fix(oci): a shared host directory shows what the host has mounted inside it

oci/remote/install_generic_stack.py

@@ -25,6 +25,7 @@ import xml.etree.ElementTree as ET
 
 from allocate_private_network import allocate_network, existing_bridges, existing_networks
 import oci_instances
+import oci_nested_mounts
 from oci_ui import translate, msg_info, msg_ok, msg_warn, msg_error, msg_info2, stop_spinner, log
 
 HERE = Path(__file__).resolve().parent
@@ -446,6 +447,7 @@ def attach_mounts(service, temporary):
             current = next(line.split(': ', 1)[1] for line in config.splitlines()
                            if line.startswith(f'mp{index}: '))
             run('pct', 'set', vmid, f'--mp{index}', current + ',ro=1')
+    oci_nested_mounts.follow(vmid)
 
 
 def wait_web(primary, url):

oci/remote/install_oci.sh

@@ -2024,6 +2024,12 @@ CREDENTIALS=$(jq -c --argjson environment "$DEPLOYMENT_ENVIRONMENT" '
 ' "$TEMPLATE_FILE")
 RUNTIME_CREDENTIALS=$(jq '[.[] | select(.retrieval.method? == "container-console-pattern")] | length' <<<"$CREDENTIALS")
 
+# A host directory is mounted with what the host has mounted inside it. Written
+# once every mount point is set and after the host monitor profile, which
+# clears the mount hooks.
+python3 "${SCRIPT_DIR}/oci_nested_mounts.py" follow "$VMID" \
+  || die "$(translate "Could not add the mount point:") CT $VMID"
+
 # The console of the container is kept as its log, the way `docker logs` keeps
 # it, and the Proxmox console opens a shell when the image has one. Both are
 # set before the configuration is recorded, so the record carries them and an

oci/remote/oci_instance_transaction.py

@@ -30,6 +30,7 @@ import oci_instances as instances
 from oci_installation_state import image_from_archive, parse_config, private_directory, sha
 from verify_oci_archive import verify_archive
 import oci_host_mounts as host_mounts
+import oci_nested_mounts as nested_mounts
 import oci_accelerators as gpu_devices
 import oci_runtime_settings as runtime_settings
 import oci_image_cache as image_cache
@@ -473,7 +474,10 @@ def preflight(record, candidate, config, coordinated=None):
             raise ValueError(translate('The imported OCI user or group is not numeric'))
     if 'lxc.init.groups' in cfg and not re.fullmatch(r'(?:[0-9]+(?:[ ,][0-9]+)*)?', cfg['lxc.init.groups']):
         raise ValueError(translate('The imported OCI groups are not numeric'))
-    keys = {line.split(': ', 1)[0] for line in config.decode().splitlines() if ': ' in line}
+    # What shows the content mounted inside a host directory follows the mount
+    # points: it is checked against them and is not a setting of its own.
+    nested_mounts.check(config)
+    keys = {line.split(': ', 1)[0] for line in nested_mounts.without(config).decode().splitlines() if ': ' in line}
     runtime_keys = {'lxc.mount.entry'} if gpu_devices.nvidia.enabled(deployment) else set()
     if gpu_devices.dynamic_mode(deployment):
         runtime_keys = {'lxc.hook.mount', 'lxc.environment'}
@@ -597,6 +601,7 @@ def freeze_host_sources(record, candidate, acknowledge_external_data, files=()):
 
 def check_runtime_mounts(config, deployment):
     runtime_settings.check(config, deployment, deployment['vmid'])
+    nested_mounts.check(config)
     actual = mounts(config)
     files = runtime_settings.file_binds(config)
     if files != {m['source']: bool(m.get('read_only')) for m in deployment.get('mounts', []) if is_file_bind(m, files)}:

oci/remote/oci_native_stack.py

@@ -13,6 +13,7 @@ import time
 
 import oci_console
 import oci_instances as instances
+import oci_nested_mounts
 from oci_installation_state import command, image_from_archive, sha
 import oci_stack_replay
 from oci_ui import translate
@@ -84,6 +85,7 @@ def create(root, args):
         # console as a shell, set before the stack records its configuration
         # so an update rebuilds them the same way.
         oci_console.configure(vmid)
+        oci_nested_mounts.follow(vmid)
     return code
 
 

oci/remote/oci_nested_mounts.py

@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+"""Show in a container what the host has mounted inside a shared directory.
+
+Proxmox mounts the directory of a mount point alone: a filesystem mounted
+inside it, as a network share under /mnt/pve, is left out and the container
+finds an empty folder in its place. Next to each of those mount points goes a
+raw LXC entry that mounts the same directory with everything inside it, and
+keeps following what the host mounts and unmounts while the container runs.
+"""
+import argparse
+import os
+from pathlib import Path
+import re
+
+from oci_installation_state import parse_config
+from oci_ui import translate
+
+ENTRY = re.compile(r'lxc\.mount\.entry: (/\S*) (\S+) none rbind,rslave(,ro)?,create=dir 0 0')
+# Read only reaches the directory alone; this takes it to what is inside.
+HOOK = re.compile(r"""lxc\.hook\.mount: /bin/sh -c 'mount -o remount,bind,ro=recursive "\$LXC_ROOTFS_MOUNT/(\S+)"'""")
+READ_ONLY = "lxc.hook.mount: /bin/sh -c 'mount -o remount,bind,ro=recursive \"$LXC_ROOTFS_MOUNT/{}\"'"
+# Trees of the kernel, mounted as the recipe of a host monitor lays them out.
+SYSTEM_TREES = ('/proc', '/sys', '/dev', '/run')
+# Characters with a meaning of their own in an LXC entry or in the hook.
+UNSAFE = re.compile(r'''[\s\\'"$`]''')
+
+
+def own(line):
+    """Whether a configuration line is one of those written here."""
+    return bool(ENTRY.fullmatch(line) or HOOK.fullmatch(line))
+
+
+def groups(config):
+    """The lines of each host directory among the mount points, a directory
+    before the ones mounted inside it."""
+    found = {}
+    for key, value in parse_config(config).items():
+        if not re.fullmatch(r'mp[0-9]+', key) or not value.startswith('/'):
+            continue
+        source, *rest = value.split(',')
+        options = dict(item.split('=', 1) for item in rest if '=' in item)
+        target = options.get('mp', '').strip('/')
+        if (not target or UNSAFE.search(source) or UNSAFE.search(target) or source == '/'
+                or any(source == tree or source.startswith(tree + '/') for tree in SYSTEM_TREES)):
+            continue
+        if options.get('ro') == '1':
+            found[target] = [f'lxc.mount.entry: {source} {target} none rbind,rslave,ro,create=dir 0 0',
+                             READ_ONLY.format(target)]
+        else:
+            found[target] = [f'lxc.mount.entry: {source} {target} none rbind,rslave,create=dir 0 0']
+    return [found[target] for target in sorted(found)]
+
+
+def without(config):
+    """The configuration without the lines written here."""
+    return b''.join(line for line in config.splitlines(keepends=True) if not own(line.decode().strip()))
+
+
+def check(config):
+    """Every line of this kind belongs to a mount point of the container, and
+    none of a mount point is missing while another of it is there."""
+    present = [line for line in config.decode().splitlines() if own(line)]
+    expected = [line for group in groups(config) if any(line in present for line in group) for line in group]
+    if sorted(present) != sorted(expected):
+        raise ValueError(translate('A container mount has a source, backup or permission different from the saved record'))
+
+
+def follow(vmid):
+    """Write the lines of the mount points the container has now, and no
+    others. Any snapshot section is left as it is."""
+    conf = Path(f'/etc/pve/lxc/{int(vmid)}.conf')
+    text = conf.read_text()
+    current, _, snapshots = text.partition('\n[')
+    kept = [line for line in current.splitlines() if not own(line)]
+    wanted = [line for group in groups(current.encode()) for line in group]
+    rebuilt = '\n'.join(kept + wanted) + '\n'
+    if snapshots:
+        rebuilt += '\n[' + snapshots
+    if rebuilt != text:
+        conf.write_text(rebuilt)
+    return wanted
+
+
+def main(argv=None):
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('action', choices=('follow',))
+    parser.add_argument('vmid', type=int)
+    args = parser.parse_args(argv)
+    if os.geteuid() != 0:
+        parser.error(translate('Root privileges are required'))
+    follow(args.vmid)
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())

oci/remote/oci_restore_recovery.py

@@ -33,6 +33,7 @@ import oci_carried_record as carried
 import oci_console
 import oci_instances as instances
 import oci_instance_transaction as transaction
+import oci_nested_mounts
 import oci_nvidia_refresh
 import oci_nvidia_runtime as nvidia
 import oci_runtime_settings as runtime_settings
@@ -460,7 +461,7 @@ def check_host_resources(plan):
                     f"{label}: {translate('it uses an NVIDIA GPU and this host has no working NVIDIA driver and Container Toolkit. Install them and run the recovery again.')}")
         for line in lines:
             key, _, value = line.partition(': ')
-            if key == 'lxc.hook.mount' and value:
+            if key == 'lxc.hook.mount' and value and not oci_nested_mounts.own(line):
                 hook = carried.NVIDIA_HOOK.fullmatch(value)
                 if not hook:
                     plan['blockers'].append(f"{label}: {translate('unknown mount hook:')} {value}")
@@ -737,7 +738,8 @@ def restore_host_files(entry):
                 runtime_settings.restore(deployment, vmid)
             else:
                 path.write_text(runtime_settings.sysctl_content(deployment))
-        elif key == 'lxc.hook.mount' and value and not Path(value).is_file():
+        elif (key == 'lxc.hook.mount' and value and not oci_nested_mounts.own(line)
+              and not Path(value).is_file()):
             content = nvidia_hook_source(entry, carried.NVIDIA_HOOK.fullmatch(value)[1])
             Path(value).parent.mkdir(parents=True, exist_ok=True, mode=0o755)
             Path(value).write_text(content)

oci/remote/oci_runtime_settings.py

@@ -5,6 +5,7 @@ import re
 import stat
 import tempfile
 
+import oci_nested_mounts as nested_mounts
 from oci_ui import translate
 
 
@@ -125,6 +126,7 @@ def file_binds(config):
 def filter_config(config, deployment):
     """The configuration without the entries of the mounts the deployment
     declares itself: what is left belongs to the acceleration profile."""
+    config = nested_mounts.without(config)
     declared = set(tmpfs_lines(deployment))
     files = {m['source'] for m in deployment.get('mounts', []) if m.get('type') == 'host-bind'}
 

oci/remote/oci_stack_modify.py

@@ -16,6 +16,7 @@ import oci_gpu_devices as gpu_devices
 import oci_host_mounts as host_mounts
 import oci_instance_reconcile as reconcile
 import oci_instances as instances
+import oci_nested_mounts
 import oci_stack_replay as replay
 from oci_ui import msg_error, msg_info, msg_ok, translate
 
@@ -129,6 +130,8 @@ def apply(vmid, changes):
             if unused:
                 run('pct', 'set', str(vmid), '--delete', unused)
         msg_ok(f"{translate('Path removed:')} {path}")
+    if changes['add_mounts'] or changes['remove_mounts']:
+        oci_nested_mounts.follow(vmid)
 
 
 def recorded_mounts(vmid):

oci/remote/oci_stack_replay.py

@@ -9,6 +9,7 @@ import os
 
 import oci_console
 import oci_gpu_devices
+import oci_nested_mounts
 from oci_ui import translate
 
 
@@ -617,12 +618,13 @@ def normalize(record):
     preserved = {key: single(key) for key in ('arch', 'ostype', 'cmode', 'console', 'tty', 'cpuunits',
                  'net0', 'net1', 'startup', 'hookscript', 'features', 'tags') if key in values}
     devices = [{'key': key, 'value': single(key)} for key in values if re.fullmatch(r'dev[0-9]+', key)]
-    # The console log line and its start hook are not replayed: the installer
-    # that rebuilds the member sets them itself, and replaying them too would
-    # leave two of each.
+    # The console log line, its start hook and what shows the content mounted
+    # inside a host directory are not replayed: the installer that rebuilds the
+    # member sets them itself, and replaying them too would leave two of each.
     raw_runtime = [line for line in config.splitlines() if line.startswith('lxc.')
                    and line.partition(': ')[0] not in ('lxc.environment.runtime', 'lxc.init.cwd',
                                                        'lxc.signal.halt', 'lxc.console.logfile')
+                   and not oci_nested_mounts.own(line)
                    and not oci_console.is_start_mark_hook(line)]
     return {'schema_version': 1, 'deployment': plan, 'runtime': runtime,
             'preserved_native': preserved, 'generated_files': copy.deepcopy(files),

oci/remote/oci_ui.sh

@@ -190,6 +190,8 @@ oci_apply_extra_mounts() {
     count=$((count + 1))
   done < <(jq -r '.extra_mounts[]? | [.type, .container_path, .source, (.size_gb // "-"), (.read_only // false)] | @tsv' "$DEPLOYMENT_FILE")
   if (( count > 0 )); then
+    python3 "${SCRIPT_DIR}/oci_nested_mounts.py" follow "$vmid" \
+      || die "$(translate "Could not add the mount point:") CT $vmid"
     msg_ok "$(translate "Mount points added:") $count"
   fi
   return 0