Summary
ShellStream.ExpectRegex decodes the read buffer to a string, matches against that string, and then
discards the byte count of the re-encoded match:
var bufferText = GetString(_readBuffer.ActiveLength);
...
var returnText = bufferText.Substring(0, match.Index + match.Length);
_readBuffer.Discard(_encoding.GetByteCount(returnText));
GetByteCount(returnText) equals the number of buffer bytes returnText was decoded from only when
the decode was lossless. It is not, whenever the consumed region contained a byte the encoding could
not decode: GetString replaces each maximal invalid subpart with U+FFFD, which re-encodes to
3 bytes regardless of how many bytes it replaced.
So Discard is asked for more bytes than the match occupied. And ArrayBuffer.Discard only checks
that bound with a Debug.Assert:
public void Discard(int byteCount)
{
Debug.Assert(byteCount <= ActiveLength, $"Expected {byteCount} <= {ActiveLength}");
_activeStart += byteCount;
if (_activeStart == _availableStart)
{
_activeStart = 0;
_availableStart = 0;
}
}
In a Release build the assert is gone, _activeStart overshoots past _availableStart, and
ActiveLength (_availableStart - _activeStart) goes negative. The _activeStart == _availableStart
reset cannot fire either, because the overshoot skipped equality — so the buffer never recovers.
The next iteration then calls GetString(_readBuffer.ActiveLength) with a negative length, and throws.
-
|
public void Discard(int byteCount) |
What it looks like in production
System.ArgumentOutOfRangeException: Non-negative number required. (Parameter 'count')
at Renci.SshNet.ShellStream.ExpectRegex(TimeSpan timeout, Int32 lookback, ExpectAction[] expectActions)
at <our shell wrapper>.ExecuteCommandAsync(String command)
Reproduced across several runs and several hosts, against Junos network devices, reading a large
configuration and then issuing a second command on the same ShellStream.
Note this is a Release-only symptom. Under Debug the assert fires instead, so a test run would
report the over-discard at its origin rather than as a later crash.
Minimal repro of the arithmetic (no SSH server needed)
The defect reduces to this identity. With the default Encoding.UTF8:
foreach (var bytes in new[]
{
new byte[] { 0x61, 0xF8, 0x62 }, // 'a', a lone 0xF8, 'b'
new byte[] { 0x61, 0xE2, 0x82, 0x62 }, // a truncated 3-byte sequence
new byte[] { 0x61, 0xFF, 0xFF, 0x62 },
new byte[] { 0x61, 0xC3, 0xB8, 0x62 }, // valid UTF-8, for contrast
})
{
var text = Encoding.UTF8.GetString(bytes);
Console.WriteLine($"in={bytes.Length} out={Encoding.UTF8.GetByteCount(text)}");
}
in=3 out=5 <- Discard() is asked for 2 bytes more than the match occupied
in=4 out=5 <- 1 byte more
in=4 out=8 <- 4 bytes more
in=4 out=4 <- lossless decode, correct
Every case where the two disagree drives ActiveLength negative by the difference.
ExpectRegex is the only one of the ten Discard(...) call sites in ShellStream that computes its
count by re-encoding; the others count in byte space (ActiveLength, indexOf…, bytesRead) and are
not affected.
Suggested direction
The number of bytes to discard is a fact about the input bytes, so re-encoding the output string
cannot recover it. Options, roughly by intrusiveness:
- Decode with a
Decoder and track how many input bytes produced the consumed prefix.
- Keep a char-index → byte-offset mapping while building
bufferText.
- Independently, consider making
ArrayBuffer.Discard clamp or throw rather than Debug.Assert, so an
over-discard surfaces where it happens instead of as an ArgumentOutOfRangeException one call later.
Happy to test a patch — we have hosts that reproduce it reliably.
Environment
- SSH.NET 2026.0.0. The two lines above are byte-identical in 2025.1.0, 2026.0.0 and
develop.
- .NET 10, Release build, Linux container.
Summary
ShellStream.ExpectRegexdecodes the read buffer to a string, matches against that string, and thendiscards the byte count of the re-encoded match:
GetByteCount(returnText)equals the number of buffer bytesreturnTextwas decoded from only whenthe decode was lossless. It is not, whenever the consumed region contained a byte the encoding could
not decode:
GetStringreplaces each maximal invalid subpart with U+FFFD, which re-encodes to3 bytes regardless of how many bytes it replaced.
So
Discardis asked for more bytes than the match occupied. AndArrayBuffer.Discardonly checksthat bound with a
Debug.Assert:In a Release build the assert is gone,
_activeStartovershoots past_availableStart, andActiveLength(_availableStart - _activeStart) goes negative. The_activeStart == _availableStartreset cannot fire either, because the overshoot skipped equality — so the buffer never recovers.
The next iteration then calls
GetString(_readBuffer.ActiveLength)with a negative length, and throws.SSH.NET/src/Renci.SshNet/ShellStream.cs
Line 548 in f099365
SSH.NET/src/Renci.SshNet/Common/ArrayBuffer.cs
Line 106 in f099365
What it looks like in production
Reproduced across several runs and several hosts, against Junos network devices, reading a large
configuration and then issuing a second command on the same
ShellStream.Note this is a Release-only symptom. Under Debug the assert fires instead, so a test run would
report the over-discard at its origin rather than as a later crash.
Minimal repro of the arithmetic (no SSH server needed)
The defect reduces to this identity. With the default
Encoding.UTF8:Every case where the two disagree drives
ActiveLengthnegative by the difference.ExpectRegexis the only one of the tenDiscard(...)call sites inShellStreamthat computes itscount by re-encoding; the others count in byte space (
ActiveLength,indexOf…,bytesRead) and arenot affected.
Suggested direction
The number of bytes to discard is a fact about the input bytes, so re-encoding the output string
cannot recover it. Options, roughly by intrusiveness:
Decoderand track how many input bytes produced the consumed prefix.bufferText.ArrayBuffer.Discardclamp or throw rather thanDebug.Assert, so anover-discard surfaces where it happens instead of as an
ArgumentOutOfRangeExceptionone call later.Happy to test a patch — we have hosts that reproduce it reliably.
Environment
develop.