[pve-devel] [PATCH v2 container 1/1] Add device passthrough
Filip Schauer
f.schauer at proxmox.com
Thu Nov 2 15:28:22 CET 2023
On 30/10/2023 14:34, Wolfgang Bumiller wrote:
> On Tue, Oct 24, 2023 at 02:55:53PM +0200, Filip Schauer wrote:
>> Add a dev[n] argument to the container config to pass devices through to
>> a container. A device can be passed by its path. Alternatively a mapped
>> USB device can be passed through with usbmapping=<name>.
>>
>> Signed-off-by: Filip Schauer<f.schauer at proxmox.com>
>> ---
>> src/PVE/LXC.pm | 34 +++++++++++++++++++++++-
>> src/PVE/LXC/Config.pm | 60 +++++++++++++++++++++++++++++++++++++++++++
>> 2 files changed, 93 insertions(+), 1 deletion(-)
>>
>> diff --git a/src/PVE/LXC.pm b/src/PVE/LXC.pm
>> index c9b5ba7..a3ddb62 100644
>> --- a/src/PVE/LXC.pm
>> +++ b/src/PVE/LXC.pm
>> @@ -5,7 +5,8 @@ use warnings;
>>
>> use Cwd qw();
>> use Errno qw(ELOOP ENOTDIR EROFS ECONNREFUSED EEXIST);
>> -use Fcntl qw(O_RDONLY O_WRONLY O_NOFOLLOW O_DIRECTORY);
>> +use Fcntl qw(O_RDONLY O_WRONLY O_NOFOLLOW O_DIRECTORY :mode);
>> +use File::Basename;
>> use File::Path;
>> use File::Spec;
>> use IO::Poll qw(POLLIN POLLHUP);
>> @@ -639,6 +640,37 @@ sub update_lxc_config {
>> $raw .= "lxc.mount.auto = sys:mixed\n";
>> }
>>
>> + # Clear passthrough directory from previous run
>> + my $passthrough_dir = "/var/lib/lxc/$vmid/passthrough";
>> + File::Path::rmtree($passthrough_dir);
> I think we need to make a few changes here.
>
> First: we don't necessarily need this directory.
> Having a device list would certainly be nice, but it makes more sense to
> just have a file we can easily parse (possibly even just a json hash),
> like the `devices` file we already create in the pre-start hook, except
> prepared *for* the pre-start hook, which *should* be able to just
> `mknod` the devices right into the container's `/dev` on startup.
Devices mknoded into the container's /dev directory in the pre-start
hook will not be visible in the container once it is fully started.
Meanwhile mknoding a device to a different path inside the container
works fine. It seems that LXC mounts over the /dev directory. This can
be solved by calling mknod in lxc-pve-autodev-hook, but this does not
work with unprivileged containers without the mknod capability.
So are bind mounts our only option without modifying LXC,
or am I overlooking something?
> We'd also avoid "lingering" device nodes with potentially harmful
> uid/permissions in /var, which is certainly better from a security POV.
>
> But note that we do need the `lxc.cgroup2.*` entries before starting the
> container in order to ensure the devices cgroup has the right
> permissions.
>
>> +
>> + PVE::LXC::Config->foreach_passthrough_device($conf, sub {
>> + my ($key, $sanitized_path) = @_;
>> +
>> + my $absolute_path = "/$sanitized_path";
>> + my ($mode, $rdev) = (stat($absolute_path))[2, 6];
>> + die "Could not find major and minor ids of device $absolute_path.\n"
>> + unless ($mode && $rdev);
>> +
>> + my $major = PVE::Tools::dev_t_major($rdev);
>> + my $minor = PVE::Tools::dev_t_minor($rdev);
>> + my $device_type_char = S_ISBLK($mode) ? 'b' : 'c';
>> + my $passthrough_device_path = "$passthrough_dir/$sanitized_path";
>> + File::Path::make_path(dirname($passthrough_device_path));
>> + PVE::Tools::run_command([
>> + '/usr/bin/mknod',
>> + '-m', '0660',
>> + $passthrough_device_path,
>> + $device_type_char,
>> + $major,
>> + $minor
>> + ]);
> It's probably worth adding a helper for the mknod syscall to
> `PVE::Tools`, there are a bunch of syscalls in there already.
>
>> + chown 100000, 100000, $passthrough_device_path if ($unprivileged);
> ^ This isn't necessarily the correct id. Users may have custom id
> mappings.
> `PVE::LXC::parse_id_maps($conf)` returns the mapping alongside the root
> uid and gid. (See for example `sub mount_all` for how it's used.
>
>> +
>> + $raw .= "lxc.cgroup2.devices.allow = $device_type_char $major:$minor rw\n";
>> + $raw .= "lxc.mount.entry = $passthrough_device_path $sanitized_path none bind,create=file\n";
>> + });
>> +
>> # WARNING: DO NOT REMOVE this without making sure that loop device nodes
>> # cannot be exposed to the container with r/w access (cgroup perms).
>> # When this is enabled mounts will still remain in the monitor's namespace
>> diff --git a/src/PVE/LXC/Config.pm b/src/PVE/LXC/Config.pm
>> index 56e1f10..edd813e 100644
>> --- a/src/PVE/LXC/Config.pm
>> +++ b/src/PVE/LXC/Config.pm
>> @@ -29,6 +29,7 @@ mkdir $lockdir;
>> mkdir "/etc/pve/nodes/$nodename/lxc";
>> my $MAX_MOUNT_POINTS = 256;
>> my $MAX_UNUSED_DISKS = $MAX_MOUNT_POINTS;
>> +my $MAX_DEVICES = 256;
>>
>> # BEGIN implemented abstract methods from PVE::AbstractConfig
>>
>> @@ -908,6 +909,49 @@ for (my $i = 0; $i < $MAX_UNUSED_DISKS; $i++) {
>> }
>> }
>>
>> +PVE::JSONSchema::register_format('pve-lxc-dev-string', \&verify_lxc_dev_string);
>> +sub verify_lxc_dev_string {
>> + my ($dev, $noerr) = @_;
>> +
>> + if (
>> + $dev =~m@/\.\.?/@ ||
>> + $dev =~m@/\.\.?$@ ||
>> + $dev !~ m!^/dev/!
>> + ) {
>> + return undef if $noerr;
>> + die "$dev is not a valid device path\n";
>> + }
>> +
>> + return $dev;
>> +}
>> +
>> +my $dev_desc = {
>> + path => {
>> + optional => 1,
>> + type => 'string',
>> + default_key => 1,
>> + format => 'pve-lxc-dev-string',
>> + format_description => 'Path',
>> + description => 'Device to pass through to the container',
>> + verbose_description => 'Path to the device to pass through to the container'
>> + },
>> + usbmapping => {
>> + optional => 1,
>> + type => 'string',
>> + format => 'pve-configid',
>> + format_description => 'mapping-id',
>> + description => 'The ID of a cluster wide USB mapping.'
>> + }
>> +};
>> +
>> +for (my $i = 0; $i < $MAX_DEVICES; $i++) {
>> + $confdesc->{"dev$i"} = {
>> + optional => 1,
>> + type => 'string', format => $dev_desc,
>> + description => "Device to pass through to the container",
>> + }
>> +}
>> +
>> sub parse_pct_config {
>> my ($filename, $raw, $strict) = @_;
>>
>> @@ -1255,6 +1299,22 @@ sub parse_volume {
>> return;
>> }
>>
>> +sub parse_device {
>> + my ($class, $device_string, $noerr) = @_;
>> +
>> + my $res;
>> + eval { $res = PVE::JSONSchema::parse_property_string($dev_desc, $device_string) };
>> + if ($@) {
>> + return undef if $noerr;
>> + die $@;
>> + }
>> +
>> + die "Either path or usbmapping has to be defined"
>> + unless (defined($res->{path}) || defined($res->{usbmapping}));
>> +
>> + return $res;
>> +}
>> +
>> sub print_volume {
>> my ($class, $key, $volume) = @_;
>>
>> --
>> 2.39.2
>>
>>
>>
>> _______________________________________________
>> pve-devel mailing list
>> pve-devel at lists.proxmox.com
>> https://lists.proxmox.com/cgi-bin/mailman/listinfo/pve-devel
>>
>>
More information about the pve-devel
mailing list