[LU-12233] Deadlock on LNet shutdown

• Type: Bug
• Resolution: Fixed
• Priority: Major
• Fix Version/s: Lustre 2.14.0, Lustre 2.12.6
• Affects Version/s: Lustre 2.13.0, Lustre 2.12.1, Lustre 2.12.3
• Severity: 3
• Assignee: Serguei Smirnov
• Reporter: Chris Horn

Description

I reproduced this issue with master and Cray's 2.12 branch. For completeness I'll note that my master was slightly modified so that I can configure LNet on Cray's hardware, and I also applied the fix from LU-11756.

Here's the relevant git log. Commit '8cb7ccf54e' is on master.

86ef522cac LU-11756 o2iblnd: kib_conn leak
888adb9340 MRP-342 lnet: add config file support
d661c584c6 Revert "LU-11838 lnet: change lnet_ipaddr_enumerate() to use for_each_netdev()"
4c681cf4ee Revert "LU-11838 o2iblnd: get IP address more directly."
f4fe014620 Revert "LU-6399 lnet: socket cleanup"
8cb7ccf54e LU-11986 lnet: properly cleanup lnet debugfs files

LNetNIFini() takes the ln_api_mutex and then shuts down LNet. It doesn't release the mutex until all teardown functions have returned.

The message receive path also takes the ln_api_mutex in lnet_nid2peerni_locked().
kgnilnd_check_fma_rx->lnet_parse->lnet_nid2peerni_locked
kiblnd_handle_rx->lnet_parse->lnet_nid2peerni_locked
ksocknal_process_receive->lnet_parse->lnet_nid2peerni_locked

/*
 * Get a peer_ni for the given nid, create it if necessary. Takes a
 * hold on the peer_ni.
 */
struct lnet_peer_ni *
lnet_nid2peerni_locked(lnet_nid_t nid, lnet_nid_t pref, int cpt)
{
        struct lnet_peer_ni *lpni = NULL;
        int rc;

        if (the_lnet.ln_state != LNET_STATE_RUNNING)
                return ERR_PTR(-ESHUTDOWN);

        /*
         * find if a peer_ni already exists.
         * If so then just return that.
         */
        lpni = lnet_find_peer_ni_locked(nid);
        if (lpni)
                return lpni;

        /*
         * Slow path:
         * use the lnet_api_mutex to serialize the creation of the peer_ni
         * and the creation/deletion of the local ni/net. When a local ni is
         * created, if there exists a set of peer_nis on that network,
         * they need to be traversed and updated. When a local NI is
         * deleted, which could result in a network being deleted, then
         * all peer nis on that network need to be removed as well.
         *
         * Creation through traffic should also be serialized with
         * creation through DLC.
         */
        lnet_net_unlock(cpt);
        mutex_lock(&the_lnet.ln_api_mutex);

int
LNetNIFini()
{
        mutex_lock(&the_lnet.ln_api_mutex);

        LASSERT(the_lnet.ln_refcount > 0);

        if (the_lnet.ln_refcount != 1) {
                the_lnet.ln_refcount--;
        } else {
                LASSERT(!the_lnet.ln_niinit_self);

                lnet_fault_fini();

                lnet_router_debugfs_init();
                lnet_peer_discovery_stop();
                lnet_push_target_fini();
                lnet_monitor_thr_stop();
                lnet_ping_target_fini();

                /* Teardown fns that use my own API functions BEFORE here */
                the_lnet.ln_refcount = 0;

                lnet_acceptor_stop();
                lnet_destroy_routes();
                lnet_shutdown_lndnets(); <<<  the_lnet.ln_state = LNET_STATE_STOPPING; happens here
                lnet_unprepare();
        }

        mutex_unlock(&the_lnet.ln_api_mutex);
        return 0;
}
EXPORT_SYMBOL(LNetNIFini);

We can see there is a decent sized window where the deadlock can be hit.

It is easy to reproduce for me.

[root@snx11922n000 ~]# pdsh -g lustre modprobe lnet; lctl net up ; lctl list_nids ; lctl ping 10.12.0.50@o2ib40 ; lctl net down ; lustre_rmmod

Sometimes the command needs to be repeated a couple of times.

I believe this regression was introduced by:

commit fa8b4e6357c53ea457ef6624b0b19bece0b0fdde
Author: Amir Shehata <[e-mail]>
Date:   Thu May 26 15:42:39 2016 -0700

    LU-7734 lnet: peer/peer_ni handling adjustments

Issue Links

• is related to: LU-13807 LNet: Net delete dead lock deadlock