Re: [PATCH ALT4 V2 1/2] audit: show fstype:pathname for entries with anonymous parents

[Richard Guy Briggs]

On 2017-05-30 17:21, Paul Moore wrote:
> On Tue, Apr 4, 2017 at 5:21 AM, Richard Guy Briggs <rgb@xxxxxxxxxx> wrote:
> > Tracefs or debugfs were causing hundreds to thousands of null PATH
> > records to be associated with the init_module and finit_module SYSCALL
> > records on a few modules when the following rule was in place for
> > startup:
> >         -a always,exit -F arch=x86_64 -S init_module -F key=mod-load
> >
> > This happens because the parent inode is not found in the task's
> > audit_names list and hence treats it as anonymous.  This gives us no
> > information other than a numerical device number that may no longer be
> > visible upon log inspeciton, and an inode number.
> >
> > Fill in the filesystem type, filesystem magic number and full pathname
> > from the filesystem mount point on previously null PATH records from
> > entries that have an anonymous parent from the child dentry using
> > dentry_path_raw().
> >
> > Make the dentry argument of __audit_inode_child() non-const so that we
> > can take a reference to it in the case of an anonymous parent with
> > dget() and dget_parent() to be able to later print a partial path from
> > the host filesystem rather than null.
> >
> > Since all we are given is an inode of the parent and the dentry of the
> > child, finding the path from the mount point to the root of the
> > filesystem is more challenging that would involve searching all
> > vfsmounts from "/" until a matching dentry is found for that
> > filesystem's root dentry.  Even if one is found, there may be more than
> > one mount point.  At this point the gain seems marginal since
> > knowing the filesystem type and path are a significant help in tracking
> > down the source of the PATH records and being to address them.
> >
> > Sample output:
> > type=PROCTITLE msg=audit(1488317694.446:143): proctitle=2F7362696E2F6D6F6470726F6265002D71002D2D006E66737634
> > type=PATH msg=audit(1488317694.446:143): item=797 name=tracefs(74726163):/events/nfs4/nfs4_setclientid/format inode=15969 dev=00:09 mode=0100444 ouid=0 ogid=0 rdev=00:00 obj=system_u:object_r:tracefs_t:s0 nametype=CREATE
> > type=PATH msg=audit(1488317694.446:143): item=796 name=tracefs(74726163):/events/nfs4/nfs4_setclientid inode=15964 dev=00:09 mode=040755 ouid=0 ogid=0 rdev=00:00 obj=system_u:object_r:tracefs_t:s0 nametype=PARENT
> > ...
> > type=PATH msg=audit(1488317694.446:143): item=1 name=tracefs(74726163):/events/nfs4 inode=15571 dev=00:09 mode=040755 ouid=0 ogid=0 rdev=00:00 obj=system_u:object_r:tracefs_t:s0 nametype=CREATE
> > type=PATH msg=audit(1488317694.446:143): item=0 name=tracefs(74726163):/events inode=119 dev=00:09 mode=040755 ouid=0 ogid=0 rdev=00:00 obj=system_u:object_r:tracefs_t:s0 nametype=PARENT
> > type=UNKNOWN[1330] msg=audit(1488317694.446:143): name="nfsv4"
> > type=SYSCALL msg=audit(1488317694.446:143): arch=c000003e syscall=313 success=yes exit=0 a0=1 a1=55d5a35ce106 a2=0 a3=1 items=798 ppid=6 pid=528 auid=4294967295 uid=0 gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0 fsgid=0 tty=(none) ses=4294967295 comm="modprobe" exe="/usr/bin/kmod" subj=system_u:system_r:insmod_t:s0 key="mod-load"
> >
> > See: https://github.com/linux-audit/audit-kernel/issues/8
> > Test case: https://github.com/linux-audit/audit-testsuite/issues/42
> >
> > Signed-off-by: Richard Guy Briggs <rgb@xxxxxxxxxx>
> > ---
> >  include/linux/audit.h |    8 ++++----
> >  kernel/audit.c        |   16 ++++++++++++++++
> >  kernel/audit.h        |    1 +
> >  kernel/auditsc.c      |    8 +++++++-
> >  4 files changed, 28 insertions(+), 5 deletions(-)
> 
> ...
> 
> > diff --git a/kernel/audit.c b/kernel/audit.c
> > index 25dd70a..7d83c5a 100644
> > --- a/kernel/audit.c
> > +++ b/kernel/audit.c
> > @@ -66,6 +66,7 @@
> >  #include <linux/freezer.h>
> >  #include <linux/pid_namespace.h>
> >  #include <net/netns/generic.h>
> > +#include <linux/dcache.h>
> >
> >  #include "audit.h"
> >
> > @@ -1884,6 +1885,10 @@ void audit_copy_inode(struct audit_names *name, const struct dentry *dentry,
> >         name->gid   = inode->i_gid;
> >         name->rdev  = inode->i_rdev;
> >         security_inode_getsecid(inode, &name->osid);
> > +       if (name->dentry) {
> > +               dput(name->dentry);
> > +               name->dentry = NULL;
> > +       }
> 
> Out of curiosity, what terrible things happen if we take a reference
> to a non-NULL dentry passed to audit_copy_inode() and store it in
> name->dentry?  Does performance tank?

Interesting idea.  Right now it is optimized to only take a reference to
the dentry's parent dentry in the case we're handed an anonymous entry.
Most of the time it will never be used even though we invest in the
overhead of taking a reference count.  Besides, __audit_inode_child()
hands in a NULL for the dentry parameter to audit_copy_inode().  I'm
assuming you are hinting at also using that dentry to compare the
audit_names entry, which I think it a bad idea since there could be
multiple paths to access a dentry.  I did orignially have another patch
that would have tried to use that as well, which didn't seem to hurt,
but I didn't think was worth upstreaming.

> Also out of curiosity, why do we want to drop a dentry reference here
> if one already exists?

I think we want to drop a dentry reference here because this inode child
could be a subsequent access to the same dentry with a full path,
removing the need to cache this dentry information in the first place.

> >         audit_copy_fcaps(name, dentry);
> >  }
> >
> > @@ -1925,6 +1930,17 @@ void audit_log_name(struct audit_context *context, struct audit_names *n,
> >                         audit_log_n_untrustedstring(ab, n->name->name,
> >                                                     n->name_len);
> >                 }
> > +       } else if (n->dentry) {
> > +               char *fullpath;
> > +               const char *fullpathp;
> > +
> > +               fullpath = kmalloc(PATH_MAX, GFP_KERNEL);
> > +               if (!fullpath)
> > +                       return;
> 
> I'm wondering if there is some value in still emitting the record if
> the kmalloc() fails, just with the name field set as the unset "?"
> value, e.g. "name=?".  Thoughts?

Possibly.  We've got much bigger problems if that happens, but this
sounds like a good defensive coding approach.  I'm even tempted to call
audit_panic().

> paul moore

- RGB

--
Richard Guy Briggs <rgb@xxxxxxxxxx>
Sr. S/W Engineer, Kernel Security, Base Operating Systems
Remote, Ottawa, Red Hat Canada
IRC: rgb, SunRaycer
Voice: +1.647.777.2635, Internal: (81) 32635