Repository navigation
asset.dump_folder: bound the synchronous build/write phase per asset #369
Description
Remainder of #244 (asset.dump_folder per-asset time bound) after its load-phase slice landed.
What #244 shipped: a package that is not in memory is loaded with LoadPackageAsync and the sweep pumps the loader at most 50 ms per tick (AsyncLoadPumpSliceSeconds), publishing currentPhase: "loading"; a load past 120 s is skipped with ASSET_LOAD_TIMEOUT; slowAssets[] lists every asset whose synchronous dump call held the editor 5 s or more, with its slowest phase (loading / building / writing); asset.dump refuses packages in dump-stalled.txt and journals its own call.
Still unbounded, in order:
- Build/write phase. After the load,
DumpSingleAssetrunsBuildAllFilesForAsset(every aspect builder) andAssetDumpWriter::WriteAssetDumpas one game-thread call (Handlers/Asset/AssetDumpHandler.cpp,DumpSingleAsset). Split it into per-aspect steps the ticker can interleave across ticks (holding the asset with a strong pointer, re-checking cancellation, publishing the aspect ascurrentPhase), or give non-yielding aspects a bounded failure path.slowAssets[].phase == "building"is the signal for which assets need it. - Abandoned loads stop release steps. A timed-out load keeps running;
TickFolderDumpReleaseskips the release step (itsFlushAsyncLoadingwould block on that load) until it completes. A load that never completes therefore stops memory release for the rest of the sweep. Releasing around in-flight loads needs to know which packages they import. - Registry preflight.
BuildFolderDumpCandidatescallsIAssetRegistry::GetAssetsonce, synchronously. Page it per sub-path if large trees show it as a long tick. - Pathological fixture. A fixture with a deliberately slow build phase, asserting per-tick latency, progress,
system.job_statusand cancellation stay responsive.PinWright.asset.dump.AsyncFolderDump.UnloadedPackageLoadsAcrossTickscovers only the load phase. - Stalled list has no verb. A user deletes the
dump-stalled.txtline by hand to retry.
Activity
SSS135 commentedon Oct 6, 2026
Review follow-ups from #244 that land here (not fixed in #244):
bResidentinTickFolderDump). ItsDumpSingleAssetLoadObjectthen finishes the rest of that load in one journaled but unbounded call, and the package is not added toSweepLoadedPackages, so it is never released. Fix direction: treat a package with an in-flight load (GetAsyncLoadPercentage >= 0, or the loader's own state) as not resident and wait on it through the pump.FlushAsyncLoading(editor map open, PIE start, other PinWright verbs) finishes it synchronously. This is a stall, not a hang.StaticTick,EditorEngine.cpp:1788, 5 ms per frame) advances abandoned loads with no.dump-inflight.tmpmarker, so a freeze there journals nothing. The pump journal now lists the active path plus every still-running abandoned path, so a freeze inside a pump quarantines all of them, including possibly innocent ones.ProcessAsyncLoadingUntilComplete(TFunctionRef<bool()>, double),LoadPackageAsync(const FString&, FLoadPackageAsyncDelegate, ...),FLoadPackageAsyncDelegateandIsAsyncLoadingSuspendedare verified against 5.8 source only. Older engines may takefloat TimeLimit; passing theconstexpr doublemay raise an MSVC C4305/C4244 warning. asset.dump_folder needs a real per-asset time bound #244 should not be closed until a version-matrix run has covered 5.3-5.7.bAbortReleaseGcWait, restoring the clearedRF_Standaloneflags), so tests prove the step runs (releaseStepCount,releasedPackageCount) but not that the package is freed. A live-editor check that a swept package is unloaded after the job completes is still needed.Bound asset.dump_folder package loads and run the final release