We've seen some cases in the wild where waiting for unmount/deactivate
of devmapper devices taking a long time (several seconds). So, we increase
the sleeps to 10 seconds before we timeout. For instance:
https://github.com/dotcloud/docker/issues/4389
But, in order to not keep other processes blocked we unlock the global
dm lock while waiting to allow other devices to continue working.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
We currently use a global lock to protect global data (like the
Devices map) as well as device data itself and access to
(non-threadsafe) libdevmapper.
This commit also adds a per-device lock, which will allow per-device
operations to temporarily release the global lock while e.g. waiting.
The per-device lock will make sure that nothing else accesses that
device while we're operating on it.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
We've seen situations where removal of "ID-init" failed during
container deletion (EBUSY), after removal of "ID" has succeeded. This
caused the container delete operation to fail, and on the next delete
attempt the removal of "ID" failed immediately with "does not exist".
Ideally we should not fail the ID-init removal, but its also non-ideal
to allow a state where the container is half-removed and we cannot
make progress deleting the container. So, we silently ignore not-exist
errors on device removal.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
We no longer pass "pool" anywhere that uses byHash() per the last
commit, so we can now remove this hack.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
Shutdown contains debug warnings like:
[debug] deviceset.go:699 [deviceset docker-0:33-17945897] waitRemove(/dev/mapper/docker-0:33-17945897-pool)
[debug] deviceset.go:380 libdevmapper(3): libdm-common.c:552 (-1) Device /dev/mapper/docker-0:33-17945897-pool not found
This is because shutdown is using removeDeviceAndWait() to remove the pool device and the
wait part fails because the pool is gone.
We fix this by adding a pool specific removal function which avoids all the trickiness of the
normal remove.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
If /dev/loop-control exists on the system running the test then
ioctlLoopCtlGetFree() will be called, but if not it won't. It does
not exist in the standard docker build environment, so the tests
currently require this to not be called. This makes it instead
optional, allowing the tests to run on e.g. Fedora.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
We're currently leaving around lots of empty directories in
/var/lib/docker/devicemapper/mnt/ for removed images and containers.
Fix this by removing the directory when the device is removed.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
This makes the device mapper not zero out blocks allocated on the
thinp device. This is safe in our use case, as we access the device
via a filesystem that doesn't leak any uninitialized data to userspace.
This partially helps with https://github.com/dotcloud/docker/issues/3280
and should generally improve preformance on the devicemapper backend.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
For some reason we seem to get transient EBUSY when removing
thinp devices, which prohibit removing containers. When
this happens we retry a few times which seems to fix the
issue for me.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
This makes sure the device is removed just like in deactivateDevice.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
We already have some kind of refcounting in DeviceSet, this fleshes
it out to allow it to completely subsume the refcounting in
devmapper.Driver.
This allows us to drop the double refcounting, and the locking inside
devmapper.Driver. This, in particular the locking simplification will
make it easier in the future to parallelize the device mapper.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
This adds a function that calls the lowlevel removeDevice
and then waits for it to finish.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
"Remove" is a bit overloaded, as it is also what deactivating a
device mapper device is called. Using "delete" is more clear here.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
This implements the new Put() operation such that
Get()/Put() maintains a refcount for each ID, mounting
only on first Get() and unmounting on the last Get().
This means we avoid littering the system with lots of mounts
and active devicemapper devices and free resources related
to them.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
asdfljk
This makes all users of Put() have a corresponding call
to Get() which means we will be able to track whether
any particular ID is in use and if not unmount it.
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
This works around the fact that deleting a device in a thin pool
doesn't discard the free space. Unfortunately even this is not perfect,
as it seems discards are respected only for blocks that has never been
shared in the thin device code. However, this has been fixed in the
upstream kernel device-mapper tree:
http://git.kernel.org/cgit/linux/kernel/git/device-mapper/linux-dm.git/commit/?h=for-next&id=0ab1c92ff748b745c1ed7cde31bb37ad2c5f901a
When this hits the kernel I belive this will fully return space
for removed images/containers to the host FS. For now it only
helps partially (which is better than nothing).
Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)
