[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-ppc] [PATCH] hw/ppc: disable hotplug before CAS is completed

From: Daniel Henrique Barboza
Subject: Re: [Qemu-ppc] [PATCH] hw/ppc: disable hotplug before CAS is completed
Date: Wed, 23 Aug 2017 09:58:20 -0300
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.2.1

On 08/22/2017 09:24 PM, David Gibson wrote:
On Tue, Aug 22, 2017 at 08:50:44PM -0300, Daniel Henrique Barboza wrote:

On 08/18/2017 01:14 AM, David Gibson wrote:
On Thu, Aug 17, 2017 at 06:31:28PM -0300, Daniel Henrique Barboza wrote:
On 08/17/2017 04:52 AM, David Gibson wrote:
On Tue, Aug 15, 2017 at 05:28:46PM -0300, Daniel Henrique Barboza wrote:
This patch is a follow up on the discussions that started with
Laurent's patch series "spapr: disable hotplugging without OS" [1]
and discussions made at patch "spapr: reset DRCs on migration
pre_load" [2].

At this moment, we do not support CPU/memory hotplug in early
boot stages, before CAS. The reason is that the hotplug event
can't be handled at SLOF level (or even in PRELAUNCH runstate) and
at the same time can't be canceled. This leads to devices being
unable to be hot unplugged and, in some cases, guest kernel Ooops.
After CAS, with the FDT in place, the guest can handle the hotplug
events and everything works as usual.

An attempt to try to support hotplug before CAS was made, but not
successful. The key difference in the current code flow between a
coldplugged and a hotplugged device, in the PRELAUNCH state, is that
the coldplugged device is registered at the base FDT, allowing its
DRC to go straight to CONFIGURED state. In theory, this can also be
done with a hotplugged device if we can add it to the base of the
existing FDT. However, tampering with the FDT after writing in the
guest memory, besides being a dubitable idea, is also not
possible. The FDT is written in ppc_spapr_reset and there is no
way to retrieve it - we can calculate the fdt_address but the
fdt_size isn't stored. Storing the fdt_size to allow for
later retrieval is yet another state that would need to be
migrated. In short, it is not worth the trouble.

All this said, this patch opted to disable CPU/mem hotplug at early
boot stages. CAS detection is made by checking if there are
any bits set in ov5_cas to avoid adding an extra state that
would need tracking/migration. The patch also makes sure that
it doesn't interfere with hotplug in the INMIGRATE state.

[1] https://lists.gnu.org/archive/html/qemu-devel/2017-05/msg05226.html
[2] https://lists.gnu.org/archive/html/qemu-devel/2017-07/msg01989.html

Signed-off-by: Daniel Henrique Barboza <address@hidden>
I don't think this is a good idea.

1) After my DRC cleanups, early hotplug works just fine for me.  I'm
not sure why it isn't for you: we need to understand that before

2) libvirt actually uses early hotplug fairly often (before even
starting the firmware).  At the moment this works - at least in some
cases (see above), though there are some wrinkles to work out.  This
will break it completely and require an entirely different approach to
fix again.
Now that you mentioned I remember having this same discussion with you,
about the same topic. Back then we decided to leave it alone, since you
reproduce the behavior but I could.

I still can reproduce this bug and ended up investigating a bit more today:

- one difference in QEMU between hotplugging before and after CAS is here:

hw/ppc/spapr_events.c - rtas_event_log_to_source

      switch (log_type) {
          source = spapr_event_sources_get_source(spapr->event_sources,
          if (spapr_ovec_test(spapr->ov5_cas, OV5_HP_EVT)) {
          /* fall back to epow for legacy hotplug interrupt source */
      case RTAS_LOG_TYPE_EPOW:
          source = spapr_event_sources_get_source(spapr->event_sources,

Note the ovec_test for OV5_HP_EVT. When hotplugging a CPU In early boot,
doesn't have anything set, making this check fails and due to the 'break'
position there
(that I believe it is intended), it falls back to log the event as EPOW
instead of HOT_PLUG.
Ah.. I'm not sure if this is the cause of the problem you're seeing,
but it's a good point regardless.  We're queuing an event before we
know how to queue events, which is definitely wrong.

Urgh.. how to fix this.  Banning hotplug until after CAS seems
logically sound, but will break the libvirt use case.  So we could..

1) Queue events internally, but only deliver them to the right actual
RTAS queue after CAS.  Blech.

2) At CAS time, reset all the DRCs, and emit extra DT information for
anything hotplugged since init.  I've suggested this before, but
dropped it when early hotplug seemed to be working for me.
Update: I've been trying to do the (2) solution here by adding extra
DT info of the hotplugged CPU, while resetting its DRC. The problem I am
facing is that the kernel panics at parse_numa_properties when trying to
find the node id of the added CPU. This is intriguing because the extra DT
info is enough to make the kernel start the added CPU at smp_release_cpus().

The DT information is being added at the CAS response, in
and it's the same information that is added for the coldplugged CPUs. I have
theory that the guest (SLOF?) isn't prepared to handle a new CPU being added
this point, which might be leading to this kernel bug I am seeing. I
appreciate if
anyone with good CAS/SLOF knowledge can weight in here - there is a good
that adding a new CPU element at this point in the DT isn't supported at

Now, in other news ....

3) Work around in libvirt by having it issue a system_reset after the
... talking with Mike about this bug he proposed a similar solution to (3),
but with
a CAS induced reset. If there are hotplugged devs at that point, set
to true. Doing that, QEMU reboots automatically and the hotplugged devs are
coldplugged. It works.

I am mentioning this because this is a simpler solution than relying on
libvirt issuing a system_reset. This is also solution that I would like to
avoid though - I
am not looking forward to see people bugs complaining about the machine
rebooting in
the middle of the boot. I will exhaust my attempts at (2) before proposing
with this CAS reboot solution.
Ah!  Actually I think that's a great idea.  I mean, it's an ugly hack,
but it should also get us something that more or less works with
*much* less work than other approaches.  So I think it's a good
interim solution.
Alright then!

Do you have a patch?  It's probably too late for qemu 2.10, but if we
can get it in as soon as 2.11 opens, then we should be able to get it
into the RH releases at least, which will be very helpful.

I'll test the code with PCI hotplug (have tested only with Logical DRCs so far),
make changes if required and send it.




In the short to medium term at least, I think (2) is probably our
best option.

I tried to hack this code by adding another break and ensure that the event
got logged
as HOT_PLUG (like it happens in the post-CAS) but then I got a kernel panic
at boot. So
I am not sure if this code needs any change or afterthought.

- hotplugging the CPU at early stage gives me a warning message in SLOF:


Calling ibm,client-architecture-support...Node not supported
Node not supported
   not implemented
memory layout at init:

The code that gives the 'Node not supported' message is related to the
function of board-qemu/slof/fdt.fs . The code is looking for either
"memory@" or
"ibm,dynamic-reconfiguration-memory" nodes, giving this error when finding a
CPU node.

- if I hotplug another CPU after the guest completes the boot, the
previously added
CPU suddenly turns online too:
Right the interrupt from the new event is unjamming the queue, so the
guest kernel / drmgr is then able to add the earlier cpu as well.

[ started VM with -S ]

(qemu) device_add host-spapr-cpu-core,id=core1,core-id=1
(qemu) cont

[ guest finishes boot ]

address@hidden:~$ lscpu
Architecture:        ppc64le
Byte Order:          Little Endian
CPU(s):              1
On-line CPU(s) list: 0
Thread(s) per core:  1
Core(s) per socket:  1
Socket(s):           1
NUMA node(s):        1
Model:               2.1 (pvr 004b 0201)
Model name:          POWER8E (raw), altivec supported
Hypervisor vendor:   KVM
Virtualization type: para
L1d cache:           64K
L1i cache:           32K
NUMA node0 CPU(s):   0
address@hidden:~$ (qemu)
(qemu) info cpus
* CPU #0: nip=0xc0000000000a464c thread_id=131946
    CPU #1: nip=0x0000000000000000 (halted) thread_id=131954
(qemu) device_add host-spapr-cpu-core,id=core2,core-id=2
(qemu) info cpus
* CPU #0: nip=0xc0000000000a464c thread_id=131946
    CPU #1: nip=0xc0000000000a464c thread_id=131954
    CPU #2: nip=0xc0000000000a464c thread_id=132144

address@hidden:~$ lscpu
Architecture:        ppc64le
Byte Order:          Little Endian
CPU(s):              3
On-line CPU(s) list: 0-2
Thread(s) per core:  1
Core(s) per socket:  3
Socket(s):           1
NUMA node(s):        1
Model:               2.1 (pvr 004b 0201)
Model name:          POWER8E (raw), altivec supported
Hypervisor vendor:   KVM
Virtualization type: para
L1d cache:           64K
L1i cache:           32K
NUMA node0 CPU(s):   0-2

This makes me believe that the issue is that the guest isn't aware of the
presence, making me wonder whether this has something to do with the
in the end of spapr_hotplug_req_event being lost. In the second hotplug, we
re-assert the IRQ in the end of check-exception and the guest is made aware
the queued hotplug event that was ignored at first.
Yeah, I think that's it.

3) There's no fundamental reason early hotplug shouldn't work - the
event will just be queued until the OS boots and processes it.

I know I suggested disabling early hotplug earlier, but that was
before I'd dug into the DRC layer and properly understood what was
going on here.

reply via email to

[Prev in Thread] Current Thread [Next in Thread]