Re: [PATCH 1/7][QUOTA] Move sysctl management code under ifdefCONFIG_SYSCTL

[Andrew Morton]

On Tue, 04 Dec 2007 04:40:26 -0700 ebiederm@xxxxxxxxxxxx (Eric W. Biederman) wrote:

> Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> writes:
> 
> > On Tue, 04 Dec 2007 12:31:37 +0300 Pavel Emelyanov <xemul@xxxxxxxxxx> wrote:
> >
> >> Andrew Morton wrote:
> >> > On Tue, 04 Dec 2007 11:58:30 +0300 Pavel Emelyanov <xemul@xxxxxxxxxx> wrote:
> >> > 
> >> >>>> +#ifdef CONFIG_SYSCTL
> >> >>>>  	register_sysctl_table(sys_table);
> >> >>>> +#endif
> >> >>>>  
> >> >>>>  	dquot_cachep = kmem_cache_create("dquot",
> >> >>>> sizeof(struct dquot), sizeof(unsigned long) * 4,
> >> >>> We should avoid the ifdefs around the register_sysctl_table() call.
> >> >>>
> >> >>> At present the !CONFIG_SYSCTL implementation of register_sysctl_table() is
> >> >>> a non-inlined NULL-returning stub.  All we have to do is to inline that
> > stub
> >> >>> then these ifdefs can go away.
> >> >> What if some code checks for the return value to be not-NULL? In case
> >> >> CONFIG_SYSCTL=n this code will always think, that the registration failed.
> >> > 
> >> > The stub function should return success?
> >> 
> >> Well, I think yes. If some functionality is turned off, then the 
> >> caller should think that everything is going fine (or he should
> >> explicitly removes the call to it with some other ifdef). 
> >> 
> >> At least this is true for stubs that return the error code, not 
> >> the pointer. E.g. copy_semundo() always returns success if SYSVIPC 
> >> is off, or namespaces cloning routines act in a similar way.
> >> 
> >> Thus I though, that routines, that return pointers should better
> >> report that everything is OK (somehow) to reduce the number of 
> >> "helpers" in the outer code. No?
> >> 
> >
> > Dunno.  Returning NULL should be OK.  If anyone is dereferenceing that
> > pointer with CONFIG_SYSCTL=n then they might need some attention?  
> 
> We do have some current code in the network stack that fails miserably
> when  register_sysctl_table returns NULL, and there are explicit
> checks for that.

So that code would be failing today with CONFIG_SYSCTL=n?  Unless the
failing code is itself under #ifdef CONFIG_SYSCTL, in which case we don't
need to change anything?

> Grr.
> 
> I had forgotten about that.
> 
> I expect the right answer is to simply have code ignore the fact
> that register_sysctl_xxxx returns NULL, and not error on it.
> 
> The alternative is to get fancy and have everyone check the
> return code and make the return type an IS_ERR thing.  That seems
> a lot more trouble then it is worth.
> 
> We can probably define it as register_sysctl_xxxx always returns
> a token that must be passed to unregister_sysctl, and no errors
> will be reported except to dmesg.  That at sounds simple sane
> and supportable from where we are now.
> 
> Eric
> 
> 
--
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/