Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 6 additions & 9 deletions docs/backends.md
Original file line number Diff line number Diff line change
Expand Up @@ -52,15 +52,12 @@ often around 10 GiB, which is why the backend prints how to grow it.

**Not every operator is root.** An Incus host is routinely driven by an
unprivileged user. That is why `community_scripts_dir()` exists at all, and why
the header cache, logs and defaults all go through it. It also means
`kernel.keys.maxkeys` cannot be written from an unprivileged user namespace, so
that sysctl is set only for privileged containers, and only when the host value
is actually below the target.

**Optional settings must not brick the container.** A rejected `linux.sysctl.*`
key stops the instance from starting, so they are applied softly and rolled back
on failure — and the rollback confirms the container really is running via
`incus list -c s`, rather than trusting the exit status of a start.
the header cache, logs and defaults all go through it.

**No per-container `kernel.keys.maxkeys`.** lxc applies `linux.sysctl.*` with
`/proc/sys` already read-only, so that key stopped privileged and unprivileged
containers alike from starting. A failed restart is still recovered by a forced
stop and start, confirmed via `incus list -c s` rather than the exit status.

**No pty.** The install runs through `incus exec` *without* `-t`. Allocating one
looks harmless and is not: apt's dpkg progress does a `tcsetattr` on stdin, gets
Expand Down
52 changes: 8 additions & 44 deletions incus/backend.func
Original file line number Diff line number Diff line change
Expand Up @@ -84,19 +84,14 @@ _incus_apply_addresses() {
return "$changed"
}

# Restart the container, tolerating optional boot-time configs (e.g.
# linux.sysctl.kernel.keys.maxkeys) that some hosts/kernels reject for the
# container's privilege level. Such configs pass 'config set' but abort the
# container at *start* time, so a plain restart would leave it stopped. On
# failure, roll those configs back and retry so the container is never left
# dead. Returns non-zero only if it still cannot start.
# Restart the container; when that fails, force a stop and start so it is not
# left dead. Returns non-zero only if it still cannot start.
_incus_restart_ct() {
local logf="${LOGFILE:-$INCUS_BUILD_LOG}"
if incus restart "${CT_NAME}" >>"$logf" 2>&1; then
return 0
fi
msg_warn "Container failed to start with current config — rolling back optional sysctls"
_incus_soft incus config unset "${CT_NAME}" linux.sysctl.kernel.keys.maxkeys >>"$logf" 2>&1
msg_warn "Container restart failed - forcing a stop and start"
_incus_soft incus stop -f "${CT_NAME}" >>"$logf" 2>&1
_incus_soft incus start "${CT_NAME}" >>"$logf" 2>&1
# A failed restart leaves incus disagreeing with itself - the same recovery
Expand Down Expand Up @@ -513,26 +508,6 @@ incus_create_lxc_container() {
INCUS_BUILD_OK=0
export INCUS_CT_CREATED INCUS_ORPHAN_HANDLED INCUS_BUILD_OK CT_NAME CTID

# Optional boot-time configs that some hosts/kernels reject for this
# container type. They fail when the container *starts*, not at 'config set'
# (e.g. kernel.keys.maxkeys is not writable from an unprivileged user
# namespace), so a plain set can silently persist a config that aborts the
# next boot. Apply best-effort and verify with a self-healing restart below.
# kernel.keys.maxkeys cannot be written from an unprivileged user namespace -
# lxc aborts the start with "Permission denied - Failed to setup sysctl
# parameters kernel.keys.maxkeys to 2000". This was set for CT_TYPE=1, which
# IS the unprivileged case, so every unprivileged create paid a failed start
# plus a rollback to reach the state it would have had anyway. Privileged
# only, and only when the host default is actually below what we want.
local optional_sysctl_pending=0
# The sysctl probe reads this machine; against a remote, just try the setting.
if [[ "${CT_TYPE:-1}" == "0" ]] &&
{ [[ "${INCUS_REMOTE:-local}" != "local" ]] || (($(cat /proc/sys/kernel/keys/maxkeys 2>/dev/null || echo 2000) < 2000)); }; then
if incus config set "${CT_NAME}" linux.sysctl.kernel.keys.maxkeys=2000 >>"$LOGFILE" 2>&1; then
optional_sysctl_pending=1
fi
fi

msg_ok "Container ${CT_NAME} created"

msg_info "Waiting for container to initialize"
Expand Down Expand Up @@ -582,21 +557,10 @@ incus_create_lxc_container() {
_incus_soft incus config set "${CT_NAME}" "user.tags=${TAGS}" >>"$LOGFILE" 2>&1
fi

# Restart when network devices changed, or to verify an optional boot-time
# sysctl we set above (so a rejected one is rolled back now rather than
# bricking a later reboot with boot.autostart=true).
if [[ "$net_reconfigure" -eq 1 || "$optional_sysctl_pending" -eq 1 ]]; then
if [[ "$net_reconfigure" -eq 1 ]]; then
msg_info "Reconfiguring network settings"
else
msg_info "Verifying container configuration"
fi
if [[ "$net_reconfigure" -eq 1 ]]; then
msg_info "Reconfiguring network settings"
if _incus_restart_ct; then
if [[ "$net_reconfigure" -eq 1 ]]; then
msg_ok "Network settings applied"
else
msg_ok "Container configuration verified"
fi
msg_ok "Network settings applied"
else
# Everything past this point needs incus exec. Warning and carrying on
# over a stopped container is what turned one failed restart into three
Expand Down Expand Up @@ -1041,7 +1005,7 @@ http://dl-cdn.alpinelinux.org/alpine/${alpine_rel}/community
EOF"
if ! incus_ct_exec /bin/sh -c 'getent hosts dl-cdn.alpinelinux.org >/dev/null 2>&1'; then
msg_warn "Alpine repository DNS resolution failed in container, injecting public DNS servers"
incus_ct_exec /bin/sh -c 'printf "nameserver 8.8.8.8\nnameserver 1.1.1.1\n" >/etc/resolv.conf'
incus_ct_exec /bin/sh -c 'printf "nameserver 8.8.8.8\nnameserver 1.1.1.1\n" >/etc/resolv.conf' || true
fi
incus_ct_exec ash -c "apk add bash newt curl openssh nano mc ncurses jq >/dev/null" || msg_warn "Alpine base package install had issues"
elif [[ "${var_os:-debian}" =~ ^(debian|ubuntu|devuan)$ ]]; then
Expand All @@ -1062,7 +1026,7 @@ EOF"

if ! incus_ct_exec bash -c "${_ct_env} getent hosts deb.debian.org >/dev/null 2>&1 && getent hosts archive.ubuntu.com >/dev/null 2>&1"; then
msg_warn "APT repository DNS resolution failed in container, injecting public DNS servers"
incus_ct_exec bash -c "echo -e 'nameserver 8.8.8.8\nnameserver 1.1.1.1' >/etc/resolv.conf"
incus_ct_exec bash -c "echo -e 'nameserver 8.8.8.8\nnameserver 1.1.1.1' >/etc/resolv.conf" || true
fi

incus_ct_exec bash -c "${_ct_env} apt-get update >/dev/null 2>&1 && apt-get install -y ${_base_pkgs} >/dev/null 2>&1" || {
Expand Down
Loading