Update

vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash.go

@@ -52,6 +52,11 @@ func (d *Digest) Reset() {
 	d.n = 0
 }
 
+// maxAsmSize bounds the input handed to a single writeBlocks call. Assembly is
+// never preemptible, so an unbounded call holds every P in stop-the-world for
+// its duration. Whole 32-byte blocks, so Write can feed it in pieces.
+const maxAsmSize = 128 << 10 // 4096 whole 32-byte blocks
+
 // Size always returns 8 bytes.
 func (d *Digest) Size() int { return 8 }
 
@@ -83,6 +88,13 @@ func (d *Digest) Write(b []byte) (n int, err error) {
 		d.n = 0
 	}
 
+	for len(b) >= maxAsmSize {
+		// Assembly is not preemptible, so hashing a large buffer in one call
+		// blocks every stop-the-world for the whole call. Feed it in chunks.
+		writeBlocks(d, b[:maxAsmSize])
+		b = b[maxAsmSize:]
+	}
+
 	if len(b) >= 32 {
 		// One or more full blocks left.
 		nw := writeBlocks(d, b)

vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_amd64.s

@@ -175,7 +175,11 @@ finalize:
 	RET
 
 // func writeBlocks(d *Digest, b []byte) int
-TEXT ·writeBlocks(SB), NOSPLIT|NOFRAME, $0-40
+//
+// Deliberately not NOSPLIT: the stack check is the preemption point that lets a
+// stop-the-world proceed between chunks. The frame must be >= abi.StackSmall
+// (128), or the assembler marks this leaf NOSPLIT anyway and drops that check.
+TEXT ·writeBlocks(SB), $128-40
 	// Load fixed primes needed for round.
 	MOVQ ·primes+0(SB), prime1
 	MOVQ ·primes+8(SB), prime2

vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_arm64.s

@@ -163,7 +163,11 @@ finalize:
 	RET
 
 // func writeBlocks(s *Digest, b []byte) int
-TEXT ·writeBlocks(SB), NOSPLIT|NOFRAME, $0-40
+//
+// Deliberately not NOSPLIT: the stack check is the preemption point that lets a
+// stop-the-world proceed between chunks. The frame must be >= abi.StackSmall
+// (128), or the assembler marks this leaf NOSPLIT anyway and drops that check.
+TEXT ·writeBlocks(SB), $128-40
 	LDP ·primes+0(SB), (prime1, prime2)
 
 	// Load state. Assume v[1-4] are stored contiguously.

vendor/github.com/klauspost/compress/zstd/internal/xxhash/xxhash_other.go

@@ -1,5 +1,4 @@
 //go:build (!amd64 && !arm64) || appengine || !gc || purego || noasm
-// +build !amd64,!arm64 appengine !gc purego noasm
 
 package xxhash