Michal Koutný <mkoutny@xxxxxxxx> writes: > On Thu, Feb 10, 2022 at 08:13:20PM -0600, "Eric W. Biederman" <ebiederm@xxxxxxxxxxxx> wrote: >> @@ -1881,7 +1881,7 @@ static int do_execveat_common(int fd, struct filename *filename, > [...] >> - (current_user() != INIT_USER) && >> + (current_ucounts() != &init_ucounts) && > [...] >> @@ -2027,7 +2027,7 @@ static __latent_entropy struct task_struct *copy_process( > [...] >> - if (p->real_cred->user != INIT_USER && >> + if ((task_ucounts(p) != &init_ucounts) && > > These substitutions make sense to me. > >> !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) >> goto bad_fork_cleanup_count; >> } >> diff --git a/kernel/user_namespace.c b/kernel/user_namespace.c >> index 6b2e3ca7ee99..f0c04073403d 100644 >> --- a/kernel/user_namespace.c >> +++ b/kernel/user_namespace.c >> @@ -123,6 +123,8 @@ int create_user_ns(struct cred *new) >> ns->ucount_max[i] = INT_MAX; >> } >> set_rlimit_ucount_max(ns, UCOUNT_RLIMIT_NPROC, rlimit(RLIMIT_NPROC)); >> + if (new->ucounts == &init_ucounts) >> + set_rlimit_ucount_max(ns, UCOUNT_RLIMIT_NPROC, RLIMIT_INFINITY); >> set_rlimit_ucount_max(ns, UCOUNT_RLIMIT_MSGQUEUE, rlimit(RLIMIT_MSGQUEUE)); >> set_rlimit_ucount_max(ns, UCOUNT_RLIMIT_SIGPENDING, rlimit(RLIMIT_SIGPENDING)); >> set_rlimit_ucount_max(ns, UCOUNT_RLIMIT_MEMLOCK, rlimit(RLIMIT_MEMLOCK)); > > First, I wanted to object this double fork_init() but I realized it's > relevant for newly created user_ns. > > Second, I think new->ucounts would be correct at this point and the > check should be > >> if (ucounts == &init_ucounts) > > i.e. before set_cred_ucounts() new->ucounts may not be correct. > > I'd suggest also a comment in the create_user_ns() explaining the > reason is to exempt global root from RLIMINT_NRPOC also indirectly via > descendant user_nss. Yes. This one got culled from my next version of the patchset as it is not conservative enough. I think it is probably the right general direction. On further reflection I am not convinced that it makes sense to test user or ucounts. They are really not fields designed to support permission checks. I think if we want to exempt the root user's children from the root users rlimit using the second set_rlimit_ucount_max is the way to go. Someone filed a bug that strongly suggests that we want the second set_rlimit_ucount_max: https://bugzilla.kernel.org/show_bug.cgi?id=215596 I am still trying to understand that case. Eric