Re: [PATCH] LANDLOCK: use kmem_cache for landlock_object

The subject should start with "landlock: Use" instead of "LANDLOCK: use"

On Thu, Mar 28, 2024 at 01:23:17PM +0530, Ayush Tiwari wrote:
> Hello Paul
> Thanks a lot for the feedback. Apologies for the mistakes. Could you
> help me in some places so that I can correct the errors, like:
> On Wed, Mar 27, 2024 at 07:43:36PM -0400, Paul Moore wrote:
> > On Wed, Mar 27, 2024 at 7:26 PM Ayush Tiwari <[email protected]> wrote:
> > >
> > > Use kmem_cache replace kzalloc() calls with kmem_cache_zalloc() for
> > > struct landlock_object and update the related dependencies.
> > >
> > > Signed-off-by: Ayush Tiwari <[email protected]>
> > > ---
> > >  security/landlock/fs.c     |  2 +-
> > >  security/landlock/object.c | 14 ++++++++++++--
> > >  security/landlock/object.h |  4 ++++
> > >  security/landlock/setup.c  |  2 ++
> > >  4 files changed, 19 insertions(+), 3 deletions(-)
> > 
> > Hi Ayush,
> > 
> > Mickaël has the final say on Landlock patches, but I had a few
> > comments that I've included below ...
> > 
> > > diff --git a/security/landlock/fs.c b/security/landlock/fs.c
> > > index fc520a06f9af..227dd67dd902 100644
> > > --- a/security/landlock/fs.c
> > > +++ b/security/landlock/fs.c
> > > @@ -124,7 +124,7 @@ static struct landlock_object *get_inode_object(struct inode *const inode)
> > >         if (unlikely(rcu_access_pointer(inode_sec->object))) {
> > >                 /* Someone else just created the object, bail out and retry. */
> > >                 spin_unlock(&inode->i_lock);
> > > -               kfree(new_object);
> > > +               kmem_cache_free(landlock_object_cache, new_object);
> > 
> > See my comment below, but you may want to wrap this in a Landlock
> > object API function.
> Sure. I will definitely implement this.
> > 
> > >                 rcu_read_lock();
> > >                 goto retry;
> > > diff --git a/security/landlock/object.c b/security/landlock/object.c
> > > index 1f50612f0185..df1354215617 100644
> > > --- a/security/landlock/object.c
> > > +++ b/security/landlock/object.c
> > > @@ -17,6 +17,15 @@
> > >
> > >  #include "object.h"
> > >
> > > +struct kmem_cache *landlock_object_cache;
> > > +
> > > +void __init landlock_object_init(void)
> > > +{
> > > +       landlock_object_cache = kmem_cache_create(
> > > +               "landlock_object_cache", sizeof(struct landlock_object), 0,

No need for the "_cache" name suffix.

> > > +               SLAB_PANIC, NULL);
> > 
> > The comments in include/linux/slab.h suggest using the KMEM_CACHE()
> > macro, instead of kmem_cache_create(), as a best practice for creating
> > slab caches.
> > 
> Sure. Apologies I didn't see that, I tried to implement it from scratch
> using the reference from linux memory management APIs.
> > > +}
> > > +
> > >  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);
> > 
> > If the line is too long, you might want to consider splitting the
> > function parameters like this:
> > 
> >   new_object = kmem_cache_zalloc(landlock_object_cache,
> >                                  GFP_KERNEL_ACCOUNT);
> > 
> 
> Sure. I didn't do as it was below the 100 columns limit, but will
> definitely implement it.

Please just use clang-format.

> > >         if (!new_object)
> > >                 return ERR_PTR(-ENOMEM);
> > >         refcount_set(&new_object->usage, 1);
> > > @@ -62,6 +72,6 @@ void landlock_put_object(struct landlock_object *const object)
> > >                  * @object->underobj to @object (if it still exists).
> > >                  */
> > >                 object->underops->release(object);
> > > -               kfree_rcu(object, rcu_free);

Is it safe?

According to commit ae65a5211d90 ("mm/slab: document kfree() as allowed
for kmem_cache_alloc() objects"), no change should be needed (and it
must not be backported to kernels older than 6.4 with CONFIG_SLOB). This
way we can avoid exporting landlock_object_cache.  Please add a note
about this commit and the related warning in the commit message.

> > > +               kmem_cache_free(landlock_object_cache, object);
> > >         }
> > >  }
> > > diff --git a/security/landlock/object.h b/security/landlock/object.h
> > > index 5f28c35e8aa8..8ba1af3ddc2e 100644
> > > --- a/security/landlock/object.h
> > > +++ b/security/landlock/object.h
> > > @@ -13,6 +13,10 @@
> > >  #include <linux/refcount.h>
> > >  #include <linux/spinlock.h>
> > >
> > > +extern struct kmem_cache *landlock_object_cache;
> > 
> > This really is a decision for Mickaël, but you may want to make
> > @landlock_object_cache private to object.c and create functions to
> > manage it as needed, e.g. put/free operations.
> > 
> Okay. I didn't make it private as I was using it in fs.c to use
> kmem_cache_free, but if this is supposed to be private, I can modify the
> approach and expose it via some function, not directly exposing
> landlock_object_cache.

Yes, that would be better.

> > > +void __init landlock_object_init(void);
> > > +
> > >  struct landlock_object;
> > >
> > >  /**
> > > diff --git a/security/landlock/setup.c b/security/landlock/setup.c
> > > index f6dd33143b7f..a5fca4582ee1 100644
> > 
> > -- 
> > paul-moore.com
> I will make all the changes you mentioned, and as you said, I will
> wait for Mickael's say.

Agree with Paul and Greg unless commented otherwise. Thanks