Re: 2.6.12-rc5-mm1

[Mikael Pettersson <mikpe@csd.uu.se>]

2.6.12-rc5-mm1 includes Andrea's seccomp-disable-tsc patch,
which I believe is broken on SMP. In process.c we find:

 /*
+ * This function selects if the context switch from prev to next
+ * has to tweak the TSC disable bit in the cr4.
+ */
+static void disable_tsc(struct thread_info *prev,
+			struct thread_info *next)
+{
+	if (unlikely(has_secure_computing(prev) ||
+		     has_secure_computing(next))) {
+		/* slow path here */
+		if (has_secure_computing(prev) &&
+		    !has_secure_computing(next)) {
+			clear_in_cr4(X86_CR4_TSD);
+		} else if (!has_secure_computing(prev) &&
+			   has_secure_computing(next))
+			set_in_cr4(X86_CR4_TSD);
+	}
+}

which it calls from __switch_to().

The problem is that {set,clear}_in_cr4() both update a single
global mmu_cr4_features variable, which is asynchronously written
to all CPUs by {,__}flush_tlb_all(). Hence, the CR4.TSD setting
is at best probabilistic.

I spotted this because perfctr used to flip CR4.PCE in __switch_to()
ages ago, but I had to abandon that when kernel 2.3.40 changed to
the current scheme with a global mmu_cr4_features.
(Another reason was that CR4 writes were and still are very slow.)

/Mikael