Re: [patch] block layer: kmemcheck fixes

[Jens Axboe <jens.axboe@oracle.com>]

On Thu, Feb 07 2008, Ingo Molnar wrote:
> 
> * Linus Torvalds <torvalds@linux-foundation.org> wrote:
> 
> > On Thu, 7 Feb 2008, Ingo Molnar wrote:
> > >  	INIT_HLIST_NODE(&rq->hash);
> > >  	RB_CLEAR_NODE(&rq->rb_node);
> > > -	rq->ioprio = 0;
> > > -	rq->buffer = NULL;
> > > -	rq->ref_count = 1;
> > > -	rq->q = q;
> > > -	rq->special = NULL;
> > > -	rq->data_len = 0;
> > > -	rq->data = NULL;
> > > -	rq->nr_phys_segments = 0;
> > > -	rq->sense = NULL;
> > > -	rq->end_io = NULL;
> > > -	rq->end_io_data = NULL;
> > > -	rq->completion_data = NULL;
> > > -	rq->next_rq = NULL;
> > > +	rq->completion_data		= NULL;
> > > +	/* rq->elevator_private			*/
> > > +	/* rq->elevator_private2		*/
> > > +	/* rq->rq_disk				*/
> > > +	/* rq->start_time			*/
> > > +	rq->nr_phys_segments		= 0;
> > > +	/* rq->nr_hw_segments			*/
> > > +	rq->ioprio			= 0;
> > > +	rq->special			= NULL;
> > > +	rq->buffer			= NULL;
> > ...
> > 
> > Can we please just stop doing these one-by-one assignments, and just do 
> > something like
> > 
> > 	memset(rq, 0, sizeof(*rq));
> > 	rq->q = q;
> > 	rq->ref_count = 1;
> > 	INIT_HLIST_NODE(&rq->hash);
> > 	RB_CLEAR_NODE(&rq->rb_node);
> > 
> > instead?
> > 
> > The memset() is likely faster and smaller than one-by-one assignments 
> > anyway, even if the one-by-ones can avoid initializing some field or 
> > there ends up being a double initialization..
> 
> i definitely agree and do that for all code i write.
> 
> But if someone does item by item initialization for some crazy 
> performance reason (networking folks tend to have such constructs), it 
> should be done i think how i've done it in the patch: by systematically 
> listing _every_ field in the structure, in the same order, and 
> indicating it clearly when it is not initialized and why.

That assumes that people find the references in two places when adding
members to a structure, not very likely (people are lazy!).

> and there it already shows that we do not initialize a few other members 
> that could cause problems later on:
> 
> +       rq->data_len                    = 0;
> +       /* rq->sense_len                        */
> +       rq->data                        = NULL;
> +       rq->sense                       = NULL;
> 
> why is sense_len not initialized - while data_len is? In any case, these 

because sense isn't set, when someone sets ->sense they should set
sense_len as well.

> days the memclear instructions are dirt cheap and we should just always 
> initialize everything to zero by default, especially if it's almost all 
> zero-initialized anyway.

Completely agree, some of these are just dormant bugs waiting to happen.
Clearing everything is the sanest approach.

-- 
Jens Axboe