clean up latest ipipe-tracing changes
PATCH 1/1 clean up latest ipipe-tracing changes
Hi,
here is a suggestion to do a tiny cleanup of Gilles' cpuid-fix for the tracer. Sorry that I didn't comment on this earlier. Nothing critical though.
I think it's better to drag the include/linux/compiler.h changes into the main patch instead of starting to use __attribute__((no_instrument_function)) all over the place just to avoid notrace (especially as I expect that Ingo's tracer will get merged anyway in the future). Moreover, local_irq_restore_hw_notrace & frieds are available even for !CONFIG_IPIPE_TRACE, so no need to differentiate here. Finally, I do not see a need for marking a static inline function non-traced.
The patch does not seem to produce different binary code (I checked kernel/ipipe/tracer.o) - as expected.
Jan
arch/i386/kernel/ipipe-root.c +1 -14
@@ -85,25 +85,13 @@ static void __ipipe_null_handler(unsigne#ifdef CONFIG_SMP -static __attribute__((no_instrument_function)) int __ipipe_boot_cpuid(void) +static notrace int __ipipe_boot_cpuid(void) {return 0 ; }u8 __ipipe_apicid_2_cpu[IPIPE_NR_CPUS ]; -#ifndef CONFIG_IPIPE_TRACE -static int __ipipe_hard_cpuid(void) -{ - unsigned long flags; - int cpu; - - local_irq_save_hw(flags); - cpu = __ipipe_apicid_2_cpu[GET_APIC_ID(apic_read(APIC_ID))]; - local_irq_restore_hw(flags); - return cpu; -} -#else /* CONFIG_IPIPE_TRACE */ static notrace int __ipipe_hard_cpuid(void) {unsigned long flags;
@@ -114,7 +102,6 @@ static notrace int __ipipe_hard_cpuid(volocal_irq_restore_hw_notrace (flags);return cpu; } -#endif /* CONFIG_IPIPE_TRACE */ int (*__ipipe_logical_cpuid)(void ) = &__ipipe_boot_cpuid;
include/asm-i386/ipipe.h +1 -1
@@ -97,7 +97,7 @@#define IPIPE_CRITICAL_VECTOR 0xf9/* Used by ipipe_critical_enter/exit() */ #define IPIPE_CRITICAL_IPI (IPIPE_CRITICAL_VECTOR - FIRST_EXTERNAL_VECTOR) -static inline __attribute__((no_instrument_function))int ipipe_processor_id (void ) +static inline int ipipe_processor_id (void ) {extern int (*__ipipe_logical_cpuid)(void );return __ipipe_logical_cpuid ();
Quoted 2 lines · Jan Kiszka
(...). Finally, I do not see a need for marking a static inline function non-traced. (...)
Marking a function inline does not guarantee that the said function will always be inlined, especially since with 2.6.16, inline no longer necessary means __attribute__((always_inline)).
Signature · 1 line
Gilles Chanteperdrix.
Quoted 8 lines · Gilles Chanteperdrix
Quoted 2 lines · Jan Kiszka
(...). Finally, I do not see a need for marking a static inline function non-traced. (...)
Marking a function inline does not guarantee that the said function will always be inlined, especially since with 2.6.16, inline no longer necessary means __attribute__((always_inline)).
Yep, correct, though this particular function should only be non-inlined by a confused compiler. So, shouldn't we rather convert this function into a macro or mark it as always_inline? A review of the tracer for further clarifications is required, I guess.
Jan
Quoted 10 lines · Jan Kiszka
Quoted 8 lines · Gilles Chanteperdrix
Quoted 2 lines · Jan Kiszka
(...). Finally, I do not see a need for marking a static inline function non-traced. (...)
Marking a function inline does not guarantee that the said function will always be inlined, especially since with 2.6.16, inline no longer necessary means __attribute__((always_inline)).
Quoted 4 lines
Yep, correct, though this particular function should only be non-inlined by a confused compiler. So, shouldn't we rather convert this function into a macro or mark it as
always_inline? A review of the tracer for further clarifications is required, I guess.
In my humble opinion, the function should not be instrumented. This way, the tracer will continue to work if the function happens to not been inlined.
Signature · 1 line
Gilles Chanteperdrix.
Philippe Gerum
AppliedQuoted 15 lines · Jan Kiszka
Hi,
here is a suggestion to do a tiny cleanup of Gilles' cpuid-fix for the tracer. Sorry that I didn't comment on this earlier. Nothing critical though.
I think it's better to drag the
include/linux/compiler.hchanges into the main patch instead of starting to use __attribute__((no_instrument_function)) all over the place just to avoid notrace (especially as I expect that Ingo's tracer will get merged anyway in the future). Moreover,local_irq_restore_hw_notrace& frieds are available even for !CONFIG_IPIPE_TRACE, so no need to differentiate here. Finally, I do not see a need for marking a static inline function non-traced.
Quoted 3 lines
The patch does not seem to produce different binary code (I checked
kernel/ipipe/tracer.o) - as expected.
Applied, leaving the ipipe_processor_id() unchanged though. Thanks.
Signature · 1 line
Philippe.
Quoted 16 lines · Philippe Gerum
Quoted 15 lines · Jan Kiszka
Hi,
here is a suggestion to do a tiny cleanup of Gilles' cpuid-fix for the tracer. Sorry that I didn't comment on this earlier. Nothing critical though.
I think it's better to drag the
include/linux/compiler.hchanges into the main patch instead of starting to use __attribute__((no_instrument_function)) all over the place just to avoid notrace (especially as I expect that Ingo's tracer will get merged anyway in the future). Moreover,local_irq_restore_hw_notrace& frieds are available even for !CONFIG_IPIPE_TRACE, so no need to differentiate here. Finally, I do not see a need for marking a static inline function non-traced.
Quoted 6 lines
Quoted 3 lines
The patch does not seem to produce different binary code (I checked
kernel/ipipe/tracer.o) - as expected.Applied, leaving the
ipipe_processor_id()unchanged though. Thanks.
I still think this makes more sense :) (otherwise, at least switch to "notrace").
Jan
include/asm-i386/ipipe.h +5 -5
@@ -97,11 +97,11 @@#define IPIPE_CRITICAL_VECTOR 0xf9/* Used by ipipe_critical_enter/exit() */ #define IPIPE_CRITICAL_IPI (IPIPE_CRITICAL_VECTOR - FIRST_EXTERNAL_VECTOR) -static inline __attribute__((no_instrument_function))int ipipe_processor_id (void ) -{ -extern int (*__ipipe_logical_cpuid)(void ); -return __ipipe_logical_cpuid (); -} +#define ipipe_processor_id () \ +({ \ + extern int (*__ipipe_logical_cpuid)(void); \ + __ipipe_logical_cpuid(); \ +})extern u8 __ipipe_apicid_2_cpu[];
Quoted 43 lines · Jan Kiszka
Quoted 17 lines · Philippe Gerum
Quoted 15 lines · Jan Kiszka
Hi,
here is a suggestion to do a tiny cleanup of Gilles' cpuid-fix for the tracer. Sorry that I didn't comment on this earlier. Nothing critical though.
I think it's better to drag the
include/linux/compiler.hchanges into the main patch instead of starting to use __attribute__((no_instrument_function)) all over the place just to avoid notrace (especially as I expect that Ingo's tracer will get merged anyway in the future). Moreover,local_irq_restore_hw_notrace& frieds are available even for !CONFIG_IPIPE_TRACE, so no need to differentiate here. Finally, I do not see a need for marking a static inline function non-traced.Quoted 6 lines
Quoted 3 lines
The patch does not seem to produce different binary code (I checked
kernel/ipipe/tracer.o) - as expected.Applied, leaving the
ipipe_processor_id()unchanged though. Thanks.I still think this makes more sense :) (otherwise, at least switch to "notrace").
Jan
------------------------------------------------------------------------
Index:
linux-2.6.16-ipipe/include/asm-i386/ipipe.h
===================================================================include/asm-i386/ipipe.h +0
-0@@ -97,11 +97,11 @@#define IPIPE_CRITICAL_VECTOR 0xf9/* Used by ipipe_critical_enter/exit() */ #define IPIPE_CRITICAL_IPI (IPIPE_CRITICAL_VECTOR - FIRST_EXTERNAL_VECTOR)
include/asm-i386/ipipe.h line 104
Quoted 10 lines
include/asm-i386/ipipe.h +5
-5-static inline __attribute__((no_instrument_function))int ipipe_processor_id (void ) -{ -extern int (*__ipipe_logical_cpuid)(void ); -return __ipipe_logical_cpuid (); -} +#define ipipe_processor_id () \ +({ \ + extern int (*__ipipe_logical_cpuid)(void); \ + __ipipe_logical_cpuid(); \ +})
Applied - moving the declaration out of the macro - Thanks.
Signature · 1 line
Philippe.