Re: Xen unstability on HP Moonshot m400

[Mark Salter <msalter@redhat.com>]

On Tue, 2015-03-24 at 09:54 -0400, Mark Salter wrote:
> On Mon, 2015-03-23 at 23:58 +0000, Stefano Stabellini wrote:
> > On Mon, 23 Mar 2015, Christoffer Dall wrote:
> > > On Mon, Mar 23, 2015 at 1:36 PM, Ian Campbell <ian.campbell@citrix.com> wrote:
> > >       On Sat, 2015-03-21 at 13:34 +0100, Christoffer Dall wrote:
> > >       > Hi,
> > >       >
> > >       > I have been experiencing a problematic crash running Xen on m400 over
> > >       > the last few days.  I already spoke to Ian and Stefano about this, but
> > >       > thought I'd summarize what I've seen so far and loop in a wider
> > >       > audience.
> > >       >
> > >       > The basic setup is this:
> > >       >  - Two m400 nodes, one running Linux bare-metal, the other running
> > >       > Xen.
> > >       >  - The Xen node runs Dom0 and 1 DomU
> > >       >  - The m400 has a Mellanox Connectx-3 PCIe 10G ethernet card with two
> > >       > parts on it
> > >       >  - Dom0 uses NAT forwarding from Dom0's eth0 (which is connected to
> > >       > the internet) and regular bridging to eth1 which is connected to a
> > >       > private VLAN to the bare-metal node
> > >       >  - Dom0 and DomU are configured with 14GB of ram, 4 cpus each
> > >       >  - DomU runs apache2 serving the GCC manual (see
> > >       > https://github.com/chazy/kvmperf/blob/master/cmdline_tests/apache_install.sh)
> > >       >
> > >       > The bare-metal node runs apache bench, like this: "ab -n 100000 -c 100
> > >       >http://secure-web.cisco.com/1r5tZ8-7RF8gHRANwFdizEZzgeMsjxVO0yKbYiV4zy7LeiUfYBXMkFq7FGW_SZ1x-VxdzyK-ErDsOUiQ9z2x-N
> > > y7XkL_loHP8ene_BuNFscGyWmQ3r6CtXAYaZCY4xRmmPT1uJOsZDLMu7j-LfCOGmQDSdBwgW7QYukI2bCtTrXM/http%3A%2F%2F10.10.1.120%2F
> > >       gcc%2Findex.html"
> > >       >
> > >       > (10.10.1.120 is the DomU IP address of the bridged interface to eth1)
> > >       >
> > >       > What happens now is that the entire Xen node goes down.  I see various
> > >       > errors in the kernel log, some examples:
> > >       > http://pastebin.ubuntu.com/10642148/
> > >       > http://pastebin.ubuntu.com/10642177/
> > >       > http://pastebin.ubuntu.com/10642181/
> > >       > http://pastebin.ubuntu.com/10635573/
> > >       >
> > >       >
> > >       > All Linux kernels are 3.18 plus some tweaks for the m400 cartridge:
> > >       > https://github.com/columbia/linux-kvm-arm/tree/columbia-armvirt-3.18
> > > 
> > >       Is it worth adding
> > >       https://git.kernel.org/cgit/linux/kernel/git/arm64/linux.git/commit/?id=285994a62c80f1d72c6924282bcb59608098d5ec
> > >       to your kernel? It isn't Xen specific but it's perhaps possible that Xen opens the window wider.
> 
> You definitely want that one. Without it, the page table walker could
> end up using a stale pointer to a page being used for something other
> than page tables.
> 
> > > 
> > >       How confident are you in
> > >       https://github.com/columbia/linux-kvm-arm/commit/5e29cb0478f3d90e4f568d6bea6840960331bcbb ?
> > >       (although I suppose you aren't running in ACPI mode if you are running
> > >       Xen?)
> > > 
> > > 
> > > I'm not confident at all, but Linux (last I checked was v3.19) doesn't boot without it, so not sure if there's an
> > > alternative?  Mark?
> > 
> > This patch is key: it doesn't look like it is setting
> > dev->archdata.dma_coherent appropriately, see the implementation of
> > set_arch_dma_coherent_ops.
> 
> You'd want this if booting with ACPI. You might also need it for
> enumerated PCI devices even if booting with devicetree.

There's an updated version of this patch for newer kernels in the
devel branch of git.fedorahosted.org/git/kernel-arm64.git

There is also this one in Linus' tree which may be of interest to you:

commit 7132813c384515c9dede1ae20e56f3895feb7f1e
Author: Suzuki K. Poulose <suzuki.poulose@arm.com>
Date:   Thu Mar 19 18:17:09 2015 +0000

    arm64: Honor __GFP_ZERO in dma allocations

> > 
> > 
> > > 
> > >       If we think the issue might be to do with coherency of foreign mappings
> > >       undergoing i/o from dom0 and we've already ruled out disk (by using a
> > >       loopback mounted rootfs) then it might be worth bodging netback to
> > >       always copy too.
> > > 
> > >       Adding a call to skb_orphan_frags right before the netif_receive_skb in
> > >       drivers/net/xen-netback/netback.c:xenvif_tx_submit is a simple but
> > >       rather inefficient way of doing that (so I hope it doesn't perturb the
> > >       issue).
> > > 
> > > 
> > > I'll be happy to try this.
> > 
> > If we are right and the problem is due to the commit above not setting
> > dma_coherent to true (the kernel will think that actually the network
> > card is not coherent), then Ian's workaround should hide the problem.
> > 
> > 
> > > 
> > >       Stefano (who is more familiar with the Linux swiotlb side of things than
> > >       me) is travelling this week so he'll be on West coast time, not sure
> > >       when he gets off a plane nor if he's on email anyway (he's at ELC + this
> > >       ARM ACPI thing)
> > > 
> > > 
> > > ok, we'll see what happens.
> > > 
> > > -Christoffer
> > > 
> > > 
>