Re: [PATCH RFC V2 3/5] jump_label: if a key has already been initialized, don't nop it out

[Jason Baron <jbaron@redhat.com>]

On Mon, Oct 03, 2011 at 09:27:56AM -0700, Jeremy Fitzhardinge wrote:
> On 10/03/2011 08:02 AM, Jason Baron wrote:
> > Hi,
> >
> > (Sorry for the late reply - I was away for a few days).
> >
> > The early enable is really nice - it means there are not restrictions on
> > when jump_label_inc()/dec() can be called which is nice.
> >
> > comments below.
> >
> >
> > On Sat, Oct 01, 2011 at 02:55:35PM -0700, Jeremy Fitzhardinge wrote:
> >> From: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
> >>
> >> If a key has been enabled before jump_label_init() is called, don't
> >> nop it out.
> >>
> >> This removes arch_jump_label_text_poke_early() (which can only nop
> >> out a site) and uses arch_jump_label_transform() instead.
> >>
> >> Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
> >> ---
> >>  include/linux/jump_label.h |    3 ++-
> >>  kernel/jump_label.c        |   20 ++++++++------------
> >>  2 files changed, 10 insertions(+), 13 deletions(-)
> >>
> >> diff --git a/include/linux/jump_label.h b/include/linux/jump_label.h
> >> index 1213e9d..c8fb1b3 100644
> >> --- a/include/linux/jump_label.h
> >> +++ b/include/linux/jump_label.h
> >> @@ -45,7 +45,8 @@ extern void jump_label_lock(void);
> >>  extern void jump_label_unlock(void);
> >>  extern void arch_jump_label_transform(struct jump_entry *entry,
> >>  				 enum jump_label_type type);
> >> -extern void arch_jump_label_text_poke_early(jump_label_t addr);
> >> +extern void arch_jump_label_transform_early(struct jump_entry *entry,
> >> +				 enum jump_label_type type);
> >>  extern int jump_label_text_reserved(void *start, void *end);
> >>  extern void jump_label_inc(struct jump_label_key *key);
> >>  extern void jump_label_dec(struct jump_label_key *key);
> >> diff --git a/kernel/jump_label.c b/kernel/jump_label.c
> >> index a8ce450..059202d5 100644
> >> --- a/kernel/jump_label.c
> >> +++ b/kernel/jump_label.c
> >> @@ -121,13 +121,6 @@ static void __jump_label_update(struct jump_label_key *key,
> >>  	}
> >>  }
> >>  
> >> -/*
> >> - * Not all archs need this.
> >> - */
> >> -void __weak arch_jump_label_text_poke_early(jump_label_t addr)
> >> -{
> >> -}
> >> -
> >>  static __init int jump_label_init(void)
> >>  {
> >>  	struct jump_entry *iter_start = __start___jump_table;
> >> @@ -139,12 +132,15 @@ static __init int jump_label_init(void)
> >>  	jump_label_sort_entries(iter_start, iter_stop);
> >>  
> >>  	for (iter = iter_start; iter < iter_stop; iter++) {
> >> -		arch_jump_label_text_poke_early(iter->code);
> >> -		if (iter->key == (jump_label_t)(unsigned long)key)
> >> +		struct jump_label_key *iterk;
> >> +
> >> +		iterk = (struct jump_label_key *)(unsigned long)iter->key;
> >> +		arch_jump_label_transform(iter, jump_label_enabled(iterk) ?
> >> +					  JUMP_LABEL_ENABLE : JUMP_LABEL_DISABLE);
> > The only reason I called this at boot-time was that the 'ideal' x86
> > no-op isn't known until boot time. Thus, in the enabled case we could
> > skip the the arch_jump_label_transform() call. ie:
> >
> > if (!enabled)
> > 	arch_jump_label_transform(iter, JUMP_LABEL_DISABLE);
> 
> 
> Yep, fair enough.
> 
> >
> >
> >> +		if (iterk == key)
> >>  			continue;
> >>  
> >> -		key = (struct jump_label_key *)(unsigned long)iter->key;
> >> -		atomic_set(&key->enabled, 0);
> >> +		key = iterk;
> >>  		key->entries = iter;
> >>  #ifdef CONFIG_MODULES
> >>  		key->next = NULL;
> >> @@ -212,7 +208,7 @@ void jump_label_apply_nops(struct module *mod)
> >>  		return;
> >>  
> >>  	for (iter = iter_start; iter < iter_stop; iter++)
> >> -		arch_jump_label_text_poke_early(iter->code);
> >> +		arch_jump_label_transform(iter, JUMP_LABEL_DISABLE);
> >>  }
> >>  
> >>  static int jump_label_add_module(struct module *mod)
> >> -- 
> >> 1.7.6.2
> >>
> > hmmm...this is used on module load in smp - so this would introduce a number of
> > calls to stop_machine() where we didn't have them before. Yes, module
> > load is a very slow path to begin with, but I think its at least worth
> > pointing out...
> 
> Ah, that explains it - the module stuff certainly isn't "early" except -
> I guess - in the module's lifetime.
> 
> Well, I suppose I could introduce either second variant of the function,
> or add a "live" flag (ie, may be updating code that a processor is
> executing), which requires a stop_machine, or direct update if it doesn't.
> 
> But is there any reason why we couldn't just generate a reasonably
> efficient 5-byte atomic nop in the first place, and get rid of all that
> fooling around?  It looks like x86 is the only arch where it makes any
> difference at all, and how much difference does it really make?  Or is
> there no one 5-byte atomic nop that works on all x86 variants, aside
> from jmp +0?
> 
>     J

Yes, there are really two reasons for the initial no-op patching pass:

1) The jmp +0, is a 'safe' no-op that I know is going to initially
boot for all x86. I'm not sure if there is a 5-byte nop that works on
all x86 variants - but by using jmp +0, we make it much easier to debug
cases where we may be using broken no-ops.

2) This optimization is about as close to a 0 cost off case as possible.
I know there have been various no-op benchmarks posted on lkml in the
past, so the choice of no-op does seem to make a difference. see:
http://lkml.indiana.edu/hypermail/linux/kernel/0808.1/2416.html, for
example. So at least to me, if we are not using the lowest cost no-op,
we are at least in-part defeating the point of this optimization.

I like the "live" flag suggestion mentioned above. Less functions is
better, and non-x86 arches can simply ignore the flag.

Thanks,

-Jason