Re: [PATCH 2/3] jump label: move jump table to r/w section

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Steven Rostedt (rostedt@goodmis.org) wrote:
> On Tue, 2010-11-23 at 18:55 -0500, Mathieu Desnoyers wrote:
> > * Jason Baron (jbaron@redhat.com) wrote:
> > > Since we writing the jump table it should be be in R/W kernel
> > > section. Move it to DATA_DATA
> > > 
> > > Signed-off-by: Jason Baron <jbaron@redhat.com>
> > > ---
> > >  include/asm-generic/vmlinux.lds.h |   14 ++++----------
> > >  1 files changed, 4 insertions(+), 10 deletions(-)
> > > 
> > > diff --git a/include/asm-generic/vmlinux.lds.h b/include/asm-generic/vmlinux.lds.h
> > > index bd69d79..9ca894d 100644
> > > --- a/include/asm-generic/vmlinux.lds.h
> > > +++ b/include/asm-generic/vmlinux.lds.h
> > > @@ -161,6 +161,10 @@
> > >  	VMLINUX_SYMBOL(__start___tracepoints) = .;			\
> > >  	*(__tracepoints)						\
> > >  	VMLINUX_SYMBOL(__stop___tracepoints) = .;			\
> > > +	. = ALIGN(8);							\
> > 
> > Past churn with various architectures and compiler with tracepoints,
> > markers and immediate values lead me to hint at the following approach
> > for jump label structure alignment:
> > 
> > . = ALIGN(32);
> > 
> > and to modify jump_label.h to have:
> > 
> > struct jump_entry {
> > 	jump_label_t code;
> > 	jump_label_t target;
> > 	jump_label_t key;
> > } __attribute__((aligned(32)));
> > 
> > Otherwise, the compiler is free to choose on which value it prefers to
> > align the jump_entry structures, which might not match the address at
> > which the linker scripts puts the beginning of the jump table.
> > 
> > In this case, given that we put put the jump label table after the
> > tracepoint table, we should be already aligned on 32 bytes. But I would
> > recommend to put the . = ALIGN(32) in the linker script anyway, just for
> > documentation purpose (and it should not add any padding in this case).
> > 
> > This is not a problem introduced by this patch, it also applies to the
> > current jump label code.
> 
> Agreed, but this change could probably wait for 2.6.38.
> 
> Also, if this is done, then an it should be wrapped in a
> #ifdef CONFIG_JUMP_LABEL and only inserted if we are using jump labels.
> Otherwise we may add a 32 byte hole for nothing. I know it's small, but
> why waste it if you don't need to.

The lack of proper alignment did cause nasty hard-to-identify/reproduce
regressions based on the compiler used, target platform and data layout.
It's up to you, but I'd be much more comfortable merging this in 2.6.37.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com