[PATCH 09/30] x86_64: ifdef out struct thread_struct::ip

[PATCH 09/30] x86_64: ifdef out struct thread_struct::ip

[Alexey Dobriyan]

struct thread_struct::ip isn't used on x86_64, struct pt_regs::ip is used
instead. 

kgdb should be reading 0, but I can't check it.

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
---

 arch/x86/include/asm/processor.h |    2 ++
 arch/x86/kernel/kgdb.c           |    2 +-
 2 files changed, 3 insertions(+), 1 deletion(-)

--- a/arch/x86/include/asm/processor.h
+++ b/arch/x86/include/asm/processor.h
@@ -421,7 +421,9 @@ struct thread_struct {
 	unsigned short		fsindex;
 	unsigned short		gsindex;
 #endif
+#ifdef CONFIG_X86_32
 	unsigned long		ip;
+#endif
 #ifdef CONFIG_X86_64
 	unsigned long		fs;
 #endif
--- a/arch/x86/kernel/kgdb.c
+++ b/arch/x86/kernel/kgdb.c
@@ -141,7 +141,7 @@ void sleeping_thread_to_gdb_regs(unsigned long *gdb_regs, struct task_struct *p)
 	gdb_regs32[GDB_PS]	= *(unsigned long *)(p->thread.sp + 8);
 	gdb_regs32[GDB_CS]	= __KERNEL_CS;
 	gdb_regs32[GDB_SS]	= __KERNEL_DS;
-	gdb_regs[GDB_PC]	= p->thread.ip;
+	gdb_regs[GDB_PC]	= 0;
 	gdb_regs[GDB_R8]	= 0;
 	gdb_regs[GDB_R9]	= 0;
 	gdb_regs[GDB_R10]	= 0;

Re: [PATCH 09/30] x86_64: ifdef out struct thread_struct::ip

[Matt Helsley]

On Fri, Apr 10, 2009 at 06:35:22AM +0400, Alexey Dobriyan wrote:
> struct thread_struct::ip isn't used on x86_64, struct pt_regs::ip is used
> instead. 
> 
> kgdb should be reading 0, but I can't check it.
> 
> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
> ---
> 
>  arch/x86/include/asm/processor.h |    2 ++
>  arch/x86/kernel/kgdb.c           |    2 +-
>  2 files changed, 3 insertions(+), 1 deletion(-)
> 
> --- a/arch/x86/include/asm/processor.h
> +++ b/arch/x86/include/asm/processor.h
> @@ -421,7 +421,9 @@ struct thread_struct {
>  	unsigned short		fsindex;
>  	unsigned short		gsindex;
>  #endif
> +#ifdef CONFIG_X86_32
>  	unsigned long		ip;
> +#endif
>  #ifdef CONFIG_X86_64
>  	unsigned long		fs;
>  #endif

Do these make struct thread_struct behave better in cachelines (smaller,
less aliasing)? Can we really fit more in the slab du jour?

Otherwise it seems like we're littering these structs with #ifdefs
and not really saving anything. If these #ifdefs don't save any space why not 
just put in a comment:

>  	unsigned long		ip; /* Used only on i386 */

Or maybe even:

	union {
	  	unsigned long		ip; /* Used only on i386 */
	  	unsigned long		fs; /* Used only on x86_64 */
	};

Cheers,
	-Matt Helsley

Re: [PATCH 09/30] x86_64: ifdef out struct thread_struct::ip

[Jaswinder Singh Rajput]

On Thu, 2009-04-09 at 20:53 -0700, Matt Helsley wrote:
> On Fri, Apr 10, 2009 at 06:35:22AM +0400, Alexey Dobriyan wrote:
> > struct thread_struct::ip isn't used on x86_64, struct pt_regs::ip is used
> > instead. 
> > 
> > kgdb should be reading 0, but I can't check it.
> > 
> > Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
> > ---
> > 
> >  arch/x86/include/asm/processor.h |    2 ++
> >  arch/x86/kernel/kgdb.c           |    2 +-
> >  2 files changed, 3 insertions(+), 1 deletion(-)
> > 
> > --- a/arch/x86/include/asm/processor.h
> > +++ b/arch/x86/include/asm/processor.h
> > @@ -421,7 +421,9 @@ struct thread_struct {
> >  	unsigned short		fsindex;
> >  	unsigned short		gsindex;
> >  #endif
> > +#ifdef CONFIG_X86_32
> >  	unsigned long		ip;
> > +#endif
> >  #ifdef CONFIG_X86_64
> >  	unsigned long		fs;
> >  #endif
> 
> Do these make struct thread_struct behave better in cachelines (smaller,
> less aliasing)? Can we really fit more in the slab du jour?
> 
> Otherwise it seems like we're littering these structs with #ifdefs
> and not really saving anything. If these #ifdefs don't save any space why not 
> just put in a comment:
> 
> >  	unsigned long		ip; /* Used only on i386 */
> 
> Or maybe even:
> 
> 	union {
> 	  	unsigned long		ip; /* Used only on i386 */
> 	  	unsigned long		fs; /* Used only on x86_64 */
> 	};
> 

Can we do it like this:
  	unsigned long		ip_fs; /* ip: i386, fs: x86_64 */

I am using same variable for both cases, or we can use some better name than ip_fs.
I am assuming either it is i386 or x86_64 machine ;-)

--
JSR

Re: [PATCH 09/30] x86_64: ifdef out struct thread_struct::ip

[Ingo Molnar]

* Matt Helsley <matthltc@us.ibm.com> wrote:

> On Fri, Apr 10, 2009 at 06:35:22AM +0400, Alexey Dobriyan wrote:
> > struct thread_struct::ip isn't used on x86_64, struct pt_regs::ip is used
> > instead. 
> > 
> > kgdb should be reading 0, but I can't check it.
> > 
> > Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
> > ---
> > 
> >  arch/x86/include/asm/processor.h |    2 ++
> >  arch/x86/kernel/kgdb.c           |    2 +-
> >  2 files changed, 3 insertions(+), 1 deletion(-)
> > 
> > --- a/arch/x86/include/asm/processor.h
> > +++ b/arch/x86/include/asm/processor.h
> > @@ -421,7 +421,9 @@ struct thread_struct {
> >  	unsigned short		fsindex;
> >  	unsigned short		gsindex;
> >  #endif
> > +#ifdef CONFIG_X86_32
> >  	unsigned long		ip;
> > +#endif
> >  #ifdef CONFIG_X86_64
> >  	unsigned long		fs;
> >  #endif
> 
> Do these make struct thread_struct behave better in cachelines 
> (smaller, less aliasing)? Can we really fit more in the slab du 
> jour?
> 
> Otherwise it seems like we're littering these structs with #ifdefs 
> and not really saving anything. [...]

Removing fields always saves memory (even if it does not show up 
currently due to allocators cache-aligning sizes).

But the #ifdef ugliness is a real worry.

> [...] If these #ifdefs don't save any 
> space why not just put in a comment:
> 
> >  	unsigned long		ip; /* Used only on i386 */

Yes.

> Or maybe even:
> 
> 	union {
> 	  	unsigned long		ip; /* Used only on i386 */
> 	  	unsigned long		fs; /* Used only on x86_64 */
> 	};

Maybe. If this ever gets misunderstood somewhere in platform code we 
will get ugly failure modes and zero compiler help.

	Ingo

Re: [PATCH 09/30] x86_64: ifdef out struct thread_struct::ip

[Ingo Molnar]

* Jaswinder Singh Rajput <jaswinder@kernel.org> wrote:

> On Thu, 2009-04-09 at 20:53 -0700, Matt Helsley wrote:
> > On Fri, Apr 10, 2009 at 06:35:22AM +0400, Alexey Dobriyan wrote:
> > > struct thread_struct::ip isn't used on x86_64, struct pt_regs::ip is used
> > > instead. 
> > > 
> > > kgdb should be reading 0, but I can't check it.
> > > 
> > > Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
> > > ---
> > > 
> > >  arch/x86/include/asm/processor.h |    2 ++
> > >  arch/x86/kernel/kgdb.c           |    2 +-
> > >  2 files changed, 3 insertions(+), 1 deletion(-)
> > > 
> > > --- a/arch/x86/include/asm/processor.h
> > > +++ b/arch/x86/include/asm/processor.h
> > > @@ -421,7 +421,9 @@ struct thread_struct {
> > >  	unsigned short		fsindex;
> > >  	unsigned short		gsindex;
> > >  #endif
> > > +#ifdef CONFIG_X86_32
> > >  	unsigned long		ip;
> > > +#endif
> > >  #ifdef CONFIG_X86_64
> > >  	unsigned long		fs;
> > >  #endif
> > 
> > Do these make struct thread_struct behave better in cachelines (smaller,
> > less aliasing)? Can we really fit more in the slab du jour?
> > 
> > Otherwise it seems like we're littering these structs with #ifdefs
> > and not really saving anything. If these #ifdefs don't save any space why not 
> > just put in a comment:
> > 
> > >  	unsigned long		ip; /* Used only on i386 */
> > 
> > Or maybe even:
> > 
> > 	union {
> > 	  	unsigned long		ip; /* Used only on i386 */
> > 	  	unsigned long		fs; /* Used only on x86_64 */
> > 	};
> > 
> 
> Can we do it like this:
>   	unsigned long		ip_fs; /* ip: i386, fs: x86_64 */
> 
> I am using same variable for both cases, or we can use some better 
> name than ip_fs. I am assuming either it is i386 or x86_64 machine 
> ;-)

This is the least clean variant amongst all the suggestions.

	Ingo

Re: [PATCH 09/30] x86_64: ifdef out struct thread_struct::ip

[Ingo Molnar]

* Ingo Molnar <mingo@elte.hu> wrote:

> 
> * Matt Helsley <matthltc@us.ibm.com> wrote:
> 
> > On Fri, Apr 10, 2009 at 06:35:22AM +0400, Alexey Dobriyan wrote:
> > > struct thread_struct::ip isn't used on x86_64, struct pt_regs::ip is used
> > > instead. 
> > > 
> > > kgdb should be reading 0, but I can't check it.
> > > 
> > > Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
> > > ---
> > > 
> > >  arch/x86/include/asm/processor.h |    2 ++
> > >  arch/x86/kernel/kgdb.c           |    2 +-
> > >  2 files changed, 3 insertions(+), 1 deletion(-)
> > > 
> > > --- a/arch/x86/include/asm/processor.h
> > > +++ b/arch/x86/include/asm/processor.h
> > > @@ -421,7 +421,9 @@ struct thread_struct {
> > >  	unsigned short		fsindex;
> > >  	unsigned short		gsindex;
> > >  #endif
> > > +#ifdef CONFIG_X86_32
> > >  	unsigned long		ip;
> > > +#endif
> > >  #ifdef CONFIG_X86_64
> > >  	unsigned long		fs;
> > >  #endif
> > 
> > Do these make struct thread_struct behave better in cachelines 
> > (smaller, less aliasing)? Can we really fit more in the slab du 
> > jour?
> > 
> > Otherwise it seems like we're littering these structs with #ifdefs 
> > and not really saving anything. [...]
> 
> Removing fields always saves memory (even if it does not show up 
> currently due to allocators cache-aligning sizes).

I mean: we always try to consider structure size reductions as if 
they saved real memory, even if they dont do so in the current 
layout and allocation patterns.

In reality only every 8th 8-byte reduction will result in a true 
64-byte cacheline reduction - but still we have to continue small 
reductions otherwise we wont ever get to the larger reductions in 
the first place.

I.e. Alexey's patch shows a real item worth checking, regardless of 
the current size and alignment scenario.

	Ingo

Re: [PATCH 09/30] x86_64: ifdef out struct thread_struct::ip

[Jaswinder Singh Rajput]

On Fri, 2009-04-10 at 11:20 +0200, Ingo Molnar wrote:
> * Jaswinder Singh Rajput <jaswinder@kernel.org> wrote:
> 
> > On Thu, 2009-04-09 at 20:53 -0700, Matt Helsley wrote:
> > > On Fri, Apr 10, 2009 at 06:35:22AM +0400, Alexey Dobriyan wrote:
> > > > struct thread_struct::ip isn't used on x86_64, struct pt_regs::ip is used
> > > > instead. 
> > > > 
> > > > kgdb should be reading 0, but I can't check it.
> > > > 
> > > > Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
> > > > ---
> > > > 
> > > >  arch/x86/include/asm/processor.h |    2 ++
> > > >  arch/x86/kernel/kgdb.c           |    2 +-
> > > >  2 files changed, 3 insertions(+), 1 deletion(-)
> > > > 
> > > > --- a/arch/x86/include/asm/processor.h
> > > > +++ b/arch/x86/include/asm/processor.h
> > > > @@ -421,7 +421,9 @@ struct thread_struct {
> > > >  	unsigned short		fsindex;
> > > >  	unsigned short		gsindex;
> > > >  #endif
> > > > +#ifdef CONFIG_X86_32
> > > >  	unsigned long		ip;
> > > > +#endif
> > > >  #ifdef CONFIG_X86_64
> > > >  	unsigned long		fs;
> > > >  #endif
> > > 
> > > Do these make struct thread_struct behave better in cachelines (smaller,
> > > less aliasing)? Can we really fit more in the slab du jour?
> > > 
> > > Otherwise it seems like we're littering these structs with #ifdefs
> > > and not really saving anything. If these #ifdefs don't save any space why not 
> > > just put in a comment:
> > > 
> > > >  	unsigned long		ip; /* Used only on i386 */
> > > 
> > > Or maybe even:
> > > 
> > > 	union {
> > > 	  	unsigned long		ip; /* Used only on i386 */
> > > 	  	unsigned long		fs; /* Used only on x86_64 */
> > > 	};
> > > 
> > 
> > Can we do it like this:
> >   	unsigned long		ip_fs; /* ip: i386, fs: x86_64 */
> > 
> > I am using same variable for both cases, or we can use some better 
> > name than ip_fs. I am assuming either it is i386 or x86_64 machine 
> > ;-)
> 
> This is the least clean variant amongst all the suggestions.
> 

Yes, this was a wakeful call for you.

I send dozen of emails in last 24 hours to you for your feedback.

I do not need reply for this email. Please send reply for other emails.

Thanks,

--
JSR

Re: [PATCH 09/30] x86_64: ifdef out struct thread_struct::ip

[Alexey Dobriyan]

On Fri, Apr 10, 2009 at 11:19:31AM +0200, Ingo Molnar wrote:
> 
> * Matt Helsley <matthltc@us.ibm.com> wrote:
> 
> > On Fri, Apr 10, 2009 at 06:35:22AM +0400, Alexey Dobriyan wrote:
> > > struct thread_struct::ip isn't used on x86_64, struct pt_regs::ip is used
> > > instead. 
> > > 
> > > kgdb should be reading 0, but I can't check it.
> > > 
> > > Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
> > > ---
> > > 
> > >  arch/x86/include/asm/processor.h |    2 ++
> > >  arch/x86/kernel/kgdb.c           |    2 +-
> > >  2 files changed, 3 insertions(+), 1 deletion(-)
> > > 
> > > --- a/arch/x86/include/asm/processor.h
> > > +++ b/arch/x86/include/asm/processor.h
> > > @@ -421,7 +421,9 @@ struct thread_struct {
> > >  	unsigned short		fsindex;
> > >  	unsigned short		gsindex;
> > >  #endif
> > > +#ifdef CONFIG_X86_32
> > >  	unsigned long		ip;
> > > +#endif
> > >  #ifdef CONFIG_X86_64
> > >  	unsigned long		fs;
> > >  #endif
> > 
> > Do these make struct thread_struct behave better in cachelines 
> > (smaller, less aliasing)? Can we really fit more in the slab du 
> > jour?
> > 
> > Otherwise it seems like we're littering these structs with #ifdefs 
> > and not really saving anything. [...]
> 
> Removing fields always saves memory (even if it does not show up 
> currently due to allocators cache-aligning sizes).
> 
> But the #ifdef ugliness is a real worry.

You should have thought about it when i386/x86_64 unification was
introduced.

pagefault code was full of ifdefs (it's less now), and this is trivial
ifdef in a header.

> > [...] If these #ifdefs don't save any 
> > space why not just put in a comment:
> > 
> > >  	unsigned long		ip; /* Used only on i386 */
> 
> Yes.
> 
> > Or maybe even:
> > 
> > 	union {
> > 	  	unsigned long		ip; /* Used only on i386 */
> > 	  	unsigned long		fs; /* Used only on x86_64 */
> > 	};
> 
> Maybe. If this ever gets misunderstood somewhere in platform code we 
> will get ugly failure modes and zero compiler help.