Automatic merge from submit-queue
Enable lazy initialization of ext3/ext4 filesystems
**What this PR does / why we need it**: It enables lazy inode table and journal initialization in ext3 and ext4.
**Which issue this PR fixes** *(optional, in `fixes #<issue number>(, fixes #<issue_number>, ...)` format, will close that issue when PR gets merged)*: fixes#30752, fixes#30240
**Release note**:
```release-note
Enable lazy inode table and journal initialization for ext3 and ext4
```
**Special notes for your reviewer**:
This PR removes the extended options to mkfs.ext3/mkfs.ext4, so that the defaults (enabled) for lazy initialization are used.
These extended options come from a script that was historically located at */usr/share/google/safe_format_and_mount* and later ported to GO so this dependency to the script could be removed. After some search, I found the original script here: https://github.com/GoogleCloudPlatform/compute-image-packages/blob/legacy/google-startup-scripts/usr/share/google/safe_format_and_mount
Checking the history of this script, I found the commit [Disable lazy init of inode table and journal.](4d7346f7f5). This one introduces the extended flags with this description:
```
Now that discard with guaranteed zeroing is supported by PD,
initializing them is really fast and prevents perf from being affected
when the filesystem is first mounted.
```
The problem is, that this is not true for all cloud providers and all disk types, e.g. Azure and AWS. I only tested with magnetic disks on Azure and AWS, so maybe it's different for SSDs on these cloud providers. The result is that this performance optimization dramatically increases the time needed to format a disk in such cases.
When mkfs.ext4 is told to not lazily initialize the inode tables and the check for guaranteed zeroing on discard fails, it falls back to a very naive implementation that simply loops and writes zeroed buffers to the disk. Performance on this highly depends on free memory and also uses up all this free memory for write caching, reducing performance of everything else in the system.
As of https://github.com/kubernetes/kubernetes/issues/30752, there is also something inside kubelet that somehow degrades performance of all this. It's however not exactly known what it is but I'd assume it has something to do with cgroups throttling IO or memory.
I checked the kernel code for lazy inode table initialization. The nice thing is, that the kernel also does the guaranteed zeroing on discard check. If it is guaranteed, the kernel uses discard for the lazy initialization, which should finish in a just few seconds. If it is not guaranteed, it falls back to using *bio*s, which does not require the use of the write cache. The result is, that free memory is not required and not touched, thus performance is maxed and the system does not suffer.
As the original reason for disabling lazy init was a performance optimization and the kernel already does this optimization by default (and in a much better way), I'd suggest to completely remove these flags and rely on the kernel to do it in the best way.
This is a fix on top #38124. In this fix, we move the logic to filter
out shared mount references into operation_executor's UnmountDevice
function to avoid this part is being used by other types volumes such as
rdb, azure etc. This filter function should be only needed during
unmount device for GCI image.
this is a workaround for the unmount device issue caused by gci mounter. In GCI cluster, if gci mounter is used for mounting, the container started by mounter script will cause additional mounts created in the container. Since these mounts are irrelavant to the original mounts, they should be not considered when checking the mount references. By comparing the mount path prefix, those additional mounts can be filtered out.
Plan to work on better approach to solve this issue.
This change is to only enable containerized mounter for nfs and
glusterfs types. For other types such as tmpfs, ext2/3/4 or empty type,
we should still use mount from $PATH
This PR is to fix the issue in converting aws volume id from mount
paths. Currently there are three aws volume id formats supported. The
following lists example of those three formats and their corresponding
global mount paths:
1. aws:///vol-123456
(/var/lib/kubelet/plugins/kubernetes.io/aws-ebs/mounts/aws/vol-123456)
2. aws://us-east-1/vol-123456
(/var/lib/kubelet/plugins/kubernetes.io/mounts/aws/us-est-1/vol-123455)
3. vol-123456
(/var/lib/kubelet/plugins/kubernetes.io/mounts/aws/us-est-1/vol-123455)
For the first two cases, we need to check the mount path and convert
them back to the original format.
When kubelet restarts, all the information about the volumes will be
gone from actual/desired states. When update node status with mounted
volumes, the volume list might be empty although there are still volumes
are mounted and in turn causing master to detach those volumes since
they are not in the mounted volumes list. This fix is to make sure only
update mounted volumes list after reconciler starts sync states process.
This sync state process will scan the existing volume directories and
reconstruct actual states if they are missing.
This PR also fixes the problem during orphaned pods' directories. In
case of the pod directory is unmounted but has not yet deleted (e.g.,
interrupted with kubelet restarts), clean up routine will delete the
directory so that the pod directoriy could be cleaned up (it is safe to
delete directory since it is no longer mounted)
The third issue this PR fixes is that during reconstruct volume in
actual state, mounter could not be nil since it is required for creating
container.VolumeMap. If it is nil, it might cause nil pointer exception
in kubelet.
Details are in proposal PR #33203
In order to be able to use new mounter library, this PR adds the
mounterPath flag to kubelet which passes the flag to the mount
interface. If flag is empty, mount uses default mount path.
Currently kubelet volume management works on the concept of desired
and actual world of states. The volume manager periodically compares the
two worlds and perform volume mount/unmount and/or attach/detach
operations. When kubelet restarts, the cache of those two worlds are
gone. Although desired world can be recovered through apiserver, actual
world can not be recovered which may cause some volumes cannot be cleaned
up if their information is deleted by apiserver. This change adds the
reconstruction of the actual world by reading the pod directories from
disk. The reconstructed volume information is added to both desired
world and actual world if it cannot be found in either world. The rest
logic would be as same as before, desired world populator may clean up
the volume entry if it is no longer in apiserver, and then volume
manager should invoke unmount to clean it up.
In nsenter_mount.go/isLikelyNotMountPoint function, the returned output
from findmnt command misses the last letter. Modify the code to make sure
that output has the full target path. fix#26421#25056#22911
fix https://github.com/kubernetes/kubernetes/issues/24717
If kubelet root-dir is a symlink, the pod's cinder volume dir can't be
umounted even after pod is deleted.
This patch reads target path of symlink before comparing with entries in
proc mounts.
Kubelet was not able to mount volumes when running inside a container and
using nsenter mounter,
NsenterMounter.IsLikelyNotMountPoint() should return ErrNotExist when the
checked directory does not exists as the regular mounted does this and
some volume plugins depend on this behavior.
Add a mutex to guard SetUpAt() and TearDownAt() calls - they should not
run in parallel. There is a race in these calls when there are two pods
using the same volume, one of them is dying and the other one starting.
TearDownAt() checks that a volume is not needed by any pods and detaches the
volume. It does so by counting how many times is the volume mounted
(GetMountRefs() call below).
When SetUpAt() of the starting pod already attached the volume and did not mount
it yet, TearDownAt() of the dying pod will detach it - GetMountRefs() does not
count with this volume.
These two threads run in parallel:
dying pod.TearDownAt("myVolume") starting pod.SetUpAt("myVolume")
| |
| AttachDisk("myVolume")
refs, err := mount.GetMountRefs() |
Unmount("myDir") |
if refs == 1 { |
| | Mount("myVolume", "myDir")
| | |
| DetachDisk("myVolume") |
| start containers - OOPS! The volume is detached!
|
finish the pod cleanup
Also, add some logs to cinder plugin for easier debugging in the future, add
a test and update the fake mounter to know about bind mounts.
Kubernetes Submit Queue