fix(installer): release the target disk before archinstall partitions it

The live ISO auto-activates LVM/md/swap found on attached drives. The
first hardware bench's SSD still carried a Proxmox VE 'pve' VG, so six
device-mapper nodes sat on sda3 and archinstall died at 12 % with
'unable to inform the kernel of the change ... in use'.

install_run now calls diskprep.release_disk() first: swapoff/umount
leaves-first, dmsetup remove / mdadm --stop for every dm/md node stacked
on the disk (retrying until the stack is gone), wipefs -a on partitions
and disk, blockdev --rereadpt. Steps are logged at the top of the install
log (spawn_archinstall opens it in append mode now). The progress page
names this failure explicitly if it still happens.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MScAinbyMdeNc7H2BZdnnG

CHANGELOG.md

@@ -9,6 +9,17 @@ This project uses calendar versioning: `YY.N-stage` (e.g. `26.0-alpha` = 2026, r
 
 ### Fixed
 
+- **Installer frees the target disk before partitioning.** The live ISO
+  auto-activates whatever an attached drive carries (LVM volume groups, md
+  arrays, swap) — the first hardware bench's SSD still had a Proxmox VE
+  `pve` VG on it — and archinstall then failed at 12 % with `unable to
+  inform the kernel of the change, probably because it/they are in use`.
+  `install_run` now runs `diskprep.release_disk()` first: swapoff/umount
+  leaves-first, `dmsetup remove` / `mdadm --stop` for every dm/md node
+  stacked on the disk, `wipefs -a` on partitions + disk, then
+  `blockdev --rereadpt`. Every step is written to the top of the install
+  log. The progress page also names this failure explicitly instead of a
+  bare "hit a snag" if it does still happen.
 - **Console banner always shows the IP fallback.** On both the live ISO and
   the installed system the `/etc/issue` welcome was written once, after
   `network-online.target`, and simply omitted the `http://<ip>` line when no

tests/test_diskprep.py

@@ -0,0 +1,146 @@
+"""release_disk() against a recorded `lsblk -P` picture of the first
+hardware bench: an SSD with an old Proxmox VE `pve` volume group stacked on
+sda3 (swap, root, thin pool). Commands are captured through the injected
+runner — nothing touches real devices."""
+
+import subprocess
+from pathlib import Path
+
+import pytest
+from diskprep import release_disk
+
+# What the bench looked like before cleanup (lsblk -lnP -o NAME,TYPE,MOUNTPOINTS /dev/sda).
+BENCH_BEFORE = """NAME="sda" TYPE="disk" MOUNTPOINTS=""
+NAME="sda1" TYPE="part" MOUNTPOINTS=""
+NAME="sda3" TYPE="part" MOUNTPOINTS=""
+NAME="pve-swap" TYPE="lvm" MOUNTPOINTS="[SWAP]"
+NAME="pve-root" TYPE="lvm" MOUNTPOINTS="/mnt/old"
+NAME="pve-data_tmeta" TYPE="lvm" MOUNTPOINTS=""
+NAME="pve-data-tpool" TYPE="lvm" MOUNTPOINTS=""
+NAME="pve-data" TYPE="lvm" MOUNTPOINTS=""
+NAME="pve-data_tdata" TYPE="lvm" MOUNTPOINTS=""
+NAME="pve-data-tpool" TYPE="lvm" MOUNTPOINTS=""
+NAME="pve-data" TYPE="lvm" MOUNTPOINTS=""
+"""
+
+BENCH_CLEAN = """NAME="sda" TYPE="disk" MOUNTPOINTS=""
+NAME="sda1" TYPE="part" MOUNTPOINTS=""
+NAME="sda3" TYPE="part" MOUNTPOINTS=""
+"""
+
+EMPTY = """NAME="sda" TYPE="disk" MOUNTPOINTS=""
+"""
+
+
+class FakeRunner:
+    """Answers lsblk from a script of snapshots (one per call, last one
+    sticks) and records everything else."""
+
+    def __init__(self, lsblk_outputs):
+        self.lsblk_outputs = list(lsblk_outputs)
+        self.calls = []
+
+    def __call__(self, cmd, capture_output=True, text=True):
+        if cmd[0] == "lsblk":
+            if len(self.lsblk_outputs) > 1:
+                out = self.lsblk_outputs.pop(0)
+            else:
+                out = self.lsblk_outputs[0]
+            return subprocess.CompletedProcess(cmd, 0, out, "")
+        self.calls.append(cmd)
+        return subprocess.CompletedProcess(cmd, 0, "", "")
+
+
+def test_release_tears_down_lvm_stack_leaves_first():
+    # lsblk #1 (inventory) and #2 (first removal pass) see the full stack;
+    # everything after that is clean — the dm removals "worked".
+    runner = FakeRunner([BENCH_BEFORE, BENCH_BEFORE, BENCH_CLEAN])
+    lines = release_disk("/dev/sda", run=runner)
+    calls = runner.calls
+
+    assert ["swapoff", "/dev/pve-swap"] in calls
+    assert ["umount", "--all-targets", "/dev/pve-root"] in calls
+    removed = [c[2] for c in calls if c[:2] == ["dmsetup", "remove"]]
+    # Thin LV before its pool, pool before tmeta/tdata.
+    assert removed.index("pve-data") < removed.index("pve-data-tpool")
+    assert removed.index("pve-data-tpool") < removed.index("pve-data_tmeta")
+    assert set(removed) == {
+        "pve-swap",
+        "pve-root",
+        "pve-data_tmeta",
+        "pve-data-tpool",
+        "pve-data",
+        "pve-data_tdata",
+    }
+    assert ["wipefs", "-a", "/dev/sda3"] in calls
+    assert ["wipefs", "-a", "/dev/sda1"] in calls
+    assert ["wipefs", "-a", "/dev/sda"] in calls
+    assert calls[-1] == ["blockdev", "--rereadpt", "/dev/sda"]
+    assert not any("WARNING" in line for line in lines)
+    assert lines[0].startswith("Releasing /dev/sda")
+
+
+def test_release_warns_when_stack_survives():
+    runner = FakeRunner([BENCH_BEFORE])  # never becomes clean
+    lines = release_disk("/dev/sda", run=runner)
+    assert any("WARNING: still in use" in line for line in lines)
+    # Still tries to give archinstall a chance.
+    assert ["blockdev", "--rereadpt", "/dev/sda"] in runner.calls
+
+
+def test_release_empty_disk_only_wipes_and_rereads():
+    runner = FakeRunner([EMPTY])
+    lines = release_disk("sda", run=runner)  # bare name is accepted too
+    assert runner.calls == [["wipefs", "-a", "/dev/sda"], ["blockdev", "--rereadpt", "/dev/sda"]]
+    assert any("nothing stacked" in line for line in lines)
+
+
+def test_release_survives_missing_tools():
+    def runner(cmd, capture_output=True, text=True):
+        if cmd[0] == "lsblk":
+            return subprocess.CompletedProcess(cmd, 0, BENCH_BEFORE, "")
+        raise FileNotFoundError(cmd[0])
+
+    lines = release_disk("/dev/sda", run=runner)
+    assert any("command not found" in line for line in lines)
+
+
+@pytest.fixture
+def app_module(monkeypatch, tmp_path):
+    import app as app_module
+
+    monkeypatch.setattr(app_module, "INSTALL_LOG", tmp_path / "install.log")
+    monkeypatch.setattr(app_module, "STATE_DIR", tmp_path / "state")
+    return app_module
+
+
+def test_install_run_releases_disk_before_spawning(app_module, monkeypatch):
+    order = []
+
+    def fake_release(disk):
+        order.append(("release", disk))
+        return ["Releasing X"]
+
+    monkeypatch.setattr(app_module, "release_disk", fake_release)
+    monkeypatch.setattr(app_module, "spawn_archinstall", lambda *a: order.append(("spawn",)))
+    monkeypatch.setattr(app_module, "build_archinstall_config", lambda s: {})
+    monkeypatch.setattr(app_module, "build_archinstall_creds", lambda s: {})
+    monkeypatch.delenv("FURTKA_DRY_RUN", raising=False)
+    app_module.settings.update(boot_drive="/dev/sda", username="u", password="pw12345678")
+
+    client = app_module.app.test_client()
+    resp = client.post("/install/run")
+
+    assert resp.status_code == 302
+    assert order == [("release", "/dev/sda"), ("spawn",)]
+    assert Path(app_module.INSTALL_LOG).read_text().startswith("Releasing X")
+
+
+def test_progress_names_disk_in_use_failure(app_module):
+    log = (
+        "Traceback (most recent call last):\n"
+        "_ped.IOException: ... unable to inform the kernel of the change ..."
+    )
+    progress = app_module.parse_install_progress(log)
+    assert progress["status"] == "error"
+    assert "still in use" in progress["phase"]

webinstaller/app.py

@@ -11,6 +11,7 @@ import sys
 from datetime import UTC
 from pathlib import Path
 
+from diskprep import release_disk
 from drives import list_scored_devices
 from flask import Flask, jsonify, redirect, render_template, request, url_for
 
@@ -96,6 +97,16 @@ PROGRESS_PHASES = [
 
 PROGRESS_ERROR_MARKERS = ("Traceback (most recent call last)", "archinstall: error:")
 
+# Known failure signatures → a phase label that tells the user what to do
+# instead of a bare "open Show details". Checked in order; first hit wins.
+PROGRESS_ERROR_HINTS = [
+    (
+        "unable to inform the kernel of the change",
+        "Installation failed — the disk was still in use (old LVM/RAID/swap "
+        "from a previous system). Reboot the installer and try again",
+    ),
+]
+
 
 def parse_install_progress(log):
     percent = 2
@@ -110,6 +121,10 @@ def parse_install_progress(log):
     elif any(m in log for m in PROGRESS_ERROR_MARKERS):
         status = "error"
         phase = "Installation failed — open Show details below"
+        for marker, hint in PROGRESS_ERROR_HINTS:
+            if marker in log:
+                phase = hint
+                break
     else:
         status = "running"
 
@@ -264,6 +279,9 @@ _FURTKA_UNITS = (
     "furtka-status.service",
     "furtka-status.timer",
     "furtka-welcome.service",
+    # Re-runs the welcome banner every few seconds so the console always
+    # shows the current IP fallback (or a "no IP yet" hint).
+    "furtka-welcome.timer",
     # Daily apps-catalog pull. Timer drives the service; the .service itself
     # is oneshot and also callable ad-hoc via `furtka catalog sync`.
     "furtka-catalog-sync.service",
@@ -279,9 +297,6 @@ def _resource_manager_commands():
     file isn't present (dev box without an ISO build), returns [] so the rest
     of the install still works — the resource manager just won't be installed,
     and nothing else on the system references furtka-* units.
-    # Re-runs the welcome banner every few seconds so the console always
-    # shows the current IP fallback (or a "no IP yet" hint).
-    "furtka-welcome.timer",
     """
     if not RESOURCE_MANAGER_PAYLOAD.exists():
         print(
@@ -537,7 +552,9 @@ def write_install_files(s, state_dir):
 
 
 def spawn_archinstall(config_path, creds_path, log_path):
-    log_fh = open(log_path, "wb")
+    # Append: install_run() already truncated the log and may have written
+    # the disk-release lines we want to keep visible above archinstall's own.
+    log_fh = open(log_path, "ab")
     return subprocess.Popen(
         [
             "archinstall",
@@ -609,6 +626,11 @@ def install_run():
             f"--creds {creds_path} --silent\n"
         )
     else:
+        # Free the target disk first: udev auto-activates old LVM/RAID/swap
+        # on the live ISO, and archinstall can't re-read a partition table
+        # while those hold the partitions open (first hardware bench, 26.19).
+        with INSTALL_LOG.open("a") as fh:
+            fh.write("\n".join(release_disk(settings["boot_drive"])) + "\n\n")
         spawn_archinstall(config_path, creds_path, INSTALL_LOG)
     return redirect(url_for("install_log_view"))
 

webinstaller/diskprep.py

@@ -0,0 +1,125 @@
+"""Release a target disk before handing it to archinstall.
+
+The live ISO ships lvm2, mdadm, dmraid and cryptsetup (releng defaults), and
+udev happily auto-activates whatever it finds on attached disks at boot. A
+drive that used to live in another machine — the first hardware bench had an
+old Proxmox VE install on its SSD — therefore comes up with device-mapper
+volumes (or swap, or md arrays) stacked on top of its partitions. archinstall
+then writes the new partition table fine but `BLKRRPART` fails with
+
+    Partition(s) 2, 3 on /dev/sda have been written, but we have been unable
+    to inform the kernel of the change, probably because it/they are in use.
+
+and the install dies at 12 %. Nothing on a fresh VM ever triggers this,
+which is why it survived every smoke run.
+
+`release_disk()` tears that stack down leaves-first (swapoff, umount, dmsetup
+remove / mdadm --stop), wipes signatures and asks the kernel to re-read the
+partition table. Every step is logged so the install log shows what was
+done. All commands are best-effort: the point is to leave archinstall a
+disk the kernel can re-partition, not to be a full disk-management tool.
+"""
+
+import re
+import subprocess
+
+_PAIR_RE = re.compile(r'(\w+)="([^"]*)"')
+
+
+def _run(cmd, run):
+    try:
+        result = run(cmd, capture_output=True, text=True)
+    except FileNotFoundError:
+        return 127, f"{cmd[0]}: command not found"
+    out = ((result.stdout or "") + (result.stderr or "")).strip()
+    return result.returncode, out
+
+
+def _parse_pairs(output):
+    """Parse `lsblk -P` key="value" lines into dicts."""
+    rows = []
+    for line in output.splitlines():
+        row = {}
+        for match in _PAIR_RE.finditer(line):
+            row[match.group(1)] = match.group(2)
+        if row.get("NAME"):
+            rows.append(row)
+    return rows
+
+
+def _children(disk, run):
+    """Everything stacked on `disk`, in lsblk tree order, disk itself excluded."""
+    rc, out = _run(["lsblk", "-lnP", "-o", "NAME,TYPE,MOUNTPOINTS", disk], run)
+    if rc != 0:
+        return []
+    rows = _parse_pairs(out)
+    seen = set()
+    children = []
+    for row in rows:
+        if row.get("TYPE") == "disk" or row["NAME"] in seen:
+            continue
+        seen.add(row["NAME"])
+        children.append(row)
+    return children
+
+
+def release_disk(disk, run=subprocess.run):
+    """Free `disk` (e.g. "/dev/sda") of anything that would block a
+    partition-table reload. Returns the log lines describing what happened."""
+    if not disk.startswith("/dev/"):
+        disk = f"/dev/{disk}"
+    lines = [f"Releasing {disk} before install…"]
+
+    children = _children(disk, run)
+    if not children:
+        lines.append("  nothing stacked on the disk — good.")
+    else:
+        lines.append("  found: " + ", ".join(f"{c['NAME']} ({c['TYPE']})" for c in children))
+
+    # Leaves first: mountpoints and swap sit on the outermost devices.
+    for child in reversed(children):
+        dev = f"/dev/{child['NAME']}"
+        mounts = child.get("MOUNTPOINTS", "")
+        if "[SWAP]" in mounts:
+            rc, out = _run(["swapoff", dev], run)
+            lines.append(f"  swapoff {dev}: {'ok' if rc == 0 else out}")
+        elif mounts:
+            rc, out = _run(["umount", "--all-targets", dev], run)
+            lines.append(f"  umount {dev}: {'ok' if rc == 0 else out}")
+
+    # Device-mapper (LVM, thin pools, LUKS) and md stacks. Thin pools need
+    # their thin LVs gone before the pool goes, and lsblk lists shared nodes
+    # more than once, so go leaves-first and retry a couple of passes until
+    # nothing dm/md-shaped is left.
+    for _ in range(3):
+        stacked = [c for c in reversed(_children(disk, run)) if c["TYPE"] != "part"]
+        if not stacked:
+            break
+        for child in stacked:
+            name, typ = child["NAME"], child["TYPE"]
+            if typ.startswith("raid") or name.startswith("md"):
+                cmd = ["mdadm", "--stop", f"/dev/{name}"]
+            else:
+                cmd = ["dmsetup", "remove", name]
+            rc, out = _run(cmd, run)
+            lines.append(f"  {' '.join(cmd)}: {'ok' if rc == 0 else out}")
+
+    leftovers = [c["NAME"] for c in _children(disk, run) if c["TYPE"] != "part"]
+    if leftovers:
+        lines.append(
+            "  WARNING: still in use after cleanup: "
+            + ", ".join(leftovers)
+            + " — the install may fail to re-read the partition table."
+        )
+
+    # Old signatures (LVM PV headers, RAID superblocks) would otherwise be
+    # re-activated by udev the moment the kernel re-reads the table.
+    for child in reversed(_children(disk, run)):
+        if child["TYPE"] == "part":
+            rc, out = _run(["wipefs", "-a", f"/dev/{child['NAME']}"], run)
+            lines.append(f"  wipefs /dev/{child['NAME']}: {'ok' if rc == 0 else out}")
+    rc, out = _run(["wipefs", "-a", disk], run)
+    lines.append(f"  wipefs {disk}: {'ok' if rc == 0 else out}")
+    rc, out = _run(["blockdev", "--rereadpt", disk], run)
+    lines.append(f"  re-read partition table: {'ok' if rc == 0 else out}")
+    return lines