[ovirt-devel] Re: Improving VM behavior in case of IO errors

Am 11.08.2020 um 17:44 hat Nir Soffer geschrieben: > On Mon, Aug 10, 2020 at 11:53 AM Kevin Wolf <kwolf(a)redhat.com&gt; wrote: > > > Am 09.08.2020 um 23:50 hat Nir Soffer geschrieben: > > > On Wed, Jul 29, 2020 at 2:30 PM Shubha Kulkarni > > > <shubha.kulkarni(a)oracle.com&gt; wrote: > > > > > > > > Thanks for the feedback Nir. > > > > > > > > I agree in general that having an additional engine config for disk > > > > level error handling default would be the right way. It would be good > > to > > > > decide the granularity. Would it make sense to have this for a specific > > > > disk type like lun or would you prefer to make it generic for all > > types? > > > > > > This must be for a specific disk type, since for thin images on block > > > storage we cannot support propagating errors to the guest. This will > > > break thin provisioning. > > > > Is werror=enospc not enough for thin provisioning to work? This will > > still stop the guest for any other kinds of I/O errors. > > > > Right, this should work, and what we actually use now for propagating > errors for anything but cdrom. Hm, wait, the options you quote below are all either 'stop' or 'report', but never 'enospc'. Is 'enospc' used for yet another kind of disk?

For LUN using werror=enospc,rerror=enospc seems wrong, but we do this for > many years. > > This is how we handle cdrom: > > -device > ide-cd,bus=ide.2,id=ua-346e176c-f983-4510-af4b-786b368efdd6,bootindex=2,werror=report,rerror=report Makes sense to me. This is read-only and removable media. Stopping the guest usually makes sense so that it won't assume the disk is broken, but if it happens with removable media, you can just eject and re-insert the same image and it's fixed.

Image: > > -device > virtio-blk-pci,iothread=iothread1,scsi=off,bus=pci.5,addr=0x0,drive=libvirt-2-format,id=ua-1d93fa9e-1665-40d7-9ffc-770513242795,bootindex=1,write-cache=on,serial=1d93fa9e-1665-40d7-9ffc-770513242795,werror=stop,rerror=stop I assume this is the one that could use 'enospc'?

> > LUN: > > -device > virtio-blk-pci,iothread=iothread2,scsi=off,bus=pci.7,addr=0x0,drive=libvirt-1-format,id=ua-19b06845-2c54-422d-921b-6ec0ee2e935b,write-cache=on,werror=stop,rerror=stop > \ > > Kevin, any reason not to use werror=report,rerror=report for LUN when > we want to propagate errors to the guest? If you want to propagate errors, then 'report' is the right setting. What does "LUN" mean exactly?

it just that you have some restriction like that it's always raw?

Maybe I would use 'enospc' for consistency even though you never expect this error to happen. But 'report' is fine, too.

propagating errors to the guest means that it will give up on the disk. Whether this is the desired behaviour should probably be configured by the user. Kevin > > > Kevin > > > > > Handling the LUN use case first seems like the best way, since in this > > case we > > > don't manage the LUN and we don't support resuming paused using LUNs yet, > > > so propagating the error may be more useful. > > > > > > Managed Block Storage (cinderlib based disks) are very much like > > > direct LUN. In this > > > case we do manage the disks on the server, but otherwise we don't > > > support anything > > > on the host (e.g. monitoring, resuming paused VMs) so propagating the > > error like > > > direct LUNs may be more useful. > > > > > > Images are a bigger problem since thin disks cannot support > > > propagating errors but > > > preallocated disks can. But once you create a snapshot prealocated disks > > behave > > > exactly like thin disks because they are the same. > > > > > > Snapshots are also created automatically in for preallocated images, > > > for example during > > > live storage migration, and deleted automatically after the migration. > > > So you cannot > > > assume that having only preallocated disks is good for propagating > > errors. > > > > > > Even if you limit this option to file based storage, this is going to > > > break when you migrate > > > the disks to block storage. > > > > > > Nir > > > > > > > > > > > Thanks, > > > > > > > > Shubha > > > > > > > > On 7/28/2020 2:03 PM, Nir Soffer wrote: > > > > > On Tue, Jul 28, 2020 at 4:58 AM Shubha Kulkarni > > > > > <shubha.kulkarni(a)oracle.com&gt; wrote: > > > > >> Hello, > > > > >> > > > > >> In OVirt, we have a property propagate_error at the disk level that > > > > >> decides in case of an error, how this error be propagated to the VM. > > > > >> This value is maintained in the database table with the default > > value > > > > >> set as Off. The default setting(Off) results in a policy that ends > > up > > > > >> pausing the VM rather than propagating the errors to VM. There is > > no > > > > >> provision in the UI currently to configure this property for disk > > > > >> (images or luns). So there is no easy way to set this value. > > Further, > > > > >> even if the value is manually set to "On" in db, it gets > > overwriiten by > > > > >> UI everytime some other property is updated as described here - > > > > >> https://bugzilla.redhat.com/show_bug.cgi?id=1669367 > > > > >> > > > > >> Setting the value to "Off" is not ideal for multipath devices where > > a > > > > >> single path failure causes vm to pause. > > > > > Single path failure should be transparent to qemu. multipath will > > fail over > > > > > the I/O to another path. The I/O will fail only if all paths are > > down, and > > > > > (with the default configuration), multipath path checkers failed 4 > > times. > > > > > > > > > >> It puts serious restrictions for > > > > >> the DR situation and unlike VMWare * Hyper-V, oVirt is not able to > > > > >> support the DR functionality - > > > > >> https://bugzilla.redhat.com/show_bug.cgi?id=1314160 > > > > > Alghouth in this bug we see that failover that looks successful from > > multipath > > > > > and vdsm point of view ended in paused VM: > > > > > https://bugzilla.redhat.com/1860377 > > > > > > > > > > Maybe Ben can explain how this can happen. > > > > > > > > > > I hope that qemu will provide more info on errors in the future. If > > we had a log > > > > > about the failure I/O it could be helpful. > > > > > > > > > >> While we wait for RFE, the proposal here is to revise the out of > > the box > > > > >> behavior for LUNs. For LUNs, we should propagate the errors to VM > > rather > > > > >> than directly stopping those. This will allow us to handle > > short-term > > > > >> multipath outages and improve availability. This is a simple change > > in > > > > >> behavior but will have good positive impact. I would like to seek > > > > >> feedback about this to make sure that everyone is ok with the > > proposal. > > > > > I think it makes sense, but this is just a default, and it cannot > > work > > > > > for all cases. > > > > > > > > > > This can end in broken VM with read only file system that must be > > > > > rebooted, while > > > > > with error_policy="stop", failover may be transparent to the VM even > > > > > if it was paused > > > > > for a short time. > > > > > > > > > > I would start by making engine defaults configurable using engine > > > > > config, so different > > > > > oVirt distributions can use different defaults. > > > > > > > > > > Nir > > > > > > > > > > > > > > > >