kern_jail_set() Error Scenario Question

[James Gritton]

On 08/22/13 17:46, Matt Miller wrote:
> We ran into the following scenario in an application recently and were
> wondering if the behavior of kern_jail_set() is as expected here.
>
> This was an application bug where we were in, say, the JID 1 context
> and tried to call jailparam_set() with the flags (JAIL_CREATE |
> JAIL_UPDATE) and the "jid" param set to 1.  The basic idea was to
> create or update JID 1 with some params, but the error was we were
> already in the JID 1 context.  So, our understanding is this shouldn't
> work since JID 1 already exists and you can only modify it from a
> proper ancestor.
>
> However, rather than getting an error back from jailparam_set(), it
> ended up creating a second prison with JID 1, so there were two
> prisons existing with JID 1 at that point.  This is based on 8.2.0
> code, but, at first glance, it looks like the logic causing this may
> be the same in head.
>
> Looking at kern_jail_set(), what happens here is:
>
> 1. We find a prison with JID=1, however since it's not a proper child
> we set pr = NULL in line 1024:
>
> 1011                 pr = prison_find(jid);
> 1012                 if (pr != NULL) {
> 1013                         ppr = pr->pr_parent;
> 1014                         /* Create: jid must not exist. */
> 1015                         if (cuflags == JAIL_CREATE) {
> 1016                                 mtx_unlock(&pr->pr_mtx);
> 1017                                 error = EEXIST;
> 1018                                 vfs_opterror(opts, "jail %d
> already exists",
> 1019                                     jid);
> 1020                                 goto done_unlock_list;
> 1021                         }
> 1022                         if (!prison_ischild(mypr, pr)) {
> 1023                                 mtx_unlock(&pr->pr_mtx);
> 1024                                 pr = NULL;
> 1025                         } else if (pr->pr_uref == 0) {
>
> 2. Since pr is NULL, we create a new prison.  Since the jid is not
> zero, we insert it in the list and set its pr_id.  At this point, we
> have two prisons with a JID of 1 and the same parent prison.
>
> 1166         /* If there's no prison to update, create a new one and
> link it in. */
> 1167         if (pr == NULL) {
> ...
> 1185                 pr = malloc(sizeof(*pr), M_PRISON, M_WAITOK | M_ZERO);
> 1186                 if (jid == 0) {
> ...
> 1212                 } else {
> 1213                         /*
> 1214                          * The jail already has a jid (that did
> not yet exist),
> 1215                          * so just find where to insert it.
> 1216                          */
> 1217                         TAILQ_FOREACH(tpr, &allprison, pr_list)
> 1218                                 if (tpr->pr_id >= jid) {
> 1219                                         TAILQ_INSERT_BEFORE(tpr,
> pr, pr_list);
> 1220                                         break;
> 1221                                 }
> 1222                 }
> ...
> 1229                 pr->pr_parent = ppr;
> 1230                 pr->pr_id = jid;
>
> We wanted to see if this is per design or a situation that should
> avoid creating the second prison and return an error.

That's definitely not per design.  I'll try reproducing this, and put in
correct logic.  The proper response is indeed an error: ENOENT, because
from inside the JID 1 context you shouldn't be able to see jail #1 (you
can't operate on your own jail).

- Jamie