Re: module: fix module_refcount() return when running in a module exit routine

[James Bottomley]

On Thu, 2015-01-22 at 08:50 -0800, Christoph Hellwig wrote:
> On Tue, Jan 20, 2015 at 09:23:35AM -0800, James Bottomley wrote:
> > On Tue, 2015-01-20 at 11:15 +1030, Rusty Russell wrote:
> > > James Bottomley <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> writes:
> > > > On Mon, 2015-01-19 at 16:21 +1030, Rusty Russell wrote:
> > > >> Masami Hiramatsu <masami.hiramatsu.pt@xxxxxxxxxxx> writes:
> > > >> > (2015/01/19 1:55), James Bottomley wrote:
> > > >> >> From: James Bottomley <JBottomley@xxxxxxxxxxxxx>
> > > >> >> 
> > > >> >> After e513cc1 module: Remove stop_machine from module unloading,
> > > >> >> module_refcount() is returning (unsigned long)-1 when called from within
> > > >> >> a routine that runs in module_exit.  This is confusing the scsi device
> > > >> >> put code which is coded to detect a module_refcount() of zero for
> > > >> >> running within a module exit routine and not try to do another
> > > >> >> module_put.  The fix is to restore the original behaviour of
> > > >> >> module_refcount() and return zero if we're running inside an exit
> > > >> >> routine.
> > > >> >> 
> > > >> >> Fixes: e513cc1c07e2ab93a4514eec9833e031df3e30bb
> > > >> >> Reported-by: Bart Van Assche <bvanassche@xxxxxxx>
> > > >> >> Signed-off-by: James Bottomley <JBottomley@xxxxxxxxxxxxx>
> > > >> >> 
> > > >> >
> > > >> > Yes, this should be fixed as you said, since it must return "unsigned long" value.
> > > >> 
> > > >> But there are only three non-module callers:
> > > >> 
> > > >> drivers/scsi/scsi.c:1012:	if (module && module_refcount(module) != 0)
> > > >> drivers/staging/lustre/lustre/obdclass/lu_object.c:1359:			LINVRNT(module_refcount(key->lct_owner) > 0);
> > > >> include/linux/module.h:447:unsigned long module_refcount(struct module *mod);
> > > >> kernel/debug/kdb/kdb_main.c:2026:		kdb_printf("%4ld ", module_refcount(mod));
> > > >> kernel/module.c:775:unsigned long module_refcount(struct module *mod)
> > > >> kernel/module.c:779:EXPORT_SYMBOL(module_refcount);
> > > >> kernel/module.c:859:	seq_printf(m, " %lu ", module_refcount(mod));
> > > >> kernel/module.c:911:	return sprintf(buffer, "%lu\n", module_refcount(mk->mod));
> > > >> 
> > > >> The first one I think should be eliminated, and the second one is simply
> > > >> an assertion before calling module_put() (which should probably be
> > > >> eliminated).  The others are just printing information.
> > > >
> > > > If you really want to insist on module_reference() returning -1 when the
> > > > module is in it's exit phase, OK, but in that case, I think it should
> > > > return a signed value, not an unsigned one.
> > > 
> > > Sure; I just didn't want to paper over the problem here.  And I'm not
> > > sure we want to lose information, eg. in kgdb we're presumably looking
> > > at it because something went wrong...
> > > 
> > > Thanks,
> > > Rusty.
> > > 
> > > Subject: module: make module_refcount() a signed integer.
> > > 
> > > James Bottomley points out that it will be -1 during unload.  It's
> > > only used for diagnostics, so let's not hide that as it could be a
> > > clue as to what's gone wrong.
> > > 
> > > Cc: Jason Wessel <jason.wessel@xxxxxxxxxxxxx>
> > > Signed-off-by: Rusty Russell <rusty@xxxxxxxxxxxxxxx>
> > > 
> > > diff --git a/include/linux/module.h b/include/linux/module.h
> > > index ebfb0e153c6a..b653d7c0a05a 100644
> > > --- a/include/linux/module.h
> > > +++ b/include/linux/module.h
> > > @@ -444,7 +444,7 @@ extern void __module_put_and_exit(struct module *mod, long code)
> > >  #define module_put_and_exit(code) __module_put_and_exit(THIS_MODULE, code)
> > >  
> > >  #ifdef CONFIG_MODULE_UNLOAD
> > > -unsigned long module_refcount(struct module *mod);
> > > +int module_refcount(struct module *mod);
> > >  void __symbol_put(const char *symbol);
> > >  #define symbol_put(x) __symbol_put(VMLINUX_SYMBOL_STR(x))
> > >  void symbol_put_addr(void *addr);
> > > diff --git a/kernel/debug/kdb/kdb_main.c b/kernel/debug/kdb/kdb_main.c
> > > index f191bddf64b8..7b40c5f07dce 100644
> > > --- a/kernel/debug/kdb/kdb_main.c
> > > +++ b/kernel/debug/kdb/kdb_main.c
> > > @@ -2023,7 +2023,7 @@ static int kdb_lsmod(int argc, const char **argv)
> > >  		kdb_printf("%-20s%8u  0x%p ", mod->name,
> > >  			   mod->core_size, (void *)mod);
> > >  #ifdef CONFIG_MODULE_UNLOAD
> > > -		kdb_printf("%4ld ", module_refcount(mod));
> > > +		kdb_printf("%4d ", module_refcount(mod));
> > >  #endif
> > >  		if (mod->state == MODULE_STATE_GOING)
> > >  			kdb_printf(" (Unloading)");
> > > diff --git a/kernel/module.c b/kernel/module.c
> > > index 3965511ae133..2387c98347c1 100644
> > > --- a/kernel/module.c
> > > +++ b/kernel/module.c
> > > @@ -772,9 +772,9 @@ static int try_stop_module(struct module *mod, int flags, int *forced)
> > >  	return 0;
> > >  }
> > >  
> > > -unsigned long module_refcount(struct module *mod)
> > > +int module_refcount(struct module *mod)
> > >  {
> > > -	return (unsigned long)atomic_read(&mod->refcnt) - MODULE_REF_BASE;
> > > +	return atomic_read(&mod->refcnt) - MODULE_REF_BASE;
> > >  }
> > >  EXPORT_SYMBOL(module_refcount);
> > >  
> > > @@ -856,7 +856,7 @@ static inline void print_unload_info(struct seq_file *m, struct module *mod)
> > >  	struct module_use *use;
> > >  	int printed_something = 0;
> > >  
> > > -	seq_printf(m, " %lu ", module_refcount(mod));
> > > +	seq_printf(m, " %i ", module_refcount(mod));
> > >  
> > >  	/*
> > >  	 * Always include a trailing , so userspace can differentiate
> > > @@ -908,7 +908,7 @@ EXPORT_SYMBOL_GPL(symbol_put_addr);
> > >  static ssize_t show_refcnt(struct module_attribute *mattr,
> > >  			   struct module_kobject *mk, char *buffer)
> > >  {
> > > -	return sprintf(buffer, "%lu\n", module_refcount(mk->mod));
> > > +	return sprintf(buffer, "%i\n", module_refcount(mk->mod));
> > >  }
> > 
> > Actually, I don't think this is enough.  Some Australian once came up
> > with a guide to APIs, and lectured on it at length, one of which was
> > that the name should be the obvious use and it is unexpected that a
> > refcount would go negative.  I think we could raise it from -6 on the
> > API scale to +3 if we add some documentation like below.
> 
> We'll also still need to change scsi_device_put to deal with
> a negative refcount..

I don't believe so ... we never call module_refcount() now, and the
actual module ref count never goes negative.  That's the point of
Rusty's change.  In the unload routines, the actual module refcount is
zero.  __module_get() will increment this and the final module_put()
will decrement it without triggering a warning.

In the old code, we expected try_module_get() to fail but then used
module_refcount() to detect the failure and not do a put.

James


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