This is a better fix for futimes() on kernels not supporting O_PATH.
The previous fix broke when copying a device, as it tried to open it
and got and error.
Older kernel can't handle O_PATH in open() so this will
fail on dirs and symlinks. For dirs wa can fallback to
the normal Utimes, but for symlinks there is not much to do
but ignore their timestamps.
This creates a container by copying the corresponding files
from the layers into the containers. This is not gonna be very useful
on a developer setup, as there is no copy-on-write or general diskspace
sharing. It also makes container instantiation slower.
However, it may be useful in deployment where we don't always have a lot
of containers running (long-running daemons) and where we don't
do a lot of docker commits.
Change the comparison to better handle files that are copied during
container creation but not actually changed:
* Inode - this will change during a copy
* ctime - this will change during a copy (as we can't set it back)
* blocksize - this will change for sparse files during copy
* size for directories - this can change anytime but doesn't
necessarily reflect an actual contents change
* Compare mtimes at microsecond precision (as this is what utimes has)
There are some changes here that make the file metadata better match
the layer files:
* Set the mode of the file after the chown, as otherwise the per-group/uid
specific flags and e.g. sticky bit is lost
* Use lchown instead of chown
* Delay mtime updates to after all other changes so that later file
creation doesn't change the mtime for the parent directory
* Use Futimes in combination with O_PATH|O_NOFOLLOW to set mtime on symlinks
Rather than scan the files in the old directory twice to detect the
deletions we now scan both directories twice and then do all the
diffing on the in-memory structure.
This is more efficient, but it also lets us diff more complex things
later that are not exact on-disk trees.
The init layer needs to be topmost to make sure certain files
are always there (for instance, the ubuntu:12.10 image wrongly
has /dev/shm being a symlink to /run/shm, and we need to override
that). However, previously the devmapper code implemented the
init layer by putting it in the base devmapper device, which meant
layers above it could override these files (so that ubuntu:12.10
broke).
So, instead we put the base layer in *each* images devmapper device.
This is "safe" because we still have the pristine layer data
in the layer directory. Also, it means we diff the container
against the image with the init layer applied, so it won't show
up in diffs/commits.
lxc-start requires / to be mounted private, otherwise the changes
it does inside the container (both mounts and unmounts) will propagate
out to the host.
We work around this by starting up lxc-start in its own namespace where
we set / to rprivate.
Unfortunately go can't really execute any code between clone and exec,
so we can't do this in a nice way. Instead we have a horrible hack that
use the unshare command, the shell and the mount command...
Solomon Hykes