Re: [BUG RFD]selinux: sidtab_context_to_sid return -NOMEM when concurrent with security_load_policy

[Li Kun <hw.likun@xxxxxxxxxx>]

On 2018/1/2 12:16, Stephen Smalley wrote:

On Jan 1, 2018 8:40 PM, "Li Kun" <hw.likun@xxxxxxxxxx> wrote:

在 2017/12/30 1:25, Stephen Smalley 写道:

On Dec 28, 2017 10:14 PM, "Li Kun" <hw.likun@xxxxxxxxxx> wrote:

在 2017/12/28 22:57, Stephen Smalley 写道:

On Sat, Dec 23, 2017 at 3:03 AM, Li Kun <hw.likun@xxxxxxxxxx> wrote:

Hi all,
When i start a docker container, the runc will call selinux_setprocattr to set the exec_sid before start the container.
Meanwhile if i use "semodule -i" to load a policy pp, the old sidtab will be shutdown before switch to the new sidtab, and cause
sidtab_context_to_sid  failed with the errno -NOMEM.

Error message:
Dec 14 09:18:18 wuwenming_vudfua_0 docker: time="2017-12-14T09:18:18.549862564+08:00" level=error msg="Handler for GET /v1.23/containers/192.168.0.2:9082/dfs_d:V100R013C11SPC000BCD2/json returned error: No such container: 192.168.0.2:9082/dfs_d:V100R013C11SPC000BCD2"
Dec 14 09:18:18 wuwenming_vudfua_0 kernel: [   48.463179] SELinux: security_context_to_sid(system_u:object_r:svirt_sandbox_file_t:s0:c290,c371) failed for (dev overlay, type overlay) errno=-12
Dec 14 09:18:18 wuwenming_vudfua_0 dbus-daemon: dbus[720]: avc:  received policyload notice (seqno=2)
Dec 14 09:18:18 wuwenming_vudfua_0 dbus[720]: avc:  received policyload notice (seqno=2)


runc                                                                                                                                    semodule
->selinux_setprocattr                                                                                                           ->security_load_policy
    ->security_context_to_sid                                                                                                        ->sidtab_shutdown(oldsidtab)
        ->read_lock(&policy_rwlock);                                                                                              ->sidtab_map(&oldsidtab, clone_sid, &newsidtab);
       
        ->sidtab_context_to_sid                                                       
            {
                        ....
                        if (s->next_sid == UINT_MAX || s->shutdown) {
                        ret = -ENOMEM;
                        ....
            }
        ->read_unlock(&policy_rwlock);                                                                                               
                                                                                                                                                ->write_lock_irq(&policy_rwlock);
                                                                                                                                                ->"switch to new policydb"
                                                                                                                                                ->write_unlock_irq(&policy_rwlock);

I wonder that if this is the intention or a bug?
If it is the intention, what should the application do when it get -ENOMEM error,  to try again?
If it is a bug, may the two options below suitable to solve the issue?
Option1:
    use policy_rwlock to protect the "sidtab_shutdown" & "sidtab_map" , load_policy is rarely to be called after system bringing up, 
so i think it will not impact much on the performance.
Option2:
    Use a temp list to store the sid in oldsidtab after it is shutdown, and deal with it
after cloning the major sids from oldsidtab to newsidtab and getting the policy_rwlock.

The current logic is intentional.  One might argue that a different error code might be more suitable in this situation (e.g. EAGAIN or something), but ENOMEM is already a possible error code when loading policy upon transient OOM conditions, so the caller might have to retry regardless.  The policy write lock must only be held for the minimal critical section as otherwise all system activity could be blocked; it is only held when making the new policydb and sidtab active not during their allocation and setup - ideally it would be further reduced to a couple of pointer updates. Is this scenario something you are encountering in actual practice or just a theoretically possible one? 

Thank you very much for your reply.
Yes, we encounter this issue in our implementation.
We have an intertion in our design that we can divide the policy into several parts to load separately.
So after the red-hat linux bringing up, which has loaded the system policy db released by redhat, we begin to load the policy released by our product ,
and meanwhile someone want to start a docker container, and the runc get an ENOMEM error like i mentioned above.
Should we merge all the policy into the system policy db and just load once?

-- 
Best Regards
Li Kun

Typically you would install your policy module when your package is installed and not need to do it at runtime.

Would you think it's suibable to change the error code to EAGAIN？

You likely want it to return - EINTR or -ERESTARTSYS if you want existing callers of setprocattr to automatically restart the call instead of failing.

AFAIU, the EINTR & ERESTARTSYS are releated to sigpending under most circumstances, is it proper to use them here?
And as i mentioned above, most apps(like runc, bind)do not retry when they meet ENOMEM, so i think maybe EAGAIN is more suitable for the transient race conditions?

-- 
Best Regards
Li Kun