fix: fr32: Correct multithreaded Unpad chunk boundary alignment and simplify unpadReader (#13455)

* reader tests, new unpad reader

* fix: fr32: multithreaded Unpad chunk boundary alignment

* fr32: avoid allocation in NewUnpadReaderBuf

* add changelog entry

* pieceReader: use readPaddedSize for buffer sizing

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* address important review issue

* fix lint

---------

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Author: magik6k

CHANGELOG.md

@@ -45,6 +45,7 @@ The Lotus and Lotus-Miner v1.34.2 release includes numerous bug fixes, CLI enhan
 - fix(miner): ensure sender account exists ([filecoin-project/lotus#13348](https://github.com/filecoin-project/lotus/pull/13348))
 - fix(eth): properly return vm error in all gas estimation methods ([filecoin-project/lotus#13389](https://github.com/filecoin-project/lotus/pull/13389))
 - chore: all actor cmd support --actor ([filecoin-project/lotus#13391](https://github.com/filecoin-project/lotus/pull/13391))
+- fix(fr32): fix data corruption in multithreaded Pad/Unpad for non-aligned sizes
 
 ## 📝 Changelog
 

storage/sealer/fr32/fr32.go

@@ -41,7 +41,14 @@ func mtChunkCount(usz abi.PaddedPieceSize) uint64 {
 
 func mt(in, out []byte, padLen int, op func(unpadded, padded []byte)) {
 	threads := mtChunkCount(abi.PaddedPieceSize(padLen))
-	threadBytes := abi.PaddedPieceSize(padLen / int(threads))
+
+	// Ensure threadBytes is aligned to 128-byte chunk boundaries.
+	// Each fr32 chunk is 128 padded bytes / 127 unpadded bytes.
+	chunksPerThread := (padLen / int(threads)) / 128
+	if chunksPerThread == 0 {
+		chunksPerThread = 1
+	}
+	threadBytes := abi.PaddedPieceSize(chunksPerThread * 128)
 
 	var wg sync.WaitGroup
 	wg.Add(int(threads))
@@ -53,6 +60,16 @@ func mt(in, out []byte, padLen int, op func(unpadded, padded []byte)) {
 			start := threadBytes * abi.PaddedPieceSize(thread)
 			end := start + threadBytes
 
+			// Last thread takes any remainder
+			if thread == int(threads)-1 {
+				end = abi.PaddedPieceSize(padLen)
+			}
+
+			// Skip if this thread has no work
+			if start >= abi.PaddedPieceSize(padLen) {
+				return
+			}
+
 			op(in[start.Unpadded():end.Unpadded()], out[start:end])
 		}(i)
 	}

storage/sealer/fr32/fr32_test.go

@@ -3,6 +3,7 @@ package fr32_test
 import (
 	"bytes"
 	"crypto/rand"
+	"fmt"
 	"io"
 	"os"
 	"testing"
@@ -149,6 +150,134 @@ func TestRoundtrip16MRand(t *testing.T) {
 	require.Equal(t, ffi, buf)
 }
 
+// TestRoundtripMisalignedSizes tests the multithreaded Pad/Unpad with sizes that
+// previously caused data corruption due to thread boundary misalignment.
+// The bug occurred when (padLen / threads) was not a multiple of 128 bytes,
+// causing partial chunks at thread boundaries to be skipped.
+func TestRoundtripMisalignedSizes(t *testing.T) {
+	// These sizes are chosen to trigger the multithreaded path (> 512KB)
+	// and create thread boundaries that don't align to 128-byte chunks.
+	testCases := []struct {
+		name      string
+		numChunks int
+	}{
+		// 66061 chunks = 8455808 padded bytes
+		// With 16 threads: 8455808/16 = 528488 bytes per thread
+		// 528488/128 = 4128.5 - NOT aligned! This was the original bug case.
+		{"66061_chunks_8MiB_boundary", 66061},
+
+		// Various sizes that create misaligned thread boundaries
+		{"prime_chunks_1009", 1009 * 8},  // ~1MB, prime-ish number of chunks
+		{"odd_chunks_8193", 8193},        // Just over 8192 (power of 2)
+		{"odd_chunks_65537", 65537},      // Just over 65536 (power of 2)
+		{"odd_chunks_100003", 100003},    // Large prime
+		{"boundary_chunks_66000", 66000}, // Near the original bug size
+		{"boundary_chunks_70000", 70000}, // Larger odd size
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			unpaddedSize := tc.numChunks * 127
+			paddedSize := tc.numChunks * 128
+
+			// Skip if too large for this test
+			if paddedSize > 64<<20 {
+				t.Skip("Size too large for this test")
+			}
+
+			input := make([]byte, unpaddedSize)
+			_, err := rand.Read(input)
+			require.NoError(t, err)
+
+			padded := make([]byte, paddedSize)
+			fr32.Pad(input, padded)
+
+			output := make([]byte, unpaddedSize)
+			fr32.Unpad(padded, output)
+
+			require.Equal(t, input, output, "Roundtrip failed for %d chunks", tc.numChunks)
+		})
+	}
+}
+
+// TestUnpadMisalignedThreadBoundaries specifically tests the fix for the
+// multithreaded Unpad bug where thread boundaries weren't aligned to
+// 128-byte fr32 chunks, causing data loss.
+func TestUnpadMisalignedThreadBoundaries(t *testing.T) {
+	// Create data that's just over 8MiB to trigger the original bug
+	// 66061 chunks * 127 bytes = 8389747 unpadded bytes
+	// 66061 chunks * 128 bytes = 8455808 padded bytes
+	numChunks := 66061
+	unpaddedSize := numChunks * 127
+	paddedSize := numChunks * 128
+
+	// Create sequential data so we can detect exactly where corruption occurs
+	input := make([]byte, unpaddedSize)
+	for i := range input {
+		input[i] = byte(i & 0xFF)
+	}
+
+	padded := make([]byte, paddedSize)
+	fr32.Pad(input, padded)
+
+	output := make([]byte, unpaddedSize)
+	fr32.Unpad(padded, output)
+
+	// Check for corruption at thread boundaries
+	// With the original bug, corruption occurred at offsets like:
+	// 528384 (thread 0/1 boundary), 1056768 (thread 1/2 boundary), etc.
+
+	// First verify total length
+	require.Equal(t, len(input), len(output), "Output length mismatch")
+
+	// Check every byte
+	for i := 0; i < len(input); i++ {
+		if input[i] != output[i] {
+			// Find the extent of the corruption
+			corruptStart := i
+			corruptEnd := i
+			for corruptEnd < len(input) && input[corruptEnd] != output[corruptEnd] {
+				corruptEnd++
+			}
+			t.Fatalf("Data corruption at offset %d (0x%x) to %d (0x%x): expected 0x%02x, got 0x%02x (corrupt bytes: %d)",
+				corruptStart, corruptStart, corruptEnd, corruptEnd,
+				input[i], output[i], corruptEnd-corruptStart)
+		}
+	}
+}
+
+// TestPadUnpadVariousSizesAboveMTTresh tests Pad/Unpad roundtrip for various
+// sizes above the MTTresh (512KB) threshold that triggers multithreading.
+func TestPadUnpadVariousSizesAboveMTTresh(t *testing.T) {
+	// Test sizes from just above MTTresh to several MB
+	// These should all use the multithreaded path
+	sizes := []int{
+		513 * 1024 / 127 * 127,        // Just above 512KB, aligned to chunks
+		1 * 1024 * 1024 / 127 * 127,   // ~1MB aligned
+		2*1024*1024/127*127 + 127*100, // ~2MB + extra chunks
+		4*1024*1024/127*127 + 127*333, // ~4MB + odd chunks
+		8*1024*1024/127*127 + 127*777, // ~8MB + odd chunks
+	}
+
+	for _, unpaddedSize := range sizes {
+		paddedSize := unpaddedSize / 127 * 128
+
+		t.Run(fmt.Sprintf("%d_bytes", unpaddedSize), func(t *testing.T) {
+			input := make([]byte, unpaddedSize)
+			_, err := rand.Read(input)
+			require.NoError(t, err)
+
+			padded := make([]byte, paddedSize)
+			fr32.Pad(input, padded)
+
+			output := make([]byte, unpaddedSize)
+			fr32.Unpad(padded, output)
+
+			require.Equal(t, input, output)
+		})
+	}
+}
+
 func BenchmarkPadChunk(b *testing.B) {
 	var buf [128]byte
 	in := bytes.Repeat([]byte{0xff}, 127)