On Fri, Nov 11, 2022 at 05:23:48PM +0100, Borislav Petkov wrote:
> On Mon, Nov 07, 2022 at 02:53:18PM -0800, Jithu Joseph wrote:
> > static int scan_chunks_sanity_check(struct device *dev)
> > {
> > - int metadata_size, curr_pkg, cpu, ret = -ENOMEM;
> > struct ifs_data *ifsd = ifs_get_data(dev);
> > + int curr_pkg, cpu, ret = -ENOMEM;
> > bool *package_authenticated;
> > struct ifs_work local_work;
> > - char *test_ptr;
> >
> > package_authenticated = kcalloc(topology_max_packages(), sizeof(bool), GFP_KERNEL);
> > if (!package_authenticated)
> > return ret;
>
> Bah, how big is that thing so that you can't simply do a bitfield on the
> stack here instead of kcalloc-ing?
Kernel is built with gcc options that prevent a variable sized local
array. So:
DECLARE_BITMAP(auth, topology_max_packages());
grumbles:
drivers/platform/x86/intel/ifs/load.c:196:2: warning: ISO C90 forbids variable length array ‘ifs_auth2’ [-Wvla]
We could pick a likely big enough static number:
#define MAX_SUPPORTED_PACKAGES 128
DECLARE_BITMAP(auth, MAX_SUPPORTED_PACKAGES);
and error out of this code if SGI or someone build a monster machine:
if (topology_max_packages() > MAX_SUPPORTED_PACKAGES) {
pr_error_once("IFS driver needs update to support this machine\n");
return -E2BIG;
}
That avoids the kcalloc() and making sure to kfree() in all error paths.
But seems a bit hacky. Other ideas?
-Tony
