If dir_fd wasn't passed, create_device() openend it to `/dev` and forgot
about closing it. To fix this, it would have to gain logic to only close
the fd if it wasn't passed in.
Side-step the problem by removing dir_fd, since nothing is using it
right now. We can add it back if something needs it.
Closing the socket is the responsibility of whoever opened it.
Fix this in the only user (qemu assembler) by using socket() in a `with`
block, which closes the socket on exit.
Storytime! I tried to run multiple osbuilds at once. It failed when
unmounting the buildtree. Weird. It turned out the buildtree was not
there anymore when osbuild tried to unmount it. But who unmounted it?
We need to deep dive into mount-types.
Nowadays, the / directory is shared-mounted by systemd. See:
https://serverfault.com/questions/868682/implications-of-mount-make-private
This has interesting implications, see the following example:
we start osbuild1 with /var/tmp/os1 as its store
osbuild1 creates /var/tmp/os1/tmp
osbuild1 bind-mounts / onto /var/tmp/os1/tmp
we start osbuild2 with /var/tmp/os2 as its store
osbuild2 creates /var/tmp/os2/tmp
osbuild2 bind-mounts / onto /var/tmp/os2/tmp
Now, the shared-mounting goes into effect:
The second mount-event gets propagated into the first mount, where it
creates another mount, so we get something like this:
/var/tmp/os1/tmp/var/tmp/os2/tmp
But this is just a start! Imagine running three osbuilds at once.
The event would get propagated to those 3 mounts created by two
osbuilds, creating 3 extra mounts, 7 in total.
It turns out this mounting strategy creates an *exponential number* of
mounts. Crazy, right?
This commit mounts the root inside build root using private bind, which
doesn't propagate bind-events. This solves the problem with the
exponential growth.
But the original problem was different, mount points were disappearing.
So how does this fix solve the problem?
Honestly, I don't know. Something with mount-event propagation is
probably responsible, but I cannot imagine how it is actually affecting
the unbinding.
Treat outputs like we treat trees: store them in the object store. This
simplifies using osbuild and allows returning a cached version if one is
available.
This makes the `--output` parameter redundant. Remove it.
`osbuild --json [ARGS]` will suppress the normal output and print its
result as JSON. For now, it only does this when it returns 0. Otherwise,
it prints the error from the latest stage.
This is useful for other tools to call it and get machine-readable
output.
Introduce and output id, which is the checksum over a full pipeline,
including all stages and the assembler. The id of a pipeline did not
include assemblers before. To be less confusing, rename the existing id
to "tree id".
In BuildRoot a new mount /var pointing to temporary directory in host's
/var/tmp is created. This enables us to have temporary storage inside
the container which is not hosted on tmpfs. Thanks to that we can move
larger files out of the part of filesystem which is hosted on tmpfs to
save up memory on machines with low memory capacity.
The best practice for creating a pipeline should be to include at least
one level of build-pipelines. This makes sure that the tools used to
generate the target image are well-defined.
In principle one could add several layers, though in pracite, one would
hope that the envinment used to build the buildroot does not affect the
final image (and as we anyway cannot recurr indefinitely, we fall back
to simply using the host system in this case).
This only makes sense, if the contents of the host system truly does not
affect the generated image, and as such we do not include any information
about the host when computing the hash that identifies a pipeline.
In fact, any image could be used in its place, as long as the required
tools are present. This commit takes advantage of that fact. Rather than
run a pipeline with the host as the build root, take a second pipeline
to generate the buildroot, but do not include this when computing the
pipeline id (so it is different from simply editing the original JSON).
This is necessary so we can use the same pipelines on significantly
different host systems (run with different --bulid-pipeline arguments).
In particular, it allows our test pipelines that generate f30 images
to be run unmodified on Travis (which runs Ubuntu).
Signed-off-by: Tom Gundersen <teg@jklm.no>
Import modules between files using the syntax `from . import foobar`,
renaming what used to be `FooBar` to `foobar.FooBar` when moved to a
separate file.
In __init__.py only import what is meant to be public API.
Signed-off-by: Tom Gundersen <teg@jklm.no>
Make the order of argumnets in line with how it is used (and also
how it is conceptionally closer to the pipeline json document).
This makes no practical difference as the two arguments were both
just used for computing the hash.
Signed-off-by: Tom Gundersen <teg@jklm.no>
Each pipeline is now self-contained without references to another.
However, as the final stage in a pipeline is saved to the content
store, we are able to reuse it if one pipeline is the prefix of
another, as described in the previous commit. This makes the
concept of a base redundant.
The ObjectStore must take a directory as argument, never None, so
the conditional assertion for this in Pipeline.run() is ok to
remove.
Signed-off-by: Tom Gundersen <teg@jklm.no>
Don't do this only for the base, but for any prefix of the current
pipeline.
Note that if two pipelines share a prefix, but one is not the prefix
of another, no sharing is possible. Only a proper prefix can be
reused by another pipeline, as only the result of the last pipeline
is saved to the object store (this restriction could be changed in
the future).
Signed-off-by: Tom Gundersen <teg@jklm.no>
Take this as an argumnet to __init__ in the same way that `base`
is.
This avoids us having to deal with the case of someone setting a
stage before the build, which does not work as the stage id will
be wrong.
Signed-off-by: Tom Gundersen <teg@jklm.no>
Renaming a directory over an existing one is only an error if the
existing one is not empty, in which case ENOEMPTY is thrown.
Tested with:
>>> os.mkdir("foo")
>>> os.mkdir("bar")
>>> os.rename("foo", "bar")
# no error
>>> open("foo/a", "w").write("a")
1
>>> try: os.rename("bar", "foo")
... except OSError as e: e.errno == errno.ENOTEMPTY
...
True
The build pipeline, is a sub-pipeline used to generate the build
tree to use rather than the current root directory. This can be
nested arbitrarily deep, but ultimately we will fall back to the
current logic when no build property is found.
Just like the tree after the last stage of a regular pipeline ends
up in the object store, so does currently each build tree (as the
build sub-pipeline really is just a regular pipeline in its own
right). We may want to avoid both these instances of the implicit
storing semantics, and rather make it something the caller opts-in
to. However, for now that is left as a future optimization.
Signed-off-by: Tom Gundersen <teg@jklm.no>
This also changes the structure of the object store, though the
basic idea is the same.
The object store contains a directory of objects, which are content
addressable filesystem trees. Currently we only ever use their
content-hash internally, but the idea for this is basically Lars
Karlitski and Kay Sievers' `treesum()`. We may exopse this in the
future.
Moreover, it contains a directory of refs, which are symlinks named
by the stage id they correspond to (as before), pointing to an object
generated from that stage-id.
The ObjectStore exposes three method:
`has_tree()`: This checks if the content store contains the given tree.
If so, we can rely on the tree remaining there.
`get_tree()`: This is meant to be used with a `with` block and yields
the path to a read-only instance of the tree with the given id. If the
tree_id is passed in as None, an empty directory is given instead.
`new_tree()`: This is meant to be used with a `with` block and yields
the path to a directory in which the tree by the given id should be
created. If a base_id is passed in, the tree is initialized with the
tree with the given id. Only when the block is exited successfully
is the tree written to the content store, referenced by the id in
question.
Use this in Pipeline.run() to avoid regenerating trees unneccessarily.
In order to trigger a regeneration, the content store must currently
be manually flushed.
Update the travis test to run the noop pipeline twice, verifying that
the stage is only run the first time.
Signed-off-by: Tom Gundersen <teg@jklm.no>
Rather than hard-coding this to /, let the caller provide the
directory path to use.
In the past, we needed to give special treatment to /, as it had
to be bind-mounted before being used by nspawn, to work around a
check they had, refusing to use the host root in the container.
We no longer pass the directory directly to nspawn, but rather
mount the subdirs we want ourselves, so that no longer applies.
The callers pass in /, so the behavior is unchanged.
Signed-off-by: Tom Gundersen <teg@jklm.no>
We want the same functionality, but we now impleent it ourselves.
In addition to bind-mounting in /usr into the target container
(which is all nspawn does), we also add /bin, /sbin, /lib and
/lib64, if they exist and are not symlinks (presuambly into
/usr).
This means we can work on distros who have not implemented the
usr-move, like Ubuntu Bionic (used by Travis).
Signed-off-by: Tom Gundersen <teg@jklm.no>
The underlying filesystem was mounted in __init__ and unmonuted in
__exit__/__del__. This meant that if the same object was reused in
several `with` clauses, only the first one would work as intended.
Signed-off-by: Tom Gundersen <teg@jklm.no>
Support the LOOP_SET_DIRECT_IO ioctl, which alows us to control
whether or not a loopback device should perform its own buffering
or rely on the one done by the underlying backing file.
Enabling this should improve both throughput and memory consumption,
it is not currently hooked up as more testing would be required.
Stop guessing if we're in the source directory by looking if a `stages`
subdirectory exists. Instead, assume that osbuild is installed on the
host.
If `--libdir` is given, mount the libdir into `/run/osbuild/lib` (alas,
we can't overwrite `/usr/libexec/osbuild`) and run osbuild from there.
Thus, running from source must now be done like this:
# python3 -m osbuild --libdir . [other args]
This really only makes sense if we are running systemd as PID1
inside the container, but we are not booting a system, just using
it as a glorified chroot.
This means entering the namespaces from the outside will be a bit
more cumbersome, but that was not used much and was never reliable
to begin with.
Signed-off-by: Tom Gundersen <teg@jklm.no>
loop.py is a simple wrapper around the kernel loop API. remoteloop.py
uses this to create a server/clinet pair that communicates over an
AF_UNIX/SOCK_DGRAM socket to allow the server to create loop devices
for the client.
The client passes a fd that should be bound to the resulting loop
device, and a dir-fd where the loop device node should be created.
The server returns the name of the device node to the client.
The idea is that the client is run from whithin a container without
access to devtmpfs (and hence /dev/loop-control), and the server
runs on the host. The client would typically pass its (fake) /dev
as the output directory.
For the client this will be similar to `losetup -f foo.img --show`.
[@larskarlitski: pylint: ignore the new LoopInfo class, because it
only has dynamic attributes. Also disable attribute-defined-outside-init,
which (among other problems) is not ignored for that class.]
Signed-off-by: Tom Gundersen <teg@jklm.no>
Add a directory to each BuildRoot potentially containing a set of
sockets. Also add a helper to create a named bound socket in a given
BuildRoot.
Signed-off-by: Tom Gundersen <teg@jklm.no>
