This makes sure fsdiff doesn't try to unmount things that shouldn't be.
**Note**: This is intended as a temporary solution to have as minor a
change as possible for 1.11.1. A bigger change will be required in order
to support container re-attach.
Signed-off-by: Brian Goff <cpuguy83@gmail.com>
Right now there is no way to know what's the minimum free space threshold
daemon is applying. It would be good to export it through docker info and
then user knows what's the current value. Also this could be useful to
higher level management tools which can look at this value and setup their
own internal thresholds for image garbage collection etc.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
People have reported following issue with overlay
$ docker run -ti --name=foo -v /dev/:/dev fedora bash
$ docker cp foo:/bin/bash /tmp
$ exit container
Upon container exit, /dev/pts gets unmounted too. This happens because
docker cp volume mounts get propagated to /run/docker/libcontainer/....
and when container exits, it must be tearing down mount point under
/run/docker/libcontainerd/... and as these are "shared" mounts it
propagates events to /dev/pts and it gets unmounted too.
One way to solve this problem is to make sure "docker cp" volume mounts
don't become visible under /run/docker/libcontainerd/..
Here are more details of what is actually happening.
Make overlay home directory (/var/lib/docker/overlay) private mount when
docker starts and unmount it when docker stops. Following is the reason
to do it.
In fedora and some other distributions / is "shared". That means when
docker creates a container and mounts it root in /var/lib/docker/overlay/...
that mount point is "shared".
Looks like after that containerd/runc bind mounts that rootfs into
/runc/docker/libcontainerd/container-id/rootfs. And this puts both source
and destination mounts points in shared group and they both are setup
to propagate mount events to each other.
Later when "docker cp" is run it sets up container volumes under
/var/lib/dokcer/overlay/container-id/... And all these mounts propagate
to /runc/docker/libcontainerd/... Now mountVolumes() makes these new
mount points private but by that time propagation already has happened
and private only takes affect when unmount happens.
So to stop this propagation of volumes by docker cp, make
/var/lib/docker/overlay a private mount point. That means when a container
rootfs is created, that mount point will be private too (it will inherit
property from parent). And that means when bind mount happens in /runc/
dir, overlay mount point will not propagate mounts to /runc/.
Other graphdrivers like devicemapper are already doing it and they don't
face this issue.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
In TP5, Hyper-V containers need all image files ACLed so that the virtual
machine process can access them. This was fixed post-TP5 in Windows, but
for TP5 we need to explicitly add these ACLs.
Signed-off-by: John Starks <jostarks@microsoft.com>
If aufs is already modprobe'd but we are in a user namespace, the
aufs driver will happily load but then get eperm when it actually tries
to do something. So detect that condition.
Signed-off-by: Serge Hallyn <serge.hallyn@ubuntu.com>
This adds support to the Windows graph driver for ApplyDiff on a base
layer. It also adds support for hard links, which are needed because the
Windows base layers double in size without hard link support.
Signed-off-by: John Starks <jostarks@microsoft.com>
Fixes an issue that prevents nano server images from loading properly. Also updates logic for custom image loading to avoid preventing daemon start because an image failed to load.
Signed-off-by: Stefan J. Wernli <swernli@microsoft.com>
Overlay tests were failing when /var/tmp was an overlay mount with a misleading message.
Now overlay tests will be skipped when attempting to be run on overlay.
Tests will now use the TMPDIR environment variable instead of only /var/tmp
Fixes#21686
Signed-off-by: Derek McGowan <derek@mcgstyle.net> (github: dmcgowan)
On aufs, auplink is run before the Unmount. Irrespective of the
result, we proceed to issue a Unmount syscall. In which case,
demote erros on auplink to warning.
Signed-off-by: Anusha Ragunathan <anusha@docker.com>
Since the layer store was introduced, the level above the graphdriver
now differentiates between read/write and read-only layers. This
distinction is useful for graphdrivers that need to take special steps
when creating a layer based on whether it is read-only or not.
Adding this parameter allows the graphdrivers to differentiate, which
in the case of the Windows graphdriver, removes our dependence on parsing
the id of the parent for "-init" in order to infer this information.
This will also set the stage for unblocking some of the layer store
unit tests in the next preview build of Windows.
Signed-off-by: Stefan J. Wernli <swernli@microsoft.com>
These fields are needed to specify the exact version of Windows that an
image can run on. They may be useful for other platforms in the future.
This also changes image.store.Create to validate that the loaded image is
supported on the current machine. This change affects Linux as well, since
it now validates the architecture and OS fields.
Signed-off-by: John Starks <jostarks@microsoft.com>
btrfs-progs-4.5 introduces device delete by devid
for this reason btrfs_ioctl_vol_args_v2's name was encapsulated
in a union
this patch is for setting btrfs_ioctl_vol_args_v2's name
using a C function in order to preserve compatibility
with all btrfs-progs versions
Signed-off-by: Julio Montes <imc.coder@gmail.com>
Instead of implementing refcounts at each graphdriver, implement this in
the layer package which is what the engine actually interacts with now.
This means interacting directly with the graphdriver is no longer
explicitly safe with regard to Get/Put calls being refcounted.
In addition, with the containerd, layers may still be mounted after
a daemon restart since we will no longer explicitly kill containers when
we shutdown or startup engine.
Because of this ref counts would need to be repopulated.
Signed-off-by: Brian Goff <cpuguy83@gmail.com>
Instead of implementing refcounts at each graphdriver, implement this in
the layer package which is what the engine actually interacts with now.
This means interacting directly with the graphdriver is no longer
explicitly safe with regard to Get/Put calls being refcounted.
In addition, with the containerd, layers may still be mounted after
a daemon restart since we will no longer explicitly kill containers when
we shutdown or startup engine.
Because of this ref counts would need to be repopulated.
Signed-off-by: Brian Goff <cpuguy83@gmail.com>
Now what we provide dynamic binaries for all plaforms,
we shouldn't try to run docker without udev sync support.
This change changes the previous warning to an Error,
unless the user explicitly overrides the warning, in
which case they're at their own risk.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Once thin pool gets full, bad things can happen. Especially in case of xfs
it is possible that xfs keeps on retrying IO infinitely (for certain kind
of IO) and container hangs.
One way to mitigate the problem is that once thin pool is about to get full,
start failing some of the docker operations like pulling new images or
creation of new containers. That way user will get warning ahead of time
and can try to rectify it by creating more free space in thin pool. This
can be done either by deleting existing images/containers or by adding more
free space to thin pool.
This patch adds a new option dm.min_free_space to devicemapper graph
driver. Say one specifies dm.min_free_space=10%. This means atleast
10% of data and metadata blocks should be free in pool before new device
creation is allowed, otherwise operation will fail.
By default min_free_space is 10%. User can change it by specifying
dm.min_free_space=X% on command line. A value of 0% will disable the
check.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Check whether or not the file system type of a mountpoint is aufs
by calling statfs() instead of parsing mountinfo. This assumes
that aufs graph driver does not allow aufs as a backing file
system.
Signed-off-by: Tatsushi Inagaki <e29253@jp.ibm.com>
Previously, Windows layer diffs were written using a Windows-internal
format based on the BackupRead/BackupWrite Win32 APIs. This caused
problems with tar-split and tarsum and led to performance problems
in implementing methods such as DiffPath. It also was just an
unnecessary differentiation point between Windows and Linux.
With this change, Windows layer diffs look much more like their
Linux counterparts. They use AUFS-style whiteout files for files
that have been removed, and they encode all metadata directly in
the tar file.
This change only affects Windows post-TP4, since changes to the Windows
container storage APIs were necessary to make this possible.
Signed-off-by: John Starks <jostarks@microsoft.com>
This allows a graph driver to provide a custom FileGetter for tar-split
to use. Windows will use this to provide a more efficient implementation
in a follow-up change.
Signed-off-by: John Starks <jostarks@microsoft.com>
Fixes issues with layer remounting (e.g. a running container which then
has `docker cp` used to copy files in or out) by applying the same
refcounting implementation that exists in other graphdrivers like
overlay and aufs.
Docker-DCO-1.1-Signed-off-by: Phil Estes <estesp@linux.vnet.ibm.com> (github: estesp)
Save was failing file integrity checksums due to bugs in both
Windows and Docker. This commit includes fixes to file time handling
in tarexport and system.chtimes that are necessary along with
the Windows platform fixes to correctly support save. With this
change, sysfile_backups for windowsfilter driver are no longer
needed, so that code is removed.
Signed-off-by: Stefan J. Wernli <swernli@microsoft.com>
Fix root directory of the mountpoint being owned by real root. This is
unique to ZFS because of the way file mountpoints are created using the
ZFS tooling, and the remapping that happens at layer unpack doesn't
impact this root (already created) holding directory for the layer.
Docker-DCO-1.1-Signed-off-by: Phil Estes <estesp@linux.vnet.ibm.com> (github: estesp)
For btrfs driver, in d.Create(), Get() of parentDir is called but not followed
by Put().
If we apply SElinux mount label, we need to mount btrfs subvolumes in d.Get(),
without a Put() would end up with a later Remove() failure on
"Device resourse is busy".
This calls the subvolume helper function directly in d.Create().
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Most storage drivers call graphdriver.GetFSMagic(home),
it is more clean to easy to maintain. So btrfs need to
adopt such change.
Signed-off-by: Kai Qiang Wu(Kennan) <wkqwu@cn.ibm.com>
Right now if somebody has enabled deferred device deletion, then
deleteTransaction() returns success even if device could not be deleted. It
has been marked for deferred deletion. Right now we will mark device ID free
and potentially use it again when somebody tries to create new container. And
that's wrong. Device ID is not free yet. It will become free once devices
has actually been deleted by the goroutine later.
So move the location of call to markDeviceIDFree() to a place where we know
device actually got deleted and was not marked for deferred deletion.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Make sure btrfs mounted subvolumes are owned properly when a remapped
root exists (user namespaces are enabled, for example)
Docker-DCO-1.1-Signed-off-by: Phil Estes <estesp@linux.vnet.ibm.com> (github: estesp)
Instead of creating a "0.0" subdirectory and migrating graphroot
metadata into it when user namespaces are available in the daemon
(currently only in experimental), change the graphroot dir permissions
to only include the execute bit for "other" users.
This allows easy migration to and from user namespaces and will allow
easier integration of user namespace support into the master build.
Docker-DCO-1.1-Signed-off-by: Phil Estes <estesp@linux.vnet.ibm.com> (github: estesp)
Support restoreCustomImage for windows with a new interface to extract
the graph driver from the LayerStore.
Signed-off-by: Daniel Nephin <dnephin@docker.com>
The loopback logic is not technically exclusive to the devicemapper
driver. This reorganizes the code such that the loopback code is usable
outside of the devicemapper package and driver.
Signed-off-by: Vincent Batts <vbatts@redhat.com>
Really fixing 2 things:
1. Panic when any error is detected while walking the btrfs graph dir on
removal due to no error check.
2. Nested subvolumes weren't actually being removed due to passing in
the wrong path
On point 2, for a path detected as a nested subvolume, we were calling
`subvolDelete("/path/to/subvol", "subvol")`, where the last part of the
path was duplicated due to a logic error, and as such actually causing
point #1 since `subvolDelete` joins the two arguemtns, and
`/path/to/subvol/subvol` (the joined version) doesn't exist.
Also adds a test for nested subvol delete.
Signed-off-by: Brian Goff <cpuguy83@gmail.com>
After the very first init of the graph `docker info` correctly shows the
base fs type under `Backing Filesystem`. This information isn't stored
anywhere. After a restart (w/o erasing `/var/lib/docker`) `docker info`
shows an empty string under `Backing Filesystem`.
This patch records the base fs type after the first run in the metadata
or, to fix old devices that don't have this info in the metadata, just
probe the fs type of the base device at graph startup.
Signed-off-by: Antonio Murdaca <runcom@redhat.com>
All underlay dirs need proper remapped ownership. This bug was masked by the
fact that the setupInitLayer code was chown'ing the dirs at startup
time. Since that bug is now fixed, it revealed this permissions issue.
Docker-DCO-1.1-Signed-off-by: Phil Estes <estesp@linux.vnet.ibm.com> (github: estesp)
ext4 filesystem creation can take a long time on 100G thin device and
systemd might time out and kill docker service. Often user is left thinking
why docker is taking so long and logs don't give any hint. Log an info
message in journal for start and end of filesystem creation. That way
a user can look at logs and figure out that filesystem creation is
taking long time.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Add distribution package for managing pulls and pushes. This is based on
the old code in the graph package, with major changes to work with the
new image/layer model.
Add v1 migration code.
Update registry, api/*, and daemon packages to use the reference
package's types where applicable.
Update daemon package to use image/layer/tag stores instead of the graph
package
Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
Signed-off-by: Tonis Tiigi <tonistiigi@gmail.com>
This change will allow us to run SELinux in a container with
BTRFS back end. We continue to work on fixing the kernel/BTRFS
but this change will allow SELinux Security separation on BTRFS.
It basically relabels the content on container creation.
Just relabling -init directory in BTRFS use case. Everything looks like it
works. I don't believe tar/achive stores the SELinux labels, so we are good
as far as docker commit.
Tested Speed on startup with BTRFS on top of loopback directory. BTRFS
not on loopback should get even better perfomance on startup time. The
more inodes inside of the container image will increase the relabel time.
This patch will give people who care more about security the option of
runnin BTRFS with SELinux. Those who don't want to take the slow down
can disable SELinux either in individual containers or for all containers
by continuing to disable SELinux in the daemon.
Without relabel:
> time docker run --security-opt label:disable fedora echo test
test
real 0m0.918s
user 0m0.009s
sys 0m0.026s
With Relabel
test
real 0m1.942s
user 0m0.007s
sys 0m0.030s
Signed-off-by: Dan Walsh <dwalsh@redhat.com>
Signed-off-by: Dan Walsh <dwalsh@redhat.com>
If platform supports xfs filesystem then use xfs as default filesystem
for container rootfs instead of ext4. Reason being that ext4 is pre-allcating
lot of metadata (around 1.8GB on 100G thin volume) and that can take long
enough on AWS storage that systemd times out and docker fails to start.
If one disables pre-allocation of ext4 metadata, then it will be allocated
when containers are mounted and we will have multiple copies of metadata
per container. For a 100G thin device, it was around 1.5GB of metadata
per container.
ext4 has an optimization to skip zeroing if discards are issued and
underlying device guarantees that zero will be returned when discarded
blocks are read back. devicemapper thin devices don't offer that guarantee
so ext4 optimization does not kick in. In fact given discards are optional
and can be dropped on the floor if need be, it looks like it might not be
possible to guarantee that all the blocks got discarded and if read back
zero will be returned.
Signed-off-by: Anusha Ragunathan <anusha@docker.com>
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
If user wants to use a filesystem it can be specified using dm.fs=<filesystem>
option. It is possible that docker already had base image and a filesystem
on that. Later if user wants to change file system using dm.fs= option
and restarts docker, that's not possible. Warn user about it.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
