The refCounter used for sharing temporary decompressed log files and
tracking when the files can be deleted is keyed off the source file's
path. But the path of a log file is not stable: it is renamed on each
rotation. Consequently, when logging is configured with both rotation
and compression, multiple concurrent readers of a container's logs could
read logs out of order, see duplicates or decompress a log file which
has already been decompressed.
Replace refCounter with a new implementation, sharedTempFileConverter,
which is agnostic to the file path, keying off the source file's
identity instead. Additionally, sharedTempFileConverter handles the full
lifecycle of the temporary file, from creation to deletion. This is all
abstracted from the consumer: all the bookkeeping and cleanup is handled
behind the scenes when Close() is called on the returned reader value.
Only one file descriptor is used per temporary file, which is shared by
all readers.
A channel is used for concurrency control so that the lock can be
acquired inside a select statement. While not currently utilized, this
makes it possible to add support for cancellation to
sharedTempFileConverter in the future.
Signed-off-by: Cory Snider <csnider@mirantis.com>
Truncating the current log file while a reader is still reading through
it results in log lines getting missed. In contrast, rotating the file
allows readers who have the file open can continue to read from it
undisturbed. Rotating frees up the file name for the logger to create a
new file in its place. This remains true even when max-file=1; the
current log file is "rotated" from its name without giving it a new one.
On POSIXy filesystem APIs, rotating the last file is straightforward:
unlink()ing a file name immediately deletes the name from the filesystem
and makes it available for reuse, even if processes have the file open
at the time. Windows on the other hand only makes the name available
for reuse once the file itself is deleted, which only happens when no
processes have it open. To reuse the file name while the file is still
in use, the file needs to be renamed. So that's what we have to do:
rotate the file to a temporary name before marking it for deletion.
Signed-off-by: Cory Snider <csnider@mirantis.com>
The json-file driver appends a newline character to log messages with
PLogMetaData.Last set, but the local driver did not. Alter the behavior
of the local driver to match that of the json-file driver.
Signed-off-by: Cory Snider <csnider@mirantis.com>
The LogFile follower would stop immediately upon the producer closing.
The close signal would race the file watcher; if a message were to be
logged and the logger immediately closed, the follower could miss that
last message if the close signal (formerly ProducerGone) was to win the
race. Add logic to perform one more round of reading when the producer
is closed to catch up on any final logs.
Signed-off-by: Cory Snider <csnider@mirantis.com>
Whether or not the logger has been closed is a property of the logger,
and only of concern to its log reading implementation, not log watchers.
The loggers and their reader implementations can communicate as they see
fit. A single channel per logger which is closed when the logger is
closed is plenty sufficient to broadcast the state to log readers, with
no extra bookeeping or synchronization required.
Signed-off-by: Cory Snider <csnider@mirantis.com>
The asynchronous startup of the log-reading goroutine made the
follow-tail tests nondeterministic. The Log calls in the tests which
were supposed to happen after the reader started reading would sometimes
execute before the reader, throwing off the counts. Tweak the ReadLogs
implementation so that the order of operations is deterministic.
Signed-off-by: Cory Snider <csnider@mirantis.com>
Add an extensive test suite for validating the behavior of any
LogReader. Test the current LogFile-based implementations against it.
Signed-off-by: Cory Snider <csnider@mirantis.com>
The jsonfilelog read benchmark was incorrectly reusing the same message
pointer in the producer loop. The message value would be reset after the
first call to jsonlogger.Log, resulting in all subsequent calls logging
a zero-valued message. This is not a representative workload for
benchmarking and throws off the throughput metric.
Reduce variation between benchmark runs by using a constant timestamp.
Write to the producer goroutine's error channel only on a non-nil error
to eliminate spurious synchronization between producer and consumer
goroutines external to the logger being benchmarked.
Signed-off-by: Cory Snider <csnider@mirantis.com>
On Linux the daemon was not respecting the HostConfig.ConsoleSize
property and relied on cli initializing the tty size after the container
was created. This caused a delay between container creation and
the tty actually being resized.
This is also a small change to the api description, because
HostConfig.ConsoleSize is no longer Windows-only.
Signed-off-by: Paweł Gronowski <pawel.gronowski@docker.com>
Similar to the (now removed) `apparmor` build tag, this build-time toggle existed for users who needed to build without the `libseccomp` library. That's no longer necessary, and given the importance of seccomp to the overall default security profile of Docker containers, it makes sense that any binary built for Linux should support (and use by default) seccomp if the underlying host does.
Signed-off-by: Tianon Gravi <admwiggin@gmail.com>
strings.ReplaceAll(s, old, new) is a wrapper function for
strings.Replace(s, old, new, -1). But strings.ReplaceAll is more
readable and removes the hardcoded -1.
Signed-off-by: Eng Zer Jun <engzerjun@gmail.com>
- remove isErrNoSuchProcess() in favor of a plain errors.As()
- errNoSuchProcess.Error(): remove punctuation
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
This allows the postContainersKill() handler to pass values as-is. As part of
the rewrite, I also moved the daemon.GetContainer(name) call later in the
function, so that we can fail early if an invalid signal is passed, before
doing the (heavier) fetching of the container.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
The libtrust trust-key is only used for pushing legacy image manifests;
pushing these images has been deprecated, and we only need to be able
to push them in our CI.
This patch disables generating the trust-key (and related paths) unless
the DOCKER_ALLOW_SCHEMA1_PUSH_DONOTUSE env-var is set (which we do in
our CI).
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
This change is in preparation of deprecating support for old manifests.
Currently the daemon's ID is based on the trust-key ID, which will be
removed once we fully deprecate support for old manifests (the trust
key is currently only used in tests).
This patch:
- looks if a trust-key is present; if so, it migrates the trust-key
ID to the new "engine-id" file within the daemon's root.
- if no trust-key is present (so in case it's a "fresh" install), we
generate a UUID instead and use that as ID.
The migration is to prevent engines from getting a new ID on upgrades;
while we don't provide any guarantees on the engine's ID, users may
expect the ID to be "stable" (not change) between upgrades.
A test has been added, which can be ran with;
make DOCKER_GRAPHDRIVER=vfs TEST_FILTER='TestConfigDaemonID' test-integration
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Make sure we validate the default address given before using it, and
combine the parsing/validation logic so that it can be reused.
This patch also makes the errors more consistent, and uses pkg/errors
for generating them.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Use the default (0) value to indicate "not set", which simplifies
working with these configuration options, preventing the need to
use intermediate variables etc.
While changing this code, also making some small cleanups, such
as replacing "fmt.Sprintf()" for "strconv" variants.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
This method returned the network controller, only to set it on the daemon.
While making this change, also;
- update some error messages to be in the correct format
- use errors.Wrap() where possible
- extract configuring networks into a separate function to make the flow
slightly easier to follow.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
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>
reloadMaxDownloadAttempts() is used to reload the configuration,
but validation happened before merging the config with the defaults.
This removes the validation from this function, instead centralizing
validation in config.Validate().
NOTE:
Currently this validation is "ok", as it checks for "nil" values;
I am working on changes to reduce the use of pointers in the config,
and instead provide a mechanism to fill in defaults. This change is
in preparation of that.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
The Reload logic is problematic and needs a rewrite.
Currently, config.Reload() is validating newConfig before the reload callback
is executed. At that point, newConfig may be a partial configuration, yet to be
merged with the existing configuration (in the "reload()" callback). Validating
this config before it's merged can result in incorrect validation errors.
However, the current "reload()" callback we use is DaemonCli.reloadConfig(),
which includes a call to Daemon.Reload(), which both performs "merging" and
validation, as well as actually updating the daemon configuration. Calling
DaemonCli.reloadConfig() *before* validation, could thus lead to a failure in
that function (making the reload non-atomic).
While *some* errors could always occur when applying/updating the config, we
should make it more atomic, and;
1. get (a copy of) the active configuration
2. get the new configuration
3. apply the (reloadable) options from the new configuration
4. validate the merged results
5. apply the new configuration.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
MergeDaemonConfigurations was validating the configs before and after
merging. However, the "fileConfig" configuration may contain only a
"partial" configuration (options to apply to / override the existing
config). This means that some options may not be set and contain default
or empty values.
Validating such partial configurations can produce validation failures,
so to prevent those, we should validate the configuration _after_
merging, to validate the "final" state.
There's more cleaning up / improvements to be made in this area; for
example, we currently use our "self crafted" `getConflictFreeConfiguration()`
function, which is used to detect options that are not allowed to
be overridden, and which could potentially be handled by mergo.Merge(),
but leaving those changes for a future exercise.
This patch removes the first validation step, changing the function
to only validate the resulting configuration after merging.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
Log-level validation was previously performed when configuring the daemon-logs;
this moves the validation to config.Validate() so that we can catch invalid
settings when running dockerd --validate.
Signed-off-by: Sebastiaan van Stijn <github@gone.nl>
