Support for debugging Fibers has been partially implemented. Fibers will appear to the debugger as ordinary threads. When you view the list of threads from within the debugger, you will see both fibers and the carrier threads they run on.
If a fiber is mounted, both the fiber and its carrier thread are basically the same execution context. If one is at a breakpoint, they both will appear to be at the same breakpoint. They will have the same stack traces, and single stepping in one advances the PC in the other. They are essentially two different views of the same Thread (with some minor exceptions in debugger behavior noted below).
You can tell which carrier thread a fiber is mounted on by the name of the fiber, which will be fiber@<hashcode>[<carrier_thread_name>]
. If it is not mounted, you will instead see fiber@<hashcode>[<no carrier thread>]
. Here is an example debugger threads list with 2 carrier threads and 3 fibers, two of which are mounted:
ForkJoinPool-1-worker-3
ForkJoinPool-1-worker-1
Fiber@5a1fbaa6[<no carrier thread>]
Fiber@50e7d1d6[ForkJoinPool-1-worker-3,CarrierThreads]
Fiber@5e2035ba[ForkJoinPool-1-worker-1,CarrierThreads]
Fiber Stack Traces
Stack traces for fibers will include any scheduler related frames. For example, only the top free frames below belong to the Fiber's implementation. The rest are considered to be scheduler related:
[1] Main.taker (Main.java:29)
[2] Main.lambda$static$2 (Main.java:35)
[3] Main$$Lambda$7.1705736037.run (null)
[4] java.lang.Fiber.lambda$new$0 (Fiber.java:161)
[5] java.lang.Fiber$$Lambda$9.250421012.run (null)
[6] java.lang.Continuation.enter0 (Continuation.java:233)
[7] java.lang.Continuation.enter (Continuation.java:220)
[8] java.lang.Continuation.run (Continuation.java:178)
[9] java.lang.Fiber.runContinuation (Fiber.java:309)
[10] java.lang.Fiber$$Lambda$10.295530567.run (null)
[11] java.util.concurrent.ForkJoinTask$RunnableExecuteAction.exec (ForkJoinTask.java:1,425)
[12] java.util.concurrent.ForkJoinTask.doExec (ForkJoinTask.java:290)
[13] java.util.concurrent.ForkJoinPool$WorkQueue.topLevelExec (ForkJoinPool.java:1,020)
[14] java.util.concurrent.ForkJoinPool.scan (ForkJoinPool.java:1,656)
[15] java.util.concurrent.ForkJoinPool.runWorker (ForkJoinPool.java:1,594)
[16] java.util.concurrent.ForkJoinWorkerThread.run (ForkJoinWorkerThread.java:189)
If the fiber is unmounted, currently the stack trace is a bit confusing because the fiber will temporarily be mounted on a suspended helper thread. So you get something like:
[1] java.lang.Thread.suspend0 (native method)
[2] java.lang.Thread.suspendThread (Thread.java:1,229)
[3] java.lang.Fiber.lambda$tryMountAndSuspend$3 (Fiber.java:1,225)
[4] java.lang.Fiber$$Lambda$14.660017696.run (null)
[5] java.lang.Fiber.maybePark (Fiber.java:529)
[6] java.lang.Fiber.parkNanos (Fiber.java:488)
[7] java.lang.Thread.sleep (Thread.java:364)
[8] Main.taker (Main.java:27)
[9] Main.lambda$static$2 (Main.java:35)
[10] Main$$Lambda$7.1705736037.run (null)
[11] java.lang.Fiber.lambda$new$0 (Fiber.java:161)
[12] java.lang.Fiber$$Lambda$9.250421012.run (null)
[13] java.lang.Continuation.enter0 (Continuation.java:233)
[14] java.lang.Continuation.enter (Continuation.java:220)
[15] java.lang.Continuation.run (Continuation.java:178)
[16] java.lang.Fiber.tryRun (Fiber.java:1,173)
[17] java.lang.Fiber.lambda$tryMountAndSuspend$4 (Fiber.java:1,223)
[18] java.lang.Fiber$$Lambda$13.153425756.run (null)
[19] java.lang.Thread.run (Thread.java:935)
[20] jdk.internal.misc.InnocuousThread.run (InnocuousThread.java:134)
In this case frames 7-10 (and actually also 5-6) are considered fiber frames. The frames above it are just there to support temporarily mounting the fiber. This will be improved in the future.
Single Stepping
If you single step while in the context of a fiber, execution flow will remain in the fiber even if you step over a method call that results in a yield, causing the fiber to be unmounted and later remounted on a different carrier thread. If you are in the context of the carrier thread when the fiber yields, the behavior is not well defined yet, and there are also known issues, especially when single stepping in and around Continuation.yield()
.
Missing Fiber Support
There are a number of debugger features that are known to not work (or likely won't work) if you try them. They include (but are probably not limited to):
- Suspending or resuming a specific fiber. When the debugger asks for all threads to be suspended (which is usually what happens at breakpoints), the fibers will also be considered suspended. However, do not try to suspend a specific fiber, or resume any fiber that was suspended as a result of suspending all threads.
- Setting local variables in fibers (you can view them). This actually applies to fiber frames, even if the current thread is a carrier thread. More specifically, you cannot set locals in continuation frames.
- Dealing with a large number of fibers. For one, you will overwhelm the debugger because it will see them all. Also, fibers are not tracked in an efficient manner by the debug agent, so it can also become bogged down.
- Forcing a fiber frame to return ("Drop Frame" in Intellij), although this should work fine if executed from the carrier thread context instead of the fiber context.
Plans for Different Fiber Debugging Modes
There are plans to be able to pass a flag to the debug agent to control how it supports fibers. Likely options are:
fiberdebug=on
: The debugger will see all fibers (this is what is currently implemented)fiberdebug=off
: The debugger will not see any fibers.fiberdebug=partial
: The debugger will only see interesting fibers. For example, fibers that have reached a breakpoint. This would allow you to single step in a fiber that you hit a breakpoint in, but would not bog the debugger down with a large number of fibers.
Miscellaneous Implementation Details.
For this first version of Fiber debugger support, we decided not to make any changes the JDWP or JDI, and therefore no changes are needed in the debuggers. There are some JVMTI additions that provide the Debug Agent with events to indicate Fiber start/terminate and also mount/unmount. There are also new JVMTI APIs that allow getting the Fiber running on a Carrier Thread (if one is running), and to get the Carrier Thread a Fiber is mounted on (if it is mounted).
For JDWP, JDI, and debuggers, Fibers appear to be ordinary threads, and no changes have been made in these layers of the debugging support. The heavy lifting of Fiber debugger support is in the Debug Agent, which manages the mapping between a Fiber and its Carrier Thread. For example:
- If the Debug Agent gets a BREAKPOINT event from JVMTI on a carrier thread, it will deliver that event to the debugger on the Fiber instead.
- If you single step in a Fiber, this results in the Debug Agent enabling single stepping on the Carrier Thread that the Fiber is running on. When the JVMTI delivers the SINGLE_STEP event to the Debug Agent on the Carrier Thread, it is relayed to the debugger on the Fiber instead of the Carrier Thread.