Re: [PATCH v7 2/5] softmmu/vl: Let -fw_cfg option take a 'gen_id' argument

[Laszlo Ersek]

Gerd, Corey: there's a question for you near the end, please.

On 05/28/20 19:31, Philippe Mathieu-Daudé wrote:
> The 'gen_id' argument refers to a QOM object able to produce
> data consumable by the fw_cfg device. The producer object must
> implement the FW_CFG_DATA_GENERATOR interface.
>
> Signed-off-by: Philippe Mathieu-Daudé <philmd@redhat.com>
> ---
> v7:
> - renamed 'blob_id' -> 'gen_id' (danpb)
> - update comment in code (lersek)
> - fixed CODING_STYLE (lersek)
> - use Laszlo's if (SUM(options)) != 1 { error } form
> ---
>  softmmu/vl.c | 22 ++++++++++++++--------
>  1 file changed, 14 insertions(+), 8 deletions(-)
>
> diff --git a/softmmu/vl.c b/softmmu/vl.c
> index ae5451bc23..cdb1d187ed 100644
> --- a/softmmu/vl.c
> +++ b/softmmu/vl.c
> @@ -489,6 +489,11 @@ static QemuOptsList qemu_fw_cfg_opts = {
>              .name = "string",
>              .type = QEMU_OPT_STRING,
>              .help = "Sets content of the blob to be inserted from a string",
> +        }, {
> +            .name = "gen_id",
> +            .type = QEMU_OPT_STRING,
> +            .help = "Sets id of the object generating the fw_cfg blob "
> +                    "to be inserted",
>          },
>          { /* end of list */ }
>      },
> @@ -2020,7 +2025,7 @@ static int parse_fw_cfg(void *opaque, QemuOpts *opts, 
> Error **errp)
>  {
>      gchar *buf;
>      size_t size;
> -    const char *name, *file, *str;
> +    const char *name, *file, *str, *gen_id;
>      FWCfgState *fw_cfg = (FWCfgState *) opaque;
>
>      if (fw_cfg == NULL) {
> @@ -2030,14 +2035,13 @@ static int parse_fw_cfg(void *opaque, QemuOpts *opts, 
> Error **errp)
>      name = qemu_opt_get(opts, "name");
>      file = qemu_opt_get(opts, "file");
>      str = qemu_opt_get(opts, "string");
> +    gen_id = qemu_opt_get(opts, "gen_id");
>
> -    /* we need name and either a file or the content string */
> -    if (!(nonempty_str(name) && (nonempty_str(file) || nonempty_str(str)))) {
> -        error_setg(errp, "invalid argument(s)");
> -        return -1;
> -    }
> -    if (nonempty_str(file) && nonempty_str(str)) {
> -        error_setg(errp, "file and string are mutually exclusive");
> +    /* we need the name, and exactly one of: file, content string, gen_id */
> +    if (!nonempty_str(name) ||
> +          nonempty_str(file) + nonempty_str(str) + nonempty_str(gen_id) != 
> 1) {

(1) I believe the indentation of this line is not correct. I think it
should be out-dented by 2 spaces.

> +        error_setg(errp, "name, plus exactly one of file,"
> +                         " string and gen_id, are needed");
>          return -1;
>      }
>      if (strlen(name) > FW_CFG_MAX_FILE_PATH - 1) {
> @@ -2052,6 +2056,8 @@ static int parse_fw_cfg(void *opaque, QemuOpts *opts, 
> Error **errp)
>      if (nonempty_str(str)) {
>          size = strlen(str); /* NUL terminator NOT included in fw_cfg blob */
>          buf = g_memdup(str, size);
> +    } else if (nonempty_str(gen_id)) {
> +        return fw_cfg_add_from_generator(fw_cfg, name, gen_id, errp);

(2) This is no longer correct: fw_cfg_add_from_generator() now returns 0
on failure, but parse_fw_cfg() is supposed to return nonzer on failure.
See the comment on qemu_opts_foreach() -- "parse_fw_cfg" is passed as
the loop callback to qemu_opts_foreach().

Technically, we could simply write

        return !fw_cfg_add_from_generator(fw_cfg, name, gen_id, errp);

but that wouldn't be consistent with the -1 error codes returned
elsewhere from parse_fw_cfg(). So how about:

        size_t fw_cfg_size;

        fw_cfg_size = fw_cfg_add_from_generator(fw_cfg, name, gen_id, errp);
        return (fw_cfg_size > 0) ? 0 : -1;

I think your testing may have missed this because the problem is only
visible if you have *another* -fw_cfg option on the QEMU command line.
Returning the wrong status code from here terminates the
qemu_opts_foreach() loop, without attempting to set "error_fatal".
Therefore the loop is silently terminated, thus the only symptom would
be that -fw_cfg options beyond the "gen_id" one wouldn't take effect.


(3) I've noticed another *potential* issue, from looking at the larger
context. I apologize for missing it in v6.

See commit bab47d9a75a3 ("Sort the fw_cfg file list", 2016-04-07). (I'm
copying Corey; Gerd is already copied.) From that commit, we have, at
the end of this function:

    /* For legacy, keep user files in a specific global order. */
    fw_cfg_set_order_override(fw_cfg, FW_CFG_ORDER_OVERRIDE_USER);
    fw_cfg_add_file(fw_cfg, name, buf, size);
    fw_cfg_reset_order_override(fw_cfg);

This takes effect for "file" and "string", but not for "gen_id". Should
we apply it to "gen_id" as well? (Sorry, I really don't understand what
commit bab47d9a75a3 is about!)

*IF* we want to apply the same logic to "gen_id", then we should
*perhaps* do, on the "nonempty_str(gen_id)" branch:

        size_t fw_cfg_size;

        fw_cfg_set_order_override(fw_cfg, FW_CFG_ORDER_OVERRIDE_USER);
        fw_cfg_size = fw_cfg_add_from_generator(fw_cfg, name, gen_id, errp);
        fw_cfg_reset_order_override(fw_cfg);
        return (fw_cfg_size > 0) ? 0 : -1;

I think???

Or maybe even use FW_CFG_ORDER_OVERRIDE_DEVICE rather than
FW_CFG_ORDER_OVERRIDE_USER? I don't have the slightest clue.

(I guess if I understood what commit bab47d9a75a3 was about, I'd be less
in doubt now. But that commit only hints at "avoid[ing] any future
issues of moving the file creation" -- I don't know what those issues
were in the first place!)

With (1) optionally fixed, and (2) fixed, I'd be willing to R-b this
patch; but I'm really thrown off by (3).

Thanks,
Laszlo


>      } else {
>          GError *err = NULL;
>          if (!g_file_get_contents(file, &buf, &size, &err)) {
>