L0_diff = (.L0_repl_e1 - .L0_repl_s1) - (...
                                    ^
<instantiation>:1:60: error: invalid reassignment of non-absolute variable '.L0_diff'
.L0_orig_s: call *genapic+64(%rip); .L0_orig_e: .L0_diff = (.L0_repl_e1 - .L0_repl_s1) - (...
                                                           ^
<inline asm>:1:2: note: while in macro instantiation
        ALTERNATIVE "call *genapic+64(%rip)", "call .", X86_FEATURE_LM
        ^
<instantiation>:1:156: error: invalid symbol redefinition
  ...- (.L0_orig_e - .L0_orig_s); mknops ((-(.L0_diff > 0)) * .L0_diff); .L0_orig_p:
                                                                         ^
<instantiation>:18:5: error: invalid symbol redefinition
    .L0_repl_s1: call .; .L0_repl_e1:
    ^
<instantiation>:18:26: error: invalid symbol redefinition
    .L0_repl_s1: call .; .L0_repl_e1:
                         ^
<instantiation>:1:1: error: invalid symbol redefinition
.L0_orig_s: call *genapic+64(%rip); .L0_orig_e: .L0_diff = (.L0_repl_e1 - .L0_repl_s1) - (...
^
<instantiation>:1:37: error: invalid symbol redefinition
.L0_orig_s: call *genapic+64(%rip); .L0_orig_e: .L0_diff = (.L0_repl_e1 - .L0_repl_s1) - (...
                                    ^
<instantiation>:1:60: error: invalid reassignment of non-absolute variable '.L0_diff'
.L0_orig_s: call *genapic+64(%rip); .L0_orig_e: .L0_diff = (.L0_repl_e1 - .L0_repl_s1) - (...
                                                           ^
<inline asm>:1:2: note: while in macro instantiation
        ALTERNATIVE "call *genapic+64(%rip)", "call .", X86_FEATURE_LM
        ^
<instantiation>:1:156: error: invalid symbol redefinition
  ...- (.L0_orig_e - .L0_orig_s); mknops ((-(.L0_diff > 0)) * .L0_diff); .L0_orig_p:
                                                                         ^
<instantiation>:18:5: error: invalid symbol redefinition
    .L0_repl_s1: call .; .L0_repl_e1:
    ^
<instantiation>:18:26: error: invalid symbol redefinition
    .L0_repl_s1: call .; .L0_repl_e1:
                         ^

This is a bug in llvm that needs to be fixed before switching to use
the alternative assembly macros in inline assembly call sites.

Fixes: 67d01cdb5 ("x86: infrastructure to allow converting certain indirect calls to direct ones")
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel