Re: [Qemu-devel] [PATCH for-2.11 v3 01/25] qom: cpu: fix parsed feature string length

[Igor Mammedov]

On Thu, 24 Aug 2017 14:00:21 -0300
Philippe Mathieu-Daudé <address@hidden> wrote:

> Hi Igor,
> 
> On 08/24/2017 01:31 PM, Igor Mammedov wrote:
> > since commit ( 9262685b cpu: Factor out cpu_generic_init() )
> > features parsed by it were truncated only to the 1st feature
> > after CPU name due to fact that
> > 
> >     featurestr = strtok(NULL, ",");
> >     cc->parse_features(cpu, featurestr, &err);
> > 
> > would extract exactly one feature and parse_features() callback
> > would parse it and only it leaving the rest of features ignored.
> > 
> > Reuse approach from x86 custom impl. i.e. replace strtok() token
> > parsing with g_strsplit(), which would split feature string in
> > 2 parts name and features list and pass the later to
> > parse_features() callback.
> > 
> > Signed-off-by: Igor Mammedov <address@hidden>
> > ---
> > CC: Eduardo Habkost <address@hidden>
> > CC: Andreas Färber <address@hidden>
> > 
> > Probably due to existing users not actualy using/having any
> > features to parse bug were unnoticed for 2 years but switching
> > from custom cpu_foo_init() to cpu_generic_init() triggered it.
> > ---
> >   qom/cpu.c | 14 ++++++--------
> >   1 file changed, 6 insertions(+), 8 deletions(-)
> > 
> > diff --git a/qom/cpu.c b/qom/cpu.c
> > index 4f38db0..caf5c14 100644
> > --- a/qom/cpu.c
> > +++ b/qom/cpu.c
> > @@ -50,28 +50,26 @@ bool cpu_exists(int64_t id)
> >   
> >   CPUState *cpu_generic_init(const char *typename, const char *cpu_model)
> >   {
> > -    char *str, *name, *featurestr;
> >       CPUState *cpu = NULL;
> >       ObjectClass *oc;
> >       CPUClass *cc;
> >       Error *err = NULL;
> > +    gchar **model_pieces;
> >   
> > -    str = g_strdup(cpu_model);
> > -    name = strtok(str, ",");
> > +    model_pieces = g_strsplit(cpu_model, ",", 2);
> >   
> > -    oc = cpu_class_by_name(typename, name);
> > +    oc = cpu_class_by_name(typename, model_pieces[0]);
> >       if (oc == NULL) {
> > -        g_free(str);
> > +        g_strfreev(model_pieces);
> >           return NULL;
> >       }
> >   
> >       cc = CPU_CLASS(oc);
> > -    featurestr = strtok(NULL, ",");
> >       /* TODO: all callers of cpu_generic_init() need to be converted to
> >        * call parse_features() only once, before calling cpu_generic_init().
> >        */
> > -    cc->parse_features(object_class_get_name(oc), featurestr, &err);
> > -    g_free(str);  
> 
> I feel safer adding:
> 
>         if (g_strv_length(model_pieces) > 1) {
all current callbacks (i386/sparc/cpu_common_parse_features) deal with

  cc->parse_features(,NULL,)

by explicitly checking for NULL, so there is no need for check here.

Though, it's possible to avoid calling callback at all
if model_pieces[1] == NULL

Would you like to post a patch on top?

> 
> > +    cc->parse_features(object_class_get_name(oc), model_pieces[1], &err);  
> 
>         }
> 
> > +    g_strfreev(model_pieces);
> >       if (err != NULL) {
> >           goto out;
> >       }
> >   
>