we experienced false errors in the first MB in test #7 with SMP. There is simple way to reproduce it:

  1. Turn on SMP (F2)
  2. Select test #7 (c 1 3 7 <enter>)
  3. Limit address to 1 MB (2 2 1m <enter> 0)

But the problem only appears on some systems. Fortunately, I was able to find a bug and fix it.

Memtest86+ contains USB workaround (USB_WAR). This workaround should prohibit writing to the lowest 1280 Bytes of the memmory by changing the start address from 0 to 1280 (0x500). This checking is done only by one (master) core just before each test starts. This change is not fully propagated to other cores due to a race-condition.

This can be easily fixed by adding memmory barrier just after this workaround code (patch at the end of the post).

I am not sure, if this also fixes other (already reported) bugs with SMP and test #7. I was not able to reproduce them.

Interresting fact: The chances of false detection are much higher with precompiled memtest86+5.01 downloaded form the official site. If I compile it with newer gcc (or use compiled version from Debian repository), the race-condition appears less often.

--- a/main.c
+++ b/main.c
@@ -852,6 +852,7 @@
         cprint(LINE_RANGE, COL_MID+30, " of ");
         aprint(LINE_RANGE, COL_MID+34, v->selected_pages);
+    s_barrier();
     switch(tseq[test].pat) {