Re: [ovs-dev] openvswitch crash on i386

From: Juerg Haefliger
Date: Wed Mar 06 2019 - 10:31:35 EST


On Tue, 5 Mar 2019 11:58:42 -0800
Joe Stringer <joe@xxxxxxx> wrote:

> On Tue, Mar 5, 2019 at 2:12 AM Christian Ehrhardt
> <christian.ehrhardt@xxxxxxxxxxxxx> wrote:
> >
> > On Tue, Mar 5, 2019 at 10:58 AM Juerg Haefliger
> > <juerg.haefliger@xxxxxxxxxxxxx> wrote:
> > >
> > > Hi,
> > >
> > > Running the following commands in a loop will crash an i386 5.0 kernel
> > > typically within a few iterations:
> > >
> > > ovs-vsctl add-br test
> > > ovs-vsctl del-br test
> > >
> > > [ 106.215748] BUG: unable to handle kernel paging request at e8a35f3b
> > > [ 106.216733] #PF error: [normal kernel read fault]
> > > [ 106.217464] *pdpt = 0000000019a76001 *pde = 0000000000000000
> > > [ 106.218346] Oops: 0000 [#1] SMP PTI
> > > [ 106.218911] CPU: 0 PID: 2050 Comm: systemd-udevd Tainted: G E 5.0.0 #25
> > > [ 106.220103] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.11.1-1ubuntu1 04/01/2014
> > > [ 106.221447] EIP: kmem_cache_alloc_trace+0x7a/0x1b0
> > > [ 106.222178] Code: 01 00 00 8b 07 64 8b 50 04 64 03 05 28 61 e8 d2 8b 08 89 4d ec 85 c9 0f 84 03 01 00 00 8b 45 ec 8b 5f 14 8d 4a 01 8b 37 01 c3 <33> 1b 33 9f b4 00 00 00 64 0f c7 0e 75 cb 8b 75 ec 8b 47 14 0f 18
> > > [ 106.224752] EAX: e8a35f3b EBX: e8a35f3b ECX: 0000869f EDX: 0000869e
> > > [ 106.225683] ESI: d2e96ef0 EDI: da401a00 EBP: d9b85dd0 ESP: d9b85db0
> > > [ 106.226662] DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068 EFLAGS: 00010282
> > > [ 106.227710] CR0: 80050033 CR2: e8a35f3b CR3: 185b8000 CR4: 000006f0
> > > [ 106.228703] DR0: 00000000 DR1: 00000000 DR2: 00000000 DR3: 00000000
> > > [ 106.229604] DR6: fffe0ff0 DR7: 00000400
> > > [ 106.230114] Call Trace:
> > > [ 106.230525] ? kernfs_fop_open+0xb4/0x390
> > > [ 106.231176] kernfs_fop_open+0xb4/0x390
> > > [ 106.231856] ? security_file_open+0x7c/0xc0
> > > [ 106.232562] do_dentry_open+0x131/0x370
> > > [ 106.233229] ? kernfs_fop_write+0x180/0x180
> > > [ 106.233905] vfs_open+0x25/0x30
> > > [ 106.234432] path_openat+0x2fd/0x1450
> > > [ 106.235084] ? cp_new_stat64+0x115/0x140
> > > [ 106.235754] ? cp_new_stat64+0x115/0x140
> > > [ 106.236427] do_filp_open+0x6a/0xd0
> > > [ 106.237026] ? cp_new_stat64+0x115/0x140
> > > [ 106.237748] ? strncpy_from_user+0x3d/0x180
> > > [ 106.238539] ? __alloc_fd+0x36/0x120
> > > [ 106.239256] do_sys_open+0x175/0x210
> > > [ 106.239955] sys_openat+0x1b/0x20
> > > [ 106.240596] do_fast_syscall_32+0x7f/0x1e0
> > > [ 106.241313] entry_SYSENTER_32+0x6b/0xbe
> > > [ 106.242017] EIP: 0xb7fae871
> > > [ 106.242559] Code: 8b 98 58 cd ff ff 89 c8 85 d2 74 02 89 0a 5b 5d c3 8b 04 24 c3 8b 14 24 c3 8b 34 24 c3 8b 3c 24 c3 51 52 55 89 e5 0f 34 cd 80 <5d> 5a 59 c3 90 90 90 90 8d 76 00 58 b8 77 00 00 00 cd 80 90 8d 76
> > > [ 106.245551] EAX: ffffffda EBX: ffffff9c ECX: bffdcb60 EDX: 00088000
> > > [ 106.246651] ESI: 00000000 EDI: b7f9e000 EBP: 00088000 ESP: bffdc970
> > > [ 106.247706] DS: 007b ES: 007b FS: 0000 GS: 0033 SS: 007b EFLAGS: 00000246
> > > [ 106.248851] Modules linked in: openvswitch(E)
> > > [ 106.249621] CR2: 00000000e8a35f3b
> > > [ 106.250218] ---[ end trace 6a8d05679a59cda7 ]---
> > >
> > > I've bisected this down to the following commit that seems to have introduced
> > > the issue:
> > >
> > > commit 120645513f55a4ac5543120d9e79925d30a0156f (refs/bisect/bad)
> > > Author: Jarno Rajahalme <jarno@xxxxxxx>
> > > Date: Fri Apr 21 16:48:06 2017 -0700
> > >
> > > openvswitch: Add eventmask support to CT action.
> > >
> > > Add a new optional conntrack action attribute OVS_CT_ATTR_EVENTMASK,
> > > which can be used in conjunction with the commit flag
> > > (OVS_CT_ATTR_COMMIT) to set the mask of bits specifying which
> > > conntrack events (IPCT_*) should be delivered via the Netfilter
> > > netlink multicast groups. Default behavior depends on the system
> > > configuration, but typically a lot of events are delivered. This can be
> > > very chatty for the NFNLGRP_CONNTRACK_UPDATE group, even if only some
> > > types of events are of interest.
> > >
> > > Netfilter core init_conntrack() adds the event cache extension, so we
> > > only need to set the ctmask value. However, if the system is
> > > configured without support for events, the setting will be skipped due
> > > to extension not being found.
> > >
> > > Signed-off-by: Jarno Rajahalme <jarno@xxxxxxx>
> > > Reviewed-by: Greg Rose <gvrose8192@xxxxxxxxx>
> > > Acked-by: Joe Stringer <joe@xxxxxxx>
> > > Signed-off-by: David S. Miller <davem@xxxxxxxxxxxxx>
> >
> > Hi Juerg,
> > the symptom, the identified breaking commit and actually all of it
> > seems to be [1] which James, Joseph and I worked on already.
> > I wanted to make you aware of the past context that already exists.
> >
> > Back then we already reverted the change, found it to be working then.
> > Afterwards Joseph brought it up with Jarno [2] and got some patch it
> > seems, but that (whatever change it was - I have never seen it) wasn't
> > enough and still crashing.
> > Then we lost traction on the case and now you had to re-debug it I'm afraid :-/
> >
> > [1]: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1736390
> > [2]: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1736390/comments/55
> >
> > > Reverting that commit from 5.0 makes the problem go away. I'm not able to
> > > reproduce the crash on x86_64.
> > >
> > > ...Juerg
> > > _______________________________________________
> > > dev mailing list
> > > dev@xxxxxxxxxxxxxxx
> > > https://mail.openvswitch.org/mailman/listinfo/ovs-dev
> >
> >
> >
> > --
> > Christian Ehrhardt
> > Software Engineer, Ubuntu Server
> > Canonical Ltd
>
> Somehow the patch was shared on three different open source lists (the
> bug, linux-kernel, and ovs-dev), yet only one of the three actually
> retained the message:
>
> https://mail.openvswitch.org/pipermail/ovs-dev/2018-September/352395.html

Thanks for the link, Jarno, but that patch doesn't help at all.

FWIW, removing the following case statement makes the test pass:

case OVS_CT_ATTR_EVENTMASK:
info->have_eventmask = true;
info->eventmask = nla_get_u32(a);
break;

But changing it to something like the below also blows up the machine:

case OVS_CT_ATTR_EVENTMASK:
info->have_eventmask = false;
info->eventmask = nla_get_u32(a);
break;

Also bad:

case OVS_CT_ATTR_EVENTMASK:
break;

I'm not saying these tests make any sense, just saying :-)

...Juerg

Attachment: pgpFnQRN5DZAR.pgp
Description: OpenPGP digital signature