Cache<'Key, 'Value> in Caches.fs has a finalizer that calls Dispose (L397-L398). I measured two effects of it with FCS 43.12.201. Main has the same code in the places linked below.
1. A dropped check survives one extra full GC
MemoizationTable creates its cache with NoEviction (illib.fs L1103), and InfoReader has 11 of these tables. With NoEviction and Immediate, Dispose has no eviction processor to stop (L308-L309). It only counts the disposal for the metrics (L387-L390). MemoizationTable is not disposable, so the finalizer runs for each table.
An object with a finalizer survives the collection that finds it dead, and so does everything it references, until the finalizer has run and a later collection of its generation happens. A table references the infos it computed, and through them the type checker state. So when a check is dropped, that state survives the first full GC and is promoted to gen 2. A process that goes idle can wait a long time for the next gen 2 GC.
Measured:
-
A Cache created through reflection, holding a 50 MB value. Long weak references, which stay set while an object waits for its finalizer:
| Mode |
Finalizer |
Value alive after one full GC |
After WaitForPendingFinalizers and a second full GC |
NoEviction |
kept |
yes |
no |
NoEviction |
GC.SuppressFinalize |
no |
no |
Immediate |
kept |
yes |
no |
Immediate |
GC.SuppressFinalize |
no |
no |
-
FsAutoComplete on the Fantomas solution, transparent compiler. I called ClearCache for every project, then ran one GC.Collect(2, GCCollectionMode.Aggressive, true, true) and took a heap dump. No root reached 62 MB of the heap, and 44.2 MB of that was reachable only from dead Cache instances. An earlier run left 135 MB unreachable. With GC.WaitForPendingFinalizers() and a second collection, 0.8 MB was left. FsAutoComplete now does that as a workaround.
2. A cache in MailboxProcessor mode is never collected
MailboxProcessor is the default eviction mode (L157). The eviction loop waits in mb.Receive() (L289-L301). FSharp.Core waits there with Async.AwaitWaitHandle, which registers the wait with the thread pool, and the loop references the cache. The root path from a dump:
root Stack System.Threading.PortableThreadPool+WaitThread
._registeredWaits -> System.Threading.RegisteredWaitHandle[]
.[] -> System.Threading.RegisteredWaitHandle
._callbackHelper -> System.Threading._ThreadPoolWaitOrTimerCallback
._waitOrTimerCallback -> System.Threading.WaitOrTimerCallback
._target -> <StartupCode$FSharp-Core>.$Async+AwaitWaitHandle@1918-3
.ctxt -> Microsoft.FSharp.Control.AsyncActivationContents<System.Boolean>
.cont@ -> Microsoft.FSharp.Control.AsyncPrimitives+Bind@589<EvictionQueueMessage<...>, ...>
.ctxt -> Microsoft.FSharp.Control.AsyncActivationContents<EvictionQueueMessage<...>>
.cont@ -> Microsoft.FSharp.Control.AsyncPrimitives+Bind@589<Unit, EvictionQueueMessage<...>>
.part2 -> <StartupCode$FSharp-Compiler-Service>.$Caches+processNext@295-1<...>
.this -> FSharp.Compiler.Caches.Cache<...>
So the cache stays reachable until something disposes it. The finalizer is meant to stop the loop when Dispose was not called, but it can never run. In the reflection test above, a MailboxProcessor cache and its 50 MB value were still alive after three full GCs, with or without GC.SuppressFinalize.
Outside CompilationMode.OneOff, getTypeSubsumptionCache (TypeRelations.fs L42-L49) and getOverloadResolutionCache (OverloadResolutionCache.fs L48-L63) create a cache of this kind for each TcGlobals, kept in a WeakMap. Nothing disposes them, so every TcGlobals leaves its cache and a thread pool wait behind.
Measured: an fsi script creates five FSharpCheckers one after the other, checks a script with a type test with each, and drops them. Then three full GCs, each followed by WaitForPendingFinalizers, and a heap dump:
| Type |
Instances a root reaches |
TcGlobals |
1 (fsi's own) |
FSharpChecker |
1 (fsi's own) |
Cache<TTypeCacheKey, bool> (typeSubsumptionCache) |
6 |
RegisteredWaitHandle |
9 |
The background compiler and the transparent compiler give the same counts. The entries do not reference the typed tree, and in FsAutoComplete the caches held 0.3 to 2 MB each. So each TcGlobals leaks little, but nothing ever frees it. My script did not use the overload resolution cache, which is created the same way.
Proposal
- Remove the finalizer. With
NoEviction and Immediate it has nothing to clean up, and with MailboxProcessor it cannot run.
- Give the caches that nobody disposes,
typeSubsumptionCache and overloadResolutionCache, a lifetime they can end. Two options: use EvictionMode.Immediate for them, or keep the eviction loop from making the cache reachable while it waits. Disposing them would need an owner, and the WeakMap is not one.
Repro: Cache lifetime per eviction mode (reflection)
#r "nuget: FSharp.Compiler.Service, 43.12.201"
open System
open System.Reflection
open System.Runtime.CompilerServices
open System.Collections.Generic
open Microsoft.FSharp.Reflection
let asm = typeof<FSharp.Compiler.CodeAnalysis.FSharpChecker>.Assembly
let flags = BindingFlags.Public ||| BindingFlags.NonPublic ||| BindingFlags.Static ||| BindingFlags.Instance
let cacheType = asm.GetType("FSharp.Compiler.Caches.Cache`2", true).MakeGenericType(typeof<string>, typeof<byte[]>)
let getDefault = asm.GetType("FSharp.Compiler.Caches.CacheOptions", true).GetMethod("getDefault", flags).MakeGenericMethod(typeof<string>)
[<MethodImpl(MethodImplOptions.NoInlining)>]
let create (mode: string) (suppressFinalizer: bool) =
let defaults = getDefault.Invoke(null, [| box (EqualityComparer<string>.Default :> IEqualityComparer<string>) |])
let optionsType = defaults.GetType()
let fields = FSharpValue.GetRecordFields(defaults, flags)
let modeCase = FSharpType.GetUnionCases(asm.GetType("FSharp.Compiler.Caches.EvictionMode", true), flags) |> Array.find (fun c -> c.Name = mode)
fields[FSharpType.GetRecordFields(optionsType, flags) |> Array.findIndex (fun p -> p.Name = "EvictionMode")] <- FSharpValue.MakeUnion(modeCase, [||], flags)
let cache = (cacheType.GetConstructors(flags) |> Array.exactlyOne).Invoke([| FSharpValue.MakeRecord(optionsType, fields, flags); box (Some "repro") |])
let value = Array.zeroCreate<byte> (50 * 1024 * 1024)
cacheType.GetMethod("TryAdd", flags).Invoke(cache, [| box "key"; box value |]) |> ignore
Threading.Thread.Sleep 200
if suppressFinalizer then GC.SuppressFinalize cache
// Long weak reference: it stays set while the object waits for its finalizer
WeakReference(value, true)
for mode in [ "NoEviction"; "Immediate"; "MailboxProcessor" ] do
for suppress in [ false; true ] do
let value = create mode suppress
GC.Collect(2, GCCollectionMode.Forced, true, true)
let afterOne = value.IsAlive
GC.WaitForPendingFinalizers()
GC.Collect(2, GCCollectionMode.Forced, true, true)
GC.WaitForPendingFinalizers()
GC.Collect(2, GCCollectionMode.Forced, true, true)
printfn "%-16s suppressed %-5b value alive after one GC %-5b after finalizers and two more GCs %b" mode suppress afterOne value.IsAlive
Repro: one typeSubsumptionCache left behind per dropped checker
Run with dotnet fsi leak.fsx background (or transparent), take a heap dump with dotnet-dump collect --type Heap -p <pid> while it sleeps, and count the instances a GC root reaches (I used ClrMD).
#r "nuget: FSharp.Compiler.Service, 43.12.201"
open System
open System.Runtime.CompilerServices
open FSharp.Compiler.CodeAnalysis
open FSharp.Compiler.Text
let useTransparentCompiler = fsi.CommandLineArgs[1] = "transparent"
let file = IO.Path.Combine(__SOURCE_DIRECTORY__, "Script.fsx")
// The type test makes the checker use the type subsumption cache
let source = SourceText.ofString "let f (o: obj) = match o with :? string as s -> s.Length | :? System.IComparable -> 1 | _ -> 0\nprintfn \"%d\" (f (box \"a\"))"
[<MethodImpl(MethodImplOptions.NoInlining)>]
let check () =
let checker = FSharpChecker.Create(useTransparentCompiler = useTransparentCompiler)
let options, _ = checker.GetProjectOptionsFromScript(file, source, assumeDotNetFramework = false) |> Async.RunSynchronously
let _, answer = checker.ParseAndCheckFileInProject(file, 0, source, options) |> Async.RunSynchronously
match answer with
| FSharpCheckFileAnswer.Succeeded _ -> ()
| FSharpCheckFileAnswer.Aborted -> failwith "aborted"
for _ in 1..5 do check ()
for _ in 1..3 do
GC.Collect(2, GCCollectionMode.Forced, true, true)
GC.WaitForPendingFinalizers()
printfn "pid %d" Environment.ProcessId
Threading.Thread.Sleep 120000
Related: #20755 (another thing the transparent compiler keeps in memory, found the same way).
Cache<'Key, 'Value>inCaches.fshas a finalizer that callsDispose(L397-L398). I measured two effects of it with FCS 43.12.201. Main has the same code in the places linked below.1. A dropped check survives one extra full GC
MemoizationTablecreates its cache withNoEviction(illib.fs L1103), andInfoReaderhas 11 of these tables. WithNoEvictionandImmediate,Disposehas no eviction processor to stop (L308-L309). It only counts the disposal for the metrics (L387-L390).MemoizationTableis not disposable, so the finalizer runs for each table.An object with a finalizer survives the collection that finds it dead, and so does everything it references, until the finalizer has run and a later collection of its generation happens. A table references the infos it computed, and through them the type checker state. So when a check is dropped, that state survives the first full GC and is promoted to gen 2. A process that goes idle can wait a long time for the next gen 2 GC.
Measured:
A
Cachecreated through reflection, holding a 50 MB value. Long weak references, which stay set while an object waits for its finalizer:WaitForPendingFinalizersand a second full GCNoEvictionNoEvictionGC.SuppressFinalizeImmediateImmediateGC.SuppressFinalizeFsAutoComplete on the Fantomas solution, transparent compiler. I called
ClearCachefor every project, then ran oneGC.Collect(2, GCCollectionMode.Aggressive, true, true)and took a heap dump. No root reached 62 MB of the heap, and 44.2 MB of that was reachable only from deadCacheinstances. An earlier run left 135 MB unreachable. WithGC.WaitForPendingFinalizers()and a second collection, 0.8 MB was left. FsAutoComplete now does that as a workaround.2. A cache in
MailboxProcessormode is never collectedMailboxProcessoris the default eviction mode (L157). The eviction loop waits inmb.Receive()(L289-L301). FSharp.Core waits there withAsync.AwaitWaitHandle, which registers the wait with the thread pool, and the loop references the cache. The root path from a dump:So the cache stays reachable until something disposes it. The finalizer is meant to stop the loop when
Disposewas not called, but it can never run. In the reflection test above, aMailboxProcessorcache and its 50 MB value were still alive after three full GCs, with or withoutGC.SuppressFinalize.Outside
CompilationMode.OneOff,getTypeSubsumptionCache(TypeRelations.fs L42-L49) andgetOverloadResolutionCache(OverloadResolutionCache.fs L48-L63) create a cache of this kind for eachTcGlobals, kept in aWeakMap. Nothing disposes them, so everyTcGlobalsleaves its cache and a thread pool wait behind.Measured: an fsi script creates five
FSharpCheckers one after the other, checks a script with a type test with each, and drops them. Then three full GCs, each followed byWaitForPendingFinalizers, and a heap dump:TcGlobalsFSharpCheckerCache<TTypeCacheKey, bool>(typeSubsumptionCache)RegisteredWaitHandleThe background compiler and the transparent compiler give the same counts. The entries do not reference the typed tree, and in FsAutoComplete the caches held 0.3 to 2 MB each. So each
TcGlobalsleaks little, but nothing ever frees it. My script did not use the overload resolution cache, which is created the same way.Proposal
NoEvictionandImmediateit has nothing to clean up, and withMailboxProcessorit cannot run.typeSubsumptionCacheandoverloadResolutionCache, a lifetime they can end. Two options: useEvictionMode.Immediatefor them, or keep the eviction loop from making the cache reachable while it waits. Disposing them would need an owner, and theWeakMapis not one.Repro:
Cachelifetime per eviction mode (reflection)Repro: one
typeSubsumptionCacheleft behind per dropped checkerRun with
dotnet fsi leak.fsx background(ortransparent), take a heap dump withdotnet-dump collect --type Heap -p <pid>while it sleeps, and count the instances a GC root reaches (I used ClrMD).Related: #20755 (another thing the transparent compiler keeps in memory, found the same way).