Re: [PATCH] bitmap: Fix optimization of bitmap_set/clear for big-endian machines

Re: [PATCH] bitmap: Fix optimization of bitmap_set/clear for big-endian machines

[Paul Mackerras]

On Wed, Oct 25, 2017 at 07:39:48AM +0000, Matthew Wilcox wrote:
> Hang on, don't tell me you found this by inspection.  Are you not running the bitmap testcase, enabled by CONFIG_TEST_BITMAP?  Either that should be producing an error, or there's a missing test case, or your inspection is wrong ...

I did find it by inspection.  I was looking for a version of the
bitmap_* API that does little-endian style bitmaps on all systems, and
the inline bitmap_set() does that in the case where it calls memset,
but not in the case where it calls __bitmap_set.

I'll fire up a big-endian system tomorrow when I get to work to run
the test case.  (PPC64 is almost entirely little-endian these days as
far as the IBM POWER systems are concerned.)

In any case, it's pretty clearly wrong as it is.  On a big-endian
64-bit system, bitmap_set(p, 56, 16) should set bytes 0 and 15 to
0xff, and there's no way a single memset can do that.

Paul.

(and yes, I stuffed up the address for lkml)

RE: [PATCH] bitmap: Fix optimization of bitmap_set/clear for big-endian machines

[Matthew Wilcox]

From: Paul Mackerras [mailto:paulus@ozlabs.org]
> On Wed, Oct 25, 2017 at 07:39:48AM +0000, Matthew Wilcox wrote:
> > Hang on, don't tell me you found this by inspection.  Are you not running the
> bitmap testcase, enabled by CONFIG_TEST_BITMAP?  Either that should be
> producing an error, or there's a missing test case, or your inspection is wrong ...
> 
> I did find it by inspection.  I was looking for a version of the
> bitmap_* API that does little-endian style bitmaps on all systems, and
> the inline bitmap_set() does that in the case where it calls memset,
> but not in the case where it calls __bitmap_set.

I do believe that you noticed it by inspection, but you shouldn't've had to.  I thought we had a comprehensive set of tests for exactly this, which means that either 01.org isn't running the right set of tests on a BE system or the tests are broken.

> I'll fire up a big-endian system tomorrow when I get to work to run
> the test case.  (PPC64 is almost entirely little-endian these days as
> far as the IBM POWER systems are concerned.)
> 
> In any case, it's pretty clearly wrong as it is.  On a big-endian
> 64-bit system, bitmap_set(p, 56, 16) should set bytes 0 and 15 to
> 0xff, and there's no way a single memset can do that.

So ... this loop should include that case, right?

        for (start = 0; start < 1024; start += 8) {
                memset(bmap1, 0x5a, sizeof(bmap1));
                memset(bmap2, 0x5a, sizeof(bmap2));
                for (nbits = 0; nbits < 1024 - start; nbits += 8) {
                        bitmap_set(bmap1, start, nbits);
                        __bitmap_set(bmap2, start, nbits);
                        if (!bitmap_equal(bmap1, bmap2, 1024))
                                printk("set not equal %d %d\n", start, nbits);
                        if (!__bitmap_equal(bmap1, bmap2, 1024))
                                printk("set not __equal %d %d\n", start, nbits);

Re: [PATCH] bitmap: Fix optimization of bitmap_set/clear for big-endian machines

[Michael Ellerman]

Paul Mackerras <paulus@ozlabs.org> writes:

> On Wed, Oct 25, 2017 at 07:39:48AM +0000, Matthew Wilcox wrote:
>> Hang on, don't tell me you found this by inspection. Are you not
>> running the bitmap testcase, enabled by CONFIG_TEST_BITMAP? Either
>> that should be producing an error, or there's a missing test case, or
>> your inspection is wrong ...

Oops.

  [    1.907533] test_bitmap: test 13: input is '0-2047:128/256' OK, Time: 1359
  [    1.907596] set not equal 0 8
  [    1.907640] set not __equal 0 8
  [    1.907684] clear not equal 0 8
  [    1.907728] clear not __equal 0 8
  [    1.907772] set not equal 0 16
  [    1.907816] set not __equal 0 16
  [    1.907861] clear not equal 0 16
  [    1.907905] clear not __equal 0 16
  [    1.907949] set not equal 0 24
  [    1.907993] set not __equal 0 24
  [    1.908038] clear not equal 0 24
  [    1.908082] clear not __equal 0 24
  ...
  [ snip ~30,000 lines ]
  ...
  [    3.345729] set not equal 1008 8
  [    3.345773] set not __equal 1008 8
  [    3.345818] clear not equal 1008 8
  [    3.345863] clear not __equal 1008 8

  [    3.345909] test_bitmap: all 460506 tests passed
                              ^^^^^^^^^^^^^^^^^^^^^^^
                              O_o

With the patch:

  [    1.904393] test_bitmap: test 13: input is '0-2047:128/256' OK, Time: 1359
  [    1.916270] test_bitmap: all 460506 tests passed

cheers

Re: [PATCH] bitmap: Fix optimization of bitmap_set/clear for big-endian machines

[Michael Ellerman]

Matthew Wilcox <willy@infradead.org> writes:

> (I don't think I can reliably send patches from outlook; sorry for
> breaking the threading)
>
> I see where we're not incrementing the failure count ... try this patch!
>
> --- 8< ---
>
> Subject: Fix bitmap optimisation tests to report errors correctly
> From: Matthew Wilcox <mawilcox@microsoft.com>
>
> I had neglected to increment the error counter when the tests failed,
> which made the tests noisy when they fail, but not actually return an
> error code.
>
> Reported-by: Michael Ellerman <mpe@ellerman.id.au>
> Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com>
> Cc: stable@kernel.org

Thanks, that works for me.

  test_bitmap: failed 31840 out of 460506 tests

cheers

Re: [PATCH] bitmap: Fix optimization of bitmap_set/clear for big-endian machines

[Matthew Wilcox]

(I don't think I can reliably send patches from outlook; sorry for
breaking the threading)

I see where we're not incrementing the failure count ... try this patch!

--- 8< ---

Subject: Fix bitmap optimisation tests to report errors correctly
From: Matthew Wilcox <mawilcox@microsoft.com>

I had neglected to increment the error counter when the tests failed,
which made the tests noisy when they fail, but not actually return an
error code.

Reported-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com>
Cc: stable@kernel.org

diff --git a/lib/test_bitmap.c b/lib/test_bitmap.c
index aa1f2669bdd5..ae8a830e4e54 100644
--- a/lib/test_bitmap.c
+++ b/lib/test_bitmap.c
@@ -430,23 +430,32 @@ static void noinline __init test_mem_optimisations(void)
 	unsigned int start, nbits;
 
 	for (start = 0; start < 1024; start += 8) {
-		memset(bmap1, 0x5a, sizeof(bmap1));
-		memset(bmap2, 0x5a, sizeof(bmap2));
 		for (nbits = 0; nbits < 1024 - start; nbits += 8) {
+			memset(bmap1, 0x5a, sizeof(bmap1));
+			memset(bmap2, 0x5a, sizeof(bmap2));
+
 			bitmap_set(bmap1, start, nbits);
 			__bitmap_set(bmap2, start, nbits);
-			if (!bitmap_equal(bmap1, bmap2, 1024))
+			if (!bitmap_equal(bmap1, bmap2, 1024)) {
 				printk("set not equal %d %d\n", start, nbits);
-			if (!__bitmap_equal(bmap1, bmap2, 1024))
+				failed_tests++;
+			}
+			if (!__bitmap_equal(bmap1, bmap2, 1024)) {
 				printk("set not __equal %d %d\n", start, nbits);
+				failed_tests++;
+			}
 
 			bitmap_clear(bmap1, start, nbits);
 			__bitmap_clear(bmap2, start, nbits);
-			if (!bitmap_equal(bmap1, bmap2, 1024))
+			if (!bitmap_equal(bmap1, bmap2, 1024)) {
 				printk("clear not equal %d %d\n", start, nbits);
-			if (!__bitmap_equal(bmap1, bmap2, 1024))
+				failed_tests++;
+			}
+			if (!__bitmap_equal(bmap1, bmap2, 1024)) {
 				printk("clear not __equal %d %d\n", start,
 									nbits);
+				failed_tests++;
+			}
 		}
 	}
 }