canary-on-1.16 block tests flaky about csi-driver-host-path HOT 8 CLOSED

Its flaky across all jobs, not just this.

I think it's related to running on kind, where hostpath volumes go to the host. There may be races due another test grabbing the same device

from csi-driver-host-path.

Although if that were the case, then intree parallel tests should be flaky too? Maybe it's a probability game since there a lot more tests running intree

from csi-driver-host-path.

There may be races due another test grabbing the same device

Creating a loop device is done by kubelet and the CSI hostpath driver with https://github.com/kubernetes/kubernetes/blob/8a09460c2f7ba8f6acd8a6fb7603ed3ac4805eb6/pkg/volume/util/volumepathhandler/volume_path_handler_linux.go#L96-L112.

The command invocation itself (= "losetup -f --show") shouldn't be racy. Let's look at its implementation:

openat(AT_FDCWD, "/dev/loop-control", O_RDWR|O_CLOEXEC) = 3
ioctl(3, LOOP_CTL_GET_FREE)             = 0
close(3)                                = 0
lstat("/tmp", {st_mode=S_IFDIR|S_ISVTX|0777, st_size=260, ...}) = 0
lstat("/tmp/test", {st_mode=S_IFREG|0644, st_size=1073741824, ...}) = 0
openat(AT_FDCWD, "/tmp/test", O_RDWR|O_CLOEXEC) = 3
openat(AT_FDCWD, "/dev/loop0", O_RDWR|O_CLOEXEC) = 4
ioctl(4, LOOP_SET_FD, 3)                = 0
ioctl(4, LOOP_SET_STATUS64, {lo_offset=0, lo_number=0, lo_flags=0, lo_file_name="/tmp/test", ...}) = 0

losetup first determines the next available loop device, then opens it and configures it. At first glance this looks like it might be racy, but that depends on how LOOP_SET_FD and losetup behave when someone else grabs the device in parallel.

I tested that by running losetup under gdb and pausing it in the LOOP_SET_FD ioctl, then taking the loop device with some other losetup call. The first invocation recovered gracefully from that by retrying with another loop device.

So my conclusion is that we don't have a race around loop device creation.

Perhaps KinD containers have the same issue that we also had in the CSI hostpath container where /dev is a static copy of the host and then new loop devices don't show up?

https://github.com/kubernetes/kubernetes/blob/8a09460c2f7ba8f6acd8a6fb7603ed3ac4805eb6/pkg/volume/util/volumepathhandler/volume_path_handler_linux.go#L96-L112

Indeed, when I just ran the CSI hostpath prow tests on my development machine after using all existing loop devices, the blockvolume tests fail because losetup fails:

Jan 13 08:31:51.654: INFO: At 2020-01-13 08:27:28 +0100 CET - event for security-context-6ade56c3-0fd9-4080-a3b5-e46a57584a4c: {kubelet csi-prow-worker} FailedMapVolume: MapVolume.AttachFileDevice failed for volume "pvc-228ac164-bec1-43fe-a3d1-7671ef0549f6" : exit status 1

I've not found more about it in the logs, but I can reproduce the problem manually by creating loop devices with docker exec f049e80f378a sh -c 'truncate -s 1G /tmp/testfile; losetup -f --show /tmp/testfile' until eventually all the ones that existed when the KinD container was started are in use and the command fails with:

losetup: /tmp/testfile: failed to set up loop device: No such file or directory

I'll file an issue against KinD.

from csi-driver-host-path.

I'll file an issue against KinD.

=> kubernetes-sigs/kind#1248

from csi-driver-host-path.

Issues go stale after 90d of inactivity.
Mark the issue as fresh with /remove-lifecycle stale.
Stale issues rot after an additional 30d of inactivity and eventually close.

If this issue is safe to close now please do so with /close.

Send feedback to sig-testing, kubernetes/test-infra and/or fejta.
/lifecycle stale

from csi-driver-host-path.

Stale issues rot after 30d of inactivity.
Mark the issue as fresh with /remove-lifecycle rotten.
Rotten issues close after an additional 30d of inactivity.

If this issue is safe to close now please do so with /close.

Send feedback to sig-testing, kubernetes/test-infra and/or fejta.
/lifecycle rotten

from csi-driver-host-path.

/close

from csi-driver-host-path.

@msau42: Closing this issue.

from csi-driver-host-path.