Most accesses to this map are protected. Add the missing locks (some of which
are in code paths where we wouldn't normally expect concurrent accesses, but
better to be safe).
Signed-off-by: Ian Campbell <ijc@docker.com>
Tests such as TestNoSnapshotLeak were failing in client mode (e.g. using #522)
because we weren't releasing the intermediate refs.
Resolve this by refactoring the existing code which frees the intermediate refs
from `gatewayFrontend.Solve` into a method on `llbBridgeForwarder` and as well
as the original site also call from the solver when the top-level solve (in
clientside frontend mode) completes. The original call (which is via a defer)
could likely sensibly be moved either earlier or later if desired but leave it
here it is to minimise the scope of the change.
The previous code used the `retErr` named return but the code between that
point and the end of the function already ensured that `lbf.err` is the same as
`retErr`, thus the only change in the code which has moved is
`s/retErr/lbf.err/`. Note that the `res` named return was previously unused by
name.
Fixes#581.
Signed-off-by: Ian Campbell <ijc@docker.com>
This allows builder code to be written which can be built as either a gateway
container or in a purely client side configuration, giving implementors more
flexibility.
Now when `Solve` sees a request with neither a definition nor a frontend
specified it will make the job available via new LLBBridge endpoints on the
control socket which the client can then use. These end points require the job
id to be present in the gRPC metadata and a client side object is added to
facilitate this.
The `llbBridgeForwarder` type is now exposed as a public `interface
LLBBridgeForwarder` which satisfies the underlying gRPC server interface
(`pb.LLBBridgeServer`) as well as a new `Done()` & `Result()` pair which can be
used to wait for the client to call `Return()` (using a model similar to
`context.Context`).
Signed-off-by: Ian Campbell <ijc@docker.com>
Reorder the code slightly so that an error returned by the call to the
`BuildFunc` is picked up by the `defer`d error handling.
Signed-off-by: Ian Campbell <ijc@docker.com>
Refactor the setting/getting of the result for a `llbBridgeForwarder` in order
to check that `Return` is only called once and only with the correct options.
Signed-off-by: Ian Campbell <ijc@docker.com>
Provide a `New` method and a new `Run` method on the `grpcClient` to allow
manual creation of a client in addition to the current ability to initialise
from the environment. Accordingly the existing `Run` method becomes
`RunFromEnvironment`.
Signed-off-by: Ian Campbell <ijc@docker.com>
e.g. with busybox image:
OCI runtime create failed: container_linux.go:348:
starting container process caused "process_linux.go:402:
container init caused \"rootfs_linux.go:58:
mounting \\\"proc\\\" to rootfs \\\"/.../rootfs\\\" at \\\"/proc\\\"
caused \\\"mkdir /.../rootfs/proc: read-only file system\\\"\"": unknown
This is because we were setting the underlying snapshot readonly so the various
mountpoints (here /proc) cannot be created. This would not be necessary if
those mountpoints were present in images but they typically are not.
The right way to get around this (used e.g. by `ctr`) is to use a writeable
snapshot but to set root readonly in the OCI spec. In this configuration the
rootfs is writeable when mounts are processed but is then made readonly by the
runtime (runc) just before entering the user specified binary within the
container.
This involved a surprising amount of plumbing.
Use this new found ability in the dockerfile converter's `dispatchCopy`
function.
Signed-off-by: Ian Campbell <ijc@docker.com>
Modelled after the vendor support provide a validator and an updator for files
produced by `go generate` (which today just means `*.pb.go`).
Main difference from the vendor support is that we are no longer simply nuking
and replacing a single directory, so I ended up hardcoding `*.pb.go` in a bunch
of places which I don't like but cannot see a way around which doesn't risk
nuking people's other local changes.
The generated files are placed in an unpacked form in a `FROM scratch`
container for update. Use a subdirectory and `tar --strip-components` (portable
to MacOS and Linux according to `tar(1)`) since trying to do a `docker export`
of just the root ends up adding `.dockerenv`, `sys`, `proc` and `dev` to the
source tree.
The validate container is not `FROM scratch` because we want `cat`.
The run in `frontend/gateway/pb/generate.go` was missing an include so fix
that.
The versions of `protoc` and the gogo plugins were chosen to regenerate the
existing code as closely as possible. The updates to `*.pg.go` here are all the
result of regenerating with go1.9 which fixed
https://github.com/golang/go/issues/17663 and replaced an invalid timestamp in
the gzip header of the data encoded in `fileDescriptor*`, and adopted a new
standard for marking generated files.
Finally, I noticed that my `docker run`s were missing an `--rm` which I
inherited from `validate-vendor`, so fix all those.
Closes: #322
Signed-off-by: Ian Campbell <ijc@docker.com>
Ian Campbell
Tonis Tiigi
Akihiro Suda
Akihiro Suda