Duncan Coutts pushed to branch wip/io-manager-deadlock-detection at Glasgow Haskell Compiler / GHC Commits: a23f0d0a by Duncan Coutts at 2026-06-19T08:09:20+01:00 Remove wakeupIOManager, ioManagerWakeup and setIOManagerWakeupFd We no longer need wakeupIOManager for the threaded RTS case, so we can remove it and the bits only needed to support it. This includes the pipe/eventfd fd shared between the RTS and the in-library I/O manager used for waking up the I/O manager thread. The pipe/eventfd still exists, but it no longer has to be communicated to the RTS, since the RTS no longer needs to use it. So we remove the RTS API export setIOManagerWakeupFd, and remove uses of it within the I/O managers in ghc-internal. - - - - - f4c88345 by Duncan Coutts at 2026-06-19T08:09:20+01:00 Add a new interruptIOManager API for the I/O managers It will be used to interrupt awaitCompletedTimeoutsOrIO. Also update the return type and docs for awaitCompletedTimeoutsOrIO to have it return false when it gets interrupted, and have no useful post condition in that case. - - - - - a0b052a9 by Duncan Coutts at 2026-06-19T08:09:20+01:00 Add interruptIOManager support for select I/O manager Uses the FdWakup mechanism. - - - - - 89b15fd6 by Duncan Coutts at 2026-06-19T08:09:20+01:00 Add interruptIOManager support for poll I/O manager Uses the FdWakup mechanism. A quirk we have to cope with is that we now need to poll one more fd -- the wakeup_fd_r -- but this fd has no corresponding entry in the aiop_table. This is awkward since we have set up our aiop_poll_table to be an auxilliary table with matching indicies. The solution this patch uses (and described in the comments) is to have two tables: struct pollfd *aiop_poll_table, *full_poll_table; and to have the aiop_poll_table alias the tail of the full_poll_table. The head entry in the full_poll_table is the extra fd. So we poll the full_poll_table, while the aiop_poll_table still has matching indicies with the aiop_table. Hurrah for C aliasing rules. - - - - - 9e503496 by Duncan Coutts at 2026-06-19T08:09:20+01:00 Add interruptIOManager support for win32 legacy I/O manager And remove unused related helper resetAbandonRequestWait. It is not called because the event is created in auto-reset mode, so never needs to be reset manually. - - - - - 8840ca5f by Duncan Coutts at 2026-06-19T08:09:20+01:00 Note lack of interruptIOManager support for WinIO I/O manager Though there's a plausible design, we can't sanely test it at the moment due to related WinIO bugs. Filed as issue #27403. - - - - - 26b395ed by Duncan Coutts at 2026-06-19T08:32:49+01:00 Make signal handling be a respondibility of the I/O manager(s) Previously it was scattered between I/O managers and the scheduler, and especially the scheduler's deadlock detection. Previously the scheduler would poll for pending signals each iteration of the scheduler loop. The scheduler also had some hairy signal functionality in the deadlock detection: in the non-threaded RTS (only) if there were still no threads running after deadlock detection then it would block waiting for signals. But signals can and (in my opinion) should be thought of as just a funny kind of I/O, and thus should be a responsibility of the I/O manager. So now we have the I/O managers poll for signals when they are polling for I/O completion (and removing the separate poll in the scheduler). And when I/O managers block waiting for I/O then they now also start signal handlers if they get interrupted by a signal. Crucially, if there is no pending I/O or timers, the awaitCompletedTimeoutsOrIO will still block waiting for signals. This patch puts us into an intermediate state: it temporarily breaks deadlock detection in the non-threaded RTS. The waiting on I/O currently happens before deadlock detection. This means we'll now wait forever on signals before doing deadlock detection. We need to move waiting after deadlock detection. We'll do that in a later patch. - - - - - 67e08f7e by Duncan Coutts at 2026-06-19T08:32:49+01:00 Clean up the RTS internal signal handling API Now that the I/O manager is responsible for signals, we can simplify the API we present for signal handling. We now just need startPendingSignalHandlers, which is called from the I/O managers. We can get rid of awaitUserSignals. We also don't need RtsSignals.h to re-export the platform-specific posix/Signals.h or win32/ConsoleHandler.h We can also hide more of the implementation of signals. Less has to be exposed in posix/Signals.h or win32/ConsoleHandler.h. Indeed, posix/Signals.h becomes empty and we remove it. Partly this is because we don't need inline functions (or macros) in the interface. Also remove signal_handlers from RTS ABI exported symbols list. It does not appear to have any users in the core libs, and its really an internal implementation detail. It should not be exposed unless it's really necessary. - - - - - f185da55 by Duncan Coutts at 2026-06-19T08:32:49+01:00 In the scheduler, move I/O blocking after deadlock detection To make deadlock detection effective in the non-threaded RTS when there are deadlocked threads and other unrelated threads waiting on I/O, we need to arrange to do deadlock detection before we block in scheduler to wait on I/O. The solution is to: 1. adjust scheduleFindWork, which runs before deadlock detection, to only poll for I/O and not block; and 2. add a step after deadlock detection to wait on I/O if there are still no threads to run (and there's any I/O or timeouts outstanding) The scheduleCheckBlockedThreads is now so simple that it made more sense to inline it into scheduleFindWork. - - - - - 6a0f27f4 by Duncan Coutts at 2026-06-19T08:32:49+01:00 Remove bogus anyPendingTimeoutsOrIO guard from scheduleDetectDeadlock The deadlock detection was only invoked if both of these conditions hold: 1. the run queue is empty 2. there is no pending I/O or timeouts The second condition is unnecessary. The deadlock detection mechanism can find deadlocks even if there are other threads waiting on I/O or timers. Having this extra condition means that we fail to detect blocked threads if there are any threads waiting on I/O or timers. Part of fixing issue #26408 - - - - - 46aae717 by Duncan Coutts at 2026-06-19T08:32:49+01:00 Don't consider pending I/O for early context switch optimisation Context switches are normally initiated by the timer signal. If however the user specifies "context switch as often as possible", with +RTS -C0 then the scheduler arranges for an early context switch (when it's just about to run a Haskell thread). Context switching very often is expensive, so as an optimisation there cases where we do not arrange an early context switch: 1. if there's no other threads to run 2. if there is no pending I/O or timers This patch eliminates case 2, leaving only case 1. The rationale is as follows. The use of this was inconsistent across platforms and threaded/non-threaded RTS ways. It only worked on the non-threaded RTS and on Windows only worked for the win32-legacy I/O manager. On all other combinations anyPendingTimeoutsOrIO would always return false. The fact that nobody noticed and complained about this inconsistency suggests that the feature is not relied upon. If however it turns out that applications do rely on this, then the proper thing to do is not to restore this check, but to add a new I/O manager hint function that returns if there is any pending events that are likely to happen *soon*: for example timeouts expiring within one timeslice, or I/O waits on things likely to complete soon like disk I/O, but not for example socket/pipe I/O. The motivation to avoid this use of anyPendingTimeoutsOrIO is to allow us to eliminate anyPendingTimeoutsOrIO entirely. All other uses of this are just guards on {await,poll}CompletedTimeoutsOrIO and the guards can safely be folded into those functions. This will better cope with some I/O managers having no proper implementation of anyPendingTimeoutsOrIO. Ultimately this will let us simplify the scheduler which currently has to have special #ifdef mingw32_HOST_OS cases to cope with the lack of a working anyPendingTimeoutsOrIO for some Windows I/O managers - - - - - e1cefda2 by Duncan Coutts at 2026-06-19T08:32:49+01:00 Remove anyPendingTimeoutsOrIO guarding {poll,await}CompletedTimeoutsOrIO Previously the API of the I/O manager used a two step process: check anyPendingTimeoutsOrIO and then call {poll,await}CompletedTimeoutsOrIO. This was primarily there as a performance thing, to cheaply check if we need to do anything. And then because anyPendingTimeoutsOrIO existed, it was used for other things too. We have now eliminated the other uses, and are just left with the performance pattern. But this was problematic because not all I/O managers correctly implement anyPendingTimeoutsOrIO (specifically the win32 ones), and now that we also make I/O managers responsible for signals then we need to poll/await even if there is no pending I/O or timeouts. If there is no pending I/O or timeouts then poll/await needs to degenerate to just waiting forever for any signals. - - - - - e1be565f by Duncan Coutts at 2026-06-19T08:32:49+01:00 Remove anyPendingTimeoutsOrIO, it is no longer used And this avoids the problems arising from the win32 I/O managers having had a bogus implementation. - - - - - 0051eb9f by Duncan Coutts at 2026-06-19T08:32:49+01:00 Remove second scheduler call to awaitCompletedTimeoutsOrIO Previously awaitCompletedTimeoutsOrIO was called both before and after deadlock detection in the scheduler. The reason for that was that the win32 I/O managers had a bogus implementation of anyPendingTimeoutsOrIO and this was used to guard the call of awaitCompletedTimeoutsOrIO prior to deadlock detection. This meant the first call site was never actually called when using the win32 I/O managers. This was the reason for the second call: the first one was never used. What a mess. So now we have a simple design in the scheduler: 1. poll for completed I/O, timers or signals 2. if no runnable threads: do deadlock detection 3. if still no runnable threads: block waiting for I/O, timers or signals. - - - - - ab8c5a1f by Duncan Coutts at 2026-06-19T08:32:49+01:00 Lift emptyRunQueue guard out of scheduleDetectDeadlock this improved the clarity of the logic when reading the scheduler code. - - - - - 953effcf by Duncan Coutts at 2026-06-19T08:32:49+01:00 Make non-threaded deadlock detection also rely on idle GC Only do deadlock detection GC when idle GC kicks in. This also relies on using wakeUpRts, so now do this unconditionally. Previously wakeUpRts was for the threaded rts only. - - - - - 0a463c1c by Duncan Coutts at 2026-06-19T08:32:49+01:00 Enable idle GC by default on non-threaded RTS. The behaviour is now uniform between threaded and non-threaded. The deadlock detection now relies on idle GC for both threaded and non-threaded ways. Previously deadlock detection did not rely on idle GC for the non-threaded way. - - - - - da2b0853 by Duncan Coutts at 2026-06-19T08:32:49+01:00 Add a long Note [Deadlock detection] It describes the historical and modern designs and their trade-offs. The point is we've now unified the code for deadlock detection between the threaded and non-threaded ways, by changing the non-threaded to follow the same design as the threaded. - - - - - fa0f947a by Duncan Coutts at 2026-06-19T08:32:49+01:00 Add a test for deadlock detection, issue #26408 - - - - - f2345a09 by Duncan Coutts at 2026-06-19T08:32:50+01:00 Update the user guide with the revised idle GC behaviour i.e. it's now not just for the threaded RTS, but general. Also document the fact that disabling idle GC also disables deadlock detection. And add a changelog entry. - - - - - 32 changed files: - + changelog.d/idle-gc-and-deadlock-detection - docs/users_guide/runtime_control.rst - libraries/ghc-internal/src/GHC/Internal/Event/Control.hs - libraries/ghc-internal/src/GHC/Internal/Event/Manager.hs - libraries/ghc-internal/src/GHC/Internal/Event/TimerManager.hs - rts/IOManager.c - rts/IOManager.h - rts/IOManagerInternals.h - rts/Linker.c - rts/RtsFlags.c - rts/RtsSignals.h - rts/RtsStartup.c - rts/RtsSymbols.c - rts/Schedule.c - rts/include/rts/IOInterface.h - rts/posix/MIO.c - rts/posix/MIO.h - rts/posix/Poll.c - rts/posix/Poll.h - rts/posix/Select.c - rts/posix/Select.h - rts/posix/Signals.c - − rts/posix/Signals.h - rts/win32/AsyncMIO.c - rts/win32/AsyncMIO.h - rts/win32/AwaitEvent.c - rts/win32/AwaitEvent.h - rts/win32/ConsoleHandler.c - rts/win32/ConsoleHandler.h - + testsuite/tests/rts/T26408.hs - + testsuite/tests/rts/T26408.stderr - testsuite/tests/rts/all.T The diff was not included because it is too large. View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/e248df7a530927e858a2c3e9bf1f5e1... -- View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/compare/e248df7a530927e858a2c3e9bf1f5e1... You're receiving this email because of your account on gitlab.haskell.org.