Proof-of-concept for VSS-like quiesce agent in Linux

[lavamind]

Dear XCP-ng community,

In my unending quest for bulletproof backups, I've stumbled upon a technique which might be of interest to XCP-ng developpers and users. In short, it's a method for freezing filesystem I/O inside a domU during a snapshot event. It could potentially be extended into a hook provider, allowing not just blocking of filesystem I/O but coordinating the execution of any tasks moments before a snapshot is created, like activating database locks, unmounting filesystems, etc.

Requirements

By leveraging the existing vm-snapshot-with-quiesce XAPI call in XenServer/XCP-ng, no modifications whatsoever are required in dom0. In domU, the python-based agent requires the psutil and subprocess modules, as well as a slightly-modified version of pyxs. The upstream version needs a small fix to work around a bug in XenBus. The agent uses fsfreeze, a standard tool included in util-linux, to freeze the guest filesystems. Supported filesystems are ext3, ext4, xfs, jfs and reiserfs.

Disclaimer

The code below is just a rough proof-of-concept. It's far from being well tested and is missing lots of stuff like better error handling. I'm sharing it as an experiment for the benefit of the community. If you decide to test this on your own systems, make sure you have good backups. I'm not responsible for any damage to your data and/or systems!

Code

import psutil
import subprocess
from pyxs import Client

with Client(xen_bus_path="/dev/xen/xenbus") as c:

    domid = c.read(b"domid")
    vsspath = c.read(b"vss")
    dompath = c.get_domain_path(domid)

    print("Got domain path:", dompath.decode('ascii'))
    print("Got VSS path:", vsspath.decode('ascii'))

    print("Enabling quiesce feature")
    c.write(dompath + b"/control/feature-quiesce", b"1")

    print("Establishing watches")
    m = c.monitor()
    m.watch(dompath + b"/control/snapshot/action", b"")
    m.watch(vsspath + b"/status", b"")

    print("Waiting for snapshot signal...")

    for wpath, _token in m.wait():
        if wpath == dompath + b"/control/snapshot/action" and c.exists(dompath + b"/control/snapshot/action"):
            action = c.read(dompath + b"/control/snapshot/action")

            if action == b"create-snapshot":
                print("Received snapshot-create event")

                # Acknowledge VSS request
                c.delete(dompath + b"/control/snapshot/action")
                c.write(vsspath + b"/status", b"provider-initialized")

                # Construct list of VDIs to snapshot
                devlist = []
                vdilist = []
                vbdlist = c.list(dompath + b"/device/vbd")
                for vbd in vbdlist:
                    state = c.read(dompath + b"/device/vbd/" + vbd + b"/state")
                    devtype = c.read(dompath + b"/device/vbd/" + vbd + b"/device-type")
                    if state == b"4" and devtype == b"disk":
                        backend = c.read(dompath + b"/device/vbd/" + vbd + b"/backend")
                        vdiuuid = c.read(backend + b"/sm-data/vdi-uuid")
                        devlist.append(c.read(backend + b"/dev").decode('ascii'))
                        vdilist.append(vdiuuid)
                    else:
                        continue

                # Populate VDI snapshot list
                for vdi in vdilist:
                    c.mkdir(vsspath + b"/snapshot/" + vdi)

                # Freeze filesystems
                print("Begin freezing filesystems")
                fslist = []
                for p in psutil.disk_partitions():
                    if p.fstype not in ['ext3', 'ext4', 'xfs', 'jfs', 'reiserfs']:
                        continue
                    for d in devlist:
                        if p.device.startswith("/dev/" + d):
                            r = subprocess.run(["/sbin/fsfreeze", "-f", p.mountpoint])
                            if r.returncode == 0:
                                fslist.append(p.mountpoint)
                                print("Successfully froze " + p.mountpoint)
                            else:
                                print("Error, unable to freeze " + p.mountpoint)

                # Instruct snapwatchd to create VM snapshot
                print("Sending create-snapshots signal...")
                c.write(vsspath + b"/status", b"create-snapshots")

        elif wpath == vsspath + b"/status" and c.exists(vsspath + b"/status"):
            status = c.read(vsspath + b"/status")

            if status == b"snapshots-created":
                print("Received snapshot-created event")

                # Unfreeze filesystems
                for f in fslist:
                    #r = subprocess.run(["/sbin/fsfreeze", "-u", f])
                    r = subprocess.run(["/bin/true"])
                    if r.returncode == 0:
                        print("Successfully unfroze", f)
                    else:
                        print("Error, unable to unfreeze", p.mountpoint) # this should not happen ...

                c.write(vsspath + b"/status", b"create-snapshotinfo")

            elif status == b"snapshotinfo-created":
                print("Received snapshotinfo-created event")

                # Create fake VSS transport id (Windows-only...)
                c.mkdir(dompath + b"/control/snapshot/snapid")
                c.write(dompath + b"/control/snapshot/snapid/0", b"0")
                c.write(dompath + b"/control/snapshot/snapid/1", b"1")

                # Record snapshot uuid
                snapuuid = c.read(vsspath + b"/snapuuid")
                print("Snapshot created:", snapuuid)
                c.write(dompath + b"/control/snapshot/snapuuid", snapuuid)

                # Signal snapshot creation
                print("Sending snapshot-created signal...")
                c.write(dompath + b"/control/snapshot/status", b"snapshot-created")

                # Cleanup vsspath
                print("Cleaning up")
                c.delete(vsspath + b"/snapshot")
                c.delete(vsspath + b"/snaptype")
                c.delete(vsspath + b"/snapinfo")
                c.delete(vsspath + b"/snapuuid")
                c.delete(vsspath + b"/status")

            elif status == b"snapshots-failed":
                print("Received snapshots-failed event")

                # Unfreeze filesystems
                for f in fslist:
                    r = subprocess.run(["/sbin/fsfreeze", "-u", f])
                    if r.returncode == 0:
                        print("Successfully unfroze", f)
                    else:
                        print("Error, unable to unfreeze", p.mountpoint) # this should not happen ...

                # Signal snapshot error
                print("Sending snapshot-error event...")
                c.write(dompath + b"/control/snapshot/status", b"snapshot-error")

Resources

The key resources that allowed me to develop this was the xenstore command, available in both domU and dom0, in particular xenstore ls, which is unfortunately only available in dom0. Logs in /var/log/SMlog and /var/log/xensource.log were also invaluable, as was the source code for both XAPI and the snapwatchd storage-manager (sm) component.

Having a Windows Server guest VM on hand was also useful to understand the xenstore messaging task sequence of a quiesce snapshot using the Citrix Xen VSS provider.

Development

I'm curious to see if this can actually be useful to the XCP-ng community.

If you're curious and familiar with Xen, please do test it and report back!

[olivierlambert]

@stormi should we open a repo in XCP-ng org on GH to host it or first let OP do it on GH and fork it?

[lavamind]

@olivierlambert @stormi I don't mind hosting in under XCP-ng as long as I get commit access, at least in a dev branch

[stormi]

How would we name this repository?

[borzel]

@stormi hm, maybe:

• safe snap
• snap freeze
• consistent snap
• snapshot-with-quiesce
• linux-snapshot-with-quiesce

[maxcuttins]

smell like a good feature

[lavamind]

Maybe we could just fork xe-guest-utilities and figure out how to add it in there? Or, if we want a distinct package, how about a more generic name like xcp-guest-agent, the rationale being that the agent's purpose can be broader than just snapshot/quiesce. Watches and communication with xenstore opens up lots of possibilities in domU.

[stormi]

I agree. Ideally we should also contribute such changes to the xe-guest-agent upstream repository.