This is a page tracks issues that we run into while testing the Multi-Rail Feature

Problem List

Description Priority Reporter Notes

with 17 interfaces trying to discover on the any of the interface the first time returns an error "no route to host".

Ok was able to reproduce. If I follow the steps in UT-DD-EN-0005 exactly, then the first time I try to discover any of the nids it fails

steps to reproduce

P1 && P2
  756  lnetctl lnet configure
  757  lnetctl net add --net tcp --if eth0,eth1,eth2,eth3,eth4,eth5,eth6,eth7,eth8,eth9,eth10,eth11,eth12,eth13,eth14,eth15,eth16
 
P1:
lnetctl ping --nid 192.168.122.42@tcp


P2:
lnetctl ping --nid 192.168.122.23@tcp


P1:
[root@MRtest01 Lustre]# lnetctl discover --nid 192.168.122.46@tcp
manage:
    - discover:
          errno: -1
          descr: failed to discover 192.168.122.46@tcp: No route to host


Second time it works

critical

Amir

With 17 interface discovered "show peer" hangs

When the rc from the kernel is not 0. The structure is not copied out of the kernel to user space. The code depends on that in order to pass the new size if the data to be copied out is too big for the buffer passed in by the user. Since that doesn't happen when rc == -E2BIG, user space code gets into an infinite loop sending IOCTLs to the kernel

from libcfs_ioctl()
 
145 »·······»·······»·······if (err == 0) {
146 »·······»·······»·······»·······if (copy_to_user(uparam, hdr, hdr->ioc_len))
147 »·······»·······»·······»·······»·······err = -EFAULT;
148 »·······»·······»·······}
149 »·······»·······»·······break;


The buffer is only copied to user space if the ioctl handler returns 0. Not really sure if it's safe to change that.

Amir

This has been fixed

"lnetctl discover" command hangs with discovery off. This happened once, so an intermittent issue. Will try to reproduce.

Major

Sonia

"lnetctl discover" discovers the peer even with discovery off

Major

Sonia

"lnetctl discover --force" expects a parameter (no parameter should be needed with --force). Major Doug

Doug: I configured a Parallels VM with 16 interfaces (won't let me do 17 as 16 is a limit). When I "lctl network configure" with no YAML or module parameters, I get this error from ksocklnd: "Mar 7 14:01:16 centos-7 kernel: LNet: 5111:0:(socklnd.c:2652:ksocknal_enumerate_interfaces()) Ignoring interface virbr0 (too many interfaces)". Minor Doug

Doug: When discovering a node with 16 interfaces via "lnetctl discover --nid", it works, but I am seeing this log:

Mar  7 15:48:04 centos-7 kernel: LNetError: 24769:0:(peer.c:1726:lnet_peer_push_event()) Push Put from unknown 0@<0:0> (source 0@<0:0>)

Minor

Doug

Olaf: I've seen the message before. Not sure how best to get rid of it, but it would be safe to just not emit it for the loopback case.

Doug: Tried the command "lnetctl set discovery" (no args) and got a core dump.

Major

Doug

Olaf: As I recall, the problem is parse_long() is passed a NULL pointer for its number parameter in this case, and it doesn't check for that. Easiest to fix in parse_long() rather than fix all callers. This bug affects all lnetctl set commands.

Doug: How do I query the discovery setting? There is no "lnetctl show discovery" (should there be?). I tried "lnetctl set discovery" with no parameters and that core dumped (see previous bullet). From a usability perspective, there is no obvious way to get this information.

Minor

Doug

Doug: was told it is "lnetctl global show". Don't like that (see Usability section) but see this problem as solved.

Doug: When I enter "lnetctl set" to see what options I can set, I get this:

[root@centos-7 ~]# lnetctl set
set {tiny_buffers | small_buffers | large_buffers | routing}

It does not mention "discovery" at all.

Major

Doug

Olaf: So the text in list[] for the set subcommand needs to be updated. Note that max_interfaces needs to be added there as well.

When you do "lnetctl export" global section doesn't show discovery status

Minor

Amir

Doug: Test: UT-DD-EN-0002. The first call to discover P2 fails with this:

[root@centos-7 ~]# lnetctl discover --nid 10.211.55.62@tcp
manage:
    - discover:
          errno: -1
          descr: failed to discover 10.211.55.62@tcp: No route to host

The second attempt works as expected. I repeated the test twice and got the same result each time.

Critical

Doug

Doug: Test: UT-DD-EN-0003. Same behaviour as above.

Critical

Doug

Amir: There should be a way to turn off Multi-Rail. From the code the LNET_PING_FEAT_MULTI_RAIL is not unsettable.

Critical

Amir

Doug: Test: UT-DD-DIS-0001. Test passed and worked as described. However, I decided to run lnet-selftest to see how running some traffic after the test goes. The lnet-selftest failed (never stopped running, did not show any stats). I looked at top and can see that the "lnet-discovery" thread is using 70% CPU (it was not using any CPU prior to running lnet-selftest). I suspect it is receiving all in coming traffic so lnet-selftest is not getting anything. Additional note: I just redid the this test but ran lnet-selftest "before" trying to invoke discovery. lnet-discovery thread still takes off and lnet-selftest locks. Seems that turning off discovery causes lnet-discovery thread to misbehave.

Blocker

Doug

So this doesn't have to do specifically with the test case. It's when discovery is off and you run lnet_selftest. From the logs I collected it appears that for each selftest message being sent, the same NID gets queued on the discovery thread. But It doesn't end up doing anything. So in effect it goes into a crazy loop, trying to discover, but because it's off, it doesn't and probably doesn't update the state properly, so the next time to the same peer triggers the nid to be queued on the discovery thread again. Since the discovery thread does pretty heavy locking, it drives the system into a grind. Selftest also reacts poorly and hangs the node.

I don't think we should be queuing anything on the discovery thread if discovery is off.

This has been fixed

Doug: After the previous point, I tried to kill the lnet-discovery thread. It did not stop. I then tried "kill -9". It still did not stop. I then did a reboot. Node went away and could not complete the reboot because it could not unload LNet. Had to reset the node. We need a way to stop any of our worker threads when things do not go well. Hard resetting a node in the field will be unacceptable to customers.

Blocker

Doug

This has been fixed

DD doesn't handle the case where a Mulit-Rail peer is torn down and then booted with a downrev Lustre (non-mr). This needs to be handled. Both this scenario and turning off the Multi-Rail feature are going to be handled fairly similarly

node gets configured to !MR
push with flag not set
peer receives the push tears down the peer and recreates
From now on, the peer is viewed as non-MR.
Future message exchange will setup the preferred-NID

in case of a down-rev
admin will need to trigger an explicit discover
ping response comes back with feature bit off
rest is same as above

Blocker

Amir

DD doesn't send a push when an interface is added to an existing network

Blocker

Amir

The functionality to trigger a push when the configuration is updated is missing.

Olaf: Adding an interface should cause lnet_peer_needs_push() to return true, therefore lnet_peer_is_uptodate() to return false, which in turn triggers discovery when there is traffic to the peer. Moreover, LNet-internal traffic (traffic to portal 0) will never trigger discovery, see lnet_msg_discovery().

Amir: As discussed there is the scenario where triggering discovery on traffic is not sufficient, in case one of the peers changes it's primary interface, or even all of its interfaces. The node initiating traffic will not be able to access it.

LASSERT hit with the latest timeout patch

<0>LNetError: 4706:0:(peer.c:1704:lnet_peer_discovery_complete()) ASSERTION( lp->lp_state & (1 << 4) ) failed:
<0>LNetError: 4706:0:(peer.c:1704:lnet_peer_discovery_complete()) LBUG
<4>Pid: 4706, comm: lnet_discovery
<4>
<4>Call Trace:
<4> [<ffffffffa0c8b885>] libcfs_debug_dumpstack+0x55/0x80 [libcfs]
<4> [<ffffffffa0c8b9cf>] lbug_with_loc+0x3f/0x90 [libcfs]
<4> [<ffffffffa0d2e297>] lnet_peer_discovery_complete+0x147/0x150 [lnet]
<4> [<ffffffffa0d33ffd>] lnet_peer_discovery+0xe5d/0x1440 [lnet]
<4> [<ffffffff8152a6be>] ? thread_return+0x4e/0x7d0
<4> [<ffffffff81064bd2>] ? default_wake_function+0x12/0x20
<4> [<ffffffff8109ebb0>] ? autoremove_wake_function+0x0/0x40
<4> [<ffffffffa0d331a0>] ? lnet_peer_discovery+0x0/0x1440 [lnet]
<4> [<ffffffff8109e71e>] kthread+0x9e/0xc0
<4> [<ffffffff8100c20a>] child_rip+0xa/0x20
<4> [<ffffffff8109e680>] ? kthread+0x0/0xc0
<4> [<ffffffff8100c200>] ? child_rip+0x0/0x20
<4>
<0>Kernel panic - not syncing: LBUG
<4>Pid: 4706, comm: lnet_discovery Not tainted 2.6.32.504.16.2.el6_lustre #1
<4>Call Trace:
<4> [<ffffffff81529fbc>] ? panic+0xa7/0x16f
<4> [<ffffffffa0c8b9e6>] ? lbug_with_loc+0x56/0x90 [libcfs]
<4> [<ffffffffa0d2e297>] ? lnet_peer_discovery_complete+0x147/0x150 [lnet]
<4> [<ffffffffa0d33ffd>] ? lnet_peer_discovery+0xe5d/0x1440 [lnet]
<4> [<ffffffff8152a6be>] ? thread_return+0x4e/0x7d0
<4> [<ffffffff81064bd2>] ? default_wake_function+0x12/0x20
<4> [<ffffffff8109ebb0>] ? autoremove_wake_function+0x0/0x40
<4> [<ffffffffa0d331a0>] ? lnet_peer_discovery+0x0/0x1440 [lnet]
<4> [<ffffffff8109e71e>] ? kthread+0x9e/0xc0
<4> [<ffffffff8100c20a>] ? child_rip+0xa/0x20
<4> [<ffffffff8109e680>] ? kthread+0x0/0xc0
<4> [<ffffffff8100c200>] ? child_rip+0x0/0x20

(gdb) l* lnet_peer_discovery_complete+0x147
0x2e2b7 is at /home/ashehata/LustreBuild/mr-dd/lnet/lnet/peer.c:1704.
1699    {
1700            CDEBUG(D_NET, "Dequeue peer %s\n",
1701                   libcfs_nid2str(lp->lp_primary_nid));
1702
1703            spin_lock(&lp->lp_lock);
1704            LASSERT(lp->lp_state & LNET_PEER_QUEUED);
1705            lp->lp_state &= ~LNET_PEER_QUEUED;
1706            spin_unlock(&lp->lp_lock);
1707            list_del_init(&lp->lp_dc_list);
1708            wake_up_all(&lp->lp_dc_waitq);
(gdb) l *lnet_peer_discovery+0xe5d
0x3401d is in lnet_peer_discovery (/home/ashehata/LustreBuild/mr-dd/lnet/lnet/peer.c:3013).
3008            lnet_net_lock(LNET_LOCK_EX);
3009            list_for_each_entry(lp, &the_lnet.ln_dc_request, lp_dc_list) {
3010                    lnet_peer_discovery_error(lp, -ESHUTDOWN);
3011                    lnet_peer_discovery_complete(lp);
3012            }
3013            list_for_each_entry(lp, &the_lnet.ln_dc_working, lp_dc_list) {
3014                    lnet_peer_discovery_error(lp, -ESHUTDOWN);
3015                    lnet_peer_discovery_complete(lp);
3016            }
3017            list_for_each_entry(lp, &the_lnet.ln_dc_expired, lp_dc_list) {

To reproduce

Modified the code to not send a REPLY for the GET
initiate discovery from node. 
timeout set 180
CTRL-C command
lnetctl lnet unconfigure
**crash

Blocker

This seems to be a bit of a race there. I can't reproduce again.

Doug: I have a bug which can be reproduced by requires these very specific steps to do so.

Start with a node and a peer. I have two interfaces configured for the node and 3 for the peer. I leave discovery on in both the node and peer.

In the peer, manually configure the node (as a peer) to have one interface, non-MR:

[root@centos-7 ~]# !359
lnetctl peer add --prim_nid 10.211.55.58@tcp --non_mr
[root@centos-7 ~]# lnetctl peer show
peer:
    - primary nid: 10.211.55.58@tcp
      Multi-Rail: True
      peer ni:
        - nid: 10.211.55.58@tcp
          state: NA

Usability Issues

Description	Priority	Reporter
Doug: How do I query the discovery setting? There is no "lnetctl show discovery" (should there be?). I tried "lnetctl set discovery" with no parameters and that core dumped (see previous section). From a usability perspective, there is no obvious way to get this information.		Doug
Doug: I understand that Chris Morrone pushed for us to have "lnetctl set <key> <value>". That breaks from the original paradigm to follow which was done with the Linux "ip" command. It was designed to be: "ip <object> <action> <optional params>". So, it would be more logical to have: "lnetctl discover set 0/1" than "lnetctl set discovery 0/1". Then starting discovery can be: "lnetctl discover start <nids>". Looking for discovery status can be: "lnetctl discovery show". I found myself guessing at these commands as I have given here and had no idea to look at "set".		Doug
Doug: You set the max interfaces with "lnetctl set max_interfaces" but is it shown as "max_intf" in the global settings. Should be the same for consistency.		Doug
Document the behavior of Dynamic Discovery, including what type of traffic triggers discovery.	Critical	Amir