Re: [PATCH v2] landlock: Use kmem for landlock_object

On Sun, Mar 31, 2024 at 11:38:28PM +0530, Ayush Tiwari wrote:
> On Sun, Mar 31, 2024 at 06:54:39PM +0200, Greg KH wrote:
> > On Sun, Mar 31, 2024 at 09:12:06PM +0530, Ayush Tiwari wrote:
> > > Hello Greg. Thanks for the feedback.
> > > On Sat, Mar 30, 2024 at 05:12:18PM +0100, Greg KH wrote:
> > > > On Sat, Mar 30, 2024 at 07:24:19PM +0530, Ayush Tiwari wrote:
> > > > > Use kmem_cache replace kzalloc() calls with kmem_cache_zalloc() for
> > > > > struct landlock_object and update the related dependencies to improve
> > > > > memory allocation and deallocation performance.
> > > > 
> > > > So it's faster?  Great, what are the measurements?
> > > > 
> > > Thank you for the feedback. Regarding the performance improvements, I
> > > realized I should have provided concrete measurements to support the
> > > claim. The intention behind switching to kmem_cache_zalloc() was to
> > > optimize memory management efficiency based on general principles.
> > > Could you suggest some way to get some measurements if possible?
> > 
> > If you can not measure the difference, why make the change at all?
> 
> Kindly refer to this issue: https://github.com/landlock-lsm/linux/issues/19
> I have been assigned this issue hence I am focussing on making the
> changes that have been listed.

As Greg asked, it would be good know the performance impact of such
change.  This could be measured by creating a lot of related allocations
and accessing them in non-sequential order (e.g. adding new rules,
accessing a related inode while being sandboxed).  I guess there will be
a lot of noise (because of other subsystems) but it's worth a try.  You
should look at similar commits and their related threads to see what
others did.

> > 
> > Again, you need to prove the need for this change, so far I fail to see
> > a reason why.
> > 
> > > > > +static struct kmem_cache *landlock_object_cache;
> > > > > +
> > > > > +void __init landlock_object_cache_init(void)
> > > > > +{
> > > > > +	landlock_object_cache = kmem_cache_create(
> > > > > +		"landlock_object_cache", sizeof(struct landlock_object), 0,
> > > > > +		SLAB_PANIC, NULL);
> > > > 
> > > > You really want SLAB_PANIC?  Why?
> > > >
> > > The SLAB_PANIC flag used in kmem_cache_create indicates that if the
> > > kernel is unable to create the cache, it should panic. The use of
> > > SLAB_PANIC in the creation of the landlock_object_cache is due to the
> > > critical nature of this cache for the Landlock LSM's operation. I
> > > found it to be a good choice to be used. Should I use some other
> > > altrnative?
> > 
> > Is panicing really a good idea?  Why can't you properly recover from
> > allocation failures?
> 
> I am relying on SLAB_PANIC because of the reason I mentioned earlier,
> and also because it was used in lsm_file_cache that I was asked to look
> into as reference. I could try to recover from allocation failures but
> currently my focus is on working on the changes that are listed. I will
> definitely try to look into it once I am done with all changes.

Not being able to create this kmem cache would mean that Landlock would
not be able to properly run, so we could print a warning and exit the
Landlock init function.  However, most calls to kmem_cache_create() are
init calls, and most of them (especially in security/*) set SLAB_PANIC.
I'm wondering why Landlock should do differently, if others should be
fixed, and if the extra complexity of handling several
kmem_cache_create() potential failure is worth it for init handlers?

> 
> > > > > +
> > > > >  struct landlock_object *
> > > > >  landlock_create_object(const struct landlock_object_underops *const underops,
> > > > >  		       void *const underobj)
> > > > > @@ -25,7 +34,8 @@ landlock_create_object(const struct landlock_object_underops *const underops,
> > > > >  
> > > > >  	if (WARN_ON_ONCE(!underops || !underobj))
> > > > >  		return ERR_PTR(-ENOENT);
> > > > > -	new_object = kzalloc(sizeof(*new_object), GFP_KERNEL_ACCOUNT);
> > > > > +	new_object =
> > > > > +		kmem_cache_zalloc(landlock_object_cache, GFP_KERNEL_ACCOUNT);
> > > > 
> > > > Odd indentation, why?
> > > > 
> > > This indentation is due to formatting introduced by running
> > > clang-format.
> > 
> > Why not keep it all on one line?
> > 
> I kept it all in one line in v1, but Paul and Mickael asked me to use
> clang-format, hence it is this way.

Yes, it may looks weird but we format everything with clang-format to
not waste time discussing about style.

> > thanks,
> > 
> > greg k-h
>