x86/mm/cpa: Avoid wbinvd() for PREEMPT

Although wbinvd() is faster than flushing many individual pages, it blocks
the memory bus for "long" periods of time (>100us), thus directly causing
unusually large latencies on all CPUs, regardless of any CPU isolation
features that may be active. This is an unpriviledged operatation as it is
exposed to user space via the graphics subsystem.

For 1024 pages, flushing those pages individually can take up to 2200us,
but the task remains fully preemptible during that time.
Signed-off-by: John Ogness <john.ogness@linutronix.de>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Sebastian Siewior <bigeasy@linutronix.de>
Cc: linux-rt-users <linux-rt-users@vger.kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

arch/x86/mm/pageattr.c

@@ -214,7 +214,20 @@ static void cpa_flush_array(unsigned long *start, int numpages, int cache,
 			    int in_flags, struct page **pages)
 {
 	unsigned int i, level;
+#ifdef CONFIG_PREEMPT
+	/*
+	 * Avoid wbinvd() because it causes latencies on all CPUs,
+	 * regardless of any CPU isolation that may be in effect.
+	 *
+	 * This should be extended for CAT enabled systems independent of
+	 * PREEMPT because wbinvd() does not respect the CAT partitions and
+	 * this is exposed to unpriviledged users through the graphics
+	 * subsystem.
+	 */
+	unsigned long do_wbinvd = 0;
+#else
 	unsigned long do_wbinvd = cache && numpages >= 1024; /* 4M threshold */
+#endif
 
 	BUG_ON(irqs_disabled());