Bug 1499492

Hi,

the following part of the code has raised the error:

exec.c
------------------------------------------------------------------------
+static void qemu_ram_setup_dump(void *addr, ram_addr_t size)
+{
+    int ret;
+    QemuOpts *machine_opts;
+
+    /* Use MADV_DONTDUMP, if user doesn't want the guest memory in the core */
+    machine_opts = qemu_opts_find(qemu_find_opts("machine"), 0);
+    if (machine_opts &&
+        !qemu_opt_get_bool(machine_opts, "dump-guest-core", true)) {
+        ret = madvise(addr, size, MADV_DONTDUMP);
+        if (ret) {
+            perror("qemu_madvise");
+            fprintf(stderr, "madvise doesn't support MADV_DONTDUMP, "
+                            "but dump_guest_core=off specified\n");
+        }
+    }
+}
------------------------------------------------------------------------

Please also review and comment the function madvise_behavior_valid(int behavior) and send me the link to the unit test.

I think the bug is caused by missing documentation/wrong understanding or wrong implementation.


BR, Mota

Please provide the version of RHEL running on the host. Is it RHEL 7.3? 

In particular, what is the host kernel version? 

For completeness, please also provide the libvirt XML for this VM and the libvirt version. Thanks.

Created attachment 1336133 [details]
XML_KVM_for_SAP_Performance_testing

Yes, they are both RHEL 7.3.

Host and VM kernel versions are:

3.10.0-514.32.2.el7.x86_64 #1 SMP Mon Sep 11 05:36:21 EDT 2017 x86_64 x86_64 x86_64 GNU/Linux

# libvirtd --version
libvirtd (libvirt) 3.2.0


I attached to the Bug the XML for VM:
XML_KVM_for_SAP_Performance_testing

More Info:

I'm using the following qemu versions:

- qemu-img-rhev 2.6.0-28.el7_3.12
- qemu-kvm-common-rhev 2.6.0-28.el7_3.12
- qemu-kvm-rhev 2.6.0-28.el7_3.12
- qemu-kvm-tools-rhev 2.6.0-28.el7_3.12


Thanks,
BR, Mota

Hi Mota,

> memory-backend-file,id=ram-node0,prealloc=yes,mem-path=/dev/hugepages/
> libvirt/qemu/2-ls3059v1,size=387352363008,host-nodes=0,policy=bind: cannot
> bind memory to host NUMA nodes: Invalid argument

as already discussed in email, this error probably is related to your cmdline setting for disabling memory cgroups. As such the NUMA pinning can't be applied correctly.

Please remove the setting
  cgroup_disable=memory
from your kernel cmdline and try starting the VM again.

Thanks!
Martin

(In reply to Martin Tessun from comment #6)
> Hi Mota,
> 
> > memory-backend-file,id=ram-node0,prealloc=yes,mem-path=/dev/hugepages/
> > libvirt/qemu/2-ls3059v1,size=387352363008,host-nodes=0,policy=bind: cannot
> > bind memory to host NUMA nodes: Invalid argument
> 
> as already discussed in email, this error probably is related to your
> cmdline setting for disabling memory cgroups. As such the NUMA pinning can't
> be applied correctly.
> 
> Please remove the setting
>   cgroup_disable=memory
> from your kernel cmdline and try starting the VM again.
> 
> Thanks!
> Martin

After doing the math correctly the error is quite obvious:
1. We have 1GB hugepages defined
2. We align 1443GB to a total of 4 Numa Nodes
3. As we distribute equally, we end up with 360.75GB per Numa Node.

So increasing the memory by 1GB and adjusting the NUMA Node memory to 361GB per cell, got the machine starting. So at least the last message could be solved this way and the VM does start now.

Hi All,

first of all my special thanks to Martin Tessun. He did a great job, great to have you in the team. 

After Martin has changed the NUMA configuration as he described, the VM is on running again and I can continue with the Performance testing.

We will adjust the "SAP HANA on KVM: Best Practices Resource Guide".

Thanks and best regards,
Mota

Closing as not a bug. If you feel otherwise, please reopen. Thanks!