Motivation:
We should better use JNI to lookup constants so we are sure we not mess
things up
Modifications:
Use JNI calls to lookup constants once
Result:
Less error prone code
Motivation:
We should submit multiple IO ops at once to reduce the syscall overhead.
Modifications:
- Submit multiple IO ops in batches
- Adjust default ringsize
Result:
Much better performance
Motivation:
To ensure we use the correct values when passing values from Java to C and the other way around it is better to use JNI to lookup the values.
Modifications:
Add NativeStaticallyRefererencedJniMethods and use it (just as we do in kqueue / epoll)
Results:
More robust code
Motivation:
We did have a bug in how we calculated the values for the timespec which lead to incorrect wakeups. Beside this we also missed to always call runAllTasks() which is needed to fetch the ready to be executed scheduled tasks.
Modifications:
- Fix timespec setup
- Always call runAllTasks()
- Add extra testcase
- Remove @Ignore from previous failing test
Result:
Timeouts work as expected
Motivation:
We need to respect RecvByteBufAllocator.Handle.continueReading() so settings like MAX_MESSAGES_PER_READ are respected. This also ensures that AUTO_READ is correctly working in all cases
Modifications:
- Correctly respect continueReading();
- Fix IOUringRecvByteAllocatorHandle
- Cleanup
Result:
Correctly handling reading
Motivation:
We did not correctly compute all fields when POLL_REMOVE entry was calculate. Which could lead to not finding the right operation.
Modifications:
- Correctly fill all fields
- Fix unit tests
Result:
Remove IO_POLL operations work again as expected
Motivation:
We not correctly handled errors and also had some problems with home POLL* was handled.
Modifictions:
- Cleanup
- No need to for links anymore
- Add error handling for most operations (poll still missing)
- Add better handling for RDHUP
- Correctly handle writeScheduled flag for writev
Result:
Cleaner and more correct code
Motivation:
writev which allows to write data into multiple buffers
Modification:
-Added iovec array pool to manage iov memory
-flush override to make sure that write is not called
Result:
performance is much better
Motivation:
we should remove pollIn link, as we don't use pollIn linking anymore
Modification:
-some cleanups in the tests and in IOUring
-pollIn linking was removed
Result:
clean code
Motivation:
We must correctly use the polling support of io_uring to reduce the number of events in flight + only allocate buffers if really needed. For this we should respect the different poll masks and only do the corresponding IO action once the fd becomes ready for it.
Modification:
- Correctly respect poll masks and so only schedule an IO event if the fd is ready for it
- Move some code for cleanup
Result:
More correct usage of io_uring and less memory usage
Motivation:
if connect returns EINPROGRESS we send POLL OUT and check
via socket.finishConnect if the connection is successful
Modifications:
-added new io_uring connect op
-use a directbuffer for the socket address
Result:
you are able to connect to a peer
Motivation:
when you submit a poll, io_uring still hold reference to it even if close is called
source io_uring mailing list(https://lore.kernel.org/io-uring/27657840-4E8E-422D-93BB-7F485F21341A@kernel.dk/T/#t)
Modification:
-To delete fd reference in io_uring, we use POLL_REMOVE to avoid a server socket address error
-Added a POLL_REMOVE test
Result:
server can be closed and restarted
Motivation:
We did create a lot of objects related to the completion queue and submission queue which produced a lot of GC. Beside this we also did maintain an extra map which is not really needed as we can encode everything that we need in the user_data field.
Modification:
- Reduce complexity and GC pressure by store needed informations in the user_data field
- Small refactoring of the code to move channel related logic to the channel
- Remove unused classes
- Use callback to process stuff in the completion queue and so remove all GC created by it
- Simplify by not storing channel and buffer in the event
Result:
Less GC pressure and no extra lookups for events needed
Motivation:
We use eventfd in our io_uring based transport to wakeup the eventloop. When doing so we need to be careful that we read any data previous written to it.
Modification:
- Correctly read data that was written to eventfd before submit another event related to it to the submission queue as otherwise we will see another completion event related to it asap
- Ensure we not remove the wrong event from the storted event ids (we did remove the wrong before because we reused the Event object)
- ensure we only use the submission queue from the EventLoop thread in all cases
- add another unit test
Result:
Wakeups via eventfd work as expected
Motivation:
We need to use deadlineToDelayNanos(...) to calculate the timeout for io_uring as otherwise the timeout will be scheduled at the wrong time in the future
Modifications:
Make use of deadlineToDelayNanos(...)
Result:
Correctly schedule timeou
Motivation:
There was a bug in the implemention so we missed to submit what was in the submission queue. This could lead to a deadlock. Beside this we should also process all events that we can poll without blocking and only after that process tasks. This will ensure we drain the ringbuffers in a timely manner
Modifications:
- Add missing submit() call
- Rename peek() to poll() as we consume the data so peek() is missleading
- Process all events that can be processed without blocking
Result:
Fix a bug, clarify and better performance
Motivation:
-at the moment we dont shutdown when we get a read error message
-missing autoread support
Modifications:
-even if autoread is disable, should do check if the read event is already submitted
-added new Handle exception method to shutdown the channels
Result:
EL read event can handle read errors
Motivation:
no checks for non writeable sockets
Modifications:
-Added a linked write poll to make sure that the socket does not write if it is not writeable
-Added a new boolean to avoid to submit a second write operation
Result:
writeable socket check
Motivation:
when the channel connection is lost, we dont get any notification(unless the customer has not submitted a writer or read event)
Modifications:
add rhup polling to be notified when the connection is lost
Result:
the eventloop is notified on connection losts
Motivation:
The problem is that if io_uring accept/read non blocking doesnt return -EAGAIN for non-blocking sockets
in general, then it removes a way for the application to tell if
there's ever any data there.
There is a fix in Kernel 5.8 https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v5.8&id=e697deed834de15d2322d0619d51893022c90ea2 which means we need to add poll before the accept/read event(poll<link>read/accept) to fix in netty as well
Modification:
-add poll before the accept/read event with this flag IOSQE_IO_LINK
Result:
netty prototype works on Kernel 5.8
Motivation:
wake up the blocking call io_uring which is called when a new task is added(not from the eventloop thread)
Modification:
-Added timeout operation for scheduled task(not tested yet)
-Added Poll operation
-Added two tests to reproduce the polling signal issue
Result:
io_uring_enter doesnt get any polling signal from eventFdWrite if both functions are executed in different threads
Motivation:
unnecessary use of LinuxSocket class, missing CRLF etc.
Modification:
-Add CRLF
-remove IOUringChannelConfig and LinuxSocket class
Result:
less code and cleanup
Motivation:
fix checkstyle errors and many classes are unnecessarily public
Modification:
-fixed maven checkstyle errors
-using package-private and final classes
Result:
better code quality
Motivation:
missing eventLoop completionQueue logic
Modification:
-mapping between channel and eventloop
-added new prototype blocking example
Result:
First running prototype
Motivation:
-missing channel configs
-we dont store byteBuf and channel for the read operation to execute channel.fireChannelReadComplete
Modifications:
-add channels configs
-new Event properties for the processing completionQueue
Result:
with these changes it will be much easier to implement the eventloop
Motivation:
JNI ring buffer implementation is inflexible and not really efficient
Modifications:
-RingBuffer instance is created in JNI which contains io_uring ring buffer information
-move the JNI ring buffer logic to Java
-added Todos
-using unsafe memory barriers loadFence and storeFence
-revert epoll file
Result:
this java ring buffer implementation is more flexible and efficient
Motivation:
prototype is not buildable and JNI io_uring implementation is missing
Modifications:
-added io_uring implementation(source from https://github.com/axboe/liburing)
-eventloop stores netty io_uring pointer which is used for two ring buffers to execute events like read and write operations in JNI
-memory barriers already included in JNI(will be changed in the future)
-pom file adopted from native epoll
Result:
prototype can finally be built