Re: [PATCH 0/2] jump label: 2.6.38 updates

[Jason Baron <jbaron@redhat.com>]

On Mon, Feb 14, 2011 at 04:57:04PM +0100, Peter Zijlstra wrote:
> On Mon, 2011-02-14 at 10:51 -0500, Jason Baron wrote:
> > On Sat, Feb 12, 2011 at 07:47:45PM +0100, Peter Zijlstra wrote:
> > > On Fri, 2011-02-11 at 22:38 +0100, Peter Zijlstra wrote:
> > > > 
> > > > So why can't we make that jump_label_entry::refcount and
> > > > jump_label_key::state an atomic_t and be done with it? 
> > > 
> > > So I had a bit of a poke at this because I didn't quite understand why
> > > all that stuff was as it was. I applied both Jason's patches and then
> > > basically rewrote kernel/jump_label.c just for kicks ;-)
> > > 
> > > I haven't tried compiling this, let alone running it, but provided I
> > > didn't actually forget anything the storage per key is now 16 bytes when
> > > modules are disabled and 24 * (1 + mods) bytes for when they are
> > > enabled. The old code had 64 + 40 * mods bytes.
> > > 
> > > I still need to clean up the static_branch_else bits and look at !x86
> > > aside from the already mentioned bits.. but what do people think?
> > > 
> > > ---
> > 
> > Generally, I really like this! Its the direction I think the jump label
> > code should be going. The complete removal of the hash table, makes the
> > design a lot better and simpler. We just need to get some of the details
> > cleaned up, and of course we need this to compile :) But I don't see any
> > fundamental problems with this approach. 
> > 
> > Things that still need to be sorted out:
> > 
> > 1) Since jump_label.h, are included in kernel.h, (indirectly via the
> > dynamic_debug.h) the atomic_t definitions could be problematic, since
> > atomic.h includes kernel.h indirectly...so we might need some header
> > magic.
> 
> Yes, I remember running into that when I did the jump_label_ref stuff,
> some head-scratching is in order there.
> 

yes. i suspect this might be the hardest bit of this...

> > 2) I had some code to disallow writing to module __init section, by
> > setting the 'key' value to 0, after the module->init was run, but
> > before, the memory was freed. And then I check for a non-zero key value
> > when the jump label is updated. In this way we can't corrupt some random
> > piece of memory. I had this done via the 'MODULE_STATE_LIVE' notifier.
> 
> AH! I wondered what that was about.. that wouldn't work now since we
> actually rely on iter->key to remain what it was.
> 

we could just use iter->code, or iter->target -> 0 to indicate that the
entry is not valid, and leave iter->key as it is.