Re: [PATCH 1/2] kmemleak: Inform kmemleak about kernel stack allocation

[Catalin Marinas <catalin.marinas@arm.com>]

On Mon, 2009-08-31 at 10:45 +0200, Ingo Molnar wrote:
> * Catalin Marinas <catalin.marinas@arm.com> wrote:
> 
> > On Sat, 2009-08-29 at 15:29 +0200, Ingo Molnar wrote:
> > > * Catalin Marinas <catalin.marinas@arm.com> wrote:
[...]
> > > > diff --git a/arch/x86/include/asm/thread_info.h b/arch/x86/include/asm/thread_info.h
> > > > index fad7d40..f26432a 100644
> > > > --- a/arch/x86/include/asm/thread_info.h
> > > > +++ b/arch/x86/include/asm/thread_info.h
> > > > @@ -162,7 +162,12 @@ struct thread_info {
> > > >  #define __HAVE_ARCH_THREAD_INFO_ALLOCATOR
> > > >  
> > > >  #define alloc_thread_info(tsk)						\
> > > > -	((struct thread_info *)__get_free_pages(THREAD_FLAGS, THREAD_ORDER))
> > > > +({									\
> > > > +	struct thread_info *ti = (struct thread_info *)			\
> > > > +		__get_free_pages(THREAD_FLAGS, THREAD_ORDER);		\
> > > > +	kmemleak_alloc(ti, THREAD_SIZE, 1, THREAD_FLAGS);		\
> > > > +	ti;								\
> > > > +})
> > > 
> > > Sidenote:this used to be a trivial wrapper to gfp so it was 
> > > borderline OK as a CPP macro - now it's a non-trivial CPP wrapper 
> > > macro which is not OK. Mind converting it to an inline function?
> > 
> > I tried this first but got compilation errors in files that didn't 
> > even call this function. To make it workable, thread_info.h would 
> > need to include additional headers. If that's acceptable, I can 
> > post an updated patch.
> 
> I havent tried the patch myself, but by your description those build 
> problems seem to be pre-existing include file dependency problems 
> that should be tracked down and resolved - instead of widening them 
> by adding even more hidden dependencies via CPP macros.

I tried to move to an inline function and linux/gfp.h is needed for
__get_free_pages() and GFP_* macros. This leads to some complicated
circular dependencies like below:

  CC      arch/x86/kernel/asm-offsets.s
In file included from /work/Linux/2.6/linux-2.6-arm/include/linux/mmzone.h:9,
                 from /work/Linux/2.6/linux-2.6-arm/include/linux/gfp.h:4,
                 from /work/Linux/2.6/linux-2.6-arm/arch/x86/include/asm/thread_info.h:22,
                 from /work/Linux/2.6/linux-2.6-arm/include/linux/thread_info.h:56,
                 from /work/Linux/2.6/linux-2.6-arm/include/linux/preempt.h:9,
                 from /work/Linux/2.6/linux-2.6-arm/include/linux/spinlock.h:50,
                 from /work/Linux/2.6/linux-2.6-arm/include/linux/seqlock.h:29,
                 from /work/Linux/2.6/linux-2.6-arm/include/linux/time.h:8,
                 from /work/Linux/2.6/linux-2.6-arm/include/linux/stat.h:60,
                 from /work/Linux/2.6/linux-2.6-arm/include/linux/module.h:10,
                 from /work/Linux/2.6/linux-2.6-arm/include/linux/crypto.h:21,
                 from /work/Linux/2.6/linux-2.6-arm/arch/x86/kernel/asm-offsets_32.c:7,
                 from /work/Linux/2.6/linux-2.6-arm/arch/x86/kernel/asm-offsets.c:2:
include/linux/wait.h|51| error: expected specifier-qualifier-list before ‘spinlock_t’

The linux/mmzone.h file normally includes linux/spinlock.h but the
reverse is also true when the latter includes the former via
thread_info.h and gfp.h. Because of the (correct) guards in the header
files, the spinlock_t definition is no longer available in mmzone.h.


Anyway, what I'd like with this patch is to reduce the latency caused by
holding the tasklist_lock while scanning the stacks and also allow
rescheduling. If you have 500 threads in a system with 8K stacks,
kmemleak would need to scan about 4MB of stacks. While I think I can use
rcu_read_lock/unlock around do_each_thread..while_each_thread,
scheduling still isn't possible.

An alternative would be to traverse the tasks list for every stack
scanned. This traversing should be negligible compared to a stack
scanning time (looking up pointers in the prio tree).

Any opinion/suggestion here?

Thanks.

-- 
Catalin