Re: [PATCH v2] LSM: add SafeSetID module that gates setid calls

[Micah Morton <mortonm@chromium.org>]

On Mon, Jan 14, 2019 at 4:38 PM Kees Cook <keescook@chromium.org> wrote:
>
> On Fri, Jan 11, 2019 at 9:13 AM <mortonm@chromium.org> wrote:
> >
> > From: Micah Morton <mortonm@chromium.org>
> >
> > SafeSetID gates the setid family of syscalls to restrict UID/GID
> > transitions from a given UID/GID to only those approved by a
> > system-wide whitelist. These restrictions also prohibit the given
> > UIDs/GIDs from obtaining auxiliary privileges associated with
> > CAP_SET{U/G}ID, such as allowing a user to set up user namespace UID
> > mappings. For now, only gating the set*uid family of syscalls is
> > supported, with support for set*gid coming in a future patch set.
> >
> > Signed-off-by: Micah Morton <mortonm@chromium.org>
> > ---
> > Changes since the last patch set: Rebase after commit
> > a35ce66b801631823fc78c8a78d104f9c0976867 got applied to next-general.
> > As a result of that commit, we can remove the changes in arch/ and the
> > setuid_syscall function in lsm.c, since this code no longer needs to do
> > arch-specific operations to see if security_capable is being called from
> > a setid syscall. Instead, we add the ns_capable_insetid function and
> > call it from the setid syscalls in kernel/sys.c (rather than calling the
> > original ns_capable function), which allows us to signal to the
> > security_capable hooks whether the hook is being called from within a
> > setid syscall.
>
> I would split this patch into two halfs: the "no op" change that
> "marks" all the setid call sites in the first patch, then the LSM
> itself in the second patch.

Done.

>
> > +bool ns_capable_insetid(struct user_namespace *ns, int cap)
> > +{
> > +       return ns_capable_common(ns, cap, CAP_OPT_INSETID);
> > +}
> > +EXPORT_SYMBOL(ns_capable_insetid);
>
> Since we have the noaudit helper still, using this one seems fine to
> me. I might bikeshed the name to "ns_capable_setid()". If others don't
> want a new helper, then it should be fine to just change the callsites
> to the direct ns_capable_common(ns, cap, CAP_OPT_INSETID).

Done.

>
> > +static int safesetid_security_capable(const struct cred *cred,
> > +                                     struct user_namespace *ns,
> > +                                     int cap,
> > +                                     unsigned int opts)
> > +{
> > +       if (cap == CAP_SETUID &&
> > +           check_setuid_policy_hashtable_key(cred->uid)) {
> > +               if (!(opts & CAP_OPT_INSETID)) {
> > +                       /*
> > +                        * Deny if we're not in a set*uid() syscall to avoid
> > +                        * giving powers gated by CAP_SETUID that are related
> > +                        * to functionality other than calling set*uid() (e.g.
> > +                        * allowing user to set up userns uid mappings).
> > +                        */
> > +                       pr_warn("Operation requires CAP_SETUID, which is not available to UID %u for operations besides approved set*uid transitions",
> > +                               __kuid_val(cred->uid));
> > +                       return -1;
> > +                }
> > +       }
> > +       return 0;
> > +}
>
> Much cleaner than the per-arch syscall tests. :)
>
> > +static void setuid_policy_violation(kuid_t parent, kuid_t child)
> > +{
> > +       pr_warn("UID transition (%d -> %d) blocked",
> > +               __kuid_val(parent),
> > +               __kuid_val(child));
> > +        /*
> > +         * Kill this process to avoid potential security vulnerabilities
> > +         * that could arise from a missing whitelist entry preventing a
> > +         * privileged process from dropping to a lesser-privileged one.
> > +         */
> > +        do_exit(SIGKILL);
>
> I think I asked earlier if this should be an unblockable signal raise
> instead of a do_exit(). I don't remember if that got answered?

Could you elaborate on this a bit, or share a pointer to some code/doc
that explains the difference? I don't recall this point being raised
before (might have missed it), and I'm no expert on the different
approaches to killing a process in this way.

>
> > +}
> > +
> > +static int check_uid_transition(kuid_t parent, kuid_t child)
> > +{
> > +       if (check_setuid_policy_hashtable_key_value(parent, child))
> > +               return 0;
> > +       setuid_policy_violation(parent, child);
> > +       return -1;
> > +}
>
> Any reason not to just collapse setuid_policy_violation() into this function?

Done.

>
> > +
> > +/*
> > + * Check whether there is either an exception for user under old cred struct to
> > + * set*uid to user under new cred struct, or the UID transition is allowed (by
> > + * Linux set*uid rules) even without CAP_SETUID.
> > + */
> > +static int safesetid_task_fix_setuid(struct cred *new,
> > +                                    const struct cred *old,
> > +                                    int flags)
> > +{
> > +
> > +       /* Do nothing if there are no setuid restrictions for this UID. */
> > +       if (!check_setuid_policy_hashtable_key(old->uid))
> > +               return 0;
> > +
> > +       switch (flags) {
> > +       case LSM_SETID_RE:
> > +               /*
> > +                * Users for which setuid restrictions exist can only set the
> > +                * real UID to the real UID or the effective UID, unless an
> > +                * explicit whitelist policy allows the transition.
> > +                */
> > +               if (!uid_eq(old->uid, new->uid) &&
> > +                       !uid_eq(old->euid, new->uid)) {
> > +                       return check_uid_transition(old->uid, new->uid);
> > +               }
> > +               /*
> > +                * Users for which setuid restrictions exist can only set the
> > +                * effective UID to the real UID, the effective UID, or the
> > +                * saved set-UID, unless an explicit whitelist policy allows
> > +                * the transition.
> > +                */
> > +               if (!uid_eq(old->uid, new->euid) &&
> > +                       !uid_eq(old->euid, new->euid) &&
> > +                       !uid_eq(old->suid, new->euid)) {
> > +                       return check_uid_transition(old->euid, new->euid);
> > +               }
> > +               break;
> > +       case LSM_SETID_ID:
> > +               /*
> > +                * Users for which setuid restrictions exist cannot change the
> > +                * real UID or saved set-UID unless an explicit whitelist
> > +                * policy allows the transition.
> > +                */
> > +               if (!uid_eq(old->uid, new->uid))
> > +                       return check_uid_transition(old->uid, new->uid);
> > +               if (!uid_eq(old->suid, new->suid))
> > +                       return check_uid_transition(old->suid, new->suid);
> > +               break;
> > +       case LSM_SETID_RES:
> > +               /*
> > +                * Users for which setuid restrictions exist cannot change the
> > +                * real UID, effective UID, or saved set-UID to anything but
> > +                * one of: the current real UID, the current effective UID or
> > +                * the current saved set-user-ID unless an explicit whitelist
> > +                * policy allows the transition.
> > +                */
> > +               if (!uid_eq(new->uid, old->uid) &&
> > +                       !uid_eq(new->uid, old->euid) &&
> > +                       !uid_eq(new->uid, old->suid)) {
> > +                       return check_uid_transition(old->uid, new->uid);
> > +               }
> > +               if (!uid_eq(new->euid, old->uid) &&
> > +                       !uid_eq(new->euid, old->euid) &&
> > +                       !uid_eq(new->euid, old->suid)) {
> > +                       return check_uid_transition(old->euid, new->euid);
> > +               }
> > +               if (!uid_eq(new->suid, old->uid) &&
> > +                       !uid_eq(new->suid, old->euid) &&
> > +                       !uid_eq(new->suid, old->suid)) {
> > +                       return check_uid_transition(old->suid, new->suid);
> > +               }
> > +               break;
> > +       case LSM_SETID_FS:
> > +               /*
> > +                * Users for which setuid restrictions exist cannot change the
> > +                * filesystem UID to anything but one of: the current real UID,
> > +                * the current effective UID or the current saved set-UID
> > +                * unless an explicit whitelist policy allows the transition.
> > +                */
> > +               if (!uid_eq(new->fsuid, old->uid)  &&
> > +                       !uid_eq(new->fsuid, old->euid)  &&
> > +                       !uid_eq(new->fsuid, old->suid) &&
> > +                       !uid_eq(new->fsuid, old->fsuid)) {
> > +                       return check_uid_transition(old->fsuid, new->fsuid);
> > +               }
> > +               break;
> > +       }
> > +       return 0;
> > +}
> > +
> > +int add_safesetid_whitelist_entry(kuid_t parent, kuid_t child)
> > +{
> > +       struct entry *new;
> > +
> > +       /* Return if entry already exists */
> > +       if (check_setuid_policy_hashtable_key_value(parent, child))
> > +               return 0;
> > +
> > +       new = kzalloc(sizeof(struct entry), GFP_KERNEL);
> > +       if (!new)
> > +               return -ENOMEM;
> > +       new->parent_kuid = __kuid_val(parent);
> > +       new->child_kuid = __kuid_val(child);
> > +       spin_lock(&safesetid_whitelist_hashtable_spinlock);
> > +       hash_add_rcu(safesetid_whitelist_hashtable,
> > +                    &new->next,
> > +                    __kuid_val(parent));
> > +       spin_unlock(&safesetid_whitelist_hashtable_spinlock);
> > +       return 0;
> > +}
> > +
> > +void flush_safesetid_whitelist_entries(void)
> > +{
> > +       struct entry *entry;
> > +       struct hlist_node *hlist_node;
> > +       unsigned int bkt_loop_cursor;
> > +       HLIST_HEAD(free_list);
> > +
> > +       /*
> > +        * Could probably use hash_for_each_rcu here instead, but this should
> > +        * be fine as well.
> > +        */
> > +       spin_lock(&safesetid_whitelist_hashtable_spinlock);
> > +       hash_for_each_safe(safesetid_whitelist_hashtable, bkt_loop_cursor,
> > +                          hlist_node, entry, next) {
> > +               hash_del_rcu(&entry->next);
> > +               hlist_add_head(&entry->dlist, &free_list);
> > +       }
> > +       spin_unlock(&safesetid_whitelist_hashtable_spinlock);
> > +       synchronize_rcu();
> > +       hlist_for_each_entry_safe(entry, hlist_node, &free_list, dlist) {
> > +               hlist_del(&entry->dlist);
> > +               kfree(entry);
> > +       }
> > +}
> > +
> > +static struct security_hook_list safesetid_security_hooks[] = {
> > +       LSM_HOOK_INIT(task_fix_setuid, safesetid_task_fix_setuid),
> > +       LSM_HOOK_INIT(capable, safesetid_security_capable)
> > +};
> > +
> > +static int __init safesetid_security_init(void)
> > +{
> > +       security_add_hooks(safesetid_security_hooks,
> > +                          ARRAY_SIZE(safesetid_security_hooks), "safesetid");
> > +
> > +       return 0;
> > +}
> > +
> > +DEFINE_LSM(safesetid_security_init) = {
> > +       .init = safesetid_security_init,
> > +};
> > diff --git a/security/safesetid/lsm.h b/security/safesetid/lsm.h
> > new file mode 100644
> > index 000000000000..bf78af9bf314
> > --- /dev/null
> > +++ b/security/safesetid/lsm.h
> > @@ -0,0 +1,30 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +/*
> > + * SafeSetID Linux Security Module
> > + *
> > + * Author: Micah Morton <mortonm@chromium.org>
> > + *
> > + * Copyright (C) 2018 The Chromium OS Authors.
> > + *
> > + * This program is free software; you can redistribute it and/or modify
> > + * it under the terms of the GNU General Public License version 2, as
> > + * published by the Free Software Foundation.
> > + *
> > + */
> > +#ifndef _SAFESETID_H
> > +#define _SAFESETID_H
> > +
> > +#include <linux/types.h>
> > +
> > +/* Function type. */
> > +enum safesetid_whitelist_file_write_type {
> > +       SAFESETID_WHITELIST_ADD, /* Add whitelist policy. */
> > +       SAFESETID_WHITELIST_FLUSH, /* Flush whitelist policies. */
> > +};
> > +
> > +/* Add entry to safesetid whitelist to allow 'parent' to setid to 'child'. */
> > +int add_safesetid_whitelist_entry(kuid_t parent, kuid_t child);
> > +
> > +void flush_safesetid_whitelist_entries(void);
> > +
> > +#endif /* _SAFESETID_H */
> > diff --git a/security/safesetid/securityfs.c b/security/safesetid/securityfs.c
> > new file mode 100644
> > index 000000000000..ff5fcf2c1b37
> > --- /dev/null
> > +++ b/security/safesetid/securityfs.c
> > @@ -0,0 +1,189 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * SafeSetID Linux Security Module
> > + *
> > + * Author: Micah Morton <mortonm@chromium.org>
> > + *
> > + * Copyright (C) 2018 The Chromium OS Authors.
> > + *
> > + * This program is free software; you can redistribute it and/or modify
> > + * it under the terms of the GNU General Public License version 2, as
> > + * published by the Free Software Foundation.
> > + *
> > + */
> > +#include <linux/security.h>
> > +#include <linux/cred.h>
> > +
> > +#include "lsm.h"
> > +
> > +static struct dentry *safesetid_policy_dir;
> > +
> > +struct safesetid_file_entry {
> > +       const char *name;
> > +       enum safesetid_whitelist_file_write_type type;
> > +       struct dentry *dentry;
> > +};
> > +
> > +static struct safesetid_file_entry safesetid_files[] = {
> > +       {.name = "add_whitelist_policy",
> > +        .type = SAFESETID_WHITELIST_ADD},
> > +       {.name = "flush_whitelist_policies",
> > +        .type = SAFESETID_WHITELIST_FLUSH},
> > +};
> > +
> > +/*
> > + * In the case the input buffer contains one or more invalid UIDs, the kuid_t
> > + * variables pointed to by 'parent' and 'child' will get updated but this
> > + * function will return an error.
> > + */
> > +static int parse_safesetid_whitelist_policy(const char __user *buf,
> > +                                           size_t len,
> > +                                           kuid_t *parent,
> > +                                           kuid_t *child)
> > +{
> > +       char *kern_buf;
> > +       char *parent_buf;
> > +       char *child_buf;
> > +       const char separator[] = ":";
> > +       int ret;
> > +       size_t first_substring_length;
> > +       long parsed_parent;
> > +       long parsed_child;
> > +
> > +       /* Duplicate string from user memory and NULL-terminate */
> > +       kern_buf = memdup_user_nul(buf, len);
> > +       if (IS_ERR(kern_buf))
> > +               return PTR_ERR(kern_buf);
> > +
> > +       /*
> > +        * Format of |buf| string should be <UID>:<UID>.
> > +        * Find location of ":" in kern_buf (copied from |buf|).
> > +        */
> > +       first_substring_length = strcspn(kern_buf, separator);
> > +       if (first_substring_length == 0 || first_substring_length == len) {
> > +               ret = -EINVAL;
> > +               goto free_kern;
> > +       }
> > +
> > +       parent_buf = kmemdup_nul(kern_buf, first_substring_length, GFP_KERNEL);
> > +       if (!parent_buf) {
> > +               ret = -ENOMEM;
> > +               goto free_kern;
> > +       }
> > +
> > +       ret = kstrtol(parent_buf, 0, &parsed_parent);
> > +       if (ret)
> > +               goto free_both;
> > +
> > +       child_buf = kern_buf + first_substring_length + 1;
> > +       ret = kstrtol(child_buf, 0, &parsed_child);
> > +       if (ret)
> > +               goto free_both;
> > +
> > +       *parent = make_kuid(current_user_ns(), parsed_parent);
> > +       if (!uid_valid(*parent)) {
> > +               ret = -EINVAL;
> > +               goto free_both;
> > +       }
> > +
> > +       *child = make_kuid(current_user_ns(), parsed_child);
> > +       if (!uid_valid(*child)) {
> > +               ret = -EINVAL;
> > +               goto free_both;
> > +       }
> > +
> > +free_both:
> > +       kfree(parent_buf);
> > +free_kern:
> > +       kfree(kern_buf);
> > +       return ret;
> > +}
> > +
> > +static ssize_t safesetid_file_write(struct file *file,
> > +                                   const char __user *buf,
> > +                                   size_t len,
> > +                                   loff_t *ppos)
> > +{
> > +       struct safesetid_file_entry *file_entry =
> > +               file->f_inode->i_private;
> > +       kuid_t parent;
> > +       kuid_t child;
> > +       int ret;
> > +
> > +       if (!ns_capable(current_user_ns(), CAP_SYS_ADMIN))
>
> Maybe CAP_MAC_ADMIN instead of (the overloaded) CAP_SYS_ADMIN?

Makes sense. Done.

>
> > +               return -EPERM;
> > +
> > +       if (*ppos != 0)
> > +               return -EINVAL;
> > +
> > +       if (file_entry->type == SAFESETID_WHITELIST_FLUSH) {
> > +               flush_safesetid_whitelist_entries();
> > +               return len;
> > +       }
> > +
> > +       /*
> > +        * If we get to here, must be the case that file_entry->type equals
> > +        * SAFESETID_WHITELIST_ADD
>
> It seems a bit silly with only two options here, but it'll change for
> gids, yes? How about just building a switch() around file_entry->type
> instead and avoid needing to refactor this later?

Yes, there will likely be more entries when gids are introduced. Done.

>
> > +        */
> > +       ret = parse_safesetid_whitelist_policy(buf, len, &parent,
> > +                                                        &child);
> > +       if (ret)
> > +               return ret;
> > +
> > +       ret = add_safesetid_whitelist_entry(parent, child);
> > +       if (ret)
> > +               return ret;
> > +
> > +       /* Return len on success so caller won't keep trying to write */
> > +       return len;
> > +}
> > +
> > +static const struct file_operations safesetid_file_fops = {
> > +       .write = safesetid_file_write,
> > +};
> > +
> > +static void safesetid_shutdown_securityfs(void)
> > +{
> > +       int i;
> > +
> > +       for (i = 0; i < ARRAY_SIZE(safesetid_files); ++i) {
> > +               struct safesetid_file_entry *entry =
> > +                       &safesetid_files[i];
> > +               securityfs_remove(entry->dentry);
> > +               entry->dentry = NULL;
> > +       }
> > +
> > +       securityfs_remove(safesetid_policy_dir);
> > +       safesetid_policy_dir = NULL;
> > +}
> > +
> > +static int __init safesetid_init_securityfs(void)
> > +{
> > +       int i;
> > +       int ret;
> > +
> > +       safesetid_policy_dir = securityfs_create_dir("safesetid", NULL);
> > +       if (!safesetid_policy_dir) {
> > +               ret = PTR_ERR(safesetid_policy_dir);
> > +               goto error;
> > +       }
> > +
> > +       for (i = 0; i < ARRAY_SIZE(safesetid_files); ++i) {
> > +               struct safesetid_file_entry *entry =
> > +                       &safesetid_files[i];
> > +               entry->dentry = securityfs_create_file(
> > +                       entry->name, 0200, safesetid_policy_dir,
> > +                       entry, &safesetid_file_fops);
> > +               if (IS_ERR(entry->dentry)) {
> > +                       ret = PTR_ERR(entry->dentry);
> > +                       goto error;
> > +               }
> > +       }
> > +
> > +       return 0;
> > +
> > +error:
> > +       safesetid_shutdown_securityfs();
> > +       return ret;
> > +}
> > +fs_initcall(safesetid_init_securityfs);
> > --
> > 2.20.1.97.g81188d93c3-goog
> >
>
> But overall, it looks good to me. :)
>
> --
> Kees Cook