[Devel] Re: [PATCH 2/3] user.c: use kmem_cache_zalloc()

[Satyam Sharma]

On Fri, 21 Sep 2007, Andrew Morton wrote:
> 
> On Fri, 21 Sep 2007 13:39:06 +0400
> Alexey Dobriyan <adobriyan at sw.ru> wrote:
> 
> > Quite a few fields are zeroed during user_struct creation, so use
> > kmem_cache_zalloc() --  save a few lines and #ifdef. Also will help avoid
> >  #ifdef CONFIG_POSIX_MQUEUE in next patch.
> > 
> > Signed-off-by: Alexey Dobriyan <adobriyan at sw.ru>
> > ---
> > 
> >  kernel/user.c |   13 +------------
> >  1 file changed, 1 insertion(+), 12 deletions(-)
> > 
> > --- a/kernel/user.c
> > +++ b/kernel/user.c
> > @@ -129,21 +129,11 @@ struct user_struct * alloc_uid(struct user_namespace *ns, uid_t uid)
> >  	if (!up) {
> >  		struct user_struct *new;
> >  
> > -		new = kmem_cache_alloc(uid_cachep, GFP_KERNEL);
> > +		new = kmem_cache_zalloc(uid_cachep, GFP_KERNEL);
> >  		if (!new)
> >  			return NULL;
> >  		new->uid = uid;
> >  		atomic_set(&new->__count, 1);
> > -		atomic_set(&new->processes, 0);
> > -		atomic_set(&new->files, 0);
> > -		atomic_set(&new->sigpending, 0);
> > -#ifdef CONFIG_INOTIFY_USER
> > -		atomic_set(&new->inotify_watches, 0);
> > -		atomic_set(&new->inotify_devs, 0);
> > -#endif
> > -
> > -		new->mq_bytes = 0;
> > -		new->locked_shm = 0;
> 
> 
> This assumes that setting an atomic_t to the all-zeroes pattern is
> equivalent to atomic_set(v, 0).
> 
> This happens to be true for all present architectures, afaik.  But an
> architecture which has crappy primitives could quite legitimately implement
> its atomic_t as:
> 
> typedef struct {
> 	int counter;
> 	spinlock_t lock;
> } atomic_t;
> 
> in which case your assumption breaks.

Agreed, and this (implementing atomic ops using spinlocks) is already true
for the CRIS platform.

However, cris' implementation explicitly takes care to ensure that
atomic_t contains just a solitary int member, and no spinlock_t's
inside the atomic_t itself. [ include/asm-cris/arch-v32/atomic.h ]

Of course, that "128" limits scalability, so no more than 128 CPUs can be
executing atomic ops at any given instant of time, but admittedly I'm
getting anal here myself ... (but probably that's often perfectly the
right attitude to have too)


> So it's all a bit theoretical and a bit anal, and I'm sure we're making the
> same mistake in other places, but it's not a change I particularly like..

Hmm, it's borderline.

Such changes make text smaller (in terms of LOC as well vmlinux size).

But they also hurt grepping. Often we (at least I) want to grep for when
is a variable/struct member/etc getting initialized or getting
set/assigned to. Take this case, for example -- I bet it's important (for
overall logic) that those variables get initialized to zero. But *zalloc()
functions do that implicitly, so it wastes precious seconds or minutes of
developer time when grepping that code.

OTOH, we could make it standard practise to put a little comment on top
of such *zalloc() usages, explicitly enumerating the struct members that
that the *zalloc() is assumed to initialize to zero.

<runs away>


Satyam