Re: [Qemu-devel] QEMU terminates during reboot after memory unplug with vhost=on

[Michael S. Tsirkin]

On Thu, Nov 16, 2017 at 01:23:39PM +0100, Greg Kurz wrote:
> Hi,
> 
> I'm resurrecting a thread about a QEMU crash we're still hitting on ppc64. It
> was reported to the list by Bharata 2 months ago:
> 
> https://lists.gnu.org/archive/html/qemu-devel/2017-09/msg03685.html
> 
> "Hi,
> 
> QEMU hits the below assert
> 
> qemu-system-ppc64: used ring relocated for ring 2
> qemu-system-ppc64: qemu/hw/virtio/vhost.c:649: vhost_commit: Assertion `r >= 
> 0' failed.
> 
> in the following scenario:
> 
> 1. Boot guest with vhost=on
>   -netdev tap,id=mynet0,script=qemu-ifup,downscript=qemu-ifdown,vhost=on 
> -device virtio-net-pci,netdev=mynet0
> 2. Hot add a DIMM device 
> 3. Reboot
>    When the guest reboots, we can see
>    vhost_virtqueue_start:vq->used_phys getting assigned an address that
>    falls in the hotplugged memory range.
> 4. Remove the DIMM device
>    Guest refuses the removal as the hotplugged memory is under use.
> 5. Reboot
>    QEMU forces the removal of the DIMM device during reset and that's
>    when we hit the above assert.
> 
> Any pointers on why we are hitting this assert ? Shouldn't vhost be
> done with using the hotplugged memory when we hit reset ?
> 
> Regards,
> Bharata."
> 
> #0  0x00007ffff760eff0 in raise () from /lib64/libc.so.6
> #1  0x00007ffff761136c in abort () from /lib64/libc.so.6
> #2  0x00007ffff7604c44 in __assert_fail_base () from /lib64/libc.so.6
> #3  0x00007ffff7604d34 in __assert_fail () from /lib64/libc.so.6
> #4  0x0000000010161138 in vhost_commit (listener=0x11469e88) at 
> /home/greg/Work/qemu/qemu-spapr/hw/virtio/vhost.c:650
> #5  0x00000000100917fc in memory_region_transaction_commit () at 
> /home/greg/Work/qemu/qemu-spapr/memory.c:1094
> #6  0x0000000010096748 in memory_region_del_subregion (mr=0x1143eed0, 
> subregion=0x116f1920) at /home/greg/Work/qemu/qemu-spapr/memory.c:2337
> #7  0x00000000104a9aec in pc_dimm_memory_unplug (dev=0x11445c50, 
> hpms=0x1143eec0, mr=0x116f1920) at 
> /home/greg/Work/qemu/qemu-spapr/hw/mem/pc-dimm.c:126
> #8  0x0000000010180454 in spapr_lmb_release (dev=0x11445c50) at 
> /home/greg/Work/qemu/qemu-spapr/hw/ppc/spapr.c:3151
> #9  0x00000000101a397c in spapr_drc_release (drc=0x116b9cc0) at 
> /home/greg/Work/qemu/qemu-spapr/hw/ppc/spapr_drc.c:401
> #10 0x00000000101a3ba0 in spapr_drc_reset (drc=0x116b9cc0) at 
> /home/greg/Work/qemu/qemu-spapr/hw/ppc/spapr_drc.c:439
> #11 0x00000000101a3c88 in drc_reset (opaque=0x116b9cc0) at 
> /home/greg/Work/qemu/qemu-spapr/hw/ppc/spapr_drc.c:460
> #12 0x0000000010447380 in qemu_devices_reset () at 
> /home/greg/Work/qemu/qemu-spapr/hw/core/reset.c:69
> #13 0x000000001017ae80 in ppc_spapr_reset () at 
> /home/greg/Work/qemu/qemu-spapr/hw/ppc/spapr.c:1445
> #14 0x0000000010377c60 in qemu_system_reset (reason=SHUTDOWN_CAUSE_HOST_QMP) 
> at /home/greg/Work/qemu/qemu-spapr/vl.c:1788
> #15 0x00000000103785ac in main_loop_should_exit () at 
> /home/greg/Work/qemu/qemu-spapr/vl.c:1962
> #16 0x0000000010378708 in main_loop () at 
> /home/greg/Work/qemu/qemu-spapr/vl.c:1999
> #17 0x0000000010382c54 in main (argc=21, argv=0x7ffffffff098, 
> envp=0x7ffffffff148) at /home/greg/Work/qemu/qemu-spapr/vl.c:4897
> 
> 
> This basically happens because on pseries, like x86, we usually wait
> for the guest to eject the DIMM before actually removing it, BUT,
> unlike x86, we force the removal on reset. This is handled by a DRC
> object which registers a handler with qemu_register_reset().
> 
> At reset time, the machine calls qemu_devices_reset() but unfortunately,
> the DRC reset handler gets called BEFORE the VirtIONet device one. The
> vhost device is still active and it doesn't like the ring addresses to
> change while in this state.
> 
> Michael,
> 
> The assert() has been around since the beginning, at a time I believe there 
> was
> no such thing as memory hot-unplug. Now that memory can go away at reset time,
> is it really legitimate to crash QEMU if vhost detects a ring address change ?

It's just a symptom of a problem though. If memory is going away
while vhost backend is running, things are not going to
end well. Less scary for a network device, more scary for a block
device. VFIO probably has the same issue, it just does not
have an assert.

> David,
> 
> Depending on Michael's answer (or anyone else that knows vhost well enough),
> I'm not sure we can fix that properly for 2.11. A possible fix/workaround
> could be to reset DRCs (at least the LMB ones) after qemu_devices_reset(),
> so that we're sure vhost is stopped.
> 
> Thoughts ?

I think the release of DIMM on reset trick must happen after all
devices have been reset.
> 
> --
> Greg