diff --git a/semaphore/semaphore.go b/semaphore/semaphore.go
index 30f632c..b618162 100644
--- a/semaphore/semaphore.go
+++ b/semaphore/semaphore.go
@@ -35,11 +35,25 @@
 // Acquire acquires the semaphore with a weight of n, blocking until resources
 // are available or ctx is done. On success, returns nil. On failure, returns
 // ctx.Err() and leaves the semaphore unchanged.
-//
-// If ctx is already done, Acquire may still succeed without blocking.
 func (s *Weighted) Acquire(ctx context.Context, n int64) error {
+	done := ctx.Done()
+
 	s.mu.Lock()
+	select {
+	case <-done:
+		// ctx becoming done has "happened before" acquiring the semaphore,
+		// whether it became done before the call began or while we were
+		// waiting for the mutex. We prefer to fail even if we could acquire
+		// the mutex without blocking.
+		s.mu.Unlock()
+		return ctx.Err()
+	default:
+	}
 	if s.size-s.cur >= n && s.waiters.Len() == 0 {
+		// Since we hold s.mu and haven't synchronized since checking done, if
+		// ctx becomes done before we return here, it becoming done must have
+		// "happened concurrently" with this call - it cannot "happen before"
+		// we return in this branch. So, we're ok to always acquire here.
 		s.cur += n
 		s.mu.Unlock()
 		return nil
@@ -48,7 +62,7 @@
 	if n > s.size {
 		// Don't make other Acquire calls block on one that's doomed to fail.
 		s.mu.Unlock()
-		<-ctx.Done()
+		<-done
 		return ctx.Err()
 	}
 
@@ -58,14 +72,14 @@
 	s.mu.Unlock()
 
 	select {
-	case <-ctx.Done():
-		err := ctx.Err()
+	case <-done:
 		s.mu.Lock()
 		select {
 		case <-ready:
-			// Acquired the semaphore after we were canceled.  Rather than trying to
-			// fix up the queue, just pretend we didn't notice the cancelation.
-			err = nil
+			// Acquired the semaphore after we were canceled.
+			// Pretend we didn't and put the tokens back.
+			s.cur -= n
+			s.notifyWaiters()
 		default:
 			isFront := s.waiters.Front() == elem
 			s.waiters.Remove(elem)
@@ -75,9 +89,19 @@
 			}
 		}
 		s.mu.Unlock()
-		return err
+		return ctx.Err()
 
 	case <-ready:
+		// Acquired the semaphore. Check that ctx isn't already done.
+		// We check the done channel instead of calling ctx.Err because we
+		// already have the channel, and ctx.Err is O(n) with the nesting
+		// depth of ctx.
+		select {
+		case <-done:
+			s.Release(n)
+			return ctx.Err()
+		default:
+		}
 		return nil
 	}
 }
diff --git a/semaphore/semaphore_test.go b/semaphore/semaphore_test.go
index 6e8eca2..61012d6 100644
--- a/semaphore/semaphore_test.go
+++ b/semaphore/semaphore_test.go
@@ -200,3 +200,38 @@
 	}
 	sem.Release(1)
 }
+
+func TestWeightedAcquireCanceled(t *testing.T) {
+	// https://go.dev/issue/63615
+	sem := semaphore.NewWeighted(2)
+	ctx, cancel := context.WithCancel(context.Background())
+	sem.Acquire(context.Background(), 1)
+	ch := make(chan struct{})
+	go func() {
+		// Synchronize with the Acquire(2) below.
+		for sem.TryAcquire(1) {
+			sem.Release(1)
+		}
+		// Now cancel ctx, and then release the token.
+		cancel()
+		sem.Release(1)
+		close(ch)
+	}()
+	// Since the context closing happens before enough tokens become available,
+	// this Acquire must fail.
+	if err := sem.Acquire(ctx, 2); err != context.Canceled {
+		t.Errorf("Acquire with canceled context returned wrong error: want context.Canceled, got %v", err)
+	}
+	// There must always be two tokens in the semaphore after the other
+	// goroutine releases the one we held at the start.
+	<-ch
+	if !sem.TryAcquire(2) {
+		t.Fatal("TryAcquire after canceled Acquire failed")
+	}
+	// Additionally verify that we don't acquire with a done context even when
+	// we wouldn't need to block to do so.
+	sem.Release(2)
+	if err := sem.Acquire(ctx, 1); err != context.Canceled {
+		t.Errorf("Acquire with canceled context returned wrong error: want context.Canceled, got %v", err)
+	}
+}