Re: [PATCHv5 08/19] x86/mm: Introduce variables to store number, shift and mask of KeyIDs

On Thu, 19 Jul 2018, Kirill A. Shutemov wrote:

On Thu, Jul 19, 2018 at 03:18:03PM +0200, Thomas Gleixner wrote:

quoted

On Thu, 19 Jul 2018, Kirill A. Shutemov wrote:

quoted

On Thu, Jul 19, 2018 at 02:37:35PM +0200, Thomas Gleixner wrote:

quoted

On Thu, 19 Jul 2018, Kirill A. Shutemov wrote:

quoted

On Wed, Jul 18, 2018 at 04:19:10PM -0700, Dave Hansen wrote:

quoted

quoted

 	} else {
 		/*
 		 * Reset __PHYSICAL_MASK.

@@ -591,6 +592,9 @@ static void detect_tme(struct cpuinfo_x86 *c)
 		 * between CPUs.
 		 */
 		physical_mask = (1ULL << __PHYSICAL_MASK_SHIFT) - 1;
+		mktme_keyid_mask = 0;
+		mktme_keyid_shift = 0;
+		mktme_nr_keyids = 0;
 	}

Should be unnecessary.  These are zeroed by the compiler.

No. detect_tme() called for each CPU in the system.

And then the variables are cleared out while other CPUs can access them?
How is that supposed to work?

This code path only matter in patalogical case: when MKTME configuation is
inconsitent between CPUs. Basically if BIOS screwed things up we disable
MKTME.

I still don't see how that's supposed to work.

When the inconsistent CPU is brought up _AFTER_ MKTME is enabled, then how
does clearing the variables help? It does not magically make all the other
stuff go away.

We don't actually enable MKTME in kernel. BIOS does. Kernel makes choose
to use it or not. Current design targeted to be used by userspace.
So until init we don't have any other stuff to go away. We can just
pretend that MKTME was never there.

Hotplug is not guaranteed to happen _BEFORE_ init. Think about physical
hotplug.

Thanks,

	tglx