The Container.State struct holds the container's state, and most of
its fields are expected to change dynamically. Some o these state-changes
are explicit, for example, setting the container to be "stopped". Other
state changes can be more explicit, for example due to the containers'
process exiting or being "OOM" killed by the kernel.
The distinction between explicit ("desired") state changes and "state"
("actual state") is sometimes vague; for some properties, we clearly
separated them, for example if a user requested the container to be
stopped or restarted, we store state in the Container object itself;
HasBeenManuallyStopped bool // used for unless-stopped restart policy
HasBeenManuallyRestarted bool `json:"-"` // used to distinguish restart caused by restart policy from the manual one
Other properties are more ambiguous. such as "HasBeenStartedBefore" and
"RestartCount", which are stored on the Container (and persisted to
disk), but may be more related to "actual" state, and likely should
not be persisted;
RestartCount int
HasBeenStartedBefore bool
Given that (per the above) concurrency must be taken into account, most
changes to the `container.State` struct should be protected; here's where
things get blurry. While the `State` type provides various accessor methods,
only some of them take concurrency into account; for example, [State.IsRunning]
and [State.GetPID] acquire a lock, whereas [State.ExitCodeValue] does not.
Even the (commonly used) [State.StateString] has no locking at all.
The way to handle this is error-prone; [container.State] contains a mutex,
and it's exported. Given that its embedded in the [container.Container]
struct, it's also exposed as an exported mutex for the container. The
assumption here is that by "merging" the two, the caller to acquire a lock
when either the container _or_ its state must be mutated. However, because
some methods on `container.State` handle their own locking, consumers must
be deeply familiar with the internals; if both changes to the `Container`
AND `Container.State` must be made. This gets amplified more as some
(exported!) methods, such as [container.SetRunning] mutate multiple fields,
but don't acquire a lock (so expect the caller to hold one), but their
(also exported) counterpart (e.g. [State.IsRunning]) do.
It should be clear from the above, that this needs some architectural
changes; a clearer separation between "desired" and "actual" state (opening
the potential to update the container's config without manually touching
its `State`), possibly a method to obtain a read-only copy of the current
state (for those querying state), and reviewing which fields belong where
(and should be persisted to disk, or only remain in memory).
This PR preserves the status quo; it makes no structural changes, other
than exposing where we access the container's state. Where previously the
State fields and methods were referred to as "part of the container"
(e.g. `ctr.IsRunning()` or `ctr.Running`), we now explicitly reference
the embedded `State` (`ctr.State.IsRunning`, `ctr.State.Running`).
The exception (for now) is the mutex, which is still referenced through
the embedded struct (`ctr.Lock()` instead of `ctr.State.Lock()`), as this
is (mostly) by design to protect the container, and what's in it (including
its `State`).
[State.IsRunning]: c4afa77157/daemon/container/state.go (L205-L209)
[State.GetPID]: c4afa77157/daemon/container/state.go (L211-L216)
[State.ExitCodeValue]: c4afa77157/daemon/container/state.go (L218-L228)
[State.StateString]: c4afa77157/daemon/container/state.go (L102-L131)
[container.State]: c4afa77157/daemon/container/state.go (L15-L23)
[container.Container]: c4afa77157/daemon/container/container.go (L67-L75)
[container.SetRunning]: c4afa77157/daemon/container/state.go (L230-L277)
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
The "backend" types in API were designed to decouple the API server
implementation from the daemon, or other parts of the code that
back the API server. This would allow the daemon to evolve (e.g.
functionality moved to different subsystems) without that impacting
the API server's implementation.
Now that the API server is no longer part of the API package (module),
there is no benefit to having it in the API module. The API server
may evolve (and require changes in the backend), which has no direct
relation with the API module (types, responses); the backend definition
is, however, coupled to the API server implementation.
It's worth noting that, while "technically" possible to use the API
server package, and implement an alternative backend implementation,
this has never been a prime objective. The backend definition was
never considered "stable", and we don't expect external users to
(attempt) to use it as such.
This patch moves the backend types to the daemon/server package,
so that they can evolve with the daemon and API server implementation
without that impacting the API module (which we intend to be stable,
following SemVer).
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
These comments were added to enforce using the correct import path for
our packages ("github.com/docker/docker", not "github.com/moby/moby").
However, when working in go module mode (not GOPATH / vendor), they have
no effect, so their impact is limited.
Remove these imports in preparation of migrating our code to become an
actual go module.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
This function had some LCOW remnants, where the assumption was made
the only on Windows, the image's OS could potentially not match the
host's OS (see 3e6a13ccb8).
While we currently are not able to run a Windows image on Linux (or
vice versa), this function doesn't have to take into account;
- If a shell is configured; use whatever is configured
- otherwise, use "cmd.exe" for Windows images, and "/bin/sh" otherwise
(likely Linux, but the existing code did not account for other platforms).
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
It was not using the daemon, so can be a regular function. While at it,
also changed the parameter type to accept a regular string-slice, as
we don't need strslice.StrSlice's json.Unmarshaler implementation, and
reversed the logic for the early return.
Finally, for uses where the entrypoint was always nil, this patch removes
the use of this utility altogether.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Change the persistent container metadata to store the whole platform
(as defined by OCI) instead of only the operating system.
Signed-off-by: Paweł Gronowski <pawel.gronowski@docker.com>
This moves the `Health` and `HealthcheckResult` types to the container package,
as well as the related `NoHealthcheck`, `Starting`, `Healthy`, and `Unhealthy`
consts.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
The health status and probe log of containers are not mission-criticial
data which must survive a crash. It is not worth prematrely wearing out
consumer-grade flash storage by overwriting and fsync()ing the container
config on after every probe. Update only the live Container object and
the ViewDB replica on every container health probe instead. It will
eventually get checkpointed along with some other state (or config)
change. Running containers will not be checkpointed on daemon shutdown
when live-restore is enabled, but it does not matter: the health status
and probe log will be zeroed out when the daemon starts back up.
Signed-off-by: Cory Snider <csnider@mirantis.com>
When the start interval is 0 we should treat that as unset.
This is especially important for older API versions where we reset the
value to 0.
Instead of using the default probe value we should be using the
configured `interval` value (which may be a default as well) which gives
us back the old behavior before support for start interval was added.
Signed-off-by: Brian Goff <cpuguy83@gmail.com>
The github.com/containerd/containerd/log package was moved to a separate
module, which will also be used by upcoming (patch) releases of containerd.
This patch moves our own uses of the package to use the new module.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Define consts for the Actions we use for events, instead of "ad-hoc" strings.
Having these consts makes it easier to find where specific events are triggered,
makes the events less error-prone, and allows documenting each Action (if needed).
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
This adds an additional interval to be used by healthchecks during the
start period.
Typically when a container is just starting you want to check if it is
ready more quickly than a typical healthcheck might run. Without this
users have to balance between running healthchecks to frequently vs
taking a very long time to mark a container as healthy for the first
time.
Signed-off-by: Brian Goff <cpuguy83@gmail.com>
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
We have integration tests which assert the invariant that a
GET /containers/{id}/json response lists only IDs of execs which are in
the Running state, according to GET /exec/{id}/json. The invariant could
be violated if those requests were to race the handling of the exec's
task-exit event. The coarse-grained locking of the container ExecStore
when starting an exec task was accidentally synchronizing
(*Daemon).ProcessEvent and (*Daemon).ContainerExecInspect to it just
enough to make it improbable for the integration tests to catch the
invariant violation on execs which exit immediately. Removing the
unnecessary locking made the underlying race condition more likely for
the tests to hit.
Maintain the invariant by deleting the exec from its container's
ExecCommands before clearing its Running flag. Additionally, fix other
potential data races with execs by ensuring that the ExecConfig lock is
held whenever a mutable field is read from or written to.
Signed-off-by: Cory Snider <csnider@mirantis.com>
The containerd client is very chatty at the best of times. Because the
libcontained API is stateless and references containers and processes by
string ID for every method call, the implementation is essentially
forced to use the containerd client in a way which amplifies the number
of redundant RPCs invoked to perform any operation. The libcontainerd
remote implementation has to reload the containerd container, task
and/or process metadata for nearly every operation. This in turn
amplifies the number of context switches between dockerd and containerd
to perform any container operation or handle a containerd event,
increasing the load on the system which could otherwise be allocated to
workloads.
Overhaul the libcontainerd interface to reduce the impedance mismatch
with the containerd client so that the containerd client can be used
more efficiently. Split the API out into container, task and process
interfaces which the consumer is expected to retain so that
libcontainerd can retain state---especially the analogous containerd
client objects---without having to manage any state-store inside the
libcontainerd client.
Signed-off-by: Cory Snider <csnider@mirantis.com>
Add an integration test to verify that health checks are killed on
timeout and that the output is captured.
Co-authored-by: Nicolas De Loof <nicolas.deloof@gmail.com>
Signed-off-by: Cory Snider <csnider@mirantis.com>
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Terminating the exec process when the context is canceled has been
broken since Docker v17.11 so nobody has been able to depend upon that
behaviour in five years of releases. We are thus free from backwards-
compatibility constraints.
Co-authored-by: Nicolas De Loof <nicolas.deloof@gmail.com>
Co-authored-by: Sebastiaan van Stijn <github@gone.nl>
Signed-off-by: Nicolas De Loof <nicolas.deloof@gmail.com>
Signed-off-by: Cory Snider <csnider@mirantis.com>
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Now client have the possibility to set the console size of the executed
process immediately at the creation. This makes a difference for example
when executing commands that output some kind of text user interface
which is bounded by the console dimensions.
Signed-off-by: Paweł Gronowski <pawel.gronowski@docker.com>
Starting an exec can take a significant amount of time while under heavy
container operation load. In extreme cases the time to start the process
can take upwards of a second, which is a significant fraction of the
default health probe timeout (30s). With a shorter timeout, the exec
start delay could make the difference between a successful probe and a
probe timeout! Mitigate the impact of excessive exec start latencies by
only starting the probe timeout timer after the exec'ed process has
started.
Add a metric to sample the latency of starting health-check exec probes.
Signed-off-by: Cory Snider <csnider@mirantis.com>
As reported in docker/compose#6445, when deploying a Linux
container on Windows (LCOW), the daemon made the wrong assumption
when deciding which shell to use to execute the healthcheck, looking
at the host's platform instead of the container's platform.
This patch adds a check for the container's platform when deploying
on Windows, and sets the correct shell.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
`time.After` keeps a timer running until the specified duration is
completed. It also allocates a new timer on each call. This can wind up
leaving lots of uneccessary timers running in the background that are
not needed and consume resources.
Instead of `time.After`, use `time.NewTimer` so the timer can actually
be stopped.
In some of these cases it's not a big deal since the duraiton is really
short, but in others it is much worse.
Signed-off-by: Brian Goff <cpuguy83@gmail.com>
Since Go 1.7, context is a standard package. Since Go 1.9, everything
that is provided by "x/net/context" is a couple of type aliases to
types in "context".
Many vendored packages still use x/net/context, so vendor entry remains
for now.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
Adds a mutex to protect the status, as well. When running the race
detector with the unit test, we can see that the Status field is written
without holding this lock. Adding a mutex to read and set status
addresses the issue.
Signed-off-by: Stephen J Day <stephen.day@docker.com>
Replicate relevant mutations to the in-memory ACID store. Readers will
then be able to query container state without locking.
Signed-off-by: Fabio Kung <fabio.kung@gmail.com>
