Re: [RFC PATCH 02/18] qemu-storage-daemon: Add --object option

[Markus Armbruster]

Kevin Wolf <address@hidden> writes:

> Add a command line option to create user-creatable QOM objects.
>
> Signed-off-by: Kevin Wolf <address@hidden>
> ---
>  qemu-storage-daemon.c | 35 +++++++++++++++++++++++++++++++++++
>  1 file changed, 35 insertions(+)
>
> diff --git a/qemu-storage-daemon.c b/qemu-storage-daemon.c
> index a251dc255c..48d6af43a6 100644
> --- a/qemu-storage-daemon.c
> +++ b/qemu-storage-daemon.c
> @@ -35,6 +35,8 @@
>  #include "qemu/log.h"
>  #include "qemu/main-loop.h"
>  #include "qemu/module.h"
> +#include "qemu/option.h"
> +#include "qom/object_interfaces.h"
>  
>  #include "trace/control.h"
>  
> @@ -51,10 +53,26 @@ static void help(void)
>  "                         specify tracing options\n"
>  "  -V, --version          output version information and exit\n"
>  "\n"
> +"  --object <properties>  define a QOM object such as 'secret' for\n"
> +"                         passwords and/or encryption keys\n"

This is less helpful than qemu-system-FOO's help:

-object TYPENAME[,PROP1=VALUE1,...]
                create a new object of type TYPENAME setting properties
                in the order they are specified.  Note that the 'id'
                property must be set.  These objects are placed in the
                '/objects' path.

> +"\n"
>  QEMU_HELP_BOTTOM "\n",
>      error_get_progname());
>  }
>  
> +enum {
> +    OPTION_OBJECT = 256,
> +};
> +
> +static QemuOptsList qemu_object_opts = {
> +    .name = "object",
> +    .implied_opt_name = "qom-type",
> +    .head = QTAILQ_HEAD_INITIALIZER(qemu_object_opts.head),
> +    .desc = {
> +        { }
> +    },
> +};
> +

Note for later: copied from vl.c.

>  static int process_options(int argc, char *argv[], Error **errp)
>  {
>      int c;
> @@ -63,6 +81,7 @@ static int process_options(int argc, char *argv[], Error 
> **errp)
>  
>      static const struct option long_options[] = {
>          {"help", no_argument, 0, 'h'},
> +        {"object", required_argument, 0, OPTION_OBJECT},
>          {"version", no_argument, 0, 'V'},
>          {"trace", required_argument, NULL, 'T'},
>          {0, 0, 0, 0}
> @@ -88,6 +107,22 @@ static int process_options(int argc, char *argv[], Error 
> **errp)
>              g_free(trace_file);
>              trace_file = trace_opt_parse(optarg);
>              break;
> +        case OPTION_OBJECT:
> +            {
> +                QemuOpts *opts;
> +                const char *type;
> +
> +                opts = qemu_opts_parse(&qemu_object_opts,
> +                                       optarg, true, &error_fatal);
> +                type = qemu_opt_get(opts, "qom-type");
> +
> +                if (user_creatable_print_help(type, opts)) {
> +                    exit(EXIT_SUCCESS);
> +                }
> +                user_creatable_add_opts(opts, &error_fatal);
> +                qemu_opts_del(opts);
> +                break;
> +            }
>          }
>      }
>      if (optind != argc) {

PATCH 01 duplicates case QEMU_OPTION_trace pretty much verbatim.  Makes
sense, as qemu-storage-daemon is basically qemu-system-FOO with "FOO"
and most "system" cut away.

This patch adds vl.c's case QEMU_OPTION_object in a much simpler form.
This is one of my least favourite options, and I'll tell you why below.
Let's compare the two versions.

vl.c:

            case QEMU_OPTION_object:
                opts = qemu_opts_parse_noisily(qemu_find_opts("object"),
                                               optarg, true);
                if (!opts) {
                    exit(1);
                }
                break;

Further down:

    qemu_opts_foreach(qemu_find_opts("object"),
                      user_creatable_add_opts_foreach,
                      object_create_initial, &error_fatal);

Still further down:

    qemu_opts_foreach(qemu_find_opts("object"),
                      user_creatable_add_opts_foreach,
                      object_create_delayed, &error_fatal);

These are basically

    for opts in qemu_object_opts {
        type = qemu_opt_get(opts, "qom-type");
        if (type) {
            if (user_creatable_print_help(type, opts)) {
                exit(0);
            }
            if (!predicate(type)) {
                continue;
            }
        }
        obj = user_creatable_add_opts(opts, &error_fatal);
        object_unref(obj);
    }

where predicate(type) is true in exactly one of the two places for each
QOM type.

The reason for these gymnastics is to create objects at the right time
during startup, except there is no right time, but two.

Differences:

* Options are processed left to right without gymnastics.  Getting their
  order right is the user's problem.  I consider this an improvement.

* You use &qemu_object_opts instead of qemu_find_opts("object").  Also
  an improvement.

* You use qemu_opts_parse() instead of qemu_opts_parse_noisily().
  The latter can print help.  I failed to find a case where we lose help
  compared to qemu-system-FOO.  I didn't try very hard.

* You neglect to guard user_creatable_print_help():

    $ qemu-storage-daemon --object wrong=1,help
    Segmentation fault (core dumped)

* You neglect to object_unref().  I just double-checked the final
  reference count: it's 2.

These bugs shouldn't be hard to fix.


At this point you might wonder why I dislike this option so much.
vl.c's gymnastics are ugly, but not unusually ugly, and they're gone
here.  To explain my distaste, I have to go back a little bit.

Like quite a few options, --object is paired with QMP command, namely
object-add.  Both have the same parameters: QOM type, object ID, and
additional type-specific object properties.  There's a difference,
though: object-add wraps the latter in a 'props' object, while --object
does not.

QAPI schema:

    { 'command': 'object-add',
      'data': {'qom-type': 'str', 'id': 'str', '*props': 'any'} }

QAPIfying this part of the CLI isn't easy.

The obvious QAPIfied CLI buddy of object-add is incompatible to current
--object.  That's not a concern for the storage daemon.  But it's also
ugly, because object-add's nesting of the type-specific properties
within @props is.  In QMP, it's merely yet another pair of curlies.  In
the CLI, we get to prefix props. to each type-specific property.

If we want to give the storage daemon a QAPIfied command line from the
start (and I think we do), we'll have to decide how to address this
issue, and possibly more (I'm only at PATCH 02/18).

We have a long history of rather careless interface design, and now some
of these chickens come home to roost.