Problem:
On remap/attach of device (i.e. nodeplugin restart), there is no way
for rbd-nbd to defend if the backend storage is matching with the initial
backend storage.
Say, if an initial map request for backend "pool1/image1" got mapped to
/dev/nbd0 and the userspace process is terminated (on nodeplugin restart).
A next remap/attach (nodeplugin start) request within reattach-timeout is
allowed to use /dev/nbd0 for a different backend "pool1/image2"
For example, an operation like below could be dangerous:
$ sudo rbd-nbd map --try-netlink rbd-pool/ext4-image
/dev/nbd0
$ sudo blkid /dev/nbd0
/dev/nbd0: UUID="bfc444b4-64b1-418f-8b36-6e0d170cfc04" TYPE="ext4"
$ sudo pkill -15 rbd-nbd <-- nodeplugin terminate
$ sudo rbd-nbd attach --try-netlink --device /dev/nbd0 rbd-pool/xfs-image
/dev/nbd0
$ sudo blkid /dev/nbd0
/dev/nbd0: UUID="d29bf343-6570-4069-a9ea-2fa156ced908" TYPE="xfs"
Solution:
rbd-nbd/kernel now provides a way to keep some metadata in sysfs to identify
between the device and the backend, so that when a remap/attach request is
made, rbd-nbd can compare and avoid such dangerous operations.
With the provided solution, as part of the initial map request, backend
cookie (ceph-csi VOLID) can be stored in the sysfs per device config, so
that on a remap/attach request rbd-nbd will check and validate if the
backend per device cookie matches with the initial map backend with the help
of cookie.
At Ceph-csi we use VOLID as device cookie, which will be unique, we pass
the VOLID as cookie at map and use the same at the time of attach, that
way rbd-nbd can identify backends and their matching devices.
Requires:
https://github.com/ceph/ceph/pull/41323https://lkml.org/lkml/2021/4/29/274
Signed-off-by: Prasanna Kumar Kalever <prasanna.kalever@redhat.com>
This change allows the user to choose not to fallback to NBD mounter
when some ImageFeatures are absent with krbd driver, rather just fail
the NodeStage call.
Signed-off-by: Prasanna Kumar Kalever <prasanna.kalever@redhat.com>
Currently, we recognize and warn for the provided image features based on
our prior intelligence at ceph-csi (i.e based on supportedFeatures map
and validateImageFeatures) at image/PV creation time. It might be very
much possible that the cluster is heterogeneous i.e. the PV creation and
application container might both be on different nodes with different
kernel versions (krbd driver versions).
This PR adds a mechanism to check for the supported krbd features during
mount time, if the krbd driver doesn't have the specified image feature
then it will fall back to rbd-nbd mounter.
Fixes: #478
Signed-off-by: Prasanna Kumar Kalever <prasanna.kalever@redhat.com>
this commit make use of the migration request secret parsing and set
the required fields for further nodestage operations
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
parseAndDeleteMigratedVolume() prviously clubbed the logic of
parsing of migration volume handle and then continued with the
deletion of the volume. however this commit split this
logic into two, ie parsing has been done in parseMigrationVolID()
and DeleteMigratedVolume() deletes the backend volume.
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
This commit adds a couple of helper functions to parse the migration
request secret and set it for further csi driver operations.
More details:
The intree secret has a data field called "key" which is the base64
admin secret key. The ceph CSI driver currently expect the secret to
contain data field "UserKey" for the equivalant. The CSI driver also
expect the "UserID" field which is not available in the in-tree secret
by deafult. This missing userID will be filled (if the username differ
than 'admin') in the migration secret as 'adminId' field in the
migration request, this commit adds the logic to parse this migration
secret as below:
"key" field value will be picked up from the migraion secret to "UserKey"
field.
"adminId" field value will be picked up from the migration secret to "UserID"
field
if `adminId` field is nil or not set, `UserID` field will be filled with
default value ie `admin`.The above logic get activated only when the secret
is a migration secret, otherwise skipped to the normal workflow as we have
today.
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
Thick-provisioning was introduced to make accounting of assigned space
for volumes easier. When thick-provisioned volumes are the only consumer
of the Ceph cluster, this works fine. However, it is unlikely that this
is the case. Instead, accounting of the requested (thin-provisioned)
size of volumes is much more practical as different types of volumes can
be tracked.
OpenShift already provides cluster-wide quotas, which can combine
accounting of requested volumes by grouping different StorageClasses.
In addition to the difficult practise of allowing only thick-provisioned
RBD backed volumes, the performance makes thick-provisioning
troublesome. As volumes need to be completely allocated, data needs to
be written to the volume. This can take a long time, depending on the
size of the volume. Provisioning, cloning and snapshotting becomes very
much noticeable, and because of the additional time consumption, more
prone to failures.
Signed-off-by: Niels de Vos <ndevos@redhat.com>
for comparing the image states use the states
defined in the go-ceph avoid creating of the
deplicate const in cephcsi.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
below are the local states of the mirrored image
"unknown" -> If the image is in an error state
means data is completely synced
"error" -> If the image is in an error state
means it needs resync
"syncing"
"starting_replay"
"replaying"
"stopping_replay"
"stopped"
If the resync is successfully started which
means the image will be in "replaying" state.
we can consider "replaying" state to report
resync succesfully going on state.
we are discarding the intermediate states like
"syncing", "starting_replay" and "stopping_replay".
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
After moving moving image to trash, if `trash remove` step fails,
then external-provisioner will issue subsequent requests, in which
image will be absent in pool( will be in trash) and omap cleanup will
be done with stale image left in trash with no `trash remove` step on it.
To avoid this scenario list trash images and find corresponding id for given
image name and add a task to flatten when we encounter a ErrImageNotFound.
Fixes: #1728
Signed-off-by: Rakshith R <rar@redhat.com>
In some corner case like `re-player shutdown` the
local image will not be in error state. It would
be also worth considering `description` field to
make sure about split-brain.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
previously we were retriving clusterID using the monitors field
in the volume context at node stage code path. however it is possible to
retrieve or use clusterID directly from the volume context. This
commit also remove the getClusterIDFromMigrationVolume() function
which was used previously and its tests
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
we reuse or overload the variable name in the test execution at present.
This commit use a different variable name as initialized in each run
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
For static volume, the user will manually mounts
already existing image as a volume to the application
pods. As its a rbd Image, if the PVC is of type
fileSystem the image will be mapped, formatted
and mounted on the node,
If the user resizes the image on the ceph cluster.
User cannot not automatically resize the filesystem
created on the rbd image. Even if deletes and
recreates the kubernetes objects, the new size
will not be visible on the node.
With this changes During the NodeStageVolumeRequest
the nodeplugin will check the size of the mapped rbd
image on the node using the devicePath. and also
the rbd image size on the ceph cluster.
If the size is not matching it will do the file
system resize on the node as part of the
NodeStageVolumeRequest RPC call.
The user need to do below operation to see new size
* Resize the rbd image in ceph cluster
* Scale down all the application pods using the static
PVC.
* Make sure no application pods which are using the
static PVC is running on a node.
* Scale up all the application pods.
Validate the new size in application pod mounted
volume.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
in NodeStage operation we are flattening
the image to support mounting on the older
clients. this commits moves it to a helper
function to reduce code complexity.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
During PVC snapshot/clone both kms config and passphrase needs to copied,
while for PVC restore only passphrase needs to be copied to dest rbdvol
since destination storageclass may have another kms config.
Signed-off-by: Rakshith R <rar@redhat.com>
This commit adds the logic to detect a passed in volumeID
is a migrated volume ID and if yes, the driver connect to the
backend cluster and clean/delete the image. The logic
only applied if its a migration volume ID. The migration volume ID
carry the information like mons, pool and image name which is
good enough for the driver to identify and connect to the backend
cluster for its operations.
migration volID format:
<mig>_mons-<monsHash>_image-<imageUID>_<poolHash>
Details on the hash values:
* MonsHash: this carry a hash value (md5sum) which will be acted as the
`clusterID` for the operations in this context.
* ImageUID: this is the unique UUID generated by kubernetes for the created
volume.
* PoolHash: this is an encoded string of pool name.
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
as we are refractoring the cephfs code,
Moving all the core functions to a new folder
/pkg called core. This will make things easier
to implement. For now onwards all the core
functionalities will be added to the core
package.
Signed-off-by: Madhu Rajanna <madhupr007@gmail.com>
the migration nodestage request does not carry the 'clusterID' in it
and only monitors are available with the volumeContext. The volume
context flag 'migration=true' and 'static=true' flags allow us to
fill 'clusterID' from the passed in monitors to the volume Context,so
that rest of the static operations on nodestage can be proceeded as we
do treat static volumes today.
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
Currently, we delete the ceph client log file on unmap/detach.
This patch provides additional alternatives for users who would like to
persist the log files.
Strategies:
-----------
`remove`: delete log file on unmap/detach
`compress`: compress the log file to gzip on unmap/detach
`preserve`: preserve the log file in text format
Note that the default strategy will be remove on unmap, and these options
can be tweaked from the storage class
Compression size details example:
On Map: (with debug-rbd=20)
---------
$ ls -lh
-rw-r--r-- 1 root root 526K Sep 1 18:15
rbd-nbd-0001-0024-fed5480a-f00f-417a-a51d-31d8a8144c03-0000000000000003-d2e89c87-0b4d-11ec-8ea6-160f128e682d.log
On unmap:
---------
$ ls -lh
-rw-r--r-- 1 root root 33K Sep 1 18:15
rbd-nbd-0001-0024-fed5480a-f00f-417a-a51d-31d8a8144c03-0000000000000003-d2e89c87-0b4d-11ec-8ea6-160f128e682d.gz
Signed-off-by: Prasanna Kumar Kalever <prasanna.kalever@redhat.com>
Log:
internal/rbd/rbd_attach.go:424:2: hugeParam: dArgs is heavy (88 bytes);
consider passing it by pointer (gocritic)
Signed-off-by: Prasanna Kumar Kalever <prasanna.kalever@redhat.com>
Currently we return a !ready status if an image
is not found when a replication resync is issued.
We also return a !ready just post issuing a resync.
The change is to ensure we return errors in these
cases for the caller to retry the operation till
we can determine we are actually resyncing, and then
return !ready with nil errors.
Part of addressing:
https://github.com/csi-addons/volume-replication-operator/issues/101
Signed-off-by: Shyamsundar Ranganathan <srangana@redhat.com>
This commit:
- modifies GetMonsAndClusterID() to take clusterID instead of options.
- moves out validation of clusterID is set or not out of GetMonsAndClusterID().
- defines ErrClusterIDNotSet new error for reusability.
- add GetClusterID() to obtain clusterID from options.
Signed-off-by: Rakshith R <rar@redhat.com>
This commit adds capability to genVolFromVolumeOptions() to fetch
mapped clusted-id & mon ips for mirrored PVC on secondary cluster
which may have different cluster-id.
This is required for NodeStageVolume().
We also don't need to check for mapping during volume create requests,
so it can be disabled by passing a bool checkClusterIDMapping as false.
GetMonsAndClusterID() is modified to accept bool checkClusterIDMapping
based on which clustermapping is checked to fetch mapped cluster-id and
mon-ips.
Signed-off-by: Rakshith R <rar@redhat.com>
This commit calls WriteCephConfig() in cephcsi.go to
create ceph.conf and keyring if it is not mounted to
be used by all cli calls and conn cmds.
Before this change, rbd-controller/omap-generator did not create
ceph.conf on startup.
Signed-off-by: Rakshith R <rar@redhat.com>
Thanks to the random unmap failure on my local machine:
I0901 17:08:37.841890 2617035 cephcmds.go:55] ID: 11 Req-ID:
0001-0024-fed5480a-f00f-417a-a51d-31d8a8144c03-0000000000000003-024983f3-0b47-11ec-8fcb-e671f0b9f58e
an error (exit status 22) occurred while running rbd args: [unmap
rbd-pool/csi-vol-024983f3-0b47-11ec-8fcb-e671f0b9f58e --device-type nbd
--options try-netlink --options reattach-timeout=300 --options
io-timeout=0]
Noticed the map args are also getting passed to/as unmap args, which is not
correct. We have separate things for mapOptions and unmapOptions. This PR
makes sure that the map args are not passed at the time of unmap.
Signed-off-by: Prasanna Kumar Kalever <prasanna.kalever@redhat.com>
At present the nodeStageVolume() handle many logic of filling rbdvol
struct based on the request received and this method is complex to
follow. with this patch, filling or populating volOptions has been
segregrated and handled hence make the stage functions' job easy.
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
MirrorImageState (type C.rbd_mirror_image_state_t) has a string
method which can be used while returning error in the replication
controller. Previously, we were using int return in the error which
is not the proper usage.
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
all the error check scenarios of genVolFromVolID() and unreserving
omap entries based on the error made deleteVolume method complex,
this patch create a new function which handle the error check and
unrerving omap entries accordingly and finally return the response
to deletevolume/caller.
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
When NewK8sClient() detects and error, it used to call FatalLogMsg()
which causes a panic. There are additional features that can be used on
Kubernetes clusters, but these are not a requirement for most
functionalities of the driver.
Instead of causing a panic, returning an error should suffice. This
allows using the driver on non-Kubernetes clusters again.
Fixes: #2452
Signed-off-by: Niels de Vos <ndevos@redhat.com>
This commit adds fetchMappedClusterIDAndMons() which returns
monitors and clusterID info after checking cluster mapping info.
This is required for regenerating omap entries in mirrored cluster
with different clusterID.
Signed-off-by: Rakshith R <rar@redhat.com>
This commit moves getMappedID() from rbd to util
package since it is not rbd specific and exports
it from there.
Signed-off-by: Rakshith R <rar@redhat.com>
A new "internal/kms" package is introduced, it holds the API that can be
consumed by the RBD components.
The KMS providers are currently in the same package as the API. With
later follow-up changes the providers will be placed in their own
sub-package.
Because of the name of the package "kms", the types, functions and
structs inside the package should not be prefixed with KMS anymore:
internal/kms/kms.go:213:6: type name will be used as kms.KMSInitializerArgs by other packages, and that stutters; consider calling this InitializerArgs (golint)
Updates: #852
Signed-off-by: Niels de Vos <ndevos@redhat.com>
By placing the NewK8sClient() function in its own package, the KMS API
can be split from the "internal/util" package. Some of the KMS providers
use the NewK8sClient() function, and this causes circular dependencies
between "internal/utils" -> "internal/kms" -> "internal/utils", which
are not alowed in Go.
Updates: #852
Signed-off-by: Niels de Vos <ndevos@redhat.com>
if the volumeattachment has been fetched but marked for deletion
the nbd healer dont want to process further on this pv. This patch
adds a check for pv is marked for deletion and if so, make the
healer skip processing the same
Signed-off-by: Humble Chirammal <hchiramm@redhat.com>
Moving the log functions into its own internal/util/log package makes it
possible to split out the humongous internal/util packages in further
smaller pieces. This reduces the inter-dependencies between utility
functions and components, preventing circular dependencies which are not
allowed in Go.
Updates: #852
Signed-off-by: Niels de Vos <ndevos@redhat.com>
With the tests at CI, it kind of looks like that the IO is timing out after
30 seconds (default with rbd-nbd). Since we have tweaked reattach-timeout
to 300 seconds at ceph-csi, we need to explicitly set io-timeout on the
device too, as it doesn't make any sense to keep
io-timeout < reattach-timeout
Hence we set io-timeout for rbd nbd to 0. Specifying io-timeout 0 tells
the nbd driver to not abort the request and instead see if it can be
restarted on another socket.
Signed-off-by: Prasanna Kumar Kalever <prasanna.kalever@redhat.com>
Suggested-by: Ilya Dryomov <idryomov@redhat.com>
- Update the meta stash with logDir details
- Use the same to remove logfile on unstage/unmap to be space efficient
Signed-off-by: Prasanna Kumar Kalever <prasanna.kalever@redhat.com>
