Re: [PATCH 3/3] lto-plugin: implement LDPT_GET_API_VERSION

[Rui Ueyama <rui314@gmail.com>]

I updated my code so that it now acquires a lock before calling
claim_file if v0.
https://github.com/rui314/mold/commit/54fedf005fb35acdcb0408395aa403f94262190a

On Mon, Jul 11, 2022 at 8:51 PM Martin Liška <mliska@suse.cz> wrote:
>
> On 7/11/22 14:24, Rui Ueyama wrote:
> > I updated my patch to support the proposed API:
> > https://github.com/rui314/mold/commit/22bbfa9bba9beeaf40b76481d175939ee2c62ec8
> >
> > Martin,
> >
> > I think you want to apply this patch. Currently, your API always
> > passes LAPI_V0 as the maximum API version.
>
> Are you sure?
>
> The function signature is:
>
> (*ld_plugin_get_api_version) (const char *plugin_identifier,
>                              const char *plugin_version,
>                              enum linker_api_version minimal_api_supported,
>                              enum linker_api_version maximal_api_supported,
> ...
>
> Which means the plug-in always set minimal as LAPI_V0 and maximal
> LAPI_V0/V1. That seems correct to me.

I'm sorry I misunderstood your code. It looks correct.

> Martin
>
>
> >
> > diff --git a/lto-plugin/lto-plugin.c b/lto-plugin/lto-plugin.c
> > index e9afd2fb76d..c97bda9de91 100644
> > --- a/lto-plugin/lto-plugin.c
> > +++ b/lto-plugin/lto-plugin.c
> > @@ -1441,15 +1441,15 @@ negotiate_api_version (void)
> >    const char *linker_version;
> >
> >    enum linker_api_version supported_api = LAPI_V0;
> >  #if HAVE_PTHREAD_LOCKING
> >    supported_api = LAPI_V1;
> >  #endif
> >
> > -  api_version = get_api_version ("GCC", BASE_VERSION, LAPI_V0,
> > +  api_version = get_api_version ("GCC", BASE_VERSION, LAPI_V1,
> >                                  supported_api, &linker_identifier,
> > &linker_version);
> >    if (api_version > supported_api)
> >      {
> >        fprintf (stderr, "requested an unsupported API version (%d)\n",
> > api_version);
> >        abort ();
> >      }
> >
> > On Mon, Jul 11, 2022 at 6:51 PM Martin Liška <mliska@suse.cz> wrote:
> >>
> >> On 7/11/22 11:55, Richard Biener wrote:
> >>> On Mon, Jul 11, 2022 at 11:16 AM Alexander Monakov <amonakov@ispras.ru> wrote:
> >>>>
> >>>> On Mon, 11 Jul 2022, Rui Ueyama wrote:
> >>>>
> >>>>>> but ignoring min_api_supported is wrong, and assuming max_api_supported > 0
> >>>>>> is also wrong. It really should check how given [min; max] range intersects
> >>>>>> with its own range of supported versions.
> >>>>>
> >>>>> Currently only one version is defined which is LAPI_V1. I don't think
> >>>>> LAPI_UNSPECIFIED is a version number; rather, it's an unspecified
> >>>>> value. No ordering should be defined between a defined value and an
> >>>>> unspecified value. If LAPI_UNSPECIFIED < LAPI_V1, it should be renamed
> >>>>> LAPI_V0.
> >>>>
> >>>> You still cannot rely on API guarantees of LAPI_V1 when the plugin does not
> >>>> advertise it (thread safety of claim_file in this particular case).
> >>>
> >>
> >> Hi.
> >>
> >> All right, I think we should rename LAPI_UNSPECIFIED to LAPI_V0 in order
> >> to support minimal_api_supported == LAPI_V0.
> >>
> >>> So with LAPI_UNSPECIFIED all the plugin gets is the linker name and version.
> >>> Clarifying the documentation on LAPI_UNSPECIFIED might be nice, also
> >>> what the expectation is on the linker when the plugin returns
> >>> LAPI_UNSPECIFIED when it speficied minimal_api_supported == V1.
> >>
> >> I've clarified that linker should return a value that is in range
> >> [minimal_api_supported, maximal_api_supported] and added an abort
> >> if it's not the case.
> >>
> >> Having that, mold should respect if maximal_api_supported == LAPI_V0 is returned
> >> by a plug-in (happens now as we miss locking for some targets).
> >>
> >> Martin
> >>
> >>> "minimal_api_supported == LAPI_UNSPECIFIED" does not make much
> >>> sense if using Ruis reading of this value?
> >>>
> >>>> And you still should check the intersection of supported API ranges
> >>>> and give a sane diagnostic when min_api_supported advertised by the plugin
> >>>> exceeds LAPI_V1 (though, granted, the plugin could error out as well in this
> >>>> case).
> >>>>
> >>>> Alexander
>