Re: [PATCH v4 22/24] powerpc/pseries: Implement secvars for dynamic secure boot

[Andrew Donnellan <ajd@linux.ibm.com>]

On Tue, 2023-01-24 at 15:17 +1000, Nicholas Piggin wrote:
> > +static const char * const plpks_var_names[] = {
> > +       "PK",
> > +       "KEK",
> > +       "db",
> > +       "dbx",
> > +       "grubdb",
> > +       "grubdbx",
> > +       "sbat",
> > +       "moduledb",
> > +       "trustedcadb",
> > +       NULL,
> > +};
> 
> Var and key are used somewhat interchangeably? These are keys, I
> think?
> And plpks could have other vars but we're only interested in (at
> least a
> subset of) keys here if I understood right.
> 
> I guess the terminology is like that throughout secvar so maybe
> nothing
> to be done.

The "key" terminology seems to come from OPAL, while on the PLPKS side
it's a bit of a mess but "var" follows the usage in existing code (the
spec refers more to "objects").

> 
> > +
> > +static int plpks_get_variable(const char *key, u64 key_len, u8
> > *data,
> > +                             u64 *data_size)
> > +{
> > +       struct plpks_var var = {0};
> > +       int rc = 0;
> > +
> > +       var.name = kcalloc(key_len - 1, sizeof(wchar_t),
> > GFP_KERNEL);
> > +       if (!var.name)
> > +               return -ENOMEM;
> > +       rc = utf8s_to_utf16s(key, key_len - 1, UTF16_LITTLE_ENDIAN,
> > (wchar_t *)var.name,
> > +                            key_len - 1);
> > +       if (rc < 0)
> > +               goto err;
> 
> Okay I can't work out why it's key_len - 1 rather than key_len.

The existing code in secvar-sysfs.c calls secvar_ops->get() with
key_len = strlen(name) + 1, to include the null byte, which is what
OPAL expects. For PLPKS, the variable name explicitly does not include
a trailing null byte.

I'll add a comment indicating as such, perhaps at some later point it
can be reworked.

> 
> > +       var.namelen = rc * 2;
> > +
> > +       var.os = PLPKS_VAR_LINUX;
> > +       if (data) {
> > +               var.data = data;
> > +               var.datalen = *data_size;
> > +       }
> > +       rc = plpks_read_os_var(&var);
> > +
> > +       if (rc)
> > +               goto err;
> > +
> > +       *data_size = var.datalen;
> > +
> > +err:
> > +       kfree(var.name);
> > +       if (rc && rc != -ENOENT) {
> > +               pr_err("Failed to read variable '%s': %d\n", key,
> > rc);
> > +               // Return -EIO since userspace probably doesn't
> > care about the
> > +               // specific error
> > +               rc = -EIO;
> > +       }
> > +       return rc;
> > +}
> > +
> > +static int plpks_set_variable(const char *key, u64 key_len, u8
> > *data,
> > +                             u64 data_size)
> > +{
> > +       struct plpks_var var = {0};
> > +       int rc = 0;
> > +       u64 flags;
> > +
> > +       // Secure variables need to be prefixed with 8 bytes of
> > flags.
> > +       // We only want to perform the write if we have at least
> > one byte of data.
> > +       if (data_size <= sizeof(flags))
> 
> So it's unstructured 8 byte of flags, not a u64 integer? Why not u8
> flags[8] then?

No, it's a u64 and it's passed in the hcall as a single u64.

> 
> > +               return -EINVAL;
> > +
> > +       var.name = kcalloc(key_len - 1, sizeof(wchar_t),
> > GFP_KERNEL);
> > +       if (!var.name)
> > +               return -ENOMEM;
> > +       rc = utf8s_to_utf16s(key, key_len - 1, UTF16_LITTLE_ENDIAN,
> > (wchar_t *)var.name,
> > +                            key_len - 1);
> > +       if (rc < 0)
> > +               goto err;
> > +       var.namelen = rc * 2;
> > +
> > +       memcpy(&flags, data, sizeof(flags));
> > +
> > +       var.datalen = data_size - sizeof(flags);
> > +       var.data = data + sizeof(flags);
> > +       var.os = PLPKS_VAR_LINUX;
> > +       var.policy = get_policy(key);
> > +
> > +       // Unlike in the read case, the plpks error code can be
> > useful to
> > +       // userspace on write, so we return it rather than just -
> > EIO
> > +       rc = plpks_signed_update_var(&var, flags);
> > +
> > +err:
> > +       kfree(var.name);
> > +       return rc;
> > +}
> > +
> > +// PLPKS dynamic secure boot doesn't give us a format string in
> > the same way OPAL does.
> > +// Instead, report the format using the SB_VERSION variable in the
> > keystore.
> > +static ssize_t plpks_secvar_format(char *buf, size_t bufsize)
> > +{
> > +       struct plpks_var var = {0};
> > +       ssize_t ret;
> > +
> > +       var.component = NULL;
> > +       // Only the signed variables have null bytes in their
> > names, this one doesn't
> > +       var.name = "SB_VERSION";
> > +       var.namelen = 10;
> 
> Could you make that strlen(var.name) for the benefit of those of us
> with
> missing fingers?

Will do.

> 
> > +       var.datalen = 1;
> > +       var.data = kzalloc(1, GFP_KERNEL);
> 
> This could just point to a u8 on stack I think?

Until we get VMAP_STACK and we'll have to switch back.

> 
> 
> > +       if (!var.data)
> > +               return -ENOMEM;
> > +
> > +       // Unlike the other vars, SB_VERSION is owned by firmware
> > instead of the OS
> > +       ret = plpks_read_fw_var(&var);
> > +       if (ret) {
> > +               if (ret == -ENOENT) {
> > +                       ret = snprintf(buf, bufsize, "ibm,plpks-sb-
> > unknown");
> > +               } else {
> > +                       pr_err("Error %ld reading SB_VERSION from
> > firmware\n", ret);
> > +                       ret = -EIO;
> > +               }
> 
> Is there a meaningful distinction? Does anything good come of
> advertising an unknown format like this?

Our thinking was simply to distinguish between cases where the API is
otherwise working happily but for some reason simply not advertising a
version number (you might still want to try to interact with the key
store regardless) vs the case where the hypervisor is returning a real
error.

I plan to keep this as is for the next revision, but I'm happy to
change it if there's a strong objection, it could go either way.

> 
> > +               goto err;
> > +       }
> > +
> > +       // This string is made up by us - the hypervisor doesn't
> > provide us
> > +       // with a format string in the way that OPAL firmware does.
> > Hypervisor
> > +       // defines SB_VERSION as a "1 byte unsigned integer value".
> 
> I'd put the comment about SB_VERSION at the top where you use/define
> it
> or mention it in the comment.

Will fix.

> 
> > +       ret = snprintf(buf, bufsize, "ibm,plpks-sb-v%hhu",
> > var.data[0]);
> > +
> > +err:
> > +       kfree(var.data);
> > +       return ret;
> > +}
> > +
> > +static int plpks_max_size(u64 *max_size)
> > +{
> > +       // The max object size reported by the hypervisor is
> > accurate for the
> > +       // object itself, but we use the first 8 bytes of data on
> > write as the
> > +       // signed update flags, so the max size a user can write is
> > larger.
> > +       *max_size = (u64)plpks_get_maxobjectsize() + 8;
> 
> You have this 8 open coded twice (once as sizeof(u64)). You could
> make
> it a #define at the top with a brief overview of the hcall format so
> you
> don't need so much commentage for it. Although a note here that the
> objsize does not include the flags bytes is good to keep.

Will do.

-- 
Andrew Donnellan    OzLabs, ADL Canberra
ajd@linux.ibm.com   IBM Australia Limited