gopls/internal/cache: prime goimports cache asynchronously

Gopls' refresh of the goimports resolver already introduces
non-determinism into imports operations: gopls does not observe changes
until the asynchronous refresh occurs.

This change allows operations to continue to run on the stale resolver
until a new resolver is ready.

Due to inherent raciness, it's hard to benchmark the impact of this
change: one would have to catch gopls during a refresh, which occurs at
an automatically adjusted pacing.

Also update TODOs.

Fixes golang/go#59216

Change-Id: I303df998d804c9a1cd1c0e307872d1d271eed601
Reviewed-on: https://go-review.googlesource.com/c/tools/+/561235
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Alan Donovan <[email protected]>

Author: findleyr

gopls/internal/cache/imports.go

@@ -207,21 +207,23 @@ func (s *importsState) refreshProcessEnv() {
 	start := time.Now()
 
 	s.mu.Lock()
-	env := s.processEnv
-	if resolver, err := s.processEnv.GetResolver(); err == nil {
-		resolver.ClearForNewScan()
-	}
-	// TODO(rfindley): it's not clear why we're unlocking here. Shouldn't we
-	// guard the use of env below? In any case, we can prime a separate resolver.
+	resolver, err := s.processEnv.GetResolver()
 	s.mu.Unlock()
+	if err != nil {
+		return
+	}
 
 	event.Log(s.ctx, "background imports cache refresh starting")
 
-	// TODO(rfindley, golang/go#59216): do this priming with a separate resolver,
-	// and then replace, so that we never have to wait on an unprimed cache.
-	if err := imports.PrimeCache(context.Background(), env); err == nil {
+	// Prime the new resolver before updating the processEnv, so that gopls
+	// doesn't wait on an unprimed cache.
+	if err := imports.PrimeCache(context.Background(), resolver); err == nil {
 		event.Log(ctx, fmt.Sprintf("background refresh finished after %v", time.Since(start)))
 	} else {
 		event.Log(ctx, fmt.Sprintf("background refresh finished after %v", time.Since(start)), keys.Err.Of(err))
 	}
+
+	s.mu.Lock()
+	s.processEnv.UpdateResolver(resolver)
+	s.mu.Unlock()
 }

internal/imports/fix.go

@@ -701,20 +701,21 @@ func ScoreImportPaths(ctx context.Context, env *ProcessEnv, paths []string) (map
 	return result, nil
 }
 
-func PrimeCache(ctx context.Context, env *ProcessEnv) error {
+func PrimeCache(ctx context.Context, resolver Resolver) error {
 	// Fully scan the disk for directories, but don't actually read any Go files.
 	callback := &scanCallback{
-		rootFound: func(gopathwalk.Root) bool {
-			return true
+		rootFound: func(root gopathwalk.Root) bool {
+			// See getCandidatePkgs: walking GOROOT is apparently expensive and
+			// unnecessary.
+			return root.Type != gopathwalk.RootGOROOT
 		},
 		dirFound: func(pkg *pkg) bool {
 			return false
 		},
-		packageNameLoaded: func(pkg *pkg) bool {
-			return false
-		},
+		// packageNameLoaded and exportsLoaded must never be called.
 	}
-	return getCandidatePkgs(ctx, callback, "", "", env)
+
+	return resolver.scan(ctx, callback)
 }
 
 func candidateImportName(pkg *pkg) string {
@@ -1089,7 +1090,12 @@ type Resolver interface {
 	// scoreImportPath returns the relevance for an import path.
 	scoreImportPath(ctx context.Context, path string) float64
 
-	ClearForNewScan()
+	// ClearForNewScan returns a new Resolver based on the receiver that has
+	// cleared its internal caches of directory contents.
+	//
+	// The new resolver should be primed and then set via
+	// [ProcessEnv.UpdateResolver].
+	ClearForNewScan() Resolver
 }
 
 // A scanCallback controls a call to scan and receives its results.
@@ -1270,11 +1276,8 @@ func newGopathResolver(env *ProcessEnv) *gopathResolver {
 	return r
 }
 
-func (r *gopathResolver) ClearForNewScan() {
-	<-r.scanSema
-	r.cache = NewDirInfoCache()
-	r.walked = false
-	r.scanSema <- struct{}{}
+func (r *gopathResolver) ClearForNewScan() Resolver {
+	return newGopathResolver(r.env)
 }
 
 func (r *gopathResolver) loadPackageNames(importPaths []string, srcDir string) (map[string]string, error) {

internal/imports/mod.go

@@ -30,16 +30,11 @@ import (
 // both caused by populating the cache, albeit in slightly different ways.
 //
 // A high level list of TODOs:
-//  - Write an additional benchmark for refreshing the directory state.
-//  - Split scanning the module cache from other ModuleResolver functionality,
-//    as it is the source of performance woes (and inconsistency).
-//  - Allow sharing module cache state across multiple ModuleResolvers.
 //  - Optimize the scan itself, as there is some redundancy statting and
 //    reading go.mod files.
-//  - Make it possible to reuse the current state while running a refresh in
-//    the background.
-//  - Fix context cancellation (again): if the context is cancelled while a
-//    root is being walked, nothing stops that ongoing walk.
+//  - Invert the relationship between ProcessEnv and Resolver (see the
+//    docstring of ProcessEnv).
+//  - Make it easier to use an external resolver implementation.
 //
 // Smaller TODOs are annotated in the code below.
 
@@ -72,7 +67,11 @@ type ModuleResolver struct {
 	modsByDir      []*gocommand.ModuleJSON          // ...or by the number of path components in their Dir.
 
 	// Scanning state, populated by scan
-	scanSema     chan struct{}            // prevents concurrent scans and guards scannedRoots
+
+	// scanSema prevents concurrent scans, and guards scannedRoots and the cache
+	// fields below (though the caches themselves are concurrency safe).
+	// Receive to acquire, send to release.
+	scanSema     chan struct{}
 	scannedRoots map[gopathwalk.Root]bool // if true, root has been walked
 
 	// Caches of directory info, populated by scans and scan callbacks
@@ -86,12 +85,16 @@ type ModuleResolver struct {
 	otherCache       *DirInfoCache
 }
 
+// newModuleResolver returns a new module-aware goimports resolver.
+//
+// Note: use caution when modifying this constructor: changes must also be
+// reflected in ModuleResolver.ClearForNewScan.
 func newModuleResolver(e *ProcessEnv, moduleCacheCache *DirInfoCache) (*ModuleResolver, error) {
 	r := &ModuleResolver{
 		env:      e,
 		scanSema: make(chan struct{}, 1),
 	}
-	r.scanSema <- struct{}{}
+	r.scanSema <- struct{}{} // release
 
 	goenv, err := r.env.goEnv()
 	if err != nil {
@@ -265,10 +268,23 @@ func (r *ModuleResolver) initAllMods() error {
 // It preserves the set of roots, but forgets about the set of directories.
 // Though it forgets the set of module cache directories, it remembers their
 // contents, since they are assumed to be immutable.
-func (r *ModuleResolver) ClearForNewScan() {
-	<-r.scanSema
-	prevRoots := r.scannedRoots
-	r.scannedRoots = map[gopathwalk.Root]bool{}
+func (r *ModuleResolver) ClearForNewScan() Resolver {
+	<-r.scanSema // acquire r, to guard scannedRoots
+	r2 := &ModuleResolver{
+		env:            r.env,
+		dummyVendorMod: r.dummyVendorMod,
+		moduleCacheDir: r.moduleCacheDir,
+		roots:          r.roots,
+		mains:          r.mains,
+		mainByDir:      r.mainByDir,
+		modsByModPath:  r.modsByModPath,
+
+		scanSema:         make(chan struct{}, 1),
+		scannedRoots:     make(map[gopathwalk.Root]bool),
+		otherCache:       NewDirInfoCache(),
+		moduleCacheCache: r.moduleCacheCache,
+	}
+	r2.scanSema <- struct{}{} // r2 must start released
 	// Invalidate root scans. We don't need to invalidate module cache roots,
 	// because they are immutable.
 	// (We don't support a use case where GOMODCACHE is cleaned in the middle of
@@ -278,12 +294,12 @@ func (r *ModuleResolver) ClearForNewScan() {
 	// Scanning for new directories in GOMODCACHE should be handled elsewhere,
 	// via a call to ScanModuleCache.
 	for _, root := range r.roots {
-		if root.Type == gopathwalk.RootModuleCache && prevRoots[root] {
-			r.scannedRoots[root] = true
+		if root.Type == gopathwalk.RootModuleCache && r.scannedRoots[root] {
+			r2.scannedRoots[root] = true
 		}
 	}
-	r.otherCache = NewDirInfoCache()
-	r.scanSema <- struct{}{}
+	r.scanSema <- struct{}{} // release r
+	return r2
 }
 
 // ClearModuleInfo invalidates resolver state that depends on go.mod file
@@ -299,16 +315,27 @@ func (e *ProcessEnv) ClearModuleInfo() {
 	if r, ok := e.resolver.(*ModuleResolver); ok {
 		resolver, resolverErr := newModuleResolver(e, e.ModCache)
 		if resolverErr == nil {
-			<-r.scanSema // guards caches
+			<-r.scanSema // acquire (guards caches)
 			resolver.moduleCacheCache = r.moduleCacheCache
 			resolver.otherCache = r.otherCache
-			r.scanSema <- struct{}{}
+			r.scanSema <- struct{}{} // release
 		}
 		e.resolver = resolver
 		e.resolverErr = resolverErr
 	}
 }
 
+// UpdateResolver sets the resolver for the ProcessEnv to use in imports
+// operations. Only for use with the result of [Resolver.ClearForNewScan].
+//
+// TODO(rfindley): this awkward API is a result of the (arguably) inverted
+// relationship between configuration and state described in the doc comment
+// for [ProcessEnv].
+func (e *ProcessEnv) UpdateResolver(r Resolver) {
+	e.resolver = r
+	e.resolverErr = nil
+}
+
 // findPackage returns the module and directory from within the main modules
 // and their dependencies that contains the package at the given import path,
 // or returns nil, "" if no module is in scope.
@@ -580,9 +607,9 @@ func (r *ModuleResolver) scan(ctx context.Context, callback *scanCallback) error
 		select {
 		case <-ctx.Done():
 			return
-		case <-r.scanSema:
+		case <-r.scanSema: // acquire
 		}
-		defer func() { r.scanSema <- struct{}{} }()
+		defer func() { r.scanSema <- struct{}{} }() // release
 		// We have the lock on r.scannedRoots, and no other scans can run.
 		for _, root := range roots {
 			if ctx.Err() != nil {

internal/imports/mod_test.go

@@ -244,7 +244,7 @@ import _ "rsc.io/sampler"
 	// Clear out the resolver's cache, since we've changed the environment.
 	mt.env.Env["GOFLAGS"] = "-mod=vendor"
 	mt.env.ClearModuleInfo()
-	mt.env.resolver.ClearForNewScan()
+	mt.env.UpdateResolver(mt.env.resolver.ClearForNewScan())
 	mt.assertModuleFoundInDir("rsc.io/sampler", "sampler", `/vendor/`)
 }
 
@@ -1314,7 +1314,7 @@ func BenchmarkModuleResolver_RescanModCache(b *testing.B) {
 	b.ResetTimer()
 	for i := 0; i < b.N; i++ {
 		scanToSlice(resolver, exclude)
-		resolver.(*ModuleResolver).ClearForNewScan()
+		resolver = resolver.ClearForNewScan()
 	}
 }