Moved the initialization of the SETTINGS frame to a common place
shared by both the low-level HTTP2Client and the higher level
HttpClientTransportOverHTTP2.
Did the same with HTTP2 server ConnectionFactories.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
* Issue #2439 - Remove HTTP/2 data copy.
Implemented reference counting for the network buffer, with the
semantic that calling succeeded() on callbacks decrements the
reference count.
Introduced interface Retainable, used by the client when notifying
multiple application content listeners.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Fixed by start reading from the server after the session
promise has been notified, which may send the first request.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Updated the logic to avoid sending an empty data frame
when only sending the trailers in HttpTransportOverHTTP2.send().
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
* Fixes#2366 - Review HTTP/2 interleaving.
Now using a fairer algorithm that performs multiple generation passes
but without generating all frames for the current entries - it stops
after N passes over all the entries when the number of bytes to
write is large enough.
Using HttpConfiguration.outputBufferSize as the writeThreshold on
the server and a default 32 KiB on the client to stop the generation
after a pass over all the entries.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Removed usage of ternary expressions in favor of if/else statements
to improve readability of the logic for the send() method.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
The problem was caused by the reset arriving to the server
_before_ the commit callback was invoked.
Now waiting for the commit callback to complete before
sending the reset to the server.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Changed the signature of HttpParser.Listener.onBadMessage()
to take a BadMessageException and updated dependent code.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Dumping also HttpChannelOverHTTP2 when dumping the HTTP2Stream.
The channel is now stored as a field in HTTP2Stream rather than
as an attribute to save the attributes map allocation.
Consequently, IStream has now getter/setter for the property
"attachment" that is used to store the channel without having
a type dependency on HttpChannel.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Merging this PR for #2199 as we currently think it is not worse than previous and the code base is definitely simpler. While JMH has shown significant benefits for this approach, we have yet been able to demonstrate them in full scale integration tests - however that is likely due to GC and thread pool issues dominating.
A `org.eclipse.jetty.http2.PEC_MODE` System property has been added to allow the EWYK scheduling to be disabled for HTTP/2 if need be.
* Implementation of #1803 proposal 2 - EITHER dispatch type for EWYK
* made code more readable
* increase small threadpools in tests for extra reserved thread
* clean up
* minor code simplifications
* Work in progress to simplify reserved thread pool
* use a single ReservedThreadPool built into the QueuedThreadPool
* fixed javadoc
* removed more old reserved thread references
* disable EWYK for h2
* fixes from review
* use EWYK for HTTP2
* Fixed javadocs, imports and QueuedThreadPool constructors.
* fix javadoc
* EWYK avoid unnecessary pendings
* after review
* fixed merge with jetty-threadpool.xml
* alternate EWYK implementations
* added jetty copyright headers
* Simplified EWYK code after review
* fixed bad merge
* Code cleanups.
* Improved Javadocs for deprecated property "reservedThreads".
* Improved Javadocs for deprecated property "reservedThreads".
* added a system property to enable only PEC for HTTP/2
Signed-off-by: Greg Wilkins <gregw@webtide.com>
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Issue #1803 - Review usage of scheduling strategies
Use a single ReservedThreadExecutor built into the QueuedThreadPool
via new interface TryExecutor.
Signed-off-by: Greg Wilkins <gregw@webtide.com>
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Issue #2081 No idle timeout exception when dispatch is delayed
* Delegate the readtimeout handling to HttpChannel so that a delayed dispatch can be ended.
* Added unit test for delayed dispatch idle
* Now using HttpInput.onIdleTimeout() to fail the HttpInput, and then dispatching the request in case it has not been dispatched yet. This ensure consistent behavior independently of the value of HttpConfiguration.delayDispatchUntilContent.
* Fixed for both HTTP/1.1 and HTTP/2.
* Added tests for non-blocking reads.
Signed-off-by: Greg Wilkins <gregw@webtide.com>
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
* Code cleanups.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
* Improved test case handler.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
* Improved exception message.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
* Issue #1973 - Implement minimum response data rate.
Implemented response content data rate control in HttpOutput.
Introduced a WriteFlusher.Listener interface that produces events
for every flush(). These events are forwarded to the Connection
and from there to the HttpOutput so that the data rate control can
be enforced.
Both HTTP/1.1 and HTTP/2 are implemented.
Data rate control for HTTP/1.1 is approximate because it will count
also headers bytes and the chunk bytes, while for HTTP/2 is precise.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
* Issue #1973 - Implement minimum response data rate.
Addressed review comments.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Now waking up the flusher via iterate() after a reset has been received.
This ensures that frames that may have stalled are removed from the
flusher queue.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Because now the failures are asynchronous, code that was executed after
invoking the failure listener must be now executed after the
asynchronous processing done by the listener and therefore Callbacks
are introduced.
Idle timeout have a special meaning in that they become a no-operation
if the application is dispatched but idle (neither reading nor writing).
HttpChannelOverHTTP2 now forwards the idle timeout to HttpInput, which
will only change its state if it is interested in reading.
HttpInput.consumeAll() has been modified to consume all input even if
it's already failed.
Failures caused by the other peer (e.g. I/O failures or stream resets)
are now retained and will eagerly consumed any queued data to free up
the flow control windows.
Implemented support in the generic HttpChannel and Request classes.
Linked HTTP/1.1 and HTTP/2 trailers to call HttpChannel, so that trailers
are now available in Servlet APIs by casting to Jetty's Request class.
The semantic is that trailers will only be available _after_ reading
the request content.
It's common for a client to either close the connection or reset a
stream, so EofException is a better exception to report since it will
be logged at DEBUG level by the server, reducing the noise in the logs.
The fix notifies the transport when a reset frame is received,
allowing the transport to fail the write callback which then notifies
the application, either by throwing (in case of blocking writes) or
by calling error listeners.
Also added a guard, in HttpChannel.handle() for the ERROR_DISPATCH case,
that checks if the response is already committed, and if so, abort
the transport - similar to what's already there for 9.4.
* Issue #824 - Implement notifications of asynchronous error conditions for HTTP/2.
Introduced new method HttpChannelState.asyncError() to be called in
case of asynchronous errors, i.e. those errors that do not happen in
the HttpChannel.handle() loop.
Implemented HTTP/2 callbacks to call HttpChannelState.asyncError()
and plug in the existing error handling mechanism.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
* Issue #824 - Implement notifications of asynchronous error conditions for HTTP/2.
Improved implementation to ignore idle timeouts for streams and
session in case that requests are being handled, matching the HTTP/1.1
behavior.
Signed-off-by: Simone Bordet <simone.bordet@gmail.com>
Defined missing status codes (removed old/wrong javadoc)
Changed use of 413 to 431
Received SETTINGS_MAX_HEADER_LIST_SIZE is set on hpack encoded and warning generated only for large responses.
Reverted to warning if a small SETTINGS_MAX_HEADER_LIST_SIZE. is received.
Setting it on the httpConfig effects all connections, not just the one the
setting was received from.
Fixed by recycling only channels that have completed normally by
having read the request content and responded.
Channels that don't read the request content won't be recycled, thus
avoiding the NPE.
* Made ProduceExecuteConsume the default ExecutionFactory for HTTP/2.
* Made the HTTP/2 fillable callback non-blocking.
* Introduced configuration for the server initial session recv window.
* Sending a WINDOW_UPDATE frame at session setup to inform the client
about the server session recv window.
This is necessary because the HTTP/1.1 machinery will recycle the
Metadata.Request object, and it may race with a thread dispatch to
the HTTP/2 machinery that will handle the request.
Simone Bordet
WalkerWatch
Greg Wilkins
Joakim Erdfelt
John Gardiner Myers
Jesse McConnell