This commit reorders a maybe die check and a logging statement for the
following reasons:
- we should die as quickly as possible if the cause is fatal
- we do not want the JVM to be so broken that when we try to log
another exception is thrown (maybe another out of memory exception)
and then the maybe die is never invoked
- maybe die will log the cause anyway if the cause is fatal so we only
need to log if the cause is not fatal
There is a group of five settings relating to raw tcp configurations
(no_delay, buffer sizes, etc) that we have for the http transport. These
currently live in the netty module. As they are unrelated to netty
specifically, this commit moves these settings to the
`HttpTransportSettings` class in core.
Right now it is possible for the `HttpPipeliningHandler` to queue
pipelined responses. On channel close, we do not clear and release these
responses. This commit releases the responses and completes the promise.
With this commit we remove the following three previously unused
(and undocumented) Netty 4 related settings:
* transport.netty.max_cumulation_buffer_capacity,
* transport.netty.max_composite_buffer_components and
* http.netty.max_cumulation_buffer_capacity
from Elasticsearch.
Currently when we close a channel in Netty4Utils.closeChannels we
block until the closing is complete. This introduces the possibility
that a network selector thread will block while waiting until a
separate network selector thread closes a channel.
For instance: T1 closes channel 1 (which is assigned to a T1 selector).
Channel 1's close listener executes the closing of the node. That
means that T1 now tries to close channel 2. However, channel 2 is
assigned to a selector that is running on T2. T1 now must wait until T2
closes that channel at some point in the future.
This commit addresses this by adding a boolean to closeChannels
indicating if we should block on close. We only set this boolean to true
if we are closing down the server channels at shutdown. This call is
never made from a network thread. When we call the closeChannels method
with that boolean set to false, we do not block on close.
Transport profiles unfortunately have never been validated. Yet, it's very
easy to make a mistake when configuring profiles which will most likely stay
undetected since we don't validate the settings but allow almost everything
based on the wildcard in `transport.profiles.*`. This change removes the
settings subset based parsing of profiles but rather uses concrete affix settings
for the profiles which makes it easier to fall back to higher level settings since
the fallback settings are present when the profile setting is parsed. Previously, it was
unclear in the code which setting is used ie. if the profiles settings (with removed
prefixes) or the global node setting. There is no distinction anymore since we don't pull
prefix based settings.
This commit makes the use of the global network settings explicit instead
of implicit within NetworkService. It cleans up several places where we fall
back to the global settings while we should have used tcp or http ones.
In addition this change also removes unnecessary settings classes
These settings have not be working for a full major version since they
are not registered. Given that they are simply duplicates we can just remove
them.
Today TcpTransport is the de-facto base-class for transport implementations.
The need for all the callbacks we have in TransportServiceAdaptor are not necessary
anymore since we can simply have the logic inside the base class itself. This change
moves the stats metrics directly into TcpTransport removing the need for low level
bytes send / received callbacks.
Today if a channel gets closed due to a disconnect we notify the response
handler that the connection is closed and the node is disconnected. Unfortunately
this is not a complete solution since it only works for published connections.
Connections that are unpublished ie. for discovery can indefinitely hang since we
never invoke their handers when we get a failure while a user is waiting for
the response. This change adds connection tracking to TcpTransport that ensures
we are notifying the corresponding connection if there is a failure on a channel.
Shared settings were added intially to allow the few common settings
names across aws plugins. However, in 6.0 these settings have been
removed. The last use was in netty, but since 6.0 also has the netty 3
modules removed, there is no longer a need for the shared property. This
commit removes the shared setting property.
Netty removed a logging guarded we added to prevent a scary logging
message. We added a hack to work around this. They've added the guard
back, so we can remove the hack now.
Today we prune transport handlers in TransportService when a node is disconnected.
This can cause connections to starve in the TransportService if the connection is
opened as a short living connection ie. without sharing the connection to a node
via registering in the transport itself. This change now moves to pruning based
on the connections cache key to ensure we notify handlers as soon as the connection
is closed for all connections not just for registered connections.
Relates to #24632
Relates to #24575
Relates to #24557
In pre-release versions of Elasticsearch 5.0.0, users were subject to
log messages of the form "your platform does not.*reliably.*potential
system instability". This is because we disable Netty from being unsafe,
and Netty throws up this scary info-level message when unsafe is
unavailable, even if it was unavailable because the user requested that
it be unavailabe. Users were rightly confused, and concerned. So, we
contributed a guard to Netty to prevent this log message from showing up
when unsafe was explicitly disabled. This guard shipped with all
versions of Netty that shipped starting with Elasticsearch
5.0.0. Unfortunately, this guard was lost in an unrelated refactoring
and now with the 4.1.10.Final upgrade, users will again see this
message. This commit is a hack around this until we can get a fix
upstream again.
Relates #24469
Netty uses the number of processors for sizing various resources (e.g.,
thread pools, buffer pools, etc.). However, it uses the runtime number
of available processors which might not match the configured number of
processors as set in Elasticsearch to limit the number of threads (for
example, in Docker containers). A new feature was added to Netty that
enables configuring the number of processors Netty should see for sizing
this various resources. This commit takes advantage of this feature to
set this number of available processors to be equal to the configured
number of processors set in Elasticsearch.
Relates #24420
It looks like auto-complete gave us a nasty surprise here with
Logger#equals being invoked instead of Logger#error swallowing the
absolute worst-possible level of a log message. This commit fixes the
invocation.
This commit makes closing a ReleasableBytesStreamOutput release the underlying BigArray so
that we can use try-with-resources with these streams and avoid leaking memory by not returning
the BigArray. As part of this change, the ReleasableBytesStreamOutput adds protection to only
release the BigArray once.
In order to make some of the changes cleaner, the ReleasableBytesStream interface has been
removed. The BytesStream interface is changed to a abstract class so that we can use it as a
useable return type for a new method, Streams#flushOnCloseStream. This new method wraps a
given stream and overrides the close method so that the stream is simply flushed and not closed.
This behavior is used in the TcpTransport when compression is used with a
ReleasableBytesStreamOutput as we need to close the compressed stream to ensure all of the data
is written from this stream. Closing the compressed stream will try to close the underlying stream
but we only want to flush so that all of the written bytes are available.
Additionally, an error message method added in the BytesRestResponse did not use a builder
provided by the channel and instead created its own JSON builder. This changes that method to use
the channel builder and in turn the bytes stream output that is managed by the channel.
Note, this commit differs from 6bfecdf921a1941b48273d76551872df4062cfae in that it updates
ReleasableBytesStreamOutput to handle the case of the BigArray decreasing in size, which changes
the reference to the BigArray. When the reference is changed, the releasable needs to be updated
otherwise there could be a leak of bytes and corruption of data in unrelated streams.
This reverts commit afd45c14327cd0f8d155e5ac9740f48e8e39b09c, which reverted #23572.
This commit makes closing a ReleasableBytesStreamOutput release the underlying BigArray so
that we can use try-with-resources with these streams and avoid leaking memory by not returning
the BigArray. As part of this change, the ReleasableBytesStreamOutput adds protection to only release the BigArray once.
In order to make some of the changes cleaner, the ReleasableBytesStream interface has been
removed. The BytesStream interface is changed to a abstract class so that we can use it as a
useable return type for a new method, Streams#flushOnCloseStream. This new method wraps a
given stream and overrides the close method so that the stream is simply flushed and not closed.
This behavior is used in the TcpTransport when compression is used with a
ReleasableBytesStreamOutput as we need to close the compressed stream to ensure all of the data
is written from this stream. Closing the compressed stream will try to close the underlying stream
but we only want to flush so that all of the written bytes are available.
Additionally, an error message method added in the BytesRestResponse did not use a builder
provided by the channel and instead created its own JSON builder. This changes that method to use the channel builder and in turn the bytes stream output that is managed by the channel.
This commit changes the listener passed to sendMessage from a Runnable
to a ActionListener.
This change also removes IOException from the sendMessage signature.
That signature is misleading as it allows implementers to assume an
exception will be thrown in case of failure. That does not happen due
to Netty's async nature.
This commit upgrades the Netty dependencies from version 4.1.8 to
version 4.1.9. This commit picks up a few bug fixes that impacted us:
- Netty was incorrectly ignoring interfaces with self-assigned MAC
addresses (e.g., instances running in Docker containers or on EC2)
- incorrect handling of the Expect: 100-continue header
Relates #23540
With this commit we change the default receive predictor size for Netty
from 32kB to 64kB as our testing has shown that this leads to less
allocations on smaller heaps like the default out of the box
configuration and this value also works reasonably well for larger
heaps.
Closes#23185
Previously, the RestController would stash the context prior to copying headers. However, there could be deprecation
log messages logged and in turn warning headers being added to the context prior to the stashing of the context. These
headers in the context would then be removed from the request and also leaked back into the calling thread's context.
This change moves the stashing of the context to the HttpTransport so that the network threads' context isn't
accidentally populated with warning headers and to ensure the headers added early on in the RestController are not
excluded from the response.
This commit sets the intial size of the pipeline handler queue small to
prevent waste if pipelined requests are never sent. Since the queue will
grow quickly if pipeline requests are indeed set, this should not be
problematic.
Relates #23335
When pipelined responses are sent to the pipeline handler for writing,
they are not necessarily written immediately. They must be held in a
priority queue until all responses preceding the given response are
written. This means that when write is invoked on the handler, the
promise that is attached to the write invocation will not necessarily be
the promise associated with the responses that are written while the
queue is drained. To address this, the promise associated with a
pipelined response must be held with the response and then used when the
channel context is actually written to. This was introduced when
ensuring that the releasing promise is always chained through on write
calls lest the releasing promise never be invoked. This leads to many
failing test cases, so no new test cases are needed here.
Relates #23317
When sending a response to a client, we attach a releasing listener to
the channel promise. If the client disappears before the response is
sent, the releasing listener was never notified. The reason the
listeners were never notified was due to a mistaken invocation of write
and flush on the channel which has two overrides: one that takes an
existing promise, and one that does not and instead creates a new
promise. When the client disappears, it is this latter promise that is
notified, which does not contain the releasing listener. This commit
addreses this issue by invoking the override that passes our channel
promise through.
Relates #23310
Previously we calculated Netty' receive predictor size for HTTP and transport
traffic based on available memory and worker nodes. This resulted in a receive
predictor size between 64kb and 512kb. In our benchmarks this leads to increased
GC pressure.
With this commit we set Netty's receive predictor size to 32kb. This value is in
a sweet spot between heap memory waste (-> GC pressure) and effect on request
metrics (achieved throughput and latency numbers).
Closes#23185
When Netty decodes a bad HTTP request, it marks the decoder result on
the HTTP request as a failure, and reroutes the request to GET
/bad-request. This either leads to puzzling responses when a bad request
is sent to Elasticsearch (if an index named "bad-request" does not exist
then it produces an index not found exception and otherwise responds
with the index settings for the index named "bad-request"). This commit
addresses this by inspecting the decoder result on the HTTP request and
dispatching the request to a bad request handler preserving the initial
cause of the bad request and providing an error message to the client.
Relates #23153
Netty 4.1.8 wraps connect and accept operations in doPrivileged blocks.
This means that we not need to give permissions to the entire transport
module. Additionally this commit deletes the privileged socket channel
and privileged server socket chanel.
This is related to #22116. Core no longer needs `SocketPermission`
`connect`.
This permission is relegated to these modules/plugins:
- transport-netty4 module
- reindex module
- repository-url module
- discovery-azure-classic plugin
- discovery-ec2 plugin
- discovery-gce plugin
- repository-azure plugin
- repository-gcs plugin
- repository-hdfs plugin
- repository-s3 plugin
And for tests:
- mocksocket jar
- rest client
- httpcore-nio jar
- httpasyncclient jar
This commit upgrades the checkstyle configuration from version 5.9 to
version 7.5, the latest version as of today. The main enhancement
obtained via this upgrade is better detection of redundant modifiers.
Relates #22960
This change adds a strict mode for xcontent parsing on the rest layer. The strict mode will be off by default for 5.x and in a separate commit will be enabled by default for 6.0. The strict mode, which can be enabled by setting `http.content_type.required: true` in 5.x, will require that all incoming rest requests have a valid and supported content type header before the request is dispatched. In the non-strict mode, the Content-Type header will be inspected and if it is not present or not valid, we will continue with auto detection of content like we have done previously.
The content type header is parsed to the matching XContentType value with the only exception being for plain text requests. This value is then passed on with the content bytes so that we can reduce the number of places where we need to auto-detect the content type.
As part of this, many transport requests and builders were updated to provide methods that
accepted the XContentType along with the bytes and the methods that would rely on auto-detection have been deprecated.
In the non-strict mode, deprecation warnings are issued whenever a request with body doesn't provide the Content-Type header.
See #19388
This commit adds a SpecialPermission constant and uses that constant
opposed to introducing new instances everywhere.
Additionally, this commit introduces a single static method to check that
the current code has permission. This avoids all the duplicated access
blocks that exist currently.
This is related to #22116. Core no longer needs SocketPermission
accept. This permission is relegated to the transport-netty4 module
and (for tests) to the mocksocket jar.
Previously, certain settings that could take multiple comma delimited
values would pick up incorrect values for all entries but the first if
each comma separated value was followed by a whitespace character. For
example, the multi-value "A,B,C" would be correctly parsed as
["A", "B", "C"] but the multi-value "A, B, C" would be incorrectly parsed
as ["A", " B", " C"].
This commit allows a comma separated list to have whitespace characters
after each entry. The specific settings that were affected by this are:
cluster.routing.allocation.awareness.attributes
index.routing.allocation.require.*
index.routing.allocation.include.*
index.routing.allocation.exclude.*
cluster.routing.allocation.require.*
cluster.routing.allocation.include.*
cluster.routing.allocation.exclude.*
http.cors.allow-methods
http.cors.allow-headers
For the allocation filtering related settings, this commit also provides
validation of each specified entry if the filtering is done by _ip,
_host_ip, or _publish_ip, to ensure that each entry is a valid IP
address.
Closes#22297
Today we have quite some abstractions that are essentially providing a simple
dispatch method to the plugins defining a `HttpServerTransport`. This commit
removes `HttpServer` and `HttpServerAdaptor` and introduces a simple `Dispatcher` functional
interface that delegate to `RestController` by default.
Relates to #18482
This is related to #22116. netty channels require socket `connect` and
`accept` privileges. Netty does not currently wrap these operations
with `doPrivileged` blocks. These changes extend the netty channels
and wrap calls to the relevant super methods in doPrivileged blocks.
