Date:	Tue, 26 May 2015 17:56:49 -0400
From:	bfields@...ldses.org (J. Bruce Fields)
To:	Andreas Gruenbacher <andreas.gruenbacher@...il.com>
Cc:	linux-kernel@...r.kernel.org, linux-fsdevel@...r.kernel.org,
	linux-nfs@...r.kernel.org, Andreas Gruenbacher <agruen@...nel.org>
Subject: Re: [RFC v3 29/45] richacl: Apply the file masks to a richacl

On Fri, Apr 24, 2015 at 01:04:26PM +0200, Andreas Gruenbacher wrote:
> Put all the pieces of the acl transformation puzzle together for
> computing a richacl which has the file masks "applied" so that the
> standard nfsv4 access check algorithm can be used on the richacl.
> 
> Signed-off-by: Andreas Gruenbacher <agruen@...nel.org>
> ---
>  fs/richacl_compat.c     | 103 ++++++++++++++++++++++++++++++++++++++++++++++++
>  include/linux/richacl.h |   3 ++
>  2 files changed, 106 insertions(+)
> 
> diff --git a/fs/richacl_compat.c b/fs/richacl_compat.c
> index 645917f..49af600 100644
> --- a/fs/richacl_compat.c
> +++ b/fs/richacl_compat.c
> @@ -647,3 +647,106 @@ richacl_isolate_group_class(struct richacl_alloc *x)
>  	}
>  	return 0;
>  }
> +
> +/**
> + * __richacl_apply_masks  -  apply the file masks to all aces
> + * @x:		acl and number of allocated entries
> + *
> + * Apply the owner mask to owner@ aces, the other mask to
> + * everyone@ aces, and the group mask to all other aces.
> + *
> + * The previous transformations have brought the acl into a
> + * form in which applying the masks will not lead to the
> + * accidental loss of permissions anymore.
> + */
> +static int
> +__richacl_apply_masks(struct richacl_alloc *x)
> +{
> +	struct richace *ace;
> +
> +	richacl_for_each_entry(ace, x->acl) {
> +		unsigned int mask;
> +
> +		if (richace_is_inherit_only(ace) || !richace_is_allow(ace))
> +			continue;
> +		if (richace_is_owner(ace))
> +			mask = x->acl->a_owner_mask;
> +		else if (richace_is_everyone(ace))
> +			mask = x->acl->a_other_mask;
> +		else
> +			mask = x->acl->a_group_mask;
> +		if (richace_change_mask(x, &ace, ace->e_mask & mask))
> +			return -1;
> +	}
> +	return 0;
> +}
> +
> +/**
> + * richacl_apply_masks  -  apply the masks to the acl
> + *
> + * Transform @acl so that the standard NFSv4 permission check algorithm (which
> + * is not aware of file masks) will compute the same access decisions as the
> + * richacl permission check algorithm (which looks at the acl and the file
> + * masks).
> + *
> + * This algorithm is split into several steps:
> + *
> + *   - Move everyone@ aces to the end of the acl.  This simplifies the other
> + *     transformations, and allows the everyone@ allow ace at the end of the
> + *     acl to eventually allow permissions to the other class only.
> + *
> + *   - Propagate everyone@ permissions up the acl.  This transformation makes
> + *     sure that the owner and group class aces won't lose any permissions when
> + *     we apply the other mask to the everyone@ allow ace at the end of the acl.
> + *
> + *   - Apply the file masks to all aces.
> + *
> + *   - Make sure that the owner is not granted any permissions beyond the owner
> + *     mask from group class aces or from everyone@.
> + *
> + *   - Make sure that the group class is not granted any permissions from
> + *     everyone@.
> + *
> + * The algorithm is exact except for richacls which cannot be represented as an
> + * acl alone: for example, given this acl:
> + *
> + *    group@:rw::allow
> + *
> + * when file masks corresponding to mode 0600 are applied, the owner would only
> + * get rw access if he is a member of the owning group.  This algorithm would
> + * produce an empty acl in this case.  We fix this case by modifying
> + * richacl_permission() so that the group mask is always applied to group class
> + * aces.  With this fix, the owner would not have any access (beyond the
> + * implicit permissions always granted to owners).

This discussion is really confusing.

If I understand correctly, you're saying:

	- this mapping is inaccurate in some cases.
	- therefore we've modified the permission algorithm to fix those
	  cases.
	- therefore actually, wait, the mapping is always accurate.

Do I have this right?

Please just define the group-mask behavior to be part of the richacl
permission-checking algorithm from the start, and this will be less
confusing.

I don't think I'm convinced yet that the resulting algorithm really is
correct.  I mean: I think it probably is, I just don't understand the
argument quite yet.

> + *
> + * NOTE: Depending on the acl and file masks, this algorithm can increase the
> + * number of aces by almost a factor of three in the worst case. This may make
> + * the acl too large for some purposes.

If only for the sake of reviewers, it would be helpful to have some more
detail here to convince us that this isn't going to be a big problem in
practice.

For example, can we identify a subset of richacls which would cover most
reasonable use cases and not be badly mangled by this algorithm?  It
would be nice to be able to document some simple best practices which if
followed would guarantee that the mapped acls will be sane.  (Order
denies before allows?  Stick to a single EVERYONE@ ace at the end?
Don't chmod to modes which give more permissions to other than to the
group or to the group than to the owner?).

--b.

> + */
> +int
> +richacl_apply_masks(struct richacl **acl)
> +{
> +	if ((*acl)->a_flags & RICHACL_MASKED) {
> +		struct richacl_alloc x = {
> +			.acl = richacl_clone(*acl, GFP_KERNEL),
> +			.count = (*acl)->a_count,
> +		};
> +		if (!x.acl)
> +			return -ENOMEM;
> +
> +		if (richacl_move_everyone_aces_down(&x) ||
> +		    richacl_propagate_everyone(&x) ||
> +		    __richacl_apply_masks(&x) ||
> +		    richacl_isolate_owner_class(&x) ||
> +		    richacl_isolate_group_class(&x)) {
> +			richacl_put(x.acl);
> +			return -ENOMEM;
> +		}
> +
> +		x.acl->a_flags &= ~RICHACL_MASKED;
> +		richacl_put(*acl);
> +		*acl = x.acl;
> +	}
> +	return 0;
> +}
> +EXPORT_SYMBOL_GPL(richacl_apply_masks);
> diff --git a/include/linux/richacl.h b/include/linux/richacl.h
> index bcc2b64..6a97dca 100644
> --- a/include/linux/richacl.h
> +++ b/include/linux/richacl.h
> @@ -325,4 +325,7 @@ extern int richacl_equiv_mode(const struct richacl *, mode_t *);
>  extern int richacl_permission(struct inode *, const struct richacl *, int);
>  extern struct richacl *richacl_create(struct inode *, struct inode *);
>  
> +/* richacl_compat.c */
> +extern int richacl_apply_masks(struct richacl **);
> +
>  #endif /* __RICHACL_H */
> -- 
> 2.1.0
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to majordomo@...r.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists