OPINIONATED OMARCHY
TROUBLESHOOTING CORPUS
790
RECORDS
790
AUDITED
0
UNCHECKED
2388
SOURCES
LIVE
network medium common DISPUTE

Fix SSH and HTTPS that hang mid-transfer on a VPN, PPPoE line or hotspot

omarchyarchendeavouroscachyosmanjarolaptopdesktop
CORRECTED An auditor found the fix wrong or incomplete and rewrote it.
AUDIT NOTE

Every claim in the record checks out, and two Omarchy 4 facts it ignores change the first diagnostic step, so this is corrected by addition rather than by repair. Confirmed on this machine (omarchy 4.0.2-1, kernel 7.1.9-arch1-2, iputils 20250605-1, networkmanager 1.58.1-1): ping -h lists -M <pmtud opt> taking do, dont, want or probe, and the ping binary carries both Frag needed and DF set (mtu = %u) and local error: message too long, mtu: %u. The record quotes only the first, so I added the second, because it distinguishes a local interface limit from a real path black hole. Property names verified in man 5 nm-settings-nmcli: 802-3-ethernet.mtu and 802-11-wireless.mtu both read "If non-zero, only transmit packets of the specified size or smaller, breaking larger packets up into multiple Ethernet frames", which is what the record quotes. Read the Arch WireGuard page in full: it gives 1420 as the default, states that below 1280 wg-quick may fail to create the interface, and carries the signature near verbatim as "ICMP ping works because of its low packet size, but most of TCP connections fail because of full MTU size utilization". The arithmetic all holds: 1500 minus 8 is 1492 for PPPoE, payload plus 28 is the IPv4 MTU, and the WireGuard overheads of 60 and 80 are right.

First addition, and the substantive one. Omarchy 4 ships net.ipv4.tcp_mtu_probing=1 in /etc/sysctl.d/99-omarchy-sysctl.conf, owned by omarchy-settings 4.0.2-1 per pacman -Qo, under the comment "Solve common flakiness with SSH (MTU discovery on flaky links)". That is this record's headline symptom named in Omarchy's own source, and the same line is present on the upstream quattro branch. The kernel documents the value as "1 - Disabled by default, enabled when an ICMP black hole detected", and net/ipv4/tcp_ipv4.c never initialises sysctl_tcp_mtu_probing, so the built-in default is 0 and plain Arch leaves it there. Nothing under /usr/lib/sysctl.d/ sets it, so on this machine the Omarchy drop-in is the only writer, and sysctl net.ipv4.tcp_mtu_probing returns 1. This is exactly the class of defect the brief asks for, a setting Omarchy assigns in a drop-in that the record ignores. It cuts both ways: on Omarchy the reader should check it before touching any MTU, and on plain Arch setting it is a cheaper remedy than the per-profile MTU edits the record jumps straight to. I also confirmed net.ipv4.tcp_base_mss is 1024 and net.ipv4.tcp_mtu_probe_floor is 48 locally, matching TCP_BASE_MSS and TCP_MIN_SND_MSS in the kernel source, which is why the new danger warns against value 2.

Second addition. Omarchy runs ufw with DEFAULT_INPUT_POLICY="DROP" and ENABLED=yes, which makes the local firewall the obvious suspect for the filtered ICMP the cause describes, and it is the wrong suspect. /etc/ufw/before.rules line 34 is -A ufw-before-input -p icmp --icmp-type destination-unreachable -j ACCEPT and line 40 is the same for ufw-before-forward, and fragmentation-needed is a code of destination-unreachable. /etc/ufw/sysctl.conf sets net/ipv4/icmp_echo_ignore_all=0 and touches nothing MTU related. So ufw does not cause this and the cause now says so. Read unprivileged, no changes made.

Third, smaller. The record's clamping advice is correctly scoped to a machine that routes for others, so I kept it, but I named the two Omarchy specifics it leaves out. /etc/ufw/before.rules contains only a *filter table (*filter at line 12, COMMIT at line 75), so the mangle rule cannot live there, and man ufw-framework documents /etc/ufw/before.init as the initialization customization script ufw-init runs if present and executable, which is the right home. /etc/default/ufw sets DEFAULT_FORWARD_POLICY="DROP", so clamping alone will not make this box route. I also changed the one-off ip link set example from wlan0 to wlo1, because predictable naming is on here (no net.ifnames=0 in /proc/cmdline, no .link override) and the actual device is wlo1.

Not exercised. I did not set an MTU, add or modify a connection, change a sysctl, run any ping at all, or use sudo, per the operator's instruction on a live workstation. So no path MTU black hole was induced and no remedy was observed working end to end. The MTU arithmetic, the WireGuard behaviour and the clamping rule come from the cited sources rather than from a local test. What I verified locally is version and file state: the sysctl file and its owner, the ufw rule text, the ping binary's strings and flags, the nmcli property documentation, and the interface names. All four cited URLs still resolve and all four still say what the record claims, so nothing is removed.

The cause above was rewritten on 2026-09-11 to match this note; the fix was corrected by the audit itself.
RISK

Lowering MTU costs a little throughput on paths that did not need it, so scope the change to the connection profile that is broken rather than applying it globally. Do not set an MTU below 1280 on any interface carrying IPv6, because it is below the protocol minimum and will break IPv6 outright. Set net.ipv4.tcp_mtu_probing to 1 and not 2: at 2 probing is always on and starts from net.ipv4.tcp_base_mss, which defaults to 1024, so every healthy path pays for it. Add your own drop-in under /etc/sysctl.d/ rather than editing /etc/sysctl.d/99-omarchy-sysctl.conf, which is owned by the omarchy-settings package and will be replaced on update. The iptables clamping rule is for a machine that forwards traffic, and on a workstation it is a no-op that looks like a fix.

Symptom

Small things work and big things stall. ping succeeds, DNS resolves, an SSH banner appears and then the session freezes the moment you run something that prints a lot. git clone and apt/pacman downloads hang at a few percent forever with no error. Some HTTPS sites load and others hang after the TLS handshake. It happens on a VPN (WireGuard, corporate IPsec), behind a PPPoE DSL/fibre modem, or on certain hotel and mobile hotspots, and the same machine is fine on other networks.

Cause

A path MTU black hole. Something on the path has an MTU smaller than yours, the router that needs to fragment sees the DF bit and drops the packet, and the ICMP "fragmentation needed" message that would tell your kernel to shrink is filtered out somewhere along the way. Path MTU Discovery never completes, so your host keeps firing full-size segments into a hole. Small packets (ping, DNS, the SSH banner, the TLS handshake) fit and get through. Anything at full MSS does not. PPPoE takes 8 bytes off 1500, and WireGuard takes 60 for IPv4 or 80 for IPv6. The Arch WireGuard page describes this exact signature: ICMP ping works because of its low packet size while most TCP connections fail.

The filtering is almost never on your own machine. Omarchy's ufw is default-deny incoming, which makes it the obvious suspect, but /etc/ufw/before.rules accepts ICMP destination-unreachable in both ufw-before-input and ufw-before-forward, and "fragmentation needed" is a code of that type. Look upstream instead.

Omarchy 4 also differs from plain Arch here, and it matters for the first step. It ships net.ipv4.tcp_mtu_probing=1 in /etc/sysctl.d/99-omarchy-sysctl.conf, commented "Solve common flakiness with SSH (MTU discovery on flaky links)", which is this symptom by name. The kernel documents that value as disabled by default and enabled when an ICMP black hole is detected, so on Omarchy the kernel should already probe its way down on its own. Plain Arch sets nothing and leaves the kernel default of 0, disabled. If you are on Omarchy 4 and still seeing this, either the sysctl is not in effect or the hole is on a path this host only forwards.

Fix

First, on Omarchy 4, check that the mitigation it ships is actually live. On plain Arch this is the cheapest real fix and the record's remaining steps are the fallback:


sysctl net.ipv4.tcp_mtu_probing
cat /etc/sysctl.d/99-omarchy-sysctl.conf

1 means the kernel enables packetization-layer probing once it detects a black hole. 0 means it will never try. To set it on plain Arch, in a drop-in of your own rather than by editing Omarchy's file:


printf 'net.ipv4.tcp_mtu_probing=1\n' | sudo tee /etc/sysctl.d/99-local-mtu.conf
sudo sysctl --system

Then find the largest payload that survives the path. -M do sets DF, and the IPv4 header plus ICMP header is 28 bytes, so working MTU equals payload plus 28:


ping -M do -s 1472 -c 3 1.1.1.1      # 1472 + 28 = 1500
ping -M do -s 1452 -c 3 1.1.1.1      # 1480
ping -M do -s 1392 -c 3 1.1.1.1      # 1420
ping -M do -s 1272 -c 3 1.1.1.1      # 1300

A size too large for the path prints Frag needed and DF set (mtu = NNNN) when a router tells you, or simply times out when the ICMP is being filtered, which is the black hole case. A size too large for your own interface fails immediately with local error: message too long, mtu: NNNN instead, which is your NIC and not the path. Take the largest payload that succeeds and add 28.

**Ethernet, via NetworkManager:**


nmcli connection modify "<connection-name>" 802-3-ethernet.mtu 1400
nmcli connection up "<connection-name>"

**Wi-Fi, via NetworkManager:**


nmcli connection modify "<connection-name>" 802-11-wireless.mtu 1400
nmcli connection up "<connection-name>"

(mtu is documented for both settings as "If non-zero, only transmit packets of the specified size or smaller". 0 restores the default.)

**PPPoE:** the ceiling is 1492. Anything above that will black-hole by definition.

**WireGuard:** set it in the interface, not on the underlying NIC:


# /etc/wireguard/wg0.conf
[Interface]
Address = 10.200.200.2/24
MTU = 1420
PrivateKey = <key>

sudo wg-quick down wg0 && sudo wg-quick up wg0
ip link show wg0 | grep mtu

1420 is the WireGuard default. Drop to 1380, then 1280 if the tunnel itself is riding over PPPoE or a mobile link. 1280 is the IPv6 minimum and is the safe floor, and wg-quick refuses to create the interface below it.

**Set it without a manager (one-off test):**


sudo ip link set dev wlo1 mtu 1400

**If this machine routes for others** (a Tailscale subnet router, a hotspot, a container host), clamp TCP MSS to the real path MTU instead of guessing per-client. A plain workstation behind a router is not the place for this, because nothing it sends is forwarded and the rule will never match:


sudo iptables -t mangle -A FORWARD -p tcp --tcp-flags SYN,RST SYN \
  -j TCPMSS --clamp-mss-to-pmtu

On Omarchy, make that persistent through ufw rather than a raw iptables call at boot, and note two things. /etc/ufw/before.rules holds only a *filter table, so a mangle rule cannot go there. The supported home is /etc/ufw/before.init, which man ufw-framework describes as an initialization customization script that ufw-init executes if it exists and is executable. And /etc/default/ufw sets DEFAULT_FORWARD_POLICY="DROP", so clamping alone will not make this machine route for anyone until forwarding is permitted as well.

Verify

ip link show <iface> reports the new MTU. ping -M do -s $((MTU-28)) -c 3 1.1.1.1 succeeds while one byte larger fails. sysctl net.ipv4.tcp_mtu_probing reports 1 if you took that route. Then reproduce the original failure: ssh <host> 'yes | head -100000' runs to completion, and a git clone of a real repository finishes.

Sources

pmtu-blackhole-large-transfers-hang