Ensure pfn isn't already in use before assigning to addr[i].  Retry if
in use.  The subsequent pfns were checked against the bitmap, but not
the initial selection.

This prevents some false positives on a machine with a small amount of
RAM where pfn collisions are more likely to be seen.

A Xen HVM first showed:
[ 3640.227939] CPA ffff88800869c000: bad pte 800000000869c163
[ 3640.227982] CPA ffff88800104e000: bad pte 800000000104e161

Testing this patch with a Xen PV guest with one 192MB of RAM showed
varying numbers of addresses like:
[ 2768.082971] CPA test 5 pfn 930d in use

Signed-off-by: Jason Andryuk <jandryuk@gmail.com>
---
This could infinite loop on a machine with a small amount of RAM - is
that worth handling?  Just skip an in-use pfn instead?
---
 arch/x86/mm/pat/cpa-test.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/arch/x86/mm/pat/cpa-test.c b/arch/x86/mm/pat/cpa-test.c
index ad3c1feec990..bbf337e3e246 100644
--- a/arch/x86/mm/pat/cpa-test.c
+++ b/arch/x86/mm/pat/cpa-test.c
@@ -136,7 +136,14 @@ static int pageattr_test(void)
 	failed += print_split(&sa);
 
 	for (i = 0; i < NTEST; i++) {
-		unsigned long pfn = get_random_u32_below(max_pfn_mapped);
+		unsigned long pfn;
+
+ retry:
+		pfn = get_random_u32_below(max_pfn_mapped);
+		if (test_bit(pfn, bm)) {
+			pr_debug("CPA test %d pfn %lx in use\n", i, pfn);
+			goto retry;
+		}
 
 		addr[i] = (unsigned long)__va(pfn << PAGE_SHIFT);
 		len[i] = get_random_u32_below(NPAGES);
-- 
2.43.0