Re: Testers needed for LTTng 0.6.0pre1

* Christopher Yeoh ([email protected]) wrote:
> At 2006/9/27 15:54-0400  Mathieu Desnoyers writes:
> > 
> > I would like you to test it on various architectures (and check for warnings) to
> > see if I broke something when porting the LTTng instrumentation to markers.
> 
> The marker mechanism doesn't compile on ppc64. Neither the asm-powerpc
> or asm-generic version of marker.h work.
> 
> The problem appears to be in the macro:
> 
> 		asm volatile(	".section .markers, \"a\";\n\t" \
> 					".long %1, 0f, %2, %3;\n\t" \
> 					".previous;\n\t" \
> 					".align 32;\n\t" \
> 					"0:\n\t" \
> 					"li $0,%0;\n\t" \
> 				: "=r" (condition) : \
> 				"m" (*(#name)), \
> 				"m" (__mark_call_##name), \
> 				"m" (*format)); \
> 
> For the .long part, gcc ends up outputting something like:
> 
> long 0(9), 0f, 0(11), 0(3);
> 
> ie register relative addresses which the assembler barfs on.
> 
> Would something like the following be ok? (based on asm-generic version
> - I did try to do get something working based on the asm-powerpc
> version working, but so far without success):
> 
> #define MARK(name, format, args...) \
>         do { \
>                 static marker_probe_func *__mark_call_##name = \
>                                         __mark_empty_function; \
>                 volatile static char __marker_enable_##name = 0; \
>                 static struct _mark_marker_local { \
>                         const char *local_name; \
>                         volatile char *enable; \
>                         marker_probe_func **call; \
>                         const char *local_format; \
>                 } __attribute__((packed)) i __attribute__ ((section(".markers"))) \
> 		 __attribute__((__used__))  = \
>                     { #name, &__marker_enable_##name, &__mark_call_##name, format }; \
>                 if (unlikely(__marker_enable_##name)) { \
>                         preempt_disable(); \
>                         (*__mark_call_##name)(format, ## args); \
>                         preempt_enable_no_resched(); \
>                 } \
>         } while(0)
> 
> 

The idea is interesting, however, gcc will put the "enable" pointer as a unique
value when the same marker is used more than once in a given object. We won't be
able to patch the immediate value for each marker site then.


> There also seems to be a problem with fs/compat.c (~1514) with lttng_name
> not being declared.
> 
Ok thank, I will fix it.

I'll try to come with something for ppc asm..

Mathieu

> Chris
> -- 
> [email protected]
> IBM OzLabs Linux Development Group, ADL
> Canberra, Australia
> 
OpenPGP public key:              http://krystal.dyndns.org:8080/key/compudj.gpg
Key fingerprint:     8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68