[Bug] Expanding memory balloon causes VM to freeze

[maggie-lou]

Describe the bug

Even when there should be enough free memory in the VM, expanding the balloon sometimes causes the VM to freeze.

During a sample run (using the scripts linked below), after restoring the VM from a snapshot, free -h returned:

               total        used        free      shared  buff/cache   available
Mem:           108Mi        35Mi        45Mi       2.5Mi        35Mi        72Mi
Swap:             0B          0B          0B

Originally, the balloon was initialized to 5MB. When I inflated it to 20MB, it inflated successfully. When I inflated it to 30MB, the VM froze and there were a bunch of "Failed to update balloon stats, missing descriptor." errors.

To Reproduce

You can use the scripts in this branch: #4989

1. Build firecracker with this patch: Fix unregistering memory ranges from UFFD when expanding the balloon #4988
2. (Only needs to be run once): Prepare rootfs and guest kernel: get_rootfs_guest_kernel.sh
3. Run firecracker: run_firecracker.sh
4. Initialize a VM with a balloon and snapshot it: snapshot_vm.sh
5. You will probably need to kill the former firecracker process and restart it: run_firecracker.sh
6. Start the UFFD handler with the snapshot: run_uffd_handler.sh
7. Expand the balloon. : trigger_remove_events.sh
8. Expand the balloon even more : If you edit trigger_remove_events.sh. to inflate the balloon to 40MB, the VM will freeze and there are "Failed to update balloon stats, missing descriptor." errors

Expected behavior

I expected the balloon to be able to expand to 30MB because there is 72Mi of memory available.

Environment

• Firecracker version: 1.10.0 with patch Fix unregistering memory ranges from UFFD when expanding the balloon #4988
• Host and guest kernel versions: 6.1 for both
• Rootfs used: ubuntu-24.04
• Architecture: Intel x86

Additional context

We are using UFFD to restore snapshots. The memory snapshots are quite large, so we're looking into using memory balloons with the goal of having the UFFD handler process removed memory ranges, so we don't have to save those memory ranges in the snapshot files. We've noticed that the VM will sometimes freeze when expanding the balloon, even when there should be sufficient memory.

Around the same time as the freeze, we always see the "Failed to update balloon stats, missing descriptor." errors as well as vsock connection errors VIRTIO_VSOCK_OP_RST.

I've tried disabling async page faults, in case the freezing was related to some sort of race condition in the kernel but the problem persists.

Checks

• Have you searched the Firecracker Issues database for similar problems?
• Have you read the existing relevant Firecracker documentation?
• Are you certain the bug being reported is a Firecracker issue?

No - It could be a Linux bug as well. Though we've read cases where people seem to be successfully using UFFD + the balloon, so this use case seems like it should be possible now.

[JackThomson2]

Hi @maggie-lou,

Thanks for raising the issue, with the helpful reproducer steps. I'll investigate myself and let you know what I find.

Thanks

[maggie-lou]

In case this is helpful, when testing the balloon in a more production environment, I've noticed Kernel RCU stalls. I wonder if the balloon has a bad interaction with the RCU.

[  234.560950] rcu: INFO: rcu_preempt self-detected stall on CPU
[  234.562786] rcu: 	1-....: (14819 ticks this GP) idle=bb24/1/0x4000000000000000 softirq=4838/4841 fqs=5790
[  234.565430] 	(t=14750 jiffies g=5389 q=4 ncpus=2)
[  234.566859] CPU: 1 PID: 20 Comm: kworker/1:0 Not tainted 6.2.0 #1
[  234.568654] Workqueue: virtio_vsock virtio_transport_rx_work
[  234.570932] RIP: 0010:vring_unmap_one_split+0x41/0x50

[roypat]

Hey! I've had a look into this today, and think what you're seeing is a combination of two issues: The first one is indeed that the Firecracker incorrectly handles balloon inflation events on restored VMs, and the fix from your PR for that is indeed the right one. However, the behavior you see after that (VM freezes) are not actually a bug in Firecracker, but rather a shortcoming of the simplistic UFFD handler used in our integration tests.

Essentially, the problem is with the handling of -EAGAIN when its returned by UFFDIO_COPY - it's not actually a race condition on one page (e.g. a fault and a free for the same page), but rather a race between any pair of fault and free: UFFD will have any ioctl return -EAGAIN if there's a remove message pending in the queue. So if we get EAGAIN, we cannot ignore the pagefault message its associated with, but rather we must first process all pending remove messages before we can proceed with the pagefault (and this is where the problem was: if we instead ignore the pagefault, then the guest thread that caused this page fault will never get awoken again). I've opened #5021 to fix all that up. Could you please give that a try to see if it resolves your issue?
Thanks!