Date:	Thu, 23 Sep 2010 15:11:44 -0400
From:	Jason Baron <jbaron@...hat.com>
To:	Mathieu Desnoyers <compudj@...stal.dyndns.org>
Cc:	Steven Rostedt <rostedt@...dmis.org>, linux-kernel@...r.kernel.org,
	Ingo Molnar <mingo@...e.hu>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Frederic Weisbecker <fweisbec@...il.com>,
	Andi Kleen <andi@...stfloor.org>,
	David Miller <davem@...emloft.net>,
	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
	Rusty Russell <rusty@...tcorp.com.au>,
	Masami Hiramatsu <masami.hiramatsu.pt@...achi.com>
Subject: Re: [PATCH 03/11] jump label: Base patch for jump label

On Thu, Sep 23, 2010 at 02:55:40PM -0400, Mathieu Desnoyers wrote:
> * Jason Baron (jbaron@...hat.com) wrote:
> > On Thu, Sep 23, 2010 at 11:48:52AM -0400, Mathieu Desnoyers wrote:
> > > * Jason Baron (jbaron@...hat.com) wrote:
> > > > On Thu, Sep 23, 2010 at 10:37:58AM -0400, Mathieu Desnoyers wrote:
> > > > > * Steven Rostedt (rostedt@...dmis.org) wrote:
> > > > > > From: Jason Baron <jbaron@...hat.com>
> > > > > > 
> > > > > > base patch to implement 'jump labeling'. Based on a new 'asm goto' inline
> > > > > > assembly gcc mechanism, we can now branch to labels from an 'asm goto'
> > > > > > statment. This allows us to create a 'no-op' fastpath, which can subsequently
> > > > > > be patched with a jump to the slowpath code. This is useful for code which
> > > > > > might be rarely used, but which we'd like to be able to call, if needed.
> > > > > > Tracepoints are the current usecase that these are being implemented for.
> > > > > > 
> > > > > [...]
> > > > > > +/***
> > > > > > + * jump_label_update - update jump label text
> > > > > > + * @key -  key value associated with a a jump label
> > > > > > + * @type - enum set to JUMP_LABEL_ENABLE or JUMP_LABEL_DISABLE
> > > > > > + *
> > > > > > + * Will enable/disable the jump for jump label @key, depending on the
> > > > > > + * value of @type.
> > > > > > + *
> > > > > > + */
> > > > > > +
> > > > > > +void jump_label_update(unsigned long key, enum jump_label_type type)
> > > > > > +{
> > > > > > +	struct jump_entry *iter;
> > > > > > +	struct jump_label_entry *entry;
> > > > > > +	struct hlist_node *module_node;
> > > > > > +	struct jump_label_module_entry *e_module;
> > > > > > +	int count;
> > > > > > +
> > > > > > +	mutex_lock(&jump_label_mutex);
> > > > > > +	entry = get_jump_label_entry((jump_label_t)key);
> > > > > > +	if (entry) {
> > > > > > +		count = entry->nr_entries;
> > > > > > +		iter = entry->table;
> > > > > > +		while (count--) {
> > > > > > +			if (kernel_text_address(iter->code))
> > > > > 
> > > > > As I pointed out in another thread, I'm concerned about the use of
> > > > > kernel_text_address without module mutex here. kernel_text_address calls
> > > > > is_module_text_address(), which calls __module_text_address() with
> > > > > preemption off.
> > > > > 
> > > > > __module_text_address() looks like:
> > > > > 
> > > > > struct module *__module_address(unsigned long addr)
> > > > > {
> > > > >         struct module *mod;
> > > > > 
> > > > >         if (addr < module_addr_min || addr > module_addr_max)
> > > > >                 return NULL;
> > > > > 
> > > > >         list_for_each_entry_rcu(mod, &modules, list)
> > > > >                 if (within_module_core(addr, mod)
> > > > >                     || within_module_init(addr, mod))
> > > > >                         return mod;
> > > > >         return NULL;
> > > > > }
> > > > > 
> > > > > struct module *__module_text_address(unsigned long addr)
> > > > > {
> > > > >         struct module *mod = __module_address(addr);
> > > > >         if (mod) {
> > > > >                 /* Make sure it's within the text section. */
> > > > >                 if (!within(addr, mod->module_init, mod->init_text_size)
> > > > >                     && !within(addr, mod->module_core, mod->core_text_size))
> > > > >                         mod = NULL;
> > > > >         }
> > > > >         return mod;
> > > > > }
> > > > > 
> > > > > So the test for the address being in the module core is already
> > > > > problematic, since we hold preempt off only within
> > > > > is_module_text_address(). The is_module_text_address() caller is then
> > > > > free to write to this address even after the module has been unloaded
> > > > > and the module unload grace period ended.
> > > > > 
> > > > > Even worse, such grace period is not waited for at module load time
> > > > > within:
> > > > > 
> > > > > init_module()
> > > > >        module_free(mod, mod->module_init);
> > > > >        mod->module_init = NULL;
> > > > >        mod->init_size = 0;
> > > > >        mod->init_text_size = 0;
> > > > >   (done with module_mutex held, while the module is already in the
> > > > >    module list)
> > > > > 
> > > > > We'd probably have to hold the module mutex around the
> > > > > is_module_text_address() call and address use (which can be a pain), or
> > > > > to correctly address this part of init_module() with RCU and require
> > > > > that preempt off is held across both __module_text_address() call site
> > > > > and the actual use of that pointer (which does not fit with jump label,
> > > > > which need to sleep, so we'd have to move module.c to a preemptable
> > > > > rcu_read_lock/synchronize_rcu() C.S.).
> > > > > 
> > > > > Thoughts ?
> > > > > 
> > > > 
> > > > I was thinking about the rcu_read_lock/synchronize_rcu() for this race.
> > > > We can hold the rcu_read_lock() across the is_module_text_address()
> > > > check in the jump label code, and then we can do in module.c:
> > > > 
> > > > mod->module_init = NULL;
> > > > synchronize_rcu();
> > > > module_free(mod, mod->module_init);
> > > 
> > > Beware that you need to copy the module_init address. Also make sure you
> > > audit the "module_free" (per-architecture) to make sure they don't use
> > > "mod" in ways you did not foresee.
> > > 
> > > > .
> > > > .
> > > > .
> > > > 
> > > > or we could push the rcu_read_lock() further down into
> > > > is_module_address()?
> > > 
> > > We need to pull rcu_read_lock further _up_. It needs to be held across
> > > both is_module_address() and the actual use of the address, otherwise
> > > the memory mapping can be removed underneath us.
> > > 
> > > You can see the rcu read lock as keeping the memory mapping alive for as
> > > long as the rcu read lock is held.
> > > 
> > > We'd also need to add a synchronize_rcu() in module removal.
> > > 
> > 
> > I agree that we this synchronization for the module __init section.
> > 
> > However, I believe we are ok for module removal case. free_module() is
> > called *after* blocking_notifier_call_chain() call. The
> > blocking_notifier_call_chain() is going to call back into the jump label
> > code, grab the jump_label_mutex and remove the reference to the module that
> > is about to freed. Thus, the jump label code can no longer reference it.
> 
> Ah, yes, this makes sense.
> 
> > 
> > So I think the following patch is all that is required here (lightly
> > tested).
> 
> Some comments below,
> 
> > 
> > Steve, I'll re-post as a separate patch, if we agree on this fix.
> > 
> > thanks,
> > 
> > -Jason
> > 
> > 
> > 
> > 
> > jump label: fix __init module section race
> > 
> > Jump label uses is_module_text_address() to ensure that the module
> > __init sections are valid before updating them. However, between the
> > check for a valid module __init section and the subsequent jump
> > label update, the module's __init section could be free out from under
> > us.
> > 
> > We fix this potential race putting the address check *and* the jump
> > label update under a rcu_read_lock(), and making sure a grace period
> > has completed before we free the __init section.
> > 
> > Thanks to Mathieu Desnoyers for pointing out this race condition.
> > 
> > Signed-off-by: Jason Baron <jbaron@...hat.com>
> 
> You can put my
> Reported-by: Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
> 
> > ---
> >  kernel/jump_label.c |    2 ++
> >  kernel/module.c     |    5 ++++-
> >  2 files changed, 6 insertions(+), 1 deletions(-)
> > 
> > diff --git a/kernel/jump_label.c b/kernel/jump_label.c
> > index f82878b..7830bfb 100644
> > --- a/kernel/jump_label.c
> > +++ b/kernel/jump_label.c
> > @@ -160,6 +160,7 @@ void jump_label_update(unsigned long key, enum jump_label_type type)
> >  			iter++;
> >  		}
> >  		/* eanble/disable jump labels in modules */
> 
> Separate patch for typo: eanble -> enable would be appropriate.
> 
> > +		rcu_read_lock();
> >  		hlist_for_each_entry(e_module, module_node, &(entry->modules),
> >  							hlist) {
> >  			count = e_module->nr_entries;
> > @@ -170,6 +171,7 @@ void jump_label_update(unsigned long key, enum jump_label_type type)
> >  				iter++;
> >  			}
> >  		}
> > +		rcu_read_unlock();
> 
> Please note that the impact of this read lock is that synchronize_rcu()
> cannot be called from text_poke anymore (because text_poke is called
> within a RCU read-side C.S.). So hopefully Masami did not plan to call
> synchronize_rcu() from the upcoming breakpoint-based text_poke.
> 

ok, we can re-visit this if it becomes an issue at some future
point...thanks for mentioning it.


> >  	}
> >  	mutex_unlock(&jump_label_mutex);
> >  }
> > diff --git a/kernel/module.c b/kernel/module.c
> > index eba1341..09f7e9e 100644
> > --- a/kernel/module.c
> > +++ b/kernel/module.c
> > @@ -2692,6 +2692,7 @@ SYSCALL_DEFINE3(init_module, void __user *, umod,
> >  		unsigned long, len, const char __user *, uargs)
> >  {
> >  	struct module *mod;
> > +	void *init_code;
> >  	int ret = 0;
> >  
> >  	/* Must have permission */
> > @@ -2749,8 +2750,10 @@ SYSCALL_DEFINE3(init_module, void __user *, umod,
> >  	mod->symtab = mod->core_symtab;
> >  	mod->strtab = mod->core_strtab;
> >  #endif
> > -	module_free(mod, mod->module_init);
> > +	init_code = mod->module_init;
> >  	mod->module_init = NULL;
> > +	synchronize_rcu();
> > +	module_free(mod, init_code);
> 
> I'm a bit concerned about the fact that "mod" is passed as parameter to
> module_free(). It does not seem to be used on x86, but have you reviewed
> all other architectures ? I would feel more comfortable if we removed
> this parameter from module_free() if it is indeed unused everywhere.
> 

good point.

So you're concern here is that mod->module_init is used in the
'module_free'? indeed i missed it on ia64:

void
module_free (struct module *mod, void *module_region)
{
        if (mod && mod->arch.init_unw_table &&
            module_region == mod->module_init) {
                unw_remove_unwind_table(mod->arch.init_unw_table);
                mod->arch.init_unw_table = NULL;
        }
        vfree(module_region);
}

So, I don't see why we couldn't s/mod->module_init/module_region, but
then I wonder why it wasn't written that way in the first place?

thanks,

-Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists