Re: [PATCH v4 13/15] x86/sev: Take advantage of configfs visibility support in TSM

From: Dan Williams
Date: Wed May 01 2024 - 16:16:10 EST


Kuppuswamy Sathyanarayanan wrote:
> On Wed, Apr 24, 2024 at 9:00 AM Tom Lendacky <thomas.lendacky@xxxxxxx> wrote:
> >
> > The TSM attestation report support provides multiple configfs attribute
> > types (both for standard and binary attributes) to allow for additional
> > attributes to be displayed for SNP as compared to TDX. With the ability
> > to hide attributes via configfs, consoldate the multiple attribute groups
> > into a single standard attribute group and a single binary attribute
> > group. Modify the TDX support to hide the attributes that were previously
> > "hidden" as a result of registering the selective attribute groups.
> >
> > Co-developed-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> > Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> > Signed-off-by: Tom Lendacky <thomas.lendacky@xxxxxxx>
[..]
> > + */
> > +enum tsm_attr_index {
> > + TSM_REPORT_GENERATION,
>
> Do we need an index for the generation attribute ? Since it is a core
> function, we can allow it by default.

That is up to the is_visible() callback to decide the declaration of
which index corresponds to which attribute is just static information.

>
> > + TSM_REPORT_PROVIDER,
>
> Same as above.

These numbers need to match the array indices of tsm_report_attrs.

Your suggestion makes the declaration of tsm_report_attrs more
difficult:

static struct configfs_attribute *tsm_report_attrs[] = {
[TSM_REPORT_GENERATION] = &tsm_report_attr_generation,
[TSM_REPORT_PROVIDER] = &tsm_report_attr_provider,
[TSM_REPORT_PRIVLEVEL] = &tsm_report_attr_privlevel,
[TSM_REPORT_PRIVLEVEL_FLOOR] = &tsm_report_attr_privlevel_floor,
NULL,
};

..because then the definition of TSM_REPORT_PRIVLEVEL would need to
know how many attributes precede it in the array. So, defining it this
way makes it more robust against future changes that want to
add/delete/reorder attributes in the array.

>
> > + TSM_REPORT_PRIVLEVEL,
> > + TSM_REPORT_PRIVLEVEL_FLOOR,
> > +};
> > +
> > +/**
> > + * enum tsm_bin_attr_index - index used to reference binary report attributes
> > + * @TSM_REPORT_INBLOB: index of the binary report input attribute
> > + * @TSM_REPORT_OUTBLOB: index of the binary report output attribute
> > + * @TSM_REPORT_AUXBLOB: index of the binary auxiliary data attribute
> > + */
> > +enum tsm_bin_attr_index {
> > + TSM_REPORT_INBLOB,
> > + TSM_REPORT_OUTBLOB,
> > + TSM_REPORT_AUXBLOB,
> > +};
>
> Why differentiate between bin attr and regular attr? Why not use the same enum?

I do not understand your suggestion. There are two arrays:

static struct configfs_bin_attribute *tsm_report_bin_attrs[] = {
[TSM_REPORT_INBLOB] = &tsm_report_attr_inblob,
[TSM_REPORT_OUTBLOB] = &tsm_report_attr_outblob,
[TSM_REPORT_AUXBLOB] = &tsm_report_attr_auxblob,
NULL,
};

..so there are 2 sets of indices. If only one enum is used then one of
those arrays becomes sparsely populated causing the NULL array
terminator to pop up in unexpected indices.

> > +
> > /**
> > * struct tsm_ops - attributes and operations for tsm instances
> > * @name: tsm id reflected in /sys/kernel/config/tsm/report/$report/provider
> > * @privlevel_floor: convey base privlevel for nested scenarios
> > * @report_new: Populate @report with the report blob and auxblob
> > * (optional), return 0 on successful population, or -errno otherwise
> > + * @report_attr_visible: show or hide a report attribute entry
> > + * @report_bin_attr_visible: show or hide a report binary attribute entry
> > *
> > * Implementation specific ops, only one is expected to be registered at
> > * a time i.e. only one of "sev-guest", "tdx-guest", etc.
> > @@ -56,14 +85,12 @@ struct tsm_ops {
> > const char *name;
> > unsigned int privlevel_floor;
> > int (*report_new)(struct tsm_report *report, void *data);
> > + bool (*report_attr_visible)(struct config_item *item,
> > + struct configfs_attribute *attr, int n);
> > + bool (*report_bin_attr_visible)(struct config_item *item,
> > + struct configfs_bin_attribute *attr, int n);
>
> I think we can use a single callback . We don't really gain anything
> with bin attr differentiation.

No, the goal here is symmetry with sysfs, and the arrays are separate so
the @n argument is a different index space. It also loses some type
safety for making sure that bin_attr callbacks can safely assume that
they were not passed a text attribute by mistake.