enhance(cache): scope cache entries by ref like GitHub (#1361)

Cache entries were only scoped by repository, so an entry a pull request saved could be restored on the default branch, for example through a `restore-keys` prefix.

Entries now carry the ref of the job that saved them. As on GitHub ([Restrictions for accessing a cache](https://docs.github.com/en/actions/reference/workflows-and-actions/dependency-caching#restrictions-for-accessing-a-cache)), a job restores entries of its own ref first, then of its pull request's base branch, then of the default branch. Pull request events are scoped by their pull request number.

Entries saved before this change are no longer restored and expire through `cache.retention`, so caches start cold once after the upgrade. Runners sharing a cache, and an external `cache-server`, should be upgraded together.

Co-authored-by: silverwind <[email protected]>
Co-authored-by: silverwind <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1361
Reviewed-by: silverwind <[email protected]>
Co-authored-by: JorgeDLMV <[email protected]>

act/artifactcache/handler.go

@@ -59,6 +59,10 @@ type credKey struct{}
 type JobCredential struct {
 	Repo string `json:"repo"`
 
+	Ref        string `json:"ref"`
+	BaseRef    string `json:"base_ref"`
+	DefaultRef string `json:"default_ref"`
+
 	// Results is the instance whose artifact service this server forwards for the job, and
 	// InsecureTLS how the runner reaches it; see results.go.
 	Results     string `json:"-"`
@@ -68,6 +72,10 @@ type JobCredential struct {
 	PublicURL string `json:"public_url"`
 }
 
+func (cred JobCredential) owns(cache *Cache) bool {
+	return cache.Repo == cred.Repo && cache.Ref == cred.Ref
+}
+
 // credEntry holds a registered job's credential along with an active
 // registration count. RegisterJob is reference-counted so that if two tasks
 // briefly share an ACTIONS_RUNTIME_TOKEN — e.g. a runner that retries a task
@@ -653,7 +661,7 @@ func (h *Handler) find(w http.ResponseWriter, r *http.Request) {
 	}
 	defer db.Close()
 
-	cache, err := h.lookupCache(db, cred.Repo, keys, version)
+	cache, err := h.lookupCache(db, cred, keys, version)
 	if err != nil {
 		h.responseJSON(w, r, 500, err)
 		return
@@ -669,26 +677,27 @@ func (h *Handler) find(w http.ResponseWriter, r *http.Request) {
 	})
 }
 
-// lookupCache returns the entry to restore for these keys, or (nil, nil) when there is none:
-// either nothing matched, or the match had lost its blob to a prune, in which case the dangling
-// entry is dropped on the way out.
-func (h *Handler) lookupCache(db *cacheDB, repo string, keys []string, version string) (*Cache, error) {
-	cache, err := findCache(db, repo, keys, version)
-	if err != nil || cache == nil {
-		return nil, err
+// lookupCache returns the entry to restore or (nil, nil), dropping matches whose blob was pruned.
+func (h *Handler) lookupCache(db *cacheDB, cred JobCredential, keys []string, version string) (*Cache, error) {
+	for {
+		cache, err := findCache(db, cred, keys, version)
+		if err != nil || cache == nil {
+			return nil, err
+		}
+		ok, err := h.storage.Exist(cache.ID)
+		if err != nil {
+			return nil, err
+		}
+		if ok {
+			// Handing out a download URL counts as access, or eviction could drop the entry between
+			// this call and the GET that follows it.
+			h.touch(db, cache)
+			return cache, nil
+		}
+		if !h.deleteCache(db, cache) {
+			return nil, nil //nolint:nilnil // absence is not an error here
+		}
 	}
-	ok, err := h.storage.Exist(cache.ID)
-	if err != nil {
-		return nil, err
-	}
-	if !ok {
-		h.deleteCache(db, cache)
-		return nil, nil //nolint:nilnil // absence is not an error here
-	}
-	// Handing out a download URL counts as access, or eviction could drop the entry between
-	// this call and the GET that follows it.
-	h.touch(db, cache)
-	return cache, nil
 }
 
 // POST /_apis/artifactcache/caches
@@ -700,7 +709,7 @@ func (h *Handler) reserve(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	cache := &Cache{Repo: cred.Repo, Key: api.Key, Version: api.Version, Size: api.Size}
+	cache := &Cache{Repo: cred.Repo, Ref: cred.Ref, Key: api.Key, Version: api.Version, Size: api.Size}
 	if cache.Size == 0 {
 		cache.Size = -1
 	}
@@ -749,7 +758,7 @@ func (h *Handler) upload(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	if cache.Repo != cred.Repo {
+	if !cred.owns(cache) {
 		h.responseJSON(w, r, 403, fmt.Errorf("cache %d: forbidden", id))
 		return
 	}
@@ -808,7 +817,7 @@ func (h *Handler) commit(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	if cache.Repo != cred.Repo {
+	if !cred.owns(cache) {
 		h.responseJSON(w, r, 403, fmt.Errorf("cache %d: forbidden", id))
 		return
 	}
@@ -1071,25 +1080,35 @@ func (h *Handler) signedArtifactURL(cred JobCredential, cacheID uint64, exp time
 	return h.signedURL(cred, apiPath+"/artifacts", "", cacheID, exp)
 }
 
-// if not found, return (nil, nil) instead of an error.
-func findCache(db *cacheDB, repo string, keys []string, version string) (*Cache, error) {
-	for _, prefix := range keys {
-		// if a key in the list matches exactly, don't return partial matches
-		exact, err := findExactCache(db, repo, prefix, version, true)
-		if err != nil {
-			return nil, err
+// findCache tries every key in one ref before the next ref, and returns (nil, nil) on a miss.
+func findCache(db *cacheDB, cred JobCredential, keys []string, version string) (*Cache, error) {
+	refs := []string{cred.Ref}
+	for _, ref := range []string{cred.BaseRef, cred.DefaultRef} {
+		if ref != "" && !slices.Contains(refs, ref) {
+			refs = append(refs, ref)
 		}
-		if exact != nil {
-			return exact, nil
-		}
-		caches, err := db.find(func(cache *Cache) bool {
-			return cache.Repo == repo && strings.HasPrefix(cache.Key, prefix) && cache.Version == version && cache.Complete
-		})
-		if err != nil {
-			return nil, fmt.Errorf("find cache: %w", err)
-		}
-		if cache := latest(caches, (*Cache).createdAtNano); cache != nil {
-			return cache, nil
+	}
+	caches, err := db.find(func(cache *Cache) bool {
+		return cache.Repo == cred.Repo && cache.Version == version && cache.Complete && slices.Contains(refs, cache.Ref) &&
+			slices.ContainsFunc(keys, func(prefix string) bool { return strings.HasPrefix(cache.Key, prefix) })
+	})
+	if err != nil {
+		return nil, fmt.Errorf("find cache: %w", err)
+	}
+	for _, ref := range refs {
+		for _, prefix := range keys {
+			var exact, partial []*Cache
+			for _, cache := range caches {
+				if cache.Ref == ref && cache.Key == prefix {
+					exact = append(exact, cache)
+				} else if cache.Ref == ref && strings.HasPrefix(cache.Key, prefix) {
+					partial = append(partial, cache)
+				}
+			}
+			// if a key in the list matches exactly, don't return partial matches
+			if cache := cmp.Or(latest(exact, (*Cache).createdAtNano), latest(partial, (*Cache).createdAtNano)); cache != nil {
+				return cache, nil
+			}
 		}
 	}
 	return nil, nil //nolint:nilnil // pre-existing issue from nektos/act
@@ -1111,9 +1130,9 @@ func latest(caches []*Cache, stamp func(*Cache) int64) *Cache {
 // A completed entry is the one to restore, sorted by when it was written. An incomplete one is a
 // reservation being uploaded to, sorted by when it was last written to, because the upload route
 // touches UsedAt on every part.
-func findExactCache(db *cacheDB, repo, key, version string, complete bool) (*Cache, error) {
+func findExactCache(db *cacheDB, cred JobCredential, key, version string, complete bool) (*Cache, error) {
 	caches, err := db.find(func(cache *Cache) bool {
-		return cache.Key == key && cache.Repo == repo && cache.Version == version && cache.Complete == complete
+		return cred.owns(cache) && cache.Key == key && cache.Version == version && cache.Complete == complete
 	})
 	if err != nil {
 		return nil, fmt.Errorf("find cache: %w", err)
@@ -1280,13 +1299,12 @@ func (h *Handler) evictExpired(db *cacheDB) {
 	h.sweep(db, func(cache *Cache) bool { return cache.UsedAt < expired })
 }
 
-// evictSuperseded removes entries a newer one with the same key and version replaced. The
-// grouping includes Repo so two repos sharing a (key, version) do not evict each other.
+// evictSuperseded removes entries a newer one with the same repo, ref, key and version replaced.
 func (h *Handler) evictSuperseded(db *cacheDB) {
-	type entry struct{ repo, key, version string }
+	type entry struct{ repo, ref, key, version string }
 	groups := make(map[entry][]*Cache)
 	for _, cache := range h.caches(db, func(cache *Cache) bool { return cache.Complete }) {
-		group := entry{cache.Repo, cache.Key, cache.Version}
+		group := entry{cache.Repo, cache.Ref, cache.Key, cache.Version}
 		groups[group] = append(groups[group], cache)
 	}
 	for _, caches := range groups {