Re: [BUG] Panic when systemd boot do mkdir on tmpfs mounted path with smack enabled environment

From: Al Viro
Date: Fri May 27 2016 - 14:52:00 EST


On Fri, May 27, 2016 at 04:11:41PM +0100, Al Viro wrote:

> > After commit, "b968091 security_d_instantiate(): move to the point prior to attaching dentry to inode", booting on system with
> > systemd and security smack, following kernel panic occurs.
>
> /*
> * If this is a new directory and the label was
> * transmuted when the inode was initialized
> * set the transmute attribute on the directory
> * and mark the inode.
> *
> * If there is a transmute attribute on the
> * directory mark the inode.
> */
> if (isp->smk_flags & SMK_INODE_CHANGED) {
> isp->smk_flags &= ~SMK_INODE_CHANGED;
> rc = inode->i_op->setxattr(dp,
> XATTR_NAME_SMACKTRANSMUTE,
> TRANS_TRUE, TRANS_TRUE_SIZE,
> 0);
>
> Damnation ;-/ That change (separating inode and dentry arguments of
> ->getxattr() so that security_d_instantiate() could be called before dentry
> is hashed or attached to inode) had been discussed back in early March and
> reaction of Casey back then had been basically "I believe that smack can
> live with that, will verify that in about a week". With no followup
> objections - neither immediate, nor in a week. As the matter of fact,
> your posting is the first time anyone has reported stepping into that problem.
> And that change had been present in linux-next since the beginning of May ;-/
> Sigh...
>
> > It works fine if reverting the commit, "b968091 security_d_instantiate(): move to the point prior to attaching dentry to inode", for
> > d_instantiate() like following.
>
> Can't be reverted in mainline. Not without shitloads of other stuff.
>
> There is a fairly straightforward way to handle that - do to ->setxattr()
> what we'd already done to ->getxattr(). See vfs.git#smack-fix. Warning:
> it's only build-tested. I'm going to have it go through LTP and xfstests
> shortly; _please_ check if it works on your setup, because I've no idea
> how to put together a testing setup for smack.

FWIW, that couple of commits seems to survive the testing here and is
pretty obvious. I have _NOT_ tested it on smack setups, so I really want
somebody (Casey or someone in Samsung) to check if it fixes the problem.
The change itself isn't tricky, but I fucking _hate_ doing that this late
in the merge window ;-/