x86: Fix serialization in pit_expect_msb()
authorLinus Torvalds <torvalds@linux-foundation.org>
Fri, 31 Jul 2009 19:45:41 +0000 (12:45 -0700)
committerIngo Molnar <mingo@elte.hu>
Mon, 10 Aug 2009 17:56:57 +0000 (19:56 +0200)
Wei Chong Tan reported a fast-PIT-calibration corner-case:

| pit_expect_msb() is vulnerable to SMI disturbance corner case
| in some platforms which causes /proc/cpuinfo to show wrong
| CPU MHz value when quick_pit_calibrate() jumps to success
| section.

I think that the real issue isn't even an SMI - but the fact
that in the very last iteration of the loop, there's no
serializing instruction _after_ the last 'rdtsc'. So even in
the absense of SMI's, we do have a situation where the cycle
counter was read without proper serialization.

The last check should be done outside the outer loop, since
_inside_ the outer loop, we'll be testing that the PIT has
the right MSB value has the right value in the next iteration.

So only the _last_ iteration is special, because that's the one
that will not check the PIT MSB value any more, and because the
final 'get_cycles()' isn't serialized.

In other words:

 - I'd like to move the PIT MSB check to after the last
   iteration, rather than in every iteration

 - I think we should comment on the fact that it's also a
   serializing instruction and so 'fences in' the TSC read.

Here's a suggested replacement.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Reported-by: "Tan, Wei Chong" <wei.chong.tan@intel.com>
Tested-by: "Tan, Wei Chong" <wei.chong.tan@intel.com>
LKML-Reference: <B28277FD4E0F9247A3D55704C440A140D5D683F3@pgsmsx504.gar.corp.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/x86/kernel/tsc.c

index 6e1a368..71f4368 100644 (file)
@@ -275,15 +275,20 @@ static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
  * use the TSC value at the transitions to calculate a pretty
  * good value for the TSC frequencty.
  */
+static inline int pit_verify_msb(unsigned char val)
+{
+       /* Ignore LSB */
+       inb(0x42);
+       return inb(0x42) == val;
+}
+
 static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
 {
        int count;
        u64 tsc = 0;
 
        for (count = 0; count < 50000; count++) {
-               /* Ignore LSB */
-               inb(0x42);
-               if (inb(0x42) != val)
+               if (!pit_verify_msb(val))
                        break;
                tsc = get_cycles();
        }
@@ -336,8 +341,7 @@ static unsigned long quick_pit_calibrate(void)
         * to do that is to just read back the 16-bit counter
         * once from the PIT.
         */
-       inb(0x42);
-       inb(0x42);
+       pit_verify_msb(0);
 
        if (pit_expect_msb(0xff, &tsc, &d1)) {
                for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
@@ -348,8 +352,19 @@ static unsigned long quick_pit_calibrate(void)
                         * Iterate until the error is less than 500 ppm
                         */
                        delta -= tsc;
-                       if (d1+d2 < delta >> 11)
-                               goto success;
+                       if (d1+d2 >= delta >> 11)
+                               continue;
+
+                       /*
+                        * Check the PIT one more time to verify that
+                        * all TSC reads were stable wrt the PIT.
+                        *
+                        * This also guarantees serialization of the
+                        * last cycle read ('d2') in pit_expect_msb.
+                        */
+                       if (!pit_verify_msb(0xfe - i))
+                               break;
+                       goto success;
                }
        }
        printk("Fast TSC calibration failed\n");